1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  */
  21 
  22 /*
  23  * Copyright (c) 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright (c) 2014, 2017 by Delphix. All rights reserved.
  25  */
  26 
  27 /* This file contains all TCP output processing functions. */
  28 
  29 #include <sys/types.h>
  30 #include <sys/stream.h>
  31 #include <sys/strsun.h>
  32 #include <sys/strsubr.h>
  33 #include <sys/stropts.h>
  34 #include <sys/strlog.h>
  35 #define _SUN_TPI_VERSION 2
  36 #include <sys/tihdr.h>
  37 #include <sys/suntpi.h>
  38 #include <sys/xti_inet.h>
  39 #include <sys/timod.h>
  40 #include <sys/pattr.h>
  41 #include <sys/squeue_impl.h>
  42 #include <sys/squeue.h>
  43 #include <sys/sockio.h>
  44 #include <sys/tsol/tnet.h>
  45 
  46 #include <inet/common.h>
  47 #include <inet/ip.h>
  48 #include <inet/tcp.h>
  49 #include <inet/tcp_impl.h>
  50 #include <inet/snmpcom.h>
  51 #include <inet/proto_set.h>
  52 #include <inet/ipsec_impl.h>
  53 #include <inet/ip_ndp.h>
  54 
  55 static mblk_t   *tcp_get_seg_mp(tcp_t *, uint32_t, int32_t *);
  56 static void     tcp_wput_cmdblk(queue_t *, mblk_t *);
  57 static void     tcp_wput_flush(tcp_t *, mblk_t *);
  58 static void     tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp);
  59 static int      tcp_xmit_end(tcp_t *);
  60 static int      tcp_send(tcp_t *, const int, const int, const int,
  61                     const int, int *, uint_t *, int *, mblk_t **, mblk_t *);
  62 static void     tcp_xmit_early_reset(char *, mblk_t *, uint32_t, uint32_t,
  63                     int, ip_recv_attr_t *, ip_stack_t *, conn_t *);
  64 static boolean_t        tcp_send_rst_chk(tcp_stack_t *);
  65 static void     tcp_process_shrunk_swnd(tcp_t *, uint32_t);
  66 static void     tcp_fill_header(tcp_t *, uchar_t *, int);
  67 
  68 /*
  69  * Functions called directly via squeue having a prototype of edesc_t.
  70  */
  71 static void     tcp_wput_nondata(void *, mblk_t *, void *, ip_recv_attr_t *);
  72 static void     tcp_wput_ioctl(void *, mblk_t *, void *, ip_recv_attr_t *);
  73 static void     tcp_wput_proto(void *, mblk_t *, void *, ip_recv_attr_t *);
  74 
  75 /*
  76  * This controls how tiny a write must be before we try to copy it
  77  * into the mblk on the tail of the transmit queue.  Not much
  78  * speedup is observed for values larger than sixteen.  Zero will
  79  * disable the optimisation.
  80  */
  81 static int tcp_tx_pull_len = 16;
  82 
  83 static void
  84 cc_after_idle(tcp_t *tcp)
  85 {
  86         uint32_t old_cwnd = tcp->tcp_cwnd;
  87 
  88         if (CC_ALGO(tcp)->after_idle != NULL)
  89                 CC_ALGO(tcp)->after_idle(&tcp->tcp_ccv);
  90 
  91         DTRACE_PROBE3(cwnd__cc__after__idle, tcp_t *, tcp, uint32_t, old_cwnd,
  92             uint32_t, tcp->tcp_cwnd);
  93 }
  94 
  95 void
  96 tcp_wput(queue_t *q, mblk_t *mp)
  97 {
  98         conn_t  *connp = Q_TO_CONN(q);
  99         tcp_t   *tcp;
 100         void (*output_proc)();
 101         t_scalar_t type;
 102         uchar_t *rptr;
 103         struct iocblk   *iocp;
 104         size_t size;
 105 
 106         ASSERT(connp->conn_ref >= 2);
 107 
 108         switch (DB_TYPE(mp)) {
 109         case M_DATA:
 110                 tcp = connp->conn_tcp;
 111                 ASSERT(tcp != NULL);
 112 
 113                 size = msgdsize(mp);
 114 
 115                 mutex_enter(&tcp->tcp_non_sq_lock);
 116                 tcp->tcp_squeue_bytes += size;
 117                 if (TCP_UNSENT_BYTES(tcp) > connp->conn_sndbuf) {
 118                         tcp_setqfull(tcp);
 119                 }
 120                 mutex_exit(&tcp->tcp_non_sq_lock);
 121 
 122                 CONN_INC_REF(connp);
 123                 SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_output, connp,
 124                     NULL, tcp_squeue_flag, SQTAG_TCP_OUTPUT);
 125                 return;
 126 
 127         case M_CMD:
 128                 tcp_wput_cmdblk(q, mp);
 129                 return;
 130 
 131         case M_PROTO:
 132         case M_PCPROTO:
 133                 /*
 134                  * if it is a snmp message, don't get behind the squeue
 135                  */
 136                 tcp = connp->conn_tcp;
 137                 rptr = mp->b_rptr;
 138                 if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
 139                         type = ((union T_primitives *)rptr)->type;
 140                 } else {
 141                         if (connp->conn_debug) {
 142                                 (void) strlog(TCP_MOD_ID, 0, 1,
 143                                     SL_ERROR|SL_TRACE,
 144                                     "tcp_wput_proto, dropping one...");
 145                         }
 146                         freemsg(mp);
 147                         return;
 148                 }
 149                 if (type == T_SVR4_OPTMGMT_REQ) {
 150                         /*
 151                          * All Solaris components should pass a db_credp
 152                          * for this TPI message, hence we ASSERT.
 153                          * But in case there is some other M_PROTO that looks
 154                          * like a TPI message sent by some other kernel
 155                          * component, we check and return an error.
 156                          */
 157                         cred_t  *cr = msg_getcred(mp, NULL);
 158 
 159                         ASSERT(cr != NULL);
 160                         if (cr == NULL) {
 161                                 tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
 162                                 return;
 163                         }
 164                         if (snmpcom_req(q, mp, tcp_snmp_set, ip_snmp_get,
 165                             cr)) {
 166                                 /*
 167                                  * This was a SNMP request
 168                                  */
 169                                 return;
 170                         } else {
 171                                 output_proc = tcp_wput_proto;
 172                         }
 173                 } else {
 174                         output_proc = tcp_wput_proto;
 175                 }
 176                 break;
 177         case M_IOCTL:
 178                 /*
 179                  * Most ioctls can be processed right away without going via
 180                  * squeues - process them right here. Those that do require
 181                  * squeue (currently _SIOCSOCKFALLBACK)
 182                  * are processed by tcp_wput_ioctl().
 183                  */
 184                 iocp = (struct iocblk *)mp->b_rptr;
 185                 tcp = connp->conn_tcp;
 186 
 187                 switch (iocp->ioc_cmd) {
 188                 case TCP_IOC_ABORT_CONN:
 189                         tcp_ioctl_abort_conn(q, mp);
 190                         return;
 191                 case TI_GETPEERNAME:
 192                 case TI_GETMYNAME:
 193                         mi_copyin(q, mp, NULL,
 194                             SIZEOF_STRUCT(strbuf, iocp->ioc_flag));
 195                         return;
 196 
 197                 default:
 198                         output_proc = tcp_wput_ioctl;
 199                         break;
 200                 }
 201                 break;
 202         default:
 203                 output_proc = tcp_wput_nondata;
 204                 break;
 205         }
 206 
 207         CONN_INC_REF(connp);
 208         SQUEUE_ENTER_ONE(connp->conn_sqp, mp, output_proc, connp,
 209             NULL, tcp_squeue_flag, SQTAG_TCP_WPUT_OTHER);
 210 }
 211 
 212 /*
 213  * The TCP normal data output path.
 214  * NOTE: the logic of the fast path is duplicated from this function.
 215  */
 216 void
 217 tcp_wput_data(tcp_t *tcp, mblk_t *mp, boolean_t urgent)
 218 {
 219         int             len;
 220         mblk_t          *local_time;
 221         mblk_t          *mp1;
 222         uint32_t        snxt;
 223         int             tail_unsent;
 224         int             tcpstate;
 225         int             usable = 0;
 226         mblk_t          *xmit_tail;
 227         int32_t         mss;
 228         int32_t         num_sack_blk = 0;
 229         int32_t         total_hdr_len;
 230         int32_t         tcp_hdr_len;
 231         int             rc;
 232         conn_t          *connp = tcp->tcp_connp;
 233         clock_t         now = LBOLT_FASTPATH;
 234 
 235         tcpstate = tcp->tcp_state;
 236         if (mp == NULL) {
 237                 /*
 238                  * tcp_wput_data() with NULL mp should only be called when
 239                  * there is unsent data.
 240                  */
 241                 ASSERT(tcp->tcp_unsent > 0);
 242                 /* Really tacky... but we need this for detached closes. */
 243                 len = tcp->tcp_unsent;
 244                 goto data_null;
 245         }
 246 
 247         ASSERT(mp->b_datap->db_type == M_DATA);
 248         /*
 249          * Don't allow data after T_ORDREL_REQ or T_DISCON_REQ,
 250          * or before a connection attempt has begun.
 251          */
 252         if (tcpstate < TCPS_SYN_SENT || tcpstate > TCPS_CLOSE_WAIT ||
 253             (tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
 254                 if ((tcp->tcp_valid_bits & TCP_FSS_VALID) != 0) {
 255 #ifdef DEBUG
 256                         cmn_err(CE_WARN,
 257                             "tcp_wput_data: data after ordrel, %s",
 258                             tcp_display(tcp, NULL,
 259                             DISP_ADDR_AND_PORT));
 260 #else
 261                         if (connp->conn_debug) {
 262                                 (void) strlog(TCP_MOD_ID, 0, 1,
 263                                     SL_TRACE|SL_ERROR,
 264                                     "tcp_wput_data: data after ordrel, %s\n",
 265                                     tcp_display(tcp, NULL,
 266                                     DISP_ADDR_AND_PORT));
 267                         }
 268 #endif /* DEBUG */
 269                 }
 270                 if (tcp->tcp_snd_zcopy_aware &&
 271                     (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
 272                         tcp_zcopy_notify(tcp);
 273                 freemsg(mp);
 274                 mutex_enter(&tcp->tcp_non_sq_lock);
 275                 if (tcp->tcp_flow_stopped &&
 276                     TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
 277                         tcp_clrqfull(tcp);
 278                 }
 279                 mutex_exit(&tcp->tcp_non_sq_lock);
 280                 return;
 281         }
 282 
 283         /* Strip empties */
 284         for (;;) {
 285                 ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <=
 286                     (uintptr_t)INT_MAX);
 287                 len = (int)(mp->b_wptr - mp->b_rptr);
 288                 if (len > 0)
 289                         break;
 290                 mp1 = mp;
 291                 mp = mp->b_cont;
 292                 freeb(mp1);
 293                 if (mp == NULL) {
 294                         return;
 295                 }
 296         }
 297 
 298         /* If we are the first on the list ... */
 299         if (tcp->tcp_xmit_head == NULL) {
 300                 tcp->tcp_xmit_head = mp;
 301                 tcp->tcp_xmit_tail = mp;
 302                 tcp->tcp_xmit_tail_unsent = len;
 303         } else {
 304                 /* If tiny tx and room in txq tail, pullup to save mblks. */
 305                 struct datab *dp;
 306 
 307                 mp1 = tcp->tcp_xmit_last;
 308                 if (len < tcp_tx_pull_len &&
 309                     (dp = mp1->b_datap)->db_ref == 1 &&
 310                     dp->db_lim - mp1->b_wptr >= len) {
 311                         ASSERT(len > 0);
 312                         ASSERT(!mp1->b_cont);
 313                         if (len == 1) {
 314                                 *mp1->b_wptr++ = *mp->b_rptr;
 315                         } else {
 316                                 bcopy(mp->b_rptr, mp1->b_wptr, len);
 317                                 mp1->b_wptr += len;
 318                         }
 319                         if (mp1 == tcp->tcp_xmit_tail)
 320                                 tcp->tcp_xmit_tail_unsent += len;
 321                         mp1->b_cont = mp->b_cont;
 322                         if (tcp->tcp_snd_zcopy_aware &&
 323                             (mp->b_datap->db_struioflag & STRUIO_ZCNOTIFY))
 324                                 mp1->b_datap->db_struioflag |= STRUIO_ZCNOTIFY;
 325                         freeb(mp);
 326                         mp = mp1;
 327                 } else {
 328                         tcp->tcp_xmit_last->b_cont = mp;
 329                 }
 330                 len += tcp->tcp_unsent;
 331         }
 332 
 333         /* Tack on however many more positive length mblks we have */
 334         if ((mp1 = mp->b_cont) != NULL) {
 335                 do {
 336                         int tlen;
 337                         ASSERT((uintptr_t)(mp1->b_wptr - mp1->b_rptr) <=
 338                             (uintptr_t)INT_MAX);
 339                         tlen = (int)(mp1->b_wptr - mp1->b_rptr);
 340                         if (tlen <= 0) {
 341                                 mp->b_cont = mp1->b_cont;
 342                                 freeb(mp1);
 343                         } else {
 344                                 len += tlen;
 345                                 mp = mp1;
 346                         }
 347                 } while ((mp1 = mp->b_cont) != NULL);
 348         }
 349         tcp->tcp_xmit_last = mp;
 350         tcp->tcp_unsent = len;
 351 
 352         if (urgent)
 353                 usable = 1;
 354 
 355 data_null:
 356         snxt = tcp->tcp_snxt;
 357         xmit_tail = tcp->tcp_xmit_tail;
 358         tail_unsent = tcp->tcp_xmit_tail_unsent;
 359 
 360         /*
 361          * Note that tcp_mss has been adjusted to take into account the
 362          * timestamp option if applicable.  Because SACK options do not
 363          * appear in every TCP segments and they are of variable lengths,
 364          * they cannot be included in tcp_mss.  Thus we need to calculate
 365          * the actual segment length when we need to send a segment which
 366          * includes SACK options.
 367          */
 368         if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
 369                 int32_t opt_len;
 370 
 371                 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
 372                     tcp->tcp_num_sack_blk);
 373                 opt_len = num_sack_blk * sizeof (sack_blk_t) + TCPOPT_NOP_LEN *
 374                     2 + TCPOPT_HEADER_LEN;
 375                 mss = tcp->tcp_mss - opt_len;
 376                 total_hdr_len = connp->conn_ht_iphc_len + opt_len;
 377                 tcp_hdr_len = connp->conn_ht_ulp_len + opt_len;
 378         } else {
 379                 mss = tcp->tcp_mss;
 380                 total_hdr_len = connp->conn_ht_iphc_len;
 381                 tcp_hdr_len = connp->conn_ht_ulp_len;
 382         }
 383 
 384         if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
 385             (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
 386                 cc_after_idle(tcp);
 387         }
 388         if (tcpstate == TCPS_SYN_RCVD) {
 389                 /*
 390                  * The three-way connection establishment handshake is not
 391                  * complete yet. We want to queue the data for transmission
 392                  * after entering ESTABLISHED state (RFC793). A jump to
 393                  * "done" label effectively leaves data on the queue.
 394                  */
 395                 goto done;
 396         } else {
 397                 int usable_r;
 398 
 399                 /*
 400                  * In the special case when cwnd is zero, which can only
 401                  * happen if the connection is ECN capable, return now.
 402                  * New segments is sent using tcp_timer().  The timer
 403                  * is set in tcp_input_data().
 404                  */
 405                 if (tcp->tcp_cwnd == 0) {
 406                         /*
 407                          * Note that tcp_cwnd is 0 before 3-way handshake is
 408                          * finished.
 409                          */
 410                         ASSERT(tcp->tcp_ecn_ok ||
 411                             tcp->tcp_state < TCPS_ESTABLISHED);
 412                         return;
 413                 }
 414 
 415                 /* NOTE: trouble if xmitting while SYN not acked? */
 416                 usable_r = snxt - tcp->tcp_suna;
 417                 usable_r = tcp->tcp_swnd - usable_r;
 418 
 419                 /*
 420                  * Check if the receiver has shrunk the window.  If
 421                  * tcp_wput_data() with NULL mp is called, tcp_fin_sent
 422                  * cannot be set as there is unsent data, so FIN cannot
 423                  * be sent out.  Otherwise, we need to take into account
 424                  * of FIN as it consumes an "invisible" sequence number.
 425                  */
 426                 ASSERT(tcp->tcp_fin_sent == 0);
 427                 if (usable_r < 0) {
 428                         /*
 429                          * The receiver has shrunk the window and we have sent
 430                          * -usable_r date beyond the window, re-adjust.
 431                          *
 432                          * If TCP window scaling is enabled, there can be
 433                          * round down error as the advertised receive window
 434                          * is actually right shifted n bits.  This means that
 435                          * the lower n bits info is wiped out.  It will look
 436                          * like the window is shrunk.  Do a check here to
 437                          * see if the shrunk amount is actually within the
 438                          * error in window calculation.  If it is, just
 439                          * return.  Note that this check is inside the
 440                          * shrunk window check.  This makes sure that even
 441                          * though tcp_process_shrunk_swnd() is not called,
 442                          * we will stop further processing.
 443                          */
 444                         if ((-usable_r >> tcp->tcp_snd_ws) > 0) {
 445                                 tcp_process_shrunk_swnd(tcp, -usable_r);
 446                         }
 447                         return;
 448                 }
 449 
 450                 /* usable = MIN(swnd, cwnd) - unacked_bytes */
 451                 if (tcp->tcp_swnd > tcp->tcp_cwnd)
 452                         usable_r -= tcp->tcp_swnd - tcp->tcp_cwnd;
 453 
 454                 /* usable = MIN(usable, unsent) */
 455                 if (usable_r > len)
 456                         usable_r = len;
 457 
 458                 /* usable = MAX(usable, {1 for urgent, 0 for data}) */
 459                 if (usable_r > 0) {
 460                         usable = usable_r;
 461                 } else {
 462                         /* Bypass all other unnecessary processing. */
 463                         goto done;
 464                 }
 465         }
 466 
 467 #ifdef KERNEL_32
 468         local_time = (mblk_t *)now;
 469 #else
 470         local_time = (mblk_t *)(intptr_t)gethrtime();
 471 #endif
 472 
 473         /*
 474          * "Our" Nagle Algorithm.  This is not the same as in the old
 475          * BSD.  This is more in line with the true intent of Nagle.
 476          *
 477          * The conditions are:
 478          * 1. The amount of unsent data (or amount of data which can be
 479          *    sent, whichever is smaller) is less than Nagle limit.
 480          * 2. The last sent size is also less than Nagle limit.
 481          * 3. There is unack'ed data.
 482          * 4. Urgent pointer is not set.  Send urgent data ignoring the
 483          *    Nagle algorithm.  This reduces the probability that urgent
 484          *    bytes get "merged" together.
 485          * 5. The app has not closed the connection.  This eliminates the
 486          *    wait time of the receiving side waiting for the last piece of
 487          *    (small) data.
 488          *
 489          * If all are satisified, exit without sending anything.  Note
 490          * that Nagle limit can be smaller than 1 MSS.  Nagle limit is
 491          * the smaller of 1 MSS and global tcp_naglim_def (default to be
 492          * 4095).
