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>
45 #include <sys/domain.h>
47 #include <sys/hhook.h>
49 #include <sys/kernel.h>
55 #include <sys/mutex.h>
56 #include <sys/protosw.h>
58 #include <sys/socket.h>
59 #include <sys/socketvar.h>
60 #include <sys/sysctl.h>
63 #include <net/route.h>
66 #include <netinet/in.h>
67 #include <netinet/in_kdtrace.h>
68 #include <netinet/in_systm.h>
69 #include <netinet/ip.h>
70 #include <netinet/in_pcb.h>
71 #include <netinet/ip_var.h>
72 #include <netinet/ip_options.h>
74 #include <netinet6/in6_pcb.h>
75 #include <netinet/ip6.h>
76 #include <netinet6/ip6_var.h>
78 #include <netinet/tcp.h>
80 #include <netinet/tcp_fsm.h>
81 #include <netinet/tcp_log_buf.h>
82 #include <netinet/tcp_seq.h>
83 #include <netinet/tcp_timer.h>
84 #include <netinet/tcp_var.h>
85 #include <netinet/tcpip.h>
86 #include <netinet/cc/cc.h>
87 #include <netinet/tcp_fastopen.h>
89 #include <netinet/tcp_pcap.h>
92 #include <netinet/tcp_debug.h>
95 #include <netinet/tcp_offload.h>
98 #include <netipsec/ipsec_support.h>
100 #include <machine/in_cksum.h>
102 #include <security/mac/mac_framework.h>
104 VNET_DEFINE(int, path_mtu_discovery) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
106 &VNET_NAME(path_mtu_discovery), 1,
107 "Enable Path MTU Discovery");
109 VNET_DEFINE(int, tcp_do_tso) = 1;
110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
111 &VNET_NAME(tcp_do_tso), 0,
112 "Enable TCP Segmentation Offload");
114 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
115 #define V_tcp_sendspace VNET(tcp_sendspace)
116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
117 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_do_autosndbuf), 0,
122 "Enable automatic send buffer sizing");
124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
126 &VNET_NAME(tcp_autosndbuf_inc), 0,
127 "Incrementor step size of automatic send buffer");
129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
131 &VNET_NAME(tcp_autosndbuf_max), 0,
132 "Max size of automatic send buffer");
134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
135 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
137 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
138 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
141 * Make sure that either retransmit or persist timer is set for SYN, FIN and
144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
145 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
146 tcp_timer_active((tp), TT_REXMT) || \
147 tcp_timer_active((tp), TT_PERSIST), \
148 ("neither rexmt nor persist timer is set"))
150 static void inline cc_after_idle(struct tcpcb *tp);
154 * Wrapper for the TCP established output helper hook.
157 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
158 struct tcpopt *to, uint32_t len, int tso)
160 struct tcp_hhook_data hhook_data;
162 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
166 hhook_data.len = len;
167 hhook_data.tso = tso;
169 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
176 * CC wrapper hook functions
179 cc_after_idle(struct tcpcb *tp)
181 INP_WLOCK_ASSERT(tp->t_inpcb);
183 if (CC_ALGO(tp)->after_idle != NULL)
184 CC_ALGO(tp)->after_idle(tp->ccv);
188 * Tcp output routine: figure out what should be sent and send it.
191 tcp_output(struct tcpcb *tp)
193 struct socket *so = tp->t_inpcb->inp_socket;
195 uint32_t recwin, sendwin;
196 int off, flags, error = 0; /* Keep compiler happy */
197 u_int if_hw_tsomaxsegcount = 0;
198 u_int if_hw_tsomaxsegsize = 0;
200 struct ip *ip = NULL;
202 struct ipovly *ipov = NULL;
205 u_char opt[TCP_MAXOLEN];
206 unsigned ipoptlen, optlen, hdrlen;
207 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
208 unsigned ipsec_optlen = 0;
210 int idle, sendalot, curticks;
211 int sack_rxmit, sack_bytes_rxmt;
215 unsigned int wanted_cookie = 0;
216 unsigned int dont_sendalot = 0;
218 int maxburst = TCP_MAXBURST;
221 struct ip6_hdr *ip6 = NULL;
224 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
227 const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
229 const bool hw_tls = false;
232 INP_WLOCK_ASSERT(tp->t_inpcb);
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)
260 tp->t_flags &= ~TF_LASTIDLE;
262 if (tp->t_flags & TF_MORETOCOME) {
263 tp->t_flags |= TF_LASTIDLE;
269 * If we've recently taken a timeout, snd_max will be greater than
270 * snd_nxt. There may be SACK information that allows us to avoid
271 * resending already delivered data. Adjust snd_nxt accordingly.
273 if ((tp->t_flags & TF_SACK_PERMIT) &&
274 SEQ_LT(tp->snd_nxt, tp->snd_max))
279 off = tp->snd_nxt - tp->snd_una;
280 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
282 flags = tcp_outflags[tp->t_state];
284 * Send any SACK-generated retransmissions. If we're explicitly trying
285 * to send out new data (when sendalot is 1), bypass this function.
286 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
287 * we're replacing a (future) new transmission with a retransmission
288 * now, and we previously incremented snd_cwnd in tcp_input().
291 * Still in sack recovery , reset rxmit flag to zero.
297 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
298 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
302 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
303 /* Do not retransmit SACK segments beyond snd_recover */
304 if (SEQ_GT(p->end, tp->snd_recover)) {
306 * (At least) part of sack hole extends beyond
307 * snd_recover. Check to see if we can rexmit data
310 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
312 * Can't rexmit any more data for this hole.
313 * That data will be rexmitted in the next
314 * sack recovery episode, when snd_recover
315 * moves past p->rxmit.
