xref: /freebsd/sys/netinet/tcp_output.c (revision 03836978bec158bdc0ecee7a4198962f91ce8298)
1 /*-
2  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3  *	The Regents of the University of California.  All rights reserved.
4  *
5  * Redistribution and use in source and binary forms, with or without
6  * modification, are permitted provided that the following conditions
7  * are met:
8  * 1. Redistributions of source code must retain the above copyright
9  *    notice, this list of conditions and the following disclaimer.
10  * 2. Redistributions in binary form must reproduce the above copyright
11  *    notice, this list of conditions and the following disclaimer in the
12  *    documentation and/or other materials provided with the distribution.
13  * 4. Neither the name of the University nor the names of its contributors
14  *    may be used to endorse or promote products derived from this software
15  *    without specific prior written permission.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  *
29  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include "opt_inet.h"
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
39 
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/hhook.h>
44 #include <sys/kernel.h>
45 #include <sys/lock.h>
46 #include <sys/mbuf.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
52 
53 #include <net/if.h>
54 #include <net/route.h>
55 #include <net/vnet.h>
56 
57 #include <netinet/cc.h>
58 #include <netinet/in.h>
59 #include <netinet/in_systm.h>
60 #include <netinet/ip.h>
61 #include <netinet/in_pcb.h>
62 #include <netinet/ip_var.h>
63 #include <netinet/ip_options.h>
64 #ifdef INET6
65 #include <netinet6/in6_pcb.h>
66 #include <netinet/ip6.h>
67 #include <netinet6/ip6_var.h>
68 #endif
69 #define	TCPOUTFLAGS
70 #include <netinet/tcp_fsm.h>
71 #include <netinet/tcp_seq.h>
72 #include <netinet/tcp_timer.h>
73 #include <netinet/tcp_var.h>
74 #include <netinet/tcpip.h>
75 #ifdef TCPDEBUG
76 #include <netinet/tcp_debug.h>
77 #endif
78 #ifdef TCP_OFFLOAD
79 #include <netinet/tcp_offload.h>
80 #endif
81 
82 #ifdef IPSEC
83 #include <netipsec/ipsec.h>
84 #endif /*IPSEC*/
85 
86 #include <machine/in_cksum.h>
87 
88 #include <security/mac/mac_framework.h>
89 
90 VNET_DEFINE(int, path_mtu_discovery) = 1;
91 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 	&VNET_NAME(path_mtu_discovery), 1,
93 	"Enable Path MTU Discovery");
94 
95 VNET_DEFINE(int, tcp_do_tso) = 1;
96 #define	V_tcp_do_tso		VNET(tcp_do_tso)
97 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
98 	&VNET_NAME(tcp_do_tso), 0,
99 	"Enable TCP Segmentation Offload");
100 
101 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
102 #define	V_tcp_sendspace	VNET(tcp_sendspace)
103 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
104 	&VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
105 
106 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
107 #define	V_tcp_do_autosndbuf	VNET(tcp_do_autosndbuf)
108 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
109 	&VNET_NAME(tcp_do_autosndbuf), 0,
110 	"Enable automatic send buffer sizing");
111 
112 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
113 #define	V_tcp_autosndbuf_inc	VNET(tcp_autosndbuf_inc)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
115 	&VNET_NAME(tcp_autosndbuf_inc), 0,
116 	"Incrementor step size of automatic send buffer");
117 
118 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
119 #define	V_tcp_autosndbuf_max	VNET(tcp_autosndbuf_max)
120 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
121 	&VNET_NAME(tcp_autosndbuf_max), 0,
122 	"Max size of automatic send buffer");
123 
124 static void inline	hhook_run_tcp_est_out(struct tcpcb *tp,
125 			    struct tcphdr *th, struct tcpopt *to,
126 			    long len, int tso);
127 static void inline	cc_after_idle(struct tcpcb *tp);
128 
129 /*
130  * Wrapper for the TCP established output helper hook.
131  */
132 static void inline
133 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
134     struct tcpopt *to, long len, int tso)
135 {
136 	struct tcp_hhook_data hhook_data;
137 
138 	if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
139 		hhook_data.tp = tp;
140 		hhook_data.th = th;
141 		hhook_data.to = to;
142 		hhook_data.len = len;
143 		hhook_data.tso = tso;
144 
145 		hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
146 		    tp->osd);
147 	}
148 }
149 
150 /*
151  * CC wrapper hook functions
152  */
153 static void inline
154 cc_after_idle(struct tcpcb *tp)
155 {
156 	INP_WLOCK_ASSERT(tp->t_inpcb);
157 
158 	if (CC_ALGO(tp)->after_idle != NULL)
159 		CC_ALGO(tp)->after_idle(tp->ccv);
160 }
161 
162 /*
163  * Tcp output routine: figure out what should be sent and send it.
164  */
165 int
166 tcp_output(struct tcpcb *tp)
167 {
168 	struct socket *so = tp->t_inpcb->inp_socket;
169 	long len, recwin, sendwin;
170 	int off, flags, error = 0;	/* Keep compiler happy */
171 	struct mbuf *m;
172 	struct ip *ip = NULL;
173 	struct ipovly *ipov = NULL;
174 	struct tcphdr *th;
175 	u_char opt[TCP_MAXOLEN];
176 	unsigned ipoptlen, optlen, hdrlen;
177 #ifdef IPSEC
178 	unsigned ipsec_optlen = 0;
179 #endif
180 	int idle, sendalot;
181 	int sack_rxmit, sack_bytes_rxmt;
182 	struct sackhole *p;
183 	int tso, mtu;
184 	struct tcpopt to;
185 #if 0
186 	int maxburst = TCP_MAXBURST;
187 #endif
188 #ifdef INET6
189 	struct ip6_hdr *ip6 = NULL;
190 	int isipv6;
191 
192 	isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
193 #endif
194 
195 	INP_WLOCK_ASSERT(tp->t_inpcb);
196 
197 #ifdef TCP_OFFLOAD
198 	if (tp->t_flags & TF_TOE)
199 		return (tcp_offload_output(tp));
200 #endif
201 
202 	/*
203 	 * Determine length of data that should be transmitted,
204 	 * and flags that will be used.
205 	 * If there is some data or critical controls (SYN, RST)
206 	 * to send, then transmit; otherwise, investigate further.
207 	 */
208 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
209 	if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
210 		cc_after_idle(tp);
211 	tp->t_flags &= ~TF_LASTIDLE;
212 	if (idle) {
213 		if (tp->t_flags & TF_MORETOCOME) {
214 			tp->t_flags |= TF_LASTIDLE;
215 			idle = 0;
216 		}
217 	}
218 again:
219 	/*
220 	 * If we've recently taken a timeout, snd_max will be greater than
221 	 * snd_nxt.  There may be SACK information that allows us to avoid
222 	 * resending already delivered data.  Adjust snd_nxt accordingly.
