2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
39 #include "opt_tcpdebug.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/kernel.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
49 #include <sys/socket.h>
50 #include <sys/socketvar.h>
51 #include <sys/sysctl.h>
53 #include <net/route.h>
55 #include <netinet/in.h>
56 #include <netinet/in_systm.h>
57 #include <netinet/ip.h>
58 #include <netinet/in_pcb.h>
59 #include <netinet/ip_var.h>
60 #include <netinet/ip_options.h>
62 #include <netinet6/in6_pcb.h>
63 #include <netinet/ip6.h>
64 #include <netinet6/ip6_var.h>
66 #include <netinet/tcp.h>
68 #include <netinet/tcp_fsm.h>
69 #include <netinet/tcp_seq.h>
70 #include <netinet/tcp_timer.h>
71 #include <netinet/tcp_var.h>
72 #include <netinet/tcpip.h>
74 #include <netinet/tcp_debug.h>
78 #include <netipsec/ipsec.h>
81 #include <machine/in_cksum.h>
83 #include <security/mac/mac_framework.h>
86 extern struct mbuf *m_copypack();
89 int path_mtu_discovery = 1;
90 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
91 &path_mtu_discovery, 1, "Enable Path MTU Discovery");
94 SYSCTL_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
95 &ss_fltsz, 1, "Slow start flight size");
97 int ss_fltsz_local = 4;
98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize, CTLFLAG_RW,
99 &ss_fltsz_local, 1, "Slow start flight size for local networks");
101 int tcp_do_newreno = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
103 &tcp_do_newreno, 0, "Enable NewReno Algorithms");
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
107 &tcp_do_tso, 0, "Enable TCP Segmentation Offload");
109 int tcp_do_autosndbuf = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
111 &tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
113 int tcp_autosndbuf_inc = 8*1024;
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
115 &tcp_autosndbuf_inc, 0, "Incrementor step size of automatic send buffer");
117 int tcp_autosndbuf_max = 256*1024;
118 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
119 &tcp_autosndbuf_max, 0, "Max size of automatic send buffer");
123 * Tcp output routine: figure out what should be sent and send it.
126 tcp_output(struct tcpcb *tp)
128 struct socket *so = tp->t_inpcb->inp_socket;
129 long len, recwin, sendwin;
130 int off, flags, error;
132 struct ip *ip = NULL;
133 struct ipovly *ipov = NULL;
135 u_char opt[TCP_MAXOLEN];
136 unsigned ipoptlen, optlen, hdrlen;
138 unsigned ipsec_optlen = 0;
141 int sack_rxmit, sack_bytes_rxmt;
146 int maxburst = TCP_MAXBURST;
149 struct ip6_hdr *ip6 = NULL;
152 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
155 INP_WLOCK_ASSERT(tp->t_inpcb);
158 * Determine length of data that should be transmitted,
159 * and flags that will be used.
160 * If there is some data or critical controls (SYN, RST)
161 * to send, then transmit; otherwise, investigate further.
163 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
164 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
166 * We have been idle for "a while" and no acks are
167 * expected to clock out any data we send --
168 * slow start to get ack "clock" running again.
170 * Set the slow-start flight size depending on whether
171 * this is a local network or not.
176 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
180 if (in_localaddr(tp->t_inpcb->inp_faddr))
182 tp->snd_cwnd = tp->t_maxseg * ss;
184 tp->t_flags &= ~TF_LASTIDLE;
186 if (tp->t_flags & TF_MORETOCOME) {
187 tp->t_flags |= TF_LASTIDLE;
193 * If we've recently taken a timeout, snd_max will be greater than
194 * snd_nxt. There may be SACK information that allows us to avoid
195 * resending already delivered data. Adjust snd_nxt accordingly.
197 if ((tp->t_flags & TF_SACK_PERMIT) &&
198 SEQ_LT(tp->snd_nxt, tp->snd_max))
201 off = tp->snd_nxt - tp->snd_una;
202 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
203 sendwin = min(sendwin, tp->snd_bwnd);
205 flags = tcp_outflags[tp->t_state];
207 * Send any SACK-generated retransmissions. If we're explicitly trying
208 * to send out new data (when sendalot is 1), bypass this function.
209 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
210 * we're replacing a (future) new transmission with a retransmission
211 * now, and we previously incremented snd_cwnd in tcp_input().
214 * Still in sack recovery , reset rxmit flag to zero.
