xref: /freebsd/sys/netinet/tcp_output.c (revision 603eaf792b659f91d7d1a065d82503966d1386fc)
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 < so->so_snd.sb_cc)
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(so->so_snd.sb_cc, sendwin) - off);
352 		else {
353 			long cwin;
354 
355                         /*
356 			 * We are inside of a SACK recovery episode and are
357 			 * sending new data, having retransmitted all the
358 			 * data possible in the scoreboard.
359 			 */
360 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
361 			       - off);
362 			/*
363 			 * Don't remove this (len > 0) check !
364 			 * We explicitly check for len > 0 here (although it
365 			 * isn't really necessary), to work around a gcc
366 			 * optimization issue - to force gcc to compute
367 			 * len above. Without this check, the computation
368 			 * of len is bungled by the optimizer.
369 			 */
370 			if (len > 0) {
371 				cwin = tp->snd_cwnd -
372 					(tp->snd_nxt - tp->sack_newdata) -
373 					sack_bytes_rxmt;
374 				if (cwin < 0)
375 					cwin = 0;
376 				len = lmin(len, cwin);
377 			}
378 		}
379 	}
380 
381 	/*
382 	 * Lop off SYN bit if it has already been sent.  However, if this
383 	 * is SYN-SENT state and if segment contains data and if we don't
384 	 * know that foreign host supports TAO, suppress sending segment.
385 	 */
386 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
387 		if (tp->t_state != TCPS_SYN_RECEIVED)
388 			flags &= ~TH_SYN;
389 		off--, len++;
390 	}
391 
392 	/*
393 	 * Be careful not to send data and/or FIN on SYN segments.
394 	 * This measure is needed to prevent interoperability problems
395 	 * with not fully conformant TCP implementations.
396 	 */
397 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
398 		len = 0;
399 		flags &= ~TH_FIN;
400 	}
401 
402 	if (len < 0) {
403 		/*
404 		 * If FIN has been sent but not acked,
405 		 * but we haven't been called to retransmit,
406 		 * len will be < 0.  Otherwise, window shrank
407 		 * after we sent into it.  If window shrank to 0,
408 		 * cancel pending retransmit, pull snd_nxt back
409 		 * to (closed) window, and set the persist timer
410 		 * if it isn't already going.  If the window didn't
411 		 * close completely, just wait for an ACK.
412 		 */
413 		len = 0;
414 		if (sendwin == 0) {
415 			tcp_timer_activate(tp, TT_REXMT, 0);
416 			tp->t_rxtshift = 0;
417 			tp->snd_nxt = tp->snd_una;
418 			if (!tcp_timer_active(tp, TT_PERSIST))
419 				tcp_setpersist(tp);
420 		}
421 	}
422 
423 	/* len will be >= 0 after this point. */
424 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
425 
426 	/*
427 	 * Automatic sizing of send socket buffer.  Often the send buffer
428 	 * size is not optimally adjusted to the actual network conditions
429 	 * at hand (delay bandwidth product).  Setting the buffer size too
430 	 * small limits throughput on links with high bandwidth and high
431 	 * delay (eg. trans-continental/oceanic links).  Setting the
432 	 * buffer size too big consumes too much real kernel memory,
433 	 * especially with many connections on busy servers.
434 	 *
435 	 * The criteria to step up the send buffer one notch are:
436 	 *  1. receive window of remote host is larger than send buffer
437 	 *     (with a fudge factor of 5/4th);
438 	 *  2. send buffer is filled to 7/8th with data (so we actually
439 	 *     have data to make use of it);
440 	 *  3. send buffer fill has not hit maximal automatic size;
441 	 *  4. our send window (slow start and cogestion controlled) is
442 	 *     larger than sent but unacknowledged data in send buffer.
443 	 *
444 	 * The remote host receive window scaling factor may limit the
445 	 * growing of the send buffer before it reaches its allowed
446 	 * maximum.
447 	 *
448 	 * It scales directly with slow start or congestion window
449 	 * and does at most one step per received ACK.  This fast
450 	 * scaling has the drawback of growing the send buffer beyond
451 	 * what is strictly necessary to make full use of a given
452 	 * delay*bandwith product.  However testing has shown this not
453 	 * to be much of an problem.  At worst we are trading wasting
454 	 * of available bandwith (the non-use of it) for wasting some
455 	 * socket buffer memory.
456 	 *
457 	 * TODO: Shrink send buffer during idle periods together
458 	 * with congestion window.  Requires another timer.  Has to
459 	 * wait for upcoming tcp timer rewrite.
460 	 */
461 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
462 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
463 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
464 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
465 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
466 			if (!sbreserve_locked(&so->so_snd,
467 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
468 			     V_tcp_autosndbuf_max), so, curthread))
469 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
470 		}
471 	}
472 
473 	/*
474 	 * Decide if we can use TCP Segmentation Offloading (if supported by
475 	 * hardware).
476 	 *
477 	 * TSO may only be used if we are in a pure bulk sending state.  The
478 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
479 	 * IP options prevent using TSO.  With TSO the TCP header is the same
480 	 * (except for the sequence number) for all generated packets.  This
481 	 * makes it impossible to transmit any options which vary per generated
482 	 * segment or packet.
483 	 */
484 #ifdef IPSEC
485 	/*
486 	 * Pre-calculate here as we save another lookup into the darknesses
487 	 * of IPsec that way and can actually decide if TSO is ok.