 493          */
 494         if (usable < (int)tcp->tcp_naglim &&
 495             tcp->tcp_naglim > tcp->tcp_last_sent_len &&
 496             snxt != tcp->tcp_suna &&
 497             !(tcp->tcp_valid_bits & TCP_URG_VALID) &&
 498             !(tcp->tcp_valid_bits & TCP_FSS_VALID)) {
 499                 goto done;
 500         }
 501 
 502         /*
 503          * If tcp_zero_win_probe is not set and the tcp->tcp_cork option
 504          * is set, then we have to force TCP not to send partial segment
 505          * (smaller than MSS bytes). We are calculating the usable now
 506          * based on full mss and will save the rest of remaining data for
 507          * later. When tcp_zero_win_probe is set, TCP needs to send out
 508          * something to do zero window probe.
 509          */
 510         if (tcp->tcp_cork && !tcp->tcp_zero_win_probe) {
 511                 if (usable < mss)
 512                         goto done;
 513                 usable = (usable / mss) * mss;
 514         }
 515 
 516         /* Update the latest receive window size in TCP header. */
 517         tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
 518 
 519         /* Send the packet. */
 520         rc = tcp_send(tcp, mss, total_hdr_len, tcp_hdr_len,
 521             num_sack_blk, &usable, &snxt, &tail_unsent, &xmit_tail,
 522             local_time);
 523 
 524         /* Pretend that all we were trying to send really got sent */
 525         if (rc < 0 && tail_unsent < 0) {
 526                 do {
 527                         xmit_tail = xmit_tail->b_cont;
 528                         xmit_tail->b_prev = local_time;
 529                         ASSERT((uintptr_t)(xmit_tail->b_wptr -
 530                             xmit_tail->b_rptr) <= (uintptr_t)INT_MAX);
 531                         tail_unsent += (int)(xmit_tail->b_wptr -
 532                             xmit_tail->b_rptr);
 533                 } while (tail_unsent < 0);
 534         }
 535 done:;
 536         tcp->tcp_xmit_tail = xmit_tail;
 537         tcp->tcp_xmit_tail_unsent = tail_unsent;
 538         len = tcp->tcp_snxt - snxt;
 539         if (len) {
 540                 /*
 541                  * If new data was sent, need to update the notsack
 542                  * list, which is, afterall, data blocks that have
 543                  * not been sack'ed by the receiver.  New data is
 544                  * not sack'ed.
 545                  */
 546                 if (tcp->tcp_snd_sack_ok && tcp->tcp_notsack_list != NULL) {
 547                         /* len is a negative value. */
 548                         tcp->tcp_pipe -= len;
 549                         tcp_notsack_update(&(tcp->tcp_notsack_list),
 550                             tcp->tcp_snxt, snxt,
 551                             &(tcp->tcp_num_notsack_blk),
 552                             &(tcp->tcp_cnt_notsack_list));
 553                 }
 554                 tcp->tcp_snxt = snxt + tcp->tcp_fin_sent;
 555                 tcp->tcp_rack = tcp->tcp_rnxt;
 556                 tcp->tcp_rack_cnt = 0;
 557                 if ((snxt + len) == tcp->tcp_suna) {
 558                         TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
 559                 }
 560         } else if (snxt == tcp->tcp_suna && tcp->tcp_swnd == 0) {
 561                 /*
 562                  * Didn't send anything. Make sure the timer is running
 563                  * so that we will probe a zero window.
 564                  */
 565                 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
 566         }
 567         /* Note that len is the amount we just sent but with a negative sign */
 568         tcp->tcp_unsent += len;
 569         mutex_enter(&tcp->tcp_non_sq_lock);
 570         if (tcp->tcp_flow_stopped) {
 571                 if (TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
 572                         tcp_clrqfull(tcp);
 573                 }
 574         } else if (TCP_UNSENT_BYTES(tcp) >= connp->conn_sndbuf) {
 575                 if (!(tcp->tcp_detached))
 576                         tcp_setqfull(tcp);
 577         }
 578         mutex_exit(&tcp->tcp_non_sq_lock);
 579 }
 580 
 581 /*
 582  * Initial STREAMS write side put() procedure for sockets. It tries to
 583  * handle the T_CAPABILITY_REQ which sockfs sends down while setting
 584  * up the socket without using the squeue. Non T_CAPABILITY_REQ messages
 585  * are handled by tcp_wput() as usual.
 586  *
 587  * All further messages will also be handled by tcp_wput() because we cannot
 588  * be sure that the above short cut is safe later.
 589  */
 590 void
 591 tcp_wput_sock(queue_t *wq, mblk_t *mp)
 592 {
 593         conn_t                  *connp = Q_TO_CONN(wq);
 594         tcp_t                   *tcp = connp->conn_tcp;
 595         struct T_capability_req *car = (struct T_capability_req *)mp->b_rptr;
 596 
 597         ASSERT(wq->q_qinfo == &tcp_sock_winit);
 598         wq->q_qinfo = &tcp_winit;
 599 
 600         ASSERT(IPCL_IS_TCP(connp));
 601         ASSERT(TCP_IS_SOCKET(tcp));
 602 
 603         if (DB_TYPE(mp) == M_PCPROTO &&
 604             MBLKL(mp) == sizeof (struct T_capability_req) &&
 605             car->PRIM_type == T_CAPABILITY_REQ) {
 606                 tcp_capability_req(tcp, mp);
 607                 return;
 608         }
 609 
 610         tcp_wput(wq, mp);
 611 }
 612 
 613 /* ARGSUSED */
 614 void
 615 tcp_wput_fallback(queue_t *wq, mblk_t *mp)
 616 {
 617 #ifdef DEBUG
 618         cmn_err(CE_CONT, "tcp_wput_fallback: Message during fallback \n");
 619 #endif
 620         freemsg(mp);
 621 }
 622 
 623 /*
 624  * Call by tcp_wput() to handle misc non M_DATA messages.
 625  */
 626 /* ARGSUSED */
 627 static void
 628 tcp_wput_nondata(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
 629 {
 630         conn_t  *connp = (conn_t *)arg;
 631         tcp_t   *tcp = connp->conn_tcp;
 632 
 633         ASSERT(DB_TYPE(mp) != M_IOCTL);
 634         /*
 635          * TCP is D_MP and qprocsoff() is done towards the end of the tcp_close.
 636          * Once the close starts, streamhead and sockfs will not let any data
 637          * packets come down (close ensures that there are no threads using the
 638          * queue and no new threads will come down) but since qprocsoff()
 639          * hasn't happened yet, a M_FLUSH or some non data message might
 640          * get reflected back (in response to our own FLUSHRW) and get
 641          * processed after tcp_close() is done. The conn would still be valid
 642          * because a ref would have added but we need to check the state
 643          * before actually processing the packet.
 644          */
 645         if (TCP_IS_DETACHED(tcp) || (tcp->tcp_state == TCPS_CLOSED)) {
 646                 freemsg(mp);
 647                 return;
 648         }
 649 
 650         switch (DB_TYPE(mp)) {
 651         case M_IOCDATA:
 652                 tcp_wput_iocdata(tcp, mp);
 653                 break;
 654         case M_FLUSH:
 655                 tcp_wput_flush(tcp, mp);
 656                 break;
 657         default:
 658                 ip_wput_nondata(connp->conn_wq, mp);
 659                 break;
 660         }
 661 }
 662 
 663 /* tcp_wput_flush is called by tcp_wput_nondata to handle M_FLUSH messages. */
 664 static void
 665 tcp_wput_flush(tcp_t *tcp, mblk_t *mp)
 666 {
 667         uchar_t fval = *mp->b_rptr;
 668         mblk_t  *tail;
 669         conn_t  *connp = tcp->tcp_connp;
 670         queue_t *q = connp->conn_wq;
 671 
 672         /* TODO: How should flush interact with urgent data? */
 673         if ((fval & FLUSHW) && tcp->tcp_xmit_head != NULL &&
 674             !(tcp->tcp_valid_bits & TCP_URG_VALID)) {
 675                 /*
 676                  * Flush only data that has not yet been put on the wire.  If
 677                  * we flush data that we have already transmitted, life, as we
 678                  * know it, may come to an end.
 679                  */
 680                 tail = tcp->tcp_xmit_tail;
 681                 tail->b_wptr -= tcp->tcp_xmit_tail_unsent;
 682                 tcp->tcp_xmit_tail_unsent = 0;
 683                 tcp->tcp_unsent = 0;
 684                 if (tail->b_wptr != tail->b_rptr)
 685                         tail = tail->b_cont;
 686                 if (tail) {
 687                         mblk_t **excess = &tcp->tcp_xmit_head;
 688                         for (;;) {
 689                                 mblk_t *mp1 = *excess;
 690                                 if (mp1 == tail)
 691                                         break;
 692                                 tcp->tcp_xmit_tail = mp1;
 693                                 tcp->tcp_xmit_last = mp1;
 694                                 excess = &mp1->b_cont;
 695                         }
 696                         *excess = NULL;
 697                         tcp_close_mpp(&tail);
 698                         if (tcp->tcp_snd_zcopy_aware)
 699                                 tcp_zcopy_notify(tcp);
 700                 }
 701                 /*
 702                  * We have no unsent data, so unsent must be less than
 703                  * conn_sndlowat, so re-enable flow.
 704                  */
 705                 mutex_enter(&tcp->tcp_non_sq_lock);
 706                 if (tcp->tcp_flow_stopped) {
 707                         tcp_clrqfull(tcp);
 708                 }
 709                 mutex_exit(&tcp->tcp_non_sq_lock);
 710         }
 711         /*
 712          * TODO: you can't just flush these, you have to increase rwnd for one
 713          * thing.  For another, how should urgent data interact?
 714          */
 715         if (fval & FLUSHR) {
 716                 *mp->b_rptr = fval & ~FLUSHW;
 717                 /* XXX */
 718                 qreply(q, mp);
 719                 return;
 720         }
 721         freemsg(mp);
 722 }
 723 
 724 /*
 725  * tcp_wput_iocdata is called by tcp_wput_nondata to handle all M_IOCDATA
 726  * messages.
 727  */
 728 static void
 729 tcp_wput_iocdata(tcp_t *tcp, mblk_t *mp)
 730 {
 731         mblk_t          *mp1;
 732         struct iocblk   *iocp = (struct iocblk *)mp->b_rptr;
 733         STRUCT_HANDLE(strbuf, sb);
 734         uint_t          addrlen;
 735         conn_t          *connp = tcp->tcp_connp;
 736         queue_t         *q = connp->conn_wq;
 737 
 738         /* Make sure it is one of ours. */
 739         switch (iocp->ioc_cmd) {
 740         case TI_GETMYNAME:
 741         case TI_GETPEERNAME:
 742                 break;
 743         default:
 744                 /*
 745                  * If the conn is closing, then error the ioctl here. Otherwise
 746                  * use the CONN_IOCTLREF_* macros to hold off tcp_close until
 747                  * we're done here.
 748                  */
 749                 mutex_enter(&connp->conn_lock);
 750                 if (connp->conn_state_flags & CONN_CLOSING) {
 751                         mutex_exit(&connp->conn_lock);
 752                         iocp->ioc_error = EINVAL;
 753                         mp->b_datap->db_type = M_IOCNAK;
 754                         iocp->ioc_count = 0;
 755                         qreply(q, mp);
 756                         return;
 757                 }
 758 
 759                 CONN_INC_IOCTLREF_LOCKED(connp);
 760                 ip_wput_nondata(q, mp);
 761                 CONN_DEC_IOCTLREF(connp);
 762                 return;
 763         }
 764         switch (mi_copy_state(q, mp, &mp1)) {
 765         case -1:
 766                 return;
 767         case MI_COPY_CASE(MI_COPY_IN, 1):
 768                 break;
 769         case MI_COPY_CASE(MI_COPY_OUT, 1):
 770                 /* Copy out the strbuf. */
 771                 mi_copyout(q, mp);
 772                 return;
 773         case MI_COPY_CASE(MI_COPY_OUT, 2):
 774                 /* All done. */
 775                 mi_copy_done(q, mp, 0);
 776                 return;
 777         default:
 778                 mi_copy_done(q, mp, EPROTO);
 779                 return;
 780         }
 781         /* Check alignment of the strbuf */
 782         if (!OK_32PTR(mp1->b_rptr)) {
 783                 mi_copy_done(q, mp, EINVAL);
 784                 return;
 785         }
 786 
 787         STRUCT_SET_HANDLE(sb, iocp->ioc_flag, (void *)mp1->b_rptr);
 788 
 789         if (connp->conn_family == AF_INET)
 790                 addrlen = sizeof (sin_t);
 791         else
 792                 addrlen = sizeof (sin6_t);
 793 
 794         if (STRUCT_FGET(sb, maxlen) < addrlen) {
 795                 mi_copy_done(q, mp, EINVAL);
 796                 return;
 797         }
 798 
 799         switch (iocp->ioc_cmd) {
 800         case TI_GETMYNAME:
 801                 break;
 802         case TI_GETPEERNAME:
 803                 if (tcp->tcp_state < TCPS_SYN_RCVD) {
 804                         mi_copy_done(q, mp, ENOTCONN);
 805                         return;
 806                 }
 807                 break;
 808         }
 809         mp1 = mi_copyout_alloc(q, mp, STRUCT_FGETP(sb, buf), addrlen, B_TRUE);
 810         if (!mp1)
 811                 return;
 812 
 813         STRUCT_FSET(sb, len, addrlen);
 814         switch (((struct iocblk *)mp->b_rptr)->ioc_cmd) {
 815         case TI_GETMYNAME:
 816                 (void) conn_getsockname(connp, (struct sockaddr *)mp1->b_wptr,
 817                     &addrlen);
 818                 break;
 819         case TI_GETPEERNAME:
 820                 (void) conn_getpeername(connp, (struct sockaddr *)mp1->b_wptr,
 821                     &addrlen);
 822                 break;
 823         }
 824         mp1->b_wptr += addrlen;
 825         /* Copy out the address */
 826         mi_copyout(q, mp);
 827 }
 828 
 829 /*
 830  * tcp_wput_ioctl is called by tcp_wput_nondata() to handle all M_IOCTL
 831  * messages.
 832  */
 833 /* ARGSUSED */
 834 static void
 835 tcp_wput_ioctl(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
 836 {
 837         conn_t          *connp = (conn_t *)arg;
 838         tcp_t           *tcp = connp->conn_tcp;
 839         queue_t         *q = connp->conn_wq;
 840         struct iocblk   *iocp;
 841 
 842         ASSERT(DB_TYPE(mp) == M_IOCTL);
 843         /*
 844          * Try and ASSERT the minimum possible references on the
 845          * conn early enough. Since we are executing on write side,
 846          * the connection is obviously not detached and that means
 847          * there is a ref each for TCP and IP. Since we are behind
 848          * the squeue, the minimum references needed are 3. If the
 849          * conn is in classifier hash list, there should be an
 850          * extra ref for that (we check both the possibilities).
 851          */
 852         ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
 853             (connp->conn_fanout == NULL && connp->conn_ref >= 3));
 854 
 855         iocp = (struct iocblk *)mp->b_rptr;
 856         switch (iocp->ioc_cmd) {
 857         case _SIOCSOCKFALLBACK:
 858                 /*
 859                  * Either sockmod is about to be popped and the socket
 860                  * would now be treated as a plain stream, or a module
 861                  * is about to be pushed so we could no longer use read-
 862                  * side synchronous streams for fused loopback tcp.
 863                  * Drain any queued data and disable direct sockfs
 864                  * interface from now on.
 865                  */
 866                 if (!tcp->tcp_issocket) {
 867                         DB_TYPE(mp) = M_IOCNAK;
 868                         iocp->ioc_error = EINVAL;
 869                 } else {
 870                         tcp_use_pure_tpi(tcp);
 871                         DB_TYPE(mp) = M_IOCACK;
 872                         iocp->ioc_error = 0;
 873                 }
 874                 iocp->ioc_count = 0;
 875                 iocp->ioc_rval = 0;
 876                 qreply(q, mp);
 877                 return;
 878         }
 879 
 880         /*
 881          * If the conn is closing, then error the ioctl here. Otherwise bump the
 882          * conn_ioctlref to hold off tcp_close until we're done here.
 883          */
 884         mutex_enter(&(connp)->conn_lock);
 885         if ((connp)->conn_state_flags & CONN_CLOSING) {
 886                 mutex_exit(&(connp)->conn_lock);
 887                 iocp->ioc_error = EINVAL;
 888                 mp->b_datap->db_type = M_IOCNAK;
 889                 iocp->ioc_count = 0;
 890                 qreply(q, mp);
 891                 return;
 892         }
 893 
 894         CONN_INC_IOCTLREF_LOCKED(connp);
 895         ip_wput_nondata(q, mp);
 896         CONN_DEC_IOCTLREF(connp);
 897 }
 898 
 899 /*
 900  * This routine is called by tcp_wput() to handle all TPI requests.
 901  */
 902 /* ARGSUSED */
 903 static void
 904 tcp_wput_proto(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
 905 {
 906         conn_t          *connp = (conn_t *)arg;
 907         tcp_t           *tcp = connp->conn_tcp;
 908         union T_primitives *tprim = (union T_primitives *)mp->b_rptr;
 909         uchar_t         *rptr;
 910         t_scalar_t      type;
 911         cred_t          *cr;
 912 
 913         /*
 914          * Try and ASSERT the minimum possible references on the
 915          * conn early enough. Since we are executing on write side,
 916          * the connection is obviously not detached and that means
 917          * there is a ref each for TCP and IP. Since we are behind
 918          * the squeue, the minimum references needed are 3. If the
 919          * conn is in classifier hash list, there should be an
 920          * extra ref for that (we check both the possibilities).
 921          */
 922         ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
 923             (connp->conn_fanout == NULL && connp->conn_ref >= 3));
 924 
 925         rptr = mp->b_rptr;
 926         ASSERT((uintptr_t)(mp->b_wptr - rptr) <= (uintptr_t)INT_MAX);
 927         if ((mp->b_wptr - rptr) >= sizeof (t_scalar_t)) {
 928                 type = ((union T_primitives *)rptr)->type;
 929                 if (type == T_EXDATA_REQ) {
 930                         tcp_output_urgent(connp, mp, arg2, NULL);
 931                 } else if (type != T_DATA_REQ) {
 932                         goto non_urgent_data;
 933                 } else {
 934                         /* TODO: options, flags, ... from user */
 935                         /* Set length to zero for reclamation below */
 936                         tcp_wput_data(tcp, mp->b_cont, B_TRUE);
 937                         freeb(mp);
 938                 }
 939                 return;
 940         } else {
 941                 if (connp->conn_debug) {
 942                         (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
 943                             "tcp_wput_proto, dropping one...");
 944                 }
 945                 freemsg(mp);
 946                 return;
 947         }
 948 
 949 non_urgent_data:
 950 
 951         switch ((int)tprim->type) {
 952         case O_T_BIND_REQ:      /* bind request */
 953         case T_BIND_REQ:        /* new semantics bind request */
 954                 tcp_tpi_bind(tcp, mp);
 955                 break;
 956         case T_UNBIND_REQ:      /* unbind request */
 957                 tcp_tpi_unbind(tcp, mp);
 958                 break;
 959         case O_T_CONN_RES:      /* old connection response XXX */
 960         case T_CONN_RES:        /* connection response */
 961                 tcp_tli_accept(tcp, mp);
 962                 break;
 963         case T_CONN_REQ:        /* connection request */
 964                 tcp_tpi_connect(tcp, mp);
 965                 break;
 966         case T_DISCON_REQ:      /* disconnect request */
 967                 tcp_disconnect(tcp, mp);
 968                 break;
 969         case T_CAPABILITY_REQ:
 970                 tcp_capability_req(tcp, mp);    /* capability request */
 971                 break;
 972         case T_INFO_REQ:        /* information request */
 973                 tcp_info_req(tcp, mp);
 974                 break;
 975         case T_SVR4_OPTMGMT_REQ:        /* manage options req */
 976         case T_OPTMGMT_REQ:
 977                 /*
 978                  * Note:  no support for snmpcom_req() through new
 979                  * T_OPTMGMT_REQ. See comments in ip.c
 980                  */
 981 
 982                 /*
 983                  * All Solaris components should pass a db_credp
 984                  * for this TPI message, hence we ASSERT.
