tcp: Stringent TSO segment length assertion.
[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;
177 struct ipovly *ipov;
178 struct tcphdr *th;
984263bc 179 u_char opt[TCP_MAXOLEN];
91489f6b 180 unsigned int ipoptlen, optlen, hdrlen;
5f60906c 181 int idle;
61896e3c 182 boolean_t sendalot;
46fc0bdc 183 struct ip6_hdr *ip6;
984263bc 184#ifdef INET6
0770aa24
JH
185 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
186#else
187 const boolean_t isipv6 = FALSE;
984263bc 188#endif
5f60906c
SZ
189 boolean_t can_tso = FALSE, use_tso;
190 boolean_t report_sack, idle_cwv = FALSE;
191 u_int segsz, tso_hlen;
984263bc 192
2d23a8be
MD
193 KKASSERT(so->so_port == &curthread->td_msgport);
194
984263bc
MD
195 /*
196 * Determine length of data that should be transmitted,
197 * and flags that will be used.
198 * If there is some data or critical controls (SYN, RST)
199 * to send, then transmit; otherwise, investigate further.
200 */
8d573eef
JH
201
202 /*
203 * If we have been idle for a while, the send congestion window
3d127502
SZ
204 * could be no longer representative of the current state of the
205 * link; need to validate congestion window. However, we should
206 * not perform congestion window validation here, since we could
207 * be asked to send pure ACK.
8d573eef
JH
208 */
209 if (tp->snd_max == tp->snd_una &&
3d127502 210 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
5f60906c 211 idle_cwv = TRUE;
653eb2c0
MD
212
213 /*
214 * Calculate whether the transmit stream was previously idle
215 * and adjust TF_LASTIDLE for the next time.
216 */
08525cf0
JH
217 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
218 if (idle && (tp->t_flags & TF_MORETOCOME))
219 tp->t_flags |= TF_LASTIDLE;
220 else
221 tp->t_flags &= ~TF_LASTIDLE;
79a80d9f 222
9a5b142f
JH
223 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
224 !IN_FASTRECOVERY(tp))
225 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
226
5f60906c
SZ
227 /*
228 * Find out whether TSO could be used or not
229 *
230 * For TSO capable devices, the following assumptions apply to
231 * the processing of TCP flags:
232 * - If FIN is set on the large TCP segment, the device must set
233 * FIN on the last segment that it creates from the large TCP
234 * segment.
235 * - If PUSH is set on the large TCP segment, the device must set
236 * PUSH on the last segment that it creates from the large TCP
237 * segment.
238 */
239#if !defined(IPSEC) && !defined(FAST_IPSEC)
240 if (tcp_do_tso
241#ifdef TCP_SIGNATURE
242 && (tp->t_flags & TF_SIGNATURE) == 0
243#endif
244 ) {
245 if (!isipv6) {
246 struct rtentry *rt = inp->inp_route.ro_rt;
247
248 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
249 (rt->rt_ifp->if_hwassist & CSUM_TSO))
250 can_tso = TRUE;
251 }
252 }
253#endif /* !IPSEC && !FAST_IPSEC */
254
984263bc 255again:
46fc0bdc
SZ
256 m = NULL;
257 ip = NULL;
258 ipov = NULL;
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 {
497 /*
498 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
499 * thoff), and make sure that we send equal size
500 * transfers down the stack (rather than big-small-
501 * big-small-...).
502 */
503 len = (min(len, (IP_MAXPACKET - tso_hlen)) / segsz) *
504 segsz;
505 if (len <= segsz)
506 use_tso = FALSE;
507 }
61896e3c 508 sendalot = TRUE;
5f60906c
SZ
509 } else {
510 use_tso = FALSE;
984263bc 511 }
6d49aa6f 512 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
984263bc
MD
513 flags &= ~TH_FIN;
514
6d49aa6f 515 recvwin = ssb_space(&so->so_rcv);
984263bc
MD
516
517 /*
518 * Sender silly window avoidance. We transmit under the following
519 * conditions when len is non-zero:
520 *
521 * - We have a full segment
522 * - This is the last buffer in a write()/send() and we are
523 * either idle or running NODELAY
524 * - we've timed out (e.g. persist timer)
525 * - we have more then 1/2 the maximum send window's worth of
653eb2c0 526 * data (receiver may be limiting the window size)
984263bc
MD
527 * - we need to retransmit
528 */
529 if (len) {
5f60906c 530 if (len >= segsz)
984263bc
MD
531 goto send;
532 /*
533 * NOTE! on localhost connections an 'ack' from the remote
534 * end may occur synchronously with the output and cause
535 * us to flush a buffer queued with moretocome. XXX
536 *
537 * note: the len + off check is almost certainly unnecessary.
