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                 cc_after_idle(tp);
 265         tp->t_flags &= ~TF_LASTIDLE;
 266         if (idle) {
 267                 if (tp->t_flags & TF_MORETOCOME) {
 268                         tp->t_flags |= TF_LASTIDLE;
 269                         idle = 0;
 270                 }
 271         }
 272 again:
 273         /*
 274          * If we've recently taken a timeout, snd_max will be greater than
 275          * snd_nxt.  There may be SACK information that allows us to avoid
 276          * resending already delivered data.  Adjust snd_nxt accordingly.
 277          */
 278         if ((tp->t_flags & TF_SACK_PERMIT) &&
 279             SEQ_LT(tp->snd_nxt, tp->snd_max))
 280                 tcp_sack_adjust(tp);
 281         sendalot = 0;
 282         tso = 0;
 283         mtu = 0;
 284         off = tp->snd_nxt - tp->snd_una;
 285         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
 286
 287         flags = tcp_outflags[tp->t_state];
 288         /*
 289          * Send any SACK-generated retransmissions.  If we're explicitly trying
 290          * to send out new data (when sendalot is 1), bypass this function.
 291          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
 292          * we're replacing a (future) new transmission with a retransmission
 293          * now, and we previously incremented snd_cwnd in tcp_input().
 294          */
 295         /*
 296          * Still in sack recovery , reset rxmit flag to zero.
 297          */
 298         sack_rxmit = 0;
 299         sack_bytes_rxmt = 0;
 300         len = 0;
 301         p = NULL;
 302         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
 303             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
 304                 uint32_t cwin;
 305
 306                 cwin =
 307                     imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
 308                 /* Do not retransmit SACK segments beyond snd_recover */
 309                 if (SEQ_GT(p->end, tp->snd_recover)) {
 310                         /*
 311                          * (At least) part of sack hole extends beyond
 312                          * snd_recover. Check to see if we can rexmit data
 313                          * for this hole.
 314                          */
 315                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
 316                                 /*
 317                                  * Can't rexmit any more data for this hole.
 318                                  * That data will be rexmitted in the next
 319                                  * sack recovery episode, when snd_recover
 320                                  * moves past p->rxmit.
 321                                  */
 322                                 p = NULL;
 323                                 goto after_sack_rexmit;
 324                         } else
 325                                 /* Can rexmit part of the current hole */
 326                                 len = ((int32_t)ulmin(cwin,
 327                                                    tp->snd_recover - p->rxmit));
 328                 } else
 329                         len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
 330                 off = p->rxmit - tp->snd_una;
 331                 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
 332                     __func__, off));
 333                 if (len > 0) {
 334                         sack_rxmit = 1;
 335                         sendalot = 1;
 336                         TCPSTAT_INC(tcps_sack_rexmits);
 337                         TCPSTAT_ADD(tcps_sack_rexmit_bytes,
 338                             min(len, tp->t_maxseg));
 339                 }
 340         }
 341 after_sack_rexmit:
 342         /*
 343          * Get standard flags, and add SYN or FIN if requested by 'hidden'
 344          * state flags.
 345          */
 346         if (tp->t_flags & TF_NEEDFIN)
 347                 flags |= TH_FIN;
 348         if (tp->t_flags & TF_NEEDSYN)
 349                 flags |= TH_SYN;
 350
 351         SOCKBUF_LOCK(&so->so_snd);
 352         /*
 353          * If in persist timeout with window of 0, send 1 byte.
 354          * Otherwise, if window is small but nonzero
 355          * and timer expired, we will send what we can
 356          * and go to transmit state.
 357          */
 358         if (tp->t_flags & TF_FORCEDATA) {
 359                 if (sendwin == 0) {
 360                         /*
 361                          * If we still have some data to send, then
 362                          * clear the FIN bit.  Usually this would
 363                          * happen below when it realizes that we
 364                          * aren't sending all the data.  However,
 365                          * if we have exactly 1 byte of unsent data,
 366                          * then it won't clear the FIN bit below,
 367                          * and if we are in persist state, we wind
 368                          * up sending the packet without recording
 369                          * that we sent the FIN bit.
 370                          *
 371                          * We can't just blindly clear the FIN bit,
 372                          * because if we don't have any more data
 373                          * to send then the probe will be the FIN
 374                          * itself.
 375                          */
 376                         if (off < sbused(&so->so_snd))
 377                                 flags &= ~TH_FIN;
 378                         sendwin = 1;
 379                 } else {
 380                         tcp_timer_activate(tp, TT_PERSIST, 0);
 381                         tp->t_rxtshift = 0;
 382                 }
 383         }
 384
 385         /*
 386          * If snd_nxt == snd_max and we have transmitted a FIN, the
 387          * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
 388          * a negative length.  This can also occur when TCP opens up
 389          * its congestion window while receiving additional duplicate
 390          * acks after fast-retransmit because TCP will reset snd_nxt
 391          * to snd_max after the fast-retransmit.
 392          *
 393          * In the normal retransmit-FIN-only case, however, snd_nxt will
 394          * be set to snd_una, the offset will be 0, and the length may
 395          * wind up 0.
 396          *
 397          * If sack_rxmit is true we are retransmitting from the scoreboard
 398          * in which case len is already set.
 399          */
 400         if (sack_rxmit == 0) {
 401                 if (sack_bytes_rxmt == 0)
 402                         len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
 403                             off);
 404                 else {
 405                         int32_t cwin;
 406
 407                         /*
 408                          * We are inside of a SACK recovery episode and are
 409                          * sending new data, having retransmitted all the
 410                          * data possible in the scoreboard.
 411                          */
 412                         len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
 413                             off);
 414                         /*
 415                          * Don't remove this (len > 0) check !
 416                          * We explicitly check for len > 0 here (although it
 417                          * isn't really necessary), to work around a gcc
 418                          * optimization issue - to force gcc to compute
 419                          * len above. Without this check, the computation
 420                          * of len is bungled by the optimizer.
 421                          */
 422                         if (len > 0) {
 423                                 cwin = tp->snd_cwnd -
 424                                         (tp->snd_nxt - tp->snd_recover) -
 425                                         sack_bytes_rxmt;
 426                                 if (cwin < 0)
 427                                         cwin = 0;
 428                                 len = imin(len, cwin);
 429                         }
 430                 }
 431         }
 432
 433         /*
 434          * Lop off SYN bit if it has already been sent.  However, if this
 435          * is SYN-SENT state and if segment contains data and if we don't
 436          * know that foreign host supports TAO, suppress sending segment.
 437          */
 438         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
 439                 if (tp->t_state != TCPS_SYN_RECEIVED)
 440                         flags &= ~TH_SYN;
 441                 /*
 442                  * When sending additional segments following a TFO SYN|ACK,
 443                  * do not include the SYN bit.
 444                  */
 445                 if (IS_FASTOPEN(tp->t_flags) &&
 446                     (tp->t_state == TCPS_SYN_RECEIVED))
 447                         flags &= ~TH_SYN;
 448                 off--, len++;
 449         }
 450
 451         /*
 452          * Be careful not to send data and/or FIN on SYN segments.
 453          * This measure is needed to prevent interoperability problems
 454          * with not fully conformant TCP implementations.
 455          */
 456         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
 457                 len = 0;
 458                 flags &= ~TH_FIN;
 459         }
 460
 461         /*
 462          * On TFO sockets, ensure no data is sent in the following cases:
 463          *
 464          *  - When retransmitting SYN|ACK on a passively-created socket
 465          *
 466          *  - When retransmitting SYN on an actively created socket
 467          *
 468          *  - When sending a zero-length cookie (cookie request) on an
 469          *    actively created socket
 470          *
 471          *  - When the socket is in the CLOSED state (RST is being sent)
 472          */
 473         if (IS_FASTOPEN(tp->t_flags) &&
 474             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
 475              ((tp->t_state == TCPS_SYN_SENT) &&
 476               (tp->t_tfo_client_cookie_len == 0)) ||
 477              (flags & TH_RST)))
 478                 len = 0;
 479         if (len <= 0) {
 480                 /*
 481                  * If FIN has been sent but not acked,
 482                  * but we haven't been called to retransmit,
 483                  * len will be < 0.  Otherwise, window shrank
 484                  * after we sent into it.  If window shrank to 0,
 485                  * cancel pending retransmit, pull snd_nxt back
 486                  * to (closed) window, and set the persist timer
 487                  * if it isn't already going.  If the window didn't
 488                  * close completely, just wait for an ACK.
