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))
214 off = tp->snd_nxt - tp->snd_una;
215 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
216 sendwin = min(sendwin, tp->snd_bwnd);
218 flags = tcp_outflags[tp->t_state];
220 * Send any SACK-generated retransmissions. If we're explicitly trying
221 * to send out new data (when sendalot is 1), bypass this function.
222 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
223 * we're replacing a (future) new transmission with a retransmission
224 * now, and we previously incremented snd_cwnd in tcp_input().
227 * Still in sack recovery , reset rxmit flag to zero.
233 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp) &&
234 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
237 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
240 /* Do not retransmit SACK segments beyond snd_recover */
241 if (SEQ_GT(p->end, tp->snd_recover)) {
243 * (At least) part of sack hole extends beyond
244 * snd_recover. Check to see if we can rexmit data
247 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
249 * Can't rexmit any more data for this hole.
250 * That data will be rexmitted in the next
251 * sack recovery episode, when snd_recover
252 * moves past p->rxmit.
255 goto after_sack_rexmit;
257 /* Can rexmit part of the current hole */
258 len = ((long)ulmin(cwin,
259 tp->snd_recover - p->rxmit));
261 len = ((long)ulmin(cwin, p->end - p->rxmit));
262 off = p->rxmit - tp->snd_una;
263 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
268 TCPSTAT_INC(tcps_sack_rexmits);
269 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
270 min(len, tp->t_maxseg));
275 * Get standard flags, and add SYN or FIN if requested by 'hidden'
278 if (tp->t_flags & TF_NEEDFIN)
280 if (tp->t_flags & TF_NEEDSYN)
283 SOCKBUF_LOCK(&so->so_snd);
285 * If in persist timeout with window of 0, send 1 byte.
286 * Otherwise, if window is small but nonzero
287 * and timer expired, we will send what we can
288 * and go to transmit state.
290 if (tp->t_flags & TF_FORCEDATA) {
293 * If we still have some data to send, then
294 * clear the FIN bit. Usually this would
295 * happen below when it realizes that we
296 * aren't sending all the data. However,
297 * if we have exactly 1 byte of unsent data,
298 * then it won't clear the FIN bit below,
299 * and if we are in persist state, we wind
300 * up sending the packet without recording
301 * that we sent the FIN bit.
303 * We can't just blindly clear the FIN bit,
304 * because if we don't have any more data
305 * to send then the probe will be the FIN
308 if (off < so->so_snd.sb_cc)
312 tcp_timer_activate(tp, TT_PERSIST, 0);
318 * If snd_nxt == snd_max and we have transmitted a FIN, the
319 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
320 * a negative length. This can also occur when TCP opens up
321 * its congestion window while receiving additional duplicate
322 * acks after fast-retransmit because TCP will reset snd_nxt
323 * to snd_max after the fast-retransmit.
325 * In the normal retransmit-FIN-only case, however, snd_nxt will
326 * be set to snd_una, the offset will be 0, and the length may
329 * If sack_rxmit is true we are retransmitting from the scoreboard
330 * in which case len is already set.
332 if (sack_rxmit == 0) {
333 if (sack_bytes_rxmt == 0)
334 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
339 * We are inside of a SACK recovery episode and are
340 * sending new data, having retransmitted all the
341 * data possible in the scoreboard.
343 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
346 * Don't remove this (len > 0) check !
347 * We explicitly check for len > 0 here (although it
348 * isn't really necessary), to work around a gcc
349 * optimization issue - to force gcc to compute
350 * len above. Without this check, the computation
351 * of len is bungled by the optimizer.
354 cwin = tp->snd_cwnd -
355 (tp->snd_nxt - tp->sack_newdata) -
359 len = lmin(len, cwin);
365 * Lop off SYN bit if it has already been sent. However, if this
366 * is SYN-SENT state and if segment contains data and if we don't
367 * know that foreign host supports TAO, suppress sending segment.
369 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
370 if (tp->t_state != TCPS_SYN_RECEIVED)
376 * Be careful not to send data and/or FIN on SYN segments.
