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_tcpdebug.h"
  41
  42 #include <sys/param.h>
  43 #include <sys/systm.h>
  44 #include <sys/domain.h>
  45 #ifdef TCP_HHOOK
  46 #include <sys/hhook.h>
  47 #endif
  48 #include <sys/kernel.h>
  49 #include <sys/lock.h>
  50 #include <sys/mbuf.h>
  51 #include <sys/mutex.h>
  52 #include <sys/protosw.h>
  53 #include <sys/sdt.h>
  54 #include <sys/socket.h>
  55 #include <sys/socketvar.h>
  56 #include <sys/sysctl.h>
  57
  58 #include <net/if.h>
  59 #include <net/route.h>
  60 #include <net/vnet.h>
  61
  62 #include <netinet/in.h>
  63 #include <netinet/in_kdtrace.h>
  64 #include <netinet/in_systm.h>
  65 #include <netinet/ip.h>
  66 #include <netinet/in_pcb.h>
  67 #include <netinet/ip_var.h>
  68 #include <netinet/ip_options.h>
  69 #ifdef INET6
  70 #include <netinet6/in6_pcb.h>
  71 #include <netinet/ip6.h>
  72 #include <netinet6/ip6_var.h>
  73 #endif
  74 #include <netinet/tcp.h>
  75 #include <netinet/tcp_fsm.h>
  76 #include <netinet/tcp_log_buf.h>
  77 #include <netinet/tcp_seq.h>
  78 #include <netinet/tcp_timer.h>
  79 #include <netinet/tcp_var.h>
  80 #include <netinet/tcpip.h>
  81 #include <netinet/cc/cc.h>
  82 #include <netinet/tcp_fastopen.h>
  83 #ifdef TCPPCAP
  84 #include <netinet/tcp_pcap.h>
  85 #endif
  86 #ifdef TCPDEBUG
  87 #include <netinet/tcp_debug.h>
  88 #endif
  89 #ifdef TCP_OFFLOAD
  90 #include <netinet/tcp_offload.h>
  91 #endif
  92
  93 #include <netipsec/ipsec_support.h>
  94
  95 #include <machine/in_cksum.h>
  96
  97 #include <security/mac/mac_framework.h>
  98
  99 VNET_DEFINE(int, path_mtu_discovery) = 1;
 100 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
 101         &VNET_NAME(path_mtu_discovery), 1,
 102         "Enable Path MTU Discovery");
 103
 104 VNET_DEFINE(int, tcp_do_tso) = 1;
 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
 106         &VNET_NAME(tcp_do_tso), 0,
 107         "Enable TCP Segmentation Offload");
 108
 109 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
 110 #define V_tcp_sendspace VNET(tcp_sendspace)
 111 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
 112         &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
 113
 114 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
 115 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
 116         &VNET_NAME(tcp_do_autosndbuf), 0,
 117         "Enable automatic send buffer sizing");
 118
 119 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
 121         &VNET_NAME(tcp_autosndbuf_inc), 0,
 122         "Incrementor step size of automatic send buffer");
 123
 124 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
 126         &VNET_NAME(tcp_autosndbuf_max), 0,
 127         "Max size of automatic send buffer");
 128
 129 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
 130 #define V_tcp_sendbuf_auto_lowat        VNET(tcp_sendbuf_auto_lowat)
 131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
 132         &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
 133         "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
 134
 135 /*
 136  * Make sure that either retransmit or persist timer is set for SYN, FIN and
 137  * non-ACK.
 138  */
 139 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)                        \
 140         KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
 141             tcp_timer_active((tp), TT_REXMT) ||                         \
 142             tcp_timer_active((tp), TT_PERSIST),                         \
 143             ("neither rexmt nor persist timer is set"))
 144
 145 static void inline      cc_after_idle(struct tcpcb *tp);
 146
 147 #ifdef TCP_HHOOK
 148 /*
 149  * Wrapper for the TCP established output helper hook.
 150  */
 151 void
 152 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
 153     struct tcpopt *to, uint32_t len, int tso)
 154 {
 155         struct tcp_hhook_data hhook_data;
 156
 157         if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
 158                 hhook_data.tp = tp;
 159                 hhook_data.th = th;
 160                 hhook_data.to = to;
 161                 hhook_data.len = len;
 162                 hhook_data.tso = tso;
 163
 164                 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
 165                     tp->osd);
 166         }
 167 }
 168 #endif
 169
 170 /*
 171  * Flags used when sending segments in tcp_output.  Basic flags (TH_RST,
 172  * TH_ACK,TH_SYN,TH_FIN) are totally determined by state, with the proviso
 173  * that TH_FIN is sent only if all data queued for output is included in the
 174  * segment.
 175  */
 176 u_char  tcp_outflags[TCP_NSTATES] = {
 177         TH_RST|TH_ACK,          /* 0, CLOSED */
 178         0,                      /* 1, LISTEN */
 179         TH_SYN,                 /* 2, SYN_SENT */
 180         TH_SYN|TH_ACK,          /* 3, SYN_RECEIVED */
 181         TH_ACK,                 /* 4, ESTABLISHED */
 182         TH_ACK,                 /* 5, CLOSE_WAIT */
 183         TH_FIN|TH_ACK,          /* 6, FIN_WAIT_1 */
 184         TH_FIN|TH_ACK,          /* 7, CLOSING */
 185         TH_FIN|TH_ACK,          /* 8, LAST_ACK */
 186         TH_ACK,                 /* 9, FIN_WAIT_2 */
 187         TH_ACK,                 /* 10, TIME_WAIT */
 188 };
 189
 190 /*
 191  * CC wrapper hook functions
 192  */
 193 static void inline
 194 cc_after_idle(struct tcpcb *tp)
 195 {
 196         INP_WLOCK_ASSERT(tp->t_inpcb);
 197
 198         if (CC_ALGO(tp)->after_idle != NULL)
 199                 CC_ALGO(tp)->after_idle(tp->ccv);
 200 }
 201
 202 /*
 203  * Tcp output routine: figure out what should be sent and send it.
 204  */
 205 int
 206 tcp_output(struct tcpcb *tp)
 207 {
 208         struct socket *so = tp->t_inpcb->inp_socket;
 209         int32_t len;
 210         uint32_t recwin, sendwin;
 211         int off, flags, error = 0;      /* Keep compiler happy */
 212         struct mbuf *m;
 213         struct ip *ip = NULL;
 214 #ifdef TCPDEBUG
 215         struct ipovly *ipov = NULL;
 216 #endif
 217         struct tcphdr *th;
 218         u_char opt[TCP_MAXOLEN];
 219         unsigned ipoptlen, optlen, hdrlen;
 220 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 221         unsigned ipsec_optlen = 0;
 222 #endif
 223         int idle, sendalot, curticks;
 224         int sack_rxmit, sack_bytes_rxmt;
 225         struct sackhole *p;
 226         int tso, mtu;
 227         struct tcpopt to;
 228         unsigned int wanted_cookie = 0;
 229         unsigned int dont_sendalot = 0;
 230 #if 0
 231         int maxburst = TCP_MAXBURST;
 232 #endif
 233 #ifdef INET6
 234         struct ip6_hdr *ip6 = NULL;
 235         int isipv6;
 236
 237         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
 238 #endif
 239
 240         INP_WLOCK_ASSERT(tp->t_inpcb);
 241
 242 #ifdef TCP_OFFLOAD
 243         if (tp->t_flags & TF_TOE)
 244                 return (tcp_offload_output(tp));
 245 #endif
 246
 247         /*
 248          * For TFO connections in SYN_RECEIVED, only allow the initial
 249          * SYN|ACK and those sent by the retransmit timer.
 250          */
 251         if (IS_FASTOPEN(tp->t_flags) &&
 252             (tp->t_state == TCPS_SYN_RECEIVED) &&
 253             SEQ_GT(tp->snd_max, tp->snd_una) &&    /* initial SYN|ACK sent */
 254             (tp->snd_nxt != tp->snd_una))         /* not a retransmit */
 255                 return (0);
 256
 257         /*
 258          * Determine length of data that should be transmitted,
 259          * and flags that will be used.
 260          * If there is some data or critical controls (SYN, RST)
 261          * to send, then transmit; otherwise, investigate further.
 262          */
 263         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
 264         if (idle && ticks - tp->t_rcvtime >= tp->t_rxtcur)
 265                 cc_after_idle(tp);
 266         tp->t_flags &= ~TF_LASTIDLE;
 267         if (idle) {
 268                 if (tp->t_flags & TF_MORETOCOME) {
 269                         tp->t_flags |= TF_LASTIDLE;
 270                         idle = 0;
 271                 }
 272         }
 273 again:
 274         /*
 275          * If we've recently taken a timeout, snd_max will be greater than
 276          * snd_nxt.  There may be SACK information that allows us to avoid
 277          * resending already delivered data.  Adjust snd_nxt accordingly.
