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) 1991, 2010, Oracle and/or its affiliates. All rights reserved.
  24  * Copyright 2017 Joyent, Inc.
  25  * Copyright (c) 2011 Nexenta Systems, Inc. All rights reserved.
  26  * Copyright (c) 2013, 2017 by Delphix. All rights reserved.
  27  * Copyright 2014, OmniTI Computer Consulting, Inc. All rights reserved.
  28  */
  29 /* Copyright (c) 1990 Mentat Inc. */
  30 
  31 #include <sys/types.h>
  32 #include <sys/stream.h>
  33 #include <sys/strsun.h>
  34 #include <sys/strsubr.h>
  35 #include <sys/stropts.h>
  36 #include <sys/strlog.h>
  37 #define _SUN_TPI_VERSION 2
  38 #include <sys/tihdr.h>
  39 #include <sys/timod.h>
  40 #include <sys/ddi.h>
  41 #include <sys/sunddi.h>
  42 #include <sys/suntpi.h>
  43 #include <sys/xti_inet.h>
  44 #include <sys/cmn_err.h>
  45 #include <sys/debug.h>
  46 #include <sys/sdt.h>
  47 #include <sys/vtrace.h>
  48 #include <sys/kmem.h>
  49 #include <sys/ethernet.h>
  50 #include <sys/cpuvar.h>
  51 #include <sys/dlpi.h>
  52 #include <sys/pattr.h>
  53 #include <sys/policy.h>
  54 #include <sys/priv.h>
  55 #include <sys/zone.h>
  56 #include <sys/sunldi.h>
  57 
  58 #include <sys/errno.h>
  59 #include <sys/signal.h>
  60 #include <sys/socket.h>
  61 #include <sys/socketvar.h>
  62 #include <sys/sockio.h>
  63 #include <sys/isa_defs.h>
  64 #include <sys/md5.h>
  65 #include <sys/random.h>
  66 #include <sys/uio.h>
  67 #include <sys/systm.h>
  68 #include <netinet/in.h>
  69 #include <netinet/tcp.h>
  70 #include <netinet/ip6.h>
  71 #include <netinet/icmp6.h>
  72 #include <net/if.h>
  73 #include <net/route.h>
  74 #include <inet/ipsec_impl.h>
  75 
  76 #include <inet/common.h>
  77 #include <inet/cc.h>
  78 #include <inet/ip.h>
  79 #include <inet/ip_impl.h>
  80 #include <inet/ip6.h>
  81 #include <inet/ip_ndp.h>
  82 #include <inet/proto_set.h>
  83 #include <inet/mib2.h>
  84 #include <inet/optcom.h>
  85 #include <inet/snmpcom.h>
  86 #include <inet/kstatcom.h>
  87 #include <inet/tcp.h>
  88 #include <inet/tcp_impl.h>
  89 #include <inet/tcp_cluster.h>
  90 #include <inet/udp_impl.h>
  91 #include <net/pfkeyv2.h>
  92 #include <inet/ipdrop.h>
  93 
  94 #include <inet/ipclassifier.h>
  95 #include <inet/ip_ire.h>
  96 #include <inet/ip_ftable.h>
  97 #include <inet/ip_if.h>
  98 #include <inet/ipp_common.h>
  99 #include <inet/ip_rts.h>
 100 #include <inet/ip_netinfo.h>
 101 #include <sys/squeue_impl.h>
 102 #include <sys/squeue.h>
 103 #include <sys/tsol/label.h>
 104 #include <sys/tsol/tnet.h>
 105 #include <rpc/pmap_prot.h>
 106 #include <sys/callo.h>
 107 
 108 /*
 109  * TCP Notes: aka FireEngine Phase I (PSARC 2002/433)
 110  *
 111  * (Read the detailed design doc in PSARC case directory)
 112  *
 113  * The entire tcp state is contained in tcp_t and conn_t structure
 114  * which are allocated in tandem using ipcl_conn_create() and passing
 115  * IPCL_TCPCONN as a flag. We use 'conn_ref' and 'conn_lock' to protect
 116  * the references on the tcp_t. The tcp_t structure is never compressed
 117  * and packets always land on the correct TCP perimeter from the time
 118  * eager is created till the time tcp_t dies (as such the old mentat
 119  * TCP global queue is not used for detached state and no IPSEC checking
 120  * is required). The global queue is still allocated to send out resets
 121  * for connection which have no listeners and IP directly calls
 122  * tcp_xmit_listeners_reset() which does any policy check.
 123  *
 124  * Protection and Synchronisation mechanism:
 125  *
 126  * The tcp data structure does not use any kind of lock for protecting
 127  * its state but instead uses 'squeues' for mutual exclusion from various
 128  * read and write side threads. To access a tcp member, the thread should
 129  * always be behind squeue (via squeue_enter with flags as SQ_FILL, SQ_PROCESS,
 130  * or SQ_NODRAIN). Since the squeues allow a direct function call, caller
 131  * can pass any tcp function having prototype of edesc_t as argument
 132  * (different from traditional STREAMs model where packets come in only
 133  * designated entry points). The list of functions that can be directly
 134  * called via squeue are listed before the usual function prototype.
 135  *
 136  * Referencing:
 137  *
 138  * TCP is MT-Hot and we use a reference based scheme to make sure that the
 139  * tcp structure doesn't disappear when its needed. When the application
 140  * creates an outgoing connection or accepts an incoming connection, we
 141  * start out with 2 references on 'conn_ref'. One for TCP and one for IP.
 142  * The IP reference is just a symbolic reference since ip_tcpclose()
 143  * looks at tcp structure after tcp_close_output() returns which could
 144  * have dropped the last TCP reference. So as long as the connection is
 145  * in attached state i.e. !TCP_IS_DETACHED, we have 2 references on the
 146  * conn_t. The classifier puts its own reference when the connection is
 147  * inserted in listen or connected hash. Anytime a thread needs to enter
 148  * the tcp connection perimeter, it retrieves the conn/tcp from q->ptr
 149  * on write side or by doing a classify on read side and then puts a
 150  * reference on the conn before doing squeue_enter/tryenter/fill. For
 151  * read side, the classifier itself puts the reference under fanout lock
 152  * to make sure that tcp can't disappear before it gets processed. The
 153  * squeue will drop this reference automatically so the called function
 154  * doesn't have to do a DEC_REF.
 155  *
 156  * Opening a new connection:
 157  *
 158  * The outgoing connection open is pretty simple. tcp_open() does the
 159  * work in creating the conn/tcp structure and initializing it. The
 160  * squeue assignment is done based on the CPU the application
 161  * is running on. So for outbound connections, processing is always done
 162  * on application CPU which might be different from the incoming CPU
 163  * being interrupted by the NIC. An optimal way would be to figure out
 164  * the NIC <-> CPU binding at listen time, and assign the outgoing
 165  * connection to the squeue attached to the CPU that will be interrupted
 166  * for incoming packets (we know the NIC based on the bind IP address).
 167  * This might seem like a problem if more data is going out but the
 168  * fact is that in most cases the transmit is ACK driven transmit where
 169  * the outgoing data normally sits on TCP's xmit queue waiting to be
 170  * transmitted.
 171  *
 172  * Accepting a connection:
 173  *
 174  * This is a more interesting case because of various races involved in
 175  * establishing a eager in its own perimeter. Read the meta comment on
 176  * top of tcp_input_listener(). But briefly, the squeue is picked by
 177  * ip_fanout based on the ring or the sender (if loopback).
 178  *
 179  * Closing a connection:
 180  *
 181  * The close is fairly straight forward. tcp_close() calls tcp_close_output()
 182  * via squeue to do the close and mark the tcp as detached if the connection
 183  * was in state TCPS_ESTABLISHED or greater. In the later case, TCP keep its
 184  * reference but tcp_close() drop IP's reference always. So if tcp was
 185  * not killed, it is sitting in time_wait list with 2 reference - 1 for TCP
 186  * and 1 because it is in classifier's connected hash. This is the condition
 187  * we use to determine that its OK to clean up the tcp outside of squeue
 188  * when time wait expires (check the ref under fanout and conn_lock and
 189  * if it is 2, remove it from fanout hash and kill it).
 190  *
 191  * Although close just drops the necessary references and marks the
 192  * tcp_detached state, tcp_close needs to know the tcp_detached has been
 193  * set (under squeue) before letting the STREAM go away (because a
 194  * inbound packet might attempt to go up the STREAM while the close
 195  * has happened and tcp_detached is not set). So a special lock and
 196  * flag is used along with a condition variable (tcp_closelock, tcp_closed,
 197  * and tcp_closecv) to signal tcp_close that tcp_close_out() has marked
 198  * tcp_detached.
 199  *
 200  * Special provisions and fast paths:
 201  *
 202  * We make special provisions for sockfs by marking tcp_issocket
 203  * whenever we have only sockfs on top of TCP. This allows us to skip
 204  * putting the tcp in acceptor hash since a sockfs listener can never
 205  * become acceptor and also avoid allocating a tcp_t for acceptor STREAM
 206  * since eager has already been allocated and the accept now happens
 207  * on acceptor STREAM. There is a big blob of comment on top of
 208  * tcp_input_listener explaining the new accept. When socket is POP'd,
 209  * sockfs sends us an ioctl to mark the fact and we go back to old
 210  * behaviour. Once tcp_issocket is unset, its never set for the
 211  * life of that connection.
 212  *
 213  * IPsec notes :
 214  *
 215  * Since a packet is always executed on the correct TCP perimeter
 216  * all IPsec processing is defered to IP including checking new
 217  * connections and setting IPSEC policies for new connection. The
 218  * only exception is tcp_xmit_listeners_reset() which is called
 219  * directly from IP and needs to policy check to see if TH_RST
 220  * can be sent out.
 221  */
 222 
 223 /*
 224  * Values for squeue switch:
 225  * 1: SQ_NODRAIN
 226  * 2: SQ_PROCESS
 227  * 3: SQ_FILL
 228  */
 229 int tcp_squeue_wput = 2;        /* /etc/systems */
 230 int tcp_squeue_flag;
 231 
 232 /*
 233  * To prevent memory hog, limit the number of entries in tcp_free_list
 234  * to 1% of available memory / number of cpus
 235  */
 236 uint_t tcp_free_list_max_cnt = 0;
 237 
 238 #define TIDUSZ  4096    /* transport interface data unit size */
 239 
 240 /*
 241  * Size of acceptor hash list.  It has to be a power of 2 for hashing.