538 */
539 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
540 (idle || (tp->t_flags & TF_NODELAY)) &&
6d49aa6f 541 len + off >= so->so_snd.ssb_cc &&
0770aa24 542 !(tp->t_flags & TF_NOPUSH)) {
984263bc
MD
543 goto send;
544 }
c1d3b892 545 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
984263bc
MD
546 goto send;
547 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
548 goto send;
549 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
550 goto send;
c1fabe85
SZ
551 if (tp->t_flags & TF_XMITNOW)
552 goto send;
984263bc
MD
553 }
554
555 /*
556 * Compare available window to amount of window
557 * known to peer (as advertised window less
558 * next expected input). If the difference is at least two
559 * max size segments, or at least 50% of the maximum possible
560 * window, then want to send a window update to peer.
561 */
c53f07a5 562 if (recvwin > 0) {
984263bc
MD
563 /*
564 * "adv" is the amount we can increase the window,
565 * taking into account that we are limited by
566 * TCP_MAXWIN << tp->rcv_scale.
567 */
c53f07a5 568 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
984263bc 569 (tp->rcv_adv - tp->rcv_nxt);
46e92930 570 long hiwat;
984263bc 571
ddb2e271
MD
572 /*
573 * This ack case typically occurs when the user has drained
574 * the TCP socket buffer sufficiently to warrent an ack
575 * containing a 'pure window update'... that is, an ack that
576 * ONLY updates the tcp window.
577 *
578 * It is unclear why we would need to do a pure window update
579 * past 2 segments if we are going to do one at 1/2 the high
580 * water mark anyway, especially since under normal conditions
f23061d4 581 * the user program will drain the socket buffer quickly.
ddb2e271
MD
582 * The 2-segment pure window update will often add a large
583 * number of extra, unnecessary acks to the stream.
584 *
585 * avoid_pure_win_update now defaults to 1.
586 */
46e92930
MD
587 if (avoid_pure_win_update == 0 ||
588 (tp->t_flags & TF_RXRESIZED)) {
5f60906c 589 if (adv >= (long) (2 * segsz)) {
ddb2e271 590 goto send;
46e92930 591 }
ddb2e271 592 }
46e92930
MD
593 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
594 if (hiwat > (long)so->so_rcv.ssb_hiwat)
595 hiwat = (long)so->so_rcv.ssb_hiwat;
596 if (adv >= hiwat / 2)
984263bc
MD
597 goto send;
598 }
599
600 /*
601 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
602 * is also a catch-all for the retransmit timer timeout case.
603 */
604 if (tp->t_flags & TF_ACKNOW)
605 goto send;
606 if ((flags & TH_RST) ||
0770aa24 607 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
984263bc
MD
608 goto send;
609 if (SEQ_GT(tp->snd_up, tp->snd_una))
610 goto send;
611 /*
612 * If our state indicates that FIN should be sent
613 * and we have not yet done so, then we need to send.
614 */
7462c355 615 if ((flags & TH_FIN) &&
0770aa24 616 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
984263bc
MD
617 goto send;
618
619 /*
620 * TCP window updates are not reliable, rather a polling protocol
621 * using ``persist'' packets is used to insure receipt of window
622 * updates. The three ``states'' for the output side are:
623 * idle not doing retransmits or persists
624 * persisting to move a small or zero window
625 * (re)transmitting and thereby not persisting
626 *
a48c5dd5 627 * tcp_callout_active(tp, tp->tt_persist)
984263bc 628 * is true when we are in persist state.
c1d3b892 629 * The TF_FORCE flag in tp->t_flags
984263bc 630 * is set when we are called to send a persist packet.
a48c5dd5 631 * tcp_callout_active(tp, tp->tt_rexmt)
984263bc
MD
632 * is set when we are retransmitting
633 * The output side is idle when both timers are zero.
634 *
635 * If send window is too small, there is data to transmit, and no
636 * retransmit or persist is pending, then go to persist state.
7462c355 637 *
984263bc 638 * If nothing happens soon, send when timer expires:
7462c355
MD
639 * if window is nonzero, transmit what we can, otherwise force out
640 * a byte.
641 *
642 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
643 * with data pending. This situation can occur during a
644 * simultanious connect.
984263bc 645 */
6d49aa6f 646 if (so->so_snd.ssb_cc > 0 &&
7462c355 647 tp->t_state != TCPS_SYN_RECEIVED &&
a48c5dd5
SZ
648 !tcp_callout_active(tp, tp->tt_rexmt) &&
649 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
650 tp->t_rxtshift = 0;
651 tcp_setpersist(tp);
652 }
653
654 /*
655 * No reason to send a segment, just return.
656 */
c1fabe85 657 tp->t_flags &= ~TF_XMITNOW;
984263bc
MD
658 return (0);
659
660send:
661 /*
662 * Before ESTABLISHED, force sending of initial options
663 * unless TCP set not to do any options.
