xref: /dragonfly/sys/netinet/tcp_output.c (revision 0de61e28)
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/malloc.h>	/* for M_NOWAIT */
74 #include <sys/sysctl.h>
75 #include <sys/mbuf.h>
76 #include <sys/domain.h>
77 #include <sys/protosw.h>
78 #include <sys/socket.h>
79 #include <sys/socketvar.h>
80 #include <sys/in_cksum.h>
81 #include <sys/thread.h>
82 #include <sys/globaldata.h>
83 
84 #include <net/if.h>
85 #include <net/if_var.h>
86 #include <net/route.h>
87 #include <net/netmsg2.h>
88 #include <net/netisr2.h>
89 
90 #include <netinet/in.h>
91 #include <netinet/in_systm.h>
92 #include <netinet/ip.h>
93 #include <netinet/in_pcb.h>
94 #include <netinet/ip_var.h>
95 #include <netinet6/in6_pcb.h>
96 #include <netinet/ip6.h>
97 #include <netinet6/ip6_var.h>
98 #include <netinet/tcp.h>
99 #define	TCPOUTFLAGS
100 #include <netinet/tcp_fsm.h>
101 #include <netinet/tcp_seq.h>
102 #include <netinet/tcp_timer.h>
103 #include <netinet/tcp_timer2.h>
104 #include <netinet/tcp_var.h>
105 #include <netinet/tcpip.h>
106 #ifdef TCPDEBUG
107 #include <netinet/tcp_debug.h>
108 #endif
109 
110 #ifdef notyet
111 extern struct mbuf *m_copypack();
112 #endif
113 
114 int path_mtu_discovery = 1;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
116 	&path_mtu_discovery, 1, "Enable Path MTU Discovery");
117 
118 static int avoid_pure_win_update = 1;
119 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
120 	&avoid_pure_win_update, 1, "Avoid pure window updates when possible");
121 
122 /*
123  * 1 - enabled for increasing and decreasing the buffer size
124  * 2 - enabled only for increasing the buffer size
125  */
126 int tcp_do_autosndbuf = 1;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
128     &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
129 
130 int tcp_autosndbuf_inc = 8*1024;
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
132     &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
133 
134 int tcp_autosndbuf_min = 32768;
135 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_min, CTLFLAG_RW,
136     &tcp_autosndbuf_min, 0, "Min size of automatic send buffer");
137 
138 int tcp_autosndbuf_max = 2*1024*1024;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
140     &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
141 
142 int tcp_prio_synack = 1;
143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, prio_synack, CTLFLAG_RW,
144     &tcp_prio_synack, 0, "Prioritize SYN, SYN|ACK and pure ACK");
145 
146 static int tcp_idle_cwv = 1;
147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
148     &tcp_idle_cwv, 0,
149     "Congestion window validation after idle period (part of RFC2861)");
150 
151 static int tcp_idle_restart = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
153     &tcp_idle_restart, 0, "Reset congestion window after idle period");
154 
155 static int tcp_do_tso = 1;
156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
157     &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
158 
159 static int tcp_fairsend = 4;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW,
161     &tcp_fairsend, 0,
162     "Amount of segments sent before yield to other senders or receivers");
163 
164 static void	tcp_idle_cwnd_validate(struct tcpcb *);
165 
166 static int	tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
167 static void	tcp_output_sched(struct tcpcb *tp);
168 
169 /*
170  * Tcp output routine: figure out what should be sent and send it.
171  */
172 int
173 tcp_output(struct tcpcb *tp)
174 {
175 	struct inpcb * const inp = tp->t_inpcb;
176 	struct socket *so = inp->inp_socket;
177 	long len, recvwin, sendwin;
178 	int nsacked = 0;
179 	int off, flags, error = 0;
180 #ifdef TCP_SIGNATURE
181 	int sigoff = 0;
182 #endif
183 	struct mbuf *m;
184 	struct ip *ip;
185 	struct tcphdr *th;
186 	u_char opt[TCP_MAXOLEN];
187 	unsigned int ipoptlen, optlen, hdrlen;
188 	int idle;
189 	boolean_t sendalot;
190 	struct ip6_hdr *ip6;
191 #ifdef INET6
192 	const boolean_t isipv6 = INP_ISIPV6(inp);
193 #else
194 	const boolean_t isipv6 = FALSE;
195 #endif
196 	boolean_t can_tso = FALSE, use_tso;
197 	boolean_t report_sack, idle_cwv = FALSE;
198 	u_int segsz, tso_hlen, tso_lenmax = 0;
199 	int segcnt = 0;
200 	boolean_t need_sched = FALSE;
201 
202 	KKASSERT(so->so_port == &curthread->td_msgport);
203 
204 	/*
205 	 * Determine length of data that should be transmitted,
206 	 * and flags that will be used.
207 	 * If there is some data or critical controls (SYN, RST)
208 	 * to send, then transmit; otherwise, investigate further.
209 	 */
210 
211 	/*
212 	 * If we have been idle for a while, the send congestion window
213 	 * could be no longer representative of the current state of the
214 	 * link; need to validate congestion window.  However, we should
215 	 * not perform congestion window validation here, since we could
216 	 * be asked to send pure ACK.
217 	 */
218 	if (tp->snd_max == tp->snd_una &&
219 	    (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
220 		idle_cwv = TRUE;
221 
222 	/*
223 	 * Calculate whether the transmit stream was previously idle
224 	 * and adjust TF_LASTIDLE for the next time.
225 	 */
226 	idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
227 	if (idle && (tp->t_flags & TF_MORETOCOME))
228 		tp->t_flags |= TF_LASTIDLE;
229 	else
230 		tp->t_flags &= ~TF_LASTIDLE;
231 
232 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
233 	    !IN_FASTRECOVERY(tp))
234 		nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
235 
236 	/*
237 	 * Find out whether TSO could be used or not
238 	 *
239 	 * For TSO capable devices, the following assumptions apply to
240 	 * the processing of TCP flags:
241 	 * - If FIN is set on the large TCP segment, the device must set
242 	 *   FIN on the last segment that it creates from the large TCP
243 	 *   segment.
244 	 * - If PUSH is set on the large TCP segment, the device must set
245 	 *   PUSH on the last segment that it creates from the large TCP
246 	 *   segment.
