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 $
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>
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>
100 #include <netinet/tcp_fsm.h>
101 #include <netinet/tcp_seq.h>
102 #include <netinet/tcp_timer.h>
103 #include <netinet/tcp_timer2.h>
104 #include <netinet/tcp_var.h>
105 #include <netinet/tcpip.h>
107 #include <netinet/tcp_debug.h>
111 #include <netinet6/ipsec.h>
115 #include <netproto/ipsec/ipsec.h>
117 #endif /*FAST_IPSEC*/
120 extern struct mbuf *m_copypack();
123 int path_mtu_discovery = 0;
124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
125 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
127 static int avoid_pure_win_update = 1;
128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
129 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
131 int tcp_do_autosndbuf = 1;
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
133 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
135 int tcp_autosndbuf_inc = 8*1024;
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
137 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
139 int tcp_autosndbuf_max = 2*1024*1024;
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
141 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
143 static int tcp_idle_cwv = 1;
144 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
146 "Congestion window validation after idle period (part of RFC2861)");
148 static int tcp_idle_restart = 1;
149 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
150 &tcp_idle_restart, 0, "Reset congestion window after idle period");
152 static void tcp_idle_cwnd_validate(struct tcpcb *);
155 * Tcp output routine: figure out what should be sent and send it.
158 tcp_output(struct tcpcb *tp)
160 struct inpcb * const inp = tp->t_inpcb;
161 struct socket *so = inp->inp_socket;
162 long len, recvwin, sendwin;
164 int off, flags, error = 0;
172 u_char opt[TCP_MAXOLEN];
173 unsigned int ipoptlen, optlen, hdrlen;
174 int idle, idle_cwv = 0;
178 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
180 const boolean_t isipv6 = FALSE;
183 KKASSERT(so->so_port == &curthread->td_msgport);
186 * Determine length of data that should be transmitted,
187 * and flags that will be used.
188 * If there is some data or critical controls (SYN, RST)
189 * to send, then transmit; otherwise, investigate further.
193 * If we have been idle for a while, the send congestion window
194 * could be no longer representative of the current state of the
195 * link; need to validate congestion window. However, we should
196 * not perform congestion window validation here, since we could
197 * be asked to send pure ACK.
199 if (tp->snd_max == tp->snd_una &&
200 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
204 * Calculate whether the transmit stream was previously idle
205 * and adjust TF_LASTIDLE for the next time.
207 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
208 if (idle && (tp->t_flags & TF_MORETOCOME))
209 tp->t_flags |= TF_LASTIDLE;
211 tp->t_flags &= ~TF_LASTIDLE;
213 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
214 !IN_FASTRECOVERY(tp))
215 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
224 /* Make use of SACK information when slow-starting after a RTO. */
225 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
226 !IN_FASTRECOVERY(tp)) {
227 tcp_seq old_snd_nxt = tp->snd_nxt;
229 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
230 nsacked += tp->snd_nxt - old_snd_nxt;
234 off = tp->snd_nxt - tp->snd_una;
235 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
236 sendwin = min(sendwin, tp->snd_bwnd);
238 flags = tcp_outflags[tp->t_state];
240 * Get standard flags, and add SYN or FIN if requested by 'hidden'
243 if (tp->t_flags & TF_NEEDFIN)
245 if (tp->t_flags & TF_NEEDSYN)
249 * If in persist timeout with window of 0, send 1 byte.
250 * Otherwise, if window is small but nonzero
251 * and timer expired, we will send what we can
252 * and go to transmit state.
254 if (tp->t_flags & TF_FORCE) {
257 * If we still have some data to send, then
258 * clear the FIN bit. Usually this would
259 * happen below when it realizes that we
260 * aren't sending all the data. However,
261 * if we have exactly 1 byte of unsent data,
262 * then it won't clear the FIN bit below,
263 * and if we are in persist state, we wind
264 * up sending the packet without recording
265 * that we sent the FIN bit.
