2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
46 #include <sys/hhook.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
59 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
75 #include <netinet/tcp_fastopen.h>
77 #include <netinet/tcp.h>
79 #include <netinet/tcp_fsm.h>
80 #include <netinet/tcp_seq.h>
81 #include <netinet/tcp_timer.h>
82 #include <netinet/tcp_var.h>
83 #include <netinet/tcpip.h>
84 #include <netinet/cc/cc.h>
86 #include <netinet/tcp_pcap.h>
89 #include <netinet/tcp_debug.h>
92 #include <netinet/tcp_offload.h>
95 #include <netipsec/ipsec_support.h>
97 #include <machine/in_cksum.h>
99 #include <security/mac/mac_framework.h>
101 VNET_DEFINE(int, path_mtu_discovery) = 1;
102 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
103 &VNET_NAME(path_mtu_discovery), 1,
104 "Enable Path MTU Discovery");
106 VNET_DEFINE(int, tcp_do_tso) = 1;
107 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
108 &VNET_NAME(tcp_do_tso), 0,
109 "Enable TCP Segmentation Offload");
111 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
112 #define V_tcp_sendspace VNET(tcp_sendspace)
113 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
114 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
116 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
117 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
118 &VNET_NAME(tcp_do_autosndbuf), 0,
119 "Enable automatic send buffer sizing");
121 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
122 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
123 &VNET_NAME(tcp_autosndbuf_inc), 0,
124 "Incrementor step size of automatic send buffer");
126 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
127 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
128 &VNET_NAME(tcp_autosndbuf_max), 0,
129 "Max size of automatic send buffer");
131 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
132 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
133 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
134 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
135 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
138 * Make sure that either retransmit or persist timer is set for SYN, FIN and
141 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
142 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
143 tcp_timer_active((tp), TT_REXMT) || \
144 tcp_timer_active((tp), TT_PERSIST), \
145 ("neither rexmt nor persist timer is set"))
148 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
149 struct tcphdr *th, struct tcpopt *to,
150 uint32_t len, int tso);
152 static void inline cc_after_idle(struct tcpcb *tp);
156 * Wrapper for the TCP established output helper hook.
159 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
160 struct tcpopt *to, uint32_t len, int tso)
162 struct tcp_hhook_data hhook_data;
164 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
168 hhook_data.len = len;
169 hhook_data.tso = tso;
171 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
178 * CC wrapper hook functions
181 cc_after_idle(struct tcpcb *tp)
183 INP_WLOCK_ASSERT(tp->t_inpcb);
185 if (CC_ALGO(tp)->after_idle != NULL)
186 CC_ALGO(tp)->after_idle(tp->ccv);
190 * Tcp output routine: figure out what should be sent and send it.
193 tcp_output(struct tcpcb *tp)
195 struct socket *so = tp->t_inpcb->inp_socket;
197 uint32_t recwin, sendwin;
198 int off, flags, error = 0; /* Keep compiler happy */
200 struct ip *ip = NULL;
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;
211 int sack_rxmit, sack_bytes_rxmt;
216 int maxburst = TCP_MAXBURST;
219 struct ip6_hdr *ip6 = NULL;
222 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
225 INP_WLOCK_ASSERT(tp->t_inpcb);
228 if (tp->t_flags & TF_TOE)
229 return (tcp_offload_output(tp));
234 * For TFO connections in SYN_RECEIVED, only allow the initial
235 * SYN|ACK and those sent by the retransmit timer.
237 if (IS_FASTOPEN(tp->t_flags) &&
238 (tp->t_state == TCPS_SYN_RECEIVED) &&
239 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
240 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
244 * Determine length of data that should be transmitted,
245 * and flags that will be used.
246 * If there is some data or critical controls (SYN, RST)
247 * to send, then transmit; otherwise, investigate further.
249 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
250 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
252 tp->t_flags &= ~TF_LASTIDLE;
254 if (tp->t_flags & TF_MORETOCOME) {
255 tp->t_flags |= TF_LASTIDLE;
261 * If we've recently taken a timeout, snd_max will be greater than
262 * snd_nxt. There may be SACK information that allows us to avoid
263 * resending already delivered data. Adjust snd_nxt accordingly.
265 if ((tp->t_flags & TF_SACK_PERMIT) &&
266 SEQ_LT(tp->snd_nxt, tp->snd_max))
271 off = tp->snd_nxt - tp->snd_una;
272 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
274 flags = tcp_outflags[tp->t_state];
276 * Send any SACK-generated retransmissions. If we're explicitly trying
277 * to send out new data (when sendalot is 1), bypass this function.
278 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
279 * we're replacing a (future) new transmission with a retransmission
280 * now, and we previously incremented snd_cwnd in tcp_input().
283 * Still in sack recovery , reset rxmit flag to zero.
289 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
290 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
294 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
295 /* Do not retransmit SACK segments beyond snd_recover */
296 if (SEQ_GT(p->end, tp->snd_recover)) {
298 * (At least) part of sack hole extends beyond
299 * snd_recover. Check to see if we can rexmit data
302 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
304 * Can't rexmit any more data for this hole.
305 * That data will be rexmitted in the next
306 * sack recovery episode, when snd_recover
307 * moves past p->rxmit.