664 * NOTE: we assume that the IP/TCP header plus TCP options
665 * always fit in a single mbuf, leaving room for a maximum
666 * link header, i.e.
79a80d9f 667 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
984263bc
MD
668 */
669 optlen = 0;
984263bc 670 if (isipv6)
79a80d9f 671 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
984263bc 672 else
79a80d9f 673 hdrlen = sizeof(struct tcpiphdr);
984263bc
MD
674 if (flags & TH_SYN) {
675 tp->snd_nxt = tp->iss;
0770aa24 676 if (!(tp->t_flags & TF_NOOPT)) {
984263bc
MD
677 u_short mss;
678
679 opt[0] = TCPOPT_MAXSEG;
680 opt[1] = TCPOLEN_MAXSEG;
681 mss = htons((u_short) tcp_mssopt(tp));
f23061d4 682 memcpy(opt + 2, &mss, sizeof mss);
984263bc
MD
683 optlen = TCPOLEN_MAXSEG;
684
685 if ((tp->t_flags & TF_REQ_SCALE) &&
0770aa24
JH
686 (!(flags & TH_ACK) ||
687 (tp->t_flags & TF_RCVD_SCALE))) {
984263bc
MD
688 *((u_int32_t *)(opt + optlen)) = htonl(
689 TCPOPT_NOP << 24 |
690 TCPOPT_WINDOW << 16 |
691 TCPOLEN_WINDOW << 8 |
692 tp->request_r_scale);
693 optlen += 4;
694 }
91489f6b
JH
695
696 if ((tcp_do_sack && !(flags & TH_ACK)) ||
697 tp->t_flags & TF_SACK_PERMITTED) {
698 uint32_t *lp = (uint32_t *)(opt + optlen);
699
700 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
701 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
702 }
984263bc 703 }
0770aa24 704 }
984263bc 705
0770aa24 706 /*
984263bc
MD
707 * Send a timestamp and echo-reply if this is a SYN and our side
708 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
709 * and our peer have sent timestamps in our SYN's.
0770aa24 710 */
61896e3c 711 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
0770aa24
JH
712 !(flags & TH_RST) &&
713 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
984263bc
MD
714 u_int32_t *lp = (u_int32_t *)(opt + optlen);
715
0770aa24
JH
716 /* Form timestamp option as shown in appendix A of RFC 1323. */
717 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
718 *lp++ = htonl(ticks);
719 *lp = htonl(tp->ts_recent);
720 optlen += TCPOLEN_TSTAMP_APPA;
721 }
984263bc 722
5b0b9fa5
PA
723 /* Set receive buffer autosizing timestamp. */
724 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
725 tp->rfbuf_ts = ticks;
726
0770aa24 727 /*
91489f6b
JH
728 * If this is a SACK connection and we have a block to report,
729 * fill in the SACK blocks in the TCP options.
730 */
5f60906c 731 if (report_sack)
91489f6b
JH
732 tcp_sack_fill_report(tp, opt, &optlen);
733
b1992928
MD
734#ifdef TCP_SIGNATURE
735 if (tp->t_flags & TF_SIGNATURE) {
736 int i;
737 u_char *bp;
738 /*
739 * Initialize TCP-MD5 option (RFC2385)
740 */
741 bp = (u_char *)opt + optlen;
742 *bp++ = TCPOPT_SIGNATURE;
743 *bp++ = TCPOLEN_SIGNATURE;
744 sigoff = optlen + 2;
745 for (i = 0; i < TCP_SIGLEN; i++)
746 *bp++ = 0;
747 optlen += TCPOLEN_SIGNATURE;
748 /*
749 * Terminate options list and maintain 32-bit alignment.
750 */
751 *bp++ = TCPOPT_NOP;
752 *bp++ = TCPOPT_EOL;
753 optlen += 2;
754 }
755#endif /* TCP_SIGNATURE */
91489f6b 756 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
0770aa24 757 hdrlen += optlen;
984263bc 758
61896e3c 759 if (isipv6) {
0770aa24 760 ipoptlen = ip6_optlen(inp);
61896e3c 761 } else {
0770aa24
JH
762 if (inp->inp_options) {
763 ipoptlen = inp->inp_options->m_len -
764 offsetof(struct ipoption, ipopt_list);
765 } else {
766 ipoptlen = 0;
767 }
984263bc 768 }
984263bc
MD
769#ifdef IPSEC
770 ipoptlen += ipsec_hdrsiz_tcp(tp);
771#endif
772
5f60906c 773 if (use_tso) {
243cd031
SZ
774 /* TSO segment length must be multiple of segment size */
775 KASSERT(len >= (2 * segsz) && (len % segsz == 0),
5f60906c
SZ
776 ("invalid TSO len %ld, segsz %u", len, segsz));
777 } else {
778 KASSERT(len <= segsz,
779 ("invalid len %ld, segsz %u", len, segsz));
45f20ce7 780
5f60906c
SZ
781 /*
782 * Adjust data length if insertion of options will bump
783 * the packet length beyond the t_maxopd length. Clear
784 * FIN to prevent premature closure since there is still
785 * more data to send after this (now truncated) packet.
