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) & (TH_SYN | TH_FIN)) == 0) ||\
142 tcp_timer_active((tp), TT_REXMT) || \
143 tcp_timer_active((tp), TT_PERSIST), \
144 ("neither rexmt nor persist timer is set"))
146 static void inline cc_after_idle(struct tcpcb *tp);
150 * Wrapper for the TCP established output helper hook.
153 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
154 struct tcpopt *to, uint32_t len, int tso)
156 struct tcp_hhook_data hhook_data;
158 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
162 hhook_data.len = len;
163 hhook_data.tso = tso;
165 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
172 * CC wrapper hook functions
175 cc_after_idle(struct tcpcb *tp)
177 INP_WLOCK_ASSERT(tp->t_inpcb);
179 if (CC_ALGO(tp)->after_idle != NULL)
180 CC_ALGO(tp)->after_idle(tp->ccv);
184 * Tcp output routine: figure out what should be sent and send it.
187 tcp_output(struct tcpcb *tp)
189 struct socket *so = tp->t_inpcb->inp_socket;
191 uint32_t recwin, sendwin;
192 int off, flags, error = 0; /* Keep compiler happy */
193 u_int if_hw_tsomaxsegcount = 0;
194 u_int if_hw_tsomaxsegsize;
196 struct ip *ip = NULL;
198 struct ipovly *ipov = NULL;
201 u_char opt[TCP_MAXOLEN];
202 unsigned ipoptlen, optlen, hdrlen;
203 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
204 unsigned ipsec_optlen = 0;
206 int idle, sendalot, curticks;
207 int sack_rxmit, sack_bytes_rxmt;
211 unsigned int wanted_cookie = 0;
212 unsigned int dont_sendalot = 0;
214 int maxburst = TCP_MAXBURST;
217 struct ip6_hdr *ip6 = NULL;
220 isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
223 INP_WLOCK_ASSERT(tp->t_inpcb);
226 if (tp->t_flags & TF_TOE)
227 return (tcp_offload_output(tp));
231 * For TFO connections in SYN_SENT or SYN_RECEIVED,
232 * only allow the initial SYN or SYN|ACK and those sent
233 * by the retransmit timer.
235 if (IS_FASTOPEN(tp->t_flags) &&
236 ((tp->t_state == TCPS_SYN_SENT) ||
237 (tp->t_state == TCPS_SYN_RECEIVED)) &&
238 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
239 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
243 * Determine length of data that should be transmitted,
244 * and flags that will be used.
245 * If there is some data or critical controls (SYN, RST)
246 * to send, then transmit; otherwise, investigate further.
248 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
249 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
250 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= 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, tcp_maxseg(tp)));
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 * As the TCP header options are now
581 * considered when setting up the initial
582 * window, we would not send the last segment
583 * if we skip considering the option length here.
584 * Note: this may not work when tcp headers change
585 * very dynamically in the future.
587 if ((((tp->t_flags & TF_SIGNATURE) ?
588 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
589 ((tp->t_flags & TF_RCVD_TSTMP) ?
590 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
591 len) >= tp->t_maxseg)
594 * NOTE! on localhost connections an 'ack' from the remote
595 * end may occur synchronously with the output and cause
596 * us to flush a buffer queued with moretocome. XXX
598 * note: the len + off check is almost certainly unnecessary.
600 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
601 (idle || (tp->t_flags & TF_NODELAY)) &&
602 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
603 (tp->t_flags & TF_NOPUSH) == 0) {
606 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
608 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
610 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
617 * Sending of standalone window updates.
619 * Window updates are important when we close our window due to a
620 * full socket buffer and are opening it again after the application
621 * reads data from it. Once the window has opened again and the
622 * remote end starts to send again the ACK clock takes over and
623 * provides the most current window information.
625 * We must avoid the silly window syndrome whereas every read
626 * from the receive buffer, no matter how small, causes a window
627 * update to be sent. We also should avoid sending a flurry of
628 * window updates when the socket buffer had queued a lot of data
629 * and the application is doing small reads.
631 * Prevent a flurry of pointless window updates by only sending
632 * an update when we can increase the advertized window by more
633 * than 1/4th of the socket buffer capacity. When the buffer is
634 * getting full or is very small be more aggressive and send an
635 * update whenever we can increase by two mss sized segments.
636 * In all other situations the ACK's to new incoming data will
637 * carry further window increases.
639 * Don't send an independent window update if a delayed
640 * ACK is pending (it will get piggy-backed on it) or the
641 * remote side already has done a half-close and won't send
642 * more data. Skip this if the connection is in T/TCP
645 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
646 !(tp->t_flags & TF_DELACK) &&
647 !TCPS_HAVERCVDFIN(tp->t_state)) {
649 * "adv" is the amount we could increase the window,
650 * taking into account that we are limited by
651 * TCP_MAXWIN << tp->rcv_scale.