 489                  *
 490                  * We also do a general check here to ensure that
 491                  * we will set the persist timer when we have data
 492                  * to send, but a 0-byte window. This makes sure
 493                  * the persist timer is set even if the packet
 494                  * hits one of the "goto send" lines below.
 495                  */
 496                 len = 0;
 497                 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
 498                         (off < (int) sbavail(&so->so_snd))) {
 499                         tcp_timer_activate(tp, TT_REXMT, 0);
 500                         tp->t_rxtshift = 0;
 501                         tp->snd_nxt = tp->snd_una;
 502                         if (!tcp_timer_active(tp, TT_PERSIST))
 503                                 tcp_setpersist(tp);
 504                 }
 505         }
 506
 507         /* len will be >= 0 after this point. */
 508         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 509
 510         tcp_sndbuf_autoscale(tp, so, sendwin);
 511
 512         /*
 513          * Decide if we can use TCP Segmentation Offloading (if supported by
 514          * hardware).
 515          *
 516          * TSO may only be used if we are in a pure bulk sending state.  The
 517          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
 518          * IP options prevent using TSO.  With TSO the TCP header is the same
 519          * (except for the sequence number) for all generated packets.  This
 520          * makes it impossible to transmit any options which vary per generated
 521          * segment or packet.
 522          *
 523          * IPv4 handling has a clear separation of ip options and ip header
 524          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
 525          * the right thing below to provide length of just ip options and thus
 526          * checking for ipoptlen is enough to decide if ip options are present.
 527          */
 528 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 529         /*
 530          * Pre-calculate here as we save another lookup into the darknesses
 531          * of IPsec that way and can actually decide if TSO is ok.
 532          */
 533 #ifdef INET6
 534         if (isipv6 && IPSEC_ENABLED(ipv6))
 535                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
 536 #ifdef INET
 537         else
 538 #endif
 539 #endif /* INET6 */
 540 #ifdef INET
 541         if (IPSEC_ENABLED(ipv4))
 542                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
 543 #endif /* INET */
 544 #endif /* IPSEC */
 545 #ifdef INET6
 546         if (isipv6)
 547                 ipoptlen = ip6_optlen(tp->t_inpcb);
 548         else
 549 #endif
 550         if (tp->t_inpcb->inp_options)
 551                 ipoptlen = tp->t_inpcb->inp_options->m_len -
 552                                 offsetof(struct ipoption, ipopt_list);
 553         else
 554                 ipoptlen = 0;
 555 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 556         ipoptlen += ipsec_optlen;
 557 #endif
 558
 559         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
 560             ((tp->t_flags & TF_SIGNATURE) == 0) &&
 561             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
 562             ipoptlen == 0 && !(flags & TH_SYN))
 563                 tso = 1;
 564
 565         if (sack_rxmit) {
 566                 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
 567                         flags &= ~TH_FIN;
 568         } else {
 569                 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
 570                     sbused(&so->so_snd)))
 571                         flags &= ~TH_FIN;
 572         }
 573
 574         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 575             (long)TCP_MAXWIN << tp->rcv_scale);
 576
 577         /*
 578          * Sender silly window avoidance.   We transmit under the following
 579          * conditions when len is non-zero:
 580          *
 581          *      - We have a full segment (or more with TSO)
 582          *      - This is the last buffer in a write()/send() and we are
 583          *        either idle or running NODELAY
 584          *      - we've timed out (e.g. persist timer)
 585          *      - we have more then 1/2 the maximum send window's worth of
 586          *        data (receiver may be limited the window size)
 587          *      - we need to retransmit
 588          */
 589         if (len) {
 590                 if (len >= tp->t_maxseg)
 591                         goto send;
 592                 /*
 593                  * NOTE! on localhost connections an 'ack' from the remote
 594                  * end may occur synchronously with the output and cause
 595                  * us to flush a buffer queued with moretocome.  XXX
 596                  *
 597                  * note: the len + off check is almost certainly unnecessary.
 598                  */
 599                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
 600                     (idle || (tp->t_flags & TF_NODELAY)) &&
 601                     (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
 602                     (tp->t_flags & TF_NOPUSH) == 0) {
 603                         goto send;
 604                 }
 605                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
 606                         goto send;
 607                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
 608                         goto send;
 609                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
 610                         goto send;
 611                 if (sack_rxmit)
 612                         goto send;
 613         }
 614
 615         /*
 616          * Sending of standalone window updates.
 617          *
 618          * Window updates are important when we close our window due to a
 619          * full socket buffer and are opening it again after the application
 620          * reads data from it.  Once the window has opened again and the
 621          * remote end starts to send again the ACK clock takes over and
 622          * provides the most current window information.
 623          *
 624          * We must avoid the silly window syndrome whereas every read
 625          * from the receive buffer, no matter how small, causes a window
 626          * update to be sent.  We also should avoid sending a flurry of
 627          * window updates when the socket buffer had queued a lot of data
 628          * and the application is doing small reads.
 629          *
 630          * Prevent a flurry of pointless window updates by only sending
 631          * an update when we can increase the advertized window by more
 632          * than 1/4th of the socket buffer capacity.  When the buffer is
 633          * getting full or is very small be more aggressive and send an
 634          * update whenever we can increase by two mss sized segments.
 635          * In all other situations the ACK's to new incoming data will
 636          * carry further window increases.
 637          *
 638          * Don't send an independent window update if a delayed
 639          * ACK is pending (it will get piggy-backed on it) or the
 640          * remote side already has done a half-close and won't send
 641          * more data.  Skip this if the connection is in T/TCP
 642          * half-open state.
 643          */
 644         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
 645             !(tp->t_flags & TF_DELACK) &&
 646             !TCPS_HAVERCVDFIN(tp->t_state)) {
 647                 /*
 648                  * "adv" is the amount we could increase the window,
 649                  * taking into account that we are limited by
 650                  * TCP_MAXWIN << tp->rcv_scale.
 651                  */
 652                 int32_t adv;
 653                 int oldwin;
 654
 655                 adv = recwin;
 656                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
 657                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
 658                         if (adv > oldwin)
 659                                 adv -= oldwin;
 660                         else
 661                                 adv = 0;
 662                 } else
 663                         oldwin = 0;
 664
 665                 /*
 666                  * If the new window size ends up being the same as or less
 667                  * than the old size when it is scaled, then don't force
 668                  * a window update.
 669                  */
 670                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
 671                         goto dontupdate;
 672
 673                 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
 674                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
 675                      recwin <= (so->so_rcv.sb_hiwat / 8) ||
 676                      so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
 677                      adv >= TCP_MAXWIN << tp->rcv_scale))
 678                         goto send;
 679                 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
 680                         goto send;
 681         }
 682 dontupdate:
 683
 684         /*
 685          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
 686          * is also a catch-all for the retransmit timer timeout case.
 687          */
 688         if (tp->t_flags & TF_ACKNOW)
 689                 goto send;
 690         if ((flags & TH_RST) ||
 691             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
 692                 goto send;
 693         if (SEQ_GT(tp->snd_up, tp->snd_una))
 694                 goto send;
 695         /*
 696          * If our state indicates that FIN should be sent
 697          * and we have not yet done so, then we need to send.
 698          */
 699         if (flags & TH_FIN &&
 700             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
 701                 goto send;
 702         /*
 703          * In SACK, it is possible for tcp_output to fail to send a segment
 704          * after the retransmission timer has been turned off.  Make sure
 705          * that the retransmission timer is set.
 706          */
 707         if ((tp->t_flags & TF_SACK_PERMIT) &&
 708             SEQ_GT(tp->snd_max, tp->snd_una) &&
 709             !tcp_timer_active(tp, TT_REXMT) &&
 710             !tcp_timer_active(tp, TT_PERSIST)) {
 711                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 712                 goto just_return;
 713         }
 714         /*
 715          * TCP window updates are not reliable, rather a polling protocol
 716          * using ``persist'' packets is used to insure receipt of window
 717          * updates.  The three ``states'' for the output side are:
 718          *      idle                    not doing retransmits or persists
 719          *      persisting              to move a small or zero window
 720          *      (re)transmitting        and thereby not persisting
 721          *
 722          * tcp_timer_active(tp, TT_PERSIST)
 723          *      is true when we are in persist state.