377 * This measure is needed to prevent interoperability problems
378 * with not fully conformant TCP implementations.
380 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
387 * If FIN has been sent but not acked,
388 * but we haven't been called to retransmit,
389 * len will be < 0. Otherwise, window shrank
390 * after we sent into it. If window shrank to 0,
391 * cancel pending retransmit, pull snd_nxt back
392 * to (closed) window, and set the persist timer
393 * if it isn't already going. If the window didn't
394 * close completely, just wait for an ACK.
398 tcp_timer_activate(tp, TT_REXMT, 0);
400 tp->snd_nxt = tp->snd_una;
401 if (!tcp_timer_active(tp, TT_PERSIST))
406 /* len will be >= 0 after this point. */
407 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
410 * Automatic sizing of send socket buffer. Often the send buffer
411 * size is not optimally adjusted to the actual network conditions
412 * at hand (delay bandwidth product). Setting the buffer size too
413 * small limits throughput on links with high bandwidth and high
414 * delay (eg. trans-continental/oceanic links). Setting the
415 * buffer size too big consumes too much real kernel memory,
416 * especially with many connections on busy servers.
418 * The criteria to step up the send buffer one notch are:
419 * 1. receive window of remote host is larger than send buffer
420 * (with a fudge factor of 5/4th);
421 * 2. send buffer is filled to 7/8th with data (so we actually
422 * have data to make use of it);
423 * 3. send buffer fill has not hit maximal automatic size;
424 * 4. our send window (slow start and cogestion controlled) is
425 * larger than sent but unacknowledged data in send buffer.
427 * The remote host receive window scaling factor may limit the
428 * growing of the send buffer before it reaches its allowed
431 * It scales directly with slow start or congestion window
432 * and does at most one step per received ACK. This fast
433 * scaling has the drawback of growing the send buffer beyond
434 * what is strictly necessary to make full use of a given
435 * delay*bandwith product. However testing has shown this not
436 * to be much of an problem. At worst we are trading wasting
437 * of available bandwith (the non-use of it) for wasting some
438 * socket buffer memory.
440 * TODO: Shrink send buffer during idle periods together
441 * with congestion window. Requires another timer. Has to
442 * wait for upcoming tcp timer rewrite.
444 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
445 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
446 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
447 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
448 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
449 if (!sbreserve_locked(&so->so_snd,
450 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
451 V_tcp_autosndbuf_max), so, curthread))
452 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
457 * Truncate to the maximum segment length or enable TCP Segmentation
458 * Offloading (if supported by hardware) and ensure that FIN is removed
459 * if the length no longer contains the last data byte.
461 * TSO may only be used if we are in a pure bulk sending state. The
462 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
463 * IP options prevent using TSO. With TSO the TCP header is the same
464 * (except for the sequence number) for all generated packets. This
465 * makes it impossible to transmit any options which vary per generated
468 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
469 * removal of FIN (if not already catched here) are handled later after
470 * the exact length of the TCP options are known.
474 * Pre-calculate here as we save another lookup into the darknesses
475 * of IPsec that way and can actually decide if TSO is ok.
477 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
479 if (len > tp->t_maxseg) {
480 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
481 ((tp->t_flags & TF_SIGNATURE) == 0) &&
482 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
483 tp->t_inpcb->inp_options == NULL &&
484 tp->t_inpcb->in6p_options == NULL
497 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
500 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
504 recwin = sbspace(&so->so_rcv);
507 * Sender silly window avoidance. We transmit under the following
508 * conditions when len is non-zero:
510 * - We have a full segment (or more with TSO)
511 * - This is the last buffer in a write()/send() and we are
512 * either idle or running NODELAY
513 * - we've timed out (e.g. persist timer)
514 * - we have more then 1/2 the maximum send window's worth of
515 * data (receiver may be limited the window size)
516 * - we need to retransmit
519 if (len >= tp->t_maxseg)
522 * NOTE! on localhost connections an 'ack' from the remote
523 * end may occur synchronously with the output and cause
524 * us to flush a buffer queued with moretocome. XXX
526 * note: the len + off check is almost certainly unnecessary.
