2 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
3 * The Regents of the University of California. All rights reserved.
5 * Redistribution and use in source and binary forms, with or without
6 * modification, are permitted provided that the following conditions
8 * 1. Redistributions of source code must retain the above copyright
9 * notice, this list of conditions and the following disclaimer.
10 * 2. Redistributions in binary form must reproduce the above copyright
11 * notice, this list of conditions and the following disclaimer in the
12 * documentation and/or other materials provided with the distribution.
13 * 4. Neither the name of the University nor the names of its contributors
14 * may be used to endorse or promote products derived from this software
15 * without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
21 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
29 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
36 #include "opt_inet6.h"
37 #include "opt_ipsec.h"
38 #include "opt_tcpdebug.h"
40 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/domain.h>
43 #include <sys/hhook.h>
44 #include <sys/kernel.h>
47 #include <sys/mutex.h>
48 #include <sys/protosw.h>
50 #include <sys/socket.h>
51 #include <sys/socketvar.h>
52 #include <sys/sysctl.h>
55 #include <net/route.h>
58 #include <netinet/in.h>
59 #include <netinet/in_kdtrace.h>
60 #include <netinet/in_systm.h>
61 #include <netinet/ip.h>
62 #include <netinet/in_pcb.h>
63 #include <netinet/ip_var.h>
64 #include <netinet/ip_options.h>
66 #include <netinet6/in6_pcb.h>
67 #include <netinet/ip6.h>
68 #include <netinet6/ip6_var.h>
71 #include <netinet/tcp_fastopen.h>
73 #include <netinet/tcp.h>
75 #include <netinet/tcp_fsm.h>
76 #include <netinet/tcp_seq.h>
77 #include <netinet/tcp_timer.h>
78 #include <netinet/tcp_var.h>
79 #include <netinet/tcpip.h>
80 #include <netinet/cc/cc.h>
82 #include <netinet/tcp_pcap.h>
85 #include <netinet/tcp_debug.h>
88 #include <netinet/tcp_offload.h>
91 #include <netipsec/ipsec_support.h>
93 #include <machine/in_cksum.h>
95 #include <security/mac/mac_framework.h>
97 VNET_DEFINE(int, path_mtu_discovery) = 1;
98 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
99 &VNET_NAME(path_mtu_discovery), 1,
100 "Enable Path MTU Discovery");
102 VNET_DEFINE(int, tcp_do_tso) = 1;
103 #define V_tcp_do_tso VNET(tcp_do_tso)
104 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
105 &VNET_NAME(tcp_do_tso), 0,
106 "Enable TCP Segmentation Offload");
108 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
109 #define V_tcp_sendspace VNET(tcp_sendspace)
110 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
111 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
113 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
114 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
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 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
121 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
122 &VNET_NAME(tcp_autosndbuf_inc), 0,
123 "Incrementor step size of automatic send buffer");
125 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
126 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
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");
132 * Make sure that either retransmit or persist timer is set for SYN, FIN and
135 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
136 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
137 tcp_timer_active((tp), TT_REXMT) || \
138 tcp_timer_active((tp), TT_PERSIST), \
139 ("neither rexmt nor persist timer is set"))
141 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
142 struct tcphdr *th, struct tcpopt *to,
144 static void inline cc_after_idle(struct tcpcb *tp);
147 * Wrapper for the TCP established output helper hook.
150 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
151 struct tcpopt *to, long len, int tso)
153 struct tcp_hhook_data hhook_data;
155 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
159 hhook_data.len = len;
160 hhook_data.tso = tso;
162 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
168 * CC wrapper hook functions
171 cc_after_idle(struct tcpcb *tp)
173 INP_WLOCK_ASSERT(tp->t_inpcb);
175 if (CC_ALGO(tp)->after_idle != NULL)
176 CC_ALGO(tp)->after_idle(tp->ccv);
180 * Tcp output routine: figure out what should be sent and send it.
183 tcp_output(struct tcpcb *tp)
185 struct socket *so = tp->t_inpcb->inp_socket;
186 long len, recwin, sendwin;
187 int off, flags, error = 0; /* Keep compiler happy */
189 struct ip *ip = NULL;
190 struct ipovly *ipov = NULL;
192 u_char opt[TCP_MAXOLEN];
193 unsigned ipoptlen, optlen, hdrlen;
194 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
195 unsigned ipsec_optlen = 0;
198 int sack_rxmit, sack_bytes_rxmt;
203 int maxburst = TCP_MAXBURST;
206 struct ip6_hdr *ip6 = NULL;
209 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
212 INP_WLOCK_ASSERT(tp->t_inpcb);
215 if (tp->t_flags & TF_TOE)
216 return (tcp_offload_output(tp));
221 * For TFO connections in SYN_RECEIVED, only allow the initial
222 * SYN|ACK and those sent by the retransmit timer.
224 if ((tp->t_flags & TF_FASTOPEN) &&
225 (tp->t_state == TCPS_SYN_RECEIVED) &&
226 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN|ACK sent */
227 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
231 * Determine length of data that should be transmitted,
232 * and flags that will be used.
233 * If there is some data or critical controls (SYN, RST)
234 * to send, then transmit; otherwise, investigate further.
236 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
237 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
239 tp->t_flags &= ~TF_LASTIDLE;
241 if (tp->t_flags & TF_MORETOCOME) {
242 tp->t_flags |= TF_LASTIDLE;
248 * If we've recently taken a timeout, snd_max will be greater than
249 * snd_nxt. There may be SACK information that allows us to avoid
250 * resending already delivered data. Adjust snd_nxt accordingly.
