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_kdtrace.h"
39 #include "opt_tcpdebug.h"
41 #include <sys/param.h>
42 #include <sys/systm.h>
43 #include <sys/domain.h>
44 #include <sys/hhook.h>
45 #include <sys/kernel.h>
48 #include <sys/mutex.h>
49 #include <sys/protosw.h>
51 #include <sys/socket.h>
52 #include <sys/socketvar.h>
53 #include <sys/sysctl.h>
56 #include <net/route.h>
59 #include <netinet/cc.h>
60 #include <netinet/in.h>
61 #include <netinet/in_kdtrace.h>
62 #include <netinet/in_systm.h>
63 #include <netinet/ip.h>
64 #include <netinet/in_pcb.h>
65 #include <netinet/ip_var.h>
66 #include <netinet/ip_options.h>
68 #include <netinet6/in6_pcb.h>
69 #include <netinet/ip6.h>
70 #include <netinet6/ip6_var.h>
73 #include <netinet/tcp_fastopen.h>
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/tcp_seq.h>
78 #include <netinet/tcp_timer.h>
79 #include <netinet/tcp_var.h>
80 #include <netinet/tcpip.h>
82 #include <netinet/tcp_debug.h>
85 #include <netinet/tcp_offload.h>
89 #include <netipsec/ipsec.h>
92 #include <machine/in_cksum.h>
94 #include <security/mac/mac_framework.h>
96 VNET_DEFINE(int, path_mtu_discovery) = 1;
97 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_RW,
98 &VNET_NAME(path_mtu_discovery), 1,
99 "Enable Path MTU Discovery");
101 VNET_DEFINE(int, tcp_do_tso) = 1;
102 #define V_tcp_do_tso VNET(tcp_do_tso)
103 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_RW,
104 &VNET_NAME(tcp_do_tso), 0,
105 "Enable TCP Segmentation Offload");
107 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
108 #define V_tcp_sendspace VNET(tcp_sendspace)
109 SYSCTL_VNET_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_RW,
110 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
112 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
113 #define V_tcp_do_autosndbuf VNET(tcp_do_autosndbuf)
114 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_RW,
115 &VNET_NAME(tcp_do_autosndbuf), 0,
116 "Enable automatic send buffer sizing");
118 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
119 #define V_tcp_autosndbuf_inc VNET(tcp_autosndbuf_inc)
120 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_RW,
121 &VNET_NAME(tcp_autosndbuf_inc), 0,
122 "Incrementor step size of automatic send buffer");
124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
125 #define V_tcp_autosndbuf_max VNET(tcp_autosndbuf_max)
126 SYSCTL_VNET_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_RW,
127 &VNET_NAME(tcp_autosndbuf_max), 0,
128 "Max size of automatic send buffer");
131 * Make sure that either retransmit or persist timer is set for SYN, FIN and
134 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
135 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
136 tcp_timer_active((tp), TT_REXMT) || \
137 tcp_timer_active((tp), TT_PERSIST), \
138 ("neither rexmt nor persist timer is set"))
140 static void inline hhook_run_tcp_est_out(struct tcpcb *tp,
141 struct tcphdr *th, struct tcpopt *to,
143 static void inline cc_after_idle(struct tcpcb *tp);
146 * Wrapper for the TCP established output helper hook.
149 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
150 struct tcpopt *to, long len, int tso)
152 struct tcp_hhook_data hhook_data;
154 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
158 hhook_data.len = len;
159 hhook_data.tso = tso;
161 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
167 * CC wrapper hook functions
170 cc_after_idle(struct tcpcb *tp)
172 INP_WLOCK_ASSERT(tp->t_inpcb);
174 if (CC_ALGO(tp)->after_idle != NULL)
175 CC_ALGO(tp)->after_idle(tp->ccv);
179 * Tcp output routine: figure out what should be sent and send it.
182 tcp_output(struct tcpcb *tp)
184 struct socket *so = tp->t_inpcb->inp_socket;
185 long len, recwin, sendwin;
186 int off, flags, error = 0; /* Keep compiler happy */
188 struct ip *ip = NULL;
189 struct ipovly *ipov = NULL;
191 u_char opt[TCP_MAXOLEN];
192 unsigned ipoptlen, optlen, hdrlen;
194 unsigned ipsec_optlen = 0;
197 int sack_rxmit, sack_bytes_rxmt;
202 int maxburst = TCP_MAXBURST;
205 struct ip6_hdr *ip6 = NULL;
208 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
211 INP_WLOCK_ASSERT(tp->t_inpcb);
214 if (tp->t_flags & TF_TOE)
215 return (tcp_offload_output(tp));
220 * For TFO connections in SYN_RECEIVED, only allow the initial
221 * SYN|ACK and those sent by the retransmit timer.
223 if ((tp->t_flags & TF_FASTOPEN) &&
224 (tp->t_state == TCPS_SYN_RECEIVED) &&
225 SEQ_GT(tp->snd_max, tp->snd_una) && /* inital SYN|ACK sent */
226 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
230 * Determine length of data that should be transmitted,
231 * and flags that will be used.
232 * If there is some data or critical controls (SYN, RST)
233 * to send, then transmit; otherwise, investigate further.
235 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
236 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
238 tp->t_flags &= ~TF_LASTIDLE;
240 if (tp->t_flags & TF_MORETOCOME) {
241 tp->t_flags |= TF_LASTIDLE;
247 * If we've recently taken a timeout, snd_max will be greater than
248 * snd_nxt. There may be SACK information that allows us to avoid
249 * resending already delivered data. Adjust snd_nxt accordingly.