310 goto after_sack_rexmit;
312 /* Can rexmit part of the current hole */
313 len = ((int32_t)ulmin(cwin,
314 tp->snd_recover - p->rxmit));
316 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
317 off = p->rxmit - tp->snd_una;
318 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
323 TCPSTAT_INC(tcps_sack_rexmits);
324 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
325 min(len, tp->t_maxseg));
330 * Get standard flags, and add SYN or FIN if requested by 'hidden'
333 if (tp->t_flags & TF_NEEDFIN)
335 if (tp->t_flags & TF_NEEDSYN)
338 SOCKBUF_LOCK(&so->so_snd);
340 * If in persist timeout with window of 0, send 1 byte.
341 * Otherwise, if window is small but nonzero
342 * and timer expired, we will send what we can
343 * and go to transmit state.
345 if (tp->t_flags & TF_FORCEDATA) {
348 * If we still have some data to send, then
349 * clear the FIN bit. Usually this would
350 * happen below when it realizes that we
351 * aren't sending all the data. However,
352 * if we have exactly 1 byte of unsent data,
353 * then it won't clear the FIN bit below,
354 * and if we are in persist state, we wind
355 * up sending the packet without recording
356 * that we sent the FIN bit.
358 * We can't just blindly clear the FIN bit,
359 * because if we don't have any more data
360 * to send then the probe will be the FIN
363 if (off < sbused(&so->so_snd))
367 tcp_timer_activate(tp, TT_PERSIST, 0);
373 * If snd_nxt == snd_max and we have transmitted a FIN, the
374 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
375 * a negative length. This can also occur when TCP opens up
376 * its congestion window while receiving additional duplicate
377 * acks after fast-retransmit because TCP will reset snd_nxt
378 * to snd_max after the fast-retransmit.
380 * In the normal retransmit-FIN-only case, however, snd_nxt will
381 * be set to snd_una, the offset will be 0, and the length may
384 * If sack_rxmit is true we are retransmitting from the scoreboard
385 * in which case len is already set.
387 if (sack_rxmit == 0) {
388 if (sack_bytes_rxmt == 0)
389 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
395 * We are inside of a SACK recovery episode and are
396 * sending new data, having retransmitted all the
397 * data possible in the scoreboard.
399 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
402 * Don't remove this (len > 0) check !
403 * We explicitly check for len > 0 here (although it
404 * isn't really necessary), to work around a gcc
405 * optimization issue - to force gcc to compute
406 * len above. Without this check, the computation
407 * of len is bungled by the optimizer.
410 cwin = tp->snd_cwnd -
411 (tp->snd_nxt - tp->sack_newdata) -
415 len = imin(len, cwin);
421 * Lop off SYN bit if it has already been sent. However, if this
422 * is SYN-SENT state and if segment contains data and if we don't
423 * know that foreign host supports TAO, suppress sending segment.
425 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
426 if (tp->t_state != TCPS_SYN_RECEIVED)
430 * When sending additional segments following a TFO SYN|ACK,
431 * do not include the SYN bit.
433 if (IS_FASTOPEN(tp->t_flags) &&
434 (tp->t_state == TCPS_SYN_RECEIVED))
441 * Be careful not to send data and/or FIN on SYN segments.
442 * This measure is needed to prevent interoperability problems
443 * with not fully conformant TCP implementations.
445 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
452 * When retransmitting SYN|ACK on a passively-created TFO socket,
453 * don't include data, as the presence of data may have caused the
454 * original SYN|ACK to have been dropped by a middlebox.
456 if (IS_FASTOPEN(tp->t_flags) &&
457 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
463 * If FIN has been sent but not acked,
464 * but we haven't been called to retransmit,
465 * len will be < 0. Otherwise, window shrank
466 * after we sent into it. If window shrank to 0,
467 * cancel pending retransmit, pull snd_nxt back
468 * to (closed) window, and set the persist timer
469 * if it isn't already going. If the window didn't
470 * close completely, just wait for an ACK.
472 * We also do a general check here to ensure that
473 * we will set the persist timer when we have data
474 * to send, but a 0-byte window. This makes sure
475 * the persist timer is set even if the packet
476 * hits one of the "goto send" lines below.
479 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
480 (off < (int) sbavail(&so->so_snd))) {
481 tcp_timer_activate(tp, TT_REXMT, 0);
483 tp->snd_nxt = tp->snd_una;
484 if (!tcp_timer_active(tp, TT_PERSIST))
489 /* len will be >= 0 after this point. */
490 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
492 tcp_sndbuf_autoscale(tp, so, sendwin);
495 * Decide if we can use TCP Segmentation Offloading (if supported by
498 * TSO may only be used if we are in a pure bulk sending state. The
499 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
500 * IP options prevent using TSO. With TSO the TCP header is the same
501 * (except for the sequence number) for all generated packets. This
502 * makes it impossible to transmit any options which vary per generated
505 * IPv4 handling has a clear separation of ip options and ip header
506 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
507 * the right thing below to provide length of just ip options and thus
508 * checking for ipoptlen is enough to decide if ip options are present.
510 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
512 * Pre-calculate here as we save another lookup into the darknesses
513 * of IPsec that way and can actually decide if TSO is ok.
