HAMMER 60I/Many: Mirroring
[dragonfly.git] / sys / netinet / tcp_output.c
blob8bd87abb0e5448ae3065397291757d9b8236cffc
1 /*
2 * Copyright (c) 2004 Jeffrey M. Hsu. All rights reserved.
3 * Copyright (c) 2004 The DragonFly Project. All rights reserved.
5 * This code is derived from software contributed to The DragonFly Project
6 * by Jeffrey M. Hsu.
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.
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.
35 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
36 * The Regents of the University of California. All rights reserved.
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.
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.
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 * $DragonFly: src/sys/netinet/tcp_output.c,v 1.34 2007/04/22 01:13:14 dillon Exp $
71 #include "opt_inet6.h"
72 #include "opt_ipsec.h"
73 #include "opt_tcpdebug.h"
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>
84 #include <sys/in_cksum.h>
85 #include <sys/thread.h>
86 #include <sys/globaldata.h>
88 #include <net/route.h>
90 #include <netinet/in.h>
91 #include <netinet/in_systm.h>
92 #include <netinet/ip.h>
93 #include <netinet/in_pcb.h>
94 #include <netinet/ip_var.h>
95 #include <netinet6/in6_pcb.h>
96 #include <netinet/ip6.h>
97 #include <netinet6/ip6_var.h>
98 #include <netinet/tcp.h>
99 #define TCPOUTFLAGS
100 #include <netinet/tcp_fsm.h>
101 #include <netinet/tcp_seq.h>
102 #include <netinet/tcp_timer.h>
103 #include <netinet/tcp_var.h>
104 #include <netinet/tcpip.h>
105 #ifdef TCPDEBUG
106 #include <netinet/tcp_debug.h>
107 #endif
109 #ifdef IPSEC
110 #include <netinet6/ipsec.h>
111 #endif /*IPSEC*/
113 #ifdef FAST_IPSEC
114 #include <netproto/ipsec/ipsec.h>
115 #define IPSEC
116 #endif /*FAST_IPSEC*/
118 #ifdef notyet
119 extern struct mbuf *m_copypack();
120 #endif
122 int path_mtu_discovery = 0;
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
124 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
126 static int avoid_pure_win_update = 1;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
128 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
131 * Tcp output routine: figure out what should be sent and send it.
134 tcp_output(struct tcpcb *tp)
136 struct inpcb * const inp = tp->t_inpcb;
137 struct socket *so = inp->inp_socket;
138 long len, recvwin, sendwin;
139 int nsacked = 0;
140 int off, flags, error;
141 struct mbuf *m;
142 struct ip *ip = NULL;
143 struct ipovly *ipov = NULL;
144 struct tcphdr *th;
145 u_char opt[TCP_MAXOLEN];
146 unsigned int ipoptlen, optlen, hdrlen;
147 int idle;
148 boolean_t sendalot;
149 struct ip6_hdr *ip6 = NULL;
150 #ifdef INET6
151 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
152 #else
153 const boolean_t isipv6 = FALSE;
154 #endif
155 struct rmxp_tao *taop;
158 * Determine length of data that should be transmitted,
159 * and flags that will be used.
160 * If there is some data or critical controls (SYN, RST)
161 * to send, then transmit; otherwise, investigate further.
165 * If we have been idle for a while, the send congestion window
166 * could be no longer representative of the current state of the link.
167 * So unless we are expecting more acks to come in, slow-start from
168 * scratch to re-determine the send congestion window.
170 if (tp->snd_max == tp->snd_una &&
171 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
172 if (tcp_do_rfc3390) {
173 int initial_cwnd =
174 min(4 * tp->t_maxseg, max(2 * tp->t_maxseg, 4380));
176 tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd);
177 } else {
178 tp->snd_cwnd = tp->t_maxseg;
180 tp->snd_wacked = 0;
184 * Calculate whether the transmit stream was previously idle
185 * and adjust TF_LASTIDLE for the next time.