220 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
221 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
224 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
227 /* Do not retransmit SACK segments beyond snd_recover */
228 if (SEQ_GT(p->end, tp->snd_recover)) {
230 * (At least) part of sack hole extends beyond
231 * snd_recover. Check to see if we can rexmit data
234 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
236 * Can't rexmit any more data for this hole.
237 * That data will be rexmitted in the next
238 * sack recovery episode, when snd_recover
239 * moves past p->rxmit.
242 goto after_sack_rexmit;
244 /* Can rexmit part of the current hole */
245 len = ((long)ulmin(cwin,
246 tp->snd_recover - p->rxmit));
248 len = ((long)ulmin(cwin, p->end - p->rxmit));
249 off = p->rxmit - tp->snd_una;
250 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
255 tcpstat.tcps_sack_rexmits++;
256 tcpstat.tcps_sack_rexmit_bytes +=
257 min(len, tp->t_maxseg);
262 * Get standard flags, and add SYN or FIN if requested by 'hidden'
265 if (tp->t_flags & TF_NEEDFIN)
267 if (tp->t_flags & TF_NEEDSYN)
270 SOCKBUF_LOCK(&so->so_snd);
272 * If in persist timeout with window of 0, send 1 byte.
273 * Otherwise, if window is small but nonzero
274 * and timer expired, we will send what we can
275 * and go to transmit state.
277 if (tp->t_flags & TF_FORCEDATA) {
280 * If we still have some data to send, then
281 * clear the FIN bit. Usually this would
282 * happen below when it realizes that we
283 * aren't sending all the data. However,
284 * if we have exactly 1 byte of unsent data,
285 * then it won't clear the FIN bit below,
286 * and if we are in persist state, we wind
287 * up sending the packet without recording
288 * that we sent the FIN bit.
290 * We can't just blindly clear the FIN bit,
291 * because if we don't have any more data
292 * to send then the probe will be the FIN
295 if (off < so->so_snd.sb_cc)
299 tcp_timer_activate(tp, TT_PERSIST, 0);
305 * If snd_nxt == snd_max and we have transmitted a FIN, the
306 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
307 * a negative length. This can also occur when TCP opens up
308 * its congestion window while receiving additional duplicate
309 * acks after fast-retransmit because TCP will reset snd_nxt
310 * to snd_max after the fast-retransmit.
312 * In the normal retransmit-FIN-only case, however, snd_nxt will
313 * be set to snd_una, the offset will be 0, and the length may
316 * If sack_rxmit is true we are retransmitting from the scoreboard
317 * in which case len is already set.
319 if (sack_rxmit == 0) {
320 if (sack_bytes_rxmt == 0)
321 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
326 * We are inside of a SACK recovery episode and are
327 * sending new data, having retransmitted all the
328 * data possible in the scoreboard.
330 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
333 * Don't remove this (len > 0) check !
334 * We explicitly check for len > 0 here (although it
335 * isn't really necessary), to work around a gcc
336 * optimization issue - to force gcc to compute
337 * len above. Without this check, the computation
338 * of len is bungled by the optimizer.
341 cwin = tp->snd_cwnd -
342 (tp->snd_nxt - tp->sack_newdata) -
346 len = lmin(len, cwin);
352 * Lop off SYN bit if it has already been sent. However, if this
353 * is SYN-SENT state and if segment contains data and if we don't
354 * know that foreign host supports TAO, suppress sending segment.
356 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
357 if (tp->t_state != TCPS_SYN_RECEIVED)
363 * Be careful not to send data and/or FIN on SYN segments.
364 * This measure is needed to prevent interoperability problems
365 * with not fully conformant TCP implementations.
367 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
374 * If FIN has been sent but not acked,
375 * but we haven't been called to retransmit,
376 * len will be < 0. Otherwise, window shrank
377 * after we sent into it. If window shrank to 0,
378 * cancel pending retransmit, pull snd_nxt back
379 * to (closed) window, and set the persist timer
380 * if it isn't already going. If the window didn't
381 * close completely, just wait for an ACK.
385 tcp_timer_activate(tp, TT_REXMT, 0);
387 tp->snd_nxt = tp->snd_una;
388 if (!tcp_timer_active(tp, TT_PERSIST))
393 /* len will be >= 0 after this point. */
394 KASSERT(len >= 0, ("%s: len < 0", __func__));
397 * Automatic sizing of send socket buffer. Often the send buffer
398 * size is not optimally adjusted to the actual network conditions
399 * at hand (delay bandwidth product). Setting the buffer size too
400 * small limits throughput on links with high bandwidth and high
401 * delay (eg. trans-continental/oceanic links). Setting the
402 * buffer size too big consumes too much real kernel memory,
403 * especially with many connections on busy servers.
