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>
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");
144 static int tcp_idle_cwv = 1;
145 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
147 "Congestion window validation after idle period (part of RFC2861)");
149 static int tcp_idle_restart = 1;
150 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
151 &tcp_idle_restart, 0, "Reset congestion window after idle period");
153 static int tcp_do_tso = 1;
154 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
155 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
157 static void tcp_idle_cwnd_validate(struct tcpcb *);
159 static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
162 * Tcp output routine: figure out what should be sent and send it.
165 tcp_output(struct tcpcb *tp)
167 struct inpcb * const inp = tp->t_inpcb;
168 struct socket *so = inp->inp_socket;
169 long len, recvwin, sendwin;
171 int off, flags, error = 0;
179 u_char opt[TCP_MAXOLEN];
180 unsigned int ipoptlen, optlen, hdrlen;
185 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
187 const boolean_t isipv6 = FALSE;
189 boolean_t can_tso = FALSE, use_tso;
190 boolean_t report_sack, idle_cwv = FALSE;
191 u_int segsz, tso_hlen;
193 KKASSERT(so->so_port == &curthread->td_msgport);
196 * Determine length of data that should be transmitted,
197 * and flags that will be used.
198 * If there is some data or critical controls (SYN, RST)
199 * to send, then transmit; otherwise, investigate further.
203 * If we have been idle for a while, the send congestion window
204 * could be no longer representative of the current state of the
205 * link; need to validate congestion window. However, we should
206 * not perform congestion window validation here, since we could
207 * be asked to send pure ACK.
209 if (tp->snd_max == tp->snd_una &&
210 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
214 * Calculate whether the transmit stream was previously idle
215 * and adjust TF_LASTIDLE for the next time.
217 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
218 if (idle && (tp->t_flags & TF_MORETOCOME))
219 tp->t_flags |= TF_LASTIDLE;
221 tp->t_flags &= ~TF_LASTIDLE;
223 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
224 !IN_FASTRECOVERY(tp))
225 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
228 * Find out whether TSO could be used or not
230 * For TSO capable devices, the following assumptions apply to
231 * the processing of TCP flags:
232 * - If FIN is set on the large TCP segment, the device must set
233 * FIN on the last segment that it creates from the large TCP
235 * - If PUSH is set on the large TCP segment, the device must set
236 * PUSH on the last segment that it creates from the large TCP
239 #if !defined(IPSEC) && !defined(FAST_IPSEC)
242 && (tp->t_flags & TF_SIGNATURE) == 0
246 struct rtentry *rt = inp->inp_route.ro_rt;
248 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
249 (rt->rt_ifp->if_hwassist & CSUM_TSO))
253 #endif /* !IPSEC && !FAST_IPSEC */
262 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
264 (!TAILQ_EMPTY(&tp->t_segq) ||
265 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
270 /* Make use of SACK information when slow-starting after a RTO. */
271 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
272 !IN_FASTRECOVERY(tp)) {
273 tcp_seq old_snd_nxt = tp->snd_nxt;
275 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
276 nsacked += tp->snd_nxt - old_snd_nxt;
280 off = tp->snd_nxt - tp->snd_una;
281 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
282 sendwin = min(sendwin, tp->snd_bwnd);
284 flags = tcp_outflags[tp->t_state];
286 * Get standard flags, and add SYN or FIN if requested by 'hidden'
289 if (tp->t_flags & TF_NEEDFIN)
291 if (tp->t_flags & TF_NEEDSYN)
295 * If in persist timeout with window of 0, send 1 byte.
296 * Otherwise, if window is small but nonzero
297 * and timer expired, we will send what we can
298 * and go to transmit state.
300 if (tp->t_flags & TF_FORCE) {
303 * If we still have some data to send, then
304 * clear the FIN bit. Usually this would
305 * happen below when it realizes that we
306 * aren't sending all the data. However,
307 * if we have exactly 1 byte of unsent data,
308 * then it won't clear the FIN bit below,
309 * and if we are in persist state, we wind
310 * up sending the packet without recording
311 * that we sent the FIN bit.
313 * We can't just blindly clear the FIN bit,
314 * because if we don't have any more data
315 * to send then the probe will be the FIN
318 if (off < so->so_snd.ssb_cc)
322 tcp_callout_stop(tp, tp->tt_persist);
328 * If snd_nxt == snd_max and we have transmitted a FIN, the
329 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
330 * a negative length. This can also occur when TCP opens up
331 * its congestion window while receiving additional duplicate
332 * acks after fast-retransmit because TCP will reset snd_nxt
333 * to snd_max after the fast-retransmit.