 242  */
 243 #define TCP_ACCEPTOR_FANOUT_SIZE                512
 244 
 245 #ifdef  _ILP32
 246 #define TCP_ACCEPTOR_HASH(accid)                                        \
 247                 (((uint_t)(accid) >> 8) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
 248 #else
 249 #define TCP_ACCEPTOR_HASH(accid)                                        \
 250                 ((uint_t)(accid) & (TCP_ACCEPTOR_FANOUT_SIZE - 1))
 251 #endif  /* _ILP32 */
 252 
 253 /*
 254  * Minimum number of connections which can be created per listener.  Used
 255  * when the listener connection count is in effect.
 256  */
 257 static uint32_t tcp_min_conn_listener = 2;
 258 
 259 uint32_t tcp_early_abort = 30;
 260 
 261 /* TCP Timer control structure */
 262 typedef struct tcpt_s {
 263         pfv_t   tcpt_pfv;       /* The routine we are to call */
 264         tcp_t   *tcpt_tcp;      /* The parameter we are to pass in */
 265 } tcpt_t;
 266 
 267 /*
 268  * Functions called directly via squeue having a prototype of edesc_t.
 269  */
 270 void            tcp_input_data(void *arg, mblk_t *mp, void *arg2,
 271     ip_recv_attr_t *ira);
 272 static void     tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2,
 273     ip_recv_attr_t *dummy);
 274 
 275 
 276 /* Prototype for TCP functions */
 277 static void     tcp_random_init(void);
 278 int             tcp_random(void);
 279 static int      tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp,
 280                     in_port_t dstport, uint_t srcid);
 281 static int      tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp,
 282                     in_port_t dstport, uint32_t flowinfo,
 283                     uint_t srcid, uint32_t scope_id);
 284 static void     tcp_iss_init(tcp_t *tcp);
 285 static void     tcp_reinit(tcp_t *tcp);
 286 static void     tcp_reinit_values(tcp_t *tcp);
 287 
 288 static void     tcp_wsrv(queue_t *q);
 289 static void     tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa);
 290 static void     tcp_update_zcopy(tcp_t *tcp);
 291 static void     tcp_notify(void *, ip_xmit_attr_t *, ixa_notify_type_t,
 292     ixa_notify_arg_t);
 293 static void     *tcp_stack_init(netstackid_t stackid, netstack_t *ns);
 294 static void     tcp_stack_fini(netstackid_t stackid, void *arg);
 295 
 296 static int      tcp_squeue_switch(int);
 297 
 298 static int      tcp_open(queue_t *, dev_t *, int, int, cred_t *, boolean_t);
 299 static int      tcp_openv4(queue_t *, dev_t *, int, int, cred_t *);
 300 static int      tcp_openv6(queue_t *, dev_t *, int, int, cred_t *);
 301 
 302 static void     tcp_squeue_add(squeue_t *);
 303 
 304 struct module_info tcp_rinfo =  {
 305         TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, TCP_RECV_HIWATER, TCP_RECV_LOWATER
 306 };
 307 
 308 static struct module_info tcp_winfo =  {
 309         TCP_MOD_ID, TCP_MOD_NAME, 0, INFPSZ, 127, 16
 310 };
 311 
 312 /*
 313  * Entry points for TCP as a device. The normal case which supports
 314  * the TCP functionality.
 315  * We have separate open functions for the /dev/tcp and /dev/tcp6 devices.
 316  */
 317 struct qinit tcp_rinitv4 = {
 318         NULL, (pfi_t)tcp_rsrv, tcp_openv4, tcp_tpi_close, NULL, &tcp_rinfo
 319 };
 320 
 321 struct qinit tcp_rinitv6 = {
 322         NULL, (pfi_t)tcp_rsrv, tcp_openv6, tcp_tpi_close, NULL, &tcp_rinfo
 323 };
 324 
 325 struct qinit tcp_winit = {
 326         (pfi_t)tcp_wput, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
 327 };
 328 
 329 /* Initial entry point for TCP in socket mode. */
 330 struct qinit tcp_sock_winit = {
 331         (pfi_t)tcp_wput_sock, (pfi_t)tcp_wsrv, NULL, NULL, NULL, &tcp_winfo
 332 };
 333 
 334 /* TCP entry point during fallback */
 335 struct qinit tcp_fallback_sock_winit = {
 336         (pfi_t)tcp_wput_fallback, NULL, NULL, NULL, NULL, &tcp_winfo
 337 };
 338 
 339 /*
 340  * Entry points for TCP as a acceptor STREAM opened by sockfs when doing
 341  * an accept. Avoid allocating data structures since eager has already
 342  * been created.
 343  */
 344 struct qinit tcp_acceptor_rinit = {
 345         NULL, (pfi_t)tcp_rsrv, NULL, tcp_tpi_close_accept, NULL, &tcp_winfo
 346 };
 347 
 348 struct qinit tcp_acceptor_winit = {
 349         (pfi_t)tcp_tpi_accept, NULL, NULL, NULL, NULL, &tcp_winfo
 350 };
 351 
 352 /* For AF_INET aka /dev/tcp */
 353 struct streamtab tcpinfov4 = {
 354         &tcp_rinitv4, &tcp_winit
 355 };
 356 
 357 /* For AF_INET6 aka /dev/tcp6 */
 358 struct streamtab tcpinfov6 = {
 359         &tcp_rinitv6, &tcp_winit
 360 };
 361 
 362 /*
 363  * Following assumes TPI alignment requirements stay along 32 bit
 364  * boundaries
 365  */
 366 #define ROUNDUP32(x) \
 367         (((x) + (sizeof (int32_t) - 1)) & ~(sizeof (int32_t) - 1))
 368 
 369 /* Template for response to info request. */
 370 struct T_info_ack tcp_g_t_info_ack = {
 371         T_INFO_ACK,             /* PRIM_type */
 372         0,                      /* TSDU_size */
 373         T_INFINITE,             /* ETSDU_size */
 374         T_INVALID,              /* CDATA_size */
 375         T_INVALID,              /* DDATA_size */
 376         sizeof (sin_t),         /* ADDR_size */
 377         0,                      /* OPT_size - not initialized here */
 378         TIDUSZ,                 /* TIDU_size */
 379         T_COTS_ORD,             /* SERV_type */
 380         TCPS_IDLE,              /* CURRENT_state */
 381         (XPG4_1|EXPINLINE)      /* PROVIDER_flag */
 382 };
 383 
 384 struct T_info_ack tcp_g_t_info_ack_v6 = {
 385         T_INFO_ACK,             /* PRIM_type */
 386         0,                      /* TSDU_size */
 387         T_INFINITE,             /* ETSDU_size */
 388         T_INVALID,              /* CDATA_size */
 389         T_INVALID,              /* DDATA_size */
 390         sizeof (sin6_t),        /* ADDR_size */
 391         0,                      /* OPT_size - not initialized here */
 392         TIDUSZ,         /* TIDU_size */
 393         T_COTS_ORD,             /* SERV_type */
 394         TCPS_IDLE,              /* CURRENT_state */
 395         (XPG4_1|EXPINLINE)      /* PROVIDER_flag */
 396 };
 397 
 398 /*
 399  * TCP tunables related declarations. Definitions are in tcp_tunables.c
 400  */
 401 extern mod_prop_info_t tcp_propinfo_tbl[];
 402 extern int tcp_propinfo_count;
 403 
 404 #define IS_VMLOANED_MBLK(mp) \
 405         (((mp)->b_datap->db_struioflag & STRUIO_ZC) != 0)
 406 
 407 uint32_t do_tcpzcopy = 1;               /* 0: disable, 1: enable, 2: force */
 408 
 409 /*
 410  * Forces all connections to obey the value of the tcps_maxpsz_multiplier
 411  * tunable settable via NDD.  Otherwise, the per-connection behavior is
 412  * determined dynamically during tcp_set_destination(), which is the default.
 413  */
 414 boolean_t tcp_static_maxpsz = B_FALSE;
 415 
 416 /*
 417  * If the receive buffer size is changed, this function is called to update
 418  * the upper socket layer on the new delayed receive wake up threshold.
 419  */
 420 static void
 421 tcp_set_recv_threshold(tcp_t *tcp, uint32_t new_rcvthresh)
 422 {
 423         uint32_t default_threshold = SOCKET_RECVHIWATER >> 3;
 424 
 425         if (IPCL_IS_NONSTR(tcp->tcp_connp)) {
 426                 conn_t *connp = tcp->tcp_connp;
 427                 struct sock_proto_props sopp;
 428 
 429                 /*
 430                  * only increase rcvthresh upto default_threshold
 431                  */
 432                 if (new_rcvthresh > default_threshold)
 433                         new_rcvthresh = default_threshold;
 434 
 435                 sopp.sopp_flags = SOCKOPT_RCVTHRESH;
 436                 sopp.sopp_rcvthresh = new_rcvthresh;
 437 
 438                 (*connp->conn_upcalls->su_set_proto_props)
 439                     (connp->conn_upper_handle, &sopp);
 440         }
 441 }
 442 
 443 /*
 444  * Figure out the value of window scale opton.  Note that the rwnd is
 445  * ASSUMED to be rounded up to the nearest MSS before the calculation.
 446  * We cannot find the scale value and then do a round up of tcp_rwnd
 447  * because the scale value may not be correct after that.
 448  *
 449  * Set the compiler flag to make this function inline.
 450  */
 451 void
 452 tcp_set_ws_value(tcp_t *tcp)
 453 {
 454         int i;
 455         uint32_t rwnd = tcp->tcp_rwnd;
 456 
 457         for (i = 0; rwnd > TCP_MAXWIN && i < TCP_MAX_WINSHIFT;
 458             i++, rwnd >>= 1)
 459                 ;
 460         tcp->tcp_rcv_ws = i;
 461 }
 462 
 463 /*
 464  * Remove cached/latched IPsec references.
 465  */
 466 void
 467 tcp_ipsec_cleanup(tcp_t *tcp)
 468 {
 469         conn_t          *connp = tcp->tcp_connp;
 470 
 471         ASSERT(connp->conn_flags & IPCL_TCPCONN);
 472 
 473         if (connp->conn_latch != NULL) {
 474                 IPLATCH_REFRELE(connp->conn_latch);
 475                 connp->conn_latch = NULL;
 476         }
 477         if (connp->conn_latch_in_policy != NULL) {
 478                 IPPOL_REFRELE(connp->conn_latch_in_policy);
 479                 connp->conn_latch_in_policy = NULL;
 480         }
 481         if (connp->conn_latch_in_action != NULL) {
 482                 IPACT_REFRELE(connp->conn_latch_in_action);
 483                 connp->conn_latch_in_action = NULL;
 484         }
 485         if (connp->conn_policy != NULL) {
 486                 IPPH_REFRELE(connp->conn_policy, connp->conn_netstack);
 487                 connp->conn_policy = NULL;
 488         }
 489 }
 490 
 491 /*
 492  * Cleaup before placing on free list.