405 * The criteria to step up the send buffer one notch are:
406 * 1. receive window of remote host is larger than send buffer
407 * (with a fudge factor of 5/4th);
408 * 2. send buffer is filled to 7/8th with data (so we actually
409 * have data to make use of it);
410 * 3. send buffer fill has not hit maximal automatic size;
411 * 4. our send window (slow start and cogestion controlled) is
412 * larger than sent but unacknowledged data in send buffer.
414 * The remote host receive window scaling factor may limit the
415 * growing of the send buffer before it reaches its allowed
418 * It scales directly with slow start or congestion window
419 * and does at most one step per received ACK. This fast
420 * scaling has the drawback of growing the send buffer beyond
421 * what is strictly necessary to make full use of a given
422 * delay*bandwith product. However testing has shown this not
423 * to be much of an problem. At worst we are trading wasting
424 * of available bandwith (the non-use of it) for wasting some
425 * socket buffer memory.
427 * TODO: Shrink send buffer during idle periods together
428 * with congestion window. Requires another timer. Has to
429 * wait for upcoming tcp timer rewrite.
431 if (tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
432 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
433 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
434 so->so_snd.sb_cc < tcp_autosndbuf_max &&
435 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
436 if (!sbreserve_locked(&so->so_snd,
437 min(so->so_snd.sb_hiwat + tcp_autosndbuf_inc,
438 tcp_autosndbuf_max), so, curthread))
439 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
444 * Truncate to the maximum segment length or enable TCP Segmentation
445 * Offloading (if supported by hardware) and ensure that FIN is removed
446 * if the length no longer contains the last data byte.
448 * TSO may only be used if we are in a pure bulk sending state. The
449 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
450 * IP options prevent using TSO. With TSO the TCP header is the same
451 * (except for the sequence number) for all generated packets. This
452 * makes it impossible to transmit any options which vary per generated
455 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
456 * removal of FIN (if not already catched here) are handled later after
457 * the exact length of the TCP options are known.
461 * Pre-calculate here as we save another lookup into the darknesses
462 * of IPsec that way and can actually decide if TSO is ok.
464 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
466 if (len > tp->t_maxseg) {
467 if ((tp->t_flags & TF_TSO) && tcp_do_tso &&
468 ((tp->t_flags & TF_SIGNATURE) == 0) &&
469 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
470 tp->t_inpcb->inp_options == NULL &&
471 tp->t_inpcb->in6p_options == NULL
484 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
487 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
491 recwin = sbspace(&so->so_rcv);
494 * Sender silly window avoidance. We transmit under the following
495 * conditions when len is non-zero:
497 * - We have a full segment (or more with TSO)
498 * - This is the last buffer in a write()/send() and we are
499 * either idle or running NODELAY
500 * - we've timed out (e.g. persist timer)
501 * - we have more then 1/2 the maximum send window's worth of
502 * data (receiver may be limited the window size)
503 * - we need to retransmit
506 if (len >= tp->t_maxseg)
509 * NOTE! on localhost connections an 'ack' from the remote
510 * end may occur synchronously with the output and cause
511 * us to flush a buffer queued with moretocome. XXX
513 * note: the len + off check is almost certainly unnecessary.
515 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
516 (idle || (tp->t_flags & TF_NODELAY)) &&
517 len + off >= so->so_snd.sb_cc &&
518 (tp->t_flags & TF_NOPUSH) == 0) {
521 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
523 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
525 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
532 * Compare available window to amount of window
533 * known to peer (as advertised window less
534 * next expected input). If the difference is at least two
535 * max size segments, or at least 50% of the maximum possible
536 * window, then want to send a window update to peer.
537 * Skip this if the connection is in T/TCP half-open state.
538 * Don't send pure window updates when the peer has closed
539 * the connection and won't ever send more data.
541 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
542 !TCPS_HAVERCVDFIN(tp->t_state)) {
544 * "adv" is the amount we can increase the window,
545 * taking into account that we are limited by
546 * TCP_MAXWIN << tp->rcv_scale.