267 * We can't just blindly clear the FIN bit,
268 * because if we don't have any more data
269 * to send then the probe will be the FIN
272 if (off < so->so_snd.ssb_cc)
276 tcp_callout_stop(tp, tp->tt_persist);
282 * If snd_nxt == snd_max and we have transmitted a FIN, the
283 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
284 * a negative length. This can also occur when TCP opens up
285 * its congestion window while receiving additional duplicate
286 * acks after fast-retransmit because TCP will reset snd_nxt
287 * to snd_max after the fast-retransmit.
289 * A negative length can also occur when we are in the
290 * TCPS_SYN_RECEIVED state due to a simultanious connect where
291 * our SYN has not been acked yet.
293 * In the normal retransmit-FIN-only case, however, snd_nxt will
294 * be set to snd_una, the offset will be 0, and the length may
297 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
300 * Lop off SYN bit if it has already been sent. However, if this
301 * is SYN-SENT state and if segment contains data, suppress sending
302 * segment (sending the segment would be an option if we still
303 * did TAO and the remote host supported it).
305 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
308 if (len > 0 && tp->t_state == TCPS_SYN_SENT)
313 * Be careful not to send data and/or FIN on SYN segments.
314 * This measure is needed to prevent interoperability problems
315 * with not fully conformant TCP implementations.
317 if (flags & TH_SYN) {
324 * A negative len can occur if our FIN has been sent but not
325 * acked, or if we are in a simultanious connect in the
326 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
329 * If our window has contracted to 0 in the FIN case
330 * (which can only occur if we have NOT been called to
331 * retransmit as per code a few paragraphs up) then we
332 * want to shift the retransmit timer over to the
335 * However, if we are in the TCPS_SYN_RECEIVED state
336 * (the SYN case) we will be in a simultanious connect and
337 * the window may be zero degeneratively. In this case we
338 * do not want to shift to the persist timer after the SYN
339 * or the SYN+ACK transmission.
342 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
343 tcp_callout_stop(tp, tp->tt_rexmt);
345 tp->snd_nxt = tp->snd_una;
346 if (!tcp_callout_active(tp, tp->tt_persist))
351 KASSERT(len >= 0, ("%s: len < 0", __func__));
353 * Automatic sizing of send socket buffer. Often the send buffer
354 * size is not optimally adjusted to the actual network conditions
355 * at hand (delay bandwidth product). Setting the buffer size too
356 * small limits throughput on links with high bandwidth and high
357 * delay (eg. trans-continental/oceanic links). Setting the
358 * buffer size too big consumes too much real kernel memory,
359 * especially with many connections on busy servers.
361 * The criteria to step up the send buffer one notch are:
362 * 1. receive window of remote host is larger than send buffer
363 * (with a fudge factor of 5/4th);
364 * 2. send buffer is filled to 7/8th with data (so we actually
365 * have data to make use of it);
366 * 3. send buffer fill has not hit maximal automatic size;
367 * 4. our send window (slow start and cogestion controlled) is
368 * larger than sent but unacknowledged data in send buffer.
370 * The remote host receive window scaling factor may limit the
371 * growing of the send buffer before it reaches its allowed
374 * It scales directly with slow start or congestion window
375 * and does at most one step per received ACK. This fast
376 * scaling has the drawback of growing the send buffer beyond
377 * what is strictly necessary to make full use of a given
378 * delay*bandwith product. However testing has shown this not
379 * to be much of an problem. At worst we are trading wasting
380 * of available bandwith (the non-use of it) for wasting some
381 * socket buffer memory.
383 * TODO: Shrink send buffer during idle periods together
384 * with congestion window. Requires another timer. Has to
385 * wait for upcoming tcp timer rewrite.
387 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
388 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
389 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
390 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
391 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
394 newsize = ulmin(so->so_snd.ssb_hiwat +
397 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
398 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
399 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
400 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
405 * Truncate to the maximum segment length and ensure that FIN is
406 * removed if the length no longer contains the last data byte.
