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_kern_tls.h"
42 #include <sys/param.h>
43 #include <sys/systm.h>
45 #include <sys/domain.h>
47 #include <sys/hhook.h>
49 #include <sys/kernel.h>
55 #include <sys/mutex.h>
56 #include <sys/protosw.h>
57 #include <sys/qmath.h>
59 #include <sys/socket.h>
60 #include <sys/socketvar.h>
61 #include <sys/sysctl.h>
62 #include <sys/stats.h>
65 #include <net/route.h>
66 #include <net/route/nhop.h>
69 #include <netinet/in.h>
70 #include <netinet/in_kdtrace.h>
71 #include <netinet/in_systm.h>
72 #include <netinet/ip.h>
73 #include <netinet/in_pcb.h>
74 #include <netinet/ip_var.h>
75 #include <netinet/ip_options.h>
77 #include <netinet6/in6_pcb.h>
78 #include <netinet/ip6.h>
79 #include <netinet6/ip6_var.h>
81 #include <netinet/tcp.h>
83 #include <netinet/tcp_fsm.h>
84 #include <netinet/tcp_seq.h>
85 #include <netinet/tcp_var.h>
86 #include <netinet/tcp_log_buf.h>
87 #include <netinet/tcp_syncache.h>
88 #include <netinet/tcp_timer.h>
89 #include <netinet/tcpip.h>
90 #include <netinet/cc/cc.h>
91 #include <netinet/tcp_fastopen.h>
93 #include <netinet/tcp_pcap.h>
96 #include <netinet/tcp_offload.h>
98 #include <netinet/tcp_ecn.h>
100 #include <netipsec/ipsec_support.h>
102 #include <netinet/udp.h>
103 #include <netinet/udp_var.h>
104 #include <machine/in_cksum.h>
106 #include <security/mac/mac_framework.h>
108 VNET_DEFINE(int, path_mtu_discovery) = 1;
109 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
110 &VNET_NAME(path_mtu_discovery), 1,
111 "Enable Path MTU Discovery");
113 VNET_DEFINE(int, tcp_do_tso) = 1;
114 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
115 &VNET_NAME(tcp_do_tso), 0,
116 "Enable TCP Segmentation Offload");
118 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
119 #define V_tcp_sendspace VNET(tcp_sendspace)
120 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
121 &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
123 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
124 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
125 &VNET_NAME(tcp_do_autosndbuf), 0,
126 "Enable automatic send buffer sizing");
128 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
129 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
130 &VNET_NAME(tcp_autosndbuf_inc), 0,
131 "Incrementor step size of automatic send buffer");
133 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
134 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
135 &VNET_NAME(tcp_autosndbuf_max), 0,
136 "Max size of automatic send buffer");
138 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
139 #define V_tcp_sendbuf_auto_lowat VNET(tcp_sendbuf_auto_lowat)
140 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
141 &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
142 "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
145 * Make sure that either retransmit or persist timer is set for SYN, FIN and
148 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags) \
149 KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
150 tcp_timer_active((tp), TT_REXMT) || \
151 tcp_timer_active((tp), TT_PERSIST), \
152 ("neither rexmt nor persist timer is set"))
156 * Wrapper for the TCP established output helper hook.
159 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
160 struct tcpopt *to, uint32_t len, int tso)
162 struct tcp_hhook_data hhook_data;
164 if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
168 hhook_data.len = len;
169 hhook_data.tso = tso;
171 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
178 * CC wrapper hook functions
181 cc_after_idle(struct tcpcb *tp)
183 INP_WLOCK_ASSERT(tptoinpcb(tp));
185 if (CC_ALGO(tp)->after_idle != NULL)
186 CC_ALGO(tp)->after_idle(&tp->t_ccv);
190 * Tcp output routine: figure out what should be sent and send it.
193 tcp_default_output(struct tcpcb *tp)
195 struct socket *so = tptosocket(tp);
196 struct inpcb *inp = tptoinpcb(tp);
198 uint32_t recwin, sendwin;
200 int off, error = 0; /* Keep compiler happy */
201 u_int if_hw_tsomaxsegcount = 0;
202 u_int if_hw_tsomaxsegsize = 0;
204 struct ip *ip = NULL;
206 u_char opt[TCP_MAXOLEN];
207 unsigned ipoptlen, optlen, hdrlen, ulen;
208 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
209 unsigned ipsec_optlen = 0;
211 int idle, sendalot, curticks;
212 int sack_rxmit, sack_bytes_rxmt;
216 struct udphdr *udp = NULL;
217 struct tcp_log_buffer *lgb;
218 unsigned int wanted_cookie = 0;
219 unsigned int dont_sendalot = 0;
221 int maxburst = TCP_MAXBURST;
224 struct ip6_hdr *ip6 = NULL;
225 const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
228 const bool hw_tls = tp->t_nic_ktls_xmit != 0;
230 const bool hw_tls = false;
234 INP_WLOCK_ASSERT(inp);
237 if (tp->t_flags & TF_TOE)
238 return (tcp_offload_output(tp));
242 * For TFO connections in SYN_SENT or SYN_RECEIVED,
243 * only allow the initial SYN or SYN|ACK and those sent
244 * by the retransmit timer.
246 if (IS_FASTOPEN(tp->t_flags) &&
247 ((tp->t_state == TCPS_SYN_SENT) ||
248 (tp->t_state == TCPS_SYN_RECEIVED)) &&
249 SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
250 (tp->snd_nxt != tp->snd_una)) /* not a retransmit */
254 * Determine length of data that should be transmitted,
255 * and flags that will be used.
256 * If there is some data or critical controls (SYN, RST)
257 * to send, then transmit; otherwise, investigate further.
259 idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
260 if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
261 (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
263 tp->t_flags &= ~TF_LASTIDLE;
265 if (tp->t_flags & TF_MORETOCOME) {
266 tp->t_flags |= TF_LASTIDLE;
272 * If we've recently taken a timeout, snd_max will be greater than
273 * snd_nxt. There may be SACK information that allows us to avoid
274 * resending already delivered data. Adjust snd_nxt accordingly.
276 if ((tp->t_flags & TF_SACK_PERMIT) &&
277 SEQ_LT(tp->snd_nxt, tp->snd_max))
282 off = tp->snd_nxt - tp->snd_una;
283 sendwin = min(tp->snd_wnd, tp->snd_cwnd);
285 flags = tcp_outflags[tp->t_state];
287 * Send any SACK-generated retransmissions. If we're explicitly trying
288 * to send out new data (when sendalot is 1), bypass this function.
289 * If we retransmit in fast recovery mode, decrement snd_cwnd, since
290 * we're replacing a (future) new transmission with a retransmission
291 * now, and we previously incremented snd_cwnd in tcp_input().
