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/cc.h>
57 #include <netinet/in.h>
58 #include <netinet/in_systm.h>
59 #include <netinet/ip.h>
60 #include <netinet/in_pcb.h>
61 #include <netinet/ip_var.h>
62 #include <netinet/ip_options.h>
64 #include <netinet6/in6_pcb.h>
65 #include <netinet/ip6.h>
66 #include <netinet6/ip6_var.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_tso) = 1;
106 #define V_tcp_do_tso VNET(tcp_do_tso)
107 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
108 &VNET_NAME(tcp_do_tso), 0,
109 "Enable TCP Segmentation Offload");
111 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
112 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
113 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
114 &VNET_NAME(tcp_do_autosndbuf), 0,
115 "Enable automatic send buffer sizing");
117 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
118 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
119 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
120 &VNET_NAME(tcp_autosndbuf_inc), 0,
121 "Incrementor step size of automatic send buffer");
123 VNET_DEFINE(int, tcp_autosndbuf_max) = 256*1024;
124 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
125 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
126 &VNET_NAME(tcp_autosndbuf_max), 0,
127 "Max size of automatic send buffer");
129 static void inline cc_after_idle(struct tcpcb *tp);
132 * CC wrapper hook functions
135 cc_after_idle(struct tcpcb *tp)
137 INP_WLOCK_ASSERT(tp->t_inpcb);
139 if (CC_ALGO(tp)->after_idle != NULL)
140 CC_ALGO(tp)->after_idle(tp->ccv);
144 * Tcp output routine: figure out what should be sent and send it.
147 tcp_output(struct tcpcb *tp)
149 struct socket *so = tp->t_inpcb->inp_socket;
150 long len, recwin, sendwin;
151 int off, flags, error;
153 struct ip *ip = NULL;
154 struct ipovly *ipov = NULL;
156 u_char opt[TCP_MAXOLEN];
157 unsigned ipoptlen, optlen, hdrlen;
159 unsigned ipsec_optlen = 0;
162 int sack_rxmit, sack_bytes_rxmt;
167 int maxburst = TCP_MAXBURST;
170 struct ip6_hdr *ip6 = NULL;
173 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
176 INP_WLOCK_ASSERT(tp->t_inpcb);
179 * Determine length of data that should be transmitted,
180 * and flags that will be used.
181 * If there is some data or critical controls (SYN, RST)
182 * to send, then transmit; otherwise, investigate further.
184 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
185 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
187 tp->t_flags &= ~TF_LASTIDLE;
189 if (tp->t_flags & TF_MORETOCOME) {
190 tp->t_flags |= TF_LASTIDLE;
196 * If we've recently taken a timeout, snd_max will be greater than
197 * snd_nxt. There may be SACK information that allows us to avoid
198 * resending already delivered data. Adjust snd_nxt accordingly.
200 if ((tp->t_flags & TF_SACK_PERMIT) &&
201 SEQ_LT(tp->snd_nxt, tp->snd_max))
205 off = tp->snd_nxt - tp->snd_una;
206 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
207 sendwin = min(sendwin, tp->snd_bwnd);
209 flags = tcp_outflags[tp->t_state];
211 * Send any SACK-generated retransmissions. If we're explicitly trying
212 * to send out new data (when sendalot is 1), bypass this function.
213 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
214 * we're replacing a (future) new transmission with a retransmission
215 * now, and we previously incremented snd_cwnd in tcp_input().
218 * Still in sack recovery , reset rxmit flag to zero.
224 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
225 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
228 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
231 /* Do not retransmit SACK segments beyond snd_recover */
232 if (SEQ_GT(p->end, tp->snd_recover)) {
234 * (At least) part of sack hole extends beyond
235 * snd_recover. Check to see if we can rexmit data
238 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
240 * Can't rexmit any more data for this hole.
241 * That data will be rexmitted in the next
242 * sack recovery episode, when snd_recover
243 * moves past p->rxmit.
246 goto after_sack_rexmit;
248 /* Can rexmit part of the current hole */
249 len = ((long)ulmin(cwin,
250 tp->snd_recover - p->rxmit));
252 len = ((long)ulmin(cwin, p->end - p->rxmit));
253 off = p->rxmit - tp->snd_una;
254 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
259 TCPSTAT_INC(tcps_sack_rexmits);
260 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
261 min(len, tp->t_maxseg));
266 * Get standard flags, and add SYN or FIN if requested by 'hidden'
269 if (tp->t_flags & TF_NEEDFIN)
271 if (tp->t_flags & TF_NEEDSYN)
274 SOCKBUF_LOCK(&so->so_snd);
276 * If in persist timeout with window of 0, send 1 byte.