408 if (len > tp->t_maxseg) {
412 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
415 recvwin = ssb_space(&so->so_rcv);
418 * Sender silly window avoidance. We transmit under the following
419 * conditions when len is non-zero:
421 * - We have a full segment
422 * - This is the last buffer in a write()/send() and we are
423 * either idle or running NODELAY
424 * - we've timed out (e.g. persist timer)
425 * - we have more then 1/2 the maximum send window's worth of
426 * data (receiver may be limiting the window size)
427 * - we need to retransmit
430 if (len == tp->t_maxseg)
433 * NOTE! on localhost connections an 'ack' from the remote
434 * end may occur synchronously with the output and cause
435 * us to flush a buffer queued with moretocome. XXX
437 * note: the len + off check is almost certainly unnecessary.
439 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
440 (idle || (tp->t_flags & TF_NODELAY)) &&
441 len + off >= so->so_snd.ssb_cc &&
442 !(tp->t_flags & TF_NOPUSH)) {
445 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
447 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
449 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
454 * Compare available window to amount of window
455 * known to peer (as advertised window less
456 * next expected input). If the difference is at least two
457 * max size segments, or at least 50% of the maximum possible
458 * window, then want to send a window update to peer.
462 * "adv" is the amount we can increase the window,
463 * taking into account that we are limited by
464 * TCP_MAXWIN << tp->rcv_scale.
466 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
467 (tp->rcv_adv - tp->rcv_nxt);
471 * This ack case typically occurs when the user has drained
472 * the TCP socket buffer sufficiently to warrent an ack
473 * containing a 'pure window update'... that is, an ack that
474 * ONLY updates the tcp window.
476 * It is unclear why we would need to do a pure window update
477 * past 2 segments if we are going to do one at 1/2 the high
478 * water mark anyway, especially since under normal conditions
479 * the user program will drain the socket buffer quickly.
480 * The 2-segment pure window update will often add a large
481 * number of extra, unnecessary acks to the stream.
483 * avoid_pure_win_update now defaults to 1.
485 if (avoid_pure_win_update == 0 ||
486 (tp->t_flags & TF_RXRESIZED)) {
487 if (adv >= (long) (2 * tp->t_maxseg)) {
491 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
492 if (hiwat > (long)so->so_rcv.ssb_hiwat)
493 hiwat = (long)so->so_rcv.ssb_hiwat;
494 if (adv >= hiwat / 2)
499 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
500 * is also a catch-all for the retransmit timer timeout case.
502 if (tp->t_flags & TF_ACKNOW)
504 if ((flags & TH_RST) ||
505 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
507 if (SEQ_GT(tp->snd_up, tp->snd_una))
510 * If our state indicates that FIN should be sent
511 * and we have not yet done so, then we need to send.
513 if ((flags & TH_FIN) &&
514 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
518 * TCP window updates are not reliable, rather a polling protocol
519 * using ``persist'' packets is used to insure receipt of window
520 * updates. The three ``states'' for the output side are:
521 * idle not doing retransmits or persists
522 * persisting to move a small or zero window
523 * (re)transmitting and thereby not persisting
525 * tcp_callout_active(tp, tp->tt_persist)
526 * is true when we are in persist state.
527 * The TF_FORCE flag in tp->t_flags
528 * is set when we are called to send a persist packet.
529 * tcp_callout_active(tp, tp->tt_rexmt)
530 * is set when we are retransmitting
531 * The output side is idle when both timers are zero.
533 * If send window is too small, there is data to transmit, and no
534 * retransmit or persist is pending, then go to persist state.
536 * If nothing happens soon, send when timer expires:
537 * if window is nonzero, transmit what we can, otherwise force out
540 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
541 * with data pending. This situation can occur during a
542 * simultanious connect.
544 if (so->so_snd.ssb_cc > 0 &&
545 tp->t_state != TCPS_SYN_RECEIVED &&
546 !tcp_callout_active(tp, tp->tt_rexmt) &&
547 !tcp_callout_active(tp, tp->tt_persist)) {
553 * No reason to send a segment, just return.
559 * Before ESTABLISHED, force sending of initial options
560 * unless TCP set not to do any options.