 985                  * But in case there is some other M_PROTO that looks
 986                  * like a TPI message sent by some other kernel
 987                  * component, we check and return an error.
 988                  */
 989                 cr = msg_getcred(mp, NULL);
 990                 ASSERT(cr != NULL);
 991                 if (cr == NULL) {
 992                         tcp_err_ack(tcp, mp, TSYSERR, EINVAL);
 993                         return;
 994                 }
 995                 /*
 996                  * If EINPROGRESS is returned, the request has been queued
 997                  * for subsequent processing by ip_restart_optmgmt(), which
 998                  * will do the CONN_DEC_REF().
 999                  */
1000                 if ((int)tprim->type == T_SVR4_OPTMGMT_REQ) {
1001                         svr4_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
1002                 } else {
1003                         tpi_optcom_req(connp->conn_wq, mp, cr, &tcp_opt_obj);
1004                 }
1005                 break;
1006 
1007         case T_UNITDATA_REQ:    /* unitdata request */
1008                 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
1009                 break;
1010         case T_ORDREL_REQ:      /* orderly release req */
1011                 freemsg(mp);
1012 
1013                 if (tcp->tcp_fused)
1014                         tcp_unfuse(tcp);
1015 
1016                 if (tcp_xmit_end(tcp) != 0) {
1017                         /*
1018                          * We were crossing FINs and got a reset from
1019                          * the other side. Just ignore it.
1020                          */
1021                         if (connp->conn_debug) {
1022                                 (void) strlog(TCP_MOD_ID, 0, 1,
1023                                     SL_ERROR|SL_TRACE,
1024                                     "tcp_wput_proto, T_ORDREL_REQ out of "
1025                                     "state %s",
1026                                     tcp_display(tcp, NULL,
1027                                     DISP_ADDR_AND_PORT));
1028                         }
1029                 }
1030                 break;
1031         case T_ADDR_REQ:
1032                 tcp_addr_req(tcp, mp);
1033                 break;
1034         default:
1035                 if (connp->conn_debug) {
1036                         (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1037                             "tcp_wput_proto, bogus TPI msg, type %d",
1038                             tprim->type);
1039                 }
1040                 /*
1041                  * We used to M_ERROR.  Sending TNOTSUPPORT gives the user
1042                  * to recover.
1043                  */
1044                 tcp_err_ack(tcp, mp, TNOTSUPPORT, 0);
1045                 break;
1046         }
1047 }
1048 
1049 /*
1050  * Handle special out-of-band ioctl requests (see PSARC/2008/265).
1051  */
1052 static void
1053 tcp_wput_cmdblk(queue_t *q, mblk_t *mp)
1054 {
1055         void    *data;
1056         mblk_t  *datamp = mp->b_cont;
1057         conn_t  *connp = Q_TO_CONN(q);
1058         tcp_t   *tcp = connp->conn_tcp;
1059         cmdblk_t *cmdp = (cmdblk_t *)mp->b_rptr;
1060 
1061         if (datamp == NULL || MBLKL(datamp) < cmdp->cb_len) {
1062                 cmdp->cb_error = EPROTO;
1063                 qreply(q, mp);
1064                 return;
1065         }
1066 
1067         data = datamp->b_rptr;
1068 
1069         switch (cmdp->cb_cmd) {
1070         case TI_GETPEERNAME:
1071                 if (tcp->tcp_state < TCPS_SYN_RCVD)
1072                         cmdp->cb_error = ENOTCONN;
1073                 else
1074                         cmdp->cb_error = conn_getpeername(connp, data,
1075                             &cmdp->cb_len);
1076                 break;
1077         case TI_GETMYNAME:
1078                 cmdp->cb_error = conn_getsockname(connp, data, &cmdp->cb_len);
1079                 break;
1080         default:
1081                 cmdp->cb_error = EINVAL;
1082                 break;
1083         }
1084 
1085         qreply(q, mp);
1086 }
1087 
1088 /*
1089  * The TCP fast path write put procedure.
1090  * NOTE: the logic of the fast path is duplicated from tcp_wput_data()
1091  */
1092 /* ARGSUSED */
1093 void
1094 tcp_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1095 {
1096         int             len;
1097         int             hdrlen;
1098         int             plen;
1099         mblk_t          *mp1;
1100         uchar_t         *rptr;
1101         uint32_t        snxt;
1102         tcpha_t         *tcpha;
1103         struct datab    *db;
1104         uint32_t        suna;
1105         uint32_t        mss;
1106         ipaddr_t        *dst;
1107         ipaddr_t        *src;
1108         uint32_t        sum;
1109         int             usable;
1110         conn_t          *connp = (conn_t *)arg;
1111         tcp_t           *tcp = connp->conn_tcp;
1112         uint32_t        msize;
1113         tcp_stack_t     *tcps = tcp->tcp_tcps;
1114         ip_xmit_attr_t  *ixa;
1115         clock_t         now;
1116 
1117         /*
1118          * Try and ASSERT the minimum possible references on the
1119          * conn early enough. Since we are executing on write side,
1120          * the connection is obviously not detached and that means
1121          * there is a ref each for TCP and IP. Since we are behind
1122          * the squeue, the minimum references needed are 3. If the
1123          * conn is in classifier hash list, there should be an
1124          * extra ref for that (we check both the possibilities).
1125          */
1126         ASSERT((connp->conn_fanout != NULL && connp->conn_ref >= 4) ||
1127             (connp->conn_fanout == NULL && connp->conn_ref >= 3));
1128 
1129         ASSERT(DB_TYPE(mp) == M_DATA);
1130         msize = (mp->b_cont == NULL) ? MBLKL(mp) : msgdsize(mp);
1131 
1132         mutex_enter(&tcp->tcp_non_sq_lock);
1133         tcp->tcp_squeue_bytes -= msize;
1134         mutex_exit(&tcp->tcp_non_sq_lock);
1135 
1136         /* Bypass tcp protocol for fused tcp loopback */
1137         if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1138                 return;
1139 
1140         mss = tcp->tcp_mss;
1141         /*
1142          * If ZEROCOPY has turned off, try not to send any zero-copy message
1143          * down. Do backoff, now.
1144          */
1145         if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on)
1146                 mp = tcp_zcopy_backoff(tcp, mp, B_FALSE);
1147 
1148 
1149         ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1150         len = (int)(mp->b_wptr - mp->b_rptr);
1151 
1152         /*
1153          * Criteria for fast path:
1154          *
1155          *   1. no unsent data
1156          *   2. single mblk in request
1157          *   3. connection established
1158          *   4. data in mblk
1159          *   5. len <= mss
1160          *   6. no tcp_valid bits
1161          */
1162         if ((tcp->tcp_unsent != 0) ||
1163             (tcp->tcp_cork) ||
1164             (mp->b_cont != NULL) ||
1165             (tcp->tcp_state != TCPS_ESTABLISHED) ||
1166             (len == 0) ||
1167             (len > mss) ||
1168             (tcp->tcp_valid_bits != 0)) {
1169                 tcp_wput_data(tcp, mp, B_FALSE);
1170                 return;
1171         }
1172 
1173         ASSERT(tcp->tcp_xmit_tail_unsent == 0);
1174         ASSERT(tcp->tcp_fin_sent == 0);
1175 
1176         /* queue new packet onto retransmission queue */
1177         if (tcp->tcp_xmit_head == NULL) {
1178                 tcp->tcp_xmit_head = mp;
1179         } else {
1180                 tcp->tcp_xmit_last->b_cont = mp;
1181         }
1182         tcp->tcp_xmit_last = mp;
1183         tcp->tcp_xmit_tail = mp;
1184 
1185         /* find out how much we can send */
1186         /* BEGIN CSTYLED */
1187         /*
1188          *    un-acked     usable
1189          *  |--------------|-----------------|
1190          *  tcp_suna       tcp_snxt       tcp_suna+tcp_swnd
1191          */
1192         /* END CSTYLED */
1193 
1194         /* start sending from tcp_snxt */
1195         snxt = tcp->tcp_snxt;
1196 
1197         /*
1198          * Check to see if this connection has been idled for some
1199          * time and no ACK is expected.  If it is, we need to slow
1200          * start again to get back the connection's "self-clock" as
1201          * described in VJ's paper.
1202          *
1203          * Reinitialize tcp_cwnd after idle.
1204          */
1205         now = LBOLT_FASTPATH;
1206         if ((tcp->tcp_suna == snxt) && !tcp->tcp_localnet &&
1207             (TICK_TO_MSEC(now - tcp->tcp_last_recv_time) >= tcp->tcp_rto)) {
1208                 cc_after_idle(tcp);
1209         }
1210 
1211         usable = tcp->tcp_swnd;              /* tcp window size */
1212         if (usable > tcp->tcp_cwnd)
1213                 usable = tcp->tcp_cwnd;      /* congestion window smaller */
1214         usable -= snxt;         /* subtract stuff already sent */
1215         suna = tcp->tcp_suna;
1216         usable += suna;
1217         /* usable can be < 0 if the congestion window is smaller */
1218         if (len > usable) {
1219                 /* Can't send complete M_DATA in one shot */
1220                 goto slow;
1221         }
1222 
1223         mutex_enter(&tcp->tcp_non_sq_lock);
1224         if (tcp->tcp_flow_stopped &&
1225             TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1226                 tcp_clrqfull(tcp);
1227         }
1228         mutex_exit(&tcp->tcp_non_sq_lock);
1229 
1230         /*
1231          * determine if anything to send (Nagle).
1232          *
1233          *   1. len < tcp_mss (i.e. small)
1234          *   2. unacknowledged data present
1235          *   3. len < nagle limit
1236          *   4. last packet sent < nagle limit (previous packet sent)
1237          */
1238         if ((len < mss) && (snxt != suna) &&
1239             (len < (int)tcp->tcp_naglim) &&
1240             (tcp->tcp_last_sent_len < tcp->tcp_naglim)) {
1241                 /*
1242                  * This was the first unsent packet and normally
1243                  * mss < xmit_hiwater so there is no need to worry
1244                  * about flow control. The next packet will go
1245                  * through the flow control check in tcp_wput_data().
1246                  */
1247                 /* leftover work from above */
1248                 tcp->tcp_unsent = len;
1249                 tcp->tcp_xmit_tail_unsent = len;
1250 
1251                 return;
1252         }
1253 
1254         /*
1255          * len <= tcp->tcp_mss && len == unsent so no sender silly window.  Can
1256          * send now.
1257          */
1258 
1259         if (snxt == suna) {
1260                 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1261         }
1262 
1263         /* we have always sent something */
1264         tcp->tcp_rack_cnt = 0;
1265 
1266         tcp->tcp_snxt = snxt + len;
1267         tcp->tcp_rack = tcp->tcp_rnxt;
1268 
1269         if ((mp1 = dupb(mp)) == 0)
1270                 goto no_memory;
1271 #ifdef KERNEL_32
1272         mp->b_prev = (mblk_t *)(uintptr_t)now;
1273 #else
1274         mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
1275 #endif
1276         mp->b_next = (mblk_t *)(uintptr_t)snxt;
1277 
1278         /* adjust tcp header information */
1279         tcpha = tcp->tcp_tcpha;
1280         tcpha->tha_flags = (TH_ACK|TH_PUSH);
1281 
1282         sum = len + connp->conn_ht_ulp_len + connp->conn_sum;
1283         sum = (sum >> 16) + (sum & 0xFFFF);
1284         tcpha->tha_sum = htons(sum);
1285 
1286         tcpha->tha_seq = htonl(snxt);
1287 
1288         TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1289         TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1290         TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1291         tcp->tcp_cs.tcp_out_data_segs++;
1292         tcp->tcp_cs.tcp_out_data_bytes += len;
1293 
1294         /* Update the latest receive window size in TCP header. */
1295         tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
1296 
1297         tcp->tcp_last_sent_len = (ushort_t)len;
1298 
1299         plen = len + connp->conn_ht_iphc_len;
1300 
1301         ixa = connp->conn_ixa;
1302         ixa->ixa_pktlen = plen;
1303 
1304         if (ixa->ixa_flags & IXAF_IS_IPV4) {
1305                 tcp->tcp_ipha->ipha_length = htons(plen);
1306         } else {
1307                 tcp->tcp_ip6h->ip6_plen = htons(plen - IPV6_HDR_LEN);
1308         }
1309 
1310         /* see if we need to allocate a mblk for the headers */
1311         hdrlen = connp->conn_ht_iphc_len;
1312         rptr = mp1->b_rptr - hdrlen;
1313         db = mp1->b_datap;
1314         if ((db->db_ref != 2) || rptr < db->db_base ||
1315             (!OK_32PTR(rptr))) {
1316                 /* NOTE: we assume allocb returns an OK_32PTR */
1317                 mp = allocb(hdrlen + tcps->tcps_wroff_xtra, BPRI_MED);
1318                 if (!mp) {
1319                         freemsg(mp1);
1320                         goto no_memory;
1321                 }
1322                 mp->b_cont = mp1;
1323                 mp1 = mp;
1324                 /* Leave room for Link Level header */
1325                 rptr = &mp1->b_rptr[tcps->tcps_wroff_xtra];
1326                 mp1->b_wptr = &rptr[hdrlen];
1327         }
1328         mp1->b_rptr = rptr;
1329 
1330         /* Fill in the timestamp option. */
1331         if (tcp->tcp_snd_ts_ok) {
1332                 U32_TO_BE32(now,
1333                     (char *)tcpha + TCP_MIN_HEADER_LENGTH + 4);
1334                 U32_TO_BE32(tcp->tcp_ts_recent,
1335                     (char *)tcpha + TCP_MIN_HEADER_LENGTH + 8);
1336         } else {
1337                 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
1338         }
1339 
1340         /* copy header into outgoing packet */
1341         dst = (ipaddr_t *)rptr;
1342         src = (ipaddr_t *)connp->conn_ht_iphc;
1343         dst[0] = src[0];
1344         dst[1] = src[1];
1345         dst[2] = src[2];
1346         dst[3] = src[3];
1347         dst[4] = src[4];
1348         dst[5] = src[5];
1349         dst[6] = src[6];
1350         dst[7] = src[7];
1351         dst[8] = src[8];
1352         dst[9] = src[9];
1353         if (hdrlen -= 40) {
1354                 hdrlen >>= 2;
1355                 dst += 10;
1356                 src += 10;
1357                 do {
1358                         *dst++ = *src++;
1359                 } while (--hdrlen);
1360         }
1361 
1362         /*
1363          * Set the ECN info in the TCP header.  Note that this
1364          * is not the template header.
1365          */
1366         if (tcp->tcp_ecn_ok) {
1367                 TCP_SET_ECT(tcp, rptr);
1368 
1369                 tcpha = (tcpha_t *)(rptr + ixa->ixa_ip_hdr_length);
1370                 if (tcp->tcp_ecn_echo_on)
1371                         tcpha->tha_flags |= TH_ECE;
1372                 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
1373                         tcpha->tha_flags |= TH_CWR;
1374                         tcp->tcp_ecn_cwr_sent = B_TRUE;
1375                 }
1376         }
1377 
1378         if (tcp->tcp_ip_forward_progress) {
1379                 tcp->tcp_ip_forward_progress = B_FALSE;
1380                 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
1381         } else {
1382                 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
1383         }
1384         tcp_send_data(tcp, mp1);
1385         return;
1386 
1387         /*
1388          * If we ran out of memory, we pretend to have sent the packet
1389          * and that it was lost on the wire.
1390          */
1391 no_memory:
1392         return;
1393 
1394 slow:
1395         /* leftover work from above */
1396         tcp->tcp_unsent = len;
1397         tcp->tcp_xmit_tail_unsent = len;
1398         tcp_wput_data(tcp, NULL, B_FALSE);
1399 }
1400 
1401 /* ARGSUSED2 */
1402 void
1403 tcp_output_urgent(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1404 {
1405         int len;
1406         uint32_t msize;
1407         conn_t *connp = (conn_t *)arg;
1408         tcp_t *tcp = connp->conn_tcp;
1409 
1410         msize = msgdsize(mp);
1411 
1412         len = msize - 1;
1413         if (len < 0) {
1414                 freemsg(mp);
1415                 return;
1416         }
1417 
1418         /*
1419          * Try to force urgent data out on the wire. Even if we have unsent
1420          * data this will at least send the urgent flag.
1421          * XXX does not handle more flag correctly.
1422          */
1423         len += tcp->tcp_unsent;
1424         len += tcp->tcp_snxt;
1425         tcp->tcp_urg = len;
1426         tcp->tcp_valid_bits |= TCP_URG_VALID;
1427 
1428         /* Bypass tcp protocol for fused tcp loopback */
1429         if (tcp->tcp_fused && tcp_fuse_output(tcp, mp, msize))
1430                 return;
1431 
1432         /* Strip off the T_EXDATA_REQ if the data is from TPI */
1433         if (DB_TYPE(mp) != M_DATA) {
1434                 mblk_t *mp1 = mp;
1435                 ASSERT(!IPCL_IS_NONSTR(connp));
1436                 mp = mp->b_cont;
1437                 freeb(mp1);
1438         }
1439         tcp_wput_data(tcp, mp, B_TRUE);
1440 }
1441 
1442 /*
1443  * Called by streams close routine via squeues when our client blows off its
1444  * descriptor, we take this to mean: "close the stream state NOW, close the tcp
1445  * connection politely" When SO_LINGER is set (with a non-zero linger time and
1446  * it is not a nonblocking socket) then this routine sleeps until the FIN is
1447  * acked.
1448  *
1449  * NOTE: tcp_close potentially returns error when lingering.
1450  * However, the stream head currently does not pass these errors
1451  * to the application. 4.4BSD only returns EINTR and EWOULDBLOCK
1452  * errors to the application (from tsleep()) and not errors
1453  * like ECONNRESET caused by receiving a reset packet.
1454  */
1455 
1456 /* ARGSUSED */
1457 void
1458 tcp_close_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1459 {
1460         char    *msg;
1461         conn_t  *connp = (conn_t *)arg;
1462         tcp_t   *tcp = connp->conn_tcp;
1463         clock_t delta = 0;
1464         tcp_stack_t     *tcps = tcp->tcp_tcps;
1465 
1466         /*
1467          * When a non-STREAMS socket is being closed, it does not always
1468          * stick around waiting for tcp_close_output to run and can therefore
1469          * have dropped a reference already. So adjust the asserts accordingly.