187 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
188 if (idle && (tp->t_flags & TF_MORETOCOME))
189 tp->t_flags |= TF_LASTIDLE;
190 else
191 tp->t_flags &= ~TF_LASTIDLE;
193 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
194 !IN_FASTRECOVERY(tp))
195 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
197 again:
198 /* Make use of SACK information when slow-starting after a RTO. */
199 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
200 !IN_FASTRECOVERY(tp)) {
201 tcp_seq old_snd_nxt = tp->snd_nxt;
203 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
204 nsacked += tp->snd_nxt - old_snd_nxt;
207 sendalot = FALSE;
208 off = tp->snd_nxt - tp->snd_una;
209 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
210 sendwin = min(sendwin, tp->snd_bwnd);
212 flags = tcp_outflags[tp->t_state];
214 * Get standard flags, and add SYN or FIN if requested by 'hidden'
215 * state flags.
217 if (tp->t_flags & TF_NEEDFIN)
218 flags |= TH_FIN;
219 if (tp->t_flags & TF_NEEDSYN)
220 flags |= TH_SYN;
223 * If in persist timeout with window of 0, send 1 byte.
224 * Otherwise, if window is small but nonzero
225 * and timer expired, we will send what we can
226 * and go to transmit state.
228 if (tp->t_flags & TF_FORCE) {
229 if (sendwin == 0) {
231 * If we still have some data to send, then
232 * clear the FIN bit. Usually this would
233 * happen below when it realizes that we
234 * aren't sending all the data. However,
235 * if we have exactly 1 byte of unsent data,
236 * then it won't clear the FIN bit below,
237 * and if we are in persist state, we wind
238 * up sending the packet without recording
239 * that we sent the FIN bit.
241 * We can't just blindly clear the FIN bit,
242 * because if we don't have any more data
243 * to send then the probe will be the FIN
244 * itself.
246 if (off < so->so_snd.ssb_cc)
247 flags &= ~TH_FIN;
248 sendwin = 1;
249 } else {
250 callout_stop(tp->tt_persist);
251 tp->t_rxtshift = 0;
256 * If snd_nxt == snd_max and we have transmitted a FIN, the
257 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
258 * a negative length. This can also occur when TCP opens up
259 * its congestion window while receiving additional duplicate
260 * acks after fast-retransmit because TCP will reset snd_nxt
261 * to snd_max after the fast-retransmit.
263 * In the normal retransmit-FIN-only case, however, snd_nxt will
264 * be set to snd_una, the offset will be 0, and the length may
265 * wind up 0.
267 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
270 * Lop off SYN bit if it has already been sent. However, if this
271 * is SYN-SENT state and if segment contains data and if we don't
272 * know that foreign host supports TAO, suppress sending segment.
274 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
275 flags &= ~TH_SYN;
276 off--, len++;
277 if (len > 0 && tp->t_state == TCPS_SYN_SENT &&
278 ((taop = tcp_gettaocache(&inp->inp_inc)) == NULL ||
279 taop->tao_ccsent == 0))
280 return 0;
284 * Be careful not to send data and/or FIN on SYN segments
285 * in cases when no CC option will be sent.
286 * This measure is needed to prevent interoperability problems
287 * with not fully conformant TCP implementations.
289 if ((flags & TH_SYN) &&
290 ((tp->t_flags & TF_NOOPT) || !(tp->t_flags & TF_REQ_CC) ||
291 ((flags & TH_ACK) && !(tp->t_flags & TF_RCVD_CC)))) {
292 len = 0;
293 flags &= ~TH_FIN;
296 if (len < 0) {
298 * If FIN has been sent but not acked,
299 * but we haven't been called to retransmit,
300 * len will be < 0. Otherwise, window shrank
301 * after we sent into it. If window shrank to 0,
302 * cancel pending retransmit, pull snd_nxt back
303 * to (closed) window, and set the persist timer
304 * if it isn't already going. If the window didn't
305 * close completely, just wait for an ACK.
