2 * SPDX-License-Identifier: BSD-3-Clause
4 * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
5 * The Regents of the University of California. All rights reserved.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
10 * 1. Redistributions of source code must retain the above copyright
11 * notice, this list of conditions and the following disclaimer.
12 * 2. Redistributions in binary form must reproduce the above copyright
13 * notice, this list of conditions and the following disclaimer in the
14 * documentation and/or other materials provided with the distribution.
15 * 3. Neither the name of the University nor the names of its contributors
16 * may be used to endorse or promote products derived from this software
17 * without specific prior written permission.
19 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
20 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
21 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
22 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
23 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
24 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
25 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
26 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
27 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
28 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31 * @(#)tcp_output.c 8.4 (Berkeley) 5/24/95
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
38 #include "opt_inet6.h"
39 #include "opt_ipsec.h"
40 #include "opt_tcpdebug.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
44 #include <sys/domain.h>
46 #include <sys/hhook.h>
48 #include <sys/kernel.h>
51 #include <sys/mutex.h>
52 #include <sys/protosw.h>
54 #include <sys/socket.h>
55 #include <sys/socketvar.h>
56 #include <sys/sysctl.h>
59 #include <net/route.h>
62 #include <netinet/in.h>
63 #include <netinet/in_kdtrace.h>
64 #include <netinet/in_systm.h>
65 #include <netinet/ip.h>
66 #include <netinet/in_pcb.h>
67 #include <netinet/ip_var.h>
68 #include <netinet/ip_options.h>
70 #include <netinet6/in6_pcb.h>
71 #include <netinet/ip6.h>
72 #include <netinet6/ip6_var.h>
74 #include <netinet/tcp.h>
76 #include <netinet/tcp_fsm.h>
77 #include <netinet/tcp_log_buf.h>
78 #include <netinet/tcp_seq.h>
79 #include <netinet/tcp_timer.h>
80 #include <netinet/tcp_var.h>
81 #include <netinet/tcpip.h>
82 #include <netinet/cc/cc.h>
83 #include <netinet/tcp_fastopen.h>
85 #include <netinet/tcp_pcap.h>
88 #include <netinet/tcp_debug.h>
91 #include <netinet/tcp_offload.h>
94 #include <netipsec/ipsec_support.h>
96 #include <machine/in_cksum.h>
98 #include <security/mac/mac_framework.h>
100 VNET_DEFINE(int, path_mtu_discovery) = 1;
101 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
102 &VNET_NAME(path_mtu_discovery), 1,
103 "Enable Path MTU Discovery");
105 VNET_DEFINE(int, tcp_do_tso) = 1;
106 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
107 &VNET_NAME(tcp_do_tso), 0,
108 "Enable TCP Segmentation Offload");
110 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
111 #define V_tcp_sendspace VNET(tcp_sendspace)
112 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
113 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
115 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
117 &VNET_NAME(tcp_do_autosndbuf), 0,
118 "Enable automatic send buffer sizing");
120 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
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 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
127 &VNET_NAME(tcp_autosndbuf_max), 0,
128 "Max size of automatic send buffer");
130 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
131 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
132 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
133 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
134 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
137 * Make sure that either retransmit or persist timer is set for SYN, FIN and
140 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
141 KASSERT(((len) == 0 && ((th_flags) & \
142 (TH_SYN | TH_FIN | TH_RST)) != 0) || \
143 tcp_timer_active((tp), TT_REXMT) || \
144 tcp_timer_active((tp), TT_PERSIST), \
145 ("neither rexmt nor persist timer is set"))
147 static void inline cc_after_idle(struct tcpcb *tp);
151 * Wrapper for the TCP established output helper hook.
154 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
155 struct tcpopt *to, uint32_t len, int tso)
157 struct tcp_hhook_data hhook_data;
159 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
163 hhook_data.len = len;
164 hhook_data.tso = tso;
166 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
173 * CC wrapper hook functions
176 cc_after_idle(struct tcpcb *tp)
178 INP_WLOCK_ASSERT(tp->t_inpcb);
180 if (CC_ALGO(tp)->after_idle != NULL)
181 CC_ALGO(tp)->after_idle(tp->ccv);
185 * Tcp output routine: figure out what should be sent and send it.
188 tcp_output(struct tcpcb *tp)
190 struct socket *so = tp->t_inpcb->inp_socket;
192 uint32_t recwin, sendwin;
193 int off, flags, error = 0; /* Keep compiler happy */
194 u_int if_hw_tsomaxsegcount = 0;
195 u_int if_hw_tsomaxsegsize;
197 struct ip *ip = NULL;
199 struct ipovly *ipov = NULL;
202 u_char opt[TCP_MAXOLEN];
203 unsigned ipoptlen, optlen, hdrlen;
204 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
205 unsigned ipsec_optlen = 0;
207 int idle, sendalot, curticks;
208 int sack_rxmit, sack_bytes_rxmt;
212 unsigned int wanted_cookie = 0;
213 unsigned int dont_sendalot = 0;
215 int maxburst = TCP_MAXBURST;
218 struct ip6_hdr *ip6 = NULL;
221 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
224 INP_WLOCK_ASSERT(tp->t_inpcb);
227 if (tp->t_flags & TF_TOE)
228 return (tcp_offload_output(tp));
232 * For TFO connections in SYN_SENT or SYN_RECEIVED,
233 * only allow the initial SYN or SYN|ACK and those sent
234 * by the retransmit timer.
236 if (IS_FASTOPEN(tp->t_flags) &&
237 ((tp->t_state == TCPS_SYN_SENT) ||
238 (tp->t_state == TCPS_SYN_RECEIVED)) &&
239 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
240 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
244 * Determine length of data that should be transmitted,
245 * and flags that will be used.