786 *
787 * If just the options do not fit we are in a no-win
788 * situation and we treat it as an unreachable host.
789 */
790 if (len + optlen + ipoptlen > tp->t_maxopd) {
791 if (tp->t_maxopd <= optlen + ipoptlen) {
792 static time_t last_optlen_report;
793
794 if (last_optlen_report != time_second) {
795 last_optlen_report = time_second;
796 kprintf("tcpcb %p: MSS (%d) too "
797 "small to hold options!\n",
798 tp, tp->t_maxopd);
799 }
800 error = EHOSTUNREACH;
801 goto out;
802 } else {
803 flags &= ~TH_FIN;
804 len = tp->t_maxopd - optlen - ipoptlen;
805 sendalot = TRUE;
45f20ce7 806 }
45f20ce7 807 }
984263bc
MD
808 }
809
984263bc 810#ifdef INET6
0770aa24 811 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
984263bc 812#else
0770aa24 813 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
984263bc 814#endif
984263bc
MD
815
816 /*
817 * Grab a header mbuf, attaching a copy of data to
818 * be transmitted, and initialize the header from
819 * the template for sends on this connection.
820 */
821 if (len) {
c1d3b892 822 if ((tp->t_flags & TF_FORCE) && len == 1)
984263bc
MD
823 tcpstat.tcps_sndprobe++;
824 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
91489f6b
JH
825 if (tp->snd_nxt == tp->snd_una)
826 tp->snd_max_rexmt = tp->snd_max;
02cc2f35
SZ
827 if (nsacked) {
828 tcpstat.tcps_sndsackrtopack++;
829 tcpstat.tcps_sndsackrtobyte += len;
830 }
984263bc
MD
831 tcpstat.tcps_sndrexmitpack++;
832 tcpstat.tcps_sndrexmitbyte += len;
833 } else {
834 tcpstat.tcps_sndpack++;
835 tcpstat.tcps_sndbyte += len;
836 }
3d127502 837 if (idle_cwv) {
5f60906c 838 idle_cwv = FALSE;
3d127502
SZ
839 tcp_idle_cwnd_validate(tp);
840 }
841 /* Update last send time after CWV */
842 tp->snd_last = ticks;
984263bc 843#ifdef notyet
6d49aa6f 844 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
61896e3c 845 max_linkhdr + hdrlen)) == NULL) {
984263bc 846 error = ENOBUFS;
1ff9b7d3 847 goto after_th;
984263bc
MD
848 }
849 /*
850 * m_copypack left space for our hdr; use it.
851 */
852 m->m_len += hdrlen;
853 m->m_data -= hdrlen;
854#else
50503f0f
JH
855#ifndef INET6
856 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
857#else
858 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
859 M_PKTHDR, NULL);
860#endif
984263bc
MD
861 if (m == NULL) {
862 error = ENOBUFS;
1ff9b7d3 863 goto after_th;
984263bc 864 }
984263bc
MD
865 m->m_data += max_linkhdr;
866 m->m_len = hdrlen;
867 if (len <= MHLEN - hdrlen - max_linkhdr) {
6d49aa6f 868 m_copydata(so->so_snd.ssb_mb, off, (int) len,
984263bc
MD
869 mtod(m, caddr_t) + hdrlen);
870 m->m_len += len;
871 } else {
6d49aa6f 872 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
61896e3c 873 if (m->m_next == NULL) {
f23061d4 874 m_free(m);
1ff9b7d3 875 m = NULL;
984263bc 876 error = ENOBUFS;
1ff9b7d3 877 goto after_th;
984263bc
MD
878 }
879 }
880#endif
881 /*
882 * If we're sending everything we've got, set PUSH.
883 * (This will keep happy those implementations which only
884 * give data to the user when a buffer fills or
885 * a PUSH comes in.)
