xref: /dragonfly/sys/netinet/tcp_output.c (revision 8bf5b238)
1 /*
2  * Copyright (c) 2004 Jeffrey M. Hsu.  All rights reserved.
3  * Copyright (c) 2004 The DragonFly Project.  All rights reserved.
4  *
5  * This code is derived from software contributed to The DragonFly Project
6  * by Jeffrey M. Hsu.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  * 3. Neither the name of The DragonFly Project nor the names of its
17  *    contributors may be used to endorse or promote products derived
18  *    from this software without specific, prior written permission.
19  *
20  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
21  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
22  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
23  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
24  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
25  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
26  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
27  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
28  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
29  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
30  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  */
33 
34 /*
35  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36  *	The Regents of the University of California.  All rights reserved.
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. Neither the name of the University nor the names of its contributors
47  *    may be used to endorse or promote products derived from this software
48  *    without specific prior written permission.
49  *
50  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
60  * SUCH DAMAGE.
61  *
62  *	@(#)tcp_output.c	8.4 (Berkeley) 5/24/95
63  * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
64  */
65 
66 #include "opt_inet.h"
67 #include "opt_inet6.h"
68 #include "opt_tcpdebug.h"
69 
70 #include <sys/param.h>
71 #include <sys/systm.h>
72 #include <sys/kernel.h>
73 #include <sys/sysctl.h>
74 #include <sys/mbuf.h>
75 #include <sys/domain.h>
76 #include <sys/protosw.h>
77 #include <sys/socket.h>
78 #include <sys/socketvar.h>
79 #include <sys/in_cksum.h>
80 #include <sys/thread.h>
81 #include <sys/globaldata.h>
82 
83 #include <net/if_var.h>
84 #include <net/route.h>
85 #include <net/netmsg2.h>
86 #include <net/netisr2.h>
87 
88 #include <netinet/in.h>
89 #include <netinet/in_systm.h>
90 #include <netinet/ip.h>
91 #include <netinet/in_pcb.h>
92 #include <netinet/ip_var.h>
93 #include <netinet6/in6_pcb.h>
94 #include <netinet/ip6.h>
95 #include <netinet6/ip6_var.h>
96 #include <netinet/tcp.h>
97 #define	TCPOUTFLAGS
98 #include <netinet/tcp_fsm.h>
99 #include <netinet/tcp_seq.h>
100 #include <netinet/tcp_timer.h>
101 #include <netinet/tcp_timer2.h>
102 #include <netinet/tcp_var.h>
103 #include <netinet/tcpip.h>
104 #ifdef TCPDEBUG
105 #include <netinet/tcp_debug.h>
106 #endif
107 
108 #ifdef notyet
109 extern struct mbuf *m_copypack();
110 #endif
111 
112 int path_mtu_discovery = 1;
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
114 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
115 
116 static int avoid_pure_win_update = 1;
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
118 	&avoid_pure_win_update, 1, "Avoid pure window updates when possible");
119 
120 /*
121  * 1 - enabled for increasing and decreasing the buffer size
122  * 2 - enabled only for increasing the buffer size
123  */
124 int tcp_do_autosndbuf = 1;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
126     &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
127 
128 int tcp_autosndbuf_inc = 8*1024;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
130     &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
131 
132 int tcp_autosndbuf_min = 32768;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_min, CTLFLAG_RW,
134     &tcp_autosndbuf_min, 0, "Min size of automatic send buffer");
135 
136 int tcp_autosndbuf_max = 2*1024*1024;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
138     &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
139 
140 int tcp_prio_synack = 1;
141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, prio_synack, CTLFLAG_RW,
142     &tcp_prio_synack, 0, "Prioritize SYN, SYN|ACK and pure ACK");
143 
144 static int tcp_idle_cwv = 1;
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
146     &tcp_idle_cwv, 0,
147     "Congestion window validation after idle period (part of RFC2861)");
148 
149 static int tcp_idle_restart = 1;
150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
151     &tcp_idle_restart, 0, "Reset congestion window after idle period");
152 
153 static int tcp_do_tso = 1;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
155     &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
156 
157 static int tcp_fairsend = 4;
158 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW,
159     &tcp_fairsend, 0,
160     "Amount of segments sent before yield to other senders or receivers");
161 
162 static void	tcp_idle_cwnd_validate(struct tcpcb *);
163 
164 static int	tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
165 static void	tcp_output_sched(struct tcpcb *tp);
166 
167 /*
168  * Tcp output routine: figure out what should be sent and send it.
169  */
170 int
171 tcp_output(struct tcpcb *tp)
172 {
173 	struct inpcb * const inp = tp->t_inpcb;
174 	struct socket *so = inp->inp_socket;
175 	long len, recvwin, sendwin;
176 	int nsacked = 0;
177 	int off, flags, error = 0;
178 #ifdef TCP_SIGNATURE
179 	int sigoff = 0;
180 #endif
181 	struct mbuf *m;
182 	struct ip *ip;
183 	struct tcphdr *th;
184 	u_char opt[TCP_MAXOLEN];
185 	unsigned int ipoptlen, optlen, hdrlen;
186 	int idle;
187 	boolean_t sendalot;
188 	struct ip6_hdr *ip6;
189 #ifdef INET6
190 	const boolean_t isipv6 = INP_ISIPV6(inp);
191 #else
192 	const boolean_t isipv6 = FALSE;
193 #endif
194 	boolean_t can_tso = FALSE, use_tso;
195 	boolean_t report_sack, idle_cwv = FALSE;
196 	u_int segsz, tso_hlen, tso_lenmax = 0;
197 	int segcnt = 0;
198 	boolean_t need_sched = FALSE;
199 
200 	KKASSERT(so->so_port == &curthread->td_msgport);
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 
209 	/*
210 	 * If we have been idle for a while, the send congestion window
211 	 * could be no longer representative of the current state of the
212 	 * link; need to validate congestion window.  However, we should
213 	 * not perform congestion window validation here, since we could
214 	 * be asked to send pure ACK.
215 	 */
216 	if (tp->snd_max == tp->snd_una &&
217 	    (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
218 		idle_cwv = TRUE;
219 
220 	/*
221 	 * Calculate whether the transmit stream was previously idle
222 	 * and adjust TF_LASTIDLE for the next time.
223 	 */
224 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
225 	if (idle && (tp->t_flags & TF_MORETOCOME))
226 		tp->t_flags |= TF_LASTIDLE;
227 	else
228 		tp->t_flags &= ~TF_LASTIDLE;
229 
230 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
231 	    !IN_FASTRECOVERY(tp))
232 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
233 
234 	/*
235 	 * Find out whether TSO could be used or not
236 	 *
237 	 * For TSO capable devices, the following assumptions apply to
238 	 * the processing of TCP flags:
239 	 * - If FIN is set on the large TCP segment, the device must set
240 	 *   FIN on the last segment that it creates from the large TCP
241 	 *   segment.
