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
  32 #include <sys/cdefs.h>
  33 #include "opt_inet.h"
  34 #include "opt_inet6.h"
  35 #include "opt_ipsec.h"
  36 #include "opt_kern_tls.h"
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/arb.h>
  41 #include <sys/domain.h>
  42 #ifdef TCP_HHOOK
  43 #include <sys/hhook.h>
  44 #endif
  45 #include <sys/kernel.h>
  46 #ifdef KERN_TLS
  47 #include <sys/ktls.h>
  48 #endif
  49 #include <sys/lock.h>
  50 #include <sys/mbuf.h>
  51 #include <sys/mutex.h>
  52 #include <sys/protosw.h>
  53 #include <sys/qmath.h>
  54 #include <sys/sdt.h>
  55 #include <sys/socket.h>
  56 #include <sys/socketvar.h>
  57 #include <sys/sysctl.h>
  58 #include <sys/stats.h>
  59
  60 #include <net/if.h>
  61 #include <net/route.h>
  62 #include <net/route/nhop.h>
  63 #include <net/vnet.h>
  64
  65 #include <netinet/in.h>
  66 #include <netinet/in_kdtrace.h>
  67 #include <netinet/in_systm.h>
  68 #include <netinet/ip.h>
  69 #include <netinet/in_pcb.h>
  70 #include <netinet/ip_var.h>
  71 #include <netinet/ip_options.h>
  72 #ifdef INET6
  73 #include <netinet6/in6_pcb.h>
  74 #include <netinet/ip6.h>
  75 #include <netinet6/ip6_var.h>
  76 #endif
  77 #include <netinet/tcp.h>
  78 #define TCPOUTFLAGS
  79 #include <netinet/tcp_fsm.h>
  80 #include <netinet/tcp_seq.h>
  81 #include <netinet/tcp_var.h>
  82 #include <netinet/tcp_log_buf.h>
  83 #include <netinet/tcp_syncache.h>
  84 #include <netinet/tcp_timer.h>
  85 #include <netinet/tcpip.h>
  86 #include <netinet/cc/cc.h>
  87 #include <netinet/tcp_fastopen.h>
  88 #ifdef TCPPCAP
  89 #include <netinet/tcp_pcap.h>
  90 #endif
  91 #ifdef TCP_OFFLOAD
  92 #include <netinet/tcp_offload.h>
  93 #endif
  94 #include <netinet/tcp_ecn.h>
  95
  96 #include <netipsec/ipsec_support.h>
  97
  98 #include <netinet/udp.h>
  99 #include <netinet/udp_var.h>
 100 #include <machine/in_cksum.h>
 101
 102 #include <security/mac/mac_framework.h>
 103
 104 VNET_DEFINE(int, path_mtu_discovery) = 1;
 105 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
 106         &VNET_NAME(path_mtu_discovery), 1,
 107         "Enable Path MTU Discovery");
 108
 109 VNET_DEFINE(int, tcp_do_tso) = 1;
 110 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
 111         &VNET_NAME(tcp_do_tso), 0,
 112         "Enable TCP Segmentation Offload");
 113
 114 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
 115 #define V_tcp_sendspace VNET(tcp_sendspace)
 116 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
 117         &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
 118
 119 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
 120 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
 121         &VNET_NAME(tcp_do_autosndbuf), 0,
 122         "Enable automatic send buffer sizing");
 123
 124 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
 125 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
 126         &VNET_NAME(tcp_autosndbuf_inc), 0,
 127         "Incrementor step size of automatic send buffer");
 128
 129 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
 130 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
 131         &VNET_NAME(tcp_autosndbuf_max), 0,
 132         "Max size of automatic send buffer");
 133
 134 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
 135 #define V_tcp_sendbuf_auto_lowat        VNET(tcp_sendbuf_auto_lowat)
 136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
 137         &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
 138         "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
 139
 140 /*
 141  * Make sure that either retransmit or persist timer is set for SYN, FIN and
 142  * non-ACK.
 143  */
 144 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)                        \
 145         KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
 146             tcp_timer_active((tp), TT_REXMT) ||                         \
 147             tcp_timer_active((tp), TT_PERSIST),                         \
 148             ("neither rexmt nor persist timer is set"))
 149
 150 #ifdef TCP_HHOOK
 151 /*
 152  * Wrapper for the TCP established output helper hook.
 153  */
 154 void
 155 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
 156     struct tcpopt *to, uint32_t len, int tso)
 157 {
 158         struct tcp_hhook_data hhook_data;
 159
 160         if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
 161                 hhook_data.tp = tp;
 162                 hhook_data.th = th;
 163                 hhook_data.to = to;
 164                 hhook_data.len = len;
 165                 hhook_data.tso = tso;
 166
 167                 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
 168                     &tp->t_osd);
 169         }
 170 }
 171 #endif
 172
 173 /*
 174  * CC wrapper hook functions
 175  */
 176 void
 177 cc_after_idle(struct tcpcb *tp)
 178 {
 179         INP_WLOCK_ASSERT(tptoinpcb(tp));
 180
 181         if (CC_ALGO(tp)->after_idle != NULL)
 182                 CC_ALGO(tp)->after_idle(&tp->t_ccv);
 183 }
 184
 185 /*
 186  * Tcp output routine: figure out what should be sent and send it.
 187  */
 188 int
 189 tcp_default_output(struct tcpcb *tp)
 190 {
 191         struct socket *so = tptosocket(tp);
 192         struct inpcb *inp = tptoinpcb(tp);
 193         int32_t len;
 194         uint32_t recwin, sendwin;
 195         uint16_t flags;
 196         int off, error = 0;     /* Keep compiler happy */
 197         u_int if_hw_tsomaxsegcount = 0;
 198         u_int if_hw_tsomaxsegsize = 0;
 199         struct mbuf *m;
 200         struct ip *ip = NULL;
 201         struct tcphdr *th;
 202         u_char opt[TCP_MAXOLEN];
 203         unsigned ipoptlen, optlen, hdrlen, ulen;
 204 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 205         unsigned ipsec_optlen = 0;
 206 #endif
 207         int idle, sendalot, curticks;
 208         int sack_rxmit, sack_bytes_rxmt;
 209         struct sackhole *p;
 210         int tso, mtu;
 211         struct tcpopt to;
 212         struct udphdr *udp = NULL;
 213         struct tcp_log_buffer *lgb;
 214         unsigned int wanted_cookie = 0;
 215         unsigned int dont_sendalot = 0;
 216 #if 0
 217         int maxburst = TCP_MAXBURST;
 218 #endif
 219 #ifdef INET6
 220         struct ip6_hdr *ip6 = NULL;
 221         const bool isipv6 = (inp->inp_vflag & INP_IPV6) != 0;
 222 #endif
 223 #ifdef KERN_TLS
 224         const bool hw_tls = tp->t_nic_ktls_xmit != 0;
 225 #else
 226         const bool hw_tls = false;
 227 #endif
 228
 229         NET_EPOCH_ASSERT();
 230         INP_WLOCK_ASSERT(inp);
 231
 232 #ifdef TCP_OFFLOAD
 233         if (tp->t_flags & TF_TOE)
 234                 return (tcp_offload_output(tp));
 235 #endif
 236
 237         /*
 238          * For TFO connections in SYN_SENT or SYN_RECEIVED,
 239          * only allow the initial SYN or SYN|ACK and those sent
 240          * by the retransmit timer.
 241          */
 242         if (IS_FASTOPEN(tp->t_flags) &&
 243             ((tp->t_state == TCPS_SYN_SENT) ||
 244              (tp->t_state == TCPS_SYN_RECEIVED)) &&
 245             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
 246             (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
 247                 return (0);
 248
 249         /*
 250          * Determine length of data that should be transmitted,
 251          * and flags that will be used.
 252          * If there is some data or critical controls (SYN, RST)
 253          * to send, then transmit; otherwise, investigate further.
 254          */
 255         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
 256         if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
 257             (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
 258                 cc_after_idle(tp);
 259         tp->t_flags &= ~TF_LASTIDLE;
 260         if (idle) {
 261                 if (tp->t_flags & TF_MORETOCOME) {
 262                         tp->t_flags |= TF_LASTIDLE;
 263                         idle = 0;
 264                 }
 265         }
 266 again:
 267         /*
 268          * If we've recently taken a timeout, snd_max will be greater than
 269          * snd_nxt.  There may be SACK information that allows us to avoid
 270          * resending already delivered data.  Adjust snd_nxt accordingly.
 271          */
 272         if ((tp->t_flags & TF_SACK_PERMIT) &&
 273             SEQ_LT(tp->snd_nxt, tp->snd_max))
 274                 tcp_sack_adjust(tp);
 275         sendalot = 0;
 276         tso = 0;
 277         mtu = 0;
 278         off = tp->snd_nxt - tp->snd_una;
 279         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
 280
 281         flags = tcp_outflags[tp->t_state];
 282         /*
 283          * Send any SACK-generated retransmissions.  If we're explicitly trying
 284          * to send out new data (when sendalot is 1), bypass this function.
 285          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
 286          * we're replacing a (future) new transmission with a retransmission
 287          * now, and we previously incremented snd_cwnd in tcp_input().
 288          */
 289         /*
 290          * Still in sack recovery , reset rxmit flag to zero.
 291          */
 292         sack_rxmit = 0;
 293         sack_bytes_rxmt = 0;
 294         len = 0;
 295         p = NULL;
 296         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
 297             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
 298                 uint32_t cwin;
 299
 300                 cwin =
 301                     imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
 302                 /* Do not retransmit SACK segments beyond snd_recover */
 303                 if (SEQ_GT(p->end, tp->snd_recover)) {
 304                         /*
 305                          * (At least) part of sack hole extends beyond
 306                          * snd_recover. Check to see if we can rexmit data
 307                          * for this hole.