252 if ((tp->t_flags & TF_SACK_PERMIT) &&
253 SEQ_LT(tp->snd_nxt, tp->snd_max))
258 off = tp->snd_nxt - tp->snd_una;
259 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
261 flags = tcp_outflags[tp->t_state];
263 * Send any SACK-generated retransmissions. If we're explicitly trying
264 * to send out new data (when sendalot is 1), bypass this function.
265 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
266 * we're replacing a (future) new transmission with a retransmission
267 * now, and we previously incremented snd_cwnd in tcp_input().
270 * Still in sack recovery , reset rxmit flag to zero.
276 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
277 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
280 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
283 /* Do not retransmit SACK segments beyond snd_recover */
284 if (SEQ_GT(p->end, tp->snd_recover)) {
286 * (At least) part of sack hole extends beyond
287 * snd_recover. Check to see if we can rexmit data
290 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
292 * Can't rexmit any more data for this hole.
293 * That data will be rexmitted in the next
294 * sack recovery episode, when snd_recover
295 * moves past p->rxmit.
298 goto after_sack_rexmit;
300 /* Can rexmit part of the current hole */
301 len = ((long)ulmin(cwin,
302 tp->snd_recover - p->rxmit));
304 len = ((long)ulmin(cwin, p->end - p->rxmit));
305 off = p->rxmit - tp->snd_una;
306 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
311 TCPSTAT_INC(tcps_sack_rexmits);
312 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
313 min(len, tp->t_maxseg));
318 * Get standard flags, and add SYN or FIN if requested by 'hidden'
321 if (tp->t_flags & TF_NEEDFIN)
323 if (tp->t_flags & TF_NEEDSYN)
326 SOCKBUF_LOCK(&so->so_snd);
328 * If in persist timeout with window of 0, send 1 byte.
329 * Otherwise, if window is small but nonzero
330 * and timer expired, we will send what we can
331 * and go to transmit state.
333 if (tp->t_flags & TF_FORCEDATA) {
336 * If we still have some data to send, then
337 * clear the FIN bit. Usually this would
338 * happen below when it realizes that we
339 * aren't sending all the data. However,
340 * if we have exactly 1 byte of unsent data,
341 * then it won't clear the FIN bit below,
342 * and if we are in persist state, we wind
343 * up sending the packet without recording
344 * that we sent the FIN bit.
346 * We can't just blindly clear the FIN bit,
347 * because if we don't have any more data
348 * to send then the probe will be the FIN
351 if (off < sbused(&so->so_snd))
355 tcp_timer_activate(tp, TT_PERSIST, 0);
361 * If snd_nxt == snd_max and we have transmitted a FIN, the
362 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
363 * a negative length. This can also occur when TCP opens up
364 * its congestion window while receiving additional duplicate
365 * acks after fast-retransmit because TCP will reset snd_nxt
366 * to snd_max after the fast-retransmit.
368 * In the normal retransmit-FIN-only case, however, snd_nxt will
369 * be set to snd_una, the offset will be 0, and the length may
372 * If sack_rxmit is true we are retransmitting from the scoreboard
373 * in which case len is already set.
375 if (sack_rxmit == 0) {
376 if (sack_bytes_rxmt == 0)
377 len = ((long)ulmin(sbavail(&so->so_snd), sendwin) -
383 * We are inside of a SACK recovery episode and are
384 * sending new data, having retransmitted all the
385 * data possible in the scoreboard.
387 len = ((long)ulmin(sbavail(&so->so_snd), tp->snd_wnd) -
390 * Don't remove this (len > 0) check !
391 * We explicitly check for len > 0 here (although it
392 * isn't really necessary), to work around a gcc
393 * optimization issue - to force gcc to compute
394 * len above. Without this check, the computation
395 * of len is bungled by the optimizer.
398 cwin = tp->snd_cwnd -
399 (tp->snd_nxt - tp->sack_newdata) -
403 len = lmin(len, cwin);
409 * Lop off SYN bit if it has already been sent. However, if this
410 * is SYN-SENT state and if segment contains data and if we don't
411 * know that foreign host supports TAO, suppress sending segment.
413 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
414 if (tp->t_state != TCPS_SYN_RECEIVED)
418 * When sending additional segments following a TFO SYN|ACK,
419 * do not include the SYN bit.
421 if ((tp->t_flags & TF_FASTOPEN) &&
422 (tp->t_state == TCPS_SYN_RECEIVED))
429 * Be careful not to send data and/or FIN on SYN segments.
430 * This measure is needed to prevent interoperability problems
431 * with not fully conformant TCP implementations.
433 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
440 * When retransmitting SYN|ACK on a passively-created TFO socket,
441 * don't include data, as the presence of data may have caused the
442 * original SYN|ACK to have been dropped by a middlebox.
444 if ((tp->t_flags & TF_FASTOPEN) &&
445 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
451 * If FIN has been sent but not acked,
452 * but we haven't been called to retransmit,
453 * len will be < 0. Otherwise, window shrank
454 * after we sent into it. If window shrank to 0,
455 * cancel pending retransmit, pull snd_nxt back
456 * to (closed) window, and set the persist timer
457 * if it isn't already going. If the window didn't
458 * close completely, just wait for an ACK.
460 * We also do a general check here to ensure that
461 * we will set the persist timer when we have data
462 * to send, but a 0-byte window. This makes sure
463 * the persist timer is set even if the packet
464 * hits one of the "goto send" lines below.
