sys/netinet/tcp_output.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-3-Clause
   3  *
   4  * Copyright (c) 1982, 1986, 1988, 1990, 1993, 1995
   5  *      The Regents of the University of California.  All rights reserved.
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  * 1. Redistributions of source code must retain the above copyright
  11  *    notice, this list of conditions and the following disclaimer.
  12  * 2. Redistributions in binary form must reproduce the above copyright
  13  *    notice, this list of conditions and the following disclaimer in the
  14  *    documentation and/or other materials provided with the distribution.
  15  * 3. Neither the name of the University nor the names of its contributors
  16  *    may be used to endorse or promote products derived from this software
  17  *    without specific prior written permission.
  18  *
  19  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
  20  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  21  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  22  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
  23  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  24  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  25  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  26  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  27  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  28  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  29  * SUCH DAMAGE.
  30  *
  31  *      @(#)tcp_output.c        8.4 (Berkeley) 5/24/95
  32  */
  33
  34 #include <sys/cdefs.h>
  35 __FBSDID("$FreeBSD$");
  36
  37 #include "opt_inet.h"
  38 #include "opt_inet6.h"
  39 #include "opt_ipsec.h"
  40 #include "opt_kern_tls.h"
  41 #include "opt_tcpdebug.h"
  42
  43 #include <sys/param.h>
  44 #include <sys/systm.h>
  45 #include <sys/arb.h>
  46 #include <sys/domain.h>
  47 #ifdef TCP_HHOOK
  48 #include <sys/hhook.h>
  49 #endif
  50 #include <sys/kernel.h>
  51 #ifdef KERN_TLS
  52 #include <sys/ktls.h>
  53 #endif
  54 #include <sys/lock.h>
  55 #include <sys/mbuf.h>
  56 #include <sys/mutex.h>
  57 #include <sys/protosw.h>
  58 #include <sys/qmath.h>
  59 #include <sys/sdt.h>
  60 #include <sys/socket.h>
  61 #include <sys/socketvar.h>
  62 #include <sys/sysctl.h>
  63 #include <sys/stats.h>
  64
  65 #include <net/if.h>
  66 #include <net/route.h>
  67 #include <net/route/nhop.h>
  68 #include <net/vnet.h>
  69
  70 #include <netinet/in.h>
  71 #include <netinet/in_kdtrace.h>
  72 #include <netinet/in_systm.h>
  73 #include <netinet/ip.h>
  74 #include <netinet/in_pcb.h>
  75 #include <netinet/ip_var.h>
  76 #include <netinet/ip_options.h>
  77 #ifdef INET6
  78 #include <netinet6/in6_pcb.h>
  79 #include <netinet/ip6.h>
  80 #include <netinet6/ip6_var.h>
  81 #endif
  82 #include <netinet/tcp.h>
  83 #define TCPOUTFLAGS
  84 #include <netinet/tcp_fsm.h>
  85 #include <netinet/tcp_log_buf.h>
  86 #include <netinet/tcp_seq.h>
  87 #include <netinet/tcp_timer.h>
  88 #include <netinet/tcp_var.h>
  89 #include <netinet/tcpip.h>
  90 #include <netinet/cc/cc.h>
  91 #include <netinet/tcp_fastopen.h>
  92 #ifdef TCPPCAP
  93 #include <netinet/tcp_pcap.h>
  94 #endif
  95 #ifdef TCPDEBUG
  96 #include <netinet/tcp_debug.h>
  97 #endif
  98 #ifdef TCP_OFFLOAD
  99 #include <netinet/tcp_offload.h>
 100 #endif
 101
 102 #include <netipsec/ipsec_support.h>
 103
 104 #include <netinet/udp.h>
 105 #include <netinet/udp_var.h>
 106 #include <machine/in_cksum.h>
 107
 108 #include <security/mac/mac_framework.h>
 109
 110 VNET_DEFINE(int, path_mtu_discovery) = 1;
 111 SYSCTL_INT(_net_inet_tcp, OID_AUTO, path_mtu_discovery, CTLFLAG_VNET | CTLFLAG_RW,
 112         &VNET_NAME(path_mtu_discovery), 1,
 113         "Enable Path MTU Discovery");
 114
 115 VNET_DEFINE(int, tcp_do_tso) = 1;
 116 SYSCTL_INT(_net_inet_tcp, OID_AUTO, tso, CTLFLAG_VNET | CTLFLAG_RW,
 117         &VNET_NAME(tcp_do_tso), 0,
 118         "Enable TCP Segmentation Offload");
 119
 120 VNET_DEFINE(int, tcp_sendspace) = 1024*32;
 121 #define V_tcp_sendspace VNET(tcp_sendspace)
 122 SYSCTL_INT(_net_inet_tcp, TCPCTL_SENDSPACE, sendspace, CTLFLAG_VNET | CTLFLAG_RW,
 123         &VNET_NAME(tcp_sendspace), 0, "Initial send socket buffer size");
 124
 125 VNET_DEFINE(int, tcp_do_autosndbuf) = 1;
 126 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto, CTLFLAG_VNET | CTLFLAG_RW,
 127         &VNET_NAME(tcp_do_autosndbuf), 0,
 128         "Enable automatic send buffer sizing");
 129
 130 VNET_DEFINE(int, tcp_autosndbuf_inc) = 8*1024;
 131 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_inc, CTLFLAG_VNET | CTLFLAG_RW,
 132         &VNET_NAME(tcp_autosndbuf_inc), 0,
 133         "Incrementor step size of automatic send buffer");
 134
 135 VNET_DEFINE(int, tcp_autosndbuf_max) = 2*1024*1024;
 136 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_max, CTLFLAG_VNET | CTLFLAG_RW,
 137         &VNET_NAME(tcp_autosndbuf_max), 0,
 138         "Max size of automatic send buffer");
 139
 140 VNET_DEFINE(int, tcp_sendbuf_auto_lowat) = 0;
 141 #define V_tcp_sendbuf_auto_lowat        VNET(tcp_sendbuf_auto_lowat)
 142 SYSCTL_INT(_net_inet_tcp, OID_AUTO, sendbuf_auto_lowat, CTLFLAG_VNET | CTLFLAG_RW,
 143         &VNET_NAME(tcp_sendbuf_auto_lowat), 0,
 144         "Modify threshold for auto send buffer growth to account for SO_SNDLOWAT");
 145
 146 /*
 147  * Make sure that either retransmit or persist timer is set for SYN, FIN and
 148  * non-ACK.
 149  */
 150 #define TCP_XMIT_TIMER_ASSERT(tp, len, th_flags)                        \
 151         KASSERT(((len) == 0 && ((th_flags) & (TH_SYN | TH_FIN)) == 0) ||\
 152             tcp_timer_active((tp), TT_REXMT) ||                         \
 153             tcp_timer_active((tp), TT_PERSIST),                         \
 154             ("neither rexmt nor persist timer is set"))
 155
 156 static void inline      cc_after_idle(struct tcpcb *tp);
 157
 158 #ifdef TCP_HHOOK
 159 /*
 160  * Wrapper for the TCP established output helper hook.
 161  */
 162 void
 163 hhook_run_tcp_est_out(struct tcpcb *tp, struct tcphdr *th,
 164     struct tcpopt *to, uint32_t len, int tso)
 165 {
 166         struct tcp_hhook_data hhook_data;
 167
 168         if (V_tcp_hhh[HHOOK_TCP_EST_OUT]->hhh_nhooks > 0) {
 169                 hhook_data.tp = tp;
 170                 hhook_data.th = th;
 171                 hhook_data.to = to;
 172                 hhook_data.len = len;
 173                 hhook_data.tso = tso;
 174
 175                 hhook_run_hooks(V_tcp_hhh[HHOOK_TCP_EST_OUT], &hhook_data,
 176                     tp->osd);
 177         }
 178 }
 179 #endif
 180
 181 /*
 182  * CC wrapper hook functions
 183  */
 184 static void inline
 185 cc_after_idle(struct tcpcb *tp)
 186 {
 187         INP_WLOCK_ASSERT(tp->t_inpcb);
 188
 189         if (CC_ALGO(tp)->after_idle != NULL)
 190                 CC_ALGO(tp)->after_idle(tp->ccv);
 191 }
 192
 193 /*
 194  * Tcp output routine: figure out what should be sent and send it.
 195  */
 196 int
 197 tcp_output(struct tcpcb *tp)
 198 {
 199         struct socket *so = tp->t_inpcb->inp_socket;
 200         int32_t len;
 201         uint32_t recwin, sendwin;
 202         int off, flags, error = 0;      /* Keep compiler happy */
 203         u_int if_hw_tsomaxsegcount = 0;
 204         u_int if_hw_tsomaxsegsize = 0;
 205         struct mbuf *m;
 206         struct ip *ip = NULL;
 207 #ifdef TCPDEBUG
 208         struct ipovly *ipov = NULL;
 209 #endif
 210         struct tcphdr *th;
 211         u_char opt[TCP_MAXOLEN];
 212         unsigned ipoptlen, optlen, hdrlen, ulen;
 213 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 214         unsigned ipsec_optlen = 0;
 215 #endif
 216         int idle, sendalot, curticks;
 217         int sack_rxmit, sack_bytes_rxmt;
 218         struct sackhole *p;
 219         int tso, mtu;
 220         struct tcpopt to;
 221         struct udphdr *udp = NULL;
 222         unsigned int wanted_cookie = 0;
 223         unsigned int dont_sendalot = 0;
 224 #if 0
 225         int maxburst = TCP_MAXBURST;
 226 #endif
 227 #ifdef INET6
 228         struct ip6_hdr *ip6 = NULL;
 229         int isipv6;
 230
 231         isipv6 = (tp->t_inpcb->inp_vflag & INP_IPV6) != 0;
 232 #endif
 233 #ifdef KERN_TLS
 234         const bool hw_tls = (so->so_snd.sb_flags & SB_TLS_IFNET) != 0;
 235 #else
 236         const bool hw_tls = false;
 237 #endif
 238
 239         NET_EPOCH_ASSERT();
 240         INP_WLOCK_ASSERT(tp->t_inpcb);
 241
 242 #ifdef TCP_OFFLOAD
 243         if (tp->t_flags & TF_TOE)
 244                 return (tcp_offload_output(tp));
 245 #endif
 246
 247         /*
 248          * For TFO connections in SYN_SENT or SYN_RECEIVED,
 249          * only allow the initial SYN or SYN|ACK and those sent
 250          * by the retransmit timer.
 251          */
 252         if (IS_FASTOPEN(tp->t_flags) &&
 253             ((tp->t_state == TCPS_SYN_SENT) ||
 254              (tp->t_state == TCPS_SYN_RECEIVED)) &&
 255             SEQ_GT(tp->snd_max, tp->snd_una) && /* initial SYN or SYN|ACK sent */
 256             (tp->snd_nxt != tp->snd_una))       /* not a retransmit */
 257                 return (0);
 258
 259         /*
 260          * Determine length of data that should be transmitted,
 261          * and flags that will be used.
 262          * If there is some data or critical controls (SYN, RST)
 263          * to send, then transmit; otherwise, investigate further.
 264          */
 265         idle = (tp->t_flags & TF_LASTIDLE) || (tp->snd_max == tp->snd_una);
 266         if (idle && (((ticks - tp->t_rcvtime) >= tp->t_rxtcur) ||
 267             (tp->t_sndtime && ((ticks - tp->t_sndtime) >= tp->t_rxtcur))))
 268                 cc_after_idle(tp);
 269         tp->t_flags &= ~TF_LASTIDLE;
 270         if (idle) {
 271                 if (tp->t_flags & TF_MORETOCOME) {
 272                         tp->t_flags |= TF_LASTIDLE;
 273                         idle = 0;
 274                 }
 275         }
 276 again:
 277         /*
 278          * If we've recently taken a timeout, snd_max will be greater than
 279          * snd_nxt.  There may be SACK information that allows us to avoid
 280          * resending already delivered data.  Adjust snd_nxt accordingly.
