2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
46 #include <sys/hhook.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
59 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
74 #include <netinet/tcp.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/tcp_log_buf.h>
77 #include <netinet/tcp_seq.h>
78 #include <netinet/tcp_timer.h>
79 #include <netinet/tcp_var.h>
80 #include <netinet/tcpip.h>
81 #include <netinet/cc/cc.h>
82 #include <netinet/tcp_fastopen.h>
84 #include <netinet/tcp_pcap.h>
87 #include <netinet/tcp_debug.h>
90 #include <netinet/tcp_offload.h>
93 #include <netipsec/ipsec_support.h>
95 #include <machine/in_cksum.h>
97 #include <security/mac/mac_framework.h>
99 VNET_DEFINE(int, path_mtu_discovery) = 1;
100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
101 &VNET_NAME(path_mtu_discovery), 1,
102 "Enable Path MTU Discovery");
104 VNET_DEFINE(int, tcp_do_tso) = 1;
105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
106 &VNET_NAME(tcp_do_tso), 0,
107 "Enable TCP Segmentation Offload");
109 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
110 #define V_tcp_sendspace VNET(tcp_sendspace)
111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
112 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
116 &VNET_NAME(tcp_do_autosndbuf), 0,
117 "Enable automatic send buffer sizing");
119 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_autosndbuf_inc), 0,
122 "Incrementor step size of automatic send buffer");
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
126 &VNET_NAME(tcp_autosndbuf_max), 0,
127 "Max size of automatic send buffer");
129 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
130 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
132 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
133 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
136 * Make sure that either retransmit or persist timer is set for SYN, FIN and
139 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
140 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
141 tcp_timer_active((tp), TT_REXMT) || \
142 tcp_timer_active((tp), TT_PERSIST), \
143 ("neither rexmt nor persist timer is set"))
145 static void inline cc_after_idle(struct tcpcb *tp);
149 * Wrapper for the TCP established output helper hook.
152 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
153 struct tcpopt *to, uint32_t len, int tso)
155 struct tcp_hhook_data hhook_data;
157 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
161 hhook_data.len = len;
162 hhook_data.tso = tso;
164 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
171 * Flags used when sending segments in tcp_output. Basic flags (TH_RST,
172 * TH_ACK,TH_SYN,TH_FIN) are totally determined by state, with the proviso
173 * that TH_FIN is sent only if all data queued for output is included in the
176 u_char tcp_outflags[TCP_NSTATES] = {
177 TH_RST|TH_ACK, /* 0, CLOSED */
179 TH_SYN, /* 2, SYN_SENT */
180 TH_SYN|TH_ACK, /* 3, SYN_RECEIVED */
181 TH_ACK, /* 4, ESTABLISHED */
182 TH_ACK, /* 5, CLOSE_WAIT */
183 TH_FIN|TH_ACK, /* 6, FIN_WAIT_1 */
184 TH_FIN|TH_ACK, /* 7, CLOSING */
185 TH_FIN|TH_ACK, /* 8, LAST_ACK */
186 TH_ACK, /* 9, FIN_WAIT_2 */
187 TH_ACK, /* 10, TIME_WAIT */
191 * CC wrapper hook functions
194 cc_after_idle(struct tcpcb *tp)
196 INP_WLOCK_ASSERT(tp->t_inpcb);
198 if (CC_ALGO(tp)->after_idle != NULL)
199 CC_ALGO(tp)->after_idle(tp->ccv);
203 * Tcp output routine: figure out what should be sent and send it.
206 tcp_output(struct tcpcb *tp)
208 struct socket *so = tp->t_inpcb->inp_socket;
210 uint32_t recwin, sendwin;
211 int off, flags, error = 0; /* Keep compiler happy */
213 struct ip *ip = NULL;
215 struct ipovly *ipov = NULL;
218 u_char opt[TCP_MAXOLEN];
219 unsigned ipoptlen, optlen, hdrlen;
220 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
221 unsigned ipsec_optlen = 0;
223 int idle, sendalot, curticks;
224 int sack_rxmit, sack_bytes_rxmt;
228 unsigned int wanted_cookie = 0;
229 unsigned int dont_sendalot = 0;
231 int maxburst = TCP_MAXBURST;
234 struct ip6_hdr *ip6 = NULL;
237 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
240 INP_WLOCK_ASSERT(tp->t_inpcb);
243 if (tp->t_flags & TF_TOE)
244 return (tcp_offload_output(tp));
248 * For TFO connections in SYN_RECEIVED, only allow the initial
249 * SYN|ACK and those sent by the retransmit timer.
251 if (IS_FASTOPEN(tp->t_flags) &&
252 (tp->t_state == TCPS_SYN_RECEIVED) &&
253 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
254 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
258 * Determine length of data that should be transmitted,
259 * and flags that will be used.
260 * If there is some data or critical controls (SYN, RST)
261 * to send, then transmit; otherwise, investigate further.
263 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
264 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
266 tp->t_flags &= ~TF_LASTIDLE;
268 if (tp->t_flags & TF_MORETOCOME) {
269 tp->t_flags |= TF_LASTIDLE;
275 * If we've recently taken a timeout, snd_max will be greater than
276 * snd_nxt. There may be SACK information that allows us to avoid
277 * resending already delivered data. Adjust snd_nxt accordingly.
279 if ((tp->t_flags & TF_SACK_PERMIT) &&
280 SEQ_LT(tp->snd_nxt, tp->snd_max))
285 off = tp->snd_nxt - tp->snd_una;
286 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
288 flags = tcp_outflags[tp->t_state];
290 * Send any SACK-generated retransmissions. If we're explicitly trying
291 * to send out new data (when sendalot is 1), bypass this function.
292 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
293 * we're replacing a (future) new transmission with a retransmission
294 * now, and we previously incremented snd_cwnd in tcp_input().
297 * Still in sack recovery , reset rxmit flag to zero.