488 	 */
489 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
490 #endif
491 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
492 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
493 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
494 #ifdef IPSEC
495 	    ipsec_optlen == 0 &&
496 #endif
497 	    tp->t_inpcb->inp_options == NULL &&
498 	    tp->t_inpcb->in6p_options == NULL)
499 		tso = 1;
500 
501 	if (sack_rxmit) {
502 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
503 			flags &= ~TH_FIN;
504 	} else {
505 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
506 			flags &= ~TH_FIN;
507 	}
508 
509 	recwin = sbspace(&so->so_rcv);
510 
511 	/*
512 	 * Sender silly window avoidance.   We transmit under the following
513 	 * conditions when len is non-zero:
514 	 *
515 	 *	- We have a full segment (or more with TSO)
516 	 *	- This is the last buffer in a write()/send() and we are
517 	 *	  either idle or running NODELAY
518 	 *	- we've timed out (e.g. persist timer)
519 	 *	- we have more then 1/2 the maximum send window's worth of
520 	 *	  data (receiver may be limited the window size)
521 	 *	- we need to retransmit
522 	 */
523 	if (len) {
524 		if (len >= tp->t_maxseg)
525 			goto send;
526 		/*
527 		 * NOTE! on localhost connections an 'ack' from the remote
528 		 * end may occur synchronously with the output and cause
529 		 * us to flush a buffer queued with moretocome.  XXX
530 		 *
531 		 * note: the len + off check is almost certainly unnecessary.
532 		 */
533 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
534 		    (idle || (tp->t_flags & TF_NODELAY)) &&
535 		    len + off >= so->so_snd.sb_cc &&
536 		    (tp->t_flags & TF_NOPUSH) == 0) {
537 			goto send;
538 		}
539 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
540 			goto send;
541 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
542 			goto send;
543 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
544 			goto send;
545 		if (sack_rxmit)
546 			goto send;
547 	}
548 
549 	/*
550 	 * Sending of standalone window updates.
551 	 *
552 	 * Window updates are important when we close our window due to a
553 	 * full socket buffer and are opening it again after the application
554 	 * reads data from it.  Once the window has opened again and the
555 	 * remote end starts to send again the ACK clock takes over and
556 	 * provides the most current window information.
557 	 *
558 	 * We must avoid the silly window syndrome whereas every read
559 	 * from the receive buffer, no matter how small, causes a window
560 	 * update to be sent.  We also should avoid sending a flurry of
561 	 * window updates when the socket buffer had queued a lot of data
562 	 * and the application is doing small reads.
563 	 *
564 	 * Prevent a flurry of pointless window updates by only sending
565 	 * an update when we can increase the advertized window by more
566 	 * than 1/4th of the socket buffer capacity.  When the buffer is
567 	 * getting full or is very small be more aggressive and send an
568 	 * update whenever we can increase by two mss sized segments.
569 	 * In all other situations the ACK's to new incoming data will
570 	 * carry further window increases.
571 	 *
572 	 * Don't send an independent window update if a delayed
573 	 * ACK is pending (it will get piggy-backed on it) or the
574 	 * remote side already has done a half-close and won't send
575 	 * more data.  Skip this if the connection is in T/TCP
576 	 * half-open state.
577 	 */
578 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
579 	    !(tp->t_flags & TF_DELACK) &&
580 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
581 		/*
582 		 * "adv" is the amount we could increase the window,
583 		 * taking into account that we are limited by
584 		 * TCP_MAXWIN << tp->rcv_scale.
585 		 */
586 		long adv;
587 		int oldwin;
588 
589 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
590 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
591 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
592 			adv -= oldwin;
593 		} else
594 			oldwin = 0;
595 
596 		/*
597 		 * If the new window size ends up being the same as the old
598 		 * size when it is scaled, then don't force a window update.
599 		 */
600 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
601 			goto dontupdate;
602 
603 		if (adv >= (long)(2 * tp->t_maxseg) &&
604 		    (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
605 		     recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
606 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
607 			goto send;
608 	}
609 dontupdate:
610 
611 	/*
612 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
613 	 * is also a catch-all for the retransmit timer timeout case.
614 	 */
615 	if (tp->t_flags & TF_ACKNOW)
616 		goto send;
617 	if ((flags & TH_RST) ||
618 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
619 		goto send;
620 	if (SEQ_GT(tp->snd_up, tp->snd_una))
621 		goto send;
622 	/*
623 	 * If our state indicates that FIN should be sent
624 	 * and we have not yet done so, then we need to send.
625 	 */
626 	if (flags & TH_FIN &&
627 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
628 		goto send;
629 	/*
630 	 * In SACK, it is possible for tcp_output to fail to send a segment
631 	 * after the retransmission timer has been turned off.  Make sure
632 	 * that the retransmission timer is set.
633 	 */
634 	if ((tp->t_flags & TF_SACK_PERMIT) &&
635 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
636 	    !tcp_timer_active(tp, TT_REXMT) &&
637 	    !tcp_timer_active(tp, TT_PERSIST)) {
638 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
639 		goto just_return;
640 	}
641 	/*
642 	 * TCP window updates are not reliable, rather a polling protocol
643 	 * using ``persist'' packets is used to insure receipt of window
644 	 * updates.  The three ``states'' for the output side are:
645 	 *	idle			not doing retransmits or persists
646 	 *	persisting		to move a small or zero window
647 	 *	(re)transmitting	and thereby not persisting
648 	 *
649 	 * tcp_timer_active(tp, TT_PERSIST)
650 	 *	is true when we are in persist state.
651 	 * (tp->t_flags & TF_FORCEDATA)
652 	 *	is set when we are called to send a persist packet.
653 	 * tcp_timer_active(tp, TT_REXMT)
654 	 *	is set when we are retransmitting
655 	 * The output side is idle when both timers are zero.
656 	 *
657 	 * If send window is too small, there is data to transmit, and no
658 	 * retransmit or persist is pending, then go to persist state.
659 	 * If nothing happens soon, send when timer expires:
660 	 * if window is nonzero, transmit what we can,
661 	 * otherwise force out a byte.