 281          */
 282         if ((tp->t_flags & TF_SACK_PERMIT) &&
 283             SEQ_LT(tp->snd_nxt, tp->snd_max))
 284                 tcp_sack_adjust(tp);
 285         sendalot = 0;
 286         tso = 0;
 287         mtu = 0;
 288         off = tp->snd_nxt - tp->snd_una;
 289         sendwin = min(tp->snd_wnd, tp->snd_cwnd);
 290
 291         flags = tcp_outflags[tp->t_state];
 292         /*
 293          * Send any SACK-generated retransmissions.  If we're explicitly trying
 294          * to send out new data (when sendalot is 1), bypass this function.
 295          * If we retransmit in fast recovery mode, decrement snd_cwnd, since
 296          * we're replacing a (future) new transmission with a retransmission
 297          * now, and we previously incremented snd_cwnd in tcp_input().
 298          */
 299         /*
 300          * Still in sack recovery , reset rxmit flag to zero.
 301          */
 302         sack_rxmit = 0;
 303         sack_bytes_rxmt = 0;
 304         len = 0;
 305         p = NULL;
 306         if ((tp->t_flags & TF_SACK_PERMIT) && IN_FASTRECOVERY(tp->t_flags) &&
 307             (p = tcp_sack_output(tp, &sack_bytes_rxmt))) {
 308                 uint32_t cwin;
 309
 310                 cwin =
 311                     imax(min(tp->snd_wnd, tp->snd_cwnd) - sack_bytes_rxmt, 0);
 312                 /* Do not retransmit SACK segments beyond snd_recover */
 313                 if (SEQ_GT(p->end, tp->snd_recover)) {
 314                         /*
 315                          * (At least) part of sack hole extends beyond
 316                          * snd_recover. Check to see if we can rexmit data
 317                          * for this hole.
 318                          */
 319                         if (SEQ_GEQ(p->rxmit, tp->snd_recover)) {
 320                                 /*
 321                                  * Can't rexmit any more data for this hole.
 322                                  * That data will be rexmitted in the next
 323                                  * sack recovery episode, when snd_recover
 324                                  * moves past p->rxmit.
 325                                  */
 326                                 p = NULL;
 327                                 goto after_sack_rexmit;
 328                         } else {
 329                                 /* Can rexmit part of the current hole */
 330                                 len = ((int32_t)ulmin(cwin,
 331                                     SEQ_SUB(tp->snd_recover, p->rxmit)));
 332                         }
 333                 } else {
 334                         len = ((int32_t)ulmin(cwin,
 335                             SEQ_SUB(p->end, p->rxmit)));
 336                 }
 337                 if (len > 0) {
 338                         off = SEQ_SUB(p->rxmit, tp->snd_una);
 339                         KASSERT(off >= 0,("%s: sack block to the left of una : %d",
 340                             __func__, off));
 341                         sack_rxmit = 1;
 342                         sendalot = 1;
 343                         TCPSTAT_INC(tcps_sack_rexmits);
 344                         TCPSTAT_ADD(tcps_sack_rexmit_bytes,
 345                             min(len, tcp_maxseg(tp)));
 346                 }
 347         }
 348 after_sack_rexmit:
 349         /*
 350          * Get standard flags, and add SYN or FIN if requested by 'hidden'
 351          * state flags.
 352          */
 353         if (tp->t_flags & TF_NEEDFIN)
 354                 flags |= TH_FIN;
 355         if (tp->t_flags & TF_NEEDSYN)
 356                 flags |= TH_SYN;
 357
 358         SOCKBUF_LOCK(&so->so_snd);
 359         /*
 360          * If in persist timeout with window of 0, send 1 byte.
 361          * Otherwise, if window is small but nonzero
 362          * and timer expired, we will send what we can
 363          * and go to transmit state.
 364          */
 365         if (tp->t_flags & TF_FORCEDATA) {
 366                 if (sendwin == 0) {
 367                         /*
 368                          * If we still have some data to send, then
 369                          * clear the FIN bit.  Usually this would
 370                          * happen below when it realizes that we
 371                          * aren't sending all the data.  However,
 372                          * if we have exactly 1 byte of unsent data,
 373                          * then it won't clear the FIN bit below,
 374                          * and if we are in persist state, we wind
 375                          * up sending the packet without recording
 376                          * that we sent the FIN bit.
 377                          *
 378                          * We can't just blindly clear the FIN bit,
 379                          * because if we don't have any more data
 380                          * to send then the probe will be the FIN
 381                          * itself.
 382                          */
 383                         if (off < sbused(&so->so_snd))
 384                                 flags &= ~TH_FIN;
 385                         sendwin = 1;
 386                 } else {
 387                         tcp_timer_activate(tp, TT_PERSIST, 0);
 388                         tp->t_rxtshift = 0;
 389                 }
 390         }
 391
 392         /*
 393          * If snd_nxt == snd_max and we have transmitted a FIN, the
 394          * offset will be > 0 even if so_snd.sb_cc is 0, resulting in
 395          * a negative length.  This can also occur when TCP opens up
 396          * its congestion window while receiving additional duplicate
 397          * acks after fast-retransmit because TCP will reset snd_nxt
 398          * to snd_max after the fast-retransmit.
 399          *
 400          * In the normal retransmit-FIN-only case, however, snd_nxt will
 401          * be set to snd_una, the offset will be 0, and the length may
 402          * wind up 0.
 403          *
 404          * If sack_rxmit is true we are retransmitting from the scoreboard
 405          * in which case len is already set.
 406          */
 407         if (sack_rxmit == 0) {
 408                 if (sack_bytes_rxmt == 0)
 409                         len = ((int32_t)min(sbavail(&so->so_snd), sendwin) -
 410                             off);
 411                 else {
 412                         int32_t cwin;
 413
 414                         /*
 415                          * We are inside of a SACK recovery episode and are
 416                          * sending new data, having retransmitted all the
 417                          * data possible in the scoreboard.
 418                          */
 419                         len = ((int32_t)min(sbavail(&so->so_snd), tp->snd_wnd) -
 420                             off);
 421                         /*
 422                          * Don't remove this (len > 0) check !
 423                          * We explicitly check for len > 0 here (although it
 424                          * isn't really necessary), to work around a gcc
 425                          * optimization issue - to force gcc to compute
 426                          * len above. Without this check, the computation
 427                          * of len is bungled by the optimizer.
 428                          */
 429                         if (len > 0) {
 430                                 cwin = tp->snd_cwnd - imax(0, (int32_t)
 431                                         (tp->snd_nxt - tp->snd_recover)) -
 432                                         sack_bytes_rxmt;
 433                                 if (cwin < 0)
 434                                         cwin = 0;
 435                                 len = imin(len, cwin);
 436                         }
 437                 }
 438         }
 439
 440         /*
 441          * Lop off SYN bit if it has already been sent.  However, if this
 442          * is SYN-SENT state and if segment contains data and if we don't
 443          * know that foreign host supports TAO, suppress sending segment.
 444          */
 445         if ((flags & TH_SYN) && SEQ_GT(tp->snd_nxt, tp->snd_una)) {
 446                 if (tp->t_state != TCPS_SYN_RECEIVED)
 447                         flags &= ~TH_SYN;
 448                 /*
 449                  * When sending additional segments following a TFO SYN|ACK,
 450                  * do not include the SYN bit.
 451                  */
 452                 if (IS_FASTOPEN(tp->t_flags) &&
 453                     (tp->t_state == TCPS_SYN_RECEIVED))
 454                         flags &= ~TH_SYN;
 455                 off--, len++;
 456         }
 457
 458         /*
 459          * Be careful not to send data and/or FIN on SYN segments.
 460          * This measure is needed to prevent interoperability problems
 461          * with not fully conformant TCP implementations.
 462          */
 463         if ((flags & TH_SYN) && (tp->t_flags & TF_NOOPT)) {
 464                 len = 0;
 465                 flags &= ~TH_FIN;
 466         }
 467
 468         /*
 469          * On TFO sockets, ensure no data is sent in the following cases:
 470          *
 471          *  - When retransmitting SYN|ACK on a passively-created socket
 472          *
 473          *  - When retransmitting SYN on an actively created socket
 474          *
 475          *  - When sending a zero-length cookie (cookie request) on an
 476          *    actively created socket
 477          *
 478          *  - When the socket is in the CLOSED state (RST is being sent)
 479          */
 480         if (IS_FASTOPEN(tp->t_flags) &&
 481             (((flags & TH_SYN) && (tp->t_rxtshift > 0)) ||
 482              ((tp->t_state == TCPS_SYN_SENT) &&
 483               (tp->t_tfo_client_cookie_len == 0)) ||
 484              (flags & TH_RST)))
 485                 len = 0;
 486         if (len <= 0) {
 487                 /*
 488                  * If FIN has been sent but not acked,
 489                  * but we haven't been called to retransmit,
 490                  * len will be < 0.  Otherwise, window shrank
 491                  * after we sent into it.  If window shrank to 0,
 492                  * cancel pending retransmit, pull snd_nxt back
 493                  * to (closed) window, and set the persist timer
 494                  * if it isn't already going.  If the window didn't
 495                  * close completely, just wait for an ACK.
 496                  *
 497                  * We also do a general check here to ensure that
 498                  * we will set the persist timer when we have data
 499                  * to send, but a 0-byte window. This makes sure
 500                  * the persist timer is set even if the packet
 501                  * hits one of the "goto send" lines below.
 502                  */
 503                 len = 0;
 504                 if ((sendwin == 0) && (TCPS_HAVEESTABLISHED(tp->t_state)) &&
 505                         (off < (int) sbavail(&so->so_snd))) {
 506                         tcp_timer_activate(tp, TT_REXMT, 0);
 507                         tp->t_rxtshift = 0;
 508                         tp->snd_nxt = tp->snd_una;
 509                         if (!tcp_timer_active(tp, TT_PERSIST))
 510                                 tcp_setpersist(tp);
 511                 }
 512         }
 513
 514         /* len will be >= 0 after this point. */
 515         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
 516
 517         tcp_sndbuf_autoscale(tp, so, sendwin);
 518
 519         /*
 520          * Decide if we can use TCP Segmentation Offloading (if supported by
 521          * hardware).
 522          *
 523          * TSO may only be used if we are in a pure bulk sending state.  The
 524          * presence of TCP-MD5, SACK retransmits, SACK advertizements and
 525          * IP options prevent using TSO.  With TSO the TCP header is the same
 526          * (except for the sequence number) for all generated packets.  This
 527          * makes it impossible to transmit any options which vary per generated
 528          * segment or packet.
 529          *
 530          * IPv4 handling has a clear separation of ip options and ip header
 531          * flags while IPv6 combines both in in6p_outputopts. ip6_optlen() does
 532          * the right thing below to provide length of just ip options and thus
 533          * checking for ipoptlen is enough to decide if ip options are present.
 534          */
 535 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 536         /*
 537          * Pre-calculate here as we save another lookup into the darknesses
 538          * of IPsec that way and can actually decide if TSO is ok.
