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"
38 #include "opt_tcpdebug.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/kernel.h>
46 #include <sys/mutex.h>
47 #include <sys/protosw.h>
48 #include <sys/socket.h>
49 #include <sys/socketvar.h>
50 #include <sys/sysctl.h>
53 #include <net/route.h>
56 #include <netinet/in.h>
57 #include <netinet/in_systm.h>
58 #include <netinet/ip.h>
59 #include <netinet/in_pcb.h>
60 #include <netinet/ip_var.h>
61 #include <netinet/ip_options.h>
63 #include <netinet6/in6_pcb.h>
64 #include <netinet/ip6.h>
65 #include <netinet6/ip6_var.h>
67 #include <netinet/tcp.h>
69 #include <netinet/tcp_fsm.h>
70 #include <netinet/tcp_seq.h>
71 #include <netinet/tcp_timer.h>
72 #include <netinet/tcp_var.h>
73 #include <netinet/tcpip.h>
75 #include <netinet/tcp_debug.h>
79 #include <netipsec/ipsec.h>
82 #include <machine/in_cksum.h>
84 #include <security/mac/mac_framework.h>
87 extern struct mbuf *m_copypack();
90 VNET_DEFINE(int, path_mtu_discovery) = 1;
91 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
92 &VNET_NAME(path_mtu_discovery), 1,
93 "Enable Path MTU Discovery");
95 VNET_DEFINE(int, ss_fltsz) = 1;
96 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, slowstart_flightsize, CTLFLAG_RW,
97 &VNET_NAME(ss_fltsz), 1,
98 "Slow start flight size");
100 VNET_DEFINE(int, ss_fltsz_local) = 4;
101 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, local_slowstart_flightsize,
102 CTLFLAG_RW, &VNET_NAME(ss_fltsz_local), 1,
103 "Slow start flight size for local networks");
105 VNET_DEFINE(int, tcp_do_newreno) = 1;
106 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
107 &VNET_NAME(tcp_do_newreno), 0,
108 "Enable NewReno Algorithms");
110 VNET_DEFINE(int, tcp_do_tso) = 1;
111 #define V_tcp_do_tso VNET(tcp_do_tso)
112 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
113 &VNET_NAME(tcp_do_tso), 0,
114 "Enable TCP Segmentation Offload");
116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
117 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
118 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
119 &VNET_NAME(tcp_do_autosndbuf), 0,
120 "Enable automatic send buffer sizing");
122 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
123 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
124 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
125 &VNET_NAME(tcp_autosndbuf_inc), 0,
126 "Incrementor step size of automatic send buffer");
128 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024;
129 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
130 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
131 &VNET_NAME(tcp_autosndbuf_max), 0,
132 "Max size of automatic send buffer");
136 * Tcp output routine: figure out what should be sent and send it.
139 tcp_output(struct tcpcb *tp)
141 struct socket *so = tp->t_inpcb->inp_socket;
142 long len, recwin, sendwin;
143 int off, flags, error;
145 struct ip *ip = NULL;
146 struct ipovly *ipov = NULL;
148 u_char opt[TCP_MAXOLEN];
149 unsigned ipoptlen, optlen, hdrlen;
151 unsigned ipsec_optlen = 0;
154 int sack_rxmit, sack_bytes_rxmt;
159 int maxburst = TCP_MAXBURST;
162 struct ip6_hdr *ip6 = NULL;
165 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
168 INP_WLOCK_ASSERT(tp->t_inpcb);
171 * Determine length of data that should be transmitted,
172 * and flags that will be used.
173 * If there is some data or critical controls (SYN, RST)
174 * to send, then transmit; otherwise, investigate further.
176 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
177 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur) {
179 * We have been idle for "a while" and no acks are
180 * expected to clock out any data we send --
181 * slow start to get ack "clock" running again.
183 * Set the slow-start flight size depending on whether
184 * this is a local network or not.
189 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
190 ss = V_ss_fltsz_local;
193 if (in_localaddr(tp->t_inpcb->inp_faddr))
194 ss = V_ss_fltsz_local;
195 tp->snd_cwnd = tp->t_maxseg * ss;
197 tp->t_flags &= ~TF_LASTIDLE;
199 if (tp->t_flags & TF_MORETOCOME) {
200 tp->t_flags |= TF_LASTIDLE;
206 * If we've recently taken a timeout, snd_max will be greater than
207 * snd_nxt. There may be SACK information that allows us to avoid
208 * resending already delivered data. Adjust snd_nxt accordingly.