1470          */
1471         ASSERT((connp->conn_fanout != NULL &&
1472             connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 3 : 4)) ||
1473             (connp->conn_fanout == NULL &&
1474             connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3)));
1475 
1476         mutex_enter(&tcp->tcp_eager_lock);
1477         if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
1478                 /*
1479                  * Cleanup for listener. For non-STREAM sockets sockfs will
1480                  * close all the eagers on 'q', so in that case only deal
1481                  * with 'q0'.
1482                  */
1483                 tcp_eager_cleanup(tcp, IPCL_IS_NONSTR(connp) ? 1 : 0);
1484                 tcp->tcp_wait_for_eagers = 1;
1485         }
1486         mutex_exit(&tcp->tcp_eager_lock);
1487 
1488         tcp->tcp_lso = B_FALSE;
1489 
1490         msg = NULL;
1491         switch (tcp->tcp_state) {
1492         case TCPS_CLOSED:
1493         case TCPS_IDLE:
1494                 break;
1495         case TCPS_BOUND:
1496                 if (tcp->tcp_listener != NULL) {
1497                         ASSERT(IPCL_IS_NONSTR(connp));
1498                         /*
1499                          * Unlink from the listener and drop the reference
1500                          * put on it by the eager. tcp_closei_local will not
1501                          * do it because tcp_tconnind_started is TRUE.
1502                          */
1503                         mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1504                         tcp_eager_unlink(tcp);
1505                         mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1506                         CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1507                 }
1508                 break;
1509         case TCPS_LISTEN:
1510                 break;
1511         case TCPS_SYN_SENT:
1512                 msg = "tcp_close, during connect";
1513                 break;
1514         case TCPS_SYN_RCVD:
1515                 /*
1516                  * Close during the connect 3-way handshake
1517                  * but here there may or may not be pending data
1518                  * already on queue. Process almost same as in
1519                  * the ESTABLISHED state.
1520                  */
1521                 /* FALLTHRU */
1522         default:
1523                 if (tcp->tcp_fused)
1524                         tcp_unfuse(tcp);
1525 
1526                 /*
1527                  * If SO_LINGER has set a zero linger time, abort the
1528                  * connection with a reset.
1529                  */
1530                 if (connp->conn_linger && connp->conn_lingertime == 0) {
1531                         msg = "tcp_close, zero lingertime";
1532                         break;
1533                 }
1534 
1535                 /*
1536                  * Abort connection if there is unread data queued.
1537                  */
1538                 if (tcp->tcp_rcv_list || tcp->tcp_reass_head) {
1539                         msg = "tcp_close, unread data";
1540                         break;
1541                 }
1542 
1543                 /*
1544                  * Abort connection if it is being closed without first
1545                  * being accepted. This can happen if a listening non-STREAM
1546                  * socket wants to get rid of the socket, for example, if the
1547                  * listener is closing.
1548                  */
1549                 if (tcp->tcp_listener != NULL) {
1550                         ASSERT(IPCL_IS_NONSTR(connp));
1551                         msg = "tcp_close, close before accept";
1552 
1553                         /*
1554                          * Unlink from the listener and drop the reference
1555                          * put on it by the eager. tcp_closei_local will not
1556                          * do it because tcp_tconnind_started is TRUE.
1557                          */
1558                         mutex_enter(&tcp->tcp_saved_listener->tcp_eager_lock);
1559                         tcp_eager_unlink(tcp);
1560                         mutex_exit(&tcp->tcp_saved_listener->tcp_eager_lock);
1561                         CONN_DEC_REF(tcp->tcp_saved_listener->tcp_connp);
1562                         break;
1563                 }
1564 
1565                 /*
1566                  * Transmit the FIN before detaching the tcp_t.
1567                  * After tcp_detach returns this queue/perimeter
1568                  * no longer owns the tcp_t thus others can modify it.
1569                  */
1570                 (void) tcp_xmit_end(tcp);
1571 
1572                 /*
1573                  * If lingering on close then wait until the fin is acked,
1574                  * the SO_LINGER time passes, or a reset is sent/received.
1575                  */
1576                 if (connp->conn_linger && connp->conn_lingertime > 0 &&
1577                     !(tcp->tcp_fin_acked) &&
1578                     tcp->tcp_state >= TCPS_ESTABLISHED) {
1579                         if (tcp->tcp_closeflags & (FNDELAY|FNONBLOCK)) {
1580                                 tcp->tcp_client_errno = EWOULDBLOCK;
1581                         } else if (tcp->tcp_client_errno == 0) {
1582 
1583                                 ASSERT(tcp->tcp_linger_tid == 0);
1584 
1585                                 /* conn_lingertime is in sec. */
1586                                 tcp->tcp_linger_tid = TCP_TIMER(tcp,
1587                                     tcp_close_linger_timeout,
1588                                     connp->conn_lingertime * MILLISEC);
1589 
1590                                 /* tcp_close_linger_timeout will finish close */
1591                                 if (tcp->tcp_linger_tid == 0)
1592                                         tcp->tcp_client_errno = ENOSR;
1593                                 else
1594                                         return;
1595                         }
1596 
1597                         /*
1598                          * Check if we need to detach or just close
1599                          * the instance.
1600                          */
1601                         if (tcp->tcp_state <= TCPS_LISTEN)
1602                                 break;
1603                 }
1604 
1605                 /*
1606                  * Make sure that no other thread will access the conn_rq of
1607                  * this instance (through lookups etc.) as conn_rq will go
1608                  * away shortly.
1609                  */
1610                 tcp_acceptor_hash_remove(tcp);
1611 
1612                 mutex_enter(&tcp->tcp_non_sq_lock);
1613                 if (tcp->tcp_flow_stopped) {
1614                         tcp_clrqfull(tcp);
1615                 }
1616                 mutex_exit(&tcp->tcp_non_sq_lock);
1617 
1618                 if (tcp->tcp_timer_tid != 0) {
1619                         delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
1620                         tcp->tcp_timer_tid = 0;
1621                 }
1622                 /*
1623                  * Need to cancel those timers which will not be used when
1624                  * TCP is detached.  This has to be done before the conn_wq
1625                  * is set to NULL.
1626                  */
1627                 tcp_timers_stop(tcp);
1628 
1629                 tcp->tcp_detached = B_TRUE;
1630                 if (tcp->tcp_state == TCPS_TIME_WAIT) {
1631                         tcp_time_wait_append(tcp);
1632                         TCP_DBGSTAT(tcps, tcp_detach_time_wait);
1633                         ASSERT(connp->conn_ref >=
1634                             (IPCL_IS_NONSTR(connp) ? 2 : 3));
1635                         goto finish;
1636                 }
1637 
1638                 /*
1639                  * If delta is zero the timer event wasn't executed and was
1640                  * successfully canceled. In this case we need to restart it
1641                  * with the minimal delta possible.
1642                  */
1643                 if (delta >= 0)
1644                         tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1645                             delta ? delta : 1);
1646 
1647                 ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 2 : 3));
1648                 goto finish;
1649         }
1650 
1651         /* Detach did not complete. Still need to remove q from stream. */
1652         if (msg) {
1653                 if (tcp->tcp_state == TCPS_ESTABLISHED ||
1654                     tcp->tcp_state == TCPS_CLOSE_WAIT)
1655                         TCPS_BUMP_MIB(tcps, tcpEstabResets);
1656                 if (tcp->tcp_state == TCPS_SYN_SENT ||
1657                     tcp->tcp_state == TCPS_SYN_RCVD)
1658                         TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1659                 tcp_xmit_ctl(msg, tcp,  tcp->tcp_snxt, 0, TH_RST);
1660         }
1661 
1662         tcp_closei_local(tcp);
1663         CONN_DEC_REF(connp);
1664         ASSERT(connp->conn_ref >= (IPCL_IS_NONSTR(connp) ? 1 : 2));
1665 
1666 finish:
1667         /*
1668          * Don't change the queues in the case of a listener that has
1669          * eagers in its q or q0. It could surprise the eagers.
1670          * Instead wait for the eagers outside the squeue.
1671          *
1672          * For non-STREAMS sockets tcp_wait_for_eagers implies that
1673          * we should delay the su_closed upcall until all eagers have
1674          * dropped their references.
1675          */
1676         if (!tcp->tcp_wait_for_eagers) {
1677                 tcp->tcp_detached = B_TRUE;
1678                 connp->conn_rq = NULL;
1679                 connp->conn_wq = NULL;
1680 
1681                 /* non-STREAM socket, release the upper handle */
1682                 if (IPCL_IS_NONSTR(connp)) {
1683                         ASSERT(connp->conn_upper_handle != NULL);
1684                         (*connp->conn_upcalls->su_closed)
1685                             (connp->conn_upper_handle);
1686                         connp->conn_upper_handle = NULL;
1687                         connp->conn_upcalls = NULL;
1688                 }
1689         }
1690 
1691         /* Signal tcp_close() to finish closing. */
1692         mutex_enter(&tcp->tcp_closelock);
1693         tcp->tcp_closed = 1;
1694         cv_signal(&tcp->tcp_closecv);
1695         mutex_exit(&tcp->tcp_closelock);
1696 }
1697 
1698 /* ARGSUSED */
1699 void
1700 tcp_shutdown_output(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1701 {
1702         conn_t  *connp = (conn_t *)arg;
1703         tcp_t   *tcp = connp->conn_tcp;
1704 
1705         freemsg(mp);
1706 
1707         if (tcp->tcp_fused)
1708                 tcp_unfuse(tcp);
1709 
1710         if (tcp_xmit_end(tcp) != 0) {
1711                 /*
1712                  * We were crossing FINs and got a reset from
1713                  * the other side. Just ignore it.
1714                  */
1715                 if (connp->conn_debug) {
1716                         (void) strlog(TCP_MOD_ID, 0, 1,
1717                             SL_ERROR|SL_TRACE,
1718                             "tcp_shutdown_output() out of state %s",
1719                             tcp_display(tcp, NULL, DISP_ADDR_AND_PORT));
1720                 }
1721         }
1722 }
1723 
1724 #pragma inline(tcp_send_data)
1725 
1726 void
1727 tcp_send_data(tcp_t *tcp, mblk_t *mp)
1728 {
1729         conn_t          *connp = tcp->tcp_connp;
1730 
1731         /*
1732          * Check here to avoid sending zero-copy message down to IP when
1733          * ZEROCOPY capability has turned off. We only need to deal with
1734          * the race condition between sockfs and the notification here.
1735          * Since we have tried to backoff the tcp_xmit_head when turning
1736          * zero-copy off and new messages in tcp_output(), we simply drop
1737          * the dup'ed packet here and let tcp retransmit, if tcp_xmit_zc_clean
1738          * is not true.
1739          */
1740         if (tcp->tcp_snd_zcopy_aware && !tcp->tcp_snd_zcopy_on &&
1741             !tcp->tcp_xmit_zc_clean) {
1742                 ip_drop_output("TCP ZC was disabled but not clean", mp, NULL);
1743                 freemsg(mp);
1744                 return;
1745         }
1746 
1747         DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, connp->conn_ixa,
1748             __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, tcp,
1749             __dtrace_tcp_tcph_t *,
1750             &mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
1751 
1752         ASSERT(connp->conn_ixa->ixa_notify_cookie == connp->conn_tcp);
1753         (void) conn_ip_output(mp, connp->conn_ixa);
1754 }
1755 
1756 /* ARGSUSED2 */
1757 void
1758 tcp_send_synack(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1759 {
1760         conn_t  *econnp = (conn_t *)arg;
1761         tcp_t   *tcp = econnp->conn_tcp;
1762         ip_xmit_attr_t *ixa = econnp->conn_ixa;
1763 
1764         /* Guard against a RST having blown it away while on the squeue */
1765         if (tcp->tcp_state == TCPS_CLOSED) {
1766                 freemsg(mp);
1767                 return;
1768         }
1769 
1770         /*
1771          * In the off-chance that the eager received and responded to
1772          * some other packet while the SYN|ACK was queued, we recalculate
1773          * the ixa_pktlen. It would be better to fix the SYN/accept
1774          * multithreading scheme to avoid this complexity.
1775          */
1776         ixa->ixa_pktlen = msgdsize(mp);
1777         (void) conn_ip_output(mp, ixa);
1778 }
1779 
1780 /*
1781  * tcp_send() is called by tcp_wput_data() and returns one of the following:
1782  *
1783  * -1 = failed allocation.
1784  *  0 = We've either successfully sent data, or our usable send window is too
1785  *      small and we'd rather wait until later before sending again.
1786  */
1787 static int
1788 tcp_send(tcp_t *tcp, const int mss, const int total_hdr_len,
1789     const int tcp_hdr_len, const int num_sack_blk, int *usable,
1790     uint_t *snxt, int *tail_unsent, mblk_t **xmit_tail, mblk_t *local_time)
1791 {
1792         int             num_lso_seg = 1;
1793         uint_t          lso_usable;
1794         boolean_t       do_lso_send = B_FALSE;
1795         tcp_stack_t     *tcps = tcp->tcp_tcps;
1796         conn_t          *connp = tcp->tcp_connp;
1797         ip_xmit_attr_t  *ixa = connp->conn_ixa;
1798 
1799         /*
1800          * Check LSO possibility. The value of tcp->tcp_lso indicates whether
1801          * the underlying connection is LSO capable. Will check whether having
1802          * enough available data to initiate LSO transmission in the for(){}
1803          * loops.
1804          */
1805         if (tcp->tcp_lso && (tcp->tcp_valid_bits & ~TCP_FSS_VALID) == 0)
1806                 do_lso_send = B_TRUE;
1807 
1808         for (;;) {
1809                 struct datab    *db;
1810                 tcpha_t         *tcpha;
1811                 uint32_t        sum;
1812                 mblk_t          *mp, *mp1;
1813                 uchar_t         *rptr;
1814                 int             len;
1815 
1816                 /*
1817                  * Calculate the maximum payload length we can send at one
1818                  * time.
1819                  */
1820                 if (do_lso_send) {
1821                         /*
1822                          * Determine whether or not it's possible to do LSO,
1823                          * and if so, how much data we can send.
1824                          */
1825                         if ((*usable - 1) / mss >= 1) {
1826                                 lso_usable = MIN(tcp->tcp_lso_max, *usable);
1827                                 num_lso_seg = lso_usable / mss;
1828                                 if (lso_usable % mss) {
1829                                         num_lso_seg++;
1830                                         tcp->tcp_last_sent_len = (ushort_t)
1831                                             (lso_usable % mss);
1832                                 } else {
1833                                         tcp->tcp_last_sent_len = (ushort_t)mss;
1834                                 }
1835                         } else {
1836                                 do_lso_send = B_FALSE;
1837                                 num_lso_seg = 1;
1838                                 lso_usable = mss;
1839                         }
1840                 }
1841 
1842                 ASSERT(num_lso_seg <= IP_MAXPACKET / mss + 1);
1843 
1844                 len = mss;
1845                 if (len > *usable) {
1846                         ASSERT(do_lso_send == B_FALSE);
1847 
1848                         len = *usable;
1849                         if (len <= 0) {
1850                                 /* Terminate the loop */
1851                                 break;  /* success; too small */
1852                         }
1853                         /*
1854                          * Sender silly-window avoidance.
1855                          * Ignore this if we are going to send a
1856                          * zero window probe out.
1857                          *
1858                          * TODO: force data into microscopic window?
1859                          *      ==> (!pushed || (unsent > usable))
1860                          */
1861                         if (len < (tcp->tcp_max_swnd >> 1) &&
1862                             (tcp->tcp_unsent - (*snxt - tcp->tcp_snxt)) > len &&
1863                             !((tcp->tcp_valid_bits & TCP_URG_VALID) &&
1864                             len == 1) && (! tcp->tcp_zero_win_probe)) {
1865                                 /*
1866                                  * If the retransmit timer is not running
1867                                  * we start it so that we will retransmit
1868                                  * in the case when the receiver has
1869                                  * decremented the window.
1870                                  */
1871                                 if (*snxt == tcp->tcp_snxt &&
1872                                     *snxt == tcp->tcp_suna) {
1873                                         /*
1874                                          * We are not supposed to send
1875                                          * anything.  So let's wait a little
1876                                          * bit longer before breaking SWS
1877                                          * avoidance.
1878                                          *
1879                                          * What should the value be?
1880                                          * Suggestion: MAX(init rexmit time,
1881                                          * tcp->tcp_rto)
1882                                          */
1883                                         TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
1884                                 }
1885                                 break;  /* success; too small */
1886                         }
1887                 }
1888 
1889                 tcpha = tcp->tcp_tcpha;
1890 
1891                 /*
1892                  * The reason to adjust len here is that we need to set flags
1893                  * and calculate checksum.
1894                  */
1895                 if (do_lso_send)
1896                         len = lso_usable;
1897 
1898                 *usable -= len; /* Approximate - can be adjusted later */
1899                 if (*usable > 0)
1900                         tcpha->tha_flags = TH_ACK;
1901                 else
1902                         tcpha->tha_flags = (TH_ACK | TH_PUSH);
1903 
1904                 /*
1905                  * Prime pump for IP's checksumming on our behalf.
1906                  * Include the adjustment for a source route if any.
1907                  * In case of LSO, the partial pseudo-header checksum should
1908                  * exclusive TCP length, so zero tha_sum before IP calculate
1909                  * pseudo-header checksum for partial checksum offload.
1910                  */
1911                 if (do_lso_send) {
1912                         sum = 0;
1913                 } else {
1914                         sum = len + tcp_hdr_len + connp->conn_sum;
1915                         sum = (sum >> 16) + (sum & 0xFFFF);
1916                 }
1917                 tcpha->tha_sum = htons(sum);
1918                 tcpha->tha_seq = htonl(*snxt);
1919 
1920                 /*
1921                  * Branch off to tcp_xmit_mp() if any of the VALID bits is
1922                  * set.  For the case when TCP_FSS_VALID is the only valid
1923                  * bit (normal active close), branch off only when we think
1924                  * that the FIN flag needs to be set.  Note for this case,
1925                  * that (snxt + len) may not reflect the actual seg_len,
1926                  * as len may be further reduced in tcp_xmit_mp().  If len
1927                  * gets modified, we will end up here again.
1928                  */
1929                 if (tcp->tcp_valid_bits != 0 &&
1930                     (tcp->tcp_valid_bits != TCP_FSS_VALID ||
1931                     ((*snxt + len) == tcp->tcp_fss))) {
1932                         uchar_t         *prev_rptr;
1933                         uint32_t        prev_snxt = tcp->tcp_snxt;
1934 
1935                         if (*tail_unsent == 0) {
1936                                 ASSERT((*xmit_tail)->b_cont != NULL);
1937                                 *xmit_tail = (*xmit_tail)->b_cont;
1938                                 prev_rptr = (*xmit_tail)->b_rptr;
1939                                 *tail_unsent = (int)((*xmit_tail)->b_wptr -
1940                                     (*xmit_tail)->b_rptr);
1941                         } else {
1942                                 prev_rptr = (*xmit_tail)->b_rptr;
1943                                 (*xmit_tail)->b_rptr = (*xmit_tail)->b_wptr -
1944                                     *tail_unsent;
1945                         }
1946                         mp = tcp_xmit_mp(tcp, *xmit_tail, len, NULL, NULL,
1947                             *snxt, B_FALSE, (uint32_t *)&len, B_FALSE);
1948                         /* Restore tcp_snxt so we get amount sent right. */
1949                         tcp->tcp_snxt = prev_snxt;
1950                         if (prev_rptr == (*xmit_tail)->b_rptr) {
1951                                 /*
1952                                  * If the previous timestamp is still in use,
1953                                  * don't stomp on it.