223 	 */
224 	if ((tp->t_flags & TF_SACK_PERMIT) &&
225 	    SEQ_LT(tp->snd_nxt, tp->snd_max))
226 		tcp_sack_adjust(tp);
227 	sendalot = 0;
228 	tso = 0;
229 	mtu = 0;
230 	off = tp->snd_nxt - tp->snd_una;
231 	sendwin = min(tp->snd_wnd, tp->snd_cwnd);
232 
233 	flags = tcp_outflags[tp->t_state];
234 	/*
235 	 * Send any SACK-generated retransmissions.  If we're explicitly trying
236 	 * to send out new data (when sendalot is 1), bypass this function.
237 	 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
238 	 * we're replacing a (future) new transmission with a retransmission
239 	 * now, and we previously incremented snd_cwnd in tcp_input().
240 	 */
241 	/*
242 	 * Still in sack recovery , reset rxmit flag to zero.
243 	 */
244 	sack_rxmit = 0;
245 	sack_bytes_rxmt = 0;
246 	len = 0;
247 	p = NULL;
248 	if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
249 	    (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
250 		long cwin;
251 
252 		cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
253 		if (cwin < 0)
254 			cwin = 0;
255 		/* Do not retransmit SACK segments beyond snd_recover */
256 		if (SEQ_GT(p->end, tp->snd_recover)) {
257 			/*
258 			 * (At least) part of sack hole extends beyond
259 			 * snd_recover. Check to see if we can rexmit data
260 			 * for this hole.
261 			 */
262 			if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
263 				/*
264 				 * Can't rexmit any more data for this hole.
265 				 * That data will be rexmitted in the next
266 				 * sack recovery episode, when snd_recover
267 				 * moves past p->rxmit.
268 				 */
269 				p = NULL;
270 				goto after_sack_rexmit;
271 			} else
272 				/* Can rexmit part of the current hole */
273 				len = ((long)ulmin(cwin,
274 						   tp->snd_recover - p->rxmit));
275 		} else
276 			len = ((long)ulmin(cwin, p->end - p->rxmit));
277 		off = p->rxmit - tp->snd_una;
278 		KASSERT(off >= 0,("%s: sack block to the left of una : %d",
279 		    __func__, off));
280 		if (len > 0) {
281 			sack_rxmit = 1;
282 			sendalot = 1;
283 			TCPSTAT_INC(tcps_sack_rexmits);
284 			TCPSTAT_ADD(tcps_sack_rexmit_bytes,
285 			    min(len, tp->t_maxseg));
286 		}
287 	}
288 after_sack_rexmit:
289 	/*
290 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
291 	 * state flags.
292 	 */
293 	if (tp->t_flags & TF_NEEDFIN)
294 		flags |= TH_FIN;
295 	if (tp->t_flags & TF_NEEDSYN)
296 		flags |= TH_SYN;
297 
298 	SOCKBUF_LOCK(&so->so_snd);
299 	/*
300 	 * If in persist timeout with window of 0, send 1 byte.
301 	 * Otherwise, if window is small but nonzero
302 	 * and timer expired, we will send what we can
303 	 * and go to transmit state.
304 	 */
305 	if (tp->t_flags & TF_FORCEDATA) {
306 		if (sendwin == 0) {
307 			/*
308 			 * If we still have some data to send, then
309 			 * clear the FIN bit.  Usually this would
310 			 * happen below when it realizes that we
311 			 * aren't sending all the data.  However,
312 			 * if we have exactly 1 byte of unsent data,
313 			 * then it won't clear the FIN bit below,
314 			 * and if we are in persist state, we wind
315 			 * up sending the packet without recording
316 			 * that we sent the FIN bit.
317 			 *
318 			 * We can't just blindly clear the FIN bit,
319 			 * because if we don't have any more data
320 			 * to send then the probe will be the FIN
321 			 * itself.
322 			 */
323 			if (off < so->so_snd.sb_cc)
324 				flags &= ~TH_FIN;
325 			sendwin = 1;
326 		} else {
327 			tcp_timer_activate(tp, TT_PERSIST, 0);
328 			tp->t_rxtshift = 0;
329 		}
330 	}
331 
332 	/*
333 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
334 	 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
335 	 * a negative length.  This can also occur when TCP opens up
336 	 * its congestion window while receiving additional duplicate
337 	 * acks after fast-retransmit because TCP will reset snd_nxt
338 	 * to snd_max after the fast-retransmit.
339 	 *
340 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
341 	 * be set to snd_una, the offset will be 0, and the length may
342 	 * wind up 0.
343 	 *
344 	 * If sack_rxmit is true we are retransmitting from the scoreboard
345 	 * in which case len is already set.
346 	 */
347 	if (sack_rxmit == 0) {
348 		if (sack_bytes_rxmt == 0)
349 			len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
350 		else {
351 			long cwin;
352 
353                         /*
354 			 * We are inside of a SACK recovery episode and are
355 			 * sending new data, having retransmitted all the
356 			 * data possible in the scoreboard.
357 			 */
358 			len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
359 			       - off);
360 			/*
361 			 * Don't remove this (len > 0) check !
362 			 * We explicitly check for len > 0 here (although it
363 			 * isn't really necessary), to work around a gcc
364 			 * optimization issue - to force gcc to compute
365 			 * len above. Without this check, the computation
366 			 * of len is bungled by the optimizer.
367 			 */
368 			if (len > 0) {
369 				cwin = tp->snd_cwnd -
370 					(tp->snd_nxt - tp->sack_newdata) -
371 					sack_bytes_rxmt;
372 				if (cwin < 0)
373 					cwin = 0;
374 				len = lmin(len, cwin);
375 			}
376 		}
377 	}
378 
379 	/*
380 	 * Lop off SYN bit if it has already been sent.  However, if this
381 	 * is SYN-SENT state and if segment contains data and if we don't
382 	 * know that foreign host supports TAO, suppress sending segment.
383 	 */
384 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
385 		if (tp->t_state != TCPS_SYN_RECEIVED)
386 			flags &= ~TH_SYN;
387 		off--, len++;
388 	}
389 
390 	/*
391 	 * Be careful not to send data and/or FIN on SYN segments.
392 	 * This measure is needed to prevent interoperability problems
393 	 * with not fully conformant TCP implementations.
394 	 */
395 	if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
396 		len = 0;
397 		flags &= ~TH_FIN;
398 	}
399 
400 	if (len < 0) {
401 		/*
402 		 * If FIN has been sent but not acked,
403 		 * but we haven't been called to retransmit,
404 		 * len will be < 0.  Otherwise, window shrank
405 		 * after we sent into it.  If window shrank to 0,
406 		 * cancel pending retransmit, pull snd_nxt back
407 		 * to (closed) window, and set the persist timer
408 		 * if it isn't already going.  If the window didn't
409 		 * close completely, just wait for an ACK.