528 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
529 (idle || (tp->t_flags & TF_NODELAY)) &&
530 len + off >= so->so_snd.sb_cc &&
531 (tp->t_flags & TF_NOPUSH) == 0) {
534 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
536 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
538 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
545 * Compare available window to amount of window
546 * known to peer (as advertised window less
547 * next expected input). If the difference is at least two
548 * max size segments, or at least 50% of the maximum possible
549 * window, then want to send a window update to peer.
550 * Skip this if the connection is in T/TCP half-open state.
551 * Don't send pure window updates when the peer has closed
552 * the connection and won't ever send more data.
554 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
555 !TCPS_HAVERCVDFIN(tp->t_state)) {
557 * "adv" is the amount we can increase the window,
558 * taking into account that we are limited by
559 * TCP_MAXWIN << tp->rcv_scale.
561 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
562 (tp->rcv_adv - tp->rcv_nxt);
564 if (adv >= (long) (2 * tp->t_maxseg))
566 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
571 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
572 * is also a catch-all for the retransmit timer timeout case.
574 if (tp->t_flags & TF_ACKNOW)
576 if ((flags & TH_RST) ||
577 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
579 if (SEQ_GT(tp->snd_up, tp->snd_una))
582 * If our state indicates that FIN should be sent
583 * and we have not yet done so, then we need to send.
585 if (flags & TH_FIN &&
586 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
589 * In SACK, it is possible for tcp_output to fail to send a segment
590 * after the retransmission timer has been turned off. Make sure
591 * that the retransmission timer is set.
593 if ((tp->t_flags & TF_SACK_PERMIT) &&
594 SEQ_GT(tp->snd_max, tp->snd_una) &&
595 !tcp_timer_active(tp, TT_REXMT) &&
596 !tcp_timer_active(tp, TT_PERSIST)) {
597 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
601 * TCP window updates are not reliable, rather a polling protocol
602 * using ``persist'' packets is used to insure receipt of window
603 * updates. The three ``states'' for the output side are:
604 * idle not doing retransmits or persists
605 * persisting to move a small or zero window
606 * (re)transmitting and thereby not persisting
608 * tcp_timer_active(tp, TT_PERSIST)
609 * is true when we are in persist state.
610 * (tp->t_flags & TF_FORCEDATA)
611 * is set when we are called to send a persist packet.
612 * tcp_timer_active(tp, TT_REXMT)
613 * is set when we are retransmitting
614 * The output side is idle when both timers are zero.
616 * If send window is too small, there is data to transmit, and no
617 * retransmit or persist is pending, then go to persist state.
618 * If nothing happens soon, send when timer expires:
619 * if window is nonzero, transmit what we can,
620 * otherwise force out a byte.
622 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
623 !tcp_timer_active(tp, TT_PERSIST)) {
629 * No reason to send a segment, just return.
632 SOCKBUF_UNLOCK(&so->so_snd);
636 SOCKBUF_LOCK_ASSERT(&so->so_snd);
638 * Before ESTABLISHED, force sending of initial options
639 * unless TCP set not to do any options.
640 * NOTE: we assume that the IP/TCP header plus TCP options
641 * always fit in a single mbuf, leaving room for a maximum
643 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
648 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
651 hdrlen = sizeof (struct tcpiphdr);
654 * Compute options for segment.
