tcp/tso: Add per-device TSO aggregation size limit
[dragonfly.git] / sys / netinet / tcp_output.c
CommitLineData
984263bc 1/*
86e10434 2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
66d6c637 3 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
f23061d4 4 *
66d6c637
JH
5 * This code is derived from software contributed to The DragonFly Project
6 * by Jeffrey M. Hsu.
f23061d4 7 *
66d6c637
JH
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.
f23061d4 19 *
66d6c637
JH
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/*
984263bc
MD
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. All advertising materials mentioning features or use of this software
47 * must display the following acknowledgement:
48 * This product includes software developed by the University of
49 * California, Berkeley and its contributors.
50 * 4. Neither the name of the University nor the names of its contributors
51 * may be used to endorse or promote products derived from this software
52 * without specific prior written permission.
53 *
54 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
55 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
58 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64 * SUCH DAMAGE.
65 *
66 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
67 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
68 */
69
b1992928 70#include "opt_inet.h"
984263bc
MD
71#include "opt_inet6.h"
72#include "opt_ipsec.h"
73#include "opt_tcpdebug.h"
74
75#include <sys/param.h>
76#include <sys/systm.h>
77#include <sys/kernel.h>
78#include <sys/sysctl.h>
79#include <sys/mbuf.h>
80#include <sys/domain.h>
81#include <sys/protosw.h>
82#include <sys/socket.h>
83#include <sys/socketvar.h>
3f9db7f8 84#include <sys/in_cksum.h>
2b57d013
MD
85#include <sys/thread.h>
86#include <sys/globaldata.h>
984263bc 87
5f60906c 88#include <net/if_var.h>
984263bc
MD
89#include <net/route.h>
90
91#include <netinet/in.h>
92#include <netinet/in_systm.h>
93#include <netinet/ip.h>
94#include <netinet/in_pcb.h>
95#include <netinet/ip_var.h>
984263bc
MD
96#include <netinet6/in6_pcb.h>
97#include <netinet/ip6.h>
98#include <netinet6/ip6_var.h>
984263bc
MD
99#include <netinet/tcp.h>
100#define TCPOUTFLAGS
101#include <netinet/tcp_fsm.h>
102#include <netinet/tcp_seq.h>
103#include <netinet/tcp_timer.h>
a48c5dd5 104#include <netinet/tcp_timer2.h>
984263bc
MD
105#include <netinet/tcp_var.h>
106#include <netinet/tcpip.h>
107#ifdef TCPDEBUG
108#include <netinet/tcp_debug.h>
109#endif
110
111#ifdef IPSEC
112#include <netinet6/ipsec.h>
113#endif /*IPSEC*/
114
115#ifdef FAST_IPSEC
bf844ffa 116#include <netproto/ipsec/ipsec.h>
984263bc
MD
117#define IPSEC
118#endif /*FAST_IPSEC*/
119
984263bc
MD
120#ifdef notyet
121extern struct mbuf *m_copypack();
122#endif
123
dcc39f86 124int path_mtu_discovery = 0;
984263bc
MD
125SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
126 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
127
ddb2e271
MD
128static int avoid_pure_win_update = 1;
129SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
130 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
131
5b0b9fa5
PA
132int tcp_do_autosndbuf = 1;
133SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
134 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
135
136int tcp_autosndbuf_inc = 8*1024;
137SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
138 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
139
46e92930 140int tcp_autosndbuf_max = 2*1024*1024;
5b0b9fa5
PA
141SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
142 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
143
3d127502
SZ
144static int tcp_idle_cwv = 1;
145SYSCTL_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
149static int tcp_idle_restart = 1;
150SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
151 &tcp_idle_restart, 0, "Reset congestion window after idle period");
152
5f60906c
SZ
153static int tcp_do_tso = 1;
154SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
155 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
156
3d127502
SZ
157static void tcp_idle_cwnd_validate(struct tcpcb *);
158
5f60906c
SZ
159static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
160
984263bc
MD
161/*
162 * Tcp output routine: figure out what should be sent and send it.
163 */
164int
f3f70f0d 165tcp_output(struct tcpcb *tp)
984263bc 166{
0770aa24
JH
167 struct inpcb * const inp = tp->t_inpcb;
168 struct socket *so = inp->inp_socket;
c53f07a5 169 long len, recvwin, sendwin;
9a5b142f 170 int nsacked = 0;
1ff9b7d3 171 int off, flags, error = 0;
b1992928
MD
172#ifdef TCP_SIGNATURE
173 int sigoff = 0;
174#endif
46fc0bdc
SZ
175 struct mbuf *m;
176 struct ip *ip;
46fc0bdc 177 struct tcphdr *th;
984263bc 178 u_char opt[TCP_MAXOLEN];
91489f6b 179 unsigned int ipoptlen, optlen, hdrlen;
5f60906c 180 int idle;
61896e3c 181 boolean_t sendalot;
46fc0bdc 182 struct ip6_hdr *ip6;
984263bc 183#ifdef INET6
0770aa24
JH
184 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
185#else
186 const boolean_t isipv6 = FALSE;
984263bc 187#endif
5f60906c
SZ
188 boolean_t can_tso = FALSE, use_tso;
189 boolean_t report_sack, idle_cwv = FALSE;
e41e61d5 190 u_int segsz, tso_hlen, tso_lenmax = 0;
984263bc 191
2d23a8be
MD
192 KKASSERT(so->so_port == &curthread->td_msgport);
193
984263bc
MD
194 /*
195 * Determine length of data that should be transmitted,
196 * and flags that will be used.
197 * If there is some data or critical controls (SYN, RST)
198 * to send, then transmit; otherwise, investigate further.
199 */
8d573eef
JH
200
201 /*
202 * If we have been idle for a while, the send congestion window
3d127502
SZ
203 * could be no longer representative of the current state of the
204 * link; need to validate congestion window. However, we should
205 * not perform congestion window validation here, since we could
206 * be asked to send pure ACK.
8d573eef
JH
207 */
208 if (tp->snd_max == tp->snd_una &&
3d127502 209 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
5f60906c 210 idle_cwv = TRUE;
653eb2c0
MD
211
212 /*
213 * Calculate whether the transmit stream was previously idle
214 * and adjust TF_LASTIDLE for the next time.