294 * Still in sack recovery , reset rxmit flag to zero.
300 if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
301 (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
305 imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
306 /* Do not retransmit SACK segments beyond snd_recover */
307 if (SEQ_GT(p->end, tp->snd_recover)) {
309 * (At least) part of sack hole extends beyond
310 * snd_recover. Check to see if we can rexmit data
313 if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
315 * Can't rexmit any more data for this hole.
316 * That data will be rexmitted in the next
317 * sack recovery episode, when snd_recover
318 * moves past p->rxmit.
321 goto after_sack_rexmit;
323 /* Can rexmit part of the current hole */
324 len = ((int32_t)ulmin(cwin,
325 SEQ_SUB(tp->snd_recover, p->rxmit)));
328 len = ((int32_t)ulmin(cwin,
329 SEQ_SUB(p->end, p->rxmit)));
332 off = SEQ_SUB(p->rxmit, tp->snd_una);
333 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
337 TCPSTAT_INC(tcps_sack_rexmits);
338 TCPSTAT_ADD(tcps_sack_rexmit_bytes,
339 min(len, tcp_maxseg(tp)));
344 * Get standard flags, and add SYN or FIN if requested by 'hidden'
347 if (tp->t_flags & TF_NEEDFIN)
349 if (tp->t_flags & TF_NEEDSYN)
352 SOCKBUF_LOCK(&so->so_snd);
354 * If in persist timeout with window of 0, send 1 byte.
355 * Otherwise, if window is small but nonzero
356 * and timer expired, we will send what we can
357 * and go to transmit state.
359 if (tp->t_flags & TF_FORCEDATA) {
362 * If we still have some data to send, then
363 * clear the FIN bit. Usually this would
364 * happen below when it realizes that we
365 * aren't sending all the data. However,
366 * if we have exactly 1 byte of unsent data,
367 * then it won't clear the FIN bit below,
368 * and if we are in persist state, we wind
369 * up sending the packet without recording
370 * that we sent the FIN bit.
372 * We can't just blindly clear the FIN bit,
373 * because if we don't have any more data
374 * to send then the probe will be the FIN
377 if (off < sbused(&so->so_snd))
381 tcp_timer_activate(tp, TT_PERSIST, 0);
387 * If snd_nxt == snd_max and we have transmitted a FIN, the
388 * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
389 * a negative length. This can also occur when TCP opens up
390 * its congestion window while receiving additional duplicate
391 * acks after fast-retransmit because TCP will reset snd_nxt
392 * to snd_max after the fast-retransmit.
394 * In the normal retransmit-FIN-only case, however, snd_nxt will
395 * be set to snd_una, the offset will be 0, and the length may
398 * If sack_rxmit is true we are retransmitting from the scoreboard
399 * in which case len is already set.
401 if (sack_rxmit == 0) {
402 if (sack_bytes_rxmt == 0)
403 len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
409 * We are inside of a SACK recovery episode and are
410 * sending new data, having retransmitted all the
411 * data possible in the scoreboard.
413 len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
416 * Don't remove this (len > 0) check !
417 * We explicitly check for len > 0 here (although it
418 * isn't really necessary), to work around a gcc
419 * optimization issue - to force gcc to compute
420 * len above. Without this check, the computation
421 * of len is bungled by the optimizer.
424 cwin = tp->snd_cwnd - imax(0, (int32_t)
425 (tp->snd_nxt - tp->snd_recover)) -
429 len = imin(len, cwin);
435 * Lop off SYN bit if it has already been sent. However, if this
436 * is SYN-SENT state and if segment contains data and if we don't
437 * know that foreign host supports TAO, suppress sending segment.
439 if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
440 if (tp->t_state != TCPS_SYN_RECEIVED)
443 * When sending additional segments following a TFO SYN|ACK,
444 * do not include the SYN bit.
446 if (IS_FASTOPEN(tp->t_flags) &&
447 (tp->t_state == TCPS_SYN_RECEIVED))
453 * Be careful not to send data and/or FIN on SYN segments.
454 * This measure is needed to prevent interoperability problems
455 * with not fully conformant TCP implementations.
457 if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
463 * On TFO sockets, ensure no data is sent in the following cases:
465 * - When retransmitting SYN|ACK on a passively-created socket
467 * - When retransmitting SYN on an actively created socket
469 * - When sending a zero-length cookie (cookie request) on an
470 * actively created socket
472 * - When the socket is in the CLOSED state (RST is being sent)
474 if (IS_FASTOPEN(tp->t_flags) &&
475 (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
476 ((tp->t_state == TCPS_SYN_SENT) &&
477 (tp->t_tfo_client_cookie_len == 0)) ||
482 * If FIN has been sent but not acked,
483 * but we haven't been called to retransmit,
484 * len will be < 0. Otherwise, window shrank
485 * after we sent into it. If window shrank to 0,
486 * cancel pending retransmit, pull snd_nxt back
487 * to (closed) window, and set the persist timer
488 * if it isn't already going. If the window didn't
489 * close completely, just wait for an ACK.
491 * We also do a general check here to ensure that
492 * we will set the persist timer when we have data
493 * to send, but a 0-byte window. This makes sure
494 * the persist timer is set even if the packet
495 * hits one of the "goto send" lines below.
498 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
499 (off < (int) sbavail(&so->so_snd)) &&
500 !tcp_timer_active(tp, TT_PERSIST)) {
501 tcp_timer_activate(tp, TT_REXMT, 0);
503 tp->snd_nxt = tp->snd_una;
504 if (!tcp_timer_active(tp, TT_PERSIST))
509 /* len will be >= 0 after this point. */
510 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
512 tcp_sndbuf_autoscale(tp, so, sendwin);
515 * Decide if we can use TCP Segmentation Offloading (if supported by
518 * TSO may only be used if we are in a pure bulk sending state. The
519 * presence of TCP-MD5, SACK retransmits, SACK advertizements and
520 * IP options prevent using TSO. With TSO the TCP header is the same
521 * (except for the sequence number) for all generated packets. This
522 * makes it impossible to transmit any options which vary per generated
525 * IPv4 handling has a clear separation of ip options and ip header
526 * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
527 * the right thing below to provide length of just ip options and thus
528 * checking for ipoptlen is enough to decide if ip options are present.
530 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
532 * Pre-calculate here as we save another lookup into the darknesses
533 * of IPsec that way and can actually decide if TSO is ok.