655 * We only have to care about SYN and established connection
656 * segments. Options for SYN-ACK segments are handled in TCP
659 if ((tp->t_flags & TF_NOOPT) == 0) {
661 /* Maximum segment size. */
662 if (flags & TH_SYN) {
663 tp->snd_nxt = tp->iss;
664 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
665 to.to_flags |= TOF_MSS;
667 /* Window scaling. */
668 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
669 to.to_wscale = tp->request_r_scale;
670 to.to_flags |= TOF_SCALE;
673 if ((tp->t_flags & TF_RCVD_TSTMP) ||
674 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
675 to.to_tsval = ticks + tp->ts_offset;
676 to.to_tsecr = tp->ts_recent;
677 to.to_flags |= TOF_TS;
678 /* Set receive buffer autosizing timestamp. */
679 if (tp->rfbuf_ts == 0 &&
680 (so->so_rcv.sb_flags & SB_AUTOSIZE))
681 tp->rfbuf_ts = ticks;
683 /* Selective ACK's. */
684 if (tp->t_flags & TF_SACK_PERMIT) {
686 to.to_flags |= TOF_SACKPERM;
687 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
688 (tp->t_flags & TF_SACK_PERMIT) &&
689 tp->rcv_numsacks > 0) {
690 to.to_flags |= TOF_SACK;
691 to.to_nsacks = tp->rcv_numsacks;
692 to.to_sacks = (u_char *)tp->sackblks;
696 /* TCP-MD5 (RFC2385). */
697 if (tp->t_flags & TF_SIGNATURE)
698 to.to_flags |= TOF_SIGNATURE;
699 #endif /* TCP_SIGNATURE */
701 /* Processing the options. */
702 hdrlen += optlen = tcp_addoptions(&to, opt);
707 ipoptlen = ip6_optlen(tp->t_inpcb);
710 if (tp->t_inpcb->inp_options)
711 ipoptlen = tp->t_inpcb->inp_options->m_len -
712 offsetof(struct ipoption, ipopt_list);
716 ipoptlen += ipsec_optlen;
720 * Adjust data length if insertion of options will
721 * bump the packet length beyond the t_maxopd length.
722 * Clear the FIN bit because we cut off the tail of
725 * When doing TSO limit a burst to TCP_MAXWIN minus the
726 * IP, TCP and Options length to keep ip->ip_len from
727 * overflowing. Prevent the last segment from being
728 * fractional thus making them all equal sized and set
729 * the flag to continue sending. TSO is disabled when
730 * IP options or IPSEC are present.
732 if (len + optlen + ipoptlen > tp->t_maxopd) {
735 if (len > TCP_MAXWIN - hdrlen - optlen) {
736 len = TCP_MAXWIN - hdrlen - optlen;
737 len = len - (len % (tp->t_maxopd - optlen));
739 } else if (tp->t_flags & TF_NEEDFIN)
742 len = tp->t_maxopd - optlen - ipoptlen;
747 /*#ifdef DIAGNOSTIC*/
749 if (max_linkhdr + hdrlen > MCLBYTES)
751 if (max_linkhdr + hdrlen > MHLEN)
753 panic("tcphdr too big");
757 * This KASSERT is here to catch edge cases at a well defined place.
758 * Before, those had triggered (random) panic conditions further down.
760 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
763 * Grab a header mbuf, attaching a copy of data to
764 * be transmitted, and initialize the header from
765 * the template for sends on this connection.
771 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
772 TCPSTAT_INC(tcps_sndprobe);
773 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
774 TCPSTAT_INC(tcps_sndrexmitpack);
775 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
777 TCPSTAT_INC(tcps_sndpack);
778 TCPSTAT_ADD(tcps_sndbyte, len);
781 if ((m = m_copypack(so->so_snd.sb_mb, off,
782 (int)len, max_linkhdr + hdrlen)) == 0) {
783 SOCKBUF_UNLOCK(&so->so_snd);
788 * m_copypack left space for our hdr; use it.
793 MGETHDR(m, M_DONTWAIT, MT_DATA);
795 SOCKBUF_UNLOCK(&so->so_snd);
800 if (MHLEN < hdrlen + max_linkhdr) {
801 MCLGET(m, M_DONTWAIT);
802 if ((m->m_flags & M_EXT) == 0) {
803 SOCKBUF_UNLOCK(&so->so_snd);
810 m->m_data += max_linkhdr;
814 * Start the m_copy functions from the closest mbuf
815 * to the offset in the socket buffer chain.
817 mb = sbsndptr(&so->so_snd, off, len, &moff);
819 if (len <= MHLEN - hdrlen - max_linkhdr) {
820 m_copydata(mb, moff, (int)len,
821 mtod(m, caddr_t) + hdrlen);
824 m->m_next = m_copy(mb, moff, (int)len);
825 if (m->m_next == NULL) {
826 SOCKBUF_UNLOCK(&so->so_snd);
834 * If we're sending everything we've got, set PUSH.