242 	 * - If PUSH is set on the large TCP segment, the device must set
243 	 *   PUSH on the last segment that it creates from the large TCP
244 	 *   segment.
245 	 */
246 	if (tcp_do_tso
247 #ifdef TCP_SIGNATURE
248 	    && (tp->t_flags & TF_SIGNATURE) == 0
249 #endif
250 	) {
251 		if (!isipv6) {
252 			struct rtentry *rt = inp->inp_route.ro_rt;
253 
254 			if (rt != NULL && (rt->rt_flags & RTF_UP) &&
255 			    (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
256 				can_tso = TRUE;
257 				tso_lenmax = rt->rt_ifp->if_tsolen;
258 			}
259 		}
260 	}
261 
262 again:
263 	m = NULL;
264 	ip = NULL;
265 	th = NULL;
266 	ip6 = NULL;
267 
268 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
269 		TF_SACK_PERMITTED &&
270 	    (!TAILQ_EMPTY(&tp->t_segq) ||
271 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
272 		report_sack = TRUE;
273 	else
274 		report_sack = FALSE;
275 
276 	/* Make use of SACK information when slow-starting after a RTO. */
277 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
278 	    !IN_FASTRECOVERY(tp)) {
279 		tcp_seq old_snd_nxt = tp->snd_nxt;
280 
281 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
282 		nsacked += tp->snd_nxt - old_snd_nxt;
283 	}
284 
285 	sendalot = FALSE;
286 	off = tp->snd_nxt - tp->snd_una;
287 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
288 	sendwin = min(sendwin, tp->snd_bwnd);
289 
290 	flags = tcp_outflags[tp->t_state];
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 	/*
301 	 * If in persist timeout with window of 0, send 1 byte.
302 	 * Otherwise, if window is small but nonzero
303 	 * and timer expired, we will send what we can
304 	 * and go to transmit state.
305 	 */
306 	if (tp->t_flags & TF_FORCE) {
307 		if (sendwin == 0) {
308 			/*
309 			 * If we still have some data to send, then
310 			 * clear the FIN bit.  Usually this would
311 			 * happen below when it realizes that we
312 			 * aren't sending all the data.  However,
313 			 * if we have exactly 1 byte of unsent data,
314 			 * then it won't clear the FIN bit below,
315 			 * and if we are in persist state, we wind
316 			 * up sending the packet without recording
317 			 * that we sent the FIN bit.
318 			 *
319 			 * We can't just blindly clear the FIN bit,
320 			 * because if we don't have any more data
321 			 * to send then the probe will be the FIN
322 			 * itself.
323 			 */
324 			if (off < so->so_snd.ssb_cc)
325 				flags &= ~TH_FIN;
326 			sendwin = 1;
327 		} else {
328 			tcp_callout_stop(tp, tp->tt_persist);
329 			tp->t_rxtshift = 0;
330 		}
331 	}
332 
333 	/*
334 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
335 	 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
336 	 * a negative length.  This can also occur when TCP opens up
337 	 * its congestion window while receiving additional duplicate
338 	 * acks after fast-retransmit because TCP will reset snd_nxt
339 	 * to snd_max after the fast-retransmit.
340 	 *
341 	 * A negative length can also occur when we are in the
342 	 * TCPS_SYN_RECEIVED state due to a simultanious connect where
343 	 * our SYN has not been acked yet.
344 	 *
345 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
346 	 * be set to snd_una, the offset will be 0, and the length may
347 	 * wind up 0.
348 	 */
349 	len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
350 
351 	/*
352 	 * Lop off SYN bit if it has already been sent.  However, if this
353 	 * is SYN-SENT state and if segment contains data, suppress sending
354 	 * segment (sending the segment would be an option if we still
355 	 * did TAO and the remote host supported it).
356 	 */
357 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
358 		flags &= ~TH_SYN;
359 		off--, len++;
360 		if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
361 			tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
362 			return 0;
363 		}
364 	}
365 
366 	/*
367 	 * Be careful not to send data and/or FIN on SYN segments.
368 	 * This measure is needed to prevent interoperability problems
369 	 * with not fully conformant TCP implementations.
370 	 */
371 	if (flags & TH_SYN) {
372 		len = 0;
373 		flags &= ~TH_FIN;
374 	}
375 
376 	if (len < 0) {
377 		/*
378 		 * A negative len can occur if our FIN has been sent but not
379 		 * acked, or if we are in a simultanious connect in the
380 		 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
381 		 * acked.
382 		 *
383 		 * If our window has contracted to 0 in the FIN case
384 		 * (which can only occur if we have NOT been called to
385 		 * retransmit as per code a few paragraphs up) then we
386 		 * want to shift the retransmit timer over to the
387 		 * persist timer.
388 		 *
389 		 * However, if we are in the TCPS_SYN_RECEIVED state
390 		 * (the SYN case) we will be in a simultanious connect and
391 		 * the window may be zero degeneratively.  In this case we
392 		 * do not want to shift to the persist timer after the SYN
393 		 * or the SYN+ACK transmission.
394 		 */
395 		len = 0;
396 		if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
397 			tcp_callout_stop(tp, tp->tt_rexmt);
398 			tp->t_rxtshift = 0;
399 			tp->snd_nxt = tp->snd_una;
400 			if (!tcp_callout_active(tp, tp->tt_persist))
401 				tcp_setpersist(tp);
402 		}
403 	}
404 
405 	KASSERT(len >= 0, ("%s: len < 0", __func__));
406 	/*
407 	 * Automatic sizing of send socket buffer.  Often the send buffer
408 	 * size is not optimally adjusted to the actual network conditions
409 	 * at hand (delay bandwidth product).  Setting the buffer size too
410 	 * small limits throughput on links with high bandwidth and high
411 	 * delay (eg. trans-continental/oceanic links).  Setting the
412 	 * buffer size too big consumes too much real kernel memory,
413 	 * especially with many connections on busy servers.
414 	 *
415 	 * The criteria to step up the send buffer one notch are:
416 	 *  1. receive window of remote host is larger than send buffer
417 	 *     (with a fudge factor of 5/4th);
418 	 *  2. hiwat has not significantly exceeded bwnd (inflight)
419 	 *     (bwnd is a maximal value if inflight is disabled).
420 	 *  3. send buffer is filled to 7/8th with data (so we actually
421 	 *     have data to make use of it);
422 	 *  4. hiwat has not hit maximal automatic size;
423 	 *  5. our send window (slow start and cogestion controlled) is
424 	 *     larger than sent but unacknowledged data in send buffer.
425 	 *
426 	 * The remote host receive window scaling factor may limit the
427 	 * growing of the send buffer before it reaches its allowed
428 	 * maximum.