307 len = 0;
308 if (sendwin == 0) {
309 callout_stop(tp->tt_rexmt);
310 tp->t_rxtshift = 0;
311 tp->snd_nxt = tp->snd_una;
312 if (!callout_active(tp->tt_persist))
313 tcp_setpersist(tp);
318 * len will be >= 0 after this point. Truncate to the maximum
319 * segment length and ensure that FIN is removed if the length
320 * no longer contains the last data byte.
322 if (len > tp->t_maxseg) {
323 len = tp->t_maxseg;
324 sendalot = TRUE;
326 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
327 flags &= ~TH_FIN;
329 recvwin = ssb_space(&so->so_rcv);
332 * Sender silly window avoidance. We transmit under the following
333 * conditions when len is non-zero:
335 * - We have a full segment
336 * - This is the last buffer in a write()/send() and we are
337 * either idle or running NODELAY
338 * - we've timed out (e.g. persist timer)
339 * - we have more then 1/2 the maximum send window's worth of
340 * data (receiver may be limiting the window size)
341 * - we need to retransmit
343 if (len) {
344 if (len == tp->t_maxseg)
345 goto send;
347 * NOTE! on localhost connections an 'ack' from the remote
348 * end may occur synchronously with the output and cause
349 * us to flush a buffer queued with moretocome. XXX
351 * note: the len + off check is almost certainly unnecessary.
353 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
354 (idle || (tp->t_flags & TF_NODELAY)) &&
355 len + off >= so->so_snd.ssb_cc &&
356 !(tp->t_flags & TF_NOPUSH)) {
357 goto send;
359 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
360 goto send;
361 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
362 goto send;
363 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
364 goto send;
368 * Compare available window to amount of window
369 * known to peer (as advertised window less
370 * next expected input). If the difference is at least two
371 * max size segments, or at least 50% of the maximum possible
372 * window, then want to send a window update to peer.
374 if (recvwin > 0) {
376 * "adv" is the amount we can increase the window,
377 * taking into account that we are limited by
378 * TCP_MAXWIN << tp->rcv_scale.
380 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
381 (tp->rcv_adv - tp->rcv_nxt);
384 * This ack case typically occurs when the user has drained
385 * the TCP socket buffer sufficiently to warrent an ack
386 * containing a 'pure window update'... that is, an ack that
387 * ONLY updates the tcp window.
389 * It is unclear why we would need to do a pure window update
390 * past 2 segments if we are going to do one at 1/2 the high
391 * water mark anyway, especially since under normal conditions
392 * the user program will drain the socket buffer quickly.
393 * The 2-segment pure window update will often add a large
394 * number of extra, unnecessary acks to the stream.
396 * avoid_pure_win_update now defaults to 1.
398 if (avoid_pure_win_update == 0) {
399 if (adv >= (long) (2 * tp->t_maxseg))
400 goto send;
402 if (2 * adv >= (long) so->so_rcv.ssb_hiwat)
403 goto send;
407 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
408 * is also a catch-all for the retransmit timer timeout case.
410 if (tp->t_flags & TF_ACKNOW)
411 goto send;
412 if ((flags & TH_RST) ||
413 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
414 goto send;
415 if (SEQ_GT(tp->snd_up, tp->snd_una))
416 goto send;
418 * If our state indicates that FIN should be sent
419 * and we have not yet done so, then we need to send.
421 if (flags & TH_FIN &&
422 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
423 goto send;
426 * TCP window updates are not reliable, rather a polling protocol
427 * using ``persist'' packets is used to insure receipt of window
428 * updates. The three ``states'' for the output side are:
429 * idle not doing retransmits or persists
430 * persisting to move a small or zero window
431 * (re)transmitting and thereby not persisting
433 * callout_active(tp->tt_persist)
434 * is true when we are in persist state.
435 * The TF_FORCE flag in tp->t_flags
436 * is set when we are called to send a persist packet.