657 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
658 oldwin = (tp->rcv_adv - tp->rcv_nxt);
664 * If the new window size ends up being the same as or less
665 * than the old size when it is scaled, then don't force
668 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
671 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
672 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
673 recwin <= (so->so_rcv.sb_hiwat / 8) ||
674 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
675 adv >= TCP_MAXWIN << tp->rcv_scale))
677 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
683 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
684 * is also a catch-all for the retransmit timer timeout case.
686 if (tp->t_flags & TF_ACKNOW)
688 if ((flags & TH_RST) ||
689 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
691 if (SEQ_GT(tp->snd_up, tp->snd_una))
694 * If our state indicates that FIN should be sent
695 * and we have not yet done so, then we need to send.
697 if (flags & TH_FIN &&
698 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
701 * In SACK, it is possible for tcp_output to fail to send a segment
702 * after the retransmission timer has been turned off. Make sure
703 * that the retransmission timer is set.
705 if ((tp->t_flags & TF_SACK_PERMIT) &&
706 SEQ_GT(tp->snd_max, tp->snd_una) &&
707 !tcp_timer_active(tp, TT_REXMT) &&
708 !tcp_timer_active(tp, TT_PERSIST)) {
709 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
713 * TCP window updates are not reliable, rather a polling protocol
714 * using ``persist'' packets is used to insure receipt of window
715 * updates. The three ``states'' for the output side are:
716 * idle not doing retransmits or persists
717 * persisting to move a small or zero window
718 * (re)transmitting and thereby not persisting
720 * tcp_timer_active(tp, TT_PERSIST)
721 * is true when we are in persist state.
722 * (tp->t_flags & TF_FORCEDATA)
723 * is set when we are called to send a persist packet.
724 * tcp_timer_active(tp, TT_REXMT)
725 * is set when we are retransmitting
726 * The output side is idle when both timers are zero.
728 * If send window is too small, there is data to transmit, and no
729 * retransmit or persist is pending, then go to persist state.
730 * If nothing happens soon, send when timer expires:
731 * if window is nonzero, transmit what we can,
732 * otherwise force out a byte.
734 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
735 !tcp_timer_active(tp, TT_PERSIST)) {
741 * No reason to send a segment, just return.
744 SOCKBUF_UNLOCK(&so->so_snd);
748 SOCKBUF_LOCK_ASSERT(&so->so_snd);
750 if (len >= tp->t_maxseg)
751 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
753 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
756 * Before ESTABLISHED, force sending of initial options
757 * unless TCP set not to do any options.
758 * NOTE: we assume that the IP/TCP header plus TCP options
759 * always fit in a single mbuf, leaving room for a maximum
761 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
766 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
769 hdrlen = sizeof (struct tcpiphdr);
772 * Compute options for segment.
773 * We only have to care about SYN and established connection
774 * segments. Options for SYN-ACK segments are handled in TCP
778 if ((tp->t_flags & TF_NOOPT) == 0) {
779 /* Maximum segment size. */
780 if (flags & TH_SYN) {
781 tp->snd_nxt = tp->iss;
782 to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
783 to.to_flags |= TOF_MSS;
786 * On SYN or SYN|ACK transmits on TFO connections,
787 * only include the TFO option if it is not a
788 * retransmit, as the presence of the TFO option may
789 * have caused the original SYN or SYN|ACK to have
790 * been dropped by a middlebox.
792 if (IS_FASTOPEN(tp->t_flags) &&
793 (tp->t_rxtshift == 0)) {
794 if (tp->t_state == TCPS_SYN_RECEIVED) {
795 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
797 (u_int8_t *)&tp->t_tfo_cookie.server;
798 to.to_flags |= TOF_FASTOPEN;
800 } else if (tp->t_state == TCPS_SYN_SENT) {
802 tp->t_tfo_client_cookie_len;
804 tp->t_tfo_cookie.client;
805 to.to_flags |= TOF_FASTOPEN;
808 * If we wind up having more data to
809 * send with the SYN than can fit in
810 * one segment, don't send any more
811 * until the SYN|ACK comes back from
818 /* Window scaling. */
819 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
820 to.to_wscale = tp->request_r_scale;
821 to.to_flags |= TOF_SCALE;
824 if ((tp->t_flags & TF_RCVD_TSTMP) ||
825 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
826 curticks = tcp_ts_getticks();
827 to.to_tsval = curticks + tp->ts_offset;
828 to.to_tsecr = tp->ts_recent;
829 to.to_flags |= TOF_TS;
830 if (tp->t_rxtshift == 1)
831 tp->t_badrxtwin = curticks;
834 /* Set receive buffer autosizing timestamp. */
835 if (tp->rfbuf_ts == 0 &&
836 (so->so_rcv.sb_flags & SB_AUTOSIZE))
837 tp->rfbuf_ts = tcp_ts_getticks();
839 /* Selective ACK's. */
840 if (tp->t_flags & TF_SACK_PERMIT) {
842 to.to_flags |= TOF_SACKPERM;
843 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
844 tp->rcv_numsacks > 0) {
845 to.to_flags |= TOF_SACK;
846 to.to_nsacks = tp->rcv_numsacks;
847 to.to_sacks = (u_char *)tp->sackblks;
850 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
851 /* TCP-MD5 (RFC2385). */
853 * Check that TCP_MD5SIG is enabled in tcpcb to
854 * account the size needed to set this TCP option.
