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. Neither the name of the University nor the names of its contributors
47 * may be used to endorse or promote products derived from this software
48 * without specific prior written permission.
50 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
51 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
52 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
53 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
54 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
55 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
56 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
57 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
58 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
59 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
62 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
63 * $FreeBSD: src/sys/netinet/tcp_output.c,v 1.39.2.20 2003/01/29 22:45:36 hsu Exp $
67 #include "opt_inet6.h"
68 #include "opt_ipsec.h"
69 #include "opt_tcpdebug.h"
71 #include <sys/param.h>
72 #include <sys/systm.h>
73 #include <sys/kernel.h>
74 #include <sys/sysctl.h>
76 #include <sys/domain.h>
77 #include <sys/protosw.h>
78 #include <sys/socket.h>
79 #include <sys/socketvar.h>
80 #include <sys/in_cksum.h>
81 #include <sys/thread.h>
82 #include <sys/globaldata.h>
84 #include <net/if_var.h>
85 #include <net/route.h>
86 #include <net/netmsg2.h>
87 #include <net/netisr2.h>
89 #include <netinet/in.h>
90 #include <netinet/in_systm.h>
91 #include <netinet/ip.h>
92 #include <netinet/in_pcb.h>
93 #include <netinet/ip_var.h>
94 #include <netinet6/in6_pcb.h>
95 #include <netinet/ip6.h>
96 #include <netinet6/ip6_var.h>
97 #include <netinet/tcp.h>
99 #include <netinet/tcp_fsm.h>
100 #include <netinet/tcp_seq.h>
101 #include <netinet/tcp_timer.h>
102 #include <netinet/tcp_timer2.h>
103 #include <netinet/tcp_var.h>
104 #include <netinet/tcpip.h>
106 #include <netinet/tcp_debug.h>
110 #include <netinet6/ipsec.h>
114 #include <netproto/ipsec/ipsec.h>
116 #endif /*FAST_IPSEC*/
119 extern struct mbuf *m_copypack();
122 int path_mtu_discovery = 0;
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
124 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
126 static int avoid_pure_win_update = 1;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, avoid_pure_win_update, CTLFLAG_RW,
128 &avoid_pure_win_update, 1, "Avoid pure window updates when possible");
130 int tcp_do_autosndbuf = 1;
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
132 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
134 int tcp_autosndbuf_inc = 8*1024;
135 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
136 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
138 int tcp_autosndbuf_max = 2*1024*1024;
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
140 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
142 int tcp_prio_synack = 1;
143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, prio_synack, CTLFLAG_RW,
144 &tcp_prio_synack, 0, "Prioritize SYN, SYN|ACK and pure ACK");
146 static int tcp_idle_cwv = 1;
147 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
149 "Congestion window validation after idle period (part of RFC2861)");
151 static int tcp_idle_restart = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
153 &tcp_idle_restart, 0, "Reset congestion window after idle period");
155 static int tcp_do_tso = 1;
156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
157 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
159 static int tcp_fairsend = 4;
160 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW,
162 "Amount of segments sent before yield to other senders or receivers");
164 static void tcp_idle_cwnd_validate(struct tcpcb *);
166 static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
167 static void tcp_output_sched(struct tcpcb *tp);
170 * Tcp output routine: figure out what should be sent and send it.
173 tcp_output(struct tcpcb *tp)
175 struct inpcb * const inp = tp->t_inpcb;
176 struct socket *so = inp->inp_socket;
177 long len, recvwin, sendwin;
179 int off, flags, error = 0;
186 u_char opt[TCP_MAXOLEN];
187 unsigned int ipoptlen, optlen, hdrlen;
192 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
194 const boolean_t isipv6 = FALSE;
196 boolean_t can_tso = FALSE, use_tso;
197 boolean_t report_sack, idle_cwv = FALSE;
198 u_int segsz, tso_hlen, tso_lenmax = 0;
200 boolean_t need_sched = FALSE;
202 KKASSERT(so->so_port == &curthread->td_msgport);
205 * Determine length of data that should be transmitted,
206 * and flags that will be used.
207 * If there is some data or critical controls (SYN, RST)
208 * to send, then transmit; otherwise, investigate further.
212 * If we have been idle for a while, the send congestion window
213 * could be no longer representative of the current state of the
214 * link; need to validate congestion window. However, we should
215 * not perform congestion window validation here, since we could
216 * be asked to send pure ACK.
218 if (tp->snd_max == tp->snd_una &&
219 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
223 * Calculate whether the transmit stream was previously idle
224 * and adjust TF_LASTIDLE for the next time.