246 * If there is some data or critical controls (SYN, RST)
247 * to send, then transmit; otherwise, investigate further.
249 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
250 if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
252 tp->t_flags &= ~TF_LASTIDLE;
254 if (tp->t_flags & TF_MORETOCOME) {
255 tp->t_flags |= TF_LASTIDLE;
261 * If we've recently taken a timeout, snd_max will be greater than
262 * snd_nxt. There may be SACK information that allows us to avoid
263 * resending already delivered data. Adjust snd_nxt accordingly.
265 if ((tp->t_flags & TF_SACK_PERMIT) &&
266 SEQ_LT(tp->snd_nxt, tp->snd_max))
271 off = tp->snd_nxt - tp->snd_una;
272 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
274 flags = tcp_outflags[tp->t_state];
276 * Send any SACK-generated retransmissions. If we're explicitly trying
277 * to send out new data (when sendalot is 1), bypass this function.
278 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
279 * we're replacing a (future) new transmission with a retransmission
280 * now, and we previously incremented snd_cwnd in tcp_input().
283 * Still in sack recovery , reset rxmit flag to zero.
289 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
290 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
294 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
295 /* Do not retransmit SACK segments beyond snd_recover */
296 if (SEQ_GT(p->end, tp->snd_recover)) {
298 * (At least) part of sack hole extends beyond
299 * snd_recover. Check to see if we can rexmit data
302 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
304 * Can't rexmit any more data for this hole.
305 * That data will be rexmitted in the next
306 * sack recovery episode, when snd_recover
307 * moves past p->rxmit.
310 goto after_sack_rexmit;
312 /* Can rexmit part of the current hole */
313 len = ((int32_t)ulmin(cwin,
314 tp->snd_recover - p->rxmit));
316 len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
317 off = p->rxmit - tp->snd_una;
318 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
323 TCPSTAT_INC(tcps_sack_rexmits);
324 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
325 min(len, tp->t_maxseg));
330 * Get standard flags, and add SYN or FIN if requested by 'hidden'
333 if (tp->t_flags & TF_NEEDFIN)
335 if (tp->t_flags & TF_NEEDSYN)
338 SOCKBUF_LOCK(&so->so_snd);
340 * If in persist timeout with window of 0, send 1 byte.
341 * Otherwise, if window is small but nonzero
342 * and timer expired, we will send what we can
343 * and go to transmit state.
345 if (tp->t_flags & TF_FORCEDATA) {
348 * If we still have some data to send, then
349 * clear the FIN bit. Usually this would
350 * happen below when it realizes that we
351 * aren't sending all the data. However,
352 * if we have exactly 1 byte of unsent data,
353 * then it won't clear the FIN bit below,
354 * and if we are in persist state, we wind
355 * up sending the packet without recording
356 * that we sent the FIN bit.
358 * We can't just blindly clear the FIN bit,
359 * because if we don't have any more data
360 * to send then the probe will be the FIN
363 if (off < sbused(&so->so_snd))
367 tcp_timer_activate(tp, TT_PERSIST, 0);
373 * If snd_nxt == snd_max and we have transmitted a FIN, the
374 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
375 * a negative length. This can also occur when TCP opens up
376 * its congestion window while receiving additional duplicate
377 * acks after fast-retransmit because TCP will reset snd_nxt
378 * to snd_max after the fast-retransmit.
380 * In the normal retransmit-FIN-only case, however, snd_nxt will
381 * be set to snd_una, the offset will be 0, and the length may
384 * If sack_rxmit is true we are retransmitting from the scoreboard
385 * in which case len is already set.
387 if (sack_rxmit == 0) {
388 if (sack_bytes_rxmt == 0)
389 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
395 * We are inside of a SACK recovery episode and are
396 * sending new data, having retransmitted all the
397 * data possible in the scoreboard.
399 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
402 * Don't remove this (len > 0) check !
403 * We explicitly check for len > 0 here (although it
404 * isn't really necessary), to work around a gcc
405 * optimization issue - to force gcc to compute
406 * len above. Without this check, the computation
407 * of len is bungled by the optimizer.
410 cwin = tp->snd_cwnd -
411 (tp->snd_nxt - tp->sack_newdata) -
415 len = imin(len, cwin);
421 * Lop off SYN bit if it has already been sent. However, if this
422 * is SYN-SENT state and if segment contains data and if we don't
423 * know that foreign host supports TAO, suppress sending segment.
425 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
426 if (tp->t_state != TCPS_SYN_RECEIVED)
429 * When sending additional segments following a TFO SYN|ACK,
430 * do not include the SYN bit.
432 if (IS_FASTOPEN(tp->t_flags) &&
433 (tp->t_state == TCPS_SYN_RECEIVED))
439 * Be careful not to send data and/or FIN on SYN segments.
440 * This measure is needed to prevent interoperability problems
441 * with not fully conformant TCP implementations.
443 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
449 * On TFO sockets, ensure no data is sent in the following cases:
451 * - When retransmitting SYN|ACK on a passively-created socket
453 * - When retransmitting SYN on an actively created socket
455 * - When sending a zero-length cookie (cookie request) on an
456 * actively created socket
458 * - When the socket is in the CLOSED state (RST is being sent)
460 if (IS_FASTOPEN(tp->t_flags) &&
461 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
462 ((tp->t_state == TCPS_SYN_SENT) &&
463 (tp->t_tfo_client_cookie_len == 0)) ||
468 * If FIN has been sent but not acked,
469 * but we haven't been called to retransmit,
470 * len will be < 0. Otherwise, window shrank
471 * after we sent into it. If window shrank to 0,
472 * cancel pending retransmit, pull snd_nxt back
473 * to (closed) window, and set the persist timer
474 * if it isn't already going. If the window didn't
475 * close completely, just wait for an ACK.
