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>
52 #include <sys/vimage.h>
54 #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 int path_mtu_discovery = 1;
91 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, path_mtu_discovery,
92 CTLFLAG_RW, path_mtu_discovery, 1, "Enable Path MTU Discovery");
95 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO,
96 slowstart_flightsize, CTLFLAG_RW,
97 ss_fltsz, 1, "Slow start flight size");
99 int ss_fltsz_local = 4;
100 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO,
101 local_slowstart_flightsize, CTLFLAG_RW,
102 ss_fltsz_local, 1, "Slow start flight size for local networks");
104 int tcp_do_newreno = 1;
105 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, newreno, CTLFLAG_RW,
106 tcp_do_newreno, 0, "Enable NewReno Algorithms");
109 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
110 tcp_do_tso, 0, "Enable TCP Segmentation Offload");
112 int tcp_do_autosndbuf = 1;
113 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_auto,
115 tcp_do_autosndbuf, 0, "Enable automatic send buffer sizing");
117 int tcp_autosndbuf_inc = 8*1024;
118 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_inc,
119 CTLFLAG_RW, tcp_autosndbuf_inc, 0,
120 "Incrementor step size of automatic send buffer");
122 int tcp_autosndbuf_max = 256*1024;
123 SYSCTL_V_INT(V_NET, vnet_inet, _net_inet_tcp, OID_AUTO, sendbuf_max,
124 CTLFLAG_RW, tcp_autosndbuf_max, 0,
125 "Max size of automatic send buffer");
129 * Tcp output routine: figure out what should be sent and send it.
132 tcp_output(struct tcpcb *tp)
134 INIT_VNET_INET(tp->t_inpcb->inp_vnet);
135 struct socket *so = tp->t_inpcb->inp_socket;
136 long len, recwin, sendwin;
137 int off, flags, error;
139 struct ip *ip = NULL;
140 struct ipovly *ipov = NULL;
142 u_char opt[TCP_MAXOLEN];
143 unsigned ipoptlen, optlen, hdrlen;
145 unsigned ipsec_optlen = 0;
148 int sack_rxmit, sack_bytes_rxmt;
153 int maxburst = TCP_MAXBURST;
156 struct ip6_hdr *ip6 = NULL;
159 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
162 INP_WLOCK_ASSERT(tp->t_inpcb);
165 * Determine length of data that should be transmitted,
166 * and flags that will be used.
167 * If there is some data or critical controls (SYN, RST)
168 * to send, then transmit; otherwise, investigate further.
170 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
171 if (idle && (ticks - tp->t_rcvtime) >= tp->t_rxtcur) {
173 * We have been idle for "a while" and no acks are
174 * expected to clock out any data we send --
175 * slow start to get ack "clock" running again.
177 * Set the slow-start flight size depending on whether
178 * this is a local network or not.
183 if (in6_localaddr(&tp->t_inpcb->in6p_faddr))
184 ss = V_ss_fltsz_local;
187 if (in_localaddr(tp->t_inpcb->inp_faddr))
188 ss = V_ss_fltsz_local;
189 tp->snd_cwnd = tp->t_maxseg * ss;
191 tp->t_flags &= ~TF_LASTIDLE;
193 if (tp->t_flags & TF_MORETOCOME) {
194 tp->t_flags |= TF_LASTIDLE;
200 * If we've recently taken a timeout, snd_max will be greater than
201 * snd_nxt. There may be SACK information that allows us to avoid
202 * resending already delivered data. Adjust snd_nxt accordingly.
204 if ((tp->t_flags & TF_SACK_PERMIT) &&
205 SEQ_LT(tp->snd_nxt, tp->snd_max))
208 off = tp->snd_nxt - tp->snd_una;
209 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
210 sendwin = min(sendwin, tp->snd_bwnd);
212 flags = tcp_outflags[tp->t_state];
214 * Send any SACK-generated retransmissions. If we're explicitly trying
215 * to send out new data (when sendalot is 1), bypass this function.
216 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
217 * we're replacing a (future) new transmission with a retransmission
218 * now, and we previously incremented snd_cwnd in tcp_input().
221 * Still in sack recovery , reset rxmit flag to zero.
227 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
228 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
231 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
234 /* Do not retransmit SACK segments beyond snd_recover */
235 if (SEQ_GT(p->end, tp->snd_recover)) {
237 * (At least) part of sack hole extends beyond
238 * snd_recover. Check to see if we can rexmit data
241 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
243 * Can't rexmit any more data for this hole.