210 if ((tp->t_flags & TF_SACK_PERMIT) &&
211 SEQ_LT(tp->snd_nxt, tp->snd_max))
215 off = tp->snd_nxt - tp->snd_una;
216 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
217 sendwin = min(sendwin, tp->snd_bwnd);
219 flags = tcp_outflags[tp->t_state];
221 * Send any SACK-generated retransmissions. If we're explicitly trying
222 * to send out new data (when sendalot is 1), bypass this function.
223 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
224 * we're replacing a (future) new transmission with a retransmission
225 * now, and we previously incremented snd_cwnd in tcp_input().
228 * Still in sack recovery , reset rxmit flag to zero.
234 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
235 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
238 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
241 /* Do not retransmit SACK segments beyond snd_recover */
242 if (SEQ_GT(p->end, tp->snd_recover)) {
244 * (At least) part of sack hole extends beyond
245 * snd_recover. Check to see if we can rexmit data
248 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
250 * Can't rexmit any more data for this hole.
251 * That data will be rexmitted in the next
252 * sack recovery episode, when snd_recover
253 * moves past p->rxmit.
256 goto after_sack_rexmit;
258 /* Can rexmit part of the current hole */
259 len = ((long)ulmin(cwin,
260 tp->snd_recover - p->rxmit));
262 len = ((long)ulmin(cwin, p->end - p->rxmit));
263 off = p->rxmit - tp->snd_una;
264 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
269 TCPSTAT_INC(tcps_sack_rexmits);
270 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
271 min(len, tp->t_maxseg));
276 * Get standard flags, and add SYN or FIN if requested by 'hidden'
279 if (tp->t_flags & TF_NEEDFIN)
281 if (tp->t_flags & TF_NEEDSYN)
284 SOCKBUF_LOCK(&so->so_snd);
286 * If in persist timeout with window of 0, send 1 byte.
287 * Otherwise, if window is small but nonzero
288 * and timer expired, we will send what we can
289 * and go to transmit state.
291 if (tp->t_flags & TF_FORCEDATA) {
294 * If we still have some data to send, then
295 * clear the FIN bit. Usually this would
296 * happen below when it realizes that we
297 * aren't sending all the data. However,
298 * if we have exactly 1 byte of unsent data,
299 * then it won't clear the FIN bit below,
300 * and if we are in persist state, we wind
301 * up sending the packet without recording
302 * that we sent the FIN bit.
304 * We can't just blindly clear the FIN bit,
305 * because if we don't have any more data
306 * to send then the probe will be the FIN
309 if (off < so->so_snd.sb_cc)
313 tcp_timer_activate(tp, TT_PERSIST, 0);
319 * If snd_nxt == snd_max and we have transmitted a FIN, the
320 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
321 * a negative length. This can also occur when TCP opens up
322 * its congestion window while receiving additional duplicate
323 * acks after fast-retransmit because TCP will reset snd_nxt
324 * to snd_max after the fast-retransmit.
326 * In the normal retransmit-FIN-only case, however, snd_nxt will
327 * be set to snd_una, the offset will be 0, and the length may
330 * If sack_rxmit is true we are retransmitting from the scoreboard
331 * in which case len is already set.
333 if (sack_rxmit == 0) {
334 if (sack_bytes_rxmt == 0)
335 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
340 * We are inside of a SACK recovery episode and are
341 * sending new data, having retransmitted all the
342 * data possible in the scoreboard.
344 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
347 * Don't remove this (len > 0) check !
348 * We explicitly check for len > 0 here (although it
349 * isn't really necessary), to work around a gcc
350 * optimization issue - to force gcc to compute
351 * len above. Without this check, the computation
352 * of len is bungled by the optimizer.
355 cwin = tp->snd_cwnd -
356 (tp->snd_nxt - tp->sack_newdata) -
360 len = lmin(len, cwin);
366 * Lop off SYN bit if it has already been sent. However, if this
367 * is SYN-SENT state and if segment contains data and if we don't
368 * know that foreign host supports TAO, suppress sending segment.
370 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
371 if (tp->t_state != TCPS_SYN_RECEIVED)
377 * Be careful not to send data and/or FIN on SYN segments.
378 * This measure is needed to prevent interoperability problems
379 * with not fully conformant TCP implementations.