277 * Otherwise, if window is small but nonzero
278 * and timer expired, we will send what we can
279 * and go to transmit state.
281 if (tp->t_flags & TF_FORCEDATA) {
284 * If we still have some data to send, then
285 * clear the FIN bit. Usually this would
286 * happen below when it realizes that we
287 * aren't sending all the data. However,
288 * if we have exactly 1 byte of unsent data,
289 * then it won't clear the FIN bit below,
290 * and if we are in persist state, we wind
291 * up sending the packet without recording
292 * that we sent the FIN bit.
294 * We can't just blindly clear the FIN bit,
295 * because if we don't have any more data
296 * to send then the probe will be the FIN
299 if (off < so->so_snd.sb_cc)
303 tcp_timer_activate(tp, TT_PERSIST, 0);
309 * If snd_nxt == snd_max and we have transmitted a FIN, the
310 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
311 * a negative length. This can also occur when TCP opens up
312 * its congestion window while receiving additional duplicate
313 * acks after fast-retransmit because TCP will reset snd_nxt
314 * to snd_max after the fast-retransmit.
316 * In the normal retransmit-FIN-only case, however, snd_nxt will
317 * be set to snd_una, the offset will be 0, and the length may
320 * If sack_rxmit is true we are retransmitting from the scoreboard
321 * in which case len is already set.
323 if (sack_rxmit == 0) {
324 if (sack_bytes_rxmt == 0)
325 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
330 * We are inside of a SACK recovery episode and are
331 * sending new data, having retransmitted all the
332 * data possible in the scoreboard.
334 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
337 * Don't remove this (len > 0) check !
338 * We explicitly check for len > 0 here (although it
339 * isn't really necessary), to work around a gcc
340 * optimization issue - to force gcc to compute
341 * len above. Without this check, the computation
342 * of len is bungled by the optimizer.
345 cwin = tp->snd_cwnd -
346 (tp->snd_nxt - tp->sack_newdata) -
350 len = lmin(len, cwin);
356 * Lop off SYN bit if it has already been sent. However, if this
357 * is SYN-SENT state and if segment contains data and if we don't
358 * know that foreign host supports TAO, suppress sending segment.
360 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
361 if (tp->t_state != TCPS_SYN_RECEIVED)
367 * Be careful not to send data and/or FIN on SYN segments.
368 * This measure is needed to prevent interoperability problems
369 * with not fully conformant TCP implementations.
371 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
378 * If FIN has been sent but not acked,
379 * but we haven't been called to retransmit,
380 * len will be < 0. Otherwise, window shrank
381 * after we sent into it. If window shrank to 0,
382 * cancel pending retransmit, pull snd_nxt back
383 * to (closed) window, and set the persist timer
384 * if it isn't already going. If the window didn't
385 * close completely, just wait for an ACK.
389 tcp_timer_activate(tp, TT_REXMT, 0);
391 tp->snd_nxt = tp->snd_una;
392 if (!tcp_timer_active(tp, TT_PERSIST))
397 /* len will be >= 0 after this point. */
398 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
401 * Automatic sizing of send socket buffer. Often the send buffer
402 * size is not optimally adjusted to the actual network conditions
403 * at hand (delay bandwidth product). Setting the buffer size too
404 * small limits throughput on links with high bandwidth and high
405 * delay (eg. trans-continental/oceanic links). Setting the
406 * buffer size too big consumes too much real kernel memory,
407 * especially with many connections on busy servers.
409 * The criteria to step up the send buffer one notch are:
410 * 1. receive window of remote host is larger than send buffer
411 * (with a fudge factor of 5/4th);
412 * 2. send buffer is filled to 7/8th with data (so we actually
413 * have data to make use of it);
414 * 3. send buffer fill has not hit maximal automatic size;
415 * 4. our send window (slow start and cogestion controlled) is
416 * larger than sent but unacknowledged data in send buffer.
418 * The remote host receive window scaling factor may limit the
419 * growing of the send buffer before it reaches its allowed
422 * It scales directly with slow start or congestion window
423 * and does at most one step per received ACK. This fast
424 * scaling has the drawback of growing the send buffer beyond
425 * what is strictly necessary to make full use of a given
426 * delay*bandwith product. However testing has shown this not
427 * to be much of an problem. At worst we are trading wasting
428 * of available bandwith (the non-use of it) for wasting some
429 * socket buffer memory.