251 if ((tp->t_flags & TF_SACK_PERMIT) &&
252 SEQ_LT(tp->snd_nxt, tp->snd_max))
257 off = tp->snd_nxt - tp->snd_una;
258 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
260 flags = tcp_outflags[tp->t_state];
262 * Send any SACK-generated retransmissions. If we're explicitly trying
263 * to send out new data (when sendalot is 1), bypass this function.
264 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
265 * we're replacing a (future) new transmission with a retransmission
266 * now, and we previously incremented snd_cwnd in tcp_input().
269 * Still in sack recovery , reset rxmit flag to zero.
275 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
276 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
279 cwin = min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt;
282 /* Do not retransmit SACK segments beyond snd_recover */
283 if (SEQ_GT(p->end, tp->snd_recover)) {
285 * (At least) part of sack hole extends beyond
286 * snd_recover. Check to see if we can rexmit data
289 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
291 * Can't rexmit any more data for this hole.
292 * That data will be rexmitted in the next
293 * sack recovery episode, when snd_recover
294 * moves past p->rxmit.
297 goto after_sack_rexmit;
299 /* Can rexmit part of the current hole */
300 len = ((long)ulmin(cwin,
301 tp->snd_recover - p->rxmit));
303 len = ((long)ulmin(cwin, p->end - p->rxmit));
304 off = p->rxmit - tp->snd_una;
305 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
310 TCPSTAT_INC(tcps_sack_rexmits);
311 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
312 min(len, tp->t_maxseg));
317 * Get standard flags, and add SYN or FIN if requested by 'hidden'
320 if (tp->t_flags & TF_NEEDFIN)
322 if (tp->t_flags & TF_NEEDSYN)
325 SOCKBUF_LOCK(&so->so_snd);
327 * If in persist timeout with window of 0, send 1 byte.
328 * Otherwise, if window is small but nonzero
329 * and timer expired, we will send what we can
330 * and go to transmit state.
332 if (tp->t_flags & TF_FORCEDATA) {
335 * If we still have some data to send, then
336 * clear the FIN bit. Usually this would
337 * happen below when it realizes that we
338 * aren't sending all the data. However,
339 * if we have exactly 1 byte of unsent data,
340 * then it won't clear the FIN bit below,
341 * and if we are in persist state, we wind
342 * up sending the packet without recording
343 * that we sent the FIN bit.
345 * We can't just blindly clear the FIN bit,
346 * because if we don't have any more data
347 * to send then the probe will be the FIN
350 if (off < so->so_snd.sb_cc)
354 tcp_timer_activate(tp, TT_PERSIST, 0);
360 * If snd_nxt == snd_max and we have transmitted a FIN, the
361 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
362 * a negative length. This can also occur when TCP opens up
363 * its congestion window while receiving additional duplicate
364 * acks after fast-retransmit because TCP will reset snd_nxt
365 * to snd_max after the fast-retransmit.
367 * In the normal retransmit-FIN-only case, however, snd_nxt will
368 * be set to snd_una, the offset will be 0, and the length may
371 * If sack_rxmit is true we are retransmitting from the scoreboard
372 * in which case len is already set.
374 if (sack_rxmit == 0) {
375 if (sack_bytes_rxmt == 0)
376 len = ((long)ulmin(so->so_snd.sb_cc, sendwin) - off);
381 * We are inside of a SACK recovery episode and are
382 * sending new data, having retransmitted all the
383 * data possible in the scoreboard.
385 len = ((long)ulmin(so->so_snd.sb_cc, tp->snd_wnd)
388 * Don't remove this (len > 0) check !
389 * We explicitly check for len > 0 here (although it
390 * isn't really necessary), to work around a gcc
391 * optimization issue - to force gcc to compute
392 * len above. Without this check, the computation
393 * of len is bungled by the optimizer.
396 cwin = tp->snd_cwnd -
397 (tp->snd_nxt - tp->sack_newdata) -
401 len = lmin(len, cwin);
407 * Lop off SYN bit if it has already been sent. However, if this
408 * is SYN-SENT state and if segment contains data and if we don't
409 * know that foreign host supports TAO, suppress sending segment.
411 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
412 if (tp->t_state != TCPS_SYN_RECEIVED)
416 * When sending additional segments following a TFO SYN|ACK,
417 * do not include the SYN bit.
419 if ((tp->t_flags & TF_FASTOPEN) &&
420 (tp->t_state == TCPS_SYN_RECEIVED))
427 * Be careful not to send data and/or FIN on SYN segments.
428 * This measure is needed to prevent interoperability problems
429 * with not fully conformant TCP implementations.
431 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
438 * When retransmitting SYN|ACK on a passively-created TFO socket,
439 * don't include data, as the presence of data may have caused the
440 * original SYN|ACK to have been dropped by a middlebox.
442 if ((tp->t_flags & TF_FASTOPEN) &&
443 (((tp->t_state == TCPS_SYN_RECEIVED) && (tp->t_rxtshift > 0)) ||
449 * If FIN has been sent but not acked,
450 * but we haven't been called to retransmit,
451 * len will be < 0. Otherwise, window shrank
452 * after we sent into it. If window shrank to 0,
453 * cancel pending retransmit, pull snd_nxt back
454 * to (closed) window, and set the persist timer
455 * if it isn't already going. If the window didn't
456 * close completely, just wait for an ACK.
458 * We also do a general check here to ensure that
459 * we will set the persist timer when we have data
460 * to send, but a 0-byte window. This makes sure
461 * the persist timer is set even if the packet
462 * hits one of the "goto send" lines below.
