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 * 3. 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>
44 #include <sys/hhook.h>
46 #include <sys/kernel.h>
49 #include <sys/mutex.h>
50 #include <sys/protosw.h>
52 #include <sys/socket.h>
53 #include <sys/socketvar.h>
54 #include <sys/sysctl.h>
57 #include <net/route.h>
60 #include <netinet/in.h>
61 #include <netinet/in_kdtrace.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/ip_options.h>
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/ip6_var.h>
73 #include <netinet/tcp_fastopen.h>
75 #include <netinet/tcp.h>
77 #include <netinet/tcp_fsm.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcpip.h>
82 #include <netinet/cc/cc.h>
84 #include <netinet/tcp_pcap.h>
87 #include <netinet/tcp_debug.h>
90 #include <netinet/tcp_offload.h>
93 #include <netipsec/ipsec_support.h>
95 #include <machine/in_cksum.h>
97 #include <security/mac/mac_framework.h>
99 VNET_DEFINE(int, path_mtu_discovery) = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
101 &VNET_NAME(path_mtu_discovery), 1,
102 "Enable Path MTU Discovery");
104 VNET_DEFINE(int, tcp_do_tso) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
106 &VNET_NAME(tcp_do_tso), 0,
107 "Enable TCP Segmentation Offload");
109 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
110 #define V_tcp_sendspace VNET(tcp_sendspace)
111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
112 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
116 &VNET_NAME(tcp_do_autosndbuf), 0,
117 "Enable automatic send buffer sizing");
119 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_autosndbuf_inc), 0,
122 "Incrementor step size of automatic send buffer");
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
126 &VNET_NAME(tcp_autosndbuf_max), 0,
127 "Max size of automatic send buffer");
129 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
130 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
132 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
133 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
136 * Make sure that either retransmit or persist timer is set for SYN, FIN and
139 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
140 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
141 tcp_timer_active((tp), TT_REXMT) || \
142 tcp_timer_active((tp), TT_PERSIST), \
143 ("neither rexmt nor persist timer is set"))
146 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
147 struct tcphdr *th, struct tcpopt *to,
148 uint32_t len, int tso);
150 static void inline cc_after_idle(struct tcpcb *tp);
154 * Wrapper for the TCP established output helper hook.
157 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
158 struct tcpopt *to, uint32_t len, int tso)
160 struct tcp_hhook_data hhook_data;
162 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
166 hhook_data.len = len;
167 hhook_data.tso = tso;
169 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
176 * CC wrapper hook functions
179 cc_after_idle(struct tcpcb *tp)
181 INP_WLOCK_ASSERT(tp->t_inpcb);
183 if (CC_ALGO(tp)->after_idle != NULL)
184 CC_ALGO(tp)->after_idle(tp->ccv);
188 * Tcp output routine: figure out what should be sent and send it.
191 tcp_output(struct tcpcb *tp)
193 struct socket *so = tp->t_inpcb->inp_socket;
195 uint32_t recwin, sendwin;
196 int off, flags, error = 0; /* Keep compiler happy */
198 struct ip *ip = NULL;
199 struct ipovly *ipov = NULL;
201 u_char opt[TCP_MAXOLEN];
202 unsigned ipoptlen, optlen, hdrlen;
203 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
204 unsigned ipsec_optlen = 0;
207 int sack_rxmit, sack_bytes_rxmt;
212 int maxburst = TCP_MAXBURST;
215 struct ip6_hdr *ip6 = NULL;
218 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
221 INP_WLOCK_ASSERT(tp->t_inpcb);
224 if (tp->t_flags & TF_TOE)
225 return (tcp_offload_output(tp));
230 * For TFO connections in SYN_RECEIVED, only allow the initial
231 * SYN|ACK and those sent by the retransmit timer.
233 if (IS_FASTOPEN(tp->t_flags) &&
234 (tp->t_state == TCPS_SYN_RECEIVED) &&
235 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
236 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
240 * Determine length of data that should be transmitted,
241 * and flags that will be used.
242 * If there is some data or critical controls (SYN, RST)
243 * to send, then transmit; otherwise, investigate further.
245 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
246 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
248 tp->t_flags &= ~TF_LASTIDLE;
250 if (tp->t_flags & TF_MORETOCOME) {
251 tp->t_flags |= TF_LASTIDLE;
257 * If we've recently taken a timeout, snd_max will be greater than
258 * snd_nxt. There may be SACK information that allows us to avoid
259 * resending already delivered data. Adjust snd_nxt accordingly.
261 if ((tp->t_flags & TF_SACK_PERMIT) &&
262 SEQ_LT(tp->snd_nxt, tp->snd_max))
267 off = tp->snd_nxt - tp->snd_una;
268 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
270 flags = tcp_outflags[tp->t_state];
272 * Send any SACK-generated retransmissions. If we're explicitly trying
273 * to send out new data (when sendalot is 1), bypass this function.
274 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
275 * we're replacing a (future) new transmission with a retransmission
276 * now, and we previously incremented snd_cwnd in tcp_input().
279 * Still in sack recovery , reset rxmit flag to zero.
285 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
286 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
290 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
291 /* Do not retransmit SACK segments beyond snd_recover */
292 if (SEQ_GT(p->end, tp->snd_recover)) {
294 * (At least) part of sack hole extends beyond
295 * snd_recover. Check to see if we can rexmit data
298 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
300 * Can't rexmit any more data for this hole.
