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
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
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
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
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
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 $
72 #include "opt_inet6.h"
73 #include "opt_ipsec.h"
74 #include "opt_tcpdebug.h"
76 #include <sys/param.h>
77 #include <sys/systm.h>
78 #include <sys/kernel.h>
79 #include <sys/sysctl.h>
81 #include <sys/domain.h>
82 #include <sys/protosw.h>
83 #include <sys/socket.h>
84 #include <sys/socketvar.h>
85 #include <sys/in_cksum.h>
86 #include <sys/thread.h>
87 #include <sys/globaldata.h>
89 #include <net/route.h>
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>
96 #include <netinet6/in6_pcb.h>
97 #include <netinet/ip6.h>
98 #include <netinet6/ip6_var.h>
99 #include <netinet/tcp.h>
101 #include <netinet/tcp_fsm.h>
102 #include <netinet/tcp_seq.h>
103 #include <netinet/tcp_timer.h>
104 #include <netinet/tcp_timer2.h>
105 #include <netinet/tcp_var.h>
106 #include <netinet/tcpip.h>
108 #include <netinet/tcp_debug.h>
112 #include <netinet6/ipsec.h>
116 #include <netproto/ipsec/ipsec.h>
118 #endif /*FAST_IPSEC*/
121 extern struct mbuf *m_copypack();
124 int path_mtu_discovery = 0;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
126 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
128 static int avoid_pure_win_update = 1;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
130 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
132 int tcp_do_autosndbuf = 1;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
134 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
136 int tcp_autosndbuf_inc = 8*1024;
137 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
138 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
140 int tcp_autosndbuf_max = 2*1024*1024;
141 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
142 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
145 * Tcp output routine: figure out what should be sent and send it.
148 tcp_output(struct tcpcb *tp)
150 struct inpcb * const inp = tp->t_inpcb;
151 struct socket *so = inp->inp_socket;
152 long len, recvwin, sendwin;
154 int off, flags, error = 0;
162 u_char opt[TCP_MAXOLEN];
163 unsigned int ipoptlen, optlen, hdrlen;
168 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
170 const boolean_t isipv6 = FALSE;
173 KKASSERT(so->so_port == &curthread->td_msgport);
176 * Determine length of data that should be transmitted,
177 * and flags that will be used.
178 * If there is some data or critical controls (SYN, RST)
179 * to send, then transmit; otherwise, investigate further.
183 * If we have been idle for a while, the send congestion window
184 * could be no longer representative of the current state of the link.
185 * So unless we are expecting more acks to come in, slow-start from
186 * scratch to re-determine the send congestion window.
188 if (tp->snd_max == tp->snd_una &&
189 (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
190 u_long initial_cwnd = tcp_initial_window(tp);
193 * According to RFC5681:
196 tp->snd_cwnd = min(tp->snd_cwnd, initial_cwnd);
201 * Calculate whether the transmit stream was previously idle
202 * and adjust TF_LASTIDLE for the next time.
204 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
205 if (idle && (tp->t_flags & TF_MORETOCOME))
206 tp->t_flags |= TF_LASTIDLE;
208 tp->t_flags &= ~TF_LASTIDLE;
210 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
211 !IN_FASTRECOVERY(tp))
212 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
221 /* Make use of SACK information when slow-starting after a RTO. */
222 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
223 !IN_FASTRECOVERY(tp)) {
224 tcp_seq old_snd_nxt = tp->snd_nxt;
226 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
227 nsacked += tp->snd_nxt - old_snd_nxt;
231 off = tp->snd_nxt - tp->snd_una;
232 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
233 sendwin = min(sendwin, tp->snd_bwnd);
235 flags = tcp_outflags[tp->t_state];
237 * Get standard flags, and add SYN or FIN if requested by 'hidden'
240 if (tp->t_flags & TF_NEEDFIN)
242 if (tp->t_flags & TF_NEEDSYN)
246 * If in persist timeout with window of 0, send 1 byte.