662 	 */
663 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
664 	    !tcp_timer_active(tp, TT_PERSIST)) {
665 		tp->t_rxtshift = 0;
666 		tcp_setpersist(tp);
667 	}
668 
669 	/*
670 	 * No reason to send a segment, just return.
671 	 */
672 just_return:
673 	SOCKBUF_UNLOCK(&so->so_snd);
674 	return (0);
675 
676 send:
677 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
678 	if (len > 0) {
679 		if (len >= tp->t_maxseg)
680 			tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
681 		else
682 			tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
683 	}
684 	/*
685 	 * Before ESTABLISHED, force sending of initial options
686 	 * unless TCP set not to do any options.
687 	 * NOTE: we assume that the IP/TCP header plus TCP options
688 	 * always fit in a single mbuf, leaving room for a maximum
689 	 * link header, i.e.
690 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
691 	 */
692 	optlen = 0;
693 #ifdef INET6
694 	if (isipv6)
695 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
696 	else
697 #endif
698 		hdrlen = sizeof (struct tcpiphdr);
699 
700 	/*
701 	 * Compute options for segment.
702 	 * We only have to care about SYN and established connection
703 	 * segments.  Options for SYN-ACK segments are handled in TCP
704 	 * syncache.
705 	 */
706 	if ((tp->t_flags & TF_NOOPT) == 0) {
707 		to.to_flags = 0;
708 		/* Maximum segment size. */
709 		if (flags & TH_SYN) {
710 			tp->snd_nxt = tp->iss;
711 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
712 			to.to_flags |= TOF_MSS;
713 		}
714 		/* Window scaling. */
715 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
716 			to.to_wscale = tp->request_r_scale;
717 			to.to_flags |= TOF_SCALE;
718 		}
719 		/* Timestamps. */
720 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
721 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
722 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
723 			to.to_tsecr = tp->ts_recent;
724 			to.to_flags |= TOF_TS;
725 			/* Set receive buffer autosizing timestamp. */
726 			if (tp->rfbuf_ts == 0 &&
727 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
728 				tp->rfbuf_ts = tcp_ts_getticks();
729 		}
730 		/* Selective ACK's. */
731 		if (tp->t_flags & TF_SACK_PERMIT) {
732 			if (flags & TH_SYN)
733 				to.to_flags |= TOF_SACKPERM;
734 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
735 			    (tp->t_flags & TF_SACK_PERMIT) &&
736 			    tp->rcv_numsacks > 0) {
737 				to.to_flags |= TOF_SACK;
738 				to.to_nsacks = tp->rcv_numsacks;
739 				to.to_sacks = (u_char *)tp->sackblks;
740 			}
741 		}
742 #ifdef TCP_SIGNATURE
743 		/* TCP-MD5 (RFC2385). */
744 		if (tp->t_flags & TF_SIGNATURE)
745 			to.to_flags |= TOF_SIGNATURE;
746 #endif /* TCP_SIGNATURE */
747 
748 		/* Processing the options. */
749 		hdrlen += optlen = tcp_addoptions(&to, opt);
750 	}
751 
752 #ifdef INET6
753 	if (isipv6)
754 		ipoptlen = ip6_optlen(tp->t_inpcb);
755 	else
756 #endif
757 	if (tp->t_inpcb->inp_options)
758 		ipoptlen = tp->t_inpcb->inp_options->m_len -
759 				offsetof(struct ipoption, ipopt_list);
760 	else
761 		ipoptlen = 0;
762 #ifdef IPSEC
763 	ipoptlen += ipsec_optlen;
764 #endif
765 
766 	/*
767 	 * Adjust data length if insertion of options will
768 	 * bump the packet length beyond the t_maxopd length.
769 	 * Clear the FIN bit because we cut off the tail of
770 	 * the segment.
771 	 */
772 	if (len + optlen + ipoptlen > tp->t_maxopd) {
773 		flags &= ~TH_FIN;
774 
775 		if (tso) {
776 			u_int if_hw_tsomax;
777 			u_int if_hw_tsomaxsegcount;
778 			u_int if_hw_tsomaxsegsize;
779 			struct mbuf *mb;
780 			u_int moff;
781 			int max_len;
782 
783 			/* extract TSO information */
784 			if_hw_tsomax = tp->t_tsomax;
785 			if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
786 			if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
787 
788 			/*
789 			 * Limit a TSO burst to prevent it from
790 			 * overflowing or exceeding the maximum length
791 			 * allowed by the network interface:
792 			 */
793 			KASSERT(ipoptlen == 0,
794 			    ("%s: TSO can't do IP options", __func__));
795 
796 			/*
797 			 * Check if we should limit by maximum payload
798 			 * length:
799 			 */
800 			if (if_hw_tsomax != 0) {
801 				/* compute maximum TSO length */
802 				max_len = (if_hw_tsomax - hdrlen);
803 				if (max_len <= 0) {
804 					len = 0;
805 				} else if (len > (u_int)max_len) {
806 					sendalot = 1;
807 					len = (u_int)max_len;
808 				}
809 			}
810 
811 			/*
812 			 * Check if we should limit by maximum segment
813 			 * size and count:
814 			 */
815 			if (if_hw_tsomaxsegcount != 0 &&
816 			    if_hw_tsomaxsegsize != 0) {
817 				max_len = 0;
818 				mb = sbsndmbuf(&so->so_snd, off, &moff);
819 
820 				while (mb != NULL && (u_int)max_len < len) {
821 					u_int mlen;
822 					u_int frags;
823 
824 					/*
825 					 * Get length of mbuf fragment
826 					 * and how many hardware frags,
827 					 * rounded up, it would use:
828 					 */
829 					mlen = (mb->m_len - moff);
830 					frags = howmany(mlen,
831 					    if_hw_tsomaxsegsize);
832 
833 					/* Handle special case: Zero Length Mbuf */
834 					if (frags == 0)
835 						frags = 1;
836 
837 					/*
838 					 * Check if the fragment limit
839 					 * will be reached or exceeded:
840 					 */
841 					if (frags >= if_hw_tsomaxsegcount) {
842 						max_len += min(mlen,
843 						    if_hw_tsomaxsegcount *
844 						    if_hw_tsomaxsegsize);
845 						break;
846 					}
847 					max_len += mlen;
848 					if_hw_tsomaxsegcount -= frags;
849 					moff = 0;
850 					mb = mb->m_next;
851 				}
852 				if (max_len <= 0) {
853 					len = 0;
854 				} else if (len > (u_int)max_len) {
855 					sendalot = 1;
856 					len = (u_int)max_len;
857 				}
858 			}
859 
860 			/*
861 			 * Prevent the last segment from being
862 			 * fractional unless the send sockbuf can be
863 			 * emptied:
864 			 */
865 			max_len = (tp->t_maxopd - optlen);
866 			if ((off + len) < so->so_snd.sb_cc) {
867 				moff = len % (u_int)max_len;
868 				if (moff != 0) {
869 					len -= moff;
870 					sendalot = 1;
871 				}
872 			}
873 
874 			/*
875 			 * In case there are too many small fragments
876 			 * don't use TSO:
877 			 */
878 			if (len <= (u_int)max_len) {
879 				len = (u_int)max_len;
880 				sendalot = 1;
881 				tso = 0;
882 			}
883 
884 			/*
885 			 * Send the FIN in a separate segment
886 			 * after the bulk sending is done.