301 * That data will be rexmitted in the next
302 * sack recovery episode, when snd_recover
303 * moves past p->rxmit.
306 goto after_sack_rexmit;
308 /* Can rexmit part of the current hole */
309 len = ((int32_t)ulmin(cwin,
310 tp->snd_recover - p->rxmit));
312 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
313 off = p->rxmit - tp->snd_una;
314 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
319 TCPSTAT_INC(tcps_sack_rexmits);
320 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
321 min(len, tp->t_maxseg));
326 * Get standard flags, and add SYN or FIN if requested by 'hidden'
329 if (tp->t_flags & TF_NEEDFIN)
331 if (tp->t_flags & TF_NEEDSYN)
334 SOCKBUF_LOCK(&so->so_snd);
336 * If in persist timeout with window of 0, send 1 byte.
337 * Otherwise, if window is small but nonzero
338 * and timer expired, we will send what we can
339 * and go to transmit state.
341 if (tp->t_flags & TF_FORCEDATA) {
344 * If we still have some data to send, then
345 * clear the FIN bit. Usually this would
346 * happen below when it realizes that we
347 * aren't sending all the data. However,
348 * if we have exactly 1 byte of unsent data,
349 * then it won't clear the FIN bit below,
350 * and if we are in persist state, we wind
351 * up sending the packet without recording
352 * that we sent the FIN bit.
354 * We can't just blindly clear the FIN bit,
355 * because if we don't have any more data
356 * to send then the probe will be the FIN
359 if (off < sbused(&so->so_snd))
363 tcp_timer_activate(tp, TT_PERSIST, 0);
369 * If snd_nxt == snd_max and we have transmitted a FIN, the
370 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
371 * a negative length. This can also occur when TCP opens up
372 * its congestion window while receiving additional duplicate
373 * acks after fast-retransmit because TCP will reset snd_nxt
374 * to snd_max after the fast-retransmit.
376 * In the normal retransmit-FIN-only case, however, snd_nxt will
377 * be set to snd_una, the offset will be 0, and the length may
380 * If sack_rxmit is true we are retransmitting from the scoreboard
381 * in which case len is already set.
383 if (sack_rxmit == 0) {
384 if (sack_bytes_rxmt == 0)
385 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
391 * We are inside of a SACK recovery episode and are
392 * sending new data, having retransmitted all the
393 * data possible in the scoreboard.
395 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
398 * Don't remove this (len > 0) check !
399 * We explicitly check for len > 0 here (although it
400 * isn't really necessary), to work around a gcc
401 * optimization issue - to force gcc to compute
402 * len above. Without this check, the computation
403 * of len is bungled by the optimizer.
406 cwin = tp->snd_cwnd -
407 (tp->snd_nxt - tp->sack_newdata) -
411 len = imin(len, cwin);
417 * Lop off SYN bit if it has already been sent. However, if this
418 * is SYN-SENT state and if segment contains data and if we don't
419 * know that foreign host supports TAO, suppress sending segment.
421 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
422 if (tp->t_state != TCPS_SYN_RECEIVED)
426 * When sending additional segments following a TFO SYN|ACK,
427 * do not include the SYN bit.
429 if (IS_FASTOPEN(tp->t_flags) &&
430 (tp->t_state == TCPS_SYN_RECEIVED))
437 * Be careful not to send data and/or FIN on SYN segments.
438 * This measure is needed to prevent interoperability problems
439 * with not fully conformant TCP implementations.
441 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
448 * When retransmitting SYN|ACK on a passively-created TFO socket,
449 * don't include data, as the presence of data may have caused the
450 * original SYN|ACK to have been dropped by a middlebox.
452 if (IS_FASTOPEN(tp->t_flags) &&
453 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
459 * If FIN has been sent but not acked,
460 * but we haven't been called to retransmit,
461 * len will be < 0. Otherwise, window shrank
462 * after we sent into it. If window shrank to 0,
463 * cancel pending retransmit, pull snd_nxt back
464 * to (closed) window, and set the persist timer
465 * if it isn't already going. If the window didn't
466 * close completely, just wait for an ACK.
468 * We also do a general check here to ensure that
469 * we will set the persist timer when we have data
470 * to send, but a 0-byte window. This makes sure
471 * the persist timer is set even if the packet
472 * hits one of the "goto send" lines below.
475 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
476 (off < (int) sbavail(&so->so_snd))) {
477 tcp_timer_activate(tp, TT_REXMT, 0);
479 tp->snd_nxt = tp->snd_una;
480 if (!tcp_timer_active(tp, TT_PERSIST))
485 /* len will be >= 0 after this point. */
486 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
489 * Automatic sizing of send socket buffer. Often the send buffer
490 * size is not optimally adjusted to the actual network conditions
491 * at hand (delay bandwidth product). Setting the buffer size too
492 * small limits throughput on links with high bandwidth and high
493 * delay (eg. trans-continental/oceanic links). Setting the
494 * buffer size too big consumes too much real kernel memory,
495 * especially with many connections on busy servers.
497 * The criteria to step up the send buffer one notch are:
498 * 1. receive window of remote host is larger than send buffer
499 * (with a fudge factor of 5/4th);
500 * 2. send buffer is filled to 7/8th with data (so we actually
501 * have data to make use of it);
502 * 3. send buffer fill has not hit maximal automatic size;
503 * 4. our send window (slow start and cogestion controlled) is
504 * larger than sent but unacknowledged data in send buffer.
