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 static int tcp_idle_cwv = 1;
143 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_cwv, CTLFLAG_RW,
145 "Congestion window validation after idle period (part of RFC2861)");
147 static int tcp_idle_restart = 1;
148 SYSCTL_INT(_net_inet_tcp, OID_AUTO, idle_restart, CTLFLAG_RW,
149 &tcp_idle_restart, 0, "Reset congestion window after idle period");
151 static int tcp_do_tso = 1;
152 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
153 &tcp_do_tso, 0, "Enable TCP Segmentation Offload (TSO)");
155 static int tcp_fairsend = 4;
156 SYSCTL_INT(_net_inet_tcp, OID_AUTO, fairsend, CTLFLAG_RW,
158 "Amount of segments sent before yield to other senders or receivers");
160 static void tcp_idle_cwnd_validate(struct tcpcb *);
162 static int tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen);
163 static void tcp_output_sched(struct tcpcb *tp);
166 * Tcp output routine: figure out what should be sent and send it.
169 tcp_output(struct tcpcb *tp)
171 struct inpcb * const inp = tp->t_inpcb;
172 struct socket *so = inp->inp_socket;
173 long len, recvwin, sendwin;
175 int off, flags, error = 0;
182 u_char opt[TCP_MAXOLEN];
183 unsigned int ipoptlen, optlen, hdrlen;
188 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
190 const boolean_t isipv6 = FALSE;
192 boolean_t can_tso = FALSE, use_tso;
193 boolean_t report_sack, idle_cwv = FALSE;
194 u_int segsz, tso_hlen, tso_lenmax = 0;
196 boolean_t need_sched = FALSE;
198 KKASSERT(so->so_port == &curthread->td_msgport);
201 * Determine length of data that should be transmitted,
202 * and flags that will be used.
203 * If there is some data or critical controls (SYN, RST)
204 * to send, then transmit; otherwise, investigate further.
208 * If we have been idle for a while, the send congestion window
209 * could be no longer representative of the current state of the
210 * link; need to validate congestion window. However, we should
211 * not perform congestion window validation here, since we could
212 * be asked to send pure ACK.
214 if (tp->snd_max == tp->snd_una &&
215 (ticks - tp->snd_last) >= tp->t_rxtcur && tcp_idle_restart)
219 * Calculate whether the transmit stream was previously idle
220 * and adjust TF_LASTIDLE for the next time.
222 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
223 if (idle && (tp->t_flags & TF_MORETOCOME))
224 tp->t_flags |= TF_LASTIDLE;
226 tp->t_flags &= ~TF_LASTIDLE;
228 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
229 !IN_FASTRECOVERY(tp))
230 nsacked = tcp_sack_bytes_below(&tp->scb, tp->snd_nxt);
233 * Find out whether TSO could be used or not
235 * For TSO capable devices, the following assumptions apply to
236 * the processing of TCP flags:
237 * - If FIN is set on the large TCP segment, the device must set
238 * FIN on the last segment that it creates from the large TCP
240 * - If PUSH is set on the large TCP segment, the device must set
241 * PUSH on the last segment that it creates from the large TCP
244 #if !defined(IPSEC) && !defined(FAST_IPSEC)
247 && (tp->t_flags & TF_SIGNATURE) == 0
251 struct rtentry *rt = inp->inp_route.ro_rt;
253 if (rt != NULL && (rt->rt_flags & RTF_UP) &&
254 (rt->rt_ifp->if_hwassist & CSUM_TSO)) {
256 tso_lenmax = rt->rt_ifp->if_tsolen;
260 #endif /* !IPSEC && !FAST_IPSEC */
268 if ((tp->t_flags & (TF_SACK_PERMITTED | TF_NOOPT)) ==
270 (!TAILQ_EMPTY(&tp->t_segq) ||
271 tp->reportblk.rblk_start != tp->reportblk.rblk_end))
276 /* Make use of SACK information when slow-starting after a RTO. */
277 if (TCP_DO_SACK(tp) && tp->snd_nxt != tp->snd_max &&
278 !IN_FASTRECOVERY(tp)) {
279 tcp_seq old_snd_nxt = tp->snd_nxt;
281 tcp_sack_skip_sacked(&tp->scb, &tp->snd_nxt);
282 nsacked += tp->snd_nxt - old_snd_nxt;
286 off = tp->snd_nxt - tp->snd_una;
287 sendwin = min(tp->snd_wnd, tp->snd_cwnd + nsacked);
288 sendwin = min(sendwin, tp->snd_bwnd);
290 flags = tcp_outflags[tp->t_state];
292 * Get standard flags, and add SYN or FIN if requested by 'hidden'
295 if (tp->t_flags & TF_NEEDFIN)
297 if (tp->t_flags & TF_NEEDSYN)
301 * If in persist timeout with window of 0, send 1 byte.
302 * Otherwise, if window is small but nonzero
303 * and timer expired, we will send what we can
304 * and go to transmit state.