431 * TODO: Shrink send buffer during idle periods together
432 * with congestion window. Requires another timer. Has to
433 * wait for upcoming tcp timer rewrite.
435 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
436 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
437 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
438 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
439 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
440 if (!sbreserve_locked(&so->so_snd,
441 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
442 V_tcp_autosndbuf_max), so, curthread))
443 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
448 * Truncate to the maximum segment length or enable TCP Segmentation
449 * Offloading (if supported by hardware) and ensure that FIN is removed
450 * if the length no longer contains the last data byte.
452 * TSO may only be used if we are in a pure bulk sending state. The
453 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
454 * IP options prevent using TSO. With TSO the TCP header is the same
455 * (except for the sequence number) for all generated packets. This
456 * makes it impossible to transmit any options which vary per generated
459 * The length of TSO bursts is limited to TCP_MAXWIN. That limit and
460 * removal of FIN (if not already catched here) are handled later after
461 * the exact length of the TCP options are known.
465 * Pre-calculate here as we save another lookup into the darknesses
466 * of IPsec that way and can actually decide if TSO is ok.
468 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
470 if (len > tp->t_maxseg) {
471 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso &&
472 ((tp->t_flags & TF_SIGNATURE) == 0) &&
473 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
474 tp->t_inpcb->inp_options == NULL &&
475 tp->t_inpcb->in6p_options == NULL
488 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
491 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
495 recwin = sbspace(&so->so_rcv);
498 * Sender silly window avoidance. We transmit under the following
499 * conditions when len is non-zero:
501 * - We have a full segment (or more with TSO)
502 * - This is the last buffer in a write()/send() and we are
503 * either idle or running NODELAY
504 * - we've timed out (e.g. persist timer)
505 * - we have more then 1/2 the maximum send window's worth of
506 * data (receiver may be limited the window size)
507 * - we need to retransmit
510 if (len >= tp->t_maxseg)
513 * NOTE! on localhost connections an 'ack' from the remote
514 * end may occur synchronously with the output and cause
515 * us to flush a buffer queued with moretocome. XXX
517 * note: the len + off check is almost certainly unnecessary.
519 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
520 (idle || (tp->t_flags & TF_NODELAY)) &&
521 len + off >= so->so_snd.sb_cc &&
522 (tp->t_flags & TF_NOPUSH) == 0) {
525 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
527 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
529 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
536 * Compare available window to amount of window
537 * known to peer (as advertised window less
538 * next expected input). If the difference is at least two
539 * max size segments, or at least 50% of the maximum possible
540 * window, then want to send a window update to peer.
541 * Skip this if the connection is in T/TCP half-open state.
542 * Don't send pure window updates when the peer has closed
543 * the connection and won't ever send more data.
545 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
546 !TCPS_HAVERCVDFIN(tp->t_state)) {
548 * "adv" is the amount we can increase the window,
549 * taking into account that we are limited by
550 * TCP_MAXWIN << tp->rcv_scale.
552 long adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale) -
553 (tp->rcv_adv - tp->rcv_nxt);
555 if (adv >= (long) (2 * tp->t_maxseg))
557 if (2 * adv >= (long) so->so_rcv.sb_hiwat)
562 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
563 * is also a catch-all for the retransmit timer timeout case.
565 if (tp->t_flags & TF_ACKNOW)
567 if ((flags & TH_RST) ||
568 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
570 if (SEQ_GT(tp->snd_up, tp->snd_una))
573 * If our state indicates that FIN should be sent
574 * and we have not yet done so, then we need to send.
576 if (flags & TH_FIN &&
577 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
580 * In SACK, it is possible for tcp_output to fail to send a segment
581 * after the retransmission timer has been turned off. Make sure
582 * that the retransmission timer is set.
584 if ((tp->t_flags & TF_SACK_PERMIT) &&
585 SEQ_GT(tp->snd_max, tp->snd_una) &&
586 !tcp_timer_active(tp, TT_REXMT) &&
587 !tcp_timer_active(tp, TT_PERSIST)) {
588 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
592 * TCP window updates are not reliable, rather a polling protocol
593 * using ``persist'' packets is used to insure receipt of window
594 * updates. The three ``states'' for the output side are:
595 * idle not doing retransmits or persists
596 * persisting to move a small or zero window
597 * (re)transmitting and thereby not persisting
599 * tcp_timer_active(tp, TT_PERSIST)
600 * is true when we are in persist state.