465 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
466 (off < (int) so->so_snd.sb_cc)) {
467 tcp_timer_activate(tp, TT_REXMT, 0);
469 tp->snd_nxt = tp->snd_una;
470 if (!tcp_timer_active(tp, TT_PERSIST))
475 /* len will be >= 0 after this point. */
476 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
479 * Automatic sizing of send socket buffer. Often the send buffer
480 * size is not optimally adjusted to the actual network conditions
481 * at hand (delay bandwidth product). Setting the buffer size too
482 * small limits throughput on links with high bandwidth and high
483 * delay (eg. trans-continental/oceanic links). Setting the
484 * buffer size too big consumes too much real kernel memory,
485 * especially with many connections on busy servers.
487 * The criteria to step up the send buffer one notch are:
488 * 1. receive window of remote host is larger than send buffer
489 * (with a fudge factor of 5/4th);
490 * 2. send buffer is filled to 7/8th with data (so we actually
491 * have data to make use of it);
492 * 3. send buffer fill has not hit maximal automatic size;
493 * 4. our send window (slow start and cogestion controlled) is
494 * larger than sent but unacknowledged data in send buffer.
496 * The remote host receive window scaling factor may limit the
497 * growing of the send buffer before it reaches its allowed
500 * It scales directly with slow start or congestion window
501 * and does at most one step per received ACK. This fast
502 * scaling has the drawback of growing the send buffer beyond
503 * what is strictly necessary to make full use of a given
504 * delay*bandwith product. However testing has shown this not
505 * to be much of an problem. At worst we are trading wasting
506 * of available bandwith (the non-use of it) for wasting some
507 * socket buffer memory.
509 * TODO: Shrink send buffer during idle periods together
510 * with congestion window. Requires another timer. Has to
511 * wait for upcoming tcp timer rewrite.
513 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
514 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat &&
515 so->so_snd.sb_cc >= (so->so_snd.sb_hiwat / 8 * 7) &&
516 so->so_snd.sb_cc < V_tcp_autosndbuf_max &&
517 sendwin >= (so->so_snd.sb_cc - (tp->snd_nxt - tp->snd_una))) {
518 if (!sbreserve_locked(&so->so_snd,
519 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
520 V_tcp_autosndbuf_max), so, curthread))
521 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
526 * Decide if we can use TCP Segmentation Offloading (if supported by
529 * TSO may only be used if we are in a pure bulk sending state. The
530 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
531 * IP options prevent using TSO. With TSO the TCP header is the same
532 * (except for the sequence number) for all generated packets. This
533 * makes it impossible to transmit any options which vary per generated
538 * Pre-calculate here as we save another lookup into the darknesses
539 * of IPsec that way and can actually decide if TSO is ok.
541 ipsec_optlen = ipsec_hdrsiz_tcp(tp);
543 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
544 ((tp->t_flags & TF_SIGNATURE) == 0) &&
545 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
549 tp->t_inpcb->inp_options == NULL &&
550 tp->t_inpcb->in6p_options == NULL)
554 if (SEQ_LT(p->rxmit + len, tp->snd_una + so->so_snd.sb_cc))
557 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una + so->so_snd.sb_cc))
561 recwin = sbspace(&so->so_rcv);
564 * Sender silly window avoidance. We transmit under the following
565 * conditions when len is non-zero:
567 * - We have a full segment (or more with TSO)
568 * - This is the last buffer in a write()/send() and we are
569 * either idle or running NODELAY
570 * - we've timed out (e.g. persist timer)
571 * - we have more then 1/2 the maximum send window's worth of
572 * data (receiver may be limited the window size)
573 * - we need to retransmit
576 if (len >= tp->t_maxseg)
579 * NOTE! on localhost connections an 'ack' from the remote
580 * end may occur synchronously with the output and cause
581 * us to flush a buffer queued with moretocome. XXX
583 * note: the len + off check is almost certainly unnecessary.
585 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
586 (idle || (tp->t_flags & TF_NODELAY)) &&
587 len + off >= so->so_snd.sb_cc &&
588 (tp->t_flags & TF_NOPUSH) == 0) {
591 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
593 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
595 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
602 * Sending of standalone window updates.
604 * Window updates are important when we close our window due to a
605 * full socket buffer and are opening it again after the application
606 * reads data from it. Once the window has opened again and the
607 * remote end starts to send again the ACK clock takes over and
608 * provides the most current window information.
610 * We must avoid the silly window syndrome whereas every read
611 * from the receive buffer, no matter how small, causes a window
612 * update to be sent. We also should avoid sending a flurry of
613 * window updates when the socket buffer had queued a lot of data
614 * and the application is doing small reads.
616 * Prevent a flurry of pointless window updates by only sending
617 * an update when we can increase the advertized window by more
618 * than 1/4th of the socket buffer capacity. When the buffer is
619 * getting full or is very small be more aggressive and send an
620 * update whenever we can increase by two mss sized segments.
621 * In all other situations the ACK's to new incoming data will
622 * carry further window increases.
624 * Don't send an independent window update if a delayed
625 * ACK is pending (it will get piggy-backed on it) or the
626 * remote side already has done a half-close and won't send
627 * more data. Skip this if the connection is in T/TCP
630 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
631 !(tp->t_flags & TF_DELACK) &&
632 !TCPS_HAVERCVDFIN(tp->t_state)) {
634 * "adv" is the amount we could increase the window,
635 * taking into account that we are limited by
636 * TCP_MAXWIN << tp->rcv_scale.
641 adv = min(recwin, (long)TCP_MAXWIN << tp->rcv_scale);
642 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
643 oldwin = (tp->rcv_adv - tp->rcv_nxt);
649 * If the new window size ends up being the same as the old
650 * size when it is scaled, then don't force a window update.
652 if (oldwin >> tp->rcv_scale == (adv + oldwin) >> tp->rcv_scale)
655 if (adv >= (long)(2 * tp->t_maxseg) &&
656 (adv >= (long)(so->so_rcv.sb_hiwat / 4) ||
657 recwin <= (long)(so->so_rcv.sb_hiwat / 8) ||
658 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
664 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
665 * is also a catch-all for the retransmit timer timeout case.