247 * Otherwise, if window is small but nonzero
248 * and timer expired, we will send what we can
249 * and go to transmit state.
251 if (tp->t_flags & TF_FORCE) {
254 * If we still have some data to send, then
255 * clear the FIN bit. Usually this would
256 * happen below when it realizes that we
257 * aren't sending all the data. However,
258 * if we have exactly 1 byte of unsent data,
259 * then it won't clear the FIN bit below,
260 * and if we are in persist state, we wind
261 * up sending the packet without recording
262 * that we sent the FIN bit.
264 * We can't just blindly clear the FIN bit,
265 * because if we don't have any more data
266 * to send then the probe will be the FIN
269 if (off < so->so_snd.ssb_cc)
273 tcp_callout_stop(tp, tp->tt_persist);
279 * If snd_nxt == snd_max and we have transmitted a FIN, the
280 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
281 * a negative length. This can also occur when TCP opens up
282 * its congestion window while receiving additional duplicate
283 * acks after fast-retransmit because TCP will reset snd_nxt
284 * to snd_max after the fast-retransmit.
286 * A negative length can also occur when we are in the
287 * TCPS_SYN_RECEIVED state due to a simultanious connect where
288 * our SYN has not been acked yet.
290 * In the normal retransmit-FIN-only case, however, snd_nxt will
291 * be set to snd_una, the offset will be 0, and the length may
294 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
297 * Lop off SYN bit if it has already been sent. However, if this
298 * is SYN-SENT state and if segment contains data, suppress sending
299 * segment (sending the segment would be an option if we still
300 * did TAO and the remote host supported it).
302 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
305 if (len > 0 && tp->t_state == TCPS_SYN_SENT)
310 * Be careful not to send data and/or FIN on SYN segments.
311 * This measure is needed to prevent interoperability problems
312 * with not fully conformant TCP implementations.
314 if (flags & TH_SYN) {
321 * A negative len can occur if our FIN has been sent but not
322 * acked, or if we are in a simultanious connect in the
323 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
326 * If our window has contracted to 0 in the FIN case
327 * (which can only occur if we have NOT been called to
328 * retransmit as per code a few paragraphs up) then we
329 * want to shift the retransmit timer over to the
332 * However, if we are in the TCPS_SYN_RECEIVED state
333 * (the SYN case) we will be in a simultanious connect and
334 * the window may be zero degeneratively. In this case we
335 * do not want to shift to the persist timer after the SYN
336 * or the SYN+ACK transmission.
339 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
340 tcp_callout_stop(tp, tp->tt_rexmt);
342 tp->snd_nxt = tp->snd_una;
343 if (!tcp_callout_active(tp, tp->tt_persist))
348 KASSERT(len >= 0, ("%s: len < 0", __func__));
350 * Automatic sizing of send socket buffer. Often the send buffer
351 * size is not optimally adjusted to the actual network conditions
352 * at hand (delay bandwidth product). Setting the buffer size too
353 * small limits throughput on links with high bandwidth and high
354 * delay (eg. trans-continental/oceanic links). Setting the
355 * buffer size too big consumes too much real kernel memory,
356 * especially with many connections on busy servers.
358 * The criteria to step up the send buffer one notch are:
359 * 1. receive window of remote host is larger than send buffer
360 * (with a fudge factor of 5/4th);
361 * 2. send buffer is filled to 7/8th with data (so we actually
362 * have data to make use of it);
363 * 3. send buffer fill has not hit maximal automatic size;
364 * 4. our send window (slow start and cogestion controlled) is
365 * larger than sent but unacknowledged data in send buffer.
367 * The remote host receive window scaling factor may limit the
368 * growing of the send buffer before it reaches its allowed
371 * It scales directly with slow start or congestion window
372 * and does at most one step per received ACK. This fast
373 * scaling has the drawback of growing the send buffer beyond
374 * what is strictly necessary to make full use of a given
375 * delay*bandwith product. However testing has shown this not
376 * to be much of an problem. At worst we are trading wasting
377 * of available bandwith (the non-use of it) for wasting some
378 * socket buffer memory.