381 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
388 * If FIN has been sent but not acked,
389 * but we haven't been called to retransmit,
390 * len will be < 0. Otherwise, window shrank
391 * after we sent into it. If window shrank to 0,
392 * cancel pending retransmit, pull snd_nxt back
393 * to (closed) window, and set the persist timer
394 * if it isn't already going. If the window didn't
395 * close completely, just wait for an ACK.
399 tcp_timer_activate(tp, TT_REXMT, 0);
401 tp->snd_nxt = tp->snd_una;
402 if (!tcp_timer_active(tp, TT_PERSIST))
407 /* len will be >= 0 after this point. */
408 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
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 (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
446 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
447 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
448 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
449 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
450 if (!sbreserve_locked(&so->so_snd,
451 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
452 V_tcp_autosndbuf_max), so, curthread))
453 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
458 * Truncate to the maximum segment length or enable TCP Segmentation
459 * Offloading (if supported by hardware) and ensure that FIN is removed
460 * if the length no longer contains the last data byte.
462 * TSO may only be used if we are in a pure bulk sending state. The
463 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
464 * IP options prevent using TSO. With TSO the TCP header is the same
465 * (except for the sequence number) for all generated packets. This
466 * makes it impossible to transmit any options which vary per generated
469 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
470 * removal of FIN (if not already catched here) are handled later after
471 * the exact length of the TCP options are known.
475 * Pre-calculate here as we save another lookup into the darknesses
476 * of IPsec that way and can actually decide if TSO is ok.
478 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
480 if (len > tp->t_maxseg) {
481 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
482 ((tp->t_flags & TF_SIGNATURE) == 0) &&
483 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
484 tp->t_inpcb->inp_options == NULL &&
485 tp->t_inpcb->in6p_options == NULL
498 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
501 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
505 recwin = sbspace(&so->so_rcv);
508 * Sender silly window avoidance. We transmit under the following
509 * conditions when len is non-zero:
511 * - We have a full segment (or more with TSO)
512 * - This is the last buffer in a write()/send() and we are
513 * either idle or running NODELAY
514 * - we've timed out (e.g. persist timer)
515 * - we have more then 1/2 the maximum send window's worth of
516 * data (receiver may be limited the window size)
517 * - we need to retransmit
520 if (len >= tp->t_maxseg)
523 * NOTE! on localhost connections an 'ack' from the remote
524 * end may occur synchronously with the output and cause
525 * us to flush a buffer queued with moretocome. XXX
527 * note: the len + off check is almost certainly unnecessary.
529 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
530 (idle || (tp->t_flags & TF_NODELAY)) &&
531 len + off >= so->so_snd.sb_cc &&
532 (tp->t_flags & TF_NOPUSH) == 0) {
535 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
537 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
539 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
546 * Compare available window to amount of window
547 * known to peer (as advertised window less
548 * next expected input). If the difference is at least two
549 * max size segments, or at least 50% of the maximum possible
550 * window, then want to send a window update to peer.
551 * Skip this if the connection is in T/TCP half-open state.
552 * Don't send pure window updates when the peer has closed
553 * the connection and won't ever send more data.
555 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
556 !TCPS_HAVERCVDFIN(tp->t_state)) {
558 * "adv" is the amount we can increase the window,
559 * taking into account that we are limited by
560 * TCP_MAXWIN << tp->rcv_scale.
562 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
563 (tp->rcv_adv - tp->rcv_nxt);
565 if (adv >= (long) (2 * tp->t_maxseg))
567 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
572 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
573 * is also a catch-all for the retransmit timer timeout case.
575 if (tp->t_flags & TF_ACKNOW)
577 if ((flags & TH_RST) ||
578 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
580 if (SEQ_GT(tp->snd_up, tp->snd_una))
583 * If our state indicates that FIN should be sent
584 * and we have not yet done so, then we need to send.
586 if (flags & TH_FIN &&
587 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
590 * In SACK, it is possible for tcp_output to fail to send a segment
591 * after the retransmission timer has been turned off. Make sure
592 * that the retransmission timer is set.
594 if ((tp->t_flags & TF_SACK_PERMIT) &&
595 SEQ_GT(tp->snd_max, tp->snd_una) &&
596 !tcp_timer_active(tp, TT_REXMT) &&
597 !tcp_timer_active(tp, TT_PERSIST)) {
598 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
602 * TCP window updates are not reliable, rather a polling protocol
603 * using ``persist'' packets is used to insure receipt of window
604 * updates. The three ``states'' for the output side are:
605 * idle not doing retransmits or persists
606 * persisting to move a small or zero window
607 * (re)transmitting and thereby not persisting
609 * tcp_timer_active(tp, TT_PERSIST)
610 * is true when we are in persist state.