506 * The remote host receive window scaling factor may limit the
507 * growing of the send buffer before it reaches its allowed
510 * It scales directly with slow start or congestion window
511 * and does at most one step per received ACK. This fast
512 * scaling has the drawback of growing the send buffer beyond
513 * what is strictly necessary to make full use of a given
514 * delay*bandwidth product. However testing has shown this not
515 * to be much of an problem. At worst we are trading wasting
516 * of available bandwidth (the non-use of it) for wasting some
517 * socket buffer memory.
519 * TODO: Shrink send buffer during idle periods together
520 * with congestion window. Requires another timer. Has to
521 * wait for upcoming tcp timer rewrite.
523 * XXXGL: should there be used sbused() or sbavail()?
525 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
526 int autosndbuf_mod = 0;
527 if (V_tcp_sendbuf_auto_lowat)
528 autosndbuf_mod = so->so_snd.sb_lowat;
530 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - autosndbuf_mod &&
531 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) - autosndbuf_mod &&
532 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
533 sendwin >= (sbused(&so->so_snd) -
534 (tp->snd_nxt - tp->snd_una))) {
535 if (!sbreserve_locked(&so->so_snd,
536 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
537 V_tcp_autosndbuf_max), so, curthread))
538 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
543 * Decide if we can use TCP Segmentation Offloading (if supported by
546 * TSO may only be used if we are in a pure bulk sending state. The
547 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
548 * IP options prevent using TSO. With TSO the TCP header is the same
549 * (except for the sequence number) for all generated packets. This
550 * makes it impossible to transmit any options which vary per generated
553 * IPv4 handling has a clear separation of ip options and ip header
554 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
555 * the right thing below to provide length of just ip options and thus
556 * checking for ipoptlen is enough to decide if ip options are present.
558 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
560 * Pre-calculate here as we save another lookup into the darknesses
561 * of IPsec that way and can actually decide if TSO is ok.
564 if (isipv6 && IPSEC_ENABLED(ipv6))
565 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
571 if (IPSEC_ENABLED(ipv4))
572 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
577 ipoptlen = ip6_optlen(tp->t_inpcb);
580 if (tp->t_inpcb->inp_options)
581 ipoptlen = tp->t_inpcb->inp_options->m_len -
582 offsetof(struct ipoption, ipopt_list);
585 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
586 ipoptlen += ipsec_optlen;
589 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
590 ((tp->t_flags & TF_SIGNATURE) == 0) &&
591 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
596 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
599 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
600 sbused(&so->so_snd)))
604 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
605 (long)TCP_MAXWIN << tp->rcv_scale);
608 * Sender silly window avoidance. We transmit under the following
609 * conditions when len is non-zero:
611 * - We have a full segment (or more with TSO)
612 * - This is the last buffer in a write()/send() and we are
613 * either idle or running NODELAY
614 * - we've timed out (e.g. persist timer)
615 * - we have more then 1/2 the maximum send window's worth of
616 * data (receiver may be limited the window size)
617 * - we need to retransmit
620 if (len >= tp->t_maxseg)
623 * NOTE! on localhost connections an 'ack' from the remote
624 * end may occur synchronously with the output and cause
625 * us to flush a buffer queued with moretocome. XXX
627 * note: the len + off check is almost certainly unnecessary.
629 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
630 (idle || (tp->t_flags & TF_NODELAY)) &&
631 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
632 (tp->t_flags & TF_NOPUSH) == 0) {
635 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
637 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
639 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
646 * Sending of standalone window updates.
648 * Window updates are important when we close our window due to a
649 * full socket buffer and are opening it again after the application
650 * reads data from it. Once the window has opened again and the
651 * remote end starts to send again the ACK clock takes over and
652 * provides the most current window information.
654 * We must avoid the silly window syndrome whereas every read
655 * from the receive buffer, no matter how small, causes a window
656 * update to be sent. We also should avoid sending a flurry of
657 * window updates when the socket buffer had queued a lot of data
658 * and the application is doing small reads.
660 * Prevent a flurry of pointless window updates by only sending
661 * an update when we can increase the advertized window by more
662 * than 1/4th of the socket buffer capacity. When the buffer is
663 * getting full or is very small be more aggressive and send an
664 * update whenever we can increase by two mss sized segments.
665 * In all other situations the ACK's to new incoming data will
666 * carry further window increases.
668 * Don't send an independent window update if a delayed
669 * ACK is pending (it will get piggy-backed on it) or the
670 * remote side already has done a half-close and won't send
671 * more data. Skip this if the connection is in T/TCP
674 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
675 !(tp->t_flags & TF_DELACK) &&
676 !TCPS_HAVERCVDFIN(tp->t_state)) {
678 * "adv" is the amount we could increase the window,
679 * taking into account that we are limited by
680 * TCP_MAXWIN << tp->rcv_scale.
686 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
687 oldwin = (tp->rcv_adv - tp->rcv_nxt);
693 * If the new window size ends up being the same as or less
694 * than the old size when it is scaled, then don't force
697 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
700 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
701 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
702 recwin <= (so->so_rcv.sb_hiwat / 8) ||
703 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
705 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
711 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
712 * is also a catch-all for the retransmit timer timeout case.
714 if (tp->t_flags & TF_ACKNOW)
716 if ((flags & TH_RST) ||
717 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
719 if (SEQ_GT(tp->snd_up, tp->snd_una))
722 * If our state indicates that FIN should be sent
723 * and we have not yet done so, then we need to send.
