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>
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_kern_tls.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
43 #include <sys/domain.h>
45 #include <sys/hhook.h>
47 #include <sys/kernel.h>
53 #include <sys/mutex.h>
54 #include <sys/protosw.h>
55 #include <sys/qmath.h>
57 #include <sys/socket.h>
58 #include <sys/socketvar.h>
59 #include <sys/sysctl.h>
60 #include <sys/stats.h>
63 #include <net/route.h>
64 #include <net/route/nhop.h>
67 #include <netinet/in.h>
68 #include <netinet/in_kdtrace.h>
69 #include <netinet/in_systm.h>
70 #include <netinet/ip.h>
71 #include <netinet/in_pcb.h>
72 #include <netinet/ip_var.h>
73 #include <netinet/ip_options.h>
75 #include <netinet6/in6_pcb.h>
76 #include <netinet/ip6.h>
77 #include <netinet6/ip6_var.h>
79 #include <netinet/tcp.h>
81 #include <netinet/tcp_fsm.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_var.h>
84 #include <netinet/tcp_log_buf.h>
85 #include <netinet/tcp_syncache.h>
86 #include <netinet/tcp_timer.h>
87 #include <netinet/tcpip.h>
88 #include <netinet/cc/cc.h>
89 #include <netinet/tcp_fastopen.h>
91 #include <netinet/tcp_pcap.h>
94 #include <netinet/tcp_offload.h>
96 #include <netinet/tcp_ecn.h>
98 #include <netipsec/ipsec_support.h>
100 #include <netinet/udp.h>
101 #include <netinet/udp_var.h>
102 #include <machine/in_cksum.h>
104 #include <security/mac/mac_framework.h>
106 VNET_DEFINE(int, path_mtu_discovery) = 1;
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
108 &VNET_NAME(path_mtu_discovery), 1,
109 "Enable Path MTU Discovery");
111 VNET_DEFINE(int, tcp_do_tso) = 1;
112 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
113 &VNET_NAME(tcp_do_tso), 0,
114 "Enable TCP Segmentation Offload");
116 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
117 #define V_tcp_sendspace VNET(tcp_sendspace)
118 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
119 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
121 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(tcp_do_autosndbuf), 0,
124 "Enable automatic send buffer sizing");
126 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
128 &VNET_NAME(tcp_autosndbuf_inc), 0,
129 "Incrementor step size of automatic send buffer");
131 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
133 &VNET_NAME(tcp_autosndbuf_max), 0,
134 "Max size of automatic send buffer");
136 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
137 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
138 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
139 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
140 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
143 * Make sure that either retransmit or persist timer is set for SYN, FIN and
146 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
147 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
148 tcp_timer_active((tp), TT_REXMT) || \
149 tcp_timer_active((tp), TT_PERSIST), \
150 ("neither rexmt nor persist timer is set"))
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(tptoinpcb(tp));
183 if (CC_ALGO(tp)->after_idle != NULL)
184 CC_ALGO(tp)->after_idle(&tp->t_ccv);
188 * Tcp output routine: figure out what should be sent and send it.
191 tcp_default_output(struct tcpcb *tp)
193 struct socket *so = tptosocket(tp);
194 struct inpcb *inp = tptoinpcb(tp);
196 uint32_t recwin, sendwin;
198 int off, error = 0; /* Keep compiler happy */
199 u_int if_hw_tsomaxsegcount = 0;
200 u_int if_hw_tsomaxsegsize = 0;
202 struct ip *ip = NULL;
204 u_char opt[TCP_MAXOLEN];
205 unsigned ipoptlen, optlen, hdrlen, ulen;
206 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
207 unsigned ipsec_optlen = 0;
209 int idle, sendalot, curticks;
210 int sack_rxmit, sack_bytes_rxmt;
214 struct udphdr *udp = NULL;
215 struct tcp_log_buffer *lgb;
216 unsigned int wanted_cookie = 0;
217 unsigned int dont_sendalot = 0;
219 int maxburst = TCP_MAXBURST;
222 struct ip6_hdr *ip6 = NULL;
223 const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
226 const bool hw_tls = tp->t_nic_ktls_xmit != 0;
228 const bool hw_tls = false;
232 INP_WLOCK_ASSERT(inp);
235 if (tp->t_flags & TF_TOE)
236 return (tcp_offload_output(tp));
240 * For TFO connections in SYN_SENT or SYN_RECEIVED,
241 * only allow the initial SYN or SYN|ACK and those sent
242 * by the retransmit timer.
244 if (IS_FASTOPEN(tp->t_flags) &&
245 ((tp->t_state == TCPS_SYN_SENT) ||
246 (tp->t_state == TCPS_SYN_RECEIVED)) &&
247 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
248 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
252 * Determine length of data that should be transmitted,
253 * and flags that will be used.
254 * If there is some data or critical controls (SYN, RST)
255 * to send, then transmit; otherwise, investigate further.
257 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
258 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
259 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
261 tp->t_flags &= ~TF_LASTIDLE;
263 if (tp->t_flags & TF_MORETOCOME) {
264 tp->t_flags |= TF_LASTIDLE;
270 * If we've recently taken a timeout, snd_max will be greater than
271 * snd_nxt. There may be SACK information that allows us to avoid
272 * resending already delivered data. Adjust snd_nxt accordingly.
274 if ((tp->t_flags & TF_SACK_PERMIT) &&
275 SEQ_LT(tp->snd_nxt, tp->snd_max))
280 off = tp->snd_nxt - tp->snd_una;
281 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
283 flags = tcp_outflags[tp->t_state];
285 * Send any SACK-generated retransmissions. If we're explicitly trying
286 * to send out new data (when sendalot is 1), bypass this function.
287 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
288 * we're replacing a (future) new transmission with a retransmission
289 * now, and we previously incremented snd_cwnd in tcp_input().
292 * Still in sack recovery , reset rxmit flag to zero.
298 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
299 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
303 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
304 /* Do not retransmit SACK segments beyond snd_recover */
305 if (SEQ_GT(p->end, tp->snd_recover)) {
307 * (At least) part of sack hole extends beyond
308 * snd_recover. Check to see if we can rexmit data
311 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
313 * Can't rexmit any more data for this hole.
314 * That data will be rexmitted in the next
315 * sack recovery episode, when snd_recover
316 * moves past p->rxmit.
