sys/netinet/tcp_output.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
   5  *      The Regents of the University of California.  All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  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.
  18  *
  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
  29  * SUCH DAMAGE.
  30  *
  31  *      @(#)tcp_output.c        8.4 (Berkeley) 5/24/95
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_inet.h"
  38 #include "opt_inet6.h"
  39 #include "opt_ipsec.h"
  40 #include "opt_kern_tls.h"
  41 #include "opt_tcpdebug.h"
  42
  43 #include <sys/param.h>
  44 #include <sys/systm.h>
  45 #include <sys/arb.h>
  46 #include <sys/domain.h>
  47 #ifdef TCP_HHOOK
  48 #include <sys/hhook.h>
  49 #endif
  50 #include <sys/kernel.h>
  51 #ifdef KERN_TLS
  52 #include <sys/ktls.h>
  53 #endif
  54 #include <sys/lock.h>
  55 #include <sys/mbuf.h>
  56 #include <sys/mutex.h>
  57 #include <sys/protosw.h>
  58 #include <sys/qmath.h>
  59 #include <sys/sdt.h>
  60 #include <sys/socket.h>
  61 #include <sys/socketvar.h>
  62 #include <sys/sysctl.h>
  63 #include <sys/stats.h>
  64
  65 #include <net/if.h>
  66 #include <net/route.h>
  67 #include <net/route/nhop.h>
  68 #include <net/vnet.h>
  69
  70 #include <netinet/in.h>
  71 #include <netinet/in_kdtrace.h>
  72 #include <netinet/in_systm.h>
  73 #include <netinet/ip.h>
  74 #include <netinet/in_pcb.h>
  75 #include <netinet/ip_var.h>
  76 #include <netinet/ip_options.h>
  77 #ifdef INET6
  78 #include <netinet6/in6_pcb.h>
  79 #include <netinet/ip6.h>
  80 #include <netinet6/ip6_var.h>
  81 #endif
  82 #include <netinet/tcp.h>
  83 #define TCPOUTFLAGS
  84 #include <netinet/tcp_fsm.h>
  85 #include <netinet/tcp_log_buf.h>
  86 #include <netinet/tcp_seq.h>
  87 #include <netinet/tcp_timer.h>
  88 #include <netinet/tcp_var.h>
  89 #include <netinet/tcpip.h>
  90 #include <netinet/cc/cc.h>
  91 #include <netinet/tcp_fastopen.h>
  92 #ifdef TCPPCAP
  93 #include <netinet/tcp_pcap.h>
  94 #endif
  95 #ifdef TCPDEBUG
  96 #include <netinet/tcp_debug.h>
  97 #endif
  98 #ifdef TCP_OFFLOAD
  99 #include <netinet/tcp_offload.h>
 100 #endif
 101
 102 #include <netipsec/ipsec_support.h>
 103
 104 #include <machine/in_cksum.h>
 105
 106 #include <security/mac/mac_framework.h>
 107
 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");
 112
 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");
 117
 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");
 122
 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");
 127
 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");
 132
 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");
 137
 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");
 143
 144 /*
 145  * Make sure that either retransmit or persist timer is set for SYN, FIN and
 146  * non-ACK.
 147  */
 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"))
 153
 154 static void inline      cc_after_idle(struct tcpcb *tp);
 155
 156 #ifdef TCP_HHOOK
 157 /*
 158  * Wrapper for the TCP established output helper hook.
 159  */
 160 void
 161 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
 162     struct tcpopt *to, uint32_t len, int tso)
 163 {
 164         struct tcp_hhook_data hhook_data;
 165
 166         if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
 167                 hhook_data.tp = tp;
 168                 hhook_data.th = th;
 169                 hhook_data.to = to;
 170                 hhook_data.len = len;
 171                 hhook_data.tso = tso;
 172
 173                 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
 174                     tp->osd);
 175         }
 176 }
 177 #endif
 178
 179 /*
 180  * CC wrapper hook functions
 181  */
 182 static void inline
 183 cc_after_idle(struct tcpcb *tp)
 184 {
 185         INP_WLOCK_ASSERT(tp->t_inpcb);
 186
 187         if (CC_ALGO(tp)->after_idle != NULL)
 188                 CC_ALGO(tp)->after_idle(tp->ccv);
 189 }
 190
 191 /*
 192  * Tcp output routine: figure out what should be sent and send it.
 193  */
 194 int
 195 tcp_output(struct tcpcb *tp)
 196 {
 197         struct socket *so = tp->t_inpcb->inp_socket;
 198         int32_t len;
 199         uint32_t recwin, sendwin;
 200         int off, flags, error = 0;      /* Keep compiler happy */
 201         u_int if_hw_tsomaxsegcount = 0;
 202         u_int if_hw_tsomaxsegsize = 0;
 203         struct mbuf *m;
 204         struct ip *ip = NULL;
 205 #ifdef TCPDEBUG
 206         struct ipovly *ipov = NULL;
 207 #endif
 208         struct tcphdr *th;
 209         u_char opt[TCP_MAXOLEN];
 210         unsigned ipoptlen, optlen, hdrlen;
 211 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 212         unsigned ipsec_optlen = 0;
 213 #endif
 214         int idle, sendalot, curticks;
 215         int sack_rxmit, sack_bytes_rxmt;
 216         struct sackhole *p;
 217         int tso, mtu;
 218         struct tcpopt to;
 219         unsigned int wanted_cookie = 0;
 220         unsigned int dont_sendalot = 0;
 221 #if 0
 222         int maxburst = TCP_MAXBURST;
 223 #endif
 224 #ifdef INET6
 225         struct ip6_hdr *ip6 = NULL;
 226         int isipv6;
 227
 228         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
 229 #endif
 230 #ifdef KERN_TLS
 231         const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
 232 #else
 233         const bool hw_tls = false;
 234 #endif
 235
 236         NET_EPOCH_ASSERT();
 237         INP_WLOCK_ASSERT(tp->t_inpcb);
 238
 239 #ifdef TCP_OFFLOAD
 240         if (tp->t_flags & TF_TOE)
 241                 return (tcp_offload_output(tp));
 242 #endif
 243
 244         /*
 245          * For TFO connections in SYN_SENT or SYN_RECEIVED,
 246          * only allow the initial SYN or SYN|ACK and those sent
 247          * by the retransmit timer.
 248          */
 249         if (IS_FASTOPEN(tp->t_flags) &&
 250             ((tp->t_state == TCPS_SYN_SENT) ||
 251              (tp->t_state == TCPS_SYN_RECEIVED)) &&
 252             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
 253             (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
 254                 return (0);
 255
 256         /*
 257          * Determine length of data that should be transmitted,
 258          * and flags that will be used.
 259          * If there is some data or critical controls (SYN, RST)
 260          * to send, then transmit; otherwise, investigate further.
 261          */
 262         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
 263         if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
 264             (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
 265                 cc_after_idle(tp);
 266         tp->t_flags &= ~TF_LASTIDLE;
 267         if (idle) {
 268                 if (tp->t_flags & TF_MORETOCOME) {
 269                         tp->t_flags |= TF_LASTIDLE;
 270                         idle = 0;
 271                 }
 272         }
 273 again:
 274         /*
 275          * If we've recently taken a timeout, snd_max will be greater than
 276          * snd_nxt.  There may be SACK information that allows us to avoid
 277          * resending already delivered data.  Adjust snd_nxt accordingly.
 278          */
 279         if ((tp->t_flags & TF_SACK_PERMIT) &&
 280             SEQ_LT(tp->snd_nxt, tp->snd_max))
 281                 tcp_sack_adjust(tp);
 282         sendalot = 0;
 283         tso = 0;
 284         mtu = 0;
 285         off = tp->snd_nxt - tp->snd_una;
 286         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
 287
 288         flags = tcp_outflags[tp->t_state];
 289         /*
 290          * Send any SACK-generated retransmissions.  If we're explicitly trying
 291          * to send out new data (when sendalot is 1), bypass this function.
 292          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
 293          * we're replacing a (future) new transmission with a retransmission
 294          * now, and we previously incremented snd_cwnd in tcp_input().
 295          */
 296         /*
 297          * Still in sack recovery , reset rxmit flag to zero.
 298          */
 299         sack_rxmit = 0;
 300         sack_bytes_rxmt = 0;
 301         len = 0;
 302         p = NULL;
 303         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
 304             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
 305                 uint32_t cwin;
 306
 307                 cwin =
 308                     imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
 309                 /* Do not retransmit SACK segments beyond snd_recover */
 310                 if (SEQ_GT(p->end, tp->snd_recover)) {
 311                         /*
 312                          * (At least) part of sack hole extends beyond
 313                          * snd_recover. Check to see if we can rexmit data
 314                          * for this hole.