725 if (flags & TH_FIN &&
726 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
729 * In SACK, it is possible for tcp_output to fail to send a segment
730 * after the retransmission timer has been turned off. Make sure
731 * that the retransmission timer is set.
733 if ((tp->t_flags & TF_SACK_PERMIT) &&
734 SEQ_GT(tp->snd_max, tp->snd_una) &&
735 !tcp_timer_active(tp, TT_REXMT) &&
736 !tcp_timer_active(tp, TT_PERSIST)) {
737 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
741 * TCP window updates are not reliable, rather a polling protocol
742 * using ``persist'' packets is used to insure receipt of window
743 * updates. The three ``states'' for the output side are:
744 * idle not doing retransmits or persists
745 * persisting to move a small or zero window
746 * (re)transmitting and thereby not persisting
748 * tcp_timer_active(tp, TT_PERSIST)
749 * is true when we are in persist state.
750 * (tp->t_flags & TF_FORCEDATA)
751 * is set when we are called to send a persist packet.
752 * tcp_timer_active(tp, TT_REXMT)
753 * is set when we are retransmitting
754 * The output side is idle when both timers are zero.
756 * If send window is too small, there is data to transmit, and no
757 * retransmit or persist is pending, then go to persist state.
758 * If nothing happens soon, send when timer expires:
759 * if window is nonzero, transmit what we can,
760 * otherwise force out a byte.
762 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
763 !tcp_timer_active(tp, TT_PERSIST)) {
769 * No reason to send a segment, just return.
772 SOCKBUF_UNLOCK(&so->so_snd);
776 SOCKBUF_LOCK_ASSERT(&so->so_snd);
778 if (len >= tp->t_maxseg)
779 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
781 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
784 * Before ESTABLISHED, force sending of initial options
785 * unless TCP set not to do any options.
786 * NOTE: we assume that the IP/TCP header plus TCP options
787 * always fit in a single mbuf, leaving room for a maximum
789 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
794 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
797 hdrlen = sizeof (struct tcpiphdr);
800 * Compute options for segment.
801 * We only have to care about SYN and established connection
802 * segments. Options for SYN-ACK segments are handled in TCP
806 if ((tp->t_flags & TF_NOOPT) == 0) {
807 /* Maximum segment size. */
808 if (flags & TH_SYN) {
809 tp->snd_nxt = tp->iss;
810 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
811 to.to_flags |= TOF_MSS;
814 * Only include the TFO option on the first
815 * transmission of the SYN|ACK on a
816 * passively-created TFO socket, as the presence of
817 * the TFO option may have caused the original
818 * SYN|ACK to have been dropped by a middlebox.
820 if (IS_FASTOPEN(tp->t_flags) &&
821 (tp->t_state == TCPS_SYN_RECEIVED) &&
822 (tp->t_rxtshift == 0)) {
823 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
824 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
825 to.to_flags |= TOF_FASTOPEN;
829 /* Window scaling. */
830 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
831 to.to_wscale = tp->request_r_scale;
832 to.to_flags |= TOF_SCALE;
835 if ((tp->t_flags & TF_RCVD_TSTMP) ||
836 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
837 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
838 to.to_tsecr = tp->ts_recent;
839 to.to_flags |= TOF_TS;
842 /* Set receive buffer autosizing timestamp. */
843 if (tp->rfbuf_ts == 0 &&
844 (so->so_rcv.sb_flags & SB_AUTOSIZE))
845 tp->rfbuf_ts = tcp_ts_getticks();
847 /* Selective ACK's. */
848 if (tp->t_flags & TF_SACK_PERMIT) {
850 to.to_flags |= TOF_SACKPERM;
851 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
852 (tp->t_flags & TF_SACK_PERMIT) &&
853 tp->rcv_numsacks > 0) {
854 to.to_flags |= TOF_SACK;
855 to.to_nsacks = tp->rcv_numsacks;
856 to.to_sacks = (u_char *)tp->sackblks;
859 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
860 /* TCP-MD5 (RFC2385). */
862 * Check that TCP_MD5SIG is enabled in tcpcb to
863 * account the size needed to set this TCP option.
865 if (tp->t_flags & TF_SIGNATURE)
866 to.to_flags |= TOF_SIGNATURE;
867 #endif /* TCP_SIGNATURE */
869 /* Processing the options. */
870 hdrlen += optlen = tcp_addoptions(&to, opt);
874 * Adjust data length if insertion of options will
875 * bump the packet length beyond the t_maxseg length.
876 * Clear the FIN bit because we cut off the tail of
879 if (len + optlen + ipoptlen > tp->t_maxseg) {
884 u_int if_hw_tsomaxsegcount;
885 u_int if_hw_tsomaxsegsize;
890 /* extract TSO information */
891 if_hw_tsomax = tp->t_tsomax;
892 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
893 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
896 * Limit a TSO burst to prevent it from
897 * overflowing or exceeding the maximum length
898 * allowed by the network interface:
900 KASSERT(ipoptlen == 0,
901 ("%s: TSO can't do IP options", __func__));
904 * Check if we should limit by maximum payload
907 if (if_hw_tsomax != 0) {
908 /* compute maximum TSO length */
909 max_len = (if_hw_tsomax - hdrlen -
913 } else if (len > max_len) {
920 * Check if we should limit by maximum segment
923 if (if_hw_tsomaxsegcount != 0 &&
924 if_hw_tsomaxsegsize != 0) {
926 * Subtract one segment for the LINK
927 * and TCP/IP headers mbuf that will
928 * be prepended to this mbuf chain
929 * after the code in this section
930 * limits the number of mbufs in the
931 * chain to if_hw_tsomaxsegcount.