319 goto after_sack_rexmit;
321 /* Can rexmit part of the current hole */
322 len = ((int32_t)ulmin(cwin,
323 SEQ_SUB(tp->snd_recover, p->rxmit)));
326 len = ((int32_t)ulmin(cwin,
327 SEQ_SUB(p->end, p->rxmit)));
330 off = SEQ_SUB(p->rxmit, tp->snd_una);
331 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
335 TCPSTAT_INC(tcps_sack_rexmits);
336 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
337 min(len, tcp_maxseg(tp)));
342 * Get standard flags, and add SYN or FIN if requested by 'hidden'
345 if (tp->t_flags & TF_NEEDFIN)
347 if (tp->t_flags & TF_NEEDSYN)
350 SOCKBUF_LOCK(&so->so_snd);
352 * If in persist timeout with window of 0, send 1 byte.
353 * Otherwise, if window is small but nonzero
354 * and timer expired, we will send what we can
355 * and go to transmit state.
357 if (tp->t_flags & TF_FORCEDATA) {
360 * If we still have some data to send, then
361 * clear the FIN bit. Usually this would
362 * happen below when it realizes that we
363 * aren't sending all the data. However,
364 * if we have exactly 1 byte of unsent data,
365 * then it won't clear the FIN bit below,
366 * and if we are in persist state, we wind
367 * up sending the packet without recording
368 * that we sent the FIN bit.
370 * We can't just blindly clear the FIN bit,
371 * because if we don't have any more data
372 * to send then the probe will be the FIN
375 if (off < sbused(&so->so_snd))
379 tcp_timer_activate(tp, TT_PERSIST, 0);
385 * If snd_nxt == snd_max and we have transmitted a FIN, the
386 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
387 * a negative length. This can also occur when TCP opens up
388 * its congestion window while receiving additional duplicate
389 * acks after fast-retransmit because TCP will reset snd_nxt
390 * to snd_max after the fast-retransmit.
392 * In the normal retransmit-FIN-only case, however, snd_nxt will
393 * be set to snd_una, the offset will be 0, and the length may
396 * If sack_rxmit is true we are retransmitting from the scoreboard
397 * in which case len is already set.
399 if (sack_rxmit == 0) {
400 if (sack_bytes_rxmt == 0)
401 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
407 * We are inside of a SACK recovery episode and are
408 * sending new data, having retransmitted all the
409 * data possible in the scoreboard.
411 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
414 * Don't remove this (len > 0) check !
415 * We explicitly check for len > 0 here (although it
416 * isn't really necessary), to work around a gcc
417 * optimization issue - to force gcc to compute
418 * len above. Without this check, the computation
419 * of len is bungled by the optimizer.
422 cwin = tp->snd_cwnd - imax(0, (int32_t)
423 (tp->snd_nxt - tp->snd_recover)) -
427 len = imin(len, cwin);
433 * Lop off SYN bit if it has already been sent. However, if this
434 * is SYN-SENT state and if segment contains data and if we don't
435 * know that foreign host supports TAO, suppress sending segment.
437 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
438 if (tp->t_state != TCPS_SYN_RECEIVED)
441 * When sending additional segments following a TFO SYN|ACK,
442 * do not include the SYN bit.
444 if (IS_FASTOPEN(tp->t_flags) &&
445 (tp->t_state == TCPS_SYN_RECEIVED))
451 * Be careful not to send data and/or FIN on SYN segments.
452 * This measure is needed to prevent interoperability problems
453 * with not fully conformant TCP implementations.
455 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
461 * On TFO sockets, ensure no data is sent in the following cases:
463 * - When retransmitting SYN|ACK on a passively-created socket
465 * - When retransmitting SYN on an actively created socket
467 * - When sending a zero-length cookie (cookie request) on an
468 * actively created socket
470 * - When the socket is in the CLOSED state (RST is being sent)
472 if (IS_FASTOPEN(tp->t_flags) &&
473 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
474 ((tp->t_state == TCPS_SYN_SENT) &&
475 (tp->t_tfo_client_cookie_len == 0)) ||
480 * If FIN has been sent but not acked,
481 * but we haven't been called to retransmit,
482 * len will be < 0. Otherwise, window shrank
483 * after we sent into it. If window shrank to 0,
484 * cancel pending retransmit, pull snd_nxt back
485 * to (closed) window, and set the persist timer
486 * if it isn't already going. If the window didn't
487 * close completely, just wait for an ACK.
489 * We also do a general check here to ensure that
490 * we will set the persist timer when we have data
491 * to send, but a 0-byte window. This makes sure
492 * the persist timer is set even if the packet
493 * hits one of the "goto send" lines below.
496 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
497 (off < (int) sbavail(&so->so_snd)) &&
498 !tcp_timer_active(tp, TT_PERSIST)) {
499 tcp_timer_activate(tp, TT_REXMT, 0);
501 tp->snd_nxt = tp->snd_una;
502 if (!tcp_timer_active(tp, TT_PERSIST))
507 /* len will be >= 0 after this point. */
508 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
510 tcp_sndbuf_autoscale(tp, so, sendwin);
513 * Decide if we can use TCP Segmentation Offloading (if supported by
516 * TSO may only be used if we are in a pure bulk sending state. The
517 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
518 * IP options prevent using TSO. With TSO the TCP header is the same
519 * (except for the sequence number) for all generated packets. This
520 * makes it impossible to transmit any options which vary per generated
523 * IPv4 handling has a clear separation of ip options and ip header
524 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
525 * the right thing below to provide length of just ip options and thus
526 * checking for ipoptlen is enough to decide if ip options are present.
528 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
530 * Pre-calculate here as we save another lookup into the darknesses
531 * of IPsec that way and can actually decide if TSO is ok.
534 if (isipv6 && IPSEC_ENABLED(ipv6))
535 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
541 if (IPSEC_ENABLED(ipv4))
542 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
547 ipoptlen = ip6_optlen(inp);
550 if (inp->inp_options)
551 ipoptlen = inp->inp_options->m_len -
552 offsetof(struct ipoption, ipopt_list);
555 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
556 ipoptlen += ipsec_optlen;
559 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
561 ((tp->t_flags & TF_SIGNATURE) == 0) &&
562 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
563 ipoptlen == 0 && !(flags & TH_SYN))
567 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
570 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
571 sbused(&so->so_snd)))
575 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
576 (long)TCP_MAXWIN << tp->rcv_scale);
579 * Sender silly window avoidance. We transmit under the following
580 * conditions when len is non-zero:
582 * - We have a full segment (or more with TSO)
583 * - This is the last buffer in a write()/send() and we are
584 * either idle or running NODELAY
585 * - we've timed out (e.g. persist timer)
586 * - we have more then 1/2 the maximum send window's worth of
587 * data (receiver may be limited the window size)
588 * - we need to retransmit
591 if (len >= tp->t_maxseg)
594 * As the TCP header options are now
595 * considered when setting up the initial
596 * window, we would not send the last segment
597 * if we skip considering the option length here.
