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/if_var.h>
89 #include <net/route.h>
90 #include <net/netmsg2.h>
92 #include <netinet/in.h>
93 #include <netinet/in_systm.h>
94 #include <netinet/ip.h>
95 #include <netinet/in_pcb.h>
96 #include <netinet/ip_var.h>
97 #include <netinet6/in6_pcb.h>
98 #include <netinet/ip6.h>
99 #include <netinet6/ip6_var.h>
100 #include <netinet/tcp.h>
102 #include <netinet/tcp_fsm.h>
103 #include <netinet/tcp_seq.h>
104 #include <netinet/tcp_timer.h>
105 #include <netinet/tcp_timer2.h>
106 #include <netinet/tcp_var.h>
107 #include <netinet/tcpip.h>
109 #include <netinet/tcp_debug.h>
113 #include <netinet6/ipsec.h>
117 #include <netproto/ipsec/ipsec.h>
119 #endif /*FAST_IPSEC*/
122 extern struct mbuf *m_copypack();
125 int path_mtu_discovery = 0;
126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
127 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
129 static int avoid_pure_win_update = 1;
130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
131 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
133 int tcp_do_autosndbuf = 1;
134 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
135 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
137 int tcp_autosndbuf_inc = 8*1024;
138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
139 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
141 int tcp_autosndbuf_max = 2*1024*1024;
142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
143 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
145 static int tcp_idle_cwv = 1;
146 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
148 "Congestion window validation after idle period (part of RFC2861)");
150 static int tcp_idle_restart = 1;
151 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
152 &tcp_idle_restart, 0, "Reset congestion window after idle period");
154 static int tcp_do_tso = 1;
155 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
156 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
158 static int tcp_fairsend = 4;
159 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW,
161 "Amount of segments sent before yield to other senders or receivers");
163 static void tcp_idle_cwnd_validate(struct tcpcb *);
165 static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
166 static void tcp_output_sched(struct tcpcb *tp);
169 * Tcp output routine: figure out what should be sent and send it.
172 tcp_output(struct tcpcb *tp)
174 struct inpcb * const inp = tp->t_inpcb;
175 struct socket *so = inp->inp_socket;
176 long len, recvwin, sendwin;
178 int off, flags, error = 0;
185 u_char opt[TCP_MAXOLEN];
186 unsigned int ipoptlen, optlen, hdrlen;
191 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
193 const boolean_t isipv6 = FALSE;
195 boolean_t can_tso = FALSE, use_tso;
196 boolean_t report_sack, idle_cwv = FALSE;
197 u_int segsz, tso_hlen, tso_lenmax = 0;
199 boolean_t need_sched = FALSE;
201 KKASSERT(so->so_port == &curthread->td_msgport);
204 * Determine length of data that should be transmitted,
205 * and flags that will be used.
206 * If there is some data or critical controls (SYN, RST)
207 * to send, then transmit; otherwise, investigate further.
211 * If we have been idle for a while, the send congestion window
212 * could be no longer representative of the current state of the
213 * link; need to validate congestion window. However, we should
214 * not perform congestion window validation here, since we could
215 * be asked to send pure ACK.
217 if (tp->snd_max == tp->snd_una &&
218 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
222 * Calculate whether the transmit stream was previously idle
223 * and adjust TF_LASTIDLE for the next time.
225 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
226 if (idle && (tp->t_flags & TF_MORETOCOME))
227 tp->t_flags |= TF_LASTIDLE;
229 tp->t_flags &= ~TF_LASTIDLE;
231 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
232 !IN_FASTRECOVERY(tp))
233 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
236 * Find out whether TSO could be used or not
238 * For TSO capable devices, the following assumptions apply to
239 * the processing of TCP flags:
240 * - If FIN is set on the large TCP segment, the device must set
241 * FIN on the last segment that it creates from the large TCP
243 * - If PUSH is set on the large TCP segment, the device must set
244 * PUSH on the last segment that it creates from the large TCP
247 #if !defined(IPSEC) && !defined(FAST_IPSEC)
250 && (tp->t_flags & TF_SIGNATURE) == 0
254 struct rtentry *rt = inp->inp_route.ro_rt;
256 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
257 (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
259 tso_lenmax = rt->rt_ifp->if_tsolen;
263 #endif /* !IPSEC && !FAST_IPSEC */
271 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
273 (!TAILQ_EMPTY(&tp->t_segq) ||
274 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
279 /* Make use of SACK information when slow-starting after a RTO. */
280 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
281 !IN_FASTRECOVERY(tp)) {
282 tcp_seq old_snd_nxt = tp->snd_nxt;
284 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
285 nsacked += tp->snd_nxt - old_snd_nxt;
289 off = tp->snd_nxt - tp->snd_una;
290 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
291 sendwin = min(sendwin, tp->snd_bwnd);
293 flags = tcp_outflags[tp->t_state];
295 * Get standard flags, and add SYN or FIN if requested by 'hidden'
298 if (tp->t_flags & TF_NEEDFIN)
300 if (tp->t_flags & TF_NEEDSYN)
304 * If in persist timeout with window of 0, send 1 byte.