306 if (tp->t_flags & TF_FORCE) {
309 * If we still have some data to send, then
310 * clear the FIN bit. Usually this would
311 * happen below when it realizes that we
312 * aren't sending all the data. However,
313 * if we have exactly 1 byte of unsent data,
314 * then it won't clear the FIN bit below,
315 * and if we are in persist state, we wind
316 * up sending the packet without recording
317 * that we sent the FIN bit.
319 * We can't just blindly clear the FIN bit,
320 * because if we don't have any more data
321 * to send then the probe will be the FIN
324 if (off < so->so_snd.ssb_cc)
328 tcp_callout_stop(tp, tp->tt_persist);
334 * If snd_nxt == snd_max and we have transmitted a FIN, the
335 * offset will be > 0 even if so_snd.ssb_cc is 0, resulting in
336 * a negative length. This can also occur when TCP opens up
337 * its congestion window while receiving additional duplicate
338 * acks after fast-retransmit because TCP will reset snd_nxt
339 * to snd_max after the fast-retransmit.
341 * A negative length can also occur when we are in the
342 * TCPS_SYN_RECEIVED state due to a simultanious connect where
343 * our SYN has not been acked yet.
345 * In the normal retransmit-FIN-only case, however, snd_nxt will
346 * be set to snd_una, the offset will be 0, and the length may
349 len = (long)ulmin(so->so_snd.ssb_cc, sendwin) - off;
352 * Lop off SYN bit if it has already been sent. However, if this
353 * is SYN-SENT state and if segment contains data, suppress sending
354 * segment (sending the segment would be an option if we still
355 * did TAO and the remote host supported it).
357 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
360 if (len > 0 && tp->t_state == TCPS_SYN_SENT) {
361 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
367 * Be careful not to send data and/or FIN on SYN segments.
368 * This measure is needed to prevent interoperability problems
369 * with not fully conformant TCP implementations.
371 if (flags & TH_SYN) {
378 * A negative len can occur if our FIN has been sent but not
379 * acked, or if we are in a simultanious connect in the
380 * TCPS_SYN_RECEIVED state with our SYN sent but not yet
383 * If our window has contracted to 0 in the FIN case
384 * (which can only occur if we have NOT been called to
385 * retransmit as per code a few paragraphs up) then we
386 * want to shift the retransmit timer over to the
389 * However, if we are in the TCPS_SYN_RECEIVED state
390 * (the SYN case) we will be in a simultanious connect and
391 * the window may be zero degeneratively. In this case we
392 * do not want to shift to the persist timer after the SYN
393 * or the SYN+ACK transmission.
396 if (sendwin == 0 && tp->t_state != TCPS_SYN_RECEIVED) {
397 tcp_callout_stop(tp, tp->tt_rexmt);
399 tp->snd_nxt = tp->snd_una;
400 if (!tcp_callout_active(tp, tp->tt_persist))
405 KASSERT(len >= 0, ("%s: len < 0", __func__));
407 * Automatic sizing of send socket buffer. Often the send buffer
408 * size is not optimally adjusted to the actual network conditions
409 * at hand (delay bandwidth product). Setting the buffer size too
410 * small limits throughput on links with high bandwidth and high
411 * delay (eg. trans-continental/oceanic links). Setting the
412 * buffer size too big consumes too much real kernel memory,
413 * especially with many connections on busy servers.
415 * The criteria to step up the send buffer one notch are:
416 * 1. receive window of remote host is larger than send buffer
417 * (with a fudge factor of 5/4th);
418 * 2. send buffer is filled to 7/8th with data (so we actually
419 * have data to make use of it);
420 * 3. send buffer fill has not hit maximal automatic size;
421 * 4. our send window (slow start and cogestion controlled) is
422 * larger than sent but unacknowledged data in send buffer.
424 * The remote host receive window scaling factor may limit the
425 * growing of the send buffer before it reaches its allowed
428 * It scales directly with slow start or congestion window
429 * and does at most one step per received ACK. This fast
430 * scaling has the drawback of growing the send buffer beyond
431 * what is strictly necessary to make full use of a given
432 * delay*bandwith product. However testing has shown this not
433 * to be much of an problem. At worst we are trading wasting
434 * of available bandwith (the non-use of it) for wasting some
435 * socket buffer memory.
437 * TODO: Shrink send buffer during idle periods together
438 * with congestion window. Requires another timer. Has to
439 * wait for upcoming tcp timer rewrite.