 493  * Disassociate from the netstack/tcp_stack_t since the freelist
 494  * is per squeue and not per netstack.
 495  */
 496 void
 497 tcp_cleanup(tcp_t *tcp)
 498 {
 499         mblk_t          *mp;
 500         conn_t          *connp = tcp->tcp_connp;
 501         tcp_stack_t     *tcps = tcp->tcp_tcps;
 502         netstack_t      *ns = tcps->tcps_netstack;
 503         mblk_t          *tcp_rsrv_mp;
 504 
 505         tcp_bind_hash_remove(tcp);
 506 
 507         /* Cleanup that which needs the netstack first */
 508         tcp_ipsec_cleanup(tcp);
 509         ixa_cleanup(connp->conn_ixa);
 510 
 511         if (connp->conn_ht_iphc != NULL) {
 512                 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
 513                 connp->conn_ht_iphc = NULL;
 514                 connp->conn_ht_iphc_allocated = 0;
 515                 connp->conn_ht_iphc_len = 0;
 516                 connp->conn_ht_ulp = NULL;
 517                 connp->conn_ht_ulp_len = 0;
 518                 tcp->tcp_ipha = NULL;
 519                 tcp->tcp_ip6h = NULL;
 520                 tcp->tcp_tcpha = NULL;
 521         }
 522 
 523         /* We clear any IP_OPTIONS and extension headers */
 524         ip_pkt_free(&connp->conn_xmit_ipp);
 525 
 526         tcp_free(tcp);
 527 
 528         /*
 529          * Since we will bzero the entire structure, we need to
 530          * remove it and reinsert it in global hash list. We
 531          * know the walkers can't get to this conn because we
 532          * had set CONDEMNED flag earlier and checked reference
 533          * under conn_lock so walker won't pick it and when we
 534          * go the ipcl_globalhash_remove() below, no walker
 535          * can get to it.
 536          */
 537         ipcl_globalhash_remove(connp);
 538 
 539         /* Save some state */
 540         mp = tcp->tcp_timercache;
 541 
 542         tcp_rsrv_mp = tcp->tcp_rsrv_mp;
 543 
 544         if (connp->conn_cred != NULL) {
 545                 crfree(connp->conn_cred);
 546                 connp->conn_cred = NULL;
 547         }
 548         ipcl_conn_cleanup(connp);
 549         connp->conn_flags = IPCL_TCPCONN;
 550 
 551         /*
 552          * Now it is safe to decrement the reference counts.
 553          * This might be the last reference on the netstack
 554          * in which case it will cause the freeing of the IP Instance.
 555          */
 556         connp->conn_netstack = NULL;
 557         connp->conn_ixa->ixa_ipst = NULL;
 558         netstack_rele(ns);
 559         ASSERT(tcps != NULL);
 560         tcp->tcp_tcps = NULL;
 561 
 562         bzero(tcp, sizeof (tcp_t));
 563 
 564         /* restore the state */
 565         tcp->tcp_timercache = mp;
 566 
 567         tcp->tcp_rsrv_mp = tcp_rsrv_mp;
 568 
 569         tcp->tcp_connp = connp;
 570 
 571         ASSERT(connp->conn_tcp == tcp);
 572         ASSERT(connp->conn_flags & IPCL_TCPCONN);
 573         connp->conn_state_flags = CONN_INCIPIENT;
 574         ASSERT(connp->conn_proto == IPPROTO_TCP);
 575         ASSERT(connp->conn_ref == 1);
 576 }
 577 
 578 #pragma inline(tcp_calculate_rto)
 579 
 580 /*
 581  * RTO = average estimates (sa / 8) + 4 * deviation estimates (sd)
 582  *
 583  * Add tcp_rexmit_interval extra in case of extreme environment where the
 584  * algorithm fails to work.  The default value of tcp_rexmit_interval_extra
 585  * should be 0.
 586  *
 587  * As we use a finer grained clock than BSD and update RTO for every ACKs, add
 588  * in another .25 of RTT to the deviation of RTO to accommodate burstiness of
 589  * 1/4 of window size.
 590  */
 591 clock_t
 592 tcp_calculate_rto(tcp_t *tcp, tcp_stack_t *tcps)
 593 {
 594         clock_t rto;
 595 
 596         rto = NSEC2MSEC((tcp->tcp_rtt_sa >> 3) + (tcp->tcp_rtt_sa >> 5) +
 597             tcp->tcp_rtt_sd) + tcps->tcps_rexmit_interval_extra +
 598             tcps->tcps_conn_grace_period;
 599 
 600         if (rto < tcp->tcp_rto_min)
 601                 rto = tcp->tcp_rto_min;
 602         else if (rto > tcp->tcp_rto_max)
 603                 rto = tcp->tcp_rto_max;
 604 
 605         return (rto);
 606 }
 607 
 608 /*
 609  * Adapt to the information, such as rtt and rtt_sd, provided from the
 610  * DCE and IRE maintained by IP.
 611  *
 612  * Checks for multicast and broadcast destination address.
 613  * Returns zero if ok; an errno on failure.
 614  *
 615  * Note that the MSS calculation here is based on the info given in
 616  * the DCE and IRE.  We do not do any calculation based on TCP options.  They
 617  * will be handled in tcp_input_data() when TCP knows which options to use.
 618  *
 619  * Note on how TCP gets its parameters for a connection.
 620  *
 621  * When a tcp_t structure is allocated, it gets all the default parameters.
 622  * In tcp_set_destination(), it gets those metric parameters, like rtt, rtt_sd,
 623  * spipe, rpipe, ... from the route metrics.  Route metric overrides the
 624  * default.
 625  *
 626  * An incoming SYN with a multicast or broadcast destination address is dropped
 627  * in ip_fanout_v4/v6.
 628  *
 629  * An incoming SYN with a multicast or broadcast source address is always
 630  * dropped in tcp_set_destination, since IPDF_ALLOW_MCBC is not set in
 631  * conn_connect.
 632  * The same logic in tcp_set_destination also serves to
 633  * reject an attempt to connect to a broadcast or multicast (destination)
 634  * address.
 635  */
 636 int
 637 tcp_set_destination(tcp_t *tcp)
 638 {
 639         uint32_t        mss_max;
 640         uint32_t        mss;
 641         boolean_t       tcp_detached = TCP_IS_DETACHED(tcp);
 642         conn_t          *connp = tcp->tcp_connp;
 643         tcp_stack_t     *tcps = tcp->tcp_tcps;
 644         iulp_t          uinfo;
 645         int             error;
 646         uint32_t        flags;
 647 
 648         flags = IPDF_LSO | IPDF_ZCOPY;
 649         /*
 650          * Make sure we have a dce for the destination to avoid dce_ident
 651          * contention for connected sockets.
 652          */
 653         flags |= IPDF_UNIQUE_DCE;
 654 
 655         if (!tcps->tcps_ignore_path_mtu)
 656                 connp->conn_ixa->ixa_flags |= IXAF_PMTU_DISCOVERY;
 657 
 658         /* Use conn_lock to satify ASSERT; tcp is already serialized */
 659         mutex_enter(&connp->conn_lock);
 660         error = conn_connect(connp, &uinfo, flags);
 661         mutex_exit(&connp->conn_lock);
 662         if (error != 0)
 663                 return (error);
 664 
 665         error = tcp_build_hdrs(tcp);
 666         if (error != 0)
 667                 return (error);
 668 
 669         tcp->tcp_localnet = uinfo.iulp_localnet;
 670 
 671         if (uinfo.iulp_rtt != 0) {
 672                 tcp->tcp_rtt_sa = MSEC2NSEC(uinfo.iulp_rtt);
 673                 tcp->tcp_rtt_sd = MSEC2NSEC(uinfo.iulp_rtt_sd);
 674                 tcp->tcp_rto = tcp_calculate_rto(tcp, tcps);
 675         }
 676         if (uinfo.iulp_ssthresh != 0)
 677                 tcp->tcp_cwnd_ssthresh = uinfo.iulp_ssthresh;
 678         else
 679                 tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
 680         if (uinfo.iulp_spipe > 0) {
 681                 connp->conn_sndbuf = MIN(uinfo.iulp_spipe,
 682                     tcps->tcps_max_buf);
 683                 if (tcps->tcps_snd_lowat_fraction != 0) {
 684                         connp->conn_sndlowat = connp->conn_sndbuf /
 685                             tcps->tcps_snd_lowat_fraction;
 686                 }
 687                 (void) tcp_maxpsz_set(tcp, B_TRUE);
 688         }
 689         /*
 690          * Note that up till now, acceptor always inherits receive
 691          * window from the listener.  But if there is a metrics
 692          * associated with a host, we should use that instead of
 693          * inheriting it from listener. Thus we need to pass this
 694          * info back to the caller.
 695          */
 696         if (uinfo.iulp_rpipe > 0) {
 697                 tcp->tcp_rwnd = MIN(uinfo.iulp_rpipe,
 698                     tcps->tcps_max_buf);
 699         }
 700 
 701         if (uinfo.iulp_rtomax > 0) {
 702                 tcp->tcp_second_timer_threshold =
 703                     uinfo.iulp_rtomax;
 704         }
 705 
 706         /*
 707          * Use the metric option settings, iulp_tstamp_ok and
 708          * iulp_wscale_ok, only for active open. What this means
 709          * is that if the other side uses timestamp or window
 710          * scale option, TCP will also use those options. That
 711          * is for passive open.  If the application sets a
 712          * large window, window scale is enabled regardless of
 713          * the value in iulp_wscale_ok.  This is the behavior
 714          * since 2.6.  So we keep it.
 715          * The only case left in passive open processing is the
 716          * check for SACK.
 717          * For ECN, it should probably be like SACK.  But the
 718          * current value is binary, so we treat it like the other
 719          * cases.  The metric only controls active open.For passive
 720          * open, the ndd param, tcp_ecn_permitted, controls the
 721          * behavior.
 722          */
 723         if (!tcp_detached) {
 724                 /*
 725                  * The if check means that the following can only
 726                  * be turned on by the metrics only IRE, but not off.
 727                  */
 728                 if (uinfo.iulp_tstamp_ok)
 729                         tcp->tcp_snd_ts_ok = B_TRUE;
 730                 if (uinfo.iulp_wscale_ok)
 731                         tcp->tcp_snd_ws_ok = B_TRUE;
 732                 if (uinfo.iulp_sack == 2)
 733                         tcp->tcp_snd_sack_ok = B_TRUE;
 734                 if (uinfo.iulp_ecn_ok)
 735                         tcp->tcp_ecn_ok = B_TRUE;
 736         } else {
 737                 /*
 738                  * Passive open.