303 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
304 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
308 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
309 /* Do not retransmit SACK segments beyond snd_recover */
310 if (SEQ_GT(p->end, tp->snd_recover)) {
312 * (At least) part of sack hole extends beyond
313 * snd_recover. Check to see if we can rexmit data
316 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
318 * Can't rexmit any more data for this hole.
319 * That data will be rexmitted in the next
320 * sack recovery episode, when snd_recover
321 * moves past p->rxmit.
324 goto after_sack_rexmit;
326 /* Can rexmit part of the current hole */
327 len = ((int32_t)ulmin(cwin,
328 tp->snd_recover - p->rxmit));
330 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
331 off = p->rxmit - tp->snd_una;
332 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
337 TCPSTAT_INC(tcps_sack_rexmits);
338 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
339 min(len, tp->t_maxseg));
344 * Get standard flags, and add SYN or FIN if requested by 'hidden'
347 if (tp->t_flags & TF_NEEDFIN)
349 if (tp->t_flags & TF_NEEDSYN)
352 SOCKBUF_LOCK(&so->so_snd);
354 * If in persist timeout with window of 0, send 1 byte.
355 * Otherwise, if window is small but nonzero
356 * and timer expired, we will send what we can
357 * and go to transmit state.
359 if (tp->t_flags & TF_FORCEDATA) {
362 * If we still have some data to send, then
363 * clear the FIN bit. Usually this would
364 * happen below when it realizes that we
365 * aren't sending all the data. However,
366 * if we have exactly 1 byte of unsent data,
367 * then it won't clear the FIN bit below,
368 * and if we are in persist state, we wind
369 * up sending the packet without recording
370 * that we sent the FIN bit.
372 * We can't just blindly clear the FIN bit,
373 * because if we don't have any more data
374 * to send then the probe will be the FIN
377 if (off < sbused(&so->so_snd))
381 tcp_timer_activate(tp, TT_PERSIST, 0);
387 * If snd_nxt == snd_max and we have transmitted a FIN, the
388 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
389 * a negative length. This can also occur when TCP opens up
390 * its congestion window while receiving additional duplicate
391 * acks after fast-retransmit because TCP will reset snd_nxt
392 * to snd_max after the fast-retransmit.
394 * In the normal retransmit-FIN-only case, however, snd_nxt will
395 * be set to snd_una, the offset will be 0, and the length may
398 * If sack_rxmit is true we are retransmitting from the scoreboard
399 * in which case len is already set.
401 if (sack_rxmit == 0) {
402 if (sack_bytes_rxmt == 0)
403 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
409 * We are inside of a SACK recovery episode and are
410 * sending new data, having retransmitted all the
411 * data possible in the scoreboard.
413 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
416 * Don't remove this (len > 0) check !
417 * We explicitly check for len > 0 here (although it
418 * isn't really necessary), to work around a gcc
419 * optimization issue - to force gcc to compute
420 * len above. Without this check, the computation
421 * of len is bungled by the optimizer.
424 cwin = tp->snd_cwnd -
425 (tp->snd_nxt - tp->sack_newdata) -
429 len = imin(len, cwin);
435 * Lop off SYN bit if it has already been sent. However, if this
436 * is SYN-SENT state and if segment contains data and if we don't
437 * know that foreign host supports TAO, suppress sending segment.
439 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
440 if (tp->t_state != TCPS_SYN_RECEIVED)
443 * When sending additional segments following a TFO SYN|ACK,
444 * do not include the SYN bit.
446 if (IS_FASTOPEN(tp->t_flags) &&
447 (tp->t_state == TCPS_SYN_RECEIVED))
453 * Be careful not to send data and/or FIN on SYN segments.
454 * This measure is needed to prevent interoperability problems
455 * with not fully conformant TCP implementations.
457 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
463 * On TFO sockets, ensure no data is sent in the following cases:
465 * - When retransmitting SYN|ACK on a passively-created socket
467 * - When retransmitting SYN on an actively created socket
469 * - When sending a zero-length cookie (cookie request) on an
470 * actively created socket
472 * - When the socket is in the CLOSED state (RST is being sent)
474 if (IS_FASTOPEN(tp->t_flags) &&
475 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
476 ((tp->t_state == TCPS_SYN_SENT) &&
477 (tp->t_tfo_client_cookie_len == 0)) ||
482 * If FIN has been sent but not acked,
483 * but we haven't been called to retransmit,
484 * len will be < 0. Otherwise, window shrank
485 * after we sent into it. If window shrank to 0,
486 * cancel pending retransmit, pull snd_nxt back
487 * to (closed) window, and set the persist timer
488 * if it isn't already going. If the window didn't
489 * close completely, just wait for an ACK.
491 * We also do a general check here to ensure that
492 * we will set the persist timer when we have data
493 * to send, but a 0-byte window. This makes sure
494 * the persist timer is set even if the packet
495 * hits one of the "goto send" lines below.
498 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
499 (off < (int) sbavail(&so->so_snd))) {
500 tcp_timer_activate(tp, TT_REXMT, 0);
502 tp->snd_nxt = tp->snd_una;
503 if (!tcp_timer_active(tp, TT_PERSIST))
508 /* len will be >= 0 after this point. */
509 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
511 tcp_sndbuf_autoscale(tp, so, sendwin);
514 * Decide if we can use TCP Segmentation Offloading (if supported by
517 * TSO may only be used if we are in a pure bulk sending state. The
518 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
519 * IP options prevent using TSO. With TSO the TCP header is the same
520 * (except for the sequence number) for all generated packets. This
521 * makes it impossible to transmit any options which vary per generated
524 * IPv4 handling has a clear separation of ip options and ip header
525 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
526 * the right thing below to provide length of just ip options and thus
527 * checking for ipoptlen is enough to decide if ip options are present.
529 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
531 * Pre-calculate here as we save another lookup into the darknesses
532 * of IPsec that way and can actually decide if TSO is ok.