226 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
227 if (idle && (tp->t_flags & TF_MORETOCOME))
228 tp->t_flags |= TF_LASTIDLE;
230 tp->t_flags &= ~TF_LASTIDLE;
232 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
233 !IN_FASTRECOVERY(tp))
234 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
237 * Find out whether TSO could be used or not
239 * For TSO capable devices, the following assumptions apply to
240 * the processing of TCP flags:
241 * - If FIN is set on the large TCP segment, the device must set
242 * FIN on the last segment that it creates from the large TCP
244 * - If PUSH is set on the large TCP segment, the device must set
245 * PUSH on the last segment that it creates from the large TCP
248 #if !defined(IPSEC) && !defined(FAST_IPSEC)
251 && (tp->t_flags & TF_SIGNATURE) == 0
255 struct rtentry *rt = inp->inp_route.ro_rt;
257 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
258 (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
260 tso_lenmax = rt->rt_ifp->if_tsolen;
264 #endif /* !IPSEC && !FAST_IPSEC */
272 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
274 (!TAILQ_EMPTY(&tp->t_segq) ||
275 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
280 /* Make use of SACK information when slow-starting after a RTO. */
281 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
282 !IN_FASTRECOVERY(tp)) {
283 tcp_seq old_snd_nxt = tp->snd_nxt;
285 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
286 nsacked += tp->snd_nxt - old_snd_nxt;
290 off = tp->snd_nxt - tp->snd_una;
291 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
292 sendwin = min(sendwin, tp->snd_bwnd);
294 flags = tcp_outflags[tp->t_state];
296 * Get standard flags, and add SYN or FIN if requested by 'hidden'
299 if (tp->t_flags & TF_NEEDFIN)
301 if (tp->t_flags & TF_NEEDSYN)
305 * If in persist timeout with window of 0, send 1 byte.
306 * Otherwise, if window is small but nonzero
307 * and timer expired, we will send what we can
308 * and go to transmit state.
310 if (tp->t_flags & TF_FORCE) {
313 * If we still have some data to send, then
314 * clear the FIN bit. Usually this would
315 * happen below when it realizes that we
316 * aren't sending all the data. However,
317 * if we have exactly 1 byte of unsent data,
318 * then it won't clear the FIN bit below,
319 * and if we are in persist state, we wind
320 * up sending the packet without recording
321 * that we sent the FIN bit.
323 * We can't just blindly clear the FIN bit,
324 * because if we don't have any more data
325 * to send then the probe will be the FIN
328 if (off < so->so_snd.ssb_cc)
332 tcp_callout_stop(tp, tp->tt_persist);
338 * If snd_nxt == snd_max and we have transmitted a FIN, the
339 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
340 * a negative length. This can also occur when TCP opens up
341 * its congestion window while receiving additional duplicate
342 * acks after fast-retransmit because TCP will reset snd_nxt
343 * to snd_max after the fast-retransmit.
345 * A negative length can also occur when we are in the
346 * TCPS_SYN_RECEIVED state due to a simultanious connect where
347 * our SYN has not been acked yet.
349 * In the normal retransmit-FIN-only case, however, snd_nxt will
350 * be set to snd_una, the offset will be 0, and the length may
353 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
356 * Lop off SYN bit if it has already been sent. However, if this
357 * is SYN-SENT state and if segment contains data, suppress sending
358 * segment (sending the segment would be an option if we still
359 * did TAO and the remote host supported it).
361 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
364 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
365 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
371 * Be careful not to send data and/or FIN on SYN segments.
372 * This measure is needed to prevent interoperability problems
373 * with not fully conformant TCP implementations.
375 if (flags & TH_SYN) {
382 * A negative len can occur if our FIN has been sent but not
383 * acked, or if we are in a simultanious connect in the
384 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
387 * If our window has contracted to 0 in the FIN case
388 * (which can only occur if we have NOT been called to
389 * retransmit as per code a few paragraphs up) then we
390 * want to shift the retransmit timer over to the
393 * However, if we are in the TCPS_SYN_RECEIVED state
394 * (the SYN case) we will be in a simultanious connect and
395 * the window may be zero degeneratively. In this case we
396 * do not want to shift to the persist timer after the SYN
397 * or the SYN+ACK transmission.