215 */
08525cf0
JH
216 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
217 if (idle && (tp->t_flags & TF_MORETOCOME))
218 tp->t_flags |= TF_LASTIDLE;
219 else
220 tp->t_flags &= ~TF_LASTIDLE;
79a80d9f 221
9a5b142f
JH
222 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
223 !IN_FASTRECOVERY(tp))
224 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
225
5f60906c
SZ
226 /*
227 * Find out whether TSO could be used or not
228 *
229 * For TSO capable devices, the following assumptions apply to
230 * the processing of TCP flags:
231 * - If FIN is set on the large TCP segment, the device must set
232 * FIN on the last segment that it creates from the large TCP
233 * segment.
234 * - If PUSH is set on the large TCP segment, the device must set
235 * PUSH on the last segment that it creates from the large TCP
236 * segment.
237 */
238#if !defined(IPSEC) && !defined(FAST_IPSEC)
239 if (tcp_do_tso
240#ifdef TCP_SIGNATURE
241 && (tp->t_flags & TF_SIGNATURE) == 0
242#endif
243 ) {
244 if (!isipv6) {
245 struct rtentry *rt = inp->inp_route.ro_rt;
246
247 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
e41e61d5 248 (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
5f60906c 249 can_tso = TRUE;
e41e61d5
SZ
250 tso_lenmax = rt->rt_ifp->if_tsolen;
251 }
5f60906c
SZ
252 }
253 }
254#endif /* !IPSEC && !FAST_IPSEC */
255
984263bc 256again:
46fc0bdc
SZ
257 m = NULL;
258 ip = NULL;
46fc0bdc
SZ
259 th = NULL;
260 ip6 = NULL;
261
5f60906c
SZ
262 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
263 TF_SACK_PERMITTED &&
264 (!TAILQ_EMPTY(&tp->t_segq) ||
265 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
266 report_sack = TRUE;
267 else
268 report_sack = FALSE;
269
91489f6b
JH
270 /* Make use of SACK information when slow-starting after a RTO. */
271 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
9a5b142f
JH
272 !IN_FASTRECOVERY(tp)) {
273 tcp_seq old_snd_nxt = tp->snd_nxt;
274
91489f6b 275 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
9a5b142f
JH
276 nsacked += tp->snd_nxt - old_snd_nxt;
277 }
91489f6b 278
61896e3c 279 sendalot = FALSE;
984263bc 280 off = tp->snd_nxt - tp->snd_una;
9a5b142f 281 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
c53f07a5 282 sendwin = min(sendwin, tp->snd_bwnd);
984263bc
MD
283
284 flags = tcp_outflags[tp->t_state];
285 /*
286 * Get standard flags, and add SYN or FIN if requested by 'hidden'
287 * state flags.
288 */
289 if (tp->t_flags & TF_NEEDFIN)
290 flags |= TH_FIN;
291 if (tp->t_flags & TF_NEEDSYN)
292 flags |= TH_SYN;
293
294 /*
295 * If in persist timeout with window of 0, send 1 byte.
296 * Otherwise, if window is small but nonzero
297 * and timer expired, we will send what we can
298 * and go to transmit state.
299 */
c1d3b892 300 if (tp->t_flags & TF_FORCE) {
c53f07a5 301 if (sendwin == 0) {
984263bc
MD
302 /*
303 * If we still have some data to send, then
304 * clear the FIN bit. Usually this would
305 * happen below when it realizes that we
306 * aren't sending all the data. However,
307 * if we have exactly 1 byte of unsent data,
308 * then it won't clear the FIN bit below,
309 * and if we are in persist state, we wind
310 * up sending the packet without recording
311 * that we sent the FIN bit.
312 *
313 * We can't just blindly clear the FIN bit,
314 * because if we don't have any more data
315 * to send then the probe will be the FIN
316 * itself.
317 */
6d49aa6f 318 if (off < so->so_snd.ssb_cc)
984263bc 319 flags &= ~TH_FIN;
c53f07a5 320 sendwin = 1;
984263bc 321 } else {
a48c5dd5 322 tcp_callout_stop(tp, tp->tt_persist);
984263bc
MD
323 tp->t_rxtshift = 0;
324 }
325 }
326
327 /*
0770aa24 328 * If snd_nxt == snd_max and we have transmitted a FIN, the
6d49aa6f 329 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
c53f07a5 330 * a negative length. This can also occur when TCP opens up
984263bc
MD
331 * its congestion window while receiving additional duplicate
332 * acks after fast-retransmit because TCP will reset snd_nxt
333 * to snd_max after the fast-retransmit.
334 *
55696211
MD
335 * A negative length can also occur when we are in the
336 * TCPS_SYN_RECEIVED state due to a simultanious connect where
337 * our SYN has not been acked yet.
338 *
984263bc
MD
339 * In the normal retransmit-FIN-only case, however, snd_nxt will
340 * be set to snd_una, the offset will be 0, and the length may
341 * wind up 0.
342 */
6d49aa6f 343 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
984263bc 344
984263bc
MD
345 /*
346 * Lop off SYN bit if it has already been sent. However, if this
27b8aee3
AE
347 * is SYN-SENT state and if segment contains data, suppress sending
348 * segment (sending the segment would be an option if we still
349 * did TAO and the remote host supported it).
984263bc
MD
350 */
351 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
352 flags &= ~TH_SYN;
353 off--, len++;
3718f371 354 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
c1fabe85 355 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
984263bc 356 return 0;
3718f371 357 }
984263bc
MD
358 }
359
360 /*
27b8aee3 361 * Be careful not to send data and/or FIN on SYN segments.
984263bc
MD
362 * This measure is needed to prevent interoperability problems
363 * with not fully conformant TCP implementations.
364 */
27b8aee3 365 if (flags & TH_SYN) {
984263bc
MD
366 len = 0;
367 flags &= ~TH_FIN;
368 }
369
370 if (len < 0) {
371 /*
55696211
MD
372 * A negative len can occur if our FIN has been sent but not
373 * acked, or if we are in a simultanious connect in the
374 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
375 * acked.
376 *
377 * If our window has contracted to 0 in the FIN case
378 * (which can only occur if we have NOT been called to
379 * retransmit as per code a few paragraphs up) then we
380 * want to shift the retransmit timer over to the
381 * persist timer.
382 *
383 * However, if we are in the TCPS_SYN_RECEIVED state
384 * (the SYN case) we will be in a simultanious connect and
385 * the window may be zero degeneratively. In this case we
386 * do not want to shift to the persist timer after the SYN
387 * or the SYN+ACK transmission.