410 		 */
411 		len = 0;
412 		if (sendwin == 0) {
413 			tcp_timer_activate(tp, TT_REXMT, 0);
414 			tp->t_rxtshift = 0;
415 			tp->snd_nxt = tp->snd_una;
416 			if (!tcp_timer_active(tp, TT_PERSIST))
417 				tcp_setpersist(tp);
418 		}
419 	}
420 
421 	/* len will be >= 0 after this point. */
422 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
423 
424 	/*
425 	 * Automatic sizing of send socket buffer.  Often the send buffer
426 	 * size is not optimally adjusted to the actual network conditions
427 	 * at hand (delay bandwidth product).  Setting the buffer size too
428 	 * small limits throughput on links with high bandwidth and high
429 	 * delay (eg. trans-continental/oceanic links).  Setting the
430 	 * buffer size too big consumes too much real kernel memory,
431 	 * especially with many connections on busy servers.
432 	 *
433 	 * The criteria to step up the send buffer one notch are:
434 	 *  1. receive window of remote host is larger than send buffer
435 	 *     (with a fudge factor of 5/4th);
436 	 *  2. send buffer is filled to 7/8th with data (so we actually
437 	 *     have data to make use of it);
438 	 *  3. send buffer fill has not hit maximal automatic size;
439 	 *  4. our send window (slow start and cogestion controlled) is
440 	 *     larger than sent but unacknowledged data in send buffer.
441 	 *
442 	 * The remote host receive window scaling factor may limit the
443 	 * growing of the send buffer before it reaches its allowed
444 	 * maximum.
445 	 *
446 	 * It scales directly with slow start or congestion window
447 	 * and does at most one step per received ACK.  This fast
448 	 * scaling has the drawback of growing the send buffer beyond
449 	 * what is strictly necessary to make full use of a given
450 	 * delay*bandwith product.  However testing has shown this not
451 	 * to be much of an problem.  At worst we are trading wasting
452 	 * of available bandwith (the non-use of it) for wasting some
453 	 * socket buffer memory.
454 	 *
455 	 * TODO: Shrink send buffer during idle periods together
456 	 * with congestion window.  Requires another timer.  Has to
457 	 * wait for upcoming tcp timer rewrite.
458 	 */
459 	if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
460 		if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
461 		    so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
462 		    so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
463 		    sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
464 			if (!sbreserve_locked(&so->so_snd,
465 			    min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
466 			     V_tcp_autosndbuf_max), so, curthread))
467 				so->so_snd.sb_flags &= ~SB_AUTOSIZE;
468 		}
469 	}
470 
471 	/*
472 	 * Decide if we can use TCP Segmentation Offloading (if supported by
473 	 * hardware).
474 	 *
475 	 * TSO may only be used if we are in a pure bulk sending state.  The
476 	 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
477 	 * IP options prevent using TSO.  With TSO the TCP header is the same
478 	 * (except for the sequence number) for all generated packets.  This
479 	 * makes it impossible to transmit any options which vary per generated
480 	 * segment or packet.
481 	 */
482 #ifdef IPSEC
483 	/*
484 	 * Pre-calculate here as we save another lookup into the darknesses
485 	 * of IPsec that way and can actually decide if TSO is ok.
486 	 */
487 	ipsec_optlen = ipsec_hdrsiz_tcp(tp);
488 #endif
489 	if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
490 	    ((tp->t_flags & TF_SIGNATURE) == 0) &&
491 	    tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
492 #ifdef IPSEC
493 	    ipsec_optlen == 0 &&
494 #endif
495 	    tp->t_inpcb->inp_options == NULL &&
496 	    tp->t_inpcb->in6p_options == NULL)
497 		tso = 1;
498 
499 	if (sack_rxmit) {
500 		if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
501 			flags &= ~TH_FIN;
502 	} else {
503 		if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
504 			flags &= ~TH_FIN;
505 	}
506 
507 	recwin = sbspace(&so->so_rcv);
508 
509 	/*
510 	 * Sender silly window avoidance.   We transmit under the following
511 	 * conditions when len is non-zero:
512 	 *
513 	 *	- We have a full segment (or more with TSO)
514 	 *	- This is the last buffer in a write()/send() and we are
515 	 *	  either idle or running NODELAY
516 	 *	- we've timed out (e.g. persist timer)
517 	 *	- we have more then 1/2 the maximum send window's worth of
518 	 *	  data (receiver may be limited the window size)
519 	 *	- we need to retransmit
520 	 */
521 	if (len) {
522 		if (len >= tp->t_maxseg)
523 			goto send;
524 		/*
525 		 * NOTE! on localhost connections an 'ack' from the remote
526 		 * end may occur synchronously with the output and cause
527 		 * us to flush a buffer queued with moretocome.  XXX
528 		 *
529 		 * note: the len + off check is almost certainly unnecessary.
530 		 */
531 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
532 		    (idle || (tp->t_flags & TF_NODELAY)) &&
533 		    len + off >= so->so_snd.sb_cc &&
534 		    (tp->t_flags & TF_NOPUSH) == 0) {
535 			goto send;
536 		}
537 		if (tp->t_flags & TF_FORCEDATA)		/* typ. timeout case */
538 			goto send;
539 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
540 			goto send;
541 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
542 			goto send;
543 		if (sack_rxmit)
544 			goto send;
545 	}
546 
547 	/*
548 	 * Sending of standalone window updates.
549 	 *
550 	 * Window updates are important when we close our window due to a
551 	 * full socket buffer and are opening it again after the application
552 	 * reads data from it.  Once the window has opened again and the
553 	 * remote end starts to send again the ACK clock takes over and
554 	 * provides the most current window information.
555 	 *
556 	 * We must avoid the silly window syndrome whereas every read
557 	 * from the receive buffer, no matter how small, causes a window
558 	 * update to be sent.  We also should avoid sending a flurry of
559 	 * window updates when the socket buffer had queued a lot of data
560 	 * and the application is doing small reads.
561 	 *
562 	 * Prevent a flurry of pointless window updates by only sending
563 	 * an update when we can increase the advertized window by more
564 	 * than 1/4th of the socket buffer capacity.  When the buffer is
565 	 * getting full or is very small be more aggressive and send an
566 	 * update whenever we can increase by two mss sized segments.
567 	 * In all other situations the ACK's to new incoming data will
568 	 * carry further window increases.
569 	 *
570 	 * Don't send an independent window update if a delayed
571 	 * ACK is pending (it will get piggy-backed on it) or the
572 	 * remote side already has done a half-close and won't send
573 	 * more data.  Skip this if the connection is in T/TCP
574 	 * half-open state.
575 	 */
576 	if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
577 	    !(tp->t_flags & TF_DELACK) &&
578 	    !TCPS_HAVERCVDFIN(tp->t_state)) {
579 		/*
580 		 * "adv" is the amount we could increase the window,
581 		 * taking into account that we are limited by
582 		 * TCP_MAXWIN << tp->rcv_scale.
583 		 */
584 		long adv;
585 		int oldwin;
586 
587 		adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
588 		if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
589 			oldwin = (tp->rcv_adv - tp->rcv_nxt);
590 			adv -= oldwin;
591 		} else
592 			oldwin = 0;
593 
594 		/*
595 		 * If the new window size ends up being the same as the old
596 		 * size when it is scaled, then don't force a window update.