 315                          */
 316                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
 317                                 /*
 318                                  * Can't rexmit any more data for this hole.
 319                                  * That data will be rexmitted in the next
 320                                  * sack recovery episode, when snd_recover
 321                                  * moves past p->rxmit.
 322                                  */
 323                                 p = NULL;
 324                                 goto after_sack_rexmit;
 325                         } else
 326                                 /* Can rexmit part of the current hole */
 327                                 len = ((int32_t)ulmin(cwin,
 328                                                    tp->snd_recover - p->rxmit));
 329                 } else
 330                         len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
 331                 off = p->rxmit - tp->snd_una;
 332                 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
 333                     __func__, off));
 334                 if (len > 0) {
 335                         sack_rxmit = 1;
 336                         sendalot = 1;
 337                         TCPSTAT_INC(tcps_sack_rexmits);
 338                         TCPSTAT_ADD(tcps_sack_rexmit_bytes,
 339                             min(len, tp->t_maxseg));
 340                 }
 341         }
 342 after_sack_rexmit:
 343         /*
 344          * Get standard flags, and add SYN or FIN if requested by 'hidden'
 345          * state flags.
 346          */
 347         if (tp->t_flags & TF_NEEDFIN)
 348                 flags |= TH_FIN;
 349         if (tp->t_flags & TF_NEEDSYN)
 350                 flags |= TH_SYN;
 351
 352         SOCKBUF_LOCK(&so->so_snd);
 353         /*
 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.
 358          */
 359         if (tp->t_flags & TF_FORCEDATA) {
 360                 if (sendwin == 0) {
 361                         /*
 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.
 371                          *
 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
 375                          * itself.
 376                          */
 377                         if (off < sbused(&so->so_snd))
 378                                 flags &= ~TH_FIN;
 379                         sendwin = 1;
 380                 } else {
 381                         tcp_timer_activate(tp, TT_PERSIST, 0);
 382                         tp->t_rxtshift = 0;
 383                 }
 384         }
 385
 386         /*
 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.
 393          *
 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
 396          * wind up 0.
 397          *
 398          * If sack_rxmit is true we are retransmitting from the scoreboard
 399          * in which case len is already set.
 400          */
 401         if (sack_rxmit == 0) {
 402                 if (sack_bytes_rxmt == 0)
 403                         len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
 404                             off);
 405                 else {
 406                         int32_t cwin;
 407
 408                         /*
 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.
 412                          */
 413                         len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
 414                             off);
 415                         /*
 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.
 422                          */
 423                         if (len > 0) {
 424                                 cwin = tp->snd_cwnd -
 425                                         (tp->snd_nxt - tp->snd_recover) -
 426                                         sack_bytes_rxmt;
 427                                 if (cwin < 0)
 428                                         cwin = 0;
 429                                 len = imin(len, cwin);
 430                         }
 431                 }
 432         }
 433
 434         /*
 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.
 438          */
 439         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
 440                 if (tp->t_state != TCPS_SYN_RECEIVED)
 441                         flags &= ~TH_SYN;
 442                 /*
 443                  * When sending additional segments following a TFO SYN|ACK,
 444                  * do not include the SYN bit.
 445                  */
 446                 if (IS_FASTOPEN(tp->t_flags) &&
 447                     (tp->t_state == TCPS_SYN_RECEIVED))
 448                         flags &= ~TH_SYN;
 449                 off--, len++;
 450         }
 451
 452         /*
 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.
 456          */
 457         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
 458                 len = 0;
 459                 flags &= ~TH_FIN;
 460         }
 461
 462         /*
 463          * On TFO sockets, ensure no data is sent in the following cases:
 464          *
 465          *  - When retransmitting SYN|ACK on a passively-created socket
 466          *
 467          *  - When retransmitting SYN on an actively created socket
 468          *
 469          *  - When sending a zero-length cookie (cookie request) on an
 470          *    actively created socket
 471          *
 472          *  - When the socket is in the CLOSED state (RST is being sent)
 473          */
 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)) ||
 478              (flags & TH_RST)))
 479                 len = 0;
 480         if (len <= 0) {
 481                 /*
 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.
 490                  *
 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.
 496                  */
 497                 len = 0;
 498                 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
 499                         (off < (int) sbavail(&so->so_snd))) {
 500                         tcp_timer_activate(tp, TT_REXMT, 0);
 501                         tp->t_rxtshift = 0;
 502                         tp->snd_nxt = tp->snd_una;
 503                         if (!tcp_timer_active(tp, TT_PERSIST))
 504                                 tcp_setpersist(tp);
 505                 }
 506         }
 507
 508         /* len will be >= 0 after this point. */
 509         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 510
 511         tcp_sndbuf_autoscale(tp, so, sendwin);
 512
 513         /*
 514          * Decide if we can use TCP Segmentation Offloading (if supported by
 515          * hardware).
 516          *
 517          * TSO may only be used if we are in a pure bulk sending state.  The
 518          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
 519          * IP options prevent using TSO.  With TSO the TCP header is the same
 520          * (except for the sequence number) for all generated packets.  This
 521          * makes it impossible to transmit any options which vary per generated
 522          * segment or packet.
 523          *
 524          * IPv4 handling has a clear separation of ip options and ip header
 525          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
 526          * the right thing below to provide length of just ip options and thus
 527          * checking for ipoptlen is enough to decide if ip options are present.
 528          */
 529 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 530         /*
 531          * Pre-calculate here as we save another lookup into the darknesses
 532          * of IPsec that way and can actually decide if TSO is ok.
 533          */
 534 #ifdef INET6
 535         if (isipv6 && IPSEC_ENABLED(ipv6))
 536                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
 537 #ifdef INET
 538         else
 539 #endif
 540 #endif /* INET6 */
 541 #ifdef INET
 542         if (IPSEC_ENABLED(ipv4))
 543                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
 544 #endif /* INET */
 545 #endif /* IPSEC */
 546 #ifdef INET6
 547         if (isipv6)
 548                 ipoptlen = ip6_optlen(tp->t_inpcb);
 549         else
 550 #endif
 551         if (tp->t_inpcb->inp_options)
 552                 ipoptlen = tp->t_inpcb->inp_options->m_len -
 553                                 offsetof(struct ipoption, ipopt_list);
 554         else
 555                 ipoptlen = 0;
 556 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 557         ipoptlen += ipsec_optlen;
 558 #endif
 559
 560         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
 561             ((tp->t_flags & TF_SIGNATURE) == 0) &&
 562             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
 563             ipoptlen == 0 && !(flags & TH_SYN))
 564                 tso = 1;
 565
 566         if (sack_rxmit) {
 567                 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
 568                         flags &= ~TH_FIN;
 569         } else {
 570                 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
 571                     sbused(&so->so_snd)))
 572                         flags &= ~TH_FIN;
 573         }
 574
 575         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 576             (long)TCP_MAXWIN << tp->rcv_scale);
 577
 578         /*
 579          * Sender silly window avoidance.   We transmit under the following
 580          * conditions when len is non-zero:
 581          *
 582          *      - We have a full segment (or more with TSO)
 583          *      - This is the last buffer in a write()/send() and we are
 584          *        either idle or running NODELAY
 585          *      - we've timed out (e.g. persist timer)
 586          *      - we have more then 1/2 the maximum send window's worth of
 587          *        data (receiver may be limited the window size)
 588          *      - we need to retransmit
 589          */
 590         if (len) {
 591                 if (len >= tp->t_maxseg)
 592                         goto send;
 593                 /*
 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
 597                  *
 598                  * note: the len + off check is almost certainly unnecessary.
 599                  */
 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) {
 604                         goto send;
 605                 }
 606                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
 607                         goto send;
 608                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
 609                         goto send;
 610                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
 611                         goto send;
 612                 if (sack_rxmit)
 613                         goto send;
 614         }
 615
 616         /*
 617          * Sending of standalone window updates.
 618          *
 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.
 624          *
 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.
 630          *
 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.
 638          *
 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
 643          * half-open state.
 644          */
 645         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
 646             !(tp->t_flags & TF_DELACK) &&
 647             !TCPS_HAVERCVDFIN(tp->t_state)) {
 648                 /*
 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.