467 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
468 (off < (int) sbavail(&so->so_snd))) {
469 tcp_timer_activate(tp, TT_REXMT, 0);
471 tp->snd_nxt = tp->snd_una;
472 if (!tcp_timer_active(tp, TT_PERSIST))
477 /* len will be >= 0 after this point. */
478 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
481 * Automatic sizing of send socket buffer. Often the send buffer
482 * size is not optimally adjusted to the actual network conditions
483 * at hand (delay bandwidth product). Setting the buffer size too
484 * small limits throughput on links with high bandwidth and high
485 * delay (eg. trans-continental/oceanic links). Setting the
486 * buffer size too big consumes too much real kernel memory,
487 * especially with many connections on busy servers.
489 * The criteria to step up the send buffer one notch are:
490 * 1. receive window of remote host is larger than send buffer
491 * (with a fudge factor of 5/4th);
492 * 2. send buffer is filled to 7/8th with data (so we actually
493 * have data to make use of it);
494 * 3. send buffer fill has not hit maximal automatic size;
495 * 4. our send window (slow start and cogestion controlled) is
496 * larger than sent but unacknowledged data in send buffer.
498 * The remote host receive window scaling factor may limit the
499 * growing of the send buffer before it reaches its allowed
502 * It scales directly with slow start or congestion window
503 * and does at most one step per received ACK. This fast
504 * scaling has the drawback of growing the send buffer beyond
505 * what is strictly necessary to make full use of a given
506 * delay*bandwidth product. However testing has shown this not
507 * to be much of an problem. At worst we are trading wasting
508 * of available bandwidth (the non-use of it) for wasting some
509 * socket buffer memory.
511 * TODO: Shrink send buffer during idle periods together
512 * with congestion window. Requires another timer. Has to
513 * wait for upcoming tcp timer rewrite.
515 * XXXGL: should there be used sbused() or sbavail()?
517 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
518 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
519 sbused(&so->so_snd) >= (so->so_snd.sb_hiwat / 8 * 7) &&
520 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
521 sendwin >= (sbused(&so->so_snd) -
522 (tp->snd_nxt - tp->snd_una))) {
523 if (!sbreserve_locked(&so->so_snd,
524 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
525 V_tcp_autosndbuf_max), so, curthread))
526 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
531 * Decide if we can use TCP Segmentation Offloading (if supported by
534 * TSO may only be used if we are in a pure bulk sending state. The
535 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
536 * IP options prevent using TSO. With TSO the TCP header is the same
537 * (except for the sequence number) for all generated packets. This
538 * makes it impossible to transmit any options which vary per generated
541 * IPv4 handling has a clear separation of ip options and ip header
542 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
543 * the right thing below to provide length of just ip options and thus
544 * checking for ipoptlen is enough to decide if ip options are present.
546 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
548 * Pre-calculate here as we save another lookup into the darknesses
549 * of IPsec that way and can actually decide if TSO is ok.
552 if (isipv6 && IPSEC_ENABLED(ipv6))
553 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
559 if (IPSEC_ENABLED(ipv4))
560 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
565 ipoptlen = ip6_optlen(tp->t_inpcb);
568 if (tp->t_inpcb->inp_options)
569 ipoptlen = tp->t_inpcb->inp_options->m_len -
570 offsetof(struct ipoption, ipopt_list);
573 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
574 ipoptlen += ipsec_optlen;
577 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
578 ((tp->t_flags & TF_SIGNATURE) == 0) &&
579 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
584 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
587 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
588 sbused(&so->so_snd)))
592 recwin = sbspace(&so->so_rcv);
595 * Sender silly window avoidance. We transmit under the following
596 * conditions when len is non-zero:
598 * - We have a full segment (or more with TSO)
599 * - This is the last buffer in a write()/send() and we are
600 * either idle or running NODELAY
601 * - we've timed out (e.g. persist timer)
602 * - we have more then 1/2 the maximum send window's worth of
603 * data (receiver may be limited the window size)
604 * - we need to retransmit
607 if (len >= tp->t_maxseg)
610 * NOTE! on localhost connections an 'ack' from the remote
611 * end may occur synchronously with the output and cause
612 * us to flush a buffer queued with moretocome. XXX
614 * note: the len + off check is almost certainly unnecessary.
616 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
617 (idle || (tp->t_flags & TF_NODELAY)) &&
618 len + off >= sbavail(&so->so_snd) &&
619 (tp->t_flags & TF_NOPUSH) == 0) {
622 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
624 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
626 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
633 * Sending of standalone window updates.
635 * Window updates are important when we close our window due to a
636 * full socket buffer and are opening it again after the application
637 * reads data from it. Once the window has opened again and the
638 * remote end starts to send again the ACK clock takes over and
639 * provides the most current window information.
641 * We must avoid the silly window syndrome whereas every read
642 * from the receive buffer, no matter how small, causes a window
643 * update to be sent. We also should avoid sending a flurry of
644 * window updates when the socket buffer had queued a lot of data
645 * and the application is doing small reads.
647 * Prevent a flurry of pointless window updates by only sending
648 * an update when we can increase the advertized window by more
649 * than 1/4th of the socket buffer capacity. When the buffer is
650 * getting full or is very small be more aggressive and send an
651 * update whenever we can increase by two mss sized segments.
652 * In all other situations the ACK's to new incoming data will
653 * carry further window increases.
655 * Don't send an independent window update if a delayed
656 * ACK is pending (it will get piggy-backed on it) or the
657 * remote side already has done a half-close and won't send
658 * more data. Skip this if the connection is in T/TCP
661 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
662 !(tp->t_flags & TF_DELACK) &&
663 !TCPS_HAVERCVDFIN(tp->t_state)) {
665 * "adv" is the amount we could increase the window,
666 * taking into account that we are limited by
667 * TCP_MAXWIN << tp->rcv_scale.