611 * (tp->t_flags & TF_FORCEDATA)
612 * is set when we are called to send a persist packet.
613 * tcp_timer_active(tp, TT_REXMT)
614 * is set when we are retransmitting
615 * The output side is idle when both timers are zero.
617 * If send window is too small, there is data to transmit, and no
618 * retransmit or persist is pending, then go to persist state.
619 * If nothing happens soon, send when timer expires:
620 * if window is nonzero, transmit what we can,
621 * otherwise force out a byte.
623 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
624 !tcp_timer_active(tp, TT_PERSIST)) {
630 * No reason to send a segment, just return.
633 SOCKBUF_UNLOCK(&so->so_snd);
637 SOCKBUF_LOCK_ASSERT(&so->so_snd);
639 * Before ESTABLISHED, force sending of initial options
640 * unless TCP set not to do any options.
641 * NOTE: we assume that the IP/TCP header plus TCP options
642 * always fit in a single mbuf, leaving room for a maximum
644 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
649 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
652 hdrlen = sizeof (struct tcpiphdr);
655 * Compute options for segment.
656 * We only have to care about SYN and established connection
657 * segments. Options for SYN-ACK segments are handled in TCP
660 if ((tp->t_flags & TF_NOOPT) == 0) {
662 /* Maximum segment size. */
663 if (flags & TH_SYN) {
664 tp->snd_nxt = tp->iss;
665 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
666 to.to_flags |= TOF_MSS;
668 /* Window scaling. */
669 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
670 to.to_wscale = tp->request_r_scale;
671 to.to_flags |= TOF_SCALE;
674 if ((tp->t_flags & TF_RCVD_TSTMP) ||
675 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
676 to.to_tsval = ticks + tp->ts_offset;
677 to.to_tsecr = tp->ts_recent;
678 to.to_flags |= TOF_TS;
679 /* Set receive buffer autosizing timestamp. */
680 if (tp->rfbuf_ts == 0 &&
681 (so->so_rcv.sb_flags & SB_AUTOSIZE))
682 tp->rfbuf_ts = ticks;
684 /* Selective ACK's. */
685 if (tp->t_flags & TF_SACK_PERMIT) {
687 to.to_flags |= TOF_SACKPERM;
688 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
689 (tp->t_flags & TF_SACK_PERMIT) &&
690 tp->rcv_numsacks > 0) {
691 to.to_flags |= TOF_SACK;
692 to.to_nsacks = tp->rcv_numsacks;
693 to.to_sacks = (u_char *)tp->sackblks;
697 /* TCP-MD5 (RFC2385). */
698 if (tp->t_flags & TF_SIGNATURE)
699 to.to_flags |= TOF_SIGNATURE;
700 #endif /* TCP_SIGNATURE */
702 /* Processing the options. */
703 hdrlen += optlen = tcp_addoptions(&to, opt);
708 ipoptlen = ip6_optlen(tp->t_inpcb);
711 if (tp->t_inpcb->inp_options)
712 ipoptlen = tp->t_inpcb->inp_options->m_len -
713 offsetof(struct ipoption, ipopt_list);
717 ipoptlen += ipsec_optlen;
721 * Adjust data length if insertion of options will
722 * bump the packet length beyond the t_maxopd length.
723 * Clear the FIN bit because we cut off the tail of
726 * When doing TSO limit a burst to TCP_MAXWIN minus the
727 * IP, TCP and Options length to keep ip->ip_len from
728 * overflowing. Prevent the last segment from being
729 * fractional thus making them all equal sized and set
730 * the flag to continue sending. TSO is disabled when
731 * IP options or IPSEC are present.
733 if (len + optlen + ipoptlen > tp->t_maxopd) {
736 if (len > TCP_MAXWIN - hdrlen - optlen) {
737 len = TCP_MAXWIN - hdrlen - optlen;
738 len = len - (len % (tp->t_maxopd - optlen));
740 } else if (tp->t_flags & TF_NEEDFIN)
743 len = tp->t_maxopd - optlen - ipoptlen;
748 /*#ifdef DIAGNOSTIC*/
750 if (max_linkhdr + hdrlen > MCLBYTES)
752 if (max_linkhdr + hdrlen > MHLEN)
754 panic("tcphdr too big");
758 * This KASSERT is here to catch edge cases at a well defined place.
759 * Before, those had triggered (random) panic conditions further down.