535 if (isipv6 && IPSEC_ENABLED(ipv6))
536 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
542 if (IPSEC_ENABLED(ipv4))
543 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
548 ipoptlen = ip6_optlen(tp->t_inpcb);
551 if (tp->t_inpcb->inp_options)
552 ipoptlen = tp->t_inpcb->inp_options->m_len -
553 offsetof(struct ipoption, ipopt_list);
556 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
557 ipoptlen += ipsec_optlen;
560 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
561 ((tp->t_flags & TF_SIGNATURE) == 0) &&
562 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
563 ipoptlen == 0 && !(flags & TH_SYN))
567 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
570 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
571 sbused(&so->so_snd)))
575 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
576 (long)TCP_MAXWIN << tp->rcv_scale);
579 * Sender silly window avoidance. We transmit under the following
580 * conditions when len is non-zero:
582 * - We have a full segment (or more with TSO)
583 * - This is the last buffer in a write()/send() and we are
584 * either idle or running NODELAY
585 * - we've timed out (e.g. persist timer)
586 * - we have more then 1/2 the maximum send window's worth of
587 * data (receiver may be limited the window size)
588 * - we need to retransmit
591 if (len >= tp->t_maxseg)
594 * NOTE! on localhost connections an 'ack' from the remote
595 * end may occur synchronously with the output and cause
596 * us to flush a buffer queued with moretocome. XXX
598 * note: the len + off check is almost certainly unnecessary.
600 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
601 (idle || (tp->t_flags & TF_NODELAY)) &&
602 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
603 (tp->t_flags & TF_NOPUSH) == 0) {
606 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
608 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
610 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
617 * Sending of standalone window updates.
619 * Window updates are important when we close our window due to a
620 * full socket buffer and are opening it again after the application
621 * reads data from it. Once the window has opened again and the
622 * remote end starts to send again the ACK clock takes over and
623 * provides the most current window information.
625 * We must avoid the silly window syndrome whereas every read
626 * from the receive buffer, no matter how small, causes a window
627 * update to be sent. We also should avoid sending a flurry of
628 * window updates when the socket buffer had queued a lot of data
629 * and the application is doing small reads.
631 * Prevent a flurry of pointless window updates by only sending
632 * an update when we can increase the advertized window by more
633 * than 1/4th of the socket buffer capacity. When the buffer is
634 * getting full or is very small be more aggressive and send an
635 * update whenever we can increase by two mss sized segments.
636 * In all other situations the ACK's to new incoming data will
637 * carry further window increases.
639 * Don't send an independent window update if a delayed
640 * ACK is pending (it will get piggy-backed on it) or the
641 * remote side already has done a half-close and won't send
642 * more data. Skip this if the connection is in T/TCP
645 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
646 !(tp->t_flags & TF_DELACK) &&
647 !TCPS_HAVERCVDFIN(tp->t_state)) {
649 * "adv" is the amount we could increase the window,
650 * taking into account that we are limited by
651 * TCP_MAXWIN << tp->rcv_scale.
657 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
658 oldwin = (tp->rcv_adv - tp->rcv_nxt);
664 * If the new window size ends up being the same as or less
665 * than the old size when it is scaled, then don't force
668 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
671 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
672 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
673 recwin <= (so->so_rcv.sb_hiwat / 8) ||
674 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
676 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
682 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
683 * is also a catch-all for the retransmit timer timeout case.
685 if (tp->t_flags & TF_ACKNOW)
687 if ((flags & TH_RST) ||
688 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
690 if (SEQ_GT(tp->snd_up, tp->snd_una))
693 * If our state indicates that FIN should be sent
694 * and we have not yet done so, then we need to send.
696 if (flags & TH_FIN &&
697 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
700 * In SACK, it is possible for tcp_output to fail to send a segment
701 * after the retransmission timer has been turned off. Make sure
702 * that the retransmission timer is set.
704 if ((tp->t_flags & TF_SACK_PERMIT) &&
705 SEQ_GT(tp->snd_max, tp->snd_una) &&
706 !tcp_timer_active(tp, TT_REXMT) &&
707 !tcp_timer_active(tp, TT_PERSIST)) {
708 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
712 * TCP window updates are not reliable, rather a polling protocol
713 * using ``persist'' packets is used to insure receipt of window
714 * updates. The three ``states'' for the output side are:
715 * idle not doing retransmits or persists
716 * persisting to move a small or zero window
717 * (re)transmitting and thereby not persisting
719 * tcp_timer_active(tp, TT_PERSIST)
720 * is true when we are in persist state.
721 * (tp->t_flags & TF_FORCEDATA)
722 * is set when we are called to send a persist packet.
723 * tcp_timer_active(tp, TT_REXMT)
724 * is set when we are retransmitting
725 * The output side is idle when both timers are zero.
727 * If send window is too small, there is data to transmit, and no
728 * retransmit or persist is pending, then go to persist state.
729 * If nothing happens soon, send when timer expires:
730 * if window is nonzero, transmit what we can,
731 * otherwise force out a byte.
733 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
734 !tcp_timer_active(tp, TT_PERSIST)) {
740 * No reason to send a segment, just return.
743 SOCKBUF_UNLOCK(&so->so_snd);
747 SOCKBUF_LOCK_ASSERT(&so->so_snd);
749 if (len >= tp->t_maxseg)
750 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
752 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
755 * Before ESTABLISHED, force sending of initial options
756 * unless TCP set not to do any options.
757 * NOTE: we assume that the IP/TCP header plus TCP options
758 * always fit in a single mbuf, leaving room for a maximum
760 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
765 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
768 hdrlen = sizeof (struct tcpiphdr);
771 * Compute options for segment.
772 * We only have to care about SYN and established connection
773 * segments. Options for SYN-ACK segments are handled in TCP
777 if ((tp->t_flags & TF_NOOPT) == 0) {
778 /* Maximum segment size. */
779 if (flags & TH_SYN) {
780 tp->snd_nxt = tp->iss;
781 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
782 to.to_flags |= TOF_MSS;
785 * On SYN or SYN|ACK transmits on TFO connections,
786 * only include the TFO option if it is not a
787 * retransmit, as the presence of the TFO option may
788 * have caused the original SYN or SYN|ACK to have
789 * been dropped by a middlebox.