536 if (isipv6 && IPSEC_ENABLED(ipv6))
537 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
543 if (IPSEC_ENABLED(ipv4))
544 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
549 ipoptlen = ip6_optlen(inp);
552 if (inp->inp_options)
553 ipoptlen = inp->inp_options->m_len -
554 offsetof(struct ipoption, ipopt_list);
557 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
558 ipoptlen += ipsec_optlen;
561 if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
563 ((tp->t_flags & TF_SIGNATURE) == 0) &&
564 tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
565 ipoptlen == 0 && !(flags & TH_SYN))
569 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
572 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
573 sbused(&so->so_snd)))
577 recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
578 (long)TCP_MAXWIN << tp->rcv_scale);
581 * Sender silly window avoidance. We transmit under the following
582 * conditions when len is non-zero:
584 * - We have a full segment (or more with TSO)
585 * - This is the last buffer in a write()/send() and we are
586 * either idle or running NODELAY
587 * - we've timed out (e.g. persist timer)
588 * - we have more then 1/2 the maximum send window's worth of
589 * data (receiver may be limited the window size)
590 * - we need to retransmit
593 if (len >= tp->t_maxseg)
596 * As the TCP header options are now
597 * considered when setting up the initial
598 * window, we would not send the last segment
599 * if we skip considering the option length here.
600 * Note: this may not work when tcp headers change
601 * very dynamically in the future.
603 if ((((tp->t_flags & TF_SIGNATURE) ?
604 PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
605 ((tp->t_flags & TF_RCVD_TSTMP) ?
606 PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
607 len) >= tp->t_maxseg)
610 * NOTE! on localhost connections an 'ack' from the remote
611 * end may occur synchronously with the output and cause
612 * us to flush a buffer queued with moretocome. XXX
614 * note: the len + off check is almost certainly unnecessary.
616 if (!(tp->t_flags & TF_MORETOCOME) && /* normal case */
617 (idle || (tp->t_flags & TF_NODELAY)) &&
618 (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
619 (tp->t_flags & TF_NOPUSH) == 0) {
622 if (tp->t_flags & TF_FORCEDATA) /* typ. timeout case */
624 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
626 if (SEQ_LT(tp->snd_nxt, tp->snd_max)) /* retransmit case */
633 * Sending of standalone window updates.
635 * Window updates are important when we close our window due to a
636 * full socket buffer and are opening it again after the application
637 * reads data from it. Once the window has opened again and the
638 * remote end starts to send again the ACK clock takes over and
639 * provides the most current window information.
641 * We must avoid the silly window syndrome whereas every read
642 * from the receive buffer, no matter how small, causes a window
643 * update to be sent. We also should avoid sending a flurry of
644 * window updates when the socket buffer had queued a lot of data
645 * and the application is doing small reads.
647 * Prevent a flurry of pointless window updates by only sending
648 * an update when we can increase the advertized window by more
649 * than 1/4th of the socket buffer capacity. When the buffer is
650 * getting full or is very small be more aggressive and send an
651 * update whenever we can increase by two mss sized segments.
652 * In all other situations the ACK's to new incoming data will
653 * carry further window increases.
655 * Don't send an independent window update if a delayed
656 * ACK is pending (it will get piggy-backed on it) or the
657 * remote side already has done a half-close and won't send
658 * more data. Skip this if the connection is in T/TCP
661 if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
662 !(tp->t_flags & TF_DELACK) &&
663 !TCPS_HAVERCVDFIN(tp->t_state)) {
665 * "adv" is the amount we could increase the window,
666 * taking into account that we are limited by
667 * TCP_MAXWIN << tp->rcv_scale.
673 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
674 oldwin = (tp->rcv_adv - tp->rcv_nxt);
683 * If the new window size ends up being the same as or less
684 * than the old size when it is scaled, then don't force
687 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
690 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
691 (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
692 recwin <= (so->so_rcv.sb_hiwat / 8) ||
693 so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
694 adv >= TCP_MAXWIN << tp->rcv_scale))
696 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
702 * Send if we owe the peer an ACK, RST, SYN, or urgent data. ACKNOW
703 * is also a catch-all for the retransmit timer timeout case.
705 if (tp->t_flags & TF_ACKNOW)
707 if ((flags & TH_RST) ||
708 ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
710 if (SEQ_GT(tp->snd_up, tp->snd_una))
713 * If our state indicates that FIN should be sent
714 * and we have not yet done so, then we need to send.
716 if (flags & TH_FIN &&
717 ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
720 * In SACK, it is possible for tcp_output to fail to send a segment
721 * after the retransmission timer has been turned off. Make sure
722 * that the retransmission timer is set.
724 if ((tp->t_flags & TF_SACK_PERMIT) &&
725 SEQ_GT(tp->snd_max, tp->snd_una) &&
726 !tcp_timer_active(tp, TT_REXMT) &&
727 !tcp_timer_active(tp, TT_PERSIST)) {
728 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
732 * TCP window updates are not reliable, rather a polling protocol
733 * using ``persist'' packets is used to insure receipt of window
734 * updates. The three ``states'' for the output side are:
735 * idle not doing retransmits or persists
736 * persisting to move a small or zero window
737 * (re)transmitting and thereby not persisting
739 * tcp_timer_active(tp, TT_PERSIST)
740 * is true when we are in persist state.
741 * (tp->t_flags & TF_FORCEDATA)
742 * is set when we are called to send a persist packet.
743 * tcp_timer_active(tp, TT_REXMT)
744 * is set when we are retransmitting
745 * The output side is idle when both timers are zero.
747 * If send window is too small, there is data to transmit, and no
748 * retransmit or persist is pending, then go to persist state.
749 * If nothing happens soon, send when timer expires:
750 * if window is nonzero, transmit what we can,
751 * otherwise force out a byte.
753 if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
754 !tcp_timer_active(tp, TT_PERSIST)) {
760 * No reason to send a segment, just return.
763 SOCKBUF_UNLOCK(&so->so_snd);
767 SOCKBUF_LOCK_ASSERT(&so->so_snd);
769 if (len >= tp->t_maxseg)
770 tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
772 tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
775 * Before ESTABLISHED, force sending of initial options
776 * unless TCP set not to do any options.
777 * NOTE: we assume that the IP/TCP header plus TCP options
778 * always fit in a single mbuf, leaving room for a maximum
780 * max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
785 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
788 hdrlen = sizeof (struct tcpiphdr);
790 if (flags & TH_SYN) {
791 tp->snd_nxt = tp->iss;
795 * Compute options for segment.
796 * We only have to care about SYN and established connection
797 * segments. Options for SYN-ACK segments are handled in TCP
801 if ((tp->t_flags & TF_NOOPT) == 0) {
802 /* Maximum segment size. */
803 if (flags & TH_SYN) {
804 to.to_mss = tcp_mssopt(&inp->inp_inc);
806 to.to_mss -= V_tcp_udp_tunneling_overhead;
807 to.to_flags |= TOF_MSS;
810 * On SYN or SYN|ACK transmits on TFO connections,
811 * only include the TFO option if it is not a
812 * retransmit, as the presence of the TFO option may
813 * have caused the original SYN or SYN|ACK to have
814 * been dropped by a middlebox.