429 	 *
430 	 * It scales directly with slow start or congestion window
431 	 * and does at most one step per received ACK.  This fast
432 	 * scaling has the drawback of growing the send buffer beyond
433 	 * what is strictly necessary to make full use of a given
434 	 * delay*bandwith product.  However testing has shown this not
435 	 * to be much of an problem.  At worst we are trading wasting
436 	 * of available bandwith (the non-use of it) for wasting some
437 	 * socket buffer memory.
438 	 *
439 	 * The criteria for shrinking the buffer is based solely on
440 	 * the inflight code (snd_bwnd).  If inflight is disabled,
441 	 * the buffer will not be shrinked.  Note that snd_bwnd already
442 	 * has a fudge factor.  Our test adds a little hysteresis.
443 	 */
444 	if (tcp_do_autosndbuf && (so->so_snd.ssb_flags & SSB_AUTOSIZE)) {
445 		const int asbinc = tcp_autosndbuf_inc;
446 		const int hiwat = so->so_snd.ssb_hiwat;
447 		const int lowat = so->so_snd.ssb_lowat;
448 		u_long newsize;
449 
450 		if ((tp->snd_wnd / 4 * 5) >= hiwat &&
451 		    so->so_snd.ssb_cc >= (hiwat / 8 * 7) &&
452 		    hiwat < tp->snd_bwnd + hiwat / 10 &&
453 		    hiwat + asbinc < tcp_autosndbuf_max &&
454 		    hiwat < (TCP_MAXWIN << tp->snd_scale) &&
455 		    sendwin >= (so->so_snd.ssb_cc -
456 				(tp->snd_nxt - tp->snd_una))) {
457 			newsize = ulmin(hiwat + asbinc, tcp_autosndbuf_max);
458 			if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
459 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
460 #if 0
461 			if (newsize >= (TCP_MAXWIN << tp->snd_scale))
462 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
463 #endif
464 		} else if ((long)tp->snd_bwnd <
465 			   (long)(hiwat * 3 / 4 - lowat - asbinc) &&
466 			   hiwat > tp->t_maxseg * 2 + asbinc &&
467 			   hiwat + asbinc >= tcp_autosndbuf_min &&
468 			   tcp_do_autosndbuf == 1) {
469 			newsize = ulmax(hiwat - asbinc, tp->t_maxseg * 2);
470 			ssb_reserve(&so->so_snd, newsize, so, NULL);
471 		}
472 	}
473 
474 	/*
475 	 * Don't use TSO, if:
476 	 * - Congestion window needs validation
477 	 * - There are SACK blocks to report
478 	 * - RST or SYN flags is set
479 	 * - URG will be set
480 	 *
481 	 * XXX
482 	 * Checking for SYN|RST looks overkill, just to be safe than sorry
483 	 */
484 	use_tso = can_tso;
485 	if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
486 		use_tso = FALSE;
487 	if (use_tso) {
488 		tcp_seq ugr_nxt = tp->snd_nxt;
489 
490 		if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
491 		    tp->snd_nxt == tp->snd_max)
492 			--ugr_nxt;
493 
494 		if (SEQ_GT(tp->snd_up, ugr_nxt))
495 			use_tso = FALSE;
496 	}
497 
498 	if (use_tso) {
499 		/*
500 		 * Find out segment size and header length for TSO
501 		 */
502 		error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
503 		if (error)
504 			use_tso = FALSE;
505 	}
506 	if (!use_tso) {
507 		segsz = tp->t_maxseg;
508 		tso_hlen = 0; /* not used */
509 	}
510 
511 	/*
512 	 * Truncate to the maximum segment length if not TSO, and ensure that
513 	 * FIN is removed if the length no longer contains the last data byte.
514 	 */
515 	if (len > segsz) {
516 		if (!use_tso) {
517 			len = segsz;
518 			++segcnt;
519 		} else {
520 			int nsegs;
521 
522 			if (__predict_false(tso_lenmax < segsz))
523 				tso_lenmax = segsz << 1;
524 
525 			/*
526 			 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
527 			 * thoff), and make sure that we send equal size
528 			 * transfers down the stack (rather than big-small-
529 			 * big-small-...).
530 			 */
531 			len = min(len, tso_lenmax);
532 			nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz;
533 			KKASSERT(nsegs > 0);
534 
535 			len = nsegs * segsz;
536 
537 			if (len <= segsz) {
538 				use_tso = FALSE;
539 				++segcnt;
540 			} else {
541 				segcnt += nsegs;
542 			}
543 		}
544 		sendalot = TRUE;
545 	} else {
546 		use_tso = FALSE;
547 		if (len > 0)
548 			++segcnt;
549 	}
550 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
551 		flags &= ~TH_FIN;
552 
553 	recvwin = ssb_space(&so->so_rcv);
554 
555 	/*
556 	 * Sender silly window avoidance.   We transmit under the following
557 	 * conditions when len is non-zero:
558 	 *
559 	 *	- We have a full segment
560 	 *	- This is the last buffer in a write()/send() and we are
561 	 *	  either idle or running NODELAY
562 	 *	- we've timed out (e.g. persist timer)
563 	 *	- we have more then 1/2 the maximum send window's worth of
564 	 *	  data (receiver may be limiting the window size)
565 	 *	- we need to retransmit
566 	 */
567 	if (len) {
568 		if (len >= segsz)
569 			goto send;
570 		/*
571 		 * NOTE! on localhost connections an 'ack' from the remote
572 		 * end may occur synchronously with the output and cause
573 		 * us to flush a buffer queued with moretocome.  XXX
574 		 *
575 		 * note: the len + off check is almost certainly unnecessary.
576 		 */
577 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
578 		    (idle || (tp->t_flags & TF_NODELAY)) &&
579 		    len + off >= so->so_snd.ssb_cc &&
580 		    !(tp->t_flags & TF_NOPUSH)) {
581 			goto send;
582 		}
583 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
584 			goto send;
585 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
586 			goto send;
587 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
588 			goto send;
589 		if (tp->t_flags & TF_XMITNOW)
590 			goto send;
591 	}
592 
593 	/*
594 	 * Compare available window to amount of window
595 	 * known to peer (as advertised window less
596 	 * next expected input).  If the difference is at least two
597 	 * max size segments, or at least 50% of the maximum possible
598 	 * window, then want to send a window update to peer.
599 	 */
600 	if (recvwin > 0) {
601 		/*
602 		 * "adv" is the amount we can increase the window,
603 		 * taking into account that we are limited by
604 		 * TCP_MAXWIN << tp->rcv_scale.
605 		 */
606 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
607 			(tp->rcv_adv - tp->rcv_nxt);
608 		long hiwat;
609 
610 		/*
611 		 * This ack case typically occurs when the user has drained
612 		 * the TCP socket buffer sufficiently to warrent an ack
613 		 * containing a 'pure window update'... that is, an ack that
614 		 * ONLY updates the tcp window.