791 if (IS_FASTOPEN(tp->t_flags) &&
792 (tp->t_rxtshift == 0)) {
793 if (tp->t_state == TCPS_SYN_RECEIVED) {
794 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
796 (u_int8_t *)&tp->t_tfo_cookie.server;
797 to.to_flags |= TOF_FASTOPEN;
799 } else if (tp->t_state == TCPS_SYN_SENT) {
801 tp->t_tfo_client_cookie_len;
803 tp->t_tfo_cookie.client;
804 to.to_flags |= TOF_FASTOPEN;
807 * If we wind up having more data to
808 * send with the SYN than can fit in
809 * one segment, don't send any more
810 * until the SYN|ACK comes back from
817 /* Window scaling. */
818 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
819 to.to_wscale = tp->request_r_scale;
820 to.to_flags |= TOF_SCALE;
823 if ((tp->t_flags & TF_RCVD_TSTMP) ||
824 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
825 curticks = tcp_ts_getticks();
826 to.to_tsval = curticks + tp->ts_offset;
827 to.to_tsecr = tp->ts_recent;
828 to.to_flags |= TOF_TS;
829 if (tp->t_rxtshift == 1)
830 tp->t_badrxtwin = curticks;
833 /* Set receive buffer autosizing timestamp. */
834 if (tp->rfbuf_ts == 0 &&
835 (so->so_rcv.sb_flags & SB_AUTOSIZE))
836 tp->rfbuf_ts = tcp_ts_getticks();
838 /* Selective ACK's. */
839 if (tp->t_flags & TF_SACK_PERMIT) {
841 to.to_flags |= TOF_SACKPERM;
842 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
843 (tp->t_flags & TF_SACK_PERMIT) &&
844 tp->rcv_numsacks > 0) {
845 to.to_flags |= TOF_SACK;
846 to.to_nsacks = tp->rcv_numsacks;
847 to.to_sacks = (u_char *)tp->sackblks;
850 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
851 /* TCP-MD5 (RFC2385). */
853 * Check that TCP_MD5SIG is enabled in tcpcb to
854 * account the size needed to set this TCP option.
856 if (tp->t_flags & TF_SIGNATURE)
857 to.to_flags |= TOF_SIGNATURE;
858 #endif /* TCP_SIGNATURE */
860 /* Processing the options. */
861 hdrlen += optlen = tcp_addoptions(&to, opt);
863 * If we wanted a TFO option to be added, but it was unable
864 * to fit, ensure no data is sent.
866 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
867 !(to.to_flags & TOF_FASTOPEN))
872 * Adjust data length if insertion of options will
873 * bump the packet length beyond the t_maxseg length.
874 * Clear the FIN bit because we cut off the tail of
877 if (len + optlen + ipoptlen > tp->t_maxseg) {
882 u_int if_hw_tsomaxsegcount;
883 u_int if_hw_tsomaxsegsize;
888 /* extract TSO information */
889 if_hw_tsomax = tp->t_tsomax;
890 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
891 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
894 * Limit a TSO burst to prevent it from
895 * overflowing or exceeding the maximum length
896 * allowed by the network interface:
898 KASSERT(ipoptlen == 0,
899 ("%s: TSO can't do IP options", __func__));
902 * Check if we should limit by maximum payload
905 if (if_hw_tsomax != 0) {
906 /* compute maximum TSO length */
907 max_len = (if_hw_tsomax - hdrlen -
911 } else if (len > max_len) {
918 * Check if we should limit by maximum segment
921 if (if_hw_tsomaxsegcount != 0 &&
922 if_hw_tsomaxsegsize != 0) {
924 * Subtract one segment for the LINK
925 * and TCP/IP headers mbuf that will
926 * be prepended to this mbuf chain
927 * after the code in this section
928 * limits the number of mbufs in the
929 * chain to if_hw_tsomaxsegcount.
931 if_hw_tsomaxsegcount -= 1;
933 mb = sbsndmbuf(&so->so_snd, off, &moff);
935 while (mb != NULL && max_len < len) {
940 * Get length of mbuf fragment
941 * and how many hardware frags,
942 * rounded up, it would use:
944 mlen = (mb->m_len - moff);
945 frags = howmany(mlen,
946 if_hw_tsomaxsegsize);
948 /* Handle special case: Zero Length Mbuf */
953 * Check if the fragment limit
954 * will be reached or exceeded:
956 if (frags >= if_hw_tsomaxsegcount) {
958 if_hw_tsomaxsegcount *
959 if_hw_tsomaxsegsize);
963 if_hw_tsomaxsegcount -= frags;
969 } else if (len > max_len) {
976 * Prevent the last segment from being
977 * fractional unless the send sockbuf can be
980 max_len = (tp->t_maxseg - optlen);
981 if (((uint32_t)off + (uint32_t)len) <
982 sbavail(&so->so_snd)) {
983 moff = len % max_len;
991 * In case there are too many small fragments
994 if (len <= max_len) {
1001 * Send the FIN in a separate segment
1002 * after the bulk sending is done.
1003 * We don't trust the TSO implementations
1004 * to clear the FIN flag on all but the
1007 if (tp->t_flags & TF_NEEDFIN)
1011 len = tp->t_maxseg - optlen - ipoptlen;
1019 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1020 ("%s: len > IP_MAXPACKET", __func__));
1022 /*#ifdef DIAGNOSTIC*/
1024 if (max_linkhdr + hdrlen > MCLBYTES)
1026 if (max_linkhdr + hdrlen > MHLEN)
1028 panic("tcphdr too big");
1032 * This KASSERT is here to catch edge cases at a well defined place.
1033 * Before, those had triggered (random) panic conditions further down.
1035 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1038 * Grab a header mbuf, attaching a copy of data to
1039 * be transmitted, and initialize the header from
1040 * the template for sends on this connection.