477 * We also do a general check here to ensure that
478 * we will set the persist timer when we have data
479 * to send, but a 0-byte window. This makes sure
480 * the persist timer is set even if the packet
481 * hits one of the "goto send" lines below.
484 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
485 (off < (int) sbavail(&so->so_snd))) {
486 tcp_timer_activate(tp, TT_REXMT, 0);
488 tp->snd_nxt = tp->snd_una;
489 if (!tcp_timer_active(tp, TT_PERSIST))
494 /* len will be >= 0 after this point. */
495 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
497 tcp_sndbuf_autoscale(tp, so, sendwin);
500 * Decide if we can use TCP Segmentation Offloading (if supported by
503 * TSO may only be used if we are in a pure bulk sending state. The
504 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
505 * IP options prevent using TSO. With TSO the TCP header is the same
506 * (except for the sequence number) for all generated packets. This
507 * makes it impossible to transmit any options which vary per generated
510 * IPv4 handling has a clear separation of ip options and ip header
511 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
512 * the right thing below to provide length of just ip options and thus
513 * checking for ipoptlen is enough to decide if ip options are present.
515 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
517 * Pre-calculate here as we save another lookup into the darknesses
518 * of IPsec that way and can actually decide if TSO is ok.
521 if (isipv6 && IPSEC_ENABLED(ipv6))
522 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
528 if (IPSEC_ENABLED(ipv4))
529 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
534 ipoptlen = ip6_optlen(tp->t_inpcb);
537 if (tp->t_inpcb->inp_options)
538 ipoptlen = tp->t_inpcb->inp_options->m_len -
539 offsetof(struct ipoption, ipopt_list);
542 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
543 ipoptlen += ipsec_optlen;
546 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
547 ((tp->t_flags & TF_SIGNATURE) == 0) &&
548 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
549 ipoptlen == 0 && !(flags & TH_SYN))
553 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
556 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
557 sbused(&so->so_snd)))
561 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
562 (long)TCP_MAXWIN << tp->rcv_scale);
565 * Sender silly window avoidance. We transmit under the following
566 * conditions when len is non-zero:
568 * - We have a full segment (or more with TSO)
569 * - This is the last buffer in a write()/send() and we are
570 * either idle or running NODELAY
571 * - we've timed out (e.g. persist timer)
572 * - we have more then 1/2 the maximum send window's worth of
573 * data (receiver may be limited the window size)
574 * - we need to retransmit
577 if (len >= tp->t_maxseg)
580 * NOTE! on localhost connections an 'ack' from the remote
581 * end may occur synchronously with the output and cause
582 * us to flush a buffer queued with moretocome. XXX
584 * note: the len + off check is almost certainly unnecessary.
586 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
587 (idle || (tp->t_flags & TF_NODELAY)) &&
588 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
589 (tp->t_flags & TF_NOPUSH) == 0) {
592 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
594 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
596 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
603 * Sending of standalone window updates.
605 * Window updates are important when we close our window due to a
606 * full socket buffer and are opening it again after the application
607 * reads data from it. Once the window has opened again and the
608 * remote end starts to send again the ACK clock takes over and
609 * provides the most current window information.
611 * We must avoid the silly window syndrome whereas every read
612 * from the receive buffer, no matter how small, causes a window
613 * update to be sent. We also should avoid sending a flurry of
614 * window updates when the socket buffer had queued a lot of data
615 * and the application is doing small reads.
617 * Prevent a flurry of pointless window updates by only sending
618 * an update when we can increase the advertized window by more
619 * than 1/4th of the socket buffer capacity. When the buffer is
620 * getting full or is very small be more aggressive and send an
621 * update whenever we can increase by two mss sized segments.
622 * In all other situations the ACK's to new incoming data will
623 * carry further window increases.
625 * Don't send an independent window update if a delayed
626 * ACK is pending (it will get piggy-backed on it) or the
627 * remote side already has done a half-close and won't send
628 * more data. Skip this if the connection is in T/TCP
631 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
632 !(tp->t_flags & TF_DELACK) &&
633 !TCPS_HAVERCVDFIN(tp->t_state)) {
635 * "adv" is the amount we could increase the window,
636 * taking into account that we are limited by
637 * TCP_MAXWIN << tp->rcv_scale.
643 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
644 oldwin = (tp->rcv_adv - tp->rcv_nxt);
650 * If the new window size ends up being the same as or less
651 * than the old size when it is scaled, then don't force
654 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
657 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
658 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
659 recwin <= (so->so_rcv.sb_hiwat / 8) ||
660 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
661 adv >= TCP_MAXWIN << tp->rcv_scale))
663 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
669 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
670 * is also a catch-all for the retransmit timer timeout case.
672 if (tp->t_flags & TF_ACKNOW)
674 if ((flags & TH_RST) ||
675 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
677 if (SEQ_GT(tp->snd_up, tp->snd_una))
680 * If our state indicates that FIN should be sent
681 * and we have not yet done so, then we need to send.
683 if (flags & TH_FIN &&
684 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
687 * In SACK, it is possible for tcp_output to fail to send a segment
688 * after the retransmission timer has been turned off. Make sure
689 * that the retransmission timer is set.
691 if ((tp->t_flags & TF_SACK_PERMIT) &&
692 SEQ_GT(tp->snd_max, tp->snd_una) &&
693 !tcp_timer_active(tp, TT_REXMT) &&
694 !tcp_timer_active(tp, TT_PERSIST)) {
695 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
699 * TCP window updates are not reliable, rather a polling protocol
700 * using ``persist'' packets is used to insure receipt of window
701 * updates. The three ``states'' for the output side are:
702 * idle not doing retransmits or persists
703 * persisting to move a small or zero window
704 * (re)transmitting and thereby not persisting
706 * tcp_timer_active(tp, TT_PERSIST)
707 * is true when we are in persist state.