437 * callout_active(tp->tt_rexmt)
438 * is set when we are retransmitting
439 * The output side is idle when both timers are zero.
441 * If send window is too small, there is data to transmit, and no
442 * retransmit or persist is pending, then go to persist state.
443 * If nothing happens soon, send when timer expires:
444 * if window is nonzero, transmit what we can,
445 * otherwise force out a byte.
447 if (so->so_snd.ssb_cc > 0 &&
448 !callout_active(tp->tt_rexmt) && !callout_active(tp->tt_persist)) {
449 tp->t_rxtshift = 0;
450 tcp_setpersist(tp);
454 * No reason to send a segment, just return.
456 return (0);
458 send:
460 * Before ESTABLISHED, force sending of initial options
461 * unless TCP set not to do any options.
462 * NOTE: we assume that the IP/TCP header plus TCP options
463 * always fit in a single mbuf, leaving room for a maximum
464 * link header, i.e.
465 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
467 optlen = 0;
468 if (isipv6)
469 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
470 else
471 hdrlen = sizeof(struct tcpiphdr);
472 if (flags & TH_SYN) {
473 tp->snd_nxt = tp->iss;
474 if (!(tp->t_flags & TF_NOOPT)) {
475 u_short mss;
477 opt[0] = TCPOPT_MAXSEG;
478 opt[1] = TCPOLEN_MAXSEG;
479 mss = htons((u_short) tcp_mssopt(tp));
480 memcpy(opt + 2, &mss, sizeof mss);
481 optlen = TCPOLEN_MAXSEG;
483 if ((tp->t_flags & TF_REQ_SCALE) &&
484 (!(flags & TH_ACK) ||
485 (tp->t_flags & TF_RCVD_SCALE))) {
486 *((u_int32_t *)(opt + optlen)) = htonl(
487 TCPOPT_NOP << 24 |
488 TCPOPT_WINDOW << 16 |
489 TCPOLEN_WINDOW << 8 |
490 tp->request_r_scale);
491 optlen += 4;
494 if ((tcp_do_sack && !(flags & TH_ACK)) ||
495 tp->t_flags & TF_SACK_PERMITTED) {
496 uint32_t *lp = (uint32_t *)(opt + optlen);
498 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
499 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
505 * Send a timestamp and echo-reply if this is a SYN and our side
506 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
507 * and our peer have sent timestamps in our SYN's.
509 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
510 !(flags & TH_RST) &&
511 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
512 u_int32_t *lp = (u_int32_t *)(opt + optlen);
514 /* Form timestamp option as shown in appendix A of RFC 1323. */
515 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
516 *lp++ = htonl(ticks);
517 *lp = htonl(tp->ts_recent);
518 optlen += TCPOLEN_TSTAMP_APPA;
522 * Send `CC-family' options if our side wants to use them (TF_REQ_CC),
523 * options are allowed (!TF_NOOPT) and it's not a RST.
525 if ((tp->t_flags & (TF_REQ_CC | TF_NOOPT)) == TF_REQ_CC &&
526 !(flags & TH_RST)) {
527 switch (flags & (TH_SYN | TH_ACK)) {
529 * This is a normal ACK, send CC if we received CC before
530 * from our peer.
532 case TH_ACK:
533 if (!(tp->t_flags & TF_RCVD_CC))
534 break;
535 /*FALLTHROUGH*/
538 * We can only get here in T/TCP's SYN_SENT* state, when
539 * we're a sending a non-SYN segment without waiting for
540 * the ACK of our SYN. A check above assures that we only
541 * do this if our peer understands T/TCP.
543 case 0:
544 opt[optlen++] = TCPOPT_NOP;
545 opt[optlen++] = TCPOPT_NOP;
546 opt[optlen++] = TCPOPT_CC;
547 opt[optlen++] = TCPOLEN_CC;
548 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
549 optlen += 4;
550 break;
553 * This is our initial SYN, check whether we have to use
554 * CC or CC.new.