887 			 * We don't trust the TSO implementations
888 			 * to clear the FIN flag on all but the
889 			 * last segment.
890 			 */
891 			if (tp->t_flags & TF_NEEDFIN)
892 				sendalot = 1;
893 
894 		} else {
895 			len = tp->t_maxopd - optlen - ipoptlen;
896 			sendalot = 1;
897 		}
898 	} else
899 		tso = 0;
900 
901 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
902 	    ("%s: len > IP_MAXPACKET", __func__));
903 
904 /*#ifdef DIAGNOSTIC*/
905 #ifdef INET6
906 	if (max_linkhdr + hdrlen > MCLBYTES)
907 #else
908 	if (max_linkhdr + hdrlen > MHLEN)
909 #endif
910 		panic("tcphdr too big");
911 /*#endif*/
912 
913 	/*
914 	 * This KASSERT is here to catch edge cases at a well defined place.
915 	 * Before, those had triggered (random) panic conditions further down.
916 	 */
917 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
918 
919 	/*
920 	 * Grab a header mbuf, attaching a copy of data to
921 	 * be transmitted, and initialize the header from
922 	 * the template for sends on this connection.
923 	 */
924 	if (len) {
925 		struct mbuf *mb;
926 		u_int moff;
927 
928 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
929 			TCPSTAT_INC(tcps_sndprobe);
930 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
931 			tp->t_sndrexmitpack++;
932 			TCPSTAT_INC(tcps_sndrexmitpack);
933 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
934 		} else {
935 			TCPSTAT_INC(tcps_sndpack);
936 			TCPSTAT_ADD(tcps_sndbyte, len);
937 		}
938 #ifdef INET6
939 		if (MHLEN < hdrlen + max_linkhdr)
940 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
941 		else
942 #endif
943 			m = m_gethdr(M_NOWAIT, MT_DATA);
944 
945 		if (m == NULL) {
946 			SOCKBUF_UNLOCK(&so->so_snd);
947 			error = ENOBUFS;
948 			sack_rxmit = 0;
949 			goto out;
950 		}
951 
952 		m->m_data += max_linkhdr;
953 		m->m_len = hdrlen;
954 
955 		/*
956 		 * Start the m_copy functions from the closest mbuf
957 		 * to the offset in the socket buffer chain.
958 		 */
959 		mb = sbsndptr(&so->so_snd, off, len, &moff);
960 
961 		if (len <= MHLEN - hdrlen - max_linkhdr) {
962 			m_copydata(mb, moff, (int)len,
963 			    mtod(m, caddr_t) + hdrlen);
964 			m->m_len += len;
965 		} else {
966 			m->m_next = m_copy(mb, moff, (int)len);
967 			if (m->m_next == NULL) {
968 				SOCKBUF_UNLOCK(&so->so_snd);
969 				(void) m_free(m);
970 				error = ENOBUFS;
971 				sack_rxmit = 0;
972 				goto out;
973 			}
974 		}
975 
976 		/*
977 		 * If we're sending everything we've got, set PUSH.
978 		 * (This will keep happy those implementations which only
979 		 * give data to the user when a buffer fills or
980 		 * a PUSH comes in.)