816 if (IS_FASTOPEN(tp->t_flags) &&
817 (tp->t_rxtshift == 0)) {
818 if (tp->t_state == TCPS_SYN_RECEIVED) {
819 to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
821 (u_int8_t *)&tp->t_tfo_cookie.server;
822 to.to_flags |= TOF_FASTOPEN;
824 } else if (tp->t_state == TCPS_SYN_SENT) {
826 tp->t_tfo_client_cookie_len;
828 tp->t_tfo_cookie.client;
829 to.to_flags |= TOF_FASTOPEN;
832 * If we wind up having more data to
833 * send with the SYN than can fit in
834 * one segment, don't send any more
835 * until the SYN|ACK comes back from
842 /* Window scaling. */
843 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
844 to.to_wscale = tp->request_r_scale;
845 to.to_flags |= TOF_SCALE;
848 if ((tp->t_flags & TF_RCVD_TSTMP) ||
849 ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
850 curticks = tcp_ts_getticks();
851 to.to_tsval = curticks + tp->ts_offset;
852 to.to_tsecr = tp->ts_recent;
853 to.to_flags |= TOF_TS;
854 if (tp->t_rxtshift == 1)
855 tp->t_badrxtwin = curticks;
858 /* Set receive buffer autosizing timestamp. */
859 if (tp->rfbuf_ts == 0 &&
860 (so->so_rcv.sb_flags & SB_AUTOSIZE))
861 tp->rfbuf_ts = tcp_ts_getticks();
863 /* Selective ACK's. */
864 if (tp->t_flags & TF_SACK_PERMIT) {
866 to.to_flags |= TOF_SACKPERM;
867 else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
868 tp->rcv_numsacks > 0) {
869 to.to_flags |= TOF_SACK;
870 to.to_nsacks = tp->rcv_numsacks;
871 to.to_sacks = (u_char *)tp->sackblks;
874 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
875 /* TCP-MD5 (RFC2385). */
877 * Check that TCP_MD5SIG is enabled in tcpcb to
878 * account the size needed to set this TCP option.
880 if (tp->t_flags & TF_SIGNATURE)
881 to.to_flags |= TOF_SIGNATURE;
882 #endif /* TCP_SIGNATURE */
884 /* Processing the options. */
885 hdrlen += optlen = tcp_addoptions(&to, opt);
887 * If we wanted a TFO option to be added, but it was unable
888 * to fit, ensure no data is sent.
890 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
891 !(to.to_flags & TOF_FASTOPEN))
895 if (V_tcp_udp_tunneling_port == 0) {
896 /* The port was removed?? */
897 SOCKBUF_UNLOCK(&so->so_snd);
898 return (EHOSTUNREACH);
900 hdrlen += sizeof(struct udphdr);
903 * Adjust data length if insertion of options will
904 * bump the packet length beyond the t_maxseg length.
905 * Clear the FIN bit because we cut off the tail of
908 if (len + optlen + ipoptlen > tp->t_maxseg) {
916 /* extract TSO information */
917 if_hw_tsomax = tp->t_tsomax;
918 if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
919 if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
922 * Limit a TSO burst to prevent it from
923 * overflowing or exceeding the maximum length
924 * allowed by the network interface:
926 KASSERT(ipoptlen == 0,
927 ("%s: TSO can't do IP options", __func__));
930 * Check if we should limit by maximum payload
933 if (if_hw_tsomax != 0) {
934 /* compute maximum TSO length */
935 max_len = (if_hw_tsomax - hdrlen -
939 } else if (len > max_len) {
946 * Prevent the last segment from being
947 * fractional unless the send sockbuf can be
950 max_len = (tp->t_maxseg - optlen);
951 if (((uint32_t)off + (uint32_t)len) <
952 sbavail(&so->so_snd)) {
953 moff = len % max_len;
961 * In case there are too many small fragments
964 if (len <= max_len) {
971 * Send the FIN in a separate segment
972 * after the bulk sending is done.
973 * We don't trust the TSO implementations
974 * to clear the FIN flag on all but the
977 if (tp->t_flags & TF_NEEDFIN)
980 if (optlen + ipoptlen >= tp->t_maxseg) {
982 * Since we don't have enough space to put
983 * the IP header chain and the TCP header in
984 * one packet as required by RFC 7112, don't
985 * send it. Also ensure that at least one
986 * byte of the payload can be put into the
989 SOCKBUF_UNLOCK(&so->so_snd);
994 len = tp->t_maxseg - optlen - ipoptlen;
1002 KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1003 ("%s: len > IP_MAXPACKET", __func__));
1005 /*#ifdef DIAGNOSTIC*/
1007 if (max_linkhdr + hdrlen > MCLBYTES)
1009 if (max_linkhdr + hdrlen > MHLEN)
1011 panic("tcphdr too big");
1015 * This KASSERT is here to catch edge cases at a well defined place.
1016 * Before, those had triggered (random) panic conditions further down.
1018 KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1021 * Grab a header mbuf, attaching a copy of data to
1022 * be transmitted, and initialize the header from
1023 * the template for sends on this connection.
1027 struct sockbuf *msb;
1030 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1031 TCPSTAT_INC(tcps_sndprobe);
1033 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1034 stats_voi_update_abs_u32(tp->t_stats,
1035 VOI_TCP_RETXPB, len);
1037 stats_voi_update_abs_u64(tp->t_stats,
1040 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1041 tp->t_sndrexmitpack++;
1042 TCPSTAT_INC(tcps_sndrexmitpack);
1043 TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1045 stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1049 TCPSTAT_INC(tcps_sndpack);
1050 TCPSTAT_ADD(tcps_sndbyte, len);
1052 stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1057 if (MHLEN < hdrlen + max_linkhdr)
1058 m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1061 m = m_gethdr(M_NOWAIT, MT_DATA);
1064 SOCKBUF_UNLOCK(&so->so_snd);
1070 m->m_data += max_linkhdr;
1074 * Start the m_copy functions from the closest mbuf
1075 * to the offset in the socket buffer chain.
1077 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1078 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1079 m_copydata(mb, moff, len,
1080 mtod(m, caddr_t) + hdrlen);
1081 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1082 sbsndptr_adv(&so->so_snd, mb, len);
1085 if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1089 m->m_next = tcp_m_copym(mb, moff,
1090 &len, if_hw_tsomaxsegcount,
1091 if_hw_tsomaxsegsize, msb, hw_tls);
1092 if (len <= (tp->t_maxseg - optlen)) {
1094 * Must have ran out of mbufs for the copy
1095 * shorten it to no longer need tso. Lets
1096 * not put on sendalot since we are low on
1101 if (m->m_next == NULL) {
1102 SOCKBUF_UNLOCK(&so->so_snd);
1111 * If we're sending everything we've got, set PUSH.