598 * Note: this may not work when tcp headers change
599 * very dynamically in the future.
601 if ((((tp->t_flags & TF_SIGNATURE) ?
602 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
603 ((tp->t_flags & TF_RCVD_TSTMP) ?
604 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
605 len) >= tp->t_maxseg)
608 * NOTE! on localhost connections an 'ack' from the remote
609 * end may occur synchronously with the output and cause
610 * us to flush a buffer queued with moretocome. XXX
612 * note: the len + off check is almost certainly unnecessary.
614 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
615 (idle || (tp->t_flags & TF_NODELAY)) &&
616 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
617 (tp->t_flags & TF_NOPUSH) == 0) {
620 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
622 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
624 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
631 * Sending of standalone window updates.
633 * Window updates are important when we close our window due to a
634 * full socket buffer and are opening it again after the application
635 * reads data from it. Once the window has opened again and the
636 * remote end starts to send again the ACK clock takes over and
637 * provides the most current window information.
639 * We must avoid the silly window syndrome whereas every read
640 * from the receive buffer, no matter how small, causes a window
641 * update to be sent. We also should avoid sending a flurry of
642 * window updates when the socket buffer had queued a lot of data
643 * and the application is doing small reads.
645 * Prevent a flurry of pointless window updates by only sending
646 * an update when we can increase the advertized window by more
647 * than 1/4th of the socket buffer capacity. When the buffer is
648 * getting full or is very small be more aggressive and send an
649 * update whenever we can increase by two mss sized segments.
650 * In all other situations the ACK's to new incoming data will
651 * carry further window increases.
653 * Don't send an independent window update if a delayed
654 * ACK is pending (it will get piggy-backed on it) or the
655 * remote side already has done a half-close and won't send
656 * more data. Skip this if the connection is in T/TCP
659 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
660 !(tp->t_flags & TF_DELACK) &&
661 !TCPS_HAVERCVDFIN(tp->t_state)) {
663 * "adv" is the amount we could increase the window,
664 * taking into account that we are limited by
665 * TCP_MAXWIN << tp->rcv_scale.
671 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
672 oldwin = (tp->rcv_adv - tp->rcv_nxt);
681 * If the new window size ends up being the same as or less
682 * than the old size when it is scaled, then don't force
685 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
688 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
689 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
690 recwin <= (so->so_rcv.sb_hiwat / 8) ||
691 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
692 adv >= TCP_MAXWIN << tp->rcv_scale))
694 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
700 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
701 * is also a catch-all for the retransmit timer timeout case.
703 if (tp->t_flags & TF_ACKNOW)
705 if ((flags & TH_RST) ||
706 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
708 if (SEQ_GT(tp->snd_up, tp->snd_una))
711 * If our state indicates that FIN should be sent
712 * and we have not yet done so, then we need to send.
714 if (flags & TH_FIN &&
715 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
718 * In SACK, it is possible for tcp_output to fail to send a segment
719 * after the retransmission timer has been turned off. Make sure
720 * that the retransmission timer is set.
722 if ((tp->t_flags & TF_SACK_PERMIT) &&
723 SEQ_GT(tp->snd_max, tp->snd_una) &&
724 !tcp_timer_active(tp, TT_REXMT) &&
725 !tcp_timer_active(tp, TT_PERSIST)) {
726 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
730 * TCP window updates are not reliable, rather a polling protocol
731 * using ``persist'' packets is used to insure receipt of window
732 * updates. The three ``states'' for the output side are:
733 * idle not doing retransmits or persists
734 * persisting to move a small or zero window
735 * (re)transmitting and thereby not persisting
737 * tcp_timer_active(tp, TT_PERSIST)
738 * is true when we are in persist state.
739 * (tp->t_flags & TF_FORCEDATA)
740 * is set when we are called to send a persist packet.
741 * tcp_timer_active(tp, TT_REXMT)
742 * is set when we are retransmitting
743 * The output side is idle when both timers are zero.
745 * If send window is too small, there is data to transmit, and no
746 * retransmit or persist is pending, then go to persist state.
747 * If nothing happens soon, send when timer expires:
748 * if window is nonzero, transmit what we can,
749 * otherwise force out a byte.
751 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
752 !tcp_timer_active(tp, TT_PERSIST)) {
758 * No reason to send a segment, just return.
761 SOCKBUF_UNLOCK(&so->so_snd);
765 SOCKBUF_LOCK_ASSERT(&so->so_snd);
767 if (len >= tp->t_maxseg)
768 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
770 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
773 * Before ESTABLISHED, force sending of initial options
774 * unless TCP set not to do any options.
775 * NOTE: we assume that the IP/TCP header plus TCP options
776 * always fit in a single mbuf, leaving room for a maximum
778 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
783 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
786 hdrlen = sizeof (struct tcpiphdr);
788 if (flags & TH_SYN) {
789 tp->snd_nxt = tp->iss;
793 * Compute options for segment.
794 * We only have to care about SYN and established connection
795 * segments. Options for SYN-ACK segments are handled in TCP
799 if ((tp->t_flags & TF_NOOPT) == 0) {
800 /* Maximum segment size. */
801 if (flags & TH_SYN) {
802 to.to_mss = tcp_mssopt(&inp->inp_inc);
804 to.to_mss -= V_tcp_udp_tunneling_overhead;
805 to.to_flags |= TOF_MSS;
808 * On SYN or SYN|ACK transmits on TFO connections,
809 * only include the TFO option if it is not a
810 * retransmit, as the presence of the TFO option may
811 * have caused the original SYN or SYN|ACK to have
812 * been dropped by a middlebox.