597 		 */
598 		if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
599 			goto dontupdate;
600 
601 		if (adv >= (long)(2 * tp->t_maxseg) &&
602 		    (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
603 		     recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
604 		     so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
605 			goto send;
606 	}
607 dontupdate:
608 
609 	/*
610 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
611 	 * is also a catch-all for the retransmit timer timeout case.
612 	 */
613 	if (tp->t_flags & TF_ACKNOW)
614 		goto send;
615 	if ((flags & TH_RST) ||
616 	    ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
617 		goto send;
618 	if (SEQ_GT(tp->snd_up, tp->snd_una))
619 		goto send;
620 	/*
621 	 * If our state indicates that FIN should be sent
622 	 * and we have not yet done so, then we need to send.
623 	 */
624 	if (flags & TH_FIN &&
625 	    ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
626 		goto send;
627 	/*
628 	 * In SACK, it is possible for tcp_output to fail to send a segment
629 	 * after the retransmission timer has been turned off.  Make sure
630 	 * that the retransmission timer is set.
631 	 */
632 	if ((tp->t_flags & TF_SACK_PERMIT) &&
633 	    SEQ_GT(tp->snd_max, tp->snd_una) &&
634 	    !tcp_timer_active(tp, TT_REXMT) &&
635 	    !tcp_timer_active(tp, TT_PERSIST)) {
636 		tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
637 		goto just_return;
638 	}
639 	/*
640 	 * TCP window updates are not reliable, rather a polling protocol
641 	 * using ``persist'' packets is used to insure receipt of window
642 	 * updates.  The three ``states'' for the output side are:
643 	 *	idle			not doing retransmits or persists
644 	 *	persisting		to move a small or zero window
645 	 *	(re)transmitting	and thereby not persisting
646 	 *
647 	 * tcp_timer_active(tp, TT_PERSIST)
648 	 *	is true when we are in persist state.
649 	 * (tp->t_flags & TF_FORCEDATA)
650 	 *	is set when we are called to send a persist packet.
651 	 * tcp_timer_active(tp, TT_REXMT)
652 	 *	is set when we are retransmitting
653 	 * The output side is idle when both timers are zero.
654 	 *
655 	 * If send window is too small, there is data to transmit, and no
656 	 * retransmit or persist is pending, then go to persist state.
657 	 * If nothing happens soon, send when timer expires:
658 	 * if window is nonzero, transmit what we can,
659 	 * otherwise force out a byte.
660 	 */
661 	if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
662 	    !tcp_timer_active(tp, TT_PERSIST)) {
663 		tp->t_rxtshift = 0;
664 		tcp_setpersist(tp);
665 	}
666 
667 	/*
668 	 * No reason to send a segment, just return.
669 	 */
670 just_return:
671 	SOCKBUF_UNLOCK(&so->so_snd);
672 	return (0);
673 
674 send:
675 	SOCKBUF_LOCK_ASSERT(&so->so_snd);
676 	/*
677 	 * Before ESTABLISHED, force sending of initial options
678 	 * unless TCP set not to do any options.
679 	 * NOTE: we assume that the IP/TCP header plus TCP options
680 	 * always fit in a single mbuf, leaving room for a maximum
681 	 * link header, i.e.
682 	 *	max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
683 	 */
684 	optlen = 0;
685 #ifdef INET6
686 	if (isipv6)
687 		hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
688 	else
689 #endif
690 		hdrlen = sizeof (struct tcpiphdr);
691 
692 	/*
693 	 * Compute options for segment.
694 	 * We only have to care about SYN and established connection
695 	 * segments.  Options for SYN-ACK segments are handled in TCP
696 	 * syncache.
697 	 */
698 	if ((tp->t_flags & TF_NOOPT) == 0) {
699 		to.to_flags = 0;
700 		/* Maximum segment size. */
701 		if (flags & TH_SYN) {
702 			tp->snd_nxt = tp->iss;
703 			to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
704 			to.to_flags |= TOF_MSS;
705 		}
706 		/* Window scaling. */
707 		if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
708 			to.to_wscale = tp->request_r_scale;
709 			to.to_flags |= TOF_SCALE;
710 		}
711 		/* Timestamps. */
712 		if ((tp->t_flags & TF_RCVD_TSTMP) ||
713 		    ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
714 			to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
715 			to.to_tsecr = tp->ts_recent;
716 			to.to_flags |= TOF_TS;
717 			/* Set receive buffer autosizing timestamp. */
718 			if (tp->rfbuf_ts == 0 &&
719 			    (so->so_rcv.sb_flags & SB_AUTOSIZE))
720 				tp->rfbuf_ts = tcp_ts_getticks();
721 		}
722 		/* Selective ACK's. */
723 		if (tp->t_flags & TF_SACK_PERMIT) {
724 			if (flags & TH_SYN)
725 				to.to_flags |= TOF_SACKPERM;
726 			else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
727 			    (tp->t_flags & TF_SACK_PERMIT) &&
728 			    tp->rcv_numsacks > 0) {
729 				to.to_flags |= TOF_SACK;
730 				to.to_nsacks = tp->rcv_numsacks;
731 				to.to_sacks = (u_char *)tp->sackblks;
732 			}
733 		}
734 #ifdef TCP_SIGNATURE
735 		/* TCP-MD5 (RFC2385). */
736 		if (tp->t_flags & TF_SIGNATURE)
737 			to.to_flags |= TOF_SIGNATURE;
738 #endif /* TCP_SIGNATURE */
739 
740 		/* Processing the options. */
741 		hdrlen += optlen = tcp_addoptions(&to, opt);
742 	}
743 
744 #ifdef INET6
745 	if (isipv6)
746 		ipoptlen = ip6_optlen(tp->t_inpcb);
747 	else
748 #endif
749 	if (tp->t_inpcb->inp_options)
750 		ipoptlen = tp->t_inpcb->inp_options->m_len -
751 				offsetof(struct ipoption, ipopt_list);
752 	else
753 		ipoptlen = 0;
754 #ifdef IPSEC
755 	ipoptlen += ipsec_optlen;
756 #endif
757 
758 	/*
759 	 * Adjust data length if insertion of options will
760 	 * bump the packet length beyond the t_maxopd length.
761 	 * Clear the FIN bit because we cut off the tail of
762 	 * the segment.
763 	 */
764 	if (len + optlen + ipoptlen > tp->t_maxopd) {
765 		flags &= ~TH_FIN;
766 
767 		if (tso) {
768 			KASSERT(ipoptlen == 0,
769 			    ("%s: TSO can't do IP options", __func__));
770 
771 			/*
772 			 * Limit a burst to IP_MAXPACKET minus IP,
773 			 * TCP and options length to keep ip->ip_len
774 			 * from overflowing.
775 			 */
776 			if (len > IP_MAXPACKET - hdrlen) {
777 				len = IP_MAXPACKET - hdrlen;
778 				sendalot = 1;
779 			}
780 
781 			/*
782 			 * Prevent the last segment from being
783 			 * fractional unless the send sockbuf can
784 			 * be emptied.
