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_kern_tls.h"
41 #include "opt_tcpdebug.h"
43 #include <sys/param.h>
44 #include <sys/systm.h>
46 #include <sys/domain.h>
48 #include <sys/hhook.h>
50 #include <sys/kernel.h>
56 #include <sys/mutex.h>
57 #include <sys/protosw.h>
58 #include <sys/qmath.h>
60 #include <sys/socket.h>
61 #include <sys/socketvar.h>
62 #include <sys/sysctl.h>
63 #include <sys/stats.h>
66 #include <net/route.h>
69 #include <netinet/in.h>
70 #include <netinet/in_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
77 #include <netinet6/in6_pcb.h>
78 #include <netinet/ip6.h>
79 #include <netinet6/ip6_var.h>
81 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_log_buf.h>
85 #include <netinet/tcp_seq.h>
86 #include <netinet/tcp_timer.h>
87 #include <netinet/tcp_var.h>
88 #include <netinet/tcpip.h>
89 #include <netinet/cc/cc.h>
90 #include <netinet/tcp_fastopen.h>
92 #include <netinet/tcp_pcap.h>
95 #include <netinet/tcp_debug.h>
98 #include <netinet/tcp_offload.h>
101 #include <netipsec/ipsec_support.h>
103 #include <machine/in_cksum.h>
105 #include <security/mac/mac_framework.h>
107 VNET_DEFINE(int, path_mtu_discovery) = 1;
108 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
109 &VNET_NAME(path_mtu_discovery), 1,
110 "Enable Path MTU Discovery");
112 VNET_DEFINE(int, tcp_do_tso) = 1;
113 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
114 &VNET_NAME(tcp_do_tso), 0,
115 "Enable TCP Segmentation Offload");
117 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
118 #define V_tcp_sendspace VNET(tcp_sendspace)
119 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
120 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
122 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
123 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
124 &VNET_NAME(tcp_do_autosndbuf), 0,
125 "Enable automatic send buffer sizing");
127 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
128 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
129 &VNET_NAME(tcp_autosndbuf_inc), 0,
130 "Incrementor step size of automatic send buffer");
132 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
134 &VNET_NAME(tcp_autosndbuf_max), 0,
135 "Max size of automatic send buffer");
137 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
138 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
139 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
140 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
141 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
144 * Make sure that either retransmit or persist timer is set for SYN, FIN and
147 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
148 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
149 tcp_timer_active((tp), TT_REXMT) || \
150 tcp_timer_active((tp), TT_PERSIST), \
151 ("neither rexmt nor persist timer is set"))
153 static void inline cc_after_idle(struct tcpcb *tp);
157 * Wrapper for the TCP established output helper hook.
160 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
161 struct tcpopt *to, uint32_t len, int tso)
163 struct tcp_hhook_data hhook_data;
165 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
169 hhook_data.len = len;
170 hhook_data.tso = tso;
172 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
179 * CC wrapper hook functions
182 cc_after_idle(struct tcpcb *tp)
184 INP_WLOCK_ASSERT(tp->t_inpcb);
186 if (CC_ALGO(tp)->after_idle != NULL)
187 CC_ALGO(tp)->after_idle(tp->ccv);
191 * Tcp output routine: figure out what should be sent and send it.
194 tcp_output(struct tcpcb *tp)
196 struct socket *so = tp->t_inpcb->inp_socket;
198 uint32_t recwin, sendwin;
199 int off, flags, error = 0; /* Keep compiler happy */
200 u_int if_hw_tsomaxsegcount = 0;
201 u_int if_hw_tsomaxsegsize = 0;
203 struct ip *ip = NULL;
205 struct ipovly *ipov = NULL;
208 u_char opt[TCP_MAXOLEN];
209 unsigned ipoptlen, optlen, hdrlen;
210 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
211 unsigned ipsec_optlen = 0;
213 int idle, sendalot, curticks;
214 int sack_rxmit, sack_bytes_rxmt;
218 unsigned int wanted_cookie = 0;
219 unsigned int dont_sendalot = 0;
221 int maxburst = TCP_MAXBURST;
224 struct ip6_hdr *ip6 = NULL;
227 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
230 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
232 const bool hw_tls = false;
236 INP_WLOCK_ASSERT(tp->t_inpcb);
239 if (tp->t_flags & TF_TOE)
240 return (tcp_offload_output(tp));
244 * For TFO connections in SYN_SENT or SYN_RECEIVED,
245 * only allow the initial SYN or SYN|ACK and those sent
246 * by the retransmit timer.
248 if (IS_FASTOPEN(tp->t_flags) &&
249 ((tp->t_state == TCPS_SYN_SENT) ||
250 (tp->t_state == TCPS_SYN_RECEIVED)) &&
251 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
252 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
256 * Determine length of data that should be transmitted,
257 * and flags that will be used.
258 * If there is some data or critical controls (SYN, RST)
259 * to send, then transmit; otherwise, investigate further.
261 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
262 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
264 tp->t_flags &= ~TF_LASTIDLE;
266 if (tp->t_flags & TF_MORETOCOME) {
267 tp->t_flags |= TF_LASTIDLE;
273 * If we've recently taken a timeout, snd_max will be greater than
274 * snd_nxt. There may be SACK information that allows us to avoid
275 * resending already delivered data. Adjust snd_nxt accordingly.
277 if ((tp->t_flags & TF_SACK_PERMIT) &&
278 SEQ_LT(tp->snd_nxt, tp->snd_max))
283 off = tp->snd_nxt - tp->snd_una;
284 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
286 flags = tcp_outflags[tp->t_state];
288 * Send any SACK-generated retransmissions. If we're explicitly trying
289 * to send out new data (when sendalot is 1), bypass this function.