933 if_hw_tsomaxsegcount -= 1;
935 mb = sbsndmbuf(&so->so_snd, off, &moff);
937 while (mb != NULL && max_len < len) {
942 * Get length of mbuf fragment
943 * and how many hardware frags,
944 * rounded up, it would use:
946 mlen = (mb->m_len - moff);
947 frags = howmany(mlen,
948 if_hw_tsomaxsegsize);
950 /* Handle special case: Zero Length Mbuf */
955 * Check if the fragment limit
956 * will be reached or exceeded:
958 if (frags >= if_hw_tsomaxsegcount) {
960 if_hw_tsomaxsegcount *
961 if_hw_tsomaxsegsize);
965 if_hw_tsomaxsegcount -= frags;
971 } else if (len > max_len) {
978 * Prevent the last segment from being
979 * fractional unless the send sockbuf can be
982 max_len = (tp->t_maxseg - optlen);
983 if (((uint32_t)off + (uint32_t)len) <
984 sbavail(&so->so_snd)) {
985 moff = len % max_len;
993 * In case there are too many small fragments
996 if (len <= max_len) {
1003 * Send the FIN in a separate segment
1004 * after the bulk sending is done.
1005 * We don't trust the TSO implementations
1006 * to clear the FIN flag on all but the
1009 if (tp->t_flags & TF_NEEDFIN)
1013 len = tp->t_maxseg - optlen - ipoptlen;
1019 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1020 ("%s: len > IP_MAXPACKET", __func__));
1022 /*#ifdef DIAGNOSTIC*/
1024 if (max_linkhdr + hdrlen > MCLBYTES)
1026 if (max_linkhdr + hdrlen > MHLEN)
1028 panic("tcphdr too big");
1032 * This KASSERT is here to catch edge cases at a well defined place.
1033 * Before, those had triggered (random) panic conditions further down.
1035 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1038 * Grab a header mbuf, attaching a copy of data to
1039 * be transmitted, and initialize the header from
1040 * the template for sends on this connection.
1046 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1047 TCPSTAT_INC(tcps_sndprobe);
1048 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1049 tp->t_sndrexmitpack++;
1050 TCPSTAT_INC(tcps_sndrexmitpack);
1051 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1053 TCPSTAT_INC(tcps_sndpack);
1054 TCPSTAT_ADD(tcps_sndbyte, len);
1057 if (MHLEN < hdrlen + max_linkhdr)
1058 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1061 m = m_gethdr(M_NOWAIT, MT_DATA);
1064 SOCKBUF_UNLOCK(&so->so_snd);
1070 m->m_data += max_linkhdr;
1074 * Start the m_copy functions from the closest mbuf
1075 * to the offset in the socket buffer chain.
1077 mb = sbsndptr(&so->so_snd, off, len, &moff);
1079 if (len <= MHLEN - hdrlen - max_linkhdr) {
1080 m_copydata(mb, moff, len,
1081 mtod(m, caddr_t) + hdrlen);
1084 m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1085 if (m->m_next == NULL) {
1086 SOCKBUF_UNLOCK(&so->so_snd);
1095 * If we're sending everything we've got, set PUSH.
1096 * (This will keep happy those implementations which only
1097 * give data to the user when a buffer fills or
1100 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1103 SOCKBUF_UNLOCK(&so->so_snd);
1105 SOCKBUF_UNLOCK(&so->so_snd);
1106 if (tp->t_flags & TF_ACKNOW)
1107 TCPSTAT_INC(tcps_sndacks);
1108 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1109 TCPSTAT_INC(tcps_sndctrl);
1110 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1111 TCPSTAT_INC(tcps_sndurg);
1113 TCPSTAT_INC(tcps_sndwinup);
1115 m = m_gethdr(M_NOWAIT, MT_DATA);
1122 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1127 m->m_data += max_linkhdr;
1130 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1131 m->m_pkthdr.rcvif = (struct ifnet *)0;
1133 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1137 ip6 = mtod(m, struct ip6_hdr *);
1138 th = (struct tcphdr *)(ip6 + 1);
1139 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1143 ip = mtod(m, struct ip *);
1144 ipov = (struct ipovly *)ip;
1145 th = (struct tcphdr *)(ip + 1);
1146 tcpip_fillheaders(tp->t_inpcb, ip, th);
1150 * Fill in fields, remembering maximum advertised
1151 * window for use in delaying messages about window sizes.
1152 * If resending a FIN, be sure not to use a new sequence number.
1154 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1155 tp->snd_nxt == tp->snd_max)
1158 * If we are starting a connection, send ECN setup
1159 * SYN packet. If we are on a retransmit, we may
1160 * resend those bits a number of times as per
1163 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1164 if (tp->t_rxtshift >= 1) {
1165 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1166 flags |= TH_ECE|TH_CWR;
1168 flags |= TH_ECE|TH_CWR;
1171 if (tp->t_state == TCPS_ESTABLISHED &&
1172 (tp->t_flags & TF_ECN_PERMIT)) {
1174 * If the peer has ECN, mark data packets with
1175 * ECN capable transmission (ECT).