601 * (tp->t_flags & TF_FORCEDATA)
602 * is set when we are called to send a persist packet.
603 * tcp_timer_active(tp, TT_REXMT)
604 * is set when we are retransmitting
605 * The output side is idle when both timers are zero.
607 * If send window is too small, there is data to transmit, and no
608 * retransmit or persist is pending, then go to persist state.
609 * If nothing happens soon, send when timer expires:
610 * if window is nonzero, transmit what we can,
611 * otherwise force out a byte.
613 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
614 !tcp_timer_active(tp, TT_PERSIST)) {
620 * No reason to send a segment, just return.
623 SOCKBUF_UNLOCK(&so->so_snd);
627 SOCKBUF_LOCK_ASSERT(&so->so_snd);
629 * Before ESTABLISHED, force sending of initial options
630 * unless TCP set not to do any options.
631 * NOTE: we assume that the IP/TCP header plus TCP options
632 * always fit in a single mbuf, leaving room for a maximum
634 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
639 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
642 hdrlen = sizeof (struct tcpiphdr);
645 * Compute options for segment.
646 * We only have to care about SYN and established connection
647 * segments. Options for SYN-ACK segments are handled in TCP
650 if ((tp->t_flags & TF_NOOPT) == 0) {
652 /* Maximum segment size. */
653 if (flags & TH_SYN) {
654 tp->snd_nxt = tp->iss;
655 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
656 to.to_flags |= TOF_MSS;
658 /* Window scaling. */
659 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
660 to.to_wscale = tp->request_r_scale;
661 to.to_flags |= TOF_SCALE;
664 if ((tp->t_flags & TF_RCVD_TSTMP) ||
665 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
666 to.to_tsval = ticks + tp->ts_offset;
667 to.to_tsecr = tp->ts_recent;
668 to.to_flags |= TOF_TS;
669 /* Set receive buffer autosizing timestamp. */
670 if (tp->rfbuf_ts == 0 &&
671 (so->so_rcv.sb_flags & SB_AUTOSIZE))
672 tp->rfbuf_ts = ticks;
674 /* Selective ACK's. */
675 if (tp->t_flags & TF_SACK_PERMIT) {
677 to.to_flags |= TOF_SACKPERM;
678 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
679 (tp->t_flags & TF_SACK_PERMIT) &&
680 tp->rcv_numsacks > 0) {
681 to.to_flags |= TOF_SACK;
682 to.to_nsacks = tp->rcv_numsacks;
683 to.to_sacks = (u_char *)tp->sackblks;
687 /* TCP-MD5 (RFC2385). */
688 if (tp->t_flags & TF_SIGNATURE)
689 to.to_flags |= TOF_SIGNATURE;
690 #endif /* TCP_SIGNATURE */
692 /* Processing the options. */
693 hdrlen += optlen = tcp_addoptions(&to, opt);
698 ipoptlen = ip6_optlen(tp->t_inpcb);
701 if (tp->t_inpcb->inp_options)
702 ipoptlen = tp->t_inpcb->inp_options->m_len -
703 offsetof(struct ipoption, ipopt_list);
707 ipoptlen += ipsec_optlen;
711 * Adjust data length if insertion of options will
712 * bump the packet length beyond the t_maxopd length.
713 * Clear the FIN bit because we cut off the tail of
716 * When doing TSO limit a burst to TCP_MAXWIN minus the
717 * IP, TCP and Options length to keep ip->ip_len from
718 * overflowing. Prevent the last segment from being
719 * fractional thus making them all equal sized and set
720 * the flag to continue sending. TSO is disabled when
721 * IP options or IPSEC are present.
723 if (len + optlen + ipoptlen > tp->t_maxopd) {
726 if (len > TCP_MAXWIN - hdrlen - optlen) {
727 len = TCP_MAXWIN - hdrlen - optlen;
728 len = len - (len % (tp->t_maxopd - optlen));
730 } else if (tp->t_flags & TF_NEEDFIN)
733 len = tp->t_maxopd - optlen - ipoptlen;
738 /*#ifdef DIAGNOSTIC*/
740 if (max_linkhdr + hdrlen > MCLBYTES)
742 if (max_linkhdr + hdrlen > MHLEN)
744 panic("tcphdr too big");
748 * This KASSERT is here to catch edge cases at a well defined place.
749 * Before, those had triggered (random) panic conditions further down.
751 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
754 * Grab a header mbuf, attaching a copy of data to
755 * be transmitted, and initialize the header from
756 * the template for sends on this connection.