247 	 */
248 	if (tcp_do_tso
249 #ifdef TCP_SIGNATURE
250 	    && (tp->t_flags & TF_SIGNATURE) == 0
251 #endif
252 	) {
253 		if (!isipv6) {
254 			struct rtentry *rt = inp->inp_route.ro_rt;
255 
256 			if (rt != NULL && (rt->rt_flags & RTF_UP) &&
257 			    (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
258 				can_tso = TRUE;
259 				tso_lenmax = rt->rt_ifp->if_tsolen;
260 			}
261 		}
262 	}
263 
264 again:
265 	m = NULL;
266 	ip = NULL;
267 	th = NULL;
268 	ip6 = NULL;
269 
270 	if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
271 		TF_SACK_PERMITTED &&
272 	    (!TAILQ_EMPTY(&tp->t_segq) ||
273 	     tp->reportblk.rblk_start != tp->reportblk.rblk_end))
274 		report_sack = TRUE;
275 	else
276 		report_sack = FALSE;
277 
278 	/* Make use of SACK information when slow-starting after a RTO. */
279 	if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
280 	    !IN_FASTRECOVERY(tp)) {
281 		tcp_seq old_snd_nxt = tp->snd_nxt;
282 
283 		tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
284 		nsacked += tp->snd_nxt - old_snd_nxt;
285 	}
286 
287 	sendalot = FALSE;
288 	off = tp->snd_nxt - tp->snd_una;
289 	sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
290 	sendwin = min(sendwin, tp->snd_bwnd);
291 
292 	flags = tcp_outflags[tp->t_state];
293 	/*
294 	 * Get standard flags, and add SYN or FIN if requested by 'hidden'
295 	 * state flags.
296 	 */
297 	if (tp->t_flags & TF_NEEDFIN)
298 		flags |= TH_FIN;
299 	if (tp->t_flags & TF_NEEDSYN)
300 		flags |= TH_SYN;
301 
302 	/*
303 	 * If in persist timeout with window of 0, send 1 byte.
304 	 * Otherwise, if window is small but nonzero
305 	 * and timer expired, we will send what we can
306 	 * and go to transmit state.
307 	 */
308 	if (tp->t_flags & TF_FORCE) {
309 		if (sendwin == 0) {
310 			/*
311 			 * If we still have some data to send, then
312 			 * clear the FIN bit.  Usually this would
313 			 * happen below when it realizes that we
314 			 * aren't sending all the data.  However,
315 			 * if we have exactly 1 byte of unsent data,
316 			 * then it won't clear the FIN bit below,
317 			 * and if we are in persist state, we wind
318 			 * up sending the packet without recording
319 			 * that we sent the FIN bit.
320 			 *
321 			 * We can't just blindly clear the FIN bit,
322 			 * because if we don't have any more data
323 			 * to send then the probe will be the FIN
324 			 * itself.
325 			 */
326 			if (off < so->so_snd.ssb_cc)
327 				flags &= ~TH_FIN;
328 			sendwin = 1;
329 		} else {
330 			tcp_callout_stop(tp, tp->tt_persist);
331 			tp->t_rxtshift = 0;
332 		}
333 	}
334 
335 	/*
336 	 * If snd_nxt == snd_max and we have transmitted a FIN, the
337 	 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
338 	 * a negative length.  This can also occur when TCP opens up
339 	 * its congestion window while receiving additional duplicate
340 	 * acks after fast-retransmit because TCP will reset snd_nxt
341 	 * to snd_max after the fast-retransmit.
342 	 *
343 	 * A negative length can also occur when we are in the
344 	 * TCPS_SYN_RECEIVED state due to a simultanious connect where
345 	 * our SYN has not been acked yet.
346 	 *
347 	 * In the normal retransmit-FIN-only case, however, snd_nxt will
348 	 * be set to snd_una, the offset will be 0, and the length may
349 	 * wind up 0.
350 	 */
351 	len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
352 
353 	/*
354 	 * Lop off SYN bit if it has already been sent.  However, if this
355 	 * is SYN-SENT state and if segment contains data, suppress sending
356 	 * segment (sending the segment would be an option if we still
357 	 * did TAO and the remote host supported it).
358 	 */
359 	if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
360 		flags &= ~TH_SYN;
361 		off--, len++;
362 		if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
363 			tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
364 			return 0;
365 		}
366 	}
367 
368 	/*
369 	 * Be careful not to send data and/or FIN on SYN segments.
370 	 * This measure is needed to prevent interoperability problems
371 	 * with not fully conformant TCP implementations.
372 	 */
373 	if (flags & TH_SYN) {
374 		len = 0;
375 		flags &= ~TH_FIN;
376 	}
377 
378 	if (len < 0) {
379 		/*
380 		 * A negative len can occur if our FIN has been sent but not
381 		 * acked, or if we are in a simultanious connect in the
382 		 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
383 		 * acked.
384 		 *
385 		 * If our window has contracted to 0 in the FIN case
386 		 * (which can only occur if we have NOT been called to
387 		 * retransmit as per code a few paragraphs up) then we
388 		 * want to shift the retransmit timer over to the
389 		 * persist timer.
390 		 *
391 		 * However, if we are in the TCPS_SYN_RECEIVED state
392 		 * (the SYN case) we will be in a simultanious connect and
393 		 * the window may be zero degeneratively.  In this case we
394 		 * do not want to shift to the persist timer after the SYN
395 		 * or the SYN+ACK transmission.
396 		 */
397 		len = 0;
398 		if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
399 			tcp_callout_stop(tp, tp->tt_rexmt);
400 			tp->t_rxtshift = 0;
401 			tp->snd_nxt = tp->snd_una;
402 			if (!tcp_callout_active(tp, tp->tt_persist))
403 				tcp_setpersist(tp);
404 		}
405 	}
406 
407 	KASSERT(len >= 0, ("%s: len < 0", __func__));
408 	/*
409 	 * Automatic sizing of send socket buffer.  Often the send buffer
410 	 * size is not optimally adjusted to the actual network conditions
411 	 * at hand (delay bandwidth product).  Setting the buffer size too
412 	 * small limits throughput on links with high bandwidth and high
413 	 * delay (eg. trans-continental/oceanic links).  Setting the
414 	 * buffer size too big consumes too much real kernel memory,
415 	 * especially with many connections on busy servers.
416 	 *
417 	 * The criteria to step up the send buffer one notch are:
418 	 *  1. receive window of remote host is larger than send buffer
419 	 *     (with a fudge factor of 5/4th);
420 	 *  2. hiwat has not significantly exceeded bwnd (inflight)
421 	 *     (bwnd is a maximal value if inflight is disabled).
422 	 *  3. send buffer is filled to 7/8th with data (so we actually
423 	 *     have data to make use of it);
424 	 *  4. hiwat has not hit maximal automatic size;
425 	 *  5. our send window (slow start and cogestion controlled) is
426 	 *     larger than sent but unacknowledged data in send buffer.
427 	 *
428 	 * The remote host receive window scaling factor may limit the
429 	 * growing of the send buffer before it reaches its allowed
430 	 * maximum.
431 	 *
432 	 * It scales directly with slow start or congestion window
433 	 * and does at most one step per received ACK.  This fast
434 	 * scaling has the drawback of growing the send buffer beyond
435 	 * what is strictly necessary to make full use of a given
436 	 * delay*bandwith product.  However testing has shown this not
437 	 * to be much of an problem.  At worst we are trading wasting
438 	 * of available bandwith (the non-use of it) for wasting some
439 	 * socket buffer memory.