984263bc
MD
388 */
389 len = 0;
55696211 390 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
a48c5dd5 391 tcp_callout_stop(tp, tp->tt_rexmt);
984263bc
MD
392 tp->t_rxtshift = 0;
393 tp->snd_nxt = tp->snd_una;
a48c5dd5 394 if (!tcp_callout_active(tp, tp->tt_persist))
984263bc
MD
395 tcp_setpersist(tp);
396 }
397 }
398
5b0b9fa5 399 KASSERT(len >= 0, ("%s: len < 0", __func__));
984263bc 400 /*
5b0b9fa5
PA
401 * Automatic sizing of send socket buffer. Often the send buffer
402 * size is not optimally adjusted to the actual network conditions
403 * at hand (delay bandwidth product). Setting the buffer size too
404 * small limits throughput on links with high bandwidth and high
405 * delay (eg. trans-continental/oceanic links). Setting the
406 * buffer size too big consumes too much real kernel memory,
407 * especially with many connections on busy servers.
408 *
409 * The criteria to step up the send buffer one notch are:
410 * 1. receive window of remote host is larger than send buffer
411 * (with a fudge factor of 5/4th);
412 * 2. send buffer is filled to 7/8th with data (so we actually
413 * have data to make use of it);
414 * 3. send buffer fill has not hit maximal automatic size;
415 * 4. our send window (slow start and cogestion controlled) is
416 * larger than sent but unacknowledged data in send buffer.
417 *
418 * The remote host receive window scaling factor may limit the
419 * growing of the send buffer before it reaches its allowed
420 * maximum.
421 *
422 * It scales directly with slow start or congestion window
423 * and does at most one step per received ACK. This fast
424 * scaling has the drawback of growing the send buffer beyond
425 * what is strictly necessary to make full use of a given
426 * delay*bandwith product. However testing has shown this not
427 * to be much of an problem. At worst we are trading wasting
428 * of available bandwith (the non-use of it) for wasting some
429 * socket buffer memory.
430 *
431 * TODO: Shrink send buffer during idle periods together
432 * with congestion window. Requires another timer. Has to
433 * wait for upcoming tcp timer rewrite.
434 */
435 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
436 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
437 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
438 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
439 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
46e92930
MD
440 u_long newsize;
441
442 newsize = ulmin(so->so_snd.ssb_hiwat +
443 tcp_autosndbuf_inc,
444 tcp_autosndbuf_max);
445 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
14343ad3 446 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
46e92930 447 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
14343ad3 448 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
5b0b9fa5
PA
449 }
450 }
451
452 /*
5f60906c
SZ
453 * Don't use TSO, if:
454 * - Congestion window needs validation
455 * - There are SACK blocks to report
456 * - RST or SYN flags is set
457 * - URG will be set
458 *
459 * XXX
460 * Checking for SYN|RST looks overkill, just to be safe than sorry
461 */
462 use_tso = can_tso;
463 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
464 use_tso = FALSE;
465 if (use_tso) {
466 tcp_seq ugr_nxt = tp->snd_nxt;
467
468 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
469 tp->snd_nxt == tp->snd_max)
470 --ugr_nxt;
471
472 if (SEQ_GT(tp->snd_up, ugr_nxt))
473 use_tso = FALSE;
474 }
475
476 if (use_tso) {
477 /*
478 * Find out segment size and header length for TSO
479 */
480 error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
481 if (error)
482 use_tso = FALSE;
483 }
484 if (!use_tso) {
485 segsz = tp->t_maxseg;
486 tso_hlen = 0; /* not used */
487 }
488
489 /*
490 * Truncate to the maximum segment length if not TSO, and ensure that
491 * FIN is removed if the length no longer contains the last data byte.
984263bc 492 */
5f60906c
SZ
493 if (len > segsz) {
494 if (!use_tso) {
495 len = segsz;
496 } else {
e41e61d5
SZ
497 if (__predict_false(tso_lenmax < segsz))
498 tso_lenmax = segsz << 1;
499
5f60906c
SZ
500 /*
501 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
502 * thoff), and make sure that we send equal size
503 * transfers down the stack (rather than big-small-
504 * big-small-...).
505 */
e41e61d5 506 len = min(len, tso_lenmax);
5f60906c
SZ
507 len = (min(len, (IP_MAXPACKET - tso_hlen)) / segsz) *
508 segsz;
509 if (len <= segsz)
510 use_tso = FALSE;
511 }
61896e3c 512 sendalot = TRUE;
5f60906c
SZ
513 } else {
514 use_tso = FALSE;
984263bc 515 }
6d49aa6f 516 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
984263bc
MD
517 flags &= ~TH_FIN;
518
6d49aa6f 519 recvwin = ssb_space(&so->so_rcv);
984263bc
MD
520
521 /*
522 * Sender silly window avoidance. We transmit under the following
523 * conditions when len is non-zero:
524 *
525 * - We have a full segment
526 * - This is the last buffer in a write()/send() and we are
527 * either idle or running NODELAY
528 * - we've timed out (e.g. persist timer)
529 * - we have more then 1/2 the maximum send window's worth of
653eb2c0 530 * data (receiver may be limiting the window size)
984263bc
MD
531 * - we need to retransmit
532 */
533 if (len) {
5f60906c 534 if (len >= segsz)
984263bc
MD
535 goto send;
536 /*
537 * NOTE! on localhost connections an 'ack' from the remote
538 * end may occur synchronously with the output and cause
539 * us to flush a buffer queued with moretocome. XXX
540 *
541 * note: the len + off check is almost certainly unnecessary.
542 */
543 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
544 (idle || (tp->t_flags & TF_NODELAY)) &&
6d49aa6f 545 len + off >= so->so_snd.ssb_cc &&
0770aa24 546 !(tp->t_flags & TF_NOPUSH)) {
984263bc
MD
547 goto send;
548 }
c1d3b892 549 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
984263bc
MD
550 goto send;
551 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
552 goto send;
553 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
554 goto send;
c1fabe85
SZ
555 if (tp->t_flags & TF_XMITNOW)
556 goto send;
984263bc
MD
557 }
558
559 /*
560 * Compare available window to amount of window
561 * known to peer (as advertised window less
562 * next expected input). If the difference is at least two
563 * max size segments, or at least 50% of the maximum possible
564 * window, then want to send a window update to peer.
565 */
c53f07a5 566 if (recvwin > 0) {
984263bc
MD
567 /*
568 * "adv" is the amount we can increase the window,
569 * taking into account that we are limited by
570 * TCP_MAXWIN << tp->rcv_scale.