667 if (tp->t_flags & TF_ACKNOW)
669 if ((flags & TH_RST) ||
670 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
672 if (SEQ_GT(tp->snd_up, tp->snd_una))
675 * If our state indicates that FIN should be sent
676 * and we have not yet done so, then we need to send.
678 if (flags & TH_FIN &&
679 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
682 * In SACK, it is possible for tcp_output to fail to send a segment
683 * after the retransmission timer has been turned off. Make sure
684 * that the retransmission timer is set.
686 if ((tp->t_flags & TF_SACK_PERMIT) &&
687 SEQ_GT(tp->snd_max, tp->snd_una) &&
688 !tcp_timer_active(tp, TT_REXMT) &&
689 !tcp_timer_active(tp, TT_PERSIST)) {
690 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
694 * TCP window updates are not reliable, rather a polling protocol
695 * using ``persist'' packets is used to insure receipt of window
696 * updates. The three ``states'' for the output side are:
697 * idle not doing retransmits or persists
698 * persisting to move a small or zero window
699 * (re)transmitting and thereby not persisting
701 * tcp_timer_active(tp, TT_PERSIST)
702 * is true when we are in persist state.
703 * (tp->t_flags & TF_FORCEDATA)
704 * is set when we are called to send a persist packet.
705 * tcp_timer_active(tp, TT_REXMT)
706 * is set when we are retransmitting
707 * The output side is idle when both timers are zero.
709 * If send window is too small, there is data to transmit, and no
710 * retransmit or persist is pending, then go to persist state.
711 * If nothing happens soon, send when timer expires:
712 * if window is nonzero, transmit what we can,
713 * otherwise force out a byte.
715 if (so->so_snd.sb_cc && !tcp_timer_active(tp, TT_REXMT) &&
716 !tcp_timer_active(tp, TT_PERSIST)) {
722 * No reason to send a segment, just return.
725 SOCKBUF_UNLOCK(&so->so_snd);
729 SOCKBUF_LOCK_ASSERT(&so->so_snd);
731 if (len >= tp->t_maxseg)
732 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
734 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
737 * Before ESTABLISHED, force sending of initial options
738 * unless TCP set not to do any options.
739 * NOTE: we assume that the IP/TCP header plus TCP options
740 * always fit in a single mbuf, leaving room for a maximum
742 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
747 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
750 hdrlen = sizeof (struct tcpiphdr);
753 * Compute options for segment.
754 * We only have to care about SYN and established connection
755 * segments. Options for SYN-ACK segments are handled in TCP
759 if ((tp->t_flags & TF_NOOPT) == 0) {
760 /* Maximum segment size. */
761 if (flags & TH_SYN) {
762 tp->snd_nxt = tp->iss;
763 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
764 to.to_flags |= TOF_MSS;
767 * Only include the TFO option on the first
768 * transmission of the SYN|ACK on a
769 * passively-created TFO socket, as the presence of
770 * the TFO option may have caused the original
771 * SYN|ACK to have been dropped by a middlebox.
773 if ((tp->t_flags & TF_FASTOPEN) &&
774 (tp->t_state == TCPS_SYN_RECEIVED) &&
775 (tp->t_rxtshift == 0)) {
776 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
777 to.to_tfo_cookie = (u_char *)&tp->t_tfo_cookie;
778 to.to_flags |= TOF_FASTOPEN;
782 /* Window scaling. */
783 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
784 to.to_wscale = tp->request_r_scale;
785 to.to_flags |= TOF_SCALE;
788 if ((tp->t_flags & TF_RCVD_TSTMP) ||
789 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
790 to.to_tsval = tcp_ts_getticks() + tp->ts_offset;
791 to.to_tsecr = tp->ts_recent;
792 to.to_flags |= TOF_TS;
795 /* Set receive buffer autosizing timestamp. */
796 if (tp->rfbuf_ts == 0 &&
797 (so->so_rcv.sb_flags & SB_AUTOSIZE))
798 tp->rfbuf_ts = tcp_ts_getticks();
800 /* Selective ACK's. */
801 if (tp->t_flags & TF_SACK_PERMIT) {
803 to.to_flags |= TOF_SACKPERM;
804 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
805 (tp->t_flags & TF_SACK_PERMIT) &&
806 tp->rcv_numsacks > 0) {
807 to.to_flags |= TOF_SACK;
808 to.to_nsacks = tp->rcv_numsacks;
809 to.to_sacks = (u_char *)tp->sackblks;
813 /* TCP-MD5 (RFC2385). */
814 if (tp->t_flags & TF_SIGNATURE)
815 to.to_flags |= TOF_SIGNATURE;
816 #endif /* TCP_SIGNATURE */
818 /* Processing the options. */
819 hdrlen += optlen = tcp_addoptions(&to, opt);
824 ipoptlen = ip6_optlen(tp->t_inpcb);
827 if (tp->t_inpcb->inp_options)
828 ipoptlen = tp->t_inpcb->inp_options->m_len -
829 offsetof(struct ipoption, ipopt_list);
833 ipoptlen += ipsec_optlen;
837 * Adjust data length if insertion of options will
838 * bump the packet length beyond the t_maxopd length.