290 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
291 * we're replacing a (future) new transmission with a retransmission
292 * now, and we previously incremented snd_cwnd in tcp_input().
295 * Still in sack recovery , reset rxmit flag to zero.
301 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
302 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
306 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
307 /* Do not retransmit SACK segments beyond snd_recover */
308 if (SEQ_GT(p->end, tp->snd_recover)) {
310 * (At least) part of sack hole extends beyond
311 * snd_recover. Check to see if we can rexmit data
314 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
316 * Can't rexmit any more data for this hole.
317 * That data will be rexmitted in the next
318 * sack recovery episode, when snd_recover
319 * moves past p->rxmit.
322 goto after_sack_rexmit;
324 /* Can rexmit part of the current hole */
325 len = ((int32_t)ulmin(cwin,
326 tp->snd_recover - p->rxmit));
328 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
329 off = p->rxmit - tp->snd_una;
330 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, tp->t_maxseg));
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 -
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_activate(tp, TT_REXMT, 0);
500 tp->snd_nxt = tp->snd_una;
501 if (!tcp_timer_active(tp, TT_PERSIST))
506 /* len will be >= 0 after this point. */
507 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
509 tcp_sndbuf_autoscale(tp, so, sendwin);
512 * Decide if we can use TCP Segmentation Offloading (if supported by
515 * TSO may only be used if we are in a pure bulk sending state. The
516 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
517 * IP options prevent using TSO. With TSO the TCP header is the same
518 * (except for the sequence number) for all generated packets. This
519 * makes it impossible to transmit any options which vary per generated
522 * IPv4 handling has a clear separation of ip options and ip header
523 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
524 * the right thing below to provide length of just ip options and thus
525 * checking for ipoptlen is enough to decide if ip options are present.
527 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
529 * Pre-calculate here as we save another lookup into the darknesses
530 * of IPsec that way and can actually decide if TSO is ok.
533 if (isipv6 && IPSEC_ENABLED(ipv6))
534 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
540 if (IPSEC_ENABLED(ipv4))
541 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
546 ipoptlen = ip6_optlen(tp->t_inpcb);
549 if (tp->t_inpcb->inp_options)
550 ipoptlen = tp->t_inpcb->inp_options->m_len -
551 offsetof(struct ipoption, ipopt_list);
554 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
555 ipoptlen += ipsec_optlen;
558 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
559 ((tp->t_flags & TF_SIGNATURE) == 0) &&
560 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
561 ipoptlen == 0 && !(flags & TH_SYN))
565 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
568 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
569 sbused(&so->so_snd)))
573 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
574 (long)TCP_MAXWIN << tp->rcv_scale);
577 * Sender silly window avoidance. We transmit under the following
578 * conditions when len is non-zero:
580 * - We have a full segment (or more with TSO)
581 * - This is the last buffer in a write()/send() and we are
582 * either idle or running NODELAY
583 * - we've timed out (e.g. persist timer)
584 * - we have more then 1/2 the maximum send window's worth of
585 * data (receiver may be limited the window size)
586 * - we need to retransmit
589 if (len >= tp->t_maxseg)
592 * NOTE! on localhost connections an 'ack' from the remote
593 * end may occur synchronously with the output and cause
594 * us to flush a buffer queued with moretocome. XXX
596 * note: the len + off check is almost certainly unnecessary.
598 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
599 (idle || (tp->t_flags & TF_NODELAY)) &&
600 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
601 (tp->t_flags & TF_NOPUSH) == 0) {
604 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
606 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
608 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
615 * Sending of standalone window updates.
617 * Window updates are important when we close our window due to a
618 * full socket buffer and are opening it again after the application
619 * reads data from it. Once the window has opened again and the
620 * remote end starts to send again the ACK clock takes over and
621 * provides the most current window information.
623 * We must avoid the silly window syndrome whereas every read
624 * from the receive buffer, no matter how small, causes a window
625 * update to be sent. We also should avoid sending a flurry of
626 * window updates when the socket buffer had queued a lot of data
627 * and the application is doing small reads.
629 * Prevent a flurry of pointless window updates by only sending
630 * an update when we can increase the advertized window by more
631 * than 1/4th of the socket buffer capacity. When the buffer is
632 * getting full or is very small be more aggressive and send an
633 * update whenever we can increase by two mss sized segments.
634 * In all other situations the ACK's to new incoming data will
635 * carry further window increases.
637 * Don't send an independent window update if a delayed
638 * ACK is pending (it will get piggy-backed on it) or the
639 * remote side already has done a half-close and won't send
640 * more data. Skip this if the connection is in T/TCP
643 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
644 !(tp->t_flags & TF_DELACK) &&
645 !TCPS_HAVERCVDFIN(tp->t_state)) {
647 * "adv" is the amount we could increase the window,
648 * taking into account that we are limited by
649 * TCP_MAXWIN << tp->rcv_scale.
655 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
656 oldwin = (tp->rcv_adv - tp->rcv_nxt);
662 * If the new window size ends up being the same as or less
663 * than the old size when it is scaled, then don't force
666 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
669 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
670 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
671 recwin <= (so->so_rcv.sb_hiwat / 8) ||
672 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
673 adv >= TCP_MAXWIN << tp->rcv_scale))
675 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
681 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
682 * is also a catch-all for the retransmit timer timeout case.
684 if (tp->t_flags & TF_ACKNOW)
686 if ((flags & TH_RST) ||
687 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
689 if (SEQ_GT(tp->snd_up, tp->snd_una))
692 * If our state indicates that FIN should be sent
693 * and we have not yet done so, then we need to send.