 308                          */
 309                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
 310                                 /*
 311                                  * Can't rexmit any more data for this hole.
 312                                  * That data will be rexmitted in the next
 313                                  * sack recovery episode, when snd_recover
 314                                  * moves past p->rxmit.
 315                                  */
 316                                 p = NULL;
 317                                 goto after_sack_rexmit;
 318                         } else {
 319                                 /* Can rexmit part of the current hole */
 320                                 len = ((int32_t)ulmin(cwin,
 321                                     SEQ_SUB(tp->snd_recover, p->rxmit)));
 322                         }
 323                 } else {
 324                         len = ((int32_t)ulmin(cwin,
 325                             SEQ_SUB(p->end, p->rxmit)));
 326                 }
 327                 if (len > 0) {
 328                         off = SEQ_SUB(p->rxmit, tp->snd_una);
 329                         KASSERT(off >= 0,("%s: sack block to the left of una : %d",
 330                             __func__, off));
 331                         sack_rxmit = 1;
 332                         sendalot = 1;
 333                 }
 334         }
 335 after_sack_rexmit:
 336         /*
 337          * Get standard flags, and add SYN or FIN if requested by 'hidden'
 338          * state flags.
 339          */
 340         if (tp->t_flags & TF_NEEDFIN)
 341                 flags |= TH_FIN;
 342         if (tp->t_flags & TF_NEEDSYN)
 343                 flags |= TH_SYN;
 344
 345         SOCKBUF_LOCK(&so->so_snd);
 346         /*
 347          * If in persist timeout with window of 0, send 1 byte.
 348          * Otherwise, if window is small but nonzero
 349          * and timer expired, we will send what we can
 350          * and go to transmit state.
 351          */
 352         if (tp->t_flags & TF_FORCEDATA) {
 353                 if (sendwin == 0) {
 354                         /*
 355                          * If we still have some data to send, then
 356                          * clear the FIN bit.  Usually this would
 357                          * happen below when it realizes that we
 358                          * aren't sending all the data.  However,
 359                          * if we have exactly 1 byte of unsent data,
 360                          * then it won't clear the FIN bit below,
 361                          * and if we are in persist state, we wind
 362                          * up sending the packet without recording
 363                          * that we sent the FIN bit.
 364                          *
 365                          * We can't just blindly clear the FIN bit,
 366                          * because if we don't have any more data
 367                          * to send then the probe will be the FIN
 368                          * itself.
 369                          */
 370                         if (off < sbused(&so->so_snd))
 371                                 flags &= ~TH_FIN;
 372                         sendwin = 1;
 373                 } else {
 374                         tcp_timer_activate(tp, TT_PERSIST, 0);
 375                         tp->t_rxtshift = 0;
 376                 }
 377         }
 378
 379         /*
 380          * If snd_nxt == snd_max and we have transmitted a FIN, the
 381          * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
 382          * a negative length.  This can also occur when TCP opens up
 383          * its congestion window while receiving additional duplicate
 384          * acks after fast-retransmit because TCP will reset snd_nxt
 385          * to snd_max after the fast-retransmit.
 386          *
 387          * In the normal retransmit-FIN-only case, however, snd_nxt will
 388          * be set to snd_una, the offset will be 0, and the length may
 389          * wind up 0.
 390          *
 391          * If sack_rxmit is true we are retransmitting from the scoreboard
 392          * in which case len is already set.
 393          */
 394         if (sack_rxmit == 0) {
 395                 if (sack_bytes_rxmt == 0) {
 396                         len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
 397                             off);
 398                 } else {
 399                         int32_t cwin;
 400
 401                         /*
 402                          * We are inside of a SACK recovery episode and are
 403                          * sending new data, having retransmitted all the
 404                          * data possible in the scoreboard.
 405                          */
 406                         len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
 407                             off);
 408                         /*
 409                          * Don't remove this (len > 0) check !
 410                          * We explicitly check for len > 0 here (although it
 411                          * isn't really necessary), to work around a gcc
 412                          * optimization issue - to force gcc to compute
 413                          * len above. Without this check, the computation
 414                          * of len is bungled by the optimizer.
 415                          */
 416                         if (len > 0) {
 417                                 cwin = tp->snd_cwnd - imax(0, (int32_t)
 418                                         (tp->snd_nxt - tp->snd_recover)) -
 419                                         sack_bytes_rxmt;
 420                                 if (cwin < 0)
 421                                         cwin = 0;
 422                                 len = imin(len, cwin);
 423                         }
 424                 }
 425         }
 426
 427         /*
 428          * Lop off SYN bit if it has already been sent.  However, if this
 429          * is SYN-SENT state and if segment contains data and if we don't
 430          * know that foreign host supports TAO, suppress sending segment.
 431          */
 432         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
 433                 if (tp->t_state != TCPS_SYN_RECEIVED)
 434                         flags &= ~TH_SYN;
 435                 /*
 436                  * When sending additional segments following a TFO SYN|ACK,
 437                  * do not include the SYN bit.
 438                  */
 439                 if (IS_FASTOPEN(tp->t_flags) &&
 440                     (tp->t_state == TCPS_SYN_RECEIVED))
 441                         flags &= ~TH_SYN;
 442                 off--, len++;
 443         }
 444
 445         /*
 446          * Be careful not to send data and/or FIN on SYN segments.
 447          * This measure is needed to prevent interoperability problems
 448          * with not fully conformant TCP implementations.
 449          */
 450         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
 451                 len = 0;
 452                 flags &= ~TH_FIN;
 453         }
 454
 455         /*
 456          * On TFO sockets, ensure no data is sent in the following cases:
 457          *
 458          *  - When retransmitting SYN|ACK on a passively-created socket
 459          *
 460          *  - When retransmitting SYN on an actively created socket
 461          *
 462          *  - When sending a zero-length cookie (cookie request) on an
 463          *    actively created socket
 464          *
 465          *  - When the socket is in the CLOSED state (RST is being sent)
 466          */
 467         if (IS_FASTOPEN(tp->t_flags) &&
 468             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
 469              ((tp->t_state == TCPS_SYN_SENT) &&
 470               (tp->t_tfo_client_cookie_len == 0)) ||
 471              (flags & TH_RST)))
 472                 len = 0;
 473         if (len <= 0) {
 474                 /*
 475                  * If FIN has been sent but not acked,
 476                  * but we haven't been called to retransmit,
 477                  * len will be < 0.  Otherwise, window shrank
 478                  * after we sent into it.  If window shrank to 0,
 479                  * cancel pending retransmit, pull snd_nxt back
 480                  * to (closed) window, and set the persist timer
 481                  * if it isn't already going.  If the window didn't
 482                  * close completely, just wait for an ACK.
 483                  *
 484                  * We also do a general check here to ensure that
 485                  * we will set the persist timer when we have data
 486                  * to send, but a 0-byte window. This makes sure
 487                  * the persist timer is set even if the packet
 488                  * hits one of the "goto send" lines below.
 489                  */
 490                 len = 0;
 491                 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
 492                     (off < (int) sbavail(&so->so_snd)) &&
 493                     !tcp_timer_active(tp, TT_PERSIST)) {
 494                         tcp_timer_activate(tp, TT_REXMT, 0);
 495                         tp->t_rxtshift = 0;
 496                         tp->snd_nxt = tp->snd_una;
 497                         if (!tcp_timer_active(tp, TT_PERSIST))
 498                                 tcp_setpersist(tp);
 499                 }
 500         }
 501
 502         /* len will be >= 0 after this point. */
 503         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 504
 505         tcp_sndbuf_autoscale(tp, so, sendwin);
 506
 507         /*
 508          * Decide if we can use TCP Segmentation Offloading (if supported by
 509          * hardware).
 510          *
 511          * TSO may only be used if we are in a pure bulk sending state.  The
 512          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
 513          * IP options prevent using TSO.  With TSO the TCP header is the same
 514          * (except for the sequence number) for all generated packets.  This
 515          * makes it impossible to transmit any options which vary per generated
 516          * segment or packet.
 517          *
 518          * IPv4 handling has a clear separation of ip options and ip header
 519          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
 520          * the right thing below to provide length of just ip options and thus
 521          * checking for ipoptlen is enough to decide if ip options are present.
 522          */
 523 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 524         /*
 525          * Pre-calculate here as we save another lookup into the darknesses
 526          * of IPsec that way and can actually decide if TSO is ok.
 527          */
 528 #ifdef INET6
 529         if (isipv6 && IPSEC_ENABLED(ipv6))
 530                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, inp);
 531 #ifdef INET
 532         else
 533 #endif
 534 #endif /* INET6 */
 535 #ifdef INET
 536         if (IPSEC_ENABLED(ipv4))
 537                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, inp);
 538 #endif /* INET */
 539 #endif /* IPSEC */
 540 #ifdef INET6
 541         if (isipv6)
 542                 ipoptlen = ip6_optlen(inp);
 543         else
 544 #endif
 545         if (inp->inp_options)
 546                 ipoptlen = inp->inp_options->m_len -
 547                                 offsetof(struct ipoption, ipopt_list);
 548         else
 549                 ipoptlen = 0;
 550 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 551         ipoptlen += ipsec_optlen;
 552 #endif
 553
 554         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
 555             (tp->t_port == 0) &&
 556             ((tp->t_flags & TF_SIGNATURE) == 0) &&
 557             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
 558             ipoptlen == 0 && !(flags & TH_SYN))
 559                 tso = 1;
 560
 561         if (SEQ_LT((sack_rxmit ? p->rxmit : tp->snd_nxt) + len,
 562                     tp->snd_una + sbused(&so->so_snd))) {
 563                         flags &= ~TH_FIN;
 564         }
 565
 566         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 567             (long)TCP_MAXWIN << tp->rcv_scale);
 568
 569         /*
 570          * Sender silly window avoidance.   We transmit under the following
 571          * conditions when len is non-zero:
 572          *
 573          *      - We have a full segment (or more with TSO)
 574          *      - This is the last buffer in a write()/send() and we are
 575          *        either idle or running NODELAY
 576          *      - we've timed out (e.g. persist timer)
 577          *      - we have more then 1/2 the maximum send window's worth of
 578          *        data (receiver may be limited the window size)
 579          *      - we need to retransmit
 580          */
 581         if (len) {
 582                 if (len >= tp->t_maxseg)
 583                         goto send;
 584                 /*
 585                  * As the TCP header options are now
 586                  * considered when setting up the initial
 587                  * window, we would not send the last segment
 588                  * if we skip considering the option length here.