1176 * Ignore pure ack packets, retransmissions and window probes.
1178 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1179 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1182 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1185 ip->ip_tos |= IPTOS_ECN_ECT0;
1186 TCPSTAT_INC(tcps_ecn_ect0);
1190 * Reply with proper ECN notifications.
1192 if (tp->t_flags & TF_ECN_SND_CWR) {
1194 tp->t_flags &= ~TF_ECN_SND_CWR;
1196 if (tp->t_flags & TF_ECN_SND_ECE)
1201 * If we are doing retransmissions, then snd_nxt will
1202 * not reflect the first unsent octet. For ACK only
1203 * packets, we do not want the sequence number of the
1204 * retransmitted packet, we want the sequence number
1205 * of the next unsent octet. So, if there is no data
1206 * (and no SYN or FIN), use snd_max instead of snd_nxt
1207 * when filling in ti_seq. But if we are in persist
1208 * state, snd_max might reflect one byte beyond the
1209 * right edge of the window, so use snd_nxt in that
1210 * case, since we know we aren't doing a retransmission.
1211 * (retransmit and persist are mutually exclusive...)
1213 if (sack_rxmit == 0) {
1214 if (len || (flags & (TH_SYN|TH_FIN)) ||
1215 tcp_timer_active(tp, TT_PERSIST))
1216 th->th_seq = htonl(tp->snd_nxt);
1218 th->th_seq = htonl(tp->snd_max);
1220 th->th_seq = htonl(p->rxmit);
1222 tp->sackhint.sack_bytes_rexmit += len;
1224 th->th_ack = htonl(tp->rcv_nxt);
1226 bcopy(opt, th + 1, optlen);
1227 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1229 th->th_flags = flags;
1231 * Calculate receive window. Don't shrink window,
1232 * but avoid silly window syndrome.
1234 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1235 recwin < tp->t_maxseg)
1237 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1238 recwin < (tp->rcv_adv - tp->rcv_nxt))
1239 recwin = (tp->rcv_adv - tp->rcv_nxt);
1242 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1243 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1244 * case is handled in syncache.
1247 th->th_win = htons((u_short)
1248 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1250 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1253 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1254 * a 0 window. This may cause the remote transmitter to stall. This
1255 * flag tells soreceive() to disable delayed acknowledgements when
1256 * draining the buffer. This can occur if the receiver is attempting
1257 * to read more data than can be buffered prior to transmitting on
1260 if (th->th_win == 0) {
1262 tp->t_flags |= TF_RXWIN0SENT;
1264 tp->t_flags &= ~TF_RXWIN0SENT;
1265 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1266 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1267 th->th_flags |= TH_URG;
1270 * If no urgent pointer to send, then we pull
1271 * the urgent pointer to the left edge of the send window
1272 * so that it doesn't drift into the send window on sequence
1273 * number wraparound.
1275 tp->snd_up = tp->snd_una; /* drag it along */
1278 * Put TCP length in extended header, and then
1279 * checksum extended header and data.
1281 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1282 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1284 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1285 if (to.to_flags & TOF_SIGNATURE) {
1287 * Calculate MD5 signature and put it into the place
1288 * determined before.
1289 * NOTE: since TCP options buffer doesn't point into
1290 * mbuf's data, calculate offset and use it.
1292 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
1293 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
1295 * Do not send segment if the calculation of MD5
1296 * digest has failed.
1305 * There is no need to fill in ip6_plen right now.
1306 * It will be filled later by ip6_output.
1308 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1309 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1310 optlen + len, IPPROTO_TCP, 0);
1313 #if defined(INET6) && defined(INET)
1318 m->m_pkthdr.csum_flags = CSUM_TCP;
1319 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1320 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1322 /* IP version must be set here for ipv4/ipv6 checking later */
1323 KASSERT(ip->ip_v == IPVERSION,
1324 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1329 * Enable TSO and specify the size of the segments.
1330 * The TCP pseudo header checksum is always provided.
1333 KASSERT(len > tp->t_maxseg - optlen,
1334 ("%s: len <= tso_segsz", __func__));
1335 m->m_pkthdr.csum_flags |= CSUM_TSO;
1336 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1339 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1340 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1341 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1342 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1344 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1345 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1346 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1350 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1351 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1358 if (so->so_options & SO_DEBUG) {
1364 save = ipov->ih_len;
1365 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1367 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1371 ipov->ih_len = save;
1373 #endif /* TCPDEBUG */
1374 TCP_PROBE3(debug__output, tp, th, m);
1377 * Fill in IP length and desired time to live and
1378 * send to IP level. There should be a better way
1379 * to handle ttl and tos; we could keep them in
1380 * the template, but need a way to checksum without them.
1383 * m->m_pkthdr.len should have been set before checksum calculation,
1384 * because in6_cksum() need it.
1389 * we separately set hoplimit for every segment, since the
1390 * user might want to change the value via setsockopt.
1391 * Also, desired default hop limit might be changed via
1392 * Neighbor Discovery.
1394 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1397 * Set the packet size here for the benefit of DTrace probes.
1398 * ip6_output() will set it properly; it's supposed to include
1399 * the option header lengths as well.