318 goto after_sack_rexmit;
320 /* Can rexmit part of the current hole */
321 len = ((int32_t)ulmin(cwin,
322 tp->snd_recover - p->rxmit));
324 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
325 off = p->rxmit - tp->snd_una;
326 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
331 TCPSTAT_INC(tcps_sack_rexmits);
332 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
333 min(len, tp->t_maxseg));
338 * Get standard flags, and add SYN or FIN if requested by 'hidden'
341 if (tp->t_flags & TF_NEEDFIN)
343 if (tp->t_flags & TF_NEEDSYN)
346 SOCKBUF_LOCK(&so->so_snd);
348 * If in persist timeout with window of 0, send 1 byte.
349 * Otherwise, if window is small but nonzero
350 * and timer expired, we will send what we can
351 * and go to transmit state.
353 if (tp->t_flags & TF_FORCEDATA) {
356 * If we still have some data to send, then
357 * clear the FIN bit. Usually this would
358 * happen below when it realizes that we
359 * aren't sending all the data. However,
360 * if we have exactly 1 byte of unsent data,
361 * then it won't clear the FIN bit below,
362 * and if we are in persist state, we wind
363 * up sending the packet without recording
364 * that we sent the FIN bit.
366 * We can't just blindly clear the FIN bit,
367 * because if we don't have any more data
368 * to send then the probe will be the FIN
371 if (off < sbused(&so->so_snd))
375 tcp_timer_activate(tp, TT_PERSIST, 0);
381 * If snd_nxt == snd_max and we have transmitted a FIN, the
382 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
383 * a negative length. This can also occur when TCP opens up
384 * its congestion window while receiving additional duplicate
385 * acks after fast-retransmit because TCP will reset snd_nxt
386 * to snd_max after the fast-retransmit.
388 * In the normal retransmit-FIN-only case, however, snd_nxt will
389 * be set to snd_una, the offset will be 0, and the length may
392 * If sack_rxmit is true we are retransmitting from the scoreboard
393 * in which case len is already set.
395 if (sack_rxmit == 0) {
396 if (sack_bytes_rxmt == 0)
397 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
403 * We are inside of a SACK recovery episode and are
404 * sending new data, having retransmitted all the
405 * data possible in the scoreboard.
407 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
410 * Don't remove this (len > 0) check !
411 * We explicitly check for len > 0 here (although it
412 * isn't really necessary), to work around a gcc
413 * optimization issue - to force gcc to compute
414 * len above. Without this check, the computation
415 * of len is bungled by the optimizer.
418 cwin = tp->snd_cwnd -
419 (tp->snd_nxt - tp->sack_newdata) -
423 len = imin(len, cwin);
429 * Lop off SYN bit if it has already been sent. However, if this
430 * is SYN-SENT state and if segment contains data and if we don't
431 * know that foreign host supports TAO, suppress sending segment.
433 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
434 if (tp->t_state != TCPS_SYN_RECEIVED)
437 * When sending additional segments following a TFO SYN|ACK,
438 * do not include the SYN bit.
440 if (IS_FASTOPEN(tp->t_flags) &&
441 (tp->t_state == TCPS_SYN_RECEIVED))
447 * Be careful not to send data and/or FIN on SYN segments.
448 * This measure is needed to prevent interoperability problems
449 * with not fully conformant TCP implementations.
451 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
457 * On TFO sockets, ensure no data is sent in the following cases:
459 * - When retransmitting SYN|ACK on a passively-created socket
461 * - When retransmitting SYN on an actively created socket
463 * - When sending a zero-length cookie (cookie request) on an
464 * actively created socket
466 * - When the socket is in the CLOSED state (RST is being sent)
468 if (IS_FASTOPEN(tp->t_flags) &&
469 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
470 ((tp->t_state == TCPS_SYN_SENT) &&
471 (tp->t_tfo_client_cookie_len == 0)) ||
476 * If FIN has been sent but not acked,
477 * but we haven't been called to retransmit,
478 * len will be < 0. Otherwise, window shrank
479 * after we sent into it. If window shrank to 0,
480 * cancel pending retransmit, pull snd_nxt back
481 * to (closed) window, and set the persist timer
482 * if it isn't already going. If the window didn't
483 * close completely, just wait for an ACK.
485 * We also do a general check here to ensure that
486 * we will set the persist timer when we have data
487 * to send, but a 0-byte window. This makes sure
488 * the persist timer is set even if the packet
489 * hits one of the "goto send" lines below.
492 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
493 (off < (int) sbavail(&so->so_snd))) {
494 tcp_timer_activate(tp, TT_REXMT, 0);
496 tp->snd_nxt = tp->snd_una;
497 if (!tcp_timer_active(tp, TT_PERSIST))
502 /* len will be >= 0 after this point. */
503 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
505 tcp_sndbuf_autoscale(tp, so, sendwin);
508 * Decide if we can use TCP Segmentation Offloading (if supported by
511 * TSO may only be used if we are in a pure bulk sending state. The
512 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
513 * IP options prevent using TSO. With TSO the TCP header is the same
514 * (except for the sequence number) for all generated packets. This
515 * makes it impossible to transmit any options which vary per generated
518 * IPv4 handling has a clear separation of ip options and ip header
519 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
520 * the right thing below to provide length of just ip options and thus
521 * checking for ipoptlen is enough to decide if ip options are present.
523 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
525 * Pre-calculate here as we save another lookup into the darknesses
526 * of IPsec that way and can actually decide if TSO is ok.
529 if (isipv6 && IPSEC_ENABLED(ipv6))
530 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
536 if (IPSEC_ENABLED(ipv4))
537 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
542 ipoptlen = ip6_optlen(tp->t_inpcb);
545 if (tp->t_inpcb->inp_options)
546 ipoptlen = tp->t_inpcb->inp_options->m_len -
547 offsetof(struct ipoption, ipopt_list);
550 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
551 ipoptlen += ipsec_optlen;
554 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
555 ((tp->t_flags & TF_SIGNATURE) == 0) &&
556 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
557 ipoptlen == 0 && !(flags & TH_SYN))
561 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
564 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
565 sbused(&so->so_snd)))
569 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
570 (long)TCP_MAXWIN << tp->rcv_scale);
573 * Sender silly window avoidance. We transmit under the following
574 * conditions when len is non-zero:
576 * - We have a full segment (or more with TSO)
577 * - This is the last buffer in a write()/send() and we are
578 * either idle or running NODELAY
579 * - we've timed out (e.g. persist timer)
580 * - we have more then 1/2 the maximum send window's worth of
581 * data (receiver may be limited the window size)
582 * - we need to retransmit
585 if (len >= tp->t_maxseg)
588 * NOTE! on localhost connections an 'ack' from the remote
589 * end may occur synchronously with the output and cause
590 * us to flush a buffer queued with moretocome. XXX
592 * note: the len + off check is almost certainly unnecessary.