695 if (flags & TH_FIN &&
696 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
699 * In SACK, it is possible for tcp_output to fail to send a segment
700 * after the retransmission timer has been turned off. Make sure
701 * that the retransmission timer is set.
703 if ((tp->t_flags & TF_SACK_PERMIT) &&
704 SEQ_GT(tp->snd_max, tp->snd_una) &&
705 !tcp_timer_active(tp, TT_REXMT) &&
706 !tcp_timer_active(tp, TT_PERSIST)) {
707 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
711 * TCP window updates are not reliable, rather a polling protocol
712 * using ``persist'' packets is used to insure receipt of window
713 * updates. The three ``states'' for the output side are:
714 * idle not doing retransmits or persists
715 * persisting to move a small or zero window
716 * (re)transmitting and thereby not persisting
718 * tcp_timer_active(tp, TT_PERSIST)
719 * is true when we are in persist state.
720 * (tp->t_flags & TF_FORCEDATA)
721 * is set when we are called to send a persist packet.
722 * tcp_timer_active(tp, TT_REXMT)
723 * is set when we are retransmitting
724 * The output side is idle when both timers are zero.
726 * If send window is too small, there is data to transmit, and no
727 * retransmit or persist is pending, then go to persist state.
728 * If nothing happens soon, send when timer expires:
729 * if window is nonzero, transmit what we can,
730 * otherwise force out a byte.
732 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
733 !tcp_timer_active(tp, TT_PERSIST)) {
739 * No reason to send a segment, just return.
742 SOCKBUF_UNLOCK(&so->so_snd);
746 SOCKBUF_LOCK_ASSERT(&so->so_snd);
748 if (len >= tp->t_maxseg)
749 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
751 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
754 * Before ESTABLISHED, force sending of initial options
755 * unless TCP set not to do any options.
756 * NOTE: we assume that the IP/TCP header plus TCP options
757 * always fit in a single mbuf, leaving room for a maximum
759 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
764 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
767 hdrlen = sizeof (struct tcpiphdr);
770 * Compute options for segment.
771 * We only have to care about SYN and established connection
772 * segments. Options for SYN-ACK segments are handled in TCP
776 if ((tp->t_flags & TF_NOOPT) == 0) {
777 /* Maximum segment size. */
778 if (flags & TH_SYN) {
779 tp->snd_nxt = tp->iss;
780 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
781 to.to_flags |= TOF_MSS;
784 * On SYN or SYN|ACK transmits on TFO connections,
785 * only include the TFO option if it is not a
786 * retransmit, as the presence of the TFO option may
787 * have caused the original SYN or SYN|ACK to have
788 * been dropped by a middlebox.
790 if (IS_FASTOPEN(tp->t_flags) &&
791 (tp->t_rxtshift == 0)) {
792 if (tp->t_state == TCPS_SYN_RECEIVED) {
793 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
795 (u_int8_t *)&tp->t_tfo_cookie.server;
796 to.to_flags |= TOF_FASTOPEN;
798 } else if (tp->t_state == TCPS_SYN_SENT) {
800 tp->t_tfo_client_cookie_len;
802 tp->t_tfo_cookie.client;
803 to.to_flags |= TOF_FASTOPEN;
806 * If we wind up having more data to
807 * send with the SYN than can fit in
808 * one segment, don't send any more
809 * until the SYN|ACK comes back from
816 /* Window scaling. */
817 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
818 to.to_wscale = tp->request_r_scale;
819 to.to_flags |= TOF_SCALE;
822 if ((tp->t_flags & TF_RCVD_TSTMP) ||
823 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
824 curticks = tcp_ts_getticks();
825 to.to_tsval = curticks + tp->ts_offset;
826 to.to_tsecr = tp->ts_recent;
827 to.to_flags |= TOF_TS;
828 if (tp->t_rxtshift == 1)
829 tp->t_badrxtwin = curticks;
832 /* Set receive buffer autosizing timestamp. */
833 if (tp->rfbuf_ts == 0 &&
834 (so->so_rcv.sb_flags & SB_AUTOSIZE))
835 tp->rfbuf_ts = tcp_ts_getticks();
837 /* Selective ACK's. */
838 if (tp->t_flags & TF_SACK_PERMIT) {
840 to.to_flags |= TOF_SACKPERM;
841 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
842 (tp->t_flags & TF_SACK_PERMIT) &&
843 tp->rcv_numsacks > 0) {
844 to.to_flags |= TOF_SACK;
845 to.to_nsacks = tp->rcv_numsacks;
846 to.to_sacks = (u_char *)tp->sackblks;
849 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
850 /* TCP-MD5 (RFC2385). */
852 * Check that TCP_MD5SIG is enabled in tcpcb to
853 * account the size needed to set this TCP option.
855 if (tp->t_flags & TF_SIGNATURE)
856 to.to_flags |= TOF_SIGNATURE;
857 #endif /* TCP_SIGNATURE */
859 /* Processing the options. */
860 hdrlen += optlen = tcp_addoptions(&to, opt);
862 * If we wanted a TFO option to be added, but it was unable
863 * to fit, ensure no data is sent.
865 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
866 !(to.to_flags & TOF_FASTOPEN))
871 * Adjust data length if insertion of options will
872 * bump the packet length beyond the t_maxseg length.