835 * (This will keep happy those implementations which only
836 * give data to the user when a buffer fills or
839 if (off + len == so->so_snd.sb_cc)
841 SOCKBUF_UNLOCK(&so->so_snd);
843 SOCKBUF_UNLOCK(&so->so_snd);
844 if (tp->t_flags & TF_ACKNOW)
845 TCPSTAT_INC(tcps_sndacks);
846 else if (flags & (TH_SYN|TH_FIN|TH_RST))
847 TCPSTAT_INC(tcps_sndctrl);
848 else if (SEQ_GT(tp->snd_up, tp->snd_una))
849 TCPSTAT_INC(tcps_sndurg);
851 TCPSTAT_INC(tcps_sndwinup);
853 MGETHDR(m, M_DONTWAIT, MT_DATA);
859 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
864 m->m_data += max_linkhdr;
867 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
868 m->m_pkthdr.rcvif = (struct ifnet *)0;
870 mac_inpcb_create_mbuf(tp->t_inpcb, m);
874 ip6 = mtod(m, struct ip6_hdr *);
875 th = (struct tcphdr *)(ip6 + 1);
876 tcpip_fillheaders(tp->t_inpcb, ip6, th);
880 ip = mtod(m, struct ip *);
881 ipov = (struct ipovly *)ip;
882 th = (struct tcphdr *)(ip + 1);
883 tcpip_fillheaders(tp->t_inpcb, ip, th);
887 * Fill in fields, remembering maximum advertised
888 * window for use in delaying messages about window sizes.
889 * If resending a FIN, be sure not to use a new sequence number.
891 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
892 tp->snd_nxt == tp->snd_max)
895 * If we are starting a connection, send ECN setup
896 * SYN packet. If we are on a retransmit, we may
897 * resend those bits a number of times as per
900 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
901 if (tp->t_rxtshift >= 1) {
902 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
903 flags |= TH_ECE|TH_CWR;
905 flags |= TH_ECE|TH_CWR;
908 if (tp->t_state == TCPS_ESTABLISHED &&
909 (tp->t_flags & TF_ECN_PERMIT)) {
911 * If the peer has ECN, mark data packets with
912 * ECN capable transmission (ECT).
913 * Ignore pure ack packets, retransmissions and window probes.
915 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
916 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
919 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
922 ip->ip_tos |= IPTOS_ECN_ECT0;
923 TCPSTAT_INC(tcps_ecn_ect0);
927 * Reply with proper ECN notifications.
929 if (tp->t_flags & TF_ECN_SND_CWR) {
931 tp->t_flags &= ~TF_ECN_SND_CWR;
933 if (tp->t_flags & TF_ECN_SND_ECE)
938 * If we are doing retransmissions, then snd_nxt will
939 * not reflect the first unsent octet. For ACK only
940 * packets, we do not want the sequence number of the
941 * retransmitted packet, we want the sequence number
942 * of the next unsent octet. So, if there is no data
943 * (and no SYN or FIN), use snd_max instead of snd_nxt
944 * when filling in ti_seq. But if we are in persist
945 * state, snd_max might reflect one byte beyond the
946 * right edge of the window, so use snd_nxt in that
947 * case, since we know we aren't doing a retransmission.
948 * (retransmit and persist are mutually exclusive...)
950 if (sack_rxmit == 0) {
951 if (len || (flags & (TH_SYN|TH_FIN)) ||
952 tcp_timer_active(tp, TT_PERSIST))
953 th->th_seq = htonl(tp->snd_nxt);
955 th->th_seq = htonl(tp->snd_max);
957 th->th_seq = htonl(p->rxmit);
959 tp->sackhint.sack_bytes_rexmit += len;
961 th->th_ack = htonl(tp->rcv_nxt);
963 bcopy(opt, th + 1, optlen);
964 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
966 th->th_flags = flags;
968 * Calculate receive window. Don't shrink window,
969 * but avoid silly window syndrome.