1046 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1047 TCPSTAT_INC(tcps_sndprobe);
1048 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1049 tp->t_sndrexmitpack++;
1050 TCPSTAT_INC(tcps_sndrexmitpack);
1051 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1053 TCPSTAT_INC(tcps_sndpack);
1054 TCPSTAT_ADD(tcps_sndbyte, len);
1057 if (MHLEN < hdrlen + max_linkhdr)
1058 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1061 m = m_gethdr(M_NOWAIT, MT_DATA);
1064 SOCKBUF_UNLOCK(&so->so_snd);
1070 m->m_data += max_linkhdr;
1074 * Start the m_copy functions from the closest mbuf
1075 * to the offset in the socket buffer chain.
1077 mb = sbsndptr(&so->so_snd, off, len, &moff);
1079 if (len <= MHLEN - hdrlen - max_linkhdr) {
1080 m_copydata(mb, moff, len,
1081 mtod(m, caddr_t) + hdrlen);
1084 m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1085 if (m->m_next == NULL) {
1086 SOCKBUF_UNLOCK(&so->so_snd);
1095 * If we're sending everything we've got, set PUSH.
1096 * (This will keep happy those implementations which only
1097 * give data to the user when a buffer fills or
1100 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1103 SOCKBUF_UNLOCK(&so->so_snd);
1105 SOCKBUF_UNLOCK(&so->so_snd);
1106 if (tp->t_flags & TF_ACKNOW)
1107 TCPSTAT_INC(tcps_sndacks);
1108 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1109 TCPSTAT_INC(tcps_sndctrl);
1110 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1111 TCPSTAT_INC(tcps_sndurg);
1113 TCPSTAT_INC(tcps_sndwinup);
1115 m = m_gethdr(M_NOWAIT, MT_DATA);
1122 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1127 m->m_data += max_linkhdr;
1130 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1131 m->m_pkthdr.rcvif = (struct ifnet *)0;
1133 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1137 ip6 = mtod(m, struct ip6_hdr *);
1138 th = (struct tcphdr *)(ip6 + 1);
1139 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1143 ip = mtod(m, struct ip *);
1145 ipov = (struct ipovly *)ip;
1147 th = (struct tcphdr *)(ip + 1);
1148 tcpip_fillheaders(tp->t_inpcb, ip, th);
1152 * Fill in fields, remembering maximum advertised
1153 * window for use in delaying messages about window sizes.
1154 * If resending a FIN, be sure not to use a new sequence number.
1156 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1157 tp->snd_nxt == tp->snd_max)
1160 * If we are starting a connection, send ECN setup
1161 * SYN packet. If we are on a retransmit, we may
1162 * resend those bits a number of times as per
1165 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1166 if (tp->t_rxtshift >= 1) {
1167 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1168 flags |= TH_ECE|TH_CWR;
1170 flags |= TH_ECE|TH_CWR;
1173 if (tp->t_state == TCPS_ESTABLISHED &&
1174 (tp->t_flags & TF_ECN_PERMIT)) {
1176 * If the peer has ECN, mark data packets with
1177 * ECN capable transmission (ECT).
1178 * Ignore pure ack packets, retransmissions and window probes.
1180 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1181 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1184 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1187 ip->ip_tos |= IPTOS_ECN_ECT0;
1188 TCPSTAT_INC(tcps_ecn_ect0);
1192 * Reply with proper ECN notifications.
1194 if (tp->t_flags & TF_ECN_SND_CWR) {
1196 tp->t_flags &= ~TF_ECN_SND_CWR;
1198 if (tp->t_flags & TF_ECN_SND_ECE)
1203 * If we are doing retransmissions, then snd_nxt will
1204 * not reflect the first unsent octet. For ACK only
1205 * packets, we do not want the sequence number of the
1206 * retransmitted packet, we want the sequence number
1207 * of the next unsent octet. So, if there is no data
1208 * (and no SYN or FIN), use snd_max instead of snd_nxt
1209 * when filling in ti_seq. But if we are in persist
1210 * state, snd_max might reflect one byte beyond the
1211 * right edge of the window, so use snd_nxt in that
1212 * case, since we know we aren't doing a retransmission.
1213 * (retransmit and persist are mutually exclusive...)
1215 if (sack_rxmit == 0) {
1216 if (len || (flags & (TH_SYN|TH_FIN)) ||
1217 tcp_timer_active(tp, TT_PERSIST))
1218 th->th_seq = htonl(tp->snd_nxt);
1220 th->th_seq = htonl(tp->snd_max);
1222 th->th_seq = htonl(p->rxmit);
1224 tp->sackhint.sack_bytes_rexmit += len;
1226 th->th_ack = htonl(tp->rcv_nxt);
1228 bcopy(opt, th + 1, optlen);
1229 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1231 th->th_flags = flags;
1233 * Calculate receive window. Don't shrink window,
1234 * but avoid silly window syndrome.
1236 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1237 recwin < tp->t_maxseg)
1239 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1240 recwin < (tp->rcv_adv - tp->rcv_nxt))
1241 recwin = (tp->rcv_adv - tp->rcv_nxt);
1244 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1245 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1246 * case is handled in syncache.
1249 th->th_win = htons((u_short)
1250 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1252 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1255 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1256 * a 0 window. This may cause the remote transmitter to stall. This
1257 * flag tells soreceive() to disable delayed acknowledgements when
1258 * draining the buffer. This can occur if the receiver is attempting
1259 * to read more data than can be buffered prior to transmitting on
1262 if (th->th_win == 0) {
1264 tp->t_flags |= TF_RXWIN0SENT;
1266 tp->t_flags &= ~TF_RXWIN0SENT;
1267 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1268 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1269 th->th_flags |= TH_URG;
1272 * If no urgent pointer to send, then we pull
1273 * the urgent pointer to the left edge of the send window
1274 * so that it doesn't drift into the send window on sequence
1275 * number wraparound.
1277 tp->snd_up = tp->snd_una; /* drag it along */
1280 * Put TCP length in extended header, and then
1281 * checksum extended header and data.