 739                  *
 740                  * As above, the if check means that SACK can only be
 741                  * turned on by the metric only IRE.
 742                  */
 743                 if (uinfo.iulp_sack > 0) {
 744                         tcp->tcp_snd_sack_ok = B_TRUE;
 745                 }
 746         }
 747 
 748         /*
 749          * XXX Note that currently, iulp_mtu can be as small as 68
 750          * because of PMTUd.  So tcp_mss may go to negative if combined
 751          * length of all those options exceeds 28 bytes.  But because
 752          * of the tcp_mss_min check below, we may not have a problem if
 753          * tcp_mss_min is of a reasonable value.  The default is 1 so
 754          * the negative problem still exists.  And the check defeats PMTUd.
 755          * In fact, if PMTUd finds that the MSS should be smaller than
 756          * tcp_mss_min, TCP should turn off PMUTd and use the tcp_mss_min
 757          * value.
 758          *
 759          * We do not deal with that now.  All those problems related to
 760          * PMTUd will be fixed later.
 761          */
 762         ASSERT(uinfo.iulp_mtu != 0);
 763         mss = tcp->tcp_initial_pmtu = uinfo.iulp_mtu;
 764 
 765         /* Sanity check for MSS value. */
 766         if (connp->conn_ipversion == IPV4_VERSION)
 767                 mss_max = tcps->tcps_mss_max_ipv4;
 768         else
 769                 mss_max = tcps->tcps_mss_max_ipv6;
 770 
 771         if (tcp->tcp_ipsec_overhead == 0)
 772                 tcp->tcp_ipsec_overhead = conn_ipsec_length(connp);
 773 
 774         mss -= tcp->tcp_ipsec_overhead;
 775 
 776         if (mss < tcps->tcps_mss_min)
 777                 mss = tcps->tcps_mss_min;
 778         if (mss > mss_max)
 779                 mss = mss_max;
 780 
 781         /* Note that this is the maximum MSS, excluding all options. */
 782         tcp->tcp_mss = mss;
 783 
 784         /*
 785          * Update the tcp connection with LSO capability.
 786          */
 787         tcp_update_lso(tcp, connp->conn_ixa);
 788 
 789         /*
 790          * Initialize the ISS here now that we have the full connection ID.
 791          * The RFC 1948 method of initial sequence number generation requires
 792          * knowledge of the full connection ID before setting the ISS.
 793          */
 794         tcp_iss_init(tcp);
 795 
 796         tcp->tcp_loopback = (uinfo.iulp_loopback | uinfo.iulp_local);
 797 
 798         /*
 799          * Make sure that conn is not marked incipient
 800          * for incoming connections. A blind
 801          * removal of incipient flag is cheaper than
 802          * check and removal.
 803          */
 804         mutex_enter(&connp->conn_lock);
 805         connp->conn_state_flags &= ~CONN_INCIPIENT;
 806         mutex_exit(&connp->conn_lock);
 807         return (0);
 808 }
 809 
 810 /*
 811  * tcp_clean_death / tcp_close_detached must not be called more than once
 812  * on a tcp. Thus every function that potentially calls tcp_clean_death
 813  * must check for the tcp state before calling tcp_clean_death.
 814  * Eg. tcp_input_data, tcp_eager_kill, tcp_clean_death_wrapper,
 815  * tcp_timer_handler, all check for the tcp state.
 816  */
 817 /* ARGSUSED */
 818 void
 819 tcp_clean_death_wrapper(void *arg, mblk_t *mp, void *arg2,
 820     ip_recv_attr_t *dummy)
 821 {
 822         tcp_t   *tcp = ((conn_t *)arg)->conn_tcp;
 823 
 824         freemsg(mp);
 825         if (tcp->tcp_state > TCPS_BOUND)
 826                 (void) tcp_clean_death(((conn_t *)arg)->conn_tcp, ETIMEDOUT);
 827 }
 828 
 829 /*
 830  * We are dying for some reason.  Try to do it gracefully.  (May be called
 831  * as writer.)
 832  *
 833  * Return -1 if the structure was not cleaned up (if the cleanup had to be
 834  * done by a service procedure).
 835  * TBD - Should the return value distinguish between the tcp_t being
 836  * freed and it being reinitialized?
 837  */
 838 int
 839 tcp_clean_death(tcp_t *tcp, int err)
 840 {
 841         mblk_t  *mp;
 842         queue_t *q;
 843         conn_t  *connp = tcp->tcp_connp;
 844         tcp_stack_t     *tcps = tcp->tcp_tcps;
 845 
 846         if (tcp->tcp_fused)
 847                 tcp_unfuse(tcp);
 848 
 849         if (tcp->tcp_linger_tid != 0 &&
 850             TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
 851                 tcp_stop_lingering(tcp);
 852         }
 853 
 854         ASSERT(tcp != NULL);
 855         ASSERT((connp->conn_family == AF_INET &&
 856             connp->conn_ipversion == IPV4_VERSION) ||
 857             (connp->conn_family == AF_INET6 &&
 858             (connp->conn_ipversion == IPV4_VERSION ||
 859             connp->conn_ipversion == IPV6_VERSION)));
 860 
 861         if (TCP_IS_DETACHED(tcp)) {
 862                 if (tcp->tcp_hard_binding) {
 863                         /*
 864                          * Its an eager that we are dealing with. We close the
 865                          * eager but in case a conn_ind has already gone to the
 866                          * listener, let tcp_accept_finish() send a discon_ind
 867                          * to the listener and drop the last reference. If the
 868                          * listener doesn't even know about the eager i.e. the
 869                          * conn_ind hasn't gone up, blow away the eager and drop
 870                          * the last reference as well. If the conn_ind has gone
 871                          * up, state should be BOUND. tcp_accept_finish
 872                          * will figure out that the connection has received a
 873                          * RST and will send a DISCON_IND to the application.
 874                          */
 875                         tcp_closei_local(tcp);
 876                         if (!tcp->tcp_tconnind_started) {
 877                                 CONN_DEC_REF(connp);
 878                         } else {
 879                                 tcp->tcp_state = TCPS_BOUND;
 880                                 DTRACE_TCP6(state__change, void, NULL,
 881                                     ip_xmit_attr_t *, connp->conn_ixa,
 882                                     void, NULL, tcp_t *, tcp, void, NULL,
 883                                     int32_t, TCPS_CLOSED);
 884                         }
 885                 } else {
 886                         tcp_close_detached(tcp);
 887                 }
 888                 return (0);
 889         }
 890 
 891         TCP_STAT(tcps, tcp_clean_death_nondetached);
 892 
 893         /*
 894          * The connection is dead.  Decrement listener connection counter if
 895          * necessary.
 896          */
 897         if (tcp->tcp_listen_cnt != NULL)
 898                 TCP_DECR_LISTEN_CNT(tcp);
 899 
 900         /*
 901          * When a connection is moved to TIME_WAIT state, the connection
 902          * counter is already decremented.  So no need to decrement here
 903          * again.  See SET_TIME_WAIT() macro.
 904          */
 905         if (tcp->tcp_state >= TCPS_ESTABLISHED &&
 906             tcp->tcp_state < TCPS_TIME_WAIT) {
 907                 TCPS_CONN_DEC(tcps);
 908         }
 909 
 910         q = connp->conn_rq;
 911 
 912         /* Trash all inbound data */
 913         if (!IPCL_IS_NONSTR(connp)) {
 914                 ASSERT(q != NULL);
 915                 flushq(q, FLUSHALL);
 916         }
 917 
 918         /*
 919          * If we are at least part way open and there is error
 920          * (err==0 implies no error)
 921          * notify our client by a T_DISCON_IND.
 922          */
 923         if ((tcp->tcp_state >= TCPS_SYN_SENT) && err) {
 924                 if (tcp->tcp_state >= TCPS_ESTABLISHED &&
 925                     !TCP_IS_SOCKET(tcp)) {
 926                         /*
 927                          * Send M_FLUSH according to TPI. Because sockets will
 928                          * (and must) ignore FLUSHR we do that only for TPI
 929                          * endpoints and sockets in STREAMS mode.
 930                          */
 931                         (void) putnextctl1(q, M_FLUSH, FLUSHR);
 932                 }
 933                 if (connp->conn_debug) {
 934                         (void) strlog(TCP_MOD_ID, 0, 1, SL_TRACE|SL_ERROR,
 935                             "tcp_clean_death: discon err %d", err);
 936                 }
 937                 if (IPCL_IS_NONSTR(connp)) {
 938                         /* Direct socket, use upcall */
 939                         (*connp->conn_upcalls->su_disconnected)(
 940                             connp->conn_upper_handle, tcp->tcp_connid, err);
 941                 } else {
 942                         mp = mi_tpi_discon_ind(NULL, err, 0);
 943                         if (mp != NULL) {
 944                                 putnext(q, mp);
 945                         } else {
 946                                 if (connp->conn_debug) {
 947                                         (void) strlog(TCP_MOD_ID, 0, 1,
 948                                             SL_ERROR|SL_TRACE,
 949                                             "tcp_clean_death, sending M_ERROR");
 950                                 }
 951                                 (void) putnextctl1(q, M_ERROR, EPROTO);
 952                         }
 953                 }
 954                 if (tcp->tcp_state <= TCPS_SYN_RCVD) {
 955                         /* SYN_SENT or SYN_RCVD */
 956                         TCPS_BUMP_MIB(tcps, tcpAttemptFails);
 957                 } else if (tcp->tcp_state <= TCPS_CLOSE_WAIT) {
 958                         /* ESTABLISHED or CLOSE_WAIT */
 959                         TCPS_BUMP_MIB(tcps, tcpEstabResets);
 960                 }
 961         }
 962 
 963         /*
 964          * ESTABLISHED non-STREAMS eagers are not 'detached' because
 965          * an upper handle is obtained when the SYN-ACK comes in. So it
 966          * should receive the 'disconnected' upcall, but tcp_reinit should
 967          * not be called since this is an eager.
 968          */
 969         if (tcp->tcp_listener != NULL && IPCL_IS_NONSTR(connp)) {
 970                 tcp_closei_local(tcp);
 971                 tcp->tcp_state = TCPS_BOUND;
 972                 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
 973                     connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
 974                     int32_t, TCPS_CLOSED);
 975                 return (0);
 976         }
 977 
 978         tcp_reinit(tcp);
 979         if (IPCL_IS_NONSTR(connp))
 980                 (void) tcp_do_unbind(connp);
 981 
 982         return (-1);
 983 }
 984 
 985 /*
 986  * In case tcp is in the "lingering state" and waits for the SO_LINGER timeout
 987  * to expire, stop the wait and finish the close.