886 */
6d49aa6f 887 if (off + len == so->so_snd.ssb_cc)
984263bc
MD
888 flags |= TH_PUSH;
889 } else {
890 if (tp->t_flags & TF_ACKNOW)
891 tcpstat.tcps_sndacks++;
61896e3c 892 else if (flags & (TH_SYN | TH_FIN | TH_RST))
984263bc
MD
893 tcpstat.tcps_sndctrl++;
894 else if (SEQ_GT(tp->snd_up, tp->snd_una))
895 tcpstat.tcps_sndurg++;
896 else
897 tcpstat.tcps_sndwinup++;
898
74f1caca 899 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
984263bc
MD
900 if (m == NULL) {
901 error = ENOBUFS;
1ff9b7d3 902 goto after_th;
984263bc 903 }
0770aa24
JH
904 if (isipv6 &&
905 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
984263bc 906 MH_ALIGN(m, hdrlen);
0770aa24
JH
907 else
908 m->m_data += max_linkhdr;
984263bc
MD
909 m->m_len = hdrlen;
910 }
60233e58 911 m->m_pkthdr.rcvif = NULL;
984263bc
MD
912 if (isipv6) {
913 ip6 = mtod(m, struct ip6_hdr *);
914 th = (struct tcphdr *)(ip6 + 1);
5f60906c 915 tcp_fillheaders(tp, ip6, th, use_tso);
0770aa24
JH
916 } else {
917 ip = mtod(m, struct ip *);
918 ipov = (struct ipovly *)ip;
919 th = (struct tcphdr *)(ip + 1);
920 /* this picks up the pseudo header (w/o the length) */
5f60906c 921 tcp_fillheaders(tp, ip, th, use_tso);
0770aa24 922 }
1ff9b7d3 923after_th:
984263bc
MD
924 /*
925 * Fill in fields, remembering maximum advertised
926 * window for use in delaying messages about window sizes.
927 * If resending a FIN, be sure not to use a new sequence number.
928 */
929 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
930 tp->snd_nxt == tp->snd_max)
931 tp->snd_nxt--;
1ff9b7d3
SZ
932
933 if (th != NULL) {
934 /*
935 * If we are doing retransmissions, then snd_nxt will
936 * not reflect the first unsent octet. For ACK only
937 * packets, we do not want the sequence number of the
938 * retransmitted packet, we want the sequence number
939 * of the next unsent octet. So, if there is no data
940 * (and no SYN or FIN), use snd_max instead of snd_nxt
941 * when filling in ti_seq. But if we are in persist
942 * state, snd_max might reflect one byte beyond the
943 * right edge of the window, so use snd_nxt in that
944 * case, since we know we aren't doing a retransmission.
945 * (retransmit and persist are mutually exclusive...)
946 */
947 if (len || (flags & (TH_SYN|TH_FIN)) ||
948 tcp_callout_active(tp, tp->tt_persist))
949 th->th_seq = htonl(tp->snd_nxt);
950 else
951 th->th_seq = htonl(tp->snd_max);
952 th->th_ack = htonl(tp->rcv_nxt);
953 if (optlen) {
954 bcopy(opt, th + 1, optlen);
955 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
956 }
957 th->th_flags = flags;
984263bc 958 }
46e92930 959
984263bc 960 /*
46e92930
MD
961 * Calculate receive window. Don't shrink window, but avoid
962 * silly window syndrome by sending a 0 window if the actual
963 * window is less then one segment.
984263bc 964 */
6d49aa6f 965 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
5f60906c 966 recvwin < (long)segsz)
c53f07a5 967 recvwin = 0;
55126001
MD
968 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
969 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
c53f07a5
JH
970 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
971 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
984263bc
MD
972
973 /*
974 * Adjust the RXWIN0SENT flag - indicate that we have advertised
975 * a 0 window. This may cause the remote transmitter to stall. This
976 * flag tells soreceive() to disable delayed acknowledgements when
977 * draining the buffer. This can occur if the receiver is attempting
978 * to read more data then can be buffered prior to transmitting on
979 * the connection.
980 */
c53f07a5 981 if (recvwin == 0)
984263bc
MD
982 tp->t_flags |= TF_RXWIN0SENT;
983 else
984 tp->t_flags &= ~TF_RXWIN0SENT;
985
1ff9b7d3
SZ
986 if (th != NULL)
987 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
988
984263bc 989 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
5f60906c 990 KASSERT(!use_tso, ("URG with TSO"));
1ff9b7d3
SZ
991 if (th != NULL) {
992 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
993 th->th_flags |= TH_URG;
994 }
79a80d9f 995 } else {
984263bc
MD
996 /*
997 * If no urgent pointer to send, then we pull
998 * the urgent pointer to the left edge of the send window
999 * so that it doesn't drift into the send window on sequence
1000 * number wraparound.
1001 */
1002 tp->snd_up = tp->snd_una; /* drag it along */
79a80d9f 1003 }
984263bc 1004
1ff9b7d3 1005 if (th != NULL) {
b1992928 1006#ifdef TCP_SIGNATURE
1ff9b7d3
SZ
1007 if (tp->t_flags & TF_SIGNATURE) {
1008 tcpsignature_compute(m, len, optlen,
1009 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1010 }
b1992928
MD
1011#endif /* TCP_SIGNATURE */
1012
984263bc 1013 /*
1ff9b7d3
SZ
1014 * Put TCP length in extended header, and then
1015 * checksum extended header and data.