615 		 *
616 		 * It is unclear why we would need to do a pure window update
617 		 * past 2 segments if we are going to do one at 1/2 the high
618 		 * water mark anyway, especially since under normal conditions
619 		 * the user program will drain the socket buffer quickly.
620 		 * The 2-segment pure window update will often add a large
621 		 * number of extra, unnecessary acks to the stream.
622 		 *
623 		 * avoid_pure_win_update now defaults to 1.
624 		 */
625 		if (avoid_pure_win_update == 0 ||
626 		    (tp->t_flags & TF_RXRESIZED)) {
627 			if (adv >= (long) (2 * segsz)) {
628 				goto send;
629 			}
630 		}
631 		hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
632 		if (hiwat > (long)so->so_rcv.ssb_hiwat)
633 			hiwat = (long)so->so_rcv.ssb_hiwat;
634 		if (adv >= hiwat / 2)
635 			goto send;
636 	}
637 
638 	/*
639 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
640 	 * is also a catch-all for the retransmit timer timeout case.
641 	 */
642 	if (tp->t_flags & TF_ACKNOW)
643 		goto send;
644 	if ((flags & TH_RST) ||
645 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
646 		goto send;
647 	if (SEQ_GT(tp->snd_up, tp->snd_una))
648 		goto send;
649 	/*
650 	 * If our state indicates that FIN should be sent
651 	 * and we have not yet done so, then we need to send.
652 	 */
653 	if ((flags & TH_FIN) &&
654 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
655 		goto send;
656 
657 	/*
658 	 * TCP window updates are not reliable, rather a polling protocol
659 	 * using ``persist'' packets is used to insure receipt of window
660 	 * updates.  The three ``states'' for the output side are:
661 	 *	idle			not doing retransmits or persists
662 	 *	persisting		to move a small or zero window
663 	 *	(re)transmitting	and thereby not persisting
664 	 *
665 	 * tcp_callout_active(tp, tp->tt_persist)
666 	 *	is true when we are in persist state.
667 	 * The TF_FORCE flag in tp->t_flags
668 	 *	is set when we are called to send a persist packet.
669 	 * tcp_callout_active(tp, tp->tt_rexmt)
670 	 *	is set when we are retransmitting
671 	 * The output side is idle when both timers are zero.
672 	 *
673 	 * If send window is too small, there is data to transmit, and no
674 	 * retransmit or persist is pending, then go to persist state.
675 	 *
676 	 * If nothing happens soon, send when timer expires:
677 	 * if window is nonzero, transmit what we can, otherwise force out
678 	 * a byte.
679 	 *
680 	 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
681 	 * with data pending.  This situation can occur during a
682 	 * simultanious connect.
683 	 */
684 	if (so->so_snd.ssb_cc > 0 &&
685 	    tp->t_state != TCPS_SYN_RECEIVED &&
686 	    !tcp_callout_active(tp, tp->tt_rexmt) &&
687 	    !tcp_callout_active(tp, tp->tt_persist)) {
688 		tp->t_rxtshift = 0;
689 		tcp_setpersist(tp);
690 	}
691 
692 	/*
693 	 * No reason to send a segment, just return.
694 	 */
695 	tp->t_flags &= ~TF_XMITNOW;
696 	return (0);
697 
698 send:
699 	if (need_sched && len > 0) {
700 		tcp_output_sched(tp);
701 		return 0;
702 	}
703 
704 	/*
705 	 * Before ESTABLISHED, force sending of initial options
706 	 * unless TCP set not to do any options.
707 	 * NOTE: we assume that the IP/TCP header plus TCP options
708 	 * always fit in a single mbuf, leaving room for a maximum
709 	 * link header, i.e.
710 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
711 	 */
712 	optlen = 0;
713 	if (isipv6)
714 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
715 	else
716 		hdrlen = sizeof(struct tcpiphdr);
717 	if (flags & TH_SYN) {
718 		tp->snd_nxt = tp->iss;
719 		if (!(tp->t_flags & TF_NOOPT)) {
720 			u_short mss;
721 
722 			opt[0] = TCPOPT_MAXSEG;
723 			opt[1] = TCPOLEN_MAXSEG;
724 			mss = htons((u_short) tcp_mssopt(tp));
725 			memcpy(opt + 2, &mss, sizeof mss);
726 			optlen = TCPOLEN_MAXSEG;
727 
728 			if ((tp->t_flags & TF_REQ_SCALE) &&
729 			    (!(flags & TH_ACK) ||
730 			     (tp->t_flags & TF_RCVD_SCALE))) {
731 				*((u_int32_t *)(opt + optlen)) = htonl(
732 					TCPOPT_NOP << 24 |
733 					TCPOPT_WINDOW << 16 |
734 					TCPOLEN_WINDOW << 8 |
735 					tp->request_r_scale);
736 				optlen += 4;
737 			}
738 
739 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
740 			    tp->t_flags & TF_SACK_PERMITTED) {
741 				uint32_t *lp = (uint32_t *)(opt + optlen);
742 
743 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
744 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
745 			}
746 		}
747 	}
748 
749 	/*
750 	 * Send a timestamp and echo-reply if this is a SYN and our side
751 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
752 	 * and our peer have sent timestamps in our SYN's.
753 	 */
754 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
755 	    !(flags & TH_RST) &&
756 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
757 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
758 
759 		/* Form timestamp option as shown in appendix A of RFC 1323. */
760 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
761 		*lp++ = htonl(ticks);
762 		*lp   = htonl(tp->ts_recent);
763 		optlen += TCPOLEN_TSTAMP_APPA;
764 	}
765 
766 	/* Set receive buffer autosizing timestamp. */
767 	if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
768 		tp->rfbuf_ts = ticks;
769 
770 	/*
771 	 * If this is a SACK connection and we have a block to report,
772 	 * fill in the SACK blocks in the TCP options.
773 	 */
774 	if (report_sack)
775 		tcp_sack_fill_report(tp, opt, &optlen);
776 
777 #ifdef TCP_SIGNATURE
778 	if (tp->t_flags & TF_SIGNATURE) {
779 		int i;
780 		u_char *bp;
781 		/*
782 		 * Initialize TCP-MD5 option (RFC2385)
783 		 */
784 		bp = (u_char *)opt + optlen;
785 		*bp++ = TCPOPT_SIGNATURE;
786 		*bp++ = TCPOLEN_SIGNATURE;
787 		sigoff = optlen + 2;
788 		for (i = 0; i < TCP_SIGLEN; i++)
789 			*bp++ = 0;
790 		optlen += TCPOLEN_SIGNATURE;
791 		/*
792 		 * Terminate options list and maintain 32-bit alignment.