440 	 *
441 	 * The criteria for shrinking the buffer is based solely on
442 	 * the inflight code (snd_bwnd).  If inflight is disabled,
443 	 * the buffer will not be shrinked.  Note that snd_bwnd already
444 	 * has a fudge factor.  Our test adds a little hysteresis.
445 	 */
446 	if (tcp_do_autosndbuf && (so->so_snd.ssb_flags & SSB_AUTOSIZE)) {
447 		const int asbinc = tcp_autosndbuf_inc;
448 		const int hiwat = so->so_snd.ssb_hiwat;
449 		const int lowat = so->so_snd.ssb_lowat;
450 		u_long newsize;
451 
452 		if ((tp->snd_wnd / 4 * 5) >= hiwat &&
453 		    so->so_snd.ssb_cc >= (hiwat / 8 * 7) &&
454 		    hiwat < tp->snd_bwnd + hiwat / 10 &&
455 		    hiwat + asbinc < tcp_autosndbuf_max &&
456 		    hiwat < (TCP_MAXWIN << tp->snd_scale) &&
457 		    sendwin >= (so->so_snd.ssb_cc -
458 				(tp->snd_nxt - tp->snd_una))) {
459 			newsize = ulmin(hiwat + asbinc, tcp_autosndbuf_max);
460 			if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
461 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
462 #if 0
463 			if (newsize >= (TCP_MAXWIN << tp->snd_scale))
464 				atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
465 #endif
466 		} else if ((long)tp->snd_bwnd <
467 			   (long)(hiwat * 3 / 4 - lowat - asbinc) &&
468 			   hiwat > tp->t_maxseg * 2 + asbinc &&
469 			   hiwat + asbinc >= tcp_autosndbuf_min &&
470 			   tcp_do_autosndbuf == 1) {
471 			newsize = ulmax(hiwat - asbinc, tp->t_maxseg * 2);
472 			ssb_reserve(&so->so_snd, newsize, so, NULL);
473 		}
474 	}
475 
476 	/*
477 	 * Don't use TSO, if:
478 	 * - Congestion window needs validation
479 	 * - There are SACK blocks to report
480 	 * - RST or SYN flags is set
481 	 * - URG will be set
482 	 *
483 	 * XXX
484 	 * Checking for SYN|RST looks overkill, just to be safe than sorry
485 	 */
486 	use_tso = can_tso;
487 	if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
488 		use_tso = FALSE;
489 	if (use_tso) {
490 		tcp_seq ugr_nxt = tp->snd_nxt;
491 
492 		if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
493 		    tp->snd_nxt == tp->snd_max)
494 			--ugr_nxt;
495 
496 		if (SEQ_GT(tp->snd_up, ugr_nxt))
497 			use_tso = FALSE;
498 	}
499 
500 	if (use_tso) {
501 		/*
502 		 * Find out segment size and header length for TSO
503 		 */
504 		error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
505 		if (error)
506 			use_tso = FALSE;
507 	}
508 	if (!use_tso) {
509 		segsz = tp->t_maxseg;
510 		tso_hlen = 0; /* not used */
511 	}
512 
513 	/*
514 	 * Truncate to the maximum segment length if not TSO, and ensure that
515 	 * FIN is removed if the length no longer contains the last data byte.
516 	 */
517 	if (len > segsz) {
518 		if (!use_tso) {
519 			len = segsz;
520 			++segcnt;
521 		} else {
522 			int nsegs;
523 
524 			if (__predict_false(tso_lenmax < segsz))
525 				tso_lenmax = segsz << 1;
526 
527 			/*
528 			 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
529 			 * thoff), and make sure that we send equal size
530 			 * transfers down the stack (rather than big-small-
531 			 * big-small-...).
532 			 */
533 			len = min(len, tso_lenmax);
534 			nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz;
535 			KKASSERT(nsegs > 0);
536 
537 			len = nsegs * segsz;
538 
539 			if (len <= segsz) {
540 				use_tso = FALSE;
541 				++segcnt;
542 			} else {
543 				segcnt += nsegs;
544 			}
545 		}
546 		sendalot = TRUE;
547 	} else {
548 		use_tso = FALSE;
549 		if (len > 0)
550 			++segcnt;
551 	}
552 	if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
553 		flags &= ~TH_FIN;
554 
555 	recvwin = ssb_space(&so->so_rcv);
556 
557 	/*
558 	 * Sender silly window avoidance.   We transmit under the following
559 	 * conditions when len is non-zero:
560 	 *
561 	 *	- We have a full segment
562 	 *	- This is the last buffer in a write()/send() and we are
563 	 *	  either idle or running NODELAY
564 	 *	- we've timed out (e.g. persist timer)
565 	 *	- we have more then 1/2 the maximum send window's worth of
566 	 *	  data (receiver may be limiting the window size)
567 	 *	- we need to retransmit
568 	 */
569 	if (len) {
570 		if (len >= segsz)
571 			goto send;
572 		/*
573 		 * NOTE! on localhost connections an 'ack' from the remote
574 		 * end may occur synchronously with the output and cause
575 		 * us to flush a buffer queued with moretocome.  XXX
576 		 *
577 		 * note: the len + off check is almost certainly unnecessary.
578 		 */
579 		if (!(tp->t_flags & TF_MORETOCOME) &&	/* normal case */
580 		    (idle || (tp->t_flags & TF_NODELAY)) &&
581 		    len + off >= so->so_snd.ssb_cc &&
582 		    !(tp->t_flags & TF_NOPUSH)) {
583 			goto send;
584 		}
585 		if (tp->t_flags & TF_FORCE)		/* typ. timeout case */
586 			goto send;
587 		if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
588 			goto send;
589 		if (SEQ_LT(tp->snd_nxt, tp->snd_max))	/* retransmit case */
590 			goto send;
591 		if (tp->t_flags & TF_XMITNOW)
592 			goto send;
593 	}
594 
595 	/*
596 	 * Compare available window to amount of window
597 	 * known to peer (as advertised window less
598 	 * next expected input).  If the difference is at least two
599 	 * max size segments, or at least 50% of the maximum possible
600 	 * window, then want to send a window update to peer.
601 	 */
602 	if (recvwin > 0) {
603 		/*
604 		 * "adv" is the amount we can increase the window,
605 		 * taking into account that we are limited by
606 		 * TCP_MAXWIN << tp->rcv_scale.
607 		 */
608 		long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
609 			(tp->rcv_adv - tp->rcv_nxt);
610 		long hiwat;
611 
612 		/*
613 		 * This ack case typically occurs when the user has drained
614 		 * the TCP socket buffer sufficiently to warrent an ack
615 		 * containing a 'pure window update'... that is, an ack that
616 		 * ONLY updates the tcp window.
617 		 *
618 		 * It is unclear why we would need to do a pure window update
619 		 * past 2 segments if we are going to do one at 1/2 the high
620 		 * water mark anyway, especially since under normal conditions
621 		 * the user program will drain the socket buffer quickly.