 539          */
 540 #ifdef INET6
 541         if (isipv6 && IPSEC_ENABLED(ipv6))
 542                 ipsec_optlen = IPSEC_HDRSIZE(ipv6, tp->t_inpcb);
 543 #ifdef INET
 544         else
 545 #endif
 546 #endif /* INET6 */
 547 #ifdef INET
 548         if (IPSEC_ENABLED(ipv4))
 549                 ipsec_optlen = IPSEC_HDRSIZE(ipv4, tp->t_inpcb);
 550 #endif /* INET */
 551 #endif /* IPSEC */
 552 #ifdef INET6
 553         if (isipv6)
 554                 ipoptlen = ip6_optlen(tp->t_inpcb);
 555         else
 556 #endif
 557         if (tp->t_inpcb->inp_options)
 558                 ipoptlen = tp->t_inpcb->inp_options->m_len -
 559                                 offsetof(struct ipoption, ipopt_list);
 560         else
 561                 ipoptlen = 0;
 562 #if defined(IPSEC) || defined(IPSEC_SUPPORT)
 563         ipoptlen += ipsec_optlen;
 564 #endif
 565
 566         if ((tp->t_flags & TF_TSO) && V_tcp_do_tso && len > tp->t_maxseg &&
 567             (tp->t_port == 0) &&
 568             ((tp->t_flags & TF_SIGNATURE) == 0) &&
 569             tp->rcv_numsacks == 0 && sack_rxmit == 0 &&
 570             ipoptlen == 0 && !(flags & TH_SYN))
 571                 tso = 1;
 572
 573         if (sack_rxmit) {
 574                 if (SEQ_LT(p->rxmit + len, tp->snd_una + sbused(&so->so_snd)))
 575                         flags &= ~TH_FIN;
 576         } else {
 577                 if (SEQ_LT(tp->snd_nxt + len, tp->snd_una +
 578                     sbused(&so->so_snd)))
 579                         flags &= ~TH_FIN;
 580         }
 581
 582         recwin = lmin(lmax(sbspace(&so->so_rcv), 0),
 583             (long)TCP_MAXWIN << tp->rcv_scale);
 584
 585         /*
 586          * Sender silly window avoidance.   We transmit under the following
 587          * conditions when len is non-zero:
 588          *
 589          *      - We have a full segment (or more with TSO)
 590          *      - This is the last buffer in a write()/send() and we are
 591          *        either idle or running NODELAY
 592          *      - we've timed out (e.g. persist timer)
 593          *      - we have more then 1/2 the maximum send window's worth of
 594          *        data (receiver may be limited the window size)
 595          *      - we need to retransmit
 596          */
 597         if (len) {
 598                 if (len >= tp->t_maxseg)
 599                         goto send;
 600                 /*
 601                  * As the TCP header options are now
 602                  * considered when setting up the initial
 603                  * window, we would not send the last segment
 604                  * if we skip considering the option length here.
 605                  * Note: this may not work when tcp headers change
 606                  * very dynamically in the future.
 607                  */
 608                 if ((((tp->t_flags & TF_SIGNATURE) ?
 609                         PADTCPOLEN(TCPOLEN_SIGNATURE) : 0) +
 610                     ((tp->t_flags & TF_RCVD_TSTMP) ?
 611                         PADTCPOLEN(TCPOLEN_TIMESTAMP) : 0) +
 612                     len) >= tp->t_maxseg)
 613                         goto send;
 614                 /*
 615                  * NOTE! on localhost connections an 'ack' from the remote
 616                  * end may occur synchronously with the output and cause
 617                  * us to flush a buffer queued with moretocome.  XXX
 618                  *
 619                  * note: the len + off check is almost certainly unnecessary.
 620                  */
 621                 if (!(tp->t_flags & TF_MORETOCOME) &&   /* normal case */
 622                     (idle || (tp->t_flags & TF_NODELAY)) &&
 623                     (uint32_t)len + (uint32_t)off >= sbavail(&so->so_snd) &&
 624                     (tp->t_flags & TF_NOPUSH) == 0) {
 625                         goto send;
 626                 }
 627                 if (tp->t_flags & TF_FORCEDATA)         /* typ. timeout case */
 628                         goto send;
 629                 if (len >= tp->max_sndwnd / 2 && tp->max_sndwnd > 0)
 630                         goto send;
 631                 if (SEQ_LT(tp->snd_nxt, tp->snd_max))   /* retransmit case */
 632                         goto send;
 633                 if (sack_rxmit)
 634                         goto send;
 635         }
 636
 637         /*
 638          * Sending of standalone window updates.
 639          *
 640          * Window updates are important when we close our window due to a
 641          * full socket buffer and are opening it again after the application
 642          * reads data from it.  Once the window has opened again and the
 643          * remote end starts to send again the ACK clock takes over and
 644          * provides the most current window information.
 645          *
 646          * We must avoid the silly window syndrome whereas every read
 647          * from the receive buffer, no matter how small, causes a window
 648          * update to be sent.  We also should avoid sending a flurry of
 649          * window updates when the socket buffer had queued a lot of data
 650          * and the application is doing small reads.
 651          *
 652          * Prevent a flurry of pointless window updates by only sending
 653          * an update when we can increase the advertized window by more
 654          * than 1/4th of the socket buffer capacity.  When the buffer is
 655          * getting full or is very small be more aggressive and send an
 656          * update whenever we can increase by two mss sized segments.
 657          * In all other situations the ACK's to new incoming data will
 658          * carry further window increases.
 659          *
 660          * Don't send an independent window update if a delayed
 661          * ACK is pending (it will get piggy-backed on it) or the
 662          * remote side already has done a half-close and won't send
 663          * more data.  Skip this if the connection is in T/TCP
 664          * half-open state.
 665          */
 666         if (recwin > 0 && !(tp->t_flags & TF_NEEDSYN) &&
 667             !(tp->t_flags & TF_DELACK) &&
 668             !TCPS_HAVERCVDFIN(tp->t_state)) {
 669                 /*
 670                  * "adv" is the amount we could increase the window,
 671                  * taking into account that we are limited by
 672                  * TCP_MAXWIN << tp->rcv_scale.
 673                  */
 674                 int32_t adv;
 675                 int oldwin;
 676
 677                 adv = recwin;
 678                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt)) {
 679                         oldwin = (tp->rcv_adv - tp->rcv_nxt);
 680                         if (adv > oldwin)
 681                                 adv -= oldwin;
 682                         else
 683                                 adv = 0;
 684                 } else
 685                         oldwin = 0;
 686
 687                 /*
 688                  * If the new window size ends up being the same as or less
 689                  * than the old size when it is scaled, then don't force
 690                  * a window update.
 691                  */
 692                 if (oldwin >> tp->rcv_scale >= (adv + oldwin) >> tp->rcv_scale)
 693                         goto dontupdate;
 694
 695                 if (adv >= (int32_t)(2 * tp->t_maxseg) &&
 696                     (adv >= (int32_t)(so->so_rcv.sb_hiwat / 4) ||
 697                      recwin <= (so->so_rcv.sb_hiwat / 8) ||
 698                      so->so_rcv.sb_hiwat <= 8 * tp->t_maxseg ||
 699                      adv >= TCP_MAXWIN << tp->rcv_scale))
 700                         goto send;
 701                 if (2 * adv >= (int32_t)so->so_rcv.sb_hiwat)
 702                         goto send;
 703         }
 704 dontupdate:
 705
 706         /*
 707          * Send if we owe the peer an ACK, RST, SYN, or urgent data.  ACKNOW
 708          * is also a catch-all for the retransmit timer timeout case.
 709          */
 710         if (tp->t_flags & TF_ACKNOW)
 711                 goto send;
 712         if ((flags & TH_RST) ||
 713             ((flags & TH_SYN) && (tp->t_flags & TF_NEEDSYN) == 0))
 714                 goto send;
 715         if (SEQ_GT(tp->snd_up, tp->snd_una))
 716                 goto send;
 717         /*
 718          * If our state indicates that FIN should be sent
 719          * and we have not yet done so, then we need to send.
 720          */
 721         if (flags & TH_FIN &&
 722             ((tp->t_flags & TF_SENTFIN) == 0 || tp->snd_nxt == tp->snd_una))
 723                 goto send;
 724         /*
 725          * In SACK, it is possible for tcp_output to fail to send a segment
 726          * after the retransmission timer has been turned off.  Make sure
 727          * that the retransmission timer is set.
 728          */
 729         if ((tp->t_flags & TF_SACK_PERMIT) &&
 730             SEQ_GT(tp->snd_max, tp->snd_una) &&
 731             !tcp_timer_active(tp, TT_REXMT) &&
 732             !tcp_timer_active(tp, TT_PERSIST)) {
 733                 tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
 734                 goto just_return;
 735         }
 736         /*
 737          * TCP window updates are not reliable, rather a polling protocol
 738          * using ``persist'' packets is used to insure receipt of window
 739          * updates.  The three ``states'' for the output side are:
 740          *      idle                    not doing retransmits or persists
 741          *      persisting              to move a small or zero window
 742          *      (re)transmitting        and thereby not persisting
 743          *
 744          * tcp_timer_active(tp, TT_PERSIST)
 745          *      is true when we are in persist state.
 746          * (tp->t_flags & TF_FORCEDATA)
 747          *      is set when we are called to send a persist packet.
 748          * tcp_timer_active(tp, TT_REXMT)
 749          *      is set when we are retransmitting
 750          * The output side is idle when both timers are zero.
 751          *
 752          * If send window is too small, there is data to transmit, and no
 753          * retransmit or persist is pending, then go to persist state.
 754          * If nothing happens soon, send when timer expires:
 755          * if window is nonzero, transmit what we can,
 756          * otherwise force out a byte.
 757          */
 758         if (sbavail(&so->so_snd) && !tcp_timer_active(tp, TT_REXMT) &&
 759             !tcp_timer_active(tp, TT_PERSIST)) {
 760                 tp->t_rxtshift = 0;
 761                 tcp_setpersist(tp);
 762         }
 763
 764         /*
 765          * No reason to send a segment, just return.
 766          */
 767 just_return:
 768         SOCKBUF_UNLOCK(&so->so_snd);
 769         return (0);
 770
 771 send:
 772         SOCKBUF_LOCK_ASSERT(&so->so_snd);
 773         if (len > 0) {
 774                 if (len >= tp->t_maxseg)
 775                         tp->t_flags2 |= TF2_PLPMTU_MAXSEGSNT;
 776                 else
 777                         tp->t_flags2 &= ~TF2_PLPMTU_MAXSEGSNT;
 778         }
 779         /*
 780          * Before ESTABLISHED, force sending of initial options
 781          * unless TCP set not to do any options.
 782          * NOTE: we assume that the IP/TCP header plus TCP options
 783          * always fit in a single mbuf, leaving room for a maximum
 784          * link header, i.e.
 785          *      max_linkhdr + sizeof (struct tcpiphdr) + optlen <= MCLBYTES
 786          */
 787         optlen = 0;
 788 #ifdef INET6
 789         if (isipv6)
 790                 hdrlen = sizeof (struct ip6_hdr) + sizeof (struct tcphdr);
 791         else
 792 #endif
 793                 hdrlen = sizeof (struct tcpiphdr);
 794
 795         if (flags & TH_SYN) {
 796                 tp->snd_nxt = tp->iss;
 797         }
 798
 799         /*
 800          * Compute options for segment.
 801          * We only have to care about SYN and established connection
 802          * segments.  Options for SYN-ACK segments are handled in TCP
 803          * syncache.
 804          */
 805         to.to_flags = 0;
 806         if ((tp->t_flags & TF_NOOPT) == 0) {
 807                 /* Maximum segment size. */
 808                 if (flags & TH_SYN) {
 809                         to.to_mss = tcp_mssopt(&tp->t_inpcb->inp_inc);
 810                         if (tp->t_port)
 811                                 to.to_mss -= V_tcp_udp_tunneling_overhead;
 812                         to.to_flags |= TOF_MSS;
 813
 814                         /*
 815                          * On SYN or SYN|ACK transmits on TFO connections,
 816                          * only include the TFO option if it is not a
 817                          * retransmit, as the presence of the TFO option may
 818                          * have caused the original SYN or SYN|ACK to have
 819                          * been dropped by a middlebox.