672 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
673 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
674 oldwin = (tp->rcv_adv - tp->rcv_nxt);
680 * If the new window size ends up being the same as or less
681 * than the old size when it is scaled, then don't force
684 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
687 if (adv >= (long)(2 * tp->t_maxseg) &&
688 (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
689 recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
690 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
692 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
698 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
699 * is also a catch-all for the retransmit timer timeout case.
701 if (tp->t_flags & TF_ACKNOW)
703 if ((flags & TH_RST) ||
704 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
706 if (SEQ_GT(tp->snd_up, tp->snd_una))
709 * If our state indicates that FIN should be sent
710 * and we have not yet done so, then we need to send.
712 if (flags & TH_FIN &&
713 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
716 * In SACK, it is possible for tcp_output to fail to send a segment
717 * after the retransmission timer has been turned off. Make sure
718 * that the retransmission timer is set.
720 if ((tp->t_flags & TF_SACK_PERMIT) &&
721 SEQ_GT(tp->snd_max, tp->snd_una) &&
722 !tcp_timer_active(tp, TT_REXMT) &&
723 !tcp_timer_active(tp, TT_PERSIST)) {
724 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
728 * TCP window updates are not reliable, rather a polling protocol
729 * using ``persist'' packets is used to insure receipt of window
730 * updates. The three ``states'' for the output side are:
731 * idle not doing retransmits or persists
732 * persisting to move a small or zero window
733 * (re)transmitting and thereby not persisting
735 * tcp_timer_active(tp, TT_PERSIST)
736 * is true when we are in persist state.
737 * (tp->t_flags & TF_FORCEDATA)
738 * is set when we are called to send a persist packet.
739 * tcp_timer_active(tp, TT_REXMT)
740 * is set when we are retransmitting
741 * The output side is idle when both timers are zero.
743 * If send window is too small, there is data to transmit, and no
744 * retransmit or persist is pending, then go to persist state.
745 * If nothing happens soon, send when timer expires:
746 * if window is nonzero, transmit what we can,
747 * otherwise force out a byte.
749 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
750 !tcp_timer_active(tp, TT_PERSIST)) {
756 * No reason to send a segment, just return.
759 SOCKBUF_UNLOCK(&so->so_snd);
763 SOCKBUF_LOCK_ASSERT(&so->so_snd);
765 if (len >= tp->t_maxseg)
766 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
768 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
771 * Before ESTABLISHED, force sending of initial options
772 * unless TCP set not to do any options.
773 * NOTE: we assume that the IP/TCP header plus TCP options
774 * always fit in a single mbuf, leaving room for a maximum
776 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
781 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
784 hdrlen = sizeof (struct tcpiphdr);
787 * Compute options for segment.
788 * We only have to care about SYN and established connection
789 * segments. Options for SYN-ACK segments are handled in TCP
793 if ((tp->t_flags & TF_NOOPT) == 0) {
794 /* Maximum segment size. */
795 if (flags & TH_SYN) {
796 tp->snd_nxt = tp->iss;
797 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
798 to.to_flags |= TOF_MSS;
801 * Only include the TFO option on the first
802 * transmission of the SYN|ACK on a
803 * passively-created TFO socket, as the presence of
804 * the TFO option may have caused the original
805 * SYN|ACK to have been dropped by a middlebox.
807 if ((tp->t_flags & TF_FASTOPEN) &&
808 (tp->t_state == TCPS_SYN_RECEIVED) &&
809 (tp->t_rxtshift == 0)) {
810 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
811 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
812 to.to_flags |= TOF_FASTOPEN;
816 /* Window scaling. */
817 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
818 to.to_wscale = tp->request_r_scale;
819 to.to_flags |= TOF_SCALE;
822 if ((tp->t_flags & TF_RCVD_TSTMP) ||
823 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
824 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
825 to.to_tsecr = tp->ts_recent;
826 to.to_flags |= TOF_TS;
829 /* Set receive buffer autosizing timestamp. */
830 if (tp->rfbuf_ts == 0 &&
831 (so->so_rcv.sb_flags & SB_AUTOSIZE))
832 tp->rfbuf_ts = tcp_ts_getticks();
834 /* Selective ACK's. */
835 if (tp->t_flags & TF_SACK_PERMIT) {
837 to.to_flags |= TOF_SACKPERM;
838 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
839 (tp->t_flags & TF_SACK_PERMIT) &&
840 tp->rcv_numsacks > 0) {
841 to.to_flags |= TOF_SACK;
842 to.to_nsacks = tp->rcv_numsacks;
843 to.to_sacks = (u_char *)tp->sackblks;
846 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
847 /* TCP-MD5 (RFC2385). */
849 * Check that TCP_MD5SIG is enabled in tcpcb to
850 * account the size needed to set this TCP option.
852 if (tp->t_flags & TF_SIGNATURE)
853 to.to_flags |= TOF_SIGNATURE;
854 #endif /* TCP_SIGNATURE */
856 /* Processing the options. */
857 hdrlen += optlen = tcp_addoptions(&to, opt);
861 * Adjust data length if insertion of options will
862 * bump the packet length beyond the t_maxseg length.
