2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
46 #include <sys/hhook.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
59 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
75 #include <netinet/tcp_fastopen.h>
77 #include <netinet/tcp.h>
79 #include <netinet/tcp_fsm.h>
80 #include <netinet/tcp_seq.h>
81 #include <netinet/tcp_timer.h>
82 #include <netinet/tcp_var.h>
83 #include <netinet/tcpip.h>
84 #include <netinet/cc/cc.h>
86 #include <netinet/tcp_pcap.h>
89 #include <netinet/tcp_debug.h>
92 #include <netinet/tcp_offload.h>
95 #include <netipsec/ipsec_support.h>
97 #include <machine/in_cksum.h>
99 #include <security/mac/mac_framework.h>
101 VNET_DEFINE(int, path_mtu_discovery) = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
103 &VNET_NAME(path_mtu_discovery), 1,
104 "Enable Path MTU Discovery");
106 VNET_DEFINE(int, tcp_do_tso) = 1;
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
108 &VNET_NAME(tcp_do_tso), 0,
109 "Enable TCP Segmentation Offload");
111 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
112 #define V_tcp_sendspace VNET(tcp_sendspace)
113 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
114 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
118 &VNET_NAME(tcp_do_autosndbuf), 0,
119 "Enable automatic send buffer sizing");
121 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(tcp_autosndbuf_inc), 0,
124 "Incrementor step size of automatic send buffer");
126 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
128 &VNET_NAME(tcp_autosndbuf_max), 0,
129 "Max size of automatic send buffer");
131 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
132 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
134 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
135 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
138 * Make sure that either retransmit or persist timer is set for SYN, FIN and
141 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
142 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
143 tcp_timer_active((tp), TT_REXMT) || \
144 tcp_timer_active((tp), TT_PERSIST), \
145 ("neither rexmt nor persist timer is set"))
148 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
149 struct tcphdr *th, struct tcpopt *to,
150 uint32_t len, int tso);
152 static void inline cc_after_idle(struct tcpcb *tp);
156 * Wrapper for the TCP established output helper hook.
159 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
160 struct tcpopt *to, uint32_t len, int tso)
162 struct tcp_hhook_data hhook_data;
164 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
168 hhook_data.len = len;
169 hhook_data.tso = tso;
171 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
178 * CC wrapper hook functions
181 cc_after_idle(struct tcpcb *tp)
183 INP_WLOCK_ASSERT(tp->t_inpcb);
185 if (CC_ALGO(tp)->after_idle != NULL)
186 CC_ALGO(tp)->after_idle(tp->ccv);
190 * Tcp output routine: figure out what should be sent and send it.
193 tcp_output(struct tcpcb *tp)
195 struct socket *so = tp->t_inpcb->inp_socket;
197 uint32_t recwin, sendwin;
198 int off, flags, error = 0; /* Keep compiler happy */
200 struct ip *ip = NULL;
201 struct ipovly *ipov = NULL;
203 u_char opt[TCP_MAXOLEN];
204 unsigned ipoptlen, optlen, hdrlen;
205 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
206 unsigned ipsec_optlen = 0;
209 int sack_rxmit, sack_bytes_rxmt;
214 int maxburst = TCP_MAXBURST;
217 struct ip6_hdr *ip6 = NULL;
220 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
223 INP_WLOCK_ASSERT(tp->t_inpcb);
226 if (tp->t_flags & TF_TOE)
227 return (tcp_offload_output(tp));
232 * For TFO connections in SYN_RECEIVED, only allow the initial
233 * SYN|ACK and those sent by the retransmit timer.
235 if (IS_FASTOPEN(tp->t_flags) &&
236 (tp->t_state == TCPS_SYN_RECEIVED) &&
237 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
238 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
242 * Determine length of data that should be transmitted,
243 * and flags that will be used.
244 * If there is some data or critical controls (SYN, RST)
245 * to send, then transmit; otherwise, investigate further.
247 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
248 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
250 tp->t_flags &= ~TF_LASTIDLE;
252 if (tp->t_flags & TF_MORETOCOME) {
253 tp->t_flags |= TF_LASTIDLE;
259 * If we've recently taken a timeout, snd_max will be greater than
260 * snd_nxt. There may be SACK information that allows us to avoid
261 * resending already delivered data. Adjust snd_nxt accordingly.
263 if ((tp->t_flags & TF_SACK_PERMIT) &&
264 SEQ_LT(tp->snd_nxt, tp->snd_max))
269 off = tp->snd_nxt - tp->snd_una;
270 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
272 flags = tcp_outflags[tp->t_state];
274 * Send any SACK-generated retransmissions. If we're explicitly trying
275 * to send out new data (when sendalot is 1), bypass this function.
276 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
277 * we're replacing a (future) new transmission with a retransmission
278 * now, and we previously incremented snd_cwnd in tcp_input().
281 * Still in sack recovery , reset rxmit flag to zero.
287 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
288 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
292 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
293 /* Do not retransmit SACK segments beyond snd_recover */
294 if (SEQ_GT(p->end, tp->snd_recover)) {
296 * (At least) part of sack hole extends beyond
297 * snd_recover. Check to see if we can rexmit data
300 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
302 * Can't rexmit any more data for this hole.
303 * That data will be rexmitted in the next
304 * sack recovery episode, when snd_recover
305 * moves past p->rxmit.