380 * TODO: Shrink send buffer during idle periods together
381 * with congestion window. Requires another timer. Has to
382 * wait for upcoming tcp timer rewrite.
384 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
385 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
386 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
387 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
388 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
391 newsize = ulmin(so->so_snd.ssb_hiwat +
394 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
395 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
396 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
397 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
402 * Truncate to the maximum segment length and ensure that FIN is
403 * removed if the length no longer contains the last data byte.
405 if (len > tp->t_maxseg) {
409 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
412 recvwin = ssb_space(&so->so_rcv);
415 * Sender silly window avoidance. We transmit under the following
416 * conditions when len is non-zero:
418 * - We have a full segment
419 * - This is the last buffer in a write()/send() and we are
420 * either idle or running NODELAY
421 * - we've timed out (e.g. persist timer)
422 * - we have more then 1/2 the maximum send window's worth of
423 * data (receiver may be limiting the window size)
424 * - we need to retransmit
427 if (len == tp->t_maxseg)
430 * NOTE! on localhost connections an 'ack' from the remote
431 * end may occur synchronously with the output and cause
432 * us to flush a buffer queued with moretocome. XXX
434 * note: the len + off check is almost certainly unnecessary.
436 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
437 (idle || (tp->t_flags & TF_NODELAY)) &&
438 len + off >= so->so_snd.ssb_cc &&
439 !(tp->t_flags & TF_NOPUSH)) {
442 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
444 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
446 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
451 * Compare available window to amount of window
452 * known to peer (as advertised window less
453 * next expected input). If the difference is at least two
454 * max size segments, or at least 50% of the maximum possible
455 * window, then want to send a window update to peer.
459 * "adv" is the amount we can increase the window,
460 * taking into account that we are limited by
461 * TCP_MAXWIN << tp->rcv_scale.
463 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
464 (tp->rcv_adv - tp->rcv_nxt);
468 * This ack case typically occurs when the user has drained
469 * the TCP socket buffer sufficiently to warrent an ack
470 * containing a 'pure window update'... that is, an ack that
471 * ONLY updates the tcp window.
473 * It is unclear why we would need to do a pure window update
474 * past 2 segments if we are going to do one at 1/2 the high
475 * water mark anyway, especially since under normal conditions
476 * the user program will drain the socket buffer quickly.
477 * The 2-segment pure window update will often add a large
478 * number of extra, unnecessary acks to the stream.
480 * avoid_pure_win_update now defaults to 1.
482 if (avoid_pure_win_update == 0 ||
483 (tp->t_flags & TF_RXRESIZED)) {
484 if (adv >= (long) (2 * tp->t_maxseg)) {
488 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
489 if (hiwat > (long)so->so_rcv.ssb_hiwat)
490 hiwat = (long)so->so_rcv.ssb_hiwat;
491 if (adv >= hiwat / 2)
496 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
497 * is also a catch-all for the retransmit timer timeout case.
499 if (tp->t_flags & TF_ACKNOW)
501 if ((flags & TH_RST) ||
502 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
504 if (SEQ_GT(tp->snd_up, tp->snd_una))
507 * If our state indicates that FIN should be sent
508 * and we have not yet done so, then we need to send.
510 if ((flags & TH_FIN) &&
511 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
515 * TCP window updates are not reliable, rather a polling protocol
516 * using ``persist'' packets is used to insure receipt of window
517 * updates. The three ``states'' for the output side are:
518 * idle not doing retransmits or persists
519 * persisting to move a small or zero window
520 * (re)transmitting and thereby not persisting
522 * tcp_callout_active(tp, tp->tt_persist)
523 * is true when we are in persist state.
524 * The TF_FORCE flag in tp->t_flags
525 * is set when we are called to send a persist packet.
526 * tcp_callout_active(tp, tp->tt_rexmt)
527 * is set when we are retransmitting
528 * The output side is idle when both timers are zero.