1283 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1284 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1286 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1287 if (to.to_flags & TOF_SIGNATURE) {
1289 * Calculate MD5 signature and put it into the place
1290 * determined before.
1291 * NOTE: since TCP options buffer doesn't point into
1292 * mbuf's data, calculate offset and use it.
1294 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1295 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1297 * Do not send segment if the calculation of MD5
1298 * digest has failed.
1308 * There is no need to fill in ip6_plen right now.
1309 * It will be filled later by ip6_output.
1311 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1312 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1313 optlen + len, IPPROTO_TCP, 0);
1316 #if defined(INET6) && defined(INET)
1321 m->m_pkthdr.csum_flags = CSUM_TCP;
1322 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1323 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1325 /* IP version must be set here for ipv4/ipv6 checking later */
1326 KASSERT(ip->ip_v == IPVERSION,
1327 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1332 * Enable TSO and specify the size of the segments.
1333 * The TCP pseudo header checksum is always provided.
1336 KASSERT(len > tp->t_maxseg - optlen,
1337 ("%s: len <= tso_segsz", __func__));
1338 m->m_pkthdr.csum_flags |= CSUM_TSO;
1339 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1342 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1343 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1344 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1345 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1347 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1348 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1349 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1353 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1354 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1361 if (so->so_options & SO_DEBUG) {
1367 save = ipov->ih_len;
1368 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1370 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1374 ipov->ih_len = save;
1376 #endif /* TCPDEBUG */
1377 TCP_PROBE3(debug__output, tp, th, m);
1379 /* We're getting ready to send; log now. */
1380 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1384 * Fill in IP length and desired time to live and
1385 * send to IP level. There should be a better way
1386 * to handle ttl and tos; we could keep them in
1387 * the template, but need a way to checksum without them.
1390 * m->m_pkthdr.len should have been set before checksum calculation,
1391 * because in6_cksum() need it.
1396 * we separately set hoplimit for every segment, since the
1397 * user might want to change the value via setsockopt.
1398 * Also, desired default hop limit might be changed via
1399 * Neighbor Discovery.
1401 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1404 * Set the packet size here for the benefit of DTrace probes.
1405 * ip6_output() will set it properly; it's supposed to include
1406 * the option header lengths as well.
1408 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1410 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1411 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1413 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1415 if (tp->t_state == TCPS_SYN_SENT)
1416 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1418 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1421 /* Save packet, if requested. */
1422 tcp_pcap_add(th, m, &(tp->t_outpkts));
1425 /* TODO: IPv6 IP6TOS_ECT bit on */
1426 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1427 &tp->t_inpcb->inp_route6,
1428 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1429 NULL, NULL, tp->t_inpcb);
1431 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1432 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1435 #if defined(INET) && defined(INET6)
1440 ip->ip_len = htons(m->m_pkthdr.len);
1442 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1443 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1446 * If we do path MTU discovery, then we set DF on every packet.
1447 * This might not be the best thing to do according to RFC3390
1448 * Section 2. However the tcp hostcache migitates the problem
1449 * so it affects only the first tcp connection with a host.
1451 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1453 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1454 ip->ip_off |= htons(IP_DF);
1455 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1457 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1460 if (tp->t_state == TCPS_SYN_SENT)
1461 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1463 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1466 /* Save packet, if requested. */
1467 tcp_pcap_add(th, m, &(tp->t_outpkts));
1470 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1471 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1474 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1475 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1481 * In transmit state, time the transmission and arrange for
1482 * the retransmit. In persist state, just set snd_max.
1484 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1485 !tcp_timer_active(tp, TT_PERSIST)) {
1486 tcp_seq startseq = tp->snd_nxt;
1489 * Advance snd_nxt over sequence space of this segment.
1491 if (flags & (TH_SYN|TH_FIN)) {
1494 if (flags & TH_FIN) {
1496 tp->t_flags |= TF_SENTFIN;
1502 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1503 tp->snd_max = tp->snd_nxt;
1505 * Time this transmission if not a retransmission and
1506 * not currently timing anything.
1508 if (tp->t_rtttime == 0) {
1509 tp->t_rtttime = ticks;
1510 tp->t_rtseq = startseq;
1511 TCPSTAT_INC(tcps_segstimed);
1516 * Set retransmit timer if not currently set,
1517 * and not doing a pure ack or a keep-alive probe.
1518 * Initial value for retransmit timer is smoothed
1519 * round-trip time + 2 * round-trip time variance.
1520 * Initialize shift counter which is used for backoff
1521 * of retransmit time.
1524 if (!tcp_timer_active(tp, TT_REXMT) &&
1525 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1526 (tp->snd_nxt != tp->snd_una))) {
1527 if (tcp_timer_active(tp, TT_PERSIST)) {
1528 tcp_timer_activate(tp, TT_PERSIST, 0);
1531 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1532 } else if (len == 0 && sbavail(&so->so_snd) &&
1533 !tcp_timer_active(tp, TT_REXMT) &&
1534 !tcp_timer_active(tp, TT_PERSIST)) {
1536 * Avoid a situation where we do not set persist timer
1537 * after a zero window condition. For example:
1538 * 1) A -> B: packet with enough data to fill the window
1539 * 2) B -> A: ACK for #1 + new data (0 window
1541 * 3) A -> B: ACK for #2, 0 len packet
1543 * In this case, A will not activate the persist timer,
1544 * because it chose to send a packet. Unless tcp_output
1545 * is called for some other reason (delayed ack timer,
1546 * another input packet from B, socket syscall), A will
1547 * not send zero window probes.
1549 * So, if you send a 0-length packet, but there is data
1550 * in the socket buffer, and neither the rexmt or
1551 * persist timer is already set, then activate the
1559 * Persist case, update snd_max but since we are in
1560 * persist mode (no window) we do not update snd_nxt.
1565 if (flags & TH_FIN) {
1567 tp->t_flags |= TF_SENTFIN;
1569 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1570 tp->snd_max = tp->snd_nxt + xlen;
1574 /* Record the error. */
1575 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1576 error, 0, NULL, false);
1579 * We know that the packet was lost, so back out the
1580 * sequence number advance, if any.