1954                                  */
1955                                 if ((*xmit_tail)->b_next == NULL) {
1956                                         (*xmit_tail)->b_prev = local_time;
1957                                         (*xmit_tail)->b_next =
1958                                             (mblk_t *)(uintptr_t)(*snxt);
1959                                 }
1960                         } else
1961                                 (*xmit_tail)->b_rptr = prev_rptr;
1962 
1963                         if (mp == NULL) {
1964                                 return (-1);
1965                         }
1966                         mp1 = mp->b_cont;
1967 
1968                         if (len <= mss) /* LSO is unusable (!do_lso_send) */
1969                                 tcp->tcp_last_sent_len = (ushort_t)len;
1970                         while (mp1->b_cont) {
1971                                 *xmit_tail = (*xmit_tail)->b_cont;
1972                                 (*xmit_tail)->b_prev = local_time;
1973                                 (*xmit_tail)->b_next =
1974                                     (mblk_t *)(uintptr_t)(*snxt);
1975                                 mp1 = mp1->b_cont;
1976                         }
1977                         *snxt += len;
1978                         *tail_unsent = (*xmit_tail)->b_wptr - mp1->b_wptr;
1979                         TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1980                         TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1981                         TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1982                         tcp->tcp_cs.tcp_out_data_segs++;
1983                         tcp->tcp_cs.tcp_out_data_bytes += len;
1984                         tcp_send_data(tcp, mp);
1985                         continue;
1986                 }
1987 
1988                 *snxt += len;   /* Adjust later if we don't send all of len */
1989                 TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
1990                 TCPS_BUMP_MIB(tcps, tcpOutDataSegs);
1991                 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, len);
1992                 tcp->tcp_cs.tcp_out_data_segs++;
1993                 tcp->tcp_cs.tcp_out_data_bytes += len;
1994 
1995                 if (*tail_unsent) {
1996                         /* Are the bytes above us in flight? */
1997                         rptr = (*xmit_tail)->b_wptr - *tail_unsent;
1998                         if (rptr != (*xmit_tail)->b_rptr) {
1999                                 *tail_unsent -= len;
2000                                 if (len <= mss) /* LSO is unusable */
2001                                         tcp->tcp_last_sent_len = (ushort_t)len;
2002                                 len += total_hdr_len;
2003                                 ixa->ixa_pktlen = len;
2004 
2005                                 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2006                                         tcp->tcp_ipha->ipha_length = htons(len);
2007                                 } else {
2008                                         tcp->tcp_ip6h->ip6_plen =
2009                                             htons(len - IPV6_HDR_LEN);
2010                                 }
2011 
2012                                 mp = dupb(*xmit_tail);
2013                                 if (mp == NULL) {
2014                                         return (-1);    /* out_of_mem */
2015                                 }
2016                                 mp->b_rptr = rptr;
2017                                 /*
2018                                  * If the old timestamp is no longer in use,
2019                                  * sample a new timestamp now.
2020                                  */
2021                                 if ((*xmit_tail)->b_next == NULL) {
2022                                         (*xmit_tail)->b_prev = local_time;
2023                                         (*xmit_tail)->b_next =
2024                                             (mblk_t *)(uintptr_t)(*snxt-len);
2025                                 }
2026                                 goto must_alloc;
2027                         }
2028                 } else {
2029                         *xmit_tail = (*xmit_tail)->b_cont;
2030                         ASSERT((uintptr_t)((*xmit_tail)->b_wptr -
2031                             (*xmit_tail)->b_rptr) <= (uintptr_t)INT_MAX);
2032                         *tail_unsent = (int)((*xmit_tail)->b_wptr -
2033                             (*xmit_tail)->b_rptr);
2034                 }
2035 
2036                 (*xmit_tail)->b_prev = local_time;
2037                 (*xmit_tail)->b_next = (mblk_t *)(uintptr_t)(*snxt - len);
2038 
2039                 *tail_unsent -= len;
2040                 if (len <= mss) /* LSO is unusable (!do_lso_send) */
2041                         tcp->tcp_last_sent_len = (ushort_t)len;
2042 
2043                 len += total_hdr_len;
2044                 ixa->ixa_pktlen = len;
2045 
2046                 if (ixa->ixa_flags & IXAF_IS_IPV4) {
2047                         tcp->tcp_ipha->ipha_length = htons(len);
2048                 } else {
2049                         tcp->tcp_ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2050                 }
2051 
2052                 mp = dupb(*xmit_tail);
2053                 if (mp == NULL) {
2054                         return (-1);    /* out_of_mem */
2055                 }
2056 
2057                 len = total_hdr_len;
2058                 /*
2059                  * There are four reasons to allocate a new hdr mblk:
2060                  *  1) The bytes above us are in use by another packet
2061                  *  2) We don't have good alignment
2062                  *  3) The mblk is being shared
2063                  *  4) We don't have enough room for a header
2064                  */
2065                 rptr = mp->b_rptr - len;
2066                 if (!OK_32PTR(rptr) ||
2067                     ((db = mp->b_datap), db->db_ref != 2) ||
2068                     rptr < db->db_base) {
2069                         /* NOTE: we assume allocb returns an OK_32PTR */
2070 
2071                 must_alloc:;
2072                         mp1 = allocb(connp->conn_ht_iphc_allocated +
2073                             tcps->tcps_wroff_xtra, BPRI_MED);
2074                         if (mp1 == NULL) {
2075                                 freemsg(mp);
2076                                 return (-1);    /* out_of_mem */
2077                         }
2078                         mp1->b_cont = mp;
2079                         mp = mp1;
2080                         /* Leave room for Link Level header */
2081                         len = total_hdr_len;
2082                         rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2083                         mp->b_wptr = &rptr[len];
2084                 }
2085 
2086                 /*
2087                  * Fill in the header using the template header, and add
2088                  * options such as time-stamp, ECN and/or SACK, as needed.
2089                  */
2090                 tcp_fill_header(tcp, rptr, num_sack_blk);
2091 
2092                 mp->b_rptr = rptr;
2093 
2094                 if (*tail_unsent) {
2095                         int spill = *tail_unsent;
2096 
2097                         mp1 = mp->b_cont;
2098                         if (mp1 == NULL)
2099                                 mp1 = mp;
2100 
2101                         /*
2102                          * If we're a little short, tack on more mblks until
2103                          * there is no more spillover.
2104                          */
2105                         while (spill < 0) {
2106                                 mblk_t *nmp;
2107                                 int nmpsz;
2108 
2109                                 nmp = (*xmit_tail)->b_cont;
2110                                 nmpsz = MBLKL(nmp);
2111 
2112                                 /*
2113                                  * Excess data in mblk; can we split it?
2114                                  * If LSO is enabled for the connection,
2115                                  * keep on splitting as this is a transient
2116                                  * send path.
2117                                  */
2118                                 if (!do_lso_send && (spill + nmpsz > 0)) {
2119                                         /*
2120                                          * Don't split if stream head was
2121                                          * told to break up larger writes
2122                                          * into smaller ones.
2123                                          */
2124                                         if (tcp->tcp_maxpsz_multiplier > 0)
2125                                                 break;
2126 
2127                                         /*
2128                                          * Next mblk is less than SMSS/2
2129                                          * rounded up to nearest 64-byte;
2130                                          * let it get sent as part of the
2131                                          * next segment.
2132                                          */
2133                                         if (tcp->tcp_localnet &&
2134                                             !tcp->tcp_cork &&
2135                                             (nmpsz < roundup((mss >> 1), 64)))
2136                                                 break;
2137                                 }
2138 
2139                                 *xmit_tail = nmp;
2140                                 ASSERT((uintptr_t)nmpsz <= (uintptr_t)INT_MAX);
2141                                 /* Stash for rtt use later */
2142                                 (*xmit_tail)->b_prev = local_time;
2143                                 (*xmit_tail)->b_next =
2144                                     (mblk_t *)(uintptr_t)(*snxt - len);
2145                                 mp1->b_cont = dupb(*xmit_tail);
2146                                 mp1 = mp1->b_cont;
2147 
2148                                 spill += nmpsz;
2149                                 if (mp1 == NULL) {
2150                                         *tail_unsent = spill;
2151                                         freemsg(mp);
2152                                         return (-1);    /* out_of_mem */
2153                                 }
2154                         }
2155 
2156                         /* Trim back any surplus on the last mblk */
2157                         if (spill >= 0) {
2158                                 mp1->b_wptr -= spill;
2159                                 *tail_unsent = spill;
2160                         } else {
2161                                 /*
2162                                  * We did not send everything we could in
2163                                  * order to remain within the b_cont limit.
2164                                  */
2165                                 *usable -= spill;
2166                                 *snxt += spill;
2167                                 tcp->tcp_last_sent_len += spill;
2168                                 TCPS_UPDATE_MIB(tcps, tcpOutDataBytes, spill);
2169                                 tcp->tcp_cs.tcp_out_data_bytes += spill;
2170                                 /*
2171                                  * Adjust the checksum
2172                                  */
2173                                 tcpha = (tcpha_t *)(rptr +
2174                                     ixa->ixa_ip_hdr_length);
2175                                 sum += spill;
2176                                 sum = (sum >> 16) + (sum & 0xFFFF);
2177                                 tcpha->tha_sum = htons(sum);
2178                                 if (connp->conn_ipversion == IPV4_VERSION) {
2179                                         sum = ntohs(
2180                                             ((ipha_t *)rptr)->ipha_length) +
2181                                             spill;
2182                                         ((ipha_t *)rptr)->ipha_length =
2183                                             htons(sum);
2184                                 } else {
2185                                         sum = ntohs(
2186                                             ((ip6_t *)rptr)->ip6_plen) +
2187                                             spill;
2188                                         ((ip6_t *)rptr)->ip6_plen =
2189                                             htons(sum);
2190                                 }
2191                                 ixa->ixa_pktlen += spill;
2192                                 *tail_unsent = 0;
2193                         }
2194                 }
2195                 if (tcp->tcp_ip_forward_progress) {
2196                         tcp->tcp_ip_forward_progress = B_FALSE;
2197                         ixa->ixa_flags |= IXAF_REACH_CONF;
2198                 } else {
2199                         ixa->ixa_flags &= ~IXAF_REACH_CONF;
2200                 }
2201 
2202                 if (do_lso_send) {
2203                         /* Append LSO information to the mp. */
2204                         lso_info_set(mp, mss, HW_LSO);
2205                         ixa->ixa_fragsize = IP_MAXPACKET;
2206                         ixa->ixa_extra_ident = num_lso_seg - 1;
2207 
2208                         DTRACE_PROBE2(tcp_send_lso, int, num_lso_seg,
2209                             boolean_t, B_TRUE);
2210 
2211                         tcp_send_data(tcp, mp);
2212 
2213                         /*
2214                          * Restore values of ixa_fragsize and ixa_extra_ident.
2215                          */
2216                         ixa->ixa_fragsize = ixa->ixa_pmtu;
2217                         ixa->ixa_extra_ident = 0;
2218                         TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2219                         TCP_STAT(tcps, tcp_lso_times);
2220                         TCP_STAT_UPDATE(tcps, tcp_lso_pkt_out, num_lso_seg);
2221                 } else {
2222                         /*
2223                          * Make sure to clean up LSO information. Wherever a
2224                          * new mp uses the prepended header room after dupb(),
2225                          * lso_info_cleanup() should be called.
2226                          */
2227                         lso_info_cleanup(mp);
2228                         tcp_send_data(tcp, mp);
2229                         TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2230                 }
2231         }
2232 
2233         return (0);
2234 }
2235 
2236 /*
2237  * Initiate closedown sequence on an active connection.  (May be called as
2238  * writer.)  Return value zero for OK return, non-zero for error return.
2239  */
2240 static int
2241 tcp_xmit_end(tcp_t *tcp)
2242 {
2243         mblk_t          *mp;
2244         tcp_stack_t     *tcps = tcp->tcp_tcps;
2245         iulp_t          uinfo;
2246         ip_stack_t      *ipst = tcps->tcps_netstack->netstack_ip;
2247         conn_t          *connp = tcp->tcp_connp;
2248 
2249         if (tcp->tcp_state < TCPS_SYN_RCVD ||
2250             tcp->tcp_state > TCPS_CLOSE_WAIT) {
2251                 /*
2252                  * Invalid state, only states TCPS_SYN_RCVD,
2253                  * TCPS_ESTABLISHED and TCPS_CLOSE_WAIT are valid
2254                  */
2255                 return (-1);
2256         }
2257 
2258         tcp->tcp_fss = tcp->tcp_snxt + tcp->tcp_unsent;
2259         tcp->tcp_valid_bits |= TCP_FSS_VALID;
2260         /*
2261          * If there is nothing more unsent, send the FIN now.
2262          * Otherwise, it will go out with the last segment.
2263          */
2264         if (tcp->tcp_unsent == 0) {
2265                 mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
2266                     tcp->tcp_fss, B_FALSE, NULL, B_FALSE);
2267 
2268                 if (mp) {
2269                         tcp_send_data(tcp, mp);
2270                 } else {
2271                         /*
2272                          * Couldn't allocate msg.  Pretend we got it out.
2273                          * Wait for rexmit timeout.
2274                          */
2275                         tcp->tcp_snxt = tcp->tcp_fss + 1;
2276                         TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
2277                 }
2278 
2279                 /*
2280                  * If needed, update tcp_rexmit_snxt as tcp_snxt is
2281                  * changed.
2282                  */
2283                 if (tcp->tcp_rexmit && tcp->tcp_rexmit_nxt == tcp->tcp_fss) {
2284                         tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
2285                 }
2286         } else {
2287                 /*
2288                  * If tcp->tcp_cork is set, then the data will not get sent,
2289                  * so we have to check that and unset it first.
2290                  */
2291                 if (tcp->tcp_cork)
2292                         tcp->tcp_cork = B_FALSE;
2293                 tcp_wput_data(tcp, NULL, B_FALSE);
2294         }
2295 
2296         /*
2297          * If TCP does not get enough samples of RTT or tcp_rtt_updates
2298          * is 0, don't update the cache.
2299          */
2300         if (tcps->tcps_rtt_updates == 0 ||
2301             tcp->tcp_rtt_update < tcps->tcps_rtt_updates)
2302                 return (0);
2303 
2304         /*
2305          * We do not have a good algorithm to update ssthresh at this time.
2306          * So don't do any update.
2307          */
2308         bzero(&uinfo, sizeof (uinfo));
2309         uinfo.iulp_rtt = NSEC2MSEC(tcp->tcp_rtt_sa);
2310         uinfo.iulp_rtt_sd = NSEC2MSEC(tcp->tcp_rtt_sd);
2311 
2312         /*
2313          * Note that uinfo is kept for conn_faddr in the DCE. Could update even
2314          * if source routed but we don't.
2315          */
2316         if (connp->conn_ipversion == IPV4_VERSION) {
2317                 if (connp->conn_faddr_v4 !=  tcp->tcp_ipha->ipha_dst) {
2318                         return (0);
2319                 }
2320                 (void) dce_update_uinfo_v4(connp->conn_faddr_v4, &uinfo, ipst);
2321         } else {
2322                 uint_t ifindex;
2323 
2324                 if (!(IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6,
2325                     &tcp->tcp_ip6h->ip6_dst))) {
2326                         return (0);
2327                 }
2328                 ifindex = 0;
2329                 if (IN6_IS_ADDR_LINKSCOPE(&connp->conn_faddr_v6)) {
2330                         ip_xmit_attr_t *ixa = connp->conn_ixa;
2331 
2332                         /*
2333                          * If we are going to create a DCE we'd better have
2334                          * an ifindex
2335                          */
2336                         if (ixa->ixa_nce != NULL) {
2337                                 ifindex = ixa->ixa_nce->nce_common->ncec_ill->
2338                                     ill_phyint->phyint_ifindex;
2339                         } else {
2340                                 return (0);
2341                         }
2342                 }
2343 
2344                 (void) dce_update_uinfo(&connp->conn_faddr_v6, ifindex, &uinfo,
2345                     ipst);
2346         }
2347         return (0);
2348 }
2349 
2350 /*
2351  * Send out a control packet on the tcp connection specified.  This routine
2352  * is typically called where we need a simple ACK or RST generated.
2353  */
2354 void
2355 tcp_xmit_ctl(char *str, tcp_t *tcp, uint32_t seq, uint32_t ack, int ctl)
2356 {
2357         uchar_t         *rptr;
2358         tcpha_t         *tcpha;
2359         ipha_t          *ipha = NULL;
2360         ip6_t           *ip6h = NULL;
2361         uint32_t        sum;
2362         int             total_hdr_len;
2363         int             ip_hdr_len;
2364         mblk_t          *mp;
2365         tcp_stack_t     *tcps = tcp->tcp_tcps;
2366         conn_t          *connp = tcp->tcp_connp;
2367         ip_xmit_attr_t  *ixa = connp->conn_ixa;
2368 
2369         /*
2370          * Save sum for use in source route later.
2371          */
2372         sum = connp->conn_ht_ulp_len + connp->conn_sum;
2373         total_hdr_len = connp->conn_ht_iphc_len;
2374         ip_hdr_len = ixa->ixa_ip_hdr_length;
2375 
2376         /* If a text string is passed in with the request, pass it to strlog. */
2377         if (str != NULL && connp->conn_debug) {
2378                 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2379                     "tcp_xmit_ctl: '%s', seq 0x%x, ack 0x%x, ctl 0x%x",
2380                     str, seq, ack, ctl);
2381         }
2382         mp = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
2383             BPRI_MED);
2384         if (mp == NULL) {
2385                 return;
2386         }
2387         rptr = &mp->b_rptr[tcps->tcps_wroff_xtra];
2388         mp->b_rptr = rptr;
2389         mp->b_wptr = &rptr[total_hdr_len];
2390         bcopy(connp->conn_ht_iphc, rptr, total_hdr_len);
2391 
2392         ixa->ixa_pktlen = total_hdr_len;
2393 
2394         if (ixa->ixa_flags & IXAF_IS_IPV4) {
2395                 ipha = (ipha_t *)rptr;
2396                 ipha->ipha_length = htons(total_hdr_len);
2397         } else {
2398                 ip6h = (ip6_t *)rptr;
2399                 ip6h->ip6_plen = htons(total_hdr_len - IPV6_HDR_LEN);
2400         }
2401         tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2402         tcpha->tha_flags = (uint8_t)ctl;
2403         if (ctl & TH_RST) {
2404                 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2405                 TCPS_BUMP_MIB(tcps, tcpOutControl);
2406                 /*
2407                  * Don't send TSopt w/ TH_RST packets per RFC 1323.