400 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
401 tcp_callout_stop(tp, tp->tt_rexmt);
403 tp->snd_nxt = tp->snd_una;
404 if (!tcp_callout_active(tp, tp->tt_persist))
409 KASSERT(len >= 0, ("%s: len < 0", __func__));
411 * Automatic sizing of send socket buffer. Often the send buffer
412 * size is not optimally adjusted to the actual network conditions
413 * at hand (delay bandwidth product). Setting the buffer size too
414 * small limits throughput on links with high bandwidth and high
415 * delay (eg. trans-continental/oceanic links). Setting the
416 * buffer size too big consumes too much real kernel memory,
417 * especially with many connections on busy servers.
419 * The criteria to step up the send buffer one notch are:
420 * 1. receive window of remote host is larger than send buffer
421 * (with a fudge factor of 5/4th);
422 * 2. send buffer is filled to 7/8th with data (so we actually
423 * have data to make use of it);
424 * 3. send buffer fill has not hit maximal automatic size;
425 * 4. our send window (slow start and cogestion controlled) is
426 * larger than sent but unacknowledged data in send buffer.
428 * The remote host receive window scaling factor may limit the
429 * growing of the send buffer before it reaches its allowed
432 * It scales directly with slow start or congestion window
433 * and does at most one step per received ACK. This fast
434 * scaling has the drawback of growing the send buffer beyond
435 * what is strictly necessary to make full use of a given
436 * delay*bandwith product. However testing has shown this not
437 * to be much of an problem. At worst we are trading wasting
438 * of available bandwith (the non-use of it) for wasting some
439 * socket buffer memory.
441 * TODO: Shrink send buffer during idle periods together
442 * with congestion window. Requires another timer. Has to
443 * wait for upcoming tcp timer rewrite.
445 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
446 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
447 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
448 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
449 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
452 newsize = ulmin(so->so_snd.ssb_hiwat +
455 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
456 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
457 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
458 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
464 * - Congestion window needs validation
465 * - There are SACK blocks to report
466 * - RST or SYN flags is set
470 * Checking for SYN|RST looks overkill, just to be safe than sorry
473 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
476 tcp_seq ugr_nxt = tp->snd_nxt;
478 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
479 tp->snd_nxt == tp->snd_max)
482 if (SEQ_GT(tp->snd_up, ugr_nxt))
488 * Find out segment size and header length for TSO
490 error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
495 segsz = tp->t_maxseg;
496 tso_hlen = 0; /* not used */
500 * Truncate to the maximum segment length if not TSO, and ensure that
501 * FIN is removed if the length no longer contains the last data byte.
510 if (__predict_false(tso_lenmax < segsz))
511 tso_lenmax = segsz << 1;
514 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
515 * thoff), and make sure that we send equal size
516 * transfers down the stack (rather than big-small-
519 len = min(len, tso_lenmax);
520 nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz;
538 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
541 recvwin = ssb_space(&so->so_rcv);
544 * Sender silly window avoidance. We transmit under the following
545 * conditions when len is non-zero:
547 * - We have a full segment
548 * - This is the last buffer in a write()/send() and we are
549 * either idle or running NODELAY
550 * - we've timed out (e.g. persist timer)
551 * - we have more then 1/2 the maximum send window's worth of
552 * data (receiver may be limiting the window size)
553 * - we need to retransmit
559 * NOTE! on localhost connections an 'ack' from the remote
560 * end may occur synchronously with the output and cause
561 * us to flush a buffer queued with moretocome. XXX
563 * note: the len + off check is almost certainly unnecessary.
565 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
566 (idle || (tp->t_flags & TF_NODELAY)) &&
567 len + off >= so->so_snd.ssb_cc &&
568 !(tp->t_flags & TF_NOPUSH)) {
571 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
573 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
575 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
577 if (tp->t_flags & TF_XMITNOW)
582 * Compare available window to amount of window
583 * known to peer (as advertised window less
584 * next expected input). If the difference is at least two
585 * max size segments, or at least 50% of the maximum possible
586 * window, then want to send a window update to peer.
590 * "adv" is the amount we can increase the window,
591 * taking into account that we are limited by
592 * TCP_MAXWIN << tp->rcv_scale.
594 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
595 (tp->rcv_adv - tp->rcv_nxt);
599 * This ack case typically occurs when the user has drained
600 * the TCP socket buffer sufficiently to warrent an ack
601 * containing a 'pure window update'... that is, an ack that
602 * ONLY updates the tcp window.
604 * It is unclear why we would need to do a pure window update
605 * past 2 segments if we are going to do one at 1/2 the high
606 * water mark anyway, especially since under normal conditions
607 * the user program will drain the socket buffer quickly.
608 * The 2-segment pure window update will often add a large
609 * number of extra, unnecessary acks to the stream.