244 * That data will be rexmitted in the next
245 * sack recovery episode, when snd_recover
246 * moves past p->rxmit.
249 goto after_sack_rexmit;
251 /* Can rexmit part of the current hole */
252 len = ((long)ulmin(cwin,
253 tp->snd_recover - p->rxmit));
255 len = ((long)ulmin(cwin, p->end - p->rxmit));
256 off = p->rxmit - tp->snd_una;
257 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
262 V_tcpstat.tcps_sack_rexmits++;
263 V_tcpstat.tcps_sack_rexmit_bytes +=
264 min(len, tp->t_maxseg);
269 * Get standard flags, and add SYN or FIN if requested by 'hidden'
272 if (tp->t_flags & TF_NEEDFIN)
274 if (tp->t_flags & TF_NEEDSYN)
277 SOCKBUF_LOCK(&so->so_snd);
279 * If in persist timeout with window of 0, send 1 byte.
280 * Otherwise, if window is small but nonzero
281 * and timer expired, we will send what we can
282 * and go to transmit state.
284 if (tp->t_flags & TF_FORCEDATA) {
287 * If we still have some data to send, then
288 * clear the FIN bit. Usually this would
289 * happen below when it realizes that we
290 * aren't sending all the data. However,
291 * if we have exactly 1 byte of unsent data,
292 * then it won't clear the FIN bit below,
293 * and if we are in persist state, we wind
294 * up sending the packet without recording
295 * that we sent the FIN bit.
297 * We can't just blindly clear the FIN bit,
298 * because if we don't have any more data
299 * to send then the probe will be the FIN
302 if (off < so->so_snd.sb_cc)
306 tcp_timer_activate(tp, TT_PERSIST, 0);
312 * If snd_nxt == snd_max and we have transmitted a FIN, the
313 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
314 * a negative length. This can also occur when TCP opens up
315 * its congestion window while receiving additional duplicate
316 * acks after fast-retransmit because TCP will reset snd_nxt
317 * to snd_max after the fast-retransmit.
319 * In the normal retransmit-FIN-only case, however, snd_nxt will
320 * be set to snd_una, the offset will be 0, and the length may
323 * If sack_rxmit is true we are retransmitting from the scoreboard
324 * in which case len is already set.
326 if (sack_rxmit == 0) {
327 if (sack_bytes_rxmt == 0)
328 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
333 * We are inside of a SACK recovery episode and are
334 * sending new data, having retransmitted all the
335 * data possible in the scoreboard.
337 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
340 * Don't remove this (len > 0) check !
341 * We explicitly check for len > 0 here (although it
342 * isn't really necessary), to work around a gcc
343 * optimization issue - to force gcc to compute
344 * len above. Without this check, the computation
345 * of len is bungled by the optimizer.
348 cwin = tp->snd_cwnd -
349 (tp->snd_nxt - tp->sack_newdata) -
353 len = lmin(len, cwin);
359 * Lop off SYN bit if it has already been sent. However, if this
360 * is SYN-SENT state and if segment contains data and if we don't
361 * know that foreign host supports TAO, suppress sending segment.
363 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
364 if (tp->t_state != TCPS_SYN_RECEIVED)
370 * Be careful not to send data and/or FIN on SYN segments.
371 * This measure is needed to prevent interoperability problems
372 * with not fully conformant TCP implementations.
374 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
381 * If FIN has been sent but not acked,
382 * but we haven't been called to retransmit,
383 * len will be < 0. Otherwise, window shrank
384 * after we sent into it. If window shrank to 0,
385 * cancel pending retransmit, pull snd_nxt back
386 * to (closed) window, and set the persist timer
387 * if it isn't already going. If the window didn't
388 * close completely, just wait for an ACK.
392 tcp_timer_activate(tp, TT_REXMT, 0);
394 tp->snd_nxt = tp->snd_una;
395 if (!tcp_timer_active(tp, TT_PERSIST))
400 /* len will be >= 0 after this point. */
401 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
404 * Automatic sizing of send socket buffer. Often the send buffer
405 * size is not optimally adjusted to the actual network conditions
406 * at hand (delay bandwidth product). Setting the buffer size too
407 * small limits throughput on links with high bandwidth and high
408 * delay (eg. trans-continental/oceanic links). Setting the
409 * buffer size too big consumes too much real kernel memory,
410 * especially with many connections on busy servers.