984263bc 1016 */
1ff9b7d3
SZ
1017 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1018 if (isipv6) {
1019 /*
1020 * ip6_plen is not need to be filled now, and will be
1021 * filled in ip6_output().
1022 */
1023 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1024 sizeof(struct ip6_hdr),
1025 sizeof(struct tcphdr) + optlen + len);
1026 } else {
7df36335 1027 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
5f60906c
SZ
1028 if (use_tso) {
1029 m->m_pkthdr.csum_flags = CSUM_TSO;
f0336d39 1030 m->m_pkthdr.tso_segsz = segsz;
5f60906c
SZ
1031 } else {
1032 m->m_pkthdr.csum_flags = CSUM_TCP;
7df36335
SZ
1033 m->m_pkthdr.csum_data =
1034 offsetof(struct tcphdr, th_sum);
5f60906c
SZ
1035 if (len + optlen) {
1036 th->th_sum = in_addword(th->th_sum,
1037 htons((u_short)(optlen + len)));
1038 }
1ff9b7d3
SZ
1039 }
1040
1041 /*
1042 * IP version must be set here for ipv4/ipv6 checking
1043 * later
1044 */
1045 KASSERT(ip->ip_v == IPVERSION,
1046 ("%s: IP version incorrect: %d",
1047 __func__, ip->ip_v));
1048 }
0770aa24 1049 }
984263bc
MD
1050
1051 /*
1052 * In transmit state, time the transmission and arrange for
1053 * the retransmit. In persist state, just set snd_max.
1054 */
a48c5dd5
SZ
1055 if (!(tp->t_flags & TF_FORCE) ||
1056 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
1057 tcp_seq startseq = tp->snd_nxt;
1058
1059 /*
1060 * Advance snd_nxt over sequence space of this segment.
1061 */
61896e3c 1062 if (flags & (TH_SYN | TH_FIN)) {
984263bc
MD
1063 if (flags & TH_SYN)
1064 tp->snd_nxt++;
1065 if (flags & TH_FIN) {
1066 tp->snd_nxt++;
1067 tp->t_flags |= TF_SENTFIN;
1068 }
1069 }
1070 tp->snd_nxt += len;
1071 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1072 tp->snd_max = tp->snd_nxt;
1073 /*
1074 * Time this transmission if not a retransmission and
1075 * not currently timing anything.
1076 */
1077 if (tp->t_rtttime == 0) {
1078 tp->t_rtttime = ticks;
1079 tp->t_rtseq = startseq;
1080 tcpstat.tcps_segstimed++;
1081 }
1082 }
1083
1084 /*
1085 * Set retransmit timer if not currently set,
1086 * and not doing a pure ack or a keep-alive probe.
1087 * Initial value for retransmit timer is smoothed
1088 * round-trip time + 2 * round-trip time variance.
1089 * Initialize shift counter which is used for backoff
1090 * of retransmit time.
1091 */
a48c5dd5 1092 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
984263bc 1093 tp->snd_nxt != tp->snd_una) {
a48c5dd5
SZ
1094 if (tcp_callout_active(tp, tp->tt_persist)) {
1095 tcp_callout_stop(tp, tp->tt_persist);
984263bc
MD
1096 tp->t_rxtshift = 0;
1097 }
a48c5dd5
SZ
1098 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1099 tcp_timer_rexmt);
984263bc
MD
1100 }
1101 } else {
1102 /*
1103 * Persist case, update snd_max but since we are in
1104 * persist mode (no window) we do not update snd_nxt.
1105 */
1106 int xlen = len;
1107 if (flags & TH_SYN)
ed20d0e3 1108 panic("tcp_output: persist timer to send SYN");
984263bc
MD
1109 if (flags & TH_FIN) {
1110 ++xlen;
1111 tp->t_flags |= TF_SENTFIN;
1112 }
1113 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1114 tp->snd_max = tp->snd_nxt + xlen;
1115 }
1116
1ff9b7d3 1117 if (th != NULL) {
984263bc 1118#ifdef TCPDEBUG
1ff9b7d3
SZ
1119 /* Trace. */
1120 if (so->so_options & SO_DEBUG) {
1121 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1122 mtod(m, void *), th, 0);
1123 }
984263bc
MD
1124#endif
1125
984263bc 1126 /*
1ff9b7d3
SZ
1127 * Fill in IP length and desired time to live and
1128 * send to IP level. There should be a better way
1129 * to handle ttl and tos; we could keep them in
1130 * the template, but need a way to checksum without them.
984263bc 1131 */
1ff9b7d3
SZ
1132 /*
1133 * m->m_pkthdr.len should have been set before cksum
1134 * calcuration, because in6_cksum() need it.