611 * avoid_pure_win_update now defaults to 1.
613 if (avoid_pure_win_update == 0 ||
614 (tp->t_flags & TF_RXRESIZED)) {
615 if (adv >= (long) (2 * segsz)) {
619 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
620 if (hiwat > (long)so->so_rcv.ssb_hiwat)
621 hiwat = (long)so->so_rcv.ssb_hiwat;
622 if (adv >= hiwat / 2)
627 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
628 * is also a catch-all for the retransmit timer timeout case.
630 if (tp->t_flags & TF_ACKNOW)
632 if ((flags & TH_RST) ||
633 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
635 if (SEQ_GT(tp->snd_up, tp->snd_una))
638 * If our state indicates that FIN should be sent
639 * and we have not yet done so, then we need to send.
641 if ((flags & TH_FIN) &&
642 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
646 * TCP window updates are not reliable, rather a polling protocol
647 * using ``persist'' packets is used to insure receipt of window
648 * updates. The three ``states'' for the output side are:
649 * idle not doing retransmits or persists
650 * persisting to move a small or zero window
651 * (re)transmitting and thereby not persisting
653 * tcp_callout_active(tp, tp->tt_persist)
654 * is true when we are in persist state.
655 * The TF_FORCE flag in tp->t_flags
656 * is set when we are called to send a persist packet.
657 * tcp_callout_active(tp, tp->tt_rexmt)
658 * is set when we are retransmitting
659 * The output side is idle when both timers are zero.
661 * If send window is too small, there is data to transmit, and no
662 * retransmit or persist is pending, then go to persist state.
664 * If nothing happens soon, send when timer expires:
665 * if window is nonzero, transmit what we can, otherwise force out
668 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
669 * with data pending. This situation can occur during a
670 * simultanious connect.
672 if (so->so_snd.ssb_cc > 0 &&
673 tp->t_state != TCPS_SYN_RECEIVED &&
674 !tcp_callout_active(tp, tp->tt_rexmt) &&
675 !tcp_callout_active(tp, tp->tt_persist)) {
681 * No reason to send a segment, just return.
683 tp->t_flags &= ~TF_XMITNOW;
687 if (need_sched && len > 0) {
688 tcp_output_sched(tp);
693 * Before ESTABLISHED, force sending of initial options
694 * unless TCP set not to do any options.
695 * NOTE: we assume that the IP/TCP header plus TCP options
696 * always fit in a single mbuf, leaving room for a maximum
698 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
702 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
704 hdrlen = sizeof(struct tcpiphdr);
705 if (flags & TH_SYN) {
706 tp->snd_nxt = tp->iss;
707 if (!(tp->t_flags & TF_NOOPT)) {
710 opt[0] = TCPOPT_MAXSEG;
711 opt[1] = TCPOLEN_MAXSEG;
712 mss = htons((u_short) tcp_mssopt(tp));
713 memcpy(opt + 2, &mss, sizeof mss);
714 optlen = TCPOLEN_MAXSEG;
716 if ((tp->t_flags & TF_REQ_SCALE) &&
717 (!(flags & TH_ACK) ||
718 (tp->t_flags & TF_RCVD_SCALE))) {
719 *((u_int32_t *)(opt + optlen)) = htonl(
721 TCPOPT_WINDOW << 16 |
722 TCPOLEN_WINDOW << 8 |
723 tp->request_r_scale);
727 if ((tcp_do_sack && !(flags & TH_ACK)) ||
728 tp->t_flags & TF_SACK_PERMITTED) {
729 uint32_t *lp = (uint32_t *)(opt + optlen);
731 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
732 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
738 * Send a timestamp and echo-reply if this is a SYN and our side
739 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
740 * and our peer have sent timestamps in our SYN's.
742 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
744 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
745 u_int32_t *lp = (u_int32_t *)(opt + optlen);
747 /* Form timestamp option as shown in appendix A of RFC 1323. */
748 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
749 *lp++ = htonl(ticks);
750 *lp = htonl(tp->ts_recent);
751 optlen += TCPOLEN_TSTAMP_APPA;
754 /* Set receive buffer autosizing timestamp. */
755 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
756 tp->rfbuf_ts = ticks;
759 * If this is a SACK connection and we have a block to report,
760 * fill in the SACK blocks in the TCP options.