 820                          */
 821                         if (IS_FASTOPEN(tp->t_flags) &&
 822                             (tp->t_rxtshift == 0)) {
 823                                 if (tp->t_state == TCPS_SYN_RECEIVED) {
 824                                         to.to_tfo_len = TCP_FASTOPEN_COOKIE_LEN;
 825                                         to.to_tfo_cookie =
 826                                             (u_int8_t *)&tp->t_tfo_cookie.server;
 827                                         to.to_flags |= TOF_FASTOPEN;
 828                                         wanted_cookie = 1;
 829                                 } else if (tp->t_state == TCPS_SYN_SENT) {
 830                                         to.to_tfo_len =
 831                                             tp->t_tfo_client_cookie_len;
 832                                         to.to_tfo_cookie =
 833                                             tp->t_tfo_cookie.client;
 834                                         to.to_flags |= TOF_FASTOPEN;
 835                                         wanted_cookie = 1;
 836                                         /*
 837                                          * If we wind up having more data to
 838                                          * send with the SYN than can fit in
 839                                          * one segment, don't send any more
 840                                          * until the SYN|ACK comes back from
 841                                          * the other end.
 842                                          */
 843                                         dont_sendalot = 1;
 844                                 }
 845                         }
 846                 }
 847                 /* Window scaling. */
 848                 if ((flags & TH_SYN) && (tp->t_flags & TF_REQ_SCALE)) {
 849                         to.to_wscale = tp->request_r_scale;
 850                         to.to_flags |= TOF_SCALE;
 851                 }
 852                 /* Timestamps. */
 853                 if ((tp->t_flags & TF_RCVD_TSTMP) ||
 854                     ((flags & TH_SYN) && (tp->t_flags & TF_REQ_TSTMP))) {
 855                         curticks = tcp_ts_getticks();
 856                         to.to_tsval = curticks + tp->ts_offset;
 857                         to.to_tsecr = tp->ts_recent;
 858                         to.to_flags |= TOF_TS;
 859                         if (tp->t_rxtshift == 1)
 860                                 tp->t_badrxtwin = curticks;
 861                 }
 862
 863                 /* Set receive buffer autosizing timestamp. */
 864                 if (tp->rfbuf_ts == 0 &&
 865                     (so->so_rcv.sb_flags & SB_AUTOSIZE))
 866                         tp->rfbuf_ts = tcp_ts_getticks();
 867
 868                 /* Selective ACK's. */
 869                 if (tp->t_flags & TF_SACK_PERMIT) {
 870                         if (flags & TH_SYN)
 871                                 to.to_flags |= TOF_SACKPERM;
 872                         else if (TCPS_HAVEESTABLISHED(tp->t_state) &&
 873                             tp->rcv_numsacks > 0) {
 874                                 to.to_flags |= TOF_SACK;
 875                                 to.to_nsacks = tp->rcv_numsacks;
 876                                 to.to_sacks = (u_char *)tp->sackblks;
 877                         }
 878                 }
 879 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
 880                 /* TCP-MD5 (RFC2385). */
 881                 /*
 882                  * Check that TCP_MD5SIG is enabled in tcpcb to
 883                  * account the size needed to set this TCP option.
 884                  */
 885                 if (tp->t_flags & TF_SIGNATURE)
 886                         to.to_flags |= TOF_SIGNATURE;
 887 #endif /* TCP_SIGNATURE */
 888
 889                 /* Processing the options. */
 890                 hdrlen += optlen = tcp_addoptions(&to, opt);
 891                 /*
 892                  * If we wanted a TFO option to be added, but it was unable
 893                  * to fit, ensure no data is sent.
 894                  */
 895                 if (IS_FASTOPEN(tp->t_flags) && wanted_cookie &&
 896                     !(to.to_flags & TOF_FASTOPEN))
 897                         len = 0;
 898         }
 899         if (tp->t_port) {
 900                 if (V_tcp_udp_tunneling_port == 0) {
 901                         /* The port was removed?? */
 902                         SOCKBUF_UNLOCK(&so->so_snd);
 903                         return (EHOSTUNREACH);
 904                 }
 905                 hdrlen += sizeof(struct udphdr);
 906         }
 907         /*
 908          * Adjust data length if insertion of options will
 909          * bump the packet length beyond the t_maxseg length.
 910          * Clear the FIN bit because we cut off the tail of
 911          * the segment.
 912          */
 913         if (len + optlen + ipoptlen > tp->t_maxseg) {
 914                 flags &= ~TH_FIN;
 915
 916                 if (tso) {
 917                         u_int if_hw_tsomax;
 918                         u_int moff;
 919                         int max_len;
 920
 921                         /* extract TSO information */
 922                         if_hw_tsomax = tp->t_tsomax;
 923                         if_hw_tsomaxsegcount = tp->t_tsomaxsegcount;
 924                         if_hw_tsomaxsegsize = tp->t_tsomaxsegsize;
 925
 926                         /*
 927                          * Limit a TSO burst to prevent it from
 928                          * overflowing or exceeding the maximum length
 929                          * allowed by the network interface:
 930                          */
 931                         KASSERT(ipoptlen == 0,
 932                             ("%s: TSO can't do IP options", __func__));
 933
 934                         /*
 935                          * Check if we should limit by maximum payload
 936                          * length:
 937                          */
 938                         if (if_hw_tsomax != 0) {
 939                                 /* compute maximum TSO length */
 940                                 max_len = (if_hw_tsomax - hdrlen -
 941                                     max_linkhdr);
 942                                 if (max_len <= 0) {
 943                                         len = 0;
 944                                 } else if (len > max_len) {
 945                                         sendalot = 1;
 946                                         len = max_len;
 947                                 }
 948                         }
 949
 950                         /*
 951                          * Prevent the last segment from being
 952                          * fractional unless the send sockbuf can be
 953                          * emptied:
 954                          */
 955                         max_len = (tp->t_maxseg - optlen);
 956                         if (((uint32_t)off + (uint32_t)len) <
 957                             sbavail(&so->so_snd)) {
 958                                 moff = len % max_len;
 959                                 if (moff != 0) {
 960                                         len -= moff;
 961                                         sendalot = 1;
 962                                 }
 963                         }
 964
 965                         /*
 966                          * In case there are too many small fragments
 967                          * don't use TSO:
 968                          */
 969                         if (len <= max_len) {
 970                                 len = max_len;
 971                                 sendalot = 1;
 972                                 tso = 0;
 973                         }
 974
 975                         /*
 976                          * Send the FIN in a separate segment
 977                          * after the bulk sending is done.
 978                          * We don't trust the TSO implementations
 979                          * to clear the FIN flag on all but the
 980                          * last segment.
 981                          */
 982                         if (tp->t_flags & TF_NEEDFIN)
 983                                 sendalot = 1;
 984                 } else {
 985                         if (optlen + ipoptlen >= tp->t_maxseg) {
 986                                 /*
 987                                  * Since we don't have enough space to put
 988                                  * the IP header chain and the TCP header in
 989                                  * one packet as required by RFC 7112, don't
 990                                  * send it. Also ensure that at least one
 991                                  * byte of the payload can be put into the
 992                                  * TCP segment.
 993                                  */
 994                                 SOCKBUF_UNLOCK(&so->so_snd);
 995                                 error = EMSGSIZE;
 996                                 sack_rxmit = 0;
 997                                 goto out;
 998                         }
 999                         len = tp->t_maxseg - optlen - ipoptlen;
1000                         sendalot = 1;
1001                         if (dont_sendalot)
1002                                 sendalot = 0;
1003                 }
1004         } else
1005                 tso = 0;
1006
1007         KASSERT(len + hdrlen + ipoptlen <= IP_MAXPACKET,
1008             ("%s: len > IP_MAXPACKET", __func__));
1009
1010 /*#ifdef DIAGNOSTIC*/
1011 #ifdef INET6
1012         if (max_linkhdr + hdrlen > MCLBYTES)
1013 #else
1014         if (max_linkhdr + hdrlen > MHLEN)
1015 #endif
1016                 panic("tcphdr too big");
1017 /*#endif*/
1018
1019         /*
1020          * This KASSERT is here to catch edge cases at a well defined place.
1021          * Before, those had triggered (random) panic conditions further down.
1022          */
1023         KASSERT(len >= 0, ("[%s:%d]: len < 0", __func__, __LINE__));
1024
1025         /*
1026          * Grab a header mbuf, attaching a copy of data to
1027          * be transmitted, and initialize the header from
1028          * the template for sends on this connection.
1029          */
1030         if (len) {
1031                 struct mbuf *mb;
1032                 struct sockbuf *msb;
1033                 u_int moff;
1034
1035                 if ((tp->t_flags & TF_FORCEDATA) && len == 1) {
1036                         TCPSTAT_INC(tcps_sndprobe);
1037 #ifdef STATS
1038                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1039                                 stats_voi_update_abs_u32(tp->t_stats,
1040                                 VOI_TCP_RETXPB, len);
1041                         else
1042                                 stats_voi_update_abs_u64(tp->t_stats,
1043                                     VOI_TCP_TXPB, len);
1044 #endif /* STATS */
1045                 } else if (SEQ_LT(tp->snd_nxt, tp->snd_max) || sack_rxmit) {
1046                         tp->t_sndrexmitpack++;
1047                         TCPSTAT_INC(tcps_sndrexmitpack);
1048                         TCPSTAT_ADD(tcps_sndrexmitbyte, len);
1049 #ifdef STATS
1050                         stats_voi_update_abs_u32(tp->t_stats, VOI_TCP_RETXPB,
1051                             len);
1052 #endif /* STATS */
1053                 } else {
1054                         TCPSTAT_INC(tcps_sndpack);
1055                         TCPSTAT_ADD(tcps_sndbyte, len);
1056 #ifdef STATS
1057                         stats_voi_update_abs_u64(tp->t_stats, VOI_TCP_TXPB,
1058                             len);
1059 #endif /* STATS */
1060                 }
1061 #ifdef INET6
1062                 if (MHLEN < hdrlen + max_linkhdr)
1063                         m = m_getcl(M_NOWAIT, MT_DATA, M_PKTHDR);
1064                 else
1065 #endif
1066                         m = m_gethdr(M_NOWAIT, MT_DATA);
1067
1068                 if (m == NULL) {
1069                         SOCKBUF_UNLOCK(&so->so_snd);
1070                         error = ENOBUFS;
1071                         sack_rxmit = 0;
1072                         goto out;
1073                 }
1074
1075                 m->m_data += max_linkhdr;
1076                 m->m_len = hdrlen;
1077
1078                 /*
1079                  * Start the m_copy functions from the closest mbuf
1080                  * to the offset in the socket buffer chain.
1081                  */
1082                 mb = sbsndptr_noadv(&so->so_snd, off, &moff);
1083                 if (len <= MHLEN - hdrlen - max_linkhdr && !hw_tls) {
1084                         m_copydata(mb, moff, len,
1085                             mtod(m, caddr_t) + hdrlen);
1086                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1087                                 sbsndptr_adv(&so->so_snd, mb, len);
1088                         m->m_len += len;
1089                 } else {
1090                         if (SEQ_LT(tp->snd_nxt, tp->snd_max))
1091                                 msb = NULL;
1092                         else
1093                                 msb = &so->so_snd;
1094                         m->m_next = tcp_m_copym(mb, moff,
1095                             &len, if_hw_tsomaxsegcount,
1096                             if_hw_tsomaxsegsize, msb, hw_tls);
1097                         if (len <= (tp->t_maxseg - optlen)) {
1098                                 /*
1099                                  * Must have ran out of mbufs for the copy
1100                                  * shorten it to no longer need tso. Lets
1101                                  * not put on sendalot since we are low on
1102                                  * mbufs.
1103                                  */
1104                                 tso = 0;
1105                         }
1106                         if (m->m_next == NULL) {
1107                                 SOCKBUF_UNLOCK(&so->so_snd);
1108                                 (void) m_free(m);
1109                                 error = ENOBUFS;
1110                                 sack_rxmit = 0;
1111                                 goto out;
1112                         }
1113                 }
1114
1115                 /*
1116                  * If we're sending everything we've got, set PUSH.