561 * NOTE: we assume that the IP/TCP header plus TCP options
562 * always fit in a single mbuf, leaving room for a maximum
564 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
568 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
570 hdrlen = sizeof(struct tcpiphdr);
571 if (flags & TH_SYN) {
572 tp->snd_nxt = tp->iss;
573 if (!(tp->t_flags & TF_NOOPT)) {
576 opt[0] = TCPOPT_MAXSEG;
577 opt[1] = TCPOLEN_MAXSEG;
578 mss = htons((u_short) tcp_mssopt(tp));
579 memcpy(opt + 2, &mss, sizeof mss);
580 optlen = TCPOLEN_MAXSEG;
582 if ((tp->t_flags & TF_REQ_SCALE) &&
583 (!(flags & TH_ACK) ||
584 (tp->t_flags & TF_RCVD_SCALE))) {
585 *((u_int32_t *)(opt + optlen)) = htonl(
587 TCPOPT_WINDOW << 16 |
588 TCPOLEN_WINDOW << 8 |
589 tp->request_r_scale);
593 if ((tcp_do_sack && !(flags & TH_ACK)) ||
594 tp->t_flags & TF_SACK_PERMITTED) {
595 uint32_t *lp = (uint32_t *)(opt + optlen);
597 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
598 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
604 * Send a timestamp and echo-reply if this is a SYN and our side
605 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
606 * and our peer have sent timestamps in our SYN's.
608 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
610 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
611 u_int32_t *lp = (u_int32_t *)(opt + optlen);
613 /* Form timestamp option as shown in appendix A of RFC 1323. */
614 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
615 *lp++ = htonl(ticks);
616 *lp = htonl(tp->ts_recent);
617 optlen += TCPOLEN_TSTAMP_APPA;
620 /* Set receive buffer autosizing timestamp. */
621 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
622 tp->rfbuf_ts = ticks;
625 * If this is a SACK connection and we have a block to report,
626 * fill in the SACK blocks in the TCP options.
628 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
630 (!LIST_EMPTY(&tp->t_segq) ||
631 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
632 tcp_sack_fill_report(tp, opt, &optlen);
635 if (tp->t_flags & TF_SIGNATURE) {
639 * Initialize TCP-MD5 option (RFC2385)
641 bp = (u_char *)opt + optlen;
642 *bp++ = TCPOPT_SIGNATURE;
643 *bp++ = TCPOLEN_SIGNATURE;
645 for (i = 0; i < TCP_SIGLEN; i++)
647 optlen += TCPOLEN_SIGNATURE;
649 * Terminate options list and maintain 32-bit alignment.
655 #endif /* TCP_SIGNATURE */
656 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
660 ipoptlen = ip6_optlen(inp);
662 if (inp->inp_options) {
663 ipoptlen = inp->inp_options->m_len -
664 offsetof(struct ipoption, ipopt_list);
670 ipoptlen += ipsec_hdrsiz_tcp(tp);
674 * Adjust data length if insertion of options will bump the packet
675 * length beyond the t_maxopd length. Clear FIN to prevent premature
676 * closure since there is still more data to send after this (now
679 * If just the options do not fit we are in a no-win situation and
680 * we treat it as an unreachable host.
682 if (len + optlen + ipoptlen > tp->t_maxopd) {
683 if (tp->t_maxopd <= optlen + ipoptlen) {
684 static time_t last_optlen_report;
686 if (last_optlen_report != time_second) {
687 last_optlen_report = time_second;
688 kprintf("tcpcb %p: MSS (%d) too small to hold options!\n", tp, tp->t_maxopd);
690 error = EHOSTUNREACH;
694 len = tp->t_maxopd - optlen - ipoptlen;
700 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
702 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
706 * Grab a header mbuf, attaching a copy of data to
707 * be transmitted, and initialize the header from
708 * the template for sends on this connection.
711 if ((tp->t_flags & TF_FORCE) && len == 1)
712 tcpstat.tcps_sndprobe++;
713 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
714 if (tp->snd_nxt == tp->snd_una)
715 tp->snd_max_rexmt = tp->snd_max;
717 tcpstat.tcps_sndsackrtopack++;
718 tcpstat.tcps_sndsackrtobyte += len;
720 tcpstat.tcps_sndrexmitpack++;
721 tcpstat.tcps_sndrexmitbyte += len;
723 tcpstat.tcps_sndpack++;
724 tcpstat.tcps_sndbyte += len;
728 tcp_idle_cwnd_validate(tp);
730 /* Update last send time after CWV */
731 tp->snd_last = ticks;
733 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
734 max_linkhdr + hdrlen)) == NULL) {
739 * m_copypack left space for our hdr; use it.