763 tcp_sack_fill_report(tp, opt, &optlen);
766 if (tp->t_flags & TF_SIGNATURE) {
770 * Initialize TCP-MD5 option (RFC2385)
772 bp = (u_char *)opt + optlen;
773 *bp++ = TCPOPT_SIGNATURE;
774 *bp++ = TCPOLEN_SIGNATURE;
776 for (i = 0; i < TCP_SIGLEN; i++)
778 optlen += TCPOLEN_SIGNATURE;
780 * Terminate options list and maintain 32-bit alignment.
786 #endif /* TCP_SIGNATURE */
787 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
791 ipoptlen = ip6_optlen(inp);
793 if (inp->inp_options) {
794 ipoptlen = inp->inp_options->m_len -
795 offsetof(struct ipoption, ipopt_list);
801 ipoptlen += ipsec_hdrsiz_tcp(tp);
805 /* TSO segment length must be multiple of segment size */
806 KASSERT(len >= (2 * segsz) && (len % segsz == 0),
807 ("invalid TSO len %ld, segsz %u", len, segsz));
809 KASSERT(len <= segsz,
810 ("invalid len %ld, segsz %u", len, segsz));
813 * Adjust data length if insertion of options will bump
814 * the packet length beyond the t_maxopd length. Clear
815 * FIN to prevent premature closure since there is still
816 * more data to send after this (now truncated) packet.
818 * If just the options do not fit we are in a no-win
819 * situation and we treat it as an unreachable host.
821 if (len + optlen + ipoptlen > tp->t_maxopd) {
822 if (tp->t_maxopd <= optlen + ipoptlen) {
823 static time_t last_optlen_report;
825 if (last_optlen_report != time_uptime) {
826 last_optlen_report = time_uptime;
827 kprintf("tcpcb %p: MSS (%d) too "
828 "small to hold options!\n",
831 error = EHOSTUNREACH;
835 len = tp->t_maxopd - optlen - ipoptlen;
842 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
844 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
848 * Grab a header mbuf, attaching a copy of data to
849 * be transmitted, and initialize the header from
850 * the template for sends on this connection.
853 if ((tp->t_flags & TF_FORCE) && len == 1)
854 tcpstat.tcps_sndprobe++;
855 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
856 if (tp->snd_nxt == tp->snd_una)
857 tp->snd_max_rexmt = tp->snd_max;
859 tcpstat.tcps_sndsackrtopack++;
860 tcpstat.tcps_sndsackrtobyte += len;
862 tcpstat.tcps_sndrexmitpack++;
863 tcpstat.tcps_sndrexmitbyte += len;
865 tcpstat.tcps_sndpack++;
866 tcpstat.tcps_sndbyte += len;
870 tcp_idle_cwnd_validate(tp);
872 /* Update last send time after CWV */
873 tp->snd_last = ticks;
875 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
876 max_linkhdr + hdrlen)) == NULL) {
881 * m_copypack left space for our hdr; use it.
887 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
889 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
896 m->m_data += max_linkhdr;
898 if (len <= MHLEN - hdrlen - max_linkhdr) {
899 m_copydata(so->so_snd.ssb_mb, off, (int) len,
900 mtod(m, caddr_t) + hdrlen);
903 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
904 if (m->m_next == NULL) {
913 * If we're sending everything we've got, set PUSH.
914 * (This will keep happy those implementations which only
915 * give data to the user when a buffer fills or
918 if (off + len == so->so_snd.ssb_cc)
921 if (tp->t_flags & TF_ACKNOW)
922 tcpstat.tcps_sndacks++;
923 else if (flags & (TH_SYN | TH_FIN | TH_RST))
924 tcpstat.tcps_sndctrl++;
925 else if (SEQ_GT(tp->snd_up, tp->snd_una))
926 tcpstat.tcps_sndurg++;
928 tcpstat.tcps_sndwinup++;
930 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
936 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
939 m->m_data += max_linkhdr;
943 * Prioritize SYN, SYN|ACK and pure ACK.
944 * Leave FIN and RST as they are.
946 if (tcp_prio_synack && (flags & (TH_FIN | TH_RST)) == 0)
947 m->m_flags |= M_PRIO;
949 m->m_pkthdr.rcvif = NULL;
951 ip6 = mtod(m, struct ip6_hdr *);
952 th = (struct tcphdr *)(ip6 + 1);
953 tcp_fillheaders(tp, ip6, th, use_tso);
955 ip = mtod(m, struct ip *);
956 th = (struct tcphdr *)(ip + 1);
957 /* this picks up the pseudo header (w/o the length) */
958 tcp_fillheaders(tp, ip, th, use_tso);
962 * Fill in fields, remembering maximum advertised
963 * window for use in delaying messages about window sizes.