856 if (tp->t_flags & TF_SIGNATURE)
857 to.to_flags |= TOF_SIGNATURE;
858 #endif /* TCP_SIGNATURE */
860 /* Processing the options. */
861 hdrlen += optlen = tcp_addoptions(&to, opt);
863 * If we wanted a TFO option to be added, but it was unable
864 * to fit, ensure no data is sent.
866 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
867 !(to.to_flags & TOF_FASTOPEN))
872 * Adjust data length if insertion of options will
873 * bump the packet length beyond the t_maxseg length.
874 * Clear the FIN bit because we cut off the tail of
877 if (len + optlen + ipoptlen > tp->t_maxseg) {
885 /* extract TSO information */
886 if_hw_tsomax = tp->t_tsomax;
887 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
888 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
891 * Limit a TSO burst to prevent it from
892 * overflowing or exceeding the maximum length
893 * allowed by the network interface:
895 KASSERT(ipoptlen == 0,
896 ("%s: TSO can't do IP options", __func__));
899 * Check if we should limit by maximum payload
902 if (if_hw_tsomax != 0) {
903 /* compute maximum TSO length */
904 max_len = (if_hw_tsomax - hdrlen -
908 } else if (len > max_len) {
914 * Prevent the last segment from being
915 * fractional unless the send sockbuf can be
918 max_len = (tp->t_maxseg - optlen);
919 if (((uint32_t)off + (uint32_t)len) <
920 sbavail(&so->so_snd)) {
921 moff = len % max_len;
929 * In case there are too many small fragments
932 if (len <= max_len) {
939 * Send the FIN in a separate segment
940 * after the bulk sending is done.
941 * We don't trust the TSO implementations
942 * to clear the FIN flag on all but the
945 if (tp->t_flags & TF_NEEDFIN)
948 if (optlen + ipoptlen >= tp->t_maxseg) {
950 * Since we don't have enough space to put
951 * the IP header chain and the TCP header in
952 * one packet as required by RFC 7112, don't
953 * send it. Also ensure that at least one
954 * byte of the payload can be put into the
957 SOCKBUF_UNLOCK(&so->so_snd);
962 len = tp->t_maxseg - optlen - ipoptlen;
970 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
971 ("%s: len > IP_MAXPACKET", __func__));
973 /*#ifdef DIAGNOSTIC*/
975 if (max_linkhdr + hdrlen > MCLBYTES)
977 if (max_linkhdr + hdrlen > MHLEN)
979 panic("tcphdr too big");
983 * This KASSERT is here to catch edge cases at a well defined place.
984 * Before, those had triggered (random) panic conditions further down.
986 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
989 * Grab a header mbuf, attaching a copy of data to
990 * be transmitted, and initialize the header from
991 * the template for sends on this connection.
998 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
999 TCPSTAT_INC(tcps_sndprobe);
1000 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1001 tp->t_sndrexmitpack++;
1002 TCPSTAT_INC(tcps_sndrexmitpack);
1003 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1005 TCPSTAT_INC(tcps_sndpack);
1006 TCPSTAT_ADD(tcps_sndbyte, len);
1009 if (MHLEN < hdrlen + max_linkhdr)
1010 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1013 m = m_gethdr(M_NOWAIT, MT_DATA);
1016 SOCKBUF_UNLOCK(&so->so_snd);
1022 m->m_data += max_linkhdr;
1026 * Start the m_copy functions from the closest mbuf
1027 * to the offset in the socket buffer chain.
1029 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1030 if (len <= MHLEN - hdrlen - max_linkhdr) {
1031 m_copydata(mb, moff, len,
1032 mtod(m, caddr_t) + hdrlen);
1033 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1034 sbsndptr_adv(&so->so_snd, mb, len);
1037 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1041 m->m_next = tcp_m_copym(mb, moff,
1042 &len, if_hw_tsomaxsegcount,
1043 if_hw_tsomaxsegsize, msb);
1044 if (len <= (tp->t_maxseg - optlen)) {
1046 * Must have ran out of mbufs for the copy
1047 * shorten it to no longer need tso. Lets
1048 * not put on sendalot since we are low on
1053 if (m->m_next == NULL) {
1054 SOCKBUF_UNLOCK(&so->so_snd);
1063 * If we're sending everything we've got, set PUSH.