708 * (tp->t_flags & TF_FORCEDATA)
709 * is set when we are called to send a persist packet.
710 * tcp_timer_active(tp, TT_REXMT)
711 * is set when we are retransmitting
712 * The output side is idle when both timers are zero.
714 * If send window is too small, there is data to transmit, and no
715 * retransmit or persist is pending, then go to persist state.
716 * If nothing happens soon, send when timer expires:
717 * if window is nonzero, transmit what we can,
718 * otherwise force out a byte.
720 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
721 !tcp_timer_active(tp, TT_PERSIST)) {
727 * No reason to send a segment, just return.
730 SOCKBUF_UNLOCK(&so->so_snd);
734 SOCKBUF_LOCK_ASSERT(&so->so_snd);
736 if (len >= tp->t_maxseg)
737 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
739 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
742 * Before ESTABLISHED, force sending of initial options
743 * unless TCP set not to do any options.
744 * NOTE: we assume that the IP/TCP header plus TCP options
745 * always fit in a single mbuf, leaving room for a maximum
747 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
752 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
755 hdrlen = sizeof (struct tcpiphdr);
758 * Compute options for segment.
759 * We only have to care about SYN and established connection
760 * segments. Options for SYN-ACK segments are handled in TCP
764 if ((tp->t_flags & TF_NOOPT) == 0) {
765 /* Maximum segment size. */
766 if (flags & TH_SYN) {
767 tp->snd_nxt = tp->iss;
768 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
769 to.to_flags |= TOF_MSS;
772 * On SYN or SYN|ACK transmits on TFO connections,
773 * only include the TFO option if it is not a
774 * retransmit, as the presence of the TFO option may
775 * have caused the original SYN or SYN|ACK to have
776 * been dropped by a middlebox.
778 if (IS_FASTOPEN(tp->t_flags) &&
779 (tp->t_rxtshift == 0)) {
780 if (tp->t_state == TCPS_SYN_RECEIVED) {
781 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
783 (u_int8_t *)&tp->t_tfo_cookie.server;
784 to.to_flags |= TOF_FASTOPEN;
786 } else if (tp->t_state == TCPS_SYN_SENT) {
788 tp->t_tfo_client_cookie_len;
790 tp->t_tfo_cookie.client;
791 to.to_flags |= TOF_FASTOPEN;
794 * If we wind up having more data to
795 * send with the SYN than can fit in
796 * one segment, don't send any more
797 * until the SYN|ACK comes back from
804 /* Window scaling. */
805 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
806 to.to_wscale = tp->request_r_scale;
807 to.to_flags |= TOF_SCALE;
810 if ((tp->t_flags & TF_RCVD_TSTMP) ||
811 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
812 curticks = tcp_ts_getticks();
813 to.to_tsval = curticks + tp->ts_offset;
814 to.to_tsecr = tp->ts_recent;
815 to.to_flags |= TOF_TS;
816 if (tp->t_rxtshift == 1)
817 tp->t_badrxtwin = curticks;
820 /* Set receive buffer autosizing timestamp. */
821 if (tp->rfbuf_ts == 0 &&
822 (so->so_rcv.sb_flags & SB_AUTOSIZE))
823 tp->rfbuf_ts = tcp_ts_getticks();
825 /* Selective ACK's. */
826 if (tp->t_flags & TF_SACK_PERMIT) {
828 to.to_flags |= TOF_SACKPERM;
829 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
830 (tp->t_flags & TF_SACK_PERMIT) &&
831 tp->rcv_numsacks > 0) {
832 to.to_flags |= TOF_SACK;
833 to.to_nsacks = tp->rcv_numsacks;
834 to.to_sacks = (u_char *)tp->sackblks;
837 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
838 /* TCP-MD5 (RFC2385). */
840 * Check that TCP_MD5SIG is enabled in tcpcb to
841 * account the size needed to set this TCP option.
843 if (tp->t_flags & TF_SIGNATURE)
844 to.to_flags |= TOF_SIGNATURE;
845 #endif /* TCP_SIGNATURE */
847 /* Processing the options. */
848 hdrlen += optlen = tcp_addoptions(&to, opt);
850 * If we wanted a TFO option to be added, but it was unable
851 * to fit, ensure no data is sent.
853 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
854 !(to.to_flags & TOF_FASTOPEN))
859 * Adjust data length if insertion of options will
860 * bump the packet length beyond the t_maxseg length.
861 * Clear the FIN bit because we cut off the tail of
864 if (len + optlen + ipoptlen > tp->t_maxseg) {
872 /* extract TSO information */
873 if_hw_tsomax = tp->t_tsomax;
874 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
875 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
878 * Limit a TSO burst to prevent it from
879 * overflowing or exceeding the maximum length
880 * allowed by the network interface:
882 KASSERT(ipoptlen == 0,
883 ("%s: TSO can't do IP options", __func__));
886 * Check if we should limit by maximum payload
889 if (if_hw_tsomax != 0) {
890 /* compute maximum TSO length */
891 max_len = (if_hw_tsomax - hdrlen -
895 } else if (len > max_len) {
902 * Prevent the last segment from being
903 * fractional unless the send sockbuf can be
906 max_len = (tp->t_maxseg - optlen);
907 if (((uint32_t)off + (uint32_t)len) <
908 sbavail(&so->so_snd)) {
909 moff = len % max_len;
917 * In case there are too many small fragments
920 if (len <= max_len) {
927 * Send the FIN in a separate segment
928 * after the bulk sending is done.