 278          */
 279         if ((tp->t_flags & TF_SACK_PERMIT) &&
 280             SEQ_LT(tp->snd_nxt, tp->snd_max))
 281                 tcp_sack_adjust(tp);
 282         sendalot = 0;
 283         tso = 0;
 284         mtu = 0;
 285         off = tp->snd_nxt - tp->snd_una;
 286         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
 287
 288         flags = tcp_outflags[tp->t_state];
 289         /*
 290          * Send any SACK-generated retransmissions.  If we're explicitly trying
 291          * to send out new data (when sendalot is 1), bypass this function.
 292          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
 293          * we're replacing a (future) new transmission with a retransmission
 294          * now, and we previously incremented snd_cwnd in tcp_input().
 295          */
 296         /*
 297          * Still in sack recovery , reset rxmit flag to zero.
 298          */
 299         sack_rxmit = 0;
 300         sack_bytes_rxmt = 0;
 301         len = 0;
 302         p = NULL;
 303         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
 304             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
 305                 uint32_t cwin;
 306
 307                 cwin =
 308                     imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
 309                 /* Do not retransmit SACK segments beyond snd_recover */
 310                 if (SEQ_GT(p->end, tp->snd_recover)) {
 311                         /*
 312                          * (At least) part of sack hole extends beyond
 313                          * snd_recover. Check to see if we can rexmit data
 314                          * for this hole.
 315                          */
 316                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
 317                                 /*
 318                                  * Can't rexmit any more data for this hole.
 319                                  * That data will be rexmitted in the next
 320                                  * sack recovery episode, when snd_recover
 321                                  * moves past p->rxmit.
 322                                  */
 323                                 p = NULL;
 324                                 goto after_sack_rexmit;
 325                         } else
 326                                 /* Can rexmit part of the current hole */
 327                                 len = ((int32_t)ulmin(cwin,
 328                                                    tp->snd_recover - p->rxmit));
 329                 } else
 330                         len = ((int32_t)ulmin(cwin, p->end - p->rxmit));
 331                 off = p->rxmit - tp->snd_una;
 332                 KASSERT(off >= 0,("%s: sack block to the left of una : %d",
 333                     __func__, off));
 334                 if (len > 0) {
 335                         sack_rxmit = 1;
 336                         sendalot = 1;
 337                         TCPSTAT_INC(tcps_sack_rexmits);
 338                         TCPSTAT_ADD(tcps_sack_rexmit_bytes,
 339                             min(len, tp->t_maxseg));
 340                 }
 341         }
 342 after_sack_rexmit:
 343         /*
 344          * Get standard flags, and add SYN or FIN if requested by 'hidden'
 345          * state flags.
 346          */
 347         if (tp->t_flags & TF_NEEDFIN)
 348                 flags |= TH_FIN;
 349         if (tp->t_flags & TF_NEEDSYN)
 350                 flags |= TH_SYN;
 351
 352         SOCKBUF_LOCK(&so->so_snd);
 353         /*
 354          * If in persist timeout with window of 0, send 1 byte.
 355          * Otherwise, if window is small but nonzero
 356          * and timer expired, we will send what we can
 357          * and go to transmit state.
 358          */
 359         if (tp->t_flags & TF_FORCEDATA) {
 360                 if (sendwin == 0) {
 361                         /*
 362                          * If we still have some data to send, then
 363                          * clear the FIN bit.  Usually this would
 364                          * happen below when it realizes that we
 365                          * aren't sending all the data.  However,
 366                          * if we have exactly 1 byte of unsent data,
 367                          * then it won't clear the FIN bit below,
 368                          * and if we are in persist state, we wind
 369                          * up sending the packet without recording
 370                          * that we sent the FIN bit.
 371                          *
 372                          * We can't just blindly clear the FIN bit,
 373                          * because if we don't have any more data
 374                          * to send then the probe will be the FIN
 375                          * itself.
 376                          */
 377                         if (off < sbused(&so->so_snd))
 378                                 flags &= ~TH_FIN;
 379                         sendwin = 1;
 380                 } else {
 381                         tcp_timer_activate(tp, TT_PERSIST, 0);
 382                         tp->t_rxtshift = 0;
 383                 }
 384         }
 385
 386         /*
 387          * If snd_nxt == snd_max and we have transmitted a FIN, the
 388          * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
 389          * a negative length.  This can also occur when TCP opens up
 390          * its congestion window while receiving additional duplicate
 391          * acks after fast-retransmit because TCP will reset snd_nxt
 392          * to snd_max after the fast-retransmit.
 393          *
 394          * In the normal retransmit-FIN-only case, however, snd_nxt will
 395          * be set to snd_una, the offset will be 0, and the length may
 396          * wind up 0.
 397          *
 398          * If sack_rxmit is true we are retransmitting from the scoreboard
 399          * in which case len is already set.
 400          */
 401         if (sack_rxmit == 0) {
 402                 if (sack_bytes_rxmt == 0)
 403                         len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
 404                             off);
 405                 else {
 406                         int32_t cwin;
 407
 408                         /*
 409                          * We are inside of a SACK recovery episode and are
 410                          * sending new data, having retransmitted all the
 411                          * data possible in the scoreboard.
 412                          */
 413                         len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
 414                             off);
 415                         /*
 416                          * Don't remove this (len > 0) check !
 417                          * We explicitly check for len > 0 here (although it
 418                          * isn't really necessary), to work around a gcc
 419                          * optimization issue - to force gcc to compute
 420                          * len above. Without this check, the computation
 421                          * of len is bungled by the optimizer.
 422                          */
 423                         if (len > 0) {
 424                                 cwin = tp->snd_cwnd -
 425                                         (tp->snd_nxt - tp->sack_newdata) -
 426                                         sack_bytes_rxmt;
 427                                 if (cwin < 0)
 428                                         cwin = 0;
 429                                 len = imin(len, cwin);
 430                         }
 431                 }
 432         }
 433
 434         /*
 435          * Lop off SYN bit if it has already been sent.  However, if this
 436          * is SYN-SENT state and if segment contains data and if we don't
 437          * know that foreign host supports TAO, suppress sending segment.
 438          */
 439         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
 440                 if (tp->t_state != TCPS_SYN_RECEIVED)
 441                         flags &= ~TH_SYN;
 442                 /*
 443                  * When sending additional segments following a TFO SYN|ACK,
 444                  * do not include the SYN bit.
 445                  */
 446                 if (IS_FASTOPEN(tp->t_flags) &&
 447                     (tp->t_state == TCPS_SYN_RECEIVED))
 448                         flags &= ~TH_SYN;
 449                 off--, len++;
 450         }
 451
 452         /*
 453          * Be careful not to send data and/or FIN on SYN segments.
 454          * This measure is needed to prevent interoperability problems
 455          * with not fully conformant TCP implementations.
 456          */
 457         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
 458                 len = 0;
 459                 flags &= ~TH_FIN;
 460         }
 461
 462         /*
 463          * On TFO sockets, ensure no data is sent in the following cases:
 464          *
 465          *  - When retransmitting SYN|ACK on a passively-created socket
 466          *
 467          *  - When retransmitting SYN on an actively created socket
 468          *
 469          *  - When sending a zero-length cookie (cookie request) on an
 470          *    actively created socket
 471          *
 472          *  - When the socket is in the CLOSED state (RST is being sent)
 473          */
 474         if (IS_FASTOPEN(tp->t_flags) &&
 475             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
 476              ((tp->t_state == TCPS_SYN_SENT) &&
 477               (tp->t_tfo_client_cookie_len == 0)) ||
 478              (flags & TH_RST)))
 479                 len = 0;
 480         if (len <= 0) {
 481                 /*
 482                  * If FIN has been sent but not acked,
 483                  * but we haven't been called to retransmit,
 484                  * len will be < 0.  Otherwise, window shrank
 485                  * after we sent into it.  If window shrank to 0,
 486                  * cancel pending retransmit, pull snd_nxt back
 487                  * to (closed) window, and set the persist timer
 488                  * if it isn't already going.  If the window didn't
 489                  * close completely, just wait for an ACK.
 490                  *
 491                  * We also do a general check here to ensure that
 492                  * we will set the persist timer when we have data
 493                  * to send, but a 0-byte window. This makes sure
 494                  * the persist timer is set even if the packet
 495                  * hits one of the "goto send" lines below.
 496                  */
 497                 len = 0;
 498                 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
 499                         (off < (int) sbavail(&so->so_snd))) {
 500                         tcp_timer_activate(tp, TT_REXMT, 0);
 501                         tp->t_rxtshift = 0;
 502                         tp->snd_nxt = tp->snd_una;
 503                         if (!tcp_timer_active(tp, TT_PERSIST))
 504                                 tcp_setpersist(tp);
 505                 }
 506         }
 507
 508         /* len will be >= 0 after this point. */
 509         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 510
 511         tcp_sndbuf_autoscale(tp, so, sendwin);
 512
 513         /*
 514          * Decide if we can use TCP Segmentation Offloading (if supported by
 515          * hardware).