873 * Clear the FIN bit because we cut off the tail of
876 if (len + optlen + ipoptlen > tp->t_maxseg) {
884 /* extract TSO information */
885 if_hw_tsomax = tp->t_tsomax;
886 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
887 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
890 * Limit a TSO burst to prevent it from
891 * overflowing or exceeding the maximum length
892 * allowed by the network interface:
894 KASSERT(ipoptlen == 0,
895 ("%s: TSO can't do IP options", __func__));
898 * Check if we should limit by maximum payload
901 if (if_hw_tsomax != 0) {
902 /* compute maximum TSO length */
903 max_len = (if_hw_tsomax - hdrlen -
907 } else if (len > max_len) {
914 * Prevent the last segment from being
915 * fractional unless the send sockbuf can be
918 max_len = (tp->t_maxseg - optlen);
919 if (((uint32_t)off + (uint32_t)len) <
920 sbavail(&so->so_snd)) {
921 moff = len % max_len;
929 * In case there are too many small fragments
932 if (len <= max_len) {
939 * Send the FIN in a separate segment
940 * after the bulk sending is done.
941 * We don't trust the TSO implementations
942 * to clear the FIN flag on all but the
945 if (tp->t_flags & TF_NEEDFIN)
948 if (optlen + ipoptlen >= tp->t_maxseg) {
950 * Since we don't have enough space to put
951 * the IP header chain and the TCP header in
952 * one packet as required by RFC 7112, don't
953 * send it. Also ensure that at least one
954 * byte of the payload can be put into the
957 SOCKBUF_UNLOCK(&so->so_snd);
962 len = tp->t_maxseg - optlen - ipoptlen;
970 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
971 ("%s: len > IP_MAXPACKET", __func__));
973 /*#ifdef DIAGNOSTIC*/
975 if (max_linkhdr + hdrlen > MCLBYTES)
977 if (max_linkhdr + hdrlen > MHLEN)
979 panic("tcphdr too big");
983 * This KASSERT is here to catch edge cases at a well defined place.
984 * Before, those had triggered (random) panic conditions further down.
986 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
989 * Grab a header mbuf, attaching a copy of data to
990 * be transmitted, and initialize the header from
991 * the template for sends on this connection.
998 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
999 TCPSTAT_INC(tcps_sndprobe);
1001 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1002 stats_voi_update_abs_u32(tp->t_stats,
1003 VOI_TCP_RETXPB, len);
1005 stats_voi_update_abs_u64(tp->t_stats,
1008 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1009 tp->t_sndrexmitpack++;
1010 TCPSTAT_INC(tcps_sndrexmitpack);
1011 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1013 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1017 TCPSTAT_INC(tcps_sndpack);
1018 TCPSTAT_ADD(tcps_sndbyte, len);
1020 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1025 if (MHLEN < hdrlen + max_linkhdr)
1026 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1029 m = m_gethdr(M_NOWAIT, MT_DATA);
1032 SOCKBUF_UNLOCK(&so->so_snd);
1038 m->m_data += max_linkhdr;
1042 * Start the m_copy functions from the closest mbuf
1043 * to the offset in the socket buffer chain.
1045 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1046 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1047 m_copydata(mb, moff, len,
1048 mtod(m, caddr_t) + hdrlen);
1049 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1050 sbsndptr_adv(&so->so_snd, mb, len);
1053 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1057 m->m_next = tcp_m_copym(mb, moff,
1058 &len, if_hw_tsomaxsegcount,
1059 if_hw_tsomaxsegsize, msb, hw_tls);
1060 if (len <= (tp->t_maxseg - optlen)) {
1062 * Must have ran out of mbufs for the copy
1063 * shorten it to no longer need tso. Lets
1064 * not put on sendalot since we are low on
1069 if (m->m_next == NULL) {
1070 SOCKBUF_UNLOCK(&so->so_snd);
1079 * If we're sending everything we've got, set PUSH.
1080 * (This will keep happy those implementations which only
1081 * give data to the user when a buffer fills or
1084 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1087 SOCKBUF_UNLOCK(&so->so_snd);
1089 SOCKBUF_UNLOCK(&so->so_snd);
1090 if (tp->t_flags & TF_ACKNOW)
1091 TCPSTAT_INC(tcps_sndacks);
1092 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1093 TCPSTAT_INC(tcps_sndctrl);
1094 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1095 TCPSTAT_INC(tcps_sndurg);
1097 TCPSTAT_INC(tcps_sndwinup);
1099 m = m_gethdr(M_NOWAIT, MT_DATA);
1106 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1111 m->m_data += max_linkhdr;
1114 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1115 m->m_pkthdr.rcvif = (struct ifnet *)0;
1117 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1121 ip6 = mtod(m, struct ip6_hdr *);
1122 th = (struct tcphdr *)(ip6 + 1);
1123 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1127 ip = mtod(m, struct ip *);
1129 ipov = (struct ipovly *)ip;
1131 th = (struct tcphdr *)(ip + 1);
1132 tcpip_fillheaders(tp->t_inpcb, ip, th);
1136 * Fill in fields, remembering maximum advertised
1137 * window for use in delaying messages about window sizes.
1138 * If resending a FIN, be sure not to use a new sequence number.
1140 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1141 tp->snd_nxt == tp->snd_max)
1144 * If we are starting a connection, send ECN setup
1145 * SYN packet. If we are on a retransmit, we may
1146 * resend those bits a number of times as per
1149 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1150 if (tp->t_rxtshift >= 1) {
1151 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1152 flags |= TH_ECE|TH_CWR;
1154 flags |= TH_ECE|TH_CWR;
1157 if (tp->t_state == TCPS_ESTABLISHED &&
1158 (tp->t_flags2 & TF2_ECN_PERMIT)) {
1160 * If the peer has ECN, mark data packets with
1161 * ECN capable transmission (ECT).