335 * A negative length can also occur when we are in the
336 * TCPS_SYN_RECEIVED state due to a simultanious connect where
337 * our SYN has not been acked yet.
339 * In the normal retransmit-FIN-only case, however, snd_nxt will
340 * be set to snd_una, the offset will be 0, and the length may
343 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
346 * Lop off SYN bit if it has already been sent. However, if this
347 * is SYN-SENT state and if segment contains data, suppress sending
348 * segment (sending the segment would be an option if we still
349 * did TAO and the remote host supported it).
351 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
354 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
355 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
361 * Be careful not to send data and/or FIN on SYN segments.
362 * This measure is needed to prevent interoperability problems
363 * with not fully conformant TCP implementations.
365 if (flags & TH_SYN) {
372 * A negative len can occur if our FIN has been sent but not
373 * acked, or if we are in a simultanious connect in the
374 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
377 * If our window has contracted to 0 in the FIN case
378 * (which can only occur if we have NOT been called to
379 * retransmit as per code a few paragraphs up) then we
380 * want to shift the retransmit timer over to the
383 * However, if we are in the TCPS_SYN_RECEIVED state
384 * (the SYN case) we will be in a simultanious connect and
385 * the window may be zero degeneratively. In this case we
386 * do not want to shift to the persist timer after the SYN
387 * or the SYN+ACK transmission.
390 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
391 tcp_callout_stop(tp, tp->tt_rexmt);
393 tp->snd_nxt = tp->snd_una;
394 if (!tcp_callout_active(tp, tp->tt_persist))
399 KASSERT(len >= 0, ("%s: len < 0", __func__));
401 * Automatic sizing of send socket buffer. Often the send buffer
402 * size is not optimally adjusted to the actual network conditions
403 * at hand (delay bandwidth product). Setting the buffer size too
404 * small limits throughput on links with high bandwidth and high
405 * delay (eg. trans-continental/oceanic links). Setting the
406 * buffer size too big consumes too much real kernel memory,
407 * especially with many connections on busy servers.
409 * The criteria to step up the send buffer one notch are:
410 * 1. receive window of remote host is larger than send buffer
411 * (with a fudge factor of 5/4th);
412 * 2. send buffer is filled to 7/8th with data (so we actually
413 * have data to make use of it);
414 * 3. send buffer fill has not hit maximal automatic size;
415 * 4. our send window (slow start and cogestion controlled) is
416 * larger than sent but unacknowledged data in send buffer.
418 * The remote host receive window scaling factor may limit the
419 * growing of the send buffer before it reaches its allowed
422 * It scales directly with slow start or congestion window
423 * and does at most one step per received ACK. This fast
424 * scaling has the drawback of growing the send buffer beyond
425 * what is strictly necessary to make full use of a given
426 * delay*bandwith product. However testing has shown this not
427 * to be much of an problem. At worst we are trading wasting
428 * of available bandwith (the non-use of it) for wasting some
429 * socket buffer memory.
431 * TODO: Shrink send buffer during idle periods together
432 * with congestion window. Requires another timer. Has to
433 * wait for upcoming tcp timer rewrite.
435 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
436 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
437 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
438 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
439 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
442 newsize = ulmin(so->so_snd.ssb_hiwat +
445 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
446 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
447 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
448 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
454 * - Congestion window needs validation
455 * - There are SACK blocks to report
456 * - RST or SYN flags is set
460 * Checking for SYN|RST looks overkill, just to be safe than sorry
463 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
466 tcp_seq ugr_nxt = tp->snd_nxt;
468 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
469 tp->snd_nxt == tp->snd_max)
472 if (SEQ_GT(tp->snd_up, ugr_nxt))
478 * Find out segment size and header length for TSO
480 error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
485 segsz = tp->t_maxseg;
486 tso_hlen = 0; /* not used */
490 * Truncate to the maximum segment length if not TSO, and ensure that
491 * FIN is removed if the length no longer contains the last data byte.