548 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
549 (tp->rcv_adv - tp->rcv_nxt);
551 if (adv >= (long) (2 * tp->t_maxseg))
553 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
558 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
559 * is also a catch-all for the retransmit timer timeout case.
561 if (tp->t_flags & TF_ACKNOW)
563 if ((flags & TH_RST) ||
564 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
566 if (SEQ_GT(tp->snd_up, tp->snd_una))
569 * If our state indicates that FIN should be sent
570 * and we have not yet done so, then we need to send.
572 if (flags & TH_FIN &&
573 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
576 * In SACK, it is possible for tcp_output to fail to send a segment
577 * after the retransmission timer has been turned off. Make sure
578 * that the retransmission timer is set.
580 if ((tp->t_flags & TF_SACK_PERMIT) &&
581 SEQ_GT(tp->snd_max, tp->snd_una) &&
582 !tcp_timer_active(tp, TT_REXMT) &&
583 !tcp_timer_active(tp, TT_PERSIST)) {
584 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
588 * TCP window updates are not reliable, rather a polling protocol
589 * using ``persist'' packets is used to insure receipt of window
590 * updates. The three ``states'' for the output side are:
591 * idle not doing retransmits or persists
592 * persisting to move a small or zero window
593 * (re)transmitting and thereby not persisting
595 * tcp_timer_active(tp, TT_PERSIST)
596 * is true when we are in persist state.
597 * (tp->t_flags & TF_FORCEDATA)
598 * is set when we are called to send a persist packet.
599 * tcp_timer_active(tp, TT_REXMT)
600 * is set when we are retransmitting
601 * The output side is idle when both timers are zero.
603 * If send window is too small, there is data to transmit, and no
604 * retransmit or persist is pending, then go to persist state.
605 * If nothing happens soon, send when timer expires:
606 * if window is nonzero, transmit what we can,
607 * otherwise force out a byte.
609 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
610 !tcp_timer_active(tp, TT_PERSIST)) {
616 * No reason to send a segment, just return.
619 SOCKBUF_UNLOCK(&so->so_snd);
623 SOCKBUF_LOCK_ASSERT(&so->so_snd);
625 * Before ESTABLISHED, force sending of initial options
626 * unless TCP set not to do any options.
627 * NOTE: we assume that the IP/TCP header plus TCP options
628 * always fit in a single mbuf, leaving room for a maximum
630 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
635 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
638 hdrlen = sizeof (struct tcpiphdr);
641 * Compute options for segment.
642 * We only have to care about SYN and established connection
643 * segments. Options for SYN-ACK segments are handled in TCP
646 if ((tp->t_flags & TF_NOOPT) == 0) {
648 /* Maximum segment size. */
649 if (flags & TH_SYN) {
650 tp->snd_nxt = tp->iss;
651 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
652 to.to_flags |= TOF_MSS;
654 /* Window scaling. */
655 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
656 to.to_wscale = tp->request_r_scale;
657 to.to_flags |= TOF_SCALE;
660 if ((tp->t_flags & TF_RCVD_TSTMP) ||
661 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
662 to.to_tsval = ticks + tp->ts_offset;
663 to.to_tsecr = tp->ts_recent;
664 to.to_flags |= TOF_TS;
665 /* Set receive buffer autosizing timestamp. */
666 if (tp->rfbuf_ts == 0 &&
667 (so->so_rcv.sb_flags & SB_AUTOSIZE))
668 tp->rfbuf_ts = ticks;
670 /* Selective ACK's. */
671 if (tp->t_flags & TF_SACK_PERMIT) {
673 to.to_flags |= TOF_SACKPERM;
674 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
675 (tp->t_flags & TF_SACK_PERMIT) &&
676 tp->rcv_numsacks > 0) {
677 to.to_flags |= TOF_SACK;
678 to.to_nsacks = tp->rcv_numsacks;
679 to.to_sacks = (u_char *)tp->sackblks;
683 /* TCP-MD5 (RFC2385). */
684 if (tp->t_flags & TF_SIGNATURE)
685 to.to_flags |= TOF_SIGNATURE;
686 #endif /* TCP_SIGNATURE */
688 /* Processing the options. */
689 hdrlen += optlen = tcp_addoptions(&to, opt);
694 ipoptlen = ip6_optlen(tp->t_inpcb);
697 if (tp->t_inpcb->inp_options)
698 ipoptlen = tp->t_inpcb->inp_options->m_len -
699 offsetof(struct ipoption, ipopt_list);
703 ipoptlen += ipsec_optlen;
707 * Adjust data length if insertion of options will
708 * bump the packet length beyond the t_maxopd length.