814 if (IS_FASTOPEN(tp->t_flags) &&
815 (tp->t_rxtshift == 0)) {
816 if (tp->t_state == TCPS_SYN_RECEIVED) {
817 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
819 (u_int8_t *)&tp->t_tfo_cookie.server;
820 to.to_flags |= TOF_FASTOPEN;
822 } else if (tp->t_state == TCPS_SYN_SENT) {
824 tp->t_tfo_client_cookie_len;
826 tp->t_tfo_cookie.client;
827 to.to_flags |= TOF_FASTOPEN;
830 * If we wind up having more data to
831 * send with the SYN than can fit in
832 * one segment, don't send any more
833 * until the SYN|ACK comes back from
840 /* Window scaling. */
841 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
842 to.to_wscale = tp->request_r_scale;
843 to.to_flags |= TOF_SCALE;
846 if ((tp->t_flags & TF_RCVD_TSTMP) ||
847 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
848 curticks = tcp_ts_getticks();
849 to.to_tsval = curticks + tp->ts_offset;
850 to.to_tsecr = tp->ts_recent;
851 to.to_flags |= TOF_TS;
852 if (tp->t_rxtshift == 1)
853 tp->t_badrxtwin = curticks;
856 /* Set receive buffer autosizing timestamp. */
857 if (tp->rfbuf_ts == 0 &&
858 (so->so_rcv.sb_flags & SB_AUTOSIZE))
859 tp->rfbuf_ts = tcp_ts_getticks();
861 /* Selective ACK's. */
862 if (tp->t_flags & TF_SACK_PERMIT) {
864 to.to_flags |= TOF_SACKPERM;
865 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
866 tp->rcv_numsacks > 0) {
867 to.to_flags |= TOF_SACK;
868 to.to_nsacks = tp->rcv_numsacks;
869 to.to_sacks = (u_char *)tp->sackblks;
872 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
873 /* TCP-MD5 (RFC2385). */
875 * Check that TCP_MD5SIG is enabled in tcpcb to
876 * account the size needed to set this TCP option.
878 if (tp->t_flags & TF_SIGNATURE)
879 to.to_flags |= TOF_SIGNATURE;
880 #endif /* TCP_SIGNATURE */
882 /* Processing the options. */
883 hdrlen += optlen = tcp_addoptions(&to, opt);
885 * If we wanted a TFO option to be added, but it was unable
886 * to fit, ensure no data is sent.
888 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
889 !(to.to_flags & TOF_FASTOPEN))
893 if (V_tcp_udp_tunneling_port == 0) {
894 /* The port was removed?? */
895 SOCKBUF_UNLOCK(&so->so_snd);
896 return (EHOSTUNREACH);
898 hdrlen += sizeof(struct udphdr);
901 * Adjust data length if insertion of options will
902 * bump the packet length beyond the t_maxseg length.
903 * Clear the FIN bit because we cut off the tail of
906 if (len + optlen + ipoptlen > tp->t_maxseg) {
914 /* extract TSO information */
915 if_hw_tsomax = tp->t_tsomax;
916 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
917 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
920 * Limit a TSO burst to prevent it from
921 * overflowing or exceeding the maximum length
922 * allowed by the network interface:
924 KASSERT(ipoptlen == 0,
925 ("%s: TSO can't do IP options", __func__));
928 * Check if we should limit by maximum payload
931 if (if_hw_tsomax != 0) {
932 /* compute maximum TSO length */
933 max_len = (if_hw_tsomax - hdrlen -
937 } else if (len > max_len) {
944 * Prevent the last segment from being
945 * fractional unless the send sockbuf can be
948 max_len = (tp->t_maxseg - optlen);
949 if (((uint32_t)off + (uint32_t)len) <
950 sbavail(&so->so_snd)) {
951 moff = len % max_len;
959 * In case there are too many small fragments
962 if (len <= max_len) {
969 * Send the FIN in a separate segment
970 * after the bulk sending is done.
971 * We don't trust the TSO implementations
972 * to clear the FIN flag on all but the
975 if (tp->t_flags & TF_NEEDFIN)
978 if (optlen + ipoptlen >= tp->t_maxseg) {
980 * Since we don't have enough space to put
981 * the IP header chain and the TCP header in
982 * one packet as required by RFC 7112, don't
983 * send it. Also ensure that at least one
984 * byte of the payload can be put into the
987 SOCKBUF_UNLOCK(&so->so_snd);
992 len = tp->t_maxseg - optlen - ipoptlen;
1000 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1001 ("%s: len > IP_MAXPACKET", __func__));
1003 /*#ifdef DIAGNOSTIC*/
1005 if (max_linkhdr + hdrlen > MCLBYTES)
1007 if (max_linkhdr + hdrlen > MHLEN)
1009 panic("tcphdr too big");
1013 * This KASSERT is here to catch edge cases at a well defined place.
1014 * Before, those had triggered (random) panic conditions further down.
1016 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1019 * Grab a header mbuf, attaching a copy of data to
1020 * be transmitted, and initialize the header from
1021 * the template for sends on this connection.
1025 struct sockbuf *msb;
1028 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1029 TCPSTAT_INC(tcps_sndprobe);
1031 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1032 stats_voi_update_abs_u32(tp->t_stats,
1033 VOI_TCP_RETXPB, len);
1035 stats_voi_update_abs_u64(tp->t_stats,
1038 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1039 tp->t_sndrexmitpack++;
1040 TCPSTAT_INC(tcps_sndrexmitpack);
1041 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1043 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1047 TCPSTAT_INC(tcps_sndpack);
1048 TCPSTAT_ADD(tcps_sndbyte, len);
1050 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1055 if (MHLEN < hdrlen + max_linkhdr)
1056 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1059 m = m_gethdr(M_NOWAIT, MT_DATA);
1062 SOCKBUF_UNLOCK(&so->so_snd);
1068 m->m_data += max_linkhdr;
1072 * Start the m_copy functions from the closest mbuf
1073 * to the offset in the socket buffer chain.
1075 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1076 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1077 m_copydata(mb, moff, len,
1078 mtod(m, caddr_t) + hdrlen);
1079 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1080 sbsndptr_adv(&so->so_snd, mb, len);
1083 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1087 m->m_next = tcp_m_copym(mb, moff,
1088 &len, if_hw_tsomaxsegcount,
1089 if_hw_tsomaxsegsize, msb, hw_tls);
1090 if (len <= (tp->t_maxseg - optlen)) {
1092 * Must have ran out of mbufs for the copy
1093 * shorten it to no longer need tso. Lets
1094 * not put on sendalot since we are low on
1099 if (m->m_next == NULL) {
1100 SOCKBUF_UNLOCK(&so->so_snd);
1109 * If we're sending everything we've got, set PUSH.