441 if (tcp_do_autosndbuf && so->so_snd.ssb_flags & SSB_AUTOSIZE) {
442 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.ssb_hiwat &&
443 so->so_snd.ssb_cc >= (so->so_snd.ssb_hiwat / 8 * 7) &&
444 so->so_snd.ssb_cc < tcp_autosndbuf_max &&
445 sendwin >= (so->so_snd.ssb_cc - (tp->snd_nxt - tp->snd_una))) {
448 newsize = ulmin(so->so_snd.ssb_hiwat +
451 if (!ssb_reserve(&so->so_snd, newsize, so, NULL))
452 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
453 if (newsize >= (TCP_MAXWIN << tp->snd_scale))
454 atomic_clear_int(&so->so_snd.ssb_flags, SSB_AUTOSIZE);
460 * - Congestion window needs validation
461 * - There are SACK blocks to report
462 * - RST or SYN flags is set
466 * Checking for SYN|RST looks overkill, just to be safe than sorry
469 if (report_sack || idle_cwv || (flags & (TH_RST | TH_SYN)))
472 tcp_seq ugr_nxt = tp->snd_nxt;
474 if ((flags & TH_FIN) && (tp->t_flags & TF_SENTFIN) &&
475 tp->snd_nxt == tp->snd_max)
478 if (SEQ_GT(tp->snd_up, ugr_nxt))
484 * Find out segment size and header length for TSO
486 error = tcp_tso_getsize(tp, &segsz, &tso_hlen);
491 segsz = tp->t_maxseg;
492 tso_hlen = 0; /* not used */
496 * Truncate to the maximum segment length if not TSO, and ensure that
497 * FIN is removed if the length no longer contains the last data byte.
506 if (__predict_false(tso_lenmax < segsz))
507 tso_lenmax = segsz << 1;
510 * Truncate TSO transfers to (IP_MAXPACKET - iphlen -
511 * thoff), and make sure that we send equal size
512 * transfers down the stack (rather than big-small-
515 len = min(len, tso_lenmax);
516 nsegs = min(len, (IP_MAXPACKET - tso_hlen)) / segsz;
534 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.ssb_cc))
537 recvwin = ssb_space(&so->so_rcv);
540 * Sender silly window avoidance. We transmit under the following
541 * conditions when len is non-zero:
543 * - We have a full segment
544 * - This is the last buffer in a write()/send() and we are
545 * either idle or running NODELAY
546 * - we've timed out (e.g. persist timer)
547 * - we have more then 1/2 the maximum send window's worth of
548 * data (receiver may be limiting the window size)
549 * - we need to retransmit
555 * NOTE! on localhost connections an 'ack' from the remote
556 * end may occur synchronously with the output and cause
557 * us to flush a buffer queued with moretocome. XXX
559 * note: the len + off check is almost certainly unnecessary.
561 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
562 (idle || (tp->t_flags & TF_NODELAY)) &&
563 len + off >= so->so_snd.ssb_cc &&
564 !(tp->t_flags & TF_NOPUSH)) {
567 if (tp->t_flags & TF_FORCE) /* typ. timeout case */
569 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
571 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
573 if (tp->t_flags & TF_XMITNOW)
578 * Compare available window to amount of window
579 * known to peer (as advertised window less
580 * next expected input). If the difference is at least two
581 * max size segments, or at least 50% of the maximum possible
582 * window, then want to send a window update to peer.
586 * "adv" is the amount we can increase the window,
587 * taking into account that we are limited by
588 * TCP_MAXWIN << tp->rcv_scale.
590 long adv = min(recvwin, (long)TCP_MAXWIN << tp->rcv_scale) -
591 (tp->rcv_adv - tp->rcv_nxt);
595 * This ack case typically occurs when the user has drained
596 * the TCP socket buffer sufficiently to warrent an ack
597 * containing a 'pure window update'... that is, an ack that
598 * ONLY updates the tcp window.
600 * It is unclear why we would need to do a pure window update
601 * past 2 segments if we are going to do one at 1/2 the high
602 * water mark anyway, especially since under normal conditions
603 * the user program will drain the socket buffer quickly.
604 * The 2-segment pure window update will often add a large
605 * number of extra, unnecessary acks to the stream.
607 * avoid_pure_win_update now defaults to 1.
609 if (avoid_pure_win_update == 0 ||
610 (tp->t_flags & TF_RXRESIZED)) {
611 if (adv >= (long) (2 * segsz)) {
615 hiwat = (long)(TCP_MAXWIN << tp->rcv_scale);
616 if (hiwat > (long)so->so_rcv.ssb_hiwat)
617 hiwat = (long)so->so_rcv.ssb_hiwat;
618 if (adv >= hiwat / 2)
623 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
624 * is also a catch-all for the retransmit timer timeout case.
626 if (tp->t_flags & TF_ACKNOW)
628 if ((flags & TH_RST) ||
629 ((flags & TH_SYN) && !(tp->t_flags & TF_NEEDSYN)))
631 if (SEQ_GT(tp->snd_up, tp->snd_una))
634 * If our state indicates that FIN should be sent
635 * and we have not yet done so, then we need to send.
637 if ((flags & TH_FIN) &&
638 (!(tp->t_flags & TF_SENTFIN) || tp->snd_nxt == tp->snd_una))
642 * TCP window updates are not reliable, rather a polling protocol
643 * using ``persist'' packets is used to insure receipt of window
644 * updates. The three ``states'' for the output side are:
645 * idle not doing retransmits or persists
646 * persisting to move a small or zero window
647 * (re)transmitting and thereby not persisting
649 * tcp_callout_active(tp, tp->tt_persist)
650 * is true when we are in persist state.