1401 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1403 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1404 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1406 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1408 if (tp->t_state == TCPS_SYN_SENT)
1409 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1411 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1414 /* Save packet, if requested. */
1415 tcp_pcap_add(th, m, &(tp->t_outpkts));
1418 /* TODO: IPv6 IP6TOS_ECT bit on */
1419 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1420 &tp->t_inpcb->inp_route6,
1421 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1422 NULL, NULL, tp->t_inpcb);
1424 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1425 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1428 #if defined(INET) && defined(INET6)
1433 ip->ip_len = htons(m->m_pkthdr.len);
1435 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1436 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1439 * If we do path MTU discovery, then we set DF on every packet.
1440 * This might not be the best thing to do according to RFC3390
1441 * Section 2. However the tcp hostcache migitates the problem
1442 * so it affects only the first tcp connection with a host.
1444 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1446 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1447 ip->ip_off |= htons(IP_DF);
1448 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1450 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1453 if (tp->t_state == TCPS_SYN_SENT)
1454 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1456 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1459 /* Save packet, if requested. */
1460 tcp_pcap_add(th, m, &(tp->t_outpkts));
1463 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1464 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1467 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1468 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1474 * In transmit state, time the transmission and arrange for
1475 * the retransmit. In persist state, just set snd_max.
1477 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1478 !tcp_timer_active(tp, TT_PERSIST)) {
1479 tcp_seq startseq = tp->snd_nxt;
1482 * Advance snd_nxt over sequence space of this segment.
1484 if (flags & (TH_SYN|TH_FIN)) {
1487 if (flags & TH_FIN) {
1489 tp->t_flags |= TF_SENTFIN;
1495 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1496 tp->snd_max = tp->snd_nxt;
1498 * Time this transmission if not a retransmission and
1499 * not currently timing anything.
1501 if (tp->t_rtttime == 0) {
1502 tp->t_rtttime = ticks;
1503 tp->t_rtseq = startseq;
1504 TCPSTAT_INC(tcps_segstimed);
1509 * Set retransmit timer if not currently set,
1510 * and not doing a pure ack or a keep-alive probe.
1511 * Initial value for retransmit timer is smoothed
1512 * round-trip time + 2 * round-trip time variance.
1513 * Initialize shift counter which is used for backoff
1514 * of retransmit time.
1517 if (!tcp_timer_active(tp, TT_REXMT) &&
1518 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1519 (tp->snd_nxt != tp->snd_una))) {
1520 if (tcp_timer_active(tp, TT_PERSIST)) {
1521 tcp_timer_activate(tp, TT_PERSIST, 0);
1524 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1525 } else if (len == 0 && sbavail(&so->so_snd) &&
1526 !tcp_timer_active(tp, TT_REXMT) &&
1527 !tcp_timer_active(tp, TT_PERSIST)) {
1529 * Avoid a situation where we do not set persist timer
1530 * after a zero window condition. For example:
1531 * 1) A -> B: packet with enough data to fill the window
1532 * 2) B -> A: ACK for #1 + new data (0 window
1534 * 3) A -> B: ACK for #2, 0 len packet
1536 * In this case, A will not activate the persist timer,
1537 * because it chose to send a packet. Unless tcp_output
1538 * is called for some other reason (delayed ack timer,
1539 * another input packet from B, socket syscall), A will
1540 * not send zero window probes.
1542 * So, if you send a 0-length packet, but there is data
1543 * in the socket buffer, and neither the rexmt or
1544 * persist timer is already set, then activate the
1552 * Persist case, update snd_max but since we are in
1553 * persist mode (no window) we do not update snd_nxt.
1558 if (flags & TH_FIN) {
1560 tp->t_flags |= TF_SENTFIN;
1562 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1563 tp->snd_max = tp->snd_nxt + xlen;
1569 * We know that the packet was lost, so back out the
1570 * sequence number advance, if any.
1572 * If the error is EPERM the packet got blocked by the
1573 * local firewall. Normally we should terminate the
1574 * connection but the blocking may have been spurious
1575 * due to a firewall reconfiguration cycle. So we treat
1576 * it like a packet loss and let the retransmit timer and
1577 * timeouts do their work over time.
1578 * XXX: It is a POLA question whether calling tcp_drop right
1579 * away would be the really correct behavior instead.
1581 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1582 !tcp_timer_active(tp, TT_PERSIST)) &&
1583 ((flags & TH_SYN) == 0) &&
1587 tp->sackhint.sack_bytes_rexmit -= len;
1588 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1589 ("sackhint bytes rtx >= 0"));
1593 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1596 tp->t_softerror = error;
1599 tp->t_softerror = error;
1602 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1603 tp->snd_cwnd = tp->t_maxseg;
1607 * For some reason the interface we used initially
1608 * to send segments changed to another or lowered
1610 * If TSO was active we either got an interface
1611 * without TSO capabilits or TSO was turned off.
1612 * If we obtained mtu from ip_output() then update
1616 tp->t_flags &= ~TF_TSO;
1618 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1626 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1627 tp->t_softerror = error;
1635 TCPSTAT_INC(tcps_sndtotal);
1638 * Data sent (as far as we can tell).
1639 * If this advertises a larger window than any other segment,
1640 * then remember the size of the advertised window.
1641 * Any pending ACK has now been sent.
1643 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1644 tp->rcv_adv = tp->rcv_nxt + recwin;
1645 tp->last_ack_sent = tp->rcv_nxt;
1646 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1647 if (tcp_timer_active(tp, TT_DELACK))
1648 tcp_timer_activate(tp, TT_DELACK, 0);
1651 * This completely breaks TCP if newreno is turned on. What happens
1652 * is that if delayed-acks are turned on on the receiver, this code
1653 * on the transmitter effectively destroys the TCP window, forcing
1654 * it to four packets (1.5Kx4 = 6K window).