 589                  * Note: this may not work when tcp headers change
 590                  * very dynamically in the future.
 591                  */
 592                 if ((((tp->t_flags & TF_SIGNATURE) ?
 593                         PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
 594                     ((tp->t_flags & TF_RCVD_TSTMP) ?
 595                         PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
 596                     len) >= tp->t_maxseg)
 597                         goto send;
 598                 /*
 599                  * NOTE! on localhost connections an 'ack' from the remote
 600                  * end may occur synchronously with the output and cause
 601                  * us to flush a buffer queued with moretocome.  XXX
 602                  *
 603                  * note: the len + off check is almost certainly unnecessary.
 604                  */
 605                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
 606                     (idle || (tp->t_flags & TF_NODELAY)) &&
 607                     (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
 608                     (tp->t_flags & TF_NOPUSH) == 0) {
 609                         goto send;
 610                 }
 611                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
 612                         goto send;
 613                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
 614                         goto send;
 615                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
 616                         goto send;
 617                 if (sack_rxmit)
 618                         goto send;
 619         }
 620
 621         /*
 622          * Sending of standalone window updates.
 623          *
 624          * Window updates are important when we close our window due to a
 625          * full socket buffer and are opening it again after the application
 626          * reads data from it.  Once the window has opened again and the
 627          * remote end starts to send again the ACK clock takes over and
 628          * provides the most current window information.
 629          *
 630          * We must avoid the silly window syndrome whereas every read
 631          * from the receive buffer, no matter how small, causes a window
 632          * update to be sent.  We also should avoid sending a flurry of
 633          * window updates when the socket buffer had queued a lot of data
 634          * and the application is doing small reads.
 635          *
 636          * Prevent a flurry of pointless window updates by only sending
 637          * an update when we can increase the advertized window by more
 638          * than 1/4th of the socket buffer capacity.  When the buffer is
 639          * getting full or is very small be more aggressive and send an
 640          * update whenever we can increase by two mss sized segments.
 641          * In all other situations the ACK's to new incoming data will
 642          * carry further window increases.
 643          *
 644          * Don't send an independent window update if a delayed
 645          * ACK is pending (it will get piggy-backed on it) or the
 646          * remote side already has done a half-close and won't send
 647          * more data.
 648          */
 649         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
 650             !(tp->t_flags & TF_DELACK) &&
 651             !TCPS_HAVERCVDFIN(tp->t_state)) {
 652                 /*
 653                  * "adv" is the amount we could increase the window,
 654                  * taking into account that we are limited by
 655                  * TCP_MAXWIN << tp->rcv_scale.
 656                  */
 657                 int32_t adv;
 658                 int oldwin;
 659
 660                 adv = recwin;
 661                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
 662                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
 663                         if (adv > oldwin)
 664                                 adv -= oldwin;
 665                         else
 666                                 adv = 0;
 667                 } else
 668                         oldwin = 0;
 669
 670                 /*
 671                  * If the new window size ends up being the same as or less
 672                  * than the old size when it is scaled, then don't force
 673                  * a window update.
 674                  */
 675                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
 676                         goto dontupdate;
 677
 678                 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
 679                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
 680                      recwin <= (so->so_rcv.sb_hiwat / 8) ||
 681                      so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
 682                      adv >= TCP_MAXWIN << tp->rcv_scale))
 683                         goto send;
 684                 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
 685                         goto send;
 686         }
 687 dontupdate:
 688
 689         /*
 690          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
 691          * is also a catch-all for the retransmit timer timeout case.
 692          */
 693         if (tp->t_flags & TF_ACKNOW)
 694                 goto send;
 695         if ((flags & TH_RST) ||
 696             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
 697                 goto send;
 698         if (SEQ_GT(tp->snd_up, tp->snd_una))
 699                 goto send;
 700         /*
 701          * If our state indicates that FIN should be sent
 702          * and we have not yet done so, then we need to send.
 703          */
 704         if (flags & TH_FIN &&
 705             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
 706                 goto send;
 707         /*
 708          * In SACK, it is possible for tcp_output to fail to send a segment
 709          * after the retransmission timer has been turned off.  Make sure
 710          * that the retransmission timer is set.
 711          */
 712         if ((tp->t_flags & TF_SACK_PERMIT) &&
 713             SEQ_GT(tp->snd_max, tp->snd_una) &&
 714             !tcp_timer_active(tp, TT_REXMT) &&
 715             !tcp_timer_active(tp, TT_PERSIST)) {
 716                 tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
 717                 goto just_return;
 718         }
 719         /*
 720          * TCP window updates are not reliable, rather a polling protocol
 721          * using ``persist'' packets is used to insure receipt of window
 722          * updates.  The three ``states'' for the output side are:
 723          *      idle                    not doing retransmits or persists
 724          *      persisting              to move a small or zero window
 725          *      (re)transmitting        and thereby not persisting
 726          *
 727          * tcp_timer_active(tp, TT_PERSIST)
 728          *      is true when we are in persist state.
 729          * (tp->t_flags & TF_FORCEDATA)
 730          *      is set when we are called to send a persist packet.
 731          * tcp_timer_active(tp, TT_REXMT)
 732          *      is set when we are retransmitting
 733          * The output side is idle when both timers are zero.
 734          *
 735          * If send window is too small, there is data to transmit, and no
 736          * retransmit or persist is pending, then go to persist state.
 737          * If nothing happens soon, send when timer expires:
 738          * if window is nonzero, transmit what we can,
 739          * otherwise force out a byte.
 740          */
 741         if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
 742             !tcp_timer_active(tp, TT_PERSIST)) {
 743                 tp->t_rxtshift = 0;
 744                 tcp_setpersist(tp);
 745         }
 746
 747         /*
 748          * No reason to send a segment, just return.
 749          */
 750 just_return:
 751         SOCKBUF_UNLOCK(&so->so_snd);
 752         return (0);
 753
 754 send:
 755         SOCKBUF_LOCK_ASSERT(&so->so_snd);
 756         if (len > 0) {
 757                 if (len >= tp->t_maxseg)
 758                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
 759                 else
 760                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
 761         }
 762         /*
 763          * Before ESTABLISHED, force sending of initial options
 764          * unless TCP set not to do any options.
 765          * NOTE: we assume that the IP/TCP header plus TCP options
 766          * always fit in a single mbuf, leaving room for a maximum
 767          * link header, i.e.
 768          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
 769          */
 770         optlen = 0;
 771 #ifdef INET6
 772         if (isipv6)
 773                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
 774         else
 775 #endif
 776                 hdrlen = sizeof (struct tcpiphdr);
 777
 778         if (flags & TH_SYN) {
 779                 tp->snd_nxt = tp->iss;
 780         }
 781
 782         /*
 783          * Compute options for segment.
 784          * We only have to care about SYN and established connection
 785          * segments.  Options for SYN-ACK segments are handled in TCP
 786          * syncache.
 787          */
 788         to.to_flags = 0;
 789         if ((tp->t_flags & TF_NOOPT) == 0) {
 790                 /* Maximum segment size. */
 791                 if (flags & TH_SYN) {
 792                         to.to_mss = tcp_mssopt(&inp->inp_inc);
 793                         if (tp->t_port)
 794                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
 795                         to.to_flags |= TOF_MSS;
 796
 797                         /*
 798                          * On SYN or SYN|ACK transmits on TFO connections,
 799                          * only include the TFO option if it is not a
 800                          * retransmit, as the presence of the TFO option may
 801                          * have caused the original SYN or SYN|ACK to have
 802                          * been dropped by a middlebox.
 803                          */
 804                         if (IS_FASTOPEN(tp->t_flags) &&
 805                             (tp->t_rxtshift == 0)) {
 806                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
 807                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 808                                         to.to_tfo_cookie =
 809                                             (u_int8_t *)&tp->t_tfo_cookie.server;
 810                                         to.to_flags |= TOF_FASTOPEN;
 811                                         wanted_cookie = 1;
 812                                 } else if (tp->t_state == TCPS_SYN_SENT) {
 813                                         to.to_tfo_len =
 814                                             tp->t_tfo_client_cookie_len;
 815                                         to.to_tfo_cookie =
 816                                             tp->t_tfo_cookie.client;
 817                                         to.to_flags |= TOF_FASTOPEN;
 818                                         wanted_cookie = 1;
 819                                         /*
 820                                          * If we wind up having more data to
 821                                          * send with the SYN than can fit in
 822                                          * one segment, don't send any more
 823                                          * until the SYN|ACK comes back from
 824                                          * the other end.