 724          * (tp->t_flags & TF_FORCEDATA)
 725          *      is set when we are called to send a persist packet.
 726          * tcp_timer_active(tp, TT_REXMT)
 727          *      is set when we are retransmitting
 728          * The output side is idle when both timers are zero.
 729          *
 730          * If send window is too small, there is data to transmit, and no
 731          * retransmit or persist is pending, then go to persist state.
 732          * If nothing happens soon, send when timer expires:
 733          * if window is nonzero, transmit what we can,
 734          * otherwise force out a byte.
 735          */
 736         if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
 737             !tcp_timer_active(tp, TT_PERSIST)) {
 738                 tp->t_rxtshift = 0;
 739                 tcp_setpersist(tp);
 740         }
 741
 742         /*
 743          * No reason to send a segment, just return.
 744          */
 745 just_return:
 746         SOCKBUF_UNLOCK(&so->so_snd);
 747         return (0);
 748
 749 send:
 750         SOCKBUF_LOCK_ASSERT(&so->so_snd);
 751         if (len > 0) {
 752                 if (len >= tp->t_maxseg)
 753                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
 754                 else
 755                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
 756         }
 757         /*
 758          * Before ESTABLISHED, force sending of initial options
 759          * unless TCP set not to do any options.
 760          * NOTE: we assume that the IP/TCP header plus TCP options
 761          * always fit in a single mbuf, leaving room for a maximum
 762          * link header, i.e.
 763          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
 764          */
 765         optlen = 0;
 766 #ifdef INET6
 767         if (isipv6)
 768                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
 769         else
 770 #endif
 771                 hdrlen = sizeof (struct tcpiphdr);
 772
 773         /*
 774          * Compute options for segment.
 775          * We only have to care about SYN and established connection
 776          * segments.  Options for SYN-ACK segments are handled in TCP
 777          * syncache.
 778          */
 779         to.to_flags = 0;
 780         if ((tp->t_flags & TF_NOOPT) == 0) {
 781                 /* Maximum segment size. */
 782                 if (flags & TH_SYN) {
 783                         tp->snd_nxt = tp->iss;
 784                         to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
 785                         to.to_flags |= TOF_MSS;
 786
 787                         /*
 788                          * On SYN or SYN|ACK transmits on TFO connections,
 789                          * only include the TFO option if it is not a
 790                          * retransmit, as the presence of the TFO option may
 791                          * have caused the original SYN or SYN|ACK to have
 792                          * been dropped by a middlebox.
 793                          */
 794                         if (IS_FASTOPEN(tp->t_flags) &&
 795                             (tp->t_rxtshift == 0)) {
 796                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
 797                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 798                                         to.to_tfo_cookie =
 799                                             (u_int8_t *)&tp->t_tfo_cookie.server;
 800                                         to.to_flags |= TOF_FASTOPEN;
 801                                         wanted_cookie = 1;
 802                                 } else if (tp->t_state == TCPS_SYN_SENT) {
 803                                         to.to_tfo_len =
 804                                             tp->t_tfo_client_cookie_len;
 805                                         to.to_tfo_cookie =
 806                                             tp->t_tfo_cookie.client;
 807                                         to.to_flags |= TOF_FASTOPEN;
 808                                         wanted_cookie = 1;
 809                                         /*
 810                                          * If we wind up having more data to
 811                                          * send with the SYN than can fit in
 812                                          * one segment, don't send any more
 813                                          * until the SYN|ACK comes back from
 814                                          * the other end.
 815                                          */
 816                                         dont_sendalot = 1;
 817                                 }
 818                         }
 819                 }
 820                 /* Window scaling. */
 821                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
 822                         to.to_wscale = tp->request_r_scale;
 823                         to.to_flags |= TOF_SCALE;
 824                 }
 825                 /* Timestamps. */
 826                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
 827                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
 828                         curticks = tcp_ts_getticks();
 829                         to.to_tsval = curticks + tp->ts_offset;
 830                         to.to_tsecr = tp->ts_recent;
 831                         to.to_flags |= TOF_TS;
 832                         if (tp->t_rxtshift == 1)
 833                                 tp->t_badrxtwin = curticks;
 834                 }
 835
 836                 /* Set receive buffer autosizing timestamp. */
 837                 if (tp->rfbuf_ts == 0 &&
 838                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
 839                         tp->rfbuf_ts = tcp_ts_getticks();
 840
 841                 /* Selective ACK's. */
 842                 if (tp->t_flags & TF_SACK_PERMIT) {
 843                         if (flags & TH_SYN)
 844                                 to.to_flags |= TOF_SACKPERM;
 845                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
 846                             (tp->t_flags & TF_SACK_PERMIT) &&
 847                             tp->rcv_numsacks > 0) {
 848                                 to.to_flags |= TOF_SACK;
 849                                 to.to_nsacks = tp->rcv_numsacks;
 850                                 to.to_sacks = (u_char *)tp->sackblks;
 851                         }
 852                 }
 853 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 854                 /* TCP-MD5 (RFC2385). */
 855                 /*
 856                  * Check that TCP_MD5SIG is enabled in tcpcb to
 857                  * account the size needed to set this TCP option.
 858                  */
 859                 if (tp->t_flags & TF_SIGNATURE)
 860                         to.to_flags |= TOF_SIGNATURE;
 861 #endif /* TCP_SIGNATURE */
 862
 863                 /* Processing the options. */
 864                 hdrlen += optlen = tcp_addoptions(&to, opt);
 865                 /*
 866                  * If we wanted a TFO option to be added, but it was unable
 867                  * to fit, ensure no data is sent.
 868                  */
 869                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
 870                     !(to.to_flags & TOF_FASTOPEN))
 871                         len = 0;
 872         }
 873
 874         /*
 875          * Adjust data length if insertion of options will
 876          * bump the packet length beyond the t_maxseg length.
 877          * Clear the FIN bit because we cut off the tail of
 878          * the segment.
 879          */
 880         if (len + optlen + ipoptlen > tp->t_maxseg) {
 881                 flags &= ~TH_FIN;
 882
 883                 if (tso) {
 884                         u_int if_hw_tsomax;
 885                         u_int moff;
 886                         int max_len;
 887
 888                         /* extract TSO information */
 889                         if_hw_tsomax = tp->t_tsomax;
 890                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
 891                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
 892
 893                         /*
 894                          * Limit a TSO burst to prevent it from
 895                          * overflowing or exceeding the maximum length
 896                          * allowed by the network interface:
 897                          */
 898                         KASSERT(ipoptlen == 0,
 899                             ("%s: TSO can't do IP options", __func__));
 900
 901                         /*
 902                          * Check if we should limit by maximum payload
 903                          * length:
 904                          */
 905                         if (if_hw_tsomax != 0) {
 906                                 /* compute maximum TSO length */
 907                                 max_len = (if_hw_tsomax - hdrlen -
 908                                     max_linkhdr);
 909                                 if (max_len <= 0) {
 910                                         len = 0;
 911                                 } else if (len > max_len) {
 912                                         sendalot = 1;
 913                                         len = max_len;
 914                                 }
 915                         }
 916
 917                         /*
 918                          * Prevent the last segment from being
 919                          * fractional unless the send sockbuf can be
 920                          * emptied:
 921                          */
 922                         max_len = (tp->t_maxseg - optlen);
 923                         if (((uint32_t)off + (uint32_t)len) <
 924                             sbavail(&so->so_snd)) {
 925                                 moff = len % max_len;
 926                                 if (moff != 0) {
 927                                         len -= moff;
 928                                         sendalot = 1;
 929                                 }
 930                         }
 931
 932                         /*
 933                          * In case there are too many small fragments
 934                          * don't use TSO:
 935                          */
 936                         if (len <= max_len) {
 937                                 len = max_len;
 938                                 sendalot = 1;
 939                                 tso = 0;
 940                         }
 941
 942                         /*
 943                          * Send the FIN in a separate segment
 944                          * after the bulk sending is done.