 516          *
 517          * TSO may only be used if we are in a pure bulk sending state.  The
 518          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
 519          * IP options prevent using TSO.  With TSO the TCP header is the same
 520          * (except for the sequence number) for all generated packets.  This
 521          * makes it impossible to transmit any options which vary per generated
 522          * segment or packet.
 523          *
 524          * IPv4 handling has a clear separation of ip options and ip header
 525          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
 526          * the right thing below to provide length of just ip options and thus
 527          * checking for ipoptlen is enough to decide if ip options are present.
 528          */
 529 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 530         /*
 531          * Pre-calculate here as we save another lookup into the darknesses
 532          * of IPsec that way and can actually decide if TSO is ok.
 533          */
 534 #ifdef INET6
 535         if (isipv6 && IPSEC_ENABLED(ipv6))
 536                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
 537 #ifdef INET
 538         else
 539 #endif
 540 #endif /* INET6 */
 541 #ifdef INET
 542         if (IPSEC_ENABLED(ipv4))
 543                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
 544 #endif /* INET */
 545 #endif /* IPSEC */
 546 #ifdef INET6
 547         if (isipv6)
 548                 ipoptlen = ip6_optlen(tp->t_inpcb);
 549         else
 550 #endif
 551         if (tp->t_inpcb->inp_options)
 552                 ipoptlen = tp->t_inpcb->inp_options->m_len -
 553                                 offsetof(struct ipoption, ipopt_list);
 554         else
 555                 ipoptlen = 0;
 556 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 557         ipoptlen += ipsec_optlen;
 558 #endif
 559
 560         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
 561             ((tp->t_flags & TF_SIGNATURE) == 0) &&
 562             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
 563             ipoptlen == 0 && !(flags & TH_SYN))
 564                 tso = 1;
 565
 566         if (sack_rxmit) {
 567                 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
 568                         flags &= ~TH_FIN;
 569         } else {
 570                 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
 571                     sbused(&so->so_snd)))
 572                         flags &= ~TH_FIN;
 573         }
 574
 575         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 576             (long)TCP_MAXWIN << tp->rcv_scale);
 577
 578         /*
 579          * Sender silly window avoidance.   We transmit under the following
 580          * conditions when len is non-zero:
 581          *
 582          *      - We have a full segment (or more with TSO)
 583          *      - This is the last buffer in a write()/send() and we are
 584          *        either idle or running NODELAY
 585          *      - we've timed out (e.g. persist timer)
 586          *      - we have more then 1/2 the maximum send window's worth of
 587          *        data (receiver may be limited the window size)
 588          *      - we need to retransmit
 589          */
 590         if (len) {
 591                 if (len >= tp->t_maxseg)
 592                         goto send;
 593                 /*
 594                  * NOTE! on localhost connections an 'ack' from the remote
 595                  * end may occur synchronously with the output and cause
 596                  * us to flush a buffer queued with moretocome.  XXX
 597                  *
 598                  * note: the len + off check is almost certainly unnecessary.
 599                  */
 600                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
 601                     (idle || (tp->t_flags & TF_NODELAY)) &&
 602                     (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
 603                     (tp->t_flags & TF_NOPUSH) == 0) {
 604                         goto send;
 605                 }
 606                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
 607                         goto send;
 608                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
 609                         goto send;
 610                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
 611                         goto send;
 612                 if (sack_rxmit)
 613                         goto send;
 614         }
 615
 616         /*
 617          * Sending of standalone window updates.
 618          *
 619          * Window updates are important when we close our window due to a
 620          * full socket buffer and are opening it again after the application
 621          * reads data from it.  Once the window has opened again and the
 622          * remote end starts to send again the ACK clock takes over and
 623          * provides the most current window information.
 624          *
 625          * We must avoid the silly window syndrome whereas every read
 626          * from the receive buffer, no matter how small, causes a window
 627          * update to be sent.  We also should avoid sending a flurry of
 628          * window updates when the socket buffer had queued a lot of data
 629          * and the application is doing small reads.
 630          *
 631          * Prevent a flurry of pointless window updates by only sending
 632          * an update when we can increase the advertized window by more
 633          * than 1/4th of the socket buffer capacity.  When the buffer is
 634          * getting full or is very small be more aggressive and send an
 635          * update whenever we can increase by two mss sized segments.
 636          * In all other situations the ACK's to new incoming data will
 637          * carry further window increases.
 638          *
 639          * Don't send an independent window update if a delayed
 640          * ACK is pending (it will get piggy-backed on it) or the
 641          * remote side already has done a half-close and won't send
 642          * more data.  Skip this if the connection is in T/TCP
 643          * half-open state.
 644          */
 645         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
 646             !(tp->t_flags & TF_DELACK) &&
 647             !TCPS_HAVERCVDFIN(tp->t_state)) {
 648                 /*
 649                  * "adv" is the amount we could increase the window,
 650                  * taking into account that we are limited by
 651                  * TCP_MAXWIN << tp->rcv_scale.
 652                  */
 653                 int32_t adv;
 654                 int oldwin;
 655
 656                 adv = recwin;
 657                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
 658                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
 659                         adv -= oldwin;
 660                 } else
 661                         oldwin = 0;
 662
 663                 /*
 664                  * If the new window size ends up being the same as or less
 665                  * than the old size when it is scaled, then don't force
 666                  * a window update.
 667                  */
 668                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
 669                         goto dontupdate;
 670
 671                 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
 672                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
 673                      recwin <= (so->so_rcv.sb_hiwat / 8) ||
 674                      so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg))
 675                         goto send;
 676                 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
 677                         goto send;
 678         }
 679 dontupdate:
 680
 681         /*
 682          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
 683          * is also a catch-all for the retransmit timer timeout case.
 684          */
 685         if (tp->t_flags & TF_ACKNOW)
 686                 goto send;
 687         if ((flags & TH_RST) ||
 688             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
 689                 goto send;
 690         if (SEQ_GT(tp->snd_up, tp->snd_una))
 691                 goto send;
 692         /*
 693          * If our state indicates that FIN should be sent
 694          * and we have not yet done so, then we need to send.
 695          */
 696         if (flags & TH_FIN &&
 697             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
 698                 goto send;
 699         /*
 700          * In SACK, it is possible for tcp_output to fail to send a segment
 701          * after the retransmission timer has been turned off.  Make sure
 702          * that the retransmission timer is set.
 703          */
 704         if ((tp->t_flags & TF_SACK_PERMIT) &&
 705             SEQ_GT(tp->snd_max, tp->snd_una) &&
 706             !tcp_timer_active(tp, TT_REXMT) &&
 707             !tcp_timer_active(tp, TT_PERSIST)) {
 708                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 709                 goto just_return;
 710         }
 711         /*
 712          * TCP window updates are not reliable, rather a polling protocol
 713          * using ``persist'' packets is used to insure receipt of window
 714          * updates.  The three ``states'' for the output side are:
 715          *      idle                    not doing retransmits or persists
 716          *      persisting              to move a small or zero window
 717          *      (re)transmitting        and thereby not persisting
 718          *
 719          * tcp_timer_active(tp, TT_PERSIST)
 720          *      is true when we are in persist state.
 721          * (tp->t_flags & TF_FORCEDATA)
 722          *      is set when we are called to send a persist packet.
 723          * tcp_timer_active(tp, TT_REXMT)
 724          *      is set when we are retransmitting
 725          * The output side is idle when both timers are zero.
 726          *
 727          * If send window is too small, there is data to transmit, and no
 728          * retransmit or persist is pending, then go to persist state.
 729          * If nothing happens soon, send when timer expires:
 730          * if window is nonzero, transmit what we can,
 731          * otherwise force out a byte.
 732          */
 733         if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
 734             !tcp_timer_active(tp, TT_PERSIST)) {
 735                 tp->t_rxtshift = 0;
 736                 tcp_setpersist(tp);
 737         }
 738
 739         /*
 740          * No reason to send a segment, just return.
 741          */
 742 just_return:
 743         SOCKBUF_UNLOCK(&so->so_snd);
 744         return (0);
 745
 746 send:
 747         SOCKBUF_LOCK_ASSERT(&so->so_snd);
 748         if (len > 0) {
 749                 if (len >= tp->t_maxseg)
 750                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
 751                 else
 752                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
 753         }
 754         /*
 755          * Before ESTABLISHED, force sending of initial options
 756          * unless TCP set not to do any options.
 757          * NOTE: we assume that the IP/TCP header plus TCP options
 758          * always fit in a single mbuf, leaving room for a maximum
 759          * link header, i.e.
 760          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
 761          */
 762         optlen = 0;
 763 #ifdef INET6
 764         if (isipv6)
 765                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
 766         else
 767 #endif
 768                 hdrlen = sizeof (struct tcpiphdr);
 769
 770         /*
 771          * Compute options for segment.