761 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
764 * Grab a header mbuf, attaching a copy of data to
765 * be transmitted, and initialize the header from
766 * the template for sends on this connection.
772 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
773 TCPSTAT_INC(tcps_sndprobe);
774 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
775 TCPSTAT_INC(tcps_sndrexmitpack);
776 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
778 TCPSTAT_INC(tcps_sndpack);
779 TCPSTAT_ADD(tcps_sndbyte, len);
782 if ((m = m_copypack(so->so_snd.sb_mb, off,
783 (int)len, max_linkhdr + hdrlen)) == 0) {
784 SOCKBUF_UNLOCK(&so->so_snd);
789 * m_copypack left space for our hdr; use it.
794 MGETHDR(m, M_DONTWAIT, MT_DATA);
796 SOCKBUF_UNLOCK(&so->so_snd);
801 if (MHLEN < hdrlen + max_linkhdr) {
802 MCLGET(m, M_DONTWAIT);
803 if ((m->m_flags & M_EXT) == 0) {
804 SOCKBUF_UNLOCK(&so->so_snd);
811 m->m_data += max_linkhdr;
815 * Start the m_copy functions from the closest mbuf
816 * to the offset in the socket buffer chain.
818 mb = sbsndptr(&so->so_snd, off, len, &moff);
820 if (len <= MHLEN - hdrlen - max_linkhdr) {
821 m_copydata(mb, moff, (int)len,
822 mtod(m, caddr_t) + hdrlen);
825 m->m_next = m_copy(mb, moff, (int)len);
826 if (m->m_next == NULL) {
827 SOCKBUF_UNLOCK(&so->so_snd);
835 * If we're sending everything we've got, set PUSH.
836 * (This will keep happy those implementations which only
837 * give data to the user when a buffer fills or
840 if (off + len == so->so_snd.sb_cc)
842 SOCKBUF_UNLOCK(&so->so_snd);
844 SOCKBUF_UNLOCK(&so->so_snd);
845 if (tp->t_flags & TF_ACKNOW)
846 TCPSTAT_INC(tcps_sndacks);
847 else if (flags & (TH_SYN|TH_FIN|TH_RST))
848 TCPSTAT_INC(tcps_sndctrl);
849 else if (SEQ_GT(tp->snd_up, tp->snd_una))
850 TCPSTAT_INC(tcps_sndurg);
852 TCPSTAT_INC(tcps_sndwinup);
854 MGETHDR(m, M_DONTWAIT, MT_DATA);
860 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
865 m->m_data += max_linkhdr;
868 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
869 m->m_pkthdr.rcvif = (struct ifnet *)0;
871 mac_inpcb_create_mbuf(tp->t_inpcb, m);
875 ip6 = mtod(m, struct ip6_hdr *);
876 th = (struct tcphdr *)(ip6 + 1);
877 tcpip_fillheaders(tp->t_inpcb, ip6, th);
881 ip = mtod(m, struct ip *);
882 ipov = (struct ipovly *)ip;
883 th = (struct tcphdr *)(ip + 1);
884 tcpip_fillheaders(tp->t_inpcb, ip, th);
888 * Fill in fields, remembering maximum advertised
889 * window for use in delaying messages about window sizes.
890 * If resending a FIN, be sure not to use a new sequence number.
892 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
893 tp->snd_nxt == tp->snd_max)
896 * If we are starting a connection, send ECN setup
897 * SYN packet. If we are on a retransmit, we may
898 * resend those bits a number of times as per
901 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
902 if (tp->t_rxtshift >= 1) {
903 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
904 flags |= TH_ECE|TH_CWR;
906 flags |= TH_ECE|TH_CWR;
909 if (tp->t_state == TCPS_ESTABLISHED &&
910 (tp->t_flags & TF_ECN_PERMIT)) {
912 * If the peer has ECN, mark data packets with
913 * ECN capable transmission (ECT).
914 * Ignore pure ack packets, retransmissions and window probes.
916 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
917 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
920 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
923 ip->ip_tos |= IPTOS_ECN_ECT0;
924 TCPSTAT_INC(tcps_ecn_ect0);
928 * Reply with proper ECN notifications.