1135 */
1136 if (isipv6) {
1137 /*
1138 * we separately set hoplimit for every segment,
1139 * since the user might want to change the value
1140 * via setsockopt. Also, desired default hop
1141 * limit might be changed via Neighbor Discovery.
1142 */
1143 ip6->ip6_hlim = in6_selecthlim(inp,
0770aa24
JH
1144 (inp->in6p_route.ro_rt ?
1145 inp->in6p_route.ro_rt->rt_ifp : NULL));
1ff9b7d3
SZ
1146
1147 /* TODO: IPv6 IP6TOS_ECT bit on */
1148 error = ip6_output(m, inp->in6p_outputopts,
1149 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1150 NULL, NULL, inp);
1151 } else {
1152 struct rtentry *rt;
1153 ip->ip_len = m->m_pkthdr.len;
1154#ifdef INET6
1155 if (INP_CHECK_SOCKAF(so, AF_INET6))
1156 ip->ip_ttl = in6_selecthlim(inp,
1157 (inp->in6p_route.ro_rt ?
1158 inp->in6p_route.ro_rt->rt_ifp : NULL));
1159 else
0770aa24 1160#endif
1ff9b7d3 1161 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
0770aa24 1162
1ff9b7d3
SZ
1163 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1164 /*
1165 * See if we should do MTU discovery.
1166 * We do it only if the following are true:
1167 * 1) we have a valid route to the destination
1168 * 2) the MTU is not locked (if it is,
1169 * then discovery has been disabled)
1170 */
1171 if (path_mtu_discovery &&
1172 (rt = inp->inp_route.ro_rt) &&
1173 (rt->rt_flags & RTF_UP) &&
1174 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1175 ip->ip_off |= IP_DF;
1176
1177 error = ip_output(m, inp->inp_options, &inp->inp_route,
1178 (so->so_options & SO_DONTROUTE) |
1179 IP_DEBUGROUTE, NULL, inp);
1180 }
1181 } else {
ed20d0e3 1182 KASSERT(error != 0, ("no error, but th not set"));
0770aa24 1183 }
984263bc 1184 if (error) {
c1fabe85 1185 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
984263bc
MD
1186
1187 /*
1188 * We know that the packet was lost, so back out the
1189 * sequence number advance, if any.
1190 */
c1d3b892 1191 if (!(tp->t_flags & TF_FORCE) ||
a48c5dd5 1192 !tcp_callout_active(tp, tp->tt_persist)) {
984263bc
MD
1193 /*
1194 * No need to check for TH_FIN here because
1195 * the TF_SENTFIN flag handles that case.
1196 */
0770aa24 1197 if (!(flags & TH_SYN))
984263bc
MD
1198 tp->snd_nxt -= len;
1199 }
1200
1201out:
1202 if (error == ENOBUFS) {
ea8bd01a
JH
1203 /*
1204 * If we can't send, make sure there is something
28a9a7b6
MD
1205 * to get us going again later.
1206 *
1207 * The persist timer isn't necessarily allowed in all
1208 * states, use the rexmt timer.
ea8bd01a 1209 */
a48c5dd5
SZ
1210 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1211 !tcp_callout_active(tp, tp->tt_persist)) {
28a9a7b6
MD
1212 tcp_callout_reset(tp, tp->tt_rexmt,
1213 tp->t_rxtcur,
1214 tcp_timer_rexmt);
1215#if 0
ea8bd01a
JH
1216 tp->t_rxtshift = 0;
1217 tcp_setpersist(tp);
28a9a7b6 1218#endif
ea8bd01a 1219 }
0770aa24 1220 tcp_quench(inp, 0);
984263bc
MD
1221 return (0);
1222 }
1223 if (error == EMSGSIZE) {
1224 /*
1225 * ip_output() will have already fixed the route
1226 * for us. tcp_mtudisc() will, as its last action,
1227 * initiate retransmission, so it is important to
1228 * not do so here.
1229 */
0770aa24 1230 tcp_mtudisc(inp, 0);
984263bc
MD
1231 return 0;
1232 }
0770aa24
JH
1233 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1234 TCPS_HAVERCVDSYN(tp->t_state)) {
984263bc
MD
1235 tp->t_softerror = error;
1236 return (0);
1237 }
1238 return (error);
1239 }
1240 tcpstat.tcps_sndtotal++;
1241
1242 /*
1243 * Data sent (as far as we can tell).
46e92930 1244 *
984263bc
MD
1245 * If this advertises a larger window than any other segment,
1246 * then remember the size of the advertised window.
46e92930 1247 *
984263bc
MD
1248 * Any pending ACK has now been sent.