745 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
747 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
754 m->m_data += max_linkhdr;
756 if (len <= MHLEN - hdrlen - max_linkhdr) {
757 m_copydata(so->so_snd.ssb_mb, off, (int) len,
758 mtod(m, caddr_t) + hdrlen);
761 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
762 if (m->m_next == NULL) {
771 * If we're sending everything we've got, set PUSH.
772 * (This will keep happy those implementations which only
773 * give data to the user when a buffer fills or
776 if (off + len == so->so_snd.ssb_cc)
779 if (tp->t_flags & TF_ACKNOW)
780 tcpstat.tcps_sndacks++;
781 else if (flags & (TH_SYN | TH_FIN | TH_RST))
782 tcpstat.tcps_sndctrl++;
783 else if (SEQ_GT(tp->snd_up, tp->snd_una))
784 tcpstat.tcps_sndurg++;
786 tcpstat.tcps_sndwinup++;
788 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
794 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
797 m->m_data += max_linkhdr;
800 m->m_pkthdr.rcvif = NULL;
802 ip6 = mtod(m, struct ip6_hdr *);
803 th = (struct tcphdr *)(ip6 + 1);
804 tcp_fillheaders(tp, ip6, th);
806 ip = mtod(m, struct ip *);
807 ipov = (struct ipovly *)ip;
808 th = (struct tcphdr *)(ip + 1);
809 /* this picks up the pseudo header (w/o the length) */
810 tcp_fillheaders(tp, ip, th);
814 * Fill in fields, remembering maximum advertised
815 * window for use in delaying messages about window sizes.
816 * If resending a FIN, be sure not to use a new sequence number.
818 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
819 tp->snd_nxt == tp->snd_max)
824 * If we are doing retransmissions, then snd_nxt will
825 * not reflect the first unsent octet. For ACK only
826 * packets, we do not want the sequence number of the
827 * retransmitted packet, we want the sequence number
828 * of the next unsent octet. So, if there is no data
829 * (and no SYN or FIN), use snd_max instead of snd_nxt
830 * when filling in ti_seq. But if we are in persist
831 * state, snd_max might reflect one byte beyond the
832 * right edge of the window, so use snd_nxt in that
833 * case, since we know we aren't doing a retransmission.
834 * (retransmit and persist are mutually exclusive...)
836 if (len || (flags & (TH_SYN|TH_FIN)) ||
837 tcp_callout_active(tp, tp->tt_persist))
838 th->th_seq = htonl(tp->snd_nxt);
840 th->th_seq = htonl(tp->snd_max);
841 th->th_ack = htonl(tp->rcv_nxt);
843 bcopy(opt, th + 1, optlen);
844 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
846 th->th_flags = flags;
850 * Calculate receive window. Don't shrink window, but avoid
851 * silly window syndrome by sending a 0 window if the actual
852 * window is less then one segment.
854 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
855 recvwin < (long)tp->t_maxseg)
857 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
858 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
859 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
860 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
863 * Adjust the RXWIN0SENT flag - indicate that we have advertised
864 * a 0 window. This may cause the remote transmitter to stall. This
865 * flag tells soreceive() to disable delayed acknowledgements when
866 * draining the buffer. This can occur if the receiver is attempting
867 * to read more data then can be buffered prior to transmitting on
871 tp->t_flags |= TF_RXWIN0SENT;
873 tp->t_flags &= ~TF_RXWIN0SENT;
876 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
878 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
880 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
881 th->th_flags |= TH_URG;
885 * If no urgent pointer to send, then we pull
886 * the urgent pointer to the left edge of the send window
887 * so that it doesn't drift into the send window on sequence
890 tp->snd_up = tp->snd_una; /* drag it along */
895 if (tp->t_flags & TF_SIGNATURE) {
896 tcpsignature_compute(m, len, optlen,
897 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
899 #endif /* TCP_SIGNATURE */
902 * Put TCP length in extended header, and then
903 * checksum extended header and data.