929 * We don't trust the TSO implementations
930 * to clear the FIN flag on all but the
933 if (tp->t_flags & TF_NEEDFIN)
936 len = tp->t_maxseg - optlen - ipoptlen;
944 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
945 ("%s: len > IP_MAXPACKET", __func__));
947 /*#ifdef DIAGNOSTIC*/
949 if (max_linkhdr + hdrlen > MCLBYTES)
951 if (max_linkhdr + hdrlen > MHLEN)
953 panic("tcphdr too big");
957 * This KASSERT is here to catch edge cases at a well defined place.
958 * Before, those had triggered (random) panic conditions further down.
960 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
963 * Grab a header mbuf, attaching a copy of data to
964 * be transmitted, and initialize the header from
965 * the template for sends on this connection.
972 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
973 TCPSTAT_INC(tcps_sndprobe);
974 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
975 tp->t_sndrexmitpack++;
976 TCPSTAT_INC(tcps_sndrexmitpack);
977 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
979 TCPSTAT_INC(tcps_sndpack);
980 TCPSTAT_ADD(tcps_sndbyte, len);
983 if (MHLEN < hdrlen + max_linkhdr)
984 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
987 m = m_gethdr(M_NOWAIT, MT_DATA);
990 SOCKBUF_UNLOCK(&so->so_snd);
996 m->m_data += max_linkhdr;
1000 * Start the m_copy functions from the closest mbuf
1001 * to the offset in the socket buffer chain.
1003 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1004 if (len <= MHLEN - hdrlen - max_linkhdr) {
1005 m_copydata(mb, moff, len,
1006 mtod(m, caddr_t) + hdrlen);
1007 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1008 sbsndptr_adv(&so->so_snd, mb, len);
1011 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1015 m->m_next = tcp_m_copym(mb, moff,
1016 &len, if_hw_tsomaxsegcount,
1017 if_hw_tsomaxsegsize, msb);
1018 if (len <= (tp->t_maxseg - optlen)) {
1020 * Must have ran out of mbufs for the copy
1021 * shorten it to no longer need tso. Lets
1022 * not put on sendalot since we are low on
1027 if (m->m_next == NULL) {
1028 SOCKBUF_UNLOCK(&so->so_snd);
1037 * If we're sending everything we've got, set PUSH.
1038 * (This will keep happy those implementations which only
1039 * give data to the user when a buffer fills or
1042 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1045 SOCKBUF_UNLOCK(&so->so_snd);
1047 SOCKBUF_UNLOCK(&so->so_snd);
1048 if (tp->t_flags & TF_ACKNOW)
1049 TCPSTAT_INC(tcps_sndacks);
1050 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1051 TCPSTAT_INC(tcps_sndctrl);
1052 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1053 TCPSTAT_INC(tcps_sndurg);
1055 TCPSTAT_INC(tcps_sndwinup);
1057 m = m_gethdr(M_NOWAIT, MT_DATA);
1064 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1069 m->m_data += max_linkhdr;
1072 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1073 m->m_pkthdr.rcvif = (struct ifnet *)0;
1075 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1079 ip6 = mtod(m, struct ip6_hdr *);
1080 th = (struct tcphdr *)(ip6 + 1);
1081 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1085 ip = mtod(m, struct ip *);
1087 ipov = (struct ipovly *)ip;
1089 th = (struct tcphdr *)(ip + 1);
1090 tcpip_fillheaders(tp->t_inpcb, ip, th);
1094 * Fill in fields, remembering maximum advertised
1095 * window for use in delaying messages about window sizes.
1096 * If resending a FIN, be sure not to use a new sequence number.
1098 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1099 tp->snd_nxt == tp->snd_max)
1102 * If we are starting a connection, send ECN setup
1103 * SYN packet. If we are on a retransmit, we may
1104 * resend those bits a number of times as per
1107 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1108 if (tp->t_rxtshift >= 1) {
1109 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1110 flags |= TH_ECE|TH_CWR;
1112 flags |= TH_ECE|TH_CWR;
1115 if (tp->t_state == TCPS_ESTABLISHED &&
1116 (tp->t_flags & TF_ECN_PERMIT)) {
1118 * If the peer has ECN, mark data packets with
1119 * ECN capable transmission (ECT).
1120 * Ignore pure ack packets, retransmissions and window probes.
1122 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1123 !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1126 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1129 ip->ip_tos |= IPTOS_ECN_ECT0;
1130 TCPSTAT_INC(tcps_ecn_ect0);
1134 * Reply with proper ECN notifications.
1136 if (tp->t_flags & TF_ECN_SND_CWR) {
1138 tp->t_flags &= ~TF_ECN_SND_CWR;
1140 if (tp->t_flags & TF_ECN_SND_ECE)
1145 * If we are doing retransmissions, then snd_nxt will
1146 * not reflect the first unsent octet. For ACK only
1147 * packets, we do not want the sequence number of the
1148 * retransmitted packet, we want the sequence number
1149 * of the next unsent octet. So, if there is no data
1150 * (and no SYN or FIN), use snd_max instead of snd_nxt
1151 * when filling in ti_seq. But if we are in persist
1152 * state, snd_max might reflect one byte beyond the
1153 * right edge of the window, so use snd_nxt in that
1154 * case, since we know we aren't doing a retransmission.
1155 * (retransmit and persist are mutually exclusive...)
1157 if (sack_rxmit == 0) {
1158 if (len || (flags & (TH_SYN|TH_FIN)) ||
1159 tcp_timer_active(tp, TT_PERSIST))
1160 th->th_seq = htonl(tp->snd_nxt);
1162 th->th_seq = htonl(tp->snd_max);
1164 th->th_seq = htonl(p->rxmit);
1166 tp->sackhint.sack_bytes_rexmit += len;
1168 th->th_ack = htonl(tp->rcv_nxt);
1170 bcopy(opt, th + 1, optlen);
1171 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1173 th->th_flags = flags;
1175 * Calculate receive window. Don't shrink window,
1176 * but avoid silly window syndrome.