622 		 * The 2-segment pure window update will often add a large
623 		 * number of extra, unnecessary acks to the stream.
624 		 *
625 		 * avoid_pure_win_update now defaults to 1.
626 		 */
627 		if (avoid_pure_win_update == 0 ||
628 		    (tp->t_flags & TF_RXRESIZED)) {
629 			if (adv >= (long) (2 * segsz)) {
630 				goto send;
631 			}
632 		}
633 		hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
634 		if (hiwat > (long)so->so_rcv.ssb_hiwat)
635 			hiwat = (long)so->so_rcv.ssb_hiwat;
636 		if (adv >= hiwat / 2)
637 			goto send;
638 	}
639 
640 	/*
641 	 * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
642 	 * is also a catch-all for the retransmit timer timeout case.
643 	 */
644 	if (tp->t_flags & TF_ACKNOW)
645 		goto send;
646 	if ((flags & TH_RST) ||
647 	    ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
648 		goto send;
649 	if (SEQ_GT(tp->snd_up, tp->snd_una))
650 		goto send;
651 	/*
652 	 * If our state indicates that FIN should be sent
653 	 * and we have not yet done so, then we need to send.
654 	 */
655 	if ((flags & TH_FIN) &&
656 	    (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
657 		goto send;
658 
659 	/*
660 	 * TCP window updates are not reliable, rather a polling protocol
661 	 * using ``persist'' packets is used to insure receipt of window
662 	 * updates.  The three ``states'' for the output side are:
663 	 *	idle			not doing retransmits or persists
664 	 *	persisting		to move a small or zero window
665 	 *	(re)transmitting	and thereby not persisting
666 	 *
667 	 * tcp_callout_active(tp, tp->tt_persist)
668 	 *	is true when we are in persist state.
669 	 * The TF_FORCE flag in tp->t_flags
670 	 *	is set when we are called to send a persist packet.
671 	 * tcp_callout_active(tp, tp->tt_rexmt)
672 	 *	is set when we are retransmitting
673 	 * The output side is idle when both timers are zero.
674 	 *
675 	 * If send window is too small, there is data to transmit, and no
676 	 * retransmit or persist is pending, then go to persist state.
677 	 *
678 	 * If nothing happens soon, send when timer expires:
679 	 * if window is nonzero, transmit what we can, otherwise force out
680 	 * a byte.
681 	 *
682 	 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
683 	 * with data pending.  This situation can occur during a
684 	 * simultanious connect.
685 	 */
686 	if (so->so_snd.ssb_cc > 0 &&
687 	    tp->t_state != TCPS_SYN_RECEIVED &&
688 	    !tcp_callout_active(tp, tp->tt_rexmt) &&
689 	    !tcp_callout_active(tp, tp->tt_persist)) {
690 		tp->t_rxtshift = 0;
691 		tcp_setpersist(tp);
692 	}
693 
694 	/*
695 	 * No reason to send a segment, just return.
696 	 */
697 	tp->t_flags &= ~TF_XMITNOW;
698 	return (0);
699 
700 send:
701 	if (need_sched && len > 0) {
702 		tcp_output_sched(tp);
703 		return 0;
704 	}
705 
706 	/*
707 	 * Before ESTABLISHED, force sending of initial options
708 	 * unless TCP set not to do any options.
709 	 * NOTE: we assume that the IP/TCP header plus TCP options
710 	 * always fit in a single mbuf, leaving room for a maximum
711 	 * link header, i.e.
712 	 *	max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
713 	 */
714 	optlen = 0;
715 	if (isipv6)
716 		hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
717 	else
718 		hdrlen = sizeof(struct tcpiphdr);
719 	if (flags & TH_SYN) {
720 		tp->snd_nxt = tp->iss;
721 		if (!(tp->t_flags & TF_NOOPT)) {
722 			u_short mss;
723 
724 			opt[0] = TCPOPT_MAXSEG;
725 			opt[1] = TCPOLEN_MAXSEG;
726 			mss = htons((u_short) tcp_mssopt(tp));
727 			memcpy(opt + 2, &mss, sizeof mss);
728 			optlen = TCPOLEN_MAXSEG;
729 
730 			if ((tp->t_flags & TF_REQ_SCALE) &&
731 			    (!(flags & TH_ACK) ||
732 			     (tp->t_flags & TF_RCVD_SCALE))) {
733 				*((u_int32_t *)(opt + optlen)) = htonl(
734 					TCPOPT_NOP << 24 |
735 					TCPOPT_WINDOW << 16 |
736 					TCPOLEN_WINDOW << 8 |
737 					tp->request_r_scale);
738 				optlen += 4;
739 			}
740 
741 			if ((tcp_do_sack && !(flags & TH_ACK)) ||
742 			    tp->t_flags & TF_SACK_PERMITTED) {
743 				uint32_t *lp = (uint32_t *)(opt + optlen);
744 
745 				*lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
746 				optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
747 			}
748 		}
749 	}
750 
751 	/*
752 	 * Send a timestamp and echo-reply if this is a SYN and our side
753 	 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
754 	 * and our peer have sent timestamps in our SYN's.
755 	 */
756 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
757 	    !(flags & TH_RST) &&
758 	    (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
759 		u_int32_t *lp = (u_int32_t *)(opt + optlen);
760 
761 		/* Form timestamp option as shown in appendix A of RFC 1323. */
762 		*lp++ = htonl(TCPOPT_TSTAMP_HDR);
763 		*lp++ = htonl(ticks);
764 		*lp   = htonl(tp->ts_recent);
765 		optlen += TCPOLEN_TSTAMP_APPA;
766 	}
767 
768 	/* Set receive buffer autosizing timestamp. */
769 	if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
770 		tp->rfbuf_ts = ticks;
771 
772 	/*
773 	 * If this is a SACK connection and we have a block to report,
774 	 * fill in the SACK blocks in the TCP options.
775 	 */
776 	if (report_sack)
777 		tcp_sack_fill_report(tp, opt, &optlen);
778 
779 #ifdef TCP_SIGNATURE
780 	if (tp->t_flags & TF_SIGNATURE) {
781 		int i;
782 		u_char *bp;
783 		/*
784 		 * Initialize TCP-MD5 option (RFC2385)
785 		 */
786 		bp = (u_char *)opt + optlen;
787 		*bp++ = TCPOPT_SIGNATURE;
788 		*bp++ = TCPOLEN_SIGNATURE;
789 		sigoff = optlen + 2;
790 		for (i = 0; i < TCP_SIGLEN; i++)
791 			*bp++ = 0;
792 		optlen += TCPOLEN_SIGNATURE;
793 		/*
794 		 * Terminate options list and maintain 32-bit alignment.