905 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
908 * ip6_plen is not need to be filled now, and will be
909 * filled in ip6_output().
911 th->th_sum = in6_cksum(m, IPPROTO_TCP,
912 sizeof(struct ip6_hdr),
913 sizeof(struct tcphdr) + optlen + len);
915 m->m_pkthdr.csum_flags = CSUM_TCP;
916 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
918 th->th_sum = in_addword(th->th_sum,
919 htons((u_short)(optlen + len)));
923 * IP version must be set here for ipv4/ipv6 checking
926 KASSERT(ip->ip_v == IPVERSION,
927 ("%s: IP version incorrect: %d",
928 __func__, ip->ip_v));
933 * In transmit state, time the transmission and arrange for
934 * the retransmit. In persist state, just set snd_max.
936 if (!(tp->t_flags & TF_FORCE) ||
937 !tcp_callout_active(tp, tp->tt_persist)) {
938 tcp_seq startseq = tp->snd_nxt;
941 * Advance snd_nxt over sequence space of this segment.
943 if (flags & (TH_SYN | TH_FIN)) {
946 if (flags & TH_FIN) {
948 tp->t_flags |= TF_SENTFIN;
952 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
953 tp->snd_max = tp->snd_nxt;
955 * Time this transmission if not a retransmission and
956 * not currently timing anything.
958 if (tp->t_rtttime == 0) {
959 tp->t_rtttime = ticks;
960 tp->t_rtseq = startseq;
961 tcpstat.tcps_segstimed++;
966 * Set retransmit timer if not currently set,
967 * and not doing a pure ack or a keep-alive probe.
968 * Initial value for retransmit timer is smoothed
969 * round-trip time + 2 * round-trip time variance.
970 * Initialize shift counter which is used for backoff
971 * of retransmit time.
973 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
974 tp->snd_nxt != tp->snd_una) {
975 if (tcp_callout_active(tp, tp->tt_persist)) {
976 tcp_callout_stop(tp, tp->tt_persist);
979 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
984 * Persist case, update snd_max but since we are in
985 * persist mode (no window) we do not update snd_nxt.
989 panic("tcp_output: persist timer to send SYN");
990 if (flags & TH_FIN) {
992 tp->t_flags |= TF_SENTFIN;
994 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
995 tp->snd_max = tp->snd_nxt + xlen;
1001 if (so->so_options & SO_DEBUG) {
1002 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1003 mtod(m, void *), th, 0);
1008 * Fill in IP length and desired time to live and
1009 * send to IP level. There should be a better way
1010 * to handle ttl and tos; we could keep them in
1011 * the template, but need a way to checksum without them.
1014 * m->m_pkthdr.len should have been set before cksum
1015 * calcuration, because in6_cksum() need it.
1019 * we separately set hoplimit for every segment,
1020 * since the user might want to change the value
1021 * via setsockopt. Also, desired default hop
1022 * limit might be changed via Neighbor Discovery.
1024 ip6->ip6_hlim = in6_selecthlim(inp,
1025 (inp->in6p_route.ro_rt ?
1026 inp->in6p_route.ro_rt->rt_ifp : NULL));
1028 /* TODO: IPv6 IP6TOS_ECT bit on */
1029 error = ip6_output(m, inp->in6p_outputopts,
1030 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1034 ip->ip_len = m->m_pkthdr.len;
1036 if (INP_CHECK_SOCKAF(so, AF_INET6))
1037 ip->ip_ttl = in6_selecthlim(inp,
1038 (inp->in6p_route.ro_rt ?
1039 inp->in6p_route.ro_rt->rt_ifp : NULL));
1042 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1044 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1046 * See if we should do MTU discovery.
1047 * We do it only if the following are true:
1048 * 1) we have a valid route to the destination
1049 * 2) the MTU is not locked (if it is,
1050 * then discovery has been disabled)
1052 if (path_mtu_discovery &&
1053 (rt = inp->inp_route.ro_rt) &&
1054 (rt->rt_flags & RTF_UP) &&
1055 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1056 ip->ip_off |= IP_DF;
1058 error = ip_output(m, inp->inp_options, &inp->inp_route,
1059 (so->so_options & SO_DONTROUTE) |
1060 IP_DEBUGROUTE, NULL, inp);
1063 KASSERT(error != 0, ("no error, but th not set"));
1068 * We know that the packet was lost, so back out the
1069 * sequence number advance, if any.