308 goto after_sack_rexmit;
310 /* Can rexmit part of the current hole */
311 len = ((int32_t)ulmin(cwin,
312 tp->snd_recover - p->rxmit));
314 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
315 off = p->rxmit - tp->snd_una;
316 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
321 TCPSTAT_INC(tcps_sack_rexmits);
322 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
323 min(len, tp->t_maxseg));
328 * Get standard flags, and add SYN or FIN if requested by 'hidden'
331 if (tp->t_flags & TF_NEEDFIN)
333 if (tp->t_flags & TF_NEEDSYN)
336 SOCKBUF_LOCK(&so->so_snd);
338 * If in persist timeout with window of 0, send 1 byte.
339 * Otherwise, if window is small but nonzero
340 * and timer expired, we will send what we can
341 * and go to transmit state.
343 if (tp->t_flags & TF_FORCEDATA) {
346 * If we still have some data to send, then
347 * clear the FIN bit. Usually this would
348 * happen below when it realizes that we
349 * aren't sending all the data. However,
350 * if we have exactly 1 byte of unsent data,
351 * then it won't clear the FIN bit below,
352 * and if we are in persist state, we wind
353 * up sending the packet without recording
354 * that we sent the FIN bit.
356 * We can't just blindly clear the FIN bit,
357 * because if we don't have any more data
358 * to send then the probe will be the FIN
361 if (off < sbused(&so->so_snd))
365 tcp_timer_activate(tp, TT_PERSIST, 0);
371 * If snd_nxt == snd_max and we have transmitted a FIN, the
372 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
373 * a negative length. This can also occur when TCP opens up
374 * its congestion window while receiving additional duplicate
375 * acks after fast-retransmit because TCP will reset snd_nxt
376 * to snd_max after the fast-retransmit.
378 * In the normal retransmit-FIN-only case, however, snd_nxt will
379 * be set to snd_una, the offset will be 0, and the length may
382 * If sack_rxmit is true we are retransmitting from the scoreboard
383 * in which case len is already set.
385 if (sack_rxmit == 0) {
386 if (sack_bytes_rxmt == 0)
387 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
393 * We are inside of a SACK recovery episode and are
394 * sending new data, having retransmitted all the
395 * data possible in the scoreboard.
397 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
400 * Don't remove this (len > 0) check !
401 * We explicitly check for len > 0 here (although it
402 * isn't really necessary), to work around a gcc
403 * optimization issue - to force gcc to compute
404 * len above. Without this check, the computation
405 * of len is bungled by the optimizer.
408 cwin = tp->snd_cwnd -
409 (tp->snd_nxt - tp->sack_newdata) -
413 len = imin(len, cwin);
419 * Lop off SYN bit if it has already been sent. However, if this
420 * is SYN-SENT state and if segment contains data and if we don't
421 * know that foreign host supports TAO, suppress sending segment.
423 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
424 if (tp->t_state != TCPS_SYN_RECEIVED)
428 * When sending additional segments following a TFO SYN|ACK,
429 * do not include the SYN bit.
431 if (IS_FASTOPEN(tp->t_flags) &&
432 (tp->t_state == TCPS_SYN_RECEIVED))
439 * Be careful not to send data and/or FIN on SYN segments.
440 * This measure is needed to prevent interoperability problems
441 * with not fully conformant TCP implementations.
443 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
450 * When retransmitting SYN|ACK on a passively-created TFO socket,
451 * don't include data, as the presence of data may have caused the
452 * original SYN|ACK to have been dropped by a middlebox.
454 if (IS_FASTOPEN(tp->t_flags) &&
455 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
461 * If FIN has been sent but not acked,
462 * but we haven't been called to retransmit,
463 * len will be < 0. Otherwise, window shrank
464 * after we sent into it. If window shrank to 0,
465 * cancel pending retransmit, pull snd_nxt back
466 * to (closed) window, and set the persist timer
467 * if it isn't already going. If the window didn't
468 * close completely, just wait for an ACK.
470 * We also do a general check here to ensure that
471 * we will set the persist timer when we have data
472 * to send, but a 0-byte window. This makes sure
473 * the persist timer is set even if the packet
474 * hits one of the "goto send" lines below.
477 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
478 (off < (int) sbavail(&so->so_snd))) {
479 tcp_timer_activate(tp, TT_REXMT, 0);
481 tp->snd_nxt = tp->snd_una;
482 if (!tcp_timer_active(tp, TT_PERSIST))
487 /* len will be >= 0 after this point. */
488 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
491 * Automatic sizing of send socket buffer. Often the send buffer
492 * size is not optimally adjusted to the actual network conditions
493 * at hand (delay bandwidth product). Setting the buffer size too
494 * small limits throughput on links with high bandwidth and high
495 * delay (eg. trans-continental/oceanic links). Setting the
496 * buffer size too big consumes too much real kernel memory,
497 * especially with many connections on busy servers.
499 * The criteria to step up the send buffer one notch are:
500 * 1. receive window of remote host is larger than send buffer
501 * (with a fudge factor of 5/4th);
502 * 2. send buffer is filled to 7/8th with data (so we actually
503 * have data to make use of it);
504 * 3. send buffer fill has not hit maximal automatic size;
505 * 4. our send window (slow start and cogestion controlled) is
506 * larger than sent but unacknowledged data in send buffer.