556 case TH_SYN:
557 opt[optlen++] = TCPOPT_NOP;
558 opt[optlen++] = TCPOPT_NOP;
559 opt[optlen++] = tp->t_flags & TF_SENDCCNEW ?
560 TCPOPT_CCNEW : TCPOPT_CC;
561 opt[optlen++] = TCPOLEN_CC;
562 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
563 optlen += 4;
564 break;
567 * This is a SYN,ACK; send CC and CC.echo if we received
568 * CC from our peer.
570 case (TH_SYN | TH_ACK):
571 if (tp->t_flags & TF_RCVD_CC) {
572 opt[optlen++] = TCPOPT_NOP;
573 opt[optlen++] = TCPOPT_NOP;
574 opt[optlen++] = TCPOPT_CC;
575 opt[optlen++] = TCPOLEN_CC;
576 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_send);
577 optlen += 4;
578 opt[optlen++] = TCPOPT_NOP;
579 opt[optlen++] = TCPOPT_NOP;
580 opt[optlen++] = TCPOPT_CCECHO;
581 opt[optlen++] = TCPOLEN_CC;
582 *(u_int32_t *)&opt[optlen] = htonl(tp->cc_recv);
583 optlen += 4;
585 break;
590 * If this is a SACK connection and we have a block to report,
591 * fill in the SACK blocks in the TCP options.
593 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
594 TF_SACK_PERMITTED &&
595 (!LIST_EMPTY(&tp->t_segq) ||
596 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
597 tcp_sack_fill_report(tp, opt, &optlen);
599 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
600 hdrlen += optlen;
602 if (isipv6) {
603 ipoptlen = ip6_optlen(inp);
604 } else {
605 if (inp->inp_options) {
606 ipoptlen = inp->inp_options->m_len -
607 offsetof(struct ipoption, ipopt_list);
608 } else {
609 ipoptlen = 0;
612 #ifdef IPSEC
613 ipoptlen += ipsec_hdrsiz_tcp(tp);
614 #endif
617 * Adjust data length if insertion of options will bump the packet
618 * length beyond the t_maxopd length. Clear FIN to prevent premature
619 * closure since there is still more data to send after this (now
620 * truncated) packet.
622 * If just the options do not fit we are in a no-win situation and
623 * we treat it as an unreachable host.
625 if (len + optlen + ipoptlen > tp->t_maxopd) {
626 if (tp->t_maxopd <= optlen + ipoptlen) {
627 static time_t last_optlen_report;
629 if (last_optlen_report != time_second) {
630 last_optlen_report = time_second;
631 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd);
633 error = EHOSTUNREACH;
634 goto out;
635 } else {
636 flags &= ~TH_FIN;
637 len = tp->t_maxopd - optlen - ipoptlen;
638 sendalot = TRUE;
642 #ifdef INET6
643 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
644 #else
645 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
646 #endif
649 * Grab a header mbuf, attaching a copy of data to
650 * be transmitted, and initialize the header from
651 * the template for sends on this connection.
653 if (len) {
654 if ((tp->t_flags & TF_FORCE) && len == 1)
655 tcpstat.tcps_sndprobe++;
656 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
657 if (tp->snd_nxt == tp->snd_una)
658 tp->snd_max_rexmt = tp->snd_max;
659 tcpstat.tcps_sndrexmitpack++;
660 tcpstat.tcps_sndrexmitbyte += len;
661 } else {
662 tcpstat.tcps_sndpack++;
663 tcpstat.tcps_sndbyte += len;
665 #ifdef notyet
666 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
667 max_linkhdr + hdrlen)) == NULL) {
668 error = ENOBUFS;
669 goto out;
672 * m_copypack left space for our hdr; use it.