516 if (isipv6 && IPSEC_ENABLED(ipv6))
517 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
523 if (IPSEC_ENABLED(ipv4))
524 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
529 ipoptlen = ip6_optlen(tp->t_inpcb);
532 if (tp->t_inpcb->inp_options)
533 ipoptlen = tp->t_inpcb->inp_options->m_len -
534 offsetof(struct ipoption, ipopt_list);
537 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
538 ipoptlen += ipsec_optlen;
541 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
542 ((tp->t_flags & TF_SIGNATURE) == 0) &&
543 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
548 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
551 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
552 sbused(&so->so_snd)))
556 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
557 (long)TCP_MAXWIN << tp->rcv_scale);
560 * Sender silly window avoidance. We transmit under the following
561 * conditions when len is non-zero:
563 * - We have a full segment (or more with TSO)
564 * - This is the last buffer in a write()/send() and we are
565 * either idle or running NODELAY
566 * - we've timed out (e.g. persist timer)
567 * - we have more then 1/2 the maximum send window's worth of
568 * data (receiver may be limited the window size)
569 * - we need to retransmit
572 if (len >= tp->t_maxseg)
575 * NOTE! on localhost connections an 'ack' from the remote
576 * end may occur synchronously with the output and cause
577 * us to flush a buffer queued with moretocome. XXX
579 * note: the len + off check is almost certainly unnecessary.
581 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
582 (idle || (tp->t_flags & TF_NODELAY)) &&
583 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
584 (tp->t_flags & TF_NOPUSH) == 0) {
587 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
589 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
591 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
598 * Sending of standalone window updates.
600 * Window updates are important when we close our window due to a
601 * full socket buffer and are opening it again after the application
602 * reads data from it. Once the window has opened again and the
603 * remote end starts to send again the ACK clock takes over and
604 * provides the most current window information.
606 * We must avoid the silly window syndrome whereas every read
607 * from the receive buffer, no matter how small, causes a window
608 * update to be sent. We also should avoid sending a flurry of
609 * window updates when the socket buffer had queued a lot of data
610 * and the application is doing small reads.
612 * Prevent a flurry of pointless window updates by only sending
613 * an update when we can increase the advertized window by more
614 * than 1/4th of the socket buffer capacity. When the buffer is
615 * getting full or is very small be more aggressive and send an
616 * update whenever we can increase by two mss sized segments.
617 * In all other situations the ACK's to new incoming data will
618 * carry further window increases.
620 * Don't send an independent window update if a delayed
621 * ACK is pending (it will get piggy-backed on it) or the
622 * remote side already has done a half-close and won't send
623 * more data. Skip this if the connection is in T/TCP
626 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
627 !(tp->t_flags & TF_DELACK) &&
628 !TCPS_HAVERCVDFIN(tp->t_state)) {
630 * "adv" is the amount we could increase the window,
631 * taking into account that we are limited by
632 * TCP_MAXWIN << tp->rcv_scale.
638 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
639 oldwin = (tp->rcv_adv - tp->rcv_nxt);
645 * If the new window size ends up being the same as or less
646 * than the old size when it is scaled, then don't force
649 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
652 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
653 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
654 recwin <= (so->so_rcv.sb_hiwat / 8) ||
655 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
657 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
663 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
664 * is also a catch-all for the retransmit timer timeout case.
666 if (tp->t_flags & TF_ACKNOW)
668 if ((flags & TH_RST) ||
669 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
671 if (SEQ_GT(tp->snd_up, tp->snd_una))
674 * If our state indicates that FIN should be sent
675 * and we have not yet done so, then we need to send.
677 if (flags & TH_FIN &&
678 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
681 * In SACK, it is possible for tcp_output to fail to send a segment
682 * after the retransmission timer has been turned off. Make sure
683 * that the retransmission timer is set.
685 if ((tp->t_flags & TF_SACK_PERMIT) &&
686 SEQ_GT(tp->snd_max, tp->snd_una) &&
687 !tcp_timer_active(tp, TT_REXMT) &&
688 !tcp_timer_active(tp, TT_PERSIST)) {
689 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
693 * TCP window updates are not reliable, rather a polling protocol
694 * using ``persist'' packets is used to insure receipt of window
695 * updates. The three ``states'' for the output side are:
696 * idle not doing retransmits or persists
697 * persisting to move a small or zero window
698 * (re)transmitting and thereby not persisting
700 * tcp_timer_active(tp, TT_PERSIST)
701 * is true when we are in persist state.
702 * (tp->t_flags & TF_FORCEDATA)
703 * is set when we are called to send a persist packet.
704 * tcp_timer_active(tp, TT_REXMT)
705 * is set when we are retransmitting
706 * The output side is idle when both timers are zero.
708 * If send window is too small, there is data to transmit, and no
709 * retransmit or persist is pending, then go to persist state.
710 * If nothing happens soon, send when timer expires:
711 * if window is nonzero, transmit what we can,
712 * otherwise force out a byte.
714 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
715 !tcp_timer_active(tp, TT_PERSIST)) {
721 * No reason to send a segment, just return.
724 SOCKBUF_UNLOCK(&so->so_snd);
728 SOCKBUF_LOCK_ASSERT(&so->so_snd);
730 if (len >= tp->t_maxseg)
731 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
733 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
736 * Before ESTABLISHED, force sending of initial options
737 * unless TCP set not to do any options.
738 * NOTE: we assume that the IP/TCP header plus TCP options
739 * always fit in a single mbuf, leaving room for a maximum
741 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
746 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
749 hdrlen = sizeof (struct tcpiphdr);
752 * Compute options for segment.