 772          * We only have to care about SYN and established connection
 773          * segments.  Options for SYN-ACK segments are handled in TCP
 774          * syncache.
 775          */
 776         to.to_flags = 0;
 777         if ((tp->t_flags & TF_NOOPT) == 0) {
 778                 /* Maximum segment size. */
 779                 if (flags & TH_SYN) {
 780                         tp->snd_nxt = tp->iss;
 781                         to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
 782                         to.to_flags |= TOF_MSS;
 783
 784                         /*
 785                          * On SYN or SYN|ACK transmits on TFO connections,
 786                          * only include the TFO option if it is not a
 787                          * retransmit, as the presence of the TFO option may
 788                          * have caused the original SYN or SYN|ACK to have
 789                          * been dropped by a middlebox.
 790                          */
 791                         if (IS_FASTOPEN(tp->t_flags) &&
 792                             (tp->t_rxtshift == 0)) {
 793                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
 794                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 795                                         to.to_tfo_cookie =
 796                                             (u_int8_t *)&tp->t_tfo_cookie.server;
 797                                         to.to_flags |= TOF_FASTOPEN;
 798                                         wanted_cookie = 1;
 799                                 } else if (tp->t_state == TCPS_SYN_SENT) {
 800                                         to.to_tfo_len =
 801                                             tp->t_tfo_client_cookie_len;
 802                                         to.to_tfo_cookie =
 803                                             tp->t_tfo_cookie.client;
 804                                         to.to_flags |= TOF_FASTOPEN;
 805                                         wanted_cookie = 1;
 806                                         /*
 807                                          * If we wind up having more data to
 808                                          * send with the SYN than can fit in
 809                                          * one segment, don't send any more
 810                                          * until the SYN|ACK comes back from
 811                                          * the other end.
 812                                          */
 813                                         dont_sendalot = 1;
 814                                 }
 815                         }
 816                 }
 817                 /* Window scaling. */
 818                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
 819                         to.to_wscale = tp->request_r_scale;
 820                         to.to_flags |= TOF_SCALE;
 821                 }
 822                 /* Timestamps. */
 823                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
 824                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
 825                         curticks = tcp_ts_getticks();
 826                         to.to_tsval = curticks + tp->ts_offset;
 827                         to.to_tsecr = tp->ts_recent;
 828                         to.to_flags |= TOF_TS;
 829                         if (tp->t_rxtshift == 1)
 830                                 tp->t_badrxtwin = curticks;
 831                 }
 832
 833                 /* Set receive buffer autosizing timestamp. */
 834                 if (tp->rfbuf_ts == 0 &&
 835                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
 836                         tp->rfbuf_ts = tcp_ts_getticks();
 837
 838                 /* Selective ACK's. */
 839                 if (tp->t_flags & TF_SACK_PERMIT) {
 840                         if (flags & TH_SYN)
 841                                 to.to_flags |= TOF_SACKPERM;
 842                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
 843                             (tp->t_flags & TF_SACK_PERMIT) &&
 844                             tp->rcv_numsacks > 0) {
 845                                 to.to_flags |= TOF_SACK;
 846                                 to.to_nsacks = tp->rcv_numsacks;
 847                                 to.to_sacks = (u_char *)tp->sackblks;
 848                         }
 849                 }
 850 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 851                 /* TCP-MD5 (RFC2385). */
 852                 /*
 853                  * Check that TCP_MD5SIG is enabled in tcpcb to
 854                  * account the size needed to set this TCP option.
 855                  */
 856                 if (tp->t_flags & TF_SIGNATURE)
 857                         to.to_flags |= TOF_SIGNATURE;
 858 #endif /* TCP_SIGNATURE */
 859
 860                 /* Processing the options. */
 861                 hdrlen += optlen = tcp_addoptions(&to, opt);
 862                 /*
 863                  * If we wanted a TFO option to be added, but it was unable
 864                  * to fit, ensure no data is sent.
 865                  */
 866                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
 867                     !(to.to_flags & TOF_FASTOPEN))
 868                         len = 0;
 869         }
 870
 871         /*
 872          * Adjust data length if insertion of options will
 873          * bump the packet length beyond the t_maxseg length.
 874          * Clear the FIN bit because we cut off the tail of
 875          * the segment.
 876          */
 877         if (len + optlen + ipoptlen > tp->t_maxseg) {
 878                 flags &= ~TH_FIN;
 879
 880                 if (tso) {
 881                         u_int if_hw_tsomax;
 882                         u_int if_hw_tsomaxsegcount;
 883                         u_int if_hw_tsomaxsegsize;
 884                         struct mbuf *mb;
 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                          * Check if we should limit by maximum segment
 919                          * size and count:
 920                          */
 921                         if (if_hw_tsomaxsegcount != 0 &&
 922                             if_hw_tsomaxsegsize != 0) {
 923                                 /*
 924                                  * Subtract one segment for the LINK
 925                                  * and TCP/IP headers mbuf that will
 926                                  * be prepended to this mbuf chain
 927                                  * after the code in this section
 928                                  * limits the number of mbufs in the
 929                                  * chain to if_hw_tsomaxsegcount.
 930                                  */
 931                                 if_hw_tsomaxsegcount -= 1;
 932                                 max_len = 0;
 933                                 mb = sbsndmbuf(&so->so_snd, off, &moff);
 934
 935                                 while (mb != NULL && max_len < len) {
 936                                         u_int mlen;
 937                                         u_int frags;
 938
 939                                         /*
 940                                          * Get length of mbuf fragment
 941                                          * and how many hardware frags,
 942                                          * rounded up, it would use:
 943                                          */
 944                                         mlen = (mb->m_len - moff);
 945                                         frags = howmany(mlen,
 946                                             if_hw_tsomaxsegsize);
 947
 948                                         /* Handle special case: Zero Length Mbuf */
 949                                         if (frags == 0)
 950                                                 frags = 1;
 951
 952                                         /*
 953                                          * Check if the fragment limit
 954                                          * will be reached or exceeded:
 955                                          */
 956                                         if (frags >= if_hw_tsomaxsegcount) {
 957                                                 max_len += min(mlen,
 958                                                     if_hw_tsomaxsegcount *
 959                                                     if_hw_tsomaxsegsize);
 960                                                 break;
 961                                         }
 962                                         max_len += mlen;
 963                                         if_hw_tsomaxsegcount -= frags;
 964                                         moff = 0;
 965                                         mb = mb->m_next;
 966                                 }
 967                                 if (max_len <= 0) {
 968                                         len = 0;
 969                                 } else if (len > max_len) {
 970                                         sendalot = 1;
 971                                         len = max_len;
 972                                 }
 973                         }
 974
 975                         /*
 976                          * Prevent the last segment from being
 977                          * fractional unless the send sockbuf can be
 978                          * emptied:
 979                          */
 980                         max_len = (tp->t_maxseg - optlen);
 981                         if (((uint32_t)off + (uint32_t)len) <
 982                             sbavail(&so->so_snd)) {
 983                                 moff = len % max_len;
 984                                 if (moff != 0) {
 985                                         len -= moff;
 986                                         sendalot = 1;
 987                                 }
 988                         }
 989
 990                         /*
 991                          * In case there are too many small fragments
 992                          * don't use TSO:
 993                          */
 994                         if (len <= max_len) {
 995                                 len = max_len;
 996                                 sendalot = 1;
 997                                 tso = 0;
 998                         }
 999
1000                         /*
1001                          * Send the FIN in a separate segment
1002                          * after the bulk sending is done.
1003                          * We don't trust the TSO implementations
1004                          * to clear the FIN flag on all but the
1005                          * last segment.
1006                          */
1007                         if (tp->t_flags & TF_NEEDFIN)
1008                                 sendalot = 1;
1009
1010                 } else {
1011                         len = tp->t_maxseg - optlen - ipoptlen;
1012                         sendalot = 1;
1013                         if (dont_sendalot)
1014                                 sendalot = 0;
1015                 }
1016         } else
1017                 tso = 0;
1018
1019         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1020             ("%s: len > IP_MAXPACKET", __func__));
1021
1022 /*#ifdef DIAGNOSTIC*/
1023 #ifdef INET6
1024         if (max_linkhdr + hdrlen > MCLBYTES)
1025 #else
1026         if (max_linkhdr + hdrlen > MHLEN)
1027 #endif
1028                 panic("tcphdr too big");
1029 /*#endif*/
1030
1031         /*
1032          * This KASSERT is here to catch edge cases at a well defined place.
1033          * Before, those had triggered (random) panic conditions further down.
1034          */
1035         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1036
1037         /*
1038          * Grab a header mbuf, attaching a copy of data to
1039          * be transmitted, and initialize the header from
1040          * the template for sends on this connection.