594 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
595 (idle || (tp->t_flags & TF_NODELAY)) &&
596 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
597 (tp->t_flags & TF_NOPUSH) == 0) {
600 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
602 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
604 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
611 * Sending of standalone window updates.
613 * Window updates are important when we close our window due to a
614 * full socket buffer and are opening it again after the application
615 * reads data from it. Once the window has opened again and the
616 * remote end starts to send again the ACK clock takes over and
617 * provides the most current window information.
619 * We must avoid the silly window syndrome whereas every read
620 * from the receive buffer, no matter how small, causes a window
621 * update to be sent. We also should avoid sending a flurry of
622 * window updates when the socket buffer had queued a lot of data
623 * and the application is doing small reads.
625 * Prevent a flurry of pointless window updates by only sending
626 * an update when we can increase the advertized window by more
627 * than 1/4th of the socket buffer capacity. When the buffer is
628 * getting full or is very small be more aggressive and send an
629 * update whenever we can increase by two mss sized segments.
630 * In all other situations the ACK's to new incoming data will
631 * carry further window increases.
633 * Don't send an independent window update if a delayed
634 * ACK is pending (it will get piggy-backed on it) or the
635 * remote side already has done a half-close and won't send
636 * more data. Skip this if the connection is in T/TCP
639 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
640 !(tp->t_flags & TF_DELACK) &&
641 !TCPS_HAVERCVDFIN(tp->t_state)) {
643 * "adv" is the amount we could increase the window,
644 * taking into account that we are limited by
645 * TCP_MAXWIN << tp->rcv_scale.
651 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
652 oldwin = (tp->rcv_adv - tp->rcv_nxt);
658 * If the new window size ends up being the same as or less
659 * than the old size when it is scaled, then don't force
662 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
665 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
666 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
667 recwin <= (so->so_rcv.sb_hiwat / 8) ||
668 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
669 adv >= TCP_MAXWIN << tp->rcv_scale))
671 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
677 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
678 * is also a catch-all for the retransmit timer timeout case.
680 if (tp->t_flags & TF_ACKNOW)
682 if ((flags & TH_RST) ||
683 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
685 if (SEQ_GT(tp->snd_up, tp->snd_una))
688 * If our state indicates that FIN should be sent
689 * and we have not yet done so, then we need to send.
691 if (flags & TH_FIN &&
692 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
695 * In SACK, it is possible for tcp_output to fail to send a segment
696 * after the retransmission timer has been turned off. Make sure
697 * that the retransmission timer is set.
699 if ((tp->t_flags & TF_SACK_PERMIT) &&
700 SEQ_GT(tp->snd_max, tp->snd_una) &&
701 !tcp_timer_active(tp, TT_REXMT) &&
702 !tcp_timer_active(tp, TT_PERSIST)) {
703 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
707 * TCP window updates are not reliable, rather a polling protocol
708 * using ``persist'' packets is used to insure receipt of window
709 * updates. The three ``states'' for the output side are:
710 * idle not doing retransmits or persists
711 * persisting to move a small or zero window
712 * (re)transmitting and thereby not persisting
714 * tcp_timer_active(tp, TT_PERSIST)
715 * is true when we are in persist state.
716 * (tp->t_flags & TF_FORCEDATA)
717 * is set when we are called to send a persist packet.
718 * tcp_timer_active(tp, TT_REXMT)
719 * is set when we are retransmitting
720 * The output side is idle when both timers are zero.
722 * If send window is too small, there is data to transmit, and no
723 * retransmit or persist is pending, then go to persist state.
724 * If nothing happens soon, send when timer expires:
725 * if window is nonzero, transmit what we can,
726 * otherwise force out a byte.
728 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
729 !tcp_timer_active(tp, TT_PERSIST)) {
735 * No reason to send a segment, just return.
738 SOCKBUF_UNLOCK(&so->so_snd);
742 SOCKBUF_LOCK_ASSERT(&so->so_snd);
744 if (len >= tp->t_maxseg)
745 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
747 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
750 * Before ESTABLISHED, force sending of initial options
751 * unless TCP set not to do any options.
752 * NOTE: we assume that the IP/TCP header plus TCP options
753 * always fit in a single mbuf, leaving room for a maximum
755 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
760 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
763 hdrlen = sizeof (struct tcpiphdr);
766 * Compute options for segment.
767 * We only have to care about SYN and established connection
768 * segments. Options for SYN-ACK segments are handled in TCP
772 if ((tp->t_flags & TF_NOOPT) == 0) {
773 /* Maximum segment size. */
774 if (flags & TH_SYN) {
775 tp->snd_nxt = tp->iss;
776 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
777 to.to_flags |= TOF_MSS;
780 * On SYN or SYN|ACK transmits on TFO connections,
781 * only include the TFO option if it is not a
782 * retransmit, as the presence of the TFO option may
783 * have caused the original SYN or SYN|ACK to have
784 * been dropped by a middlebox.