753 * We only have to care about SYN and established connection
754 * segments. Options for SYN-ACK segments are handled in TCP
758 if ((tp->t_flags & TF_NOOPT) == 0) {
759 /* Maximum segment size. */
760 if (flags & TH_SYN) {
761 tp->snd_nxt = tp->iss;
762 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
763 to.to_flags |= TOF_MSS;
766 * Only include the TFO option on the first
767 * transmission of the SYN|ACK on a
768 * passively-created TFO socket, as the presence of
769 * the TFO option may have caused the original
770 * SYN|ACK to have been dropped by a middlebox.
772 if (IS_FASTOPEN(tp->t_flags) &&
773 (tp->t_state == TCPS_SYN_RECEIVED) &&
774 (tp->t_rxtshift == 0)) {
775 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
776 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
777 to.to_flags |= TOF_FASTOPEN;
781 /* Window scaling. */
782 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
783 to.to_wscale = tp->request_r_scale;
784 to.to_flags |= TOF_SCALE;
787 if ((tp->t_flags & TF_RCVD_TSTMP) ||
788 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
789 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
790 to.to_tsecr = tp->ts_recent;
791 to.to_flags |= TOF_TS;
794 /* Set receive buffer autosizing timestamp. */
795 if (tp->rfbuf_ts == 0 &&
796 (so->so_rcv.sb_flags & SB_AUTOSIZE))
797 tp->rfbuf_ts = tcp_ts_getticks();
799 /* Selective ACK's. */
800 if (tp->t_flags & TF_SACK_PERMIT) {
802 to.to_flags |= TOF_SACKPERM;
803 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
804 (tp->t_flags & TF_SACK_PERMIT) &&
805 tp->rcv_numsacks > 0) {
806 to.to_flags |= TOF_SACK;
807 to.to_nsacks = tp->rcv_numsacks;
808 to.to_sacks = (u_char *)tp->sackblks;
811 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
812 /* TCP-MD5 (RFC2385). */
814 * Check that TCP_MD5SIG is enabled in tcpcb to
815 * account the size needed to set this TCP option.
817 if (tp->t_flags & TF_SIGNATURE)
818 to.to_flags |= TOF_SIGNATURE;
819 #endif /* TCP_SIGNATURE */
821 /* Processing the options. */
822 hdrlen += optlen = tcp_addoptions(&to, opt);
826 * Adjust data length if insertion of options will
827 * bump the packet length beyond the t_maxseg length.
828 * Clear the FIN bit because we cut off the tail of
831 if (len + optlen + ipoptlen > tp->t_maxseg) {
836 u_int if_hw_tsomaxsegcount;
837 u_int if_hw_tsomaxsegsize;
842 /* extract TSO information */
843 if_hw_tsomax = tp->t_tsomax;
844 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
845 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
848 * Limit a TSO burst to prevent it from
849 * overflowing or exceeding the maximum length
850 * allowed by the network interface:
852 KASSERT(ipoptlen == 0,
853 ("%s: TSO can't do IP options", __func__));
856 * Check if we should limit by maximum payload
859 if (if_hw_tsomax != 0) {
860 /* compute maximum TSO length */
861 max_len = (if_hw_tsomax - hdrlen -
865 } else if (len > max_len) {
872 * Check if we should limit by maximum segment
875 if (if_hw_tsomaxsegcount != 0 &&
876 if_hw_tsomaxsegsize != 0) {
878 * Subtract one segment for the LINK
879 * and TCP/IP headers mbuf that will
880 * be prepended to this mbuf chain
881 * after the code in this section
882 * limits the number of mbufs in the
883 * chain to if_hw_tsomaxsegcount.
885 if_hw_tsomaxsegcount -= 1;
887 mb = sbsndmbuf(&so->so_snd, off, &moff);
889 while (mb != NULL && max_len < len) {
894 * Get length of mbuf fragment
895 * and how many hardware frags,
896 * rounded up, it would use:
898 mlen = (mb->m_len - moff);
899 frags = howmany(mlen,
900 if_hw_tsomaxsegsize);
902 /* Handle special case: Zero Length Mbuf */
907 * Check if the fragment limit
908 * will be reached or exceeded:
910 if (frags >= if_hw_tsomaxsegcount) {
912 if_hw_tsomaxsegcount *
913 if_hw_tsomaxsegsize);
917 if_hw_tsomaxsegcount -= frags;
923 } else if (len > max_len) {
930 * Prevent the last segment from being
931 * fractional unless the send sockbuf can be
934 max_len = (tp->t_maxseg - optlen);
935 if (((uint32_t)off + (uint32_t)len) <
936 sbavail(&so->so_snd)) {
937 moff = len % max_len;
945 * In case there are too many small fragments
948 if (len <= max_len) {
955 * Send the FIN in a separate segment
956 * after the bulk sending is done.
957 * We don't trust the TSO implementations
958 * to clear the FIN flag on all but the
961 if (tp->t_flags & TF_NEEDFIN)
965 len = tp->t_maxseg - optlen - ipoptlen;
971 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
972 ("%s: len > IP_MAXPACKET", __func__));
974 /*#ifdef DIAGNOSTIC*/
976 if (max_linkhdr + hdrlen > MCLBYTES)
978 if (max_linkhdr + hdrlen > MHLEN)
980 panic("tcphdr too big");
984 * This KASSERT is here to catch edge cases at a well defined place.
985 * Before, those had triggered (random) panic conditions further down.