1041          */
1042         if (len) {
1043                 struct mbuf *mb;
1044                 u_int moff;
1045
1046                 if ((tp->t_flags & TF_FORCEDATA) && len == 1)
1047                         TCPSTAT_INC(tcps_sndprobe);
1048                 else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1049                         tp->t_sndrexmitpack++;
1050                         TCPSTAT_INC(tcps_sndrexmitpack);
1051                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1052                 } else {
1053                         TCPSTAT_INC(tcps_sndpack);
1054                         TCPSTAT_ADD(tcps_sndbyte, len);
1055                 }
1056 #ifdef INET6
1057                 if (MHLEN < hdrlen + max_linkhdr)
1058                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1059                 else
1060 #endif
1061                         m = m_gethdr(M_NOWAIT, MT_DATA);
1062
1063                 if (m == NULL) {
1064                         SOCKBUF_UNLOCK(&so->so_snd);
1065                         error = ENOBUFS;
1066                         sack_rxmit = 0;
1067                         goto out;
1068                 }
1069
1070                 m->m_data += max_linkhdr;
1071                 m->m_len = hdrlen;
1072
1073                 /*
1074                  * Start the m_copy functions from the closest mbuf
1075                  * to the offset in the socket buffer chain.
1076                  */
1077                 mb = sbsndptr(&so->so_snd, off, len, &moff);
1078
1079                 if (len <= MHLEN - hdrlen - max_linkhdr) {
1080                         m_copydata(mb, moff, len,
1081                             mtod(m, caddr_t) + hdrlen);
1082                         m->m_len += len;
1083                 } else {
1084                         m->m_next = m_copym(mb, moff, len, M_NOWAIT);
1085                         if (m->m_next == NULL) {
1086                                 SOCKBUF_UNLOCK(&so->so_snd);
1087                                 (void) m_free(m);
1088                                 error = ENOBUFS;
1089                                 sack_rxmit = 0;
1090                                 goto out;
1091                         }
1092                 }
1093
1094                 /*
1095                  * If we're sending everything we've got, set PUSH.
1096                  * (This will keep happy those implementations which only
1097                  * give data to the user when a buffer fills or
1098                  * a PUSH comes in.)
1099                  */
1100                 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1101                     !(flags & TH_SYN))
1102                         flags |= TH_PUSH;
1103                 SOCKBUF_UNLOCK(&so->so_snd);
1104         } else {
1105                 SOCKBUF_UNLOCK(&so->so_snd);
1106                 if (tp->t_flags & TF_ACKNOW)
1107                         TCPSTAT_INC(tcps_sndacks);
1108                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1109                         TCPSTAT_INC(tcps_sndctrl);
1110                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1111                         TCPSTAT_INC(tcps_sndurg);
1112                 else
1113                         TCPSTAT_INC(tcps_sndwinup);
1114
1115                 m = m_gethdr(M_NOWAIT, MT_DATA);
1116                 if (m == NULL) {
1117                         error = ENOBUFS;
1118                         sack_rxmit = 0;
1119                         goto out;
1120                 }
1121 #ifdef INET6
1122                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1123                     MHLEN >= hdrlen) {
1124                         M_ALIGN(m, hdrlen);
1125                 } else
1126 #endif
1127                 m->m_data += max_linkhdr;
1128                 m->m_len = hdrlen;
1129         }
1130         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1131         m->m_pkthdr.rcvif = (struct ifnet *)0;
1132 #ifdef MAC
1133         mac_inpcb_create_mbuf(tp->t_inpcb, m);
1134 #endif
1135 #ifdef INET6
1136         if (isipv6) {
1137                 ip6 = mtod(m, struct ip6_hdr *);
1138                 th = (struct tcphdr *)(ip6 + 1);
1139                 tcpip_fillheaders(tp->t_inpcb, ip6, th);
1140         } else
1141 #endif /* INET6 */
1142         {
1143                 ip = mtod(m, struct ip *);
1144 #ifdef TCPDEBUG
1145                 ipov = (struct ipovly *)ip;
1146 #endif
1147                 th = (struct tcphdr *)(ip + 1);
1148                 tcpip_fillheaders(tp->t_inpcb, ip, th);
1149         }
1150
1151         /*
1152          * Fill in fields, remembering maximum advertised
1153          * window for use in delaying messages about window sizes.
1154          * If resending a FIN, be sure not to use a new sequence number.
1155          */
1156         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1157             tp->snd_nxt == tp->snd_max)
1158                 tp->snd_nxt--;
1159         /*
1160          * If we are starting a connection, send ECN setup
1161          * SYN packet. If we are on a retransmit, we may
1162          * resend those bits a number of times as per
1163          * RFC 3168.
1164          */
1165         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1166                 if (tp->t_rxtshift >= 1) {
1167                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1168                                 flags |= TH_ECE|TH_CWR;
1169                 } else
1170                         flags |= TH_ECE|TH_CWR;
1171         }
1172
1173         if (tp->t_state == TCPS_ESTABLISHED &&
1174             (tp->t_flags & TF_ECN_PERMIT)) {
1175                 /*
1176                  * If the peer has ECN, mark data packets with
1177                  * ECN capable transmission (ECT).
1178                  * Ignore pure ack packets, retransmissions and window probes.
1179                  */
1180                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1181                     !((tp->t_flags & TF_FORCEDATA) && len == 1)) {
1182 #ifdef INET6
1183                         if (isipv6)
1184                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1185                         else
1186 #endif
1187                                 ip->ip_tos |= IPTOS_ECN_ECT0;
1188                         TCPSTAT_INC(tcps_ecn_ect0);
1189                 }
1190
1191                 /*
1192                  * Reply with proper ECN notifications.
1193                  */
1194                 if (tp->t_flags & TF_ECN_SND_CWR) {
1195                         flags |= TH_CWR;
1196                         tp->t_flags &= ~TF_ECN_SND_CWR;
1197                 }
1198                 if (tp->t_flags & TF_ECN_SND_ECE)
1199                         flags |= TH_ECE;
1200         }
1201
1202         /*
1203          * If we are doing retransmissions, then snd_nxt will
1204          * not reflect the first unsent octet.  For ACK only
1205          * packets, we do not want the sequence number of the
1206          * retransmitted packet, we want the sequence number
1207          * of the next unsent octet.  So, if there is no data
1208          * (and no SYN or FIN), use snd_max instead of snd_nxt
1209          * when filling in ti_seq.  But if we are in persist
1210          * state, snd_max might reflect one byte beyond the
1211          * right edge of the window, so use snd_nxt in that
1212          * case, since we know we aren't doing a retransmission.
1213          * (retransmit and persist are mutually exclusive...)
1214          */
1215         if (sack_rxmit == 0) {
1216                 if (len || (flags & (TH_SYN|TH_FIN)) ||
1217                     tcp_timer_active(tp, TT_PERSIST))
1218                         th->th_seq = htonl(tp->snd_nxt);
1219                 else
1220                         th->th_seq = htonl(tp->snd_max);
1221         } else {
1222                 th->th_seq = htonl(p->rxmit);
1223                 p->rxmit += len;
1224                 tp->sackhint.sack_bytes_rexmit += len;
1225         }
1226         th->th_ack = htonl(tp->rcv_nxt);
1227         if (optlen) {
1228                 bcopy(opt, th + 1, optlen);
1229                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1230         }
1231         th->th_flags = flags;
1232         /*
1233          * Calculate receive window.  Don't shrink window,
1234          * but avoid silly window syndrome.
1235          */
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                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1253
1254         /*
1255          * Adjust the RXWIN0SENT flag - indicate that we have advertised
1256          * a 0 window.  This may cause the remote transmitter to stall.  This
1257          * flag tells soreceive() to disable delayed acknowledgements when
1258          * draining the buffer.  This can occur if the receiver is attempting
1259          * to read more data than can be buffered prior to transmitting on
1260          * the connection.
1261          */
1262         if (th->th_win == 0) {
1263                 tp->t_sndzerowin++;
1264                 tp->t_flags |= TF_RXWIN0SENT;
1265         } else
1266                 tp->t_flags &= ~TF_RXWIN0SENT;
1267         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1268                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1269                 th->th_flags |= TH_URG;
1270         } else
1271                 /*
1272                  * If no urgent pointer to send, then we pull
1273                  * the urgent pointer to the left edge of the send window
1274                  * so that it doesn't drift into the send window on sequence
1275                  * number wraparound.
1276                  */
1277                 tp->snd_up = tp->snd_una;               /* drag it along */
1278
1279         /*
1280          * Put TCP length in extended header, and then
1281          * checksum extended header and data.
1282          */
1283         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1284         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1285
1286 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1287         if (to.to_flags & TOF_SIGNATURE) {
1288                 /*
1289                  * Calculate MD5 signature and put it into the place
1290                  * determined before.
1291                  * NOTE: since TCP options buffer doesn't point into
1292                  * mbuf's data, calculate offset and use it.
1293                  */
1294                 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1295                     (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1296                         /*
1297                          * Do not send segment if the calculation of MD5
1298                          * digest has failed.