786 if (IS_FASTOPEN(tp->t_flags) &&
787 (tp->t_rxtshift == 0)) {
788 if (tp->t_state == TCPS_SYN_RECEIVED) {
789 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
791 (u_int8_t *)&tp->t_tfo_cookie.server;
792 to.to_flags |= TOF_FASTOPEN;
794 } else if (tp->t_state == TCPS_SYN_SENT) {
796 tp->t_tfo_client_cookie_len;
798 tp->t_tfo_cookie.client;
799 to.to_flags |= TOF_FASTOPEN;
802 * If we wind up having more data to
803 * send with the SYN than can fit in
804 * one segment, don't send any more
805 * until the SYN|ACK comes back from
812 /* Window scaling. */
813 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
814 to.to_wscale = tp->request_r_scale;
815 to.to_flags |= TOF_SCALE;
818 if ((tp->t_flags & TF_RCVD_TSTMP) ||
819 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
820 curticks = tcp_ts_getticks();
821 to.to_tsval = curticks + tp->ts_offset;
822 to.to_tsecr = tp->ts_recent;
823 to.to_flags |= TOF_TS;
824 if (tp->t_rxtshift == 1)
825 tp->t_badrxtwin = curticks;
828 /* Set receive buffer autosizing timestamp. */
829 if (tp->rfbuf_ts == 0 &&
830 (so->so_rcv.sb_flags & SB_AUTOSIZE))
831 tp->rfbuf_ts = tcp_ts_getticks();
833 /* Selective ACK's. */
834 if (tp->t_flags & TF_SACK_PERMIT) {
836 to.to_flags |= TOF_SACKPERM;
837 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
838 (tp->t_flags & TF_SACK_PERMIT) &&
839 tp->rcv_numsacks > 0) {
840 to.to_flags |= TOF_SACK;
841 to.to_nsacks = tp->rcv_numsacks;
842 to.to_sacks = (u_char *)tp->sackblks;
845 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
846 /* TCP-MD5 (RFC2385). */
848 * Check that TCP_MD5SIG is enabled in tcpcb to
849 * account the size needed to set this TCP option.
851 if (tp->t_flags & TF_SIGNATURE)
852 to.to_flags |= TOF_SIGNATURE;
853 #endif /* TCP_SIGNATURE */
855 /* Processing the options. */
856 hdrlen += optlen = tcp_addoptions(&to, opt);
858 * If we wanted a TFO option to be added, but it was unable
859 * to fit, ensure no data is sent.
861 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
862 !(to.to_flags & TOF_FASTOPEN))
867 * Adjust data length if insertion of options will
868 * bump the packet length beyond the t_maxseg length.
869 * Clear the FIN bit because we cut off the tail of
872 if (len + optlen + ipoptlen > tp->t_maxseg) {
880 /* extract TSO information */
881 if_hw_tsomax = tp->t_tsomax;
882 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
883 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
886 * Limit a TSO burst to prevent it from
887 * overflowing or exceeding the maximum length
888 * allowed by the network interface:
890 KASSERT(ipoptlen == 0,
891 ("%s: TSO can't do IP options", __func__));
894 * Check if we should limit by maximum payload
897 if (if_hw_tsomax != 0) {
898 /* compute maximum TSO length */
899 max_len = (if_hw_tsomax - hdrlen -
903 } else if (len > max_len) {
910 * Prevent the last segment from being
911 * fractional unless the send sockbuf can be
914 max_len = (tp->t_maxseg - optlen);
915 if (((uint32_t)off + (uint32_t)len) <
916 sbavail(&so->so_snd)) {
917 moff = len % max_len;
925 * In case there are too many small fragments
928 if (len <= max_len) {
935 * Send the FIN in a separate segment
936 * after the bulk sending is done.
937 * We don't trust the TSO implementations
938 * to clear the FIN flag on all but the
941 if (tp->t_flags & TF_NEEDFIN)
944 if (optlen + ipoptlen >= tp->t_maxseg) {
946 * Since we don't have enough space to put
947 * the IP header chain and the TCP header in
948 * one packet as required by RFC 7112, don't
949 * send it. Also ensure that at least one
950 * byte of the payload can be put into the
953 SOCKBUF_UNLOCK(&so->so_snd);
958 len = tp->t_maxseg - optlen - ipoptlen;
966 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
967 ("%s: len > IP_MAXPACKET", __func__));
969 /*#ifdef DIAGNOSTIC*/
971 if (max_linkhdr + hdrlen > MCLBYTES)
973 if (max_linkhdr + hdrlen > MHLEN)
975 panic("tcphdr too big");
979 * This KASSERT is here to catch edge cases at a well defined place.
980 * Before, those had triggered (random) panic conditions further down.
982 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
985 * Grab a header mbuf, attaching a copy of data to
986 * be transmitted, and initialize the header from
987 * the template for sends on this connection.
994 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
995 TCPSTAT_INC(tcps_sndprobe);
996 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
997 tp->t_sndrexmitpack++;
998 TCPSTAT_INC(tcps_sndrexmitpack);
999 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1001 TCPSTAT_INC(tcps_sndpack);
1002 TCPSTAT_ADD(tcps_sndbyte, len);
1005 if (MHLEN < hdrlen + max_linkhdr)
1006 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1009 m = m_gethdr(M_NOWAIT, MT_DATA);
1012 SOCKBUF_UNLOCK(&so->so_snd);
1018 m->m_data += max_linkhdr;
1022 * Start the m_copy functions from the closest mbuf
1023 * to the offset in the socket buffer chain.
1025 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1026 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1027 m_copydata(mb, moff, len,
1028 mtod(m, caddr_t) + hdrlen);
1029 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1030 sbsndptr_adv(&so->so_snd, mb, len);
1033 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1037 m->m_next = tcp_m_copym(mb, moff,
1038 &len, if_hw_tsomaxsegcount,
1039 if_hw_tsomaxsegsize, msb, hw_tls);
1040 if (len <= (tp->t_maxseg - optlen)) {
1042 * Must have ran out of mbufs for the copy
1043 * shorten it to no longer need tso. Lets
1044 * not put on sendalot since we are low on
1049 if (m->m_next == NULL) {
1050 SOCKBUF_UNLOCK(&so->so_snd);
1059 * If we're sending everything we've got, set PUSH.