 825                                          */
 826                                         dont_sendalot = 1;
 827                                 }
 828                         }
 829                 }
 830                 /* Window scaling. */
 831                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
 832                         to.to_wscale = tp->request_r_scale;
 833                         to.to_flags |= TOF_SCALE;
 834                 }
 835                 /* Timestamps. */
 836                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
 837                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
 838                         curticks = tcp_ts_getticks();
 839                         to.to_tsval = curticks + tp->ts_offset;
 840                         to.to_tsecr = tp->ts_recent;
 841                         to.to_flags |= TOF_TS;
 842                         if (tp->t_rxtshift == 1)
 843                                 tp->t_badrxtwin = curticks;
 844                 }
 845
 846                 /* Set receive buffer autosizing timestamp. */
 847                 if (tp->rfbuf_ts == 0 &&
 848                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
 849                         tp->rfbuf_ts = tcp_ts_getticks();
 850
 851                 /* Selective ACK's. */
 852                 if (tp->t_flags & TF_SACK_PERMIT) {
 853                         if (flags & TH_SYN)
 854                                 to.to_flags |= TOF_SACKPERM;
 855                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
 856                             tp->rcv_numsacks > 0) {
 857                                 to.to_flags |= TOF_SACK;
 858                                 to.to_nsacks = tp->rcv_numsacks;
 859                                 to.to_sacks = (u_char *)tp->sackblks;
 860                         }
 861                 }
 862 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 863                 /* TCP-MD5 (RFC2385). */
 864                 /*
 865                  * Check that TCP_MD5SIG is enabled in tcpcb to
 866                  * account the size needed to set this TCP option.
 867                  */
 868                 if (tp->t_flags & TF_SIGNATURE)
 869                         to.to_flags |= TOF_SIGNATURE;
 870 #endif /* TCP_SIGNATURE */
 871
 872                 /* Processing the options. */
 873                 hdrlen += optlen = tcp_addoptions(&to, opt);
 874                 /*
 875                  * If we wanted a TFO option to be added, but it was unable
 876                  * to fit, ensure no data is sent.
 877                  */
 878                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
 879                     !(to.to_flags & TOF_FASTOPEN))
 880                         len = 0;
 881         }
 882         if (tp->t_port) {
 883                 if (V_tcp_udp_tunneling_port == 0) {
 884                         /* The port was removed?? */
 885                         SOCKBUF_UNLOCK(&so->so_snd);
 886                         return (EHOSTUNREACH);
 887                 }
 888                 hdrlen += sizeof(struct udphdr);
 889         }
 890         /*
 891          * Adjust data length if insertion of options will
 892          * bump the packet length beyond the t_maxseg length.
 893          * Clear the FIN bit because we cut off the tail of
 894          * the segment.
 895          */
 896         if (len + optlen + ipoptlen > tp->t_maxseg) {
 897                 flags &= ~TH_FIN;
 898
 899                 if (tso) {
 900                         u_int if_hw_tsomax;
 901                         u_int moff;
 902                         int max_len;
 903
 904                         /* extract TSO information */
 905                         if_hw_tsomax = tp->t_tsomax;
 906                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
 907                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
 908
 909                         /*
 910                          * Limit a TSO burst to prevent it from
 911                          * overflowing or exceeding the maximum length
 912                          * allowed by the network interface:
 913                          */
 914                         KASSERT(ipoptlen == 0,
 915                             ("%s: TSO can't do IP options", __func__));
 916
 917                         /*
 918                          * Check if we should limit by maximum payload
 919                          * length:
 920                          */
 921                         if (if_hw_tsomax != 0) {
 922                                 /* compute maximum TSO length */
 923                                 max_len = (if_hw_tsomax - hdrlen -
 924                                     max_linkhdr);
 925                                 if (max_len <= 0) {
 926                                         len = 0;
 927                                 } else if (len > max_len) {
 928                                         sendalot = 1;
 929                                         len = max_len;
 930                                 }
 931                         }
 932
 933                         /*
 934                          * Prevent the last segment from being
 935                          * fractional unless the send sockbuf can be
 936                          * emptied:
 937                          */
 938                         max_len = (tp->t_maxseg - optlen);
 939                         if (((uint32_t)off + (uint32_t)len) <
 940                             sbavail(&so->so_snd)) {
 941                                 moff = len % max_len;
 942                                 if (moff != 0) {
 943                                         len -= moff;
 944                                         sendalot = 1;
 945                                 }
 946                         }
 947
 948                         /*
 949                          * In case there are too many small fragments
 950                          * don't use TSO:
 951                          */
 952                         if (len <= max_len) {
 953                                 len = max_len;
 954                                 sendalot = 1;
 955                                 tso = 0;
 956                         }
 957
 958                         /*
 959                          * Send the FIN in a separate segment
 960                          * after the bulk sending is done.
 961                          * We don't trust the TSO implementations
 962                          * to clear the FIN flag on all but the
 963                          * last segment.
 964                          */
 965                         if (tp->t_flags & TF_NEEDFIN)
 966                                 sendalot = 1;
 967                 } else {
 968                         if (optlen + ipoptlen >= tp->t_maxseg) {
 969                                 /*
 970                                  * Since we don't have enough space to put
 971                                  * the IP header chain and the TCP header in
 972                                  * one packet as required by RFC 7112, don't
 973                                  * send it. Also ensure that at least one
 974                                  * byte of the payload can be put into the
 975                                  * TCP segment.
 976                                  */
 977                                 SOCKBUF_UNLOCK(&so->so_snd);
 978                                 error = EMSGSIZE;
 979                                 sack_rxmit = 0;
 980                                 goto out;
 981                         }
 982                         len = tp->t_maxseg - optlen - ipoptlen;
 983                         sendalot = 1;
 984                         if (dont_sendalot)
 985                                 sendalot = 0;
 986                 }
 987         } else
 988                 tso = 0;
 989
 990         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
 991             ("%s: len > IP_MAXPACKET", __func__));
 992
 993 /*#ifdef DIAGNOSTIC*/
 994 #ifdef INET6
 995         if (max_linkhdr + hdrlen > MCLBYTES)
 996 #else
 997         if (max_linkhdr + hdrlen > MHLEN)
 998 #endif
 999                 panic("tcphdr too big");
1000 /*#endif*/
1001
1002         /*
1003          * This KASSERT is here to catch edge cases at a well defined place.
1004          * Before, those had triggered (random) panic conditions further down.
1005          */
1006         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1007
1008         /*
1009          * Grab a header mbuf, attaching a copy of data to
1010          * be transmitted, and initialize the header from
1011          * the template for sends on this connection.
1012          */
1013         if (len) {
1014                 struct mbuf *mb;
1015                 struct sockbuf *msb;
1016                 u_int moff;
1017
1018                 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1019                         TCPSTAT_INC(tcps_sndprobe);
1020 #ifdef STATS
1021                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1022                                 stats_voi_update_abs_u32(tp->t_stats,
1023                                 VOI_TCP_RETXPB, len);
1024                         else
1025                                 stats_voi_update_abs_u64(tp->t_stats,
1026                                     VOI_TCP_TXPB, len);
1027 #endif /* STATS */
1028                 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1029                         tp->t_sndrexmitpack++;
1030                         TCPSTAT_INC(tcps_sndrexmitpack);
1031                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1032                         if (sack_rxmit) {
1033                                 TCPSTAT_INC(tcps_sack_rexmits);
1034                                 TCPSTAT_ADD(tcps_sack_rexmit_bytes, len);
1035                         }
1036 #ifdef STATS
1037                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1038                             len);
1039 #endif /* STATS */
1040                 } else {
1041                         TCPSTAT_INC(tcps_sndpack);
1042                         TCPSTAT_ADD(tcps_sndbyte, len);
1043 #ifdef STATS
1044                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1045                             len);
1046 #endif /* STATS */
1047                 }
1048 #ifdef INET6
1049                 if (MHLEN < hdrlen + max_linkhdr)
1050                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1051                 else
1052 #endif
1053                         m = m_gethdr(M_NOWAIT, MT_DATA);
1054
1055                 if (m == NULL) {
1056                         SOCKBUF_UNLOCK(&so->so_snd);
1057                         error = ENOBUFS;
1058                         sack_rxmit = 0;
1059                         goto out;
1060                 }
1061
1062                 m->m_data += max_linkhdr;
1063                 m->m_len = hdrlen;
1064
1065                 /*
1066                  * Start the m_copy functions from the closest mbuf
1067                  * to the offset in the socket buffer chain.
1068                  */
1069                 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1070                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1071                         m_copydata(mb, moff, len,
1072                             mtod(m, caddr_t) + hdrlen);
1073                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1074                                 sbsndptr_adv(&so->so_snd, mb, len);
1075                         m->m_len += len;
1076                 } else {
1077                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1078                                 msb = NULL;
1079                         else
1080                                 msb = &so->so_snd;
1081                         m->m_next = tcp_m_copym(mb, moff,
1082                             &len, if_hw_tsomaxsegcount,
1083                             if_hw_tsomaxsegsize, msb, hw_tls);
1084                         if (len <= (tp->t_maxseg - optlen)) {
1085                                 /*
1086                                  * Must have ran out of mbufs for the copy
1087                                  * shorten it to no longer need tso. Lets
1088                                  * not put on sendalot since we are low on
1089                                  * mbufs.
1090                                  */
1091                                 tso = 0;
1092                         }
1093                         if (m->m_next == NULL) {
1094                                 SOCKBUF_UNLOCK(&so->so_snd);
1095                                 (void) m_free(m);
1096                                 error = ENOBUFS;
1097                                 sack_rxmit = 0;
1098                                 goto out;
1099                         }
1100                 }
1101
1102                 /*
1103                  * If we're sending everything we've got, set PUSH.