2408                  */
2409                 if (tcp->tcp_snd_ts_ok &&
2410                     tcp->tcp_state > TCPS_SYN_SENT) {
2411                         mp->b_wptr = &rptr[total_hdr_len - TCPOPT_REAL_TS_LEN];
2412                         *(mp->b_wptr) = TCPOPT_EOL;
2413 
2414                         ixa->ixa_pktlen = total_hdr_len - TCPOPT_REAL_TS_LEN;
2415 
2416                         if (connp->conn_ipversion == IPV4_VERSION) {
2417                                 ipha->ipha_length = htons(total_hdr_len -
2418                                     TCPOPT_REAL_TS_LEN);
2419                         } else {
2420                                 ip6h->ip6_plen = htons(total_hdr_len -
2421                                     IPV6_HDR_LEN - TCPOPT_REAL_TS_LEN);
2422                         }
2423                         tcpha->tha_offset_and_reserved -= (3 << 4);
2424                         sum -= TCPOPT_REAL_TS_LEN;
2425                 }
2426         }
2427         if (ctl & TH_ACK) {
2428                 if (tcp->tcp_snd_ts_ok) {
2429                         uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2430 
2431                         U32_TO_BE32(llbolt,
2432                             (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
2433                         U32_TO_BE32(tcp->tcp_ts_recent,
2434                             (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
2435                 }
2436 
2437                 /* Update the latest receive window size in TCP header. */
2438                 tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
2439                 /* Track what we sent to the peer */
2440                 tcp->tcp_tcpha->tha_win = tcpha->tha_win;
2441                 tcp->tcp_rack = ack;
2442                 tcp->tcp_rack_cnt = 0;
2443                 TCPS_BUMP_MIB(tcps, tcpOutAck);
2444         }
2445         TCPS_BUMP_MIB(tcps, tcpHCOutSegs);
2446         tcpha->tha_seq = htonl(seq);
2447         tcpha->tha_ack = htonl(ack);
2448         /*
2449          * Include the adjustment for a source route if any.
2450          */
2451         sum = (sum >> 16) + (sum & 0xFFFF);
2452         tcpha->tha_sum = htons(sum);
2453         tcp_send_data(tcp, mp);
2454 }
2455 
2456 /*
2457  * Generate a reset based on an inbound packet, connp is set by caller
2458  * when RST is in response to an unexpected inbound packet for which
2459  * there is active tcp state in the system.
2460  *
2461  * IPSEC NOTE : Try to send the reply with the same protection as it came
2462  * in.  We have the ip_recv_attr_t which is reversed to form the ip_xmit_attr_t.
2463  * That way the packet will go out at the same level of protection as it
2464  * came in with.
2465  */
2466 static void
2467 tcp_xmit_early_reset(char *str, mblk_t *mp, uint32_t seq, uint32_t ack, int ctl,
2468     ip_recv_attr_t *ira, ip_stack_t *ipst, conn_t *connp)
2469 {
2470         ipha_t          *ipha = NULL;
2471         ip6_t           *ip6h = NULL;
2472         ushort_t        len;
2473         tcpha_t         *tcpha;
2474         int             i;
2475         ipaddr_t        v4addr;
2476         in6_addr_t      v6addr;
2477         netstack_t      *ns = ipst->ips_netstack;
2478         tcp_stack_t     *tcps = ns->netstack_tcp;
2479         ip_xmit_attr_t  ixas, *ixa;
2480         uint_t          ip_hdr_len = ira->ira_ip_hdr_length;
2481         boolean_t       need_refrele = B_FALSE;         /* ixa_refrele(ixa) */
2482         ushort_t        port;
2483 
2484         if (!tcp_send_rst_chk(tcps)) {
2485                 TCP_STAT(tcps, tcp_rst_unsent);
2486                 freemsg(mp);
2487                 return;
2488         }
2489 
2490         /*
2491          * If connp != NULL we use conn_ixa to keep IP_NEXTHOP and other
2492          * options from the listener. In that case the caller must ensure that
2493          * we are running on the listener = connp squeue.
2494          *
2495          * We get a safe copy of conn_ixa so we don't need to restore anything
2496          * we or ip_output_simple might change in the ixa.
2497          */
2498         if (connp != NULL) {
2499                 ASSERT(connp->conn_on_sqp);
2500 
2501                 ixa = conn_get_ixa_exclusive(connp);
2502                 if (ixa == NULL) {
2503                         TCP_STAT(tcps, tcp_rst_unsent);
2504                         freemsg(mp);
2505                         return;
2506                 }
2507                 need_refrele = B_TRUE;
2508         } else {
2509                 bzero(&ixas, sizeof (ixas));
2510                 ixa = &ixas;
2511                 /*
2512                  * IXAF_VERIFY_SOURCE is overkill since we know the
2513                  * packet was for us.
2514                  */
2515                 ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE;
2516                 ixa->ixa_protocol = IPPROTO_TCP;
2517                 ixa->ixa_zoneid = ira->ira_zoneid;
2518                 ixa->ixa_ifindex = 0;
2519                 ixa->ixa_ipst = ipst;
2520                 ixa->ixa_cred = kcred;
2521                 ixa->ixa_cpid = NOPID;
2522         }
2523 
2524         if (str && tcps->tcps_dbg) {
2525                 (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE,
2526                     "tcp_xmit_early_reset: '%s', seq 0x%x, ack 0x%x, "
2527                     "flags 0x%x",
2528                     str, seq, ack, ctl);
2529         }
2530         if (mp->b_datap->db_ref != 1) {
2531                 mblk_t *mp1 = copyb(mp);
2532                 freemsg(mp);
2533                 mp = mp1;
2534                 if (mp == NULL)
2535                         goto done;
2536         } else if (mp->b_cont) {
2537                 freemsg(mp->b_cont);
2538                 mp->b_cont = NULL;
2539                 DB_CKSUMFLAGS(mp) = 0;
2540         }
2541         /*
2542          * We skip reversing source route here.
2543          * (for now we replace all IP options with EOL)
2544          */
2545         if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2546                 ipha = (ipha_t *)mp->b_rptr;
2547                 for (i = IP_SIMPLE_HDR_LENGTH; i < (int)ip_hdr_len; i++)
2548                         mp->b_rptr[i] = IPOPT_EOL;
2549                 /*
2550                  * Make sure that src address isn't flagrantly invalid.
2551                  * Not all broadcast address checking for the src address
2552                  * is possible, since we don't know the netmask of the src
2553                  * addr.  No check for destination address is done, since
2554                  * IP will not pass up a packet with a broadcast dest
2555                  * address to TCP.  Similar checks are done below for IPv6.
2556                  */
2557                 if (ipha->ipha_src == 0 || ipha->ipha_src == INADDR_BROADCAST ||
2558                     CLASSD(ipha->ipha_src)) {
2559                         BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsInDiscards);
2560                         ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2561                         freemsg(mp);
2562                         goto done;
2563                 }
2564         } else {
2565                 ip6h = (ip6_t *)mp->b_rptr;
2566 
2567                 if (IN6_IS_ADDR_UNSPECIFIED(&ip6h->ip6_src) ||
2568                     IN6_IS_ADDR_MULTICAST(&ip6h->ip6_src)) {
2569                         BUMP_MIB(&ipst->ips_ip6_mib, ipIfStatsInDiscards);
2570                         ip_drop_input("ipIfStatsInDiscards", mp, NULL);
2571                         freemsg(mp);
2572                         goto done;
2573                 }
2574 
2575                 /* Remove any extension headers assuming partial overlay */
2576                 if (ip_hdr_len > IPV6_HDR_LEN) {
2577                         uint8_t *to;
2578 
2579                         to = mp->b_rptr + ip_hdr_len - IPV6_HDR_LEN;
2580                         ovbcopy(ip6h, to, IPV6_HDR_LEN);
2581                         mp->b_rptr += ip_hdr_len - IPV6_HDR_LEN;
2582                         ip_hdr_len = IPV6_HDR_LEN;
2583                         ip6h = (ip6_t *)mp->b_rptr;
2584                         ip6h->ip6_nxt = IPPROTO_TCP;
2585                 }
2586         }
2587         tcpha = (tcpha_t *)&mp->b_rptr[ip_hdr_len];
2588         if (tcpha->tha_flags & TH_RST) {
2589                 freemsg(mp);
2590                 goto done;
2591         }
2592         tcpha->tha_offset_and_reserved = (5 << 4);
2593         len = ip_hdr_len + sizeof (tcpha_t);
2594         mp->b_wptr = &mp->b_rptr[len];
2595         if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2596                 ipha->ipha_length = htons(len);
2597                 /* Swap addresses */
2598                 v4addr = ipha->ipha_src;
2599                 ipha->ipha_src = ipha->ipha_dst;
2600                 ipha->ipha_dst = v4addr;
2601                 ipha->ipha_ident = 0;
2602                 ipha->ipha_ttl = (uchar_t)tcps->tcps_ipv4_ttl;
2603                 ixa->ixa_flags |= IXAF_IS_IPV4;
2604                 ixa->ixa_ip_hdr_length = ip_hdr_len;
2605         } else {
2606                 ip6h->ip6_plen = htons(len - IPV6_HDR_LEN);
2607                 /* Swap addresses */
2608                 v6addr = ip6h->ip6_src;
2609                 ip6h->ip6_src = ip6h->ip6_dst;
2610                 ip6h->ip6_dst = v6addr;
2611                 ip6h->ip6_hops = (uchar_t)tcps->tcps_ipv6_hoplimit;
2612                 ixa->ixa_flags &= ~IXAF_IS_IPV4;
2613 
2614                 if (IN6_IS_ADDR_LINKSCOPE(&ip6h->ip6_dst)) {
2615                         ixa->ixa_flags |= IXAF_SCOPEID_SET;
2616                         ixa->ixa_scopeid = ira->ira_ruifindex;
2617                 }
2618                 ixa->ixa_ip_hdr_length = IPV6_HDR_LEN;
2619         }
2620         ixa->ixa_pktlen = len;
2621 
2622         /* Swap the ports */
2623         port = tcpha->tha_fport;
2624         tcpha->tha_fport = tcpha->tha_lport;
2625         tcpha->tha_lport = port;
2626 
2627         tcpha->tha_ack = htonl(ack);
2628         tcpha->tha_seq = htonl(seq);
2629         tcpha->tha_win = 0;
2630         tcpha->tha_sum = htons(sizeof (tcpha_t));
2631         tcpha->tha_flags = (uint8_t)ctl;
2632         if (ctl & TH_RST) {
2633                 if (ctl & TH_ACK) {
2634                         /*
2635                          * Probe connection rejection here.
2636                          * tcp_xmit_listeners_reset() drops non-SYN segments
2637                          * that do not specify TH_ACK in their flags without
2638                          * calling this function.  As a consequence, if this
2639                          * function is called with a TH_RST|TH_ACK ctl argument,
2640                          * it is being called in response to a SYN segment
2641                          * and thus the tcp:::accept-refused probe point
2642                          * is valid here.
2643                          */
2644                         DTRACE_TCP5(accept__refused, mblk_t *, NULL,
2645                             void, NULL, void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2646                             tcph_t *, tcpha);
2647                 }
2648                 TCPS_BUMP_MIB(tcps, tcpOutRsts);
2649                 TCPS_BUMP_MIB(tcps, tcpOutControl);
2650         }
2651 
2652         /* Discard any old label */
2653         if (ixa->ixa_free_flags & IXA_FREE_TSL) {
2654                 ASSERT(ixa->ixa_tsl != NULL);
2655                 label_rele(ixa->ixa_tsl);
2656                 ixa->ixa_free_flags &= ~IXA_FREE_TSL;
2657         }
2658         ixa->ixa_tsl = ira->ira_tsl;      /* Behave as a multi-level responder */
2659 
2660         if (ira->ira_flags & IRAF_IPSEC_SECURE) {
2661                 /*
2662                  * Apply IPsec based on how IPsec was applied to
2663                  * the packet that caused the RST.
2664                  */
2665                 if (!ipsec_in_to_out(ira, ixa, mp, ipha, ip6h)) {
2666                         BUMP_MIB(&ipst->ips_ip_mib, ipIfStatsOutDiscards);
2667                         /* Note: mp already consumed and ip_drop_packet done */
2668                         goto done;
2669                 }
2670         } else {
2671                 /*
2672                  * This is in clear. The RST message we are building
2673                  * here should go out in clear, independent of our policy.
2674                  */
2675                 ixa->ixa_flags |= IXAF_NO_IPSEC;
2676         }
2677 
2678         DTRACE_TCP5(send, mblk_t *, NULL, ip_xmit_attr_t *, ixa,
2679             __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2680             __dtrace_tcp_tcph_t *, tcpha);
2681 
2682         /*
2683          * NOTE:  one might consider tracing a TCP packet here, but
2684          * this function has no active TCP state and no tcp structure
2685          * that has a trace buffer.  If we traced here, we would have
2686          * to keep a local trace buffer in tcp_record_trace().
2687          */
2688 
2689         (void) ip_output_simple(mp, ixa);
2690 done:
2691         ixa_cleanup(ixa);
2692         if (need_refrele) {
2693                 ASSERT(ixa != &ixas);
2694                 ixa_refrele(ixa);
2695         }
2696 }
2697 
2698 /*
2699  * Generate a "no listener here" RST in response to an "unknown" segment.
2700  * connp is set by caller when RST is in response to an unexpected
2701  * inbound packet for which there is active tcp state in the system.
2702  * Note that we are reusing the incoming mp to construct the outgoing RST.
2703  */
2704 void
2705 tcp_xmit_listeners_reset(mblk_t *mp, ip_recv_attr_t *ira, ip_stack_t *ipst,
2706     conn_t *connp)
2707 {
2708         uchar_t         *rptr;
2709         uint32_t        seg_len;
2710         tcpha_t         *tcpha;
2711         uint32_t        seg_seq;
2712         uint32_t        seg_ack;
2713         uint_t          flags;
2714         ipha_t          *ipha;
2715         ip6_t           *ip6h;
2716         boolean_t       policy_present;
2717         netstack_t      *ns = ipst->ips_netstack;
2718         tcp_stack_t     *tcps = ns->netstack_tcp;
2719         ipsec_stack_t   *ipss = tcps->tcps_netstack->netstack_ipsec;
2720         uint_t          ip_hdr_len = ira->ira_ip_hdr_length;
2721 
2722         TCP_STAT(tcps, tcp_no_listener);
2723 
2724         /*
2725          * DTrace this "unknown" segment as a tcp:::receive, as we did
2726          * just receive something that was TCP.
2727          */
2728         DTRACE_TCP5(receive, mblk_t *, NULL, ip_xmit_attr_t *, NULL,
2729             __dtrace_tcp_void_ip_t *, mp->b_rptr, tcp_t *, NULL,
2730             __dtrace_tcp_tcph_t *, &mp->b_rptr[ip_hdr_len]);
2731 
2732         if (IPH_HDR_VERSION(mp->b_rptr) == IPV4_VERSION) {
2733                 policy_present = ipss->ipsec_inbound_v4_policy_present;
2734                 ipha = (ipha_t *)mp->b_rptr;
2735                 ip6h = NULL;
2736         } else {
2737                 policy_present = ipss->ipsec_inbound_v6_policy_present;
2738                 ipha = NULL;
2739                 ip6h = (ip6_t *)mp->b_rptr;
2740         }
2741 
2742         if (policy_present) {
2743                 /*
2744                  * The conn_t parameter is NULL because we already know
2745                  * nobody's home.
2746                  */
2747                 mp = ipsec_check_global_policy(mp, (conn_t *)NULL, ipha, ip6h,
2748                     ira, ns);
2749                 if (mp == NULL)
2750                         return;
2751         }
2752         if (is_system_labeled() && !tsol_can_reply_error(mp, ira)) {
2753                 DTRACE_PROBE2(
2754                     tx__ip__log__error__nolistener__tcp,
2755                     char *, "Could not reply with RST to mp(1)",
2756                     mblk_t *, mp);
2757                 ip2dbg(("tcp_xmit_listeners_reset: not permitted to reply\n"));
2758                 freemsg(mp);
2759                 return;
2760         }
2761 
2762         rptr = mp->b_rptr;
2763 
2764         tcpha = (tcpha_t *)&rptr[ip_hdr_len];
2765         seg_seq = ntohl(tcpha->tha_seq);
2766         seg_ack = ntohl(tcpha->tha_ack);
2767         flags = tcpha->tha_flags;
2768 
2769         seg_len = msgdsize(mp) - (TCP_HDR_LENGTH(tcpha) + ip_hdr_len);
2770         if (flags & TH_RST) {
2771                 freemsg(mp);
2772         } else if (flags & TH_ACK) {
2773                 tcp_xmit_early_reset("no tcp, reset", mp, seg_ack, 0, TH_RST,
2774                     ira, ipst, connp);
2775         } else {
2776                 if (flags & TH_SYN) {
2777                         seg_len++;
2778                 } else {
2779                         /*
2780                          * Here we violate the RFC.  Note that a normal
2781                          * TCP will never send a segment without the ACK
2782                          * flag, except for RST or SYN segment.  This
2783                          * segment is neither.  Just drop it on the
2784                          * floor.
2785                          */
2786                         freemsg(mp);
2787                         TCP_STAT(tcps, tcp_rst_unsent);
2788                         return;
2789                 }
2790 
2791                 tcp_xmit_early_reset("no tcp, reset/ack", mp, 0,
2792                     seg_seq + seg_len, TH_RST | TH_ACK, ira, ipst, connp);
2793         }
2794 }
2795 
2796 /*
2797  * Helper function for tcp_xmit_mp() in handling connection set up flag
2798  * options setting.
2799  */
2800 static void
2801 tcp_xmit_mp_aux_iss(tcp_t *tcp, conn_t *connp, tcpha_t *tcpha, mblk_t *mp,
2802     uint_t *flags)
2803 {
2804         uint32_t u1;
2805         uint8_t *wptr = mp->b_wptr;
2806         tcp_stack_t *tcps = tcp->tcp_tcps;
2807         boolean_t add_sack = B_FALSE;
2808 
2809         /*
2810          * If TCP_ISS_VALID and the seq number is tcp_iss,
2811          * TCP can only be in SYN-SENT, SYN-RCVD or
2812          * FIN-WAIT-1 state.  It can be FIN-WAIT-1 if
2813          * our SYN is not ack'ed but the app closes this
2814          * TCP connection.
2815          */
2816         ASSERT(tcp->tcp_state == TCPS_SYN_SENT ||
2817             tcp->tcp_state == TCPS_SYN_RCVD ||
2818             tcp->tcp_state == TCPS_FIN_WAIT_1);
2819 
2820         /*
2821          * Tack on the MSS option.  It is always needed
2822          * for both active and passive open.
2823          *
2824          * MSS option value should be interface MTU - MIN
2825          * TCP/IP header according to RFC 793 as it means
2826          * the maximum segment size TCP can receive.  But
2827          * to get around some broken middle boxes/end hosts
2828          * out there, we allow the option value to be the
2829          * same as the MSS option size on the peer side.
2830          * In this way, the other side will not send
2831          * anything larger than they can receive.
2832          *
2833          * Note that for SYN_SENT state, the ndd param
2834          * tcp_use_smss_as_mss_opt has no effect as we
2835          * don't know the peer's MSS option value. So
2836          * the only case we need to take care of is in
2837          * SYN_RCVD state, which is done later.
2838          */
2839         wptr[0] = TCPOPT_MAXSEG;
2840         wptr[1] = TCPOPT_MAXSEG_LEN;
2841         wptr += 2;
2842         u1 = tcp->tcp_initial_pmtu - (connp->conn_ipversion == IPV4_VERSION ?