 945                          * We don't trust the TSO implementations
 946                          * to clear the FIN flag on all but the
 947                          * last segment.
 948                          */
 949                         if (tp->t_flags & TF_NEEDFIN)
 950                                 sendalot = 1;
 951                 } else {
 952                         if (optlen + ipoptlen >= tp->t_maxseg) {
 953                                 /*
 954                                  * Since we don't have enough space to put
 955                                  * the IP header chain and the TCP header in
 956                                  * one packet as required by RFC 7112, don't
 957                                  * send it. Also ensure that at least one
 958                                  * byte of the payload can be put into the
 959                                  * TCP segment.
 960                                  */
 961                                 SOCKBUF_UNLOCK(&so->so_snd);
 962                                 error = EMSGSIZE;
 963                                 sack_rxmit = 0;
 964                                 goto out;
 965                         }
 966                         len = tp->t_maxseg - optlen - ipoptlen;
 967                         sendalot = 1;
 968                         if (dont_sendalot)
 969                                 sendalot = 0;
 970                 }
 971         } else
 972                 tso = 0;
 973
 974         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
 975             ("%s: len > IP_MAXPACKET", __func__));
 976
 977 /*#ifdef DIAGNOSTIC*/
 978 #ifdef INET6
 979         if (max_linkhdr + hdrlen > MCLBYTES)
 980 #else
 981         if (max_linkhdr + hdrlen > MHLEN)
 982 #endif
 983                 panic("tcphdr too big");
 984 /*#endif*/
 985
 986         /*
 987          * This KASSERT is here to catch edge cases at a well defined place.
 988          * Before, those had triggered (random) panic conditions further down.
 989          */
 990         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 991
 992         /*
 993          * Grab a header mbuf, attaching a copy of data to
 994          * be transmitted, and initialize the header from
 995          * the template for sends on this connection.
 996          */
 997         if (len) {
 998                 struct mbuf *mb;
 999                 struct sockbuf *msb;
1000                 u_int moff;
1001
1002                 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1003                         TCPSTAT_INC(tcps_sndprobe);
1004 #ifdef STATS
1005                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1006                                 stats_voi_update_abs_u32(tp->t_stats,
1007                                 VOI_TCP_RETXPB, len);
1008                         else
1009                                 stats_voi_update_abs_u64(tp->t_stats,
1010                                     VOI_TCP_TXPB, len);
1011 #endif /* STATS */
1012                 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1013                         tp->t_sndrexmitpack++;
1014                         TCPSTAT_INC(tcps_sndrexmitpack);
1015                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1016 #ifdef STATS
1017                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1018                             len);
1019 #endif /* STATS */
1020                 } else {
1021                         TCPSTAT_INC(tcps_sndpack);
1022                         TCPSTAT_ADD(tcps_sndbyte, len);
1023 #ifdef STATS
1024                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1025                             len);
1026 #endif /* STATS */
1027                 }
1028 #ifdef INET6
1029                 if (MHLEN < hdrlen + max_linkhdr)
1030                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1031                 else
1032 #endif
1033                         m = m_gethdr(M_NOWAIT, MT_DATA);
1034
1035                 if (m == NULL) {
1036                         SOCKBUF_UNLOCK(&so->so_snd);
1037                         error = ENOBUFS;
1038                         sack_rxmit = 0;
1039                         goto out;
1040                 }
1041
1042                 m->m_data += max_linkhdr;
1043                 m->m_len = hdrlen;
1044
1045                 /*
1046                  * Start the m_copy functions from the closest mbuf
1047                  * to the offset in the socket buffer chain.
1048                  */
1049                 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1050                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1051                         m_copydata(mb, moff, len,
1052                             mtod(m, caddr_t) + hdrlen);
1053                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1054                                 sbsndptr_adv(&so->so_snd, mb, len);
1055                         m->m_len += len;
1056                 } else {
1057                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1058                                 msb = NULL;
1059                         else
1060                                 msb = &so->so_snd;
1061                         m->m_next = tcp_m_copym(mb, moff,
1062                             &len, if_hw_tsomaxsegcount,
1063                             if_hw_tsomaxsegsize, msb, hw_tls);
1064                         if (len <= (tp->t_maxseg - optlen)) {
1065                                 /*
1066                                  * Must have ran out of mbufs for the copy
1067                                  * shorten it to no longer need tso. Lets
1068                                  * not put on sendalot since we are low on
1069                                  * mbufs.
1070                                  */
1071                                 tso = 0;
1072                         }
1073                         if (m->m_next == NULL) {
1074                                 SOCKBUF_UNLOCK(&so->so_snd);
1075                                 (void) m_free(m);
1076                                 error = ENOBUFS;
1077                                 sack_rxmit = 0;
1078                                 goto out;
1079                         }
1080                 }
1081
1082                 /*
1083                  * If we're sending everything we've got, set PUSH.
1084                  * (This will keep happy those implementations which only
1085                  * give data to the user when a buffer fills or
1086                  * a PUSH comes in.)
1087                  */
1088                 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1089                     !(flags & TH_SYN))
1090                         flags |= TH_PUSH;
1091                 SOCKBUF_UNLOCK(&so->so_snd);
1092         } else {
1093                 SOCKBUF_UNLOCK(&so->so_snd);
1094                 if (tp->t_flags & TF_ACKNOW)
1095                         TCPSTAT_INC(tcps_sndacks);
1096                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1097                         TCPSTAT_INC(tcps_sndctrl);
1098                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1099                         TCPSTAT_INC(tcps_sndurg);
1100                 else
1101                         TCPSTAT_INC(tcps_sndwinup);
1102
1103                 m = m_gethdr(M_NOWAIT, MT_DATA);
1104                 if (m == NULL) {
1105                         error = ENOBUFS;
1106                         sack_rxmit = 0;
1107                         goto out;
1108                 }
1109 #ifdef INET6
1110                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1111                     MHLEN >= hdrlen) {
1112                         M_ALIGN(m, hdrlen);
1113                 } else
1114 #endif
1115                 m->m_data += max_linkhdr;
1116                 m->m_len = hdrlen;
1117         }
1118         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1119         m->m_pkthdr.rcvif = (struct ifnet *)0;
1120 #ifdef MAC
1121         mac_inpcb_create_mbuf(tp->t_inpcb, m);
1122 #endif
1123 #ifdef INET6
1124         if (isipv6) {
1125                 ip6 = mtod(m, struct ip6_hdr *);
1126                 th = (struct tcphdr *)(ip6 + 1);
1127                 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1128         } else
1129 #endif /* INET6 */
1130         {
1131                 ip = mtod(m, struct ip *);
1132 #ifdef TCPDEBUG
1133                 ipov = (struct ipovly *)ip;
1134 #endif
1135                 th = (struct tcphdr *)(ip + 1);
1136                 tcpip_fillheaders(tp->t_inpcb, ip, th);
1137         }
1138
1139         /*
1140          * Fill in fields, remembering maximum advertised
1141          * window for use in delaying messages about window sizes.
1142          * If resending a FIN, be sure not to use a new sequence number.
1143          */
1144         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1145             tp->snd_nxt == tp->snd_max)
1146                 tp->snd_nxt--;
1147         /*
1148          * If we are starting a connection, send ECN setup
1149          * SYN packet. If we are on a retransmit, we may
1150          * resend those bits a number of times as per
1151          * RFC 3168.
1152          */
1153         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1154                 if (tp->t_rxtshift >= 1) {
1155                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1156                                 flags |= TH_ECE|TH_CWR;
1157                 } else
1158                         flags |= TH_ECE|TH_CWR;
1159         }
1160         /* Handle parallel SYN for ECN */
1161         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1162             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1163                         flags |= TH_ECE;
1164                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1165         }
1166
1167         if (tp->t_state == TCPS_ESTABLISHED &&
1168             (tp->t_flags2 & TF2_ECN_PERMIT)) {
1169                 /*
1170                  * If the peer has ECN, mark data packets with
1171                  * ECN capable transmission (ECT).