863 * Clear the FIN bit because we cut off the tail of
866 if (len + optlen + ipoptlen > tp->t_maxseg) {
871 u_int if_hw_tsomaxsegcount;
872 u_int if_hw_tsomaxsegsize;
877 /* extract TSO information */
878 if_hw_tsomax = tp->t_tsomax;
879 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
880 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
883 * Limit a TSO burst to prevent it from
884 * overflowing or exceeding the maximum length
885 * allowed by the network interface:
887 KASSERT(ipoptlen == 0,
888 ("%s: TSO can't do IP options", __func__));
891 * Check if we should limit by maximum payload
894 if (if_hw_tsomax != 0) {
895 /* compute maximum TSO length */
896 max_len = (if_hw_tsomax - hdrlen -
900 } else if (len > max_len) {
907 * Check if we should limit by maximum segment
910 if (if_hw_tsomaxsegcount != 0 &&
911 if_hw_tsomaxsegsize != 0) {
913 * Subtract one segment for the LINK
914 * and TCP/IP headers mbuf that will
915 * be prepended to this mbuf chain
916 * after the code in this section
917 * limits the number of mbufs in the
918 * chain to if_hw_tsomaxsegcount.
920 if_hw_tsomaxsegcount -= 1;
922 mb = sbsndmbuf(&so->so_snd, off, &moff);
924 while (mb != NULL && max_len < len) {
929 * Get length of mbuf fragment
930 * and how many hardware frags,
931 * rounded up, it would use:
933 mlen = (mb->m_len - moff);
934 frags = howmany(mlen,
935 if_hw_tsomaxsegsize);
937 /* Handle special case: Zero Length Mbuf */
942 * Check if the fragment limit
943 * will be reached or exceeded:
945 if (frags >= if_hw_tsomaxsegcount) {
947 if_hw_tsomaxsegcount *
948 if_hw_tsomaxsegsize);
952 if_hw_tsomaxsegcount -= frags;
958 } else if (len > max_len) {
965 * Prevent the last segment from being
966 * fractional unless the send sockbuf can be
969 max_len = (tp->t_maxseg - optlen);
970 if ((off + len) < sbavail(&so->so_snd)) {
971 moff = len % max_len;
979 * In case there are too many small fragments
982 if (len <= max_len) {
989 * Send the FIN in a separate segment
990 * after the bulk sending is done.
991 * We don't trust the TSO implementations
992 * to clear the FIN flag on all but the
995 if (tp->t_flags & TF_NEEDFIN)
999 len = tp->t_maxseg - optlen - ipoptlen;
1005 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1006 ("%s: len > IP_MAXPACKET", __func__));
1008 /*#ifdef DIAGNOSTIC*/
1010 if (max_linkhdr + hdrlen > MCLBYTES)
1012 if (max_linkhdr + hdrlen > MHLEN)
1014 panic("tcphdr too big");
1018 * This KASSERT is here to catch edge cases at a well defined place.
1019 * Before, those had triggered (random) panic conditions further down.
1021 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1024 * Grab a header mbuf, attaching a copy of data to
1025 * be transmitted, and initialize the header from
1026 * the template for sends on this connection.
1032 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1033 TCPSTAT_INC(tcps_sndprobe);
1034 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1035 tp->t_sndrexmitpack++;
1036 TCPSTAT_INC(tcps_sndrexmitpack);
1037 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1039 TCPSTAT_INC(tcps_sndpack);
1040 TCPSTAT_ADD(tcps_sndbyte, len);
1043 if (MHLEN < hdrlen + max_linkhdr)
1044 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1047 m = m_gethdr(M_NOWAIT, MT_DATA);
1050 SOCKBUF_UNLOCK(&so->so_snd);
1056 m->m_data += max_linkhdr;
1060 * Start the m_copy functions from the closest mbuf
1061 * to the offset in the socket buffer chain.
1063 mb = sbsndptr(&so->so_snd, off, len, &moff);
1065 if (len <= MHLEN - hdrlen - max_linkhdr) {
1066 m_copydata(mb, moff, (int)len,
1067 mtod(m, caddr_t) + hdrlen);
1070 m->m_next = m_copy(mb, moff, (int)len);
1071 if (m->m_next == NULL) {
1072 SOCKBUF_UNLOCK(&so->so_snd);
1081 * If we're sending everything we've got, set PUSH.
1082 * (This will keep happy those implementations which only
1083 * give data to the user when a buffer fills or
1086 if ((off + len == sbused(&so->so_snd)) && !(flags & TH_SYN))
1088 SOCKBUF_UNLOCK(&so->so_snd);
1090 SOCKBUF_UNLOCK(&so->so_snd);
1091 if (tp->t_flags & TF_ACKNOW)
1092 TCPSTAT_INC(tcps_sndacks);
1093 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1094 TCPSTAT_INC(tcps_sndctrl);
1095 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1096 TCPSTAT_INC(tcps_sndurg);
1098 TCPSTAT_INC(tcps_sndwinup);
1100 m = m_gethdr(M_NOWAIT, MT_DATA);
1107 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1112 m->m_data += max_linkhdr;
1115 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1116 m->m_pkthdr.rcvif = (struct ifnet *)0;
1118 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1122 ip6 = mtod(m, struct ip6_hdr *);
1123 th = (struct tcphdr *)(ip6 + 1);
1124 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1128 ip = mtod(m, struct ip *);
1129 ipov = (struct ipovly *)ip;
1130 th = (struct tcphdr *)(ip + 1);
1131 tcpip_fillheaders(tp->t_inpcb, ip, th);
1135 * Fill in fields, remembering maximum advertised
1136 * window for use in delaying messages about window sizes.
1137 * If resending a FIN, be sure not to use a new sequence number.