1162 * Ignore pure ack packets, retransmissions and window probes.
1164 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1165 (sack_rxmit == 0) &&
1166 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1169 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1172 ip->ip_tos |= IPTOS_ECN_ECT0;
1173 TCPSTAT_INC(tcps_ecn_ect0);
1177 * Reply with proper ECN notifications.
1179 if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1181 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1183 if (tp->t_flags2 & TF2_ECN_SND_ECE)
1188 * If we are doing retransmissions, then snd_nxt will
1189 * not reflect the first unsent octet. For ACK only
1190 * packets, we do not want the sequence number of the
1191 * retransmitted packet, we want the sequence number
1192 * of the next unsent octet. So, if there is no data
1193 * (and no SYN or FIN), use snd_max instead of snd_nxt
1194 * when filling in ti_seq. But if we are in persist
1195 * state, snd_max might reflect one byte beyond the
1196 * right edge of the window, so use snd_nxt in that
1197 * case, since we know we aren't doing a retransmission.
1198 * (retransmit and persist are mutually exclusive...)
1200 if (sack_rxmit == 0) {
1201 if (len || (flags & (TH_SYN|TH_FIN)) ||
1202 tcp_timer_active(tp, TT_PERSIST))
1203 th->th_seq = htonl(tp->snd_nxt);
1205 th->th_seq = htonl(tp->snd_max);
1207 th->th_seq = htonl(p->rxmit);
1209 tp->sackhint.sack_bytes_rexmit += len;
1211 th->th_ack = htonl(tp->rcv_nxt);
1213 bcopy(opt, th + 1, optlen);
1214 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1216 th->th_flags = flags;
1218 * Calculate receive window. Don't shrink window,
1219 * but avoid silly window syndrome.
1220 * If a RST segment is sent, advertise a window of zero.
1222 if (flags & TH_RST) {
1225 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1226 recwin < tp->t_maxseg)
1228 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1229 recwin < (tp->rcv_adv - tp->rcv_nxt))
1230 recwin = (tp->rcv_adv - tp->rcv_nxt);
1233 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1234 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1235 * case is handled in syncache.
1238 th->th_win = htons((u_short)
1239 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1241 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1244 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1245 * a 0 window. This may cause the remote transmitter to stall. This
1246 * flag tells soreceive() to disable delayed acknowledgements when
1247 * draining the buffer. This can occur if the receiver is attempting
1248 * to read more data than can be buffered prior to transmitting on
1251 if (th->th_win == 0) {
1253 tp->t_flags |= TF_RXWIN0SENT;
1255 tp->t_flags &= ~TF_RXWIN0SENT;
1256 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1257 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1258 th->th_flags |= TH_URG;
1261 * If no urgent pointer to send, then we pull
1262 * the urgent pointer to the left edge of the send window
1263 * so that it doesn't drift into the send window on sequence
1264 * number wraparound.
1266 tp->snd_up = tp->snd_una; /* drag it along */
1269 * Put TCP length in extended header, and then
1270 * checksum extended header and data.
1272 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1273 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1275 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1276 if (to.to_flags & TOF_SIGNATURE) {
1278 * Calculate MD5 signature and put it into the place
1279 * determined before.
1280 * NOTE: since TCP options buffer doesn't point into
1281 * mbuf's data, calculate offset and use it.
1283 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1284 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1286 * Do not send segment if the calculation of MD5
1287 * digest has failed.
1297 * There is no need to fill in ip6_plen right now.
1298 * It will be filled later by ip6_output.
1300 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1301 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1302 optlen + len, IPPROTO_TCP, 0);
1305 #if defined(INET6) && defined(INET)
1310 m->m_pkthdr.csum_flags = CSUM_TCP;
1311 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1312 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1314 /* IP version must be set here for ipv4/ipv6 checking later */
1315 KASSERT(ip->ip_v == IPVERSION,
1316 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1321 * Enable TSO and specify the size of the segments.
1322 * The TCP pseudo header checksum is always provided.
1325 KASSERT(len > tp->t_maxseg - optlen,
1326 ("%s: len <= tso_segsz", __func__));
1327 m->m_pkthdr.csum_flags |= CSUM_TSO;
1328 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1331 KASSERT(len + hdrlen == m_length(m, NULL),
1332 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1333 __func__, len, hdrlen, m_length(m, NULL)));
1336 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1337 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1344 if (so->so_options & SO_DEBUG) {
1350 save = ipov->ih_len;
1351 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1353 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1357 ipov->ih_len = save;
1359 #endif /* TCPDEBUG */
1360 TCP_PROBE3(debug__output, tp, th, m);
1362 /* We're getting ready to send; log now. */
1363 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1367 * Fill in IP length and desired time to live and
1368 * send to IP level. There should be a better way
1369 * to handle ttl and tos; we could keep them in
1370 * the template, but need a way to checksum without them.
1373 * m->m_pkthdr.len should have been set before checksum calculation,
1374 * because in6_cksum() need it.
1379 * we separately set hoplimit for every segment, since the
1380 * user might want to change the value via setsockopt.
1381 * Also, desired default hop limit might be changed via
1382 * Neighbor Discovery.
1384 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1387 * Set the packet size here for the benefit of DTrace probes.
1388 * ip6_output() will set it properly; it's supposed to include
1389 * the option header lengths as well.