508 * The remote host receive window scaling factor may limit the
509 * growing of the send buffer before it reaches its allowed
512 * It scales directly with slow start or congestion window
513 * and does at most one step per received ACK. This fast
514 * scaling has the drawback of growing the send buffer beyond
515 * what is strictly necessary to make full use of a given
516 * delay*bandwidth product. However testing has shown this not
517 * to be much of an problem. At worst we are trading wasting
518 * of available bandwidth (the non-use of it) for wasting some
519 * socket buffer memory.
521 * TODO: Shrink send buffer during idle periods together
522 * with congestion window. Requires another timer. Has to
523 * wait for upcoming tcp timer rewrite.
525 * XXXGL: should there be used sbused() or sbavail()?
527 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
530 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
531 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
532 sbused(&so->so_snd) >=
533 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
534 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
535 sendwin >= (sbused(&so->so_snd) -
536 (tp->snd_nxt - tp->snd_una))) {
537 if (!sbreserve_locked(&so->so_snd,
538 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
539 V_tcp_autosndbuf_max), so, curthread))
540 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
545 * Decide if we can use TCP Segmentation Offloading (if supported by
548 * TSO may only be used if we are in a pure bulk sending state. The
549 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
550 * IP options prevent using TSO. With TSO the TCP header is the same
551 * (except for the sequence number) for all generated packets. This
552 * makes it impossible to transmit any options which vary per generated
555 * IPv4 handling has a clear separation of ip options and ip header
556 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
557 * the right thing below to provide length of just ip options and thus
558 * checking for ipoptlen is enough to decide if ip options are present.
560 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
562 * Pre-calculate here as we save another lookup into the darknesses
563 * of IPsec that way and can actually decide if TSO is ok.
566 if (isipv6 && IPSEC_ENABLED(ipv6))
567 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
573 if (IPSEC_ENABLED(ipv4))
574 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
579 ipoptlen = ip6_optlen(tp->t_inpcb);
582 if (tp->t_inpcb->inp_options)
583 ipoptlen = tp->t_inpcb->inp_options->m_len -
584 offsetof(struct ipoption, ipopt_list);
587 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
588 ipoptlen += ipsec_optlen;
591 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
592 ((tp->t_flags & TF_SIGNATURE) == 0) &&
593 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
598 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
601 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
602 sbused(&so->so_snd)))
606 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
607 (long)TCP_MAXWIN << tp->rcv_scale);
610 * Sender silly window avoidance. We transmit under the following
611 * conditions when len is non-zero:
613 * - We have a full segment (or more with TSO)
614 * - This is the last buffer in a write()/send() and we are
615 * either idle or running NODELAY
616 * - we've timed out (e.g. persist timer)
617 * - we have more then 1/2 the maximum send window's worth of
618 * data (receiver may be limited the window size)
619 * - we need to retransmit
622 if (len >= tp->t_maxseg)
625 * NOTE! on localhost connections an 'ack' from the remote
626 * end may occur synchronously with the output and cause
627 * us to flush a buffer queued with moretocome. XXX
629 * note: the len + off check is almost certainly unnecessary.
631 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
632 (idle || (tp->t_flags & TF_NODELAY)) &&
633 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
634 (tp->t_flags & TF_NOPUSH) == 0) {
637 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
639 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
641 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
648 * Sending of standalone window updates.
650 * Window updates are important when we close our window due to a
651 * full socket buffer and are opening it again after the application
652 * reads data from it. Once the window has opened again and the
653 * remote end starts to send again the ACK clock takes over and
654 * provides the most current window information.
656 * We must avoid the silly window syndrome whereas every read
657 * from the receive buffer, no matter how small, causes a window
658 * update to be sent. We also should avoid sending a flurry of
659 * window updates when the socket buffer had queued a lot of data
660 * and the application is doing small reads.
662 * Prevent a flurry of pointless window updates by only sending
663 * an update when we can increase the advertized window by more
664 * than 1/4th of the socket buffer capacity. When the buffer is
665 * getting full or is very small be more aggressive and send an
666 * update whenever we can increase by two mss sized segments.
667 * In all other situations the ACK's to new incoming data will
668 * carry further window increases.
670 * Don't send an independent window update if a delayed
671 * ACK is pending (it will get piggy-backed on it) or the
672 * remote side already has done a half-close and won't send
673 * more data. Skip this if the connection is in T/TCP
676 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
677 !(tp->t_flags & TF_DELACK) &&
678 !TCPS_HAVERCVDFIN(tp->t_state)) {
680 * "adv" is the amount we could increase the window,
681 * taking into account that we are limited by
682 * TCP_MAXWIN << tp->rcv_scale.
688 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
689 oldwin = (tp->rcv_adv - tp->rcv_nxt);
695 * If the new window size ends up being the same as or less
696 * than the old size when it is scaled, then don't force
699 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
702 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
703 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
704 recwin <= (so->so_rcv.sb_hiwat / 8) ||
705 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
707 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
713 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
714 * is also a catch-all for the retransmit timer timeout case.
716 if (tp->t_flags & TF_ACKNOW)
718 if ((flags & TH_RST) ||
719 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
721 if (SEQ_GT(tp->snd_up, tp->snd_una))
724 * If our state indicates that FIN should be sent
725 * and we have not yet done so, then we need to send.
727 if (flags & TH_FIN &&
728 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
731 * In SACK, it is possible for tcp_output to fail to send a segment
732 * after the retransmission timer has been turned off. Make sure
733 * that the retransmission timer is set.