785 			 */
786 			if (sendalot && off + len < so->so_snd.sb_cc) {
787 				len -= len % (tp->t_maxopd - optlen);
788 				sendalot = 1;
789 			}
790 
791 			/*
792 			 * Send the FIN in a separate segment
793 			 * after the bulk sending is done.
794 			 * We don't trust the TSO implementations
795 			 * to clear the FIN flag on all but the
796 			 * last segment.
797 			 */
798 			if (tp->t_flags & TF_NEEDFIN)
799 				sendalot = 1;
800 
801 		} else {
802 			len = tp->t_maxopd - optlen - ipoptlen;
803 			sendalot = 1;
804 		}
805 	} else
806 		tso = 0;
807 
808 	KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
809 	    ("%s: len > IP_MAXPACKET", __func__));
810 
811 /*#ifdef DIAGNOSTIC*/
812 #ifdef INET6
813 	if (max_linkhdr + hdrlen > MCLBYTES)
814 #else
815 	if (max_linkhdr + hdrlen > MHLEN)
816 #endif
817 		panic("tcphdr too big");
818 /*#endif*/
819 
820 	/*
821 	 * This KASSERT is here to catch edge cases at a well defined place.
822 	 * Before, those had triggered (random) panic conditions further down.
823 	 */
824 	KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
825 
826 	/*
827 	 * Grab a header mbuf, attaching a copy of data to
828 	 * be transmitted, and initialize the header from
829 	 * the template for sends on this connection.
830 	 */
831 	if (len) {
832 		struct mbuf *mb;
833 		u_int moff;
834 
835 		if ((tp->t_flags & TF_FORCEDATA) && len == 1)
836 			TCPSTAT_INC(tcps_sndprobe);
837 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
838 			tp->t_sndrexmitpack++;
839 			TCPSTAT_INC(tcps_sndrexmitpack);
840 			TCPSTAT_ADD(tcps_sndrexmitbyte, len);
841 		} else {
842 			TCPSTAT_INC(tcps_sndpack);
843 			TCPSTAT_ADD(tcps_sndbyte, len);
844 		}
845 #ifdef INET6
846 		if (MHLEN < hdrlen + max_linkhdr)
847 			m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
848 		else
849 #endif
850 			m = m_gethdr(M_NOWAIT, MT_DATA);
851 
852 		if (m == NULL) {
853 			SOCKBUF_UNLOCK(&so->so_snd);
854 			error = ENOBUFS;
855 			sack_rxmit = 0;
856 			goto out;
857 		}
858 
859 		m->m_data += max_linkhdr;
860 		m->m_len = hdrlen;
861 
862 		/*
863 		 * Start the m_copy functions from the closest mbuf
864 		 * to the offset in the socket buffer chain.
865 		 */
866 		mb = sbsndptr(&so->so_snd, off, len, &moff);
867 
868 		if (len <= MHLEN - hdrlen - max_linkhdr) {
869 			m_copydata(mb, moff, (int)len,
870 			    mtod(m, caddr_t) + hdrlen);
871 			m->m_len += len;
872 		} else {
873 			m->m_next = m_copy(mb, moff, (int)len);
874 			if (m->m_next == NULL) {
875 				SOCKBUF_UNLOCK(&so->so_snd);
876 				(void) m_free(m);
877 				error = ENOBUFS;
878 				sack_rxmit = 0;
879 				goto out;
880 			}
881 		}
882 
883 		/*
884 		 * If we're sending everything we've got, set PUSH.
885 		 * (This will keep happy those implementations which only
886 		 * give data to the user when a buffer fills or
887 		 * a PUSH comes in.)
888 		 */
889 		if (off + len == so->so_snd.sb_cc)
890 			flags |= TH_PUSH;
891 		SOCKBUF_UNLOCK(&so->so_snd);
892 	} else {
893 		SOCKBUF_UNLOCK(&so->so_snd);
894 		if (tp->t_flags & TF_ACKNOW)
895 			TCPSTAT_INC(tcps_sndacks);
896 		else if (flags & (TH_SYN|TH_FIN|TH_RST))
897 			TCPSTAT_INC(tcps_sndctrl);
898 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
899 			TCPSTAT_INC(tcps_sndurg);
900 		else
901 			TCPSTAT_INC(tcps_sndwinup);
902 
903 		m = m_gethdr(M_NOWAIT, MT_DATA);
904 		if (m == NULL) {
905 			error = ENOBUFS;
906 			sack_rxmit = 0;
907 			goto out;
908 		}
909 #ifdef INET6
910 		if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
911 		    MHLEN >= hdrlen) {
912 			MH_ALIGN(m, hdrlen);
913 		} else
914 #endif
915 		m->m_data += max_linkhdr;
916 		m->m_len = hdrlen;
917 	}
918 	SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
919 	m->m_pkthdr.rcvif = (struct ifnet *)0;
920 #ifdef MAC
921 	mac_inpcb_create_mbuf(tp->t_inpcb, m);
922 #endif
923 #ifdef INET6
924 	if (isipv6) {
925 		ip6 = mtod(m, struct ip6_hdr *);
926 		th = (struct tcphdr *)(ip6 + 1);
927 		tcpip_fillheaders(tp->t_inpcb, ip6, th);
928 	} else
929 #endif /* INET6 */
930 	{
931 		ip = mtod(m, struct ip *);
932 		ipov = (struct ipovly *)ip;
933 		th = (struct tcphdr *)(ip + 1);
934 		tcpip_fillheaders(tp->t_inpcb, ip, th);
935 	}
936 
937 	/*
938 	 * Fill in fields, remembering maximum advertised
939 	 * window for use in delaying messages about window sizes.
940 	 * If resending a FIN, be sure not to use a new sequence number.
941 	 */
942 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
943 	    tp->snd_nxt == tp->snd_max)
944 		tp->snd_nxt--;
945 	/*
946 	 * If we are starting a connection, send ECN setup
947 	 * SYN packet. If we are on a retransmit, we may
948 	 * resend those bits a number of times as per
949 	 * RFC 3168.
950 	 */
951 	if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
952 		if (tp->t_rxtshift >= 1) {
953 			if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
954 				flags |= TH_ECE|TH_CWR;
955 		} else
956 			flags |= TH_ECE|TH_CWR;
957 	}
958 
959 	if (tp->t_state == TCPS_ESTABLISHED &&
960 	    (tp->t_flags & TF_ECN_PERMIT)) {
961 		/*
962 		 * If the peer has ECN, mark data packets with
963 		 * ECN capable transmission (ECT).
964 		 * Ignore pure ack packets, retransmissions and window probes.
965 		 */
966 		if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
967 		    !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
968 #ifdef INET6
969 			if (isipv6)
970 				ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
971 			else
972 #endif
973 				ip->ip_tos |= IPTOS_ECN_ECT0;
974 			TCPSTAT_INC(tcps_ecn_ect0);
975 		}
976 
977 		/*
978 		 * Reply with proper ECN notifications.