1104                  * (This will keep happy those implementations which only
1105                  * give data to the user when a buffer fills or
1106                  * a PUSH comes in.)
1107                  */
1108                 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1109                     !(flags & TH_SYN))
1110                         flags |= TH_PUSH;
1111                 SOCKBUF_UNLOCK(&so->so_snd);
1112         } else {
1113                 SOCKBUF_UNLOCK(&so->so_snd);
1114                 if (tp->t_flags & TF_ACKNOW)
1115                         TCPSTAT_INC(tcps_sndacks);
1116                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1117                         TCPSTAT_INC(tcps_sndctrl);
1118                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1119                         TCPSTAT_INC(tcps_sndurg);
1120                 else
1121                         TCPSTAT_INC(tcps_sndwinup);
1122
1123                 m = m_gethdr(M_NOWAIT, MT_DATA);
1124                 if (m == NULL) {
1125                         error = ENOBUFS;
1126                         sack_rxmit = 0;
1127                         goto out;
1128                 }
1129 #ifdef INET6
1130                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1131                     MHLEN >= hdrlen) {
1132                         M_ALIGN(m, hdrlen);
1133                 } else
1134 #endif
1135                 m->m_data += max_linkhdr;
1136                 m->m_len = hdrlen;
1137         }
1138         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1139         m->m_pkthdr.rcvif = (struct ifnet *)0;
1140 #ifdef MAC
1141         mac_inpcb_create_mbuf(inp, m);
1142 #endif
1143 #ifdef INET6
1144         if (isipv6) {
1145                 ip6 = mtod(m, struct ip6_hdr *);
1146                 if (tp->t_port) {
1147                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1148                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1149                         udp->uh_dport = tp->t_port;
1150                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
1151                         udp->uh_ulen = htons(ulen);
1152                         th = (struct tcphdr *)(udp + 1);
1153                 } else {
1154                         th = (struct tcphdr *)(ip6 + 1);
1155                 }
1156                 tcpip_fillheaders(inp, tp->t_port, ip6, th);
1157         } else
1158 #endif /* INET6 */
1159         {
1160                 ip = mtod(m, struct ip *);
1161                 if (tp->t_port) {
1162                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1163                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1164                         udp->uh_dport = tp->t_port;
1165                         ulen = hdrlen + len - sizeof(struct ip);
1166                         udp->uh_ulen = htons(ulen);
1167                         th = (struct tcphdr *)(udp + 1);
1168                 } else
1169                         th = (struct tcphdr *)(ip + 1);
1170                 tcpip_fillheaders(inp, tp->t_port, ip, th);
1171         }
1172
1173         /*
1174          * Fill in fields, remembering maximum advertised
1175          * window for use in delaying messages about window sizes.
1176          * If resending a FIN, be sure not to use a new sequence number.
1177          */
1178         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1179             tp->snd_nxt == tp->snd_max)
1180                 tp->snd_nxt--;
1181         /*
1182          * If we are starting a connection, send ECN setup
1183          * SYN packet. If we are on a retransmit, we may
1184          * resend those bits a number of times as per
1185          * RFC 3168.
1186          */
1187         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn) {
1188                 flags |= tcp_ecn_output_syn_sent(tp);
1189         }
1190         /* Also handle parallel SYN for ECN */
1191         if ((TCPS_HAVERCVDSYN(tp->t_state)) &&
1192             (tp->t_flags2 & (TF2_ECN_PERMIT | TF2_ACE_PERMIT))) {
1193                 int ect = tcp_ecn_output_established(tp, &flags, len, sack_rxmit);
1194                 if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1195                     (tp->t_flags2 & TF2_ECN_SND_ECE))
1196                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1197 #ifdef INET6
1198                 if (isipv6) {
1199                         ip6->ip6_flow &= ~htonl(IPTOS_ECN_MASK << IPV6_FLOWLABEL_LEN);
1200                         ip6->ip6_flow |= htonl(ect << IPV6_FLOWLABEL_LEN);
1201                 }
1202                 else
1203 #endif
1204                 {
1205                         ip->ip_tos &= ~IPTOS_ECN_MASK;
1206                         ip->ip_tos |= ect;
1207                 }
1208         }
1209
1210         /*
1211          * If we are doing retransmissions, then snd_nxt will
1212          * not reflect the first unsent octet.  For ACK only
1213          * packets, we do not want the sequence number of the
1214          * retransmitted packet, we want the sequence number
1215          * of the next unsent octet.  So, if there is no data
1216          * (and no SYN or FIN), use snd_max instead of snd_nxt
1217          * when filling in ti_seq.  But if we are in persist
1218          * state, snd_max might reflect one byte beyond the
1219          * right edge of the window, so use snd_nxt in that
1220          * case, since we know we aren't doing a retransmission.
1221          * (retransmit and persist are mutually exclusive...)
1222          */
1223         if (sack_rxmit == 0) {
1224                 if (len || (flags & (TH_SYN|TH_FIN)) ||
1225                     tcp_timer_active(tp, TT_PERSIST))
1226                         th->th_seq = htonl(tp->snd_nxt);
1227                 else
1228                         th->th_seq = htonl(tp->snd_max);
1229         } else {
1230                 th->th_seq = htonl(p->rxmit);
1231                 p->rxmit += len;
1232                 /*
1233                  * Lost Retransmission Detection
1234                  * trigger resending of a (then
1235                  * still existing) hole, when
1236                  * fack acks recoverypoint.
1237                  */
1238                 if ((tp->t_flags & TF_LRD) && SEQ_GEQ(p->rxmit, p->end))
1239                         p->rxmit = tp->snd_recover;
1240                 tp->sackhint.sack_bytes_rexmit += len;
1241         }
1242         if (IN_RECOVERY(tp->t_flags)) {
1243                 /*
1244                  * Account all bytes transmitted while
1245                  * IN_RECOVERY, simplifying PRR and
1246                  * Lost Retransmit Detection
1247                  */
1248                 tp->sackhint.prr_out += len;
1249         }
1250         th->th_ack = htonl(tp->rcv_nxt);
1251         if (optlen) {
1252                 bcopy(opt, th + 1, optlen);
1253                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1254         }
1255         tcp_set_flags(th, flags);
1256         /*
1257          * Calculate receive window.  Don't shrink window,
1258          * but avoid silly window syndrome.
1259          * If a RST segment is sent, advertise a window of zero.
1260          */
1261         if (flags & TH_RST) {
1262                 recwin = 0;
1263         } else {
1264                 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1265                     recwin < tp->t_maxseg)
1266                         recwin = 0;
1267                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1268                     recwin < (tp->rcv_adv - tp->rcv_nxt))
1269                         recwin = (tp->rcv_adv - tp->rcv_nxt);
1270         }
1271         /*
1272          * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1273          * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1274          * case is handled in syncache.
1275          */
1276         if (flags & TH_SYN)
1277                 th->th_win = htons((u_short)
1278                                 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1279         else {
1280                 /* Avoid shrinking window with window scaling. */
1281                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1282                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1283         }
1284
1285         /*
1286          * Adjust the RXWIN0SENT flag - indicate that we have advertised
1287          * a 0 window.  This may cause the remote transmitter to stall.  This
1288          * flag tells soreceive() to disable delayed acknowledgements when
1289          * draining the buffer.  This can occur if the receiver is attempting
1290          * to read more data than can be buffered prior to transmitting on
1291          * the connection.
1292          */
1293         if (th->th_win == 0) {
1294                 tp->t_sndzerowin++;
1295                 tp->t_flags |= TF_RXWIN0SENT;
1296         } else
1297                 tp->t_flags &= ~TF_RXWIN0SENT;
1298         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1299                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1300                 th->th_flags |= TH_URG;
1301         } else
1302                 /*
1303                  * If no urgent pointer to send, then we pull
1304                  * the urgent pointer to the left edge of the send window
1305                  * so that it doesn't drift into the send window on sequence
1306                  * number wraparound.
1307                  */
1308                 tp->snd_up = tp->snd_una;               /* drag it along */
1309
1310         /*
1311          * Put TCP length in extended header, and then
1312          * checksum extended header and data.
1313          */
1314         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1315
1316 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1317         if (to.to_flags & TOF_SIGNATURE) {
1318                 /*
1319                  * Calculate MD5 signature and put it into the place
1320                  * determined before.
1321                  * NOTE: since TCP options buffer doesn't point into
1322                  * mbuf's data, calculate offset and use it.
1323                  */
1324                 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1325                     (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1326                         /*
1327                          * Do not send segment if the calculation of MD5
1328                          * digest has failed.
1329                          */
1330                         m_freem(m);
1331                         goto out;
1332                 }
1333         }
1334 #endif
1335 #ifdef INET6
1336         if (isipv6) {
1337                 /*
1338                  * There is no need to fill in ip6_plen right now.
1339                  * It will be filled later by ip6_output.
1340                  */
1341                 if (tp->t_port) {
1342                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1343                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1344                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1345                         th->th_sum = htons(0);
1346                         UDPSTAT_INC(udps_opackets);
1347                 } else {
1348                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1349                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1350                         th->th_sum = in6_cksum_pseudo(ip6,
1351                             sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1352                             0);
1353                 }
1354         }
1355 #endif
1356 #if defined(INET6) && defined(INET)
1357         else
1358 #endif
1359 #ifdef INET
1360         {
1361                 if (tp->t_port) {
1362                         m->m_pkthdr.csum_flags = CSUM_UDP;
1363                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1364                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1365                            ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1366                         th->th_sum = htons(0);
1367                         UDPSTAT_INC(udps_opackets);
1368                 } else {
1369                         m->m_pkthdr.csum_flags = CSUM_TCP;
1370                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1371                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
1372                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1373                             IPPROTO_TCP + len + optlen));
1374                 }
1375
1376                 /* IP version must be set here for ipv4/ipv6 checking later */
1377                 KASSERT(ip->ip_v == IPVERSION,
1378                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1379         }
1380 #endif
1381
1382         /*
1383          * Enable TSO and specify the size of the segments.