1172                  * Ignore pure ack packets, retransmissions and window probes.
1173                  */
1174                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1175                     (sack_rxmit == 0) &&
1176                     !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1177                     SEQ_LT(tp->snd_una, tp->snd_max))) {
1178 #ifdef INET6
1179                         if (isipv6)
1180                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1181                         else
1182 #endif
1183                                 ip->ip_tos |= IPTOS_ECN_ECT0;
1184                         TCPSTAT_INC(tcps_ecn_ect0);
1185                         /*
1186                          * Reply with proper ECN notifications.
1187                          * Only set CWR on new data segments.
1188                          */
1189                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1190                                 flags |= TH_CWR;
1191                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1192                         }
1193                 }
1194                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
1195                         flags |= TH_ECE;
1196         }
1197
1198         /*
1199          * If we are doing retransmissions, then snd_nxt will
1200          * not reflect the first unsent octet.  For ACK only
1201          * packets, we do not want the sequence number of the
1202          * retransmitted packet, we want the sequence number
1203          * of the next unsent octet.  So, if there is no data
1204          * (and no SYN or FIN), use snd_max instead of snd_nxt
1205          * when filling in ti_seq.  But if we are in persist
1206          * state, snd_max might reflect one byte beyond the
1207          * right edge of the window, so use snd_nxt in that
1208          * case, since we know we aren't doing a retransmission.
1209          * (retransmit and persist are mutually exclusive...)
1210          */
1211         if (sack_rxmit == 0) {
1212                 if (len || (flags & (TH_SYN|TH_FIN)) ||
1213                     tcp_timer_active(tp, TT_PERSIST))
1214                         th->th_seq = htonl(tp->snd_nxt);
1215                 else
1216                         th->th_seq = htonl(tp->snd_max);
1217         } else {
1218                 th->th_seq = htonl(p->rxmit);
1219                 p->rxmit += len;
1220                 tp->sackhint.sack_bytes_rexmit += len;
1221         }
1222         th->th_ack = htonl(tp->rcv_nxt);
1223         if (optlen) {
1224                 bcopy(opt, th + 1, optlen);
1225                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1226         }
1227         th->th_flags = flags;
1228         /*
1229          * Calculate receive window.  Don't shrink window,
1230          * but avoid silly window syndrome.
1231          * If a RST segment is sent, advertise a window of zero.
1232          */
1233         if (flags & TH_RST) {
1234                 recwin = 0;
1235         } else {
1236                 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1237                     recwin < tp->t_maxseg)
1238                         recwin = 0;
1239                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1240                     recwin < (tp->rcv_adv - tp->rcv_nxt))
1241                         recwin = (tp->rcv_adv - tp->rcv_nxt);
1242         }
1243         /*
1244          * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1245          * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1246          * case is handled in syncache.
1247          */
1248         if (flags & TH_SYN)
1249                 th->th_win = htons((u_short)
1250                                 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1251         else {
1252                 /* Avoid shrinking window with window scaling. */
1253                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1254                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1255         }
1256
1257         /*
1258          * Adjust the RXWIN0SENT flag - indicate that we have advertised
1259          * a 0 window.  This may cause the remote transmitter to stall.  This
1260          * flag tells soreceive() to disable delayed acknowledgements when
1261          * draining the buffer.  This can occur if the receiver is attempting
1262          * to read more data than can be buffered prior to transmitting on
1263          * the connection.
1264          */
1265         if (th->th_win == 0) {
1266                 tp->t_sndzerowin++;
1267                 tp->t_flags |= TF_RXWIN0SENT;
1268         } else
1269                 tp->t_flags &= ~TF_RXWIN0SENT;
1270         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1271                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1272                 th->th_flags |= TH_URG;
1273         } else
1274                 /*
1275                  * If no urgent pointer to send, then we pull
1276                  * the urgent pointer to the left edge of the send window
1277                  * so that it doesn't drift into the send window on sequence
1278                  * number wraparound.
1279                  */
1280                 tp->snd_up = tp->snd_una;               /* drag it along */
1281
1282         /*
1283          * Put TCP length in extended header, and then
1284          * checksum extended header and data.
1285          */
1286         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1287         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1288
1289 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1290         if (to.to_flags & TOF_SIGNATURE) {
1291                 /*
1292                  * Calculate MD5 signature and put it into the place
1293                  * determined before.
1294                  * NOTE: since TCP options buffer doesn't point into
1295                  * mbuf's data, calculate offset and use it.
1296                  */
1297                 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1298                     (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1299                         /*
1300                          * Do not send segment if the calculation of MD5
1301                          * digest has failed.
1302                          */
1303                         m_freem(m);
1304                         goto out;
1305                 }
1306         }
1307 #endif
1308 #ifdef INET6
1309         if (isipv6) {
1310                 /*
1311                  * There is no need to fill in ip6_plen right now.
1312                  * It will be filled later by ip6_output.
1313                  */
1314                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1315                 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1316                     optlen + len, IPPROTO_TCP, 0);
1317         }
1318 #endif
1319 #if defined(INET6) && defined(INET)
1320         else
1321 #endif
1322 #ifdef INET
1323         {
1324                 m->m_pkthdr.csum_flags = CSUM_TCP;
1325                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1326                     htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1327
1328                 /* IP version must be set here for ipv4/ipv6 checking later */
1329                 KASSERT(ip->ip_v == IPVERSION,
1330                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1331         }
1332 #endif
1333
1334         /*
1335          * Enable TSO and specify the size of the segments.
1336          * The TCP pseudo header checksum is always provided.
1337          */
1338         if (tso) {
1339                 KASSERT(len > tp->t_maxseg - optlen,
1340                     ("%s: len <= tso_segsz", __func__));
1341                 m->m_pkthdr.csum_flags |= CSUM_TSO;
1342                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1343         }
1344
1345         KASSERT(len + hdrlen == m_length(m, NULL),
1346             ("%s: mbuf chain shorter than expected: %d + %u != %u",
1347             __func__, len, hdrlen, m_length(m, NULL)));
1348
1349 #ifdef TCP_HHOOK
1350         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1351         hhook_run_tcp_est_out(tp, th, &to, len, tso);
1352 #endif
1353
1354 #ifdef TCPDEBUG
1355         /*
1356          * Trace.
1357          */
1358         if (so->so_options & SO_DEBUG) {
1359                 u_short save = 0;
1360 #ifdef INET6
1361                 if (!isipv6)
1362 #endif
1363                 {
1364                         save = ipov->ih_len;
1365                         ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1366                 }
1367                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1368 #ifdef INET6
1369                 if (!isipv6)
1370 #endif
1371                 ipov->ih_len = save;
1372         }
1373 #endif /* TCPDEBUG */
1374         TCP_PROBE3(debug__output, tp, th, m);
1375
1376         /* We're getting ready to send; log now. */
1377         TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1378             len, NULL, false);
1379
1380         /*
1381          * Fill in IP length and desired time to live and
1382          * send to IP level.  There should be a better way
1383          * to handle ttl and tos; we could keep them in
1384          * the template, but need a way to checksum without them.
1385          */
1386         /*
1387          * m->m_pkthdr.len should have been set before checksum calculation,
1388          * because in6_cksum() need it.
1389          */
1390 #ifdef INET6
1391         if (isipv6) {
1392                 /*
1393                  * we separately set hoplimit for every segment, since the
1394                  * user might want to change the value via setsockopt.
1395                  * Also, desired default hop limit might be changed via
1396                  * Neighbor Discovery.
1397                  */
1398                 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1399
1400                 /*
1401                  * Set the packet size here for the benefit of DTrace probes.
1402                  * ip6_output() will set it properly; it's supposed to include
1403                  * the option header lengths as well.