651 * The TF_FORCE flag in tp->t_flags
652 * is set when we are called to send a persist packet.
653 * tcp_callout_active(tp, tp->tt_rexmt)
654 * is set when we are retransmitting
655 * The output side is idle when both timers are zero.
657 * If send window is too small, there is data to transmit, and no
658 * retransmit or persist is pending, then go to persist state.
660 * If nothing happens soon, send when timer expires:
661 * if window is nonzero, transmit what we can, otherwise force out
664 * Don't try to set the persist state if we are in TCPS_SYN_RECEIVED
665 * with data pending. This situation can occur during a
666 * simultanious connect.
668 if (so->so_snd.ssb_cc > 0 &&
669 tp->t_state != TCPS_SYN_RECEIVED &&
670 !tcp_callout_active(tp, tp->tt_rexmt) &&
671 !tcp_callout_active(tp, tp->tt_persist)) {
677 * No reason to send a segment, just return.
679 tp->t_flags &= ~TF_XMITNOW;
683 if (need_sched && len > 0) {
684 tcp_output_sched(tp);
689 * Before ESTABLISHED, force sending of initial options
690 * unless TCP set not to do any options.
691 * NOTE: we assume that the IP/TCP header plus TCP options
692 * always fit in a single mbuf, leaving room for a maximum
694 * max_linkhdr + sizeof(struct tcpiphdr) + optlen <= MCLBYTES
698 hdrlen = sizeof(struct ip6_hdr) + sizeof(struct tcphdr);
700 hdrlen = sizeof(struct tcpiphdr);
701 if (flags & TH_SYN) {
702 tp->snd_nxt = tp->iss;
703 if (!(tp->t_flags & TF_NOOPT)) {
706 opt[0] = TCPOPT_MAXSEG;
707 opt[1] = TCPOLEN_MAXSEG;
708 mss = htons((u_short) tcp_mssopt(tp));
709 memcpy(opt + 2, &mss, sizeof mss);
710 optlen = TCPOLEN_MAXSEG;
712 if ((tp->t_flags & TF_REQ_SCALE) &&
713 (!(flags & TH_ACK) ||
714 (tp->t_flags & TF_RCVD_SCALE))) {
715 *((u_int32_t *)(opt + optlen)) = htonl(
717 TCPOPT_WINDOW << 16 |
718 TCPOLEN_WINDOW << 8 |
719 tp->request_r_scale);
723 if ((tcp_do_sack && !(flags & TH_ACK)) ||
724 tp->t_flags & TF_SACK_PERMITTED) {
725 uint32_t *lp = (uint32_t *)(opt + optlen);
727 *lp = htonl(TCPOPT_SACK_PERMITTED_ALIGNED);
728 optlen += TCPOLEN_SACK_PERMITTED_ALIGNED;
734 * Send a timestamp and echo-reply if this is a SYN and our side
735 * wants to use timestamps (TF_REQ_TSTMP is set) or both our side
736 * and our peer have sent timestamps in our SYN's.
738 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
740 (!(flags & TH_ACK) || (tp->t_flags & TF_RCVD_TSTMP))) {
741 u_int32_t *lp = (u_int32_t *)(opt + optlen);
743 /* Form timestamp option as shown in appendix A of RFC 1323. */
744 *lp++ = htonl(TCPOPT_TSTAMP_HDR);
745 *lp++ = htonl(ticks);
746 *lp = htonl(tp->ts_recent);
747 optlen += TCPOLEN_TSTAMP_APPA;
750 /* Set receive buffer autosizing timestamp. */
751 if (tp->rfbuf_ts == 0 && (so->so_rcv.ssb_flags & SSB_AUTOSIZE))
752 tp->rfbuf_ts = ticks;
755 * If this is a SACK connection and we have a block to report,
756 * fill in the SACK blocks in the TCP options.
759 tcp_sack_fill_report(tp, opt, &optlen);
762 if (tp->t_flags & TF_SIGNATURE) {
766 * Initialize TCP-MD5 option (RFC2385)
768 bp = (u_char *)opt + optlen;
769 *bp++ = TCPOPT_SIGNATURE;
770 *bp++ = TCPOLEN_SIGNATURE;
772 for (i = 0; i < TCP_SIGLEN; i++)
774 optlen += TCPOLEN_SIGNATURE;
776 * Terminate options list and maintain 32-bit alignment.
782 #endif /* TCP_SIGNATURE */
783 KASSERT(optlen <= TCP_MAXOLEN, ("too many TCP options"));
787 ipoptlen = ip6_optlen(inp);
789 if (inp->inp_options) {
790 ipoptlen = inp->inp_options->m_len -
791 offsetof(struct ipoption, ipopt_list);
797 ipoptlen += ipsec_hdrsiz_tcp(tp);
801 /* TSO segment length must be multiple of segment size */
802 KASSERT(len >= (2 * segsz) && (len % segsz == 0),
803 ("invalid TSO len %ld, segsz %u", len, segsz));
805 KASSERT(len <= segsz,
806 ("invalid len %ld, segsz %u", len, segsz));
809 * Adjust data length if insertion of options will bump
810 * the packet length beyond the t_maxopd length. Clear
811 * FIN to prevent premature closure since there is still
812 * more data to send after this (now truncated) packet.