2843             IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN) - TCP_MIN_HEADER_LENGTH;
2844         U16_TO_BE16(u1, wptr);
2845         wptr += 2;
2846 
2847         /* Update the offset to cover the additional word */
2848         tcpha->tha_offset_and_reserved += (1 << 4);
2849 
2850         switch (tcp->tcp_state) {
2851         case TCPS_SYN_SENT:
2852                 *flags = TH_SYN;
2853 
2854                 if (tcp->tcp_snd_sack_ok)
2855                         add_sack = B_TRUE;
2856 
2857                 if (tcp->tcp_snd_ts_ok) {
2858                         uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
2859 
2860                         if (add_sack) {
2861                                 wptr[0] = TCPOPT_SACK_PERMITTED;
2862                                 wptr[1] = TCPOPT_SACK_OK_LEN;
2863                                 add_sack = B_FALSE;
2864                         } else {
2865                                 wptr[0] = TCPOPT_NOP;
2866                                 wptr[1] = TCPOPT_NOP;
2867                         }
2868                         wptr[2] = TCPOPT_TSTAMP;
2869                         wptr[3] = TCPOPT_TSTAMP_LEN;
2870                         wptr += 4;
2871                         U32_TO_BE32(llbolt, wptr);
2872                         wptr += 4;
2873                         ASSERT(tcp->tcp_ts_recent == 0);
2874                         U32_TO_BE32(0L, wptr);
2875                         wptr += 4;
2876                         tcpha->tha_offset_and_reserved += (3 << 4);
2877                 }
2878 
2879                 /*
2880                  * Set up all the bits to tell other side
2881                  * we are ECN capable.
2882                  */
2883                 if (tcp->tcp_ecn_ok)
2884                         *flags |= (TH_ECE | TH_CWR);
2885 
2886                 break;
2887 
2888         case TCPS_SYN_RCVD:
2889                 *flags |= TH_SYN;
2890 
2891                 /*
2892                  * Reset the MSS option value to be SMSS
2893                  * We should probably add back the bytes
2894                  * for timestamp option and IPsec.  We
2895                  * don't do that as this is a workaround
2896                  * for broken middle boxes/end hosts, it
2897                  * is better for us to be more cautious.
2898                  * They may not take these things into
2899                  * account in their SMSS calculation.  Thus
2900                  * the peer's calculated SMSS may be smaller
2901                  * than what it can be.  This should be OK.
2902                  */
2903                 if (tcps->tcps_use_smss_as_mss_opt) {
2904                         u1 = tcp->tcp_mss;
2905                         /*
2906                          * Note that wptr points just past the MSS
2907                          * option value.
2908                          */
2909                         U16_TO_BE16(u1, wptr - 2);
2910                 }
2911 
2912                 /*
2913                  * tcp_snd_ts_ok can only be set in TCPS_SYN_RCVD
2914                  * when the peer also uses timestamps option.  And
2915                  * the TCP header template must have already been
2916                  * updated to include the timestamps option.
2917                  */
2918                 if (tcp->tcp_snd_sack_ok) {
2919                         if (tcp->tcp_snd_ts_ok) {
2920                                 uint8_t *tmp_wptr;
2921 
2922                                 /*
2923                                  * Use the NOP in the header just
2924                                  * before timestamps opton.
2925                                  */
2926                                 tmp_wptr = (uint8_t *)tcpha +
2927                                     TCP_MIN_HEADER_LENGTH;
2928                                 ASSERT(tmp_wptr[0] == TCPOPT_NOP &&
2929                                     tmp_wptr[1] == TCPOPT_NOP);
2930                                 tmp_wptr[0] = TCPOPT_SACK_PERMITTED;
2931                                 tmp_wptr[1] = TCPOPT_SACK_OK_LEN;
2932                         } else {
2933                                 add_sack = B_TRUE;
2934                         }
2935                 }
2936 
2937 
2938                 /*
2939                  * If the other side is ECN capable, reply
2940                  * that we are also ECN capable.
2941                  */
2942                 if (tcp->tcp_ecn_ok)
2943                         *flags |= TH_ECE;
2944                 break;
2945 
2946         default:
2947                 /*
2948                  * The above ASSERT() makes sure that this
2949                  * must be FIN-WAIT-1 state.  Our SYN has
2950                  * not been ack'ed so retransmit it.
2951                  */
2952                 *flags |= TH_SYN;
2953                 break;
2954         }
2955 
2956         if (add_sack) {
2957                 wptr[0] = TCPOPT_NOP;
2958                 wptr[1] = TCPOPT_NOP;
2959                 wptr[2] = TCPOPT_SACK_PERMITTED;
2960                 wptr[3] = TCPOPT_SACK_OK_LEN;
2961                 wptr += TCPOPT_REAL_SACK_OK_LEN;
2962                 tcpha->tha_offset_and_reserved += (1 << 4);
2963         }
2964 
2965         if (tcp->tcp_snd_ws_ok) {
2966                 wptr[0] =  TCPOPT_NOP;
2967                 wptr[1] =  TCPOPT_WSCALE;
2968                 wptr[2] =  TCPOPT_WS_LEN;
2969                 wptr[3] = (uchar_t)tcp->tcp_rcv_ws;
2970                 wptr += TCPOPT_REAL_WS_LEN;
2971                 tcpha->tha_offset_and_reserved += (1 << 4);
2972         }
2973 
2974         mp->b_wptr = wptr;
2975         u1 = (int)(mp->b_wptr - mp->b_rptr);
2976         /*
2977          * Get IP set to checksum on our behalf
2978          * Include the adjustment for a source route if any.
2979          */
2980         u1 += connp->conn_sum;
2981         u1 = (u1 >> 16) + (u1 & 0xFFFF);
2982         tcpha->tha_sum = htons(u1);
2983         TCPS_BUMP_MIB(tcps, tcpOutControl);
2984 }
2985 
2986 /*
2987  * Helper function for tcp_xmit_mp() in handling connection tear down
2988  * flag setting and state changes.
2989  */
2990 static void
2991 tcp_xmit_mp_aux_fss(tcp_t *tcp, ip_xmit_attr_t *ixa, uint_t *flags)
2992 {
2993         if (!tcp->tcp_fin_acked) {
2994                 *flags |= TH_FIN;
2995                 TCPS_BUMP_MIB(tcp->tcp_tcps, tcpOutControl);
2996         }
2997         if (!tcp->tcp_fin_sent) {
2998                 tcp->tcp_fin_sent = B_TRUE;
2999                 switch (tcp->tcp_state) {
3000                 case TCPS_SYN_RCVD:
3001                         tcp->tcp_state = TCPS_FIN_WAIT_1;
3002                         DTRACE_TCP6(state__change, void, NULL,
3003                             ip_xmit_attr_t *, ixa, void, NULL,
3004                             tcp_t *, tcp, void, NULL,
3005                             int32_t, TCPS_SYN_RCVD);
3006                         break;
3007                 case TCPS_ESTABLISHED:
3008                         tcp->tcp_state = TCPS_FIN_WAIT_1;
3009                         DTRACE_TCP6(state__change, void, NULL,
3010                             ip_xmit_attr_t *, ixa, void, NULL,
3011                             tcp_t *, tcp, void, NULL,
3012                             int32_t, TCPS_ESTABLISHED);
3013                         break;
3014                 case TCPS_CLOSE_WAIT:
3015                         tcp->tcp_state = TCPS_LAST_ACK;
3016                         DTRACE_TCP6(state__change, void, NULL,
3017                             ip_xmit_attr_t *, ixa, void, NULL,
3018                             tcp_t *, tcp, void, NULL,
3019                             int32_t, TCPS_CLOSE_WAIT);
3020                         break;
3021                 }
3022                 if (tcp->tcp_suna == tcp->tcp_snxt)
3023                         TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3024                 tcp->tcp_snxt = tcp->tcp_fss + 1;
3025         }
3026 }
3027 
3028 /*
3029  * tcp_xmit_mp is called to return a pointer to an mblk chain complete with
3030  * ip and tcp header ready to pass down to IP.  If the mp passed in is
3031  * non-NULL, then up to max_to_send bytes of data will be dup'ed off that
3032  * mblk. (If sendall is not set the dup'ing will stop at an mblk boundary
3033  * otherwise it will dup partial mblks.)
3034  * Otherwise, an appropriate ACK packet will be generated.  This
3035  * routine is not usually called to send new data for the first time.  It
3036  * is mostly called out of the timer for retransmits, and to generate ACKs.
3037  *
3038  * If offset is not NULL, the returned mblk chain's first mblk's b_rptr will
3039  * be adjusted by *offset.  And after dupb(), the offset and the ending mblk
3040  * of the original mblk chain will be returned in *offset and *end_mp.
3041  */
3042 mblk_t *
3043 tcp_xmit_mp(tcp_t *tcp, mblk_t *mp, int32_t max_to_send, int32_t *offset,
3044     mblk_t **end_mp, uint32_t seq, boolean_t sendall, uint32_t *seg_len,
3045     boolean_t rexmit)
3046 {
3047         int     data_length;
3048         int32_t off = 0;
3049         uint_t  flags;
3050         mblk_t  *mp1;
3051         mblk_t  *mp2;
3052         uchar_t *rptr;
3053         tcpha_t *tcpha;
3054         int32_t num_sack_blk = 0;
3055         int32_t sack_opt_len = 0;
3056         tcp_stack_t     *tcps = tcp->tcp_tcps;
3057         conn_t          *connp = tcp->tcp_connp;
3058         ip_xmit_attr_t  *ixa = connp->conn_ixa;
3059 
3060         /* Allocate for our maximum TCP header + link-level */
3061         mp1 = allocb(connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra,
3062             BPRI_MED);
3063         if (mp1 == NULL)
3064                 return (NULL);
3065         data_length = 0;
3066 
3067         /*
3068          * Note that tcp_mss has been adjusted to take into account the
3069          * timestamp option if applicable.  Because SACK options do not
3070          * appear in every TCP segments and they are of variable lengths,
3071          * they cannot be included in tcp_mss.  Thus we need to calculate
3072          * the actual segment length when we need to send a segment which
3073          * includes SACK options.
3074          */
3075         if (tcp->tcp_snd_sack_ok && tcp->tcp_num_sack_blk > 0) {
3076                 num_sack_blk = MIN(tcp->tcp_max_sack_blk,
3077                     tcp->tcp_num_sack_blk);
3078                 sack_opt_len = num_sack_blk * sizeof (sack_blk_t) +
3079                     TCPOPT_NOP_LEN * 2 + TCPOPT_HEADER_LEN;
3080                 if (max_to_send + sack_opt_len > tcp->tcp_mss)
3081                         max_to_send -= sack_opt_len;
3082         }
3083 
3084         if (offset != NULL) {
3085                 off = *offset;
3086                 /* We use offset as an indicator that end_mp is not NULL. */
3087                 *end_mp = NULL;
3088         }
3089         for (mp2 = mp1; mp && data_length != max_to_send; mp = mp->b_cont) {
3090                 /* This could be faster with cooperation from downstream */
3091                 if (mp2 != mp1 && !sendall &&
3092                     data_length + (int)(mp->b_wptr - mp->b_rptr) >
3093                     max_to_send)
3094                         /*
3095                          * Don't send the next mblk since the whole mblk
3096                          * does not fit.
3097                          */
3098                         break;
3099                 mp2->b_cont = dupb(mp);
3100                 mp2 = mp2->b_cont;
3101                 if (!mp2) {
3102                         freemsg(mp1);
3103                         return (NULL);
3104                 }
3105                 mp2->b_rptr += off;
3106                 ASSERT((uintptr_t)(mp2->b_wptr - mp2->b_rptr) <=
3107                     (uintptr_t)INT_MAX);
3108 
3109                 data_length += (int)(mp2->b_wptr - mp2->b_rptr);
3110                 if (data_length > max_to_send) {
3111                         mp2->b_wptr -= data_length - max_to_send;
3112                         data_length = max_to_send;
3113                         off = mp2->b_wptr - mp->b_rptr;
3114                         break;
3115                 } else {
3116                         off = 0;
3117                 }
3118         }
3119         if (offset != NULL) {
3120                 *offset = off;
3121                 *end_mp = mp;
3122         }
3123         if (seg_len != NULL) {
3124                 *seg_len = data_length;
3125         }
3126 
3127         /* Update the latest receive window size in TCP header. */
3128         tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3129 
3130         rptr = mp1->b_rptr + tcps->tcps_wroff_xtra;
3131         mp1->b_rptr = rptr;
3132         mp1->b_wptr = rptr + connp->conn_ht_iphc_len + sack_opt_len;
3133         bcopy(connp->conn_ht_iphc, rptr, connp->conn_ht_iphc_len);
3134         tcpha = (tcpha_t *)&rptr[ixa->ixa_ip_hdr_length];
3135         tcpha->tha_seq = htonl(seq);
3136 
3137         /*
3138          * Use tcp_unsent to determine if the PUSH bit should be used assumes
3139          * that this function was called from tcp_wput_data. Thus, when called
3140          * to retransmit data the setting of the PUSH bit may appear some
3141          * what random in that it might get set when it should not. This
3142          * should not pose any performance issues.
3143          */
3144         if (data_length != 0 && (tcp->tcp_unsent == 0 ||
3145             tcp->tcp_unsent == data_length)) {
3146                 flags = TH_ACK | TH_PUSH;
3147         } else {
3148                 flags = TH_ACK;
3149         }
3150 
3151         if (tcp->tcp_ecn_ok) {
3152                 if (tcp->tcp_ecn_echo_on)
3153                         flags |= TH_ECE;
3154 
3155                 /*
3156                  * Only set ECT bit and ECN_CWR if a segment contains new data.
3157                  * There is no TCP flow control for non-data segments, and
3158                  * only data segment is transmitted reliably.
3159                  */
3160                 if (data_length > 0 && !rexmit) {
3161                         TCP_SET_ECT(tcp, rptr);
3162                         if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3163                                 flags |= TH_CWR;
3164                                 tcp->tcp_ecn_cwr_sent = B_TRUE;
3165                         }
3166                 }
3167         }
3168 
3169         /* Check if there is any special processing needs to be done. */
3170         if (tcp->tcp_valid_bits) {
3171                 uint32_t u1;
3172 
3173                 /* We don't allow having SYN and FIN in the same segment... */
3174                 if ((tcp->tcp_valid_bits & TCP_ISS_VALID) &&
3175                     seq == tcp->tcp_iss) {
3176                         /* Need to do connection set up processing. */
3177                         tcp_xmit_mp_aux_iss(tcp, connp, tcpha, mp1, &flags);
3178                 } else if ((tcp->tcp_valid_bits & TCP_FSS_VALID) &&
3179                     (seq + data_length) == tcp->tcp_fss) {
3180                         /* Need to do connection tear down processing. */
3181                         tcp_xmit_mp_aux_fss(tcp, ixa, &flags);
3182                 }
3183 
3184                 /*
3185                  * Need to do urgent pointer processing.
3186                  *
3187                  * Note the trick here.  u1 is unsigned.  When tcp_urg
3188                  * is smaller than seq, u1 will become a very huge value.
3189                  * So the comparison will fail.  Also note that tcp_urp
3190                  * should be positive, see RFC 793 page 17.
3191                  */
3192                 u1 = tcp->tcp_urg - seq + TCP_OLD_URP_INTERPRETATION;
3193                 if ((tcp->tcp_valid_bits & TCP_URG_VALID) && u1 != 0 &&
3194                     u1 < (uint32_t)(64 * 1024)) {
3195                         flags |= TH_URG;
3196                         TCPS_BUMP_MIB(tcps, tcpOutUrg);
3197                         tcpha->tha_urp = htons(u1);
3198                 }
3199         }
3200         tcpha->tha_flags = (uchar_t)flags;
3201         tcp->tcp_rack = tcp->tcp_rnxt;
3202         tcp->tcp_rack_cnt = 0;
3203 
3204         /* Fill in the current value of timestamps option. */
3205         if (tcp->tcp_snd_ts_ok) {
3206                 if (tcp->tcp_state != TCPS_SYN_SENT) {
3207                         uint32_t llbolt = (uint32_t)LBOLT_FASTPATH;
3208 
3209                         U32_TO_BE32(llbolt,
3210                             (char *)tcpha + TCP_MIN_HEADER_LENGTH+4);
3211                         U32_TO_BE32(tcp->tcp_ts_recent,
3212                             (char *)tcpha + TCP_MIN_HEADER_LENGTH+8);
3213                 }
3214         }
3215 
3216         /* Fill in the SACK blocks. */
3217         if (num_sack_blk > 0) {
3218                 uchar_t *wptr = (uchar_t *)tcpha + connp->conn_ht_ulp_len;
3219                 sack_blk_t *tmp;
3220                 int32_t i;
3221 
3222                 wptr[0] = TCPOPT_NOP;
3223                 wptr[1] = TCPOPT_NOP;
3224                 wptr[2] = TCPOPT_SACK;
3225                 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3226                     sizeof (sack_blk_t);
3227                 wptr += TCPOPT_REAL_SACK_LEN;
3228 
3229                 tmp = tcp->tcp_sack_list;
3230                 for (i = 0; i < num_sack_blk; i++) {
3231                         U32_TO_BE32(tmp[i].begin, wptr);
3232                         wptr += sizeof (tcp_seq);
3233                         U32_TO_BE32(tmp[i].end, wptr);
3234                         wptr += sizeof (tcp_seq);
3235                 }
3236                 tcpha->tha_offset_and_reserved += ((num_sack_blk * 2 + 1) << 4);
3237         }
3238         ASSERT((uintptr_t)(mp1->b_wptr - rptr) <= (uintptr_t)INT_MAX);
3239         data_length += (int)(mp1->b_wptr - rptr);
3240 
3241         ixa->ixa_pktlen = data_length;
3242 
3243         if (ixa->ixa_flags & IXAF_IS_IPV4) {
3244                 ((ipha_t *)rptr)->ipha_length = htons(data_length);
3245         } else {
3246                 ip6_t *ip6 = (ip6_t *)rptr;
3247 
3248                 ip6->ip6_plen = htons(data_length - IPV6_HDR_LEN);
3249         }
3250 
3251         /*
3252          * Prime pump for IP
3253          * Include the adjustment for a source route if any.
3254          */
3255         data_length -= ixa->ixa_ip_hdr_length;
3256         data_length += connp->conn_sum;
3257         data_length = (data_length >> 16) + (data_length & 0xFFFF);
3258         tcpha->tha_sum = htons(data_length);
3259         if (tcp->tcp_ip_forward_progress) {
3260                 tcp->tcp_ip_forward_progress = B_FALSE;
3261                 connp->conn_ixa->ixa_flags |= IXAF_REACH_CONF;
3262         } else {
3263                 connp->conn_ixa->ixa_flags &= ~IXAF_REACH_CONF;
3264         }
3265         return (mp1);
3266 }
3267 
3268 /*
3269  * If this routine returns B_TRUE, TCP can generate a RST in response
3270  * to a segment.  If it returns B_FALSE, TCP should not respond.
3271  */
3272 static boolean_t
3273 tcp_send_rst_chk(tcp_stack_t *tcps)
3274 {
3275         int64_t now;
3276 
3277         /*
3278          * TCP needs to protect itself from generating too many RSTs.
3279          * This can be a DoS attack by sending us random segments
3280          * soliciting RSTs.
3281          *
3282          * What we do here is to have a limit of tcp_rst_sent_rate RSTs
3283          * in each 1 second interval.  In this way, TCP still generate
3284          * RSTs in normal cases but when under attack, the impact is
3285          * limited.