930 if (tp->t_flags & TF_ECN_SND_CWR) {
932 tp->t_flags &= ~TF_ECN_SND_CWR;
934 if (tp->t_flags & TF_ECN_SND_ECE)
939 * If we are doing retransmissions, then snd_nxt will
940 * not reflect the first unsent octet. For ACK only
941 * packets, we do not want the sequence number of the
942 * retransmitted packet, we want the sequence number
943 * of the next unsent octet. So, if there is no data
944 * (and no SYN or FIN), use snd_max instead of snd_nxt
945 * when filling in ti_seq. But if we are in persist
946 * state, snd_max might reflect one byte beyond the
947 * right edge of the window, so use snd_nxt in that
948 * case, since we know we aren't doing a retransmission.
949 * (retransmit and persist are mutually exclusive...)
951 if (sack_rxmit == 0) {
952 if (len || (flags & (TH_SYN|TH_FIN)) ||
953 tcp_timer_active(tp, TT_PERSIST))
954 th->th_seq = htonl(tp->snd_nxt);
956 th->th_seq = htonl(tp->snd_max);
958 th->th_seq = htonl(p->rxmit);
960 tp->sackhint.sack_bytes_rexmit += len;
962 th->th_ack = htonl(tp->rcv_nxt);
964 bcopy(opt, th + 1, optlen);
965 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
967 th->th_flags = flags;
969 * Calculate receive window. Don't shrink window,
970 * but avoid silly window syndrome.
972 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
973 recwin < (long)tp->t_maxseg)
975 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
976 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
977 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
978 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
981 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
982 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
983 * case is handled in syncache.
986 th->th_win = htons((u_short)
987 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
989 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
992 * Adjust the RXWIN0SENT flag - indicate that we have advertised
993 * a 0 window. This may cause the remote transmitter to stall. This
994 * flag tells soreceive() to disable delayed acknowledgements when
995 * draining the buffer. This can occur if the receiver is attempting
996 * to read more data than can be buffered prior to transmitting on
1000 tp->t_flags |= TF_RXWIN0SENT;
1002 tp->t_flags &= ~TF_RXWIN0SENT;
1003 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1004 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1005 th->th_flags |= TH_URG;
1008 * If no urgent pointer to send, then we pull
1009 * the urgent pointer to the left edge of the send window
1010 * so that it doesn't drift into the send window on sequence
1011 * number wraparound.
1013 tp->snd_up = tp->snd_una; /* drag it along */
1015 #ifdef TCP_SIGNATURE
1016 if (tp->t_flags & TF_SIGNATURE) {
1017 int sigoff = to.to_signature - opt;
1018 tcp_signature_compute(m, 0, len, optlen,
1019 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1024 * Put TCP length in extended header, and then
1025 * checksum extended header and data.
1027 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1031 * ip6_plen is not need to be filled now, and will be filled
1034 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1035 sizeof(struct tcphdr) + optlen + len);
1039 m->m_pkthdr.csum_flags = CSUM_TCP;
1040 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1041 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1042 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1044 /* IP version must be set here for ipv4/ipv6 checking later */
1045 KASSERT(ip->ip_v == IPVERSION,
1046 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1050 * Enable TSO and specify the size of the segments.
1051 * The TCP pseudo header checksum is always provided.
1052 * XXX: Fixme: This is currently not the case for IPv6.
1055 KASSERT(len > tp->t_maxopd - optlen,
1056 ("%s: len <= tso_segsz", __func__));
1057 m->m_pkthdr.csum_flags |= CSUM_TSO;
1058 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1062 * In transmit state, time the transmission and arrange for
1063 * the retransmit. In persist state, just set snd_max.
1065 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1066 !tcp_timer_active(tp, TT_PERSIST)) {
1067 tcp_seq startseq = tp->snd_nxt;
1070 * Advance snd_nxt over sequence space of this segment.
1072 if (flags & (TH_SYN|TH_FIN)) {
1075 if (flags & TH_FIN) {
1077 tp->t_flags |= TF_SENTFIN;
1083 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1084 tp->snd_max = tp->snd_nxt;
1086 * Time this transmission if not a retransmission and
1087 * not currently timing anything.
1089 if (tp->t_rtttime == 0) {
1090 tp->t_rtttime = ticks;
1091 tp->t_rtseq = startseq;
1092 TCPSTAT_INC(tcps_segstimed);
1097 * Set retransmit timer if not currently set,
1098 * and not doing a pure ack or a keep-alive probe.
1099 * Initial value for retransmit timer is smoothed
1100 * round-trip time + 2 * round-trip time variance.
1101 * Initialize shift counter which is used for backoff
1102 * of retransmit time.