1177 * If a RST segment is sent, advertise a window of zero.
1179 if (flags & TH_RST) {
1182 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1183 recwin < tp->t_maxseg)
1185 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1186 recwin < (tp->rcv_adv - tp->rcv_nxt))
1187 recwin = (tp->rcv_adv - tp->rcv_nxt);
1190 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1191 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1192 * case is handled in syncache.
1195 th->th_win = htons((u_short)
1196 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1198 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1201 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1202 * a 0 window. This may cause the remote transmitter to stall. This
1203 * flag tells soreceive() to disable delayed acknowledgements when
1204 * draining the buffer. This can occur if the receiver is attempting
1205 * to read more data than can be buffered prior to transmitting on
1208 if (th->th_win == 0) {
1210 tp->t_flags |= TF_RXWIN0SENT;
1212 tp->t_flags &= ~TF_RXWIN0SENT;
1213 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1214 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1215 th->th_flags |= TH_URG;
1218 * If no urgent pointer to send, then we pull
1219 * the urgent pointer to the left edge of the send window
1220 * so that it doesn't drift into the send window on sequence
1221 * number wraparound.
1223 tp->snd_up = tp->snd_una; /* drag it along */
1226 * Put TCP length in extended header, and then
1227 * checksum extended header and data.
1229 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1230 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1232 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1233 if (to.to_flags & TOF_SIGNATURE) {
1235 * Calculate MD5 signature and put it into the place
1236 * determined before.
1237 * NOTE: since TCP options buffer doesn't point into
1238 * mbuf's data, calculate offset and use it.
1240 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1241 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1243 * Do not send segment if the calculation of MD5
1244 * digest has failed.
1254 * There is no need to fill in ip6_plen right now.
1255 * It will be filled later by ip6_output.
1257 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1258 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1259 optlen + len, IPPROTO_TCP, 0);
1262 #if defined(INET6) && defined(INET)
1267 m->m_pkthdr.csum_flags = CSUM_TCP;
1268 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1269 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1271 /* IP version must be set here for ipv4/ipv6 checking later */
1272 KASSERT(ip->ip_v == IPVERSION,
1273 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1278 * Enable TSO and specify the size of the segments.
1279 * The TCP pseudo header checksum is always provided.
1282 KASSERT(len > tp->t_maxseg - optlen,
1283 ("%s: len <= tso_segsz", __func__));
1284 m->m_pkthdr.csum_flags |= CSUM_TSO;
1285 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1288 KASSERT(len + hdrlen == m_length(m, NULL),
1289 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1290 __func__, len, hdrlen, m_length(m, NULL)));
1293 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1294 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1301 if (so->so_options & SO_DEBUG) {
1307 save = ipov->ih_len;
1308 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1310 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1314 ipov->ih_len = save;
1316 #endif /* TCPDEBUG */
1317 TCP_PROBE3(debug__output, tp, th, m);
1319 /* We're getting ready to send; log now. */
1320 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1324 * Fill in IP length and desired time to live and
1325 * send to IP level. There should be a better way
1326 * to handle ttl and tos; we could keep them in
1327 * the template, but need a way to checksum without them.
1330 * m->m_pkthdr.len should have been set before checksum calculation,
1331 * because in6_cksum() need it.
1336 * we separately set hoplimit for every segment, since the
1337 * user might want to change the value via setsockopt.
1338 * Also, desired default hop limit might be changed via
1339 * Neighbor Discovery.
1341 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1344 * Set the packet size here for the benefit of DTrace probes.
1345 * ip6_output() will set it properly; it's supposed to include
1346 * the option header lengths as well.
1348 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1350 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1351 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1353 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1355 if (tp->t_state == TCPS_SYN_SENT)
1356 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1358 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1361 /* Save packet, if requested. */
1362 tcp_pcap_add(th, m, &(tp->t_outpkts));
1365 /* TODO: IPv6 IP6TOS_ECT bit on */
1366 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1367 &tp->t_inpcb->inp_route6,
1368 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1369 NULL, NULL, tp->t_inpcb);
1371 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1372 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1375 #if defined(INET) && defined(INET6)
1380 ip->ip_len = htons(m->m_pkthdr.len);
1382 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1383 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1386 * If we do path MTU discovery, then we set DF on every packet.
1387 * This might not be the best thing to do according to RFC3390
1388 * Section 2. However the tcp hostcache migitates the problem
1389 * so it affects only the first tcp connection with a host.
1391 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1393 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1394 ip->ip_off |= htons(IP_DF);
1395 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1397 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1400 if (tp->t_state == TCPS_SYN_SENT)
1401 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1403 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1406 /* Save packet, if requested. */
1407 tcp_pcap_add(th, m, &(tp->t_outpkts));
1410 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1411 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1414 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1415 mtu = tp->t_inpcb->inp_route.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 && sbavail(&so->so_snd) &&
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 + xlen;
1514 /* Record the error. */
1515 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1516 error, 0, NULL, false);
1519 * We know that the packet was lost, so back out the
1520 * sequence number advance, if any.
1522 * If the error is EPERM the packet got blocked by the
1523 * local firewall. Normally we should terminate the
1524 * connection but the blocking may have been spurious
1525 * due to a firewall reconfiguration cycle. So we treat
1526 * it like a packet loss and let the retransmit timer and
1527 * timeouts do their work over time.
1528 * XXX: It is a POLA question whether calling tcp_drop right
1529 * away would be the really correct behavior instead.