981 		 */
982 		if (off + len == so->so_snd.sb_cc)
983 			flags |= TH_PUSH;
984 		SOCKBUF_UNLOCK(&so->so_snd);
985 	} else {
986 		SOCKBUF_UNLOCK(&so->so_snd);
987 		if (tp->t_flags & TF_ACKNOW)
988 			TCPSTAT_INC(tcps_sndacks);
989 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
990 			TCPSTAT_INC(tcps_sndctrl);
991 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
992 			TCPSTAT_INC(tcps_sndurg);
993 		else
994 			TCPSTAT_INC(tcps_sndwinup);
995 
996 		m = m_gethdr(M_NOWAIT, MT_DATA);
997 		if (m == NULL) {
998 			error = ENOBUFS;
999 			sack_rxmit = 0;
1000 			goto out;
1001 		}
1002 #ifdef INET6
1003 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1004 		    MHLEN >= hdrlen) {
1005 			MH_ALIGN(m, hdrlen);
1006 		} else
1007 #endif
1008 		m->m_data += max_linkhdr;
1009 		m->m_len = hdrlen;
1010 	}
1011 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1012 	m->m_pkthdr.rcvif = (struct ifnet *)0;
1013 #ifdef MAC
1014 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
1015 #endif
1016 #ifdef INET6
1017 	if (isipv6) {
1018 		ip6 = mtod(m, struct ip6_hdr *);
1019 		th = (struct tcphdr *)(ip6 + 1);
1020 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
1021 	} else
1022 #endif /* INET6 */
1023 	{
1024 		ip = mtod(m, struct ip *);
1025 		ipov = (struct ipovly *)ip;
1026 		th = (struct tcphdr *)(ip + 1);
1027 		tcpip_fillheaders(tp->t_inpcb, ip, th);
1028 	}
1029 
1030 	/*
1031 	 * Fill in fields, remembering maximum advertised
1032 	 * window for use in delaying messages about window sizes.
1033 	 * If resending a FIN, be sure not to use a new sequence number.
1034 	 */
1035 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1036 	    tp->snd_nxt == tp->snd_max)
1037 		tp->snd_nxt--;
1038 	/*
1039 	 * If we are starting a connection, send ECN setup
1040 	 * SYN packet. If we are on a retransmit, we may
1041 	 * resend those bits a number of times as per
1042 	 * RFC 3168.
1043 	 */
1044 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1045 		if (tp->t_rxtshift >= 1) {
1046 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1047 				flags |= TH_ECE|TH_CWR;
1048 		} else
1049 			flags |= TH_ECE|TH_CWR;
1050 	}
1051 
1052 	if (tp->t_state == TCPS_ESTABLISHED &&
1053 	    (tp->t_flags & TF_ECN_PERMIT)) {
1054 		/*
1055 		 * If the peer has ECN, mark data packets with
1056 		 * ECN capable transmission (ECT).
1057 		 * Ignore pure ack packets, retransmissions and window probes.
1058 		 */
1059 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1060 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1061 #ifdef INET6
1062 			if (isipv6)
1063 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1064 			else
1065 #endif
1066 				ip->ip_tos |= IPTOS_ECN_ECT0;
1067 			TCPSTAT_INC(tcps_ecn_ect0);
1068 		}
1069 
1070 		/*
1071 		 * Reply with proper ECN notifications.
1072 		 */
1073 		if (tp->t_flags & TF_ECN_SND_CWR) {
1074 			flags |= TH_CWR;
1075 			tp->t_flags &= ~TF_ECN_SND_CWR;
1076 		}
1077 		if (tp->t_flags & TF_ECN_SND_ECE)
1078 			flags |= TH_ECE;
1079 	}
1080 
1081 	/*
1082 	 * If we are doing retransmissions, then snd_nxt will
1083 	 * not reflect the first unsent octet.  For ACK only
1084 	 * packets, we do not want the sequence number of the
1085 	 * retransmitted packet, we want the sequence number
1086 	 * of the next unsent octet.  So, if there is no data
1087 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
1088 	 * when filling in ti_seq.  But if we are in persist
1089 	 * state, snd_max might reflect one byte beyond the
1090 	 * right edge of the window, so use snd_nxt in that
1091 	 * case, since we know we aren't doing a retransmission.
1092 	 * (retransmit and persist are mutually exclusive...)
1093 	 */
1094 	if (sack_rxmit == 0) {
1095 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1096 		    tcp_timer_active(tp, TT_PERSIST))
1097 			th->th_seq = htonl(tp->snd_nxt);
1098 		else
1099 			th->th_seq = htonl(tp->snd_max);
1100 	} else {
1101 		th->th_seq = htonl(p->rxmit);
1102 		p->rxmit += len;
1103 		tp->sackhint.sack_bytes_rexmit += len;
1104 	}
1105 	th->th_ack = htonl(tp->rcv_nxt);
1106 	if (optlen) {
1107 		bcopy(opt, th + 1, optlen);
1108 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1109 	}
1110 	th->th_flags = flags;
1111 	/*
1112 	 * Calculate receive window.  Don't shrink window,
1113 	 * but avoid silly window syndrome.
1114 	 */
1115 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1116 	    recwin < (long)tp->t_maxseg)
1117 		recwin = 0;
1118 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1119 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1120 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1121 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1122 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1123 
1124 	/*
1125 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1126 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1127 	 * case is handled in syncache.
1128 	 */
1129 	if (flags & TH_SYN)
1130 		th->th_win = htons((u_short)
1131 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1132 	else
1133 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1134 
1135 	/*
1136 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1137 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1138 	 * flag tells soreceive() to disable delayed acknowledgements when
1139 	 * draining the buffer.  This can occur if the receiver is attempting
1140 	 * to read more data than can be buffered prior to transmitting on
1141 	 * the connection.
1142 	 */
1143 	if (th->th_win == 0) {
1144 		tp->t_sndzerowin++;
1145 		tp->t_flags |= TF_RXWIN0SENT;
1146 	} else
1147 		tp->t_flags &= ~TF_RXWIN0SENT;
1148 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1149 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1150 		th->th_flags |= TH_URG;
1151 	} else
1152 		/*
1153 		 * If no urgent pointer to send, then we pull
1154 		 * the urgent pointer to the left edge of the send window
1155 		 * so that it doesn't drift into the send window on sequence
1156 		 * number wraparound.
1157 		 */
1158 		tp->snd_up = tp->snd_una;		/* drag it along */
1159 
1160 #ifdef TCP_SIGNATURE
1161 	if (tp->t_flags & TF_SIGNATURE) {
1162 		int sigoff = to.to_signature - opt;
1163 		tcp_signature_compute(m, 0, len, optlen,
1164 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1165 	}
1166 #endif
1167 
1168 	/*
1169 	 * Put TCP length in extended header, and then
1170 	 * checksum extended header and data.