 652                  */
 653                 int32_t adv;
 654                 int oldwin;
 655
 656                 adv = recwin;
 657                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
 658                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
 659                         if (adv > oldwin)
 660                                 adv -= oldwin;
 661                         else
 662                                 adv = 0;
 663                 } else
 664                         oldwin = 0;
 665
 666                 /*
 667                  * If the new window size ends up being the same as or less
 668                  * than the old size when it is scaled, then don't force
 669                  * a window update.
 670                  */
 671                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
 672                         goto dontupdate;
 673
 674                 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
 675                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
 676                      recwin <= (so->so_rcv.sb_hiwat / 8) ||
 677                      so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
 678                      adv >= TCP_MAXWIN << tp->rcv_scale))
 679                         goto send;
 680                 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
 681                         goto send;
 682         }
 683 dontupdate:
 684
 685         /*
 686          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
 687          * is also a catch-all for the retransmit timer timeout case.
 688          */
 689         if (tp->t_flags & TF_ACKNOW)
 690                 goto send;
 691         if ((flags & TH_RST) ||
 692             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
 693                 goto send;
 694         if (SEQ_GT(tp->snd_up, tp->snd_una))
 695                 goto send;
 696         /*
 697          * If our state indicates that FIN should be sent
 698          * and we have not yet done so, then we need to send.
 699          */
 700         if (flags & TH_FIN &&
 701             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
 702                 goto send;
 703         /*
 704          * In SACK, it is possible for tcp_output to fail to send a segment
 705          * after the retransmission timer has been turned off.  Make sure
 706          * that the retransmission timer is set.
 707          */
 708         if ((tp->t_flags & TF_SACK_PERMIT) &&
 709             SEQ_GT(tp->snd_max, tp->snd_una) &&
 710             !tcp_timer_active(tp, TT_REXMT) &&
 711             !tcp_timer_active(tp, TT_PERSIST)) {
 712                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 713                 goto just_return;
 714         }
 715         /*
 716          * TCP window updates are not reliable, rather a polling protocol
 717          * using ``persist'' packets is used to insure receipt of window
 718          * updates.  The three ``states'' for the output side are:
 719          *      idle                    not doing retransmits or persists
 720          *      persisting              to move a small or zero window
 721          *      (re)transmitting        and thereby not persisting
 722          *
 723          * tcp_timer_active(tp, TT_PERSIST)
 724          *      is true when we are in persist state.
 725          * (tp->t_flags & TF_FORCEDATA)
 726          *      is set when we are called to send a persist packet.
 727          * tcp_timer_active(tp, TT_REXMT)
 728          *      is set when we are retransmitting
 729          * The output side is idle when both timers are zero.
 730          *
 731          * If send window is too small, there is data to transmit, and no
 732          * retransmit or persist is pending, then go to persist state.
 733          * If nothing happens soon, send when timer expires:
 734          * if window is nonzero, transmit what we can,
 735          * otherwise force out a byte.
 736          */
 737         if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
 738             !tcp_timer_active(tp, TT_PERSIST)) {
 739                 tp->t_rxtshift = 0;
 740                 tcp_setpersist(tp);
 741         }
 742
 743         /*
 744          * No reason to send a segment, just return.
 745          */
 746 just_return:
 747         SOCKBUF_UNLOCK(&so->so_snd);
 748         return (0);
 749
 750 send:
 751         SOCKBUF_LOCK_ASSERT(&so->so_snd);
 752         if (len > 0) {
 753                 if (len >= tp->t_maxseg)
 754                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
 755                 else
 756                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
 757         }
 758         /*
 759          * Before ESTABLISHED, force sending of initial options
 760          * unless TCP set not to do any options.
 761          * NOTE: we assume that the IP/TCP header plus TCP options
 762          * always fit in a single mbuf, leaving room for a maximum
 763          * link header, i.e.
 764          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
 765          */
 766         optlen = 0;
 767 #ifdef INET6
 768         if (isipv6)
 769                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
 770         else
 771 #endif
 772                 hdrlen = sizeof (struct tcpiphdr);
 773
 774         /*
 775          * Compute options for segment.
 776          * We only have to care about SYN and established connection
 777          * segments.  Options for SYN-ACK segments are handled in TCP
 778          * syncache.
 779          */
 780         to.to_flags = 0;
 781         if ((tp->t_flags & TF_NOOPT) == 0) {
 782                 /* Maximum segment size. */
 783                 if (flags & TH_SYN) {
 784                         tp->snd_nxt = tp->iss;
 785                         to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
 786                         to.to_flags |= TOF_MSS;
 787
 788                         /*
 789                          * On SYN or SYN|ACK transmits on TFO connections,
 790                          * only include the TFO option if it is not a
 791                          * retransmit, as the presence of the TFO option may
 792                          * have caused the original SYN or SYN|ACK to have
 793                          * been dropped by a middlebox.
 794                          */
 795                         if (IS_FASTOPEN(tp->t_flags) &&
 796                             (tp->t_rxtshift == 0)) {
 797                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
 798                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 799                                         to.to_tfo_cookie =
 800                                             (u_int8_t *)&tp->t_tfo_cookie.server;
 801                                         to.to_flags |= TOF_FASTOPEN;
 802                                         wanted_cookie = 1;
 803                                 } else if (tp->t_state == TCPS_SYN_SENT) {
 804                                         to.to_tfo_len =
 805                                             tp->t_tfo_client_cookie_len;
 806                                         to.to_tfo_cookie =
 807                                             tp->t_tfo_cookie.client;
 808                                         to.to_flags |= TOF_FASTOPEN;
 809                                         wanted_cookie = 1;
 810                                         /*
 811                                          * If we wind up having more data to
 812                                          * send with the SYN than can fit in
 813                                          * one segment, don't send any more
 814                                          * until the SYN|ACK comes back from
 815                                          * the other end.
 816                                          */
 817                                         dont_sendalot = 1;
 818                                 }
 819                         }
 820                 }
 821                 /* Window scaling. */
 822                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
 823                         to.to_wscale = tp->request_r_scale;
 824                         to.to_flags |= TOF_SCALE;
 825                 }
 826                 /* Timestamps. */
 827                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
 828                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
 829                         curticks = tcp_ts_getticks();
 830                         to.to_tsval = curticks + tp->ts_offset;
 831                         to.to_tsecr = tp->ts_recent;
 832                         to.to_flags |= TOF_TS;
 833                         if (tp->t_rxtshift == 1)
 834                                 tp->t_badrxtwin = curticks;
 835                 }
 836
 837                 /* Set receive buffer autosizing timestamp. */
 838                 if (tp->rfbuf_ts == 0 &&
 839                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
 840                         tp->rfbuf_ts = tcp_ts_getticks();
 841
 842                 /* Selective ACK's. */
 843                 if (tp->t_flags & TF_SACK_PERMIT) {
 844                         if (flags & TH_SYN)
 845                                 to.to_flags |= TOF_SACKPERM;
 846                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
 847                             (tp->t_flags & TF_SACK_PERMIT) &&
 848                             tp->rcv_numsacks > 0) {
 849                                 to.to_flags |= TOF_SACK;
 850                                 to.to_nsacks = tp->rcv_numsacks;
 851                                 to.to_sacks = (u_char *)tp->sackblks;
 852                         }
 853                 }
 854 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 855                 /* TCP-MD5 (RFC2385). */
 856                 /*
 857                  * Check that TCP_MD5SIG is enabled in tcpcb to
 858                  * account the size needed to set this TCP option.
 859                  */
 860                 if (tp->t_flags & TF_SIGNATURE)
 861                         to.to_flags |= TOF_SIGNATURE;
 862 #endif /* TCP_SIGNATURE */
 863
 864                 /* Processing the options. */
 865                 hdrlen += optlen = tcp_addoptions(&to, opt);
 866                 /*
 867                  * If we wanted a TFO option to be added, but it was unable
 868                  * to fit, ensure no data is sent.
 869                  */
 870                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
 871                     !(to.to_flags & TOF_FASTOPEN))
 872                         len = 0;
 873         }
 874
 875         /*
 876          * Adjust data length if insertion of options will
 877          * bump the packet length beyond the t_maxseg length.
 878          * Clear the FIN bit because we cut off the tail of
 879          * the segment.