305 * Otherwise, if window is small but nonzero
306 * and timer expired, we will send what we can
307 * and go to transmit state.
309 if (tp->t_flags & TF_FORCE) {
312 * If we still have some data to send, then
313 * clear the FIN bit. Usually this would
314 * happen below when it realizes that we
315 * aren't sending all the data. However,
316 * if we have exactly 1 byte of unsent data,
317 * then it won't clear the FIN bit below,
318 * and if we are in persist state, we wind
319 * up sending the packet without recording
320 * that we sent the FIN bit.
322 * We can't just blindly clear the FIN bit,
323 * because if we don't have any more data
324 * to send then the probe will be the FIN
327 if (off < so->so_snd.ssb_cc)
331 tcp_callout_stop(tp, tp->tt_persist);
337 * If snd_nxt == snd_max and we have transmitted a FIN, the
338 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
339 * a negative length. This can also occur when TCP opens up
340 * its congestion window while receiving additional duplicate
341 * acks after fast-retransmit because TCP will reset snd_nxt
342 * to snd_max after the fast-retransmit.
344 * A negative length can also occur when we are in the
345 * TCPS_SYN_RECEIVED state due to a simultanious connect where
346 * our SYN has not been acked yet.
348 * In the normal retransmit-FIN-only case, however, snd_nxt will
349 * be set to snd_una, the offset will be 0, and the length may
352 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
355 * Lop off SYN bit if it has already been sent. However, if this
356 * is SYN-SENT state and if segment contains data, suppress sending
357 * segment (sending the segment would be an option if we still
358 * did TAO and the remote host supported it).
360 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
363 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
364 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
370 * Be careful not to send data and/or FIN on SYN segments.
371 * This measure is needed to prevent interoperability problems
372 * with not fully conformant TCP implementations.
374 if (flags & TH_SYN) {
381 * A negative len can occur if our FIN has been sent but not
382 * acked, or if we are in a simultanious connect in the
383 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
386 * If our window has contracted to 0 in the FIN case
387 * (which can only occur if we have NOT been called to
388 * retransmit as per code a few paragraphs up) then we
389 * want to shift the retransmit timer over to the
392 * However, if we are in the TCPS_SYN_RECEIVED state
393 * (the SYN case) we will be in a simultanious connect and
394 * the window may be zero degeneratively. In this case we
395 * do not want to shift to the persist timer after the SYN
396 * or the SYN+ACK transmission.
399 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
400 tcp_callout_stop(tp, tp->tt_rexmt);
402 tp->snd_nxt = tp->snd_una;
403 if (!tcp_callout_active(tp, tp->tt_persist))
408 KASSERT(len >= 0, ("%s: len < 0", __func__));
410 * Automatic sizing of send socket buffer. Often the send buffer
411 * size is not optimally adjusted to the actual network conditions
412 * at hand (delay bandwidth product). Setting the buffer size too
413 * small limits throughput on links with high bandwidth and high
414 * delay (eg. trans-continental/oceanic links). Setting the
415 * buffer size too big consumes too much real kernel memory,
416 * especially with many connections on busy servers.
418 * The criteria to step up the send buffer one notch are:
419 * 1. receive window of remote host is larger than send buffer
420 * (with a fudge factor of 5/4th);
421 * 2. send buffer is filled to 7/8th with data (so we actually
422 * have data to make use of it);
423 * 3. send buffer fill has not hit maximal automatic size;
424 * 4. our send window (slow start and cogestion controlled) is
425 * larger than sent but unacknowledged data in send buffer.