1064 * (This will keep happy those implementations which only
1065 * give data to the user when a buffer fills or
1068 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1071 SOCKBUF_UNLOCK(&so->so_snd);
1073 SOCKBUF_UNLOCK(&so->so_snd);
1074 if (tp->t_flags & TF_ACKNOW)
1075 TCPSTAT_INC(tcps_sndacks);
1076 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1077 TCPSTAT_INC(tcps_sndctrl);
1078 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1079 TCPSTAT_INC(tcps_sndurg);
1081 TCPSTAT_INC(tcps_sndwinup);
1083 m = m_gethdr(M_NOWAIT, MT_DATA);
1090 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1095 m->m_data += max_linkhdr;
1098 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1099 m->m_pkthdr.rcvif = (struct ifnet *)0;
1101 mac_inpcb_create_mbuf(tp->t_inpcb, m);
1105 ip6 = mtod(m, struct ip6_hdr *);
1106 th = (struct tcphdr *)(ip6 + 1);
1107 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1111 ip = mtod(m, struct ip *);
1113 ipov = (struct ipovly *)ip;
1115 th = (struct tcphdr *)(ip + 1);
1116 tcpip_fillheaders(tp->t_inpcb, ip, th);
1120 * Fill in fields, remembering maximum advertised
1121 * window for use in delaying messages about window sizes.
1122 * If resending a FIN, be sure not to use a new sequence number.
1124 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1125 tp->snd_nxt == tp->snd_max)
1128 * If we are starting a connection, send ECN setup
1129 * SYN packet. If we are on a retransmit, we may
1130 * resend those bits a number of times as per
1133 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1134 if (tp->t_rxtshift >= 1) {
1135 if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1136 flags |= TH_ECE|TH_CWR;
1138 flags |= TH_ECE|TH_CWR;
1141 if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1142 (tp->t_flags & TF_ECN_PERMIT)) {
1144 * If the peer has ECN, mark data packets with
1145 * ECN capable transmission (ECT).
1146 * Ignore pure ack packets, retransmissions and window probes.
1148 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1149 (sack_rxmit == 0) &&
1150 !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1151 SEQ_LT(tp->snd_una, tp->snd_max))) {
1154 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1157 ip->ip_tos |= IPTOS_ECN_ECT0;
1158 TCPSTAT_INC(tcps_ecn_ect0);
1160 * Reply with proper ECN notifications.
1161 * Only set CWR on new data segments.
1163 if (tp->t_flags & TF_ECN_SND_CWR) {
1165 tp->t_flags &= ~TF_ECN_SND_CWR;
1168 if (tp->t_flags & TF_ECN_SND_ECE)
1173 * If we are doing retransmissions, then snd_nxt will
1174 * not reflect the first unsent octet. For ACK only
1175 * packets, we do not want the sequence number of the
1176 * retransmitted packet, we want the sequence number
1177 * of the next unsent octet. So, if there is no data
1178 * (and no SYN or FIN), use snd_max instead of snd_nxt
1179 * when filling in ti_seq. But if we are in persist
1180 * state, snd_max might reflect one byte beyond the
1181 * right edge of the window, so use snd_nxt in that
1182 * case, since we know we aren't doing a retransmission.
1183 * (retransmit and persist are mutually exclusive...)
1185 if (sack_rxmit == 0) {
1186 if (len || (flags & (TH_SYN|TH_FIN)) ||
1187 tcp_timer_active(tp, TT_PERSIST))
1188 th->th_seq = htonl(tp->snd_nxt);
1190 th->th_seq = htonl(tp->snd_max);
1192 th->th_seq = htonl(p->rxmit);
1194 tp->sackhint.sack_bytes_rexmit += len;
1196 th->th_ack = htonl(tp->rcv_nxt);
1198 bcopy(opt, th + 1, optlen);
1199 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1201 th->th_flags = flags;
1203 * Calculate receive window. Don't shrink window,
1204 * but avoid silly window syndrome.
1205 * If a RST segment is sent, advertise a window of zero.
1207 if (flags & TH_RST) {
1210 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1211 recwin < tp->t_maxseg)
1213 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1214 recwin < (tp->rcv_adv - tp->rcv_nxt))
1215 recwin = (tp->rcv_adv - tp->rcv_nxt);
1218 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1219 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1220 * case is handled in syncache.