1384          * The TCP pseudo header checksum is always provided.
1385          */
1386         if (tso) {
1387                 KASSERT(len > tp->t_maxseg - optlen,
1388                     ("%s: len <= tso_segsz", __func__));
1389                 m->m_pkthdr.csum_flags |= CSUM_TSO;
1390                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1391         }
1392
1393         KASSERT(len + hdrlen == m_length(m, NULL),
1394             ("%s: mbuf chain shorter than expected: %d + %u != %u",
1395             __func__, len, hdrlen, m_length(m, NULL)));
1396
1397 #ifdef TCP_HHOOK
1398         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1399         hhook_run_tcp_est_out(tp, th, &to, len, tso);
1400 #endif
1401
1402         TCP_PROBE3(debug__output, tp, th, m);
1403
1404         /* We're getting ready to send; log now. */
1405         /* XXXMT: We are not honoring verbose logging. */
1406
1407         if (tcp_bblogging_on(tp))
1408                 lgb = tcp_log_event(tp, th, &so->so_rcv, &so->so_snd,
1409                     TCP_LOG_OUT, ERRNO_UNK, len, NULL, false, NULL, NULL, 0,
1410                     NULL);
1411         else
1412                 lgb = NULL;
1413
1414         /*
1415          * Fill in IP length and desired time to live and
1416          * send to IP level.  There should be a better way
1417          * to handle ttl and tos; we could keep them in
1418          * the template, but need a way to checksum without them.
1419          */
1420         /*
1421          * m->m_pkthdr.len should have been set before checksum calculation,
1422          * because in6_cksum() need it.
1423          */
1424 #ifdef INET6
1425         if (isipv6) {
1426                 /*
1427                  * we separately set hoplimit for every segment, since the
1428                  * user might want to change the value via setsockopt.
1429                  * Also, desired default hop limit might be changed via
1430                  * Neighbor Discovery.
1431                  */
1432                 ip6->ip6_hlim = in6_selecthlim(inp, NULL);
1433
1434                 /*
1435                  * Set the packet size here for the benefit of DTrace probes.
1436                  * ip6_output() will set it properly; it's supposed to include
1437                  * the option header lengths as well.
1438                  */
1439                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1440
1441                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1442                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1443                 else
1444                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1445
1446                 if (tp->t_state == TCPS_SYN_SENT)
1447                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1448
1449                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1450
1451 #ifdef TCPPCAP
1452                 /* Save packet, if requested. */
1453                 tcp_pcap_add(th, m, &(tp->t_outpkts));
1454 #endif
1455
1456                 /* TODO: IPv6 IP6TOS_ECT bit on */
1457                 error = ip6_output(m, inp->in6p_outputopts, &inp->inp_route6,
1458                     ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1459                     NULL, NULL, inp);
1460
1461                 if (error == EMSGSIZE && inp->inp_route6.ro_nh != NULL)
1462                         mtu = inp->inp_route6.ro_nh->nh_mtu;
1463         }
1464 #endif /* INET6 */
1465 #if defined(INET) && defined(INET6)
1466         else
1467 #endif
1468 #ifdef INET
1469     {
1470         ip->ip_len = htons(m->m_pkthdr.len);
1471 #ifdef INET6
1472         if (inp->inp_vflag & INP_IPV6PROTO)
1473                 ip->ip_ttl = in6_selecthlim(inp, NULL);
1474 #endif /* INET6 */
1475         /*
1476          * If we do path MTU discovery, then we set DF on every packet.
1477          * This might not be the best thing to do according to RFC3390
1478          * Section 2. However the tcp hostcache migitates the problem
1479          * so it affects only the first tcp connection with a host.
1480          *
1481          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1482          */
1483         if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1484                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1485                 if (tp->t_port == 0 || len < V_tcp_minmss) {
1486                         ip->ip_off |= htons(IP_DF);
1487                 }
1488         } else {
1489                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1490         }
1491
1492         if (tp->t_state == TCPS_SYN_SENT)
1493                 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1494
1495         TCP_PROBE5(send, NULL, tp, ip, tp, th);
1496
1497 #ifdef TCPPCAP
1498         /* Save packet, if requested. */
1499         tcp_pcap_add(th, m, &(tp->t_outpkts));
1500 #endif
1501
1502         error = ip_output(m, inp->inp_options, &inp->inp_route,
1503             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0, inp);
1504
1505         if (error == EMSGSIZE && inp->inp_route.ro_nh != NULL)
1506                 mtu = inp->inp_route.ro_nh->nh_mtu;
1507     }
1508 #endif /* INET */
1509
1510         if (lgb != NULL) {
1511                 lgb->tlb_errno = error;
1512                 lgb = NULL;
1513         }
1514 out:
1515         if (error == 0)
1516                 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1517         /*
1518          * In transmit state, time the transmission and arrange for
1519          * the retransmit.  In persist state, just set snd_max.  In a closed
1520          * state just return.
1521          */
1522         if (flags & TH_RST) {
1523                 TCPSTAT_INC(tcps_sndtotal);
1524                 return (0);
1525         } else if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1526             !tcp_timer_active(tp, TT_PERSIST)) {
1527                 tcp_seq startseq = tp->snd_nxt;
1528
1529                 /*
1530                  * Advance snd_nxt over sequence space of this segment.
1531                  */
1532                 if (flags & (TH_SYN|TH_FIN)) {
1533                         if (flags & TH_SYN)
1534                                 tp->snd_nxt++;
1535                         if (flags & TH_FIN) {
1536                                 tp->snd_nxt++;
1537                                 tp->t_flags |= TF_SENTFIN;
1538                         }
1539                 }
1540                 if (sack_rxmit)
1541                         goto timer;
1542                 tp->snd_nxt += len;
1543                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1544                         /*
1545                          * Update "made progress" indication if we just
1546                          * added new data to an empty socket buffer.
1547                          */
1548                         if (tp->snd_una == tp->snd_max)
1549                                 tp->t_acktime = ticks;
1550                         tp->snd_max = tp->snd_nxt;
1551                         /*
1552                          * Time this transmission if not a retransmission and
1553                          * not currently timing anything.
1554                          */
1555                         tp->t_sndtime = ticks;
1556                         if (tp->t_rtttime == 0) {
1557                                 tp->t_rtttime = ticks;
1558                                 tp->t_rtseq = startseq;
1559                                 TCPSTAT_INC(tcps_segstimed);
1560                         }
1561 #ifdef STATS
1562                         if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1563                                 tp->t_flags |= TF_GPUTINPROG;
1564                                 tp->gput_seq = startseq;
1565                                 tp->gput_ack = startseq +
1566                                     ulmin(sbavail(&so->so_snd) - off, sendwin);
1567                                 tp->gput_ts = tcp_ts_getticks();
1568                         }
1569 #endif /* STATS */
1570                 }
1571
1572                 /*
1573                  * Set retransmit timer if not currently set,
1574                  * and not doing a pure ack or a keep-alive probe.
1575                  * Initial value for retransmit timer is smoothed
1576                  * round-trip time + 2 * round-trip time variance.
1577                  * Initialize shift counter which is used for backoff
1578                  * of retransmit time.
1579                  */
1580 timer:
1581                 if (!tcp_timer_active(tp, TT_REXMT) &&
1582                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1583                      (tp->snd_nxt != tp->snd_una))) {
1584                         if (tcp_timer_active(tp, TT_PERSIST)) {
1585                                 tcp_timer_activate(tp, TT_PERSIST, 0);
1586                                 tp->t_rxtshift = 0;
1587                         }
1588                         tcp_timer_activate(tp, TT_REXMT, TP_RXTCUR(tp));
1589                 } else if (len == 0 && sbavail(&so->so_snd) &&
1590                     !tcp_timer_active(tp, TT_REXMT) &&
1591                     !tcp_timer_active(tp, TT_PERSIST)) {
1592                         /*
1593                          * Avoid a situation where we do not set persist timer
1594                          * after a zero window condition. For example:
1595                          * 1) A -> B: packet with enough data to fill the window
1596                          * 2) B -> A: ACK for #1 + new data (0 window
1597                          *    advertisement)
1598                          * 3) A -> B: ACK for #2, 0 len packet
1599                          *
1600                          * In this case, A will not activate the persist timer,
1601                          * because it chose to send a packet. Unless tcp_output
1602                          * is called for some other reason (delayed ack timer,
1603                          * another input packet from B, socket syscall), A will
1604                          * not send zero window probes.
1605                          *
1606                          * So, if you send a 0-length packet, but there is data
1607                          * in the socket buffer, and neither the rexmt or
1608                          * persist timer is already set, then activate the
1609                          * persist timer.
1610                          */
1611                         tp->t_rxtshift = 0;
1612                         tcp_setpersist(tp);
1613                 }
1614         } else {
1615                 /*
1616                  * Persist case, update snd_max but since we are in
1617                  * persist mode (no window) we do not update snd_nxt.