1112 * (This will keep happy those implementations which only
1113 * give data to the user when a buffer fills or
1116 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1119 SOCKBUF_UNLOCK(&so->so_snd);
1121 SOCKBUF_UNLOCK(&so->so_snd);
1122 if (tp->t_flags & TF_ACKNOW)
1123 TCPSTAT_INC(tcps_sndacks);
1124 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1125 TCPSTAT_INC(tcps_sndctrl);
1126 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1127 TCPSTAT_INC(tcps_sndurg);
1129 TCPSTAT_INC(tcps_sndwinup);
1131 m = m_gethdr(M_NOWAIT, MT_DATA);
1138 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1143 m->m_data += max_linkhdr;
1146 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1147 m->m_pkthdr.rcvif = (struct ifnet *)0;
1149 mac_inpcb_create_mbuf(inp, m);
1153 ip6 = mtod(m, struct ip6_hdr *);
1155 udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1156 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1157 udp->uh_dport = tp->t_port;
1158 ulen = hdrlen + len - sizeof(struct ip6_hdr);
1159 udp->uh_ulen = htons(ulen);
1160 th = (struct tcphdr *)(udp + 1);
1162 th = (struct tcphdr *)(ip6 + 1);
1164 tcpip_fillheaders(inp, tp->t_port, ip6, th);
1168 ip = mtod(m, struct ip *);
1170 udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1171 udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1172 udp->uh_dport = tp->t_port;
1173 ulen = hdrlen + len - sizeof(struct ip);
1174 udp->uh_ulen = htons(ulen);
1175 th = (struct tcphdr *)(udp + 1);
1177 th = (struct tcphdr *)(ip + 1);
1178 tcpip_fillheaders(inp, tp->t_port, ip, th);
1182 * Fill in fields, remembering maximum advertised
1183 * window for use in delaying messages about window sizes.
1184 * If resending a FIN, be sure not to use a new sequence number.
1186 if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1187 tp->snd_nxt == tp->snd_max)
1190 * If we are starting a connection, send ECN setup
1191 * SYN packet. If we are on a retransmit, we may
1192 * resend those bits a number of times as per
1195 if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1196 flags |= tcp_ecn_output_syn_sent(tp);
1198 /* Also handle parallel SYN for ECN */
1199 if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1200 (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1201 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1202 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1203 (tp->t_flags2 & TF2_ECN_SND_ECE))
1204 tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1207 ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1208 ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1213 ip->ip_tos &= ~IPTOS_ECN_MASK;
1219 * If we are doing retransmissions, then snd_nxt will
1220 * not reflect the first unsent octet. For ACK only
1221 * packets, we do not want the sequence number of the
1222 * retransmitted packet, we want the sequence number
1223 * of the next unsent octet. So, if there is no data
1224 * (and no SYN or FIN), use snd_max instead of snd_nxt
1225 * when filling in ti_seq. But if we are in persist
1226 * state, snd_max might reflect one byte beyond the
1227 * right edge of the window, so use snd_nxt in that
1228 * case, since we know we aren't doing a retransmission.
1229 * (retransmit and persist are mutually exclusive...)
1231 if (sack_rxmit == 0) {
1232 if (len || (flags & (TH_SYN|TH_FIN)) ||
1233 tcp_timer_active(tp, TT_PERSIST))
1234 th->th_seq = htonl(tp->snd_nxt);
1236 th->th_seq = htonl(tp->snd_max);
1238 th->th_seq = htonl(p->rxmit);
1241 * Lost Retransmission Detection
1242 * trigger resending of a (then
1243 * still existing) hole, when
1244 * fack acks recoverypoint.
1246 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1247 p->rxmit = tp->snd_recover;
1248 tp->sackhint.sack_bytes_rexmit += len;
1250 if (IN_RECOVERY(tp->t_flags)) {
1252 * Account all bytes transmitted while
1253 * IN_RECOVERY, simplifying PRR and
1254 * Lost Retransmit Detection
1256 tp->sackhint.prr_out += len;
1258 th->th_ack = htonl(tp->rcv_nxt);
1260 bcopy(opt, th + 1, optlen);
1261 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1263 tcp_set_flags(th, flags);
1265 * Calculate receive window. Don't shrink window,
1266 * but avoid silly window syndrome.
1267 * If a RST segment is sent, advertise a window of zero.
1269 if (flags & TH_RST) {
1272 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1273 recwin < tp->t_maxseg)
1275 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1276 recwin < (tp->rcv_adv - tp->rcv_nxt))
1277 recwin = (tp->rcv_adv - tp->rcv_nxt);
1280 * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1281 * or <SYN,ACK>) segment itself is never scaled. The <SYN,ACK>
1282 * case is handled in syncache.
1285 th->th_win = htons((u_short)
1286 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1288 /* Avoid shrinking window with window scaling. */
1289 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1290 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1294 * Adjust the RXWIN0SENT flag - indicate that we have advertised
1295 * a 0 window. This may cause the remote transmitter to stall. This
1296 * flag tells soreceive() to disable delayed acknowledgements when
1297 * draining the buffer. This can occur if the receiver is attempting
1298 * to read more data than can be buffered prior to transmitting on
1301 if (th->th_win == 0) {
1303 tp->t_flags |= TF_RXWIN0SENT;
1305 tp->t_flags &= ~TF_RXWIN0SENT;
1306 if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1307 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1308 th->th_flags |= TH_URG;
1311 * If no urgent pointer to send, then we pull
1312 * the urgent pointer to the left edge of the send window
1313 * so that it doesn't drift into the send window on sequence
1314 * number wraparound.
1316 tp->snd_up = tp->snd_una; /* drag it along */
1319 * Put TCP length in extended header, and then
1320 * checksum extended header and data.
1322 m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1324 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1325 if (to.to_flags & TOF_SIGNATURE) {
1327 * Calculate MD5 signature and put it into the place
1328 * determined before.
1329 * NOTE: since TCP options buffer doesn't point into
1330 * mbuf's data, calculate offset and use it.
1332 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1333 (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1335 * Do not send segment if the calculation of MD5
1336 * digest has failed.
1346 * There is no need to fill in ip6_plen right now.
1347 * It will be filled later by ip6_output.