1404                  */
1405                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1406
1407                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1408                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1409                 else
1410                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1411
1412                 if (tp->t_state == TCPS_SYN_SENT)
1413                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1414
1415                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1416
1417 #ifdef TCPPCAP
1418                 /* Save packet, if requested. */
1419                 tcp_pcap_add(th, m, &(tp->t_outpkts));
1420 #endif
1421
1422                 /* TODO: IPv6 IP6TOS_ECT bit on */
1423                 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1424                     &tp->t_inpcb->inp_route6,
1425                     ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1426                     NULL, NULL, tp->t_inpcb);
1427
1428                 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1429                         mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1430         }
1431 #endif /* INET6 */
1432 #if defined(INET) && defined(INET6)
1433         else
1434 #endif
1435 #ifdef INET
1436     {
1437         ip->ip_len = htons(m->m_pkthdr.len);
1438 #ifdef INET6
1439         if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1440                 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1441 #endif /* INET6 */
1442         /*
1443          * If we do path MTU discovery, then we set DF on every packet.
1444          * This might not be the best thing to do according to RFC3390
1445          * Section 2. However the tcp hostcache migitates the problem
1446          * so it affects only the first tcp connection with a host.
1447          *
1448          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1449          */
1450         if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1451                 ip->ip_off |= htons(IP_DF);
1452                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1453         } else {
1454                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1455         }
1456
1457         if (tp->t_state == TCPS_SYN_SENT)
1458                 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1459
1460         TCP_PROBE5(send, NULL, tp, ip, tp, th);
1461
1462 #ifdef TCPPCAP
1463         /* Save packet, if requested. */
1464         tcp_pcap_add(th, m, &(tp->t_outpkts));
1465 #endif
1466
1467         error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1468             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1469             tp->t_inpcb);
1470
1471         if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1472                 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1473     }
1474 #endif /* INET */
1475
1476 out:
1477         /*
1478          * In transmit state, time the transmission and arrange for
1479          * the retransmit.  In persist state, just set snd_max.
1480          */
1481         if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1482             !tcp_timer_active(tp, TT_PERSIST)) {
1483                 tcp_seq startseq = tp->snd_nxt;
1484
1485                 /*
1486                  * Advance snd_nxt over sequence space of this segment.
1487                  */
1488                 if (flags & (TH_SYN|TH_FIN)) {
1489                         if (flags & TH_SYN)
1490                                 tp->snd_nxt++;
1491                         if (flags & TH_FIN) {
1492                                 tp->snd_nxt++;
1493                                 tp->t_flags |= TF_SENTFIN;
1494                         }
1495                 }
1496                 if (sack_rxmit)
1497                         goto timer;
1498                 tp->snd_nxt += len;
1499                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1500                         tp->snd_max = tp->snd_nxt;
1501                         /*
1502                          * Time this transmission if not a retransmission and
1503                          * not currently timing anything.
1504                          */
1505                         if (tp->t_rtttime == 0) {
1506                                 tp->t_rtttime = ticks;
1507                                 tp->t_rtseq = startseq;
1508                                 TCPSTAT_INC(tcps_segstimed);
1509                         }
1510 #ifdef STATS
1511                         if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1512                                 tp->t_flags |= TF_GPUTINPROG;
1513                                 tp->gput_seq = startseq;
1514                                 tp->gput_ack = startseq +
1515                                     ulmin(sbavail(&so->so_snd) - off, sendwin);
1516                                 tp->gput_ts = tcp_ts_getticks();
1517                         }
1518 #endif /* STATS */
1519                 }
1520
1521                 /*
1522                  * Set retransmit timer if not currently set,
1523                  * and not doing a pure ack or a keep-alive probe.
1524                  * Initial value for retransmit timer is smoothed
1525                  * round-trip time + 2 * round-trip time variance.
1526                  * Initialize shift counter which is used for backoff
1527                  * of retransmit time.
1528                  */
1529 timer:
1530                 if (!tcp_timer_active(tp, TT_REXMT) &&
1531                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1532                      (tp->snd_nxt != tp->snd_una))) {
1533                         if (tcp_timer_active(tp, TT_PERSIST)) {
1534                                 tcp_timer_activate(tp, TT_PERSIST, 0);
1535                                 tp->t_rxtshift = 0;
1536                         }
1537                         tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1538                 } else if (len == 0 && sbavail(&so->so_snd) &&
1539                     !tcp_timer_active(tp, TT_REXMT) &&
1540                     !tcp_timer_active(tp, TT_PERSIST)) {
1541                         /*
1542                          * Avoid a situation where we do not set persist timer
1543                          * after a zero window condition. For example:
1544                          * 1) A -> B: packet with enough data to fill the window
1545                          * 2) B -> A: ACK for #1 + new data (0 window
1546                          *    advertisement)
1547                          * 3) A -> B: ACK for #2, 0 len packet
1548                          *
1549                          * In this case, A will not activate the persist timer,
1550                          * because it chose to send a packet. Unless tcp_output
1551                          * is called for some other reason (delayed ack timer,
1552                          * another input packet from B, socket syscall), A will
1553                          * not send zero window probes.
1554                          *
1555                          * So, if you send a 0-length packet, but there is data
1556                          * in the socket buffer, and neither the rexmt or
1557                          * persist timer is already set, then activate the
1558                          * persist timer.
1559                          */
1560                         tp->t_rxtshift = 0;
1561                         tcp_setpersist(tp);
1562                 }
1563         } else {
1564                 /*
1565                  * Persist case, update snd_max but since we are in
1566                  * persist mode (no window) we do not update snd_nxt.
1567                  */
1568                 int xlen = len;
1569                 if (flags & TH_SYN)
1570                         ++xlen;
1571                 if (flags & TH_FIN) {
1572                         ++xlen;
1573                         tp->t_flags |= TF_SENTFIN;
1574                 }
1575                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1576                         tp->snd_max = tp->snd_nxt + xlen;
1577         }
1578         if ((error == 0) &&
1579             (TCPS_HAVEESTABLISHED(tp->t_state) &&
1580              (tp->t_flags & TF_SACK_PERMIT) &&
1581              tp->rcv_numsacks > 0)) {
1582                     /* Clean up any DSACK's sent */
1583                     tcp_clean_dsack_blocks(tp);
1584         }
1585         if (error) {
1586                 /* Record the error. */
1587                 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1588                     error, 0, NULL, false);
1589
1590                 /*
1591                  * We know that the packet was lost, so back out the
1592                  * sequence number advance, if any.
1593                  *
1594                  * If the error is EPERM the packet got blocked by the
1595                  * local firewall.  Normally we should terminate the
1596                  * connection but the blocking may have been spurious
1597                  * due to a firewall reconfiguration cycle.  So we treat
1598                  * it like a packet loss and let the retransmit timer and
1599                  * timeouts do their work over time.
1600                  * XXX: It is a POLA question whether calling tcp_drop right
1601                  * away would be the really correct behavior instead.
1602                  */
1603                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1604                     !tcp_timer_active(tp, TT_PERSIST)) &&
1605                     ((flags & TH_SYN) == 0) &&
1606                     (error != EPERM)) {
1607                         if (sack_rxmit) {
1608                                 p->rxmit -= len;
1609                                 tp->sackhint.sack_bytes_rexmit -= len;
1610                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1611                                     ("sackhint bytes rtx >= 0"));
1612                         } else
1613                                 tp->snd_nxt -= len;
1614                 }
1615                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
1616                 switch (error) {
1617                 case EACCES:
1618                 case EPERM:
1619                         tp->t_softerror = error;
1620                         return (error);
1621                 case ENOBUFS:
1622                         TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1623                         tp->snd_cwnd = tp->t_maxseg;
1624                         return (0);
1625                 case EMSGSIZE:
1626                         /*
1627                          * For some reason the interface we used initially
1628                          * to send segments changed to another or lowered
1629                          * its MTU.
1630                          * If TSO was active we either got an interface
1631                          * without TSO capabilits or TSO was turned off.
1632                          * If we obtained mtu from ip_output() then update
1633                          * it and try again.
1634                          */
1635                         if (tso)
1636                                 tp->t_flags &= ~TF_TSO;
1637                         if (mtu != 0) {
1638                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1639                                 goto again;
1640                         }
1641                         return (error);
1642                 case EHOSTDOWN:
1643                 case EHOSTUNREACH:
1644                 case ENETDOWN:
1645                 case ENETUNREACH:
1646                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
1647                                 tp->t_softerror = error;
1648                                 return (0);
1649                         }
1650                         /* FALLTHROUGH */
1651                 default:
1652                         return (error);
1653                 }
1654         }
1655         TCPSTAT_INC(tcps_sndtotal);
1656
1657         /*
1658          * Data sent (as far as we can tell).