530 * If send window is too small, there is data to transmit, and no
531 * retransmit or persist is pending, then go to persist state.
533 * If nothing happens soon, send when timer expires:
534 * if window is nonzero, transmit what we can, otherwise force out
537 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
538 * with data pending. This situation can occur during a
539 * simultanious connect.
541 if (so->so_snd.ssb_cc > 0 &&
542 tp->t_state != TCPS_SYN_RECEIVED &&
543 !tcp_callout_active(tp, tp->tt_rexmt) &&
544 !tcp_callout_active(tp, tp->tt_persist)) {
550 * No reason to send a segment, just return.
556 * Before ESTABLISHED, force sending of initial options
557 * unless TCP set not to do any options.
558 * NOTE: we assume that the IP/TCP header plus TCP options
559 * always fit in a single mbuf, leaving room for a maximum
561 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
565 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
567 hdrlen = sizeof(struct tcpiphdr);
568 if (flags & TH_SYN) {
569 tp->snd_nxt = tp->iss;
570 if (!(tp->t_flags & TF_NOOPT)) {
573 opt[0] = TCPOPT_MAXSEG;
574 opt[1] = TCPOLEN_MAXSEG;
575 mss = htons((u_short) tcp_mssopt(tp));
576 memcpy(opt + 2, &mss, sizeof mss);
577 optlen = TCPOLEN_MAXSEG;
579 if ((tp->t_flags & TF_REQ_SCALE) &&
580 (!(flags & TH_ACK) ||
581 (tp->t_flags & TF_RCVD_SCALE))) {
582 *((u_int32_t *)(opt + optlen)) = htonl(
584 TCPOPT_WINDOW << 16 |
585 TCPOLEN_WINDOW << 8 |
586 tp->request_r_scale);
590 if ((tcp_do_sack && !(flags & TH_ACK)) ||
591 tp->t_flags & TF_SACK_PERMITTED) {
592 uint32_t *lp = (uint32_t *)(opt + optlen);
594 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
595 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
601 * Send a timestamp and echo-reply if this is a SYN and our side
602 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
603 * and our peer have sent timestamps in our SYN's.
605 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
607 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
608 u_int32_t *lp = (u_int32_t *)(opt + optlen);
610 /* Form timestamp option as shown in appendix A of RFC 1323. */
611 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
612 *lp++ = htonl(ticks);
613 *lp = htonl(tp->ts_recent);
614 optlen += TCPOLEN_TSTAMP_APPA;
617 /* Set receive buffer autosizing timestamp. */
618 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
619 tp->rfbuf_ts = ticks;
622 * If this is a SACK connection and we have a block to report,
623 * fill in the SACK blocks in the TCP options.
625 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
627 (!LIST_EMPTY(&tp->t_segq) ||
628 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
629 tcp_sack_fill_report(tp, opt, &optlen);
632 if (tp->t_flags & TF_SIGNATURE) {
636 * Initialize TCP-MD5 option (RFC2385)
638 bp = (u_char *)opt + optlen;
639 *bp++ = TCPOPT_SIGNATURE;
640 *bp++ = TCPOLEN_SIGNATURE;
642 for (i = 0; i < TCP_SIGLEN; i++)
644 optlen += TCPOLEN_SIGNATURE;
646 * Terminate options list and maintain 32-bit alignment.
652 #endif /* TCP_SIGNATURE */
653 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
657 ipoptlen = ip6_optlen(inp);
659 if (inp->inp_options) {
660 ipoptlen = inp->inp_options->m_len -
661 offsetof(struct ipoption, ipopt_list);
667 ipoptlen += ipsec_hdrsiz_tcp(tp);
671 * Adjust data length if insertion of options will bump the packet
672 * length beyond the t_maxopd length. Clear FIN to prevent premature
673 * closure since there is still more data to send after this (now
676 * If just the options do not fit we are in a no-win situation and
677 * we treat it as an unreachable host.