1299                          */
1300                         m_freem(m);
1301                         goto out;
1302                 }
1303         }
1304 #endif
1305 #ifdef INET6
1306         if (isipv6) {
1307                 /*
1308                  * There is no need to fill in ip6_plen right now.
1309                  * It will be filled later by ip6_output.
1310                  */
1311                 m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1312                 th->th_sum = in6_cksum_pseudo(ip6, sizeof(struct tcphdr) +
1313                     optlen + len, IPPROTO_TCP, 0);
1314         }
1315 #endif
1316 #if defined(INET6) && defined(INET)
1317         else
1318 #endif
1319 #ifdef INET
1320         {
1321                 m->m_pkthdr.csum_flags = CSUM_TCP;
1322                 th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr,
1323                     htons(sizeof(struct tcphdr) + IPPROTO_TCP + len + optlen));
1324
1325                 /* IP version must be set here for ipv4/ipv6 checking later */
1326                 KASSERT(ip->ip_v == IPVERSION,
1327                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1328         }
1329 #endif
1330
1331         /*
1332          * Enable TSO and specify the size of the segments.
1333          * The TCP pseudo header checksum is always provided.
1334          */
1335         if (tso) {
1336                 KASSERT(len > tp->t_maxseg - optlen,
1337                     ("%s: len <= tso_segsz", __func__));
1338                 m->m_pkthdr.csum_flags |= CSUM_TSO;
1339                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1340         }
1341
1342 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
1343         KASSERT(len + hdrlen + ipoptlen - ipsec_optlen == m_length(m, NULL),
1344             ("%s: mbuf chain shorter than expected: %d + %u + %u - %u != %u",
1345             __func__, len, hdrlen, ipoptlen, ipsec_optlen, m_length(m, NULL)));
1346 #else
1347         KASSERT(len + hdrlen + ipoptlen == m_length(m, NULL),
1348             ("%s: mbuf chain shorter than expected: %d + %u + %u != %u",
1349             __func__, len, hdrlen, ipoptlen, m_length(m, NULL)));
1350 #endif
1351
1352 #ifdef TCP_HHOOK
1353         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1354         hhook_run_tcp_est_out(tp, th, &to, len, tso);
1355 #endif
1356
1357 #ifdef TCPDEBUG
1358         /*
1359          * Trace.
1360          */
1361         if (so->so_options & SO_DEBUG) {
1362                 u_short save = 0;
1363 #ifdef INET6
1364                 if (!isipv6)
1365 #endif
1366                 {
1367                         save = ipov->ih_len;
1368                         ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1369                 }
1370                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1371 #ifdef INET6
1372                 if (!isipv6)
1373 #endif
1374                 ipov->ih_len = save;
1375         }
1376 #endif /* TCPDEBUG */
1377         TCP_PROBE3(debug__output, tp, th, m);
1378
1379         /* We're getting ready to send; log now. */
1380         TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1381             len, NULL, false);
1382
1383         /*
1384          * Fill in IP length and desired time to live and
1385          * send to IP level.  There should be a better way
1386          * to handle ttl and tos; we could keep them in
1387          * the template, but need a way to checksum without them.
1388          */
1389         /*
1390          * m->m_pkthdr.len should have been set before checksum calculation,
1391          * because in6_cksum() need it.
1392          */
1393 #ifdef INET6
1394         if (isipv6) {
1395                 /*
1396                  * we separately set hoplimit for every segment, since the
1397                  * user might want to change the value via setsockopt.
1398                  * Also, desired default hop limit might be changed via
1399                  * Neighbor Discovery.
1400                  */
1401                 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1402
1403                 /*
1404                  * Set the packet size here for the benefit of DTrace probes.
1405                  * ip6_output() will set it properly; it's supposed to include
1406                  * the option header lengths as well.
1407                  */
1408                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1409
1410                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1411                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1412                 else
1413                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1414
1415                 if (tp->t_state == TCPS_SYN_SENT)
1416                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1417
1418                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1419
1420 #ifdef TCPPCAP
1421                 /* Save packet, if requested. */
1422                 tcp_pcap_add(th, m, &(tp->t_outpkts));
1423 #endif
1424
1425                 /* TODO: IPv6 IP6TOS_ECT bit on */
1426                 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1427                     &tp->t_inpcb->inp_route6,
1428                     ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1429                     NULL, NULL, tp->t_inpcb);
1430
1431                 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_rt != NULL)
1432                         mtu = tp->t_inpcb->inp_route6.ro_rt->rt_mtu;
1433         }
1434 #endif /* INET6 */
1435 #if defined(INET) && defined(INET6)
1436         else
1437 #endif
1438 #ifdef INET
1439     {
1440         ip->ip_len = htons(m->m_pkthdr.len);
1441 #ifdef INET6
1442         if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1443                 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1444 #endif /* INET6 */
1445         /*
1446          * If we do path MTU discovery, then we set DF on every packet.
1447          * This might not be the best thing to do according to RFC3390
1448          * Section 2. However the tcp hostcache migitates the problem
1449          * so it affects only the first tcp connection with a host.
1450          *
1451          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1452          */
1453         if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1454                 ip->ip_off |= htons(IP_DF);
1455                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1456         } else {
1457                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1458         }
1459
1460         if (tp->t_state == TCPS_SYN_SENT)
1461                 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1462
1463         TCP_PROBE5(send, NULL, tp, ip, tp, th);
1464
1465 #ifdef TCPPCAP
1466         /* Save packet, if requested. */
1467         tcp_pcap_add(th, m, &(tp->t_outpkts));
1468 #endif
1469
1470         error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1471             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1472             tp->t_inpcb);
1473
1474         if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_rt != NULL)
1475                 mtu = tp->t_inpcb->inp_route.ro_rt->rt_mtu;
1476     }
1477 #endif /* INET */
1478
1479 out:
1480         /*
1481          * In transmit state, time the transmission and arrange for
1482          * the retransmit.  In persist state, just set snd_max.
1483          */
1484         if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1485             !tcp_timer_active(tp, TT_PERSIST)) {
1486                 tcp_seq startseq = tp->snd_nxt;
1487
1488                 /*
1489                  * Advance snd_nxt over sequence space of this segment.
1490                  */
1491                 if (flags & (TH_SYN|TH_FIN)) {
1492                         if (flags & TH_SYN)
1493                                 tp->snd_nxt++;
1494                         if (flags & TH_FIN) {
1495                                 tp->snd_nxt++;
1496                                 tp->t_flags |= TF_SENTFIN;
1497                         }
1498                 }
1499                 if (sack_rxmit)
1500                         goto timer;
1501                 tp->snd_nxt += len;
1502                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1503                         tp->snd_max = tp->snd_nxt;
1504                         /*
1505                          * Time this transmission if not a retransmission and
1506                          * not currently timing anything.
1507                          */
1508                         if (tp->t_rtttime == 0) {
1509                                 tp->t_rtttime = ticks;
1510                                 tp->t_rtseq = startseq;
1511                                 TCPSTAT_INC(tcps_segstimed);
1512                         }
1513                 }
1514
1515                 /*
1516                  * Set retransmit timer if not currently set,
1517                  * and not doing a pure ack or a keep-alive probe.
1518                  * Initial value for retransmit timer is smoothed
1519                  * round-trip time + 2 * round-trip time variance.
1520                  * Initialize shift counter which is used for backoff
1521                  * of retransmit time.
1522                  */
1523 timer:
1524                 if (!tcp_timer_active(tp, TT_REXMT) &&
1525                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1526                      (tp->snd_nxt != tp->snd_una))) {
1527                         if (tcp_timer_active(tp, TT_PERSIST)) {
1528                                 tcp_timer_activate(tp, TT_PERSIST, 0);
1529                                 tp->t_rxtshift = 0;
1530                         }
1531                         tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1532                 } else if (len == 0 && sbavail(&so->so_snd) &&
1533                     !tcp_timer_active(tp, TT_REXMT) &&
1534                     !tcp_timer_active(tp, TT_PERSIST)) {
1535                         /*
1536                          * Avoid a situation where we do not set persist timer
1537                          * after a zero window condition. For example:
1538                          * 1) A -> B: packet with enough data to fill the window
1539                          * 2) B -> A: ACK for #1 + new data (0 window
1540                          *    advertisement)
1541                          * 3) A -> B: ACK for #2, 0 len packet
1542                          *
1543                          * In this case, A will not activate the persist timer,
1544                          * because it chose to send a packet. Unless tcp_output
1545                          * is called for some other reason (delayed ack timer,
1546                          * another input packet from B, socket syscall), A will
1547                          * not send zero window probes.
1548                          *
1549                          * So, if you send a 0-length packet, but there is data
1550                          * in the socket buffer, and neither the rexmt or
1551                          * persist timer is already set, then activate the
1552                          * persist timer.
1553                          */
1554                         tp->t_rxtshift = 0;
1555                         tcp_setpersist(tp);
1556                 }
1557         } else {
1558                 /*
1559                  * Persist case, update snd_max but since we are in
1560                  * persist mode (no window) we do not update snd_nxt.