 880          */
 881         if (len + optlen + ipoptlen > tp->t_maxseg) {
 882                 flags &= ~TH_FIN;
 883
 884                 if (tso) {
 885                         u_int if_hw_tsomax;
 886                         u_int moff;
 887                         int max_len;
 888
 889                         /* extract TSO information */
 890                         if_hw_tsomax = tp->t_tsomax;
 891                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
 892                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
 893
 894                         /*
 895                          * Limit a TSO burst to prevent it from
 896                          * overflowing or exceeding the maximum length
 897                          * allowed by the network interface:
 898                          */
 899                         KASSERT(ipoptlen == 0,
 900                             ("%s: TSO can't do IP options", __func__));
 901
 902                         /*
 903                          * Check if we should limit by maximum payload
 904                          * length:
 905                          */
 906                         if (if_hw_tsomax != 0) {
 907                                 /* compute maximum TSO length */
 908                                 max_len = (if_hw_tsomax - hdrlen -
 909                                     max_linkhdr);
 910                                 if (max_len <= 0) {
 911                                         len = 0;
 912                                 } else if (len > max_len) {
 913                                         sendalot = 1;
 914                                         len = max_len;
 915                                 }
 916                         }
 917
 918                         /*
 919                          * Prevent the last segment from being
 920                          * fractional unless the send sockbuf can be
 921                          * emptied:
 922                          */
 923                         max_len = (tp->t_maxseg - optlen);
 924                         if (((uint32_t)off + (uint32_t)len) <
 925                             sbavail(&so->so_snd)) {
 926                                 moff = len % max_len;
 927                                 if (moff != 0) {
 928                                         len -= moff;
 929                                         sendalot = 1;
 930                                 }
 931                         }
 932
 933                         /*
 934                          * In case there are too many small fragments
 935                          * don't use TSO:
 936                          */
 937                         if (len <= max_len) {
 938                                 len = max_len;
 939                                 sendalot = 1;
 940                                 tso = 0;
 941                         }
 942
 943                         /*
 944                          * Send the FIN in a separate segment
 945                          * after the bulk sending is done.
 946                          * We don't trust the TSO implementations
 947                          * to clear the FIN flag on all but the
 948                          * last segment.
 949                          */
 950                         if (tp->t_flags & TF_NEEDFIN)
 951                                 sendalot = 1;
 952                 } else {
 953                         if (optlen + ipoptlen >= tp->t_maxseg) {
 954                                 /*
 955                                  * Since we don't have enough space to put
 956                                  * the IP header chain and the TCP header in
 957                                  * one packet as required by RFC 7112, don't
 958                                  * send it. Also ensure that at least one
 959                                  * byte of the payload can be put into the
 960                                  * TCP segment.
 961                                  */
 962                                 SOCKBUF_UNLOCK(&so->so_snd);
 963                                 error = EMSGSIZE;
 964                                 sack_rxmit = 0;
 965                                 goto out;
 966                         }
 967                         len = tp->t_maxseg - optlen - ipoptlen;
 968                         sendalot = 1;
 969                         if (dont_sendalot)
 970                                 sendalot = 0;
 971                 }
 972         } else
 973                 tso = 0;
 974
 975         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
 976             ("%s: len > IP_MAXPACKET", __func__));
 977
 978 /*#ifdef DIAGNOSTIC*/
 979 #ifdef INET6
 980         if (max_linkhdr + hdrlen > MCLBYTES)
 981 #else
 982         if (max_linkhdr + hdrlen > MHLEN)
 983 #endif
 984                 panic("tcphdr too big");
 985 /*#endif*/
 986
 987         /*
 988          * This KASSERT is here to catch edge cases at a well defined place.
 989          * Before, those had triggered (random) panic conditions further down.
 990          */
 991         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 992
 993         /*
 994          * Grab a header mbuf, attaching a copy of data to
 995          * be transmitted, and initialize the header from
 996          * the template for sends on this connection.
 997          */
 998         if (len) {
 999                 struct mbuf *mb;
1000                 struct sockbuf *msb;
1001                 u_int moff;
1002
1003                 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1004                         TCPSTAT_INC(tcps_sndprobe);
1005 #ifdef STATS
1006                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1007                                 stats_voi_update_abs_u32(tp->t_stats,
1008                                 VOI_TCP_RETXPB, len);
1009                         else
1010                                 stats_voi_update_abs_u64(tp->t_stats,
1011                                     VOI_TCP_TXPB, len);
1012 #endif /* STATS */
1013                 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1014                         tp->t_sndrexmitpack++;
1015                         TCPSTAT_INC(tcps_sndrexmitpack);
1016                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1017 #ifdef STATS
1018                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1019                             len);
1020 #endif /* STATS */
1021                 } else {
1022                         TCPSTAT_INC(tcps_sndpack);
1023                         TCPSTAT_ADD(tcps_sndbyte, len);
1024 #ifdef STATS
1025                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1026                             len);
1027 #endif /* STATS */
1028                 }
1029 #ifdef INET6
1030                 if (MHLEN < hdrlen + max_linkhdr)
1031                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1032                 else
1033 #endif
1034                         m = m_gethdr(M_NOWAIT, MT_DATA);
1035
1036                 if (m == NULL) {
1037                         SOCKBUF_UNLOCK(&so->so_snd);
1038                         error = ENOBUFS;
1039                         sack_rxmit = 0;
1040                         goto out;
1041                 }
1042
1043                 m->m_data += max_linkhdr;
1044                 m->m_len = hdrlen;
1045
1046                 /*
1047                  * Start the m_copy functions from the closest mbuf
1048                  * to the offset in the socket buffer chain.
1049                  */
1050                 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1051                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1052                         m_copydata(mb, moff, len,
1053                             mtod(m, caddr_t) + hdrlen);
1054                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1055                                 sbsndptr_adv(&so->so_snd, mb, len);
1056                         m->m_len += len;
1057                 } else {
1058                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1059                                 msb = NULL;
1060                         else
1061                                 msb = &so->so_snd;
1062                         m->m_next = tcp_m_copym(mb, moff,
1063                             &len, if_hw_tsomaxsegcount,
1064                             if_hw_tsomaxsegsize, msb, hw_tls);
1065                         if (len <= (tp->t_maxseg - optlen)) {
1066                                 /*
1067                                  * Must have ran out of mbufs for the copy
1068                                  * shorten it to no longer need tso. Lets
1069                                  * not put on sendalot since we are low on
1070                                  * mbufs.
1071                                  */
1072                                 tso = 0;
1073                         }
1074                         if (m->m_next == NULL) {
1075                                 SOCKBUF_UNLOCK(&so->so_snd);
1076                                 (void) m_free(m);
1077                                 error = ENOBUFS;
1078                                 sack_rxmit = 0;
1079                                 goto out;
1080                         }
1081                 }
1082
1083                 /*
1084                  * If we're sending everything we've got, set PUSH.
1085                  * (This will keep happy those implementations which only
1086                  * give data to the user when a buffer fills or
1087                  * a PUSH comes in.)
1088                  */
1089                 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1090                     !(flags & TH_SYN))
1091                         flags |= TH_PUSH;
1092                 SOCKBUF_UNLOCK(&so->so_snd);
1093         } else {
1094                 SOCKBUF_UNLOCK(&so->so_snd);
1095                 if (tp->t_flags & TF_ACKNOW)
1096                         TCPSTAT_INC(tcps_sndacks);
1097                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1098                         TCPSTAT_INC(tcps_sndctrl);
1099                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1100                         TCPSTAT_INC(tcps_sndurg);
1101                 else
1102                         TCPSTAT_INC(tcps_sndwinup);
1103
1104                 m = m_gethdr(M_NOWAIT, MT_DATA);
1105                 if (m == NULL) {
1106                         error = ENOBUFS;
1107                         sack_rxmit = 0;
1108                         goto out;
1109                 }
1110 #ifdef INET6
1111                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1112                     MHLEN >= hdrlen) {
1113                         M_ALIGN(m, hdrlen);
1114                 } else
1115 #endif
1116                 m->m_data += max_linkhdr;
1117                 m->m_len = hdrlen;
1118         }
1119         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1120         m->m_pkthdr.rcvif = (struct ifnet *)0;
1121 #ifdef MAC
1122         mac_inpcb_create_mbuf(tp->t_inpcb, m);
1123 #endif
1124 #ifdef INET6
1125         if (isipv6) {
1126                 ip6 = mtod(m, struct ip6_hdr *);
1127                 th = (struct tcphdr *)(ip6 + 1);
1128                 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1129         } else
1130 #endif /* INET6 */
1131         {
1132                 ip = mtod(m, struct ip *);
1133 #ifdef TCPDEBUG
1134                 ipov = (struct ipovly *)ip;
1135 #endif
1136                 th = (struct tcphdr *)(ip + 1);
1137                 tcpip_fillheaders(tp->t_inpcb, ip, th);
1138         }
1139
1140         /*
1141          * Fill in fields, remembering maximum advertised
1142          * window for use in delaying messages about window sizes.