427 * The remote host receive window scaling factor may limit the
428 * growing of the send buffer before it reaches its allowed
431 * It scales directly with slow start or congestion window
432 * and does at most one step per received ACK. This fast
433 * scaling has the drawback of growing the send buffer beyond
434 * what is strictly necessary to make full use of a given
435 * delay*bandwith product. However testing has shown this not
436 * to be much of an problem. At worst we are trading wasting
437 * of available bandwith (the non-use of it) for wasting some
438 * socket buffer memory.
440 * TODO: Shrink send buffer during idle periods together
441 * with congestion window. Requires another timer. Has to
442 * wait for upcoming tcp timer rewrite.
444 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
445 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
446 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
447 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
448 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
451 newsize = ulmin(so->so_snd.ssb_hiwat +
454 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
455 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
456 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
457 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
463 * - Congestion window needs validation
464 * - There are SACK blocks to report
465 * - RST or SYN flags is set
469 * Checking for SYN|RST looks overkill, just to be safe than sorry
472 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
475 tcp_seq ugr_nxt = tp->snd_nxt;
477 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
478 tp->snd_nxt == tp->snd_max)
481 if (SEQ_GT(tp->snd_up, ugr_nxt))
487 * Find out segment size and header length for TSO
489 error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
494 segsz = tp->t_maxseg;
495 tso_hlen = 0; /* not used */
499 * Truncate to the maximum segment length if not TSO, and ensure that
500 * FIN is removed if the length no longer contains the last data byte.
509 if (__predict_false(tso_lenmax < segsz))
510 tso_lenmax = segsz << 1;
513 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
514 * thoff), and make sure that we send equal size
515 * transfers down the stack (rather than big-small-
518 len = min(len, tso_lenmax);
519 nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz;
537 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
540 recvwin = ssb_space(&so->so_rcv);
543 * Sender silly window avoidance. We transmit under the following
544 * conditions when len is non-zero:
546 * - We have a full segment
547 * - This is the last buffer in a write()/send() and we are
548 * either idle or running NODELAY
549 * - we've timed out (e.g. persist timer)
550 * - we have more then 1/2 the maximum send window's worth of
551 * data (receiver may be limiting the window size)
552 * - we need to retransmit
558 * NOTE! on localhost connections an 'ack' from the remote
559 * end may occur synchronously with the output and cause
560 * us to flush a buffer queued with moretocome. XXX
562 * note: the len + off check is almost certainly unnecessary.
564 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
565 (idle || (tp->t_flags & TF_NODELAY)) &&
566 len + off >= so->so_snd.ssb_cc &&
567 !(tp->t_flags & TF_NOPUSH)) {
570 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
572 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
574 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
576 if (tp->t_flags & TF_XMITNOW)
581 * Compare available window to amount of window
582 * known to peer (as advertised window less
583 * next expected input). If the difference is at least two
584 * max size segments, or at least 50% of the maximum possible
585 * window, then want to send a window update to peer.
589 * "adv" is the amount we can increase the window,
590 * taking into account that we are limited by
591 * TCP_MAXWIN << tp->rcv_scale.
593 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
594 (tp->rcv_adv - tp->rcv_nxt);
598 * This ack case typically occurs when the user has drained
599 * the TCP socket buffer sufficiently to warrent an ack
600 * containing a 'pure window update'... that is, an ack that
601 * ONLY updates the tcp window.
603 * It is unclear why we would need to do a pure window update
604 * past 2 segments if we are going to do one at 1/2 the high
605 * water mark anyway, especially since under normal conditions
606 * the user program will drain the socket buffer quickly.
607 * The 2-segment pure window update will often add a large
608 * number of extra, unnecessary acks to the stream.
610 * avoid_pure_win_update now defaults to 1.
612 if (avoid_pure_win_update == 0 ||
613 (tp->t_flags & TF_RXRESIZED)) {
614 if (adv >= (long) (2 * segsz)) {
618 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
619 if (hiwat > (long)so->so_rcv.ssb_hiwat)
620 hiwat = (long)so->so_rcv.ssb_hiwat;
621 if (adv >= hiwat / 2)
626 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
627 * is also a catch-all for the retransmit timer timeout case.
629 if (tp->t_flags & TF_ACKNOW)
631 if ((flags & TH_RST) ||
632 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
634 if (SEQ_GT(tp->snd_up, tp->snd_una))
637 * If our state indicates that FIN should be sent
638 * and we have not yet done so, then we need to send.