1060 * (This will keep happy those implementations which only
1061 * give data to the user when a buffer fills or
1064 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1067 SOCKBUF_UNLOCK(&so->so_snd);
1069 SOCKBUF_UNLOCK(&so->so_snd);
1070 if (tp->t_flags & TF_ACKNOW)
1071 TCPSTAT_INC(tcps_sndacks);
1072 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1073 TCPSTAT_INC(tcps_sndctrl);
1074 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1075 TCPSTAT_INC(tcps_sndurg);
1077 TCPSTAT_INC(tcps_sndwinup);
1079 m = m_gethdr(M_NOWAIT, MT_DATA);
1086 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1091 m->m_data += max_linkhdr;
1094 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1095 m->m_pkthdr.rcvif = (struct ifnet *)0;
1097 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1101 ip6 = mtod(m, struct ip6_hdr *);
1102 th = (struct tcphdr *)(ip6 + 1);
1103 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1107 ip = mtod(m, struct ip *);
1109 ipov = (struct ipovly *)ip;
1111 th = (struct tcphdr *)(ip + 1);
1112 tcpip_fillheaders(tp->t_inpcb, ip, th);
1116 * Fill in fields, remembering maximum advertised
1117 * window for use in delaying messages about window sizes.
1118 * If resending a FIN, be sure not to use a new sequence number.
1120 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1121 tp->snd_nxt == tp->snd_max)
1124 * If we are starting a connection, send ECN setup
1125 * SYN packet. If we are on a retransmit, we may
1126 * resend those bits a number of times as per
1129 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1130 if (tp->t_rxtshift >= 1) {
1131 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1132 flags |= TH_ECE|TH_CWR;
1134 flags |= TH_ECE|TH_CWR;
1137 if (tp->t_state == TCPS_ESTABLISHED &&
1138 (tp->t_flags & TF_ECN_PERMIT)) {
1140 * If the peer has ECN, mark data packets with
1141 * ECN capable transmission (ECT).
1142 * Ignore pure ack packets, retransmissions and window probes.
1144 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1145 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1148 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1151 ip->ip_tos |= IPTOS_ECN_ECT0;
1152 TCPSTAT_INC(tcps_ecn_ect0);
1156 * Reply with proper ECN notifications.
1158 if (tp->t_flags & TF_ECN_SND_CWR) {
1160 tp->t_flags &= ~TF_ECN_SND_CWR;
1162 if (tp->t_flags & TF_ECN_SND_ECE)
1167 * If we are doing retransmissions, then snd_nxt will
1168 * not reflect the first unsent octet. For ACK only
1169 * packets, we do not want the sequence number of the
1170 * retransmitted packet, we want the sequence number
1171 * of the next unsent octet. So, if there is no data
1172 * (and no SYN or FIN), use snd_max instead of snd_nxt
1173 * when filling in ti_seq. But if we are in persist
1174 * state, snd_max might reflect one byte beyond the
1175 * right edge of the window, so use snd_nxt in that
1176 * case, since we know we aren't doing a retransmission.
1177 * (retransmit and persist are mutually exclusive...)
1179 if (sack_rxmit == 0) {
1180 if (len || (flags & (TH_SYN|TH_FIN)) ||
1181 tcp_timer_active(tp, TT_PERSIST))
1182 th->th_seq = htonl(tp->snd_nxt);
1184 th->th_seq = htonl(tp->snd_max);
1186 th->th_seq = htonl(p->rxmit);
1188 tp->sackhint.sack_bytes_rexmit += len;
1190 th->th_ack = htonl(tp->rcv_nxt);
1192 bcopy(opt, th + 1, optlen);
1193 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1195 th->th_flags = flags;
1197 * Calculate receive window. Don't shrink window,
1198 * but avoid silly window syndrome.
1199 * If a RST segment is sent, advertise a window of zero.
1201 if (flags & TH_RST) {
1204 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1205 recwin < tp->t_maxseg)
1207 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1208 recwin < (tp->rcv_adv - tp->rcv_nxt))
1209 recwin = (tp->rcv_adv - tp->rcv_nxt);
1212 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1213 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1214 * case is handled in syncache.
1217 th->th_win = htons((u_short)
1218 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1220 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1223 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1224 * a 0 window. This may cause the remote transmitter to stall. This
1225 * flag tells soreceive() to disable delayed acknowledgements when
1226 * draining the buffer. This can occur if the receiver is attempting
1227 * to read more data than can be buffered prior to transmitting on
1230 if (th->th_win == 0) {
1232 tp->t_flags |= TF_RXWIN0SENT;
1234 tp->t_flags &= ~TF_RXWIN0SENT;
1235 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1236 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1237 th->th_flags |= TH_URG;
1240 * If no urgent pointer to send, then we pull
1241 * the urgent pointer to the left edge of the send window
1242 * so that it doesn't drift into the send window on sequence
1243 * number wraparound.
1245 tp->snd_up = tp->snd_una; /* drag it along */
1248 * Put TCP length in extended header, and then
1249 * checksum extended header and data.
1251 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1252 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1254 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1255 if (to.to_flags & TOF_SIGNATURE) {
1257 * Calculate MD5 signature and put it into the place
1258 * determined before.
1259 * NOTE: since TCP options buffer doesn't point into
1260 * mbuf's data, calculate offset and use it.
1262 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1263 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1265 * Do not send segment if the calculation of MD5
1266 * digest has failed.
1276 * There is no need to fill in ip6_plen right now.
1277 * It will be filled later by ip6_output.
1279 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1280 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1281 optlen + len, IPPROTO_TCP, 0);
1284 #if defined(INET6) && defined(INET)
1289 m->m_pkthdr.csum_flags = CSUM_TCP;
1290 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1291 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1293 /* IP version must be set here for ipv4/ipv6 checking later */
1294 KASSERT(ip->ip_v == IPVERSION,
1295 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1300 * Enable TSO and specify the size of the segments.
1301 * The TCP pseudo header checksum is always provided.