412 * The criteria to step up the send buffer one notch are:
413 * 1. receive window of remote host is larger than send buffer
414 * (with a fudge factor of 5/4th);
415 * 2. send buffer is filled to 7/8th with data (so we actually
416 * have data to make use of it);
417 * 3. send buffer fill has not hit maximal automatic size;
418 * 4. our send window (slow start and cogestion controlled) is
419 * larger than sent but unacknowledged data in send buffer.
421 * The remote host receive window scaling factor may limit the
422 * growing of the send buffer before it reaches its allowed
425 * It scales directly with slow start or congestion window
426 * and does at most one step per received ACK. This fast
427 * scaling has the drawback of growing the send buffer beyond
428 * what is strictly necessary to make full use of a given
429 * delay*bandwith product. However testing has shown this not
430 * to be much of an problem. At worst we are trading wasting
431 * of available bandwith (the non-use of it) for wasting some
432 * socket buffer memory.
434 * TODO: Shrink send buffer during idle periods together
435 * with congestion window. Requires another timer. Has to
436 * wait for upcoming tcp timer rewrite.
438 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
439 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
440 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
441 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
442 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
443 if (!sbreserve_locked(&so->so_snd,
444 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
445 V_tcp_autosndbuf_max), so, curthread))
446 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
451 * Truncate to the maximum segment length or enable TCP Segmentation
452 * Offloading (if supported by hardware) and ensure that FIN is removed
453 * if the length no longer contains the last data byte.
455 * TSO may only be used if we are in a pure bulk sending state. The
456 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
457 * IP options prevent using TSO. With TSO the TCP header is the same
458 * (except for the sequence number) for all generated packets. This
459 * makes it impossible to transmit any options which vary per generated
462 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
463 * removal of FIN (if not already catched here) are handled later after
464 * the exact length of the TCP options are known.
468 * Pre-calculate here as we save another lookup into the darknesses
469 * of IPsec that way and can actually decide if TSO is ok.
471 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
473 if (len > tp->t_maxseg) {
474 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
475 ((tp->t_flags & TF_SIGNATURE) == 0) &&
476 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
477 tp->t_inpcb->inp_options == NULL &&
478 tp->t_inpcb->in6p_options == NULL
491 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
494 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
498 recwin = sbspace(&so->so_rcv);
501 * Sender silly window avoidance. We transmit under the following
502 * conditions when len is non-zero:
504 * - We have a full segment (or more with TSO)
505 * - This is the last buffer in a write()/send() and we are
506 * either idle or running NODELAY
507 * - we've timed out (e.g. persist timer)
508 * - we have more then 1/2 the maximum send window's worth of
509 * data (receiver may be limited the window size)
510 * - we need to retransmit
513 if (len >= tp->t_maxseg)
516 * NOTE! on localhost connections an 'ack' from the remote
517 * end may occur synchronously with the output and cause
518 * us to flush a buffer queued with moretocome. XXX
520 * note: the len + off check is almost certainly unnecessary.
522 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
523 (idle || (tp->t_flags & TF_NODELAY)) &&
524 len + off >= so->so_snd.sb_cc &&
525 (tp->t_flags & TF_NOPUSH) == 0) {
528 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
530 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
532 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
539 * Compare available window to amount of window
540 * known to peer (as advertised window less
541 * next expected input). If the difference is at least two
542 * max size segments, or at least 50% of the maximum possible
543 * window, then want to send a window update to peer.
544 * Skip this if the connection is in T/TCP half-open state.
545 * Don't send pure window updates when the peer has closed
546 * the connection and won't ever send more data.
548 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
549 !TCPS_HAVERCVDFIN(tp->t_state)) {
551 * "adv" is the amount we can increase the window,
552 * taking into account that we are limited by
553 * TCP_MAXWIN << tp->rcv_scale.
555 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
556 (tp->rcv_adv - tp->rcv_nxt);
558 if (adv >= (long) (2 * tp->t_maxseg))
560 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
565 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
566 * is also a catch-all for the retransmit timer timeout case.
568 if (tp->t_flags & TF_ACKNOW)
570 if ((flags & TH_RST) ||
571 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
573 if (SEQ_GT(tp->snd_up, tp->snd_una))
576 * If our state indicates that FIN should be sent
577 * and we have not yet done so, then we need to send.
579 if (flags & TH_FIN &&
580 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
583 * In SACK, it is possible for tcp_output to fail to send a segment
584 * after the retransmission timer has been turned off. Make sure
585 * that the retransmission timer is set.