814 * If just the options do not fit we are in a no-win
815 * situation and we treat it as an unreachable host.
817 if (len + optlen + ipoptlen > tp->t_maxopd) {
818 if (tp->t_maxopd <= optlen + ipoptlen) {
819 static time_t last_optlen_report;
821 if (last_optlen_report != time_second) {
822 last_optlen_report = time_second;
823 kprintf("tcpcb %p: MSS (%d) too "
824 "small to hold options!\n",
827 error = EHOSTUNREACH;
831 len = tp->t_maxopd - optlen - ipoptlen;
838 KASSERT(max_linkhdr + hdrlen <= MCLBYTES, ("tcphdr too big"));
840 KASSERT(max_linkhdr + hdrlen <= MHLEN, ("tcphdr too big"));
844 * Grab a header mbuf, attaching a copy of data to
845 * be transmitted, and initialize the header from
846 * the template for sends on this connection.
849 if ((tp->t_flags & TF_FORCE) && len == 1)
850 tcpstat.tcps_sndprobe++;
851 else if (SEQ_LT(tp->snd_nxt, tp->snd_max)) {
852 if (tp->snd_nxt == tp->snd_una)
853 tp->snd_max_rexmt = tp->snd_max;
855 tcpstat.tcps_sndsackrtopack++;
856 tcpstat.tcps_sndsackrtobyte += len;
858 tcpstat.tcps_sndrexmitpack++;
859 tcpstat.tcps_sndrexmitbyte += len;
861 tcpstat.tcps_sndpack++;
862 tcpstat.tcps_sndbyte += len;
866 tcp_idle_cwnd_validate(tp);
868 /* Update last send time after CWV */
869 tp->snd_last = ticks;
871 if ((m = m_copypack(so->so_snd.ssb_mb, off, (int)len,
872 max_linkhdr + hdrlen)) == NULL) {
877 * m_copypack left space for our hdr; use it.
883 m = m_gethdr(MB_DONTWAIT, MT_HEADER);
885 m = m_getl(hdrlen + max_linkhdr, MB_DONTWAIT, MT_HEADER,
892 m->m_data += max_linkhdr;
894 if (len <= MHLEN - hdrlen - max_linkhdr) {
895 m_copydata(so->so_snd.ssb_mb, off, (int) len,
896 mtod(m, caddr_t) + hdrlen);
899 m->m_next = m_copy(so->so_snd.ssb_mb, off, (int) len);
900 if (m->m_next == NULL) {
909 * If we're sending everything we've got, set PUSH.
910 * (This will keep happy those implementations which only
911 * give data to the user when a buffer fills or
914 if (off + len == so->so_snd.ssb_cc)
917 if (tp->t_flags & TF_ACKNOW)
918 tcpstat.tcps_sndacks++;
919 else if (flags & (TH_SYN | TH_FIN | TH_RST))
920 tcpstat.tcps_sndctrl++;
921 else if (SEQ_GT(tp->snd_up, tp->snd_una))
922 tcpstat.tcps_sndurg++;
924 tcpstat.tcps_sndwinup++;
926 MGETHDR(m, MB_DONTWAIT, MT_HEADER);
932 (hdrlen + max_linkhdr > MHLEN) && hdrlen <= MHLEN)
935 m->m_data += max_linkhdr;
938 m->m_pkthdr.rcvif = NULL;
940 ip6 = mtod(m, struct ip6_hdr *);
941 th = (struct tcphdr *)(ip6 + 1);
942 tcp_fillheaders(tp, ip6, th, use_tso);
944 ip = mtod(m, struct ip *);
945 th = (struct tcphdr *)(ip + 1);
946 /* this picks up the pseudo header (w/o the length) */
947 tcp_fillheaders(tp, ip, th, use_tso);
951 * Fill in fields, remembering maximum advertised
952 * window for use in delaying messages about window sizes.
953 * If resending a FIN, be sure not to use a new sequence number.
955 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
956 tp->snd_nxt == tp->snd_max)
961 * If we are doing retransmissions, then snd_nxt will
962 * not reflect the first unsent octet. For ACK only
963 * packets, we do not want the sequence number of the
964 * retransmitted packet, we want the sequence number
965 * of the next unsent octet. So, if there is no data
966 * (and no SYN or FIN), use snd_max instead of snd_nxt
967 * when filling in ti_seq. But if we are in persist
968 * state, snd_max might reflect one byte beyond the
969 * right edge of the window, so use snd_nxt in that
970 * case, since we know we aren't doing a retransmission.
971 * (retransmit and persist are mutually exclusive...)