1110 * (This will keep happy those implementations which only
1111 * give data to the user when a buffer fills or
1114 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1117 SOCKBUF_UNLOCK(&so->so_snd);
1119 SOCKBUF_UNLOCK(&so->so_snd);
1120 if (tp->t_flags & TF_ACKNOW)
1121 TCPSTAT_INC(tcps_sndacks);
1122 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1123 TCPSTAT_INC(tcps_sndctrl);
1124 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1125 TCPSTAT_INC(tcps_sndurg);
1127 TCPSTAT_INC(tcps_sndwinup);
1129 m = m_gethdr(M_NOWAIT, MT_DATA);
1136 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1141 m->m_data += max_linkhdr;
1144 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1145 m->m_pkthdr.rcvif = (struct ifnet *)0;
1147 mac_inpcb_create_mbuf(inp, m);
1151 ip6 = mtod(m, struct ip6_hdr *);
1153 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1154 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1155 udp->uh_dport = tp->t_port;
1156 ulen = hdrlen + len - sizeof(struct ip6_hdr);
1157 udp->uh_ulen = htons(ulen);
1158 th = (struct tcphdr *)(udp + 1);
1160 th = (struct tcphdr *)(ip6 + 1);
1162 tcpip_fillheaders(inp, tp->t_port, ip6, th);
1166 ip = mtod(m, struct ip *);
1168 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1169 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1170 udp->uh_dport = tp->t_port;
1171 ulen = hdrlen + len - sizeof(struct ip);
1172 udp->uh_ulen = htons(ulen);
1173 th = (struct tcphdr *)(udp + 1);
1175 th = (struct tcphdr *)(ip + 1);
1176 tcpip_fillheaders(inp, tp->t_port, ip, th);
1180 * Fill in fields, remembering maximum advertised
1181 * window for use in delaying messages about window sizes.
1182 * If resending a FIN, be sure not to use a new sequence number.
1184 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1185 tp->snd_nxt == tp->snd_max)
1188 * If we are starting a connection, send ECN setup
1189 * SYN packet. If we are on a retransmit, we may
1190 * resend those bits a number of times as per
1193 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1194 flags |= tcp_ecn_output_syn_sent(tp);
1196 /* Also handle parallel SYN for ECN */
1197 if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1198 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1199 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1200 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1201 (tp->t_flags2 & TF2_ECN_SND_ECE))
1202 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1205 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1206 ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1211 ip->ip_tos &= ~IPTOS_ECN_MASK;
1217 * If we are doing retransmissions, then snd_nxt will
1218 * not reflect the first unsent octet. For ACK only
1219 * packets, we do not want the sequence number of the
1220 * retransmitted packet, we want the sequence number
1221 * of the next unsent octet. So, if there is no data
1222 * (and no SYN or FIN), use snd_max instead of snd_nxt
1223 * when filling in ti_seq. But if we are in persist
1224 * state, snd_max might reflect one byte beyond the
1225 * right edge of the window, so use snd_nxt in that
1226 * case, since we know we aren't doing a retransmission.
1227 * (retransmit and persist are mutually exclusive...)
1229 if (sack_rxmit == 0) {
1230 if (len || (flags & (TH_SYN|TH_FIN)) ||
1231 tcp_timer_active(tp, TT_PERSIST))
1232 th->th_seq = htonl(tp->snd_nxt);
1234 th->th_seq = htonl(tp->snd_max);
1236 th->th_seq = htonl(p->rxmit);
1239 * Lost Retransmission Detection
1240 * trigger resending of a (then
1241 * still existing) hole, when
1242 * fack acks recoverypoint.
1244 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1245 p->rxmit = tp->snd_recover;
1246 tp->sackhint.sack_bytes_rexmit += len;
1248 if (IN_RECOVERY(tp->t_flags)) {
1250 * Account all bytes transmitted while
1251 * IN_RECOVERY, simplifying PRR and
1252 * Lost Retransmit Detection
1254 tp->sackhint.prr_out += len;
1256 th->th_ack = htonl(tp->rcv_nxt);
1258 bcopy(opt, th + 1, optlen);
1259 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1261 tcp_set_flags(th, flags);
1263 * Calculate receive window. Don't shrink window,
1264 * but avoid silly window syndrome.
1265 * If a RST segment is sent, advertise a window of zero.
1267 if (flags & TH_RST) {
1270 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1271 recwin < tp->t_maxseg)
1273 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1274 recwin < (tp->rcv_adv - tp->rcv_nxt))
1275 recwin = (tp->rcv_adv - tp->rcv_nxt);
1278 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1279 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1280 * case is handled in syncache.
1283 th->th_win = htons((u_short)
1284 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1286 /* Avoid shrinking window with window scaling. */
1287 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1288 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1292 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1293 * a 0 window. This may cause the remote transmitter to stall. This
1294 * flag tells soreceive() to disable delayed acknowledgements when
1295 * draining the buffer. This can occur if the receiver is attempting
1296 * to read more data than can be buffered prior to transmitting on
1299 if (th->th_win == 0) {
1301 tp->t_flags |= TF_RXWIN0SENT;
1303 tp->t_flags &= ~TF_RXWIN0SENT;
1304 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1305 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1306 th->th_flags |= TH_URG;
1309 * If no urgent pointer to send, then we pull
1310 * the urgent pointer to the left edge of the send window
1311 * so that it doesn't drift into the send window on sequence
1312 * number wraparound.
1314 tp->snd_up = tp->snd_una; /* drag it along */
1317 * Put TCP length in extended header, and then
1318 * checksum extended header and data.
1320 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1322 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1323 if (to.to_flags & TOF_SIGNATURE) {
1325 * Calculate MD5 signature and put it into the place
1326 * determined before.
1327 * NOTE: since TCP options buffer doesn't point into
1328 * mbuf's data, calculate offset and use it.
1330 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1331 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1333 * Do not send segment if the calculation of MD5
1334 * digest has failed.
1344 * There is no need to fill in ip6_plen right now.
1345 * It will be filled later by ip6_output.
1348 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1349 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1350 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1351 th->th_sum = htons(0);
1352 UDPSTAT_INC(udps_opackets);
1354 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1355 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1356 th->th_sum = in6_cksum_pseudo(ip6,
1357 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1362 #if defined(INET6) && defined(INET)
1368 m->m_pkthdr.csum_flags = CSUM_UDP;
1369 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1370 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1371 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1372 th->th_sum = htons(0);
1373 UDPSTAT_INC(udps_opackets);
1375 m->m_pkthdr.csum_flags = CSUM_TCP;
1376 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1377 th->th_sum = in_pseudo(ip->ip_src.s_addr,
1378 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1379 IPPROTO_TCP + len + optlen));
1382 /* IP version must be set here for ipv4/ipv6 checking later */
1383 KASSERT(ip->ip_v == IPVERSION,
1384 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1389 * Enable TSO and specify the size of the segments.