1249 */
46e92930 1250 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
c53f07a5 1251 tp->rcv_adv = tp->rcv_nxt + recvwin;
46e92930
MD
1252 tp->t_flags &= ~TF_RXRESIZED;
1253 }
984263bc 1254 tp->last_ack_sent = tp->rcv_nxt;
c1fabe85 1255 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
984263bc 1256 if (tcp_delack_enabled)
a48c5dd5 1257 tcp_callout_stop(tp, tp->tt_delack);
46fc0bdc 1258 if (sendalot)
984263bc
MD
1259 goto again;
1260 return (0);
1261}
1262
1263void
f3f70f0d 1264tcp_setpersist(struct tcpcb *tp)
984263bc
MD
1265{
1266 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1267 int tt;
1268
c529614e
SZ
1269 if (tp->t_state == TCPS_SYN_SENT ||
1270 tp->t_state == TCPS_SYN_RECEIVED) {
ed20d0e3 1271 panic("tcp_setpersist: not established yet, current %s",
c529614e
SZ
1272 tp->t_state == TCPS_SYN_SENT ?
1273 "SYN_SENT" : "SYN_RECEIVED");
1274 }
94603b1b 1275
a48c5dd5 1276 if (tcp_callout_active(tp, tp->tt_rexmt))
984263bc
MD
1277 panic("tcp_setpersist: retransmit pending");
1278 /*
1279 * Start/restart persistance timer.
1280 */
5fe66e68
JH
1281 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1282 TCPTV_PERSMAX);
a48c5dd5 1283 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
984263bc
MD
1284 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1285 tp->t_rxtshift++;
1286}
3d127502
SZ
1287
1288static void
1289tcp_idle_cwnd_validate(struct tcpcb *tp)
1290{
1291 u_long initial_cwnd = tcp_initial_window(tp);
1292 u_long min_cwnd;
1293
1294 tcpstat.tcps_sndidle++;
1295
1296 /* According to RFC5681: RW=min(IW,cwnd) */
1297 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1298
1299 if (tcp_idle_cwv) {
1300 u_long idle_time, decay_cwnd;
1301
1302 /*
1303 * RFC2861, but only after idle period.
1304 */
1305
1306 /*
1307 * Before the congestion window is reduced, ssthresh
1308 * is set to the maximum of its current value and 3/4
1309 * cwnd. If the sender then has more data to send
1310 * than the decayed cwnd allows, the TCP will slow-
1311 * start (perform exponential increase) at least
1312 * half-way back up to the old value of cwnd.
1313 */
1314 tp->snd_ssthresh = max(tp->snd_ssthresh,
1315 (3 * tp->snd_cwnd) / 4);
1316
1317 /*
1318 * Decay the congestion window by half for every RTT
1319 * that the flow remains inactive.
1320 *
1321 * The difference between our implementation and
1322 * RFC2861 is that we don't allow cwnd to go below
1323 * the value allowed by RFC5681 (min_cwnd).
1324 */
1325 idle_time = ticks - tp->snd_last;
1326 decay_cwnd = tp->snd_cwnd;
1327 while (idle_time >= tp->t_rxtcur &&
1328 decay_cwnd > min_cwnd) {
1329 decay_cwnd >>= 1;
1330 idle_time -= tp->t_rxtcur;
1331 }
1332 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1333 } else {
1334 /*
1335 * Slow-start from scratch to re-determine the send
1336 * congestion window.
1337 */
1338 tp->snd_cwnd = min_cwnd;
1339 }
1340
1341 /* Restart ABC counting during congestion avoidance */
1342 tp->snd_wacked = 0;
1343}
5f60906c
SZ
1344
1345static int
1346tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1347{
1348 struct inpcb * const inp = tp->t_inpcb;
1349#ifdef INET6
1350 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1351#else
1352 const boolean_t isipv6 = FALSE;
1353#endif
1354 unsigned int ipoptlen, optlen;
1355 u_int hlen;
1356
1357 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1358
1359 if (isipv6) {
1360 ipoptlen = ip6_optlen(inp);
1361 } else {
1362 if (inp->inp_options) {
1363 ipoptlen = inp->inp_options->m_len -
1364 offsetof(struct ipoption, ipopt_list);
1365 } else {
1366 ipoptlen = 0;
1367 }
1368 }
1369#ifdef IPSEC
1370 ipoptlen += ipsec_hdrsiz_tcp(tp);
1371#endif
1372 hlen += ipoptlen;
1373
1374 optlen = 0;
1375 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1376 (tp->t_flags & TF_RCVD_TSTMP))
1377 optlen += TCPOLEN_TSTAMP_APPA;
1378 hlen += optlen;
1379
1380 if (tp->t_maxopd <= optlen + ipoptlen)
1381 return EHOSTUNREACH;
1382
1383 *segsz = tp->t_maxopd - optlen - ipoptlen;
1384 *hlen0 = hlen;
1385 return 0;
1386}