1350 m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1351 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1352 udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1353 th->th_sum = htons(0);
1354 UDPSTAT_INC(udps_opackets);
1356 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1357 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1358 th->th_sum = in6_cksum_pseudo(ip6,
1359 sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1364 #if defined(INET6) && defined(INET)
1370 m->m_pkthdr.csum_flags = CSUM_UDP;
1371 m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1372 udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1373 ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1374 th->th_sum = htons(0);
1375 UDPSTAT_INC(udps_opackets);
1377 m->m_pkthdr.csum_flags = CSUM_TCP;
1378 m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1379 th->th_sum = in_pseudo(ip->ip_src.s_addr,
1380 ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1381 IPPROTO_TCP + len + optlen));
1384 /* IP version must be set here for ipv4/ipv6 checking later */
1385 KASSERT(ip->ip_v == IPVERSION,
1386 ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1391 * Enable TSO and specify the size of the segments.
1392 * The TCP pseudo header checksum is always provided.
1395 KASSERT(len > tp->t_maxseg - optlen,
1396 ("%s: len <= tso_segsz", __func__));
1397 m->m_pkthdr.csum_flags |= CSUM_TSO;
1398 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1401 KASSERT(len + hdrlen == m_length(m, NULL),
1402 ("%s: mbuf chain shorter than expected: %d + %u != %u",
1403 __func__, len, hdrlen, m_length(m, NULL)));
1406 /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1407 hhook_run_tcp_est_out(tp, th, &to, len, tso);
1410 TCP_PROBE3(debug__output, tp, th, m);
1412 /* We're getting ready to send; log now. */
1413 /* XXXMT: We are not honoring verbose logging. */
1415 if (tcp_bblogging_on(tp))
1416 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1417 TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1423 * Fill in IP length and desired time to live and
1424 * send to IP level. There should be a better way
1425 * to handle ttl and tos; we could keep them in
1426 * the template, but need a way to checksum without them.
1429 * m->m_pkthdr.len should have been set before checksum calculation,
1430 * because in6_cksum() need it.
1435 * we separately set hoplimit for every segment, since the
1436 * user might want to change the value via setsockopt.
1437 * Also, desired default hop limit might be changed via
1438 * Neighbor Discovery.
1440 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1443 * Set the packet size here for the benefit of DTrace probes.
1444 * ip6_output() will set it properly; it's supposed to include
1445 * the option header lengths as well.
1447 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1449 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1450 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1452 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1454 if (tp->t_state == TCPS_SYN_SENT)
1455 TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1457 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1460 /* Save packet, if requested. */
1461 tcp_pcap_add(th, m, &(tp->t_outpkts));
1464 /* TODO: IPv6 IP6TOS_ECT bit on */
1465 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1466 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0),
1469 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1470 mtu = inp->inp_route6.ro_nh->nh_mtu;
1473 #if defined(INET) && defined(INET6)
1478 ip->ip_len = htons(m->m_pkthdr.len);
1480 if (inp->inp_vflag & INP_IPV6PROTO)
1481 ip->ip_ttl = in6_selecthlim(inp, NULL);
1484 * If we do path MTU discovery, then we set DF on every packet.
1485 * This might not be the best thing to do according to RFC3390
1486 * Section 2. However the tcp hostcache migitates the problem
1487 * so it affects only the first tcp connection with a host.
1489 * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1491 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1492 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1493 if (tp->t_port == 0 || len < V_tcp_minmss) {
1494 ip->ip_off |= htons(IP_DF);
1497 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1500 if (tp->t_state == TCPS_SYN_SENT)
1501 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1503 TCP_PROBE5(send, NULL, tp, ip, tp, th);
1506 /* Save packet, if requested. */
1507 tcp_pcap_add(th, m, &(tp->t_outpkts));
1510 error = ip_output(m, inp->inp_options, &inp->inp_route,
1511 ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1513 if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1514 mtu = inp->inp_route.ro_nh->nh_mtu;
1519 lgb->tlb_errno = error;
1524 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1526 * In transmit state, time the transmission and arrange for
1527 * the retransmit. In persist state, just set snd_max.
1529 if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1530 !tcp_timer_active(tp, TT_PERSIST)) {
1531 tcp_seq startseq = tp->snd_nxt;
1534 * Advance snd_nxt over sequence space of this segment.
1536 if (flags & (TH_SYN|TH_FIN)) {
1539 if (flags & TH_FIN) {
1541 tp->t_flags |= TF_SENTFIN;
1547 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1549 * Update "made progress" indication if we just
1550 * added new data to an empty socket buffer.
1552 if (tp->snd_una == tp->snd_max)
1553 tp->t_acktime = ticks;
1554 tp->snd_max = tp->snd_nxt;
1556 * Time this transmission if not a retransmission and
1557 * not currently timing anything.
1559 tp->t_sndtime = ticks;
1560 if (tp->t_rtttime == 0) {
1561 tp->t_rtttime = ticks;
1562 tp->t_rtseq = startseq;
1563 TCPSTAT_INC(tcps_segstimed);
1566 if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1567 tp->t_flags |= TF_GPUTINPROG;
1568 tp->gput_seq = startseq;
1569 tp->gput_ack = startseq +
1570 ulmin(sbavail(&so->so_snd) - off, sendwin);
1571 tp->gput_ts = tcp_ts_getticks();
1577 * Set retransmit timer if not currently set,
1578 * and not doing a pure ack or a keep-alive probe.
1579 * Initial value for retransmit timer is smoothed
1580 * round-trip time + 2 * round-trip time variance.
1581 * Initialize shift counter which is used for backoff
1582 * of retransmit time.
1585 if (!tcp_timer_active(tp, TT_REXMT) &&
1586 ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1587 (tp->snd_nxt != tp->snd_una))) {
1588 if (tcp_timer_active(tp, TT_PERSIST)) {
1589 tcp_timer_activate(tp, TT_PERSIST, 0);
1592 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1593 } else if (len == 0 && sbavail(&so->so_snd) &&
1594 !tcp_timer_active(tp, TT_REXMT) &&
1595 !tcp_timer_active(tp, TT_PERSIST)) {
1597 * Avoid a situation where we do not set persist timer
1598 * after a zero window condition. For example:
1599 * 1) A -> B: packet with enough data to fill the window
1600 * 2) B -> A: ACK for #1 + new data (0 window
1602 * 3) A -> B: ACK for #2, 0 len packet
1604 * In this case, A will not activate the persist timer,
1605 * because it chose to send a packet. Unless tcp_output
1606 * is called for some other reason (delayed ack timer,
1607 * another input packet from B, socket syscall), A will
1608 * not send zero window probes.
1610 * So, if you send a 0-length packet, but there is data
1611 * in the socket buffer, and neither the rexmt or
1612 * persist timer is already set, then activate the
1620 * Persist case, update snd_max but since we are in
1621 * persist mode (no window) we do not update snd_nxt.