735 if ((tp->t_flags & TF_SACK_PERMIT) &&
736 SEQ_GT(tp->snd_max, tp->snd_una) &&
737 !tcp_timer_active(tp, TT_REXMT) &&
738 !tcp_timer_active(tp, TT_PERSIST)) {
739 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
743 * TCP window updates are not reliable, rather a polling protocol
744 * using ``persist'' packets is used to insure receipt of window
745 * updates. The three ``states'' for the output side are:
746 * idle not doing retransmits or persists
747 * persisting to move a small or zero window
748 * (re)transmitting and thereby not persisting
750 * tcp_timer_active(tp, TT_PERSIST)
751 * is true when we are in persist state.
752 * (tp->t_flags & TF_FORCEDATA)
753 * is set when we are called to send a persist packet.
754 * tcp_timer_active(tp, TT_REXMT)
755 * is set when we are retransmitting
756 * The output side is idle when both timers are zero.
758 * If send window is too small, there is data to transmit, and no
759 * retransmit or persist is pending, then go to persist state.
760 * If nothing happens soon, send when timer expires:
761 * if window is nonzero, transmit what we can,
762 * otherwise force out a byte.
764 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
765 !tcp_timer_active(tp, TT_PERSIST)) {
771 * No reason to send a segment, just return.
774 SOCKBUF_UNLOCK(&so->so_snd);
778 SOCKBUF_LOCK_ASSERT(&so->so_snd);
780 if (len >= tp->t_maxseg)
781 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
783 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
786 * Before ESTABLISHED, force sending of initial options
787 * unless TCP set not to do any options.
788 * NOTE: we assume that the IP/TCP header plus TCP options
789 * always fit in a single mbuf, leaving room for a maximum
791 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
796 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
799 hdrlen = sizeof (struct tcpiphdr);
802 * Compute options for segment.
803 * We only have to care about SYN and established connection
804 * segments. Options for SYN-ACK segments are handled in TCP
808 if ((tp->t_flags & TF_NOOPT) == 0) {
809 /* Maximum segment size. */
810 if (flags & TH_SYN) {
811 tp->snd_nxt = tp->iss;
812 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
813 to.to_flags |= TOF_MSS;
816 * Only include the TFO option on the first
817 * transmission of the SYN|ACK on a
818 * passively-created TFO socket, as the presence of
819 * the TFO option may have caused the original
820 * SYN|ACK to have been dropped by a middlebox.
822 if (IS_FASTOPEN(tp->t_flags) &&
823 (tp->t_state == TCPS_SYN_RECEIVED) &&
824 (tp->t_rxtshift == 0)) {
825 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
826 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
827 to.to_flags |= TOF_FASTOPEN;
831 /* Window scaling. */
832 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
833 to.to_wscale = tp->request_r_scale;
834 to.to_flags |= TOF_SCALE;
837 if ((tp->t_flags & TF_RCVD_TSTMP) ||
838 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
839 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
840 to.to_tsecr = tp->ts_recent;
841 to.to_flags |= TOF_TS;
844 /* Set receive buffer autosizing timestamp. */
845 if (tp->rfbuf_ts == 0 &&
846 (so->so_rcv.sb_flags & SB_AUTOSIZE))
847 tp->rfbuf_ts = tcp_ts_getticks();
849 /* Selective ACK's. */
850 if (tp->t_flags & TF_SACK_PERMIT) {
852 to.to_flags |= TOF_SACKPERM;
853 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
854 (tp->t_flags & TF_SACK_PERMIT) &&
855 tp->rcv_numsacks > 0) {
856 to.to_flags |= TOF_SACK;
857 to.to_nsacks = tp->rcv_numsacks;
858 to.to_sacks = (u_char *)tp->sackblks;
861 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
862 /* TCP-MD5 (RFC2385). */
864 * Check that TCP_MD5SIG is enabled in tcpcb to
865 * account the size needed to set this TCP option.
867 if (tp->t_flags & TF_SIGNATURE)
868 to.to_flags |= TOF_SIGNATURE;
869 #endif /* TCP_SIGNATURE */
871 /* Processing the options. */
872 hdrlen += optlen = tcp_addoptions(&to, opt);
876 * Adjust data length if insertion of options will
877 * bump the packet length beyond the t_maxseg length.
878 * Clear the FIN bit because we cut off the tail of
881 if (len + optlen + ipoptlen > tp->t_maxseg) {
886 u_int if_hw_tsomaxsegcount;
887 u_int if_hw_tsomaxsegsize;
892 /* extract TSO information */
893 if_hw_tsomax = tp->t_tsomax;
894 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
895 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
898 * Limit a TSO burst to prevent it from
899 * overflowing or exceeding the maximum length
900 * allowed by the network interface:
902 KASSERT(ipoptlen == 0,
903 ("%s: TSO can't do IP options", __func__));
906 * Check if we should limit by maximum payload
909 if (if_hw_tsomax != 0) {
910 /* compute maximum TSO length */
911 max_len = (if_hw_tsomax - hdrlen -
915 } else if (len > max_len) {
922 * Check if we should limit by maximum segment
925 if (if_hw_tsomaxsegcount != 0 &&
926 if_hw_tsomaxsegsize != 0) {
928 * Subtract one segment for the LINK
929 * and TCP/IP headers mbuf that will
930 * be prepended to this mbuf chain
931 * after the code in this section
932 * limits the number of mbufs in the
933 * chain to if_hw_tsomaxsegcount.