1582 * If the error is EPERM the packet got blocked by the
1583 * local firewall. Normally we should terminate the
1584 * connection but the blocking may have been spurious
1585 * due to a firewall reconfiguration cycle. So we treat
1586 * it like a packet loss and let the retransmit timer and
1587 * timeouts do their work over time.
1588 * XXX: It is a POLA question whether calling tcp_drop right
1589 * away would be the really correct behavior instead.
1591 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1592 !tcp_timer_active(tp, TT_PERSIST)) &&
1593 ((flags & TH_SYN) == 0) &&
1597 tp->sackhint.sack_bytes_rexmit -= len;
1598 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1599 ("sackhint bytes rtx >= 0"));
1603 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1607 tp->t_softerror = error;
1610 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1611 tp->snd_cwnd = tp->t_maxseg;
1615 * For some reason the interface we used initially
1616 * to send segments changed to another or lowered
1618 * If TSO was active we either got an interface
1619 * without TSO capabilits or TSO was turned off.
1620 * If we obtained mtu from ip_output() then update
1624 tp->t_flags &= ~TF_TSO;
1626 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1634 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1635 tp->t_softerror = error;
1643 TCPSTAT_INC(tcps_sndtotal);
1646 * Data sent (as far as we can tell).
1647 * If this advertises a larger window than any other segment,
1648 * then remember the size of the advertised window.
1649 * Any pending ACK has now been sent.
1651 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1652 tp->rcv_adv = tp->rcv_nxt + recwin;
1653 tp->last_ack_sent = tp->rcv_nxt;
1654 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1655 if (tcp_timer_active(tp, TT_DELACK))
1656 tcp_timer_activate(tp, TT_DELACK, 0);
1659 * This completely breaks TCP if newreno is turned on. What happens
1660 * is that if delayed-acks are turned on on the receiver, this code
1661 * on the transmitter effectively destroys the TCP window, forcing
1662 * it to four packets (1.5Kx4 = 6K window).
1664 if (sendalot && --maxburst)
1673 tcp_setpersist(struct tcpcb *tp)
1675 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1678 tp->t_flags &= ~TF_PREVVALID;
1679 if (tcp_timer_active(tp, TT_REXMT))
1680 panic("tcp_setpersist: retransmit pending");
1682 * Start/restart persistence timer.
1684 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1685 tcp_persmin, tcp_persmax);
1686 tcp_timer_activate(tp, TT_PERSIST, tt);
1687 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1692 * Insert TCP options according to the supplied parameters to the place
1693 * optp in a consistent way. Can handle unaligned destinations.
1695 * The order of the option processing is crucial for optimal packing and
1696 * alignment for the scarce option space.
1698 * The optimal order for a SYN/SYN-ACK segment is:
1699 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1700 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1702 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1703 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1704 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1705 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1706 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1709 tcp_addoptions(struct tcpopt *to, u_char *optp)
1711 u_int32_t mask, optlen = 0;
1713 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1714 if ((to->to_flags & mask) != mask)
1716 if (optlen == TCP_MAXOLEN)
1718 switch (to->to_flags & mask) {
1720 while (optlen % 4) {
1721 optlen += TCPOLEN_NOP;
1722 *optp++ = TCPOPT_NOP;
1724 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1726 optlen += TCPOLEN_MAXSEG;
1727 *optp++ = TCPOPT_MAXSEG;
1728 *optp++ = TCPOLEN_MAXSEG;
1729 to->to_mss = htons(to->to_mss);
1730 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1731 optp += sizeof(to->to_mss);
1734 while (!optlen || optlen % 2 != 1) {
1735 optlen += TCPOLEN_NOP;
1736 *optp++ = TCPOPT_NOP;
1738 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1740 optlen += TCPOLEN_WINDOW;
1741 *optp++ = TCPOPT_WINDOW;
1742 *optp++ = TCPOLEN_WINDOW;
1743 *optp++ = to->to_wscale;
1746 while (optlen % 2) {
1747 optlen += TCPOLEN_NOP;
1748 *optp++ = TCPOPT_NOP;
1750 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1752 optlen += TCPOLEN_SACK_PERMITTED;
1753 *optp++ = TCPOPT_SACK_PERMITTED;
1754 *optp++ = TCPOLEN_SACK_PERMITTED;
1757 while (!optlen || optlen % 4 != 2) {
1758 optlen += TCPOLEN_NOP;
1759 *optp++ = TCPOPT_NOP;
1761 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1763 optlen += TCPOLEN_TIMESTAMP;
1764 *optp++ = TCPOPT_TIMESTAMP;
1765 *optp++ = TCPOLEN_TIMESTAMP;
1766 to->to_tsval = htonl(to->to_tsval);
1767 to->to_tsecr = htonl(to->to_tsecr);
1768 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1769 optp += sizeof(to->to_tsval);
1770 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1771 optp += sizeof(to->to_tsecr);
1775 int siglen = TCPOLEN_SIGNATURE - 2;
1777 while (!optlen || optlen % 4 != 2) {
1778 optlen += TCPOLEN_NOP;
1779 *optp++ = TCPOPT_NOP;
1781 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1782 to->to_flags &= ~TOF_SIGNATURE;
1785 optlen += TCPOLEN_SIGNATURE;
1786 *optp++ = TCPOPT_SIGNATURE;
1787 *optp++ = TCPOLEN_SIGNATURE;
1788 to->to_signature = optp;
1796 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1799 while (!optlen || optlen % 4 != 2) {
1800 optlen += TCPOLEN_NOP;
1801 *optp++ = TCPOPT_NOP;
1803 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1805 optlen += TCPOLEN_SACKHDR;
1806 *optp++ = TCPOPT_SACK;
1807 sackblks = min(to->to_nsacks,
1808 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1809 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1810 while (sackblks--) {
1811 sack_seq = htonl(sack->start);
1812 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1813 optp += sizeof(sack_seq);
1814 sack_seq = htonl(sack->end);
1815 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1816 optp += sizeof(sack_seq);
1817 optlen += TCPOLEN_SACK;
1820 TCPSTAT_INC(tcps_sack_send_blocks);
1827 /* XXX is there any point to aligning this option? */
1828 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1829 if (TCP_MAXOLEN - optlen < total_len) {
1830 to->to_flags &= ~TOF_FASTOPEN;
1833 *optp++ = TCPOPT_FAST_OPEN;
1834 *optp++ = total_len;
1835 if (to->to_tfo_len > 0) {
1836 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1837 optp += to->to_tfo_len;
1839 optlen += total_len;
1843 panic("%s: unknown TCP option type", __func__);
1848 /* Terminate and pad TCP options to a 4 byte boundary. */
1850 optlen += TCPOLEN_EOL;
1851 *optp++ = TCPOPT_EOL;
1854 * According to RFC 793 (STD0007):
1855 * "The content of the header beyond the End-of-Option option
1856 * must be header padding (i.e., zero)."