709 * Clear the FIN bit because we cut off the tail of
712 * When doing TSO limit a burst to TCP_MAXWIN minus the
713 * IP, TCP and Options length to keep ip->ip_len from
714 * overflowing. Prevent the last segment from being
715 * fractional thus making them all equal sized and set
716 * the flag to continue sending. TSO is disabled when
717 * IP options or IPSEC are present.
719 if (len + optlen + ipoptlen > tp->t_maxopd) {
722 if (len > TCP_MAXWIN - hdrlen - optlen) {
723 len = TCP_MAXWIN - hdrlen - optlen;
724 len = len - (len % (tp->t_maxopd - optlen));
726 } else if (tp->t_flags & TF_NEEDFIN)
729 len = tp->t_maxopd - optlen - ipoptlen;
734 /*#ifdef DIAGNOSTIC*/
736 if (max_linkhdr + hdrlen > MCLBYTES)
738 if (max_linkhdr + hdrlen > MHLEN)
740 panic("tcphdr too big");
744 * Grab a header mbuf, attaching a copy of data to
745 * be transmitted, and initialize the header from
746 * the template for sends on this connection.
752 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
753 tcpstat.tcps_sndprobe++;
754 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
755 tcpstat.tcps_sndrexmitpack++;
756 tcpstat.tcps_sndrexmitbyte += len;
758 tcpstat.tcps_sndpack++;
759 tcpstat.tcps_sndbyte += len;
762 if ((m = m_copypack(so->so_snd.sb_mb, off,
763 (int)len, max_linkhdr + hdrlen)) == 0) {
764 SOCKBUF_UNLOCK(&so->so_snd);
769 * m_copypack left space for our hdr; use it.
774 MGETHDR(m, M_DONTWAIT, MT_DATA);
776 SOCKBUF_UNLOCK(&so->so_snd);
781 if (MHLEN < hdrlen + max_linkhdr) {
782 MCLGET(m, M_DONTWAIT);
783 if ((m->m_flags & M_EXT) == 0) {
784 SOCKBUF_UNLOCK(&so->so_snd);
791 m->m_data += max_linkhdr;
795 * Start the m_copy functions from the closest mbuf
796 * to the offset in the socket buffer chain.
798 mb = sbsndptr(&so->so_snd, off, len, &moff);
800 if (len <= MHLEN - hdrlen - max_linkhdr) {
801 m_copydata(mb, moff, (int)len,
802 mtod(m, caddr_t) + hdrlen);
805 m->m_next = m_copy(mb, moff, (int)len);
806 if (m->m_next == NULL) {
807 SOCKBUF_UNLOCK(&so->so_snd);
815 * If we're sending everything we've got, set PUSH.
816 * (This will keep happy those implementations which only
817 * give data to the user when a buffer fills or
820 if (off + len == so->so_snd.sb_cc)
822 SOCKBUF_UNLOCK(&so->so_snd);
824 SOCKBUF_UNLOCK(&so->so_snd);
825 if (tp->t_flags & TF_ACKNOW)
826 tcpstat.tcps_sndacks++;
827 else if (flags & (TH_SYN|TH_FIN|TH_RST))
828 tcpstat.tcps_sndctrl++;
829 else if (SEQ_GT(tp->snd_up, tp->snd_una))
830 tcpstat.tcps_sndurg++;
832 tcpstat.tcps_sndwinup++;
834 MGETHDR(m, M_DONTWAIT, MT_DATA);
840 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
845 m->m_data += max_linkhdr;
848 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
849 m->m_pkthdr.rcvif = (struct ifnet *)0;
851 mac_create_mbuf_from_inpcb(tp->t_inpcb, m);
855 ip6 = mtod(m, struct ip6_hdr *);
856 th = (struct tcphdr *)(ip6 + 1);
857 tcpip_fillheaders(tp->t_inpcb, ip6, th);
861 ip = mtod(m, struct ip *);
862 ipov = (struct ipovly *)ip;
863 th = (struct tcphdr *)(ip + 1);
864 tcpip_fillheaders(tp->t_inpcb, ip, th);
868 * Fill in fields, remembering maximum advertised
869 * window for use in delaying messages about window sizes.
870 * If resending a FIN, be sure not to use a new sequence number.