1139 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1140 tp->snd_nxt == tp->snd_max)
1143 * If we are starting a connection, send ECN setup
1144 * SYN packet. If we are on a retransmit, we may
1145 * resend those bits a number of times as per
1148 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1149 if (tp->t_rxtshift >= 1) {
1150 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1151 flags |= TH_ECE|TH_CWR;
1153 flags |= TH_ECE|TH_CWR;
1156 if (tp->t_state == TCPS_ESTABLISHED &&
1157 (tp->t_flags & TF_ECN_PERMIT)) {
1159 * If the peer has ECN, mark data packets with
1160 * ECN capable transmission (ECT).
1161 * Ignore pure ack packets, retransmissions and window probes.
1163 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1164 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1167 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1170 ip->ip_tos |= IPTOS_ECN_ECT0;
1171 TCPSTAT_INC(tcps_ecn_ect0);
1175 * Reply with proper ECN notifications.
1177 if (tp->t_flags & TF_ECN_SND_CWR) {
1179 tp->t_flags &= ~TF_ECN_SND_CWR;
1181 if (tp->t_flags & TF_ECN_SND_ECE)
1186 * If we are doing retransmissions, then snd_nxt will
1187 * not reflect the first unsent octet. For ACK only
1188 * packets, we do not want the sequence number of the
1189 * retransmitted packet, we want the sequence number
1190 * of the next unsent octet. So, if there is no data
1191 * (and no SYN or FIN), use snd_max instead of snd_nxt
1192 * when filling in ti_seq. But if we are in persist
1193 * state, snd_max might reflect one byte beyond the
1194 * right edge of the window, so use snd_nxt in that
1195 * case, since we know we aren't doing a retransmission.
1196 * (retransmit and persist are mutually exclusive...)
1198 if (sack_rxmit == 0) {
1199 if (len || (flags & (TH_SYN|TH_FIN)) ||
1200 tcp_timer_active(tp, TT_PERSIST))
1201 th->th_seq = htonl(tp->snd_nxt);
1203 th->th_seq = htonl(tp->snd_max);
1205 th->th_seq = htonl(p->rxmit);
1207 tp->sackhint.sack_bytes_rexmit += len;
1209 th->th_ack = htonl(tp->rcv_nxt);
1211 bcopy(opt, th + 1, optlen);
1212 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1214 th->th_flags = flags;
1216 * Calculate receive window. Don't shrink window,
1217 * but avoid silly window syndrome.
1219 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1220 recwin < (long)tp->t_maxseg)
1222 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1223 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1224 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1225 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1226 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1229 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1230 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1231 * case is handled in syncache.
1234 th->th_win = htons((u_short)
1235 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1237 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1240 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1241 * a 0 window. This may cause the remote transmitter to stall. This
1242 * flag tells soreceive() to disable delayed acknowledgements when
1243 * draining the buffer. This can occur if the receiver is attempting
1244 * to read more data than can be buffered prior to transmitting on
1247 if (th->th_win == 0) {
1249 tp->t_flags |= TF_RXWIN0SENT;
1251 tp->t_flags &= ~TF_RXWIN0SENT;
1252 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1253 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1254 th->th_flags |= TH_URG;
1257 * If no urgent pointer to send, then we pull
1258 * the urgent pointer to the left edge of the send window
1259 * so that it doesn't drift into the send window on sequence
1260 * number wraparound.
1262 tp->snd_up = tp->snd_una; /* drag it along */
1265 * Put TCP length in extended header, and then
1266 * checksum extended header and data.
1268 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1269 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1271 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1272 if (to.to_flags & TOF_SIGNATURE) {
1274 * Calculate MD5 signature and put it into the place
1275 * determined before.
1276 * NOTE: since TCP options buffer doesn't point into
1277 * mbuf's data, calculate offset and use it.
1279 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1280 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1282 * Do not send segment if the calculation of MD5
1283 * digest has failed.
1293 * There is no need to fill in ip6_plen right now.
1294 * It will be filled later by ip6_output.
1296 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1297 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1298 optlen + len, IPPROTO_TCP, 0);
1301 #if defined(INET6) && defined(INET)
1306 m->m_pkthdr.csum_flags = CSUM_TCP;
1307 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1308 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1310 /* IP version must be set here for ipv4/ipv6 checking later */
1311 KASSERT(ip->ip_v == IPVERSION,
1312 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1317 * Enable TSO and specify the size of the segments.
1318 * The TCP pseudo header checksum is always provided.
1321 KASSERT(len > tp->t_maxseg - optlen,
1322 ("%s: len <= tso_segsz", __func__));
1323 m->m_pkthdr.csum_flags |= CSUM_TSO;
1324 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1327 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1328 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1329 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1330 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1332 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1333 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1334 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1337 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1338 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1344 if (so->so_options & SO_DEBUG) {
1350 save = ipov->ih_len;
1351 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1353 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1357 ipov->ih_len = save;
1359 #endif /* TCPDEBUG */
1360 TCP_PROBE3(debug__output, tp, th, m);
1363 * Fill in IP length and desired time to live and
1364 * send to IP level. There should be a better way
1365 * to handle ttl and tos; we could keep them in
1366 * the template, but need a way to checksum without them.
1369 * m->m_pkthdr.len should have been set before checksum calculation,
1370 * because in6_cksum() need it.
1375 * we separately set hoplimit for every segment, since the
1376 * user might want to change the value via setsockopt.
1377 * Also, desired default hop limit might be changed via
1378 * Neighbor Discovery.