979 		 */
980 		if (tp->t_flags & TF_ECN_SND_CWR) {
981 			flags |= TH_CWR;
982 			tp->t_flags &= ~TF_ECN_SND_CWR;
983 		}
984 		if (tp->t_flags & TF_ECN_SND_ECE)
985 			flags |= TH_ECE;
986 	}
987 
988 	/*
989 	 * If we are doing retransmissions, then snd_nxt will
990 	 * not reflect the first unsent octet.  For ACK only
991 	 * packets, we do not want the sequence number of the
992 	 * retransmitted packet, we want the sequence number
993 	 * of the next unsent octet.  So, if there is no data
994 	 * (and no SYN or FIN), use snd_max instead of snd_nxt
995 	 * when filling in ti_seq.  But if we are in persist
996 	 * state, snd_max might reflect one byte beyond the
997 	 * right edge of the window, so use snd_nxt in that
998 	 * case, since we know we aren't doing a retransmission.
999 	 * (retransmit and persist are mutually exclusive...)
1000 	 */
1001 	if (sack_rxmit == 0) {
1002 		if (len || (flags & (TH_SYN|TH_FIN)) ||
1003 		    tcp_timer_active(tp, TT_PERSIST))
1004 			th->th_seq = htonl(tp->snd_nxt);
1005 		else
1006 			th->th_seq = htonl(tp->snd_max);
1007 	} else {
1008 		th->th_seq = htonl(p->rxmit);
1009 		p->rxmit += len;
1010 		tp->sackhint.sack_bytes_rexmit += len;
1011 	}
1012 	th->th_ack = htonl(tp->rcv_nxt);
1013 	if (optlen) {
1014 		bcopy(opt, th + 1, optlen);
1015 		th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1016 	}
1017 	th->th_flags = flags;
1018 	/*
1019 	 * Calculate receive window.  Don't shrink window,
1020 	 * but avoid silly window syndrome.
1021 	 */
1022 	if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1023 	    recwin < (long)tp->t_maxseg)
1024 		recwin = 0;
1025 	if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1026 	    recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1027 		recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1028 	if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1029 		recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1030 
1031 	/*
1032 	 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1033 	 * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1034 	 * case is handled in syncache.
1035 	 */
1036 	if (flags & TH_SYN)
1037 		th->th_win = htons((u_short)
1038 				(min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1039 	else
1040 		th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1041 
1042 	/*
1043 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1044 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1045 	 * flag tells soreceive() to disable delayed acknowledgements when
1046 	 * draining the buffer.  This can occur if the receiver is attempting
1047 	 * to read more data than can be buffered prior to transmitting on
1048 	 * the connection.
1049 	 */
1050 	if (th->th_win == 0) {
1051 		tp->t_sndzerowin++;
1052 		tp->t_flags |= TF_RXWIN0SENT;
1053 	} else
1054 		tp->t_flags &= ~TF_RXWIN0SENT;
1055 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1056 		th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1057 		th->th_flags |= TH_URG;
1058 	} else
1059 		/*
1060 		 * If no urgent pointer to send, then we pull
1061 		 * the urgent pointer to the left edge of the send window
1062 		 * so that it doesn't drift into the send window on sequence
1063 		 * number wraparound.
1064 		 */
1065 		tp->snd_up = tp->snd_una;		/* drag it along */
1066 
1067 #ifdef TCP_SIGNATURE
1068 	if (tp->t_flags & TF_SIGNATURE) {
1069 		int sigoff = to.to_signature - opt;
1070 		tcp_signature_compute(m, 0, len, optlen,
1071 		    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1072 	}
1073 #endif
1074 
1075 	/*
1076 	 * Put TCP length in extended header, and then
1077 	 * checksum extended header and data.
1078 	 */
1079 	m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1080 	m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1081 #ifdef INET6
1082 	if (isipv6) {
1083 		/*
1084 		 * ip6_plen is not need to be filled now, and will be filled
1085 		 * in ip6_output.
1086 		 */
1087 		m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1088 		th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1089 		    optlen + len, IPPROTO_TCP, 0);
1090 	}
1091 #endif
1092 #if defined(INET6) && defined(INET)
1093 	else
1094 #endif
1095 #ifdef INET
1096 	{
1097 		m->m_pkthdr.csum_flags = CSUM_TCP;
1098 		th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1099 		    htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1100 
1101 		/* IP version must be set here for ipv4/ipv6 checking later */
1102 		KASSERT(ip->ip_v == IPVERSION,
1103 		    ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1104 	}
1105 #endif
1106 
1107 	/*
1108 	 * Enable TSO and specify the size of the segments.
1109 	 * The TCP pseudo header checksum is always provided.
1110 	 * XXX: Fixme: This is currently not the case for IPv6.
1111 	 */
1112 	if (tso) {
1113 		KASSERT(len > tp->t_maxopd - optlen,
1114 		    ("%s: len <= tso_segsz", __func__));
1115 		m->m_pkthdr.csum_flags |= CSUM_TSO;
1116 		m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1117 	}
1118 
1119 #ifdef IPSEC
1120 	KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1121 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1122 	    __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1123 #else
1124 	KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1125 	    ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1126 	    __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1127 #endif
1128 
1129 	/* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1130 	hhook_run_tcp_est_out(tp, th, &to, len, tso);
1131 
1132 #ifdef TCPDEBUG
1133 	/*
1134 	 * Trace.
1135 	 */
1136 	if (so->so_options & SO_DEBUG) {
1137 		u_short save = 0;
1138 #ifdef INET6
1139 		if (!isipv6)
1140 #endif
1141 		{
1142 			save = ipov->ih_len;
1143 			ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1144 		}
1145 		tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1146 #ifdef INET6
1147 		if (!isipv6)
1148 #endif
1149 		ipov->ih_len = save;
1150 	}
1151 #endif /* TCPDEBUG */
1152 
1153 	/*
1154 	 * Fill in IP length and desired time to live and
1155 	 * send to IP level.  There should be a better way
1156 	 * to handle ttl and tos; we could keep them in
1157 	 * the template, but need a way to checksum without them.
1158 	 */
1159 	/*
1160 	 * m->m_pkthdr.len should have been set before cksum calcuration,
1161 	 * because in6_cksum() need it.
1162 	 */
1163 #ifdef INET6
1164 	if (isipv6) {
1165 		struct route_in6 ro;
1166 
1167 		bzero(&ro, sizeof(ro));
1168 		/*
1169 		 * we separately set hoplimit for every segment, since the
1170 		 * user might want to change the value via setsockopt.
1171 		 * Also, desired default hop limit might be changed via
1172 		 * Neighbor Discovery.