872 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
873 tp->snd_nxt == tp->snd_max)
876 * If we are doing retransmissions, then snd_nxt will
877 * not reflect the first unsent octet. For ACK only
878 * packets, we do not want the sequence number of the
879 * retransmitted packet, we want the sequence number
880 * of the next unsent octet. So, if there is no data
881 * (and no SYN or FIN), use snd_max instead of snd_nxt
882 * when filling in ti_seq. But if we are in persist
883 * state, snd_max might reflect one byte beyond the
884 * right edge of the window, so use snd_nxt in that
885 * case, since we know we aren't doing a retransmission.
886 * (retransmit and persist are mutually exclusive...)
888 if (sack_rxmit == 0) {
889 if (len || (flags & (TH_SYN|TH_FIN)) ||
890 tcp_timer_active(tp, TT_PERSIST))
891 th->th_seq = htonl(tp->snd_nxt);
893 th->th_seq = htonl(tp->snd_max);
895 th->th_seq = htonl(p->rxmit);
897 tp->sackhint.sack_bytes_rexmit += len;
899 th->th_ack = htonl(tp->rcv_nxt);
901 bcopy(opt, th + 1, optlen);
902 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
904 th->th_flags = flags;
906 * Calculate receive window. Don't shrink window,
907 * but avoid silly window syndrome.
909 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
910 recwin < (long)tp->t_maxseg)
912 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
913 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
914 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
915 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
918 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
919 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
920 * case is handled in syncache.
923 th->th_win = htons((u_short)
924 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
926 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
929 * Adjust the RXWIN0SENT flag - indicate that we have advertised
930 * a 0 window. This may cause the remote transmitter to stall. This
931 * flag tells soreceive() to disable delayed acknowledgements when
932 * draining the buffer. This can occur if the receiver is attempting
933 * to read more data then can be buffered prior to transmitting on
937 tp->t_flags |= TF_RXWIN0SENT;
939 tp->t_flags &= ~TF_RXWIN0SENT;
940 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
941 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
942 th->th_flags |= TH_URG;
945 * If no urgent pointer to send, then we pull
946 * the urgent pointer to the left edge of the send window
947 * so that it doesn't drift into the send window on sequence
950 tp->snd_up = tp->snd_una; /* drag it along */
953 if (tp->t_flags & TF_SIGNATURE) {
954 int sigoff = to.to_signature - opt;
955 tcp_signature_compute(m, 0, len, optlen,
956 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
961 * Put TCP length in extended header, and then
962 * checksum extended header and data.
964 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
968 * ip6_plen is not need to be filled now, and will be filled
971 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
972 sizeof(struct tcphdr) + optlen + len);
976 m->m_pkthdr.csum_flags = CSUM_TCP;
977 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
978 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
979 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
981 /* IP version must be set here for ipv4/ipv6 checking later */
982 KASSERT(ip->ip_v == IPVERSION,
983 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
987 * Enable TSO and specify the size of the segments.
988 * The TCP pseudo header checksum is always provided.
989 * XXX: Fixme: This is currently not the case for IPv6.
992 m->m_pkthdr.csum_flags = CSUM_TSO;
993 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
997 * In transmit state, time the transmission and arrange for
998 * the retransmit. In persist state, just set snd_max.
1000 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1001 !tcp_timer_active(tp, TT_PERSIST)) {
1002 tcp_seq startseq = tp->snd_nxt;
1005 * Advance snd_nxt over sequence space of this segment.
1007 if (flags & (TH_SYN|TH_FIN)) {
1010 if (flags & TH_FIN) {
1012 tp->t_flags |= TF_SENTFIN;
1018 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1019 tp->snd_max = tp->snd_nxt;
1021 * Time this transmission if not a retransmission and
1022 * not currently timing anything.
1024 if (tp->t_rtttime == 0) {
1025 tp->t_rtttime = ticks;
1026 tp->t_rtseq = startseq;
1027 tcpstat.tcps_segstimed++;
1032 * Set retransmit timer if not currently set,
1033 * and not doing a pure ack or a keep-alive probe.
1034 * Initial value for retransmit timer is smoothed
1035 * round-trip time + 2 * round-trip time variance.
1036 * Initialize shift counter which is used for backoff
1037 * of retransmit time.