1390 * The TCP pseudo header checksum is always provided.
1393 KASSERT(len > tp->t_maxseg - optlen,
1394 ("%s: len <= tso_segsz", __func__));
1395 m->m_pkthdr.csum_flags |= CSUM_TSO;
1396 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1399 KASSERT(len + hdrlen == m_length(m, NULL),
1400 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1401 __func__, len, hdrlen, m_length(m, NULL)));
1404 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1405 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1408 TCP_PROBE3(debug__output, tp, th, m);
1410 /* We're getting ready to send; log now. */
1411 /* XXXMT: We are not honoring verbose logging. */
1413 if (tcp_bblogging_on(tp))
1414 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1415 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1421 * Fill in IP length and desired time to live and
1422 * send to IP level. There should be a better way
1423 * to handle ttl and tos; we could keep them in
1424 * the template, but need a way to checksum without them.
1427 * m->m_pkthdr.len should have been set before checksum calculation,
1428 * because in6_cksum() need it.
1433 * we separately set hoplimit for every segment, since the
1434 * user might want to change the value via setsockopt.
1435 * Also, desired default hop limit might be changed via
1436 * Neighbor Discovery.
1438 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1441 * Set the packet size here for the benefit of DTrace probes.
1442 * ip6_output() will set it properly; it's supposed to include
1443 * the option header lengths as well.
1445 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1447 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1448 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1450 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1452 if (tp->t_state == TCPS_SYN_SENT)
1453 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1455 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1458 /* Save packet, if requested. */
1459 tcp_pcap_add(th, m, &(tp->t_outpkts));
1462 /* TODO: IPv6 IP6TOS_ECT bit on */
1463 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1464 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1467 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1468 mtu = inp->inp_route6.ro_nh->nh_mtu;
1471 #if defined(INET) && defined(INET6)
1476 ip->ip_len = htons(m->m_pkthdr.len);
1478 if (inp->inp_vflag & INP_IPV6PROTO)
1479 ip->ip_ttl = in6_selecthlim(inp, NULL);
1482 * If we do path MTU discovery, then we set DF on every packet.
1483 * This might not be the best thing to do according to RFC3390
1484 * Section 2. However the tcp hostcache migitates the problem
1485 * so it affects only the first tcp connection with a host.
1487 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1489 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1490 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1491 if (tp->t_port == 0 || len < V_tcp_minmss) {
1492 ip->ip_off |= htons(IP_DF);
1495 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1498 if (tp->t_state == TCPS_SYN_SENT)
1499 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1501 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1504 /* Save packet, if requested. */
1505 tcp_pcap_add(th, m, &(tp->t_outpkts));
1508 error = ip_output(m, inp->inp_options, &inp->inp_route,
1509 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1511 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1512 mtu = inp->inp_route.ro_nh->nh_mtu;
1517 lgb->tlb_errno = error;
1522 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1524 * In transmit state, time the transmission and arrange for
1525 * the retransmit. In persist state, just set snd_max.
1527 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1528 !tcp_timer_active(tp, TT_PERSIST)) {
1529 tcp_seq startseq = tp->snd_nxt;
1532 * Advance snd_nxt over sequence space of this segment.
1534 if (flags & (TH_SYN|TH_FIN)) {
1537 if (flags & TH_FIN) {
1539 tp->t_flags |= TF_SENTFIN;
1545 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1547 * Update "made progress" indication if we just
1548 * added new data to an empty socket buffer.
1550 if (tp->snd_una == tp->snd_max)
1551 tp->t_acktime = ticks;
1552 tp->snd_max = tp->snd_nxt;
1554 * Time this transmission if not a retransmission and
1555 * not currently timing anything.
1557 tp->t_sndtime = ticks;
1558 if (tp->t_rtttime == 0) {
1559 tp->t_rtttime = ticks;
1560 tp->t_rtseq = startseq;
1561 TCPSTAT_INC(tcps_segstimed);
1564 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1565 tp->t_flags |= TF_GPUTINPROG;
1566 tp->gput_seq = startseq;
1567 tp->gput_ack = startseq +
1568 ulmin(sbavail(&so->so_snd) - off, sendwin);
1569 tp->gput_ts = tcp_ts_getticks();
1575 * Set retransmit timer if not currently set,
1576 * and not doing a pure ack or a keep-alive probe.
1577 * Initial value for retransmit timer is smoothed
1578 * round-trip time + 2 * round-trip time variance.
1579 * Initialize shift counter which is used for backoff
1580 * of retransmit time.
1583 if (!tcp_timer_active(tp, TT_REXMT) &&
1584 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1585 (tp->snd_nxt != tp->snd_una))) {
1586 if (tcp_timer_active(tp, TT_PERSIST)) {
1587 tcp_timer_activate(tp, TT_PERSIST, 0);
1590 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1591 } else if (len == 0 && sbavail(&so->so_snd) &&
1592 !tcp_timer_active(tp, TT_REXMT) &&
1593 !tcp_timer_active(tp, TT_PERSIST)) {
1595 * Avoid a situation where we do not set persist timer
1596 * after a zero window condition. For example:
1597 * 1) A -> B: packet with enough data to fill the window
1598 * 2) B -> A: ACK for #1 + new data (0 window
1600 * 3) A -> B: ACK for #2, 0 len packet
1602 * In this case, A will not activate the persist timer,
1603 * because it chose to send a packet. Unless tcp_output
1604 * is called for some other reason (delayed ack timer,
1605 * another input packet from B, socket syscall), A will
1606 * not send zero window probes.
1608 * So, if you send a 0-length packet, but there is data
1609 * in the socket buffer, and neither the rexmt or
1610 * persist timer is already set, then activate the
1618 * Persist case, update snd_max but since we are in
1619 * persist mode (no window) we do not update snd_nxt.