839 * Clear the FIN bit because we cut off the tail of
842 if (len + optlen + ipoptlen > tp->t_maxopd) {
847 u_int if_hw_tsomaxsegcount;
848 u_int if_hw_tsomaxsegsize;
853 /* extract TSO information */
854 if_hw_tsomax = tp->t_tsomax;
855 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
856 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
859 * Limit a TSO burst to prevent it from
860 * overflowing or exceeding the maximum length
861 * allowed by the network interface:
863 KASSERT(ipoptlen == 0,
864 ("%s: TSO can't do IP options", __func__));
867 * Check if we should limit by maximum payload
870 if (if_hw_tsomax != 0) {
871 /* compute maximum TSO length */
872 max_len = (if_hw_tsomax - hdrlen -
876 } else if (len > max_len) {
883 * Check if we should limit by maximum segment
886 if (if_hw_tsomaxsegcount != 0 &&
887 if_hw_tsomaxsegsize != 0) {
889 * Subtract one segment for the LINK
890 * and TCP/IP headers mbuf that will
891 * be prepended to this mbuf chain
892 * after the code in this section
893 * limits the number of mbufs in the
894 * chain to if_hw_tsomaxsegcount.
896 if_hw_tsomaxsegcount -= 1;
898 mb = sbsndmbuf(&so->so_snd, off, &moff);
900 while (mb != NULL && max_len < len) {
905 * Get length of mbuf fragment
906 * and how many hardware frags,
907 * rounded up, it would use:
909 mlen = (mb->m_len - moff);
910 frags = howmany(mlen,
911 if_hw_tsomaxsegsize);
913 /* Handle special case: Zero Length Mbuf */
918 * Check if the fragment limit
919 * will be reached or exceeded:
921 if (frags >= if_hw_tsomaxsegcount) {
923 if_hw_tsomaxsegcount *
924 if_hw_tsomaxsegsize);
928 if_hw_tsomaxsegcount -= frags;
934 } else if (len > max_len) {
941 * Prevent the last segment from being
942 * fractional unless the send sockbuf can be
945 max_len = (tp->t_maxopd - optlen);
946 if ((off + len) < so->so_snd.sb_cc) {
947 moff = len % max_len;
955 * In case there are too many small fragments
958 if (len <= max_len) {
965 * Send the FIN in a separate segment
966 * after the bulk sending is done.
967 * We don't trust the TSO implementations
968 * to clear the FIN flag on all but the
971 if (tp->t_flags & TF_NEEDFIN)
975 len = tp->t_maxopd - optlen - ipoptlen;
981 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
982 ("%s: len > IP_MAXPACKET", __func__));
984 /*#ifdef DIAGNOSTIC*/
986 if (max_linkhdr + hdrlen > MCLBYTES)
988 if (max_linkhdr + hdrlen > MHLEN)
990 panic("tcphdr too big");
994 * This KASSERT is here to catch edge cases at a well defined place.
995 * Before, those had triggered (random) panic conditions further down.
997 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1000 * Grab a header mbuf, attaching a copy of data to
1001 * be transmitted, and initialize the header from
1002 * the template for sends on this connection.
1008 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1009 TCPSTAT_INC(tcps_sndprobe);
1010 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1011 tp->t_sndrexmitpack++;
1012 TCPSTAT_INC(tcps_sndrexmitpack);
1013 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1015 TCPSTAT_INC(tcps_sndpack);
1016 TCPSTAT_ADD(tcps_sndbyte, len);
1019 if (MHLEN < hdrlen + max_linkhdr)
1020 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1023 m = m_gethdr(M_NOWAIT, MT_DATA);
1026 SOCKBUF_UNLOCK(&so->so_snd);
1032 m->m_data += max_linkhdr;
1036 * Start the m_copy functions from the closest mbuf
1037 * to the offset in the socket buffer chain.
1039 mb = sbsndptr(&so->so_snd, off, len, &moff);
1041 if (len <= MHLEN - hdrlen - max_linkhdr) {
1042 m_copydata(mb, moff, (int)len,
1043 mtod(m, caddr_t) + hdrlen);
1046 m->m_next = m_copy(mb, moff, (int)len);
1047 if (m->m_next == NULL) {
1048 SOCKBUF_UNLOCK(&so->so_snd);
1057 * If we're sending everything we've got, set PUSH.
1058 * (This will keep happy those implementations which only
1059 * give data to the user when a buffer fills or
1062 if ((off + len == so->so_snd.sb_cc) && !(flags & TH_SYN))
1064 SOCKBUF_UNLOCK(&so->so_snd);
1066 SOCKBUF_UNLOCK(&so->so_snd);
1067 if (tp->t_flags & TF_ACKNOW)
1068 TCPSTAT_INC(tcps_sndacks);
1069 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1070 TCPSTAT_INC(tcps_sndctrl);
1071 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1072 TCPSTAT_INC(tcps_sndurg);
1074 TCPSTAT_INC(tcps_sndwinup);
1076 m = m_gethdr(M_NOWAIT, MT_DATA);
1083 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1085 MH_ALIGN(m, hdrlen);
1088 m->m_data += max_linkhdr;
1091 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1092 m->m_pkthdr.rcvif = (struct ifnet *)0;
1094 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1098 ip6 = mtod(m, struct ip6_hdr *);
1099 th = (struct tcphdr *)(ip6 + 1);
1100 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1104 ip = mtod(m, struct ip *);
1105 ipov = (struct ipovly *)ip;
1106 th = (struct tcphdr *)(ip + 1);
1107 tcpip_fillheaders(tp->t_inpcb, ip, th);
1111 * Fill in fields, remembering maximum advertised
1112 * window for use in delaying messages about window sizes.
1113 * If resending a FIN, be sure not to use a new sequence number.
1115 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1116 tp->snd_nxt == tp->snd_max)
1119 * If we are starting a connection, send ECN setup
1120 * SYN packet. If we are on a retransmit, we may
1121 * resend those bits a number of times as per
1124 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1125 if (tp->t_rxtshift >= 1) {
1126 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1127 flags |= TH_ECE|TH_CWR;
1129 flags |= TH_ECE|TH_CWR;
1132 if (tp->t_state == TCPS_ESTABLISHED &&
1133 (tp->t_flags & TF_ECN_PERMIT)) {
1135 * If the peer has ECN, mark data packets with
1136 * ECN capable transmission (ECT).