679 if (len + optlen + ipoptlen > tp->t_maxopd) {
680 if (tp->t_maxopd <= optlen + ipoptlen) {
681 static time_t last_optlen_report;
683 if (last_optlen_report != time_second) {
684 last_optlen_report = time_second;
685 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd);
687 error = EHOSTUNREACH;
691 len = tp->t_maxopd - optlen - ipoptlen;
697 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
699 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
703 * Grab a header mbuf, attaching a copy of data to
704 * be transmitted, and initialize the header from
705 * the template for sends on this connection.
708 if ((tp->t_flags & TF_FORCE) && len == 1)
709 tcpstat.tcps_sndprobe++;
710 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
711 if (tp->snd_nxt == tp->snd_una)
712 tp->snd_max_rexmt = tp->snd_max;
714 tcpstat.tcps_sndsackrtopack++;
715 tcpstat.tcps_sndsackrtobyte += len;
717 tcpstat.tcps_sndrexmitpack++;
718 tcpstat.tcps_sndrexmitbyte += len;
720 tcpstat.tcps_sndpack++;
721 tcpstat.tcps_sndbyte += len;
724 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
725 max_linkhdr + hdrlen)) == NULL) {
730 * m_copypack left space for our hdr; use it.
736 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
738 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
745 m->m_data += max_linkhdr;
747 if (len <= MHLEN - hdrlen - max_linkhdr) {
748 m_copydata(so->so_snd.ssb_mb, off, (int) len,
749 mtod(m, caddr_t) + hdrlen);
752 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
753 if (m->m_next == NULL) {
762 * If we're sending everything we've got, set PUSH.
763 * (This will keep happy those implementations which only
764 * give data to the user when a buffer fills or
767 if (off + len == so->so_snd.ssb_cc)
770 if (tp->t_flags & TF_ACKNOW)
771 tcpstat.tcps_sndacks++;
772 else if (flags & (TH_SYN | TH_FIN | TH_RST))
773 tcpstat.tcps_sndctrl++;
774 else if (SEQ_GT(tp->snd_up, tp->snd_una))
775 tcpstat.tcps_sndurg++;
777 tcpstat.tcps_sndwinup++;
779 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
785 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
788 m->m_data += max_linkhdr;
791 m->m_pkthdr.rcvif = NULL;
793 ip6 = mtod(m, struct ip6_hdr *);
794 th = (struct tcphdr *)(ip6 + 1);
795 tcp_fillheaders(tp, ip6, th);
797 ip = mtod(m, struct ip *);
798 ipov = (struct ipovly *)ip;
799 th = (struct tcphdr *)(ip + 1);
800 /* this picks up the pseudo header (w/o the length) */
801 tcp_fillheaders(tp, ip, th);
805 * Fill in fields, remembering maximum advertised
806 * window for use in delaying messages about window sizes.
807 * If resending a FIN, be sure not to use a new sequence number.
809 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
810 tp->snd_nxt == tp->snd_max)
815 * If we are doing retransmissions, then snd_nxt will
816 * not reflect the first unsent octet. For ACK only
817 * packets, we do not want the sequence number of the
818 * retransmitted packet, we want the sequence number
819 * of the next unsent octet. So, if there is no data
820 * (and no SYN or FIN), use snd_max instead of snd_nxt
821 * when filling in ti_seq. But if we are in persist
822 * state, snd_max might reflect one byte beyond the
823 * right edge of the window, so use snd_nxt in that
824 * case, since we know we aren't doing a retransmission.
825 * (retransmit and persist are mutually exclusive...)
827 if (len || (flags & (TH_SYN|TH_FIN)) ||
828 tcp_callout_active(tp, tp->tt_persist))
829 th->th_seq = htonl(tp->snd_nxt);
831 th->th_seq = htonl(tp->snd_max);
832 th->th_ack = htonl(tp->rcv_nxt);
834 bcopy(opt, th + 1, optlen);
835 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
837 th->th_flags = flags;
841 * Calculate receive window. Don't shrink window, but avoid
842 * silly window syndrome by sending a 0 window if the actual
843 * window is less then one segment.