795 		 */
796 		*bp++ = TCPOPT_NOP;
797 		*bp++ = TCPOPT_EOL;
798 		optlen += 2;
799 	}
800 #endif /* TCP_SIGNATURE */
801 	KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
802 	hdrlen += optlen;
803 
804 	if (isipv6) {
805 		ipoptlen = ip6_optlen(inp);
806 	} else {
807 		if (inp->inp_options) {
808 			ipoptlen = inp->inp_options->m_len -
809 			    offsetof(struct ipoption, ipopt_list);
810 		} else {
811 			ipoptlen = 0;
812 		}
813 	}
814 
815 	if (use_tso) {
816 		/* TSO segment length must be multiple of segment size */
817 		KASSERT(len >= (2 * segsz) && (len % segsz == 0),
818 		    ("invalid TSO len %ld, segsz %u", len, segsz));
819 	} else {
820 		KASSERT(len <= segsz,
821 		    ("invalid len %ld, segsz %u", len, segsz));
822 
823 		/*
824 		 * Adjust data length if insertion of options will bump
825 		 * the packet length beyond the t_maxopd length.  Clear
826 		 * FIN to prevent premature closure since there is still
827 		 * more data to send after this (now truncated) packet.
828 		 *
829 		 * If just the options do not fit we are in a no-win
830 		 * situation and we treat it as an unreachable host.
831 		 */
832 		if (len + optlen + ipoptlen > tp->t_maxopd) {
833 			if (tp->t_maxopd <= optlen + ipoptlen) {
834 				static time_t last_optlen_report;
835 
836 				if (last_optlen_report != time_uptime) {
837 					last_optlen_report = time_uptime;
838 					kprintf("tcpcb %p: MSS (%d) too "
839 					    "small to hold options!\n",
840 					    tp, tp->t_maxopd);
841 				}
842 				error = EHOSTUNREACH;
843 				goto out;
844 			} else {
845 				flags &= ~TH_FIN;
846 				len = tp->t_maxopd - optlen - ipoptlen;
847 				sendalot = TRUE;
848 			}
849 		}
850 	}
851 
852 #ifdef INET6
853 	KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
854 #else
855 	KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
856 #endif
857 
858 	/*
859 	 * Grab a header mbuf, attaching a copy of data to
860 	 * be transmitted, and initialize the header from
861 	 * the template for sends on this connection.
862 	 */
863 	if (len) {
864 		if ((tp->t_flags & TF_FORCE) && len == 1)
865 			tcpstat.tcps_sndprobe++;
866 		else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
867 			if (tp->snd_nxt == tp->snd_una)
868 				tp->snd_max_rexmt = tp->snd_max;
869 			if (nsacked) {
870 				tcpstat.tcps_sndsackrtopack++;
871 				tcpstat.tcps_sndsackrtobyte += len;
872 			}
873 			tcpstat.tcps_sndrexmitpack++;
874 			tcpstat.tcps_sndrexmitbyte += len;
875 		} else {
876 			tcpstat.tcps_sndpack++;
877 			tcpstat.tcps_sndbyte += len;
878 		}
879 		if (idle_cwv) {
880 			idle_cwv = FALSE;
881 			tcp_idle_cwnd_validate(tp);
882 		}
883 		/* Update last send time after CWV */
884 		tp->snd_last = ticks;
885 #ifdef notyet
886 		if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
887 		    max_linkhdr + hdrlen)) == NULL) {
888 			error = ENOBUFS;
889 			goto after_th;
890 		}
891 		/*
892 		 * m_copypack left space for our hdr; use it.
893 		 */
894 		m->m_len += hdrlen;
895 		m->m_data -= hdrlen;
896 #else
897 #ifndef INET6
898 		m = m_gethdr(M_NOWAIT, MT_HEADER);
899 #else
900 		m = m_getl(hdrlen + max_linkhdr, M_NOWAIT, MT_HEADER,
901 			   M_PKTHDR, NULL);
902 #endif
903 		if (m == NULL) {
904 			error = ENOBUFS;
905 			goto after_th;
906 		}
907 		m->m_data += max_linkhdr;
908 		m->m_len = hdrlen;
909 		if (len <= MHLEN - hdrlen - max_linkhdr) {
910 			m_copydata(so->so_snd.ssb_mb, off, (int) len,
911 			    mtod(m, caddr_t) + hdrlen);
912 			m->m_len += len;
913 		} else {
914 			m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
915 			if (m->m_next == NULL) {
916 				m_free(m);
917 				m = NULL;
918 				error = ENOBUFS;
919 				goto after_th;
920 			}
921 		}
922 #endif
923 		/*
924 		 * If we're sending everything we've got, set PUSH.
925 		 * (This will keep happy those implementations which only
926 		 * give data to the user when a buffer fills or
927 		 * a PUSH comes in.)
928 		 */
929 		if (off + len == so->so_snd.ssb_cc)
930 			flags |= TH_PUSH;
931 	} else {
932 		if (tp->t_flags & TF_ACKNOW)
933 			tcpstat.tcps_sndacks++;
934 		else if (flags & (TH_SYN | TH_FIN | TH_RST))
935 			tcpstat.tcps_sndctrl++;
936 		else if (SEQ_GT(tp->snd_up, tp->snd_una))
937 			tcpstat.tcps_sndurg++;
938 		else
939 			tcpstat.tcps_sndwinup++;
940 
941 		MGETHDR(m, M_NOWAIT, MT_HEADER);
942 		if (m == NULL) {
943 			error = ENOBUFS;
944 			goto after_th;
945 		}
946 		if (isipv6 &&
947 		    (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
948 			MH_ALIGN(m, hdrlen);
949 		else
950 			m->m_data += max_linkhdr;
951 		m->m_len = hdrlen;
952 
953 		/*
954 		 * Prioritize SYN, SYN|ACK and pure ACK.
955 		 * Leave FIN and RST as they are.
956 		 */
957 		if (tcp_prio_synack && (flags & (TH_FIN | TH_RST)) == 0)
958 			m->m_flags |= M_PRIO;
959 	}
960 	m->m_pkthdr.rcvif = NULL;
961 	if (isipv6) {
962 		ip6 = mtod(m, struct ip6_hdr *);
963 		th = (struct tcphdr *)(ip6 + 1);
964 		tcp_fillheaders(tp, ip6, th, use_tso);
965 	} else {
966 		ip = mtod(m, struct ip *);
967 		th = (struct tcphdr *)(ip + 1);
968 		/* this picks up the pseudo header (w/o the length) */
969 		tcp_fillheaders(tp, ip, th, use_tso);
970 	}
971 after_th:
972 	/*
973 	 * Fill in fields, remembering maximum advertised
974 	 * window for use in delaying messages about window sizes.
975 	 * If resending a FIN, be sure not to use a new sequence number.