498 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
499 * thoff), and make sure that we send equal size
500 * transfers down the stack (rather than big-small-
503 len = (min(len, (IP_MAXPACKET - tso_hlen)) / segsz) *
512 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
515 recvwin = ssb_space(&so->so_rcv);
518 * Sender silly window avoidance. We transmit under the following
519 * conditions when len is non-zero:
521 * - We have a full segment
522 * - This is the last buffer in a write()/send() and we are
523 * either idle or running NODELAY
524 * - we've timed out (e.g. persist timer)
525 * - we have more then 1/2 the maximum send window's worth of
526 * data (receiver may be limiting the window size)
527 * - we need to retransmit
533 * NOTE! on localhost connections an 'ack' from the remote
534 * end may occur synchronously with the output and cause
535 * us to flush a buffer queued with moretocome. XXX
537 * note: the len + off check is almost certainly unnecessary.
539 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
540 (idle || (tp->t_flags & TF_NODELAY)) &&
541 len + off >= so->so_snd.ssb_cc &&
542 !(tp->t_flags & TF_NOPUSH)) {
545 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
547 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
549 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
551 if (tp->t_flags & TF_XMITNOW)
556 * Compare available window to amount of window
557 * known to peer (as advertised window less
558 * next expected input). If the difference is at least two
559 * max size segments, or at least 50% of the maximum possible
560 * window, then want to send a window update to peer.
564 * "adv" is the amount we can increase the window,
565 * taking into account that we are limited by
566 * TCP_MAXWIN << tp->rcv_scale.
568 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
569 (tp->rcv_adv - tp->rcv_nxt);
573 * This ack case typically occurs when the user has drained
574 * the TCP socket buffer sufficiently to warrent an ack
575 * containing a 'pure window update'... that is, an ack that
576 * ONLY updates the tcp window.
578 * It is unclear why we would need to do a pure window update
579 * past 2 segments if we are going to do one at 1/2 the high
580 * water mark anyway, especially since under normal conditions
581 * the user program will drain the socket buffer quickly.
582 * The 2-segment pure window update will often add a large
583 * number of extra, unnecessary acks to the stream.
585 * avoid_pure_win_update now defaults to 1.
587 if (avoid_pure_win_update == 0 ||
588 (tp->t_flags & TF_RXRESIZED)) {
589 if (adv >= (long) (2 * segsz)) {
593 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
594 if (hiwat > (long)so->so_rcv.ssb_hiwat)
595 hiwat = (long)so->so_rcv.ssb_hiwat;
596 if (adv >= hiwat / 2)
601 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
602 * is also a catch-all for the retransmit timer timeout case.
604 if (tp->t_flags & TF_ACKNOW)
606 if ((flags & TH_RST) ||
607 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
609 if (SEQ_GT(tp->snd_up, tp->snd_una))
612 * If our state indicates that FIN should be sent
613 * and we have not yet done so, then we need to send.
615 if ((flags & TH_FIN) &&
616 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
620 * TCP window updates are not reliable, rather a polling protocol
621 * using ``persist'' packets is used to insure receipt of window
622 * updates. The three ``states'' for the output side are:
623 * idle not doing retransmits or persists
624 * persisting to move a small or zero window
625 * (re)transmitting and thereby not persisting
627 * tcp_callout_active(tp, tp->tt_persist)
628 * is true when we are in persist state.
629 * The TF_FORCE flag in tp->t_flags
630 * is set when we are called to send a persist packet.
631 * tcp_callout_active(tp, tp->tt_rexmt)
632 * is set when we are retransmitting
633 * The output side is idle when both timers are zero.
635 * If send window is too small, there is data to transmit, and no
636 * retransmit or persist is pending, then go to persist state.
638 * If nothing happens soon, send when timer expires:
639 * if window is nonzero, transmit what we can, otherwise force out
642 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
643 * with data pending. This situation can occur during a
644 * simultanious connect.
646 if (so->so_snd.ssb_cc > 0 &&
647 tp->t_state != TCPS_SYN_RECEIVED &&
648 !tcp_callout_active(tp, tp->tt_rexmt) &&
649 !tcp_callout_active(tp, tp->tt_persist)) {
655 * No reason to send a segment, just return.
657 tp->t_flags &= ~TF_XMITNOW;
662 * Before ESTABLISHED, force sending of initial options
663 * unless TCP set not to do any options.