1173 		 */
1174 		ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1175 
1176 		/* TODO: IPv6 IP6TOS_ECT bit on */
1177 		error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1178 		    ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1179 		    NULL, NULL, tp->t_inpcb);
1180 
1181 		if (error == EMSGSIZE && ro.ro_rt != NULL)
1182 			mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1183 		RO_RTFREE(&ro);
1184 	}
1185 #endif /* INET6 */
1186 #if defined(INET) && defined(INET6)
1187 	else
1188 #endif
1189 #ifdef INET
1190     {
1191 	struct route ro;
1192 
1193 	bzero(&ro, sizeof(ro));
1194 	ip->ip_len = htons(m->m_pkthdr.len);
1195 #ifdef INET6
1196 	if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1197 		ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1198 #endif /* INET6 */
1199 	/*
1200 	 * If we do path MTU discovery, then we set DF on every packet.
1201 	 * This might not be the best thing to do according to RFC3390
1202 	 * Section 2. However the tcp hostcache migitates the problem
1203 	 * so it affects only the first tcp connection with a host.
1204 	 *
1205 	 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1206 	 */
1207 	if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1208 		ip->ip_off |= htons(IP_DF);
1209 
1210 	error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1211 	    ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1212 	    tp->t_inpcb);
1213 
1214 	if (error == EMSGSIZE && ro.ro_rt != NULL)
1215 		mtu = ro.ro_rt->rt_rmx.rmx_mtu;
1216 	RO_RTFREE(&ro);
1217     }
1218 #endif /* INET */
1219 
1220 out:
1221 	/*
1222 	 * In transmit state, time the transmission and arrange for
1223 	 * the retransmit.  In persist state, just set snd_max.
1224 	 */
1225 	if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1226 	    !tcp_timer_active(tp, TT_PERSIST)) {
1227 		tcp_seq startseq = tp->snd_nxt;
1228 
1229 		/*
1230 		 * Advance snd_nxt over sequence space of this segment.
1231 		 */
1232 		if (flags & (TH_SYN|TH_FIN)) {
1233 			if (flags & TH_SYN)
1234 				tp->snd_nxt++;
1235 			if (flags & TH_FIN) {
1236 				tp->snd_nxt++;
1237 				tp->t_flags |= TF_SENTFIN;
1238 			}
1239 		}
1240 		if (sack_rxmit)
1241 			goto timer;
1242 		tp->snd_nxt += len;
1243 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1244 			tp->snd_max = tp->snd_nxt;
1245 			/*
1246 			 * Time this transmission if not a retransmission and
1247 			 * not currently timing anything.
1248 			 */
1249 			if (tp->t_rtttime == 0) {
1250 				tp->t_rtttime = ticks;
1251 				tp->t_rtseq = startseq;
1252 				TCPSTAT_INC(tcps_segstimed);
1253 			}
1254 		}
1255 
1256 		/*
1257 		 * Set retransmit timer if not currently set,
1258 		 * and not doing a pure ack or a keep-alive probe.
1259 		 * Initial value for retransmit timer is smoothed
1260 		 * round-trip time + 2 * round-trip time variance.
1261 		 * Initialize shift counter which is used for backoff
1262 		 * of retransmit time.
1263 		 */
1264 timer:
1265 		if (!tcp_timer_active(tp, TT_REXMT) &&
1266 		    ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1267 		     (tp->snd_nxt != tp->snd_una))) {
1268 			if (tcp_timer_active(tp, TT_PERSIST)) {
1269 				tcp_timer_activate(tp, TT_PERSIST, 0);
1270 				tp->t_rxtshift = 0;
1271 			}
1272 			tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1273 		}
1274 	} else {
1275 		/*
1276 		 * Persist case, update snd_max but since we are in
1277 		 * persist mode (no window) we do not update snd_nxt.
1278 		 */
1279 		int xlen = len;
1280 		if (flags & TH_SYN)
1281 			++xlen;
1282 		if (flags & TH_FIN) {
1283 			++xlen;
1284 			tp->t_flags |= TF_SENTFIN;
1285 		}
1286 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1287 			tp->snd_max = tp->snd_nxt + len;
1288 	}
1289 
1290 	if (error) {
1291 
1292 		/*
1293 		 * We know that the packet was lost, so back out the
1294 		 * sequence number advance, if any.
1295 		 *
1296 		 * If the error is EPERM the packet got blocked by the
1297 		 * local firewall.  Normally we should terminate the
1298 		 * connection but the blocking may have been spurious
1299 		 * due to a firewall reconfiguration cycle.  So we treat
1300 		 * it like a packet loss and let the retransmit timer and
1301 		 * timeouts do their work over time.
1302 		 * XXX: It is a POLA question whether calling tcp_drop right
1303 		 * away would be the really correct behavior instead.
1304 		 */
1305 		if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1306 		    !tcp_timer_active(tp, TT_PERSIST)) &&
1307 		    ((flags & TH_SYN) == 0) &&
1308 		    (error != EPERM)) {
1309 			if (sack_rxmit) {
1310 				p->rxmit -= len;
1311 				tp->sackhint.sack_bytes_rexmit -= len;
1312 				KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1313 				    ("sackhint bytes rtx >= 0"));
1314 			} else
1315 				tp->snd_nxt -= len;
1316 		}
1317 		SOCKBUF_UNLOCK_ASSERT(&so->so_snd);	/* Check gotos. */
1318 		switch (error) {
1319 		case EPERM:
1320 			tp->t_softerror = error;
1321 			return (error);
1322 		case ENOBUFS:
1323 	                if (!tcp_timer_active(tp, TT_REXMT) &&
1324 			    !tcp_timer_active(tp, TT_PERSIST))
1325 	                        tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1326 			tp->snd_cwnd = tp->t_maxseg;
1327 			return (0);
1328 		case EMSGSIZE:
1329 			/*
1330 			 * For some reason the interface we used initially
1331 			 * to send segments changed to another or lowered
1332 			 * its MTU.
1333 			 * If TSO was active we either got an interface
1334 			 * without TSO capabilits or TSO was turned off.
1335 			 * If we obtained mtu from ip_output() then update
1336 			 * it and try again.
1337 			 */
1338 			if (tso)
1339 				tp->t_flags &= ~TF_TSO;
1340 			if (mtu != 0) {
1341 				tcp_mss_update(tp, -1, mtu, NULL, NULL);
1342 				goto again;
1343 			}
1344 			return (error);
1345 		case EHOSTDOWN:
1346 		case EHOSTUNREACH:
1347 		case ENETDOWN:
1348 		case ENETUNREACH:
1349 			if (TCPS_HAVERCVDSYN(tp->t_state)) {
1350 				tp->t_softerror = error;
1351 				return (0);
1352 			}
1353 			/* FALLTHROUGH */
1354 		default:
1355 			return (error);
1356 		}
1357 	}
1358 	TCPSTAT_INC(tcps_sndtotal);
1359 
1360 	/*
1361 	 * Data sent (as far as we can tell).
1362 	 * If this advertises a larger window than any other segment,
1363 	 * then remember the size of the advertised window.
1364 	 * Any pending ACK has now been sent.