 988  */
 989 void
 990 tcp_stop_lingering(tcp_t *tcp)
 991 {
 992         clock_t delta = 0;
 993         conn_t  *connp = tcp->tcp_connp;
 994 
 995         tcp->tcp_linger_tid = 0;
 996         if (tcp->tcp_state > TCPS_LISTEN) {
 997                 tcp_acceptor_hash_remove(tcp);
 998                 mutex_enter(&tcp->tcp_non_sq_lock);
 999                 if (tcp->tcp_flow_stopped) {
1000                         tcp_clrqfull(tcp);
1001                 }
1002                 mutex_exit(&tcp->tcp_non_sq_lock);
1003 
1004                 if (tcp->tcp_timer_tid != 0) {
1005                         delta = TCP_TIMER_CANCEL(tcp, tcp->tcp_timer_tid);
1006                         tcp->tcp_timer_tid = 0;
1007                 }
1008                 /*
1009                  * Need to cancel those timers which will not be used when
1010                  * TCP is detached.  This has to be done before the conn_wq
1011                  * is cleared.
1012                  */
1013                 tcp_timers_stop(tcp);
1014 
1015                 tcp->tcp_detached = B_TRUE;
1016                 connp->conn_rq = NULL;
1017                 connp->conn_wq = NULL;
1018 
1019                 if (tcp->tcp_state == TCPS_TIME_WAIT) {
1020                         tcp_time_wait_append(tcp);
1021                         TCP_DBGSTAT(tcp->tcp_tcps, tcp_detach_time_wait);
1022                         goto finish;
1023                 }
1024 
1025                 /*
1026                  * If delta is zero the timer event wasn't executed and was
1027                  * successfully canceled. In this case we need to restart it
1028                  * with the minimal delta possible.
1029                  */
1030                 if (delta >= 0) {
1031                         tcp->tcp_timer_tid = TCP_TIMER(tcp, tcp_timer,
1032                             delta ? delta : 1);
1033                 }
1034         } else {
1035                 tcp_closei_local(tcp);
1036                 CONN_DEC_REF(connp);
1037         }
1038 finish:
1039         tcp->tcp_detached = B_TRUE;
1040         connp->conn_rq = NULL;
1041         connp->conn_wq = NULL;
1042 
1043         /* Signal closing thread that it can complete close */
1044         mutex_enter(&tcp->tcp_closelock);
1045         tcp->tcp_closed = 1;
1046         cv_signal(&tcp->tcp_closecv);
1047         mutex_exit(&tcp->tcp_closelock);
1048 
1049         /* If we have an upper handle (socket), release it */
1050         if (IPCL_IS_NONSTR(connp)) {
1051                 ASSERT(connp->conn_upper_handle != NULL);
1052                 (*connp->conn_upcalls->su_closed)(connp->conn_upper_handle);
1053                 connp->conn_upper_handle = NULL;
1054                 connp->conn_upcalls = NULL;
1055         }
1056 }
1057 
1058 void
1059 tcp_close_common(conn_t *connp, int flags)
1060 {
1061         tcp_t           *tcp = connp->conn_tcp;
1062         mblk_t          *mp = &tcp->tcp_closemp;
1063         boolean_t       conn_ioctl_cleanup_reqd = B_FALSE;
1064         mblk_t          *bp;
1065 
1066         ASSERT(connp->conn_ref >= 2);
1067 
1068         /*
1069          * Mark the conn as closing. ipsq_pending_mp_add will not
1070          * add any mp to the pending mp list, after this conn has
1071          * started closing.
1072          */
1073         mutex_enter(&connp->conn_lock);
1074         connp->conn_state_flags |= CONN_CLOSING;
1075         if (connp->conn_oper_pending_ill != NULL)
1076                 conn_ioctl_cleanup_reqd = B_TRUE;
1077         CONN_INC_REF_LOCKED(connp);
1078         mutex_exit(&connp->conn_lock);
1079         tcp->tcp_closeflags = (uint8_t)flags;
1080         ASSERT(connp->conn_ref >= 3);
1081 
1082         /*
1083          * tcp_closemp_used is used below without any protection of a lock
1084          * as we don't expect any one else to use it concurrently at this
1085          * point otherwise it would be a major defect.
1086          */
1087 
1088         if (mp->b_prev == NULL)
1089                 tcp->tcp_closemp_used = B_TRUE;
1090         else
1091                 cmn_err(CE_PANIC, "tcp_close: concurrent use of tcp_closemp: "
1092                     "connp %p tcp %p\n", (void *)connp, (void *)tcp);
1093 
1094         TCP_DEBUG_GETPCSTACK(tcp->tcmp_stk, 15);
1095 
1096         /*
1097          * Cleanup any queued ioctls here. This must be done before the wq/rq
1098          * are re-written by tcp_close_output().
1099          */
1100         if (conn_ioctl_cleanup_reqd)
1101                 conn_ioctl_cleanup(connp);
1102 
1103         /*
1104          * As CONN_CLOSING is set, no further ioctls should be passed down to
1105          * IP for this conn (see the guards in tcp_ioctl, tcp_wput_ioctl and
1106          * tcp_wput_iocdata). If the ioctl was queued on an ipsq,
1107          * conn_ioctl_cleanup should have found it and removed it. If the ioctl
1108          * was still in flight at the time, we wait for it here. See comments
1109          * for CONN_INC_IOCTLREF in ip.h for details.
1110          */
1111         mutex_enter(&connp->conn_lock);
1112         while (connp->conn_ioctlref > 0)
1113                 cv_wait(&connp->conn_cv, &connp->conn_lock);
1114         ASSERT(connp->conn_ioctlref == 0);
1115         ASSERT(connp->conn_oper_pending_ill == NULL);
1116         mutex_exit(&connp->conn_lock);
1117 
1118         SQUEUE_ENTER_ONE(connp->conn_sqp, mp, tcp_close_output, connp,
1119             NULL, tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1120 
1121         /*
1122          * For non-STREAMS sockets, the normal case is that the conn makes
1123          * an upcall when it's finally closed, so there is no need to wait
1124          * in the protocol. But in case of SO_LINGER the thread sleeps here
1125          * so it can properly deal with the thread being interrupted.
1126          */
1127         if (IPCL_IS_NONSTR(connp) && connp->conn_linger == 0)
1128                 goto nowait;
1129 
1130         mutex_enter(&tcp->tcp_closelock);
1131         while (!tcp->tcp_closed) {
1132                 if (!cv_wait_sig(&tcp->tcp_closecv, &tcp->tcp_closelock)) {
1133                         /*
1134                          * The cv_wait_sig() was interrupted. We now do the
1135                          * following:
1136                          *
1137                          * 1) If the endpoint was lingering, we allow this
1138                          * to be interrupted by cancelling the linger timeout
1139                          * and closing normally.
1140                          *
1141                          * 2) Revert to calling cv_wait()
1142                          *
1143                          * We revert to using cv_wait() to avoid an
1144                          * infinite loop which can occur if the calling
1145                          * thread is higher priority than the squeue worker
1146                          * thread and is bound to the same cpu.
1147                          */
1148                         if (connp->conn_linger && connp->conn_lingertime > 0) {
1149                                 mutex_exit(&tcp->tcp_closelock);
1150                                 /* Entering squeue, bump ref count. */
1151                                 CONN_INC_REF(connp);
1152                                 bp = allocb_wait(0, BPRI_HI, STR_NOSIG, NULL);
1153                                 SQUEUE_ENTER_ONE(connp->conn_sqp, bp,
1154                                     tcp_linger_interrupted, connp, NULL,
1155                                     tcp_squeue_flag, SQTAG_IP_TCP_CLOSE);
1156                                 mutex_enter(&tcp->tcp_closelock);
1157                         }
1158                         break;
1159                 }
1160         }
1161         while (!tcp->tcp_closed)
1162                 cv_wait(&tcp->tcp_closecv, &tcp->tcp_closelock);
1163         mutex_exit(&tcp->tcp_closelock);
1164 
1165         /*
1166          * In the case of listener streams that have eagers in the q or q0
1167          * we wait for the eagers to drop their reference to us. conn_rq and
1168          * conn_wq of the eagers point to our queues. By waiting for the
1169          * refcnt to drop to 1, we are sure that the eagers have cleaned
1170          * up their queue pointers and also dropped their references to us.
1171          *
1172          * For non-STREAMS sockets we do not have to wait here; the
1173          * listener will instead make a su_closed upcall when the last
1174          * reference is dropped.
1175          */
1176         if (tcp->tcp_wait_for_eagers && !IPCL_IS_NONSTR(connp)) {
1177                 mutex_enter(&connp->conn_lock);
1178                 while (connp->conn_ref != 1) {
1179                         cv_wait(&connp->conn_cv, &connp->conn_lock);
1180                 }
1181                 mutex_exit(&connp->conn_lock);
1182         }
1183 
1184 nowait:
1185         connp->conn_cpid = NOPID;
1186 }
1187 
1188 /*
1189  * Called by tcp_close() routine via squeue when lingering is
1190  * interrupted by a signal.
1191  */
1192 
1193 /* ARGSUSED */
1194 static void
1195 tcp_linger_interrupted(void *arg, mblk_t *mp, void *arg2, ip_recv_attr_t *dummy)
1196 {
1197         conn_t  *connp = (conn_t *)arg;
1198         tcp_t   *tcp = connp->conn_tcp;
1199 
1200         freeb(mp);
1201         if (tcp->tcp_linger_tid != 0 &&
1202             TCP_TIMER_CANCEL(tcp, tcp->tcp_linger_tid) >= 0) {
1203                 tcp_stop_lingering(tcp);
1204                 tcp->tcp_client_errno = EINTR;
1205         }
1206 }
1207 
1208 /*
1209  * Clean up the b_next and b_prev fields of every mblk pointed at by *mpp.
1210  * Some stream heads get upset if they see these later on as anything but NULL.
1211  */
1212 void
1213 tcp_close_mpp(mblk_t **mpp)
1214 {
1215         mblk_t  *mp;
1216 
1217         if ((mp = *mpp) != NULL) {
1218                 do {
1219                         mp->b_next = NULL;
1220                         mp->b_prev = NULL;
1221                 } while ((mp = mp->b_cont) != NULL);
1222 
1223                 mp = *mpp;
1224                 *mpp = NULL;
1225                 freemsg(mp);
1226         }
1227 }
1228 
1229 /* Do detached close. */
1230 void
1231 tcp_close_detached(tcp_t *tcp)
1232 {
1233         if (tcp->tcp_fused)
1234                 tcp_unfuse(tcp);
1235 
1236         /*
1237          * Clustering code serializes TCP disconnect callbacks and
1238          * cluster tcp list walks by blocking a TCP disconnect callback
1239          * if a cluster tcp list walk is in progress. This ensures
1240          * accurate accounting of TCPs in the cluster code even though
1241          * the TCP list walk itself is not atomic.