793 		 */
794 		*bp++ = TCPOPT_NOP;
795 		*bp++ = TCPOPT_EOL;
796 		optlen += 2;
797 	}
798 #endif /* TCP_SIGNATURE */
799 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
800 	hdrlen += optlen;
801 
802 	if (isipv6) {
803 		ipoptlen = ip6_optlen(inp);
804 	} else {
805 		if (inp->inp_options) {
806 			ipoptlen = inp->inp_options->m_len -
807 			    offsetof(struct ipoption, ipopt_list);
808 		} else {
809 			ipoptlen = 0;
810 		}
811 	}
812 
813 	if (use_tso) {
814 		/* TSO segment length must be multiple of segment size */
815 		KASSERT(len >= (2 * segsz) && (len % segsz == 0),
816 		    ("invalid TSO len %ld, segsz %u", len, segsz));
817 	} else {
818 		KASSERT(len <= segsz,
819 		    ("invalid len %ld, segsz %u", len, segsz));
820 
821 		/*
822 		 * Adjust data length if insertion of options will bump
823 		 * the packet length beyond the t_maxopd length.  Clear
824 		 * FIN to prevent premature closure since there is still
825 		 * more data to send after this (now truncated) packet.
826 		 *
827 		 * If just the options do not fit we are in a no-win
828 		 * situation and we treat it as an unreachable host.
829 		 */
830 		if (len + optlen + ipoptlen > tp->t_maxopd) {
831 			if (tp->t_maxopd <= optlen + ipoptlen) {
832 				static time_t last_optlen_report;
833 
834 				if (last_optlen_report != time_uptime) {
835 					last_optlen_report = time_uptime;
836 					kprintf("tcpcb %p: MSS (%d) too "
837 					    "small to hold options!\n",
838 					    tp, tp->t_maxopd);
839 				}
840 				error = EHOSTUNREACH;
841 				goto out;
842 			} else {
843 				flags &= ~TH_FIN;
844 				len = tp->t_maxopd - optlen - ipoptlen;
845 				sendalot = TRUE;
846 			}
847 		}
848 	}
849 
850 #ifdef INET6
851 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
852 #else
853 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
854 #endif
855 
856 	/*
857 	 * Grab a header mbuf, attaching a copy of data to
858 	 * be transmitted, and initialize the header from
859 	 * the template for sends on this connection.
860 	 */
861 	if (len) {
862 		if ((tp->t_flags & TF_FORCE) && len == 1)
863 			tcpstat.tcps_sndprobe++;
864 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
865 			if (tp->snd_nxt == tp->snd_una)
866 				tp->snd_max_rexmt = tp->snd_max;
867 			if (nsacked) {
868 				tcpstat.tcps_sndsackrtopack++;
869 				tcpstat.tcps_sndsackrtobyte += len;
870 			}
871 			tcpstat.tcps_sndrexmitpack++;
872 			tcpstat.tcps_sndrexmitbyte += len;
873 		} else {
874 			tcpstat.tcps_sndpack++;
875 			tcpstat.tcps_sndbyte += len;
876 		}
877 		if (idle_cwv) {
878 			idle_cwv = FALSE;
879 			tcp_idle_cwnd_validate(tp);
880 		}
881 		/* Update last send time after CWV */
882 		tp->snd_last = ticks;
883 #ifdef notyet
884 		if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
885 		    max_linkhdr + hdrlen)) == NULL) {
886 			error = ENOBUFS;
887 			goto after_th;
888 		}
889 		/*
890 		 * m_copypack left space for our hdr; use it.
891 		 */
892 		m->m_len += hdrlen;
893 		m->m_data -= hdrlen;
894 #else
895 #ifndef INET6
896 		m = m_gethdr(M_NOWAIT, MT_HEADER);
897 #else
898 		m = m_getl(hdrlen + max_linkhdr, M_NOWAIT, MT_HEADER,
899 			   M_PKTHDR, NULL);
900 #endif
901 		if (m == NULL) {
902 			error = ENOBUFS;
903 			goto after_th;
904 		}
905 		m->m_data += max_linkhdr;
906 		m->m_len = hdrlen;
907 		if (len <= MHLEN - hdrlen - max_linkhdr) {
908 			m_copydata(so->so_snd.ssb_mb, off, (int) len,
909 			    mtod(m, caddr_t) + hdrlen);
910 			m->m_len += len;
911 		} else {
912 			m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
913 			if (m->m_next == NULL) {
914 				m_free(m);
915 				m = NULL;
916 				error = ENOBUFS;
917 				goto after_th;
918 			}
919 		}
920 #endif
921 		/*
922 		 * If we're sending everything we've got, set PUSH.
923 		 * (This will keep happy those implementations which only
924 		 * give data to the user when a buffer fills or
925 		 * a PUSH comes in.)
926 		 */
927 		if (off + len == so->so_snd.ssb_cc)
928 			flags |= TH_PUSH;
929 	} else {
930 		if (tp->t_flags & TF_ACKNOW)
931 			tcpstat.tcps_sndacks++;
932 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
933 			tcpstat.tcps_sndctrl++;
934 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
935 			tcpstat.tcps_sndurg++;
936 		else
937 			tcpstat.tcps_sndwinup++;
938 
939 		MGETHDR(m, M_NOWAIT, MT_HEADER);
940 		if (m == NULL) {
941 			error = ENOBUFS;
942 			goto after_th;
943 		}
944 		if (isipv6 &&
945 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
946 			MH_ALIGN(m, hdrlen);
947 		else
948 			m->m_data += max_linkhdr;
949 		m->m_len = hdrlen;
950 
951 		/*
952 		 * Prioritize SYN, SYN|ACK and pure ACK.
953 		 * Leave FIN and RST as they are.
954 		 */
955 		if (tcp_prio_synack && (flags & (TH_FIN | TH_RST)) == 0)
956 			m->m_flags |= M_PRIO;
957 	}
958 	m->m_pkthdr.rcvif = NULL;
959 	if (isipv6) {
960 		ip6 = mtod(m, struct ip6_hdr *);
961 		th = (struct tcphdr *)(ip6 + 1);
962 		tcp_fillheaders(tp, ip6, th, use_tso);
963 	} else {
964 		ip = mtod(m, struct ip *);
965 		th = (struct tcphdr *)(ip + 1);
966 		/* this picks up the pseudo header (w/o the length) */
967 		tcp_fillheaders(tp, ip, th, use_tso);
968 	}
969 after_th:
970 	/*
971 	 * Fill in fields, remembering maximum advertised
972 	 * window for use in delaying messages about window sizes.
973 	 * If resending a FIN, be sure not to use a new sequence number.