1143          * If resending a FIN, be sure not to use a new sequence number.
1144          */
1145         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1146             tp->snd_nxt == tp->snd_max)
1147                 tp->snd_nxt--;
1148         /*
1149          * If we are starting a connection, send ECN setup
1150          * SYN packet. If we are on a retransmit, we may
1151          * resend those bits a number of times as per
1152          * RFC 3168.
1153          */
1154         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1155                 if (tp->t_rxtshift >= 1) {
1156                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1157                                 flags |= TH_ECE|TH_CWR;
1158                 } else
1159                         flags |= TH_ECE|TH_CWR;
1160         }
1161         /* Handle parallel SYN for ECN */
1162         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1163             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1164                         flags |= TH_ECE;
1165                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1166         }
1167
1168         if (tp->t_state == TCPS_ESTABLISHED &&
1169             (tp->t_flags2 & TF2_ECN_PERMIT)) {
1170                 /*
1171                  * If the peer has ECN, mark data packets with
1172                  * ECN capable transmission (ECT).
1173                  * Ignore pure ack packets, retransmissions and window probes.
1174                  */
1175                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1176                     (sack_rxmit == 0) &&
1177                     !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1178                     SEQ_LT(tp->snd_una, tp->snd_max))) {
1179 #ifdef INET6
1180                         if (isipv6)
1181                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1182                         else
1183 #endif
1184                                 ip->ip_tos |= IPTOS_ECN_ECT0;
1185                         TCPSTAT_INC(tcps_ecn_ect0);
1186                         /*
1187                          * Reply with proper ECN notifications.
1188                          * Only set CWR on new data segments.
1189                          */
1190                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1191                                 flags |= TH_CWR;
1192                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1193                         }
1194                 }
1195                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
1196                         flags |= TH_ECE;
1197         }
1198
1199         /*
1200          * If we are doing retransmissions, then snd_nxt will
1201          * not reflect the first unsent octet.  For ACK only
1202          * packets, we do not want the sequence number of the
1203          * retransmitted packet, we want the sequence number
1204          * of the next unsent octet.  So, if there is no data
1205          * (and no SYN or FIN), use snd_max instead of snd_nxt
1206          * when filling in ti_seq.  But if we are in persist
1207          * state, snd_max might reflect one byte beyond the
1208          * right edge of the window, so use snd_nxt in that
1209          * case, since we know we aren't doing a retransmission.
1210          * (retransmit and persist are mutually exclusive...)
1211          */
1212         if (sack_rxmit == 0) {
1213                 if (len || (flags & (TH_SYN|TH_FIN)) ||
1214                     tcp_timer_active(tp, TT_PERSIST))
1215                         th->th_seq = htonl(tp->snd_nxt);
1216                 else
1217                         th->th_seq = htonl(tp->snd_max);
1218         } else {
1219                 th->th_seq = htonl(p->rxmit);
1220                 p->rxmit += len;
1221                 tp->sackhint.sack_bytes_rexmit += len;
1222         }
1223         th->th_ack = htonl(tp->rcv_nxt);
1224         if (optlen) {
1225                 bcopy(opt, th + 1, optlen);
1226                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1227         }
1228         th->th_flags = flags;
1229         /*
1230          * Calculate receive window.  Don't shrink window,
1231          * but avoid silly window syndrome.
1232          * If a RST segment is sent, advertise a window of zero.
1233          */
1234         if (flags & TH_RST) {
1235                 recwin = 0;
1236         } else {
1237                 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1238                     recwin < tp->t_maxseg)
1239                         recwin = 0;
1240                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1241                     recwin < (tp->rcv_adv - tp->rcv_nxt))
1242                         recwin = (tp->rcv_adv - tp->rcv_nxt);
1243         }
1244         /*
1245          * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1246          * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1247          * case is handled in syncache.
1248          */
1249         if (flags & TH_SYN)
1250                 th->th_win = htons((u_short)
1251                                 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1252         else {
1253                 /* Avoid shrinking window with window scaling. */
1254                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1255                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1256         }
1257
1258         /*
1259          * Adjust the RXWIN0SENT flag - indicate that we have advertised
1260          * a 0 window.  This may cause the remote transmitter to stall.  This
1261          * flag tells soreceive() to disable delayed acknowledgements when
1262          * draining the buffer.  This can occur if the receiver is attempting
1263          * to read more data than can be buffered prior to transmitting on
1264          * the connection.
1265          */
1266         if (th->th_win == 0) {
1267                 tp->t_sndzerowin++;
1268                 tp->t_flags |= TF_RXWIN0SENT;
1269         } else
1270                 tp->t_flags &= ~TF_RXWIN0SENT;
1271         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1272                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1273                 th->th_flags |= TH_URG;
1274         } else
1275                 /*
1276                  * If no urgent pointer to send, then we pull
1277                  * the urgent pointer to the left edge of the send window
1278                  * so that it doesn't drift into the send window on sequence
1279                  * number wraparound.
1280                  */
1281                 tp->snd_up = tp->snd_una;               /* drag it along */
1282
1283         /*
1284          * Put TCP length in extended header, and then
1285          * checksum extended header and data.
1286          */
1287         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1288         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1289
1290 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1291         if (to.to_flags & TOF_SIGNATURE) {
1292                 /*
1293                  * Calculate MD5 signature and put it into the place
1294                  * determined before.
1295                  * NOTE: since TCP options buffer doesn't point into
1296                  * mbuf's data, calculate offset and use it.
1297                  */
1298                 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1299                     (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1300                         /*
1301                          * Do not send segment if the calculation of MD5
1302                          * digest has failed.
1303                          */
1304                         m_freem(m);
1305                         goto out;
1306                 }
1307         }
1308 #endif
1309 #ifdef INET6
1310         if (isipv6) {
1311                 /*
1312                  * There is no need to fill in ip6_plen right now.
1313                  * It will be filled later by ip6_output.
1314                  */
1315                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1316                 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1317                     optlen + len, IPPROTO_TCP, 0);
1318         }
1319 #endif
1320 #if defined(INET6) && defined(INET)
1321         else
1322 #endif
1323 #ifdef INET
1324         {
1325                 m->m_pkthdr.csum_flags = CSUM_TCP;
1326                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1327                     htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1328
1329                 /* IP version must be set here for ipv4/ipv6 checking later */
1330                 KASSERT(ip->ip_v == IPVERSION,
1331                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1332         }
1333 #endif
1334
1335         /*
1336          * Enable TSO and specify the size of the segments.
1337          * The TCP pseudo header checksum is always provided.
1338          */
1339         if (tso) {
1340                 KASSERT(len > tp->t_maxseg - optlen,
1341                     ("%s: len <= tso_segsz", __func__));
1342                 m->m_pkthdr.csum_flags |= CSUM_TSO;
1343                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1344         }
1345
1346         KASSERT(len + hdrlen == m_length(m, NULL),
1347             ("%s: mbuf chain shorter than expected: %d + %u != %u",
1348             __func__, len, hdrlen, m_length(m, NULL)));
1349
1350 #ifdef TCP_HHOOK
1351         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1352         hhook_run_tcp_est_out(tp, th, &to, len, tso);
1353 #endif
1354
1355 #ifdef TCPDEBUG
1356         /*
1357          * Trace.
1358          */
1359         if (so->so_options & SO_DEBUG) {
1360                 u_short save = 0;
1361 #ifdef INET6
1362                 if (!isipv6)
1363 #endif
1364                 {
1365                         save = ipov->ih_len;
1366                         ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1367                 }
1368                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1369 #ifdef INET6
1370                 if (!isipv6)
1371 #endif
1372                 ipov->ih_len = save;
1373         }
1374 #endif /* TCPDEBUG */
1375         TCP_PROBE3(debug__output, tp, th, m);
1376
1377         /* We're getting ready to send; log now. */
1378         TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1379             len, NULL, false);
1380
1381         /*
1382          * Fill in IP length and desired time to live and
1383          * send to IP level.  There should be a better way
1384          * to handle ttl and tos; we could keep them in
1385          * the template, but need a way to checksum without them.
1386          */
1387         /*
1388          * m->m_pkthdr.len should have been set before checksum calculation,
1389          * because in6_cksum() need it.