1391 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1393 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1394 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1396 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1398 if (tp->t_state == TCPS_SYN_SENT)
1399 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1401 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1404 /* Save packet, if requested. */
1405 tcp_pcap_add(th, m, &(tp->t_outpkts));
1408 /* TODO: IPv6 IP6TOS_ECT bit on */
1409 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1410 &tp->t_inpcb->inp_route6,
1411 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1412 NULL, NULL, tp->t_inpcb);
1414 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1415 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1418 #if defined(INET) && defined(INET6)
1423 ip->ip_len = htons(m->m_pkthdr.len);
1425 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1426 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1429 * If we do path MTU discovery, then we set DF on every packet.
1430 * This might not be the best thing to do according to RFC3390
1431 * Section 2. However the tcp hostcache migitates the problem
1432 * so it affects only the first tcp connection with a host.
1434 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1436 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1437 ip->ip_off |= htons(IP_DF);
1438 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1440 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1443 if (tp->t_state == TCPS_SYN_SENT)
1444 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1446 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1449 /* Save packet, if requested. */
1450 tcp_pcap_add(th, m, &(tp->t_outpkts));
1453 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1454 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1457 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1458 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1464 * In transmit state, time the transmission and arrange for
1465 * the retransmit. In persist state, just set snd_max.
1467 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1468 !tcp_timer_active(tp, TT_PERSIST)) {
1469 tcp_seq startseq = tp->snd_nxt;
1472 * Advance snd_nxt over sequence space of this segment.
1474 if (flags & (TH_SYN|TH_FIN)) {
1477 if (flags & TH_FIN) {
1479 tp->t_flags |= TF_SENTFIN;
1485 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1486 tp->snd_max = tp->snd_nxt;
1488 * Time this transmission if not a retransmission and
1489 * not currently timing anything.
1491 if (tp->t_rtttime == 0) {
1492 tp->t_rtttime = ticks;
1493 tp->t_rtseq = startseq;
1494 TCPSTAT_INC(tcps_segstimed);
1497 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1498 tp->t_flags |= TF_GPUTINPROG;
1499 tp->gput_seq = startseq;
1500 tp->gput_ack = startseq +
1501 ulmin(sbavail(&so->so_snd) - off, sendwin);
1502 tp->gput_ts = tcp_ts_getticks();
1508 * Set retransmit timer if not currently set,
1509 * and not doing a pure ack or a keep-alive probe.
1510 * Initial value for retransmit timer is smoothed
1511 * round-trip time + 2 * round-trip time variance.
1512 * Initialize shift counter which is used for backoff
1513 * of retransmit time.
1516 if (!tcp_timer_active(tp, TT_REXMT) &&
1517 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1518 (tp->snd_nxt != tp->snd_una))) {
1519 if (tcp_timer_active(tp, TT_PERSIST)) {
1520 tcp_timer_activate(tp, TT_PERSIST, 0);
1523 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1524 } else if (len == 0 && sbavail(&so->so_snd) &&
1525 !tcp_timer_active(tp, TT_REXMT) &&
1526 !tcp_timer_active(tp, TT_PERSIST)) {
1528 * Avoid a situation where we do not set persist timer
1529 * after a zero window condition. For example:
1530 * 1) A -> B: packet with enough data to fill the window
1531 * 2) B -> A: ACK for #1 + new data (0 window
1533 * 3) A -> B: ACK for #2, 0 len packet
1535 * In this case, A will not activate the persist timer,
1536 * because it chose to send a packet. Unless tcp_output
1537 * is called for some other reason (delayed ack timer,
1538 * another input packet from B, socket syscall), A will
1539 * not send zero window probes.
1541 * So, if you send a 0-length packet, but there is data
1542 * in the socket buffer, and neither the rexmt or
1543 * persist timer is already set, then activate the
1551 * Persist case, update snd_max but since we are in
1552 * persist mode (no window) we do not update snd_nxt.
1557 if (flags & TH_FIN) {
1559 tp->t_flags |= TF_SENTFIN;
1561 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1562 tp->snd_max = tp->snd_nxt + xlen;
1565 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1566 (tp->t_flags & TF_SACK_PERMIT) &&
1567 tp->rcv_numsacks > 0)) {
1568 /* Clean up any DSACK's sent */
1569 tcp_clean_dsack_blocks(tp);
1572 /* Record the error. */
1573 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1574 error, 0, NULL, false);
1577 * We know that the packet was lost, so back out the
1578 * sequence number advance, if any.
1580 * If the error is EPERM the packet got blocked by the
1581 * local firewall. Normally we should terminate the
1582 * connection but the blocking may have been spurious
1583 * due to a firewall reconfiguration cycle. So we treat
1584 * it like a packet loss and let the retransmit timer and
1585 * timeouts do their work over time.
1586 * XXX: It is a POLA question whether calling tcp_drop right
1587 * away would be the really correct behavior instead.
1589 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1590 !tcp_timer_active(tp, TT_PERSIST)) &&
1591 ((flags & TH_SYN) == 0) &&
1595 tp->sackhint.sack_bytes_rexmit -= len;
1596 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1597 ("sackhint bytes rtx >= 0"));
1601 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1605 tp->t_softerror = error;
1608 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1609 tp->snd_cwnd = tp->t_maxseg;
1613 * For some reason the interface we used initially
1614 * to send segments changed to another or lowered
1616 * If TSO was active we either got an interface
1617 * without TSO capabilits or TSO was turned off.
1618 * If we obtained mtu from ip_output() then update
1622 tp->t_flags &= ~TF_TSO;
1624 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1632 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1633 tp->t_softerror = error;
1641 TCPSTAT_INC(tcps_sndtotal);
1644 * Data sent (as far as we can tell).
1645 * If this advertises a larger window than any other segment,
1646 * then remember the size of the advertised window.