1531 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1532 !tcp_timer_active(tp, TT_PERSIST)) &&
1533 ((flags & TH_SYN) == 0) &&
1537 tp->sackhint.sack_bytes_rexmit -= len;
1538 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1539 ("sackhint bytes rtx >= 0"));
1543 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1547 tp->t_softerror = error;
1550 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1551 tp->snd_cwnd = tp->t_maxseg;
1555 * For some reason the interface we used initially
1556 * to send segments changed to another or lowered
1558 * If TSO was active we either got an interface
1559 * without TSO capabilits or TSO was turned off.
1560 * If we obtained mtu from ip_output() then update
1564 tp->t_flags &= ~TF_TSO;
1566 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1574 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1575 tp->t_softerror = error;
1583 TCPSTAT_INC(tcps_sndtotal);
1586 * Data sent (as far as we can tell).
1587 * If this advertises a larger window than any other segment,
1588 * then remember the size of the advertised window.
1589 * Any pending ACK has now been sent.
1591 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1592 tp->rcv_adv = tp->rcv_nxt + recwin;
1593 tp->last_ack_sent = tp->rcv_nxt;
1594 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1595 if (tcp_timer_active(tp, TT_DELACK))
1596 tcp_timer_activate(tp, TT_DELACK, 0);
1599 * This completely breaks TCP if newreno is turned on. What happens
1600 * is that if delayed-acks are turned on on the receiver, this code
1601 * on the transmitter effectively destroys the TCP window, forcing
1602 * it to four packets (1.5Kx4 = 6K window).
1604 if (sendalot && --maxburst)
1613 tcp_setpersist(struct tcpcb *tp)
1615 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1618 tp->t_flags &= ~TF_PREVVALID;
1619 if (tcp_timer_active(tp, TT_REXMT))
1620 panic("tcp_setpersist: retransmit pending");
1622 * Start/restart persistence timer.
1624 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1625 tcp_persmin, tcp_persmax);
1626 tcp_timer_activate(tp, TT_PERSIST, tt);
1627 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1632 * Insert TCP options according to the supplied parameters to the place
1633 * optp in a consistent way. Can handle unaligned destinations.
1635 * The order of the option processing is crucial for optimal packing and
1636 * alignment for the scarce option space.
1638 * The optimal order for a SYN/SYN-ACK segment is:
1639 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1640 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1642 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1643 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1644 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1645 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1646 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1649 tcp_addoptions(struct tcpopt *to, u_char *optp)
1651 u_int32_t mask, optlen = 0;
1653 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1654 if ((to->to_flags & mask) != mask)
1656 if (optlen == TCP_MAXOLEN)
1658 switch (to->to_flags & mask) {
1660 while (optlen % 4) {
1661 optlen += TCPOLEN_NOP;
1662 *optp++ = TCPOPT_NOP;
1664 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1666 optlen += TCPOLEN_MAXSEG;
1667 *optp++ = TCPOPT_MAXSEG;
1668 *optp++ = TCPOLEN_MAXSEG;
1669 to->to_mss = htons(to->to_mss);
1670 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1671 optp += sizeof(to->to_mss);
1674 while (!optlen || optlen % 2 != 1) {
1675 optlen += TCPOLEN_NOP;
1676 *optp++ = TCPOPT_NOP;
1678 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1680 optlen += TCPOLEN_WINDOW;
1681 *optp++ = TCPOPT_WINDOW;
1682 *optp++ = TCPOLEN_WINDOW;
1683 *optp++ = to->to_wscale;
1686 while (optlen % 2) {
1687 optlen += TCPOLEN_NOP;
1688 *optp++ = TCPOPT_NOP;
1690 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1692 optlen += TCPOLEN_SACK_PERMITTED;
1693 *optp++ = TCPOPT_SACK_PERMITTED;
1694 *optp++ = TCPOLEN_SACK_PERMITTED;
1697 while (!optlen || optlen % 4 != 2) {
1698 optlen += TCPOLEN_NOP;
1699 *optp++ = TCPOPT_NOP;
1701 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1703 optlen += TCPOLEN_TIMESTAMP;
1704 *optp++ = TCPOPT_TIMESTAMP;
1705 *optp++ = TCPOLEN_TIMESTAMP;
1706 to->to_tsval = htonl(to->to_tsval);
1707 to->to_tsecr = htonl(to->to_tsecr);
1708 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1709 optp += sizeof(to->to_tsval);
1710 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1711 optp += sizeof(to->to_tsecr);
1715 int siglen = TCPOLEN_SIGNATURE - 2;
1717 while (!optlen || optlen % 4 != 2) {
1718 optlen += TCPOLEN_NOP;
1719 *optp++ = TCPOPT_NOP;
1721 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1722 to->to_flags &= ~TOF_SIGNATURE;
1725 optlen += TCPOLEN_SIGNATURE;
1726 *optp++ = TCPOPT_SIGNATURE;
1727 *optp++ = TCPOLEN_SIGNATURE;
1728 to->to_signature = optp;
1736 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1739 while (!optlen || optlen % 4 != 2) {
1740 optlen += TCPOLEN_NOP;
1741 *optp++ = TCPOPT_NOP;
1743 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1745 optlen += TCPOLEN_SACKHDR;
1746 *optp++ = TCPOPT_SACK;
1747 sackblks = min(to->to_nsacks,
1748 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1749 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1750 while (sackblks--) {
1751 sack_seq = htonl(sack->start);
1752 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1753 optp += sizeof(sack_seq);
1754 sack_seq = htonl(sack->end);
1755 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1756 optp += sizeof(sack_seq);
1757 optlen += TCPOLEN_SACK;
1760 TCPSTAT_INC(tcps_sack_send_blocks);
1767 /* XXX is there any point to aligning this option? */
1768 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1769 if (TCP_MAXOLEN - optlen < total_len) {
1770 to->to_flags &= ~TOF_FASTOPEN;
1773 *optp++ = TCPOPT_FAST_OPEN;
1774 *optp++ = total_len;
1775 if (to->to_tfo_len > 0) {
1776 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1777 optp += to->to_tfo_len;
1779 optlen += total_len;
1783 panic("%s: unknown TCP option type", __func__);
1788 /* Terminate and pad TCP options to a 4 byte boundary. */
1790 optlen += TCPOLEN_EOL;
1791 *optp++ = TCPOPT_EOL;
1794 * According to RFC 793 (STD0007):
1795 * "The content of the header beyond the End-of-Option option
1796 * must be header padding (i.e., zero)."