664 * NOTE: we assume that the IP/TCP header plus TCP options
665 * always fit in a single mbuf, leaving room for a maximum
667 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
671 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
673 hdrlen = sizeof(struct tcpiphdr);
674 if (flags & TH_SYN) {
675 tp->snd_nxt = tp->iss;
676 if (!(tp->t_flags & TF_NOOPT)) {
679 opt[0] = TCPOPT_MAXSEG;
680 opt[1] = TCPOLEN_MAXSEG;
681 mss = htons((u_short) tcp_mssopt(tp));
682 memcpy(opt + 2, &mss, sizeof mss);
683 optlen = TCPOLEN_MAXSEG;
685 if ((tp->t_flags & TF_REQ_SCALE) &&
686 (!(flags & TH_ACK) ||
687 (tp->t_flags & TF_RCVD_SCALE))) {
688 *((u_int32_t *)(opt + optlen)) = htonl(
690 TCPOPT_WINDOW << 16 |
691 TCPOLEN_WINDOW << 8 |
692 tp->request_r_scale);
696 if ((tcp_do_sack && !(flags & TH_ACK)) ||
697 tp->t_flags & TF_SACK_PERMITTED) {
698 uint32_t *lp = (uint32_t *)(opt + optlen);
700 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
701 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
707 * Send a timestamp and echo-reply if this is a SYN and our side
708 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
709 * and our peer have sent timestamps in our SYN's.
711 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
713 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
714 u_int32_t *lp = (u_int32_t *)(opt + optlen);
716 /* Form timestamp option as shown in appendix A of RFC 1323. */
717 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
718 *lp++ = htonl(ticks);
719 *lp = htonl(tp->ts_recent);
720 optlen += TCPOLEN_TSTAMP_APPA;
723 /* Set receive buffer autosizing timestamp. */
724 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
725 tp->rfbuf_ts = ticks;
728 * If this is a SACK connection and we have a block to report,
729 * fill in the SACK blocks in the TCP options.
732 tcp_sack_fill_report(tp, opt, &optlen);
735 if (tp->t_flags & TF_SIGNATURE) {
739 * Initialize TCP-MD5 option (RFC2385)
741 bp = (u_char *)opt + optlen;
742 *bp++ = TCPOPT_SIGNATURE;
743 *bp++ = TCPOLEN_SIGNATURE;
745 for (i = 0; i < TCP_SIGLEN; i++)
747 optlen += TCPOLEN_SIGNATURE;
749 * Terminate options list and maintain 32-bit alignment.
755 #endif /* TCP_SIGNATURE */
756 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
760 ipoptlen = ip6_optlen(inp);
762 if (inp->inp_options) {
763 ipoptlen = inp->inp_options->m_len -
764 offsetof(struct ipoption, ipopt_list);
770 ipoptlen += ipsec_hdrsiz_tcp(tp);
775 ("invalid TSO len %ld, segsz %u", len, segsz));
777 KASSERT(len <= segsz,
778 ("invalid len %ld, segsz %u", len, segsz));
781 * Adjust data length if insertion of options will bump
782 * the packet length beyond the t_maxopd length. Clear
783 * FIN to prevent premature closure since there is still
784 * more data to send after this (now truncated) packet.
786 * If just the options do not fit we are in a no-win
787 * situation and we treat it as an unreachable host.
789 if (len + optlen + ipoptlen > tp->t_maxopd) {
790 if (tp->t_maxopd <= optlen + ipoptlen) {
791 static time_t last_optlen_report;
793 if (last_optlen_report != time_second) {
794 last_optlen_report = time_second;
795 kprintf("tcpcb %p: MSS (%d) too "
796 "small to hold options!\n",
799 error = EHOSTUNREACH;
803 len = tp->t_maxopd - optlen - ipoptlen;
810 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
812 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
816 * Grab a header mbuf, attaching a copy of data to
817 * be transmitted, and initialize the header from
818 * the template for sends on this connection.
821 if ((tp->t_flags & TF_FORCE) && len == 1)
822 tcpstat.tcps_sndprobe++;
823 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
824 if (tp->snd_nxt == tp->snd_una)
825 tp->snd_max_rexmt = tp->snd_max;
827 tcpstat.tcps_sndsackrtopack++;
828 tcpstat.tcps_sndsackrtobyte += len;
830 tcpstat.tcps_sndrexmitpack++;
831 tcpstat.tcps_sndrexmitbyte += len;
833 tcpstat.tcps_sndpack++;
834 tcpstat.tcps_sndbyte += len;
838 tcp_idle_cwnd_validate(tp);
840 /* Update last send time after CWV */
841 tp->snd_last = ticks;
843 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
844 max_linkhdr + hdrlen)) == NULL) {
849 * m_copypack left space for our hdr; use it.