571 */
c53f07a5 572 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
984263bc 573 (tp->rcv_adv - tp->rcv_nxt);
46e92930 574 long hiwat;
984263bc 575
ddb2e271
MD
576 /*
577 * This ack case typically occurs when the user has drained
578 * the TCP socket buffer sufficiently to warrent an ack
579 * containing a 'pure window update'... that is, an ack that
580 * ONLY updates the tcp window.
581 *
582 * It is unclear why we would need to do a pure window update
583 * past 2 segments if we are going to do one at 1/2 the high
584 * water mark anyway, especially since under normal conditions
f23061d4 585 * the user program will drain the socket buffer quickly.
ddb2e271
MD
586 * The 2-segment pure window update will often add a large
587 * number of extra, unnecessary acks to the stream.
588 *
589 * avoid_pure_win_update now defaults to 1.
590 */
46e92930
MD
591 if (avoid_pure_win_update == 0 ||
592 (tp->t_flags & TF_RXRESIZED)) {
5f60906c 593 if (adv >= (long) (2 * segsz)) {
ddb2e271 594 goto send;
46e92930 595 }
ddb2e271 596 }
46e92930
MD
597 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
598 if (hiwat > (long)so->so_rcv.ssb_hiwat)
599 hiwat = (long)so->so_rcv.ssb_hiwat;
600 if (adv >= hiwat / 2)
984263bc
MD
601 goto send;
602 }
603
604 /*
605 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
606 * is also a catch-all for the retransmit timer timeout case.
607 */
608 if (tp->t_flags & TF_ACKNOW)
609 goto send;
610 if ((flags & TH_RST) ||
0770aa24 611 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
984263bc
MD
612 goto send;
613 if (SEQ_GT(tp->snd_up, tp->snd_una))
614 goto send;
615 /*
616 * If our state indicates that FIN should be sent
617 * and we have not yet done so, then we need to send.
618 */
7462c355 619 if ((flags & TH_FIN) &&
0770aa24 620 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
984263bc
MD
621 goto send;
622
623 /*
624 * TCP window updates are not reliable, rather a polling protocol
625 * using ``persist'' packets is used to insure receipt of window
626 * updates. The three ``states'' for the output side are:
627 * idle not doing retransmits or persists
628 * persisting to move a small or zero window
629 * (re)transmitting and thereby not persisting
630 *
a48c5dd5 631 * tcp_callout_active(tp, tp->tt_persist)
984263bc 632 * is true when we are in persist state.
c1d3b892 633 * The TF_FORCE flag in tp->t_flags
984263bc 634 * is set when we are called to send a persist packet.
a48c5dd5 635 * tcp_callout_active(tp, tp->tt_rexmt)
984263bc
MD
636 * is set when we are retransmitting
637 * The output side is idle when both timers are zero.
638 *
639 * If send window is too small, there is data to transmit, and no
640 * retransmit or persist is pending, then go to persist state.
7462c355 641 *
984263bc 642 * If nothing happens soon, send when timer expires:
7462c355
MD
643 * if window is nonzero, transmit what we can, otherwise force out
644 * a byte.
645 *
646 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
647 * with data pending. This situation can occur during a
648 * simultanious connect.
984263bc 649 */
6d49aa6f 650 if (so->so_snd.ssb_cc > 0 &&
7462c355 651 tp->t_state != TCPS_SYN_RECEIVED &&
a48c5dd5
SZ
652 !tcp_callout_active(tp, tp->tt_rexmt) &&
653 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
654 tp->t_rxtshift = 0;
655 tcp_setpersist(tp);
656 }
657
658 /*
659 * No reason to send a segment, just return.
660 */
c1fabe85 661 tp->t_flags &= ~TF_XMITNOW;
984263bc
MD
662 return (0);
663
664send:
665 /*
666 * Before ESTABLISHED, force sending of initial options
667 * unless TCP set not to do any options.
668 * NOTE: we assume that the IP/TCP header plus TCP options
669 * always fit in a single mbuf, leaving room for a maximum
670 * link header, i.e.
79a80d9f 671 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
984263bc
MD
672 */
673 optlen = 0;
984263bc 674 if (isipv6)
79a80d9f 675 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
984263bc 676 else
79a80d9f 677 hdrlen = sizeof(struct tcpiphdr);
984263bc
MD
678 if (flags & TH_SYN) {
679 tp->snd_nxt = tp->iss;
0770aa24 680 if (!(tp->t_flags & TF_NOOPT)) {
984263bc
MD
681 u_short mss;
682
683 opt[0] = TCPOPT_MAXSEG;
684 opt[1] = TCPOLEN_MAXSEG;
685 mss = htons((u_short) tcp_mssopt(tp));
f23061d4 686 memcpy(opt + 2, &mss, sizeof mss);
984263bc
MD
687 optlen = TCPOLEN_MAXSEG;
688
689 if ((tp->t_flags & TF_REQ_SCALE) &&
0770aa24
JH
690 (!(flags & TH_ACK) ||
691 (tp->t_flags & TF_RCVD_SCALE))) {
984263bc
MD
692 *((u_int32_t *)(opt + optlen)) = htonl(
693 TCPOPT_NOP << 24 |
694 TCPOPT_WINDOW << 16 |
695 TCPOLEN_WINDOW << 8 |
696 tp->request_r_scale);
697 optlen += 4;
698 }
91489f6b
JH
699
700 if ((tcp_do_sack && !(flags & TH_ACK)) ||
701 tp->t_flags & TF_SACK_PERMITTED) {
702 uint32_t *lp = (uint32_t *)(opt + optlen);
703
704 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
705 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
706 }
984263bc 707 }
0770aa24 708 }
984263bc 709
0770aa24 710 /*
984263bc
MD
711 * Send a timestamp and echo-reply if this is a SYN and our side
712 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
713 * and our peer have sent timestamps in our SYN's.
0770aa24 714 */
61896e3c 715 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
0770aa24
JH
716 !(flags & TH_RST) &&
717 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
984263bc
MD
718 u_int32_t *lp = (u_int32_t *)(opt + optlen);
719
0770aa24
JH
720 /* Form timestamp option as shown in appendix A of RFC 1323. */
721 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
722 *lp++ = htonl(ticks);
723 *lp = htonl(tp->ts_recent);
724 optlen += TCPOLEN_TSTAMP_APPA;
725 }
984263bc 726
5b0b9fa5
PA
727 /* Set receive buffer autosizing timestamp. */
728 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
729 tp->rfbuf_ts = ticks;
730
0770aa24 731 /*
91489f6b
JH
732 * If this is a SACK connection and we have a block to report,
733 * fill in the SACK blocks in the TCP options.