845 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
846 recvwin < (long)tp->t_maxseg)
848 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
849 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
850 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
851 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
854 * Adjust the RXWIN0SENT flag - indicate that we have advertised
855 * a 0 window. This may cause the remote transmitter to stall. This
856 * flag tells soreceive() to disable delayed acknowledgements when
857 * draining the buffer. This can occur if the receiver is attempting
858 * to read more data then can be buffered prior to transmitting on
862 tp->t_flags |= TF_RXWIN0SENT;
864 tp->t_flags &= ~TF_RXWIN0SENT;
867 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
869 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
871 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
872 th->th_flags |= TH_URG;
876 * If no urgent pointer to send, then we pull
877 * the urgent pointer to the left edge of the send window
878 * so that it doesn't drift into the send window on sequence
881 tp->snd_up = tp->snd_una; /* drag it along */
886 if (tp->t_flags & TF_SIGNATURE) {
887 tcpsignature_compute(m, len, optlen,
888 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
890 #endif /* TCP_SIGNATURE */
893 * Put TCP length in extended header, and then
894 * checksum extended header and data.
896 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
899 * ip6_plen is not need to be filled now, and will be
900 * filled in ip6_output().
902 th->th_sum = in6_cksum(m, IPPROTO_TCP,
903 sizeof(struct ip6_hdr),
904 sizeof(struct tcphdr) + optlen + len);
906 m->m_pkthdr.csum_flags = CSUM_TCP;
907 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
909 th->th_sum = in_addword(th->th_sum,
910 htons((u_short)(optlen + len)));
914 * IP version must be set here for ipv4/ipv6 checking
917 KASSERT(ip->ip_v == IPVERSION,
918 ("%s: IP version incorrect: %d",
919 __func__, ip->ip_v));
924 * In transmit state, time the transmission and arrange for
925 * the retransmit. In persist state, just set snd_max.
927 if (!(tp->t_flags & TF_FORCE) ||
928 !tcp_callout_active(tp, tp->tt_persist)) {
929 tcp_seq startseq = tp->snd_nxt;
932 * Advance snd_nxt over sequence space of this segment.
934 if (flags & (TH_SYN | TH_FIN)) {
937 if (flags & TH_FIN) {
939 tp->t_flags |= TF_SENTFIN;
943 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
944 tp->snd_max = tp->snd_nxt;
946 * Time this transmission if not a retransmission and
947 * not currently timing anything.
949 if (tp->t_rtttime == 0) {
950 tp->t_rtttime = ticks;
951 tp->t_rtseq = startseq;
952 tcpstat.tcps_segstimed++;
957 * Set retransmit timer if not currently set,
958 * and not doing a pure ack or a keep-alive probe.
959 * Initial value for retransmit timer is smoothed
960 * round-trip time + 2 * round-trip time variance.
961 * Initialize shift counter which is used for backoff
962 * of retransmit time.
964 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
965 tp->snd_nxt != tp->snd_una) {
966 if (tcp_callout_active(tp, tp->tt_persist)) {
967 tcp_callout_stop(tp, tp->tt_persist);
970 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
975 * Persist case, update snd_max but since we are in
976 * persist mode (no window) we do not update snd_nxt.
980 panic("tcp_output: persist timer to send SYN\n");
981 if (flags & TH_FIN) {
983 tp->t_flags |= TF_SENTFIN;
985 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
986 tp->snd_max = tp->snd_nxt + xlen;
992 if (so->so_options & SO_DEBUG) {
993 tcp_trace(TA_OUTPUT, tp->t_state, tp,
994 mtod(m, void *), th, 0);
999 * Fill in IP length and desired time to live and
1000 * send to IP level. There should be a better way
1001 * to handle ttl and tos; we could keep them in
1002 * the template, but need a way to checksum without them.
1005 * m->m_pkthdr.len should have been set before cksum
1006 * calcuration, because in6_cksum() need it.