1242          */
1243         tcp_closei_local(tcp);
1244         CONN_DEC_REF(tcp->tcp_connp);
1245 }
1246 
1247 /*
1248  * The tcp_t is going away. Remove it from all lists and set it
1249  * to TCPS_CLOSED. The freeing up of memory is deferred until
1250  * tcp_inactive. This is needed since a thread in tcp_rput might have
1251  * done a CONN_INC_REF on this structure before it was removed from the
1252  * hashes.
1253  */
1254 void
1255 tcp_closei_local(tcp_t *tcp)
1256 {
1257         conn_t          *connp = tcp->tcp_connp;
1258         tcp_stack_t     *tcps = tcp->tcp_tcps;
1259         int32_t         oldstate;
1260 
1261         if (!TCP_IS_SOCKET(tcp))
1262                 tcp_acceptor_hash_remove(tcp);
1263 
1264         /*
1265          * This can be called via tcp_time_wait_processing() if TCP gets a
1266          * SYN with sequence number outside the TIME-WAIT connection's
1267          * window.  So we need to check for TIME-WAIT state here as the
1268          * connection counter is already decremented.  See SET_TIME_WAIT()
1269          * macro
1270          */
1271         if (tcp->tcp_state >= TCPS_ESTABLISHED &&
1272             tcp->tcp_state < TCPS_TIME_WAIT) {
1273                 TCPS_CONN_DEC(tcps);
1274         }
1275 
1276         /*
1277          * If we are an eager connection hanging off a listener that
1278          * hasn't formally accepted the connection yet, get off its
1279          * list and blow off any data that we have accumulated.
1280          */
1281         if (tcp->tcp_listener != NULL) {
1282                 tcp_t   *listener = tcp->tcp_listener;
1283                 mutex_enter(&listener->tcp_eager_lock);
1284                 /*
1285                  * tcp_tconnind_started == B_TRUE means that the
1286                  * conn_ind has already gone to listener. At
1287                  * this point, eager will be closed but we
1288                  * leave it in listeners eager list so that
1289                  * if listener decides to close without doing
1290                  * accept, we can clean this up. In tcp_tli_accept
1291                  * we take care of the case of accept on closed
1292                  * eager.
1293                  */
1294                 if (!tcp->tcp_tconnind_started) {
1295                         tcp_eager_unlink(tcp);
1296                         mutex_exit(&listener->tcp_eager_lock);
1297                         /*
1298                          * We don't want to have any pointers to the
1299                          * listener queue, after we have released our
1300                          * reference on the listener
1301                          */
1302                         ASSERT(tcp->tcp_detached);
1303                         connp->conn_rq = NULL;
1304                         connp->conn_wq = NULL;
1305                         CONN_DEC_REF(listener->tcp_connp);
1306                 } else {
1307                         mutex_exit(&listener->tcp_eager_lock);
1308                 }
1309         }
1310 
1311         /* Stop all the timers */
1312         tcp_timers_stop(tcp);
1313 
1314         if (tcp->tcp_state == TCPS_LISTEN) {
1315                 if (tcp->tcp_ip_addr_cache) {
1316                         kmem_free((void *)tcp->tcp_ip_addr_cache,
1317                             IP_ADDR_CACHE_SIZE * sizeof (ipaddr_t));
1318                         tcp->tcp_ip_addr_cache = NULL;
1319                 }
1320         }
1321 
1322         /* Decrement listerner connection counter if necessary. */
1323         if (tcp->tcp_listen_cnt != NULL)
1324                 TCP_DECR_LISTEN_CNT(tcp);
1325 
1326         mutex_enter(&tcp->tcp_non_sq_lock);
1327         if (tcp->tcp_flow_stopped)
1328                 tcp_clrqfull(tcp);
1329         mutex_exit(&tcp->tcp_non_sq_lock);
1330 
1331         tcp_bind_hash_remove(tcp);
1332         /*
1333          * If the tcp_time_wait_collector (which runs outside the squeue)
1334          * is trying to remove this tcp from the time wait list, we will
1335          * block in tcp_time_wait_remove while trying to acquire the
1336          * tcp_time_wait_lock. The logic in tcp_time_wait_collector also
1337          * requires the ipcl_hash_remove to be ordered after the
1338          * tcp_time_wait_remove for the refcnt checks to work correctly.
1339          */
1340         if (tcp->tcp_state == TCPS_TIME_WAIT)
1341                 (void) tcp_time_wait_remove(tcp, NULL);
1342         CL_INET_DISCONNECT(connp);
1343         ipcl_hash_remove(connp);
1344         oldstate = tcp->tcp_state;
1345         tcp->tcp_state = TCPS_CLOSED;
1346         /* Need to probe before ixa_cleanup() is called */
1347         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1348             connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
1349             int32_t, oldstate);
1350         ixa_cleanup(connp->conn_ixa);
1351 
1352         /*
1353          * Mark the conn as CONDEMNED
1354          */
1355         mutex_enter(&connp->conn_lock);
1356         connp->conn_state_flags |= CONN_CONDEMNED;
1357         mutex_exit(&connp->conn_lock);
1358 
1359         ASSERT(tcp->tcp_time_wait_next == NULL);
1360         ASSERT(tcp->tcp_time_wait_prev == NULL);
1361         ASSERT(tcp->tcp_time_wait_expire == 0);
1362 
1363         tcp_ipsec_cleanup(tcp);
1364 }
1365 
1366 /*
1367  * tcp is dying (called from ipcl_conn_destroy and error cases).
1368  * Free the tcp_t in either case.
1369  */
1370 void
1371 tcp_free(tcp_t *tcp)
1372 {
1373         mblk_t          *mp;
1374         conn_t          *connp = tcp->tcp_connp;
1375 
1376         ASSERT(tcp != NULL);
1377         ASSERT(tcp->tcp_ptpahn == NULL && tcp->tcp_acceptor_hash == NULL);
1378 
1379         connp->conn_rq = NULL;
1380         connp->conn_wq = NULL;
1381 
1382         tcp_close_mpp(&tcp->tcp_xmit_head);
1383         tcp_close_mpp(&tcp->tcp_reass_head);
1384         if (tcp->tcp_rcv_list != NULL) {
1385                 /* Free b_next chain */
1386                 tcp_close_mpp(&tcp->tcp_rcv_list);
1387         }
1388         if ((mp = tcp->tcp_urp_mp) != NULL) {
1389                 freemsg(mp);
1390         }
1391         if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1392                 freemsg(mp);
1393         }
1394 
1395         if (tcp->tcp_fused_sigurg_mp != NULL) {
1396                 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1397                 freeb(tcp->tcp_fused_sigurg_mp);
1398                 tcp->tcp_fused_sigurg_mp = NULL;
1399         }
1400 
1401         if (tcp->tcp_ordrel_mp != NULL) {
1402                 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1403                 freeb(tcp->tcp_ordrel_mp);
1404                 tcp->tcp_ordrel_mp = NULL;
1405         }
1406 
1407         TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
1408         bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
1409 
1410         if (tcp->tcp_hopopts != NULL) {
1411                 mi_free(tcp->tcp_hopopts);
1412                 tcp->tcp_hopopts = NULL;
1413                 tcp->tcp_hopoptslen = 0;
1414         }
1415         ASSERT(tcp->tcp_hopoptslen == 0);
1416         if (tcp->tcp_dstopts != NULL) {
1417                 mi_free(tcp->tcp_dstopts);
1418                 tcp->tcp_dstopts = NULL;
1419                 tcp->tcp_dstoptslen = 0;
1420         }
1421         ASSERT(tcp->tcp_dstoptslen == 0);
1422         if (tcp->tcp_rthdrdstopts != NULL) {
1423                 mi_free(tcp->tcp_rthdrdstopts);
1424                 tcp->tcp_rthdrdstopts = NULL;
1425                 tcp->tcp_rthdrdstoptslen = 0;
1426         }
1427         ASSERT(tcp->tcp_rthdrdstoptslen == 0);
1428         if (tcp->tcp_rthdr != NULL) {
1429                 mi_free(tcp->tcp_rthdr);
1430                 tcp->tcp_rthdr = NULL;
1431                 tcp->tcp_rthdrlen = 0;
1432         }
1433         ASSERT(tcp->tcp_rthdrlen == 0);
1434 
1435         /*
1436          * Following is really a blowing away a union.
1437          * It happens to have exactly two members of identical size
1438          * the following code is enough.
1439          */
1440         tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1441 
1442         /* Allow the CC algorithm to clean up after itself. */
1443         if (tcp->tcp_cc_algo != NULL && tcp->tcp_cc_algo->cb_destroy != NULL)
1444                 tcp->tcp_cc_algo->cb_destroy(&tcp->tcp_ccv);
1445 
1446         /*
1447          * Destroy any association with SO_REUSEPORT group.
1448          */
1449         if (tcp->tcp_rg_bind != NULL) {
1450                 /*
1451                  * This is only necessary for connections which enabled
1452                  * SO_REUSEPORT but were never bound.  Such connections should
1453                  * be the one and only member of the tcp_rg_tp to which they
1454                  * have been associated.
1455                  */
1456                 VERIFY(tcp_rg_remove(tcp->tcp_rg_bind, tcp));
1457                 tcp_rg_destroy(tcp->tcp_rg_bind);
1458                 tcp->tcp_rg_bind = NULL;
1459         }
1460 
1461         /*
1462          * If this is a non-STREAM socket still holding on to an upper
1463          * handle, release it. As a result of fallback we might also see
1464          * STREAMS based conns with upper handles, in which case there is
1465          * nothing to do other than clearing the field.
1466          */
1467         if (connp->conn_upper_handle != NULL) {
1468                 if (IPCL_IS_NONSTR(connp)) {
1469                         (*connp->conn_upcalls->su_closed)(
1470                             connp->conn_upper_handle);
1471                         tcp->tcp_detached = B_TRUE;
1472                 }
1473                 connp->conn_upper_handle = NULL;
1474                 connp->conn_upcalls = NULL;
1475         }
1476 }
1477 
1478 /*
1479  * tcp_get_conn/tcp_free_conn
1480  *
1481  * tcp_get_conn is used to get a clean tcp connection structure.
1482  * It tries to reuse the connections put on the freelist by the
1483  * time_wait_collector failing which it goes to kmem_cache. This
1484  * way has two benefits compared to just allocating from and
1485  * freeing to kmem_cache.