1223 th->th_win = htons((u_short)
1224 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1226 /* Avoid shrinking window with window scaling. */
1227 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1228 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1232 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1233 * a 0 window. This may cause the remote transmitter to stall. This
1234 * flag tells soreceive() to disable delayed acknowledgements when
1235 * draining the buffer. This can occur if the receiver is attempting
1236 * to read more data than can be buffered prior to transmitting on
1239 if (th->th_win == 0) {
1241 tp->t_flags |= TF_RXWIN0SENT;
1243 tp->t_flags &= ~TF_RXWIN0SENT;
1244 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1245 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1246 th->th_flags |= TH_URG;
1249 * If no urgent pointer to send, then we pull
1250 * the urgent pointer to the left edge of the send window
1251 * so that it doesn't drift into the send window on sequence
1252 * number wraparound.
1254 tp->snd_up = tp->snd_una; /* drag it along */
1257 * Put TCP length in extended header, and then
1258 * checksum extended header and data.
1260 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1261 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1263 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1264 if (to.to_flags & TOF_SIGNATURE) {
1266 * Calculate MD5 signature and put it into the place
1267 * determined before.
1268 * NOTE: since TCP options buffer doesn't point into
1269 * mbuf's data, calculate offset and use it.
1271 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1272 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1274 * Do not send segment if the calculation of MD5
1275 * digest has failed.
1285 * There is no need to fill in ip6_plen right now.
1286 * It will be filled later by ip6_output.
1288 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1289 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1290 optlen + len, IPPROTO_TCP, 0);
1293 #if defined(INET6) && defined(INET)
1298 m->m_pkthdr.csum_flags = CSUM_TCP;
1299 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1300 htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1302 /* IP version must be set here for ipv4/ipv6 checking later */
1303 KASSERT(ip->ip_v == IPVERSION,
1304 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1309 * Enable TSO and specify the size of the segments.
1310 * The TCP pseudo header checksum is always provided.
1313 KASSERT(len > tp->t_maxseg - optlen,
1314 ("%s: len <= tso_segsz", __func__));
1315 m->m_pkthdr.csum_flags |= CSUM_TSO;
1316 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1319 KASSERT(len + hdrlen == m_length(m, NULL),
1320 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1321 __func__, len, hdrlen, m_length(m, NULL)));
1324 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1325 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1332 if (so->so_options & SO_DEBUG) {
1338 save = ipov->ih_len;
1339 ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1341 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1345 ipov->ih_len = save;
1347 #endif /* TCPDEBUG */
1348 TCP_PROBE3(debug__output, tp, th, m);
1350 /* We're getting ready to send; log now. */
1351 TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1355 * Fill in IP length and desired time to live and
1356 * send to IP level. There should be a better way
1357 * to handle ttl and tos; we could keep them in
1358 * the template, but need a way to checksum without them.
1361 * m->m_pkthdr.len should have been set before checksum calculation,
1362 * because in6_cksum() need it.
1367 * we separately set hoplimit for every segment, since the
1368 * user might want to change the value via setsockopt.
1369 * Also, desired default hop limit might be changed via
1370 * Neighbor Discovery.
1372 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1375 * Set the packet size here for the benefit of DTrace probes.
1376 * ip6_output() will set it properly; it's supposed to include
1377 * the option header lengths as well.
1379 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1381 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1382 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1384 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1386 if (tp->t_state == TCPS_SYN_SENT)
1387 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1389 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1392 /* Save packet, if requested. */
1393 tcp_pcap_add(th, m, &(tp->t_outpkts));
1396 /* TODO: IPv6 IP6TOS_ECT bit on */
1397 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1398 &tp->t_inpcb->inp_route6,
1399 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1400 NULL, NULL, tp->t_inpcb);
1402 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1403 mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1406 #if defined(INET) && defined(INET6)
1411 ip->ip_len = htons(m->m_pkthdr.len);
1413 if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1414 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1417 * If we do path MTU discovery, then we set DF on every packet.
1418 * This might not be the best thing to do according to RFC3390
1419 * Section 2. However the tcp hostcache migitates the problem
1420 * so it affects only the first tcp connection with a host.
1422 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1424 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1425 ip->ip_off |= htons(IP_DF);
1426 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1428 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1431 if (tp->t_state == TCPS_SYN_SENT)
1432 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1434 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1437 /* Save packet, if requested. */
1438 tcp_pcap_add(th, m, &(tp->t_outpkts));
1441 error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1442 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1445 if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1446 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1452 * In transmit state, time the transmission and arrange for
1453 * the retransmit. In persist state, just set snd_max.
1455 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1456 !tcp_timer_active(tp, TT_PERSIST)) {
1457 tcp_seq startseq = tp->snd_nxt;
1460 * Advance snd_nxt over sequence space of this segment.
1462 if (flags & (TH_SYN|TH_FIN)) {
1465 if (flags & TH_FIN) {
1467 tp->t_flags |= TF_SENTFIN;
1473 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1474 tp->snd_max = tp->snd_nxt;
1476 * Time this transmission if not a retransmission and
1477 * not currently timing anything.