1010 * we separately set hoplimit for every segment,
1011 * since the user might want to change the value
1012 * via setsockopt. Also, desired default hop
1013 * limit might be changed via Neighbor Discovery.
1015 ip6->ip6_hlim = in6_selecthlim(inp,
1016 (inp->in6p_route.ro_rt ?
1017 inp->in6p_route.ro_rt->rt_ifp : NULL));
1019 /* TODO: IPv6 IP6TOS_ECT bit on */
1020 error = ip6_output(m, inp->in6p_outputopts,
1021 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1025 ip->ip_len = m->m_pkthdr.len;
1027 if (INP_CHECK_SOCKAF(so, AF_INET6))
1028 ip->ip_ttl = in6_selecthlim(inp,
1029 (inp->in6p_route.ro_rt ?
1030 inp->in6p_route.ro_rt->rt_ifp : NULL));
1033 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1035 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1037 * See if we should do MTU discovery.
1038 * We do it only if the following are true:
1039 * 1) we have a valid route to the destination
1040 * 2) the MTU is not locked (if it is,
1041 * then discovery has been disabled)
1043 if (path_mtu_discovery &&
1044 (rt = inp->inp_route.ro_rt) &&
1045 (rt->rt_flags & RTF_UP) &&
1046 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1047 ip->ip_off |= IP_DF;
1049 error = ip_output(m, inp->inp_options, &inp->inp_route,
1050 (so->so_options & SO_DONTROUTE) |
1051 IP_DEBUGROUTE, NULL, inp);
1054 KASSERT(error != 0, ("no error, but th not set\n"));
1059 * We know that the packet was lost, so back out the
1060 * sequence number advance, if any.
1062 if (!(tp->t_flags & TF_FORCE) ||
1063 !tcp_callout_active(tp, tp->tt_persist)) {
1065 * No need to check for TH_FIN here because
1066 * the TF_SENTFIN flag handles that case.
1068 if (!(flags & TH_SYN))
1073 if (error == ENOBUFS) {
1075 * If we can't send, make sure there is something
1076 * to get us going again later.
1078 * The persist timer isn't necessarily allowed in all
1079 * states, use the rexmt timer.
1081 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1082 !tcp_callout_active(tp, tp->tt_persist)) {
1083 tcp_callout_reset(tp, tp->tt_rexmt,
1094 if (error == EMSGSIZE) {
1096 * ip_output() will have already fixed the route
1097 * for us. tcp_mtudisc() will, as its last action,
1098 * initiate retransmission, so it is important to
1101 tcp_mtudisc(inp, 0);
1104 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1105 TCPS_HAVERCVDSYN(tp->t_state)) {
1106 tp->t_softerror = error;
1111 tcpstat.tcps_sndtotal++;
1114 * Data sent (as far as we can tell).
1116 * If this advertises a larger window than any other segment,
1117 * then remember the size of the advertised window.
1119 * Any pending ACK has now been sent.
1121 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1122 tp->rcv_adv = tp->rcv_nxt + recvwin;
1123 tp->t_flags &= ~TF_RXRESIZED;
1125 tp->last_ack_sent = tp->rcv_nxt;
1126 tp->t_flags &= ~TF_ACKNOW;
1127 if (tcp_delack_enabled)
1128 tcp_callout_stop(tp, tp->tt_delack);
1135 tcp_setpersist(struct tcpcb *tp)
1137 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1140 if (tp->t_state == TCPS_SYN_SENT ||
1141 tp->t_state == TCPS_SYN_RECEIVED) {
1142 panic("tcp_setpersist: not established yet, current %s\n",
1143 tp->t_state == TCPS_SYN_SENT ?
1144 "SYN_SENT" : "SYN_RECEIVED");
1147 if (tcp_callout_active(tp, tp->tt_rexmt))
1148 panic("tcp_setpersist: retransmit pending");
1150 * Start/restart persistance timer.
1152 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1154 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1155 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)