987 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
990 * Grab a header mbuf, attaching a copy of data to
991 * be transmitted, and initialize the header from
992 * the template for sends on this connection.
998 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
999 TCPSTAT_INC(tcps_sndprobe);
1000 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1001 tp->t_sndrexmitpack++;
1002 TCPSTAT_INC(tcps_sndrexmitpack);
1003 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1005 TCPSTAT_INC(tcps_sndpack);
1006 TCPSTAT_ADD(tcps_sndbyte, len);
1009 if (MHLEN < hdrlen + max_linkhdr)
1010 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1013 m = m_gethdr(M_NOWAIT, MT_DATA);
1016 SOCKBUF_UNLOCK(&so->so_snd);
1022 m->m_data += max_linkhdr;
1026 * Start the m_copy functions from the closest mbuf
1027 * to the offset in the socket buffer chain.
1029 mb = sbsndptr(&so->so_snd, off, len, &moff);
1031 if (len <= MHLEN - hdrlen - max_linkhdr) {
1032 m_copydata(mb, moff, len,
1033 mtod(m, caddr_t) + hdrlen);
1036 m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1037 if (m->m_next == NULL) {
1038 SOCKBUF_UNLOCK(&so->so_snd);
1047 * If we're sending everything we've got, set PUSH.
1048 * (This will keep happy those implementations which only
1049 * give data to the user when a buffer fills or
1052 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1055 SOCKBUF_UNLOCK(&so->so_snd);
1057 SOCKBUF_UNLOCK(&so->so_snd);
1058 if (tp->t_flags & TF_ACKNOW)
1059 TCPSTAT_INC(tcps_sndacks);
1060 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1061 TCPSTAT_INC(tcps_sndctrl);
1062 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1063 TCPSTAT_INC(tcps_sndurg);
1065 TCPSTAT_INC(tcps_sndwinup);
1067 m = m_gethdr(M_NOWAIT, MT_DATA);
1074 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1079 m->m_data += max_linkhdr;
1082 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1083 m->m_pkthdr.rcvif = (struct ifnet *)0;
1085 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1089 ip6 = mtod(m, struct ip6_hdr *);
1090 th = (struct tcphdr *)(ip6 + 1);
1091 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1095 ip = mtod(m, struct ip *);
1097 ipov = (struct ipovly *)ip;
1099 th = (struct tcphdr *)(ip + 1);
1100 tcpip_fillheaders(tp->t_inpcb, ip, th);
1104 * Fill in fields, remembering maximum advertised
1105 * window for use in delaying messages about window sizes.
1106 * If resending a FIN, be sure not to use a new sequence number.
1108 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1109 tp->snd_nxt == tp->snd_max)
1112 * If we are starting a connection, send ECN setup
1113 * SYN packet. If we are on a retransmit, we may
1114 * resend those bits a number of times as per
1117 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1118 if (tp->t_rxtshift >= 1) {
1119 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1120 flags |= TH_ECE|TH_CWR;
1122 flags |= TH_ECE|TH_CWR;
1125 if (tp->t_state == TCPS_ESTABLISHED &&
1126 (tp->t_flags & TF_ECN_PERMIT)) {
1128 * If the peer has ECN, mark data packets with
1129 * ECN capable transmission (ECT).
1130 * Ignore pure ack packets, retransmissions and window probes.
1132 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1133 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1136 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1139 ip->ip_tos |= IPTOS_ECN_ECT0;
1140 TCPSTAT_INC(tcps_ecn_ect0);
1144 * Reply with proper ECN notifications.
1146 if (tp->t_flags & TF_ECN_SND_CWR) {
1148 tp->t_flags &= ~TF_ECN_SND_CWR;
1150 if (tp->t_flags & TF_ECN_SND_ECE)
1155 * If we are doing retransmissions, then snd_nxt will
1156 * not reflect the first unsent octet. For ACK only
1157 * packets, we do not want the sequence number of the
1158 * retransmitted packet, we want the sequence number
1159 * of the next unsent octet. So, if there is no data
1160 * (and no SYN or FIN), use snd_max instead of snd_nxt
1161 * when filling in ti_seq. But if we are in persist
1162 * state, snd_max might reflect one byte beyond the
1163 * right edge of the window, so use snd_nxt in that
1164 * case, since we know we aren't doing a retransmission.
1165 * (retransmit and persist are mutually exclusive...)
1167 if (sack_rxmit == 0) {
1168 if (len || (flags & (TH_SYN|TH_FIN)) ||
1169 tcp_timer_active(tp, TT_PERSIST))
1170 th->th_seq = htonl(tp->snd_nxt);
1172 th->th_seq = htonl(tp->snd_max);
1174 th->th_seq = htonl(p->rxmit);
1176 tp->sackhint.sack_bytes_rexmit += len;
1178 th->th_ack = htonl(tp->rcv_nxt);
1180 bcopy(opt, th + 1, optlen);
1181 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1183 th->th_flags = flags;
1185 * Calculate receive window. Don't shrink window,
1186 * but avoid silly window syndrome.
1188 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1189 recwin < tp->t_maxseg)
1191 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1192 recwin < (tp->rcv_adv - tp->rcv_nxt))
1193 recwin = (tp->rcv_adv - tp->rcv_nxt);
1196 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1197 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1198 * case is handled in syncache.