1380 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1383 * Set the packet size here for the benefit of DTrace probes.
1384 * ip6_output() will set it properly; it's supposed to include
1385 * the option header lengths as well.
1387 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1389 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1390 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1392 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1394 if (tp->t_state == TCPS_SYN_SENT)
1395 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1397 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1400 /* Save packet, if requested. */
1401 tcp_pcap_add(th, m, &(tp->t_outpkts));
1404 /* TODO: IPv6 IP6TOS_ECT bit on */
1405 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1406 &tp->t_inpcb->inp_route6,
1407 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1408 NULL, NULL, tp->t_inpcb);
1410 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1411 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1414 #if defined(INET) && defined(INET6)
1419 ip->ip_len = htons(m->m_pkthdr.len);
1421 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1422 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1425 * If we do path MTU discovery, then we set DF on every packet.
1426 * This might not be the best thing to do according to RFC3390
1427 * Section 2. However the tcp hostcache migitates the problem
1428 * so it affects only the first tcp connection with a host.
1430 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1432 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1433 ip->ip_off |= htons(IP_DF);
1434 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1436 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1439 if (tp->t_state == TCPS_SYN_SENT)
1440 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1442 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1445 /* Save packet, if requested. */
1446 tcp_pcap_add(th, m, &(tp->t_outpkts));
1449 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1450 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1453 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1454 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1460 * In transmit state, time the transmission and arrange for
1461 * the retransmit. In persist state, just set snd_max.
1463 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1464 !tcp_timer_active(tp, TT_PERSIST)) {
1465 tcp_seq startseq = tp->snd_nxt;
1468 * Advance snd_nxt over sequence space of this segment.
1470 if (flags & (TH_SYN|TH_FIN)) {
1473 if (flags & TH_FIN) {
1475 tp->t_flags |= TF_SENTFIN;
1481 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1482 tp->snd_max = tp->snd_nxt;
1484 * Time this transmission if not a retransmission and
1485 * not currently timing anything.
1487 if (tp->t_rtttime == 0) {
1488 tp->t_rtttime = ticks;
1489 tp->t_rtseq = startseq;
1490 TCPSTAT_INC(tcps_segstimed);
1495 * Set retransmit timer if not currently set,
1496 * and not doing a pure ack or a keep-alive probe.
1497 * Initial value for retransmit timer is smoothed
1498 * round-trip time + 2 * round-trip time variance.
1499 * Initialize shift counter which is used for backoff
1500 * of retransmit time.
1503 if (!tcp_timer_active(tp, TT_REXMT) &&
1504 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1505 (tp->snd_nxt != tp->snd_una))) {
1506 if (tcp_timer_active(tp, TT_PERSIST)) {
1507 tcp_timer_activate(tp, TT_PERSIST, 0);
1510 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1511 } else if (len == 0 && sbavail(&so->so_snd) &&
1512 !tcp_timer_active(tp, TT_REXMT) &&
1513 !tcp_timer_active(tp, TT_PERSIST)) {
1515 * Avoid a situation where we do not set persist timer
1516 * after a zero window condition. For example:
1517 * 1) A -> B: packet with enough data to fill the window
1518 * 2) B -> A: ACK for #1 + new data (0 window
1520 * 3) A -> B: ACK for #2, 0 len packet
1522 * In this case, A will not activate the persist timer,
1523 * because it chose to send a packet. Unless tcp_output
1524 * is called for some other reason (delayed ack timer,
1525 * another input packet from B, socket syscall), A will
1526 * not send zero window probes.
1528 * So, if you send a 0-length packet, but there is data
1529 * in the socket buffer, and neither the rexmt or
1530 * persist timer is already set, then activate the
1538 * Persist case, update snd_max but since we are in
1539 * persist mode (no window) we do not update snd_nxt.
1544 if (flags & TH_FIN) {
1546 tp->t_flags |= TF_SENTFIN;
1548 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1549 tp->snd_max = tp->snd_nxt + xlen;
1555 * We know that the packet was lost, so back out the
1556 * sequence number advance, if any.
1558 * If the error is EPERM the packet got blocked by the
1559 * local firewall. Normally we should terminate the
1560 * connection but the blocking may have been spurious
1561 * due to a firewall reconfiguration cycle. So we treat
1562 * it like a packet loss and let the retransmit timer and
1563 * timeouts do their work over time.
1564 * XXX: It is a POLA question whether calling tcp_drop right
1565 * away would be the really correct behavior instead.
1567 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1568 !tcp_timer_active(tp, TT_PERSIST)) &&
1569 ((flags & TH_SYN) == 0) &&
1573 tp->sackhint.sack_bytes_rexmit -= len;
1574 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1575 ("sackhint bytes rtx >= 0"));
1579 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1583 tp->t_softerror = error;
1586 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1587 tp->snd_cwnd = tp->t_maxseg;
1591 * For some reason the interface we used initially
1592 * to send segments changed to another or lowered
1594 * If TSO was active we either got an interface
1595 * without TSO capabilits or TSO was turned off.
1596 * If we obtained mtu from ip_output() then update
1600 tp->t_flags &= ~TF_TSO;
1602 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1610 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1611 tp->t_softerror = error;
1619 TCPSTAT_INC(tcps_sndtotal);
1622 * Data sent (as far as we can tell).
1623 * If this advertises a larger window than any other segment,
1624 * then remember the size of the advertised window.
1625 * Any pending ACK has now been sent.