674 m->m_len += hdrlen;
675 m->m_data -= hdrlen;
676 #else
677 #ifndef INET6
678 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
679 #else
680 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
681 M_PKTHDR, NULL);
682 #endif
683 if (m == NULL) {
684 error = ENOBUFS;
685 goto out;
687 m->m_data += max_linkhdr;
688 m->m_len = hdrlen;
689 if (len <= MHLEN - hdrlen - max_linkhdr) {
690 m_copydata(so->so_snd.ssb_mb, off, (int) len,
691 mtod(m, caddr_t) + hdrlen);
692 m->m_len += len;
693 } else {
694 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
695 if (m->m_next == NULL) {
696 m_free(m);
697 error = ENOBUFS;
698 goto out;
701 #endif
703 * If we're sending everything we've got, set PUSH.
704 * (This will keep happy those implementations which only
705 * give data to the user when a buffer fills or
706 * a PUSH comes in.)
708 if (off + len == so->so_snd.ssb_cc)
709 flags |= TH_PUSH;
710 } else {
711 if (tp->t_flags & TF_ACKNOW)
712 tcpstat.tcps_sndacks++;
713 else if (flags & (TH_SYN | TH_FIN | TH_RST))
714 tcpstat.tcps_sndctrl++;
715 else if (SEQ_GT(tp->snd_up, tp->snd_una))
716 tcpstat.tcps_sndurg++;
717 else
718 tcpstat.tcps_sndwinup++;
720 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
721 if (m == NULL) {
722 error = ENOBUFS;
723 goto out;
725 if (isipv6 &&
726 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
727 MH_ALIGN(m, hdrlen);
728 else
729 m->m_data += max_linkhdr;
730 m->m_len = hdrlen;
732 m->m_pkthdr.rcvif = (struct ifnet *)0;
733 if (isipv6) {
734 ip6 = mtod(m, struct ip6_hdr *);
735 th = (struct tcphdr *)(ip6 + 1);
736 tcp_fillheaders(tp, ip6, th);
737 } else {
738 ip = mtod(m, struct ip *);
739 ipov = (struct ipovly *)ip;
740 th = (struct tcphdr *)(ip + 1);
741 /* this picks up the pseudo header (w/o the length) */
742 tcp_fillheaders(tp, ip, th);
746 * Fill in fields, remembering maximum advertised
747 * window for use in delaying messages about window sizes.
748 * If resending a FIN, be sure not to use a new sequence number.
750 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
751 tp->snd_nxt == tp->snd_max)
752 tp->snd_nxt--;
754 * If we are doing retransmissions, then snd_nxt will
755 * not reflect the first unsent octet. For ACK only
756 * packets, we do not want the sequence number of the
757 * retransmitted packet, we want the sequence number
758 * of the next unsent octet. So, if there is no data
759 * (and no SYN or FIN), use snd_max instead of snd_nxt
760 * when filling in ti_seq. But if we are in persist
761 * state, snd_max might reflect one byte beyond the
762 * right edge of the window, so use snd_nxt in that
763 * case, since we know we aren't doing a retransmission.
764 * (retransmit and persist are mutually exclusive...)
766 if (len || (flags & (TH_SYN|TH_FIN)) || callout_active(tp->tt_persist))
767 th->th_seq = htonl(tp->snd_nxt);
768 else
769 th->th_seq = htonl(tp->snd_max);
770 th->th_ack = htonl(tp->rcv_nxt);
771 if (optlen) {
772 bcopy(opt, th + 1, optlen);
773 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
775 th->th_flags = flags;
777 * Calculate receive window. Don't shrink window,
778 * but avoid silly window syndrome.
780 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
781 recvwin < (long)tp->t_maxseg)
782 recvwin = 0;
783 if (recvwin < (long)(tp->rcv_adv - tp->rcv_nxt))
784 recvwin = (long)(tp->rcv_adv - tp->rcv_nxt);
785 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
786 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
787 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
790 * Adjust the RXWIN0SENT flag - indicate that we have advertised
791 * a 0 window. This may cause the remote transmitter to stall. This
792 * flag tells soreceive() to disable delayed acknowledgements when
793 * draining the buffer. This can occur if the receiver is attempting
794 * to read more data then can be buffered prior to transmitting on
795 * the connection.