1137 * Ignore pure ack packets, retransmissions and window probes.
1139 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1140 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1143 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1146 ip->ip_tos |= IPTOS_ECN_ECT0;
1147 TCPSTAT_INC(tcps_ecn_ect0);
1151 * Reply with proper ECN notifications.
1153 if (tp->t_flags & TF_ECN_SND_CWR) {
1155 tp->t_flags &= ~TF_ECN_SND_CWR;
1157 if (tp->t_flags & TF_ECN_SND_ECE)
1162 * If we are doing retransmissions, then snd_nxt will
1163 * not reflect the first unsent octet. For ACK only
1164 * packets, we do not want the sequence number of the
1165 * retransmitted packet, we want the sequence number
1166 * of the next unsent octet. So, if there is no data
1167 * (and no SYN or FIN), use snd_max instead of snd_nxt
1168 * when filling in ti_seq. But if we are in persist
1169 * state, snd_max might reflect one byte beyond the
1170 * right edge of the window, so use snd_nxt in that
1171 * case, since we know we aren't doing a retransmission.
1172 * (retransmit and persist are mutually exclusive...)
1174 if (sack_rxmit == 0) {
1175 if (len || (flags & (TH_SYN|TH_FIN)) ||
1176 tcp_timer_active(tp, TT_PERSIST))
1177 th->th_seq = htonl(tp->snd_nxt);
1179 th->th_seq = htonl(tp->snd_max);
1181 th->th_seq = htonl(p->rxmit);
1183 tp->sackhint.sack_bytes_rexmit += len;
1185 th->th_ack = htonl(tp->rcv_nxt);
1187 bcopy(opt, th + 1, optlen);
1188 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1190 th->th_flags = flags;
1192 * Calculate receive window. Don't shrink window,
1193 * but avoid silly window syndrome.
1195 if (recwin < (long)(so->so_rcv.sb_hiwat / 4) &&
1196 recwin < (long)tp->t_maxseg)
1198 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1199 recwin < (long)(tp->rcv_adv - tp->rcv_nxt))
1200 recwin = (long)(tp->rcv_adv - tp->rcv_nxt);
1201 if (recwin > (long)TCP_MAXWIN << tp->rcv_scale)
1202 recwin = (long)TCP_MAXWIN << tp->rcv_scale;
1205 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1206 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1207 * case is handled in syncache.
1210 th->th_win = htons((u_short)
1211 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1213 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1216 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1217 * a 0 window. This may cause the remote transmitter to stall. This
1218 * flag tells soreceive() to disable delayed acknowledgements when
1219 * draining the buffer. This can occur if the receiver is attempting
1220 * to read more data than can be buffered prior to transmitting on
1223 if (th->th_win == 0) {
1225 tp->t_flags |= TF_RXWIN0SENT;
1227 tp->t_flags &= ~TF_RXWIN0SENT;
1228 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1229 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1230 th->th_flags |= TH_URG;
1233 * If no urgent pointer to send, then we pull
1234 * the urgent pointer to the left edge of the send window
1235 * so that it doesn't drift into the send window on sequence
1236 * number wraparound.
1238 tp->snd_up = tp->snd_una; /* drag it along */
1240 #ifdef TCP_SIGNATURE
1241 if (to.to_flags & TOF_SIGNATURE) {
1242 int sigoff = to.to_signature - opt;
1243 tcp_signature_compute(m, 0, len, optlen,
1244 (u_char *)(th + 1) + sigoff, IPSEC_DIR_OUTBOUND);
1249 * Put TCP length in extended header, and then
1250 * checksum extended header and data.
1252 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1253 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1257 * There is no need to fill in ip6_plen right now.
1258 * It will be filled later by ip6_output.
1260 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1261 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1262 optlen + len, IPPROTO_TCP, 0);
1265 #if defined(INET6) && defined(INET)
1270 m->m_pkthdr.csum_flags = CSUM_TCP;
1271 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1272 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1274 /* IP version must be set here for ipv4/ipv6 checking later */
1275 KASSERT(ip->ip_v == IPVERSION,
1276 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1281 * Enable TSO and specify the size of the segments.
1282 * The TCP pseudo header checksum is always provided.
1283 * XXX: Fixme: This is currently not the case for IPv6.
1286 KASSERT(len > tp->t_maxopd - optlen,
1287 ("%s: len <= tso_segsz", __func__));
1288 m->m_pkthdr.csum_flags |= CSUM_TSO;
1289 m->m_pkthdr.tso_segsz = tp->t_maxopd - optlen;
1293 KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1294 ("%s: mbuf chain shorter than expected: %ld + %u + %u - %u != %u",
1295 __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1297 KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1298 ("%s: mbuf chain shorter than expected: %ld + %u + %u != %u",
1299 __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1302 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1303 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1309 if (so->so_options & SO_DEBUG) {
1315 save = ipov->ih_len;
1316 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1318 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1322 ipov->ih_len = save;
1324 #endif /* TCPDEBUG */
1327 * Fill in IP length and desired time to live and
1328 * send to IP level. There should be a better way
1329 * to handle ttl and tos; we could keep them in
1330 * the template, but need a way to checksum without them.
1333 * m->m_pkthdr.len should have been set before cksum calcuration,
1334 * because in6_cksum() need it.
1338 struct route_in6 ro;
1340 bzero(&ro, sizeof(ro));
1342 * we separately set hoplimit for every segment, since the
1343 * user might want to change the value via setsockopt.