855 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
857 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
864 m->m_data += max_linkhdr;
866 if (len <= MHLEN - hdrlen - max_linkhdr) {
867 m_copydata(so->so_snd.ssb_mb, off, (int) len,
868 mtod(m, caddr_t) + hdrlen);
871 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
872 if (m->m_next == NULL) {
881 * If we're sending everything we've got, set PUSH.
882 * (This will keep happy those implementations which only
883 * give data to the user when a buffer fills or
886 if (off + len == so->so_snd.ssb_cc)
889 if (tp->t_flags & TF_ACKNOW)
890 tcpstat.tcps_sndacks++;
891 else if (flags & (TH_SYN | TH_FIN | TH_RST))
892 tcpstat.tcps_sndctrl++;
893 else if (SEQ_GT(tp->snd_up, tp->snd_una))
894 tcpstat.tcps_sndurg++;
896 tcpstat.tcps_sndwinup++;
898 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
904 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
907 m->m_data += max_linkhdr;
910 m->m_pkthdr.rcvif = NULL;
912 ip6 = mtod(m, struct ip6_hdr *);
913 th = (struct tcphdr *)(ip6 + 1);
914 tcp_fillheaders(tp, ip6, th, use_tso);
916 ip = mtod(m, struct ip *);
917 ipov = (struct ipovly *)ip;
918 th = (struct tcphdr *)(ip + 1);
919 /* this picks up the pseudo header (w/o the length) */
920 tcp_fillheaders(tp, ip, th, use_tso);
924 * Fill in fields, remembering maximum advertised
925 * window for use in delaying messages about window sizes.
926 * If resending a FIN, be sure not to use a new sequence number.
928 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
929 tp->snd_nxt == tp->snd_max)
934 * If we are doing retransmissions, then snd_nxt will
935 * not reflect the first unsent octet. For ACK only
936 * packets, we do not want the sequence number of the
937 * retransmitted packet, we want the sequence number
938 * of the next unsent octet. So, if there is no data
939 * (and no SYN or FIN), use snd_max instead of snd_nxt
940 * when filling in ti_seq. But if we are in persist
941 * state, snd_max might reflect one byte beyond the
942 * right edge of the window, so use snd_nxt in that
943 * case, since we know we aren't doing a retransmission.
944 * (retransmit and persist are mutually exclusive...)
946 if (len || (flags & (TH_SYN|TH_FIN)) ||
947 tcp_callout_active(tp, tp->tt_persist))
948 th->th_seq = htonl(tp->snd_nxt);
950 th->th_seq = htonl(tp->snd_max);
951 th->th_ack = htonl(tp->rcv_nxt);
953 bcopy(opt, th + 1, optlen);
954 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
956 th->th_flags = flags;
960 * Calculate receive window. Don't shrink window, but avoid
961 * silly window syndrome by sending a 0 window if the actual
962 * window is less then one segment.
964 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
965 recvwin < (long)segsz)
967 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
968 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
969 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
970 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
973 * Adjust the RXWIN0SENT flag - indicate that we have advertised
974 * a 0 window. This may cause the remote transmitter to stall. This
975 * flag tells soreceive() to disable delayed acknowledgements when
976 * draining the buffer. This can occur if the receiver is attempting
977 * to read more data then can be buffered prior to transmitting on
981 tp->t_flags |= TF_RXWIN0SENT;
983 tp->t_flags &= ~TF_RXWIN0SENT;
986 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
988 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
989 KASSERT(!use_tso, ("URG with TSO"));
991 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
992 th->th_flags |= TH_URG;
996 * If no urgent pointer to send, then we pull
997 * the urgent pointer to the left edge of the send window
998 * so that it doesn't drift into the send window on sequence
1001 tp->snd_up = tp->snd_una; /* drag it along */
1005 #ifdef TCP_SIGNATURE
1006 if (tp->t_flags & TF_SIGNATURE) {
1007 tcpsignature_compute(m, len, optlen,
1008 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1010 #endif /* TCP_SIGNATURE */
1013 * Put TCP length in extended header, and then
1014 * checksum extended header and data.
1016 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1019 * ip6_plen is not need to be filled now, and will be
1020 * filled in ip6_output().