1117                  * (This will keep happy those implementations which only
1118                  * give data to the user when a buffer fills or
1119                  * a PUSH comes in.)
1120                  */
1121                 if (((uint32_t)off + (uint32_t)len == sbused(&so->so_snd)) &&
1122                     !(flags & TH_SYN))
1123                         flags |= TH_PUSH;
1124                 SOCKBUF_UNLOCK(&so->so_snd);
1125         } else {
1126                 SOCKBUF_UNLOCK(&so->so_snd);
1127                 if (tp->t_flags & TF_ACKNOW)
1128                         TCPSTAT_INC(tcps_sndacks);
1129                 else if (flags & (TH_SYN|TH_FIN|TH_RST))
1130                         TCPSTAT_INC(tcps_sndctrl);
1131                 else if (SEQ_GT(tp->snd_up, tp->snd_una))
1132                         TCPSTAT_INC(tcps_sndurg);
1133                 else
1134                         TCPSTAT_INC(tcps_sndwinup);
1135
1136                 m = m_gethdr(M_NOWAIT, MT_DATA);
1137                 if (m == NULL) {
1138                         error = ENOBUFS;
1139                         sack_rxmit = 0;
1140                         goto out;
1141                 }
1142 #ifdef INET6
1143                 if (isipv6 && (MHLEN < hdrlen + max_linkhdr) &&
1144                     MHLEN >= hdrlen) {
1145                         M_ALIGN(m, hdrlen);
1146                 } else
1147 #endif
1148                 m->m_data += max_linkhdr;
1149                 m->m_len = hdrlen;
1150         }
1151         SOCKBUF_UNLOCK_ASSERT(&so->so_snd);
1152         m->m_pkthdr.rcvif = (struct ifnet *)0;
1153 #ifdef MAC
1154         mac_inpcb_create_mbuf(tp->t_inpcb, m);
1155 #endif
1156 #ifdef INET6
1157         if (isipv6) {
1158                 ip6 = mtod(m, struct ip6_hdr *);
1159                 if (tp->t_port) {
1160                         udp = (struct udphdr *)((caddr_t)ip6 + sizeof(struct ip6_hdr));
1161                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1162                         udp->uh_dport = tp->t_port;
1163                         ulen = hdrlen + len - sizeof(struct ip6_hdr);
1164                         udp->uh_ulen = htons(ulen);
1165                         th = (struct tcphdr *)(udp + 1);
1166                 } else {
1167                         th = (struct tcphdr *)(ip6 + 1);
1168                 }
1169                 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip6, th);
1170         } else
1171 #endif /* INET6 */
1172         {
1173                 ip = mtod(m, struct ip *);
1174 #ifdef TCPDEBUG
1175                 ipov = (struct ipovly *)ip;
1176 #endif
1177                 if (tp->t_port) {
1178                         udp = (struct udphdr *)((caddr_t)ip + sizeof(struct ip));
1179                         udp->uh_sport = htons(V_tcp_udp_tunneling_port);
1180                         udp->uh_dport = tp->t_port;
1181                         ulen = hdrlen + len - sizeof(struct ip);
1182                         udp->uh_ulen = htons(ulen);
1183                         th = (struct tcphdr *)(udp + 1);
1184                 } else
1185                         th = (struct tcphdr *)(ip + 1);
1186                 tcpip_fillheaders(tp->t_inpcb, tp->t_port, ip, th);
1187         }
1188
1189         /*
1190          * Fill in fields, remembering maximum advertised
1191          * window for use in delaying messages about window sizes.
1192          * If resending a FIN, be sure not to use a new sequence number.
1193          */
1194         if (flags & TH_FIN && tp->t_flags & TF_SENTFIN &&
1195             tp->snd_nxt == tp->snd_max)
1196                 tp->snd_nxt--;
1197         /*
1198          * If we are starting a connection, send ECN setup
1199          * SYN packet. If we are on a retransmit, we may
1200          * resend those bits a number of times as per
1201          * RFC 3168.
1202          */
1203         if (tp->t_state == TCPS_SYN_SENT && V_tcp_do_ecn == 1) {
1204                 if (tp->t_rxtshift >= 1) {
1205                         if (tp->t_rxtshift <= V_tcp_ecn_maxretries)
1206                                 flags |= TH_ECE|TH_CWR;
1207                 } else
1208                         flags |= TH_ECE|TH_CWR;
1209         }
1210         /* Handle parallel SYN for ECN */
1211         if ((tp->t_state == TCPS_SYN_RECEIVED) &&
1212             (tp->t_flags2 & TF2_ECN_SND_ECE)) {
1213                         flags |= TH_ECE;
1214                         tp->t_flags2 &= ~TF2_ECN_SND_ECE;
1215         }
1216
1217         if (TCPS_HAVEESTABLISHED(tp->t_state) &&
1218             (tp->t_flags2 & TF2_ECN_PERMIT)) {
1219                 /*
1220                  * If the peer has ECN, mark data packets with
1221                  * ECN capable transmission (ECT).
1222                  * Ignore pure ack packets, retransmissions and window probes.
1223                  */
1224                 if (len > 0 && SEQ_GEQ(tp->snd_nxt, tp->snd_max) &&
1225                     (sack_rxmit == 0) &&
1226                     !((tp->t_flags & TF_FORCEDATA) && len == 1 &&
1227                     SEQ_LT(tp->snd_una, tp->snd_max))) {
1228 #ifdef INET6
1229                         if (isipv6)
1230                                 ip6->ip6_flow |= htonl(IPTOS_ECN_ECT0 << 20);
1231                         else
1232 #endif
1233                                 ip->ip_tos |= IPTOS_ECN_ECT0;
1234                         TCPSTAT_INC(tcps_ecn_ect0);
1235                         /*
1236                          * Reply with proper ECN notifications.
1237                          * Only set CWR on new data segments.
1238                          */
1239                         if (tp->t_flags2 & TF2_ECN_SND_CWR) {
1240                                 flags |= TH_CWR;
1241                                 tp->t_flags2 &= ~TF2_ECN_SND_CWR;
1242                         }
1243                 }
1244                 if (tp->t_flags2 & TF2_ECN_SND_ECE)
1245                         flags |= TH_ECE;
1246         }
1247
1248         /*
1249          * If we are doing retransmissions, then snd_nxt will
1250          * not reflect the first unsent octet.  For ACK only
1251          * packets, we do not want the sequence number of the
1252          * retransmitted packet, we want the sequence number
1253          * of the next unsent octet.  So, if there is no data
1254          * (and no SYN or FIN), use snd_max instead of snd_nxt
1255          * when filling in ti_seq.  But if we are in persist
1256          * state, snd_max might reflect one byte beyond the
1257          * right edge of the window, so use snd_nxt in that
1258          * case, since we know we aren't doing a retransmission.
1259          * (retransmit and persist are mutually exclusive...)
1260          */
1261         if (sack_rxmit == 0) {
1262                 if (len || (flags & (TH_SYN|TH_FIN)) ||
1263                     tcp_timer_active(tp, TT_PERSIST))
1264                         th->th_seq = htonl(tp->snd_nxt);
1265                 else
1266                         th->th_seq = htonl(tp->snd_max);
1267         } else {
1268                 th->th_seq = htonl(p->rxmit);
1269                 p->rxmit += len;
1270                 tp->sackhint.sack_bytes_rexmit += len;
1271         }
1272         if (IN_RECOVERY(tp->t_flags)) {
1273                 /*
1274                  * Account all bytes transmitted while
1275                  * IN_RECOVERY, simplifying PRR and
1276                  * Lost Retransmit Detection
1277                  */
1278                 tp->sackhint.prr_out += len;
1279         }
1280         th->th_ack = htonl(tp->rcv_nxt);
1281         if (optlen) {
1282                 bcopy(opt, th + 1, optlen);
1283                 th->th_off = (sizeof (struct tcphdr) + optlen) >> 2;
1284         }
1285         th->th_flags = flags;
1286         /*
1287          * Calculate receive window.  Don't shrink window,
1288          * but avoid silly window syndrome.
1289          * If a RST segment is sent, advertise a window of zero.
1290          */
1291         if (flags & TH_RST) {
1292                 recwin = 0;
1293         } else {
1294                 if (recwin < (so->so_rcv.sb_hiwat / 4) &&
1295                     recwin < tp->t_maxseg)
1296                         recwin = 0;
1297                 if (SEQ_GT(tp->rcv_adv, tp->rcv_nxt) &&
1298                     recwin < (tp->rcv_adv - tp->rcv_nxt))
1299                         recwin = (tp->rcv_adv - tp->rcv_nxt);
1300         }
1301         /*
1302          * According to RFC1323 the window field in a SYN (i.e., a <SYN>
1303          * or <SYN,ACK>) segment itself is never scaled.  The <SYN,ACK>
1304          * case is handled in syncache.
1305          */
1306         if (flags & TH_SYN)
1307                 th->th_win = htons((u_short)
1308                                 (min(sbspace(&so->so_rcv), TCP_MAXWIN)));
1309         else {
1310                 /* Avoid shrinking window with window scaling. */
1311                 recwin = roundup2(recwin, 1 << tp->rcv_scale);
1312                 th->th_win = htons((u_short)(recwin >> tp->rcv_scale));
1313         }
1314
1315         /*
1316          * Adjust the RXWIN0SENT flag - indicate that we have advertised
1317          * a 0 window.  This may cause the remote transmitter to stall.  This
1318          * flag tells soreceive() to disable delayed acknowledgements when
1319          * draining the buffer.  This can occur if the receiver is attempting
1320          * to read more data than can be buffered prior to transmitting on
1321          * the connection.
1322          */
1323         if (th->th_win == 0) {
1324                 tp->t_sndzerowin++;
1325                 tp->t_flags |= TF_RXWIN0SENT;
1326         } else
1327                 tp->t_flags &= ~TF_RXWIN0SENT;
1328         if (SEQ_GT(tp->snd_up, tp->snd_nxt)) {
1329                 th->th_urp = htons((u_short)(tp->snd_up - tp->snd_nxt));
1330                 th->th_flags |= TH_URG;
1331         } else
1332                 /*
1333                  * If no urgent pointer to send, then we pull
1334                  * the urgent pointer to the left edge of the send window
1335                  * so that it doesn't drift into the send window on sequence
1336                  * number wraparound.
1337                  */
1338                 tp->snd_up = tp->snd_una;               /* drag it along */
1339
1340         /*
1341          * Put TCP length in extended header, and then
1342          * checksum extended header and data.
1343          */
1344         m->m_pkthdr.len = hdrlen + len; /* in6_cksum() need this */
1345
1346 #if defined(IPSEC_SUPPORT) || defined(TCP_SIGNATURE)
1347         if (to.to_flags & TOF_SIGNATURE) {
1348                 /*
1349                  * Calculate MD5 signature and put it into the place
1350                  * determined before.
1351                  * NOTE: since TCP options buffer doesn't point into
1352                  * mbuf's data, calculate offset and use it.
1353                  */
1354                 if (!TCPMD5_ENABLED() || (error = TCPMD5_OUTPUT(m, th,
1355                     (u_char *)(th + 1) + (to.to_signature - opt))) != 0) {
1356                         /*
1357                          * Do not send segment if the calculation of MD5
1358                          * digest has failed.
1359                          */
1360                         m_freem(m);
1361                         goto out;
1362                 }
1363         }
1364 #endif
1365 #ifdef INET6
1366         if (isipv6) {
1367                 /*
1368                  * There is no need to fill in ip6_plen right now.
1369                  * It will be filled later by ip6_output.