762 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
763 TCPSTAT_INC(tcps_sndprobe);
764 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
765 tp->t_sndrexmitpack++;
766 TCPSTAT_INC(tcps_sndrexmitpack);
767 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
769 TCPSTAT_INC(tcps_sndpack);
770 TCPSTAT_ADD(tcps_sndbyte, len);
773 if ((m = m_copypack(so->so_snd.sb_mb, off,
774 (int)len, max_linkhdr + hdrlen)) == 0) {
775 SOCKBUF_UNLOCK(&so->so_snd);
780 * m_copypack left space for our hdr; use it.
785 MGETHDR(m, M_DONTWAIT, MT_DATA);
787 SOCKBUF_UNLOCK(&so->so_snd);
792 if (MHLEN < hdrlen + max_linkhdr) {
793 MCLGET(m, M_DONTWAIT);
794 if ((m->m_flags & M_EXT) == 0) {
795 SOCKBUF_UNLOCK(&so->so_snd);
802 m->m_data += max_linkhdr;
806 * Start the m_copy functions from the closest mbuf
807 * to the offset in the socket buffer chain.
809 mb = sbsndptr(&so->so_snd, off, len, &moff);
811 if (len <= MHLEN - hdrlen - max_linkhdr) {
812 m_copydata(mb, moff, (int)len,
813 mtod(m, caddr_t) + hdrlen);
816 m->m_next = m_copy(mb, moff, (int)len);
817 if (m->m_next == NULL) {
818 SOCKBUF_UNLOCK(&so->so_snd);
826 * If we're sending everything we've got, set PUSH.
827 * (This will keep happy those implementations which only
828 * give data to the user when a buffer fills or
831 if (off + len == so->so_snd.sb_cc)
833 SOCKBUF_UNLOCK(&so->so_snd);
835 SOCKBUF_UNLOCK(&so->so_snd);
836 if (tp->t_flags & TF_ACKNOW)
837 TCPSTAT_INC(tcps_sndacks);
838 else if (flags & (TH_SYN|TH_FIN|TH_RST))
839 TCPSTAT_INC(tcps_sndctrl);
840 else if (SEQ_GT(tp->snd_up, tp->snd_una))
841 TCPSTAT_INC(tcps_sndurg);
843 TCPSTAT_INC(tcps_sndwinup);
845 MGETHDR(m, M_DONTWAIT, MT_DATA);
851 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
856 m->m_data += max_linkhdr;
859 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
860 m->m_pkthdr.rcvif = (struct ifnet *)0;
862 mac_inpcb_create_mbuf(tp->t_inpcb, m);
866 ip6 = mtod(m, struct ip6_hdr *);
867 th = (struct tcphdr *)(ip6 + 1);
868 tcpip_fillheaders(tp->t_inpcb, ip6, th);
872 ip = mtod(m, struct ip *);
873 ipov = (struct ipovly *)ip;
874 th = (struct tcphdr *)(ip + 1);
875 tcpip_fillheaders(tp->t_inpcb, ip, th);
879 * Fill in fields, remembering maximum advertised
880 * window for use in delaying messages about window sizes.
881 * If resending a FIN, be sure not to use a new sequence number.
883 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
884 tp->snd_nxt == tp->snd_max)
887 * If we are starting a connection, send ECN setup
888 * SYN packet. If we are on a retransmit, we may
889 * resend those bits a number of times as per
892 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
893 if (tp->t_rxtshift >= 1) {
894 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
895 flags |= TH_ECE|TH_CWR;
897 flags |= TH_ECE|TH_CWR;
900 if (tp->t_state == TCPS_ESTABLISHED &&
901 (tp->t_flags & TF_ECN_PERMIT)) {
903 * If the peer has ECN, mark data packets with
904 * ECN capable transmission (ECT).
905 * Ignore pure ack packets, retransmissions and window probes.
907 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
908 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
911 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
914 ip->ip_tos |= IPTOS_ECN_ECT0;
915 TCPSTAT_INC(tcps_ecn_ect0);
919 * Reply with proper ECN notifications.
921 if (tp->t_flags & TF_ECN_SND_CWR) {
923 tp->t_flags &= ~TF_ECN_SND_CWR;
925 if (tp->t_flags & TF_ECN_SND_ECE)
930 * If we are doing retransmissions, then snd_nxt will
931 * not reflect the first unsent octet. For ACK only
932 * packets, we do not want the sequence number of the
933 * retransmitted packet, we want the sequence number
934 * of the next unsent octet. So, if there is no data
935 * (and no SYN or FIN), use snd_max instead of snd_nxt
936 * when filling in ti_seq. But if we are in persist
937 * state, snd_max might reflect one byte beyond the
938 * right edge of the window, so use snd_nxt in that
939 * case, since we know we aren't doing a retransmission.