935 if_hw_tsomaxsegcount -= 1;
937 mb = sbsndmbuf(&so->so_snd, off, &moff);
939 while (mb != NULL && max_len < len) {
944 * Get length of mbuf fragment
945 * and how many hardware frags,
946 * rounded up, it would use:
948 mlen = (mb->m_len - moff);
949 frags = howmany(mlen,
950 if_hw_tsomaxsegsize);
952 /* Handle special case: Zero Length Mbuf */
957 * Check if the fragment limit
958 * will be reached or exceeded:
960 if (frags >= if_hw_tsomaxsegcount) {
962 if_hw_tsomaxsegcount *
963 if_hw_tsomaxsegsize);
967 if_hw_tsomaxsegcount -= frags;
973 } else if (len > max_len) {
980 * Prevent the last segment from being
981 * fractional unless the send sockbuf can be
984 max_len = (tp->t_maxseg - optlen);
985 if (((uint32_t)off + (uint32_t)len) <
986 sbavail(&so->so_snd)) {
987 moff = len % max_len;
995 * In case there are too many small fragments
998 if (len <= max_len) {
1005 * Send the FIN in a separate segment
1006 * after the bulk sending is done.
1007 * We don't trust the TSO implementations
1008 * to clear the FIN flag on all but the
1011 if (tp->t_flags & TF_NEEDFIN)
1015 len = tp->t_maxseg - optlen - ipoptlen;
1021 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1022 ("%s: len > IP_MAXPACKET", __func__));
1024 /*#ifdef DIAGNOSTIC*/
1026 if (max_linkhdr + hdrlen > MCLBYTES)
1028 if (max_linkhdr + hdrlen > MHLEN)
1030 panic("tcphdr too big");
1034 * This KASSERT is here to catch edge cases at a well defined place.
1035 * Before, those had triggered (random) panic conditions further down.
1037 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1040 * Grab a header mbuf, attaching a copy of data to
1041 * be transmitted, and initialize the header from
1042 * the template for sends on this connection.
1048 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1049 TCPSTAT_INC(tcps_sndprobe);
1050 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1051 tp->t_sndrexmitpack++;
1052 TCPSTAT_INC(tcps_sndrexmitpack);
1053 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1055 TCPSTAT_INC(tcps_sndpack);
1056 TCPSTAT_ADD(tcps_sndbyte, len);
1059 if (MHLEN < hdrlen + max_linkhdr)
1060 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1063 m = m_gethdr(M_NOWAIT, MT_DATA);
1066 SOCKBUF_UNLOCK(&so->so_snd);
1072 m->m_data += max_linkhdr;
1076 * Start the m_copy functions from the closest mbuf
1077 * to the offset in the socket buffer chain.
1079 mb = sbsndptr(&so->so_snd, off, len, &moff);
1081 if (len <= MHLEN - hdrlen - max_linkhdr) {
1082 m_copydata(mb, moff, len,
1083 mtod(m, caddr_t) + hdrlen);
1086 m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1087 if (m->m_next == NULL) {
1088 SOCKBUF_UNLOCK(&so->so_snd);
1097 * If we're sending everything we've got, set PUSH.
1098 * (This will keep happy those implementations which only
1099 * give data to the user when a buffer fills or
1102 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1105 SOCKBUF_UNLOCK(&so->so_snd);
1107 SOCKBUF_UNLOCK(&so->so_snd);
1108 if (tp->t_flags & TF_ACKNOW)
1109 TCPSTAT_INC(tcps_sndacks);
1110 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1111 TCPSTAT_INC(tcps_sndctrl);
1112 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1113 TCPSTAT_INC(tcps_sndurg);
1115 TCPSTAT_INC(tcps_sndwinup);
1117 m = m_gethdr(M_NOWAIT, MT_DATA);
1124 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1129 m->m_data += max_linkhdr;
1132 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1133 m->m_pkthdr.rcvif = (struct ifnet *)0;
1135 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1139 ip6 = mtod(m, struct ip6_hdr *);
1140 th = (struct tcphdr *)(ip6 + 1);
1141 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1145 ip = mtod(m, struct ip *);
1146 ipov = (struct ipovly *)ip;
1147 th = (struct tcphdr *)(ip + 1);
1148 tcpip_fillheaders(tp->t_inpcb, ip, th);
1152 * Fill in fields, remembering maximum advertised
1153 * window for use in delaying messages about window sizes.
1154 * If resending a FIN, be sure not to use a new sequence number.
1156 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1157 tp->snd_nxt == tp->snd_max)
1160 * If we are starting a connection, send ECN setup
1161 * SYN packet. If we are on a retransmit, we may
1162 * resend those bits a number of times as per
1165 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1166 if (tp->t_rxtshift >= 1) {
1167 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1168 flags |= TH_ECE|TH_CWR;
1170 flags |= TH_ECE|TH_CWR;
1173 if (tp->t_state == TCPS_ESTABLISHED &&
1174 (tp->t_flags & TF_ECN_PERMIT)) {
1176 * If the peer has ECN, mark data packets with
1177 * ECN capable transmission (ECT).
1178 * Ignore pure ack packets, retransmissions and window probes.