587 if ((tp->t_flags & TF_SACK_PERMIT) &&
588 SEQ_GT(tp->snd_max, tp->snd_una) &&
589 !tcp_timer_active(tp, TT_REXMT) &&
590 !tcp_timer_active(tp, TT_PERSIST)) {
591 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
595 * TCP window updates are not reliable, rather a polling protocol
596 * using ``persist'' packets is used to insure receipt of window
597 * updates. The three ``states'' for the output side are:
598 * idle not doing retransmits or persists
599 * persisting to move a small or zero window
600 * (re)transmitting and thereby not persisting
602 * tcp_timer_active(tp, TT_PERSIST)
603 * is true when we are in persist state.
604 * (tp->t_flags & TF_FORCEDATA)
605 * is set when we are called to send a persist packet.
606 * tcp_timer_active(tp, TT_REXMT)
607 * is set when we are retransmitting
608 * The output side is idle when both timers are zero.
610 * If send window is too small, there is data to transmit, and no
611 * retransmit or persist is pending, then go to persist state.
612 * If nothing happens soon, send when timer expires:
613 * if window is nonzero, transmit what we can,
614 * otherwise force out a byte.
616 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
617 !tcp_timer_active(tp, TT_PERSIST)) {
623 * No reason to send a segment, just return.
626 SOCKBUF_UNLOCK(&so->so_snd);
630 SOCKBUF_LOCK_ASSERT(&so->so_snd);
632 * Before ESTABLISHED, force sending of initial options
633 * unless TCP set not to do any options.
634 * NOTE: we assume that the IP/TCP header plus TCP options
635 * always fit in a single mbuf, leaving room for a maximum
637 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
642 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
645 hdrlen = sizeof (struct tcpiphdr);
648 * Compute options for segment.
649 * We only have to care about SYN and established connection
650 * segments. Options for SYN-ACK segments are handled in TCP
653 if ((tp->t_flags & TF_NOOPT) == 0) {
655 /* Maximum segment size. */
656 if (flags & TH_SYN) {
657 tp->snd_nxt = tp->iss;
658 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
659 to.to_flags |= TOF_MSS;
661 /* Window scaling. */
662 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
663 to.to_wscale = tp->request_r_scale;
664 to.to_flags |= TOF_SCALE;
667 if ((tp->t_flags & TF_RCVD_TSTMP) ||
668 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
669 to.to_tsval = ticks + tp->ts_offset;
670 to.to_tsecr = tp->ts_recent;
671 to.to_flags |= TOF_TS;
672 /* Set receive buffer autosizing timestamp. */
673 if (tp->rfbuf_ts == 0 &&
674 (so->so_rcv.sb_flags & SB_AUTOSIZE))
675 tp->rfbuf_ts = ticks;
677 /* Selective ACK's. */
678 if (tp->t_flags & TF_SACK_PERMIT) {
680 to.to_flags |= TOF_SACKPERM;
681 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
682 (tp->t_flags & TF_SACK_PERMIT) &&
683 tp->rcv_numsacks > 0) {
684 to.to_flags |= TOF_SACK;
685 to.to_nsacks = tp->rcv_numsacks;
686 to.to_sacks = (u_char *)tp->sackblks;
690 /* TCP-MD5 (RFC2385). */
691 if (tp->t_flags & TF_SIGNATURE)
692 to.to_flags |= TOF_SIGNATURE;
693 #endif /* TCP_SIGNATURE */
695 /* Processing the options. */
696 hdrlen += optlen = tcp_addoptions(&to, opt);
701 ipoptlen = ip6_optlen(tp->t_inpcb);
704 if (tp->t_inpcb->inp_options)
705 ipoptlen = tp->t_inpcb->inp_options->m_len -
706 offsetof(struct ipoption, ipopt_list);
710 ipoptlen += ipsec_optlen;
714 * Adjust data length if insertion of options will
715 * bump the packet length beyond the t_maxopd length.
716 * Clear the FIN bit because we cut off the tail of
719 * When doing TSO limit a burst to TCP_MAXWIN minus the
720 * IP, TCP and Options length to keep ip->ip_len from
721 * overflowing. Prevent the last segment from being
722 * fractional thus making them all equal sized and set
723 * the flag to continue sending. TSO is disabled when
724 * IP options or IPSEC are present.
726 if (len + optlen + ipoptlen > tp->t_maxopd) {
729 if (len > TCP_MAXWIN - hdrlen - optlen) {
730 len = TCP_MAXWIN - hdrlen - optlen;
731 len = len - (len % (tp->t_maxopd - optlen));
733 } else if (tp->t_flags & TF_NEEDFIN)
736 len = tp->t_maxopd - optlen - ipoptlen;
741 /*#ifdef DIAGNOSTIC*/
743 if (max_linkhdr + hdrlen > MCLBYTES)
745 if (max_linkhdr + hdrlen > MHLEN)
747 panic("tcphdr too big");
751 * This KASSERT is here to catch edge cases at a well defined place.