734 */
5f60906c 735 if (report_sack)
91489f6b
JH
736 tcp_sack_fill_report(tp, opt, &optlen);
737
b1992928
MD
738#ifdef TCP_SIGNATURE
739 if (tp->t_flags & TF_SIGNATURE) {
740 int i;
741 u_char *bp;
742 /*
743 * Initialize TCP-MD5 option (RFC2385)
744 */
745 bp = (u_char *)opt + optlen;
746 *bp++ = TCPOPT_SIGNATURE;
747 *bp++ = TCPOLEN_SIGNATURE;
748 sigoff = optlen + 2;
749 for (i = 0; i < TCP_SIGLEN; i++)
750 *bp++ = 0;
751 optlen += TCPOLEN_SIGNATURE;
752 /*
753 * Terminate options list and maintain 32-bit alignment.
754 */
755 *bp++ = TCPOPT_NOP;
756 *bp++ = TCPOPT_EOL;
757 optlen += 2;
758 }
759#endif /* TCP_SIGNATURE */
91489f6b 760 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
0770aa24 761 hdrlen += optlen;
984263bc 762
61896e3c 763 if (isipv6) {
0770aa24 764 ipoptlen = ip6_optlen(inp);
61896e3c 765 } else {
0770aa24
JH
766 if (inp->inp_options) {
767 ipoptlen = inp->inp_options->m_len -
768 offsetof(struct ipoption, ipopt_list);
769 } else {
770 ipoptlen = 0;
771 }
984263bc 772 }
984263bc
MD
773#ifdef IPSEC
774 ipoptlen += ipsec_hdrsiz_tcp(tp);
775#endif
776
5f60906c 777 if (use_tso) {
243cd031
SZ
778 /* TSO segment length must be multiple of segment size */
779 KASSERT(len >= (2 * segsz) && (len % segsz == 0),
5f60906c
SZ
780 ("invalid TSO len %ld, segsz %u", len, segsz));
781 } else {
782 KASSERT(len <= segsz,
783 ("invalid len %ld, segsz %u", len, segsz));
45f20ce7 784
5f60906c
SZ
785 /*
786 * Adjust data length if insertion of options will bump
787 * the packet length beyond the t_maxopd length. Clear
788 * FIN to prevent premature closure since there is still
789 * more data to send after this (now truncated) packet.
790 *
791 * If just the options do not fit we are in a no-win
792 * situation and we treat it as an unreachable host.
793 */
794 if (len + optlen + ipoptlen > tp->t_maxopd) {
795 if (tp->t_maxopd <= optlen + ipoptlen) {
796 static time_t last_optlen_report;
797
798 if (last_optlen_report != time_second) {
799 last_optlen_report = time_second;
800 kprintf("tcpcb %p: MSS (%d) too "
801 "small to hold options!\n",
802 tp, tp->t_maxopd);
803 }
804 error = EHOSTUNREACH;
805 goto out;
806 } else {
807 flags &= ~TH_FIN;
808 len = tp->t_maxopd - optlen - ipoptlen;
809 sendalot = TRUE;
45f20ce7 810 }
45f20ce7 811 }
984263bc
MD
812 }
813
984263bc 814#ifdef INET6
0770aa24 815 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
984263bc 816#else
0770aa24 817 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
984263bc 818#endif
984263bc
MD
819
820 /*
821 * Grab a header mbuf, attaching a copy of data to
822 * be transmitted, and initialize the header from
823 * the template for sends on this connection.
824 */
825 if (len) {
c1d3b892 826 if ((tp->t_flags & TF_FORCE) && len == 1)
984263bc
MD
827 tcpstat.tcps_sndprobe++;
828 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
91489f6b
JH
829 if (tp->snd_nxt == tp->snd_una)
830 tp->snd_max_rexmt = tp->snd_max;
02cc2f35
SZ
831 if (nsacked) {
832 tcpstat.tcps_sndsackrtopack++;
833 tcpstat.tcps_sndsackrtobyte += len;
834 }
984263bc
MD
835 tcpstat.tcps_sndrexmitpack++;
836 tcpstat.tcps_sndrexmitbyte += len;
837 } else {
838 tcpstat.tcps_sndpack++;
839 tcpstat.tcps_sndbyte += len;
840 }
3d127502 841 if (idle_cwv) {
5f60906c 842 idle_cwv = FALSE;
3d127502
SZ
843 tcp_idle_cwnd_validate(tp);
844 }
845 /* Update last send time after CWV */
846 tp->snd_last = ticks;
984263bc 847#ifdef notyet
6d49aa6f 848 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
61896e3c 849 max_linkhdr + hdrlen)) == NULL) {
984263bc 850 error = ENOBUFS;
1ff9b7d3 851 goto after_th;
984263bc
MD
852 }
853 /*
854 * m_copypack left space for our hdr; use it.
855 */
856 m->m_len += hdrlen;
857 m->m_data -= hdrlen;
858#else
50503f0f
JH
859#ifndef INET6
860 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
861#else
862 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
863 M_PKTHDR, NULL);
864#endif
984263bc
MD
865 if (m == NULL) {
866 error = ENOBUFS;
1ff9b7d3 867 goto after_th;
984263bc 868 }
984263bc
MD
869 m->m_data += max_linkhdr;
870 m->m_len = hdrlen;
871 if (len <= MHLEN - hdrlen - max_linkhdr) {
6d49aa6f 872 m_copydata(so->so_snd.ssb_mb, off, (int) len,
984263bc
MD
873 mtod(m, caddr_t) + hdrlen);
874 m->m_len += len;
875 } else {
6d49aa6f 876 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
61896e3c 877 if (m->m_next == NULL) {
f23061d4 878 m_free(m);
1ff9b7d3 879 m = NULL;
984263bc 880 error = ENOBUFS;
1ff9b7d3 881 goto after_th;
984263bc
MD
882 }
883 }
884#endif
885 /*
886 * If we're sending everything we've got, set PUSH.
887 * (This will keep happy those implementations which only
888 * give data to the user when a buffer fills or
889 * a PUSH comes in.)