640 if ((flags & TH_FIN) &&
641 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
645 * TCP window updates are not reliable, rather a polling protocol
646 * using ``persist'' packets is used to insure receipt of window
647 * updates. The three ``states'' for the output side are:
648 * idle not doing retransmits or persists
649 * persisting to move a small or zero window
650 * (re)transmitting and thereby not persisting
652 * tcp_callout_active(tp, tp->tt_persist)
653 * is true when we are in persist state.
654 * The TF_FORCE flag in tp->t_flags
655 * is set when we are called to send a persist packet.
656 * tcp_callout_active(tp, tp->tt_rexmt)
657 * is set when we are retransmitting
658 * The output side is idle when both timers are zero.
660 * If send window is too small, there is data to transmit, and no
661 * retransmit or persist is pending, then go to persist state.
663 * If nothing happens soon, send when timer expires:
664 * if window is nonzero, transmit what we can, otherwise force out
667 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
668 * with data pending. This situation can occur during a
669 * simultanious connect.
671 if (so->so_snd.ssb_cc > 0 &&
672 tp->t_state != TCPS_SYN_RECEIVED &&
673 !tcp_callout_active(tp, tp->tt_rexmt) &&
674 !tcp_callout_active(tp, tp->tt_persist)) {
680 * No reason to send a segment, just return.
682 tp->t_flags &= ~TF_XMITNOW;
686 if (need_sched && len > 0) {
687 tcp_output_sched(tp);
692 * Before ESTABLISHED, force sending of initial options
693 * unless TCP set not to do any options.
694 * NOTE: we assume that the IP/TCP header plus TCP options
695 * always fit in a single mbuf, leaving room for a maximum
697 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
701 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
703 hdrlen = sizeof(struct tcpiphdr);
704 if (flags & TH_SYN) {
705 tp->snd_nxt = tp->iss;
706 if (!(tp->t_flags & TF_NOOPT)) {
709 opt[0] = TCPOPT_MAXSEG;
710 opt[1] = TCPOLEN_MAXSEG;
711 mss = htons((u_short) tcp_mssopt(tp));
712 memcpy(opt + 2, &mss, sizeof mss);
713 optlen = TCPOLEN_MAXSEG;
715 if ((tp->t_flags & TF_REQ_SCALE) &&
716 (!(flags & TH_ACK) ||
717 (tp->t_flags & TF_RCVD_SCALE))) {
718 *((u_int32_t *)(opt + optlen)) = htonl(
720 TCPOPT_WINDOW << 16 |
721 TCPOLEN_WINDOW << 8 |
722 tp->request_r_scale);
726 if ((tcp_do_sack && !(flags & TH_ACK)) ||
727 tp->t_flags & TF_SACK_PERMITTED) {
728 uint32_t *lp = (uint32_t *)(opt + optlen);
730 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
731 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
737 * Send a timestamp and echo-reply if this is a SYN and our side
738 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
739 * and our peer have sent timestamps in our SYN's.
741 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
743 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
744 u_int32_t *lp = (u_int32_t *)(opt + optlen);
746 /* Form timestamp option as shown in appendix A of RFC 1323. */
747 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
748 *lp++ = htonl(ticks);
749 *lp = htonl(tp->ts_recent);
750 optlen += TCPOLEN_TSTAMP_APPA;
753 /* Set receive buffer autosizing timestamp. */
754 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
755 tp->rfbuf_ts = ticks;
758 * If this is a SACK connection and we have a block to report,
759 * fill in the SACK blocks in the TCP options.
762 tcp_sack_fill_report(tp, opt, &optlen);
765 if (tp->t_flags & TF_SIGNATURE) {
769 * Initialize TCP-MD5 option (RFC2385)
771 bp = (u_char *)opt + optlen;
772 *bp++ = TCPOPT_SIGNATURE;
773 *bp++ = TCPOLEN_SIGNATURE;
775 for (i = 0; i < TCP_SIGLEN; i++)
777 optlen += TCPOLEN_SIGNATURE;
779 * Terminate options list and maintain 32-bit alignment.