976 	 */
977 	if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
978 	    tp->snd_nxt == tp->snd_max)
979 		tp->snd_nxt--;
980 
981 	if (th != NULL) {
982 		/*
983 		 * If we are doing retransmissions, then snd_nxt will
984 		 * not reflect the first unsent octet.  For ACK only
985 		 * packets, we do not want the sequence number of the
986 		 * retransmitted packet, we want the sequence number
987 		 * of the next unsent octet.  So, if there is no data
988 		 * (and no SYN or FIN), use snd_max instead of snd_nxt
989 		 * when filling in ti_seq.  But if we are in persist
990 		 * state, snd_max might reflect one byte beyond the
991 		 * right edge of the window, so use snd_nxt in that
992 		 * case, since we know we aren't doing a retransmission.
993 		 * (retransmit and persist are mutually exclusive...)
994 		 */
995 		if (len || (flags & (TH_SYN|TH_FIN)) ||
996 		    tcp_callout_active(tp, tp->tt_persist))
997 			th->th_seq = htonl(tp->snd_nxt);
998 		else
999 			th->th_seq = htonl(tp->snd_max);
1000 		th->th_ack = htonl(tp->rcv_nxt);
1001 		if (optlen) {
1002 			bcopy(opt, th + 1, optlen);
1003 			th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
1004 		}
1005 		th->th_flags = flags;
1006 	}
1007 
1008 	/*
1009 	 * Calculate receive window.  Don't shrink window, but avoid
1010 	 * silly window syndrome by sending a 0 window if the actual
1011 	 * window is less then one segment.
1012 	 */
1013 	if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
1014 	    recvwin < (long)segsz)
1015 		recvwin = 0;
1016 	if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
1017 		recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
1018 	if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
1019 		recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
1020 
1021 	/*
1022 	 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1023 	 * a 0 window.  This may cause the remote transmitter to stall.  This
1024 	 * flag tells soreceive() to disable delayed acknowledgements when
1025 	 * draining the buffer.  This can occur if the receiver is attempting
1026 	 * to read more data then can be buffered prior to transmitting on
1027 	 * the connection.
1028 	 */
1029 	if (recvwin == 0)
1030 		tp->t_flags |= TF_RXWIN0SENT;
1031 	else
1032 		tp->t_flags &= ~TF_RXWIN0SENT;
1033 
1034 	if (th != NULL)
1035 		th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
1036 
1037 	if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1038 		KASSERT(!use_tso, ("URG with TSO"));
1039 		if (th != NULL) {
1040 			th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1041 			th->th_flags |= TH_URG;
1042 		}
1043 	} else {
1044 		/*
1045 		 * If no urgent pointer to send, then we pull
1046 		 * the urgent pointer to the left edge of the send window
1047 		 * so that it doesn't drift into the send window on sequence
1048 		 * number wraparound.
1049 		 */
1050 		tp->snd_up = tp->snd_una;		/* drag it along */
1051 	}
1052 
1053 	if (th != NULL) {
1054 #ifdef TCP_SIGNATURE
1055 		if (tp->t_flags & TF_SIGNATURE) {
1056 			tcpsignature_compute(m, len, optlen,
1057 			    (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1058 		}
1059 #endif /* TCP_SIGNATURE */
1060 
1061 		/*
1062 		 * Put TCP length in extended header, and then
1063 		 * checksum extended header and data.
1064 		 */
1065 		m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1066 		if (isipv6) {
1067 			/*
1068 			 * ip6_plen is not need to be filled now, and will be
1069 			 * filled in ip6_output().
1070 			 */
1071 			th->th_sum = in6_cksum(m, IPPROTO_TCP,
1072 			    sizeof(struct ip6_hdr),
1073 			    sizeof(struct tcphdr) + optlen + len);
1074 		} else {
1075 			m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1076 			if (use_tso) {
1077 				m->m_pkthdr.csum_flags = CSUM_TSO;
1078 				m->m_pkthdr.tso_segsz = segsz;
1079 			} else {
1080 				m->m_pkthdr.csum_flags = CSUM_TCP;
1081 				m->m_pkthdr.csum_data =
1082 				    offsetof(struct tcphdr, th_sum);
1083 				if (len + optlen) {
1084 					th->th_sum = in_addword(th->th_sum,
1085 					    htons((u_short)(optlen + len)));
1086 				}
1087 			}
1088 
1089 			/*
1090 			 * IP version must be set here for ipv4/ipv6 checking
1091 			 * later
1092 			 */
1093 			KASSERT(ip->ip_v == IPVERSION,
1094 			    ("%s: IP version incorrect: %d",
1095 			     __func__, ip->ip_v));
1096 		}
1097 	}
1098 
1099 	/*
1100 	 * In transmit state, time the transmission and arrange for
1101 	 * the retransmit.  In persist state, just set snd_max.
1102 	 */
1103 	if (!(tp->t_flags & TF_FORCE) ||
1104 	    !tcp_callout_active(tp, tp->tt_persist)) {
1105 		tcp_seq startseq = tp->snd_nxt;
1106 
1107 		/*
1108 		 * Advance snd_nxt over sequence space of this segment.
1109 		 */
1110 		if (flags & (TH_SYN | TH_FIN)) {
1111 			if (flags & TH_SYN)
1112 				tp->snd_nxt++;
1113 			if (flags & TH_FIN) {
1114 				tp->snd_nxt++;
1115 				tp->t_flags |= TF_SENTFIN;
1116 			}
1117 		}
1118 		tp->snd_nxt += len;
1119 		if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1120 			tp->snd_max = tp->snd_nxt;
1121 			/*
1122 			 * Time this transmission if not a retransmission and
1123 			 * not currently timing anything.
1124 			 */
1125 			if (tp->t_rtttime == 0) {
1126 				tp->t_rtttime = ticks;
1127 				tp->t_rtseq = startseq;
1128 				tcpstat.tcps_segstimed++;
1129 			}
1130 		}
1131 
1132 		/*
1133 		 * Set retransmit timer if not currently set,
1134 		 * and not doing a pure ack or a keep-alive probe.
1135 		 * Initial value for retransmit timer is smoothed
1136 		 * round-trip time + 2 * round-trip time variance.
1137 		 * Initialize shift counter which is used for backoff
1138 		 * of retransmit time.
1139 		 */
1140 		if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1141 		    tp->snd_nxt != tp->snd_una) {
1142 			if (tcp_callout_active(tp, tp->tt_persist)) {
1143 				tcp_callout_stop(tp, tp->tt_persist);
1144 				tp->t_rxtshift = 0;
1145 			}
1146 			tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1147 			    tcp_timer_rexmt);
1148 		} else if (len == 0 && so->so_snd.ssb_cc &&
1149 			   tp->t_state > TCPS_SYN_RECEIVED &&
1150 			   !tcp_callout_active(tp, tp->tt_rexmt) &&
1151 			   !tcp_callout_active(tp, tp->tt_persist)) {
1152 			/*
1153 			 * Avoid a situation where we do not set persist timer
1154 			 * after a zero window condition. For example:
1155 			 * 1) A -> B: packet with enough data to fill the window
1156 			 * 2) B -> A: ACK for #1 + new data (0 window
1157 			 *    advertisement)
1158 			 * 3) A -> B: ACK for #2, 0 len packet
1159 			 *
1160 			 * In this case, A will not activate the persist timer,
1161 			 * because it chose to send a packet. Unless tcp_output
1162 			 * is called for some other reason (delayed ack timer,
1163 			 * another input packet from B, socket syscall), A will
1164 			 * not send zero window probes.