1022 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1023 sizeof(struct ip6_hdr),
1024 sizeof(struct tcphdr) + optlen + len);
1026 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1028 m->m_pkthdr.csum_flags = CSUM_TSO;
1029 m->m_pkthdr.segsz = segsz;
1031 m->m_pkthdr.csum_flags = CSUM_TCP;
1033 th->th_sum = in_addword(th->th_sum,
1034 htons((u_short)(optlen + len)));
1039 * IP version must be set here for ipv4/ipv6 checking
1042 KASSERT(ip->ip_v == IPVERSION,
1043 ("%s: IP version incorrect: %d",
1044 __func__, ip->ip_v));
1049 * In transmit state, time the transmission and arrange for
1050 * the retransmit. In persist state, just set snd_max.
1052 if (!(tp->t_flags & TF_FORCE) ||
1053 !tcp_callout_active(tp, tp->tt_persist)) {
1054 tcp_seq startseq = tp->snd_nxt;
1057 * Advance snd_nxt over sequence space of this segment.
1059 if (flags & (TH_SYN | TH_FIN)) {
1062 if (flags & TH_FIN) {
1064 tp->t_flags |= TF_SENTFIN;
1068 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1069 tp->snd_max = tp->snd_nxt;
1071 * Time this transmission if not a retransmission and
1072 * not currently timing anything.
1074 if (tp->t_rtttime == 0) {
1075 tp->t_rtttime = ticks;
1076 tp->t_rtseq = startseq;
1077 tcpstat.tcps_segstimed++;
1082 * Set retransmit timer if not currently set,
1083 * and not doing a pure ack or a keep-alive probe.
1084 * Initial value for retransmit timer is smoothed
1085 * round-trip time + 2 * round-trip time variance.
1086 * Initialize shift counter which is used for backoff
1087 * of retransmit time.
1089 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1090 tp->snd_nxt != tp->snd_una) {
1091 if (tcp_callout_active(tp, tp->tt_persist)) {
1092 tcp_callout_stop(tp, tp->tt_persist);
1095 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1100 * Persist case, update snd_max but since we are in
1101 * persist mode (no window) we do not update snd_nxt.
1105 panic("tcp_output: persist timer to send SYN");
1106 if (flags & TH_FIN) {
1108 tp->t_flags |= TF_SENTFIN;
1110 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1111 tp->snd_max = tp->snd_nxt + xlen;
1117 if (so->so_options & SO_DEBUG) {
1118 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1119 mtod(m, void *), th, 0);
1124 * Fill in IP length and desired time to live and
1125 * send to IP level. There should be a better way
1126 * to handle ttl and tos; we could keep them in
1127 * the template, but need a way to checksum without them.
1130 * m->m_pkthdr.len should have been set before cksum
1131 * calcuration, because in6_cksum() need it.
1135 * we separately set hoplimit for every segment,
1136 * since the user might want to change the value
1137 * via setsockopt. Also, desired default hop
1138 * limit might be changed via Neighbor Discovery.
1140 ip6->ip6_hlim = in6_selecthlim(inp,
1141 (inp->in6p_route.ro_rt ?
1142 inp->in6p_route.ro_rt->rt_ifp : NULL));
1144 /* TODO: IPv6 IP6TOS_ECT bit on */
1145 error = ip6_output(m, inp->in6p_outputopts,
1146 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1150 ip->ip_len = m->m_pkthdr.len;
1152 if (INP_CHECK_SOCKAF(so, AF_INET6))
1153 ip->ip_ttl = in6_selecthlim(inp,
1154 (inp->in6p_route.ro_rt ?
1155 inp->in6p_route.ro_rt->rt_ifp : NULL));
1158 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1160 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1162 * See if we should do MTU discovery.
1163 * We do it only if the following are true:
1164 * 1) we have a valid route to the destination
1165 * 2) the MTU is not locked (if it is,
1166 * then discovery has been disabled)
1168 if (path_mtu_discovery &&
1169 (rt = inp->inp_route.ro_rt) &&
1170 (rt->rt_flags & RTF_UP) &&
1171 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1172 ip->ip_off |= IP_DF;
1174 error = ip_output(m, inp->inp_options, &inp->inp_route,
1175 (so->so_options & SO_DONTROUTE) |
1176 IP_DEBUGROUTE, NULL, inp);
1179 KASSERT(error != 0, ("no error, but th not set"));
1182 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1185 * We know that the packet was lost, so back out the
1186 * sequence number advance, if any.
1188 if (!(tp->t_flags & TF_FORCE) ||
1189 !tcp_callout_active(tp, tp->tt_persist)) {
1191 * No need to check for TH_FIN here because
1192 * the TF_SENTFIN flag handles that case.