890 */
6d49aa6f 891 if (off + len == so->so_snd.ssb_cc)
984263bc
MD
892 flags |= TH_PUSH;
893 } else {
894 if (tp->t_flags & TF_ACKNOW)
895 tcpstat.tcps_sndacks++;
61896e3c 896 else if (flags & (TH_SYN | TH_FIN | TH_RST))
984263bc
MD
897 tcpstat.tcps_sndctrl++;
898 else if (SEQ_GT(tp->snd_up, tp->snd_una))
899 tcpstat.tcps_sndurg++;
900 else
901 tcpstat.tcps_sndwinup++;
902
74f1caca 903 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
984263bc
MD
904 if (m == NULL) {
905 error = ENOBUFS;
1ff9b7d3 906 goto after_th;
984263bc 907 }
0770aa24
JH
908 if (isipv6 &&
909 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
984263bc 910 MH_ALIGN(m, hdrlen);
0770aa24
JH
911 else
912 m->m_data += max_linkhdr;
984263bc
MD
913 m->m_len = hdrlen;
914 }
60233e58 915 m->m_pkthdr.rcvif = NULL;
984263bc
MD
916 if (isipv6) {
917 ip6 = mtod(m, struct ip6_hdr *);
918 th = (struct tcphdr *)(ip6 + 1);
5f60906c 919 tcp_fillheaders(tp, ip6, th, use_tso);
0770aa24
JH
920 } else {
921 ip = mtod(m, struct ip *);
0770aa24
JH
922 th = (struct tcphdr *)(ip + 1);
923 /* this picks up the pseudo header (w/o the length) */
5f60906c 924 tcp_fillheaders(tp, ip, th, use_tso);
0770aa24 925 }
1ff9b7d3 926after_th:
984263bc
MD
927 /*
928 * Fill in fields, remembering maximum advertised
929 * window for use in delaying messages about window sizes.
930 * If resending a FIN, be sure not to use a new sequence number.
931 */
932 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
933 tp->snd_nxt == tp->snd_max)
934 tp->snd_nxt--;
1ff9b7d3
SZ
935
936 if (th != NULL) {
937 /*
938 * If we are doing retransmissions, then snd_nxt will
939 * not reflect the first unsent octet. For ACK only
940 * packets, we do not want the sequence number of the
941 * retransmitted packet, we want the sequence number
942 * of the next unsent octet. So, if there is no data
943 * (and no SYN or FIN), use snd_max instead of snd_nxt
944 * when filling in ti_seq. But if we are in persist
945 * state, snd_max might reflect one byte beyond the
946 * right edge of the window, so use snd_nxt in that
947 * case, since we know we aren't doing a retransmission.
948 * (retransmit and persist are mutually exclusive...)
949 */
950 if (len || (flags & (TH_SYN|TH_FIN)) ||
951 tcp_callout_active(tp, tp->tt_persist))
952 th->th_seq = htonl(tp->snd_nxt);
953 else
954 th->th_seq = htonl(tp->snd_max);
955 th->th_ack = htonl(tp->rcv_nxt);
956 if (optlen) {
957 bcopy(opt, th + 1, optlen);
958 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
959 }
960 th->th_flags = flags;
984263bc 961 }
46e92930 962
984263bc 963 /*
46e92930
MD
964 * Calculate receive window. Don't shrink window, but avoid
965 * silly window syndrome by sending a 0 window if the actual
966 * window is less then one segment.
984263bc 967 */
6d49aa6f 968 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
5f60906c 969 recvwin < (long)segsz)
c53f07a5 970 recvwin = 0;
55126001
MD
971 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
972 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
c53f07a5
JH
973 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
974 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
984263bc
MD
975
976 /*
977 * Adjust the RXWIN0SENT flag - indicate that we have advertised
978 * a 0 window. This may cause the remote transmitter to stall. This
979 * flag tells soreceive() to disable delayed acknowledgements when
980 * draining the buffer. This can occur if the receiver is attempting
981 * to read more data then can be buffered prior to transmitting on
982 * the connection.
983 */
c53f07a5 984 if (recvwin == 0)
984263bc
MD
985 tp->t_flags |= TF_RXWIN0SENT;
986 else
987 tp->t_flags &= ~TF_RXWIN0SENT;
988
1ff9b7d3
SZ
989 if (th != NULL)
990 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
991
984263bc 992 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
5f60906c 993 KASSERT(!use_tso, ("URG with TSO"));
1ff9b7d3
SZ
994 if (th != NULL) {
995 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
996 th->th_flags |= TH_URG;
997 }
79a80d9f 998 } else {
984263bc
MD
999 /*
1000 * If no urgent pointer to send, then we pull
1001 * the urgent pointer to the left edge of the send window
1002 * so that it doesn't drift into the send window on sequence
1003 * number wraparound.
1004 */
1005 tp->snd_up = tp->snd_una; /* drag it along */
79a80d9f 1006 }
984263bc 1007
1ff9b7d3 1008 if (th != NULL) {
b1992928 1009#ifdef TCP_SIGNATURE
1ff9b7d3
SZ
1010 if (tp->t_flags & TF_SIGNATURE) {
1011 tcpsignature_compute(m, len, optlen,
1012 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1013 }
b1992928
MD
1014#endif /* TCP_SIGNATURE */
1015
984263bc 1016 /*
1ff9b7d3
SZ
1017 * Put TCP length in extended header, and then
1018 * checksum extended header and data.
984263bc 1019 */
1ff9b7d3
SZ
1020 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1021 if (isipv6) {
1022 /*
1023 * ip6_plen is not need to be filled now, and will be
1024 * filled in ip6_output().
1025 */
1026 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1027 sizeof(struct ip6_hdr),
1028 sizeof(struct tcphdr) + optlen + len);
1029 } else {
7df36335 1030 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
5f60906c
SZ
1031 if (use_tso) {
1032 m->m_pkthdr.csum_flags = CSUM_TSO;
f0336d39 1033 m->m_pkthdr.tso_segsz = segsz;
5f60906c
SZ
1034 } else {
1035 m->m_pkthdr.csum_flags = CSUM_TCP;
7df36335
SZ
1036 m->m_pkthdr.csum_data =
1037 offsetof(struct tcphdr, th_sum);
5f60906c
SZ
1038 if (len + optlen) {
1039 th->th_sum = in_addword(th->th_sum,
1040 htons((u_short)(optlen + len)));
1041 }
1ff9b7d3
SZ
1042 }
1043
1044 /*
1045 * IP version must be set here for ipv4/ipv6 checking
1046 * later
1047 */
1048 KASSERT(ip->ip_v == IPVERSION,
1049 ("%s: IP version incorrect: %d",
1050 __func__, ip->ip_v));
1051 }
0770aa24 1052 }
984263bc
MD
1053
1054 /*
1055 * In transmit state, time the transmission and arrange for
1056 * the retransmit. In persist state, just set snd_max.