1659          * If this advertises a larger window than any other segment,
1660          * then remember the size of the advertised window.
1661          * Any pending ACK has now been sent.
1662          */
1663         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1664                 tp->rcv_adv = tp->rcv_nxt + recwin;
1665         tp->last_ack_sent = tp->rcv_nxt;
1666         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1667         if (tcp_timer_active(tp, TT_DELACK))
1668                 tcp_timer_activate(tp, TT_DELACK, 0);
1669 #if 0
1670         /*
1671          * This completely breaks TCP if newreno is turned on.  What happens
1672          * is that if delayed-acks are turned on on the receiver, this code
1673          * on the transmitter effectively destroys the TCP window, forcing
1674          * it to four packets (1.5Kx4 = 6K window).
1675          */
1676         if (sendalot && --maxburst)
1677                 goto again;
1678 #endif
1679         if (sendalot)
1680                 goto again;
1681         return (0);
1682 }
1683
1684 void
1685 tcp_setpersist(struct tcpcb *tp)
1686 {
1687         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1688         int tt;
1689
1690         tp->t_flags &= ~TF_PREVVALID;
1691         if (tcp_timer_active(tp, TT_REXMT))
1692                 panic("tcp_setpersist: retransmit pending");
1693         /*
1694          * Start/restart persistence timer.
1695          */
1696         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1697                       tcp_persmin, tcp_persmax);
1698         tcp_timer_activate(tp, TT_PERSIST, tt);
1699         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1700                 tp->t_rxtshift++;
1701 }
1702
1703 /*
1704  * Insert TCP options according to the supplied parameters to the place
1705  * optp in a consistent way.  Can handle unaligned destinations.
1706  *
1707  * The order of the option processing is crucial for optimal packing and
1708  * alignment for the scarce option space.
1709  *
1710  * The optimal order for a SYN/SYN-ACK segment is:
1711  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1712  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1713  *
1714  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1715  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1716  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1717  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1718  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1719  */
1720 int
1721 tcp_addoptions(struct tcpopt *to, u_char *optp)
1722 {
1723         u_int32_t mask, optlen = 0;
1724
1725         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1726                 if ((to->to_flags & mask) != mask)
1727                         continue;
1728                 if (optlen == TCP_MAXOLEN)
1729                         break;
1730                 switch (to->to_flags & mask) {
1731                 case TOF_MSS:
1732                         while (optlen % 4) {
1733                                 optlen += TCPOLEN_NOP;
1734                                 *optp++ = TCPOPT_NOP;
1735                         }
1736                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1737                                 continue;
1738                         optlen += TCPOLEN_MAXSEG;
1739                         *optp++ = TCPOPT_MAXSEG;
1740                         *optp++ = TCPOLEN_MAXSEG;
1741                         to->to_mss = htons(to->to_mss);
1742                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1743                         optp += sizeof(to->to_mss);
1744                         break;
1745                 case TOF_SCALE:
1746                         while (!optlen || optlen % 2 != 1) {
1747                                 optlen += TCPOLEN_NOP;
1748                                 *optp++ = TCPOPT_NOP;
1749                         }
1750                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1751                                 continue;
1752                         optlen += TCPOLEN_WINDOW;
1753                         *optp++ = TCPOPT_WINDOW;
1754                         *optp++ = TCPOLEN_WINDOW;
1755                         *optp++ = to->to_wscale;
1756                         break;
1757                 case TOF_SACKPERM:
1758                         while (optlen % 2) {
1759                                 optlen += TCPOLEN_NOP;
1760                                 *optp++ = TCPOPT_NOP;
1761                         }
1762                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1763                                 continue;
1764                         optlen += TCPOLEN_SACK_PERMITTED;
1765                         *optp++ = TCPOPT_SACK_PERMITTED;
1766                         *optp++ = TCPOLEN_SACK_PERMITTED;
1767                         break;
1768                 case TOF_TS:
1769                         while (!optlen || optlen % 4 != 2) {
1770                                 optlen += TCPOLEN_NOP;
1771                                 *optp++ = TCPOPT_NOP;
1772                         }
1773                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1774                                 continue;
1775                         optlen += TCPOLEN_TIMESTAMP;
1776                         *optp++ = TCPOPT_TIMESTAMP;
1777                         *optp++ = TCPOLEN_TIMESTAMP;
1778                         to->to_tsval = htonl(to->to_tsval);
1779                         to->to_tsecr = htonl(to->to_tsecr);
1780                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1781                         optp += sizeof(to->to_tsval);
1782                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1783                         optp += sizeof(to->to_tsecr);
1784                         break;
1785                 case TOF_SIGNATURE:
1786                         {
1787                         int siglen = TCPOLEN_SIGNATURE - 2;
1788
1789                         while (!optlen || optlen % 4 != 2) {
1790                                 optlen += TCPOLEN_NOP;
1791                                 *optp++ = TCPOPT_NOP;
1792                         }
1793                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1794                                 to->to_flags &= ~TOF_SIGNATURE;
1795                                 continue;
1796                         }
1797                         optlen += TCPOLEN_SIGNATURE;
1798                         *optp++ = TCPOPT_SIGNATURE;
1799                         *optp++ = TCPOLEN_SIGNATURE;
1800                         to->to_signature = optp;
1801                         while (siglen--)
1802                                  *optp++ = 0;
1803                         break;
1804                         }
1805                 case TOF_SACK:
1806                         {
1807                         int sackblks = 0;
1808                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
1809                         tcp_seq sack_seq;
1810
1811                         while (!optlen || optlen % 4 != 2) {
1812                                 optlen += TCPOLEN_NOP;
1813                                 *optp++ = TCPOPT_NOP;
1814                         }
1815                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1816                                 continue;
1817                         optlen += TCPOLEN_SACKHDR;
1818                         *optp++ = TCPOPT_SACK;
1819                         sackblks = min(to->to_nsacks,
1820                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1821                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1822                         while (sackblks--) {
1823                                 sack_seq = htonl(sack->start);
1824                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1825                                 optp += sizeof(sack_seq);
1826                                 sack_seq = htonl(sack->end);
1827                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1828                                 optp += sizeof(sack_seq);
1829                                 optlen += TCPOLEN_SACK;
1830                                 sack++;
1831                         }
1832                         TCPSTAT_INC(tcps_sack_send_blocks);
1833                         break;
1834                         }
1835                 case TOF_FASTOPEN:
1836                         {
1837                         int total_len;
1838
1839                         /* XXX is there any point to aligning this option? */
1840                         total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1841                         if (TCP_MAXOLEN - optlen < total_len) {
1842                                 to->to_flags &= ~TOF_FASTOPEN;
1843                                 continue;
1844                         }
1845                         *optp++ = TCPOPT_FAST_OPEN;
1846                         *optp++ = total_len;
1847                         if (to->to_tfo_len > 0) {
1848                                 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1849                                 optp += to->to_tfo_len;
1850                         }
1851                         optlen += total_len;
1852                         break;
1853                         }
1854                 default:
1855                         panic("%s: unknown TCP option type", __func__);
1856                         break;
1857                 }
1858         }
1859
1860         /* Terminate and pad TCP options to a 4 byte boundary. */
1861         if (optlen % 4) {
1862                 optlen += TCPOLEN_EOL;
1863                 *optp++ = TCPOPT_EOL;
1864         }
1865         /*
1866          * According to RFC 793 (STD0007):
1867          *   "The content of the header beyond the End-of-Option option
1868          *    must be header padding (i.e., zero)."
1869          *   and later: "The padding is composed of zeros."
1870          */
1871         while (optlen % 4) {
1872                 optlen += TCPOLEN_PAD;
1873                 *optp++ = TCPOPT_PAD;
1874         }
1875
1876         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1877         return (optlen);
1878 }
1879
1880 /*
1881  * This is a copy of m_copym(), taking the TSO segment size/limit
1882  * constraints into account, and advancing the sndptr as it goes.