1171 	 */
1172 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1173 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1174 #ifdef INET6
1175 	if (isipv6) {
1176 		/*
1177 		 * ip6_plen is not need to be filled now, and will be filled
1178 		 * in ip6_output.
1179 		 */
1180 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1181 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1182 		    optlen + len, IPPROTO_TCP, 0);
1183 	}
1184 #endif
1185 #if defined(INET6) && defined(INET)
1186 	else
1187 #endif
1188 #ifdef INET
1189 	{
1190 		m->m_pkthdr.csum_flags = CSUM_TCP;
1191 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1192 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1193 
1194 		/* IP version must be set here for ipv4/ipv6 checking later */
1195 		KASSERT(ip->ip_v == IPVERSION,
1196 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1197 	}
1198 #endif
1199 
1200 	/*
1201 	 * Enable TSO and specify the size of the segments.
1202 	 * The TCP pseudo header checksum is always provided.
1203 	 * XXX: Fixme: This is currently not the case for IPv6.
1204 	 */
1205 	if (tso) {
1206 		KASSERT(len > tp->t_maxopd - optlen,
1207 		    ("%s: len <= tso_segsz", __func__));
1208 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1209 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1210 	}
1211 
1212 #ifdef IPSEC
1213 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1214 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1215 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1216 #else
1217 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1218 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1219 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1220 #endif
1221 
1222 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1223 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1224 
1225 #ifdef TCPDEBUG
1226 	/*
1227 	 * Trace.
1228 	 */
1229 	if (so->so_options & SO_DEBUG) {
1230 		u_short save = 0;
1231 #ifdef INET6
1232 		if (!isipv6)
1233 #endif
1234 		{
1235 			save = ipov->ih_len;
1236 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1237 		}
1238 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1239 #ifdef INET6
1240 		if (!isipv6)
1241 #endif
1242 		ipov->ih_len = save;
1243 	}
1244 #endif /* TCPDEBUG */
1245 
1246 	/*
1247 	 * Fill in IP length and desired time to live and
1248 	 * send to IP level.  There should be a better way
1249 	 * to handle ttl and tos; we could keep them in
1250 	 * the template, but need a way to checksum without them.
1251 	 */
1252 	/*
1253 	 * m->m_pkthdr.len should have been set before checksum calculation,
1254 	 * because in6_cksum() need it.
1255 	 */
1256 #ifdef INET6
1257 	if (isipv6) {
1258 		struct route_in6 ro;
1259 
1260 		bzero(&ro, sizeof(ro));
1261 		/*
1262 		 * we separately set hoplimit for every segment, since the
1263 		 * user might want to change the value via setsockopt.
1264 		 * Also, desired default hop limit might be changed via
1265 		 * Neighbor Discovery.
1266 		 */
1267 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1268 
1269 		/*
1270 		 * Set the packet size here for the benefit of DTrace probes.
1271 		 * ip6_output() will set it properly; it's supposed to include
1272 		 * the option header lengths as well.
1273 		 */
1274 		ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1275 
1276 		if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1277 			tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1278 		else
1279 			tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1280 
1281 		if (tp->t_state == TCPS_SYN_SENT)
1282 			TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1283 
1284 		TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1285 
1286 		/* TODO: IPv6 IP6TOS_ECT bit on */
1287 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1288 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1289 		    NULL, NULL, tp->t_inpcb);
1290 
1291 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1292 			mtu = ro.ro_rt->rt_mtu;
1293 		RO_RTFREE(&ro);
1294 	}
1295 #endif /* INET6 */
1296 #if defined(INET) && defined(INET6)
1297 	else
1298 #endif
1299 #ifdef INET
1300     {
1301 	struct route ro;
1302 
1303 	bzero(&ro, sizeof(ro));
1304 	ip->ip_len = htons(m->m_pkthdr.len);
1305 #ifdef INET6
1306 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1307 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1308 #endif /* INET6 */
1309 	/*
1310 	 * If we do path MTU discovery, then we set DF on every packet.
1311 	 * This might not be the best thing to do according to RFC3390
1312 	 * Section 2. However the tcp hostcache migitates the problem
1313 	 * so it affects only the first tcp connection with a host.
1314 	 *
1315 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1316 	 */
1317 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) {
1318 		ip->ip_off |= htons(IP_DF);
1319 		tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1320 	} else {
1321 		tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1322 	}
1323 
1324 	if (tp->t_state == TCPS_SYN_SENT)
1325 		TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1326 
1327 	TCP_PROBE5(send, NULL, tp, ip, tp, th);
1328 
1329 	error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1330 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1331 	    tp->t_inpcb);
1332 
1333 	if (error == EMSGSIZE && ro.ro_rt != NULL)
1334 		mtu = ro.ro_rt->rt_mtu;
1335 	RO_RTFREE(&ro);
1336     }
1337 #endif /* INET */
1338 
1339 out:
1340 	/*
1341 	 * In transmit state, time the transmission and arrange for
1342 	 * the retransmit.  In persist state, just set snd_max.
1343 	 */
1344 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1345 	    !tcp_timer_active(tp, TT_PERSIST)) {
1346 		tcp_seq startseq = tp->snd_nxt;
1347 
1348 		/*
1349 		 * Advance snd_nxt over sequence space of this segment.