1201 th->th_win = htons((u_short)
1202 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1204 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1207 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1208 * a 0 window. This may cause the remote transmitter to stall. This
1209 * flag tells soreceive() to disable delayed acknowledgements when
1210 * draining the buffer. This can occur if the receiver is attempting
1211 * to read more data than can be buffered prior to transmitting on
1214 if (th->th_win == 0) {
1216 tp->t_flags |= TF_RXWIN0SENT;
1218 tp->t_flags &= ~TF_RXWIN0SENT;
1219 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1220 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1221 th->th_flags |= TH_URG;
1224 * If no urgent pointer to send, then we pull
1225 * the urgent pointer to the left edge of the send window
1226 * so that it doesn't drift into the send window on sequence
1227 * number wraparound.
1229 tp->snd_up = tp->snd_una; /* drag it along */
1232 * Put TCP length in extended header, and then
1233 * checksum extended header and data.
1235 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1236 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1238 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1239 if (to.to_flags & TOF_SIGNATURE) {
1241 * Calculate MD5 signature and put it into the place
1242 * determined before.
1243 * NOTE: since TCP options buffer doesn't point into
1244 * mbuf's data, calculate offset and use it.
1246 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1247 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1249 * Do not send segment if the calculation of MD5
1250 * digest has failed.
1260 * There is no need to fill in ip6_plen right now.
1261 * It will be filled later by ip6_output.
1263 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1264 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1265 optlen + len, IPPROTO_TCP, 0);
1268 #if defined(INET6) && defined(INET)
1273 m->m_pkthdr.csum_flags = CSUM_TCP;
1274 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1275 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1277 /* IP version must be set here for ipv4/ipv6 checking later */
1278 KASSERT(ip->ip_v == IPVERSION,
1279 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1284 * Enable TSO and specify the size of the segments.
1285 * The TCP pseudo header checksum is always provided.
1288 KASSERT(len > tp->t_maxseg - optlen,
1289 ("%s: len <= tso_segsz", __func__));
1290 m->m_pkthdr.csum_flags |= CSUM_TSO;
1291 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1294 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1295 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1296 ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1297 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1299 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1300 ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1301 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1305 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1306 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1313 if (so->so_options & SO_DEBUG) {
1319 save = ipov->ih_len;
1320 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1322 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1326 ipov->ih_len = save;
1328 #endif /* TCPDEBUG */
1329 TCP_PROBE3(debug__output, tp, th, m);
1332 * Fill in IP length and desired time to live and
1333 * send to IP level. There should be a better way
1334 * to handle ttl and tos; we could keep them in
1335 * the template, but need a way to checksum without them.
1338 * m->m_pkthdr.len should have been set before checksum calculation,
1339 * because in6_cksum() need it.
1344 * we separately set hoplimit for every segment, since the
1345 * user might want to change the value via setsockopt.
1346 * Also, desired default hop limit might be changed via
1347 * Neighbor Discovery.
1349 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1352 * Set the packet size here for the benefit of DTrace probes.
1353 * ip6_output() will set it properly; it's supposed to include
1354 * the option header lengths as well.
1356 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1358 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1359 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1361 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1363 if (tp->t_state == TCPS_SYN_SENT)
1364 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1366 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1369 /* Save packet, if requested. */
1370 tcp_pcap_add(th, m, &(tp->t_outpkts));
1373 /* TODO: IPv6 IP6TOS_ECT bit on */
1374 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1375 &tp->t_inpcb->inp_route6,
1376 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1377 NULL, NULL, tp->t_inpcb);
1379 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1380 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1383 #if defined(INET) && defined(INET6)
1388 ip->ip_len = htons(m->m_pkthdr.len);
1390 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1391 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1394 * If we do path MTU discovery, then we set DF on every packet.
1395 * This might not be the best thing to do according to RFC3390
1396 * Section 2. However the tcp hostcache migitates the problem
1397 * so it affects only the first tcp connection with a host.
1399 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1401 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1402 ip->ip_off |= htons(IP_DF);
1403 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1405 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1408 if (tp->t_state == TCPS_SYN_SENT)
1409 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1411 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1414 /* Save packet, if requested. */
1415 tcp_pcap_add(th, m, &(tp->t_outpkts));
1418 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1419 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1422 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1423 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1429 * In transmit state, time the transmission and arrange for
1430 * the retransmit. In persist state, just set snd_max.
1432 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1433 !tcp_timer_active(tp, TT_PERSIST)) {
1434 tcp_seq startseq = tp->snd_nxt;
1437 * Advance snd_nxt over sequence space of this segment.
1439 if (flags & (TH_SYN|TH_FIN)) {
1442 if (flags & TH_FIN) {
1444 tp->t_flags |= TF_SENTFIN;
1450 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1451 tp->snd_max = tp->snd_nxt;
1453 * Time this transmission if not a retransmission and
1454 * not currently timing anything.
1456 if (tp->t_rtttime == 0) {
1457 tp->t_rtttime = ticks;
1458 tp->t_rtseq = startseq;
1459 TCPSTAT_INC(tcps_segstimed);
1464 * Set retransmit timer if not currently set,
1465 * and not doing a pure ack or a keep-alive probe.
1466 * Initial value for retransmit timer is smoothed
1467 * round-trip time + 2 * round-trip time variance.
1468 * Initialize shift counter which is used for backoff
1469 * of retransmit time.