973 if (len || (flags & (TH_SYN|TH_FIN)) ||
974 tcp_callout_active(tp, tp->tt_persist))
975 th->th_seq = htonl(tp->snd_nxt);
977 th->th_seq = htonl(tp->snd_max);
978 th->th_ack = htonl(tp->rcv_nxt);
980 bcopy(opt, th + 1, optlen);
981 th->th_off = (sizeof(struct tcphdr) + optlen) >> 2;
983 th->th_flags = flags;
987 * Calculate receive window. Don't shrink window, but avoid
988 * silly window syndrome by sending a 0 window if the actual
989 * window is less then one segment.
991 if (recvwin < (long)(so->so_rcv.ssb_hiwat / 4) &&
992 recvwin < (long)segsz)
994 if (recvwin < (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt))
995 recvwin = (tcp_seq_diff_t)(tp->rcv_adv - tp->rcv_nxt);
996 if (recvwin > (long)TCP_MAXWIN << tp->rcv_scale)
997 recvwin = (long)TCP_MAXWIN << tp->rcv_scale;
1000 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1001 * a 0 window. This may cause the remote transmitter to stall. This
1002 * flag tells soreceive() to disable delayed acknowledgements when
1003 * draining the buffer. This can occur if the receiver is attempting
1004 * to read more data then can be buffered prior to transmitting on
1008 tp->t_flags |= TF_RXWIN0SENT;
1010 tp->t_flags &= ~TF_RXWIN0SENT;
1013 th->th_win = htons((u_short) (recvwin>>tp->rcv_scale));
1015 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1016 KASSERT(!use_tso, ("URG with TSO"));
1018 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1019 th->th_flags |= TH_URG;
1023 * If no urgent pointer to send, then we pull
1024 * the urgent pointer to the left edge of the send window
1025 * so that it doesn't drift into the send window on sequence
1026 * number wraparound.
1028 tp->snd_up = tp->snd_una; /* drag it along */
1032 #ifdef TCP_SIGNATURE
1033 if (tp->t_flags & TF_SIGNATURE) {
1034 tcpsignature_compute(m, len, optlen,
1035 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1037 #endif /* TCP_SIGNATURE */
1040 * Put TCP length in extended header, and then
1041 * checksum extended header and data.
1043 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1046 * ip6_plen is not need to be filled now, and will be
1047 * filled in ip6_output().
1049 th->th_sum = in6_cksum(m, IPPROTO_TCP,
1050 sizeof(struct ip6_hdr),
1051 sizeof(struct tcphdr) + optlen + len);
1053 m->m_pkthdr.csum_thlen = sizeof(struct tcphdr) + optlen;
1055 m->m_pkthdr.csum_flags = CSUM_TSO;
1056 m->m_pkthdr.tso_segsz = segsz;
1058 m->m_pkthdr.csum_flags = CSUM_TCP;
1059 m->m_pkthdr.csum_data =
1060 offsetof(struct tcphdr, th_sum);
1062 th->th_sum = in_addword(th->th_sum,
1063 htons((u_short)(optlen + len)));
1068 * IP version must be set here for ipv4/ipv6 checking
1071 KASSERT(ip->ip_v == IPVERSION,
1072 ("%s: IP version incorrect: %d",
1073 __func__, ip->ip_v));
1078 * In transmit state, time the transmission and arrange for
1079 * the retransmit. In persist state, just set snd_max.
1081 if (!(tp->t_flags & TF_FORCE) ||
1082 !tcp_callout_active(tp, tp->tt_persist)) {
1083 tcp_seq startseq = tp->snd_nxt;
1086 * Advance snd_nxt over sequence space of this segment.
1088 if (flags & (TH_SYN | TH_FIN)) {
1091 if (flags & TH_FIN) {
1093 tp->t_flags |= TF_SENTFIN;
1097 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1098 tp->snd_max = tp->snd_nxt;
1100 * Time this transmission if not a retransmission and
1101 * not currently timing anything.
1103 if (tp->t_rtttime == 0) {
1104 tp->t_rtttime = ticks;
1105 tp->t_rtseq = startseq;
1106 tcpstat.tcps_segstimed++;
1111 * Set retransmit timer if not currently set,
1112 * and not doing a pure ack or a keep-alive probe.
1113 * Initial value for retransmit timer is smoothed
1114 * round-trip time + 2 * round-trip time variance.
1115 * Initialize shift counter which is used for backoff
1116 * of retransmit time.
1118 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1119 tp->snd_nxt != tp->snd_una) {
1120 if (tcp_callout_active(tp, tp->tt_persist)) {
1121 tcp_callout_stop(tp, tp->tt_persist);
1124 tcp_callout_reset(tp, tp->tt_rexmt, tp->t_rxtcur,
1129 * Persist case, update snd_max but since we are in
1130 * persist mode (no window) we do not update snd_nxt.
1134 panic("tcp_output: persist timer to send SYN");
1135 if (flags & TH_FIN) {
1137 tp->t_flags |= TF_SENTFIN;
1139 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1140 tp->snd_max = tp->snd_nxt + xlen;
1146 if (so->so_options & SO_DEBUG) {
1147 tcp_trace(TA_OUTPUT, tp->t_state, tp,
1148 mtod(m, void *), th, 0);
1153 * Fill in IP length and desired time to live and
1154 * send to IP level. There should be a better way
1155 * to handle ttl and tos; we could keep them in
1156 * the template, but need a way to checksum without them.