1180 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1181 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1184 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1187 ip->ip_tos |= IPTOS_ECN_ECT0;
1188 TCPSTAT_INC(tcps_ecn_ect0);
1192 * Reply with proper ECN notifications.
1194 if (tp->t_flags & TF_ECN_SND_CWR) {
1196 tp->t_flags &= ~TF_ECN_SND_CWR;
1198 if (tp->t_flags & TF_ECN_SND_ECE)
1203 * If we are doing retransmissions, then snd_nxt will
1204 * not reflect the first unsent octet. For ACK only
1205 * packets, we do not want the sequence number of the
1206 * retransmitted packet, we want the sequence number
1207 * of the next unsent octet. So, if there is no data
1208 * (and no SYN or FIN), use snd_max instead of snd_nxt
1209 * when filling in ti_seq. But if we are in persist
1210 * state, snd_max might reflect one byte beyond the
1211 * right edge of the window, so use snd_nxt in that
1212 * case, since we know we aren't doing a retransmission.
1213 * (retransmit and persist are mutually exclusive...)
1215 if (sack_rxmit == 0) {
1216 if (len || (flags & (TH_SYN|TH_FIN)) ||
1217 tcp_timer_active(tp, TT_PERSIST))
1218 th->th_seq = htonl(tp->snd_nxt);
1220 th->th_seq = htonl(tp->snd_max);
1222 th->th_seq = htonl(p->rxmit);
1224 tp->sackhint.sack_bytes_rexmit += len;
1226 th->th_ack = htonl(tp->rcv_nxt);
1228 bcopy(opt, th + 1, optlen);
1229 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1231 th->th_flags = flags;
1233 * Calculate receive window. Don't shrink window,
1234 * but avoid silly window syndrome.
1236 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1237 recwin < tp->t_maxseg)
1239 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1240 recwin < (tp->rcv_adv - tp->rcv_nxt))
1241 recwin = (tp->rcv_adv - tp->rcv_nxt);
1244 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1245 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1246 * case is handled in syncache.
1249 th->th_win = htons((u_short)
1250 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1252 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1255 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1256 * a 0 window. This may cause the remote transmitter to stall. This
1257 * flag tells soreceive() to disable delayed acknowledgements when
1258 * draining the buffer. This can occur if the receiver is attempting
1259 * to read more data than can be buffered prior to transmitting on
1262 if (th->th_win == 0) {
1264 tp->t_flags |= TF_RXWIN0SENT;
1266 tp->t_flags &= ~TF_RXWIN0SENT;
1267 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1268 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1269 th->th_flags |= TH_URG;
1272 * If no urgent pointer to send, then we pull
1273 * the urgent pointer to the left edge of the send window
1274 * so that it doesn't drift into the send window on sequence
1275 * number wraparound.
1277 tp->snd_up = tp->snd_una; /* drag it along */
1280 * Put TCP length in extended header, and then
1281 * checksum extended header and data.
1283 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1284 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1286 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1287 if (to.to_flags & TOF_SIGNATURE) {
1289 * Calculate MD5 signature and put it into the place
1290 * determined before.
1291 * NOTE: since TCP options buffer doesn't point into
1292 * mbuf's data, calculate offset and use it.
1294 if (!TCPMD5_ENABLED() || TCPMD5_OUTPUT(m, th,
1295 (u_char *)(th + 1) + (to.to_signature - opt)) != 0) {
1297 * Do not send segment if the calculation of MD5
1298 * digest has failed.
1307 * There is no need to fill in ip6_plen right now.
1308 * It will be filled later by ip6_output.
1310 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1311 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1312 optlen + len, IPPROTO_TCP, 0);
1315 #if defined(INET6) && defined(INET)
1320 m->m_pkthdr.csum_flags = CSUM_TCP;
1321 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1322 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1324 /* IP version must be set here for ipv4/ipv6 checking later */
1325 KASSERT(ip->ip_v == IPVERSION,
1326 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1331 * Enable TSO and specify the size of the segments.
1332 * The TCP pseudo header checksum is always provided.
1335 KASSERT(len > tp->t_maxseg - optlen,
1336 ("%s: len <= tso_segsz", __func__));
1337 m->m_pkthdr.csum_flags |= CSUM_TSO;
1338 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1341 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1342 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1343 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1344 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1346 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1347 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1348 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1352 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1353 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1360 if (so->so_options & SO_DEBUG) {
1366 save = ipov->ih_len;
1367 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1369 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1373 ipov->ih_len = save;
1375 #endif /* TCPDEBUG */
1376 TCP_PROBE3(debug__output, tp, th, m);
1379 * Fill in IP length and desired time to live and
1380 * send to IP level. There should be a better way
1381 * to handle ttl and tos; we could keep them in
1382 * the template, but need a way to checksum without them.
1385 * m->m_pkthdr.len should have been set before checksum calculation,
1386 * because in6_cksum() need it.
1391 * we separately set hoplimit for every segment, since the
1392 * user might want to change the value via setsockopt.
1393 * Also, desired default hop limit might be changed via
1394 * Neighbor Discovery.
1396 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1399 * Set the packet size here for the benefit of DTrace probes.
1400 * ip6_output() will set it properly; it's supposed to include
1401 * the option header lengths as well.