1797 * and later: "The padding is composed of zeros."
1799 while (optlen % 4) {
1800 optlen += TCPOLEN_PAD;
1801 *optp++ = TCPOPT_PAD;
1804 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1809 * This is a copy of m_copym(), taking the TSO segment size/limit
1810 * constraints into account, and advancing the sndptr as it goes.
1813 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1814 int32_t seglimit, int32_t segsize, struct sockbuf *sb)
1816 struct mbuf *n, **np;
1819 int32_t len = *plen;
1823 uint32_t mlen, frags;
1827 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1828 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1829 if (off == 0 && m->m_flags & M_PKTHDR)
1834 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1838 if ((sb) && (m == sb->sb_sndptr)) {
1839 sb->sb_sndptroff += m->m_len;
1840 sb->sb_sndptr = m->m_next;
1849 KASSERT(len == M_COPYALL,
1850 ("tcp_m_copym, length > size of mbuf chain"));
1852 if (pkthdrlen != NULL)
1853 *pkthdrlen = len_cp;
1856 mlen = min(len, m->m_len - off);
1859 * For M_NOMAP mbufs, add 3 segments
1860 * + 1 in case we are crossing page boundaries
1861 * + 2 in case the TLS hdr/trailer are used
1862 * It is cheaper to just add the segments
1863 * than it is to take the cache miss to look
1864 * at the mbuf ext_pgs state in detail.
1866 if (m->m_flags & M_NOMAP) {
1867 fragsize = min(segsize, PAGE_SIZE);
1874 /* Break if we really can't fit anymore. */
1875 if ((frags + 1) >= seglimit) {
1877 if (pkthdrlen != NULL)
1878 *pkthdrlen = len_cp;
1883 * Reduce size if you can't copy the whole
1884 * mbuf. If we can't copy the whole mbuf, also
1885 * adjust len so the loop will end after this
1888 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1889 mlen = (seglimit - frags - 1) * fragsize;
1891 *plen = len_cp + len;
1892 if (pkthdrlen != NULL)
1895 frags += howmany(mlen, fragsize);
1899 KASSERT(seglimit > 0,
1900 ("%s: seglimit went too low", __func__));
1903 n = m_gethdr(M_NOWAIT, m->m_type);
1905 n = m_get(M_NOWAIT, m->m_type);
1910 if (!m_dup_pkthdr(n, m, M_NOWAIT))
1912 if (len == M_COPYALL)
1913 n->m_pkthdr.len -= off0;
1915 n->m_pkthdr.len = len;
1916 pkthdrlen = &n->m_pkthdr.len;
1921 if (m->m_flags & M_EXT) {
1922 n->m_data = m->m_data + off;
1925 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1928 if (sb && (sb->sb_sndptr == m) &&
1929 ((n->m_len + off) >= m->m_len) && m->m_next) {
1930 sb->sb_sndptroff += m->m_len;
1931 sb->sb_sndptr = m->m_next;
1934 if (len != M_COPYALL) {
1947 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1951 * Automatic sizing of send socket buffer. Often the send buffer
1952 * size is not optimally adjusted to the actual network conditions
1953 * at hand (delay bandwidth product). Setting the buffer size too
1954 * small limits throughput on links with high bandwidth and high
1955 * delay (eg. trans-continental/oceanic links). Setting the
1956 * buffer size too big consumes too much real kernel memory,
1957 * especially with many connections on busy servers.
1959 * The criteria to step up the send buffer one notch are:
1960 * 1. receive window of remote host is larger than send buffer
1961 * (with a fudge factor of 5/4th);
1962 * 2. send buffer is filled to 7/8th with data (so we actually
1963 * have data to make use of it);
1964 * 3. send buffer fill has not hit maximal automatic size;
1965 * 4. our send window (slow start and cogestion controlled) is
1966 * larger than sent but unacknowledged data in send buffer.
1968 * The remote host receive window scaling factor may limit the
1969 * growing of the send buffer before it reaches its allowed
1972 * It scales directly with slow start or congestion window
1973 * and does at most one step per received ACK. This fast
1974 * scaling has the drawback of growing the send buffer beyond
1975 * what is strictly necessary to make full use of a given
1976 * delay*bandwidth product. However testing has shown this not
1977 * to be much of an problem. At worst we are trading wasting
1978 * of available bandwidth (the non-use of it) for wasting some
1979 * socket buffer memory.
1981 * TODO: Shrink send buffer during idle periods together
1982 * with congestion window. Requires another timer. Has to
1983 * wait for upcoming tcp timer rewrite.
1985 * XXXGL: should there be used sbused() or sbavail()?
1987 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
1990 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
1991 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
1992 sbused(&so->so_snd) >=
1993 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
1994 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
1995 sendwin >= (sbused(&so->so_snd) -
1996 (tp->snd_nxt - tp->snd_una))) {
1997 if (!sbreserve_locked(&so->so_snd,
1998 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
1999 V_tcp_autosndbuf_max), so, curthread))
2000 so->so_snd.sb_flags &= ~SB_AUTOSIZE;