1344 * Also, desired default hop limit might be changed via
1345 * Neighbor Discovery.
1347 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1350 * Set the packet size here for the benefit of DTrace probes.
1351 * ip6_output() will set it properly; it's supposed to include
1352 * the option header lengths as well.
1354 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1356 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss)
1357 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1359 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1361 if (tp->t_state == TCPS_SYN_SENT)
1362 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1364 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1366 /* TODO: IPv6 IP6TOS_ECT bit on */
1367 error = ip6_output(m, tp->t_inpcb->in6p_outputopts, &ro,
1368 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1369 NULL, NULL, tp->t_inpcb);
1371 if (error == EMSGSIZE && ro.ro_rt != NULL)
1372 mtu = ro.ro_rt->rt_mtu;
1376 #if defined(INET) && defined(INET6)
1383 bzero(&ro, sizeof(ro));
1384 ip->ip_len = htons(m->m_pkthdr.len);
1386 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1387 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1390 * If we do path MTU discovery, then we set DF on every packet.
1391 * This might not be the best thing to do according to RFC3390
1392 * Section 2. However the tcp hostcache migitates the problem
1393 * so it affects only the first tcp connection with a host.
1395 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1397 if (V_path_mtu_discovery && tp->t_maxopd > V_tcp_minmss) {
1398 ip->ip_off |= htons(IP_DF);
1399 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1401 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1404 if (tp->t_state == TCPS_SYN_SENT)
1405 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1407 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1409 error = ip_output(m, tp->t_inpcb->inp_options, &ro,
1410 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1413 if (error == EMSGSIZE && ro.ro_rt != NULL)
1414 mtu = ro.ro_rt->rt_mtu;
1421 * In transmit state, time the transmission and arrange for
1422 * the retransmit. In persist state, just set snd_max.
1424 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1425 !tcp_timer_active(tp, TT_PERSIST)) {
1426 tcp_seq startseq = tp->snd_nxt;
1429 * Advance snd_nxt over sequence space of this segment.
1431 if (flags & (TH_SYN|TH_FIN)) {
1434 if (flags & TH_FIN) {
1436 tp->t_flags |= TF_SENTFIN;
1442 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1443 tp->snd_max = tp->snd_nxt;
1445 * Time this transmission if not a retransmission and
1446 * not currently timing anything.
1448 if (tp->t_rtttime == 0) {
1449 tp->t_rtttime = ticks;
1450 tp->t_rtseq = startseq;
1451 TCPSTAT_INC(tcps_segstimed);
1456 * Set retransmit timer if not currently set,
1457 * and not doing a pure ack or a keep-alive probe.
1458 * Initial value for retransmit timer is smoothed
1459 * round-trip time + 2 * round-trip time variance.
1460 * Initialize shift counter which is used for backoff
1461 * of retransmit time.
1464 if (!tcp_timer_active(tp, TT_REXMT) &&
1465 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1466 (tp->snd_nxt != tp->snd_una))) {
1467 if (tcp_timer_active(tp, TT_PERSIST)) {
1468 tcp_timer_activate(tp, TT_PERSIST, 0);
1471 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1472 } else if (len == 0 && so->so_snd.sb_cc &&
1473 !tcp_timer_active(tp, TT_REXMT) &&
1474 !tcp_timer_active(tp, TT_PERSIST)) {
1476 * Avoid a situation where we do not set persist timer
1477 * after a zero window condition. For example:
1478 * 1) A -> B: packet with enough data to fill the window
1479 * 2) B -> A: ACK for #1 + new data (0 window
1481 * 3) A -> B: ACK for #2, 0 len packet
1483 * In this case, A will not activate the persist timer,
1484 * because it chose to send a packet. Unless tcp_output
1485 * is called for some other reason (delayed ack timer,
1486 * another input packet from B, socket syscall), A will
1487 * not send zero window probes.
1489 * So, if you send a 0-length packet, but there is data
1490 * in the socket buffer, and neither the rexmt or
1491 * persist timer is already set, then activate the
1499 * Persist case, update snd_max but since we are in
1500 * persist mode (no window) we do not update snd_nxt.
1505 if (flags & TH_FIN) {
1507 tp->t_flags |= TF_SENTFIN;
1509 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1510 tp->snd_max = tp->snd_nxt + len;
1516 * We know that the packet was lost, so back out the
1517 * sequence number advance, if any.
1519 * If the error is EPERM the packet got blocked by the
1520 * local firewall. Normally we should terminate the
1521 * connection but the blocking may have been spurious
1522 * due to a firewall reconfiguration cycle. So we treat
1523 * it like a packet loss and let the retransmit timer and
1524 * timeouts do their work over time.
1525 * XXX: It is a POLA question whether calling tcp_drop right
1526 * away would be the really correct behavior instead.
1528 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1529 !tcp_timer_active(tp, TT_PERSIST)) &&
1530 ((flags & TH_SYN) == 0) &&
1534 tp->sackhint.sack_bytes_rexmit -= len;
1535 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1536 ("sackhint bytes rtx >= 0"));
1540 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1543 tp->t_softerror = error;
1546 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1547 tp->snd_cwnd = tp->t_maxseg;
1551 * For some reason the interface we used initially
1552 * to send segments changed to another or lowered
1554 * If TSO was active we either got an interface
1555 * without TSO capabilits or TSO was turned off.
1556 * If we obtained mtu from ip_output() then update
1560 tp->t_flags &= ~TF_TSO;
1562 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1570 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1571 tp->t_softerror = error;
1579 TCPSTAT_INC(tcps_sndtotal);
1582 * Data sent (as far as we can tell).
1583 * If this advertises a larger window than any other segment,
1584 * then remember the size of the advertised window.