974 	 */
975 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
976 	    tp->snd_nxt == tp->snd_max)
977 		tp->snd_nxt--;
978 
979 	if (th != NULL) {
980 		/*
981 		 * If we are doing retransmissions, then snd_nxt will
982 		 * not reflect the first unsent octet.  For ACK only
983 		 * packets, we do not want the sequence number of the
984 		 * retransmitted packet, we want the sequence number
985 		 * of the next unsent octet.  So, if there is no data
986 		 * (and no SYN or FIN), use snd_max instead of snd_nxt
987 		 * when filling in ti_seq.  But if we are in persist
988 		 * state, snd_max might reflect one byte beyond the
989 		 * right edge of the window, so use snd_nxt in that
990 		 * case, since we know we aren't doing a retransmission.
991 		 * (retransmit and persist are mutually exclusive...)
992 		 */
993 		if (len || (flags & (TH_SYN|TH_FIN)) ||
994 		    tcp_callout_active(tp, tp->tt_persist))
995 			th->th_seq = htonl(tp->snd_nxt);
996 		else
997 			th->th_seq = htonl(tp->snd_max);
998 		th->th_ack = htonl(tp->rcv_nxt);
999 		if (optlen) {
1000 			bcopy(opt, th + 1, optlen);
1001 			th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1002 		}
1003 		th->th_flags = flags;
1004 	}
1005 
1006 	/*
1007 	 * Calculate receive window.  Don't shrink window, but avoid
1008 	 * silly window syndrome by sending a 0 window if the actual
1009 	 * window is less then one segment.
1010 	 */
1011 	if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
1012 	    recvwin < (long)segsz)
1013 		recvwin = 0;
1014 	if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
1015 		recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
1016 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
1017 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
1018 
1019 	/*
1020 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1021 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1022 	 * flag tells soreceive() to disable delayed acknowledgements when
1023 	 * draining the buffer.  This can occur if the receiver is attempting
1024 	 * to read more data then can be buffered prior to transmitting on
1025 	 * the connection.
1026 	 */
1027 	if (recvwin == 0)
1028 		tp->t_flags |= TF_RXWIN0SENT;
1029 	else
1030 		tp->t_flags &= ~TF_RXWIN0SENT;
1031 
1032 	if (th != NULL)
1033 		th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
1034 
1035 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1036 		KASSERT(!use_tso, ("URG with TSO"));
1037 		if (th != NULL) {
1038 			th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1039 			th->th_flags |= TH_URG;
1040 		}
1041 	} else {
1042 		/*
1043 		 * If no urgent pointer to send, then we pull
1044 		 * the urgent pointer to the left edge of the send window
1045 		 * so that it doesn't drift into the send window on sequence
1046 		 * number wraparound.
1047 		 */
1048 		tp->snd_up = tp->snd_una;		/* drag it along */
1049 	}
1050 
1051 	if (th != NULL) {
1052 #ifdef TCP_SIGNATURE
1053 		if (tp->t_flags & TF_SIGNATURE) {
1054 			tcpsignature_compute(m, len, optlen,
1055 			    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1056 		}
1057 #endif /* TCP_SIGNATURE */
1058 
1059 		/*
1060 		 * Put TCP length in extended header, and then
1061 		 * checksum extended header and data.
1062 		 */
1063 		m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1064 		if (isipv6) {
1065 			/*
1066 			 * ip6_plen is not need to be filled now, and will be
1067 			 * filled in ip6_output().
1068 			 */
1069 			th->th_sum = in6_cksum(m, IPPROTO_TCP,
1070 			    sizeof(struct ip6_hdr),
1071 			    sizeof(struct tcphdr) + optlen + len);
1072 		} else {
1073 			m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1074 			if (use_tso) {
1075 				m->m_pkthdr.csum_flags = CSUM_TSO;
1076 				m->m_pkthdr.tso_segsz = segsz;
1077 			} else {
1078 				m->m_pkthdr.csum_flags = CSUM_TCP;
1079 				m->m_pkthdr.csum_data =
1080 				    offsetof(struct tcphdr, th_sum);
1081 				if (len + optlen) {
1082 					th->th_sum = in_addword(th->th_sum,
1083 					    htons((u_short)(optlen + len)));
1084 				}
1085 			}
1086 
1087 			/*
1088 			 * IP version must be set here for ipv4/ipv6 checking
1089 			 * later
1090 			 */
1091 			KASSERT(ip->ip_v == IPVERSION,
1092 			    ("%s: IP version incorrect: %d",
1093 			     __func__, ip->ip_v));
1094 		}
1095 	}
1096 
1097 	/*
1098 	 * In transmit state, time the transmission and arrange for
1099 	 * the retransmit.  In persist state, just set snd_max.
1100 	 */
1101 	if (!(tp->t_flags & TF_FORCE) ||
1102 	    !tcp_callout_active(tp, tp->tt_persist)) {
1103 		tcp_seq startseq = tp->snd_nxt;
1104 
1105 		/*
1106 		 * Advance snd_nxt over sequence space of this segment.
1107 		 */
1108 		if (flags & (TH_SYN | TH_FIN)) {
1109 			if (flags & TH_SYN)
1110 				tp->snd_nxt++;
1111 			if (flags & TH_FIN) {
1112 				tp->snd_nxt++;
1113 				tp->t_flags |= TF_SENTFIN;
1114 			}
1115 		}
1116 		tp->snd_nxt += len;
1117 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1118 			tp->snd_max = tp->snd_nxt;
1119 			/*
1120 			 * Time this transmission if not a retransmission and
1121 			 * not currently timing anything.
1122 			 */
1123 			if (tp->t_rtttime == 0) {
1124 				tp->t_rtttime = ticks;
1125 				tp->t_rtseq = startseq;
1126 				tcpstat.tcps_segstimed++;
1127 			}
1128 		}
1129 
1130 		/*
1131 		 * Set retransmit timer if not currently set,
1132 		 * and not doing a pure ack or a keep-alive probe.
1133 		 * Initial value for retransmit timer is smoothed
1134 		 * round-trip time + 2 * round-trip time variance.
1135 		 * Initialize shift counter which is used for backoff
1136 		 * of retransmit time.
1137 		 */
1138 		if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1139 		    tp->snd_nxt != tp->snd_una) {
1140 			if (tcp_callout_active(tp, tp->tt_persist)) {
1141 				tcp_callout_stop(tp, tp->tt_persist);
1142 				tp->t_rxtshift = 0;
1143 			}
1144 			tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1145 			    tcp_timer_rexmt);
1146 		} else if (len == 0 && so->so_snd.ssb_cc &&
1147 			   tp->t_state > TCPS_SYN_RECEIVED &&
1148 			   !tcp_callout_active(tp, tp->tt_rexmt) &&
1149 			   !tcp_callout_active(tp, tp->tt_persist)) {
1150 			/*
1151 			 * Avoid a situation where we do not set persist timer
1152 			 * after a zero window condition. For example:
1153 			 * 1) A -> B: packet with enough data to fill the window
1154 			 * 2) B -> A: ACK for #1 + new data (0 window
1155 			 *    advertisement)
1156 			 * 3) A -> B: ACK for #2, 0 len packet
1157 			 *
1158 			 * In this case, A will not activate the persist timer,
1159 			 * because it chose to send a packet. Unless tcp_output
1160 			 * is called for some other reason (delayed ack timer,
1161 			 * another input packet from B, socket syscall), A will
1162 			 * not send zero window probes.