1486  * 1) The time_wait_collector can free (which includes the cleanup)
1487  * outside the squeue. So when the interrupt comes, we have a clean
1488  * connection sitting in the freelist. Obviously, this buys us
1489  * performance.
1490  *
1491  * 2) Defence against DOS attack. Allocating a tcp/conn in tcp_input_listener
1492  * has multiple disadvantages - tying up the squeue during alloc.
1493  * But allocating the conn/tcp in IP land is also not the best since
1494  * we can't check the 'q' and 'q0' which are protected by squeue and
1495  * blindly allocate memory which might have to be freed here if we are
1496  * not allowed to accept the connection. By using the freelist and
1497  * putting the conn/tcp back in freelist, we don't pay a penalty for
1498  * allocating memory without checking 'q/q0' and freeing it if we can't
1499  * accept the connection.
1500  *
1501  * Care should be taken to put the conn back in the same squeue's freelist
1502  * from which it was allocated. Best results are obtained if conn is
1503  * allocated from listener's squeue and freed to the same. Time wait
1504  * collector will free up the freelist is the connection ends up sitting
1505  * there for too long.
1506  */
1507 conn_t *
1508 tcp_get_conn(void *arg, tcp_stack_t *tcps)
1509 {
1510         tcp_t                   *tcp = NULL;
1511         conn_t                  *connp = NULL;
1512         squeue_t                *sqp = (squeue_t *)arg;
1513         tcp_squeue_priv_t       *tcp_time_wait;
1514         netstack_t              *ns;
1515         mblk_t                  *tcp_rsrv_mp = NULL;
1516 
1517         tcp_time_wait =
1518             *((tcp_squeue_priv_t **)squeue_getprivate(sqp, SQPRIVATE_TCP));
1519 
1520         mutex_enter(&tcp_time_wait->tcp_time_wait_lock);
1521         tcp = tcp_time_wait->tcp_free_list;
1522         ASSERT((tcp != NULL) ^ (tcp_time_wait->tcp_free_list_cnt == 0));
1523         if (tcp != NULL) {
1524                 tcp_time_wait->tcp_free_list = tcp->tcp_time_wait_next;
1525                 tcp_time_wait->tcp_free_list_cnt--;
1526                 mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1527                 tcp->tcp_time_wait_next = NULL;
1528                 connp = tcp->tcp_connp;
1529                 connp->conn_flags |= IPCL_REUSED;
1530 
1531                 ASSERT(tcp->tcp_tcps == NULL);
1532                 ASSERT(connp->conn_netstack == NULL);
1533                 ASSERT(tcp->tcp_rsrv_mp != NULL);
1534                 ns = tcps->tcps_netstack;
1535                 netstack_hold(ns);
1536                 connp->conn_netstack = ns;
1537                 connp->conn_ixa->ixa_ipst = ns->netstack_ip;
1538                 tcp->tcp_tcps = tcps;
1539                 ipcl_globalhash_insert(connp);
1540 
1541                 connp->conn_ixa->ixa_notify_cookie = tcp;
1542                 ASSERT(connp->conn_ixa->ixa_notify == tcp_notify);
1543                 connp->conn_recv = tcp_input_data;
1544                 ASSERT(connp->conn_recvicmp == tcp_icmp_input);
1545                 ASSERT(connp->conn_verifyicmp == tcp_verifyicmp);
1546                 return (connp);
1547         }
1548         mutex_exit(&tcp_time_wait->tcp_time_wait_lock);
1549         /*
1550          * Pre-allocate the tcp_rsrv_mp. This mblk will not be freed until
1551          * this conn_t/tcp_t is freed at ipcl_conn_destroy().
1552          */
1553         tcp_rsrv_mp = allocb(0, BPRI_HI);
1554         if (tcp_rsrv_mp == NULL)
1555                 return (NULL);
1556 
1557         if ((connp = ipcl_conn_create(IPCL_TCPCONN, KM_NOSLEEP,
1558             tcps->tcps_netstack)) == NULL) {
1559                 freeb(tcp_rsrv_mp);
1560                 return (NULL);
1561         }
1562 
1563         tcp = connp->conn_tcp;
1564         tcp->tcp_rsrv_mp = tcp_rsrv_mp;
1565         mutex_init(&tcp->tcp_rsrv_mp_lock, NULL, MUTEX_DEFAULT, NULL);
1566 
1567         tcp->tcp_tcps = tcps;
1568 
1569         connp->conn_recv = tcp_input_data;
1570         connp->conn_recvicmp = tcp_icmp_input;
1571         connp->conn_verifyicmp = tcp_verifyicmp;
1572 
1573         /*
1574          * Register tcp_notify to listen to capability changes detected by IP.
1575          * This upcall is made in the context of the call to conn_ip_output
1576          * thus it is inside the squeue.
1577          */
1578         connp->conn_ixa->ixa_notify = tcp_notify;
1579         connp->conn_ixa->ixa_notify_cookie = tcp;
1580 
1581         return (connp);
1582 }
1583 
1584 /*
1585  * Handle connect to IPv4 destinations, including connections for AF_INET6
1586  * sockets connecting to IPv4 mapped IPv6 destinations.
1587  * Returns zero if OK, a positive errno, or a negative TLI error.
1588  */
1589 static int
1590 tcp_connect_ipv4(tcp_t *tcp, ipaddr_t *dstaddrp, in_port_t dstport,
1591     uint_t srcid)
1592 {
1593         ipaddr_t        dstaddr = *dstaddrp;
1594         uint16_t        lport;
1595         conn_t          *connp = tcp->tcp_connp;
1596         tcp_stack_t     *tcps = tcp->tcp_tcps;
1597         int             error;
1598 
1599         ASSERT(connp->conn_ipversion == IPV4_VERSION);
1600 
1601         /* Check for attempt to connect to INADDR_ANY */
1602         if (dstaddr == INADDR_ANY)  {
1603                 /*
1604                  * SunOS 4.x and 4.3 BSD allow an application
1605                  * to connect a TCP socket to INADDR_ANY.
1606                  * When they do this, the kernel picks the
1607                  * address of one interface and uses it
1608                  * instead.  The kernel usually ends up
1609                  * picking the address of the loopback
1610                  * interface.  This is an undocumented feature.
1611                  * However, we provide the same thing here
1612                  * in order to have source and binary
1613                  * compatibility with SunOS 4.x.
1614                  * Update the T_CONN_REQ (sin/sin6) since it is used to
1615                  * generate the T_CONN_CON.
1616                  */
1617                 dstaddr = htonl(INADDR_LOOPBACK);
1618                 *dstaddrp = dstaddr;
1619         }
1620 
1621         /* Handle __sin6_src_id if socket not bound to an IP address */
1622         if (srcid != 0 && connp->conn_laddr_v4 == INADDR_ANY) {
1623                 if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1624                     IPCL_ZONEID(connp), B_TRUE, tcps->tcps_netstack)) {
1625                         /* Mismatch - conn_laddr_v6 would be v6 address. */
1626                         return (EADDRNOTAVAIL);
1627                 }
1628                 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1629         }
1630 
1631         IN6_IPADDR_TO_V4MAPPED(dstaddr, &connp->conn_faddr_v6);
1632         connp->conn_fport = dstport;
1633 
1634         /*
1635          * At this point the remote destination address and remote port fields
1636          * in the tcp-four-tuple have been filled in the tcp structure. Now we
1637          * have to see which state tcp was in so we can take appropriate action.
1638          */
1639         if (tcp->tcp_state == TCPS_IDLE) {
1640                 /*
1641                  * We support a quick connect capability here, allowing
1642                  * clients to transition directly from IDLE to SYN_SENT
1643                  * tcp_bindi will pick an unused port, insert the connection
1644                  * in the bind hash and transition to BOUND state.
1645                  */
1646                 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1647                     tcp, B_TRUE);
1648                 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1649                     B_FALSE, B_FALSE);
1650                 if (lport == 0)
1651                         return (-TNOADDR);
1652         }
1653 
1654         /*
1655          * Lookup the route to determine a source address and the uinfo.
1656          * Setup TCP parameters based on the metrics/DCE.
1657          */
1658         error = tcp_set_destination(tcp);
1659         if (error != 0)
1660                 return (error);
1661 
1662         /*
1663          * Don't let an endpoint connect to itself.
1664          */
1665         if (connp->conn_faddr_v4 == connp->conn_laddr_v4 &&
1666             connp->conn_fport == connp->conn_lport)
1667                 return (-TBADADDR);
1668 
1669         tcp->tcp_state = TCPS_SYN_SENT;
1670 
1671         return (ipcl_conn_insert_v4(connp));
1672 }
1673 
1674 /*
1675  * Handle connect to IPv6 destinations.
1676  * Returns zero if OK, a positive errno, or a negative TLI error.
1677  */
1678 static int
1679 tcp_connect_ipv6(tcp_t *tcp, in6_addr_t *dstaddrp, in_port_t dstport,
1680     uint32_t flowinfo, uint_t srcid, uint32_t scope_id)
1681 {
1682         uint16_t        lport;
1683         conn_t          *connp = tcp->tcp_connp;
1684         tcp_stack_t     *tcps = tcp->tcp_tcps;
1685         int             error;
1686 
1687         ASSERT(connp->conn_family == AF_INET6);
1688 
1689         /*
1690          * If we're here, it means that the destination address is a native
1691          * IPv6 address.  Return an error if conn_ipversion is not IPv6.  A
1692          * reason why it might not be IPv6 is if the socket was bound to an
1693          * IPv4-mapped IPv6 address.
1694          */
1695         if (connp->conn_ipversion != IPV6_VERSION)
1696                 return (-TBADADDR);
1697 
1698         /*
1699          * Interpret a zero destination to mean loopback.
1700          * Update the T_CONN_REQ (sin/sin6) since it is used to
1701          * generate the T_CONN_CON.
1702          */
1703         if (IN6_IS_ADDR_UNSPECIFIED(dstaddrp))
1704                 *dstaddrp = ipv6_loopback;
1705 
1706         /* Handle __sin6_src_id if socket not bound to an IP address */
1707         if (srcid != 0 && IN6_IS_ADDR_UNSPECIFIED(&connp->conn_laddr_v6)) {
1708                 if (!ip_srcid_find_id(srcid, &connp->conn_laddr_v6,
1709                     IPCL_ZONEID(connp), B_FALSE, tcps->tcps_netstack)) {
1710                         /* Mismatch - conn_laddr_v6 would be v4-mapped. */
1711                         return (EADDRNOTAVAIL);
1712                 }
1713                 connp->conn_saddr_v6 = connp->conn_laddr_v6;
1714         }
1715 
1716         /*
1717          * Take care of the scope_id now.