1624 if (flags & TH_FIN) {
1626 tp->t_flags |= TF_SENTFIN;
1628 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1629 tp->snd_max = tp->snd_nxt + xlen;
1632 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1633 (tp->t_flags & TF_SACK_PERMIT) &&
1634 tp->rcv_numsacks > 0)) {
1635 /* Clean up any DSACK's sent */
1636 tcp_clean_dsack_blocks(tp);
1640 * We know that the packet was lost, so back out the
1641 * sequence number advance, if any.
1643 * If the error is EPERM the packet got blocked by the
1644 * local firewall. Normally we should terminate the
1645 * connection but the blocking may have been spurious
1646 * due to a firewall reconfiguration cycle. So we treat
1647 * it like a packet loss and let the retransmit timer and
1648 * timeouts do their work over time.
1649 * XXX: It is a POLA question whether calling tcp_drop right
1650 * away would be the really correct behavior instead.
1652 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1653 !tcp_timer_active(tp, TT_PERSIST)) &&
1654 ((flags & TH_SYN) == 0) &&
1658 tp->sackhint.sack_bytes_rexmit -= len;
1659 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1660 ("sackhint bytes rtx >= 0"));
1661 KASSERT((flags & TH_FIN) == 0,
1662 ("error while FIN with SACK rxmit"));
1669 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1673 tp->t_softerror = error;
1676 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1677 tp->snd_cwnd = tp->t_maxseg;
1681 * For some reason the interface we used initially
1682 * to send segments changed to another or lowered
1684 * If TSO was active we either got an interface
1685 * without TSO capabilits or TSO was turned off.
1686 * If we obtained mtu from ip_output() then update
1690 tp->t_flags &= ~TF_TSO;
1692 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1700 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1701 tp->t_softerror = error;
1709 TCPSTAT_INC(tcps_sndtotal);
1712 * Data sent (as far as we can tell).
1713 * If this advertises a larger window than any other segment,
1714 * then remember the size of the advertised window.
1715 * Any pending ACK has now been sent.
1717 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1718 tp->rcv_adv = tp->rcv_nxt + recwin;
1719 tp->last_ack_sent = tp->rcv_nxt;
1720 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1721 if (tcp_timer_active(tp, TT_DELACK))
1722 tcp_timer_activate(tp, TT_DELACK, 0);
1725 * This completely breaks TCP if newreno is turned on. What happens
1726 * is that if delayed-acks are turned on on the receiver, this code
1727 * on the transmitter effectively destroys the TCP window, forcing
1728 * it to four packets (1.5Kx4 = 6K window).
1730 if (sendalot && --maxburst)
1739 tcp_setpersist(struct tcpcb *tp)
1741 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1745 tp->t_flags &= ~TF_PREVVALID;
1746 if (tcp_timer_active(tp, TT_REXMT))
1747 panic("tcp_setpersist: retransmit pending");
1749 * If the state is already closed, don't bother.
1751 if (tp->t_state == TCPS_CLOSED)
1755 * Start/restart persistence timer.
1757 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1758 tcp_persmin, tcp_persmax);
1759 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1760 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1761 if (maxunacktime < 1)
1763 if (maxunacktime < tt)
1766 tcp_timer_activate(tp, TT_PERSIST, tt);
1767 if (tp->t_rxtshift < V_tcp_retries)
1772 * Insert TCP options according to the supplied parameters to the place
1773 * optp in a consistent way. Can handle unaligned destinations.
1775 * The order of the option processing is crucial for optimal packing and
1776 * alignment for the scarce option space.
1778 * The optimal order for a SYN/SYN-ACK segment is:
1779 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1780 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1782 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1783 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1784 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1785 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1786 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1789 tcp_addoptions(struct tcpopt *to, u_char *optp)
1791 u_int32_t mask, optlen = 0;
1793 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1794 if ((to->to_flags & mask) != mask)
1796 if (optlen == TCP_MAXOLEN)
1798 switch (to->to_flags & mask) {
1800 while (optlen % 4) {
1801 optlen += TCPOLEN_NOP;
1802 *optp++ = TCPOPT_NOP;
1804 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1806 optlen += TCPOLEN_MAXSEG;
1807 *optp++ = TCPOPT_MAXSEG;
1808 *optp++ = TCPOLEN_MAXSEG;
1809 to->to_mss = htons(to->to_mss);
1810 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1811 optp += sizeof(to->to_mss);
1814 while (!optlen || optlen % 2 != 1) {
1815 optlen += TCPOLEN_NOP;
1816 *optp++ = TCPOPT_NOP;
1818 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1820 optlen += TCPOLEN_WINDOW;
1821 *optp++ = TCPOPT_WINDOW;
1822 *optp++ = TCPOLEN_WINDOW;
1823 *optp++ = to->to_wscale;
1826 while (optlen % 2) {
1827 optlen += TCPOLEN_NOP;
1828 *optp++ = TCPOPT_NOP;
1830 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1832 optlen += TCPOLEN_SACK_PERMITTED;
1833 *optp++ = TCPOPT_SACK_PERMITTED;
1834 *optp++ = TCPOLEN_SACK_PERMITTED;
1837 while (!optlen || optlen % 4 != 2) {
1838 optlen += TCPOLEN_NOP;
1839 *optp++ = TCPOPT_NOP;
1841 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1843 optlen += TCPOLEN_TIMESTAMP;
1844 *optp++ = TCPOPT_TIMESTAMP;
1845 *optp++ = TCPOLEN_TIMESTAMP;
1846 to->to_tsval = htonl(to->to_tsval);
1847 to->to_tsecr = htonl(to->to_tsecr);
1848 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1849 optp += sizeof(to->to_tsval);
1850 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1851 optp += sizeof(to->to_tsecr);
1855 int siglen = TCPOLEN_SIGNATURE - 2;
1857 while (!optlen || optlen % 4 != 2) {
1858 optlen += TCPOLEN_NOP;
1859 *optp++ = TCPOPT_NOP;
1861 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1862 to->to_flags &= ~TOF_SIGNATURE;
1865 optlen += TCPOLEN_SIGNATURE;
1866 *optp++ = TCPOPT_SIGNATURE;
1867 *optp++ = TCPOLEN_SIGNATURE;
1868 to->to_signature = optp;
1876 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1879 while (!optlen || optlen % 4 != 2) {
1880 optlen += TCPOLEN_NOP;
1881 *optp++ = TCPOPT_NOP;
1883 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1885 optlen += TCPOLEN_SACKHDR;
1886 *optp++ = TCPOPT_SACK;
1887 sackblks = min(to->to_nsacks,
1888 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1889 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1890 while (sackblks--) {
1891 sack_seq = htonl(sack->start);
1892 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1893 optp += sizeof(sack_seq);
1894 sack_seq = htonl(sack->end);
1895 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1896 optp += sizeof(sack_seq);
1897 optlen += TCPOLEN_SACK;
1900 TCPSTAT_INC(tcps_sack_send_blocks);
1907 /* XXX is there any point to aligning this option? */
1908 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1909 if (TCP_MAXOLEN - optlen < total_len) {
1910 to->to_flags &= ~TOF_FASTOPEN;
1913 *optp++ = TCPOPT_FAST_OPEN;
1914 *optp++ = total_len;
1915 if (to->to_tfo_len > 0) {
1916 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1917 optp += to->to_tfo_len;
1919 optlen += total_len;
1923 panic("%s: unknown TCP option type", __func__);
1928 /* Terminate and pad TCP options to a 4 byte boundary. */
1930 optlen += TCPOLEN_EOL;
1931 *optp++ = TCPOPT_EOL;
1934 * According to RFC 793 (STD0007):
1935 * "The content of the header beyond the End-of-Option option
1936 * must be header padding (i.e., zero)."