940 * (retransmit and persist are mutually exclusive...)
942 if (sack_rxmit == 0) {
943 if (len || (flags & (TH_SYN|TH_FIN)) ||
944 tcp_timer_active(tp, TT_PERSIST))
945 th->th_seq = htonl(tp->snd_nxt);
947 th->th_seq = htonl(tp->snd_max);
949 th->th_seq = htonl(p->rxmit);
951 tp->sackhint.sack_bytes_rexmit += len;
953 th->th_ack = htonl(tp->rcv_nxt);
955 bcopy(opt, th + 1, optlen);
956 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
958 th->th_flags = flags;
960 * Calculate receive window. Don't shrink window,
961 * but avoid silly window syndrome.
963 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
964 recwin < (long)tp->t_maxseg)
966 if (recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
967 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
968 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
969 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
972 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
973 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
974 * case is handled in syncache.
977 th->th_win = htons((u_short)
978 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
980 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
983 * Adjust the RXWIN0SENT flag - indicate that we have advertised
984 * a 0 window. This may cause the remote transmitter to stall. This
985 * flag tells soreceive() to disable delayed acknowledgements when
986 * draining the buffer. This can occur if the receiver is attempting
987 * to read more data than can be buffered prior to transmitting on
990 if (th->th_win == 0) {
992 tp->t_flags |= TF_RXWIN0SENT;
994 tp->t_flags &= ~TF_RXWIN0SENT;
995 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
996 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
997 th->th_flags |= TH_URG;
1000 * If no urgent pointer to send, then we pull
1001 * the urgent pointer to the left edge of the send window
1002 * so that it doesn't drift into the send window on sequence
1003 * number wraparound.
1005 tp->snd_up = tp->snd_una; /* drag it along */
1007 #ifdef TCP_SIGNATURE
1008 if (tp->t_flags & TF_SIGNATURE) {
1009 int sigoff = to.to_signature - opt;
1010 tcp_signature_compute(m, 0, len, optlen,
1011 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1016 * Put TCP length in extended header, and then
1017 * checksum extended header and data.
1019 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1023 * ip6_plen is not need to be filled now, and will be filled
1026 th->th_sum = in6_cksum(m, IPPROTO_TCP, sizeof(struct ip6_hdr),
1027 sizeof(struct tcphdr) + optlen + len);
1031 m->m_pkthdr.csum_flags = CSUM_TCP;
1032 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1033 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1034 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1036 /* IP version must be set here for ipv4/ipv6 checking later */
1037 KASSERT(ip->ip_v == IPVERSION,
1038 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1042 * Enable TSO and specify the size of the segments.
1043 * The TCP pseudo header checksum is always provided.
1044 * XXX: Fixme: This is currently not the case for IPv6.
1047 KASSERT(len > tp->t_maxopd - optlen,
1048 ("%s: len <= tso_segsz", __func__));
1049 m->m_pkthdr.csum_flags |= CSUM_TSO;
1050 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1054 * In transmit state, time the transmission and arrange for
1055 * the retransmit. In persist state, just set snd_max.
1057 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1058 !tcp_timer_active(tp, TT_PERSIST)) {
1059 tcp_seq startseq = tp->snd_nxt;
1062 * Advance snd_nxt over sequence space of this segment.
1064 if (flags & (TH_SYN|TH_FIN)) {
1067 if (flags & TH_FIN) {
1069 tp->t_flags |= TF_SENTFIN;
1075 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1076 tp->snd_max = tp->snd_nxt;
1078 * Time this transmission if not a retransmission and
1079 * not currently timing anything.
1081 if (tp->t_rtttime == 0) {
1082 tp->t_rtttime = ticks;
1083 tp->t_rtseq = startseq;
1084 TCPSTAT_INC(tcps_segstimed);
1089 * Set retransmit timer if not currently set,
1090 * and not doing a pure ack or a keep-alive probe.
1091 * Initial value for retransmit timer is smoothed
1092 * round-trip time + 2 * round-trip time variance.
1093 * Initialize shift counter which is used for backoff
1094 * of retransmit time.