1585 * Any pending ACK has now been sent.
1587 if (recwin >= 0 && SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1588 tp->rcv_adv = tp->rcv_nxt + recwin;
1589 tp->last_ack_sent = tp->rcv_nxt;
1590 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1591 if (tcp_timer_active(tp, TT_DELACK))
1592 tcp_timer_activate(tp, TT_DELACK, 0);
1595 * This completely breaks TCP if newreno is turned on. What happens
1596 * is that if delayed-acks are turned on on the receiver, this code
1597 * on the transmitter effectively destroys the TCP window, forcing
1598 * it to four packets (1.5Kx4 = 6K window).
1600 if (sendalot && --maxburst)
1609 tcp_setpersist(struct tcpcb *tp)
1611 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1614 tp->t_flags &= ~TF_PREVVALID;
1615 if (tcp_timer_active(tp, TT_REXMT))
1616 panic("tcp_setpersist: retransmit pending");
1618 * Start/restart persistance timer.
1620 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1621 tcp_persmin, tcp_persmax);
1622 tcp_timer_activate(tp, TT_PERSIST, tt);
1623 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1628 * Insert TCP options according to the supplied parameters to the place
1629 * optp in a consistent way. Can handle unaligned destinations.
1631 * The order of the option processing is crucial for optimal packing and
1632 * alignment for the scarce option space.
1634 * The optimal order for a SYN/SYN-ACK segment is:
1635 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1636 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1638 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1639 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1640 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1641 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1642 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1645 tcp_addoptions(struct tcpopt *to, u_char *optp)
1647 u_int mask, optlen = 0;
1649 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1650 if ((to->to_flags & mask) != mask)
1652 if (optlen == TCP_MAXOLEN)
1654 switch (to->to_flags & mask) {
1656 while (optlen % 4) {
1657 optlen += TCPOLEN_NOP;
1658 *optp++ = TCPOPT_NOP;
1660 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1662 optlen += TCPOLEN_MAXSEG;
1663 *optp++ = TCPOPT_MAXSEG;
1664 *optp++ = TCPOLEN_MAXSEG;
1665 to->to_mss = htons(to->to_mss);
1666 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1667 optp += sizeof(to->to_mss);
1670 while (!optlen || optlen % 2 != 1) {
1671 optlen += TCPOLEN_NOP;
1672 *optp++ = TCPOPT_NOP;
1674 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1676 optlen += TCPOLEN_WINDOW;
1677 *optp++ = TCPOPT_WINDOW;
1678 *optp++ = TCPOLEN_WINDOW;
1679 *optp++ = to->to_wscale;
1682 while (optlen % 2) {
1683 optlen += TCPOLEN_NOP;
1684 *optp++ = TCPOPT_NOP;
1686 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1688 optlen += TCPOLEN_SACK_PERMITTED;
1689 *optp++ = TCPOPT_SACK_PERMITTED;
1690 *optp++ = TCPOLEN_SACK_PERMITTED;
1693 while (!optlen || optlen % 4 != 2) {
1694 optlen += TCPOLEN_NOP;
1695 *optp++ = TCPOPT_NOP;
1697 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1699 optlen += TCPOLEN_TIMESTAMP;
1700 *optp++ = TCPOPT_TIMESTAMP;
1701 *optp++ = TCPOLEN_TIMESTAMP;
1702 to->to_tsval = htonl(to->to_tsval);
1703 to->to_tsecr = htonl(to->to_tsecr);
1704 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1705 optp += sizeof(to->to_tsval);
1706 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1707 optp += sizeof(to->to_tsecr);
1709 #ifdef TCP_SIGNATURE
1712 int siglen = TCPOLEN_SIGNATURE - 2;
1714 while (!optlen || optlen % 4 != 2) {
1715 optlen += TCPOLEN_NOP;
1716 *optp++ = TCPOPT_NOP;
1718 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE)
1720 optlen += TCPOLEN_SIGNATURE;
1721 *optp++ = TCPOPT_SIGNATURE;
1722 *optp++ = TCPOLEN_SIGNATURE;
1723 to->to_signature = optp;
1732 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1735 while (!optlen || optlen % 4 != 2) {
1736 optlen += TCPOLEN_NOP;
1737 *optp++ = TCPOPT_NOP;
1739 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1741 optlen += TCPOLEN_SACKHDR;
1742 *optp++ = TCPOPT_SACK;
1743 sackblks = min(to->to_nsacks,
1744 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1745 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1746 while (sackblks--) {
1747 sack_seq = htonl(sack->start);
1748 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1749 optp += sizeof(sack_seq);
1750 sack_seq = htonl(sack->end);
1751 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1752 optp += sizeof(sack_seq);
1753 optlen += TCPOLEN_SACK;
1756 TCPSTAT_INC(tcps_sack_send_blocks);
1764 /* XXX is there any point to aligning this option? */
1765 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1766 if (TCP_MAXOLEN - optlen < total_len)
1768 *optp++ = TCPOPT_FAST_OPEN;
1769 *optp++ = total_len;
1770 if (to->to_tfo_len > 0) {
1771 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1772 optp += to->to_tfo_len;
1774 optlen += total_len;
1779 panic("%s: unknown TCP option type", __func__);
1784 /* Terminate and pad TCP options to a 4 byte boundary. */
1786 optlen += TCPOLEN_EOL;
1787 *optp++ = TCPOPT_EOL;
1790 * According to RFC 793 (STD0007):
1791 * "The content of the header beyond the End-of-Option option
1792 * must be header padding (i.e., zero)."
1793 * and later: "The padding is composed of zeros."
1795 while (optlen % 4) {
1796 optlen += TCPOLEN_PAD;
1797 *optp++ = TCPOPT_PAD;
1800 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));