1350 		 */
1351 		if (flags & (TH_SYN|TH_FIN)) {
1352 			if (flags & TH_SYN)
1353 				tp->snd_nxt++;
1354 			if (flags & TH_FIN) {
1355 				tp->snd_nxt++;
1356 				tp->t_flags |= TF_SENTFIN;
1357 			}
1358 		}
1359 		if (sack_rxmit)
1360 			goto timer;
1361 		tp->snd_nxt += len;
1362 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1363 			tp->snd_max = tp->snd_nxt;
1364 			/*
1365 			 * Time this transmission if not a retransmission and
1366 			 * not currently timing anything.
1367 			 */
1368 			if (tp->t_rtttime == 0) {
1369 				tp->t_rtttime = ticks;
1370 				tp->t_rtseq = startseq;
1371 				TCPSTAT_INC(tcps_segstimed);
1372 			}
1373 		}
1374 
1375 		/*
1376 		 * Set retransmit timer if not currently set,
1377 		 * and not doing a pure ack or a keep-alive probe.
1378 		 * Initial value for retransmit timer is smoothed
1379 		 * round-trip time + 2 * round-trip time variance.
1380 		 * Initialize shift counter which is used for backoff
1381 		 * of retransmit time.
1382 		 */
1383 timer:
1384 		if (!tcp_timer_active(tp, TT_REXMT) &&
1385 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1386 		     (tp->snd_nxt != tp->snd_una))) {
1387 			if (tcp_timer_active(tp, TT_PERSIST)) {
1388 				tcp_timer_activate(tp, TT_PERSIST, 0);
1389 				tp->t_rxtshift = 0;
1390 			}
1391 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1392 		}
1393 	} else {
1394 		/*
1395 		 * Persist case, update snd_max but since we are in
1396 		 * persist mode (no window) we do not update snd_nxt.
1397 		 */
1398 		int xlen = len;
1399 		if (flags & TH_SYN)
1400 			++xlen;
1401 		if (flags & TH_FIN) {
1402 			++xlen;
1403 			tp->t_flags |= TF_SENTFIN;
1404 		}
1405 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1406 			tp->snd_max = tp->snd_nxt + len;
1407 	}
1408 
1409 	if (error) {
1410 
1411 		/*
1412 		 * We know that the packet was lost, so back out the
1413 		 * sequence number advance, if any.
1414 		 *
1415 		 * If the error is EPERM the packet got blocked by the
1416 		 * local firewall.  Normally we should terminate the
1417 		 * connection but the blocking may have been spurious
1418 		 * due to a firewall reconfiguration cycle.  So we treat
1419 		 * it like a packet loss and let the retransmit timer and
1420 		 * timeouts do their work over time.
1421 		 * XXX: It is a POLA question whether calling tcp_drop right
1422 		 * away would be the really correct behavior instead.
1423 		 */
1424 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1425 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1426 		    ((flags & TH_SYN) == 0) &&
1427 		    (error != EPERM)) {
1428 			if (sack_rxmit) {
1429 				p->rxmit -= len;
1430 				tp->sackhint.sack_bytes_rexmit -= len;
1431 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1432 				    ("sackhint bytes rtx >= 0"));
1433 			} else
1434 				tp->snd_nxt -= len;
1435 		}
1436 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1437 		switch (error) {
1438 		case EPERM:
1439 			tp->t_softerror = error;
1440 			return (error);
1441 		case ENOBUFS:
1442 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1443 			    !tcp_timer_active(tp, TT_PERSIST))
1444 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1445 			tp->snd_cwnd = tp->t_maxseg;
1446 			return (0);
1447 		case EMSGSIZE:
1448 			/*
1449 			 * For some reason the interface we used initially
1450 			 * to send segments changed to another or lowered
1451 			 * its MTU.
1452 			 * If TSO was active we either got an interface
1453 			 * without TSO capabilits or TSO was turned off.
1454 			 * If we obtained mtu from ip_output() then update
1455 			 * it and try again.
1456 			 */
1457 			if (tso)
1458 				tp->t_flags &= ~TF_TSO;
1459 			if (mtu != 0) {
1460 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1461 				goto again;
1462 			}
1463 			return (error);
1464 		case EHOSTDOWN:
1465 		case EHOSTUNREACH:
1466 		case ENETDOWN:
1467 		case ENETUNREACH:
1468 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1469 				tp->t_softerror = error;
1470 				return (0);
1471 			}
1472 			/* FALLTHROUGH */
1473 		default:
1474 			return (error);
1475 		}
1476 	}
1477 	TCPSTAT_INC(tcps_sndtotal);
1478 
1479 	/*
1480 	 * Data sent (as far as we can tell).
1481 	 * If this advertises a larger window than any other segment,
1482 	 * then remember the size of the advertised window.
1483 	 * Any pending ACK has now been sent.
1484 	 */
1485 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1486 		tp->rcv_adv = tp->rcv_nxt + recwin;
1487 	tp->last_ack_sent = tp->rcv_nxt;
1488 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1489 	if (tcp_timer_active(tp, TT_DELACK))
1490 		tcp_timer_activate(tp, TT_DELACK, 0);
1491 #if 0
1492 	/*
1493 	 * This completely breaks TCP if newreno is turned on.  What happens
1494 	 * is that if delayed-acks are turned on on the receiver, this code
1495 	 * on the transmitter effectively destroys the TCP window, forcing
1496 	 * it to four packets (1.5Kx4 = 6K window).
1497 	 */
1498 	if (sendalot && --maxburst)
1499 		goto again;
1500 #endif
1501 	if (sendalot)
1502 		goto again;
1503 	return (0);
1504 }
1505 
1506 void
1507 tcp_setpersist(struct tcpcb *tp)
1508 {
1509 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1510 	int tt;
1511 
1512 	tp->t_flags &= ~TF_PREVVALID;
1513 	if (tcp_timer_active(tp, TT_REXMT))
1514 		panic("tcp_setpersist: retransmit pending");
1515 	/*
1516 	 * Start/restart persistance timer.