785 #endif /* TCP_SIGNATURE */
786 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
790 ipoptlen = ip6_optlen(inp);
792 if (inp->inp_options) {
793 ipoptlen = inp->inp_options->m_len -
794 offsetof(struct ipoption, ipopt_list);
800 ipoptlen += ipsec_hdrsiz_tcp(tp);
804 /* TSO segment length must be multiple of segment size */
805 KASSERT(len >= (2 * segsz) && (len % segsz == 0),
806 ("invalid TSO len %ld, segsz %u", len, segsz));
808 KASSERT(len <= segsz,
809 ("invalid len %ld, segsz %u", len, segsz));
812 * Adjust data length if insertion of options will bump
813 * the packet length beyond the t_maxopd length. Clear
814 * FIN to prevent premature closure since there is still
815 * more data to send after this (now truncated) packet.
817 * If just the options do not fit we are in a no-win
818 * situation and we treat it as an unreachable host.
820 if (len + optlen + ipoptlen > tp->t_maxopd) {
821 if (tp->t_maxopd <= optlen + ipoptlen) {
822 static time_t last_optlen_report;
824 if (last_optlen_report != time_second) {
825 last_optlen_report = time_second;
826 kprintf("tcpcb %p: MSS (%d) too "
827 "small to hold options!\n",
830 error = EHOSTUNREACH;
834 len = tp->t_maxopd - optlen - ipoptlen;
841 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
843 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
847 * Grab a header mbuf, attaching a copy of data to
848 * be transmitted, and initialize the header from
849 * the template for sends on this connection.
852 if ((tp->t_flags & TF_FORCE) && len == 1)
853 tcpstat.tcps_sndprobe++;
854 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
855 if (tp->snd_nxt == tp->snd_una)
856 tp->snd_max_rexmt = tp->snd_max;
858 tcpstat.tcps_sndsackrtopack++;
859 tcpstat.tcps_sndsackrtobyte += len;
861 tcpstat.tcps_sndrexmitpack++;
862 tcpstat.tcps_sndrexmitbyte += len;
864 tcpstat.tcps_sndpack++;
865 tcpstat.tcps_sndbyte += len;
869 tcp_idle_cwnd_validate(tp);
871 /* Update last send time after CWV */
872 tp->snd_last = ticks;
874 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
875 max_linkhdr + hdrlen)) == NULL) {
880 * m_copypack left space for our hdr; use it.
886 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
888 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
895 m->m_data += max_linkhdr;
897 if (len <= MHLEN - hdrlen - max_linkhdr) {
898 m_copydata(so->so_snd.ssb_mb, off, (int) len,
899 mtod(m, caddr_t) + hdrlen);
902 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
903 if (m->m_next == NULL) {
912 * If we're sending everything we've got, set PUSH.
913 * (This will keep happy those implementations which only
914 * give data to the user when a buffer fills or
917 if (off + len == so->so_snd.ssb_cc)
920 if (tp->t_flags & TF_ACKNOW)
921 tcpstat.tcps_sndacks++;
922 else if (flags & (TH_SYN | TH_FIN | TH_RST))
923 tcpstat.tcps_sndctrl++;
924 else if (SEQ_GT(tp->snd_up, tp->snd_una))
925 tcpstat.tcps_sndurg++;
927 tcpstat.tcps_sndwinup++;
929 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
935 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
938 m->m_data += max_linkhdr;
941 m->m_pkthdr.rcvif = NULL;
943 ip6 = mtod(m, struct ip6_hdr *);
944 th = (struct tcphdr *)(ip6 + 1);
945 tcp_fillheaders(tp, ip6, th, use_tso);
947 ip = mtod(m, struct ip *);
948 th = (struct tcphdr *)(ip + 1);
949 /* this picks up the pseudo header (w/o the length) */
950 tcp_fillheaders(tp, ip, th, use_tso);
954 * Fill in fields, remembering maximum advertised
955 * window for use in delaying messages about window sizes.
956 * If resending a FIN, be sure not to use a new sequence number.
958 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
959 tp->snd_nxt == tp->snd_max)
964 * If we are doing retransmissions, then snd_nxt will
965 * not reflect the first unsent octet. For ACK only
966 * packets, we do not want the sequence number of the
967 * retransmitted packet, we want the sequence number
968 * of the next unsent octet. So, if there is no data
969 * (and no SYN or FIN), use snd_max instead of snd_nxt
970 * when filling in ti_seq. But if we are in persist
971 * state, snd_max might reflect one byte beyond the
972 * right edge of the window, so use snd_nxt in that
973 * case, since we know we aren't doing a retransmission.