964 * If resending a FIN, be sure not to use a new sequence number.
966 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
967 tp->snd_nxt == tp->snd_max)
972 * If we are doing retransmissions, then snd_nxt will
973 * not reflect the first unsent octet. For ACK only
974 * packets, we do not want the sequence number of the
975 * retransmitted packet, we want the sequence number
976 * of the next unsent octet. So, if there is no data
977 * (and no SYN or FIN), use snd_max instead of snd_nxt
978 * when filling in ti_seq. But if we are in persist
979 * state, snd_max might reflect one byte beyond the
980 * right edge of the window, so use snd_nxt in that
981 * case, since we know we aren't doing a retransmission.
982 * (retransmit and persist are mutually exclusive...)
984 if (len || (flags & (TH_SYN|TH_FIN)) ||
985 tcp_callout_active(tp, tp->tt_persist))
986 th->th_seq = htonl(tp->snd_nxt);
988 th->th_seq = htonl(tp->snd_max);
989 th->th_ack = htonl(tp->rcv_nxt);
991 bcopy(opt, th + 1, optlen);
992 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
994 th->th_flags = flags;
998 * Calculate receive window. Don't shrink window, but avoid
999 * silly window syndrome by sending a 0 window if the actual
1000 * window is less then one segment.
1002 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
1003 recvwin < (long)segsz)
1005 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
1006 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
1007 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
1008 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
1011 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1012 * a 0 window. This may cause the remote transmitter to stall. This
1013 * flag tells soreceive() to disable delayed acknowledgements when
1014 * draining the buffer. This can occur if the receiver is attempting
1015 * to read more data then can be buffered prior to transmitting on
1019 tp->t_flags |= TF_RXWIN0SENT;
1021 tp->t_flags &= ~TF_RXWIN0SENT;
1024 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
1026 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1027 KASSERT(!use_tso, ("URG with TSO"));
1029 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1030 th->th_flags |= TH_URG;
1034 * If no urgent pointer to send, then we pull
1035 * the urgent pointer to the left edge of the send window
1036 * so that it doesn't drift into the send window on sequence
1037 * number wraparound.
1039 tp->snd_up = tp->snd_una; /* drag it along */
1043 #ifdef TCP_SIGNATURE
1044 if (tp->t_flags & TF_SIGNATURE) {
1045 tcpsignature_compute(m, len, optlen,
1046 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1048 #endif /* TCP_SIGNATURE */
1051 * Put TCP length in extended header, and then
1052 * checksum extended header and data.
1054 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1057 * ip6_plen is not need to be filled now, and will be
1058 * filled in ip6_output().
1060 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1061 sizeof(struct ip6_hdr),
1062 sizeof(struct tcphdr) + optlen + len);
1064 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1066 m->m_pkthdr.csum_flags = CSUM_TSO;
1067 m->m_pkthdr.tso_segsz = segsz;
1069 m->m_pkthdr.csum_flags = CSUM_TCP;
1070 m->m_pkthdr.csum_data =
1071 offsetof(struct tcphdr, th_sum);
1073 th->th_sum = in_addword(th->th_sum,
1074 htons((u_short)(optlen + len)));
1079 * IP version must be set here for ipv4/ipv6 checking
1082 KASSERT(ip->ip_v == IPVERSION,
1083 ("%s: IP version incorrect: %d",
1084 __func__, ip->ip_v));
1089 * In transmit state, time the transmission and arrange for
1090 * the retransmit. In persist state, just set snd_max.
1092 if (!(tp->t_flags & TF_FORCE) ||
1093 !tcp_callout_active(tp, tp->tt_persist)) {
1094 tcp_seq startseq = tp->snd_nxt;
1097 * Advance snd_nxt over sequence space of this segment.
1099 if (flags & (TH_SYN | TH_FIN)) {
1102 if (flags & TH_FIN) {
1104 tp->t_flags |= TF_SENTFIN;
1108 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1109 tp->snd_max = tp->snd_nxt;
1111 * Time this transmission if not a retransmission and
1112 * not currently timing anything.
1114 if (tp->t_rtttime == 0) {
1115 tp->t_rtttime = ticks;
1116 tp->t_rtseq = startseq;
1117 tcpstat.tcps_segstimed++;
1122 * Set retransmit timer if not currently set,
1123 * and not doing a pure ack or a keep-alive probe.
1124 * Initial value for retransmit timer is smoothed
1125 * round-trip time + 2 * round-trip time variance.
1126 * Initialize shift counter which is used for backoff
1127 * of retransmit time.