1517 	 */
1518 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1519 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1520 	tcp_timer_activate(tp, TT_PERSIST, tt);
1521 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1522 		tp->t_rxtshift++;
1523 }
1524 
1525 /*
1526  * Insert TCP options according to the supplied parameters to the place
1527  * optp in a consistent way.  Can handle unaligned destinations.
1528  *
1529  * The order of the option processing is crucial for optimal packing and
1530  * alignment for the scarce option space.
1531  *
1532  * The optimal order for a SYN/SYN-ACK segment is:
1533  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1534  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1535  *
1536  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1537  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1538  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1539  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1540  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1541  */
1542 int
1543 tcp_addoptions(struct tcpopt *to, u_char *optp)
1544 {
1545 	u_int mask, optlen = 0;
1546 
1547 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1548 		if ((to->to_flags & mask) != mask)
1549 			continue;
1550 		if (optlen == TCP_MAXOLEN)
1551 			break;
1552 		switch (to->to_flags & mask) {
1553 		case TOF_MSS:
1554 			while (optlen % 4) {
1555 				optlen += TCPOLEN_NOP;
1556 				*optp++ = TCPOPT_NOP;
1557 			}
1558 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1559 				continue;
1560 			optlen += TCPOLEN_MAXSEG;
1561 			*optp++ = TCPOPT_MAXSEG;
1562 			*optp++ = TCPOLEN_MAXSEG;
1563 			to->to_mss = htons(to->to_mss);
1564 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1565 			optp += sizeof(to->to_mss);
1566 			break;
1567 		case TOF_SCALE:
1568 			while (!optlen || optlen % 2 != 1) {
1569 				optlen += TCPOLEN_NOP;
1570 				*optp++ = TCPOPT_NOP;
1571 			}
1572 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1573 				continue;
1574 			optlen += TCPOLEN_WINDOW;
1575 			*optp++ = TCPOPT_WINDOW;
1576 			*optp++ = TCPOLEN_WINDOW;
1577 			*optp++ = to->to_wscale;
1578 			break;
1579 		case TOF_SACKPERM:
1580 			while (optlen % 2) {
1581 				optlen += TCPOLEN_NOP;
1582 				*optp++ = TCPOPT_NOP;
1583 			}
1584 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1585 				continue;
1586 			optlen += TCPOLEN_SACK_PERMITTED;
1587 			*optp++ = TCPOPT_SACK_PERMITTED;
1588 			*optp++ = TCPOLEN_SACK_PERMITTED;
1589 			break;
1590 		case TOF_TS:
1591 			while (!optlen || optlen % 4 != 2) {
1592 				optlen += TCPOLEN_NOP;
1593 				*optp++ = TCPOPT_NOP;
1594 			}
1595 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1596 				continue;
1597 			optlen += TCPOLEN_TIMESTAMP;
1598 			*optp++ = TCPOPT_TIMESTAMP;
1599 			*optp++ = TCPOLEN_TIMESTAMP;
1600 			to->to_tsval = htonl(to->to_tsval);
1601 			to->to_tsecr = htonl(to->to_tsecr);
1602 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1603 			optp += sizeof(to->to_tsval);
1604 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1605 			optp += sizeof(to->to_tsecr);
1606 			break;
1607 		case TOF_SIGNATURE:
1608 			{
1609 			int siglen = TCPOLEN_SIGNATURE - 2;
1610 
1611 			while (!optlen || optlen % 4 != 2) {
1612 				optlen += TCPOLEN_NOP;
1613 				*optp++ = TCPOPT_NOP;
1614 			}
1615 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1616 				continue;
1617 			optlen += TCPOLEN_SIGNATURE;
1618 			*optp++ = TCPOPT_SIGNATURE;
1619 			*optp++ = TCPOLEN_SIGNATURE;
1620 			to->to_signature = optp;
1621 			while (siglen--)
1622 				 *optp++ = 0;
1623 			break;
1624 			}
1625 		case TOF_SACK:
1626 			{
1627 			int sackblks = 0;
1628 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1629 			tcp_seq sack_seq;
1630 
1631 			while (!optlen || optlen % 4 != 2) {
1632 				optlen += TCPOLEN_NOP;
1633 				*optp++ = TCPOPT_NOP;
1634 			}
1635 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1636 				continue;
1637 			optlen += TCPOLEN_SACKHDR;
1638 			*optp++ = TCPOPT_SACK;
1639 			sackblks = min(to->to_nsacks,
1640 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1641 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1642 			while (sackblks--) {
1643 				sack_seq = htonl(sack->start);
1644 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1645 				optp += sizeof(sack_seq);
1646 				sack_seq = htonl(sack->end);
1647 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1648 				optp += sizeof(sack_seq);
1649 				optlen += TCPOLEN_SACK;
1650 				sack++;
1651 			}
1652 			TCPSTAT_INC(tcps_sack_send_blocks);
1653 			break;
1654 			}
1655 		default:
1656 			panic("%s: unknown TCP option type", __func__);
1657 			break;
1658 		}
1659 	}
1660 
1661 	/* Terminate and pad TCP options to a 4 byte boundary. */
1662 	if (optlen % 4) {
1663 		optlen += TCPOLEN_EOL;
1664 		*optp++ = TCPOPT_EOL;
1665 	}
1666 	/*
1667 	 * According to RFC 793 (STD0007):
1668 	 *   "The content of the header beyond the End-of-Option option
1669 	 *    must be header padding (i.e., zero)."
1670 	 *   and later: "The padding is composed of zeros."
1671 	 */
1672 	while (optlen % 4) {
1673 		optlen += TCPOLEN_PAD;
1674 		*optp++ = TCPOPT_PAD;
1675 	}
1676 
1677 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1678 	return (optlen);
1679 }
1680