1304 KASSERT(len > tp->t_maxseg - optlen,
1305 ("%s: len <= tso_segsz", __func__));
1306 m->m_pkthdr.csum_flags |= CSUM_TSO;
1307 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1310 KASSERT(len + hdrlen == m_length(m, NULL),
1311 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1312 __func__, len, hdrlen, m_length(m, NULL)));
1315 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1316 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1323 if (so->so_options & SO_DEBUG) {
1329 save = ipov->ih_len;
1330 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1332 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1336 ipov->ih_len = save;
1338 #endif /* TCPDEBUG */
1339 TCP_PROBE3(debug__output, tp, th, m);
1341 /* We're getting ready to send; log now. */
1342 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1346 * Fill in IP length and desired time to live and
1347 * send to IP level. There should be a better way
1348 * to handle ttl and tos; we could keep them in
1349 * the template, but need a way to checksum without them.
1352 * m->m_pkthdr.len should have been set before checksum calculation,
1353 * because in6_cksum() need it.
1358 * we separately set hoplimit for every segment, since the
1359 * user might want to change the value via setsockopt.
1360 * Also, desired default hop limit might be changed via
1361 * Neighbor Discovery.
1363 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1366 * Set the packet size here for the benefit of DTrace probes.
1367 * ip6_output() will set it properly; it's supposed to include
1368 * the option header lengths as well.
1370 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1372 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1373 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1375 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1377 if (tp->t_state == TCPS_SYN_SENT)
1378 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1380 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1383 /* Save packet, if requested. */
1384 tcp_pcap_add(th, m, &(tp->t_outpkts));
1387 /* TODO: IPv6 IP6TOS_ECT bit on */
1388 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1389 &tp->t_inpcb->inp_route6,
1390 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1391 NULL, NULL, tp->t_inpcb);
1393 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1394 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1397 #if defined(INET) && defined(INET6)
1402 ip->ip_len = htons(m->m_pkthdr.len);
1404 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1405 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1408 * If we do path MTU discovery, then we set DF on every packet.
1409 * This might not be the best thing to do according to RFC3390
1410 * Section 2. However the tcp hostcache migitates the problem
1411 * so it affects only the first tcp connection with a host.
1413 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1415 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1416 ip->ip_off |= htons(IP_DF);
1417 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1419 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1422 if (tp->t_state == TCPS_SYN_SENT)
1423 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1425 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1428 /* Save packet, if requested. */
1429 tcp_pcap_add(th, m, &(tp->t_outpkts));
1432 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1433 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1436 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1437 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1443 * In transmit state, time the transmission and arrange for
1444 * the retransmit. In persist state, just set snd_max.
1446 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1447 !tcp_timer_active(tp, TT_PERSIST)) {
1448 tcp_seq startseq = tp->snd_nxt;
1451 * Advance snd_nxt over sequence space of this segment.
1453 if (flags & (TH_SYN|TH_FIN)) {
1456 if (flags & TH_FIN) {
1458 tp->t_flags |= TF_SENTFIN;
1464 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1465 tp->snd_max = tp->snd_nxt;
1467 * Time this transmission if not a retransmission and
1468 * not currently timing anything.
1470 if (tp->t_rtttime == 0) {
1471 tp->t_rtttime = ticks;
1472 tp->t_rtseq = startseq;
1473 TCPSTAT_INC(tcps_segstimed);
1478 * Set retransmit timer if not currently set,
1479 * and not doing a pure ack or a keep-alive probe.
1480 * Initial value for retransmit timer is smoothed
1481 * round-trip time + 2 * round-trip time variance.
1482 * Initialize shift counter which is used for backoff
1483 * of retransmit time.
1486 if (!tcp_timer_active(tp, TT_REXMT) &&
1487 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1488 (tp->snd_nxt != tp->snd_una))) {
1489 if (tcp_timer_active(tp, TT_PERSIST)) {
1490 tcp_timer_activate(tp, TT_PERSIST, 0);
1493 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1494 } else if (len == 0 && sbavail(&so->so_snd) &&
1495 !tcp_timer_active(tp, TT_REXMT) &&
1496 !tcp_timer_active(tp, TT_PERSIST)) {
1498 * Avoid a situation where we do not set persist timer
1499 * after a zero window condition. For example:
1500 * 1) A -> B: packet with enough data to fill the window
1501 * 2) B -> A: ACK for #1 + new data (0 window
1503 * 3) A -> B: ACK for #2, 0 len packet
1505 * In this case, A will not activate the persist timer,
1506 * because it chose to send a packet. Unless tcp_output
1507 * is called for some other reason (delayed ack timer,
1508 * another input packet from B, socket syscall), A will
1509 * not send zero window probes.
1511 * So, if you send a 0-length packet, but there is data
1512 * in the socket buffer, and neither the rexmt or
1513 * persist timer is already set, then activate the
1521 * Persist case, update snd_max but since we are in
1522 * persist mode (no window) we do not update snd_nxt.
1527 if (flags & TH_FIN) {
1529 tp->t_flags |= TF_SENTFIN;
1531 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1532 tp->snd_max = tp->snd_nxt + xlen;
1535 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1536 (tp->t_flags & TF_SACK_PERMIT) &&
1537 tp->rcv_numsacks > 0)) {
1538 /* Clean up any DSACK's sent */
1539 tcp_clean_dsack_blocks(tp);
1542 /* Record the error. */
1543 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1544 error, 0, NULL, false);
1547 * We know that the packet was lost, so back out the
1548 * sequence number advance, if any.
1550 * If the error is EPERM the packet got blocked by the
1551 * local firewall. Normally we should terminate the
1552 * connection but the blocking may have been spurious
1553 * due to a firewall reconfiguration cycle. So we treat
1554 * it like a packet loss and let the retransmit timer and
1555 * timeouts do their work over time.
1556 * XXX: It is a POLA question whether calling tcp_drop right
1557 * away would be the really correct behavior instead.