1097 if (!tcp_timer_active(tp, TT_REXMT) &&
1098 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1099 (tp->snd_nxt != tp->snd_una))) {
1100 if (tcp_timer_active(tp, TT_PERSIST)) {
1101 tcp_timer_activate(tp, TT_PERSIST, 0);
1104 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1108 * Persist case, update snd_max but since we are in
1109 * persist mode (no window) we do not update snd_nxt.
1114 if (flags & TH_FIN) {
1116 tp->t_flags |= TF_SENTFIN;
1118 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1119 tp->snd_max = tp->snd_nxt + len;
1126 if (so->so_options & SO_DEBUG) {
1132 save = ipov->ih_len;
1133 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1135 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1139 ipov->ih_len = save;
1144 * Fill in IP length and desired time to live and
1145 * send to IP level. There should be a better way
1146 * to handle ttl and tos; we could keep them in
1147 * the template, but need a way to checksum without them.
1150 * m->m_pkthdr.len should have been set before cksum calcuration,
1151 * because in6_cksum() need it.
1156 * we separately set hoplimit for every segment, since the
1157 * user might want to change the value via setsockopt.
1158 * Also, desired default hop limit might be changed via
1159 * Neighbor Discovery.
1161 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1163 /* TODO: IPv6 IP6TOS_ECT bit on */
1164 error = ip6_output(m,
1165 tp->t_inpcb->in6p_outputopts, NULL,
1166 ((so->so_options & SO_DONTROUTE) ?
1167 IP_ROUTETOIF : 0), NULL, NULL, tp->t_inpcb);
1171 ip->ip_len = m->m_pkthdr.len;
1173 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1174 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1177 * If we do path MTU discovery, then we set DF on every packet.
1178 * This might not be the best thing to do according to RFC3390
1179 * Section 2. However the tcp hostcache migitates the problem
1180 * so it affects only the first tcp connection with a host.
1182 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1184 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1185 ip->ip_off |= IP_DF;
1187 error = ip_output(m, tp->t_inpcb->inp_options, NULL,
1188 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1194 * We know that the packet was lost, so back out the
1195 * sequence number advance, if any.
1197 * If the error is EPERM the packet got blocked by the
1198 * local firewall. Normally we should terminate the
1199 * connection but the blocking may have been spurious
1200 * due to a firewall reconfiguration cycle. So we treat
1201 * it like a packet loss and let the retransmit timer and
1202 * timeouts do their work over time.
1203 * XXX: It is a POLA question whether calling tcp_drop right
1204 * away would be the really correct behavior instead.
1206 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1207 !tcp_timer_active(tp, TT_PERSIST)) &&
1208 ((flags & TH_SYN) == 0) &&
1212 tp->sackhint.sack_bytes_rexmit -= len;
1213 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1214 ("sackhint bytes rtx >= 0"));
1219 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1222 tp->t_softerror = error;
1225 if (!tcp_timer_active(tp, TT_REXMT) &&
1226 !tcp_timer_active(tp, TT_PERSIST))
1227 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1228 tp->snd_cwnd = tp->t_maxseg;
1232 * For some reason the interface we used initially
1233 * to send segments changed to another or lowered
1236 * tcp_mtudisc() will find out the new MTU and as
1237 * its last action, initiate retransmission, so it
1238 * is important to not do so here.
1240 * If TSO was active we either got an interface
1241 * without TSO capabilits or TSO was turned off.
1242 * Disable it for this connection as too and
1243 * immediatly retry with MSS sized segments generated
1247 tp->t_flags &= ~TF_TSO;
1248 tcp_mtudisc(tp->t_inpcb, 0);
1254 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1255 tp->t_softerror = error;
1263 TCPSTAT_INC(tcps_sndtotal);
1266 * Data sent (as far as we can tell).
1267 * If this advertises a larger window than any other segment,
1268 * then remember the size of the advertised window.
1269 * Any pending ACK has now been sent.
1271 if (recwin > 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1272 tp->rcv_adv = tp->rcv_nxt + recwin;
1273 tp->last_ack_sent = tp->rcv_nxt;
1274 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1275 if (tcp_timer_active(tp, TT_DELACK))
1276 tcp_timer_activate(tp, TT_DELACK, 0);
1279 * This completely breaks TCP if newreno is turned on. What happens
1280 * is that if delayed-acks are turned on on the receiver, this code
1281 * on the transmitter effectively destroys the TCP window, forcing
1282 * it to four packets (1.5Kx4 = 6K window).
1284 if (sendalot && --maxburst)
1293 tcp_setpersist(struct tcpcb *tp)
1295 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1298 if (tcp_timer_active(tp, TT_REXMT))
1299 panic("tcp_setpersist: retransmit pending");
1301 * Start/restart persistance timer.