797 if (recvwin == 0)
798 tp->t_flags |= TF_RXWIN0SENT;
799 else
800 tp->t_flags &= ~TF_RXWIN0SENT;
802 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
803 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
804 th->th_flags |= TH_URG;
805 } else {
807 * If no urgent pointer to send, then we pull
808 * the urgent pointer to the left edge of the send window
809 * so that it doesn't drift into the send window on sequence
810 * number wraparound.
812 tp->snd_up = tp->snd_una; /* drag it along */
816 * Put TCP length in extended header, and then
817 * checksum extended header and data.
819 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
820 if (isipv6) {
822 * ip6_plen is not need to be filled now, and will be filled
823 * in ip6_output().
825 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
826 sizeof(struct tcphdr) + optlen + len);
827 } else {
828 m->m_pkthdr.csum_flags = CSUM_TCP;
829 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
830 if (len + optlen)
831 th->th_sum = in_addword(th->th_sum,
832 htons((u_short)(optlen + len)));
834 /* IP version must be set here for ipv4/ipv6 checking later */
835 KASSERT(ip->ip_v == IPVERSION,
836 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
840 * In transmit state, time the transmission and arrange for
841 * the retransmit. In persist state, just set snd_max.
843 if (!(tp->t_flags & TF_FORCE) || !callout_active(tp->tt_persist)) {
844 tcp_seq startseq = tp->snd_nxt;
847 * Advance snd_nxt over sequence space of this segment.
849 if (flags & (TH_SYN | TH_FIN)) {
850 if (flags & TH_SYN)
851 tp->snd_nxt++;
852 if (flags & TH_FIN) {
853 tp->snd_nxt++;
854 tp->t_flags |= TF_SENTFIN;
857 tp->snd_nxt += len;
858 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
859 tp->snd_max = tp->snd_nxt;
861 * Time this transmission if not a retransmission and
862 * not currently timing anything.
864 if (tp->t_rtttime == 0) {
865 tp->t_rtttime = ticks;
866 tp->t_rtseq = startseq;
867 tcpstat.tcps_segstimed++;
872 * Set retransmit timer if not currently set,
873 * and not doing a pure ack or a keep-alive probe.
874 * Initial value for retransmit timer is smoothed
875 * round-trip time + 2 * round-trip time variance.
876 * Initialize shift counter which is used for backoff
877 * of retransmit time.
879 if (!callout_active(tp->tt_rexmt) &&
880 tp->snd_nxt != tp->snd_una) {
881 if (callout_active(tp->tt_persist)) {
882 callout_stop(tp->tt_persist);
883 tp->t_rxtshift = 0;
885 callout_reset(tp->tt_rexmt, tp->t_rxtcur,
886 tcp_timer_rexmt, tp);
888 } else {
890 * Persist case, update snd_max but since we are in
891 * persist mode (no window) we do not update snd_nxt.
893 int xlen = len;
894 if (flags & TH_SYN)
895 ++xlen;
896 if (flags & TH_FIN) {
897 ++xlen;
898 tp->t_flags |= TF_SENTFIN;
900 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
901 tp->snd_max = tp->snd_nxt + xlen;
904 #ifdef TCPDEBUG
906 * Trace.
908 if (so->so_options & SO_DEBUG)
909 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
910 #endif
913 * Fill in IP length and desired time to live and
914 * send to IP level. There should be a better way
915 * to handle ttl and tos; we could keep them in
916 * the template, but need a way to checksum without them.
919 * m->m_pkthdr.len should have been set before cksum calcuration,
920 * because in6_cksum() need it.
922 if (isipv6) {
924 * we separately set hoplimit for every segment, since the
925 * user might want to change the value via setsockopt.
926 * Also, desired default hop limit might be changed via
927 * Neighbor Discovery.