1626 if (flags & TH_FIN) {
1628 tp->t_flags |= TF_SENTFIN;
1630 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1631 tp->snd_max = tp->snd_nxt + xlen;
1634 (TCPS_HAVEESTABLISHED(tp->t_state) &&
1635 (tp->t_flags & TF_SACK_PERMIT) &&
1636 tp->rcv_numsacks > 0)) {
1637 /* Clean up any DSACK's sent */
1638 tcp_clean_dsack_blocks(tp);
1642 * We know that the packet was lost, so back out the
1643 * sequence number advance, if any.
1645 * If the error is EPERM the packet got blocked by the
1646 * local firewall. Normally we should terminate the
1647 * connection but the blocking may have been spurious
1648 * due to a firewall reconfiguration cycle. So we treat
1649 * it like a packet loss and let the retransmit timer and
1650 * timeouts do their work over time.
1651 * XXX: It is a POLA question whether calling tcp_drop right
1652 * away would be the really correct behavior instead.
1654 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1655 !tcp_timer_active(tp, TT_PERSIST)) &&
1656 ((flags & TH_SYN) == 0) &&
1660 tp->sackhint.sack_bytes_rexmit -= len;
1661 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1662 ("sackhint bytes rtx >= 0"));
1663 KASSERT((flags & TH_FIN) == 0,
1664 ("error while FIN with SACK rxmit"));
1671 SOCKBUF_UNLOCK_ASSERT(&so->so_snd); /* Check gotos. */
1675 tp->t_softerror = error;
1678 TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1679 tp->snd_cwnd = tp->t_maxseg;
1683 * For some reason the interface we used initially
1684 * to send segments changed to another or lowered
1686 * If TSO was active we either got an interface
1687 * without TSO capabilits or TSO was turned off.
1688 * If we obtained mtu from ip_output() then update
1692 tp->t_flags &= ~TF_TSO;
1694 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1702 if (TCPS_HAVERCVDSYN(tp->t_state)) {
1703 tp->t_softerror = error;
1711 TCPSTAT_INC(tcps_sndtotal);
1714 * Data sent (as far as we can tell).
1715 * If this advertises a larger window than any other segment,
1716 * then remember the size of the advertised window.
1717 * Any pending ACK has now been sent.
1719 if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1720 tp->rcv_adv = tp->rcv_nxt + recwin;
1721 tp->last_ack_sent = tp->rcv_nxt;
1722 tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1723 if (tcp_timer_active(tp, TT_DELACK))
1724 tcp_timer_activate(tp, TT_DELACK, 0);
1727 * This completely breaks TCP if newreno is turned on. What happens
1728 * is that if delayed-acks are turned on on the receiver, this code
1729 * on the transmitter effectively destroys the TCP window, forcing
1730 * it to four packets (1.5Kx4 = 6K window).
1732 if (sendalot && --maxburst)
1741 tcp_setpersist(struct tcpcb *tp)
1743 int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1747 tp->t_flags &= ~TF_PREVVALID;
1748 if (tcp_timer_active(tp, TT_REXMT))
1749 panic("tcp_setpersist: retransmit pending");
1751 * If the state is already closed, don't bother.
1753 if (tp->t_state == TCPS_CLOSED)
1757 * Start/restart persistence timer.
1759 TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1760 tcp_persmin, tcp_persmax);
1761 if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1762 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1763 if (maxunacktime < 1)
1765 if (maxunacktime < tt)
1768 tcp_timer_activate(tp, TT_PERSIST, tt);
1769 if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1774 * Insert TCP options according to the supplied parameters to the place
1775 * optp in a consistent way. Can handle unaligned destinations.
1777 * The order of the option processing is crucial for optimal packing and
1778 * alignment for the scarce option space.
1780 * The optimal order for a SYN/SYN-ACK segment is:
1781 * MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1782 * Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1784 * The SACK options should be last. SACK blocks consume 8*n+2 bytes.
1785 * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1786 * At minimum we need 10 bytes (to generate 1 SACK block). If both
1787 * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1788 * we only have 10 bytes for SACK options (40 - (12 + 18)).
1791 tcp_addoptions(struct tcpopt *to, u_char *optp)
1793 u_int32_t mask, optlen = 0;
1795 for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1796 if ((to->to_flags & mask) != mask)
1798 if (optlen == TCP_MAXOLEN)
1800 switch (to->to_flags & mask) {
1802 while (optlen % 4) {
1803 optlen += TCPOLEN_NOP;
1804 *optp++ = TCPOPT_NOP;
1806 if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1808 optlen += TCPOLEN_MAXSEG;
1809 *optp++ = TCPOPT_MAXSEG;
1810 *optp++ = TCPOLEN_MAXSEG;
1811 to->to_mss = htons(to->to_mss);
1812 bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1813 optp += sizeof(to->to_mss);
1816 while (!optlen || optlen % 2 != 1) {
1817 optlen += TCPOLEN_NOP;
1818 *optp++ = TCPOPT_NOP;
1820 if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1822 optlen += TCPOLEN_WINDOW;
1823 *optp++ = TCPOPT_WINDOW;
1824 *optp++ = TCPOLEN_WINDOW;
1825 *optp++ = to->to_wscale;
1828 while (optlen % 2) {
1829 optlen += TCPOLEN_NOP;
1830 *optp++ = TCPOPT_NOP;
1832 if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1834 optlen += TCPOLEN_SACK_PERMITTED;
1835 *optp++ = TCPOPT_SACK_PERMITTED;
1836 *optp++ = TCPOLEN_SACK_PERMITTED;
1839 while (!optlen || optlen % 4 != 2) {
1840 optlen += TCPOLEN_NOP;
1841 *optp++ = TCPOPT_NOP;
1843 if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1845 optlen += TCPOLEN_TIMESTAMP;
1846 *optp++ = TCPOPT_TIMESTAMP;
1847 *optp++ = TCPOLEN_TIMESTAMP;
1848 to->to_tsval = htonl(to->to_tsval);
1849 to->to_tsecr = htonl(to->to_tsecr);
1850 bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1851 optp += sizeof(to->to_tsval);
1852 bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1853 optp += sizeof(to->to_tsecr);
1857 int siglen = TCPOLEN_SIGNATURE - 2;
1859 while (!optlen || optlen % 4 != 2) {
1860 optlen += TCPOLEN_NOP;
1861 *optp++ = TCPOPT_NOP;
1863 if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1864 to->to_flags &= ~TOF_SIGNATURE;
1867 optlen += TCPOLEN_SIGNATURE;
1868 *optp++ = TCPOPT_SIGNATURE;
1869 *optp++ = TCPOLEN_SIGNATURE;
1870 to->to_signature = optp;
1878 struct sackblk *sack = (struct sackblk *)to->to_sacks;
1881 while (!optlen || optlen % 4 != 2) {
1882 optlen += TCPOLEN_NOP;
1883 *optp++ = TCPOPT_NOP;
1885 if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1887 optlen += TCPOLEN_SACKHDR;
1888 *optp++ = TCPOPT_SACK;
1889 sackblks = min(to->to_nsacks,
1890 (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1891 *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1892 while (sackblks--) {
1893 sack_seq = htonl(sack->start);
1894 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1895 optp += sizeof(sack_seq);
1896 sack_seq = htonl(sack->end);
1897 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1898 optp += sizeof(sack_seq);
1899 optlen += TCPOLEN_SACK;
1902 TCPSTAT_INC(tcps_sack_send_blocks);
1909 /* XXX is there any point to aligning this option? */
1910 total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1911 if (TCP_MAXOLEN - optlen < total_len) {
1912 to->to_flags &= ~TOF_FASTOPEN;
1915 *optp++ = TCPOPT_FAST_OPEN;
1916 *optp++ = total_len;
1917 if (to->to_tfo_len > 0) {
1918 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1919 optp += to->to_tfo_len;
1921 optlen += total_len;
1925 panic("%s: unknown TCP option type", __func__);
1930 /* Terminate and pad TCP options to a 4 byte boundary. */
1932 optlen += TCPOLEN_EOL;
1933 *optp++ = TCPOPT_EOL;
1936 * According to RFC 793 (STD0007):
1937 * "The content of the header beyond the End-of-Option option
1938 * must be header padding (i.e., zero)."