1390          */
1391 #ifdef INET6
1392         if (isipv6) {
1393                 /*
1394                  * we separately set hoplimit for every segment, since the
1395                  * user might want to change the value via setsockopt.
1396                  * Also, desired default hop limit might be changed via
1397                  * Neighbor Discovery.
1398                  */
1399                 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1400
1401                 /*
1402                  * Set the packet size here for the benefit of DTrace probes.
1403                  * ip6_output() will set it properly; it's supposed to include
1404                  * the option header lengths as well.
1405                  */
1406                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1407
1408                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1409                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1410                 else
1411                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1412
1413                 if (tp->t_state == TCPS_SYN_SENT)
1414                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1415
1416                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1417
1418 #ifdef TCPPCAP
1419                 /* Save packet, if requested. */
1420                 tcp_pcap_add(th, m, &(tp->t_outpkts));
1421 #endif
1422
1423                 /* TODO: IPv6 IP6TOS_ECT bit on */
1424                 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1425                     &tp->t_inpcb->inp_route6,
1426                     ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1427                     NULL, NULL, tp->t_inpcb);
1428
1429                 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1430                         mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1431         }
1432 #endif /* INET6 */
1433 #if defined(INET) && defined(INET6)
1434         else
1435 #endif
1436 #ifdef INET
1437     {
1438         ip->ip_len = htons(m->m_pkthdr.len);
1439 #ifdef INET6
1440         if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1441                 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1442 #endif /* INET6 */
1443         /*
1444          * If we do path MTU discovery, then we set DF on every packet.
1445          * This might not be the best thing to do according to RFC3390
1446          * Section 2. However the tcp hostcache migitates the problem
1447          * so it affects only the first tcp connection with a host.
1448          *
1449          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1450          */
1451         if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1452                 ip->ip_off |= htons(IP_DF);
1453                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1454         } else {
1455                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1456         }
1457
1458         if (tp->t_state == TCPS_SYN_SENT)
1459                 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1460
1461         TCP_PROBE5(send, NULL, tp, ip, tp, th);
1462
1463 #ifdef TCPPCAP
1464         /* Save packet, if requested. */
1465         tcp_pcap_add(th, m, &(tp->t_outpkts));
1466 #endif
1467
1468         error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1469             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1470             tp->t_inpcb);
1471
1472         if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1473                 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1474     }
1475 #endif /* INET */
1476
1477 out:
1478         /*
1479          * In transmit state, time the transmission and arrange for
1480          * the retransmit.  In persist state, just set snd_max.
1481          */
1482         if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1483             !tcp_timer_active(tp, TT_PERSIST)) {
1484                 tcp_seq startseq = tp->snd_nxt;
1485
1486                 /*
1487                  * Advance snd_nxt over sequence space of this segment.
1488                  */
1489                 if (flags & (TH_SYN|TH_FIN)) {
1490                         if (flags & TH_SYN)
1491                                 tp->snd_nxt++;
1492                         if (flags & TH_FIN) {
1493                                 tp->snd_nxt++;
1494                                 tp->t_flags |= TF_SENTFIN;
1495                         }
1496                 }
1497                 if (sack_rxmit)
1498                         goto timer;
1499                 tp->snd_nxt += len;
1500                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1501                         tp->snd_max = tp->snd_nxt;
1502                         /*
1503                          * Time this transmission if not a retransmission and
1504                          * not currently timing anything.
1505                          */
1506                         tp->t_sndtime = ticks;
1507                         if (tp->t_rtttime == 0) {
1508                                 tp->t_rtttime = ticks;
1509                                 tp->t_rtseq = startseq;
1510                                 TCPSTAT_INC(tcps_segstimed);
1511                         }
1512 #ifdef STATS
1513                         if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1514                                 tp->t_flags |= TF_GPUTINPROG;
1515                                 tp->gput_seq = startseq;
1516                                 tp->gput_ack = startseq +
1517                                     ulmin(sbavail(&so->so_snd) - off, sendwin);
1518                                 tp->gput_ts = tcp_ts_getticks();
1519                         }
1520 #endif /* STATS */
1521                 }
1522
1523                 /*
1524                  * Set retransmit timer if not currently set,
1525                  * and not doing a pure ack or a keep-alive probe.
1526                  * Initial value for retransmit timer is smoothed
1527                  * round-trip time + 2 * round-trip time variance.
1528                  * Initialize shift counter which is used for backoff
1529                  * of retransmit time.
1530                  */
1531 timer:
1532                 if (!tcp_timer_active(tp, TT_REXMT) &&
1533                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1534                      (tp->snd_nxt != tp->snd_una))) {
1535                         if (tcp_timer_active(tp, TT_PERSIST)) {
1536                                 tcp_timer_activate(tp, TT_PERSIST, 0);
1537                                 tp->t_rxtshift = 0;
1538                         }
1539                         tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1540                 } else if (len == 0 && sbavail(&so->so_snd) &&
1541                     !tcp_timer_active(tp, TT_REXMT) &&
1542                     !tcp_timer_active(tp, TT_PERSIST)) {
1543                         /*
1544                          * Avoid a situation where we do not set persist timer
1545                          * after a zero window condition. For example:
1546                          * 1) A -> B: packet with enough data to fill the window
1547                          * 2) B -> A: ACK for #1 + new data (0 window
1548                          *    advertisement)
1549                          * 3) A -> B: ACK for #2, 0 len packet
1550                          *
1551                          * In this case, A will not activate the persist timer,
1552                          * because it chose to send a packet. Unless tcp_output
1553                          * is called for some other reason (delayed ack timer,
1554                          * another input packet from B, socket syscall), A will
1555                          * not send zero window probes.
1556                          *
1557                          * So, if you send a 0-length packet, but there is data
1558                          * in the socket buffer, and neither the rexmt or
1559                          * persist timer is already set, then activate the
1560                          * persist timer.
1561                          */
1562                         tp->t_rxtshift = 0;
1563                         tcp_setpersist(tp);
1564                 }
1565         } else {
1566                 /*
1567                  * Persist case, update snd_max but since we are in
1568                  * persist mode (no window) we do not update snd_nxt.
1569                  */
1570                 int xlen = len;
1571                 if (flags & TH_SYN)
1572                         ++xlen;
1573                 if (flags & TH_FIN) {
1574                         ++xlen;
1575                         tp->t_flags |= TF_SENTFIN;
1576                 }
1577                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1578                         tp->snd_max = tp->snd_nxt + xlen;
1579         }
1580         if ((error == 0) &&
1581             (TCPS_HAVEESTABLISHED(tp->t_state) &&
1582              (tp->t_flags & TF_SACK_PERMIT) &&
1583              tp->rcv_numsacks > 0)) {
1584                     /* Clean up any DSACK's sent */
1585                     tcp_clean_dsack_blocks(tp);
1586         }
1587         if (error) {
1588                 /* Record the error. */
1589                 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1590                     error, 0, NULL, false);
1591
1592                 /*
1593                  * We know that the packet was lost, so back out the
1594                  * sequence number advance, if any.
1595                  *
1596                  * If the error is EPERM the packet got blocked by the
1597                  * local firewall.  Normally we should terminate the
1598                  * connection but the blocking may have been spurious
1599                  * due to a firewall reconfiguration cycle.  So we treat
1600                  * it like a packet loss and let the retransmit timer and
1601                  * timeouts do their work over time.
1602                  * XXX: It is a POLA question whether calling tcp_drop right
1603                  * away would be the really correct behavior instead.
1604                  */
1605                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1606                     !tcp_timer_active(tp, TT_PERSIST)) &&
1607                     ((flags & TH_SYN) == 0) &&
1608                     (error != EPERM)) {
1609                         if (sack_rxmit) {
1610                                 p->rxmit -= len;
1611                                 tp->sackhint.sack_bytes_rexmit -= len;
1612                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1613                                     ("sackhint bytes rtx >= 0"));
1614                         } else
1615                                 tp->snd_nxt -= len;
1616                 }
1617                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
1618                 switch (error) {
1619                 case EACCES:
1620                 case EPERM:
1621                         tp->t_softerror = error;
1622                         return (error);
1623                 case ENOBUFS:
1624                         TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1625                         tp->snd_cwnd = tp->t_maxseg;
1626                         return (0);
1627                 case EMSGSIZE:
1628                         /*
1629                          * For some reason the interface we used initially
1630                          * to send segments changed to another or lowered
1631                          * its MTU.
1632                          * If TSO was active we either got an interface
1633                          * without TSO capabilits or TSO was turned off.
1634                          * If we obtained mtu from ip_output() then update
1635                          * it and try again.