1057 */
a48c5dd5
SZ
1058 if (!(tp->t_flags & TF_FORCE) ||
1059 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
1060 tcp_seq startseq = tp->snd_nxt;
1061
1062 /*
1063 * Advance snd_nxt over sequence space of this segment.
1064 */
61896e3c 1065 if (flags & (TH_SYN | TH_FIN)) {
984263bc
MD
1066 if (flags & TH_SYN)
1067 tp->snd_nxt++;
1068 if (flags & TH_FIN) {
1069 tp->snd_nxt++;
1070 tp->t_flags |= TF_SENTFIN;
1071 }
1072 }
1073 tp->snd_nxt += len;
1074 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1075 tp->snd_max = tp->snd_nxt;
1076 /*
1077 * Time this transmission if not a retransmission and
1078 * not currently timing anything.
1079 */
1080 if (tp->t_rtttime == 0) {
1081 tp->t_rtttime = ticks;
1082 tp->t_rtseq = startseq;
1083 tcpstat.tcps_segstimed++;
1084 }
1085 }
1086
1087 /*
1088 * Set retransmit timer if not currently set,
1089 * and not doing a pure ack or a keep-alive probe.
1090 * Initial value for retransmit timer is smoothed
1091 * round-trip time + 2 * round-trip time variance.
1092 * Initialize shift counter which is used for backoff
1093 * of retransmit time.
1094 */
a48c5dd5 1095 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
984263bc 1096 tp->snd_nxt != tp->snd_una) {
a48c5dd5
SZ
1097 if (tcp_callout_active(tp, tp->tt_persist)) {
1098 tcp_callout_stop(tp, tp->tt_persist);
984263bc
MD
1099 tp->t_rxtshift = 0;
1100 }
a48c5dd5
SZ
1101 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1102 tcp_timer_rexmt);
984263bc
MD
1103 }
1104 } else {
1105 /*
1106 * Persist case, update snd_max but since we are in
1107 * persist mode (no window) we do not update snd_nxt.
1108 */
1109 int xlen = len;
1110 if (flags & TH_SYN)
ed20d0e3 1111 panic("tcp_output: persist timer to send SYN");
984263bc
MD
1112 if (flags & TH_FIN) {
1113 ++xlen;
1114 tp->t_flags |= TF_SENTFIN;
1115 }
1116 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1117 tp->snd_max = tp->snd_nxt + xlen;
1118 }
1119
1ff9b7d3 1120 if (th != NULL) {
984263bc 1121#ifdef TCPDEBUG
1ff9b7d3
SZ
1122 /* Trace. */
1123 if (so->so_options & SO_DEBUG) {
1124 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1125 mtod(m, void *), th, 0);
1126 }
984263bc
MD
1127#endif
1128
984263bc 1129 /*
1ff9b7d3
SZ
1130 * Fill in IP length and desired time to live and
1131 * send to IP level. There should be a better way
1132 * to handle ttl and tos; we could keep them in
1133 * the template, but need a way to checksum without them.
984263bc 1134 */
1ff9b7d3
SZ
1135 /*
1136 * m->m_pkthdr.len should have been set before cksum
1137 * calcuration, because in6_cksum() need it.
1138 */
1139 if (isipv6) {
1140 /*
1141 * we separately set hoplimit for every segment,
1142 * since the user might want to change the value
1143 * via setsockopt. Also, desired default hop
1144 * limit might be changed via Neighbor Discovery.
1145 */
1146 ip6->ip6_hlim = in6_selecthlim(inp,
0770aa24
JH
1147 (inp->in6p_route.ro_rt ?
1148 inp->in6p_route.ro_rt->rt_ifp : NULL));
1ff9b7d3
SZ
1149
1150 /* TODO: IPv6 IP6TOS_ECT bit on */
1151 error = ip6_output(m, inp->in6p_outputopts,
1152 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1153 NULL, NULL, inp);
1154 } else {
1155 struct rtentry *rt;
1156 ip->ip_len = m->m_pkthdr.len;
1157#ifdef INET6
1158 if (INP_CHECK_SOCKAF(so, AF_INET6))
1159 ip->ip_ttl = in6_selecthlim(inp,
1160 (inp->in6p_route.ro_rt ?
1161 inp->in6p_route.ro_rt->rt_ifp : NULL));
1162 else
0770aa24 1163#endif
1ff9b7d3 1164 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
0770aa24 1165
1ff9b7d3
SZ
1166 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1167 /*
1168 * See if we should do MTU discovery.
1169 * We do it only if the following are true:
1170 * 1) we have a valid route to the destination
1171 * 2) the MTU is not locked (if it is,
1172 * then discovery has been disabled)
1173 */
1174 if (path_mtu_discovery &&
1175 (rt = inp->inp_route.ro_rt) &&
1176 (rt->rt_flags & RTF_UP) &&
1177 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1178 ip->ip_off |= IP_DF;
1179
1180 error = ip_output(m, inp->inp_options, &inp->inp_route,
1181 (so->so_options & SO_DONTROUTE) |
1182 IP_DEBUGROUTE, NULL, inp);
1183 }
1184 } else {
ed20d0e3 1185 KASSERT(error != 0, ("no error, but th not set"));
0770aa24 1186 }
984263bc 1187 if (error) {
c1fabe85 1188 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
984263bc
MD
1189
1190 /*
1191 * We know that the packet was lost, so back out the
1192 * sequence number advance, if any.
1193 */
c1d3b892 1194 if (!(tp->t_flags & TF_FORCE) ||
a48c5dd5 1195 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
1196 /*
1197 * No need to check for TH_FIN here because
1198 * the TF_SENTFIN flag handles that case.
1199 */
0770aa24 1200 if (!(flags & TH_SYN))
984263bc
MD
1201 tp->snd_nxt -= len;
1202 }
1203
1204out:
1205 if (error == ENOBUFS) {
ea8bd01a
JH
1206 /*
1207 * If we can't send, make sure there is something
28a9a7b6
MD
1208 * to get us going again later.