1403 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1405 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1406 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1408 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1410 if (tp->t_state == TCPS_SYN_SENT)
1411 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1413 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1416 /* Save packet, if requested. */
1417 tcp_pcap_add(th, m, &(tp->t_outpkts));
1420 /* TODO: IPv6 IP6TOS_ECT bit on */
1421 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1422 &tp->t_inpcb->inp_route6,
1423 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1424 NULL, NULL, tp->t_inpcb);
1426 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1427 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1430 #if defined(INET) && defined(INET6)
1435 ip->ip_len = htons(m->m_pkthdr.len);
1437 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1438 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1441 * If we do path MTU discovery, then we set DF on every packet.
1442 * This might not be the best thing to do according to RFC3390
1443 * Section 2. However the tcp hostcache migitates the problem
1444 * so it affects only the first tcp connection with a host.
1446 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1448 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1449 ip->ip_off |= htons(IP_DF);
1450 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1452 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1455 if (tp->t_state == TCPS_SYN_SENT)
1456 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1458 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1461 /* Save packet, if requested. */
1462 tcp_pcap_add(th, m, &(tp->t_outpkts));
1465 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1466 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1469 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1470 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1476 * In transmit state, time the transmission and arrange for
1477 * the retransmit. In persist state, just set snd_max.
1479 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1480 !tcp_timer_active(tp, TT_PERSIST)) {
1481 tcp_seq startseq = tp->snd_nxt;
1484 * Advance snd_nxt over sequence space of this segment.
1486 if (flags & (TH_SYN|TH_FIN)) {
1489 if (flags & TH_FIN) {
1491 tp->t_flags |= TF_SENTFIN;
1497 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1498 tp->snd_max = tp->snd_nxt;
1500 * Time this transmission if not a retransmission and
1501 * not currently timing anything.
1503 if (tp->t_rtttime == 0) {
1504 tp->t_rtttime = ticks;
1505 tp->t_rtseq = startseq;
1506 TCPSTAT_INC(tcps_segstimed);
1511 * Set retransmit timer if not currently set,
1512 * and not doing a pure ack or a keep-alive probe.
1513 * Initial value for retransmit timer is smoothed
1514 * round-trip time + 2 * round-trip time variance.
1515 * Initialize shift counter which is used for backoff
1516 * of retransmit time.
1519 if (!tcp_timer_active(tp, TT_REXMT) &&
1520 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1521 (tp->snd_nxt != tp->snd_una))) {
1522 if (tcp_timer_active(tp, TT_PERSIST)) {
1523 tcp_timer_activate(tp, TT_PERSIST, 0);
1526 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1527 } else if (len == 0 && sbavail(&so->so_snd) &&
1528 !tcp_timer_active(tp, TT_REXMT) &&
1529 !tcp_timer_active(tp, TT_PERSIST)) {
1531 * Avoid a situation where we do not set persist timer
1532 * after a zero window condition. For example:
1533 * 1) A -> B: packet with enough data to fill the window
1534 * 2) B -> A: ACK for #1 + new data (0 window
1536 * 3) A -> B: ACK for #2, 0 len packet
1538 * In this case, A will not activate the persist timer,
1539 * because it chose to send a packet. Unless tcp_output
1540 * is called for some other reason (delayed ack timer,
1541 * another input packet from B, socket syscall), A will
1542 * not send zero window probes.
1544 * So, if you send a 0-length packet, but there is data
1545 * in the socket buffer, and neither the rexmt or
1546 * persist timer is already set, then activate the
1554 * Persist case, update snd_max but since we are in
1555 * persist mode (no window) we do not update snd_nxt.
1560 if (flags & TH_FIN) {
1562 tp->t_flags |= TF_SENTFIN;
1564 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1565 tp->snd_max = tp->snd_nxt + xlen;
1571 * We know that the packet was lost, so back out the
1572 * sequence number advance, if any.
1574 * If the error is EPERM the packet got blocked by the
1575 * local firewall. Normally we should terminate the
1576 * connection but the blocking may have been spurious
1577 * due to a firewall reconfiguration cycle. So we treat
1578 * it like a packet loss and let the retransmit timer and
1579 * timeouts do their work over time.
1580 * XXX: It is a POLA question whether calling tcp_drop right
1581 * away would be the really correct behavior instead.
1583 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1584 !tcp_timer_active(tp, TT_PERSIST)) &&
1585 ((flags & TH_SYN) == 0) &&
1589 tp->sackhint.sack_bytes_rexmit -= len;
1590 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1591 ("sackhint bytes rtx >= 0"));
1595 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1598 tp->t_softerror = error;
1601 tp->t_softerror = error;
1604 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1605 tp->snd_cwnd = tp->t_maxseg;
1609 * For some reason the interface we used initially
1610 * to send segments changed to another or lowered
1612 * If TSO was active we either got an interface
1613 * without TSO capabilits or TSO was turned off.
1614 * If we obtained mtu from ip_output() then update
1618 tp->t_flags &= ~TF_TSO;
1620 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1628 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1629 tp->t_softerror = error;
1637 TCPSTAT_INC(tcps_sndtotal);
1640 * Data sent (as far as we can tell).
1641 * If this advertises a larger window than any other segment,
1642 * then remember the size of the advertised window.
1643 * Any pending ACK has now been sent.
1645 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1646 tp->rcv_adv = tp->rcv_nxt + recwin;
1647 tp->last_ack_sent = tp->rcv_nxt;
1648 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1649 if (tcp_timer_active(tp, TT_DELACK))
1650 tcp_timer_activate(tp, TT_DELACK, 0);
1653 * This completely breaks TCP if newreno is turned on. What happens
1654 * is that if delayed-acks are turned on on the receiver, this code
1655 * on the transmitter effectively destroys the TCP window, forcing
1656 * it to four packets (1.5Kx4 = 6K window).