1472 if (!tcp_timer_active(tp, TT_REXMT) &&
1473 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1474 (tp->snd_nxt != tp->snd_una))) {
1475 if (tcp_timer_active(tp, TT_PERSIST)) {
1476 tcp_timer_activate(tp, TT_PERSIST, 0);
1479 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1480 } else if (len == 0 && sbavail(&so->so_snd) &&
1481 !tcp_timer_active(tp, TT_REXMT) &&
1482 !tcp_timer_active(tp, TT_PERSIST)) {
1484 * Avoid a situation where we do not set persist timer
1485 * after a zero window condition. For example:
1486 * 1) A -> B: packet with enough data to fill the window
1487 * 2) B -> A: ACK for #1 + new data (0 window
1489 * 3) A -> B: ACK for #2, 0 len packet
1491 * In this case, A will not activate the persist timer,
1492 * because it chose to send a packet. Unless tcp_output
1493 * is called for some other reason (delayed ack timer,
1494 * another input packet from B, socket syscall), A will
1495 * not send zero window probes.
1497 * So, if you send a 0-length packet, but there is data
1498 * in the socket buffer, and neither the rexmt or
1499 * persist timer is already set, then activate the
1507 * Persist case, update snd_max but since we are in
1508 * persist mode (no window) we do not update snd_nxt.
1513 if (flags & TH_FIN) {
1515 tp->t_flags |= TF_SENTFIN;
1517 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1518 tp->snd_max = tp->snd_nxt + xlen;
1524 * We know that the packet was lost, so back out the
1525 * sequence number advance, if any.
1527 * If the error is EPERM the packet got blocked by the
1528 * local firewall. Normally we should terminate the
1529 * connection but the blocking may have been spurious
1530 * due to a firewall reconfiguration cycle. So we treat
1531 * it like a packet loss and let the retransmit timer and
1532 * timeouts do their work over time.
1533 * XXX: It is a POLA question whether calling tcp_drop right
1534 * away would be the really correct behavior instead.
1536 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1537 !tcp_timer_active(tp, TT_PERSIST)) &&
1538 ((flags & TH_SYN) == 0) &&
1542 tp->sackhint.sack_bytes_rexmit -= len;
1543 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1544 ("sackhint bytes rtx >= 0"));
1548 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1551 tp->t_softerror = error;
1554 tp->t_softerror = error;
1557 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1558 tp->snd_cwnd = tp->t_maxseg;
1562 * For some reason the interface we used initially
1563 * to send segments changed to another or lowered
1565 * If TSO was active we either got an interface
1566 * without TSO capabilits or TSO was turned off.
1567 * If we obtained mtu from ip_output() then update
1571 tp->t_flags &= ~TF_TSO;
1573 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1581 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1582 tp->t_softerror = error;
1590 TCPSTAT_INC(tcps_sndtotal);
1593 * Data sent (as far as we can tell).
1594 * If this advertises a larger window than any other segment,
1595 * then remember the size of the advertised window.
1596 * Any pending ACK has now been sent.
1598 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1599 tp->rcv_adv = tp->rcv_nxt + recwin;
1600 tp->last_ack_sent = tp->rcv_nxt;
1601 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1602 if (tcp_timer_active(tp, TT_DELACK))
1603 tcp_timer_activate(tp, TT_DELACK, 0);
1606 * This completely breaks TCP if newreno is turned on. What happens
1607 * is that if delayed-acks are turned on on the receiver, this code
1608 * on the transmitter effectively destroys the TCP window, forcing
1609 * it to four packets (1.5Kx4 = 6K window).
1611 if (sendalot && --maxburst)
1620 tcp_setpersist(struct tcpcb *tp)
1622 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1625 tp->t_flags &= ~TF_PREVVALID;
1626 if (tcp_timer_active(tp, TT_REXMT))
1627 panic("tcp_setpersist: retransmit pending");
1629 * Start/restart persistence timer.
1631 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1632 tcp_persmin, tcp_persmax);
1633 tcp_timer_activate(tp, TT_PERSIST, tt);
1634 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1639 * Insert TCP options according to the supplied parameters to the place
1640 * optp in a consistent way. Can handle unaligned destinations.
1642 * The order of the option processing is crucial for optimal packing and
1643 * alignment for the scarce option space.