1636                          */
1637                         if (tso)
1638                                 tp->t_flags &= ~TF_TSO;
1639                         if (mtu != 0) {
1640                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1641                                 goto again;
1642                         }
1643                         return (error);
1644                 case EHOSTDOWN:
1645                 case EHOSTUNREACH:
1646                 case ENETDOWN:
1647                 case ENETUNREACH:
1648                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
1649                                 tp->t_softerror = error;
1650                                 return (0);
1651                         }
1652                         /* FALLTHROUGH */
1653                 default:
1654                         return (error);
1655                 }
1656         }
1657         TCPSTAT_INC(tcps_sndtotal);
1658
1659         /*
1660          * Data sent (as far as we can tell).
1661          * If this advertises a larger window than any other segment,
1662          * then remember the size of the advertised window.
1663          * Any pending ACK has now been sent.
1664          */
1665         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1666                 tp->rcv_adv = tp->rcv_nxt + recwin;
1667         tp->last_ack_sent = tp->rcv_nxt;
1668         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1669         if (tcp_timer_active(tp, TT_DELACK))
1670                 tcp_timer_activate(tp, TT_DELACK, 0);
1671 #if 0
1672         /*
1673          * This completely breaks TCP if newreno is turned on.  What happens
1674          * is that if delayed-acks are turned on on the receiver, this code
1675          * on the transmitter effectively destroys the TCP window, forcing
1676          * it to four packets (1.5Kx4 = 6K window).
1677          */
1678         if (sendalot && --maxburst)
1679                 goto again;
1680 #endif
1681         if (sendalot)
1682                 goto again;
1683         return (0);
1684 }
1685
1686 void
1687 tcp_setpersist(struct tcpcb *tp)
1688 {
1689         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1690         int tt;
1691
1692         tp->t_flags &= ~TF_PREVVALID;
1693         if (tcp_timer_active(tp, TT_REXMT))
1694                 panic("tcp_setpersist: retransmit pending");
1695         /*
1696          * Start/restart persistence timer.
1697          */
1698         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1699                       tcp_persmin, tcp_persmax);
1700         tcp_timer_activate(tp, TT_PERSIST, tt);
1701         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1702                 tp->t_rxtshift++;
1703 }
1704
1705 /*
1706  * Insert TCP options according to the supplied parameters to the place
1707  * optp in a consistent way.  Can handle unaligned destinations.
1708  *
1709  * The order of the option processing is crucial for optimal packing and
1710  * alignment for the scarce option space.
1711  *
1712  * The optimal order for a SYN/SYN-ACK segment is:
1713  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1714  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1715  *
1716  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1717  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1718  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1719  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1720  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1721  */
1722 int
1723 tcp_addoptions(struct tcpopt *to, u_char *optp)
1724 {
1725         u_int32_t mask, optlen = 0;
1726
1727         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1728                 if ((to->to_flags & mask) != mask)
1729                         continue;
1730                 if (optlen == TCP_MAXOLEN)
1731                         break;
1732                 switch (to->to_flags & mask) {
1733                 case TOF_MSS:
1734                         while (optlen % 4) {
1735                                 optlen += TCPOLEN_NOP;
1736                                 *optp++ = TCPOPT_NOP;
1737                         }
1738                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1739                                 continue;
1740                         optlen += TCPOLEN_MAXSEG;
1741                         *optp++ = TCPOPT_MAXSEG;
1742                         *optp++ = TCPOLEN_MAXSEG;
1743                         to->to_mss = htons(to->to_mss);
1744                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1745                         optp += sizeof(to->to_mss);
1746                         break;
1747                 case TOF_SCALE:
1748                         while (!optlen || optlen % 2 != 1) {
1749                                 optlen += TCPOLEN_NOP;
1750                                 *optp++ = TCPOPT_NOP;
1751                         }
1752                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1753                                 continue;
1754                         optlen += TCPOLEN_WINDOW;
1755                         *optp++ = TCPOPT_WINDOW;
1756                         *optp++ = TCPOLEN_WINDOW;
1757                         *optp++ = to->to_wscale;
1758                         break;
1759                 case TOF_SACKPERM:
1760                         while (optlen % 2) {
1761                                 optlen += TCPOLEN_NOP;
1762                                 *optp++ = TCPOPT_NOP;
1763                         }
1764                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1765                                 continue;
1766                         optlen += TCPOLEN_SACK_PERMITTED;
1767                         *optp++ = TCPOPT_SACK_PERMITTED;
1768                         *optp++ = TCPOLEN_SACK_PERMITTED;
1769                         break;
1770                 case TOF_TS:
1771                         while (!optlen || optlen % 4 != 2) {
1772                                 optlen += TCPOLEN_NOP;
1773                                 *optp++ = TCPOPT_NOP;
1774                         }
1775                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1776                                 continue;
1777                         optlen += TCPOLEN_TIMESTAMP;
1778                         *optp++ = TCPOPT_TIMESTAMP;
1779                         *optp++ = TCPOLEN_TIMESTAMP;
1780                         to->to_tsval = htonl(to->to_tsval);
1781                         to->to_tsecr = htonl(to->to_tsecr);
1782                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1783                         optp += sizeof(to->to_tsval);
1784                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1785                         optp += sizeof(to->to_tsecr);
1786                         break;
1787                 case TOF_SIGNATURE:
1788                         {
1789                         int siglen = TCPOLEN_SIGNATURE - 2;
1790
1791                         while (!optlen || optlen % 4 != 2) {
1792                                 optlen += TCPOLEN_NOP;
1793                                 *optp++ = TCPOPT_NOP;
1794                         }
1795                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1796                                 to->to_flags &= ~TOF_SIGNATURE;
1797                                 continue;
1798                         }
1799                         optlen += TCPOLEN_SIGNATURE;
1800                         *optp++ = TCPOPT_SIGNATURE;
1801                         *optp++ = TCPOLEN_SIGNATURE;
1802                         to->to_signature = optp;
1803                         while (siglen--)
1804                                  *optp++ = 0;
1805                         break;
1806                         }
1807                 case TOF_SACK:
1808                         {
1809                         int sackblks = 0;
1810                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
1811                         tcp_seq sack_seq;
1812
1813                         while (!optlen || optlen % 4 != 2) {
1814                                 optlen += TCPOLEN_NOP;
1815                                 *optp++ = TCPOPT_NOP;
1816                         }
1817                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1818                                 continue;
1819                         optlen += TCPOLEN_SACKHDR;
1820                         *optp++ = TCPOPT_SACK;
1821                         sackblks = min(to->to_nsacks,
1822                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1823                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1824                         while (sackblks--) {
1825                                 sack_seq = htonl(sack->start);
1826                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1827                                 optp += sizeof(sack_seq);
1828                                 sack_seq = htonl(sack->end);
1829                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1830                                 optp += sizeof(sack_seq);
1831                                 optlen += TCPOLEN_SACK;
1832                                 sack++;
1833                         }
1834                         TCPSTAT_INC(tcps_sack_send_blocks);
1835                         break;
1836                         }
1837                 case TOF_FASTOPEN:
1838                         {
1839                         int total_len;
1840
1841                         /* XXX is there any point to aligning this option? */
1842                         total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1843                         if (TCP_MAXOLEN - optlen < total_len) {
1844                                 to->to_flags &= ~TOF_FASTOPEN;
1845                                 continue;
1846                         }
1847                         *optp++ = TCPOPT_FAST_OPEN;
1848                         *optp++ = total_len;
1849                         if (to->to_tfo_len > 0) {
1850                                 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1851                                 optp += to->to_tfo_len;
1852                         }
1853                         optlen += total_len;
1854                         break;
1855                         }
1856                 default:
1857                         panic("%s: unknown TCP option type", __func__);
1858                         break;
1859                 }
1860         }
1861
1862         /* Terminate and pad TCP options to a 4 byte boundary. */
1863         if (optlen % 4) {
1864                 optlen += TCPOLEN_EOL;
1865                 *optp++ = TCPOPT_EOL;
1866         }
1867         /*
1868          * According to RFC 793 (STD0007):
1869          *   "The content of the header beyond the End-of-Option option
1870          *    must be header padding (i.e., zero)."
1871          *   and later: "The padding is composed of zeros."
1872          */
1873         while (optlen % 4) {
1874                 optlen += TCPOLEN_PAD;
1875                 *optp++ = TCPOPT_PAD;
1876         }
1877
1878         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1879         return (optlen);
1880 }
1881
1882 /*
1883  * This is a copy of m_copym(), taking the TSO segment size/limit
1884  * constraints into account, and advancing the sndptr as it goes.