3286          */
3287         if (tcps->tcps_rst_sent_rate_enabled != 0) {
3288                 now = ddi_get_lbolt64();
3289                 if (TICK_TO_MSEC(now - tcps->tcps_last_rst_intrvl) >
3290                     1*SECONDS) {
3291                         tcps->tcps_last_rst_intrvl = now;
3292                         tcps->tcps_rst_cnt = 1;
3293                 } else if (++tcps->tcps_rst_cnt > tcps->tcps_rst_sent_rate) {
3294                         return (B_FALSE);
3295                 }
3296         }
3297         return (B_TRUE);
3298 }
3299 
3300 /*
3301  * This function handles all retransmissions if SACK is enabled for this
3302  * connection.  First it calculates how many segments can be retransmitted
3303  * based on tcp_pipe.  Then it goes thru the notsack list to find eligible
3304  * segments.  A segment is eligible if sack_cnt for that segment is greater
3305  * than or equal tcp_dupack_fast_retransmit.  After it has retransmitted
3306  * all eligible segments, it checks to see if TCP can send some new segments
3307  * (fast recovery).  If it can, set the appropriate flag for tcp_input_data().
3308  *
3309  * Parameters:
3310  *      tcp_t *tcp: the tcp structure of the connection.
3311  *      uint_t *flags: in return, appropriate value will be set for
3312  *      tcp_input_data().
3313  */
3314 void
3315 tcp_sack_rexmit(tcp_t *tcp, uint_t *flags)
3316 {
3317         notsack_blk_t   *notsack_blk;
3318         int32_t         usable_swnd;
3319         int32_t         mss;
3320         uint32_t        seg_len;
3321         mblk_t          *xmit_mp;
3322         tcp_stack_t     *tcps = tcp->tcp_tcps;
3323 
3324         ASSERT(tcp->tcp_notsack_list != NULL);
3325         ASSERT(tcp->tcp_rexmit == B_FALSE);
3326 
3327         /* Defensive coding in case there is a bug... */
3328         if (tcp->tcp_notsack_list == NULL) {
3329                 return;
3330         }
3331         notsack_blk = tcp->tcp_notsack_list;
3332         mss = tcp->tcp_mss;
3333 
3334         /*
3335          * Limit the num of outstanding data in the network to be
3336          * tcp_cwnd_ssthresh, which is half of the original congestion wnd.
3337          */
3338         usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3339 
3340         /* At least retransmit 1 MSS of data. */
3341         if (usable_swnd <= 0) {
3342                 usable_swnd = mss;
3343         }
3344 
3345         /* Make sure no new RTT samples will be taken. */
3346         tcp->tcp_csuna = tcp->tcp_snxt;
3347 
3348         notsack_blk = tcp->tcp_notsack_list;
3349         while (usable_swnd > 0) {
3350                 mblk_t          *snxt_mp, *tmp_mp;
3351                 tcp_seq         begin = tcp->tcp_sack_snxt;
3352                 tcp_seq         end;
3353                 int32_t         off;
3354 
3355                 for (; notsack_blk != NULL; notsack_blk = notsack_blk->next) {
3356                         if (SEQ_GT(notsack_blk->end, begin) &&
3357                             (notsack_blk->sack_cnt >=
3358                             tcps->tcps_dupack_fast_retransmit)) {
3359                                 end = notsack_blk->end;
3360                                 if (SEQ_LT(begin, notsack_blk->begin)) {
3361                                         begin = notsack_blk->begin;
3362                                 }
3363                                 break;
3364                         }
3365                 }
3366                 /*
3367                  * All holes are filled.  Manipulate tcp_cwnd to send more
3368                  * if we can.  Note that after the SACK recovery, tcp_cwnd is
3369                  * set to tcp_cwnd_ssthresh.
3370                  */
3371                 if (notsack_blk == NULL) {
3372                         usable_swnd = tcp->tcp_cwnd_ssthresh - tcp->tcp_pipe;
3373                         if (usable_swnd <= 0 || tcp->tcp_unsent == 0) {
3374                                 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna;
3375                                 ASSERT(tcp->tcp_cwnd > 0);
3376                                 return;
3377                         } else {
3378                                 usable_swnd = usable_swnd / mss;
3379                                 tcp->tcp_cwnd = tcp->tcp_snxt - tcp->tcp_suna +
3380                                     MAX(usable_swnd * mss, mss);
3381                                 *flags |= TH_XMIT_NEEDED;
3382                                 return;
3383                         }
3384                 }
3385 
3386                 /*
3387                  * Note that we may send more than usable_swnd allows here
3388                  * because of round off, but no more than 1 MSS of data.
3389                  */
3390                 seg_len = end - begin;
3391                 if (seg_len > mss)
3392                         seg_len = mss;
3393                 snxt_mp = tcp_get_seg_mp(tcp, begin, &off);
3394                 ASSERT(snxt_mp != NULL);
3395                 /* This should not happen.  Defensive coding again... */
3396                 if (snxt_mp == NULL) {
3397                         return;
3398                 }
3399 
3400                 xmit_mp = tcp_xmit_mp(tcp, snxt_mp, seg_len, &off,
3401                     &tmp_mp, begin, B_TRUE, &seg_len, B_TRUE);
3402                 if (xmit_mp == NULL)
3403                         return;
3404 
3405                 usable_swnd -= seg_len;
3406                 tcp->tcp_pipe += seg_len;
3407                 tcp->tcp_sack_snxt = begin + seg_len;
3408 
3409                 tcp_send_data(tcp, xmit_mp);
3410 
3411                 /*
3412                  * Update the send timestamp to avoid false retransmission.
3413                  */
3414 #ifdef KERNEL_32
3415                 snxt_mp->b_prev = (mblk_t *)ddi_get_lbolt();
3416 #else
3417                 snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3418 #endif
3419 
3420                 TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3421                 TCPS_UPDATE_MIB(tcps, tcpRetransBytes, seg_len);
3422                 TCPS_BUMP_MIB(tcps, tcpOutSackRetransSegs);
3423                 tcp->tcp_cs.tcp_out_retrans_segs++;
3424                 tcp->tcp_cs.tcp_out_retrans_bytes += seg_len;
3425                 /*
3426                  * Update tcp_rexmit_max to extend this SACK recovery phase.
3427                  * This happens when new data sent during fast recovery is
3428                  * also lost.  If TCP retransmits those new data, it needs
3429                  * to extend SACK recover phase to avoid starting another
3430                  * fast retransmit/recovery unnecessarily.
3431                  */
3432                 if (SEQ_GT(tcp->tcp_sack_snxt, tcp->tcp_rexmit_max)) {
3433                         tcp->tcp_rexmit_max = tcp->tcp_sack_snxt;
3434                 }
3435         }
3436 }
3437 
3438 /*
3439  * tcp_ss_rexmit() is called to do slow start retransmission after a timeout
3440  * or ICMP errors.
3441  */
3442 void
3443 tcp_ss_rexmit(tcp_t *tcp)
3444 {
3445         uint32_t        snxt;
3446         uint32_t        smax;
3447         int32_t         win;
3448         int32_t         mss;
3449         int32_t         off;
3450         mblk_t          *snxt_mp;
3451         tcp_stack_t     *tcps = tcp->tcp_tcps;
3452 
3453         /*
3454          * Note that tcp_rexmit can be set even though TCP has retransmitted
3455          * all unack'ed segments.
3456          */
3457         if (SEQ_LT(tcp->tcp_rexmit_nxt, tcp->tcp_rexmit_max)) {
3458                 smax = tcp->tcp_rexmit_max;
3459                 snxt = tcp->tcp_rexmit_nxt;
3460                 if (SEQ_LT(snxt, tcp->tcp_suna)) {
3461                         snxt = tcp->tcp_suna;
3462                 }
3463                 win = MIN(tcp->tcp_cwnd, tcp->tcp_swnd);
3464                 win -= snxt - tcp->tcp_suna;
3465                 mss = tcp->tcp_mss;
3466                 snxt_mp = tcp_get_seg_mp(tcp, snxt, &off);
3467 
3468                 while (SEQ_LT(snxt, smax) && (win > 0) && (snxt_mp != NULL)) {
3469                         mblk_t  *xmit_mp;
3470                         mblk_t  *old_snxt_mp = snxt_mp;
3471                         uint32_t cnt = mss;
3472 
3473                         if (win < cnt) {
3474                                 cnt = win;
3475                         }
3476                         if (SEQ_GT(snxt + cnt, smax)) {
3477                                 cnt = smax - snxt;
3478                         }
3479                         xmit_mp = tcp_xmit_mp(tcp, snxt_mp, cnt, &off,
3480                             &snxt_mp, snxt, B_TRUE, &cnt, B_TRUE);
3481                         if (xmit_mp == NULL)
3482                                 return;
3483 
3484                         tcp_send_data(tcp, xmit_mp);
3485 
3486                         snxt += cnt;
3487                         win -= cnt;
3488                         /*
3489                          * Update the send timestamp to avoid false
3490                          * retransmission.
3491                          */
3492 #ifdef KERNEL_32
3493                         old_snxt_mp->b_prev = (mblk_t *)ddi_get_lbolt();
3494 #else
3495                         old_snxt_mp->b_prev = (mblk_t *)(intptr_t)gethrtime();
3496 #endif
3497                         TCPS_BUMP_MIB(tcps, tcpRetransSegs);
3498                         TCPS_UPDATE_MIB(tcps, tcpRetransBytes, cnt);
3499                         tcp->tcp_cs.tcp_out_retrans_segs++;
3500                         tcp->tcp_cs.tcp_out_retrans_bytes += cnt;
3501 
3502                         tcp->tcp_rexmit_nxt = snxt;
3503                 }
3504                 /*
3505                  * If we have transmitted all we have at the time
3506                  * we started the retranmission, we can leave
3507                  * the rest of the job to tcp_wput_data().  But we
3508                  * need to check the send window first.  If the
3509                  * win is not 0, go on with tcp_wput_data().
3510                  */
3511                 if (SEQ_LT(snxt, smax) || win == 0) {
3512                         return;
3513                 }
3514         }
3515         /* Only call tcp_wput_data() if there is data to be sent. */
3516         if (tcp->tcp_unsent) {
3517                 tcp_wput_data(tcp, NULL, B_FALSE);
3518         }
3519 }
3520 
3521 /*
3522  * Do slow start retransmission after ICMP errors of PMTU changes.
3523  */
3524 void
3525 tcp_rexmit_after_error(tcp_t *tcp)
3526 {
3527         /*
3528          * All sent data has been acknowledged or no data left to send, just
3529          * to return.
3530          */
3531         if (!SEQ_LT(tcp->tcp_suna, tcp->tcp_snxt) ||
3532             (tcp->tcp_xmit_head == NULL))
3533                 return;
3534 
3535         if ((tcp->tcp_valid_bits & TCP_FSS_VALID) && (tcp->tcp_unsent == 0))
3536                 tcp->tcp_rexmit_max = tcp->tcp_fss;
3537         else
3538                 tcp->tcp_rexmit_max = tcp->tcp_snxt;
3539 
3540         tcp->tcp_rexmit_nxt = tcp->tcp_suna;
3541         tcp->tcp_rexmit = B_TRUE;
3542         tcp->tcp_dupack_cnt = 0;
3543         tcp_ss_rexmit(tcp);
3544 }
3545 
3546 /*
3547  * tcp_get_seg_mp() is called to get the pointer to a segment in the
3548  * send queue which starts at the given sequence number. If the given
3549  * sequence number is equal to last valid sequence number (tcp_snxt), the
3550  * returned mblk is the last valid mblk, and off is set to the length of
3551  * that mblk.
3552  *
3553  * send queue which starts at the given seq. no.
3554  *
3555  * Parameters:
3556  *      tcp_t *tcp: the tcp instance pointer.
3557  *      uint32_t seq: the starting seq. no of the requested segment.
3558  *      int32_t *off: after the execution, *off will be the offset to
3559  *              the returned mblk which points to the requested seq no.
3560  *              It is the caller's responsibility to send in a non-null off.
3561  *
3562  * Return:
3563  *      A mblk_t pointer pointing to the requested segment in send queue.
3564  */
3565 static mblk_t *
3566 tcp_get_seg_mp(tcp_t *tcp, uint32_t seq, int32_t *off)
3567 {
3568         int32_t cnt;
3569         mblk_t  *mp;
3570 
3571         /* Defensive coding.  Make sure we don't send incorrect data. */
3572         if (SEQ_LT(seq, tcp->tcp_suna) || SEQ_GT(seq, tcp->tcp_snxt))
3573                 return (NULL);
3574 
3575         cnt = seq - tcp->tcp_suna;
3576         mp = tcp->tcp_xmit_head;
3577         while (cnt > 0 && mp != NULL) {
3578                 cnt -= mp->b_wptr - mp->b_rptr;
3579                 if (cnt <= 0) {
3580                         cnt += mp->b_wptr - mp->b_rptr;
3581                         break;
3582                 }
3583                 mp = mp->b_cont;
3584         }
3585         ASSERT(mp != NULL);
3586         *off = cnt;
3587         return (mp);
3588 }
3589 
3590 /*
3591  * This routine adjusts next-to-send sequence number variables, in the
3592  * case where the reciever has shrunk it's window.
3593  */
3594 void
3595 tcp_update_xmit_tail(tcp_t *tcp, uint32_t snxt)
3596 {
3597         mblk_t *xmit_tail;
3598         int32_t offset;
3599 
3600         tcp->tcp_snxt = snxt;
3601 
3602         /* Get the mblk, and the offset in it, as per the shrunk window */
3603         xmit_tail = tcp_get_seg_mp(tcp, snxt, &offset);
3604         ASSERT(xmit_tail != NULL);
3605         tcp->tcp_xmit_tail = xmit_tail;
3606         tcp->tcp_xmit_tail_unsent = xmit_tail->b_wptr -
3607             xmit_tail->b_rptr - offset;
3608 }
3609 
3610 /*
3611  * This handles the case when the receiver has shrunk its win. Per RFC 1122
3612  * if the receiver shrinks the window, i.e. moves the right window to the
3613  * left, the we should not send new data, but should retransmit normally the
3614  * old unacked data between suna and suna + swnd. We might has sent data
3615  * that is now outside the new window, pretend that we didn't send  it.
3616  */
3617 static void
3618 tcp_process_shrunk_swnd(tcp_t *tcp, uint32_t shrunk_count)
3619 {
3620         uint32_t        snxt = tcp->tcp_snxt;
3621 
3622         ASSERT(shrunk_count > 0);
3623 
3624         if (!tcp->tcp_is_wnd_shrnk) {
3625                 tcp->tcp_snxt_shrunk = snxt;
3626                 tcp->tcp_is_wnd_shrnk = B_TRUE;
3627         } else if (SEQ_GT(snxt, tcp->tcp_snxt_shrunk)) {
3628                 tcp->tcp_snxt_shrunk = snxt;
3629         }
3630 
3631         /* Pretend we didn't send the data outside the window */
3632         snxt -= shrunk_count;
3633 
3634         /* Reset all the values per the now shrunk window */
3635         tcp_update_xmit_tail(tcp, snxt);
3636         tcp->tcp_unsent += shrunk_count;
3637 
3638         /*
3639          * If the SACK option is set, delete the entire list of
3640          * notsack'ed blocks.
3641          */
3642         TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
3643 
3644         if (tcp->tcp_suna == tcp->tcp_snxt && tcp->tcp_swnd == 0)
3645                 /*
3646                  * Make sure the timer is running so that we will probe a zero
3647                  * window.
3648                  */
3649                 TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
3650 }
3651 
3652 /*
3653  * tcp_fill_header is called by tcp_send() to fill the outgoing TCP header
3654  * with the template header, as well as other options such as time-stamp,
3655  * ECN and/or SACK.
3656  */
3657 static void
3658 tcp_fill_header(tcp_t *tcp, uchar_t *rptr, int num_sack_blk)
3659 {
3660         tcpha_t *tcp_tmpl, *tcpha;
3661         uint32_t *dst, *src;
3662         int hdrlen;
3663         conn_t *connp = tcp->tcp_connp;
3664 
3665         ASSERT(OK_32PTR(rptr));
3666 
3667         /* Template header */
3668         tcp_tmpl = tcp->tcp_tcpha;
3669 
3670         /* Header of outgoing packet */
3671         tcpha = (tcpha_t *)(rptr + connp->conn_ixa->ixa_ip_hdr_length);
3672 
3673         /* dst and src are opaque 32-bit fields, used for copying */
3674         dst = (uint32_t *)rptr;
3675         src = (uint32_t *)connp->conn_ht_iphc;
3676         hdrlen = connp->conn_ht_iphc_len;
3677 
3678         /* Fill time-stamp option if needed */
3679         if (tcp->tcp_snd_ts_ok) {
3680                 U32_TO_BE32(LBOLT_FASTPATH,
3681                     (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 4);
3682                 U32_TO_BE32(tcp->tcp_ts_recent,
3683                     (char *)tcp_tmpl + TCP_MIN_HEADER_LENGTH + 8);
3684         } else {
3685                 ASSERT(connp->conn_ht_ulp_len == TCP_MIN_HEADER_LENGTH);
3686         }
3687 
3688         /*
3689          * Copy the template header; is this really more efficient than
3690          * calling bcopy()?  For simple IPv4/TCP, it may be the case,
3691          * but perhaps not for other scenarios.
3692          */
3693         dst[0] = src[0];
3694         dst[1] = src[1];
3695         dst[2] = src[2];
3696         dst[3] = src[3];
3697         dst[4] = src[4];
3698         dst[5] = src[5];
3699         dst[6] = src[6];
3700         dst[7] = src[7];
3701         dst[8] = src[8];
3702         dst[9] = src[9];
3703         if (hdrlen -= 40) {
3704                 hdrlen >>= 2;
3705                 dst += 10;
3706                 src += 10;
3707                 do {
3708                         *dst++ = *src++;
3709                 } while (--hdrlen);
3710         }
3711 
3712         /*
3713          * Set the ECN info in the TCP header if it is not a zero
3714          * window probe.  Zero window probe is only sent in
3715          * tcp_wput_data() and tcp_timer().
3716          */
3717         if (tcp->tcp_ecn_ok && !tcp->tcp_zero_win_probe) {
3718                 TCP_SET_ECT(tcp, rptr);
3719 
3720                 if (tcp->tcp_ecn_echo_on)
3721                         tcpha->tha_flags |= TH_ECE;
3722                 if (tcp->tcp_cwr && !tcp->tcp_ecn_cwr_sent) {
3723                         tcpha->tha_flags |= TH_CWR;
3724                         tcp->tcp_ecn_cwr_sent = B_TRUE;
3725                 }
3726         }
3727 
3728         /* Fill in SACK options */
3729         if (num_sack_blk > 0) {
3730                 uchar_t *wptr = rptr + connp->conn_ht_iphc_len;
3731                 sack_blk_t *tmp;
3732                 int32_t i;
3733 
3734                 wptr[0] = TCPOPT_NOP;
3735                 wptr[1] = TCPOPT_NOP;
3736                 wptr[2] = TCPOPT_SACK;
3737                 wptr[3] = TCPOPT_HEADER_LEN + num_sack_blk *
3738                     sizeof (sack_blk_t);
3739                 wptr += TCPOPT_REAL_SACK_LEN;
3740 
3741                 tmp = tcp->tcp_sack_list;
3742                 for (i = 0; i < num_sack_blk; i++) {
3743                         U32_TO_BE32(tmp[i].begin, wptr);
3744                         wptr += sizeof (tcp_seq);
3745                         U32_TO_BE32(tmp[i].end, wptr);
3746                         wptr += sizeof (tcp_seq);
3747                 }
3748                 tcpha->tha_offset_and_reserved +=
3749                     ((num_sack_blk * 2 + 1) << 4);
3750         }
3751 }