1163 			 *
1164 			 * So, if you send a 0-length packet, but there is data
1165 			 * in the socket buffer, and neither the rexmt or
1166 			 * persist timer is already set, then activate the
1167 			 * persist timer.
1168 			 */
1169 			tp->t_rxtshift = 0;
1170 			tcp_setpersist(tp);
1171 		}
1172 	} else {
1173 		/*
1174 		 * Persist case, update snd_max but since we are in
1175 		 * persist mode (no window) we do not update snd_nxt.
1176 		 */
1177 		int xlen = len;
1178 		if (flags & TH_SYN)
1179 			panic("tcp_output: persist timer to send SYN");
1180 		if (flags & TH_FIN) {
1181 			++xlen;
1182 			tp->t_flags |= TF_SENTFIN;
1183 		}
1184 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1185 			tp->snd_max = tp->snd_nxt + xlen;
1186 	}
1187 
1188 	if (th != NULL) {
1189 #ifdef TCPDEBUG
1190 		/* Trace. */
1191 		if (so->so_options & SO_DEBUG) {
1192 			tcp_trace(TA_OUTPUT, tp->t_state, tp,
1193 			    mtod(m, void *), th, 0);
1194 		}
1195 #endif
1196 
1197 		/*
1198 		 * Fill in IP length and desired time to live and
1199 		 * send to IP level.  There should be a better way
1200 		 * to handle ttl and tos; we could keep them in
1201 		 * the template, but need a way to checksum without them.
1202 		 */
1203 		/*
1204 		 * m->m_pkthdr.len should have been set before cksum
1205 		 * calcuration, because in6_cksum() need it.
1206 		 */
1207 		if (isipv6) {
1208 			/*
1209 			 * we separately set hoplimit for every segment,
1210 			 * since the user might want to change the value
1211 			 * via setsockopt.  Also, desired default hop
1212 			 * limit might be changed via Neighbor Discovery.
1213 			 */
1214 			ip6->ip6_hlim = in6_selecthlim(inp,
1215 			    (inp->in6p_route.ro_rt ?
1216 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
1217 
1218 			/* TODO: IPv6 IP6TOS_ECT bit on */
1219 			error = ip6_output(m, inp->in6p_outputopts,
1220 			    &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1221 			    NULL, NULL, inp);
1222 		} else {
1223 			struct rtentry *rt;
1224 
1225 			KASSERT(!INP_CHECK_SOCKAF(so, AF_INET6), ("inet6 pcb"));
1226 
1227 			ip->ip_len = m->m_pkthdr.len;
1228 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
1229 			ip->ip_tos = inp->inp_ip_tos;	/* XXX */
1230 			/*
1231 			 * See if we should do MTU discovery.
1232 			 * We do it only if the following are true:
1233 			 *	1) we have a valid route to the destination
1234 			 *	2) the MTU is not locked (if it is,
1235 			 *	   then discovery has been disabled)
1236 			 */
1237 			if (path_mtu_discovery &&
1238 			    (rt = inp->inp_route.ro_rt) &&
1239 			    (rt->rt_flags & RTF_UP) &&
1240 			    !(rt->rt_rmx.rmx_locks & RTV_MTU))
1241 				ip->ip_off |= IP_DF;
1242 
1243 			KASSERT(inp->inp_flags & INP_HASH,
1244 			    ("inpcb has no hash"));
1245 			m_sethash(m, inp->inp_hashval);
1246 			error = ip_output(m, inp->inp_options, &inp->inp_route,
1247 					  (so->so_options & SO_DONTROUTE) |
1248 					  IP_DEBUGROUTE, NULL, inp);
1249 		}
1250 	} else {
1251 		KASSERT(error != 0, ("no error, but th not set"));
1252 	}
1253 	if (error) {
1254 		tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1255 
1256 		/*
1257 		 * We know that the packet was lost, so back out the
1258 		 * sequence number advance, if any.
1259 		 */
1260 		if (!(tp->t_flags & TF_FORCE) ||
1261 		    !tcp_callout_active(tp, tp->tt_persist)) {
1262 			/*
1263 			 * No need to check for TH_FIN here because
1264 			 * the TF_SENTFIN flag handles that case.
1265 			 */
1266 			if (!(flags & TH_SYN))
1267 				tp->snd_nxt -= len;
1268 		}
1269 
1270 out:
1271 		if (error == ENOBUFS) {
1272 			KASSERT((len == 0 && (flags & (TH_SYN | TH_FIN)) == 0) ||
1273 			    tcp_callout_active(tp, tp->tt_rexmt) ||
1274 			    tcp_callout_active(tp, tp->tt_persist),
1275 			    ("neither rexmt nor persist timer is set"));
1276 			return (0);
1277 		}
1278 		if (error == EMSGSIZE) {
1279 			/*
1280 			 * ip_output() will have already fixed the route
1281 			 * for us.  tcp_mtudisc() will, as its last action,
1282 			 * initiate retransmission, so it is important to
1283 			 * not do so here.
1284 			 */
1285 			tcp_mtudisc(inp, 0);
1286 			return 0;
1287 		}
1288 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1289 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1290 			tp->t_softerror = error;
1291 			return (0);
1292 		}
1293 		return (error);
1294 	}
1295 	tcpstat.tcps_sndtotal++;
1296 
1297 	/*
1298 	 * Data sent (as far as we can tell).
1299 	 *
1300 	 * If this advertises a larger window than any other segment,
1301 	 * then remember the size of the advertised window.
1302 	 *
1303 	 * Any pending ACK has now been sent.
1304 	 */
1305 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1306 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1307 		tp->t_flags &= ~TF_RXRESIZED;
1308 	}
1309 	tp->last_ack_sent = tp->rcv_nxt;
1310 	tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1311 	if (tcp_delack_enabled)
1312 		tcp_callout_stop(tp, tp->tt_delack);
1313 	if (sendalot) {
1314 		if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) &&
1315 		    segcnt >= tcp_fairsend)
1316 			need_sched = TRUE;
1317 		goto again;
1318 	}
1319 	return (0);
1320 }
1321 
1322 void
1323 tcp_setpersist(struct tcpcb *tp)
1324 {
1325 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1326 	int tt;
1327 
1328 	if (tp->t_state == TCPS_SYN_SENT ||
1329 	    tp->t_state == TCPS_SYN_RECEIVED) {
1330 		panic("tcp_setpersist: not established yet, current %s",
1331 		      tp->t_state == TCPS_SYN_SENT ?