752 * Before, those had triggered (random) panic conditions further down.
754 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
757 * Grab a header mbuf, attaching a copy of data to
758 * be transmitted, and initialize the header from
759 * the template for sends on this connection.
765 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
766 V_tcpstat.tcps_sndprobe++;
767 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
768 V_tcpstat.tcps_sndrexmitpack++;
769 V_tcpstat.tcps_sndrexmitbyte += len;
771 V_tcpstat.tcps_sndpack++;
772 V_tcpstat.tcps_sndbyte += len;
775 if ((m = m_copypack(so->so_snd.sb_mb, off,
776 (int)len, max_linkhdr + hdrlen)) == 0) {
777 SOCKBUF_UNLOCK(&so->so_snd);
782 * m_copypack left space for our hdr; use it.
787 MGETHDR(m, M_DONTWAIT, MT_DATA);
789 SOCKBUF_UNLOCK(&so->so_snd);
794 if (MHLEN < hdrlen + max_linkhdr) {
795 MCLGET(m, M_DONTWAIT);
796 if ((m->m_flags & M_EXT) == 0) {
797 SOCKBUF_UNLOCK(&so->so_snd);
804 m->m_data += max_linkhdr;
808 * Start the m_copy functions from the closest mbuf
809 * to the offset in the socket buffer chain.
811 mb = sbsndptr(&so->so_snd, off, len, &moff);
813 if (len <= MHLEN - hdrlen - max_linkhdr) {
814 m_copydata(mb, moff, (int)len,
815 mtod(m, caddr_t) + hdrlen);
818 m->m_next = m_copy(mb, moff, (int)len);
819 if (m->m_next == NULL) {
820 SOCKBUF_UNLOCK(&so->so_snd);
828 * If we're sending everything we've got, set PUSH.
829 * (This will keep happy those implementations which only
830 * give data to the user when a buffer fills or
833 if (off + len == so->so_snd.sb_cc)
835 SOCKBUF_UNLOCK(&so->so_snd);
837 SOCKBUF_UNLOCK(&so->so_snd);
838 if (tp->t_flags & TF_ACKNOW)
839 V_tcpstat.tcps_sndacks++;
840 else if (flags & (TH_SYN|TH_FIN|TH_RST))
841 V_tcpstat.tcps_sndctrl++;
842 else if (SEQ_GT(tp->snd_up, tp->snd_una))
843 V_tcpstat.tcps_sndurg++;
845 V_tcpstat.tcps_sndwinup++;
847 MGETHDR(m, M_DONTWAIT, MT_DATA);
853 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
858 m->m_data += max_linkhdr;
861 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
862 m->m_pkthdr.rcvif = (struct ifnet *)0;
864 mac_inpcb_create_mbuf(tp->t_inpcb, m);
868 ip6 = mtod(m, struct ip6_hdr *);
869 th = (struct tcphdr *)(ip6 + 1);
870 tcpip_fillheaders(tp->t_inpcb, ip6, th);
874 ip = mtod(m, struct ip *);
875 ipov = (struct ipovly *)ip;
876 th = (struct tcphdr *)(ip + 1);
877 tcpip_fillheaders(tp->t_inpcb, ip, th);
881 * Fill in fields, remembering maximum advertised
882 * window for use in delaying messages about window sizes.
883 * If resending a FIN, be sure not to use a new sequence number.
885 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
886 tp->snd_nxt == tp->snd_max)
889 * If we are starting a connection, send ECN setup
890 * SYN packet. If we are on a retransmit, we may
891 * resend those bits a number of times as per
894 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
895 if (tp->t_rxtshift >= 1) {
896 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
897 flags |= TH_ECE|TH_CWR;
899 flags |= TH_ECE|TH_CWR;
902 if (tp->t_state == TCPS_ESTABLISHED &&
903 (tp->t_flags & TF_ECN_PERMIT)) {
905 * If the peer has ECN, mark data packets with
906 * ECN capable transmission (ECT).
907 * Ignore pure ack packets, retransmissions and window probes.
909 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
910 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
913 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
916 ip->ip_tos |= IPTOS_ECN_ECT0;
917 V_tcpstat.tcps_ecn_ect0++;
921 * Reply with proper ECN notifications.