1159 * m->m_pkthdr.len should have been set before cksum
1160 * calcuration, because in6_cksum() need it.
1164 * we separately set hoplimit for every segment,
1165 * since the user might want to change the value
1166 * via setsockopt. Also, desired default hop
1167 * limit might be changed via Neighbor Discovery.
1169 ip6->ip6_hlim = in6_selecthlim(inp,
1170 (inp->in6p_route.ro_rt ?
1171 inp->in6p_route.ro_rt->rt_ifp : NULL));
1173 /* TODO: IPv6 IP6TOS_ECT bit on */
1174 error = ip6_output(m, inp->in6p_outputopts,
1175 &inp->in6p_route, (so->so_options & SO_DONTROUTE),
1179 ip->ip_len = m->m_pkthdr.len;
1181 if (INP_CHECK_SOCKAF(so, AF_INET6))
1182 ip->ip_ttl = in6_selecthlim(inp,
1183 (inp->in6p_route.ro_rt ?
1184 inp->in6p_route.ro_rt->rt_ifp : NULL));
1187 ip->ip_ttl = inp->inp_ip_ttl; /* XXX */
1189 ip->ip_tos = inp->inp_ip_tos; /* XXX */
1191 * See if we should do MTU discovery.
1192 * We do it only if the following are true:
1193 * 1) we have a valid route to the destination
1194 * 2) the MTU is not locked (if it is,
1195 * then discovery has been disabled)
1197 if (path_mtu_discovery &&
1198 (rt = inp->inp_route.ro_rt) &&
1199 (rt->rt_flags & RTF_UP) &&
1200 !(rt->rt_rmx.rmx_locks & RTV_MTU))
1201 ip->ip_off |= IP_DF;
1203 error = ip_output(m, inp->inp_options, &inp->inp_route,
1204 (so->so_options & SO_DONTROUTE) |
1205 IP_DEBUGROUTE, NULL, inp);
1208 KASSERT(error != 0, ("no error, but th not set"));
1211 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1214 * We know that the packet was lost, so back out the
1215 * sequence number advance, if any.
1217 if (!(tp->t_flags & TF_FORCE) ||
1218 !tcp_callout_active(tp, tp->tt_persist)) {
1220 * No need to check for TH_FIN here because
1221 * the TF_SENTFIN flag handles that case.
1223 if (!(flags & TH_SYN))
1228 if (error == ENOBUFS) {
1230 * If we can't send, make sure there is something
1231 * to get us going again later.
1233 * The persist timer isn't necessarily allowed in all
1234 * states, use the rexmt timer.
1236 if (!tcp_callout_active(tp, tp->tt_rexmt) &&
1237 !tcp_callout_active(tp, tp->tt_persist)) {
1238 tcp_callout_reset(tp, tp->tt_rexmt,
1249 if (error == EMSGSIZE) {
1251 * ip_output() will have already fixed the route
1252 * for us. tcp_mtudisc() will, as its last action,
1253 * initiate retransmission, so it is important to
1256 tcp_mtudisc(inp, 0);
1259 if ((error == EHOSTUNREACH || error == ENETDOWN) &&
1260 TCPS_HAVERCVDSYN(tp->t_state)) {
1261 tp->t_softerror = error;
1266 tcpstat.tcps_sndtotal++;
1269 * Data sent (as far as we can tell).
1271 * If this advertises a larger window than any other segment,
1272 * then remember the size of the advertised window.
1274 * Any pending ACK has now been sent.
1276 if (recvwin > 0 && SEQ_GT(tp->rcv_nxt + recvwin, tp->rcv_adv)) {
1277 tp->rcv_adv = tp->rcv_nxt + recvwin;
1278 tp->t_flags &= ~TF_RXRESIZED;
1280 tp->last_ack_sent = tp->rcv_nxt;
1281 tp->t_flags &= ~(TF_ACKNOW | TF_XMITNOW);
1282 if (tcp_delack_enabled)
1283 tcp_callout_stop(tp, tp->tt_delack);
1285 if (tcp_fairsend > 0 && (tp->t_flags & TF_FAIRSEND) &&
1286 segcnt >= tcp_fairsend)
1294 tcp_setpersist(struct tcpcb *tp)
1296 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1299 if (tp->t_state == TCPS_SYN_SENT ||
1300 tp->t_state == TCPS_SYN_RECEIVED) {
1301 panic("tcp_setpersist: not established yet, current %s",
1302 tp->t_state == TCPS_SYN_SENT ?
1303 "SYN_SENT" : "SYN_RECEIVED");
1306 if (tcp_callout_active(tp, tp->tt_rexmt))
1307 panic("tcp_setpersist: retransmit pending");
1309 * Start/restart persistance timer.