971 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
972 recwin < (long)tp->t_maxseg)
974 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
975 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
976 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
977 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
980 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
981 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
982 * case is handled in syncache.
985 th->th_win = htons((u_short)
986 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
988 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
991 * Adjust the RXWIN0SENT flag - indicate that we have advertised
992 * a 0 window. This may cause the remote transmitter to stall. This
993 * flag tells soreceive() to disable delayed acknowledgements when
994 * draining the buffer. This can occur if the receiver is attempting
995 * to read more data than can be buffered prior to transmitting on
999 tp->t_flags |= TF_RXWIN0SENT;
1001 tp->t_flags &= ~TF_RXWIN0SENT;
1002 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1003 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1004 th->th_flags |= TH_URG;
1007 * If no urgent pointer to send, then we pull
1008 * the urgent pointer to the left edge of the send window
1009 * so that it doesn't drift into the send window on sequence
1010 * number wraparound.
1012 tp->snd_up = tp->snd_una; /* drag it along */
1014 #ifdef TCP_SIGNATURE
1015 if (tp->t_flags & TF_SIGNATURE) {
1016 int sigoff = to.to_signature - opt;
1017 tcp_signature_compute(m, 0, len, optlen,
1018 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1023 * Put TCP length in extended header, and then
1024 * checksum extended header and data.
1026 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1030 * ip6_plen is not need to be filled now, and will be filled
1033 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1034 sizeof(struct tcphdr) + optlen + len);
1038 m->m_pkthdr.csum_flags = CSUM_TCP;
1039 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1040 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1041 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1043 /* IP version must be set here for ipv4/ipv6 checking later */
1044 KASSERT(ip->ip_v == IPVERSION,
1045 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1049 * Enable TSO and specify the size of the segments.
1050 * The TCP pseudo header checksum is always provided.
1051 * XXX: Fixme: This is currently not the case for IPv6.
1054 m->m_pkthdr.csum_flags = CSUM_TSO;
1055 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1059 * In transmit state, time the transmission and arrange for
1060 * the retransmit. In persist state, just set snd_max.
1062 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1063 !tcp_timer_active(tp, TT_PERSIST)) {
1064 tcp_seq startseq = tp->snd_nxt;
1067 * Advance snd_nxt over sequence space of this segment.
1069 if (flags & (TH_SYN|TH_FIN)) {
1072 if (flags & TH_FIN) {
1074 tp->t_flags |= TF_SENTFIN;
1080 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1081 tp->snd_max = tp->snd_nxt;
1083 * Time this transmission if not a retransmission and
1084 * not currently timing anything.
1086 if (tp->t_rtttime == 0) {
1087 tp->t_rtttime = ticks;
1088 tp->t_rtseq = startseq;
1089 TCPSTAT_INC(tcps_segstimed);
1094 * Set retransmit timer if not currently set,
1095 * and not doing a pure ack or a keep-alive probe.
1096 * Initial value for retransmit timer is smoothed
1097 * round-trip time + 2 * round-trip time variance.
1098 * Initialize shift counter which is used for backoff
1099 * of retransmit time.
1102 if (!tcp_timer_active(tp, TT_REXMT) &&
1103 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1104 (tp->snd_nxt != tp->snd_una))) {
1105 if (tcp_timer_active(tp, TT_PERSIST)) {
1106 tcp_timer_activate(tp, TT_PERSIST, 0);
1109 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1113 * Persist case, update snd_max but since we are in
1114 * persist mode (no window) we do not update snd_nxt.
1119 if (flags & TH_FIN) {
1121 tp->t_flags |= TF_SENTFIN;
1123 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1124 tp->snd_max = tp->snd_nxt + len;
1131 if (so->so_options & SO_DEBUG) {
1137 save = ipov->ih_len;
1138 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1140 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1144 ipov->ih_len = save;
1149 * Fill in IP length and desired time to live and
1150 * send to IP level. There should be a better way
1151 * to handle ttl and tos; we could keep them in
1152 * the template, but need a way to checksum without them.