1627 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1628 tp->rcv_adv = tp->rcv_nxt + recwin;
1629 tp->last_ack_sent = tp->rcv_nxt;
1630 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1631 if (tcp_timer_active(tp, TT_DELACK))
1632 tcp_timer_activate(tp, TT_DELACK, 0);
1635 * This completely breaks TCP if newreno is turned on. What happens
1636 * is that if delayed-acks are turned on on the receiver, this code
1637 * on the transmitter effectively destroys the TCP window, forcing
1638 * it to four packets (1.5Kx4 = 6K window).
1640 if (sendalot && --maxburst)
1649 tcp_setpersist(struct tcpcb *tp)
1651 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1654 tp->t_flags &= ~TF_PREVVALID;
1655 if (tcp_timer_active(tp, TT_REXMT))
1656 panic("tcp_setpersist: retransmit pending");
1658 * Start/restart persistence timer.
1660 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1661 tcp_persmin, tcp_persmax);
1662 tcp_timer_activate(tp, TT_PERSIST, tt);
1663 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1668 * Insert TCP options according to the supplied parameters to the place
1669 * optp in a consistent way. Can handle unaligned destinations.
1671 * The order of the option processing is crucial for optimal packing and
1672 * alignment for the scarce option space.
1674 * The optimal order for a SYN/SYN-ACK segment is:
1675 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1676 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1678 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1679 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1680 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1681 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1682 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1685 tcp_addoptions(struct tcpopt *to, u_char *optp)
1687 u_int32_t mask, optlen = 0;
1689 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1690 if ((to->to_flags & mask) != mask)
1692 if (optlen == TCP_MAXOLEN)
1694 switch (to->to_flags & mask) {
1696 while (optlen % 4) {
1697 optlen += TCPOLEN_NOP;
1698 *optp++ = TCPOPT_NOP;
1700 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1702 optlen += TCPOLEN_MAXSEG;
1703 *optp++ = TCPOPT_MAXSEG;
1704 *optp++ = TCPOLEN_MAXSEG;
1705 to->to_mss = htons(to->to_mss);
1706 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1707 optp += sizeof(to->to_mss);
1710 while (!optlen || optlen % 2 != 1) {
1711 optlen += TCPOLEN_NOP;
1712 *optp++ = TCPOPT_NOP;
1714 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1716 optlen += TCPOLEN_WINDOW;
1717 *optp++ = TCPOPT_WINDOW;
1718 *optp++ = TCPOLEN_WINDOW;
1719 *optp++ = to->to_wscale;
1722 while (optlen % 2) {
1723 optlen += TCPOLEN_NOP;
1724 *optp++ = TCPOPT_NOP;
1726 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1728 optlen += TCPOLEN_SACK_PERMITTED;
1729 *optp++ = TCPOPT_SACK_PERMITTED;
1730 *optp++ = TCPOLEN_SACK_PERMITTED;
1733 while (!optlen || optlen % 4 != 2) {
1734 optlen += TCPOLEN_NOP;
1735 *optp++ = TCPOPT_NOP;
1737 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1739 optlen += TCPOLEN_TIMESTAMP;
1740 *optp++ = TCPOPT_TIMESTAMP;
1741 *optp++ = TCPOLEN_TIMESTAMP;
1742 to->to_tsval = htonl(to->to_tsval);
1743 to->to_tsecr = htonl(to->to_tsecr);
1744 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1745 optp += sizeof(to->to_tsval);
1746 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1747 optp += sizeof(to->to_tsecr);
1751 int siglen = TCPOLEN_SIGNATURE - 2;
1753 while (!optlen || optlen % 4 != 2) {
1754 optlen += TCPOLEN_NOP;
1755 *optp++ = TCPOPT_NOP;
1757 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1758 to->to_flags &= ~TOF_SIGNATURE;
1761 optlen += TCPOLEN_SIGNATURE;
1762 *optp++ = TCPOPT_SIGNATURE;
1763 *optp++ = TCPOLEN_SIGNATURE;
1764 to->to_signature = optp;
1772 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1775 while (!optlen || optlen % 4 != 2) {
1776 optlen += TCPOLEN_NOP;
1777 *optp++ = TCPOPT_NOP;
1779 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1781 optlen += TCPOLEN_SACKHDR;
1782 *optp++ = TCPOPT_SACK;
1783 sackblks = min(to->to_nsacks,
1784 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1785 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1786 while (sackblks--) {
1787 sack_seq = htonl(sack->start);
1788 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1789 optp += sizeof(sack_seq);
1790 sack_seq = htonl(sack->end);
1791 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1792 optp += sizeof(sack_seq);
1793 optlen += TCPOLEN_SACK;
1796 TCPSTAT_INC(tcps_sack_send_blocks);
1804 /* XXX is there any point to aligning this option? */
1805 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1806 if (TCP_MAXOLEN - optlen < total_len)
1808 *optp++ = TCPOPT_FAST_OPEN;
1809 *optp++ = total_len;
1810 if (to->to_tfo_len > 0) {
1811 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1812 optp += to->to_tfo_len;
1814 optlen += total_len;
1819 panic("%s: unknown TCP option type", __func__);
1824 /* Terminate and pad TCP options to a 4 byte boundary. */
1826 optlen += TCPOLEN_EOL;
1827 *optp++ = TCPOPT_EOL;
1830 * According to RFC 793 (STD0007):
1831 * "The content of the header beyond the End-of-Option option
1832 * must be header padding (i.e., zero)."
1833 * and later: "The padding is composed of zeros."
1835 while (optlen % 4) {
1836 optlen += TCPOLEN_PAD;
1837 *optp++ = TCPOPT_PAD;
1840 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));