929 ip6->ip6_hlim = in6_selecthlim(inp,
930 (inp->in6p_route.ro_rt ?
931 inp->in6p_route.ro_rt->rt_ifp : NULL));
933 /* TODO: IPv6 IP6TOS_ECT bit on */
934 error = ip6_output(m, inp->in6p_outputopts, &inp->in6p_route,
935 (so->so_options & SO_DONTROUTE), NULL, NULL,
936 inp);
937 } else {
938 struct rtentry *rt;
939 ip->ip_len = m->m_pkthdr.len;
940 #ifdef INET6
941 if (INP_CHECK_SOCKAF(so, AF_INET6))
942 ip->ip_ttl = in6_selecthlim(inp,
943 (inp->in6p_route.ro_rt ?
944 inp->in6p_route.ro_rt->rt_ifp : NULL));
945 else
946 #endif
947 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
949 ip->ip_tos = inp->inp_ip_tos; /* XXX */
951 * See if we should do MTU discovery.
952 * We do it only if the following are true:
953 * 1) we have a valid route to the destination
954 * 2) the MTU is not locked (if it is,
955 * then discovery has been disabled)
957 if (path_mtu_discovery &&
958 (rt = inp->inp_route.ro_rt) && (rt->rt_flags & RTF_UP) &&
959 !(rt->rt_rmx.rmx_locks & RTV_MTU))
960 ip->ip_off |= IP_DF;
962 error = ip_output(m, inp->inp_options, &inp->inp_route,
963 (so->so_options & SO_DONTROUTE), NULL, inp);
965 if (error) {
968 * We know that the packet was lost, so back out the
969 * sequence number advance, if any.
971 if (!(tp->t_flags & TF_FORCE) ||
972 !callout_active(tp->tt_persist)) {
974 * No need to check for TH_FIN here because
975 * the TF_SENTFIN flag handles that case.
977 if (!(flags & TH_SYN))
978 tp->snd_nxt -= len;
981 out:
982 if (error == ENOBUFS) {
984 * If we can't send, make sure there is something
985 * to get us going again later. Persist state
986 * is not necessarily right, but it is close enough.
988 if (!callout_active(tp->tt_rexmt) &&
989 !callout_active(tp->tt_persist)) {
990 tp->t_rxtshift = 0;
991 tcp_setpersist(tp);
993 tcp_quench(inp, 0);
994 return (0);
996 if (error == EMSGSIZE) {
998 * ip_output() will have already fixed the route
999 * for us. tcp_mtudisc() will, as its last action,
1000 * initiate retransmission, so it is important to
1001 * not do so here.
1003 tcp_mtudisc(inp, 0);
1004 return 0;
1006 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1007 TCPS_HAVERCVDSYN(tp->t_state)) {
1008 tp->t_softerror = error;
1009 return (0);
1011 return (error);
1013 tcpstat.tcps_sndtotal++;
1016 * Data sent (as far as we can tell).
1017 * If this advertises a larger window than any other segment,
1018 * then remember the size of the advertised window.
1019 * Any pending ACK has now been sent.
1021 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv))
1022 tp->rcv_adv = tp->rcv_nxt + recvwin;
1023 tp->last_ack_sent = tp->rcv_nxt;
1024 tp->t_flags &= ~TF_ACKNOW;
1025 if (tcp_delack_enabled)
1026 callout_stop(tp->tt_delack);
1027 if (sendalot)
1028 goto again;
1029 return (0);
1032 void
1033 tcp_setpersist(struct tcpcb *tp)
1035 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1036 int tt;
1038 if (callout_active(tp->tt_rexmt))
1039 panic("tcp_setpersist: retransmit pending");
1041 * Start/restart persistance timer.
1043 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1044 TCPTV_PERSMAX);
1045 callout_reset(tp->tt_persist, tt, tcp_timer_persist, tp);
1046 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1047 tp->t_rxtshift++;