1939 * and later: "The padding is composed of zeros."
1941 while (optlen % 4) {
1942 optlen += TCPOLEN_PAD;
1943 *optp++ = TCPOPT_PAD;
1946 KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1951 * This is a copy of m_copym(), taking the TSO segment size/limit
1952 * constraints into account, and advancing the sndptr as it goes.
1955 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1956 int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1959 struct ktls_session *tls, *ntls;
1960 struct mbuf *start __diagused;
1962 struct mbuf *n, **np;
1965 int32_t len = *plen;
1969 uint32_t mlen, frags;
1972 KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1973 KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1974 if (off == 0 && m->m_flags & M_PKTHDR)
1979 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1983 if ((sb) && (m == sb->sb_sndptr)) {
1984 sb->sb_sndptroff += m->m_len;
1985 sb->sb_sndptr = m->m_next;
1993 if (hw_tls && (m->m_flags & M_EXTPG))
2001 KASSERT(len == M_COPYALL,
2002 ("tcp_m_copym, length > size of mbuf chain"));
2004 if (pkthdrlen != NULL)
2005 *pkthdrlen = len_cp;
2010 if (m->m_flags & M_EXTPG)
2011 ntls = m->m_epg_tls;
2016 * Avoid mixing TLS records with handshake
2017 * data or TLS records from different
2023 if (pkthdrlen != NULL)
2024 *pkthdrlen = len_cp;
2029 mlen = min(len, m->m_len - off);
2032 * For M_EXTPG mbufs, add 3 segments
2033 * + 1 in case we are crossing page boundaries
2034 * + 2 in case the TLS hdr/trailer are used
2035 * It is cheaper to just add the segments
2036 * than it is to take the cache miss to look
2037 * at the mbuf ext_pgs state in detail.
2039 if (m->m_flags & M_EXTPG) {
2040 fragsize = min(segsize, PAGE_SIZE);
2047 /* Break if we really can't fit anymore. */
2048 if ((frags + 1) >= seglimit) {
2050 if (pkthdrlen != NULL)
2051 *pkthdrlen = len_cp;
2056 * Reduce size if you can't copy the whole
2057 * mbuf. If we can't copy the whole mbuf, also
2058 * adjust len so the loop will end after this
2061 if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2062 mlen = (seglimit - frags - 1) * fragsize;
2064 *plen = len_cp + len;
2065 if (pkthdrlen != NULL)
2068 frags += howmany(mlen, fragsize);
2072 KASSERT(seglimit > 0,
2073 ("%s: seglimit went too low", __func__));
2076 n = m_gethdr(M_NOWAIT, m->m_type);
2078 n = m_get(M_NOWAIT, m->m_type);
2083 if (!m_dup_pkthdr(n, m, M_NOWAIT))
2085 if (len == M_COPYALL)
2086 n->m_pkthdr.len -= off0;
2088 n->m_pkthdr.len = len;
2089 pkthdrlen = &n->m_pkthdr.len;
2094 if (m->m_flags & (M_EXT|M_EXTPG)) {
2095 n->m_data = m->m_data + off;
2098 bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2101 if (sb && (sb->sb_sndptr == m) &&
2102 ((n->m_len + off) >= m->m_len) && m->m_next) {
2103 sb->sb_sndptroff += m->m_len;
2104 sb->sb_sndptr = m->m_next;
2107 if (len != M_COPYALL) {
2120 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2124 * Automatic sizing of send socket buffer. Often the send buffer
2125 * size is not optimally adjusted to the actual network conditions
2126 * at hand (delay bandwidth product). Setting the buffer size too
2127 * small limits throughput on links with high bandwidth and high
2128 * delay (eg. trans-continental/oceanic links). Setting the
2129 * buffer size too big consumes too much real kernel memory,
2130 * especially with many connections on busy servers.
2132 * The criteria to step up the send buffer one notch are:
2133 * 1. receive window of remote host is larger than send buffer
2134 * (with a fudge factor of 5/4th);
2135 * 2. send buffer is filled to 7/8th with data (so we actually
2136 * have data to make use of it);
2137 * 3. send buffer fill has not hit maximal automatic size;
2138 * 4. our send window (slow start and cogestion controlled) is
2139 * larger than sent but unacknowledged data in send buffer.
2141 * The remote host receive window scaling factor may limit the
2142 * growing of the send buffer before it reaches its allowed
2145 * It scales directly with slow start or congestion window
2146 * and does at most one step per received ACK. This fast
2147 * scaling has the drawback of growing the send buffer beyond
2148 * what is strictly necessary to make full use of a given
2149 * delay*bandwidth product. However testing has shown this not
2150 * to be much of an problem. At worst we are trading wasting
2151 * of available bandwidth (the non-use of it) for wasting some
2152 * socket buffer memory.
2154 * TODO: Shrink send buffer during idle periods together
2155 * with congestion window. Requires another timer. Has to
2156 * wait for upcoming tcp timer rewrite.
2158 * XXXGL: should there be used sbused() or sbavail()?
2160 if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2163 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2164 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2165 sbused(&so->so_snd) >=
2166 (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2167 sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2168 sendwin >= (sbused(&so->so_snd) -
2169 (tp->snd_nxt - tp->snd_una))) {
2170 if (!sbreserve_locked(so, SO_SND,
2171 min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2172 V_tcp_autosndbuf_max), curthread))
2173 so->so_snd.sb_flags &= ~SB_AUTOSIZE;