1040 if (!tcp_timer_active(tp, TT_REXMT) &&
1041 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1042 (tp->snd_nxt != tp->snd_una))) {
1043 if (tcp_timer_active(tp, TT_PERSIST)) {
1044 tcp_timer_activate(tp, TT_PERSIST, 0);
1047 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1051 * Persist case, update snd_max but since we are in
1052 * persist mode (no window) we do not update snd_nxt.
1057 if (flags & TH_FIN) {
1059 tp->t_flags |= TF_SENTFIN;
1061 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1062 tp->snd_max = tp->snd_nxt + len;
1069 if (so->so_options & SO_DEBUG) {
1075 save = ipov->ih_len;
1076 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1078 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1082 ipov->ih_len = save;
1087 * Fill in IP length and desired time to live and
1088 * send to IP level. There should be a better way
1089 * to handle ttl and tos; we could keep them in
1090 * the template, but need a way to checksum without them.
1093 * m->m_pkthdr.len should have been set before cksum calcuration,
1094 * because in6_cksum() need it.
1099 * we separately set hoplimit for every segment, since the
1100 * user might want to change the value via setsockopt.
1101 * Also, desired default hop limit might be changed via
1102 * Neighbor Discovery.
1104 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1106 /* TODO: IPv6 IP6TOS_ECT bit on */
1107 error = ip6_output(m,
1108 tp->t_inpcb->in6p_outputopts, NULL,
1109 ((so->so_options & SO_DONTROUTE) ?
1110 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1114 ip->ip_len = m->m_pkthdr.len;
1116 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1117 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1120 * If we do path MTU discovery, then we set DF on every packet.
1121 * This might not be the best thing to do according to RFC3390
1122 * Section 2. However the tcp hostcache migitates the problem
1123 * so it affects only the first tcp connection with a host.
1125 if (path_mtu_discovery)
1126 ip->ip_off |= IP_DF;
1128 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1129 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1135 * We know that the packet was lost, so back out the
1136 * sequence number advance, if any.
1138 * If the error is EPERM the packet got blocked by the
1139 * local firewall. Normally we should terminate the
1140 * connection but the blocking may have been spurious
1141 * due to a firewall reconfiguration cycle. So we treat
1142 * it like a packet loss and let the retransmit timer and
1143 * timeouts do their work over time.
1144 * XXX: It is a POLA question whether calling tcp_drop right
1145 * away would be the really correct behavior instead.
1147 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1148 !tcp_timer_active(tp, TT_PERSIST)) &&
1149 ((flags & TH_SYN) == 0) &&
1153 tp->sackhint.sack_bytes_rexmit -= len;
1154 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1155 ("sackhint bytes rtx >= 0"));
1160 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1163 tp->t_softerror = error;
1166 if (!tcp_timer_active(tp, TT_REXMT) &&
1167 !tcp_timer_active(tp, TT_PERSIST))
1168 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1169 tp->snd_cwnd = tp->t_maxseg;
1173 * For some reason the interface we used initially
1174 * to send segments changed to another or lowered
1177 * tcp_mtudisc() will find out the new MTU and as
1178 * its last action, initiate retransmission, so it
1179 * is important to not do so here.
1181 * If TSO was active we either got an interface
1182 * without TSO capabilits or TSO was turned off.
1183 * Disable it for this connection as too and
1184 * immediatly retry with MSS sized segments generated
1188 tp->t_flags &= ~TF_TSO;
1189 tcp_mtudisc(tp->t_inpcb, 0);
1195 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1196 tp->t_softerror = error;
1204 tcpstat.tcps_sndtotal++;
1207 * Data sent (as far as we can tell).
1208 * If this advertises a larger window than any other segment,
1209 * then remember the size of the advertised window.
1210 * Any pending ACK has now been sent.
1212 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1213 tp->rcv_adv = tp->rcv_nxt + recwin;
1214 tp->last_ack_sent = tp->rcv_nxt;
1215 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1216 if (tcp_timer_active(tp, TT_DELACK))
1217 tcp_timer_activate(tp, TT_DELACK, 0);
1220 * This completely breaks TCP if newreno is turned on. What happens
1221 * is that if delayed-acks are turned on on the receiver, this code
1222 * on the transmitter effectively destroys the TCP window, forcing
1223 * it to four packets (1.5Kx4 = 6K window).
1225 if (sendalot && (!tcp_do_newreno || --maxburst))
1234 tcp_setpersist(struct tcpcb *tp)
1236 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1239 if (tcp_timer_active(tp, TT_REXMT))
1240 panic("tcp_setpersist: retransmit pending");
1242 * Start/restart persistance timer.