1479 tp->t_sndtime = ticks;
1480 if (tp->t_rtttime == 0) {
1481 tp->t_rtttime = ticks;
1482 tp->t_rtseq = startseq;
1483 TCPSTAT_INC(tcps_segstimed);
1488 * Set retransmit timer if not currently set,
1489 * and not doing a pure ack or a keep-alive probe.
1490 * Initial value for retransmit timer is smoothed
1491 * round-trip time + 2 * round-trip time variance.
1492 * Initialize shift counter which is used for backoff
1493 * of retransmit time.
1496 if (!tcp_timer_active(tp, TT_REXMT) &&
1497 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1498 (tp->snd_nxt != tp->snd_una))) {
1499 if (tcp_timer_active(tp, TT_PERSIST)) {
1500 tcp_timer_activate(tp, TT_PERSIST, 0);
1503 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1504 } else if (len == 0 && sbavail(&so->so_snd) &&
1505 !tcp_timer_active(tp, TT_REXMT) &&
1506 !tcp_timer_active(tp, TT_PERSIST)) {
1508 * Avoid a situation where we do not set persist timer
1509 * after a zero window condition. For example:
1510 * 1) A -> B: packet with enough data to fill the window
1511 * 2) B -> A: ACK for #1 + new data (0 window
1513 * 3) A -> B: ACK for #2, 0 len packet
1515 * In this case, A will not activate the persist timer,
1516 * because it chose to send a packet. Unless tcp_output
1517 * is called for some other reason (delayed ack timer,
1518 * another input packet from B, socket syscall), A will
1519 * not send zero window probes.
1521 * So, if you send a 0-length packet, but there is data
1522 * in the socket buffer, and neither the rexmt or
1523 * persist timer is already set, then activate the
1531 * Persist case, update snd_max but since we are in
1532 * persist mode (no window) we do not update snd_nxt.
1537 if (flags & TH_FIN) {
1539 tp->t_flags |= TF_SENTFIN;
1541 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1542 tp->snd_max = tp->snd_nxt + xlen;
1545 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1546 (tp->t_flags & TF_SACK_PERMIT) &&
1547 tp->rcv_numsacks > 0)) {
1548 /* Clean up any DSACK's sent */
1549 tcp_clean_dsack_blocks(tp);
1552 /* Record the error. */
1553 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1554 error, 0, NULL, false);
1557 * We know that the packet was lost, so back out the
1558 * sequence number advance, if any.
1560 * If the error is EPERM the packet got blocked by the
1561 * local firewall. Normally we should terminate the
1562 * connection but the blocking may have been spurious
1563 * due to a firewall reconfiguration cycle. So we treat
1564 * it like a packet loss and let the retransmit timer and
1565 * timeouts do their work over time.
1566 * XXX: It is a POLA question whether calling tcp_drop right
1567 * away would be the really correct behavior instead.
1569 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1570 !tcp_timer_active(tp, TT_PERSIST)) &&
1571 ((flags & TH_SYN) == 0) &&
1575 tp->sackhint.sack_bytes_rexmit -= len;
1576 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1577 ("sackhint bytes rtx >= 0"));
1581 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1585 tp->t_softerror = error;
1588 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1589 tp->snd_cwnd = tp->t_maxseg;
1593 * For some reason the interface we used initially
1594 * to send segments changed to another or lowered
1596 * If TSO was active we either got an interface
1597 * without TSO capabilits or TSO was turned off.
1598 * If we obtained mtu from ip_output() then update
1602 tp->t_flags &= ~TF_TSO;
1604 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1612 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1613 tp->t_softerror = error;
1621 TCPSTAT_INC(tcps_sndtotal);
1624 * Data sent (as far as we can tell).
1625 * If this advertises a larger window than any other segment,
1626 * then remember the size of the advertised window.
1627 * Any pending ACK has now been sent.
1629 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1630 tp->rcv_adv = tp->rcv_nxt + recwin;
1631 tp->last_ack_sent = tp->rcv_nxt;
1632 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1633 if (tcp_timer_active(tp, TT_DELACK))
1634 tcp_timer_activate(tp, TT_DELACK, 0);
1637 * This completely breaks TCP if newreno is turned on. What happens
1638 * is that if delayed-acks are turned on on the receiver, this code
1639 * on the transmitter effectively destroys the TCP window, forcing
1640 * it to four packets (1.5Kx4 = 6K window).
1642 if (sendalot && --maxburst)
1651 tcp_setpersist(struct tcpcb *tp)
1653 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1656 tp->t_flags &= ~TF_PREVVALID;
1657 if (tcp_timer_active(tp, TT_REXMT))
1658 panic("tcp_setpersist: retransmit pending");
1660 * Start/restart persistence timer.