1071 if (!(tp->t_flags & TF_FORCE) ||
1072 !tcp_callout_active(tp, tp->tt_persist)) {
1074 * No need to check for TH_FIN here because
1075 * the TF_SENTFIN flag handles that case.
1077 if (!(flags & TH_SYN))
1082 if (error == ENOBUFS) {
1084 * If we can't send, make sure there is something
1085 * to get us going again later.
1087 * The persist timer isn't necessarily allowed in all
1088 * states, use the rexmt timer.
1090 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1091 !tcp_callout_active(tp, tp->tt_persist)) {
1092 tcp_callout_reset(tp, tp->tt_rexmt,
1103 if (error == EMSGSIZE) {
1105 * ip_output() will have already fixed the route
1106 * for us. tcp_mtudisc() will, as its last action,
1107 * initiate retransmission, so it is important to
1110 tcp_mtudisc(inp, 0);
1113 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1114 TCPS_HAVERCVDSYN(tp->t_state)) {
1115 tp->t_softerror = error;
1120 tcpstat.tcps_sndtotal++;
1123 * Data sent (as far as we can tell).
1125 * If this advertises a larger window than any other segment,
1126 * then remember the size of the advertised window.
1128 * Any pending ACK has now been sent.
1130 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1131 tp->rcv_adv = tp->rcv_nxt + recvwin;
1132 tp->t_flags &= ~TF_RXRESIZED;
1134 tp->last_ack_sent = tp->rcv_nxt;
1135 tp->t_flags &= ~TF_ACKNOW;
1136 if (tcp_delack_enabled)
1137 tcp_callout_stop(tp, tp->tt_delack);
1144 tcp_setpersist(struct tcpcb *tp)
1146 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1149 if (tp->t_state == TCPS_SYN_SENT ||
1150 tp->t_state == TCPS_SYN_RECEIVED) {
1151 panic("tcp_setpersist: not established yet, current %s",
1152 tp->t_state == TCPS_SYN_SENT ?
1153 "SYN_SENT" : "SYN_RECEIVED");
1156 if (tcp_callout_active(tp, tp->tt_rexmt))
1157 panic("tcp_setpersist: retransmit pending");
1159 * Start/restart persistance timer.
1161 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1163 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1164 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1169 tcp_idle_cwnd_validate(struct tcpcb *tp)
1171 u_long initial_cwnd = tcp_initial_window(tp);
1174 tcpstat.tcps_sndidle++;
1176 /* According to RFC5681: RW=min(IW,cwnd) */
1177 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1180 u_long idle_time, decay_cwnd;
1183 * RFC2861, but only after idle period.
1187 * Before the congestion window is reduced, ssthresh
1188 * is set to the maximum of its current value and 3/4
1189 * cwnd. If the sender then has more data to send
1190 * than the decayed cwnd allows, the TCP will slow-
1191 * start (perform exponential increase) at least
1192 * half-way back up to the old value of cwnd.
1194 tp->snd_ssthresh = max(tp->snd_ssthresh,
1195 (3 * tp->snd_cwnd) / 4);
1198 * Decay the congestion window by half for every RTT
1199 * that the flow remains inactive.
1201 * The difference between our implementation and
1202 * RFC2861 is that we don't allow cwnd to go below
1203 * the value allowed by RFC5681 (min_cwnd).
1205 idle_time = ticks - tp->snd_last;
1206 decay_cwnd = tp->snd_cwnd;
1207 while (idle_time >= tp->t_rxtcur &&
1208 decay_cwnd > min_cwnd) {
1210 idle_time -= tp->t_rxtcur;
1212 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1215 * Slow-start from scratch to re-determine the send
1216 * congestion window.
1218 tp->snd_cwnd = min_cwnd;
1221 /* Restart ABC counting during congestion avoidance */