1647 * Any pending ACK has now been sent.
1649 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1650 tp->rcv_adv = tp->rcv_nxt + recwin;
1651 tp->last_ack_sent = tp->rcv_nxt;
1652 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1653 if (tcp_timer_active(tp, TT_DELACK))
1654 tcp_timer_activate(tp, TT_DELACK, 0);
1657 * This completely breaks TCP if newreno is turned on. What happens
1658 * is that if delayed-acks are turned on on the receiver, this code
1659 * on the transmitter effectively destroys the TCP window, forcing
1660 * it to four packets (1.5Kx4 = 6K window).
1662 if (sendalot && --maxburst)
1671 tcp_setpersist(struct tcpcb *tp)
1673 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1676 tp->t_flags &= ~TF_PREVVALID;
1677 if (tcp_timer_active(tp, TT_REXMT))
1678 panic("tcp_setpersist: retransmit pending");
1680 * Start/restart persistence timer.
1682 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1683 tcp_persmin, tcp_persmax);
1684 tcp_timer_activate(tp, TT_PERSIST, tt);
1685 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1690 * Insert TCP options according to the supplied parameters to the place
1691 * optp in a consistent way. Can handle unaligned destinations.
1693 * The order of the option processing is crucial for optimal packing and
1694 * alignment for the scarce option space.
1696 * The optimal order for a SYN/SYN-ACK segment is:
1697 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1698 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1700 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1701 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1702 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1703 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1704 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1707 tcp_addoptions(struct tcpopt *to, u_char *optp)
1709 u_int32_t mask, optlen = 0;
1711 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1712 if ((to->to_flags & mask) != mask)
1714 if (optlen == TCP_MAXOLEN)
1716 switch (to->to_flags & mask) {
1718 while (optlen % 4) {
1719 optlen += TCPOLEN_NOP;
1720 *optp++ = TCPOPT_NOP;
1722 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1724 optlen += TCPOLEN_MAXSEG;
1725 *optp++ = TCPOPT_MAXSEG;
1726 *optp++ = TCPOLEN_MAXSEG;
1727 to->to_mss = htons(to->to_mss);
1728 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1729 optp += sizeof(to->to_mss);
1732 while (!optlen || optlen % 2 != 1) {
1733 optlen += TCPOLEN_NOP;
1734 *optp++ = TCPOPT_NOP;
1736 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1738 optlen += TCPOLEN_WINDOW;
1739 *optp++ = TCPOPT_WINDOW;
1740 *optp++ = TCPOLEN_WINDOW;
1741 *optp++ = to->to_wscale;
1744 while (optlen % 2) {
1745 optlen += TCPOLEN_NOP;
1746 *optp++ = TCPOPT_NOP;
1748 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1750 optlen += TCPOLEN_SACK_PERMITTED;
1751 *optp++ = TCPOPT_SACK_PERMITTED;
1752 *optp++ = TCPOLEN_SACK_PERMITTED;
1755 while (!optlen || optlen % 4 != 2) {
1756 optlen += TCPOLEN_NOP;
1757 *optp++ = TCPOPT_NOP;
1759 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1761 optlen += TCPOLEN_TIMESTAMP;
1762 *optp++ = TCPOPT_TIMESTAMP;
1763 *optp++ = TCPOLEN_TIMESTAMP;
1764 to->to_tsval = htonl(to->to_tsval);
1765 to->to_tsecr = htonl(to->to_tsecr);
1766 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1767 optp += sizeof(to->to_tsval);
1768 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1769 optp += sizeof(to->to_tsecr);
1773 int siglen = TCPOLEN_SIGNATURE - 2;
1775 while (!optlen || optlen % 4 != 2) {
1776 optlen += TCPOLEN_NOP;
1777 *optp++ = TCPOPT_NOP;
1779 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1780 to->to_flags &= ~TOF_SIGNATURE;
1783 optlen += TCPOLEN_SIGNATURE;
1784 *optp++ = TCPOPT_SIGNATURE;
1785 *optp++ = TCPOLEN_SIGNATURE;
1786 to->to_signature = optp;
1794 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1797 while (!optlen || optlen % 4 != 2) {
1798 optlen += TCPOLEN_NOP;
1799 *optp++ = TCPOPT_NOP;
1801 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1803 optlen += TCPOLEN_SACKHDR;
1804 *optp++ = TCPOPT_SACK;
1805 sackblks = min(to->to_nsacks,
1806 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1807 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1808 while (sackblks--) {
1809 sack_seq = htonl(sack->start);
1810 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1811 optp += sizeof(sack_seq);
1812 sack_seq = htonl(sack->end);
1813 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1814 optp += sizeof(sack_seq);
1815 optlen += TCPOLEN_SACK;
1818 TCPSTAT_INC(tcps_sack_send_blocks);
1825 /* XXX is there any point to aligning this option? */
1826 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1827 if (TCP_MAXOLEN - optlen < total_len) {
1828 to->to_flags &= ~TOF_FASTOPEN;
1831 *optp++ = TCPOPT_FAST_OPEN;
1832 *optp++ = total_len;
1833 if (to->to_tfo_len > 0) {
1834 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1835 optp += to->to_tfo_len;
1837 optlen += total_len;
1841 panic("%s: unknown TCP option type", __func__);
1846 /* Terminate and pad TCP options to a 4 byte boundary. */
1848 optlen += TCPOLEN_EOL;
1849 *optp++ = TCPOPT_EOL;
1852 * According to RFC 793 (STD0007):
1853 * "The content of the header beyond the End-of-Option option
1854 * must be header padding (i.e., zero)."