1662 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1663 tcp_persmin, tcp_persmax);
1664 tcp_timer_activate(tp, TT_PERSIST, tt);
1665 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1670 * Insert TCP options according to the supplied parameters to the place
1671 * optp in a consistent way. Can handle unaligned destinations.
1673 * The order of the option processing is crucial for optimal packing and
1674 * alignment for the scarce option space.
1676 * The optimal order for a SYN/SYN-ACK segment is:
1677 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1678 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1680 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1681 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1682 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1683 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1684 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1687 tcp_addoptions(struct tcpopt *to, u_char *optp)
1689 u_int32_t mask, optlen = 0;
1691 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1692 if ((to->to_flags & mask) != mask)
1694 if (optlen == TCP_MAXOLEN)
1696 switch (to->to_flags & mask) {
1698 while (optlen % 4) {
1699 optlen += TCPOLEN_NOP;
1700 *optp++ = TCPOPT_NOP;
1702 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1704 optlen += TCPOLEN_MAXSEG;
1705 *optp++ = TCPOPT_MAXSEG;
1706 *optp++ = TCPOLEN_MAXSEG;
1707 to->to_mss = htons(to->to_mss);
1708 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1709 optp += sizeof(to->to_mss);
1712 while (!optlen || optlen % 2 != 1) {
1713 optlen += TCPOLEN_NOP;
1714 *optp++ = TCPOPT_NOP;
1716 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1718 optlen += TCPOLEN_WINDOW;
1719 *optp++ = TCPOPT_WINDOW;
1720 *optp++ = TCPOLEN_WINDOW;
1721 *optp++ = to->to_wscale;
1724 while (optlen % 2) {
1725 optlen += TCPOLEN_NOP;
1726 *optp++ = TCPOPT_NOP;
1728 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1730 optlen += TCPOLEN_SACK_PERMITTED;
1731 *optp++ = TCPOPT_SACK_PERMITTED;
1732 *optp++ = TCPOLEN_SACK_PERMITTED;
1735 while (!optlen || optlen % 4 != 2) {
1736 optlen += TCPOLEN_NOP;
1737 *optp++ = TCPOPT_NOP;
1739 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1741 optlen += TCPOLEN_TIMESTAMP;
1742 *optp++ = TCPOPT_TIMESTAMP;
1743 *optp++ = TCPOLEN_TIMESTAMP;
1744 to->to_tsval = htonl(to->to_tsval);
1745 to->to_tsecr = htonl(to->to_tsecr);
1746 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1747 optp += sizeof(to->to_tsval);
1748 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1749 optp += sizeof(to->to_tsecr);
1753 int siglen = TCPOLEN_SIGNATURE - 2;
1755 while (!optlen || optlen % 4 != 2) {
1756 optlen += TCPOLEN_NOP;
1757 *optp++ = TCPOPT_NOP;
1759 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1760 to->to_flags &= ~TOF_SIGNATURE;
1763 optlen += TCPOLEN_SIGNATURE;
1764 *optp++ = TCPOPT_SIGNATURE;
1765 *optp++ = TCPOLEN_SIGNATURE;
1766 to->to_signature = optp;
1774 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1777 while (!optlen || optlen % 4 != 2) {
1778 optlen += TCPOLEN_NOP;
1779 *optp++ = TCPOPT_NOP;
1781 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1783 optlen += TCPOLEN_SACKHDR;
1784 *optp++ = TCPOPT_SACK;
1785 sackblks = min(to->to_nsacks,
1786 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1787 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1788 while (sackblks--) {
1789 sack_seq = htonl(sack->start);
1790 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1791 optp += sizeof(sack_seq);
1792 sack_seq = htonl(sack->end);
1793 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1794 optp += sizeof(sack_seq);
1795 optlen += TCPOLEN_SACK;
1798 TCPSTAT_INC(tcps_sack_send_blocks);
1805 /* XXX is there any point to aligning this option? */
1806 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1807 if (TCP_MAXOLEN - optlen < total_len) {
1808 to->to_flags &= ~TOF_FASTOPEN;
1811 *optp++ = TCPOPT_FAST_OPEN;
1812 *optp++ = total_len;
1813 if (to->to_tfo_len > 0) {
1814 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1815 optp += to->to_tfo_len;
1817 optlen += total_len;
1821 panic("%s: unknown TCP option type", __func__);
1826 /* Terminate and pad TCP options to a 4 byte boundary. */
1828 optlen += TCPOLEN_EOL;
1829 *optp++ = TCPOPT_EOL;
1832 * According to RFC 793 (STD0007):
1833 * "The content of the header beyond the End-of-Option option
1834 * must be header padding (i.e., zero)."
1835 * and later: "The padding is composed of zeros."
1837 while (optlen % 4) {
1838 optlen += TCPOLEN_PAD;
1839 *optp++ = TCPOPT_PAD;
1842 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1847 * This is a copy of m_copym(), taking the TSO segment size/limit
1848 * constraints into account, and advancing the sndptr as it goes.