1129 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1130 tp->snd_nxt != tp->snd_una) {
1131 if (tcp_callout_active(tp, tp->tt_persist)) {
1132 tcp_callout_stop(tp, tp->tt_persist);
1135 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1140 * Persist case, update snd_max but since we are in
1141 * persist mode (no window) we do not update snd_nxt.
1145 panic("tcp_output: persist timer to send SYN");
1146 if (flags & TH_FIN) {
1148 tp->t_flags |= TF_SENTFIN;
1150 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1151 tp->snd_max = tp->snd_nxt + xlen;
1157 if (so->so_options & SO_DEBUG) {
1158 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1159 mtod(m, void *), th, 0);
1164 * Fill in IP length and desired time to live and
1165 * send to IP level. There should be a better way
1166 * to handle ttl and tos; we could keep them in
1167 * the template, but need a way to checksum without them.
1170 * m->m_pkthdr.len should have been set before cksum
1171 * calcuration, because in6_cksum() need it.
1175 * we separately set hoplimit for every segment,
1176 * since the user might want to change the value
1177 * via setsockopt. Also, desired default hop
1178 * limit might be changed via Neighbor Discovery.
1180 ip6->ip6_hlim = in6_selecthlim(inp,
1181 (inp->in6p_route.ro_rt ?
1182 inp->in6p_route.ro_rt->rt_ifp : NULL));
1184 /* TODO: IPv6 IP6TOS_ECT bit on */
1185 error = ip6_output(m, inp->in6p_outputopts,
1186 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1190 ip->ip_len = m->m_pkthdr.len;
1192 if (INP_CHECK_SOCKAF(so, AF_INET6))
1193 ip->ip_ttl = in6_selecthlim(inp,
1194 (inp->in6p_route.ro_rt ?
1195 inp->in6p_route.ro_rt->rt_ifp : NULL));
1198 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1200 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1202 * See if we should do MTU discovery.
1203 * We do it only if the following are true:
1204 * 1) we have a valid route to the destination
1205 * 2) the MTU is not locked (if it is,
1206 * then discovery has been disabled)
1208 if (path_mtu_discovery &&
1209 (rt = inp->inp_route.ro_rt) &&
1210 (rt->rt_flags & RTF_UP) &&
1211 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1212 ip->ip_off |= IP_DF;
1214 error = ip_output(m, inp->inp_options, &inp->inp_route,
1215 (so->so_options & SO_DONTROUTE) |
1216 IP_DEBUGROUTE, NULL, inp);
1219 KASSERT(error != 0, ("no error, but th not set"));
1222 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1225 * We know that the packet was lost, so back out the
1226 * sequence number advance, if any.
1228 if (!(tp->t_flags & TF_FORCE) ||
1229 !tcp_callout_active(tp, tp->tt_persist)) {
1231 * No need to check for TH_FIN here because
1232 * the TF_SENTFIN flag handles that case.
1234 if (!(flags & TH_SYN))
1239 if (error == ENOBUFS) {
1241 * If we can't send, make sure there is something
1242 * to get us going again later.
1244 * The persist timer isn't necessarily allowed in all
1245 * states, use the rexmt timer.
1247 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1248 !tcp_callout_active(tp, tp->tt_persist)) {
1249 tcp_callout_reset(tp, tp->tt_rexmt,
1260 if (error == EMSGSIZE) {
1262 * ip_output() will have already fixed the route
1263 * for us. tcp_mtudisc() will, as its last action,
1264 * initiate retransmission, so it is important to
1267 tcp_mtudisc(inp, 0);
1270 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1271 TCPS_HAVERCVDSYN(tp->t_state)) {
1272 tp->t_softerror = error;
1277 tcpstat.tcps_sndtotal++;
1280 * Data sent (as far as we can tell).
1282 * If this advertises a larger window than any other segment,
1283 * then remember the size of the advertised window.
1285 * Any pending ACK has now been sent.
1287 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1288 tp->rcv_adv = tp->rcv_nxt + recvwin;
1289 tp->t_flags &= ~TF_RXRESIZED;
1291 tp->last_ack_sent = tp->rcv_nxt;
1292 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1293 if (tcp_delack_enabled)
1294 tcp_callout_stop(tp, tp->tt_delack);
1296 if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) &&
1297 segcnt >= tcp_fairsend)
1305 tcp_setpersist(struct tcpcb *tp)
1307 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1310 if (tp->t_state == TCPS_SYN_SENT ||
1311 tp->t_state == TCPS_SYN_RECEIVED) {
1312 panic("tcp_setpersist: not established yet, current %s",
1313 tp->t_state == TCPS_SYN_SENT ?