1370                  */
1371                 if (tp->t_port) {
1372                         m->m_pkthdr.csum_flags = CSUM_UDP_IPV6;
1373                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1374                         udp->uh_sum = in6_cksum_pseudo(ip6, ulen, IPPROTO_UDP, 0);
1375                         th->th_sum = htons(0);
1376                         UDPSTAT_INC(udps_opackets);
1377                 } else {
1378                         m->m_pkthdr.csum_flags = CSUM_TCP_IPV6;
1379                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1380                         th->th_sum = in6_cksum_pseudo(ip6,
1381                             sizeof(struct tcphdr) + optlen + len, IPPROTO_TCP,
1382                             0);
1383                 }
1384         }
1385 #endif
1386 #if defined(INET6) && defined(INET)
1387         else
1388 #endif
1389 #ifdef INET
1390         {
1391                 if (tp->t_port) {
1392                         m->m_pkthdr.csum_flags = CSUM_UDP;
1393                         m->m_pkthdr.csum_data = offsetof(struct udphdr, uh_sum);
1394                         udp->uh_sum = in_pseudo(ip->ip_src.s_addr,
1395                            ip->ip_dst.s_addr, htons(ulen + IPPROTO_UDP));
1396                         th->th_sum = htons(0);
1397                         UDPSTAT_INC(udps_opackets);
1398                 } else {
1399                         m->m_pkthdr.csum_flags = CSUM_TCP;
1400                         m->m_pkthdr.csum_data = offsetof(struct tcphdr, th_sum);
1401                         th->th_sum = in_pseudo(ip->ip_src.s_addr,
1402                             ip->ip_dst.s_addr, htons(sizeof(struct tcphdr) +
1403                             IPPROTO_TCP + len + optlen));
1404                 }
1405
1406                 /* IP version must be set here for ipv4/ipv6 checking later */
1407                 KASSERT(ip->ip_v == IPVERSION,
1408                     ("%s: IP version incorrect: %d", __func__, ip->ip_v));
1409         }
1410 #endif
1411
1412         /*
1413          * Enable TSO and specify the size of the segments.
1414          * The TCP pseudo header checksum is always provided.
1415          */
1416         if (tso) {
1417                 KASSERT(len > tp->t_maxseg - optlen,
1418                     ("%s: len <= tso_segsz", __func__));
1419                 m->m_pkthdr.csum_flags |= CSUM_TSO;
1420                 m->m_pkthdr.tso_segsz = tp->t_maxseg - optlen;
1421         }
1422
1423         KASSERT(len + hdrlen == m_length(m, NULL),
1424             ("%s: mbuf chain shorter than expected: %d + %u != %u",
1425             __func__, len, hdrlen, m_length(m, NULL)));
1426
1427 #ifdef TCP_HHOOK
1428         /* Run HHOOK_TCP_ESTABLISHED_OUT helper hooks. */
1429         hhook_run_tcp_est_out(tp, th, &to, len, tso);
1430 #endif
1431
1432 #ifdef TCPDEBUG
1433         /*
1434          * Trace.
1435          */
1436         if (so->so_options & SO_DEBUG) {
1437                 u_short save = 0;
1438 #ifdef INET6
1439                 if (!isipv6)
1440 #endif
1441                 {
1442                         save = ipov->ih_len;
1443                         ipov->ih_len = htons(m->m_pkthdr.len /* - hdrlen + (th->th_off << 2) */);
1444                 }
1445                 tcp_trace(TA_OUTPUT, tp->t_state, tp, mtod(m, void *), th, 0);
1446 #ifdef INET6
1447                 if (!isipv6)
1448 #endif
1449                 ipov->ih_len = save;
1450         }
1451 #endif /* TCPDEBUG */
1452         TCP_PROBE3(debug__output, tp, th, m);
1453
1454         /* We're getting ready to send; log now. */
1455         TCP_LOG_EVENT(tp, th, &so->so_rcv, &so->so_snd, TCP_LOG_OUT, ERRNO_UNK,
1456             len, NULL, false);
1457
1458         /*
1459          * Fill in IP length and desired time to live and
1460          * send to IP level.  There should be a better way
1461          * to handle ttl and tos; we could keep them in
1462          * the template, but need a way to checksum without them.
1463          */
1464         /*
1465          * m->m_pkthdr.len should have been set before checksum calculation,
1466          * because in6_cksum() need it.
1467          */
1468 #ifdef INET6
1469         if (isipv6) {
1470                 /*
1471                  * we separately set hoplimit for every segment, since the
1472                  * user might want to change the value via setsockopt.
1473                  * Also, desired default hop limit might be changed via
1474                  * Neighbor Discovery.
1475                  */
1476                 ip6->ip6_hlim = in6_selecthlim(tp->t_inpcb, NULL);
1477
1478                 /*
1479                  * Set the packet size here for the benefit of DTrace probes.
1480                  * ip6_output() will set it properly; it's supposed to include
1481                  * the option header lengths as well.
1482                  */
1483                 ip6->ip6_plen = htons(m->m_pkthdr.len - sizeof(*ip6));
1484
1485                 if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss)
1486                         tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1487                 else
1488                         tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1489
1490                 if (tp->t_state == TCPS_SYN_SENT)
1491                         TCP_PROBE5(connect__request, NULL, tp, ip6, tp, th);
1492
1493                 TCP_PROBE5(send, NULL, tp, ip6, tp, th);
1494
1495 #ifdef TCPPCAP
1496                 /* Save packet, if requested. */
1497                 tcp_pcap_add(th, m, &(tp->t_outpkts));
1498 #endif
1499
1500                 /* TODO: IPv6 IP6TOS_ECT bit on */
1501                 error = ip6_output(m, tp->t_inpcb->in6p_outputopts,
1502                     &tp->t_inpcb->inp_route6,
1503                     ((so->so_options & SO_DONTROUTE) ?  IP_ROUTETOIF : 0),
1504                     NULL, NULL, tp->t_inpcb);
1505
1506                 if (error == EMSGSIZE && tp->t_inpcb->inp_route6.ro_nh != NULL)
1507                         mtu = tp->t_inpcb->inp_route6.ro_nh->nh_mtu;
1508         }
1509 #endif /* INET6 */
1510 #if defined(INET) && defined(INET6)
1511         else
1512 #endif
1513 #ifdef INET
1514     {
1515         ip->ip_len = htons(m->m_pkthdr.len);
1516 #ifdef INET6
1517         if (tp->t_inpcb->inp_vflag & INP_IPV6PROTO)
1518                 ip->ip_ttl = in6_selecthlim(tp->t_inpcb, NULL);
1519 #endif /* INET6 */
1520         /*
1521          * If we do path MTU discovery, then we set DF on every packet.
1522          * This might not be the best thing to do according to RFC3390
1523          * Section 2. However the tcp hostcache migitates the problem
1524          * so it affects only the first tcp connection with a host.
1525          *
1526          * NB: Don't set DF on small MTU/MSS to have a safe fallback.
1527          */
1528         if (V_path_mtu_discovery && tp->t_maxseg > V_tcp_minmss) {
1529                 tp->t_flags2 |= TF2_PLPMTU_PMTUD;
1530                 if (tp->t_port == 0 || len < V_tcp_minmss) {
1531                         ip->ip_off |= htons(IP_DF);
1532                 }
1533         } else {
1534                 tp->t_flags2 &= ~TF2_PLPMTU_PMTUD;
1535         }
1536
1537         if (tp->t_state == TCPS_SYN_SENT)
1538                 TCP_PROBE5(connect__request, NULL, tp, ip, tp, th);
1539
1540         TCP_PROBE5(send, NULL, tp, ip, tp, th);
1541
1542 #ifdef TCPPCAP
1543         /* Save packet, if requested. */
1544         tcp_pcap_add(th, m, &(tp->t_outpkts));
1545 #endif
1546
1547         error = ip_output(m, tp->t_inpcb->inp_options, &tp->t_inpcb->inp_route,
1548             ((so->so_options & SO_DONTROUTE) ? IP_ROUTETOIF : 0), 0,
1549             tp->t_inpcb);
1550
1551         if (error == EMSGSIZE && tp->t_inpcb->inp_route.ro_nh != NULL)
1552                 mtu = tp->t_inpcb->inp_route.ro_nh->nh_mtu;
1553     }
1554 #endif /* INET */
1555
1556 out:
1557         if (error == 0)
1558                 tcp_account_for_send(tp, len, (tp->snd_nxt != tp->snd_max), 0, hw_tls);
1559         /*
1560          * In transmit state, time the transmission and arrange for
1561          * the retransmit.  In persist state, just set snd_max.
1562          */
1563         if ((tp->t_flags & TF_FORCEDATA) == 0 ||
1564             !tcp_timer_active(tp, TT_PERSIST)) {
1565                 tcp_seq startseq = tp->snd_nxt;
1566
1567                 /*
1568                  * Advance snd_nxt over sequence space of this segment.
1569                  */
1570                 if (flags & (TH_SYN|TH_FIN)) {
1571                         if (flags & TH_SYN)
1572                                 tp->snd_nxt++;
1573                         if (flags & TH_FIN) {
1574                                 tp->snd_nxt++;
1575                                 tp->t_flags |= TF_SENTFIN;
1576                         }
1577                 }
1578                 if (sack_rxmit)
1579                         goto timer;
1580                 tp->snd_nxt += len;
1581                 if (SEQ_GT(tp->snd_nxt, tp->snd_max)) {
1582                         tp->snd_max = tp->snd_nxt;
1583                         /*
1584                          * Time this transmission if not a retransmission and
1585                          * not currently timing anything.
1586                          */
1587                         tp->t_sndtime = ticks;
1588                         if (tp->t_rtttime == 0) {
1589                                 tp->t_rtttime = ticks;
1590                                 tp->t_rtseq = startseq;
1591                                 TCPSTAT_INC(tcps_segstimed);
1592                         }
1593 #ifdef STATS
1594                         if (!(tp->t_flags & TF_GPUTINPROG) && len) {
1595                                 tp->t_flags |= TF_GPUTINPROG;
1596                                 tp->gput_seq = startseq;
1597                                 tp->gput_ack = startseq +
1598                                     ulmin(sbavail(&so->so_snd) - off, sendwin);
1599                                 tp->gput_ts = tcp_ts_getticks();
1600                         }
1601 #endif /* STATS */
1602                 }
1603
1604                 /*
1605                  * Set retransmit timer if not currently set,
1606                  * and not doing a pure ack or a keep-alive probe.
1607                  * Initial value for retransmit timer is smoothed
1608                  * round-trip time + 2 * round-trip time variance.
1609                  * Initialize shift counter which is used for backoff
1610                  * of retransmit time.
1611                  */
1612 timer:
1613                 if (!tcp_timer_active(tp, TT_REXMT) &&
1614                     ((sack_rxmit && tp->snd_nxt != tp->snd_max) ||
1615                      (tp->snd_nxt != tp->snd_una))) {
1616                         if (tcp_timer_active(tp, TT_PERSIST)) {
1617                                 tcp_timer_activate(tp, TT_PERSIST, 0);
1618                                 tp->t_rxtshift = 0;
1619                         }
1620                         tcp_timer_activate(tp, TT_REXMT, tp->t_rxtcur);
1621                 } else if (len == 0 && sbavail(&so->so_snd) &&
1622                     !tcp_timer_active(tp, TT_REXMT) &&
1623                     !tcp_timer_active(tp, TT_PERSIST)) {
1624                         /*
1625                          * Avoid a situation where we do not set persist timer
1626                          * after a zero window condition. For example:
1627                          * 1) A -> B: packet with enough data to fill the window
1628                          * 2) B -> A: ACK for #1 + new data (0 window
1629                          *    advertisement)
1630                          * 3) A -> B: ACK for #2, 0 len packet
1631                          *
1632                          * In this case, A will not activate the persist timer,
1633                          * because it chose to send a packet. Unless tcp_output
1634                          * is called for some other reason (delayed ack timer,
1635                          * another input packet from B, socket syscall), A will
1636                          * not send zero window probes.