1209 *
1210 * The persist timer isn't necessarily allowed in all
1211 * states, use the rexmt timer.
ea8bd01a 1212 */
a48c5dd5
SZ
1213 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1214 !tcp_callout_active(tp, tp->tt_persist)) {
28a9a7b6
MD
1215 tcp_callout_reset(tp, tp->tt_rexmt,
1216 tp->t_rxtcur,
1217 tcp_timer_rexmt);
1218#if 0
ea8bd01a
JH
1219 tp->t_rxtshift = 0;
1220 tcp_setpersist(tp);
28a9a7b6 1221#endif
ea8bd01a 1222 }
0770aa24 1223 tcp_quench(inp, 0);
984263bc
MD
1224 return (0);
1225 }
1226 if (error == EMSGSIZE) {
1227 /*
1228 * ip_output() will have already fixed the route
1229 * for us. tcp_mtudisc() will, as its last action,
1230 * initiate retransmission, so it is important to
1231 * not do so here.
1232 */
0770aa24 1233 tcp_mtudisc(inp, 0);
984263bc
MD
1234 return 0;
1235 }
0770aa24
JH
1236 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1237 TCPS_HAVERCVDSYN(tp->t_state)) {
984263bc
MD
1238 tp->t_softerror = error;
1239 return (0);
1240 }
1241 return (error);
1242 }
1243 tcpstat.tcps_sndtotal++;
1244
1245 /*
1246 * Data sent (as far as we can tell).
46e92930 1247 *
984263bc
MD
1248 * If this advertises a larger window than any other segment,
1249 * then remember the size of the advertised window.
46e92930 1250 *
984263bc
MD
1251 * Any pending ACK has now been sent.
1252 */
46e92930 1253 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
c53f07a5 1254 tp->rcv_adv = tp->rcv_nxt + recvwin;
46e92930
MD
1255 tp->t_flags &= ~TF_RXRESIZED;
1256 }
984263bc 1257 tp->last_ack_sent = tp->rcv_nxt;
c1fabe85 1258 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
984263bc 1259 if (tcp_delack_enabled)
a48c5dd5 1260 tcp_callout_stop(tp, tp->tt_delack);
46fc0bdc 1261 if (sendalot)
984263bc
MD
1262 goto again;
1263 return (0);
1264}
1265
1266void
f3f70f0d 1267tcp_setpersist(struct tcpcb *tp)
984263bc
MD
1268{
1269 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1270 int tt;
1271
c529614e
SZ
1272 if (tp->t_state == TCPS_SYN_SENT ||
1273 tp->t_state == TCPS_SYN_RECEIVED) {
ed20d0e3 1274 panic("tcp_setpersist: not established yet, current %s",
c529614e
SZ
1275 tp->t_state == TCPS_SYN_SENT ?
1276 "SYN_SENT" : "SYN_RECEIVED");
1277 }
94603b1b 1278
a48c5dd5 1279 if (tcp_callout_active(tp, tp->tt_rexmt))
984263bc
MD
1280 panic("tcp_setpersist: retransmit pending");
1281 /*
1282 * Start/restart persistance timer.
1283 */
5fe66e68
JH
1284 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1285 TCPTV_PERSMAX);
a48c5dd5 1286 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
984263bc
MD
1287 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1288 tp->t_rxtshift++;
1289}
3d127502
SZ
1290
1291static void
1292tcp_idle_cwnd_validate(struct tcpcb *tp)
1293{
1294 u_long initial_cwnd = tcp_initial_window(tp);
1295 u_long min_cwnd;
1296
1297 tcpstat.tcps_sndidle++;
1298
1299 /* According to RFC5681: RW=min(IW,cwnd) */
1300 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1301
1302 if (tcp_idle_cwv) {
1303 u_long idle_time, decay_cwnd;
1304
1305 /*
1306 * RFC2861, but only after idle period.
1307 */
1308
1309 /*
1310 * Before the congestion window is reduced, ssthresh
1311 * is set to the maximum of its current value and 3/4
1312 * cwnd. If the sender then has more data to send
1313 * than the decayed cwnd allows, the TCP will slow-
1314 * start (perform exponential increase) at least
1315 * half-way back up to the old value of cwnd.
1316 */
1317 tp->snd_ssthresh = max(tp->snd_ssthresh,
1318 (3 * tp->snd_cwnd) / 4);
1319
1320 /*
1321 * Decay the congestion window by half for every RTT
1322 * that the flow remains inactive.
1323 *
1324 * The difference between our implementation and
1325 * RFC2861 is that we don't allow cwnd to go below
1326 * the value allowed by RFC5681 (min_cwnd).
1327 */
1328 idle_time = ticks - tp->snd_last;
1329 decay_cwnd = tp->snd_cwnd;
1330 while (idle_time >= tp->t_rxtcur &&
1331 decay_cwnd > min_cwnd) {
1332 decay_cwnd >>= 1;
1333 idle_time -= tp->t_rxtcur;
1334 }
1335 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1336 } else {
1337 /*
1338 * Slow-start from scratch to re-determine the send
1339 * congestion window.
1340 */
1341 tp->snd_cwnd = min_cwnd;
1342 }
1343
1344 /* Restart ABC counting during congestion avoidance */
1345 tp->snd_wacked = 0;
1346}
5f60906c
SZ
1347
1348static int
1349tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1350{
1351 struct inpcb * const inp = tp->t_inpcb;
1352#ifdef INET6
1353 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1354#else
1355 const boolean_t isipv6 = FALSE;
1356#endif
1357 unsigned int ipoptlen, optlen;
1358 u_int hlen;
1359
1360 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1361
1362 if (isipv6) {
1363 ipoptlen = ip6_optlen(inp);
1364 } else {
1365 if (inp->inp_options) {
1366 ipoptlen = inp->inp_options->m_len -
1367 offsetof(struct ipoption, ipopt_list);
1368 } else {
1369 ipoptlen = 0;
1370 }
1371 }
1372#ifdef IPSEC
1373 ipoptlen += ipsec_hdrsiz_tcp(tp);
1374#endif
1375 hlen += ipoptlen;
1376
1377 optlen = 0;
1378 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1379 (tp->t_flags & TF_RCVD_TSTMP))
1380 optlen += TCPOLEN_TSTAMP_APPA;
1381 hlen += optlen;
1382
1383 if (tp->t_maxopd <= optlen + ipoptlen)
1384 return EHOSTUNREACH;
1385
1386 *segsz = tp->t_maxopd - optlen - ipoptlen;
1387 *hlen0 = hlen;
1388 return 0;
1389}