923 if (tp->t_flags & TF_ECN_SND_CWR) {
925 tp->t_flags &= ~TF_ECN_SND_CWR;
927 if (tp->t_flags & TF_ECN_SND_ECE)
932 * If we are doing retransmissions, then snd_nxt will
933 * not reflect the first unsent octet. For ACK only
934 * packets, we do not want the sequence number of the
935 * retransmitted packet, we want the sequence number
936 * of the next unsent octet. So, if there is no data
937 * (and no SYN or FIN), use snd_max instead of snd_nxt
938 * when filling in ti_seq. But if we are in persist
939 * state, snd_max might reflect one byte beyond the
940 * right edge of the window, so use snd_nxt in that
941 * case, since we know we aren't doing a retransmission.
942 * (retransmit and persist are mutually exclusive...)
944 if (sack_rxmit == 0) {
945 if (len || (flags & (TH_SYN|TH_FIN)) ||
946 tcp_timer_active(tp, TT_PERSIST))
947 th->th_seq = htonl(tp->snd_nxt);
949 th->th_seq = htonl(tp->snd_max);
951 th->th_seq = htonl(p->rxmit);
953 tp->sackhint.sack_bytes_rexmit += len;
955 th->th_ack = htonl(tp->rcv_nxt);
957 bcopy(opt, th + 1, optlen);
958 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
960 th->th_flags = flags;
962 * Calculate receive window. Don't shrink window,
963 * but avoid silly window syndrome.
965 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
966 recwin < (long)tp->t_maxseg)
968 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
969 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
970 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
971 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
974 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
975 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
976 * case is handled in syncache.
979 th->th_win = htons((u_short)
980 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
982 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
985 * Adjust the RXWIN0SENT flag - indicate that we have advertised
986 * a 0 window. This may cause the remote transmitter to stall. This
987 * flag tells soreceive() to disable delayed acknowledgements when
988 * draining the buffer. This can occur if the receiver is attempting
989 * to read more data than can be buffered prior to transmitting on
993 tp->t_flags |= TF_RXWIN0SENT;
995 tp->t_flags &= ~TF_RXWIN0SENT;
996 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
997 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
998 th->th_flags |= TH_URG;
1001 * If no urgent pointer to send, then we pull
1002 * the urgent pointer to the left edge of the send window
1003 * so that it doesn't drift into the send window on sequence
1004 * number wraparound.
1006 tp->snd_up = tp->snd_una; /* drag it along */
1008 #ifdef TCP_SIGNATURE
1009 if (tp->t_flags & TF_SIGNATURE) {
1010 int sigoff = to.to_signature - opt;
1011 tcp_signature_compute(m, 0, len, optlen,
1012 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1017 * Put TCP length in extended header, and then
1018 * checksum extended header and data.
1020 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1024 * ip6_plen is not need to be filled now, and will be filled
1027 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1028 sizeof(struct tcphdr) + optlen + len);
1032 m->m_pkthdr.csum_flags = CSUM_TCP;
1033 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1034 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1035 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1037 /* IP version must be set here for ipv4/ipv6 checking later */
1038 KASSERT(ip->ip_v == IPVERSION,
1039 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1043 * Enable TSO and specify the size of the segments.
1044 * The TCP pseudo header checksum is always provided.
1045 * XXX: Fixme: This is currently not the case for IPv6.
1048 m->m_pkthdr.csum_flags = CSUM_TSO;
1049 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1053 * In transmit state, time the transmission and arrange for
1054 * the retransmit. In persist state, just set snd_max.
1056 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1057 !tcp_timer_active(tp, TT_PERSIST)) {
1058 tcp_seq startseq = tp->snd_nxt;
1061 * Advance snd_nxt over sequence space of this segment.
1063 if (flags & (TH_SYN|TH_FIN)) {
1066 if (flags & TH_FIN) {
1068 tp->t_flags |= TF_SENTFIN;
1074 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1075 tp->snd_max = tp->snd_nxt;
1077 * Time this transmission if not a retransmission and
1078 * not currently timing anything.
1080 if (tp->t_rtttime == 0) {
1081 tp->t_rtttime = ticks;
1082 tp->t_rtseq = startseq;
1083 V_tcpstat.tcps_segstimed++;
1088 * Set retransmit timer if not currently set,
1089 * and not doing a pure ack or a keep-alive probe.
1090 * Initial value for retransmit timer is smoothed
1091 * round-trip time + 2 * round-trip time variance.
1092 * Initialize shift counter which is used for backoff
1093 * of retransmit time.