1314 "SYN_SENT" : "SYN_RECEIVED");
1317 if (tcp_callout_active(tp, tp->tt_rexmt))
1318 panic("tcp_setpersist: retransmit pending");
1320 * Start/restart persistance timer.
1322 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1324 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1325 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1330 tcp_idle_cwnd_validate(struct tcpcb *tp)
1332 u_long initial_cwnd = tcp_initial_window(tp);
1335 tcpstat.tcps_sndidle++;
1337 /* According to RFC5681: RW=min(IW,cwnd) */
1338 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1341 u_long idle_time, decay_cwnd;
1344 * RFC2861, but only after idle period.
1348 * Before the congestion window is reduced, ssthresh
1349 * is set to the maximum of its current value and 3/4
1350 * cwnd. If the sender then has more data to send
1351 * than the decayed cwnd allows, the TCP will slow-
1352 * start (perform exponential increase) at least
1353 * half-way back up to the old value of cwnd.
1355 tp->snd_ssthresh = max(tp->snd_ssthresh,
1356 (3 * tp->snd_cwnd) / 4);
1359 * Decay the congestion window by half for every RTT
1360 * that the flow remains inactive.
1362 * The difference between our implementation and
1363 * RFC2861 is that we don't allow cwnd to go below
1364 * the value allowed by RFC5681 (min_cwnd).
1366 idle_time = ticks - tp->snd_last;
1367 decay_cwnd = tp->snd_cwnd;
1368 while (idle_time >= tp->t_rxtcur &&
1369 decay_cwnd > min_cwnd) {
1371 idle_time -= tp->t_rxtcur;
1373 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1376 * Slow-start from scratch to re-determine the send
1377 * congestion window.
1379 tp->snd_cwnd = min_cwnd;
1382 /* Restart ABC counting during congestion avoidance */
1387 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1389 struct inpcb * const inp = tp->t_inpcb;
1391 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1393 const boolean_t isipv6 = FALSE;
1395 unsigned int ipoptlen, optlen;
1398 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1401 ipoptlen = ip6_optlen(inp);
1403 if (inp->inp_options) {
1404 ipoptlen = inp->inp_options->m_len -
1405 offsetof(struct ipoption, ipopt_list);
1411 ipoptlen += ipsec_hdrsiz_tcp(tp);
1416 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1417 (tp->t_flags & TF_RCVD_TSTMP))
1418 optlen += TCPOLEN_TSTAMP_APPA;
1421 if (tp->t_maxopd <= optlen + ipoptlen)
1422 return EHOSTUNREACH;
1424 *segsz = tp->t_maxopd - optlen - ipoptlen;
1430 tcp_output_sched_handler(netmsg_t nmsg)
1432 struct tcpcb *tp = nmsg->lmsg.u.ms_resultp;
1436 lwkt_replymsg(&nmsg->lmsg, 0);
1439 tcp_output_fair(tp);
1443 tcp_output_init(struct tcpcb *tp)
1445 netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE,
1446 tcp_output_sched_handler);
1447 tp->tt_sndmore->lmsg.u.ms_resultp = tp;
1451 tcp_output_cancel(struct tcpcb *tp)
1454 * This message is still pending to be processed;
1455 * drop it. Optimized.
1458 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) {
1459 lwkt_dropmsg(&tp->tt_sndmore->lmsg);
1465 tcp_output_pending(struct tcpcb *tp)
1467 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0)
1474 tcp_output_sched(struct tcpcb *tp)
1477 if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE)
1478 lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg);
1485 * Yield to other senders or receivers on the same netisr if the current
1486 * TCP stream has sent tcp_fairsend segments and is going to burst more
1487 * segments. Bursting large amount of segements in a single TCP stream
1488 * could delay other senders' segments and receivers' ACKs quite a lot,
1489 * if others segments and ACKs are queued on to the same hardware transmit
1490 * queue; thus cause unfairness between senders and suppress receiving
1493 * Fairsend should be performed at the places that do not affect segment
1494 * sending during congestion control, e.g.
1495 * - User requested output
1496 * - ACK input triggered output
1499 * For devices that are TSO capable, their TSO aggregation size limit could
1503 tcp_output_fair(struct tcpcb *tp)
1507 tp->t_flags |= TF_FAIRSEND;
1508 ret = tcp_output(tp);
1509 tp->t_flags &= ~TF_FAIRSEND;