1155 * m->m_pkthdr.len should have been set before cksum calcuration,
1156 * because in6_cksum() need it.
1161 * we separately set hoplimit for every segment, since the
1162 * user might want to change the value via setsockopt.
1163 * Also, desired default hop limit might be changed via
1164 * Neighbor Discovery.
1166 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1168 /* TODO: IPv6 IP6TOS_ECT bit on */
1169 error = ip6_output(m,
1170 tp->t_inpcb->in6p_outputopts, NULL,
1171 ((so->so_options & SO_DONTROUTE) ?
1172 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1176 ip->ip_len = m->m_pkthdr.len;
1178 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1179 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1182 * If we do path MTU discovery, then we set DF on every packet.
1183 * This might not be the best thing to do according to RFC3390
1184 * Section 2. However the tcp hostcache migitates the problem
1185 * so it affects only the first tcp connection with a host.
1187 if (V_path_mtu_discovery)
1188 ip->ip_off |= IP_DF;
1190 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1191 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1197 * We know that the packet was lost, so back out the
1198 * sequence number advance, if any.
1200 * If the error is EPERM the packet got blocked by the
1201 * local firewall. Normally we should terminate the
1202 * connection but the blocking may have been spurious
1203 * due to a firewall reconfiguration cycle. So we treat
1204 * it like a packet loss and let the retransmit timer and
1205 * timeouts do their work over time.
1206 * XXX: It is a POLA question whether calling tcp_drop right
1207 * away would be the really correct behavior instead.
1209 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1210 !tcp_timer_active(tp, TT_PERSIST)) &&
1211 ((flags & TH_SYN) == 0) &&
1215 tp->sackhint.sack_bytes_rexmit -= len;
1216 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1217 ("sackhint bytes rtx >= 0"));
1222 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1225 tp->t_softerror = error;
1228 if (!tcp_timer_active(tp, TT_REXMT) &&
1229 !tcp_timer_active(tp, TT_PERSIST))
1230 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1231 tp->snd_cwnd = tp->t_maxseg;
1235 * For some reason the interface we used initially
1236 * to send segments changed to another or lowered
1239 * tcp_mtudisc() will find out the new MTU and as
1240 * its last action, initiate retransmission, so it
1241 * is important to not do so here.
1243 * If TSO was active we either got an interface
1244 * without TSO capabilits or TSO was turned off.
1245 * Disable it for this connection as too and
1246 * immediatly retry with MSS sized segments generated
1250 tp->t_flags &= ~TF_TSO;
1251 tcp_mtudisc(tp->t_inpcb, 0);
1257 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1258 tp->t_softerror = error;
1266 TCPSTAT_INC(tcps_sndtotal);
1269 * Data sent (as far as we can tell).
1270 * If this advertises a larger window than any other segment,
1271 * then remember the size of the advertised window.
1272 * Any pending ACK has now been sent.
1274 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1275 tp->rcv_adv = tp->rcv_nxt + recwin;
1276 tp->last_ack_sent = tp->rcv_nxt;
1277 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1278 if (tcp_timer_active(tp, TT_DELACK))
1279 tcp_timer_activate(tp, TT_DELACK, 0);
1282 * This completely breaks TCP if newreno is turned on. What happens
1283 * is that if delayed-acks are turned on on the receiver, this code
1284 * on the transmitter effectively destroys the TCP window, forcing
1285 * it to four packets (1.5Kx4 = 6K window).
1287 if (sendalot && (!V_tcp_do_newreno || --maxburst))
1296 tcp_setpersist(struct tcpcb *tp)
1298 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1301 if (tcp_timer_active(tp, TT_REXMT))
1302 panic("tcp_setpersist: retransmit pending");
1304 * Start/restart persistance timer.
1306 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1307 TCPTV_PERSMIN, TCPTV_PERSMAX);
1308 tcp_timer_activate(tp, TT_PERSIST, tt);
1309 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1314 * Insert TCP options according to the supplied parameters to the place
1315 * optp in a consistent way. Can handle unaligned destinations.