1561                  */
1562                 int xlen = len;
1563                 if (flags & TH_SYN)
1564                         ++xlen;
1565                 if (flags & TH_FIN) {
1566                         ++xlen;
1567                         tp->t_flags |= TF_SENTFIN;
1568                 }
1569                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1570                         tp->snd_max = tp->snd_nxt + xlen;
1571         }
1572
1573         if (error) {
1574                 /* Record the error. */
1575                 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1576                     error, 0, NULL, false);
1577
1578                 /*
1579                  * We know that the packet was lost, so back out the
1580                  * sequence number advance, if any.
1581                  *
1582                  * If the error is EPERM the packet got blocked by the
1583                  * local firewall.  Normally we should terminate the
1584                  * connection but the blocking may have been spurious
1585                  * due to a firewall reconfiguration cycle.  So we treat
1586                  * it like a packet loss and let the retransmit timer and
1587                  * timeouts do their work over time.
1588                  * XXX: It is a POLA question whether calling tcp_drop right
1589                  * away would be the really correct behavior instead.
1590                  */
1591                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1592                     !tcp_timer_active(tp, TT_PERSIST)) &&
1593                     ((flags & TH_SYN) == 0) &&
1594                     (error != EPERM)) {
1595                         if (sack_rxmit) {
1596                                 p->rxmit -= len;
1597                                 tp->sackhint.sack_bytes_rexmit -= len;
1598                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1599                                     ("sackhint bytes rtx >= 0"));
1600                         } else
1601                                 tp->snd_nxt -= len;
1602                 }
1603                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
1604                 switch (error) {
1605                 case EACCES:
1606                 case EPERM:
1607                         tp->t_softerror = error;
1608                         return (error);
1609                 case ENOBUFS:
1610                         TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1611                         tp->snd_cwnd = tp->t_maxseg;
1612                         return (0);
1613                 case EMSGSIZE:
1614                         /*
1615                          * For some reason the interface we used initially
1616                          * to send segments changed to another or lowered
1617                          * its MTU.
1618                          * If TSO was active we either got an interface
1619                          * without TSO capabilits or TSO was turned off.
1620                          * If we obtained mtu from ip_output() then update
1621                          * it and try again.
1622                          */
1623                         if (tso)
1624                                 tp->t_flags &= ~TF_TSO;
1625                         if (mtu != 0) {
1626                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1627                                 goto again;
1628                         }
1629                         return (error);
1630                 case EHOSTDOWN:
1631                 case EHOSTUNREACH:
1632                 case ENETDOWN:
1633                 case ENETUNREACH:
1634                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
1635                                 tp->t_softerror = error;
1636                                 return (0);
1637                         }
1638                         /* FALLTHROUGH */
1639                 default:
1640                         return (error);
1641                 }
1642         }
1643         TCPSTAT_INC(tcps_sndtotal);
1644
1645         /*
1646          * Data sent (as far as we can tell).
1647          * If this advertises a larger window than any other segment,
1648          * then remember the size of the advertised window.
1649          * Any pending ACK has now been sent.
1650          */
1651         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1652                 tp->rcv_adv = tp->rcv_nxt + recwin;
1653         tp->last_ack_sent = tp->rcv_nxt;
1654         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1655         if (tcp_timer_active(tp, TT_DELACK))
1656                 tcp_timer_activate(tp, TT_DELACK, 0);
1657 #if 0
1658         /*
1659          * This completely breaks TCP if newreno is turned on.  What happens
1660          * is that if delayed-acks are turned on on the receiver, this code
1661          * on the transmitter effectively destroys the TCP window, forcing
1662          * it to four packets (1.5Kx4 = 6K window).
1663          */
1664         if (sendalot && --maxburst)
1665                 goto again;
1666 #endif
1667         if (sendalot)
1668                 goto again;
1669         return (0);
1670 }
1671
1672 void
1673 tcp_setpersist(struct tcpcb *tp)
1674 {
1675         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1676         int tt;
1677
1678         tp->t_flags &= ~TF_PREVVALID;
1679         if (tcp_timer_active(tp, TT_REXMT))
1680                 panic("tcp_setpersist: retransmit pending");
1681         /*
1682          * Start/restart persistence timer.
1683          */
1684         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1685                       tcp_persmin, tcp_persmax);
1686         tcp_timer_activate(tp, TT_PERSIST, tt);
1687         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1688                 tp->t_rxtshift++;
1689 }
1690
1691 /*
1692  * Insert TCP options according to the supplied parameters to the place
1693  * optp in a consistent way.  Can handle unaligned destinations.
1694  *
1695  * The order of the option processing is crucial for optimal packing and
1696  * alignment for the scarce option space.
1697  *
1698  * The optimal order for a SYN/SYN-ACK segment is:
1699  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1700  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1701  *
1702  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1703  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1704  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1705  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1706  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1707  */
1708 int
1709 tcp_addoptions(struct tcpopt *to, u_char *optp)
1710 {
1711         u_int32_t mask, optlen = 0;
1712
1713         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1714                 if ((to->to_flags & mask) != mask)
1715                         continue;
1716                 if (optlen == TCP_MAXOLEN)
1717                         break;
1718                 switch (to->to_flags & mask) {
1719                 case TOF_MSS:
1720                         while (optlen % 4) {
1721                                 optlen += TCPOLEN_NOP;
1722                                 *optp++ = TCPOPT_NOP;
1723                         }
1724                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1725                                 continue;
1726                         optlen += TCPOLEN_MAXSEG;
1727                         *optp++ = TCPOPT_MAXSEG;
1728                         *optp++ = TCPOLEN_MAXSEG;
1729                         to->to_mss = htons(to->to_mss);
1730                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1731                         optp += sizeof(to->to_mss);
1732                         break;
1733                 case TOF_SCALE:
1734                         while (!optlen || optlen % 2 != 1) {
1735                                 optlen += TCPOLEN_NOP;
1736                                 *optp++ = TCPOPT_NOP;
1737                         }
1738                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1739                                 continue;
1740                         optlen += TCPOLEN_WINDOW;
1741                         *optp++ = TCPOPT_WINDOW;
1742                         *optp++ = TCPOLEN_WINDOW;
1743                         *optp++ = to->to_wscale;
1744                         break;
1745                 case TOF_SACKPERM:
1746                         while (optlen % 2) {
1747                                 optlen += TCPOLEN_NOP;
1748                                 *optp++ = TCPOPT_NOP;
1749                         }
1750                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1751                                 continue;
1752                         optlen += TCPOLEN_SACK_PERMITTED;
1753                         *optp++ = TCPOPT_SACK_PERMITTED;
1754                         *optp++ = TCPOLEN_SACK_PERMITTED;
1755                         break;
1756                 case TOF_TS:
1757                         while (!optlen || optlen % 4 != 2) {
1758                                 optlen += TCPOLEN_NOP;
1759                                 *optp++ = TCPOPT_NOP;
1760                         }
1761                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1762                                 continue;
1763                         optlen += TCPOLEN_TIMESTAMP;
1764                         *optp++ = TCPOPT_TIMESTAMP;
1765                         *optp++ = TCPOLEN_TIMESTAMP;
1766                         to->to_tsval = htonl(to->to_tsval);
1767                         to->to_tsecr = htonl(to->to_tsecr);
1768                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1769                         optp += sizeof(to->to_tsval);
1770                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1771                         optp += sizeof(to->to_tsecr);
1772                         break;
1773                 case TOF_SIGNATURE:
1774                         {
1775                         int siglen = TCPOLEN_SIGNATURE - 2;
1776
1777                         while (!optlen || optlen % 4 != 2) {
1778                                 optlen += TCPOLEN_NOP;
1779                                 *optp++ = TCPOPT_NOP;
1780                         }
1781                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1782                                 to->to_flags &= ~TOF_SIGNATURE;
1783                                 continue;
1784                         }
1785                         optlen += TCPOLEN_SIGNATURE;
1786                         *optp++ = TCPOPT_SIGNATURE;
1787                         *optp++ = TCPOLEN_SIGNATURE;
1788                         to->to_signature = optp;
1789                         while (siglen--)
1790                                  *optp++ = 0;
1791                         break;
1792                         }
1793                 case TOF_SACK:
1794                         {
1795                         int sackblks = 0;
1796                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
1797                         tcp_seq sack_seq;
1798
1799                         while (!optlen || optlen % 4 != 2) {
1800                                 optlen += TCPOLEN_NOP;
1801                                 *optp++ = TCPOPT_NOP;
1802                         }
1803                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1804                                 continue;
1805                         optlen += TCPOLEN_SACKHDR;
1806                         *optp++ = TCPOPT_SACK;
1807                         sackblks = min(to->to_nsacks,
1808                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1809                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1810                         while (sackblks--) {
1811                                 sack_seq = htonl(sack->start);
1812                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1813                                 optp += sizeof(sack_seq);
1814                                 sack_seq = htonl(sack->end);
1815                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1816                                 optp += sizeof(sack_seq);
1817                                 optlen += TCPOLEN_SACK;
1818                                 sack++;
1819                         }
1820                         TCPSTAT_INC(tcps_sack_send_blocks);
1821                         break;
1822                         }
1823                 case TOF_FASTOPEN:
1824                         {
1825                         int total_len;
1826
1827                         /* XXX is there any point to aligning this option? */
1828                         total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1829                         if (TCP_MAXOLEN - optlen < total_len) {
1830                                 to->to_flags &= ~TOF_FASTOPEN;
1831                                 continue;
1832                         }
1833                         *optp++ = TCPOPT_FAST_OPEN;
1834                         *optp++ = total_len;
1835                         if (to->to_tfo_len > 0) {
1836                                 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1837                                 optp += to->to_tfo_len;
1838                         }
1839                         optlen += total_len;
1840                         break;
1841                         }
1842                 default:
1843                         panic("%s: unknown TCP option type", __func__);
1844                         break;
1845                 }
1846         }
1847
1848         /* Terminate and pad TCP options to a 4 byte boundary. */
1849         if (optlen % 4) {
1850                 optlen += TCPOLEN_EOL;
1851                 *optp++ = TCPOPT_EOL;
1852         }
1853         /*
1854          * According to RFC 793 (STD0007):
1855          *   "The content of the header beyond the End-of-Option option
1856          *    must be header padding (i.e., zero)."