1658 if (sendalot && --maxburst)
1667 tcp_setpersist(struct tcpcb *tp)
1669 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1672 tp->t_flags &= ~TF_PREVVALID;
1673 if (tcp_timer_active(tp, TT_REXMT))
1674 panic("tcp_setpersist: retransmit pending");
1676 * Start/restart persistence timer.
1678 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1679 tcp_persmin, tcp_persmax);
1680 tcp_timer_activate(tp, TT_PERSIST, tt);
1681 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1686 * Insert TCP options according to the supplied parameters to the place
1687 * optp in a consistent way. Can handle unaligned destinations.
1689 * The order of the option processing is crucial for optimal packing and
1690 * alignment for the scarce option space.
1692 * The optimal order for a SYN/SYN-ACK segment is:
1693 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1694 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1696 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1697 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1698 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1699 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1700 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1703 tcp_addoptions(struct tcpopt *to, u_char *optp)
1705 u_int32_t mask, optlen = 0;
1707 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1708 if ((to->to_flags & mask) != mask)
1710 if (optlen == TCP_MAXOLEN)
1712 switch (to->to_flags & mask) {
1714 while (optlen % 4) {
1715 optlen += TCPOLEN_NOP;
1716 *optp++ = TCPOPT_NOP;
1718 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1720 optlen += TCPOLEN_MAXSEG;
1721 *optp++ = TCPOPT_MAXSEG;
1722 *optp++ = TCPOLEN_MAXSEG;
1723 to->to_mss = htons(to->to_mss);
1724 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1725 optp += sizeof(to->to_mss);
1728 while (!optlen || optlen % 2 != 1) {
1729 optlen += TCPOLEN_NOP;
1730 *optp++ = TCPOPT_NOP;
1732 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1734 optlen += TCPOLEN_WINDOW;
1735 *optp++ = TCPOPT_WINDOW;
1736 *optp++ = TCPOLEN_WINDOW;
1737 *optp++ = to->to_wscale;
1740 while (optlen % 2) {
1741 optlen += TCPOLEN_NOP;
1742 *optp++ = TCPOPT_NOP;
1744 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1746 optlen += TCPOLEN_SACK_PERMITTED;
1747 *optp++ = TCPOPT_SACK_PERMITTED;
1748 *optp++ = TCPOLEN_SACK_PERMITTED;
1751 while (!optlen || optlen % 4 != 2) {
1752 optlen += TCPOLEN_NOP;
1753 *optp++ = TCPOPT_NOP;
1755 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1757 optlen += TCPOLEN_TIMESTAMP;
1758 *optp++ = TCPOPT_TIMESTAMP;
1759 *optp++ = TCPOLEN_TIMESTAMP;
1760 to->to_tsval = htonl(to->to_tsval);
1761 to->to_tsecr = htonl(to->to_tsecr);
1762 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1763 optp += sizeof(to->to_tsval);
1764 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1765 optp += sizeof(to->to_tsecr);
1769 int siglen = TCPOLEN_SIGNATURE - 2;
1771 while (!optlen || optlen % 4 != 2) {
1772 optlen += TCPOLEN_NOP;
1773 *optp++ = TCPOPT_NOP;
1775 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1776 to->to_flags &= ~TOF_SIGNATURE;
1779 optlen += TCPOLEN_SIGNATURE;
1780 *optp++ = TCPOPT_SIGNATURE;
1781 *optp++ = TCPOLEN_SIGNATURE;
1782 to->to_signature = optp;
1790 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1793 while (!optlen || optlen % 4 != 2) {
1794 optlen += TCPOLEN_NOP;
1795 *optp++ = TCPOPT_NOP;
1797 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1799 optlen += TCPOLEN_SACKHDR;
1800 *optp++ = TCPOPT_SACK;
1801 sackblks = min(to->to_nsacks,
1802 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1803 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1804 while (sackblks--) {
1805 sack_seq = htonl(sack->start);
1806 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1807 optp += sizeof(sack_seq);
1808 sack_seq = htonl(sack->end);
1809 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1810 optp += sizeof(sack_seq);
1811 optlen += TCPOLEN_SACK;
1814 TCPSTAT_INC(tcps_sack_send_blocks);
1822 /* XXX is there any point to aligning this option? */
1823 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1824 if (TCP_MAXOLEN - optlen < total_len)
1826 *optp++ = TCPOPT_FAST_OPEN;
1827 *optp++ = total_len;
1828 if (to->to_tfo_len > 0) {
1829 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1830 optp += to->to_tfo_len;
1832 optlen += total_len;
1837 panic("%s: unknown TCP option type", __func__);
1842 /* Terminate and pad TCP options to a 4 byte boundary. */
1844 optlen += TCPOLEN_EOL;
1845 *optp++ = TCPOPT_EOL;
1848 * According to RFC 793 (STD0007):
1849 * "The content of the header beyond the End-of-Option option
1850 * must be header padding (i.e., zero)."
1851 * and later: "The padding is composed of zeros."
1853 while (optlen % 4) {
1854 optlen += TCPOLEN_PAD;
1855 *optp++ = TCPOPT_PAD;
1858 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));