1559 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1560 !tcp_timer_active(tp, TT_PERSIST)) &&
1561 ((flags & TH_SYN) == 0) &&
1565 tp->sackhint.sack_bytes_rexmit -= len;
1566 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1567 ("sackhint bytes rtx >= 0"));
1571 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1575 tp->t_softerror = error;
1578 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1579 tp->snd_cwnd = tp->t_maxseg;
1583 * For some reason the interface we used initially
1584 * to send segments changed to another or lowered
1586 * If TSO was active we either got an interface
1587 * without TSO capabilits or TSO was turned off.
1588 * If we obtained mtu from ip_output() then update
1592 tp->t_flags &= ~TF_TSO;
1594 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1602 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1603 tp->t_softerror = error;
1611 TCPSTAT_INC(tcps_sndtotal);
1614 * Data sent (as far as we can tell).
1615 * If this advertises a larger window than any other segment,
1616 * then remember the size of the advertised window.
1617 * Any pending ACK has now been sent.
1619 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1620 tp->rcv_adv = tp->rcv_nxt + recwin;
1621 tp->last_ack_sent = tp->rcv_nxt;
1622 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1623 if (tcp_timer_active(tp, TT_DELACK))
1624 tcp_timer_activate(tp, TT_DELACK, 0);
1627 * This completely breaks TCP if newreno is turned on. What happens
1628 * is that if delayed-acks are turned on on the receiver, this code
1629 * on the transmitter effectively destroys the TCP window, forcing
1630 * it to four packets (1.5Kx4 = 6K window).
1632 if (sendalot && --maxburst)
1641 tcp_setpersist(struct tcpcb *tp)
1643 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1646 tp->t_flags &= ~TF_PREVVALID;
1647 if (tcp_timer_active(tp, TT_REXMT))
1648 panic("tcp_setpersist: retransmit pending");
1650 * Start/restart persistence timer.
1652 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1653 tcp_persmin, tcp_persmax);
1654 tcp_timer_activate(tp, TT_PERSIST, tt);
1655 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1660 * Insert TCP options according to the supplied parameters to the place
1661 * optp in a consistent way. Can handle unaligned destinations.
1663 * The order of the option processing is crucial for optimal packing and
1664 * alignment for the scarce option space.
1666 * The optimal order for a SYN/SYN-ACK segment is:
1667 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1668 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1670 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1671 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1672 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1673 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1674 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1677 tcp_addoptions(struct tcpopt *to, u_char *optp)
1679 u_int32_t mask, optlen = 0;
1681 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1682 if ((to->to_flags & mask) != mask)
1684 if (optlen == TCP_MAXOLEN)
1686 switch (to->to_flags & mask) {
1688 while (optlen % 4) {
1689 optlen += TCPOLEN_NOP;
1690 *optp++ = TCPOPT_NOP;
1692 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1694 optlen += TCPOLEN_MAXSEG;
1695 *optp++ = TCPOPT_MAXSEG;
1696 *optp++ = TCPOLEN_MAXSEG;
1697 to->to_mss = htons(to->to_mss);
1698 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1699 optp += sizeof(to->to_mss);
1702 while (!optlen || optlen % 2 != 1) {
1703 optlen += TCPOLEN_NOP;
1704 *optp++ = TCPOPT_NOP;
1706 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1708 optlen += TCPOLEN_WINDOW;
1709 *optp++ = TCPOPT_WINDOW;
1710 *optp++ = TCPOLEN_WINDOW;
1711 *optp++ = to->to_wscale;
1714 while (optlen % 2) {
1715 optlen += TCPOLEN_NOP;
1716 *optp++ = TCPOPT_NOP;
1718 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1720 optlen += TCPOLEN_SACK_PERMITTED;
1721 *optp++ = TCPOPT_SACK_PERMITTED;
1722 *optp++ = TCPOLEN_SACK_PERMITTED;
1725 while (!optlen || optlen % 4 != 2) {
1726 optlen += TCPOLEN_NOP;
1727 *optp++ = TCPOPT_NOP;
1729 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1731 optlen += TCPOLEN_TIMESTAMP;
1732 *optp++ = TCPOPT_TIMESTAMP;
1733 *optp++ = TCPOLEN_TIMESTAMP;
1734 to->to_tsval = htonl(to->to_tsval);
1735 to->to_tsecr = htonl(to->to_tsecr);
1736 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1737 optp += sizeof(to->to_tsval);
1738 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1739 optp += sizeof(to->to_tsecr);
1743 int siglen = TCPOLEN_SIGNATURE - 2;
1745 while (!optlen || optlen % 4 != 2) {
1746 optlen += TCPOLEN_NOP;
1747 *optp++ = TCPOPT_NOP;
1749 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1750 to->to_flags &= ~TOF_SIGNATURE;
1753 optlen += TCPOLEN_SIGNATURE;
1754 *optp++ = TCPOPT_SIGNATURE;
1755 *optp++ = TCPOLEN_SIGNATURE;
1756 to->to_signature = optp;
1764 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1767 while (!optlen || optlen % 4 != 2) {
1768 optlen += TCPOLEN_NOP;
1769 *optp++ = TCPOPT_NOP;
1771 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1773 optlen += TCPOLEN_SACKHDR;
1774 *optp++ = TCPOPT_SACK;
1775 sackblks = min(to->to_nsacks,
1776 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1777 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1778 while (sackblks--) {
1779 sack_seq = htonl(sack->start);
1780 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1781 optp += sizeof(sack_seq);
1782 sack_seq = htonl(sack->end);
1783 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1784 optp += sizeof(sack_seq);
1785 optlen += TCPOLEN_SACK;
1788 TCPSTAT_INC(tcps_sack_send_blocks);
1795 /* XXX is there any point to aligning this option? */
1796 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1797 if (TCP_MAXOLEN - optlen < total_len) {
1798 to->to_flags &= ~TOF_FASTOPEN;
1801 *optp++ = TCPOPT_FAST_OPEN;
1802 *optp++ = total_len;
1803 if (to->to_tfo_len > 0) {
1804 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1805 optp += to->to_tfo_len;
1807 optlen += total_len;
1811 panic("%s: unknown TCP option type", __func__);
1816 /* Terminate and pad TCP options to a 4 byte boundary. */
1818 optlen += TCPOLEN_EOL;
1819 *optp++ = TCPOPT_EOL;
1822 * According to RFC 793 (STD0007):
1823 * "The content of the header beyond the End-of-Option option
1824 * must be header padding (i.e., zero)."