1857          *   and later: "The padding is composed of zeros."
1858          */
1859         while (optlen % 4) {
1860                 optlen += TCPOLEN_PAD;
1861                 *optp++ = TCPOPT_PAD;
1862         }
1863
1864         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1865         return (optlen);
1866 }
1867
1868 /*
1869  * This is a copy of m_copym(), taking the TSO segment size/limit
1870  * constraints into account, and advancing the sndptr as it goes.
1871  */
1872 struct mbuf *
1873 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1874     int32_t seglimit, int32_t segsize, struct sockbuf *sb)
1875 {
1876         struct mbuf *n, **np;
1877         struct mbuf *top;
1878         int32_t off = off0;
1879         int32_t len = *plen;
1880         int32_t fragsize;
1881         int32_t len_cp = 0;
1882         int32_t *pkthdrlen;
1883         uint32_t mlen, frags;
1884         bool copyhdr;
1885
1886
1887         KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1888         KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1889         if (off == 0 && m->m_flags & M_PKTHDR)
1890                 copyhdr = true;
1891         else
1892                 copyhdr = false;
1893         while (off > 0) {
1894                 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1895                 if (off < m->m_len)
1896                         break;
1897                 off -= m->m_len;
1898                 if ((sb) && (m == sb->sb_sndptr)) {
1899                         sb->sb_sndptroff += m->m_len;
1900                         sb->sb_sndptr = m->m_next;
1901                 }
1902                 m = m->m_next;
1903         }
1904         np = &top;
1905         top = NULL;
1906         pkthdrlen = NULL;
1907         while (len > 0) {
1908                 if (m == NULL) {
1909                         KASSERT(len == M_COPYALL,
1910                             ("tcp_m_copym, length > size of mbuf chain"));
1911                         *plen = len_cp;
1912                         if (pkthdrlen != NULL)
1913                                 *pkthdrlen = len_cp;
1914                         break;
1915                 }
1916                 mlen = min(len, m->m_len - off);
1917                 if (seglimit) {
1918                         /*
1919                          * For M_NOMAP mbufs, add 3 segments
1920                          * + 1 in case we are crossing page boundaries
1921                          * + 2 in case the TLS hdr/trailer are used
1922                          * It is cheaper to just add the segments
1923                          * than it is to take the cache miss to look
1924                          * at the mbuf ext_pgs state in detail.
1925                          */
1926                         if (m->m_flags & M_NOMAP) {
1927                                 fragsize = min(segsize, PAGE_SIZE);
1928                                 frags = 3;
1929                         } else {
1930                                 fragsize = segsize;
1931                                 frags = 0;
1932                         }
1933
1934                         /* Break if we really can't fit anymore. */
1935                         if ((frags + 1) >= seglimit) {
1936                                 *plen = len_cp;
1937                                 if (pkthdrlen != NULL)
1938                                         *pkthdrlen = len_cp;
1939                                 break;
1940                         }
1941
1942                         /*
1943                          * Reduce size if you can't copy the whole
1944                          * mbuf. If we can't copy the whole mbuf, also
1945                          * adjust len so the loop will end after this
1946                          * mbuf.
1947                          */
1948                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
1949                                 mlen = (seglimit - frags - 1) * fragsize;
1950                                 len = mlen;
1951                                 *plen = len_cp + len;
1952                                 if (pkthdrlen != NULL)
1953                                         *pkthdrlen = *plen;
1954                         }
1955                         frags += howmany(mlen, fragsize);
1956                         if (frags == 0)
1957                                 frags++;
1958                         seglimit -= frags;
1959                         KASSERT(seglimit > 0,
1960                             ("%s: seglimit went too low", __func__));
1961                 }
1962                 if (copyhdr)
1963                         n = m_gethdr(M_NOWAIT, m->m_type);
1964                 else
1965                         n = m_get(M_NOWAIT, m->m_type);
1966                 *np = n;
1967                 if (n == NULL)
1968                         goto nospace;
1969                 if (copyhdr) {
1970                         if (!m_dup_pkthdr(n, m, M_NOWAIT))
1971                                 goto nospace;
1972                         if (len == M_COPYALL)
1973                                 n->m_pkthdr.len -= off0;
1974                         else
1975                                 n->m_pkthdr.len = len;
1976                         pkthdrlen = &n->m_pkthdr.len;
1977                         copyhdr = false;
1978                 }
1979                 n->m_len = mlen;
1980                 len_cp += n->m_len;
1981                 if (m->m_flags & M_EXT) {
1982                         n->m_data = m->m_data + off;
1983                         mb_dupcl(n, m);
1984                 } else
1985                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
1986                             (u_int)n->m_len);
1987
1988                 if (sb && (sb->sb_sndptr == m) &&
1989                     ((n->m_len + off) >= m->m_len) && m->m_next) {
1990                         sb->sb_sndptroff += m->m_len;
1991                         sb->sb_sndptr = m->m_next;
1992                 }
1993                 off = 0;
1994                 if (len != M_COPYALL) {
1995                         len -= n->m_len;
1996                 }
1997                 m = m->m_next;
1998                 np = &n->m_next;
1999         }
2000         return (top);
2001 nospace:
2002         m_freem(top);
2003         return (NULL);
2004 }
2005
2006 void
2007 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2008 {
2009
2010         /*
2011          * Automatic sizing of send socket buffer.  Often the send buffer
2012          * size is not optimally adjusted to the actual network conditions
2013          * at hand (delay bandwidth product).  Setting the buffer size too
2014          * small limits throughput on links with high bandwidth and high
2015          * delay (eg. trans-continental/oceanic links).  Setting the
2016          * buffer size too big consumes too much real kernel memory,
2017          * especially with many connections on busy servers.
2018          *
2019          * The criteria to step up the send buffer one notch are:
2020          *  1. receive window of remote host is larger than send buffer
2021          *     (with a fudge factor of 5/4th);
2022          *  2. send buffer is filled to 7/8th with data (so we actually
2023          *     have data to make use of it);
2024          *  3. send buffer fill has not hit maximal automatic size;
2025          *  4. our send window (slow start and cogestion controlled) is
2026          *     larger than sent but unacknowledged data in send buffer.
2027          *
2028          * The remote host receive window scaling factor may limit the
2029          * growing of the send buffer before it reaches its allowed
2030          * maximum.
2031          *
2032          * It scales directly with slow start or congestion window
2033          * and does at most one step per received ACK.  This fast
2034          * scaling has the drawback of growing the send buffer beyond
2035          * what is strictly necessary to make full use of a given
2036          * delay*bandwidth product.  However testing has shown this not
2037          * to be much of an problem.  At worst we are trading wasting
2038          * of available bandwidth (the non-use of it) for wasting some
2039          * socket buffer memory.
2040          *
2041          * TODO: Shrink send buffer during idle periods together
2042          * with congestion window.  Requires another timer.  Has to
2043          * wait for upcoming tcp timer rewrite.
2044          *
2045          * XXXGL: should there be used sbused() or sbavail()?
2046          */
2047         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2048                 int lowat;
2049
2050                 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2051                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2052                     sbused(&so->so_snd) >=
2053                     (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2054                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2055                     sendwin >= (sbused(&so->so_snd) -
2056                     (tp->snd_nxt - tp->snd_una))) {
2057                         if (!sbreserve_locked(&so->so_snd,
2058                             min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2059                              V_tcp_autosndbuf_max), so, curthread))
2060                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2061                 }
2062         }
2063 }