974 * (retransmit and persist are mutually exclusive...)
976 if (len || (flags & (TH_SYN|TH_FIN)) ||
977 tcp_callout_active(tp, tp->tt_persist))
978 th->th_seq = htonl(tp->snd_nxt);
980 th->th_seq = htonl(tp->snd_max);
981 th->th_ack = htonl(tp->rcv_nxt);
983 bcopy(opt, th + 1, optlen);
984 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
986 th->th_flags = flags;
990 * Calculate receive window. Don't shrink window, but avoid
991 * silly window syndrome by sending a 0 window if the actual
992 * window is less then one segment.
994 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
995 recvwin < (long)segsz)
997 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
998 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
999 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
1000 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
1003 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1004 * a 0 window. This may cause the remote transmitter to stall. This
1005 * flag tells soreceive() to disable delayed acknowledgements when
1006 * draining the buffer. This can occur if the receiver is attempting
1007 * to read more data then can be buffered prior to transmitting on
1011 tp->t_flags |= TF_RXWIN0SENT;
1013 tp->t_flags &= ~TF_RXWIN0SENT;
1016 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
1018 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1019 KASSERT(!use_tso, ("URG with TSO"));
1021 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1022 th->th_flags |= TH_URG;
1026 * If no urgent pointer to send, then we pull
1027 * the urgent pointer to the left edge of the send window
1028 * so that it doesn't drift into the send window on sequence
1029 * number wraparound.
1031 tp->snd_up = tp->snd_una; /* drag it along */
1035 #ifdef TCP_SIGNATURE
1036 if (tp->t_flags & TF_SIGNATURE) {
1037 tcpsignature_compute(m, len, optlen,
1038 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1040 #endif /* TCP_SIGNATURE */
1043 * Put TCP length in extended header, and then
1044 * checksum extended header and data.
1046 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1049 * ip6_plen is not need to be filled now, and will be
1050 * filled in ip6_output().
1052 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1053 sizeof(struct ip6_hdr),
1054 sizeof(struct tcphdr) + optlen + len);
1056 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1058 m->m_pkthdr.csum_flags = CSUM_TSO;
1059 m->m_pkthdr.tso_segsz = segsz;
1061 m->m_pkthdr.csum_flags = CSUM_TCP;
1062 m->m_pkthdr.csum_data =
1063 offsetof(struct tcphdr, th_sum);
1065 th->th_sum = in_addword(th->th_sum,
1066 htons((u_short)(optlen + len)));
1071 * IP version must be set here for ipv4/ipv6 checking
1074 KASSERT(ip->ip_v == IPVERSION,
1075 ("%s: IP version incorrect: %d",
1076 __func__, ip->ip_v));
1081 * In transmit state, time the transmission and arrange for
1082 * the retransmit. In persist state, just set snd_max.
1084 if (!(tp->t_flags & TF_FORCE) ||
1085 !tcp_callout_active(tp, tp->tt_persist)) {
1086 tcp_seq startseq = tp->snd_nxt;
1089 * Advance snd_nxt over sequence space of this segment.
1091 if (flags & (TH_SYN | TH_FIN)) {
1094 if (flags & TH_FIN) {
1096 tp->t_flags |= TF_SENTFIN;
1100 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1101 tp->snd_max = tp->snd_nxt;
1103 * Time this transmission if not a retransmission and
1104 * not currently timing anything.
1106 if (tp->t_rtttime == 0) {
1107 tp->t_rtttime = ticks;
1108 tp->t_rtseq = startseq;
1109 tcpstat.tcps_segstimed++;
1114 * Set retransmit timer if not currently set,
1115 * and not doing a pure ack or a keep-alive probe.
1116 * Initial value for retransmit timer is smoothed
1117 * round-trip time + 2 * round-trip time variance.
1118 * Initialize shift counter which is used for backoff
1119 * of retransmit time.
1121 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1122 tp->snd_nxt != tp->snd_una) {
1123 if (tcp_callout_active(tp, tp->tt_persist)) {
1124 tcp_callout_stop(tp, tp->tt_persist);
1127 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1132 * Persist case, update snd_max but since we are in
1133 * persist mode (no window) we do not update snd_nxt.