1317 * The order of the option processing is crucial for optimal packing and
1318 * alignment for the scarce option space.
1320 * The optimal order for a SYN/SYN-ACK segment is:
1321 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1322 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1324 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1325 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1326 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1327 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1328 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1331 tcp_addoptions(struct tcpopt *to, u_char *optp)
1333 u_int mask, optlen = 0;
1335 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1336 if ((to->to_flags & mask) != mask)
1338 if (optlen == TCP_MAXOLEN)
1340 switch (to->to_flags & mask) {
1342 while (optlen % 4) {
1343 optlen += TCPOLEN_NOP;
1344 *optp++ = TCPOPT_NOP;
1346 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1348 optlen += TCPOLEN_MAXSEG;
1349 *optp++ = TCPOPT_MAXSEG;
1350 *optp++ = TCPOLEN_MAXSEG;
1351 to->to_mss = htons(to->to_mss);
1352 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1353 optp += sizeof(to->to_mss);
1356 while (!optlen || optlen % 2 != 1) {
1357 optlen += TCPOLEN_NOP;
1358 *optp++ = TCPOPT_NOP;
1360 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1362 optlen += TCPOLEN_WINDOW;
1363 *optp++ = TCPOPT_WINDOW;
1364 *optp++ = TCPOLEN_WINDOW;
1365 *optp++ = to->to_wscale;
1368 while (optlen % 2) {
1369 optlen += TCPOLEN_NOP;
1370 *optp++ = TCPOPT_NOP;
1372 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1374 optlen += TCPOLEN_SACK_PERMITTED;
1375 *optp++ = TCPOPT_SACK_PERMITTED;
1376 *optp++ = TCPOLEN_SACK_PERMITTED;
1379 while (!optlen || optlen % 4 != 2) {
1380 optlen += TCPOLEN_NOP;
1381 *optp++ = TCPOPT_NOP;
1383 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1385 optlen += TCPOLEN_TIMESTAMP;
1386 *optp++ = TCPOPT_TIMESTAMP;
1387 *optp++ = TCPOLEN_TIMESTAMP;
1388 to->to_tsval = htonl(to->to_tsval);
1389 to->to_tsecr = htonl(to->to_tsecr);
1390 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1391 optp += sizeof(to->to_tsval);
1392 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1393 optp += sizeof(to->to_tsecr);
1397 int siglen = TCPOLEN_SIGNATURE - 2;
1399 while (!optlen || optlen % 4 != 2) {
1400 optlen += TCPOLEN_NOP;
1401 *optp++ = TCPOPT_NOP;
1403 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1405 optlen += TCPOLEN_SIGNATURE;
1406 *optp++ = TCPOPT_SIGNATURE;
1407 *optp++ = TCPOLEN_SIGNATURE;
1408 to->to_signature = optp;
1416 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1419 while (!optlen || optlen % 4 != 2) {
1420 optlen += TCPOLEN_NOP;
1421 *optp++ = TCPOPT_NOP;
1423 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1425 optlen += TCPOLEN_SACKHDR;
1426 *optp++ = TCPOPT_SACK;
1427 sackblks = min(to->to_nsacks,
1428 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1429 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1430 while (sackblks--) {
1431 sack_seq = htonl(sack->start);
1432 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1433 optp += sizeof(sack_seq);
1434 sack_seq = htonl(sack->end);
1435 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1436 optp += sizeof(sack_seq);
1437 optlen += TCPOLEN_SACK;
1440 TCPSTAT_INC(tcps_sack_send_blocks);
1444 panic("%s: unknown TCP option type", __func__);
1449 /* Terminate and pad TCP options to a 4 byte boundary. */
1451 optlen += TCPOLEN_EOL;
1452 *optp++ = TCPOPT_EOL;
1455 * According to RFC 793 (STD0007):
1456 * "The content of the header beyond the End-of-Option option
1457 * must be header padding (i.e., zero)."
1458 * and later: "The padding is composed of zeros."
1460 while (optlen % 4) {
1461 optlen += TCPOLEN_PAD;
1462 *optp++ = TCPOPT_PAD;
1465 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));