1637                          *
1638                          * So, if you send a 0-length packet, but there is data
1639                          * in the socket buffer, and neither the rexmt or
1640                          * persist timer is already set, then activate the
1641                          * persist timer.
1642                          */
1643                         tp->t_rxtshift = 0;
1644                         tcp_setpersist(tp);
1645                 }
1646         } else {
1647                 /*
1648                  * Persist case, update snd_max but since we are in
1649                  * persist mode (no window) we do not update snd_nxt.
1650                  */
1651                 int xlen = len;
1652                 if (flags & TH_SYN)
1653                         ++xlen;
1654                 if (flags & TH_FIN) {
1655                         ++xlen;
1656                         tp->t_flags |= TF_SENTFIN;
1657                 }
1658                 if (SEQ_GT(tp->snd_nxt + xlen, tp->snd_max))
1659                         tp->snd_max = tp->snd_nxt + xlen;
1660         }
1661         if ((error == 0) &&
1662             (TCPS_HAVEESTABLISHED(tp->t_state) &&
1663              (tp->t_flags & TF_SACK_PERMIT) &&
1664              tp->rcv_numsacks > 0)) {
1665                     /* Clean up any DSACK's sent */
1666                     tcp_clean_dsack_blocks(tp);
1667         }
1668         if (error) {
1669                 /* Record the error. */
1670                 TCP_LOG_EVENT(tp, NULL, &so->so_rcv, &so->so_snd, TCP_LOG_OUT,
1671                     error, 0, NULL, false);
1672
1673                 /*
1674                  * We know that the packet was lost, so back out the
1675                  * sequence number advance, if any.
1676                  *
1677                  * If the error is EPERM the packet got blocked by the
1678                  * local firewall.  Normally we should terminate the
1679                  * connection but the blocking may have been spurious
1680                  * due to a firewall reconfiguration cycle.  So we treat
1681                  * it like a packet loss and let the retransmit timer and
1682                  * timeouts do their work over time.
1683                  * XXX: It is a POLA question whether calling tcp_drop right
1684                  * away would be the really correct behavior instead.
1685                  */
1686                 if (((tp->t_flags & TF_FORCEDATA) == 0 ||
1687                     !tcp_timer_active(tp, TT_PERSIST)) &&
1688                     ((flags & TH_SYN) == 0) &&
1689                     (error != EPERM)) {
1690                         if (sack_rxmit) {
1691                                 p->rxmit -= len;
1692                                 tp->sackhint.sack_bytes_rexmit -= len;
1693                                 KASSERT(tp->sackhint.sack_bytes_rexmit >= 0,
1694                                     ("sackhint bytes rtx >= 0"));
1695                         } else
1696                                 tp->snd_nxt -= len;
1697                 }
1698                 SOCKBUF_UNLOCK_ASSERT(&so->so_snd);     /* Check gotos. */
1699                 switch (error) {
1700                 case EACCES:
1701                 case EPERM:
1702                         tp->t_softerror = error;
1703                         return (error);
1704                 case ENOBUFS:
1705                         TCP_XMIT_TIMER_ASSERT(tp, len, flags);
1706                         tp->snd_cwnd = tp->t_maxseg;
1707                         return (0);
1708                 case EMSGSIZE:
1709                         /*
1710                          * For some reason the interface we used initially
1711                          * to send segments changed to another or lowered
1712                          * its MTU.
1713                          * If TSO was active we either got an interface
1714                          * without TSO capabilits or TSO was turned off.
1715                          * If we obtained mtu from ip_output() then update
1716                          * it and try again.
1717                          */
1718                         if (tso)
1719                                 tp->t_flags &= ~TF_TSO;
1720                         if (mtu != 0) {
1721                                 tcp_mss_update(tp, -1, mtu, NULL, NULL);
1722                                 goto again;
1723                         }
1724                         return (error);
1725                 case EHOSTDOWN:
1726                 case EHOSTUNREACH:
1727                 case ENETDOWN:
1728                 case ENETUNREACH:
1729                         if (TCPS_HAVERCVDSYN(tp->t_state)) {
1730                                 tp->t_softerror = error;
1731                                 return (0);
1732                         }
1733                         /* FALLTHROUGH */
1734                 default:
1735                         return (error);
1736                 }
1737         }
1738         TCPSTAT_INC(tcps_sndtotal);
1739
1740         /*
1741          * Data sent (as far as we can tell).
1742          * If this advertises a larger window than any other segment,
1743          * then remember the size of the advertised window.
1744          * Any pending ACK has now been sent.
1745          */
1746         if (SEQ_GT(tp->rcv_nxt + recwin, tp->rcv_adv))
1747                 tp->rcv_adv = tp->rcv_nxt + recwin;
1748         tp->last_ack_sent = tp->rcv_nxt;
1749         tp->t_flags &= ~(TF_ACKNOW | TF_DELACK);
1750         if (tcp_timer_active(tp, TT_DELACK))
1751                 tcp_timer_activate(tp, TT_DELACK, 0);
1752 #if 0
1753         /*
1754          * This completely breaks TCP if newreno is turned on.  What happens
1755          * is that if delayed-acks are turned on on the receiver, this code
1756          * on the transmitter effectively destroys the TCP window, forcing
1757          * it to four packets (1.5Kx4 = 6K window).
1758          */
1759         if (sendalot && --maxburst)
1760                 goto again;
1761 #endif
1762         if (sendalot)
1763                 goto again;
1764         return (0);
1765 }
1766
1767 void
1768 tcp_setpersist(struct tcpcb *tp)
1769 {
1770         int t = ((tp->t_srtt >> 2) + tp->t_rttvar) >> 1;
1771         int tt;
1772
1773         tp->t_flags &= ~TF_PREVVALID;
1774         if (tcp_timer_active(tp, TT_REXMT))
1775                 panic("tcp_setpersist: retransmit pending");
1776         /*
1777          * Start/restart persistence timer.
1778          */
1779         TCPT_RANGESET(tt, t * tcp_backoff[tp->t_rxtshift],
1780                       tcp_persmin, tcp_persmax);
1781         tcp_timer_activate(tp, TT_PERSIST, tt);
1782         if (tp->t_rxtshift < TCP_MAXRXTSHIFT)
1783                 tp->t_rxtshift++;
1784 }
1785
1786 /*
1787  * Insert TCP options according to the supplied parameters to the place
1788  * optp in a consistent way.  Can handle unaligned destinations.
1789  *
1790  * The order of the option processing is crucial for optimal packing and
1791  * alignment for the scarce option space.
1792  *
1793  * The optimal order for a SYN/SYN-ACK segment is:
1794  *   MSS (4) + NOP (1) + Window scale (3) + SACK permitted (2) +
1795  *   Timestamp (10) + Signature (18) = 38 bytes out of a maximum of 40.
1796  *
1797  * The SACK options should be last.  SACK blocks consume 8*n+2 bytes.
1798  * So a full size SACK blocks option is 34 bytes (with 4 SACK blocks).
1799  * At minimum we need 10 bytes (to generate 1 SACK block).  If both
1800  * TCP Timestamps (12 bytes) and TCP Signatures (18 bytes) are present,
1801  * we only have 10 bytes for SACK options (40 - (12 + 18)).
1802  */
1803 int
1804 tcp_addoptions(struct tcpopt *to, u_char *optp)
1805 {
1806         u_int32_t mask, optlen = 0;
1807
1808         for (mask = 1; mask < TOF_MAXOPT; mask <<= 1) {
1809                 if ((to->to_flags & mask) != mask)
1810                         continue;
1811                 if (optlen == TCP_MAXOLEN)
1812                         break;
1813                 switch (to->to_flags & mask) {
1814                 case TOF_MSS:
1815                         while (optlen % 4) {
1816                                 optlen += TCPOLEN_NOP;
1817                                 *optp++ = TCPOPT_NOP;
1818                         }
1819                         if (TCP_MAXOLEN - optlen < TCPOLEN_MAXSEG)
1820                                 continue;
1821                         optlen += TCPOLEN_MAXSEG;
1822                         *optp++ = TCPOPT_MAXSEG;
1823                         *optp++ = TCPOLEN_MAXSEG;
1824                         to->to_mss = htons(to->to_mss);
1825                         bcopy((u_char *)&to->to_mss, optp, sizeof(to->to_mss));
1826                         optp += sizeof(to->to_mss);
1827                         break;
1828                 case TOF_SCALE:
1829                         while (!optlen || optlen % 2 != 1) {
1830                                 optlen += TCPOLEN_NOP;
1831                                 *optp++ = TCPOPT_NOP;
1832                         }
1833                         if (TCP_MAXOLEN - optlen < TCPOLEN_WINDOW)
1834                                 continue;
1835                         optlen += TCPOLEN_WINDOW;
1836                         *optp++ = TCPOPT_WINDOW;
1837                         *optp++ = TCPOLEN_WINDOW;
1838                         *optp++ = to->to_wscale;
1839                         break;
1840                 case TOF_SACKPERM:
1841                         while (optlen % 2) {
1842                                 optlen += TCPOLEN_NOP;
1843                                 *optp++ = TCPOPT_NOP;
1844                         }
1845                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACK_PERMITTED)
1846                                 continue;
1847                         optlen += TCPOLEN_SACK_PERMITTED;
1848                         *optp++ = TCPOPT_SACK_PERMITTED;
1849                         *optp++ = TCPOLEN_SACK_PERMITTED;
1850                         break;
1851                 case TOF_TS:
1852                         while (!optlen || optlen % 4 != 2) {
1853                                 optlen += TCPOLEN_NOP;
1854                                 *optp++ = TCPOPT_NOP;
1855                         }
1856                         if (TCP_MAXOLEN - optlen < TCPOLEN_TIMESTAMP)
1857                                 continue;
1858                         optlen += TCPOLEN_TIMESTAMP;
1859                         *optp++ = TCPOPT_TIMESTAMP;
1860                         *optp++ = TCPOLEN_TIMESTAMP;
1861                         to->to_tsval = htonl(to->to_tsval);
1862                         to->to_tsecr = htonl(to->to_tsecr);
1863                         bcopy((u_char *)&to->to_tsval, optp, sizeof(to->to_tsval));
1864                         optp += sizeof(to->to_tsval);
1865                         bcopy((u_char *)&to->to_tsecr, optp, sizeof(to->to_tsecr));
1866                         optp += sizeof(to->to_tsecr);
1867                         break;
1868                 case TOF_SIGNATURE:
1869                         {
1870                         int siglen = TCPOLEN_SIGNATURE - 2;
1871
1872                         while (!optlen || optlen % 4 != 2) {
1873                                 optlen += TCPOLEN_NOP;
1874                                 *optp++ = TCPOPT_NOP;
1875                         }
1876                         if (TCP_MAXOLEN - optlen < TCPOLEN_SIGNATURE) {
1877                                 to->to_flags &= ~TOF_SIGNATURE;
1878                                 continue;
1879                         }
1880                         optlen += TCPOLEN_SIGNATURE;
1881                         *optp++ = TCPOPT_SIGNATURE;
1882                         *optp++ = TCPOLEN_SIGNATURE;
1883                         to->to_signature = optp;
1884                         while (siglen--)
1885                                  *optp++ = 0;
1886                         break;
1887                         }
1888                 case TOF_SACK:
1889                         {
1890                         int sackblks = 0;
1891                         struct sackblk *sack = (struct sackblk *)to->to_sacks;
1892                         tcp_seq sack_seq;
1893
1894                         while (!optlen || optlen % 4 != 2) {
1895                                 optlen += TCPOLEN_NOP;
1896                                 *optp++ = TCPOPT_NOP;
1897                         }
1898                         if (TCP_MAXOLEN - optlen < TCPOLEN_SACKHDR + TCPOLEN_SACK)
1899                                 continue;
1900                         optlen += TCPOLEN_SACKHDR;
1901                         *optp++ = TCPOPT_SACK;
1902                         sackblks = min(to->to_nsacks,
1903                                         (TCP_MAXOLEN - optlen) / TCPOLEN_SACK);
1904                         *optp++ = TCPOLEN_SACKHDR + sackblks * TCPOLEN_SACK;
1905                         while (sackblks--) {
1906                                 sack_seq = htonl(sack->start);
1907                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1908                                 optp += sizeof(sack_seq);
1909                                 sack_seq = htonl(sack->end);
1910                                 bcopy((u_char *)&sack_seq, optp, sizeof(sack_seq));
1911                                 optp += sizeof(sack_seq);
1912                                 optlen += TCPOLEN_SACK;
1913                                 sack++;
1914                         }
1915                         TCPSTAT_INC(tcps_sack_send_blocks);
1916                         break;
1917                         }
1918                 case TOF_FASTOPEN:
1919                         {
1920                         int total_len;
1921
1922                         /* XXX is there any point to aligning this option? */
1923                         total_len = TCPOLEN_FAST_OPEN_EMPTY + to->to_tfo_len;
1924                         if (TCP_MAXOLEN - optlen < total_len) {
1925                                 to->to_flags &= ~TOF_FASTOPEN;
1926                                 continue;
1927                         }
1928                         *optp++ = TCPOPT_FAST_OPEN;
1929                         *optp++ = total_len;
1930                         if (to->to_tfo_len > 0) {
1931                                 bcopy(to->to_tfo_cookie, optp, to->to_tfo_len);
1932                                 optp += to->to_tfo_len;
1933                         }
1934                         optlen += total_len;
1935                         break;
1936                         }
1937                 default:
1938                         panic("%s: unknown TCP option type", __func__);
1939                         break;
1940                 }
1941         }
1942
1943         /* Terminate and pad TCP options to a 4 byte boundary. */
1944         if (optlen % 4) {
1945                 optlen += TCPOLEN_EOL;
1946                 *optp++ = TCPOPT_EOL;
1947         }
1948         /*
1949          * According to RFC 793 (STD0007):
1950          *   "The content of the header beyond the End-of-Option option
1951          *    must be header padding (i.e., zero)."