1105 if (!tcp_timer_active(tp, TT_REXMT) &&
1106 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1107 (tp->snd_nxt != tp->snd_una))) {
1108 if (tcp_timer_active(tp, TT_PERSIST)) {
1109 tcp_timer_activate(tp, TT_PERSIST, 0);
1112 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1116 * Persist case, update snd_max but since we are in
1117 * persist mode (no window) we do not update snd_nxt.
1122 if (flags & TH_FIN) {
1124 tp->t_flags |= TF_SENTFIN;
1126 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1127 tp->snd_max = tp->snd_nxt + len;
1134 if (so->so_options & SO_DEBUG) {
1140 save = ipov->ih_len;
1141 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1143 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1147 ipov->ih_len = save;
1152 * Fill in IP length and desired time to live and
1153 * send to IP level. There should be a better way
1154 * to handle ttl and tos; we could keep them in
1155 * the template, but need a way to checksum without them.
1158 * m->m_pkthdr.len should have been set before cksum calcuration,
1159 * because in6_cksum() need it.
1164 * we separately set hoplimit for every segment, since the
1165 * user might want to change the value via setsockopt.
1166 * Also, desired default hop limit might be changed via
1167 * Neighbor Discovery.
1169 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1171 /* TODO: IPv6 IP6TOS_ECT bit on */
1172 error = ip6_output(m,
1173 tp->t_inpcb->in6p_outputopts, NULL,
1174 ((so->so_options & SO_DONTROUTE) ?
1175 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1179 ip->ip_len = m->m_pkthdr.len;
1181 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1182 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1185 * If we do path MTU discovery, then we set DF on every packet.
1186 * This might not be the best thing to do according to RFC3390
1187 * Section 2. However the tcp hostcache migitates the problem
1188 * so it affects only the first tcp connection with a host.
1190 if (V_path_mtu_discovery)
1191 ip->ip_off |= IP_DF;
1193 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1194 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1200 * We know that the packet was lost, so back out the
1201 * sequence number advance, if any.
1203 * If the error is EPERM the packet got blocked by the
1204 * local firewall. Normally we should terminate the
1205 * connection but the blocking may have been spurious
1206 * due to a firewall reconfiguration cycle. So we treat
1207 * it like a packet loss and let the retransmit timer and
1208 * timeouts do their work over time.
1209 * XXX: It is a POLA question whether calling tcp_drop right
1210 * away would be the really correct behavior instead.
1212 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1213 !tcp_timer_active(tp, TT_PERSIST)) &&
1214 ((flags & TH_SYN) == 0) &&
1218 tp->sackhint.sack_bytes_rexmit -= len;
1219 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1220 ("sackhint bytes rtx >= 0"));
1225 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1228 tp->t_softerror = error;
1231 if (!tcp_timer_active(tp, TT_REXMT) &&
1232 !tcp_timer_active(tp, TT_PERSIST))
1233 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1234 tp->snd_cwnd = tp->t_maxseg;
1238 * For some reason the interface we used initially
1239 * to send segments changed to another or lowered
1242 * tcp_mtudisc() will find out the new MTU and as
1243 * its last action, initiate retransmission, so it
1244 * is important to not do so here.
1246 * If TSO was active we either got an interface
1247 * without TSO capabilits or TSO was turned off.
1248 * Disable it for this connection as too and
1249 * immediatly retry with MSS sized segments generated
1253 tp->t_flags &= ~TF_TSO;
1254 tcp_mtudisc(tp->t_inpcb, 0);
1260 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1261 tp->t_softerror = error;
1269 TCPSTAT_INC(tcps_sndtotal);
1272 * Data sent (as far as we can tell).
1273 * If this advertises a larger window than any other segment,
1274 * then remember the size of the advertised window.
1275 * Any pending ACK has now been sent.
1277 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1278 tp->rcv_adv = tp->rcv_nxt + recwin;
1279 tp->last_ack_sent = tp->rcv_nxt;
1280 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1281 if (tcp_timer_active(tp, TT_DELACK))
1282 tcp_timer_activate(tp, TT_DELACK, 0);
1285 * This completely breaks TCP if newreno is turned on. What happens
1286 * is that if delayed-acks are turned on on the receiver, this code
1287 * on the transmitter effectively destroys the TCP window, forcing
1288 * it to four packets (1.5Kx4 = 6K window).
1290 if (sendalot && (!V_tcp_do_newreno || --maxburst))
1299 tcp_setpersist(struct tcpcb *tp)
1301 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1304 if (tcp_timer_active(tp, TT_REXMT))
1305 panic("tcp_setpersist: retransmit pending");
1307 * Start/restart persistance timer.