1618                  */
1619                 int xlen = len;
1620                 if (flags & TH_SYN)
1621                         ++xlen;
1622                 if (flags & TH_FIN) {
1623                         ++xlen;
1624                         tp->t_flags |= TF_SENTFIN;
1625                 }
1626                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1627                         tp->snd_max = tp->snd_nxt + xlen;
1628         }
1629         if ((error == 0) &&
1630             (TCPS_HAVEESTABLISHED(tp->t_state) &&
1631              (tp->t_flags & TF_SACK_PERMIT) &&
1632              tp->rcv_numsacks > 0)) {
1633                     /* Clean up any DSACK's sent */
1634                     tcp_clean_dsack_blocks(tp);
1635         }
1636         if (error) {
1637                 /*
1638                  * We know that the packet was lost, so back out the
1639                  * sequence number advance, if any.
1640                  *
1641                  * If the error is EPERM the packet got blocked by the
1642                  * local firewall.  Normally we should terminate the
1643                  * connection but the blocking may have been spurious
1644                  * due to a firewall reconfiguration cycle.  So we treat
1645                  * it like a packet loss and let the retransmit timer and
1646                  * timeouts do their work over time.
1647                  * XXX: It is a POLA question whether calling tcp_drop right
1648                  * away would be the really correct behavior instead.
1649                  */
1650                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1651                     !tcp_timer_active(tp, TT_PERSIST)) &&
1652                     ((flags & TH_SYN) == 0) &&
1653                     (error != EPERM)) {
1654                         if (sack_rxmit) {
1655                                 p->rxmit = SEQ_MIN(p->end, p->rxmit) - len;
1656                                 tp->sackhint.sack_bytes_rexmit -= len;
1657                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1658                                     ("sackhint bytes rtx >= 0"));
1659                                 KASSERT((flags & TH_FIN) == 0,
1660                                     ("error while FIN with SACK rxmit"));
1661                         } else {
1662                                 tp->snd_nxt -= len;
1663                                 if (flags & TH_FIN)
1664                                         tp->snd_nxt--;
1665                         }
1666                         if (IN_RECOVERY(tp->t_flags))
1667                                 tp->sackhint.prr_out -= len;
1668                 }
1669                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
1670                 switch (error) {
1671                 case EACCES:
1672                 case EPERM:
1673                         tp->t_softerror = error;
1674                         return (error);
1675                 case ENOBUFS:
1676                         TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1677                         tp->snd_cwnd = tp->t_maxseg;
1678                         return (0);
1679                 case EMSGSIZE:
1680                         /*
1681                          * For some reason the interface we used initially
1682                          * to send segments changed to another or lowered
1683                          * its MTU.
1684                          * If TSO was active we either got an interface
1685                          * without TSO capabilits or TSO was turned off.
1686                          * If we obtained mtu from ip_output() then update
1687                          * it and try again.
1688                          */
1689                         if (tso)
1690                                 tp->t_flags &= ~TF_TSO;
1691                         if (mtu != 0) {
1692                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1693                                 goto again;
1694                         }
1695                         return (error);
1696                 case EHOSTDOWN:
1697                 case EHOSTUNREACH:
1698                 case ENETDOWN:
1699                 case ENETUNREACH:
1700                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
1701                                 tp->t_softerror = error;
1702                                 return (0);
1703                         }
1704                         /* FALLTHROUGH */
1705                 default:
1706                         return (error);
1707                 }
1708         }
1709         TCPSTAT_INC(tcps_sndtotal);
1710
1711         /*
1712          * Data sent (as far as we can tell).
1713          * If this advertises a larger window than any other segment,
1714          * then remember the size of the advertised window.
1715          * Any pending ACK has now been sent.
1716          */
1717         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1718                 tp->rcv_adv = tp->rcv_nxt + recwin;
1719         tp->last_ack_sent = tp->rcv_nxt;
1720         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1721         if (tcp_timer_active(tp, TT_DELACK))
1722                 tcp_timer_activate(tp, TT_DELACK, 0);
1723         if (sendalot)
1724                 goto again;
1725         return (0);
1726 }
1727
1728 void
1729 tcp_setpersist(struct tcpcb *tp)
1730 {
1731         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1732         int tt;
1733         int maxunacktime;
1734
1735         tp->t_flags &= ~TF_PREVVALID;
1736         if (tcp_timer_active(tp, TT_REXMT))
1737                 panic("tcp_setpersist: retransmit pending");
1738         /*
1739          * If the state is already closed, don't bother.
1740          */
1741         if (tp->t_state == TCPS_CLOSED)
1742                 return;
1743
1744         /*
1745          * Start/restart persistence timer.
1746          */
1747         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1748                       tcp_persmin, tcp_persmax);
1749         if (TP_MAXUNACKTIME(tp) && tp->t_acktime) {
1750                 maxunacktime = tp->t_acktime + TP_MAXUNACKTIME(tp) - ticks;
1751                 if (maxunacktime < 1)
1752                         maxunacktime = 1;
1753                 if (maxunacktime < tt)
1754                         tt = maxunacktime;
1755         }
1756         tcp_timer_activate(tp, TT_PERSIST, tt);
1757         if (tp->t_rxtshift < V_tcp_retries)
1758                 tp->t_rxtshift++;
1759 }
1760
1761 /*
1762  * Insert TCP options according to the supplied parameters to the place
1763  * optp in a consistent way.  Can handle unaligned destinations.
1764  *
1765  * The order of the option processing is crucial for optimal packing and
1766  * alignment for the scarce option space.
1767  *
1768  * The optimal order for a SYN/SYN-ACK segment is:
1769  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1770  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1771  *
1772  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1773  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1774  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1775  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1776  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1777  */
1778 int
1779 tcp_addoptions(struct tcpopt *to, u_char *optp)
1780 {
1781         u_int32_t mask, optlen = 0;
1782
1783         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1784                 if ((to->to_flags & mask) != mask)
1785                         continue;
1786                 if (optlen == TCP_MAXOLEN)
1787                         break;
1788                 switch (to->to_flags & mask) {
1789                 case TOF_MSS:
1790                         while (optlen % 4) {
1791                                 optlen += TCPOLEN_NOP;
1792                                 *optp++ = TCPOPT_NOP;
1793                         }
1794                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1795                                 continue;
1796                         optlen += TCPOLEN_MAXSEG;
1797                         *optp++ = TCPOPT_MAXSEG;
1798                         *optp++ = TCPOLEN_MAXSEG;
1799                         to->to_mss = htons(to->to_mss);
1800                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1801                         optp += sizeof(to->to_mss);
1802                         break;
1803                 case TOF_SCALE:
1804                         while (!optlen || optlen % 2 != 1) {
1805                                 optlen += TCPOLEN_NOP;
1806                                 *optp++ = TCPOPT_NOP;
1807                         }
1808                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1809                                 continue;
1810                         optlen += TCPOLEN_WINDOW;
1811                         *optp++ = TCPOPT_WINDOW;
1812                         *optp++ = TCPOLEN_WINDOW;
1813                         *optp++ = to->to_wscale;
1814                         break;
1815                 case TOF_SACKPERM:
1816                         while (optlen % 2) {
1817                                 optlen += TCPOLEN_NOP;
1818                                 *optp++ = TCPOPT_NOP;
1819                         }
1820                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1821                                 continue;
1822                         optlen += TCPOLEN_SACK_PERMITTED;
1823                         *optp++ = TCPOPT_SACK_PERMITTED;
1824                         *optp++ = TCPOLEN_SACK_PERMITTED;
1825                         break;
1826                 case TOF_TS:
1827                         while (!optlen || optlen % 4 != 2) {
1828                                 optlen += TCPOLEN_NOP;
1829                                 *optp++ = TCPOPT_NOP;
1830                         }
1831                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1832                                 continue;
1833                         optlen += TCPOLEN_TIMESTAMP;
1834                         *optp++ = TCPOPT_TIMESTAMP;
1835                         *optp++ = TCPOLEN_TIMESTAMP;
1836                         to->to_tsval = htonl(to->to_tsval);
1837                         to->to_tsecr = htonl(to->to_tsecr);
1838                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1839                         optp += sizeof(to->to_tsval);
1840                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1841                         optp += sizeof(to->to_tsecr);
1842                         break;
1843                 case TOF_SIGNATURE:
1844                         {
1845                         int siglen = TCPOLEN_SIGNATURE - 2;
1846
1847                         while (!optlen || optlen % 4 != 2) {
1848                                 optlen += TCPOLEN_NOP;
1849                                 *optp++ = TCPOPT_NOP;
1850                         }
1851                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1852                                 to->to_flags &= ~TOF_SIGNATURE;
1853                                 continue;
1854                         }
1855                         optlen += TCPOLEN_SIGNATURE;
1856                         *optp++ = TCPOPT_SIGNATURE;
1857                         *optp++ = TCPOLEN_SIGNATURE;
1858                         to->to_signature = optp;
1859                         while (siglen--)
1860                                  *optp++ = 0;
1861                         break;
1862                         }
1863                 case TOF_SACK:
1864                         {
1865                         int sackblks = 0;
1866                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
1867                         tcp_seq sack_seq;
1868
1869                         while (!optlen || optlen % 4 != 2) {
1870                                 optlen += TCPOLEN_NOP;
1871                                 *optp++ = TCPOPT_NOP;
1872                         }
1873                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1874                                 continue;
1875                         optlen += TCPOLEN_SACKHDR;
1876                         *optp++ = TCPOPT_SACK;
1877                         sackblks = min(to->to_nsacks,
1878                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1879                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1880                         while (sackblks--) {
1881                                 sack_seq = htonl(sack->start);
1882                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1883                                 optp += sizeof(sack_seq);
1884                                 sack_seq = htonl(sack->end);
1885                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1886                                 optp += sizeof(sack_seq);
1887                                 optlen += TCPOLEN_SACK;
1888                                 sack++;
1889                         }
1890                         TCPSTAT_INC(tcps_sack_send_blocks);
1891                         break;
1892                         }
1893                 case TOF_FASTOPEN:
1894                         {
1895                         int total_len;
1896
1897                         /* XXX is there any point to aligning this option? */
1898                         total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1899                         if (TCP_MAXOLEN - optlen < total_len) {
1900                                 to->to_flags &= ~TOF_FASTOPEN;
1901                                 continue;
1902                         }
1903                         *optp++ = TCPOPT_FAST_OPEN;
1904                         *optp++ = total_len;
1905                         if (to->to_tfo_len > 0) {
1906                                 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1907                                 optp += to->to_tfo_len;
1908                         }
1909                         optlen += total_len;
1910                         break;
1911                         }
1912                 default:
1913                         panic("%s: unknown TCP option type", __func__);
1914                         break;
1915                 }
1916         }
1917
1918         /* Terminate and pad TCP options to a 4 byte boundary. */
1919         if (optlen % 4) {
1920                 optlen += TCPOLEN_EOL;
1921                 *optp++ = TCPOPT_EOL;
1922         }
1923         /*
1924          * According to RFC 793 (STD0007):
1925          *   "The content of the header beyond the End-of-Option option
1926          *    must be header padding (i.e., zero)."