1855 * and later: "The padding is composed of zeros."
1857 while (optlen % 4) {
1858 optlen += TCPOLEN_PAD;
1859 *optp++ = TCPOPT_PAD;
1862 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1867 * This is a copy of m_copym(), taking the TSO segment size/limit
1868 * constraints into account, and advancing the sndptr as it goes.
1871 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1872 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1875 struct ktls_session *tls, *ntls;
1878 struct mbuf *n, **np;
1881 int32_t len = *plen;
1885 uint32_t mlen, frags;
1889 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1890 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1891 if (off == 0 && m->m_flags & M_PKTHDR)
1896 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1900 if ((sb) && (m == sb->sb_sndptr)) {
1901 sb->sb_sndptroff += m->m_len;
1902 sb->sb_sndptr = m->m_next;
1910 if (m->m_flags & M_NOMAP)
1911 tls = m->m_ext.ext_pgs->tls;
1918 KASSERT(len == M_COPYALL,
1919 ("tcp_m_copym, length > size of mbuf chain"));
1921 if (pkthdrlen != NULL)
1922 *pkthdrlen = len_cp;
1927 if (m->m_flags & M_NOMAP)
1928 ntls = m->m_ext.ext_pgs->tls;
1933 * Avoid mixing TLS records with handshake
1934 * data or TLS records from different
1940 if (pkthdrlen != NULL)
1941 *pkthdrlen = len_cp;
1946 * Don't end a send in the middle of a TLS
1947 * record if it spans multiple TLS records.
1949 if (tls != NULL && (m != start) && len < m->m_len) {
1951 if (pkthdrlen != NULL)
1952 *pkthdrlen = len_cp;
1957 mlen = min(len, m->m_len - off);
1960 * For M_NOMAP mbufs, add 3 segments
1961 * + 1 in case we are crossing page boundaries
1962 * + 2 in case the TLS hdr/trailer are used
1963 * It is cheaper to just add the segments
1964 * than it is to take the cache miss to look
1965 * at the mbuf ext_pgs state in detail.
1967 if (m->m_flags & M_NOMAP) {
1968 fragsize = min(segsize, PAGE_SIZE);
1975 /* Break if we really can't fit anymore. */
1976 if ((frags + 1) >= seglimit) {
1978 if (pkthdrlen != NULL)
1979 *pkthdrlen = len_cp;
1984 * Reduce size if you can't copy the whole
1985 * mbuf. If we can't copy the whole mbuf, also
1986 * adjust len so the loop will end after this
1989 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1990 mlen = (seglimit - frags - 1) * fragsize;
1992 *plen = len_cp + len;
1993 if (pkthdrlen != NULL)
1996 frags += howmany(mlen, fragsize);
2000 KASSERT(seglimit > 0,
2001 ("%s: seglimit went too low", __func__));
2004 n = m_gethdr(M_NOWAIT, m->m_type);
2006 n = m_get(M_NOWAIT, m->m_type);
2011 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2013 if (len == M_COPYALL)
2014 n->m_pkthdr.len -= off0;
2016 n->m_pkthdr.len = len;
2017 pkthdrlen = &n->m_pkthdr.len;
2022 if (m->m_flags & M_EXT) {
2023 n->m_data = m->m_data + off;
2026 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2029 if (sb && (sb->sb_sndptr == m) &&
2030 ((n->m_len + off) >= m->m_len) && m->m_next) {
2031 sb->sb_sndptroff += m->m_len;
2032 sb->sb_sndptr = m->m_next;
2035 if (len != M_COPYALL) {
2048 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2052 * Automatic sizing of send socket buffer. Often the send buffer
2053 * size is not optimally adjusted to the actual network conditions
2054 * at hand (delay bandwidth product). Setting the buffer size too
2055 * small limits throughput on links with high bandwidth and high
2056 * delay (eg. trans-continental/oceanic links). Setting the
2057 * buffer size too big consumes too much real kernel memory,
2058 * especially with many connections on busy servers.
2060 * The criteria to step up the send buffer one notch are:
2061 * 1. receive window of remote host is larger than send buffer
2062 * (with a fudge factor of 5/4th);
2063 * 2. send buffer is filled to 7/8th with data (so we actually
2064 * have data to make use of it);
2065 * 3. send buffer fill has not hit maximal automatic size;
2066 * 4. our send window (slow start and cogestion controlled) is
2067 * larger than sent but unacknowledged data in send buffer.
2069 * The remote host receive window scaling factor may limit the
2070 * growing of the send buffer before it reaches its allowed
2073 * It scales directly with slow start or congestion window
2074 * and does at most one step per received ACK. This fast
2075 * scaling has the drawback of growing the send buffer beyond
2076 * what is strictly necessary to make full use of a given
2077 * delay*bandwidth product. However testing has shown this not
2078 * to be much of an problem. At worst we are trading wasting
2079 * of available bandwidth (the non-use of it) for wasting some
2080 * socket buffer memory.
2082 * TODO: Shrink send buffer during idle periods together
2083 * with congestion window. Requires another timer. Has to
2084 * wait for upcoming tcp timer rewrite.
2086 * XXXGL: should there be used sbused() or sbavail()?
2088 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2091 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2092 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2093 sbused(&so->so_snd) >=
2094 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2095 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2096 sendwin >= (sbused(&so->so_snd) -
2097 (tp->snd_nxt - tp->snd_una))) {
2098 if (!sbreserve_locked(&so->so_snd,
2099 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2100 V_tcp_autosndbuf_max), so, curthread))
2101 so->so_snd.sb_flags &= ~SB_AUTOSIZE;