1332 		      "SYN_SENT" : "SYN_RECEIVED");
1333 	}
1334 
1335 	if (tcp_callout_active(tp, tp->tt_rexmt))
1336 		panic("tcp_setpersist: retransmit pending");
1337 	/*
1338 	 * Start/restart persistance timer.
1339 	 */
1340 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1341 		      TCPTV_PERSMAX);
1342 	tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1343 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1344 		tp->t_rxtshift++;
1345 }
1346 
1347 static void
1348 tcp_idle_cwnd_validate(struct tcpcb *tp)
1349 {
1350 	u_long initial_cwnd = tcp_initial_window(tp);
1351 	u_long min_cwnd;
1352 
1353 	tcpstat.tcps_sndidle++;
1354 
1355 	/* According to RFC5681: RW=min(IW,cwnd) */
1356 	min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1357 
1358 	if (tcp_idle_cwv) {
1359 		u_long idle_time, decay_cwnd;
1360 
1361 		/*
1362 		 * RFC2861, but only after idle period.
1363 		 */
1364 
1365 		/*
1366 		 * Before the congestion window is reduced, ssthresh
1367 		 * is set to the maximum of its current value and 3/4
1368 		 * cwnd.  If the sender then has more data to send
1369 		 * than the decayed cwnd allows, the TCP will slow-
1370 		 * start (perform exponential increase) at least
1371 		 * half-way back up to the old value of cwnd.
1372 		 */
1373 		tp->snd_ssthresh = max(tp->snd_ssthresh,
1374 		    (3 * tp->snd_cwnd) / 4);
1375 
1376 		/*
1377 		 * Decay the congestion window by half for every RTT
1378 		 * that the flow remains inactive.
1379 		 *
1380 		 * The difference between our implementation and
1381 		 * RFC2861 is that we don't allow cwnd to go below
1382 		 * the value allowed by RFC5681 (min_cwnd).
1383 		 */
1384 		idle_time = ticks - tp->snd_last;
1385 		decay_cwnd = tp->snd_cwnd;
1386 		while (idle_time >= tp->t_rxtcur &&
1387 		    decay_cwnd > min_cwnd) {
1388 			decay_cwnd >>= 1;
1389 			idle_time -= tp->t_rxtcur;
1390 		}
1391 		tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1392 	} else {
1393 		/*
1394 		 * Slow-start from scratch to re-determine the send
1395 		 * congestion window.
1396 		 */
1397 		tp->snd_cwnd = min_cwnd;
1398 	}
1399 
1400 	/* Restart ABC counting during congestion avoidance */
1401 	tp->snd_wacked = 0;
1402 }
1403 
1404 static int
1405 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1406 {
1407 	struct inpcb * const inp = tp->t_inpcb;
1408 #ifdef INET6
1409 	const boolean_t isipv6 = INP_ISIPV6(inp);
1410 #else
1411 	const boolean_t isipv6 = FALSE;
1412 #endif
1413 	unsigned int ipoptlen, optlen;
1414 	u_int hlen;
1415 
1416 	hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1417 
1418 	if (isipv6) {
1419 		ipoptlen = ip6_optlen(inp);
1420 	} else {
1421 		if (inp->inp_options) {
1422 			ipoptlen = inp->inp_options->m_len -
1423 			    offsetof(struct ipoption, ipopt_list);
1424 		} else {
1425 			ipoptlen = 0;
1426 		}
1427 	}
1428 	hlen += ipoptlen;
1429 
1430 	optlen = 0;
1431 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1432 	    (tp->t_flags & TF_RCVD_TSTMP))
1433 		optlen += TCPOLEN_TSTAMP_APPA;
1434 	hlen += optlen;
1435 
1436 	if (tp->t_maxopd <= optlen + ipoptlen)
1437 		return EHOSTUNREACH;
1438 
1439 	*segsz = tp->t_maxopd - optlen - ipoptlen;
1440 	*hlen0 = hlen;
1441 	return 0;
1442 }
1443 
1444 static void
1445 tcp_output_sched_handler(netmsg_t nmsg)
1446 {
1447 	struct tcpcb *tp = nmsg->lmsg.u.ms_resultp;
1448 
1449 	/* Reply ASAP */
1450 	crit_enter();
1451 	lwkt_replymsg(&nmsg->lmsg, 0);
1452 	crit_exit();
1453 
1454 	tcp_output_fair(tp);
1455 }
1456 
1457 void
1458 tcp_output_init(struct tcpcb *tp)
1459 {
1460 	netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE,
1461 	    tcp_output_sched_handler);
1462 	tp->tt_sndmore->lmsg.u.ms_resultp = tp;
1463 }
1464 
1465 void
1466 tcp_output_cancel(struct tcpcb *tp)
1467 {
1468 	/*
1469 	 * This message is still pending to be processed;
1470 	 * drop it.  Optimized.
1471 	 */
1472 	crit_enter();
1473 	if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) {
1474 		lwkt_dropmsg(&tp->tt_sndmore->lmsg);
1475 	}
1476 	crit_exit();
1477 }
1478 
1479 boolean_t
1480 tcp_output_pending(struct tcpcb *tp)
1481 {
1482 	if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0)
1483 		return TRUE;
1484 	else
1485 		return FALSE;
1486 }
1487 
1488 static void
1489 tcp_output_sched(struct tcpcb *tp)
1490 {
1491 	crit_enter();
1492 	if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE)
1493 		lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg);
1494 	crit_exit();
1495 }
1496 
1497 /*
1498  * Fairsend
1499  *
1500  * Yield to other senders or receivers on the same netisr if the current
1501  * TCP stream has sent tcp_fairsend segments and is going to burst more
1502  * segments.  Bursting large amount of segements in a single TCP stream
1503  * could delay other senders' segments and receivers' ACKs quite a lot,
1504  * if others segments and ACKs are queued on to the same hardware transmit
1505  * queue; thus cause unfairness between senders and suppress receiving
1506  * performance.
1507  *
1508  * Fairsend should be performed at the places that do not affect segment
1509  * sending during congestion control, e.g.
1510  * - User requested output
1511  * - ACK input triggered output
1512  *
1513  * NOTE:
1514  * For devices that are TSO capable, their TSO aggregation size limit could
1515  * affect fairsend.
1516  */
1517 int
1518 tcp_output_fair(struct tcpcb *tp)
1519 {
1520 	int ret;
1521 
1522 	tp->t_flags |= TF_FAIRSEND;
1523 	ret = tcp_output(tp);
1524 	tp->t_flags &= ~TF_FAIRSEND;
1525 
1526 	return ret;
1527 }
1528