1309 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1310 TCPTV_PERSMIN, TCPTV_PERSMAX);
1311 tcp_timer_activate(tp, TT_PERSIST, tt);
1312 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1317 * Insert TCP options according to the supplied parameters to the place
1318 * optp in a consistent way. Can handle unaligned destinations.
1320 * The order of the option processing is crucial for optimal packing and
1321 * alignment for the scarce option space.
1323 * The optimal order for a SYN/SYN-ACK segment is:
1324 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1325 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1327 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1328 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1329 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1330 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1331 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1334 tcp_addoptions(struct tcpopt *to, u_char *optp)
1336 u_int mask, optlen = 0;
1338 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1339 if ((to->to_flags & mask) != mask)
1341 if (optlen == TCP_MAXOLEN)
1343 switch (to->to_flags & mask) {
1345 while (optlen % 4) {
1346 optlen += TCPOLEN_NOP;
1347 *optp++ = TCPOPT_NOP;
1349 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1351 optlen += TCPOLEN_MAXSEG;
1352 *optp++ = TCPOPT_MAXSEG;
1353 *optp++ = TCPOLEN_MAXSEG;
1354 to->to_mss = htons(to->to_mss);
1355 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1356 optp += sizeof(to->to_mss);
1359 while (!optlen || optlen % 2 != 1) {
1360 optlen += TCPOLEN_NOP;
1361 *optp++ = TCPOPT_NOP;
1363 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1365 optlen += TCPOLEN_WINDOW;
1366 *optp++ = TCPOPT_WINDOW;
1367 *optp++ = TCPOLEN_WINDOW;
1368 *optp++ = to->to_wscale;
1371 while (optlen % 2) {
1372 optlen += TCPOLEN_NOP;
1373 *optp++ = TCPOPT_NOP;
1375 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1377 optlen += TCPOLEN_SACK_PERMITTED;
1378 *optp++ = TCPOPT_SACK_PERMITTED;
1379 *optp++ = TCPOLEN_SACK_PERMITTED;
1382 while (!optlen || optlen % 4 != 2) {
1383 optlen += TCPOLEN_NOP;
1384 *optp++ = TCPOPT_NOP;
1386 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1388 optlen += TCPOLEN_TIMESTAMP;
1389 *optp++ = TCPOPT_TIMESTAMP;
1390 *optp++ = TCPOLEN_TIMESTAMP;
1391 to->to_tsval = htonl(to->to_tsval);
1392 to->to_tsecr = htonl(to->to_tsecr);
1393 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1394 optp += sizeof(to->to_tsval);
1395 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1396 optp += sizeof(to->to_tsecr);
1400 int siglen = TCPOLEN_SIGNATURE - 2;
1402 while (!optlen || optlen % 4 != 2) {
1403 optlen += TCPOLEN_NOP;
1404 *optp++ = TCPOPT_NOP;
1406 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1408 optlen += TCPOLEN_SIGNATURE;
1409 *optp++ = TCPOPT_SIGNATURE;
1410 *optp++ = TCPOLEN_SIGNATURE;
1411 to->to_signature = optp;
1419 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1422 while (!optlen || optlen % 4 != 2) {
1423 optlen += TCPOLEN_NOP;
1424 *optp++ = TCPOPT_NOP;
1426 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1428 optlen += TCPOLEN_SACKHDR;
1429 *optp++ = TCPOPT_SACK;
1430 sackblks = min(to->to_nsacks,
1431 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1432 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1433 while (sackblks--) {
1434 sack_seq = htonl(sack->start);
1435 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1436 optp += sizeof(sack_seq);
1437 sack_seq = htonl(sack->end);
1438 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1439 optp += sizeof(sack_seq);
1440 optlen += TCPOLEN_SACK;
1443 TCPSTAT_INC(tcps_sack_send_blocks);
1447 panic("%s: unknown TCP option type", __func__);
1452 /* Terminate and pad TCP options to a 4 byte boundary. */
1454 optlen += TCPOLEN_EOL;
1455 *optp++ = TCPOPT_EOL;
1458 * According to RFC 793 (STD0007):
1459 * "The content of the header beyond the End-of-Option option
1460 * must be header padding (i.e., zero)."
1461 * and later: "The padding is composed of zeros."
1463 while (optlen % 4) {
1464 optlen += TCPOLEN_PAD;
1465 *optp++ = TCPOPT_PAD;
1468 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));