1096 if (!tcp_timer_active(tp, TT_REXMT) &&
1097 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1098 (tp->snd_nxt != tp->snd_una))) {
1099 if (tcp_timer_active(tp, TT_PERSIST)) {
1100 tcp_timer_activate(tp, TT_PERSIST, 0);
1103 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1107 * Persist case, update snd_max but since we are in
1108 * persist mode (no window) we do not update snd_nxt.
1113 if (flags & TH_FIN) {
1115 tp->t_flags |= TF_SENTFIN;
1117 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1118 tp->snd_max = tp->snd_nxt + len;
1125 if (so->so_options & SO_DEBUG) {
1131 save = ipov->ih_len;
1132 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1134 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1138 ipov->ih_len = save;
1143 * Fill in IP length and desired time to live and
1144 * send to IP level. There should be a better way
1145 * to handle ttl and tos; we could keep them in
1146 * the template, but need a way to checksum without them.
1149 * m->m_pkthdr.len should have been set before cksum calcuration,
1150 * because in6_cksum() need it.
1155 * we separately set hoplimit for every segment, since the
1156 * user might want to change the value via setsockopt.
1157 * Also, desired default hop limit might be changed via
1158 * Neighbor Discovery.
1160 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1162 /* TODO: IPv6 IP6TOS_ECT bit on */
1163 error = ip6_output(m,
1164 tp->t_inpcb->in6p_outputopts, NULL,
1165 ((so->so_options & SO_DONTROUTE) ?
1166 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1170 ip->ip_len = m->m_pkthdr.len;
1172 if (INP_CHECK_SOCKAF(so, AF_INET6))
1173 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1176 * If we do path MTU discovery, then we set DF on every packet.
1177 * This might not be the best thing to do according to RFC3390
1178 * Section 2. However the tcp hostcache migitates the problem
1179 * so it affects only the first tcp connection with a host.
1181 if (V_path_mtu_discovery)
1182 ip->ip_off |= IP_DF;
1184 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1185 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1191 * We know that the packet was lost, so back out the
1192 * sequence number advance, if any.
1194 * If the error is EPERM the packet got blocked by the
1195 * local firewall. Normally we should terminate the
1196 * connection but the blocking may have been spurious
1197 * due to a firewall reconfiguration cycle. So we treat
1198 * it like a packet loss and let the retransmit timer and
1199 * timeouts do their work over time.
1200 * XXX: It is a POLA question whether calling tcp_drop right
1201 * away would be the really correct behavior instead.
1203 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1204 !tcp_timer_active(tp, TT_PERSIST)) &&
1205 ((flags & TH_SYN) == 0) &&
1209 tp->sackhint.sack_bytes_rexmit -= len;
1210 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1211 ("sackhint bytes rtx >= 0"));
1216 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1219 tp->t_softerror = error;
1222 if (!tcp_timer_active(tp, TT_REXMT) &&
1223 !tcp_timer_active(tp, TT_PERSIST))
1224 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1225 tp->snd_cwnd = tp->t_maxseg;
1229 * For some reason the interface we used initially
1230 * to send segments changed to another or lowered
1233 * tcp_mtudisc() will find out the new MTU and as
1234 * its last action, initiate retransmission, so it
1235 * is important to not do so here.
1237 * If TSO was active we either got an interface
1238 * without TSO capabilits or TSO was turned off.
1239 * Disable it for this connection as too and
1240 * immediatly retry with MSS sized segments generated
1244 tp->t_flags &= ~TF_TSO;
1245 tcp_mtudisc(tp->t_inpcb, 0);
1251 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1252 tp->t_softerror = error;
1260 V_tcpstat.tcps_sndtotal++;
1263 * Data sent (as far as we can tell).
1264 * If this advertises a larger window than any other segment,
1265 * then remember the size of the advertised window.
1266 * Any pending ACK has now been sent.
1268 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1269 tp->rcv_adv = tp->rcv_nxt + recwin;
1270 tp->last_ack_sent = tp->rcv_nxt;
1271 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1272 if (tcp_timer_active(tp, TT_DELACK))
1273 tcp_timer_activate(tp, TT_DELACK, 0);
1276 * This completely breaks TCP if newreno is turned on. What happens
1277 * is that if delayed-acks are turned on on the receiver, this code
1278 * on the transmitter effectively destroys the TCP window, forcing
1279 * it to four packets (1.5Kx4 = 6K window).
1281 if (sendalot && (!V_tcp_do_newreno || --maxburst))
1290 tcp_setpersist(struct tcpcb *tp)
1292 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1295 if (tcp_timer_active(tp, TT_REXMT))
1296 panic("tcp_setpersist: retransmit pending");
1298 * Start/restart persistance timer.