1927          *   and later: "The padding is composed of zeros."
1928          */
1929         while (optlen % 4) {
1930                 optlen += TCPOLEN_PAD;
1931                 *optp++ = TCPOPT_PAD;
1932         }
1933
1934         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1935         return (optlen);
1936 }
1937
1938 /*
1939  * This is a copy of m_copym(), taking the TSO segment size/limit
1940  * constraints into account, and advancing the sndptr as it goes.
1941  */
1942 struct mbuf *
1943 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1944     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1945 {
1946 #ifdef KERN_TLS
1947         struct ktls_session *tls, *ntls;
1948         struct mbuf *start __diagused;
1949 #endif
1950         struct mbuf *n, **np;
1951         struct mbuf *top;
1952         int32_t off = off0;
1953         int32_t len = *plen;
1954         int32_t fragsize;
1955         int32_t len_cp = 0;
1956         int32_t *pkthdrlen;
1957         uint32_t mlen, frags;
1958         bool copyhdr;
1959
1960         KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1961         KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1962         if (off == 0 && m->m_flags & M_PKTHDR)
1963                 copyhdr = true;
1964         else
1965                 copyhdr = false;
1966         while (off > 0) {
1967                 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1968                 if (off < m->m_len)
1969                         break;
1970                 off -= m->m_len;
1971                 if ((sb) && (m == sb->sb_sndptr)) {
1972                         sb->sb_sndptroff += m->m_len;
1973                         sb->sb_sndptr = m->m_next;
1974                 }
1975                 m = m->m_next;
1976         }
1977         np = &top;
1978         top = NULL;
1979         pkthdrlen = NULL;
1980 #ifdef KERN_TLS
1981         if (hw_tls && (m->m_flags & M_EXTPG))
1982                 tls = m->m_epg_tls;
1983         else
1984                 tls = NULL;
1985         start = m;
1986 #endif
1987         while (len > 0) {
1988                 if (m == NULL) {
1989                         KASSERT(len == M_COPYALL,
1990                             ("tcp_m_copym, length > size of mbuf chain"));
1991                         *plen = len_cp;
1992                         if (pkthdrlen != NULL)
1993                                 *pkthdrlen = len_cp;
1994                         break;
1995                 }
1996 #ifdef KERN_TLS
1997                 if (hw_tls) {
1998                         if (m->m_flags & M_EXTPG)
1999                                 ntls = m->m_epg_tls;
2000                         else
2001                                 ntls = NULL;
2002
2003                         /*
2004                          * Avoid mixing TLS records with handshake
2005                          * data or TLS records from different
2006                          * sessions.
2007                          */
2008                         if (tls != ntls) {
2009                                 MPASS(m != start);
2010                                 *plen = len_cp;
2011                                 if (pkthdrlen != NULL)
2012                                         *pkthdrlen = len_cp;
2013                                 break;
2014                         }
2015                 }
2016 #endif
2017                 mlen = min(len, m->m_len - off);
2018                 if (seglimit) {
2019                         /*
2020                          * For M_EXTPG mbufs, add 3 segments
2021                          * + 1 in case we are crossing page boundaries
2022                          * + 2 in case the TLS hdr/trailer are used
2023                          * It is cheaper to just add the segments
2024                          * than it is to take the cache miss to look
2025                          * at the mbuf ext_pgs state in detail.
2026                          */
2027                         if (m->m_flags & M_EXTPG) {
2028                                 fragsize = min(segsize, PAGE_SIZE);
2029                                 frags = 3;
2030                         } else {
2031                                 fragsize = segsize;
2032                                 frags = 0;
2033                         }
2034
2035                         /* Break if we really can't fit anymore. */
2036                         if ((frags + 1) >= seglimit) {
2037                                 *plen = len_cp;
2038                                 if (pkthdrlen != NULL)
2039                                         *pkthdrlen = len_cp;
2040                                 break;
2041                         }
2042
2043                         /*
2044                          * Reduce size if you can't copy the whole
2045                          * mbuf. If we can't copy the whole mbuf, also
2046                          * adjust len so the loop will end after this
2047                          * mbuf.
2048                          */
2049                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2050                                 mlen = (seglimit - frags - 1) * fragsize;
2051                                 len = mlen;
2052                                 *plen = len_cp + len;
2053                                 if (pkthdrlen != NULL)
2054                                         *pkthdrlen = *plen;
2055                         }
2056                         frags += howmany(mlen, fragsize);
2057                         if (frags == 0)
2058                                 frags++;
2059                         seglimit -= frags;
2060                         KASSERT(seglimit > 0,
2061                             ("%s: seglimit went too low", __func__));
2062                 }
2063                 if (copyhdr)
2064                         n = m_gethdr(M_NOWAIT, m->m_type);
2065                 else
2066                         n = m_get(M_NOWAIT, m->m_type);
2067                 *np = n;
2068                 if (n == NULL)
2069                         goto nospace;
2070                 if (copyhdr) {
2071                         if (!m_dup_pkthdr(n, m, M_NOWAIT))
2072                                 goto nospace;
2073                         if (len == M_COPYALL)
2074                                 n->m_pkthdr.len -= off0;
2075                         else
2076                                 n->m_pkthdr.len = len;
2077                         pkthdrlen = &n->m_pkthdr.len;
2078                         copyhdr = false;
2079                 }
2080                 n->m_len = mlen;
2081                 len_cp += n->m_len;
2082                 if (m->m_flags & (M_EXT | M_EXTPG)) {
2083                         n->m_data = m->m_data + off;
2084                         mb_dupcl(n, m);
2085                 } else
2086                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2087                             (u_int)n->m_len);
2088
2089                 if (sb && (sb->sb_sndptr == m) &&
2090                     ((n->m_len + off) >= m->m_len) && m->m_next) {
2091                         sb->sb_sndptroff += m->m_len;
2092                         sb->sb_sndptr = m->m_next;
2093                 }
2094                 off = 0;
2095                 if (len != M_COPYALL) {
2096                         len -= n->m_len;
2097                 }
2098                 m = m->m_next;
2099                 np = &n->m_next;
2100         }
2101         return (top);
2102 nospace:
2103         m_freem(top);
2104         return (NULL);
2105 }
2106
2107 void
2108 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2109 {
2110
2111         /*
2112          * Automatic sizing of send socket buffer.  Often the send buffer
2113          * size is not optimally adjusted to the actual network conditions
2114          * at hand (delay bandwidth product).  Setting the buffer size too
2115          * small limits throughput on links with high bandwidth and high
2116          * delay (eg. trans-continental/oceanic links).  Setting the
2117          * buffer size too big consumes too much real kernel memory,
2118          * especially with many connections on busy servers.
2119          *
2120          * The criteria to step up the send buffer one notch are:
2121          *  1. receive window of remote host is larger than send buffer
2122          *     (with a fudge factor of 5/4th);
2123          *  2. send buffer is filled to 7/8th with data (so we actually
2124          *     have data to make use of it);
2125          *  3. send buffer fill has not hit maximal automatic size;
2126          *  4. our send window (slow start and cogestion controlled) is
2127          *     larger than sent but unacknowledged data in send buffer.
2128          *
2129          * The remote host receive window scaling factor may limit the
2130          * growing of the send buffer before it reaches its allowed
2131          * maximum.
2132          *
2133          * It scales directly with slow start or congestion window
2134          * and does at most one step per received ACK.  This fast
2135          * scaling has the drawback of growing the send buffer beyond
2136          * what is strictly necessary to make full use of a given
2137          * delay*bandwidth product.  However testing has shown this not
2138          * to be much of an problem.  At worst we are trading wasting
2139          * of available bandwidth (the non-use of it) for wasting some
2140          * socket buffer memory.
2141          *
2142          * TODO: Shrink send buffer during idle periods together
2143          * with congestion window.  Requires another timer.  Has to
2144          * wait for upcoming tcp timer rewrite.
2145          *
2146          * XXXGL: should there be used sbused() or sbavail()?
2147          */
2148         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2149                 int lowat;
2150
2151                 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2152                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2153                     sbused(&so->so_snd) >=
2154                     (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2155                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2156                     sendwin >= (sbused(&so->so_snd) -
2157                     (tp->snd_nxt - tp->snd_una))) {
2158                         if (!sbreserve_locked(so, SO_SND,
2159                             min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2160                              V_tcp_autosndbuf_max), curthread))
2161                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2162                 }
2163         }
2164 }