1244 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1245 TCPTV_PERSMIN, TCPTV_PERSMAX);
1246 tcp_timer_activate(tp, TT_PERSIST, tt);
1247 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1252 * Insert TCP options according to the supplied parameters to the place
1253 * optp in a consistent way. Can handle unaligned destinations.
1255 * The order of the option processing is crucial for optimal packing and
1256 * alignment for the scarce option space.
1258 * The optimal order for a SYN/SYN-ACK segment is:
1259 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1260 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1262 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1263 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1264 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1265 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1266 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1269 tcp_addoptions(struct tcpopt *to, u_char *optp)
1271 u_int mask, optlen = 0;
1273 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1274 if ((to->to_flags & mask) != mask)
1276 if (optlen == TCP_MAXOLEN)
1278 switch (to->to_flags & mask) {
1280 while (optlen % 4) {
1281 optlen += TCPOLEN_NOP;
1282 *optp++ = TCPOPT_NOP;
1284 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1286 optlen += TCPOLEN_MAXSEG;
1287 *optp++ = TCPOPT_MAXSEG;
1288 *optp++ = TCPOLEN_MAXSEG;
1289 to->to_mss = htons(to->to_mss);
1290 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1291 optp += sizeof(to->to_mss);
1294 while (!optlen || optlen % 2 != 1) {
1295 optlen += TCPOLEN_NOP;
1296 *optp++ = TCPOPT_NOP;
1298 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1300 optlen += TCPOLEN_WINDOW;
1301 *optp++ = TCPOPT_WINDOW;
1302 *optp++ = TCPOLEN_WINDOW;
1303 *optp++ = to->to_wscale;
1306 while (optlen % 2) {
1307 optlen += TCPOLEN_NOP;
1308 *optp++ = TCPOPT_NOP;
1310 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1312 optlen += TCPOLEN_SACK_PERMITTED;
1313 *optp++ = TCPOPT_SACK_PERMITTED;
1314 *optp++ = TCPOLEN_SACK_PERMITTED;
1317 while (!optlen || optlen % 4 != 2) {
1318 optlen += TCPOLEN_NOP;
1319 *optp++ = TCPOPT_NOP;
1321 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1323 optlen += TCPOLEN_TIMESTAMP;
1324 *optp++ = TCPOPT_TIMESTAMP;
1325 *optp++ = TCPOLEN_TIMESTAMP;
1326 to->to_tsval = htonl(to->to_tsval);
1327 to->to_tsecr = htonl(to->to_tsecr);
1328 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1329 optp += sizeof(to->to_tsval);
1330 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1331 optp += sizeof(to->to_tsecr);
1335 int siglen = TCPOLEN_SIGNATURE - 2;
1337 while (!optlen || optlen % 4 != 2) {
1338 optlen += TCPOLEN_NOP;
1339 *optp++ = TCPOPT_NOP;
1341 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1343 optlen += TCPOLEN_SIGNATURE;
1344 *optp++ = TCPOPT_SIGNATURE;
1345 *optp++ = TCPOLEN_SIGNATURE;
1346 to->to_signature = optp;
1354 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1357 while (!optlen || optlen % 4 != 2) {
1358 optlen += TCPOLEN_NOP;
1359 *optp++ = TCPOPT_NOP;
1361 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1363 optlen += TCPOLEN_SACKHDR;
1364 *optp++ = TCPOPT_SACK;
1365 sackblks = min(to->to_nsacks,
1366 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1367 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1368 while (sackblks--) {
1369 sack_seq = htonl(sack->start);
1370 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1371 optp += sizeof(sack_seq);
1372 sack_seq = htonl(sack->end);
1373 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1374 optp += sizeof(sack_seq);
1375 optlen += TCPOLEN_SACK;
1378 tcpstat.tcps_sack_send_blocks++;
1382 panic("%s: unknown TCP option type", __func__);
1387 /* Terminate and pad TCP options to a 4 byte boundary. */
1389 optlen += TCPOLEN_EOL;
1390 *optp++ = TCPOPT_EOL;
1393 * According to RFC 793 (STD0007):
1394 * "The content of the header beyond the End-of-Option option
1395 * must be header padding (i.e., zero)."
1396 * and later: "The padding is composed of zeros."
1398 while (optlen % 4) {
1399 optlen += TCPOLEN_PAD;
1400 *optp++ = TCPOPT_PAD;
1403 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));