1137 panic("tcp_output: persist timer to send SYN");
1138 if (flags & TH_FIN) {
1140 tp->t_flags |= TF_SENTFIN;
1142 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1143 tp->snd_max = tp->snd_nxt + xlen;
1149 if (so->so_options & SO_DEBUG) {
1150 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1151 mtod(m, void *), th, 0);
1156 * Fill in IP length and desired time to live and
1157 * send to IP level. There should be a better way
1158 * to handle ttl and tos; we could keep them in
1159 * the template, but need a way to checksum without them.
1162 * m->m_pkthdr.len should have been set before cksum
1163 * calcuration, because in6_cksum() need it.
1167 * we separately set hoplimit for every segment,
1168 * since the user might want to change the value
1169 * via setsockopt. Also, desired default hop
1170 * limit might be changed via Neighbor Discovery.
1172 ip6->ip6_hlim = in6_selecthlim(inp,
1173 (inp->in6p_route.ro_rt ?
1174 inp->in6p_route.ro_rt->rt_ifp : NULL));
1176 /* TODO: IPv6 IP6TOS_ECT bit on */
1177 error = ip6_output(m, inp->in6p_outputopts,
1178 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1182 ip->ip_len = m->m_pkthdr.len;
1184 if (INP_CHECK_SOCKAF(so, AF_INET6))
1185 ip->ip_ttl = in6_selecthlim(inp,
1186 (inp->in6p_route.ro_rt ?
1187 inp->in6p_route.ro_rt->rt_ifp : NULL));
1190 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1192 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1194 * See if we should do MTU discovery.
1195 * We do it only if the following are true:
1196 * 1) we have a valid route to the destination
1197 * 2) the MTU is not locked (if it is,
1198 * then discovery has been disabled)
1200 if (path_mtu_discovery &&
1201 (rt = inp->inp_route.ro_rt) &&
1202 (rt->rt_flags & RTF_UP) &&
1203 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1204 ip->ip_off |= IP_DF;
1206 error = ip_output(m, inp->inp_options, &inp->inp_route,
1207 (so->so_options & SO_DONTROUTE) |
1208 IP_DEBUGROUTE, NULL, inp);
1211 KASSERT(error != 0, ("no error, but th not set"));
1214 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1217 * We know that the packet was lost, so back out the
1218 * sequence number advance, if any.
1220 if (!(tp->t_flags & TF_FORCE) ||
1221 !tcp_callout_active(tp, tp->tt_persist)) {
1223 * No need to check for TH_FIN here because
1224 * the TF_SENTFIN flag handles that case.
1226 if (!(flags & TH_SYN))
1231 if (error == ENOBUFS) {
1233 * If we can't send, make sure there is something
1234 * to get us going again later.
1236 * The persist timer isn't necessarily allowed in all
1237 * states, use the rexmt timer.
1239 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1240 !tcp_callout_active(tp, tp->tt_persist)) {
1241 tcp_callout_reset(tp, tp->tt_rexmt,
1252 if (error == EMSGSIZE) {
1254 * ip_output() will have already fixed the route
1255 * for us. tcp_mtudisc() will, as its last action,
1256 * initiate retransmission, so it is important to
1259 tcp_mtudisc(inp, 0);
1262 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1263 TCPS_HAVERCVDSYN(tp->t_state)) {
1264 tp->t_softerror = error;
1269 tcpstat.tcps_sndtotal++;
1272 * Data sent (as far as we can tell).
1274 * If this advertises a larger window than any other segment,
1275 * then remember the size of the advertised window.
1277 * Any pending ACK has now been sent.
1279 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1280 tp->rcv_adv = tp->rcv_nxt + recvwin;
1281 tp->t_flags &= ~TF_RXRESIZED;
1283 tp->last_ack_sent = tp->rcv_nxt;
1284 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1285 if (tcp_delack_enabled)
1286 tcp_callout_stop(tp, tp->tt_delack);
1288 if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) &&
1289 segcnt >= tcp_fairsend)
1297 tcp_setpersist(struct tcpcb *tp)
1299 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1302 if (tp->t_state == TCPS_SYN_SENT ||
1303 tp->t_state == TCPS_SYN_RECEIVED) {
1304 panic("tcp_setpersist: not established yet, current %s",
1305 tp->t_state == TCPS_SYN_SENT ?