1883  */
1884 struct mbuf *
1885 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1886     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1887 {
1888 #ifdef KERN_TLS
1889         struct ktls_session *tls, *ntls;
1890         struct mbuf *start;
1891 #endif
1892         struct mbuf *n, **np;
1893         struct mbuf *top;
1894         int32_t off = off0;
1895         int32_t len = *plen;
1896         int32_t fragsize;
1897         int32_t len_cp = 0;
1898         int32_t *pkthdrlen;
1899         uint32_t mlen, frags;
1900         bool copyhdr;
1901
1902
1903         KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1904         KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1905         if (off == 0 && m->m_flags & M_PKTHDR)
1906                 copyhdr = true;
1907         else
1908                 copyhdr = false;
1909         while (off > 0) {
1910                 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1911                 if (off < m->m_len)
1912                         break;
1913                 off -= m->m_len;
1914                 if ((sb) && (m == sb->sb_sndptr)) {
1915                         sb->sb_sndptroff += m->m_len;
1916                         sb->sb_sndptr = m->m_next;
1917                 }
1918                 m = m->m_next;
1919         }
1920         np = &top;
1921         top = NULL;
1922         pkthdrlen = NULL;
1923 #ifdef KERN_TLS
1924         if (hw_tls && (m->m_flags & M_EXTPG))
1925                 tls = m->m_epg_tls;
1926         else
1927                 tls = NULL;
1928         start = m;
1929 #endif
1930         while (len > 0) {
1931                 if (m == NULL) {
1932                         KASSERT(len == M_COPYALL,
1933                             ("tcp_m_copym, length > size of mbuf chain"));
1934                         *plen = len_cp;
1935                         if (pkthdrlen != NULL)
1936                                 *pkthdrlen = len_cp;
1937                         break;
1938                 }
1939 #ifdef KERN_TLS
1940                 if (hw_tls) {
1941                         if (m->m_flags & M_EXTPG)
1942                                 ntls = m->m_epg_tls;
1943                         else
1944                                 ntls = NULL;
1945
1946                         /*
1947                          * Avoid mixing TLS records with handshake
1948                          * data or TLS records from different
1949                          * sessions.
1950                          */
1951                         if (tls != ntls) {
1952                                 MPASS(m != start);
1953                                 *plen = len_cp;
1954                                 if (pkthdrlen != NULL)
1955                                         *pkthdrlen = len_cp;
1956                                 break;
1957                         }
1958
1959                         /*
1960                          * Don't end a send in the middle of a TLS
1961                          * record if it spans multiple TLS records.
1962                          */
1963                         if (tls != NULL && (m != start) && len < m->m_len) {
1964                                 *plen = len_cp;
1965                                 if (pkthdrlen != NULL)
1966                                         *pkthdrlen = len_cp;
1967                                 break;
1968                         }
1969                 }
1970 #endif
1971                 mlen = min(len, m->m_len - off);
1972                 if (seglimit) {
1973                         /*
1974                          * For M_EXTPG mbufs, add 3 segments
1975                          * + 1 in case we are crossing page boundaries
1976                          * + 2 in case the TLS hdr/trailer are used
1977                          * It is cheaper to just add the segments
1978                          * than it is to take the cache miss to look
1979                          * at the mbuf ext_pgs state in detail.
1980                          */
1981                         if (m->m_flags & M_EXTPG) {
1982                                 fragsize = min(segsize, PAGE_SIZE);
1983                                 frags = 3;
1984                         } else {
1985                                 fragsize = segsize;
1986                                 frags = 0;
1987                         }
1988
1989                         /* Break if we really can't fit anymore. */
1990                         if ((frags + 1) >= seglimit) {
1991                                 *plen = len_cp;
1992                                 if (pkthdrlen != NULL)
1993                                         *pkthdrlen = len_cp;
1994                                 break;
1995                         }
1996
1997                         /*
1998                          * Reduce size if you can't copy the whole
1999                          * mbuf. If we can't copy the whole mbuf, also
2000                          * adjust len so the loop will end after this
2001                          * mbuf.
2002                          */
2003                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2004                                 mlen = (seglimit - frags - 1) * fragsize;
2005                                 len = mlen;
2006                                 *plen = len_cp + len;
2007                                 if (pkthdrlen != NULL)
2008                                         *pkthdrlen = *plen;
2009                         }
2010                         frags += howmany(mlen, fragsize);
2011                         if (frags == 0)
2012                                 frags++;
2013                         seglimit -= frags;
2014                         KASSERT(seglimit > 0,
2015                             ("%s: seglimit went too low", __func__));
2016                 }
2017                 if (copyhdr)
2018                         n = m_gethdr(M_NOWAIT, m->m_type);
2019                 else
2020                         n = m_get(M_NOWAIT, m->m_type);
2021                 *np = n;
2022                 if (n == NULL)
2023                         goto nospace;
2024                 if (copyhdr) {
2025                         if (!m_dup_pkthdr(n, m, M_NOWAIT))
2026                                 goto nospace;
2027                         if (len == M_COPYALL)
2028                                 n->m_pkthdr.len -= off0;
2029                         else
2030                                 n->m_pkthdr.len = len;
2031                         pkthdrlen = &n->m_pkthdr.len;
2032                         copyhdr = false;
2033                 }
2034                 n->m_len = mlen;
2035                 len_cp += n->m_len;
2036                 if (m->m_flags & (M_EXT|M_EXTPG)) {
2037                         n->m_data = m->m_data + off;
2038                         mb_dupcl(n, m);
2039                 } else
2040                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2041                             (u_int)n->m_len);
2042
2043                 if (sb && (sb->sb_sndptr == m) &&
2044                     ((n->m_len + off) >= m->m_len) && m->m_next) {
2045                         sb->sb_sndptroff += m->m_len;
2046                         sb->sb_sndptr = m->m_next;
2047                 }
2048                 off = 0;
2049                 if (len != M_COPYALL) {
2050                         len -= n->m_len;
2051                 }
2052                 m = m->m_next;
2053                 np = &n->m_next;
2054         }
2055         return (top);
2056 nospace:
2057         m_freem(top);
2058         return (NULL);
2059 }
2060
2061 void
2062 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2063 {
2064
2065         /*
2066          * Automatic sizing of send socket buffer.  Often the send buffer
2067          * size is not optimally adjusted to the actual network conditions
2068          * at hand (delay bandwidth product).  Setting the buffer size too
2069          * small limits throughput on links with high bandwidth and high
2070          * delay (eg. trans-continental/oceanic links).  Setting the
2071          * buffer size too big consumes too much real kernel memory,
2072          * especially with many connections on busy servers.
2073          *
2074          * The criteria to step up the send buffer one notch are:
2075          *  1. receive window of remote host is larger than send buffer
2076          *     (with a fudge factor of 5/4th);
2077          *  2. send buffer is filled to 7/8th with data (so we actually
2078          *     have data to make use of it);
2079          *  3. send buffer fill has not hit maximal automatic size;
2080          *  4. our send window (slow start and cogestion controlled) is
2081          *     larger than sent but unacknowledged data in send buffer.
2082          *
2083          * The remote host receive window scaling factor may limit the
2084          * growing of the send buffer before it reaches its allowed
2085          * maximum.
2086          *
2087          * It scales directly with slow start or congestion window
2088          * and does at most one step per received ACK.  This fast
2089          * scaling has the drawback of growing the send buffer beyond
2090          * what is strictly necessary to make full use of a given
2091          * delay*bandwidth product.  However testing has shown this not
2092          * to be much of an problem.  At worst we are trading wasting
2093          * of available bandwidth (the non-use of it) for wasting some
2094          * socket buffer memory.
2095          *
2096          * TODO: Shrink send buffer during idle periods together
2097          * with congestion window.  Requires another timer.  Has to
2098          * wait for upcoming tcp timer rewrite.
2099          *
2100          * XXXGL: should there be used sbused() or sbavail()?
2101          */
2102         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2103                 int lowat;
2104
2105                 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2106                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2107                     sbused(&so->so_snd) >=
2108                     (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2109                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2110                     sendwin >= (sbused(&so->so_snd) -
2111                     (tp->snd_nxt - tp->snd_una))) {
2112                         if (!sbreserve_locked(&so->so_snd,
2113                             min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2114                              V_tcp_autosndbuf_max), so, curthread))
2115                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2116                 }
2117         }
2118 }