1885  */
1886 struct mbuf *
1887 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1888     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1889 {
1890 #ifdef KERN_TLS
1891         struct ktls_session *tls, *ntls;
1892         struct mbuf *start;
1893 #endif
1894         struct mbuf *n, **np;
1895         struct mbuf *top;
1896         int32_t off = off0;
1897         int32_t len = *plen;
1898         int32_t fragsize;
1899         int32_t len_cp = 0;
1900         int32_t *pkthdrlen;
1901         uint32_t mlen, frags;
1902         bool copyhdr;
1903
1904
1905         KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1906         KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1907         if (off == 0 && m->m_flags & M_PKTHDR)
1908                 copyhdr = true;
1909         else
1910                 copyhdr = false;
1911         while (off > 0) {
1912                 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1913                 if (off < m->m_len)
1914                         break;
1915                 off -= m->m_len;
1916                 if ((sb) && (m == sb->sb_sndptr)) {
1917                         sb->sb_sndptroff += m->m_len;
1918                         sb->sb_sndptr = m->m_next;
1919                 }
1920                 m = m->m_next;
1921         }
1922         np = &top;
1923         top = NULL;
1924         pkthdrlen = NULL;
1925 #ifdef KERN_TLS
1926         if (hw_tls && (m->m_flags & M_EXTPG))
1927                 tls = m->m_epg_tls;
1928         else
1929                 tls = NULL;
1930         start = m;
1931 #endif
1932         while (len > 0) {
1933                 if (m == NULL) {
1934                         KASSERT(len == M_COPYALL,
1935                             ("tcp_m_copym, length > size of mbuf chain"));
1936                         *plen = len_cp;
1937                         if (pkthdrlen != NULL)
1938                                 *pkthdrlen = len_cp;
1939                         break;
1940                 }
1941 #ifdef KERN_TLS
1942                 if (hw_tls) {
1943                         if (m->m_flags & M_EXTPG)
1944                                 ntls = m->m_epg_tls;
1945                         else
1946                                 ntls = NULL;
1947
1948                         /*
1949                          * Avoid mixing TLS records with handshake
1950                          * data or TLS records from different
1951                          * sessions.
1952                          */
1953                         if (tls != ntls) {
1954                                 MPASS(m != start);
1955                                 *plen = len_cp;
1956                                 if (pkthdrlen != NULL)
1957                                         *pkthdrlen = len_cp;
1958                                 break;
1959                         }
1960
1961                         /*
1962                          * Don't end a send in the middle of a TLS
1963                          * record if it spans multiple TLS records.
1964                          */
1965                         if (tls != NULL && (m != start) && len < m->m_len) {
1966                                 *plen = len_cp;
1967                                 if (pkthdrlen != NULL)
1968                                         *pkthdrlen = len_cp;
1969                                 break;
1970                         }
1971                 }
1972 #endif
1973                 mlen = min(len, m->m_len - off);
1974                 if (seglimit) {
1975                         /*
1976                          * For M_EXTPG mbufs, add 3 segments
1977                          * + 1 in case we are crossing page boundaries
1978                          * + 2 in case the TLS hdr/trailer are used
1979                          * It is cheaper to just add the segments
1980                          * than it is to take the cache miss to look
1981                          * at the mbuf ext_pgs state in detail.
1982                          */
1983                         if (m->m_flags & M_EXTPG) {
1984                                 fragsize = min(segsize, PAGE_SIZE);
1985                                 frags = 3;
1986                         } else {
1987                                 fragsize = segsize;
1988                                 frags = 0;
1989                         }
1990
1991                         /* Break if we really can't fit anymore. */
1992                         if ((frags + 1) >= seglimit) {
1993                                 *plen = len_cp;
1994                                 if (pkthdrlen != NULL)
1995                                         *pkthdrlen = len_cp;
1996                                 break;
1997                         }
1998
1999                         /*
2000                          * Reduce size if you can't copy the whole
2001                          * mbuf. If we can't copy the whole mbuf, also
2002                          * adjust len so the loop will end after this
2003                          * mbuf.
2004                          */
2005                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2006                                 mlen = (seglimit - frags - 1) * fragsize;
2007                                 len = mlen;
2008                                 *plen = len_cp + len;
2009                                 if (pkthdrlen != NULL)
2010                                         *pkthdrlen = *plen;
2011                         }
2012                         frags += howmany(mlen, fragsize);
2013                         if (frags == 0)
2014                                 frags++;
2015                         seglimit -= frags;
2016                         KASSERT(seglimit > 0,
2017                             ("%s: seglimit went too low", __func__));
2018                 }
2019                 if (copyhdr)
2020                         n = m_gethdr(M_NOWAIT, m->m_type);
2021                 else
2022                         n = m_get(M_NOWAIT, m->m_type);
2023                 *np = n;
2024                 if (n == NULL)
2025                         goto nospace;
2026                 if (copyhdr) {
2027                         if (!m_dup_pkthdr(n, m, M_NOWAIT))
2028                                 goto nospace;
2029                         if (len == M_COPYALL)
2030                                 n->m_pkthdr.len -= off0;
2031                         else
2032                                 n->m_pkthdr.len = len;
2033                         pkthdrlen = &n->m_pkthdr.len;
2034                         copyhdr = false;
2035                 }
2036                 n->m_len = mlen;
2037                 len_cp += n->m_len;
2038                 if (m->m_flags & (M_EXT|M_EXTPG)) {
2039                         n->m_data = m->m_data + off;
2040                         mb_dupcl(n, m);
2041                 } else
2042                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2043                             (u_int)n->m_len);
2044
2045                 if (sb && (sb->sb_sndptr == m) &&
2046                     ((n->m_len + off) >= m->m_len) && m->m_next) {
2047                         sb->sb_sndptroff += m->m_len;
2048                         sb->sb_sndptr = m->m_next;
2049                 }
2050                 off = 0;
2051                 if (len != M_COPYALL) {
2052                         len -= n->m_len;
2053                 }
2054                 m = m->m_next;
2055                 np = &n->m_next;
2056         }
2057         return (top);
2058 nospace:
2059         m_freem(top);
2060         return (NULL);
2061 }
2062
2063 void
2064 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2065 {
2066
2067         /*
2068          * Automatic sizing of send socket buffer.  Often the send buffer
2069          * size is not optimally adjusted to the actual network conditions
2070          * at hand (delay bandwidth product).  Setting the buffer size too
2071          * small limits throughput on links with high bandwidth and high
2072          * delay (eg. trans-continental/oceanic links).  Setting the
2073          * buffer size too big consumes too much real kernel memory,
2074          * especially with many connections on busy servers.
2075          *
2076          * The criteria to step up the send buffer one notch are:
2077          *  1. receive window of remote host is larger than send buffer
2078          *     (with a fudge factor of 5/4th);
2079          *  2. send buffer is filled to 7/8th with data (so we actually
2080          *     have data to make use of it);
2081          *  3. send buffer fill has not hit maximal automatic size;
2082          *  4. our send window (slow start and cogestion controlled) is
2083          *     larger than sent but unacknowledged data in send buffer.
2084          *
2085          * The remote host receive window scaling factor may limit the
2086          * growing of the send buffer before it reaches its allowed
2087          * maximum.
2088          *
2089          * It scales directly with slow start or congestion window
2090          * and does at most one step per received ACK.  This fast
2091          * scaling has the drawback of growing the send buffer beyond
2092          * what is strictly necessary to make full use of a given
2093          * delay*bandwidth product.  However testing has shown this not
2094          * to be much of an problem.  At worst we are trading wasting
2095          * of available bandwidth (the non-use of it) for wasting some
2096          * socket buffer memory.
2097          *
2098          * TODO: Shrink send buffer during idle periods together
2099          * with congestion window.  Requires another timer.  Has to
2100          * wait for upcoming tcp timer rewrite.
2101          *
2102          * XXXGL: should there be used sbused() or sbavail()?
2103          */
2104         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2105                 int lowat;
2106
2107                 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2108                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2109                     sbused(&so->so_snd) >=
2110                     (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2111                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2112                     sendwin >= (sbused(&so->so_snd) -
2113                     (tp->snd_nxt - tp->snd_una))) {
2114                         if (!sbreserve_locked(&so->so_snd,
2115                             min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2116                              V_tcp_autosndbuf_max), so, curthread))
2117                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2118                 }
2119         }
2120 }