1303 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1304 TCPTV_PERSMIN, TCPTV_PERSMAX);
1305 tcp_timer_activate(tp, TT_PERSIST, tt);
1306 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1311 * Insert TCP options according to the supplied parameters to the place
1312 * optp in a consistent way. Can handle unaligned destinations.
1314 * The order of the option processing is crucial for optimal packing and
1315 * alignment for the scarce option space.
1317 * The optimal order for a SYN/SYN-ACK segment is:
1318 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1319 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1321 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1322 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1323 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1324 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1325 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1328 tcp_addoptions(struct tcpopt *to, u_char *optp)
1330 u_int mask, optlen = 0;
1332 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1333 if ((to->to_flags & mask) != mask)
1335 if (optlen == TCP_MAXOLEN)
1337 switch (to->to_flags & mask) {
1339 while (optlen % 4) {
1340 optlen += TCPOLEN_NOP;
1341 *optp++ = TCPOPT_NOP;
1343 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1345 optlen += TCPOLEN_MAXSEG;
1346 *optp++ = TCPOPT_MAXSEG;
1347 *optp++ = TCPOLEN_MAXSEG;
1348 to->to_mss = htons(to->to_mss);
1349 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1350 optp += sizeof(to->to_mss);
1353 while (!optlen || optlen % 2 != 1) {
1354 optlen += TCPOLEN_NOP;
1355 *optp++ = TCPOPT_NOP;
1357 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1359 optlen += TCPOLEN_WINDOW;
1360 *optp++ = TCPOPT_WINDOW;
1361 *optp++ = TCPOLEN_WINDOW;
1362 *optp++ = to->to_wscale;
1365 while (optlen % 2) {
1366 optlen += TCPOLEN_NOP;
1367 *optp++ = TCPOPT_NOP;
1369 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1371 optlen += TCPOLEN_SACK_PERMITTED;
1372 *optp++ = TCPOPT_SACK_PERMITTED;
1373 *optp++ = TCPOLEN_SACK_PERMITTED;
1376 while (!optlen || optlen % 4 != 2) {
1377 optlen += TCPOLEN_NOP;
1378 *optp++ = TCPOPT_NOP;
1380 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1382 optlen += TCPOLEN_TIMESTAMP;
1383 *optp++ = TCPOPT_TIMESTAMP;
1384 *optp++ = TCPOLEN_TIMESTAMP;
1385 to->to_tsval = htonl(to->to_tsval);
1386 to->to_tsecr = htonl(to->to_tsecr);
1387 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1388 optp += sizeof(to->to_tsval);
1389 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1390 optp += sizeof(to->to_tsecr);
1394 int siglen = TCPOLEN_SIGNATURE - 2;
1396 while (!optlen || optlen % 4 != 2) {
1397 optlen += TCPOLEN_NOP;
1398 *optp++ = TCPOPT_NOP;
1400 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1402 optlen += TCPOLEN_SIGNATURE;
1403 *optp++ = TCPOPT_SIGNATURE;
1404 *optp++ = TCPOLEN_SIGNATURE;
1405 to->to_signature = optp;
1413 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1416 while (!optlen || optlen % 4 != 2) {
1417 optlen += TCPOLEN_NOP;
1418 *optp++ = TCPOPT_NOP;
1420 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1422 optlen += TCPOLEN_SACKHDR;
1423 *optp++ = TCPOPT_SACK;
1424 sackblks = min(to->to_nsacks,
1425 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1426 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1427 while (sackblks--) {
1428 sack_seq = htonl(sack->start);
1429 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1430 optp += sizeof(sack_seq);
1431 sack_seq = htonl(sack->end);
1432 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1433 optp += sizeof(sack_seq);
1434 optlen += TCPOLEN_SACK;
1437 TCPSTAT_INC(tcps_sack_send_blocks);
1441 panic("%s: unknown TCP option type", __func__);
1446 /* Terminate and pad TCP options to a 4 byte boundary. */
1448 optlen += TCPOLEN_EOL;
1449 *optp++ = TCPOPT_EOL;
1452 * According to RFC 793 (STD0007):
1453 * "The content of the header beyond the End-of-Option option
1454 * must be header padding (i.e., zero)."
1455 * and later: "The padding is composed of zeros."
1457 while (optlen % 4) {
1458 optlen += TCPOLEN_PAD;
1459 *optp++ = TCPOPT_PAD;
1462 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));