1937 * and later: "The padding is composed of zeros."
1939 while (optlen % 4) {
1940 optlen += TCPOLEN_PAD;
1941 *optp++ = TCPOPT_PAD;
1944 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1949 * This is a copy of m_copym(), taking the TSO segment size/limit
1950 * constraints into account, and advancing the sndptr as it goes.
1953 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1954 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1957 struct ktls_session *tls, *ntls;
1958 struct mbuf *start __diagused;
1960 struct mbuf *n, **np;
1963 int32_t len = *plen;
1967 uint32_t mlen, frags;
1970 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1971 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1972 if (off == 0 && m->m_flags & M_PKTHDR)
1977 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1981 if ((sb) && (m == sb->sb_sndptr)) {
1982 sb->sb_sndptroff += m->m_len;
1983 sb->sb_sndptr = m->m_next;
1991 if (hw_tls && (m->m_flags & M_EXTPG))
1999 KASSERT(len == M_COPYALL,
2000 ("tcp_m_copym, length > size of mbuf chain"));
2002 if (pkthdrlen != NULL)
2003 *pkthdrlen = len_cp;
2008 if (m->m_flags & M_EXTPG)
2009 ntls = m->m_epg_tls;
2014 * Avoid mixing TLS records with handshake
2015 * data or TLS records from different
2021 if (pkthdrlen != NULL)
2022 *pkthdrlen = len_cp;
2027 mlen = min(len, m->m_len - off);
2030 * For M_EXTPG mbufs, add 3 segments
2031 * + 1 in case we are crossing page boundaries
2032 * + 2 in case the TLS hdr/trailer are used
2033 * It is cheaper to just add the segments
2034 * than it is to take the cache miss to look
2035 * at the mbuf ext_pgs state in detail.
2037 if (m->m_flags & M_EXTPG) {
2038 fragsize = min(segsize, PAGE_SIZE);
2045 /* Break if we really can't fit anymore. */
2046 if ((frags + 1) >= seglimit) {
2048 if (pkthdrlen != NULL)
2049 *pkthdrlen = len_cp;
2054 * Reduce size if you can't copy the whole
2055 * mbuf. If we can't copy the whole mbuf, also
2056 * adjust len so the loop will end after this
2059 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2060 mlen = (seglimit - frags - 1) * fragsize;
2062 *plen = len_cp + len;
2063 if (pkthdrlen != NULL)
2066 frags += howmany(mlen, fragsize);
2070 KASSERT(seglimit > 0,
2071 ("%s: seglimit went too low", __func__));
2074 n = m_gethdr(M_NOWAIT, m->m_type);
2076 n = m_get(M_NOWAIT, m->m_type);
2081 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2083 if (len == M_COPYALL)
2084 n->m_pkthdr.len -= off0;
2086 n->m_pkthdr.len = len;
2087 pkthdrlen = &n->m_pkthdr.len;
2092 if (m->m_flags & (M_EXT|M_EXTPG)) {
2093 n->m_data = m->m_data + off;
2096 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2099 if (sb && (sb->sb_sndptr == m) &&
2100 ((n->m_len + off) >= m->m_len) && m->m_next) {
2101 sb->sb_sndptroff += m->m_len;
2102 sb->sb_sndptr = m->m_next;
2105 if (len != M_COPYALL) {
2118 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2122 * Automatic sizing of send socket buffer. Often the send buffer
2123 * size is not optimally adjusted to the actual network conditions
2124 * at hand (delay bandwidth product). Setting the buffer size too
2125 * small limits throughput on links with high bandwidth and high
2126 * delay (eg. trans-continental/oceanic links). Setting the
2127 * buffer size too big consumes too much real kernel memory,
2128 * especially with many connections on busy servers.
2130 * The criteria to step up the send buffer one notch are:
2131 * 1. receive window of remote host is larger than send buffer
2132 * (with a fudge factor of 5/4th);
2133 * 2. send buffer is filled to 7/8th with data (so we actually
2134 * have data to make use of it);
2135 * 3. send buffer fill has not hit maximal automatic size;
2136 * 4. our send window (slow start and cogestion controlled) is
2137 * larger than sent but unacknowledged data in send buffer.
2139 * The remote host receive window scaling factor may limit the
2140 * growing of the send buffer before it reaches its allowed
2143 * It scales directly with slow start or congestion window
2144 * and does at most one step per received ACK. This fast
2145 * scaling has the drawback of growing the send buffer beyond
2146 * what is strictly necessary to make full use of a given
2147 * delay*bandwidth product. However testing has shown this not
2148 * to be much of an problem. At worst we are trading wasting
2149 * of available bandwidth (the non-use of it) for wasting some
2150 * socket buffer memory.
2152 * TODO: Shrink send buffer during idle periods together
2153 * with congestion window. Requires another timer. Has to
2154 * wait for upcoming tcp timer rewrite.
2156 * XXXGL: should there be used sbused() or sbavail()?
2158 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2161 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2162 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2163 sbused(&so->so_snd) >=
2164 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2165 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2166 sendwin >= (sbused(&so->so_snd) -
2167 (tp->snd_nxt - tp->snd_una))) {
2168 if (!sbreserve_locked(so, SO_SND,
2169 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2170 V_tcp_autosndbuf_max), curthread))
2171 so->so_snd.sb_flags &= ~SB_AUTOSIZE;