1300 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1301 TCPTV_PERSMIN, TCPTV_PERSMAX);
1302 tcp_timer_activate(tp, TT_PERSIST, tt);
1303 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1308 * Insert TCP options according to the supplied parameters to the place
1309 * optp in a consistent way. Can handle unaligned destinations.
1311 * The order of the option processing is crucial for optimal packing and
1312 * alignment for the scarce option space.
1314 * The optimal order for a SYN/SYN-ACK segment is:
1315 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1316 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1318 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1319 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1320 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1321 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1322 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1325 tcp_addoptions(struct tcpopt *to, u_char *optp)
1327 INIT_VNET_INET(curvnet);
1328 u_int mask, optlen = 0;
1330 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1331 if ((to->to_flags & mask) != mask)
1333 if (optlen == TCP_MAXOLEN)
1335 switch (to->to_flags & mask) {
1337 while (optlen % 4) {
1338 optlen += TCPOLEN_NOP;
1339 *optp++ = TCPOPT_NOP;
1341 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1343 optlen += TCPOLEN_MAXSEG;
1344 *optp++ = TCPOPT_MAXSEG;
1345 *optp++ = TCPOLEN_MAXSEG;
1346 to->to_mss = htons(to->to_mss);
1347 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1348 optp += sizeof(to->to_mss);
1351 while (!optlen || optlen % 2 != 1) {
1352 optlen += TCPOLEN_NOP;
1353 *optp++ = TCPOPT_NOP;
1355 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1357 optlen += TCPOLEN_WINDOW;
1358 *optp++ = TCPOPT_WINDOW;
1359 *optp++ = TCPOLEN_WINDOW;
1360 *optp++ = to->to_wscale;
1363 while (optlen % 2) {
1364 optlen += TCPOLEN_NOP;
1365 *optp++ = TCPOPT_NOP;
1367 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1369 optlen += TCPOLEN_SACK_PERMITTED;
1370 *optp++ = TCPOPT_SACK_PERMITTED;
1371 *optp++ = TCPOLEN_SACK_PERMITTED;
1374 while (!optlen || optlen % 4 != 2) {
1375 optlen += TCPOLEN_NOP;
1376 *optp++ = TCPOPT_NOP;
1378 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1380 optlen += TCPOLEN_TIMESTAMP;
1381 *optp++ = TCPOPT_TIMESTAMP;
1382 *optp++ = TCPOLEN_TIMESTAMP;
1383 to->to_tsval = htonl(to->to_tsval);
1384 to->to_tsecr = htonl(to->to_tsecr);
1385 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1386 optp += sizeof(to->to_tsval);
1387 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1388 optp += sizeof(to->to_tsecr);
1392 int siglen = TCPOLEN_SIGNATURE - 2;
1394 while (!optlen || optlen % 4 != 2) {
1395 optlen += TCPOLEN_NOP;
1396 *optp++ = TCPOPT_NOP;
1398 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1400 optlen += TCPOLEN_SIGNATURE;
1401 *optp++ = TCPOPT_SIGNATURE;
1402 *optp++ = TCPOLEN_SIGNATURE;
1403 to->to_signature = optp;
1411 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1414 while (!optlen || optlen % 4 != 2) {
1415 optlen += TCPOLEN_NOP;
1416 *optp++ = TCPOPT_NOP;
1418 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1420 optlen += TCPOLEN_SACKHDR;
1421 *optp++ = TCPOPT_SACK;
1422 sackblks = min(to->to_nsacks,
1423 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1424 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1425 while (sackblks--) {
1426 sack_seq = htonl(sack->start);
1427 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1428 optp += sizeof(sack_seq);
1429 sack_seq = htonl(sack->end);
1430 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1431 optp += sizeof(sack_seq);
1432 optlen += TCPOLEN_SACK;
1435 V_tcpstat.tcps_sack_send_blocks++;
1439 panic("%s: unknown TCP option type", __func__);
1444 /* Terminate and pad TCP options to a 4 byte boundary. */
1446 optlen += TCPOLEN_EOL;
1447 *optp++ = TCPOPT_EOL;
1450 * According to RFC 793 (STD0007):
1451 * "The content of the header beyond the End-of-Option option
1452 * must be header padding (i.e., zero)."
1453 * and later: "The padding is composed of zeros."
1455 while (optlen % 4) {
1456 optlen += TCPOLEN_PAD;
1457 *optp++ = TCPOPT_PAD;
1460 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));