1825 * and later: "The padding is composed of zeros."
1827 while (optlen % 4) {
1828 optlen += TCPOLEN_PAD;
1829 *optp++ = TCPOPT_PAD;
1832 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1837 * This is a copy of m_copym(), taking the TSO segment size/limit
1838 * constraints into account, and advancing the sndptr as it goes.
1841 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1842 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1845 struct ktls_session *tls, *ntls;
1848 struct mbuf *n, **np;
1851 int32_t len = *plen;
1855 uint32_t mlen, frags;
1859 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1860 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1861 if (off == 0 && m->m_flags & M_PKTHDR)
1866 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1870 if ((sb) && (m == sb->sb_sndptr)) {
1871 sb->sb_sndptroff += m->m_len;
1872 sb->sb_sndptr = m->m_next;
1880 if (m->m_flags & M_NOMAP)
1881 tls = m->m_ext.ext_pgs->tls;
1888 KASSERT(len == M_COPYALL,
1889 ("tcp_m_copym, length > size of mbuf chain"));
1891 if (pkthdrlen != NULL)
1892 *pkthdrlen = len_cp;
1897 if (m->m_flags & M_NOMAP)
1898 ntls = m->m_ext.ext_pgs->tls;
1903 * Avoid mixing TLS records with handshake
1904 * data or TLS records from different
1910 if (pkthdrlen != NULL)
1911 *pkthdrlen = len_cp;
1916 * Don't end a send in the middle of a TLS
1917 * record if it spans multiple TLS records.
1919 if (tls != NULL && (m != start) && len < m->m_len) {
1921 if (pkthdrlen != NULL)
1922 *pkthdrlen = len_cp;
1927 mlen = min(len, m->m_len - off);
1930 * For M_NOMAP mbufs, add 3 segments
1931 * + 1 in case we are crossing page boundaries
1932 * + 2 in case the TLS hdr/trailer are used
1933 * It is cheaper to just add the segments
1934 * than it is to take the cache miss to look
1935 * at the mbuf ext_pgs state in detail.
1937 if (m->m_flags & M_NOMAP) {
1938 fragsize = min(segsize, PAGE_SIZE);
1945 /* Break if we really can't fit anymore. */
1946 if ((frags + 1) >= seglimit) {
1948 if (pkthdrlen != NULL)
1949 *pkthdrlen = len_cp;
1954 * Reduce size if you can't copy the whole
1955 * mbuf. If we can't copy the whole mbuf, also
1956 * adjust len so the loop will end after this
1959 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1960 mlen = (seglimit - frags - 1) * fragsize;
1962 *plen = len_cp + len;
1963 if (pkthdrlen != NULL)
1966 frags += howmany(mlen, fragsize);
1970 KASSERT(seglimit > 0,
1971 ("%s: seglimit went too low", __func__));
1974 n = m_gethdr(M_NOWAIT, m->m_type);
1976 n = m_get(M_NOWAIT, m->m_type);
1981 if (!m_dup_pkthdr(n, m, M_NOWAIT))
1983 if (len == M_COPYALL)
1984 n->m_pkthdr.len -= off0;
1986 n->m_pkthdr.len = len;
1987 pkthdrlen = &n->m_pkthdr.len;
1992 if (m->m_flags & M_EXT) {
1993 n->m_data = m->m_data + off;
1996 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1999 if (sb && (sb->sb_sndptr == m) &&
2000 ((n->m_len + off) >= m->m_len) && m->m_next) {
2001 sb->sb_sndptroff += m->m_len;
2002 sb->sb_sndptr = m->m_next;
2005 if (len != M_COPYALL) {
2018 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2022 * Automatic sizing of send socket buffer. Often the send buffer
2023 * size is not optimally adjusted to the actual network conditions
2024 * at hand (delay bandwidth product). Setting the buffer size too
2025 * small limits throughput on links with high bandwidth and high
2026 * delay (eg. trans-continental/oceanic links). Setting the
2027 * buffer size too big consumes too much real kernel memory,
2028 * especially with many connections on busy servers.
2030 * The criteria to step up the send buffer one notch are:
2031 * 1. receive window of remote host is larger than send buffer
2032 * (with a fudge factor of 5/4th);
2033 * 2. send buffer is filled to 7/8th with data (so we actually
2034 * have data to make use of it);
2035 * 3. send buffer fill has not hit maximal automatic size;
2036 * 4. our send window (slow start and cogestion controlled) is
2037 * larger than sent but unacknowledged data in send buffer.
2039 * The remote host receive window scaling factor may limit the
2040 * growing of the send buffer before it reaches its allowed
2043 * It scales directly with slow start or congestion window
2044 * and does at most one step per received ACK. This fast
2045 * scaling has the drawback of growing the send buffer beyond
2046 * what is strictly necessary to make full use of a given
2047 * delay*bandwidth product. However testing has shown this not
2048 * to be much of an problem. At worst we are trading wasting
2049 * of available bandwidth (the non-use of it) for wasting some
2050 * socket buffer memory.
2052 * TODO: Shrink send buffer during idle periods together
2053 * with congestion window. Requires another timer. Has to
2054 * wait for upcoming tcp timer rewrite.
2056 * XXXGL: should there be used sbused() or sbavail()?
2058 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2061 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2062 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2063 sbused(&so->so_snd) >=
2064 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2065 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2066 sendwin >= (sbused(&so->so_snd) -
2067 (tp->snd_nxt - tp->snd_una))) {
2068 if (!sbreserve_locked(&so->so_snd,
2069 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2070 V_tcp_autosndbuf_max), so, curthread))
2071 so->so_snd.sb_flags &= ~SB_AUTOSIZE;