1311 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift], TCPTV_PERSMIN,
1313 tcp_callout_reset(tp, tp->tt_persist, tt, tcp_timer_persist);
1314 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1319 tcp_idle_cwnd_validate(struct tcpcb *tp)
1321 u_long initial_cwnd = tcp_initial_window(tp);
1324 tcpstat.tcps_sndidle++;
1326 /* According to RFC5681: RW=min(IW,cwnd) */
1327 min_cwnd = min(tp->snd_cwnd, initial_cwnd);
1330 u_long idle_time, decay_cwnd;
1333 * RFC2861, but only after idle period.
1337 * Before the congestion window is reduced, ssthresh
1338 * is set to the maximum of its current value and 3/4
1339 * cwnd. If the sender then has more data to send
1340 * than the decayed cwnd allows, the TCP will slow-
1341 * start (perform exponential increase) at least
1342 * half-way back up to the old value of cwnd.
1344 tp->snd_ssthresh = max(tp->snd_ssthresh,
1345 (3 * tp->snd_cwnd) / 4);
1348 * Decay the congestion window by half for every RTT
1349 * that the flow remains inactive.
1351 * The difference between our implementation and
1352 * RFC2861 is that we don't allow cwnd to go below
1353 * the value allowed by RFC5681 (min_cwnd).
1355 idle_time = ticks - tp->snd_last;
1356 decay_cwnd = tp->snd_cwnd;
1357 while (idle_time >= tp->t_rxtcur &&
1358 decay_cwnd > min_cwnd) {
1360 idle_time -= tp->t_rxtcur;
1362 tp->snd_cwnd = max(decay_cwnd, min_cwnd);
1365 * Slow-start from scratch to re-determine the send
1366 * congestion window.
1368 tp->snd_cwnd = min_cwnd;
1371 /* Restart ABC counting during congestion avoidance */
1376 tcp_tso_getsize(struct tcpcb *tp, u_int *segsz, u_int *hlen0)
1378 struct inpcb * const inp = tp->t_inpcb;
1380 const boolean_t isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
1382 const boolean_t isipv6 = FALSE;
1384 unsigned int ipoptlen, optlen;
1387 hlen = sizeof(struct ip) + sizeof(struct tcphdr);
1390 ipoptlen = ip6_optlen(inp);
1392 if (inp->inp_options) {
1393 ipoptlen = inp->inp_options->m_len -
1394 offsetof(struct ipoption, ipopt_list);
1400 ipoptlen += ipsec_hdrsiz_tcp(tp);
1405 if ((tp->t_flags & (TF_REQ_TSTMP | TF_NOOPT)) == TF_REQ_TSTMP &&
1406 (tp->t_flags & TF_RCVD_TSTMP))
1407 optlen += TCPOLEN_TSTAMP_APPA;
1410 if (tp->t_maxopd <= optlen + ipoptlen)
1411 return EHOSTUNREACH;
1413 *segsz = tp->t_maxopd - optlen - ipoptlen;
1419 tcp_output_sched_handler(netmsg_t nmsg)
1421 struct tcpcb *tp = nmsg->lmsg.u.ms_resultp;
1425 lwkt_replymsg(&nmsg->lmsg, 0);
1428 tcp_output_fair(tp);
1432 tcp_output_init(struct tcpcb *tp)
1434 netmsg_init(tp->tt_sndmore, NULL, &netisr_adone_rport, MSGF_DROPABLE,
1435 tcp_output_sched_handler);
1436 tp->tt_sndmore->lmsg.u.ms_resultp = tp;
1440 tcp_output_cancel(struct tcpcb *tp)
1443 * This message is still pending to be processed;
1444 * drop it. Optimized.
1447 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0) {
1448 lwkt_dropmsg(&tp->tt_sndmore->lmsg);
1454 tcp_output_pending(struct tcpcb *tp)
1456 if ((tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE) == 0)
1463 tcp_output_sched(struct tcpcb *tp)
1466 if (tp->tt_sndmore->lmsg.ms_flags & MSGF_DONE)
1467 lwkt_sendmsg(netisr_cpuport(mycpuid), &tp->tt_sndmore->lmsg);
1474 * Yield to other senders or receivers on the same netisr if the current
1475 * TCP stream has sent tcp_fairsend segments and is going to burst more
1476 * segments. Bursting large amount of segements in a single TCP stream
1477 * could delay other senders' segments and receivers' ACKs quite a lot,
1478 * if others segments and ACKs are queued on to the same hardware transmit
1479 * queue; thus cause unfairness between senders and suppress receiving
1482 * Fairsend should be performed at the places that do not affect segment
1483 * sending during congestion control, e.g.
1484 * - User requested output
1485 * - ACK input triggered output
1488 * For devices that are TSO capable, their TSO aggregation size limit could
1492 tcp_output_fair(struct tcpcb *tp)
1496 tp->t_flags |= TF_FAIRSEND;
1497 ret = tcp_output(tp);
1498 tp->t_flags &= ~TF_FAIRSEND;