1952          *   and later: "The padding is composed of zeros."
1953          */
1954         while (optlen % 4) {
1955                 optlen += TCPOLEN_PAD;
1956                 *optp++ = TCPOPT_PAD;
1957         }
1958
1959         KASSERT(optlen <= TCP_MAXOLEN, ("%s: TCP options too long", __func__));
1960         return (optlen);
1961 }
1962
1963 /*
1964  * This is a copy of m_copym(), taking the TSO segment size/limit
1965  * constraints into account, and advancing the sndptr as it goes.
1966  */
1967 struct mbuf *
1968 tcp_m_copym(struct mbuf *m, int32_t off0, int32_t *plen,
1969     int32_t seglimit, int32_t segsize, struct sockbuf *sb, bool hw_tls)
1970 {
1971 #ifdef KERN_TLS
1972         struct ktls_session *tls, *ntls;
1973         struct mbuf *start;
1974 #endif
1975         struct mbuf *n, **np;
1976         struct mbuf *top;
1977         int32_t off = off0;
1978         int32_t len = *plen;
1979         int32_t fragsize;
1980         int32_t len_cp = 0;
1981         int32_t *pkthdrlen;
1982         uint32_t mlen, frags;
1983         bool copyhdr;
1984
1985         KASSERT(off >= 0, ("tcp_m_copym, negative off %d", off));
1986         KASSERT(len >= 0, ("tcp_m_copym, negative len %d", len));
1987         if (off == 0 && m->m_flags & M_PKTHDR)
1988                 copyhdr = true;
1989         else
1990                 copyhdr = false;
1991         while (off > 0) {
1992                 KASSERT(m != NULL, ("tcp_m_copym, offset > size of mbuf chain"));
1993                 if (off < m->m_len)
1994                         break;
1995                 off -= m->m_len;
1996                 if ((sb) && (m == sb->sb_sndptr)) {
1997                         sb->sb_sndptroff += m->m_len;
1998                         sb->sb_sndptr = m->m_next;
1999                 }
2000                 m = m->m_next;
2001         }
2002         np = &top;
2003         top = NULL;
2004         pkthdrlen = NULL;
2005 #ifdef KERN_TLS
2006         if (hw_tls && (m->m_flags & M_EXTPG))
2007                 tls = m->m_epg_tls;
2008         else
2009                 tls = NULL;
2010         start = m;
2011 #endif
2012         while (len > 0) {
2013                 if (m == NULL) {
2014                         KASSERT(len == M_COPYALL,
2015                             ("tcp_m_copym, length > size of mbuf chain"));
2016                         *plen = len_cp;
2017                         if (pkthdrlen != NULL)
2018                                 *pkthdrlen = len_cp;
2019                         break;
2020                 }
2021 #ifdef KERN_TLS
2022                 if (hw_tls) {
2023                         if (m->m_flags & M_EXTPG)
2024                                 ntls = m->m_epg_tls;
2025                         else
2026                                 ntls = NULL;
2027
2028                         /*
2029                          * Avoid mixing TLS records with handshake
2030                          * data or TLS records from different
2031                          * sessions.
2032                          */
2033                         if (tls != ntls) {
2034                                 MPASS(m != start);
2035                                 *plen = len_cp;
2036                                 if (pkthdrlen != NULL)
2037                                         *pkthdrlen = len_cp;
2038                                 break;
2039                         }
2040                 }
2041 #endif
2042                 mlen = min(len, m->m_len - off);
2043                 if (seglimit) {
2044                         /*
2045                          * For M_EXTPG mbufs, add 3 segments
2046                          * + 1 in case we are crossing page boundaries
2047                          * + 2 in case the TLS hdr/trailer are used
2048                          * It is cheaper to just add the segments
2049                          * than it is to take the cache miss to look
2050                          * at the mbuf ext_pgs state in detail.
2051                          */
2052                         if (m->m_flags & M_EXTPG) {
2053                                 fragsize = min(segsize, PAGE_SIZE);
2054                                 frags = 3;
2055                         } else {
2056                                 fragsize = segsize;
2057                                 frags = 0;
2058                         }
2059
2060                         /* Break if we really can't fit anymore. */
2061                         if ((frags + 1) >= seglimit) {
2062                                 *plen = len_cp;
2063                                 if (pkthdrlen != NULL)
2064                                         *pkthdrlen = len_cp;
2065                                 break;
2066                         }
2067
2068                         /*
2069                          * Reduce size if you can't copy the whole
2070                          * mbuf. If we can't copy the whole mbuf, also
2071                          * adjust len so the loop will end after this
2072                          * mbuf.
2073                          */
2074                         if ((frags + howmany(mlen, fragsize)) >= seglimit) {
2075                                 mlen = (seglimit - frags - 1) * fragsize;
2076                                 len = mlen;
2077                                 *plen = len_cp + len;
2078                                 if (pkthdrlen != NULL)
2079                                         *pkthdrlen = *plen;
2080                         }
2081                         frags += howmany(mlen, fragsize);
2082                         if (frags == 0)
2083                                 frags++;
2084                         seglimit -= frags;
2085                         KASSERT(seglimit > 0,
2086                             ("%s: seglimit went too low", __func__));
2087                 }
2088                 if (copyhdr)
2089                         n = m_gethdr(M_NOWAIT, m->m_type);
2090                 else
2091                         n = m_get(M_NOWAIT, m->m_type);
2092                 *np = n;
2093                 if (n == NULL)
2094                         goto nospace;
2095                 if (copyhdr) {
2096                         if (!m_dup_pkthdr(n, m, M_NOWAIT))
2097                                 goto nospace;
2098                         if (len == M_COPYALL)
2099                                 n->m_pkthdr.len -= off0;
2100                         else
2101                                 n->m_pkthdr.len = len;
2102                         pkthdrlen = &n->m_pkthdr.len;
2103                         copyhdr = false;
2104                 }
2105                 n->m_len = mlen;
2106                 len_cp += n->m_len;
2107                 if (m->m_flags & (M_EXT|M_EXTPG)) {
2108                         n->m_data = m->m_data + off;
2109                         mb_dupcl(n, m);
2110                 } else
2111                         bcopy(mtod(m, caddr_t)+off, mtod(n, caddr_t),
2112                             (u_int)n->m_len);
2113
2114                 if (sb && (sb->sb_sndptr == m) &&
2115                     ((n->m_len + off) >= m->m_len) && m->m_next) {
2116                         sb->sb_sndptroff += m->m_len;
2117                         sb->sb_sndptr = m->m_next;
2118                 }
2119                 off = 0;
2120                 if (len != M_COPYALL) {
2121                         len -= n->m_len;
2122                 }
2123                 m = m->m_next;
2124                 np = &n->m_next;
2125         }
2126         return (top);
2127 nospace:
2128         m_freem(top);
2129         return (NULL);
2130 }
2131
2132 void
2133 tcp_sndbuf_autoscale(struct tcpcb *tp, struct socket *so, uint32_t sendwin)
2134 {
2135
2136         /*
2137          * Automatic sizing of send socket buffer.  Often the send buffer
2138          * size is not optimally adjusted to the actual network conditions
2139          * at hand (delay bandwidth product).  Setting the buffer size too
2140          * small limits throughput on links with high bandwidth and high
2141          * delay (eg. trans-continental/oceanic links).  Setting the
2142          * buffer size too big consumes too much real kernel memory,
2143          * especially with many connections on busy servers.
2144          *
2145          * The criteria to step up the send buffer one notch are:
2146          *  1. receive window of remote host is larger than send buffer
2147          *     (with a fudge factor of 5/4th);
2148          *  2. send buffer is filled to 7/8th with data (so we actually
2149          *     have data to make use of it);
2150          *  3. send buffer fill has not hit maximal automatic size;
2151          *  4. our send window (slow start and cogestion controlled) is
2152          *     larger than sent but unacknowledged data in send buffer.
2153          *
2154          * The remote host receive window scaling factor may limit the
2155          * growing of the send buffer before it reaches its allowed
2156          * maximum.
2157          *
2158          * It scales directly with slow start or congestion window
2159          * and does at most one step per received ACK.  This fast
2160          * scaling has the drawback of growing the send buffer beyond
2161          * what is strictly necessary to make full use of a given
2162          * delay*bandwidth product.  However testing has shown this not
2163          * to be much of an problem.  At worst we are trading wasting
2164          * of available bandwidth (the non-use of it) for wasting some
2165          * socket buffer memory.
2166          *
2167          * TODO: Shrink send buffer during idle periods together
2168          * with congestion window.  Requires another timer.  Has to
2169          * wait for upcoming tcp timer rewrite.
2170          *
2171          * XXXGL: should there be used sbused() or sbavail()?
2172          */
2173         if (V_tcp_do_autosndbuf && so->so_snd.sb_flags & SB_AUTOSIZE) {
2174                 int lowat;
2175
2176                 lowat = V_tcp_sendbuf_auto_lowat ? so->so_snd.sb_lowat : 0;
2177                 if ((tp->snd_wnd / 4 * 5) >= so->so_snd.sb_hiwat - lowat &&
2178                     sbused(&so->so_snd) >=
2179                     (so->so_snd.sb_hiwat / 8 * 7) - lowat &&
2180                     sbused(&so->so_snd) < V_tcp_autosndbuf_max &&
2181                     sendwin >= (sbused(&so->so_snd) -
2182                     (tp->snd_nxt - tp->snd_una))) {
2183                         if (!sbreserve_locked(&so->so_snd,
2184                             min(so->so_snd.sb_hiwat + V_tcp_autosndbuf_inc,
2185                              V_tcp_autosndbuf_max), so, curthread))
2186                                 so->so_snd.sb_flags &= ~SB_AUTOSIZE;
2187                 }
2188         }
2189 }