1645 * The optimal order for a SYN/SYN-ACK segment is:
1646 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1647 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1649 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1650 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1651 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1652 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1653 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1656 tcp_addoptions(struct tcpopt *to, u_char *optp)
1658 u_int32_t mask, optlen = 0;
1660 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1661 if ((to->to_flags & mask) != mask)
1663 if (optlen == TCP_MAXOLEN)
1665 switch (to->to_flags & mask) {
1667 while (optlen % 4) {
1668 optlen += TCPOLEN_NOP;
1669 *optp++ = TCPOPT_NOP;
1671 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1673 optlen += TCPOLEN_MAXSEG;
1674 *optp++ = TCPOPT_MAXSEG;
1675 *optp++ = TCPOLEN_MAXSEG;
1676 to->to_mss = htons(to->to_mss);
1677 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1678 optp += sizeof(to->to_mss);
1681 while (!optlen || optlen % 2 != 1) {
1682 optlen += TCPOLEN_NOP;
1683 *optp++ = TCPOPT_NOP;
1685 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1687 optlen += TCPOLEN_WINDOW;
1688 *optp++ = TCPOPT_WINDOW;
1689 *optp++ = TCPOLEN_WINDOW;
1690 *optp++ = to->to_wscale;
1693 while (optlen % 2) {
1694 optlen += TCPOLEN_NOP;
1695 *optp++ = TCPOPT_NOP;
1697 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1699 optlen += TCPOLEN_SACK_PERMITTED;
1700 *optp++ = TCPOPT_SACK_PERMITTED;
1701 *optp++ = TCPOLEN_SACK_PERMITTED;
1704 while (!optlen || optlen % 4 != 2) {
1705 optlen += TCPOLEN_NOP;
1706 *optp++ = TCPOPT_NOP;
1708 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1710 optlen += TCPOLEN_TIMESTAMP;
1711 *optp++ = TCPOPT_TIMESTAMP;
1712 *optp++ = TCPOLEN_TIMESTAMP;
1713 to->to_tsval = htonl(to->to_tsval);
1714 to->to_tsecr = htonl(to->to_tsecr);
1715 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1716 optp += sizeof(to->to_tsval);
1717 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1718 optp += sizeof(to->to_tsecr);
1722 int siglen = TCPOLEN_SIGNATURE - 2;
1724 while (!optlen || optlen % 4 != 2) {
1725 optlen += TCPOLEN_NOP;
1726 *optp++ = TCPOPT_NOP;
1728 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1729 to->to_flags &= ~TOF_SIGNATURE;
1732 optlen += TCPOLEN_SIGNATURE;
1733 *optp++ = TCPOPT_SIGNATURE;
1734 *optp++ = TCPOLEN_SIGNATURE;
1735 to->to_signature = optp;
1743 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1746 while (!optlen || optlen % 4 != 2) {
1747 optlen += TCPOLEN_NOP;
1748 *optp++ = TCPOPT_NOP;
1750 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1752 optlen += TCPOLEN_SACKHDR;
1753 *optp++ = TCPOPT_SACK;
1754 sackblks = min(to->to_nsacks,
1755 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1756 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1757 while (sackblks--) {
1758 sack_seq = htonl(sack->start);
1759 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1760 optp += sizeof(sack_seq);
1761 sack_seq = htonl(sack->end);
1762 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1763 optp += sizeof(sack_seq);
1764 optlen += TCPOLEN_SACK;
1767 TCPSTAT_INC(tcps_sack_send_blocks);
1775 /* XXX is there any point to aligning this option? */
1776 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1777 if (TCP_MAXOLEN - optlen < total_len)
1779 *optp++ = TCPOPT_FAST_OPEN;
1780 *optp++ = total_len;
1781 if (to->to_tfo_len > 0) {
1782 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1783 optp += to->to_tfo_len;
1785 optlen += total_len;
1790 panic("%s: unknown TCP option type", __func__);
1795 /* Terminate and pad TCP options to a 4 byte boundary. */
1797 optlen += TCPOLEN_EOL;
1798 *optp++ = TCPOPT_EOL;
1801 * According to RFC 793 (STD0007):
1802 * "The content of the header beyond the End-of-Option option
1803 * must be header padding (i.e., zero)."
1804 * and later: "The padding is composed of zeros."
1806 while (optlen % 4) {
1807 optlen += TCPOLEN_PAD;
1808 *optp++ = TCPOPT_PAD;
1811 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1816 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1820 * Automatic sizing of send socket buffer. Often the send buffer
1821 * size is not optimally adjusted to the actual network conditions
1822 * at hand (delay bandwidth product). Setting the buffer size too
1823 * small limits throughput on links with high bandwidth and high
1824 * delay (eg. trans-continental/oceanic links). Setting the
1825 * buffer size too big consumes too much real kernel memory,
1826 * especially with many connections on busy servers.
1828 * The criteria to step up the send buffer one notch are:
1829 * 1. receive window of remote host is larger than send buffer
1830 * (with a fudge factor of 5/4th);
1831 * 2. send buffer is filled to 7/8th with data (so we actually
1832 * have data to make use of it);
1833 * 3. send buffer fill has not hit maximal automatic size;
1834 * 4. our send window (slow start and cogestion controlled) is
1835 * larger than sent but unacknowledged data in send buffer.
1837 * The remote host receive window scaling factor may limit the
1838 * growing of the send buffer before it reaches its allowed
1841 * It scales directly with slow start or congestion window
1842 * and does at most one step per received ACK. This fast
1843 * scaling has the drawback of growing the send buffer beyond
1844 * what is strictly necessary to make full use of a given
1845 * delay*bandwidth product. However testing has shown this not
1846 * to be much of an problem. At worst we are trading wasting
1847 * of available bandwidth (the non-use of it) for wasting some
1848 * socket buffer memory.
1850 * TODO: Shrink send buffer during idle periods together
1851 * with congestion window. Requires another timer. Has to
1852 * wait for upcoming tcp timer rewrite.
1854 * XXXGL: should there be used sbused() or sbavail()?
1856 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
1859 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
1860 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
1861 sbused(&so->so_snd) >=
1862 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
1863 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
1864 sendwin >= (sbused(&so->so_snd) -
1865 (tp->snd_nxt - tp->snd_una))) {
1866 if (!sbreserve_locked(&so->so_snd,
1867 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
1868 V_tcp_autosndbuf_max), so, curthread))
1869 so->so_snd.sb_flags &= ~SB_AUTOSIZE;