1857 * and later: "The padding is composed of zeros."
1859 while (optlen % 4) {
1860 optlen += TCPOLEN_PAD;
1861 *optp++ = TCPOPT_PAD;
1864 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1869 * This is a copy of m_copym(), taking the TSO segment size/limit
1870 * constraints into account, and advancing the sndptr as it goes.
1873 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1874 int32_t seglimit, int32_t segsize, struct sockbuf *sb)
1876 struct mbuf *n, **np;
1879 int32_t len = *plen;
1883 uint32_t mlen, frags;
1887 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1888 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1889 if (off == 0 && m->m_flags & M_PKTHDR)
1894 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1898 if ((sb) && (m == sb->sb_sndptr)) {
1899 sb->sb_sndptroff += m->m_len;
1900 sb->sb_sndptr = m->m_next;
1909 KASSERT(len == M_COPYALL,
1910 ("tcp_m_copym, length > size of mbuf chain"));
1912 if (pkthdrlen != NULL)
1913 *pkthdrlen = len_cp;
1916 mlen = min(len, m->m_len - off);
1919 * For M_NOMAP mbufs, add 3 segments
1920 * + 1 in case we are crossing page boundaries
1921 * + 2 in case the TLS hdr/trailer are used
1922 * It is cheaper to just add the segments
1923 * than it is to take the cache miss to look
1924 * at the mbuf ext_pgs state in detail.
1926 if (m->m_flags & M_NOMAP) {
1927 fragsize = min(segsize, PAGE_SIZE);
1934 /* Break if we really can't fit anymore. */
1935 if ((frags + 1) >= seglimit) {
1937 if (pkthdrlen != NULL)
1938 *pkthdrlen = len_cp;
1943 * Reduce size if you can't copy the whole
1944 * mbuf. If we can't copy the whole mbuf, also
1945 * adjust len so the loop will end after this
1948 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1949 mlen = (seglimit - frags - 1) * fragsize;
1951 *plen = len_cp + len;
1952 if (pkthdrlen != NULL)
1955 frags += howmany(mlen, fragsize);
1959 KASSERT(seglimit > 0,
1960 ("%s: seglimit went too low", __func__));
1963 n = m_gethdr(M_NOWAIT, m->m_type);
1965 n = m_get(M_NOWAIT, m->m_type);
1970 if (!m_dup_pkthdr(n, m, M_NOWAIT))
1972 if (len == M_COPYALL)
1973 n->m_pkthdr.len -= off0;
1975 n->m_pkthdr.len = len;
1976 pkthdrlen = &n->m_pkthdr.len;
1981 if (m->m_flags & M_EXT) {
1982 n->m_data = m->m_data + off;
1985 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1988 if (sb && (sb->sb_sndptr == m) &&
1989 ((n->m_len + off) >= m->m_len) && m->m_next) {
1990 sb->sb_sndptroff += m->m_len;
1991 sb->sb_sndptr = m->m_next;
1994 if (len != M_COPYALL) {
2007 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2011 * Automatic sizing of send socket buffer. Often the send buffer
2012 * size is not optimally adjusted to the actual network conditions
2013 * at hand (delay bandwidth product). Setting the buffer size too
2014 * small limits throughput on links with high bandwidth and high
2015 * delay (eg. trans-continental/oceanic links). Setting the
2016 * buffer size too big consumes too much real kernel memory,
2017 * especially with many connections on busy servers.
2019 * The criteria to step up the send buffer one notch are:
2020 * 1. receive window of remote host is larger than send buffer
2021 * (with a fudge factor of 5/4th);
2022 * 2. send buffer is filled to 7/8th with data (so we actually
2023 * have data to make use of it);
2024 * 3. send buffer fill has not hit maximal automatic size;
2025 * 4. our send window (slow start and cogestion controlled) is
2026 * larger than sent but unacknowledged data in send buffer.
2028 * The remote host receive window scaling factor may limit the
2029 * growing of the send buffer before it reaches its allowed
2032 * It scales directly with slow start or congestion window
2033 * and does at most one step per received ACK. This fast
2034 * scaling has the drawback of growing the send buffer beyond
2035 * what is strictly necessary to make full use of a given
2036 * delay*bandwidth product. However testing has shown this not
2037 * to be much of an problem. At worst we are trading wasting
2038 * of available bandwidth (the non-use of it) for wasting some
2039 * socket buffer memory.
2041 * TODO: Shrink send buffer during idle periods together
2042 * with congestion window. Requires another timer. Has to
2043 * wait for upcoming tcp timer rewrite.
2045 * XXXGL: should there be used sbused() or sbavail()?
2047 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2050 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2051 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2052 sbused(&so->so_snd) >=
2053 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2054 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2055 sendwin >= (sbused(&so->so_snd) -
2056 (tp->snd_nxt - tp->snd_una))) {
2057 if (!sbreserve_locked(&so->so_snd,
2058 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2059 V_tcp_autosndbuf_max), so, curthread))
2060 so->so_snd.sb_flags &= ~SB_AUTOSIZE;