1851 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1852 int32_t seglimit, int32_t segsize, struct sockbuf *sb)
1854 struct mbuf *n, **np;
1857 int32_t len = *plen;
1861 uint32_t mlen, frags;
1865 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1866 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1867 if (off == 0 && m->m_flags & M_PKTHDR)
1872 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1876 if ((sb) && (m == sb->sb_sndptr)) {
1877 sb->sb_sndptroff += m->m_len;
1878 sb->sb_sndptr = m->m_next;
1887 KASSERT(len == M_COPYALL,
1888 ("tcp_m_copym, length > size of mbuf chain"));
1890 if (pkthdrlen != NULL)
1891 *pkthdrlen = len_cp;
1894 mlen = min(len, m->m_len - off);
1897 * For M_NOMAP mbufs, add 3 segments
1898 * + 1 in case we are crossing page boundaries
1899 * + 2 in case the TLS hdr/trailer are used
1900 * It is cheaper to just add the segments
1901 * than it is to take the cache miss to look
1902 * at the mbuf ext_pgs state in detail.
1904 if (m->m_flags & M_NOMAP) {
1905 fragsize = min(segsize, PAGE_SIZE);
1912 /* Break if we really can't fit anymore. */
1913 if ((frags + 1) >= seglimit) {
1915 if (pkthdrlen != NULL)
1916 *pkthdrlen = len_cp;
1921 * Reduce size if you can't copy the whole
1922 * mbuf. If we can't copy the whole mbuf, also
1923 * adjust len so the loop will end after this
1926 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1927 mlen = (seglimit - frags - 1) * fragsize;
1929 *plen = len_cp + len;
1930 if (pkthdrlen != NULL)
1933 frags += howmany(mlen, fragsize);
1937 KASSERT(seglimit > 0,
1938 ("%s: seglimit went too low", __func__));
1941 n = m_gethdr(M_NOWAIT, m->m_type);
1943 n = m_get(M_NOWAIT, m->m_type);
1948 if (!m_dup_pkthdr(n, m, M_NOWAIT))
1950 if (len == M_COPYALL)
1951 n->m_pkthdr.len -= off0;
1953 n->m_pkthdr.len = len;
1954 pkthdrlen = &n->m_pkthdr.len;
1959 if (m->m_flags & M_EXT) {
1960 n->m_data = m->m_data + off;
1963 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1966 if (sb && (sb->sb_sndptr == m) &&
1967 ((n->m_len + off) >= m->m_len) && m->m_next) {
1968 sb->sb_sndptroff += m->m_len;
1969 sb->sb_sndptr = m->m_next;
1972 if (len != M_COPYALL) {
1985 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
1989 * Automatic sizing of send socket buffer. Often the send buffer
1990 * size is not optimally adjusted to the actual network conditions
1991 * at hand (delay bandwidth product). Setting the buffer size too
1992 * small limits throughput on links with high bandwidth and high
1993 * delay (eg. trans-continental/oceanic links). Setting the
1994 * buffer size too big consumes too much real kernel memory,
1995 * especially with many connections on busy servers.
1997 * The criteria to step up the send buffer one notch are:
1998 * 1. receive window of remote host is larger than send buffer
1999 * (with a fudge factor of 5/4th);
2000 * 2. send buffer is filled to 7/8th with data (so we actually
2001 * have data to make use of it);
2002 * 3. send buffer fill has not hit maximal automatic size;
2003 * 4. our send window (slow start and cogestion controlled) is
2004 * larger than sent but unacknowledged data in send buffer.
2006 * The remote host receive window scaling factor may limit the
2007 * growing of the send buffer before it reaches its allowed
2010 * It scales directly with slow start or congestion window
2011 * and does at most one step per received ACK. This fast
2012 * scaling has the drawback of growing the send buffer beyond
2013 * what is strictly necessary to make full use of a given
2014 * delay*bandwidth product. However testing has shown this not
2015 * to be much of an problem. At worst we are trading wasting
2016 * of available bandwidth (the non-use of it) for wasting some
2017 * socket buffer memory.
2019 * TODO: Shrink send buffer during idle periods together
2020 * with congestion window. Requires another timer. Has to
2021 * wait for upcoming tcp timer rewrite.
2023 * XXXGL: should there be used sbused() or sbavail()?
2025 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2028 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2029 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2030 sbused(&so->so_snd) >=
2031 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2032 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2033 sendwin >= (sbused(&so->so_snd) -
2034 (tp->snd_nxt - tp->snd_una))) {
2035 if (!sbreserve_locked(&so->so_snd,
2036 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2037 V_tcp_autosndbuf_max), so, curthread))
2038 so->so_snd.sb_flags &= ~SB_AUTOSIZE;