1656 if (sendalot && --maxburst)
1665 tcp_setpersist(struct tcpcb *tp)
1667 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1670 tp->t_flags &= ~TF_PREVVALID;
1671 if (tcp_timer_active(tp, TT_REXMT))
1672 panic("tcp_setpersist: retransmit pending");
1674 * Start/restart persistence timer.
1676 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1677 tcp_persmin, tcp_persmax);
1678 tcp_timer_activate(tp, TT_PERSIST, tt);
1679 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1684 * Insert TCP options according to the supplied parameters to the place
1685 * optp in a consistent way. Can handle unaligned destinations.
1687 * The order of the option processing is crucial for optimal packing and
1688 * alignment for the scarce option space.
1690 * The optimal order for a SYN/SYN-ACK segment is:
1691 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1692 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1694 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1695 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1696 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1697 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1698 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1701 tcp_addoptions(struct tcpopt *to, u_char *optp)
1703 u_int32_t mask, optlen = 0;
1705 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1706 if ((to->to_flags & mask) != mask)
1708 if (optlen == TCP_MAXOLEN)
1710 switch (to->to_flags & mask) {
1712 while (optlen % 4) {
1713 optlen += TCPOLEN_NOP;
1714 *optp++ = TCPOPT_NOP;
1716 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1718 optlen += TCPOLEN_MAXSEG;
1719 *optp++ = TCPOPT_MAXSEG;
1720 *optp++ = TCPOLEN_MAXSEG;
1721 to->to_mss = htons(to->to_mss);
1722 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1723 optp += sizeof(to->to_mss);
1726 while (!optlen || optlen % 2 != 1) {
1727 optlen += TCPOLEN_NOP;
1728 *optp++ = TCPOPT_NOP;
1730 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1732 optlen += TCPOLEN_WINDOW;
1733 *optp++ = TCPOPT_WINDOW;
1734 *optp++ = TCPOLEN_WINDOW;
1735 *optp++ = to->to_wscale;
1738 while (optlen % 2) {
1739 optlen += TCPOLEN_NOP;
1740 *optp++ = TCPOPT_NOP;
1742 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1744 optlen += TCPOLEN_SACK_PERMITTED;
1745 *optp++ = TCPOPT_SACK_PERMITTED;
1746 *optp++ = TCPOLEN_SACK_PERMITTED;
1749 while (!optlen || optlen % 4 != 2) {
1750 optlen += TCPOLEN_NOP;
1751 *optp++ = TCPOPT_NOP;
1753 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1755 optlen += TCPOLEN_TIMESTAMP;
1756 *optp++ = TCPOPT_TIMESTAMP;
1757 *optp++ = TCPOLEN_TIMESTAMP;
1758 to->to_tsval = htonl(to->to_tsval);
1759 to->to_tsecr = htonl(to->to_tsecr);
1760 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1761 optp += sizeof(to->to_tsval);
1762 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1763 optp += sizeof(to->to_tsecr);
1767 int siglen = TCPOLEN_SIGNATURE - 2;
1769 while (!optlen || optlen % 4 != 2) {
1770 optlen += TCPOLEN_NOP;
1771 *optp++ = TCPOPT_NOP;
1773 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1774 to->to_flags &= ~TOF_SIGNATURE;
1777 optlen += TCPOLEN_SIGNATURE;
1778 *optp++ = TCPOPT_SIGNATURE;
1779 *optp++ = TCPOLEN_SIGNATURE;
1780 to->to_signature = optp;
1788 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1791 while (!optlen || optlen % 4 != 2) {
1792 optlen += TCPOLEN_NOP;
1793 *optp++ = TCPOPT_NOP;
1795 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1797 optlen += TCPOLEN_SACKHDR;
1798 *optp++ = TCPOPT_SACK;
1799 sackblks = min(to->to_nsacks,
1800 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1801 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1802 while (sackblks--) {
1803 sack_seq = htonl(sack->start);
1804 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1805 optp += sizeof(sack_seq);
1806 sack_seq = htonl(sack->end);
1807 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1808 optp += sizeof(sack_seq);
1809 optlen += TCPOLEN_SACK;
1812 TCPSTAT_INC(tcps_sack_send_blocks);
1820 /* XXX is there any point to aligning this option? */
1821 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1822 if (TCP_MAXOLEN - optlen < total_len)
1824 *optp++ = TCPOPT_FAST_OPEN;
1825 *optp++ = total_len;
1826 if (to->to_tfo_len > 0) {
1827 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1828 optp += to->to_tfo_len;
1830 optlen += total_len;
1835 panic("%s: unknown TCP option type", __func__);
1840 /* Terminate and pad TCP options to a 4 byte boundary. */
1842 optlen += TCPOLEN_EOL;
1843 *optp++ = TCPOPT_EOL;
1846 * According to RFC 793 (STD0007):
1847 * "The content of the header beyond the End-of-Option option
1848 * must be header padding (i.e., zero)."
1849 * and later: "The padding is composed of zeros."
1851 while (optlen % 4) {
1852 optlen += TCPOLEN_PAD;
1853 *optp++ = TCPOPT_PAD;
1856 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));