1365 	 */
1366 	if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1367 		tp->rcv_adv = tp->rcv_nxt + recwin;
1368 	tp->last_ack_sent = tp->rcv_nxt;
1369 	tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1370 	if (tcp_timer_active(tp, TT_DELACK))
1371 		tcp_timer_activate(tp, TT_DELACK, 0);
1372 #if 0
1373 	/*
1374 	 * This completely breaks TCP if newreno is turned on.  What happens
1375 	 * is that if delayed-acks are turned on on the receiver, this code
1376 	 * on the transmitter effectively destroys the TCP window, forcing
1377 	 * it to four packets (1.5Kx4 = 6K window).
1378 	 */
1379 	if (sendalot && --maxburst)
1380 		goto again;
1381 #endif
1382 	if (sendalot)
1383 		goto again;
1384 	return (0);
1385 }
1386 
1387 void
1388 tcp_setpersist(struct tcpcb *tp)
1389 {
1390 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1391 	int tt;
1392 
1393 	tp->t_flags &= ~TF_PREVVALID;
1394 	if (tcp_timer_active(tp, TT_REXMT))
1395 		panic("tcp_setpersist: retransmit pending");
1396 	/*
1397 	 * Start/restart persistance timer.
1398 	 */
1399 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1400 		      TCPTV_PERSMIN, TCPTV_PERSMAX);
1401 	tcp_timer_activate(tp, TT_PERSIST, tt);
1402 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1403 		tp->t_rxtshift++;
1404 }
1405 
1406 /*
1407  * Insert TCP options according to the supplied parameters to the place
1408  * optp in a consistent way.  Can handle unaligned destinations.
1409  *
1410  * The order of the option processing is crucial for optimal packing and
1411  * alignment for the scarce option space.
1412  *
1413  * The optimal order for a SYN/SYN-ACK segment is:
1414  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1415  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1416  *
1417  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1418  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1419  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1420  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1421  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1422  */
1423 int
1424 tcp_addoptions(struct tcpopt *to, u_char *optp)
1425 {
1426 	u_int mask, optlen = 0;
1427 
1428 	for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1429 		if ((to->to_flags & mask) != mask)
1430 			continue;
1431 		if (optlen == TCP_MAXOLEN)
1432 			break;
1433 		switch (to->to_flags & mask) {
1434 		case TOF_MSS:
1435 			while (optlen % 4) {
1436 				optlen += TCPOLEN_NOP;
1437 				*optp++ = TCPOPT_NOP;
1438 			}
1439 			if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1440 				continue;
1441 			optlen += TCPOLEN_MAXSEG;
1442 			*optp++ = TCPOPT_MAXSEG;
1443 			*optp++ = TCPOLEN_MAXSEG;
1444 			to->to_mss = htons(to->to_mss);
1445 			bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1446 			optp += sizeof(to->to_mss);
1447 			break;
1448 		case TOF_SCALE:
1449 			while (!optlen || optlen % 2 != 1) {
1450 				optlen += TCPOLEN_NOP;
1451 				*optp++ = TCPOPT_NOP;
1452 			}
1453 			if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1454 				continue;
1455 			optlen += TCPOLEN_WINDOW;
1456 			*optp++ = TCPOPT_WINDOW;
1457 			*optp++ = TCPOLEN_WINDOW;
1458 			*optp++ = to->to_wscale;
1459 			break;
1460 		case TOF_SACKPERM:
1461 			while (optlen % 2) {
1462 				optlen += TCPOLEN_NOP;
1463 				*optp++ = TCPOPT_NOP;
1464 			}
1465 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1466 				continue;
1467 			optlen += TCPOLEN_SACK_PERMITTED;
1468 			*optp++ = TCPOPT_SACK_PERMITTED;
1469 			*optp++ = TCPOLEN_SACK_PERMITTED;
1470 			break;
1471 		case TOF_TS:
1472 			while (!optlen || optlen % 4 != 2) {
1473 				optlen += TCPOLEN_NOP;
1474 				*optp++ = TCPOPT_NOP;
1475 			}
1476 			if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1477 				continue;
1478 			optlen += TCPOLEN_TIMESTAMP;
1479 			*optp++ = TCPOPT_TIMESTAMP;
1480 			*optp++ = TCPOLEN_TIMESTAMP;
1481 			to->to_tsval = htonl(to->to_tsval);
1482 			to->to_tsecr = htonl(to->to_tsecr);
1483 			bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1484 			optp += sizeof(to->to_tsval);
1485 			bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1486 			optp += sizeof(to->to_tsecr);
1487 			break;
1488 		case TOF_SIGNATURE:
1489 			{
1490 			int siglen = TCPOLEN_SIGNATURE - 2;
1491 
1492 			while (!optlen || optlen % 4 != 2) {
1493 				optlen += TCPOLEN_NOP;
1494 				*optp++ = TCPOPT_NOP;
1495 			}
1496 			if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1497 				continue;
1498 			optlen += TCPOLEN_SIGNATURE;
1499 			*optp++ = TCPOPT_SIGNATURE;
1500 			*optp++ = TCPOLEN_SIGNATURE;
1501 			to->to_signature = optp;
1502 			while (siglen--)
1503 				 *optp++ = 0;
1504 			break;
1505 			}
1506 		case TOF_SACK:
1507 			{
1508 			int sackblks = 0;
1509 			struct sackblk *sack = (struct sackblk *)to->to_sacks;
1510 			tcp_seq sack_seq;
1511 
1512 			while (!optlen || optlen % 4 != 2) {
1513 				optlen += TCPOLEN_NOP;
1514 				*optp++ = TCPOPT_NOP;
1515 			}
1516 			if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1517 				continue;
1518 			optlen += TCPOLEN_SACKHDR;
1519 			*optp++ = TCPOPT_SACK;
1520 			sackblks = min(to->to_nsacks,
1521 					(TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1522 			*optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1523 			while (sackblks--) {
1524 				sack_seq = htonl(sack->start);
1525 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1526 				optp += sizeof(sack_seq);
1527 				sack_seq = htonl(sack->end);
1528 				bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1529 				optp += sizeof(sack_seq);
1530 				optlen += TCPOLEN_SACK;
1531 				sack++;
1532 			}
1533 			TCPSTAT_INC(tcps_sack_send_blocks);
1534 			break;
1535 			}
1536 		default:
1537 			panic("%s: unknown TCP option type", __func__);
1538 			break;
1539 		}
1540 	}
1541 
1542 	/* Terminate and pad TCP options to a 4 byte boundary. */
1543 	if (optlen % 4) {
1544 		optlen += TCPOLEN_EOL;
1545 		*optp++ = TCPOPT_EOL;
1546 	}
1547 	/*
1548 	 * According to RFC 793 (STD0007):
1549 	 *   "The content of the header beyond the End-of-Option option
1550 	 *    must be header padding (i.e., zero)."
1551 	 *   and later: "The padding is composed of zeros."
1552 	 */
1553 	while (optlen % 4) {
1554 		optlen += TCPOLEN_PAD;
1555 		*optp++ = TCPOPT_PAD;
1556 	}
1557 
1558 	KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1559 	return (optlen);
1560 }
1561