1194 if (!(flags & TH_SYN))
1199 if (error == ENOBUFS) {
1201 * If we can't send, make sure there is something
1202 * to get us going again later.
1204 * The persist timer isn't necessarily allowed in all
1205 * states, use the rexmt timer.
1207 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1208 !tcp_callout_active(tp, tp->tt_persist)) {
1209 tcp_callout_reset(tp, tp->tt_rexmt,
1220 if (error == EMSGSIZE) {
1222 * ip_output() will have already fixed the route
1223 * for us. tcp_mtudisc() will, as its last action,
1224 * initiate retransmission, so it is important to
1227 tcp_mtudisc(inp, 0);
1230 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1231 TCPS_HAVERCVDSYN(tp->t_state)) {
1232 tp->t_softerror = error;
1237 tcpstat.tcps_sndtotal++;
1240 * Data sent (as far as we can tell).
1242 * If this advertises a larger window than any other segment,
1243 * then remember the size of the advertised window.
1245 * Any pending ACK has now been sent.
1247 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1248 tp->rcv_adv = tp->rcv_nxt + recvwin;
1249 tp->t_flags &= ~TF_RXRESIZED;
1251 tp->last_ack_sent = tp->rcv_nxt;
1252 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1253 if (tcp_delack_enabled)
1254 tcp_callout_stop(tp, tp->tt_delack);
1261 tcp_setpersist(struct tcpcb *tp)
1263 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1266 if (tp->t_state == TCPS_SYN_SENT ||
1267 tp->t_state == TCPS_SYN_RECEIVED) {
1268 panic("tcp_setpersist: not established yet, current %s",
1269 tp->t_state == TCPS_SYN_SENT ?
1270 "SYN_SENT" : "SYN_RECEIVED");
1273 if (tcp_callout_active(tp, tp->tt_rexmt))
1274 panic("tcp_setpersist: retransmit pending");
1276 * Start/restart persistance timer.
1278 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1280 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1281 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1286 tcp_idle_cwnd_validate(struct tcpcb *tp)
1288 u_long initial_cwnd = tcp_initial_window(tp);
1291 tcpstat.tcps_sndidle++;
1293 /* According to RFC5681: RW=min(IW,cwnd) */
1294 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1297 u_long idle_time, decay_cwnd;
1300 * RFC2861, but only after idle period.
1304 * Before the congestion window is reduced, ssthresh
1305 * is set to the maximum of its current value and 3/4
1306 * cwnd. If the sender then has more data to send
1307 * than the decayed cwnd allows, the TCP will slow-
1308 * start (perform exponential increase) at least
1309 * half-way back up to the old value of cwnd.
1311 tp->snd_ssthresh = max(tp->snd_ssthresh,
1312 (3 * tp->snd_cwnd) / 4);
1315 * Decay the congestion window by half for every RTT
1316 * that the flow remains inactive.
1318 * The difference between our implementation and
1319 * RFC2861 is that we don't allow cwnd to go below
1320 * the value allowed by RFC5681 (min_cwnd).
1322 idle_time = ticks - tp->snd_last;
1323 decay_cwnd = tp->snd_cwnd;
1324 while (idle_time >= tp->t_rxtcur &&
1325 decay_cwnd > min_cwnd) {
1327 idle_time -= tp->t_rxtcur;
1329 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1332 * Slow-start from scratch to re-determine the send
1333 * congestion window.
1335 tp->snd_cwnd = min_cwnd;
1338 /* Restart ABC counting during congestion avoidance */
1343 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1345 struct inpcb * const inp = tp->t_inpcb;
1347 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1349 const boolean_t isipv6 = FALSE;
1351 unsigned int ipoptlen, optlen;
1354 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1357 ipoptlen = ip6_optlen(inp);
1359 if (inp->inp_options) {
1360 ipoptlen = inp->inp_options->m_len -
1361 offsetof(struct ipoption, ipopt_list);
1367 ipoptlen += ipsec_hdrsiz_tcp(tp);
1372 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1373 (tp->t_flags & TF_RCVD_TSTMP))
1374 optlen += TCPOLEN_TSTAMP_APPA;
1377 if (tp->t_maxopd <= optlen + ipoptlen)
1378 return EHOSTUNREACH;
1380 *segsz = tp->t_maxopd - optlen - ipoptlen;