1165 			 *
1166 			 * So, if you send a 0-length packet, but there is data
1167 			 * in the socket buffer, and neither the rexmt or
1168 			 * persist timer is already set, then activate the
1169 			 * persist timer.
1170 			 */
1171 			tp->t_rxtshift = 0;
1172 			tcp_setpersist(tp);
1173 		}
1174 	} else {
1175 		/*
1176 		 * Persist case, update snd_max but since we are in
1177 		 * persist mode (no window) we do not update snd_nxt.
1178 		 */
1179 		int xlen = len;
1180 		if (flags & TH_SYN)
1181 			panic("tcp_output: persist timer to send SYN");
1182 		if (flags & TH_FIN) {
1183 			++xlen;
1184 			tp->t_flags |= TF_SENTFIN;
1185 		}
1186 		if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1187 			tp->snd_max = tp->snd_nxt + xlen;
1188 	}
1189 
1190 	if (th != NULL) {
1191 #ifdef TCPDEBUG
1192 		/* Trace. */
1193 		if (so->so_options & SO_DEBUG) {
1194 			tcp_trace(TA_OUTPUT, tp->t_state, tp,
1195 			    mtod(m, void *), th, 0);
1196 		}
1197 #endif
1198 
1199 		/*
1200 		 * Fill in IP length and desired time to live and
1201 		 * send to IP level.  There should be a better way
1202 		 * to handle ttl and tos; we could keep them in
1203 		 * the template, but need a way to checksum without them.
1204 		 */
1205 		/*
1206 		 * m->m_pkthdr.len should have been set before cksum
1207 		 * calcuration, because in6_cksum() need it.
1208 		 */
1209 		if (isipv6) {
1210 			/*
1211 			 * we separately set hoplimit for every segment,
1212 			 * since the user might want to change the value
1213 			 * via setsockopt.  Also, desired default hop
1214 			 * limit might be changed via Neighbor Discovery.
1215 			 */
1216 			ip6->ip6_hlim = in6_selecthlim(inp,
1217 			    (inp->in6p_route.ro_rt ?
1218 			     inp->in6p_route.ro_rt->rt_ifp : NULL));
1219 
1220 			/* TODO: IPv6 IP6TOS_ECT bit on */
1221 			error = ip6_output(m, inp->in6p_outputopts,
1222 			    &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1223 			    NULL, NULL, inp);
1224 		} else {
1225 			struct rtentry *rt;
1226 
1227 			KASSERT(!INP_CHECK_SOCKAF(so, AF_INET6), ("inet6 pcb"));
1228 
1229 			ip->ip_len = m->m_pkthdr.len;
1230 			ip->ip_ttl = inp->inp_ip_ttl;	/* XXX */
1231 			ip->ip_tos = inp->inp_ip_tos;	/* XXX */
1232 			/*
1233 			 * See if we should do MTU discovery.
1234 			 * We do it only if the following are true:
1235 			 *	1) we have a valid route to the destination
1236 			 *	2) the MTU is not locked (if it is,
1237 			 *	   then discovery has been disabled)
1238 			 */
1239 			if (path_mtu_discovery &&
1240 			    (rt = inp->inp_route.ro_rt) &&
1241 			    (rt->rt_flags & RTF_UP) &&
1242 			    !(rt->rt_rmx.rmx_locks & RTV_MTU))
1243 				ip->ip_off |= IP_DF;
1244 
1245 			KASSERT(inp->inp_flags & INP_HASH,
1246 			    ("inpcb has no hash"));
1247 			m_sethash(m, inp->inp_hashval);
1248 			error = ip_output(m, inp->inp_options, &inp->inp_route,
1249 					  (so->so_options & SO_DONTROUTE) |
1250 					  IP_DEBUGROUTE, NULL, inp);
1251 		}
1252 	} else {
1253 		KASSERT(error != 0, ("no error, but th not set"));
1254 	}
1255 	if (error) {
1256 		tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1257 
1258 		/*
1259 		 * We know that the packet was lost, so back out the
1260 		 * sequence number advance, if any.
1261 		 */
1262 		if (!(tp->t_flags & TF_FORCE) ||
1263 		    !tcp_callout_active(tp, tp->tt_persist)) {
1264 			/*
1265 			 * No need to check for TH_FIN here because
1266 			 * the TF_SENTFIN flag handles that case.
1267 			 */
1268 			if (!(flags & TH_SYN))
1269 				tp->snd_nxt -= len;
1270 		}
1271 
1272 out:
1273 		if (error == ENOBUFS) {
1274 			KASSERT((len == 0 && (flags & (TH_SYN | TH_FIN)) == 0) ||
1275 			    tcp_callout_active(tp, tp->tt_rexmt) ||
1276 			    tcp_callout_active(tp, tp->tt_persist),
1277 			    ("neither rexmt nor persist timer is set"));
1278 			return (0);
1279 		}
1280 		if (error == EMSGSIZE) {
1281 			/*
1282 			 * ip_output() will have already fixed the route
1283 			 * for us.  tcp_mtudisc() will, as its last action,
1284 			 * initiate retransmission, so it is important to
1285 			 * not do so here.
1286 			 */
1287 			tcp_mtudisc(inp, 0);
1288 			return 0;
1289 		}
1290 		if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1291 		    TCPS_HAVERCVDSYN(tp->t_state)) {
1292 			tp->t_softerror = error;
1293 			return (0);
1294 		}
1295 		return (error);
1296 	}
1297 	tcpstat.tcps_sndtotal++;
1298 
1299 	/*
1300 	 * Data sent (as far as we can tell).
1301 	 *
1302 	 * If this advertises a larger window than any other segment,
1303 	 * then remember the size of the advertised window.
1304 	 *
1305 	 * Any pending ACK has now been sent.
1306 	 */
1307 	if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1308 		tp->rcv_adv = tp->rcv_nxt + recvwin;
1309 		tp->t_flags &= ~TF_RXRESIZED;
1310 	}
1311 	tp->last_ack_sent = tp->rcv_nxt;
1312 	tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1313 	if (tcp_delack_enabled)
1314 		tcp_callout_stop(tp, tp->tt_delack);
1315 	if (sendalot) {
1316 		if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) &&
1317 		    segcnt >= tcp_fairsend)
1318 			need_sched = TRUE;
1319 		goto again;
1320 	}
1321 	return (0);
1322 }
1323 
1324 void
1325 tcp_setpersist(struct tcpcb *tp)
1326 {
1327 	int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1328 	int tt;
1329 
1330 	if (tp->t_state == TCPS_SYN_SENT ||
1331 	    tp->t_state == TCPS_SYN_RECEIVED) {
1332 		panic("tcp_setpersist: not established yet, current %s",
1333 		      tp->t_state == TCPS_SYN_SENT ?