1306 "SYN_SENT" : "SYN_RECEIVED");
1309 if (tcp_callout_active(tp, tp->tt_rexmt))
1310 panic("tcp_setpersist: retransmit pending");
1312 * Start/restart persistance timer.
1314 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1316 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1317 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1322 tcp_idle_cwnd_validate(struct tcpcb *tp)
1324 u_long initial_cwnd = tcp_initial_window(tp);
1327 tcpstat.tcps_sndidle++;
1329 /* According to RFC5681: RW=min(IW,cwnd) */
1330 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1333 u_long idle_time, decay_cwnd;
1336 * RFC2861, but only after idle period.
1340 * Before the congestion window is reduced, ssthresh
1341 * is set to the maximum of its current value and 3/4
1342 * cwnd. If the sender then has more data to send
1343 * than the decayed cwnd allows, the TCP will slow-
1344 * start (perform exponential increase) at least
1345 * half-way back up to the old value of cwnd.
1347 tp->snd_ssthresh = max(tp->snd_ssthresh,
1348 (3 * tp->snd_cwnd) / 4);
1351 * Decay the congestion window by half for every RTT
1352 * that the flow remains inactive.
1354 * The difference between our implementation and
1355 * RFC2861 is that we don't allow cwnd to go below
1356 * the value allowed by RFC5681 (min_cwnd).
1358 idle_time = ticks - tp->snd_last;
1359 decay_cwnd = tp->snd_cwnd;
1360 while (idle_time >= tp->t_rxtcur &&
1361 decay_cwnd > min_cwnd) {
1363 idle_time -= tp->t_rxtcur;
1365 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1368 * Slow-start from scratch to re-determine the send
1369 * congestion window.
1371 tp->snd_cwnd = min_cwnd;
1374 /* Restart ABC counting during congestion avoidance */
1379 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1381 struct inpcb * const inp = tp->t_inpcb;
1383 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1385 const boolean_t isipv6 = FALSE;
1387 unsigned int ipoptlen, optlen;
1390 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1393 ipoptlen = ip6_optlen(inp);
1395 if (inp->inp_options) {
1396 ipoptlen = inp->inp_options->m_len -
1397 offsetof(struct ipoption, ipopt_list);
1403 ipoptlen += ipsec_hdrsiz_tcp(tp);
1408 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1409 (tp->t_flags & TF_RCVD_TSTMP))
1410 optlen += TCPOLEN_TSTAMP_APPA;
1413 if (tp->t_maxopd <= optlen + ipoptlen)
1414 return EHOSTUNREACH;
1416 *segsz = tp->t_maxopd - optlen - ipoptlen;
1422 tcp_output_sched_handler(netmsg_t nmsg)
1424 struct tcpcb *tp = nmsg->lmsg.u.ms_resultp;
1428 lwkt_replymsg(&nmsg->lmsg, 0);
1431 tcp_output_fair(tp);
1435 tcp_output_init(struct tcpcb *tp)
1437 netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE,
1438 tcp_output_sched_handler);
1439 tp->tt_sndmore->lmsg.u.ms_resultp = tp;
1443 tcp_output_cancel(struct tcpcb *tp)
1446 * This message is still pending to be processed;
1447 * drop it. Optimized.
1450 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) {
1451 lwkt_dropmsg(&tp->tt_sndmore->lmsg);
1457 tcp_output_pending(struct tcpcb *tp)
1459 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0)
1466 tcp_output_sched(struct tcpcb *tp)
1469 if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE)
1470 lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg);
1477 * Yield to other senders or receivers on the same netisr if the current
1478 * TCP stream has sent tcp_fairsend segments and is going to burst more
1479 * segments. Bursting large amount of segements in a single TCP stream
1480 * could delay other senders' segments and receivers' ACKs quite a lot,
1481 * if others segments and ACKs are queued on to the same hardware transmit
1482 * queue; thus cause unfairness between senders and suppress receiving
1485 * Fairsend should be performed at the places that do not affect segment
1486 * sending during congestion control, e.g.
1487 * - User requested output
1488 * - ACK input triggered output
1491 * For devices that are TSO capable, their TSO aggregation size limit could
1495 tcp_output_fair(struct tcpcb *tp)
1499 tp->t_flags |= TF_FAIRSEND;
1500 ret = tcp_output(tp);
1501 tp->t_flags &= ~TF_FAIRSEND;