1334 		      "SYN_SENT" : "SYN_RECEIVED");
1335 	}
1336 
1337 	if (tcp_callout_active(tp, tp->tt_rexmt))
1338 		panic("tcp_setpersist: retransmit pending");
1339 	/*
1340 	 * Start/restart persistance timer.
1341 	 */
1342 	TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1343 		      TCPTV_PERSMAX);
1344 	tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1345 	if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1346 		tp->t_rxtshift++;
1347 }
1348 
1349 static void
1350 tcp_idle_cwnd_validate(struct tcpcb *tp)
1351 {
1352 	u_long initial_cwnd = tcp_initial_window(tp);
1353 	u_long min_cwnd;
1354 
1355 	tcpstat.tcps_sndidle++;
1356 
1357 	/* According to RFC5681: RW=min(IW,cwnd) */
1358 	min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1359 
1360 	if (tcp_idle_cwv) {
1361 		u_long idle_time, decay_cwnd;
1362 
1363 		/*
1364 		 * RFC2861, but only after idle period.
1365 		 */
1366 
1367 		/*
1368 		 * Before the congestion window is reduced, ssthresh
1369 		 * is set to the maximum of its current value and 3/4
1370 		 * cwnd.  If the sender then has more data to send
1371 		 * than the decayed cwnd allows, the TCP will slow-
1372 		 * start (perform exponential increase) at least
1373 		 * half-way back up to the old value of cwnd.
1374 		 */
1375 		tp->snd_ssthresh = max(tp->snd_ssthresh,
1376 		    (3 * tp->snd_cwnd) / 4);
1377 
1378 		/*
1379 		 * Decay the congestion window by half for every RTT
1380 		 * that the flow remains inactive.
1381 		 *
1382 		 * The difference between our implementation and
1383 		 * RFC2861 is that we don't allow cwnd to go below
1384 		 * the value allowed by RFC5681 (min_cwnd).
1385 		 */
1386 		idle_time = ticks - tp->snd_last;
1387 		decay_cwnd = tp->snd_cwnd;
1388 		while (idle_time >= tp->t_rxtcur &&
1389 		    decay_cwnd > min_cwnd) {
1390 			decay_cwnd >>= 1;
1391 			idle_time -= tp->t_rxtcur;
1392 		}
1393 		tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1394 	} else {
1395 		/*
1396 		 * Slow-start from scratch to re-determine the send
1397 		 * congestion window.
1398 		 */
1399 		tp->snd_cwnd = min_cwnd;
1400 	}
1401 
1402 	/* Restart ABC counting during congestion avoidance */
1403 	tp->snd_wacked = 0;
1404 }
1405 
1406 static int
1407 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1408 {
1409 	struct inpcb * const inp = tp->t_inpcb;
1410 #ifdef INET6
1411 	const boolean_t isipv6 = INP_ISIPV6(inp);
1412 #else
1413 	const boolean_t isipv6 = FALSE;
1414 #endif
1415 	unsigned int ipoptlen, optlen;
1416 	u_int hlen;
1417 
1418 	hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1419 
1420 	if (isipv6) {
1421 		ipoptlen = ip6_optlen(inp);
1422 	} else {
1423 		if (inp->inp_options) {
1424 			ipoptlen = inp->inp_options->m_len -
1425 			    offsetof(struct ipoption, ipopt_list);
1426 		} else {
1427 			ipoptlen = 0;
1428 		}
1429 	}
1430 	hlen += ipoptlen;
1431 
1432 	optlen = 0;
1433 	if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1434 	    (tp->t_flags & TF_RCVD_TSTMP))
1435 		optlen += TCPOLEN_TSTAMP_APPA;
1436 	hlen += optlen;
1437 
1438 	if (tp->t_maxopd <= optlen + ipoptlen)
1439 		return EHOSTUNREACH;
1440 
1441 	*segsz = tp->t_maxopd - optlen - ipoptlen;
1442 	*hlen0 = hlen;
1443 	return 0;
1444 }
1445 
1446 static void
1447 tcp_output_sched_handler(netmsg_t nmsg)
1448 {
1449 	struct tcpcb *tp = nmsg->lmsg.u.ms_resultp;
1450 
1451 	/* Reply ASAP */
1452 	crit_enter();
1453 	lwkt_replymsg(&nmsg->lmsg, 0);
1454 	crit_exit();
1455 
1456 	tcp_output_fair(tp);
1457 }
1458 
1459 void
1460 tcp_output_init(struct tcpcb *tp)
1461 {
1462 	netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE,
1463 	    tcp_output_sched_handler);
1464 	tp->tt_sndmore->lmsg.u.ms_resultp = tp;
1465 }
1466 
1467 void
1468 tcp_output_cancel(struct tcpcb *tp)
1469 {
1470 	/*
1471 	 * This message is still pending to be processed;
1472 	 * drop it.  Optimized.
1473 	 */
1474 	crit_enter();
1475 	if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) {
1476 		lwkt_dropmsg(&tp->tt_sndmore->lmsg);
1477 	}
1478 	crit_exit();
1479 }
1480 
1481 boolean_t
1482 tcp_output_pending(struct tcpcb *tp)
1483 {
1484 	if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0)
1485 		return TRUE;
1486 	else
1487 		return FALSE;
1488 }
1489 
1490 static void
1491 tcp_output_sched(struct tcpcb *tp)
1492 {
1493 	crit_enter();
1494 	if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE)
1495 		lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg);
1496 	crit_exit();
1497 }
1498 
1499 /*
1500  * Fairsend
1501  *
1502  * Yield to other senders or receivers on the same netisr if the current
1503  * TCP stream has sent tcp_fairsend segments and is going to burst more
1504  * segments.  Bursting large amount of segements in a single TCP stream
1505  * could delay other senders' segments and receivers' ACKs quite a lot,
1506  * if others segments and ACKs are queued on to the same hardware transmit
1507  * queue; thus cause unfairness between senders and suppress receiving
1508  * performance.
1509  *
1510  * Fairsend should be performed at the places that do not affect segment
1511  * sending during congestion control, e.g.
1512  * - User requested output
1513  * - ACK input triggered output
1514  *
1515  * NOTE:
1516  * For devices that are TSO capable, their TSO aggregation size limit could
1517  * affect fairsend.
1518  */
1519 int
1520 tcp_output_fair(struct tcpcb *tp)
1521 {
1522 	int ret;
1523 
1524 	tp->t_flags |= TF_FAIRSEND;
1525 	ret = tcp_output(tp);
1526 	tp->t_flags &= ~TF_FAIRSEND;
1527 
1528 	return ret;
1529 }
1530