1718          */
1719         if (scope_id != 0 && IN6_IS_ADDR_LINKSCOPE(dstaddrp)) {
1720                 connp->conn_ixa->ixa_flags |= IXAF_SCOPEID_SET;
1721                 connp->conn_ixa->ixa_scopeid = scope_id;
1722         } else {
1723                 connp->conn_ixa->ixa_flags &= ~IXAF_SCOPEID_SET;
1724         }
1725 
1726         connp->conn_flowinfo = flowinfo;
1727         connp->conn_faddr_v6 = *dstaddrp;
1728         connp->conn_fport = dstport;
1729 
1730         /*
1731          * At this point the remote destination address and remote port fields
1732          * in the tcp-four-tuple have been filled in the tcp structure. Now we
1733          * have to see which state tcp was in so we can take appropriate action.
1734          */
1735         if (tcp->tcp_state == TCPS_IDLE) {
1736                 /*
1737                  * We support a quick connect capability here, allowing
1738                  * clients to transition directly from IDLE to SYN_SENT
1739                  * tcp_bindi will pick an unused port, insert the connection
1740                  * in the bind hash and transition to BOUND state.
1741                  */
1742                 lport = tcp_update_next_port(tcps->tcps_next_port_to_try,
1743                     tcp, B_TRUE);
1744                 lport = tcp_bindi(tcp, lport, &connp->conn_laddr_v6, 0, B_TRUE,
1745                     B_FALSE, B_FALSE);
1746                 if (lport == 0)
1747                         return (-TNOADDR);
1748         }
1749 
1750         /*
1751          * Lookup the route to determine a source address and the uinfo.
1752          * Setup TCP parameters based on the metrics/DCE.
1753          */
1754         error = tcp_set_destination(tcp);
1755         if (error != 0)
1756                 return (error);
1757 
1758         /*
1759          * Don't let an endpoint connect to itself.
1760          */
1761         if (IN6_ARE_ADDR_EQUAL(&connp->conn_faddr_v6, &connp->conn_laddr_v6) &&
1762             connp->conn_fport == connp->conn_lport)
1763                 return (-TBADADDR);
1764 
1765         tcp->tcp_state = TCPS_SYN_SENT;
1766 
1767         return (ipcl_conn_insert_v6(connp));
1768 }
1769 
1770 /*
1771  * Disconnect
1772  * Note that unlike other functions this returns a positive tli error
1773  * when it fails; it never returns an errno.
1774  */
1775 static int
1776 tcp_disconnect_common(tcp_t *tcp, t_scalar_t seqnum)
1777 {
1778         conn_t          *lconnp;
1779         tcp_stack_t     *tcps = tcp->tcp_tcps;
1780         conn_t          *connp = tcp->tcp_connp;
1781 
1782         /*
1783          * Right now, upper modules pass down a T_DISCON_REQ to TCP,
1784          * when the stream is in BOUND state. Do not send a reset,
1785          * since the destination IP address is not valid, and it can
1786          * be the initialized value of all zeros (broadcast address).
1787          */
1788         if (tcp->tcp_state <= TCPS_BOUND) {
1789                 if (connp->conn_debug) {
1790                         (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
1791                             "tcp_disconnect: bad state, %d", tcp->tcp_state);
1792                 }
1793                 return (TOUTSTATE);
1794         } else if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1795                 TCPS_CONN_DEC(tcps);
1796         }
1797 
1798         if (seqnum == -1 || tcp->tcp_conn_req_max == 0) {
1799 
1800                 /*
1801                  * According to TPI, for non-listeners, ignore seqnum
1802                  * and disconnect.
1803                  * Following interpretation of -1 seqnum is historical
1804                  * and implied TPI ? (TPI only states that for T_CONN_IND,
1805                  * a valid seqnum should not be -1).
1806                  *
1807                  *      -1 means disconnect everything
1808                  *      regardless even on a listener.
1809                  */
1810 
1811                 int old_state = tcp->tcp_state;
1812                 ip_stack_t *ipst = tcps->tcps_netstack->netstack_ip;
1813 
1814                 /*
1815                  * The connection can't be on the tcp_time_wait_head list
1816                  * since it is not detached.
1817                  */
1818                 ASSERT(tcp->tcp_time_wait_next == NULL);
1819                 ASSERT(tcp->tcp_time_wait_prev == NULL);
1820                 ASSERT(tcp->tcp_time_wait_expire == 0);
1821                 /*
1822                  * If it used to be a listener, check to make sure no one else
1823                  * has taken the port before switching back to LISTEN state.
1824                  */
1825                 if (connp->conn_ipversion == IPV4_VERSION) {
1826                         lconnp = ipcl_lookup_listener_v4(connp->conn_lport,
1827                             connp->conn_laddr_v4, IPCL_ZONEID(connp), ipst);
1828                 } else {
1829                         uint_t ifindex = 0;
1830 
1831                         if (connp->conn_ixa->ixa_flags & IXAF_SCOPEID_SET)
1832                                 ifindex = connp->conn_ixa->ixa_scopeid;
1833 
1834                         /* Allow conn_bound_if listeners? */
1835                         lconnp = ipcl_lookup_listener_v6(connp->conn_lport,
1836                             &connp->conn_laddr_v6, ifindex, IPCL_ZONEID(connp),
1837                             ipst);
1838                 }
1839                 if (tcp->tcp_conn_req_max && lconnp == NULL) {
1840                         tcp->tcp_state = TCPS_LISTEN;
1841                         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1842                             connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1843                             NULL, int32_t, old_state);
1844                 } else if (old_state > TCPS_BOUND) {
1845                         tcp->tcp_conn_req_max = 0;
1846                         tcp->tcp_state = TCPS_BOUND;
1847                         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
1848                             connp->conn_ixa, void, NULL, tcp_t *, tcp, void,
1849                             NULL, int32_t, old_state);
1850 
1851                         /*
1852                          * If this end point is not going to become a listener,
1853                          * decrement the listener connection count if
1854                          * necessary.  Note that we do not do this if it is
1855                          * going to be a listner (the above if case) since
1856                          * then it may remove the counter struct.
1857                          */
1858                         if (tcp->tcp_listen_cnt != NULL)
1859                                 TCP_DECR_LISTEN_CNT(tcp);
1860                 }
1861                 if (lconnp != NULL)
1862                         CONN_DEC_REF(lconnp);
1863                 switch (old_state) {
1864                 case TCPS_SYN_SENT:
1865                 case TCPS_SYN_RCVD:
1866                         TCPS_BUMP_MIB(tcps, tcpAttemptFails);
1867                         break;
1868                 case TCPS_ESTABLISHED:
1869                 case TCPS_CLOSE_WAIT:
1870                         TCPS_BUMP_MIB(tcps, tcpEstabResets);
1871                         break;
1872                 }
1873 
1874                 if (tcp->tcp_fused)
1875                         tcp_unfuse(tcp);
1876 
1877                 mutex_enter(&tcp->tcp_eager_lock);
1878                 if ((tcp->tcp_conn_req_cnt_q0 != 0) ||
1879                     (tcp->tcp_conn_req_cnt_q != 0)) {
1880                         tcp_eager_cleanup(tcp, 0);
1881                 }
1882                 mutex_exit(&tcp->tcp_eager_lock);
1883 
1884                 tcp_xmit_ctl("tcp_disconnect", tcp, tcp->tcp_snxt,
1885                     tcp->tcp_rnxt, TH_RST | TH_ACK);
1886 
1887                 tcp_reinit(tcp);
1888 
1889                 return (0);
1890         } else if (!tcp_eager_blowoff(tcp, seqnum)) {
1891                 return (TBADSEQ);
1892         }
1893         return (0);
1894 }
1895 
1896 /*
1897  * Our client hereby directs us to reject the connection request
1898  * that tcp_input_listener() marked with 'seqnum'.  Rejection consists
1899  * of sending the appropriate RST, not an ICMP error.
1900  */
1901 void
1902 tcp_disconnect(tcp_t *tcp, mblk_t *mp)
1903 {
1904         t_scalar_t seqnum;
1905         int     error;
1906         conn_t  *connp = tcp->tcp_connp;
1907 
1908         ASSERT((uintptr_t)(mp->b_wptr - mp->b_rptr) <= (uintptr_t)INT_MAX);
1909         if ((mp->b_wptr - mp->b_rptr) < sizeof (struct T_discon_req)) {
1910                 tcp_err_ack(tcp, mp, TPROTO, 0);
1911                 return;
1912         }
1913         seqnum = ((struct T_discon_req *)mp->b_rptr)->SEQ_number;
1914         error = tcp_disconnect_common(tcp, seqnum);
1915         if (error != 0)
1916                 tcp_err_ack(tcp, mp, error, 0);
1917         else {
1918                 if (tcp->tcp_state >= TCPS_ESTABLISHED) {
1919                         /* Send M_FLUSH according to TPI */
1920                         (void) putnextctl1(connp->conn_rq, M_FLUSH, FLUSHRW);
1921                 }
1922                 mp = mi_tpi_ok_ack_alloc(mp);
1923                 if (mp != NULL)
1924                         putnext(connp->conn_rq, mp);
1925         }
1926 }
1927 
1928 /*
1929  * Handle reinitialization of a tcp structure.
1930  * Maintain "binding state" resetting the state to BOUND, LISTEN, or IDLE.
1931  */
1932 static void
1933 tcp_reinit(tcp_t *tcp)
1934 {
1935         mblk_t          *mp;
1936         tcp_stack_t     *tcps = tcp->tcp_tcps;
1937         conn_t          *connp  = tcp->tcp_connp;
1938         int32_t         oldstate;
1939 
1940         /* tcp_reinit should never be called for detached tcp_t's */
1941         ASSERT(tcp->tcp_listener == NULL);
1942         ASSERT((connp->conn_family == AF_INET &&
1943             connp->conn_ipversion == IPV4_VERSION) ||
1944             (connp->conn_family == AF_INET6 &&
1945             (connp->conn_ipversion == IPV4_VERSION ||
1946             connp->conn_ipversion == IPV6_VERSION)));
1947 
1948         /* Cancel outstanding timers */
1949         tcp_timers_stop(tcp);
1950 
1951         tcp_close_mpp(&tcp->tcp_xmit_head);
1952         if (tcp->tcp_snd_zcopy_aware)
1953                 tcp_zcopy_notify(tcp);
1954         tcp->tcp_xmit_last = tcp->tcp_xmit_tail = NULL;
1955         tcp->tcp_unsent = tcp->tcp_xmit_tail_unsent = 0;
1956         mutex_enter(&tcp->tcp_non_sq_lock);
1957         if (tcp->tcp_flow_stopped &&
1958             TCP_UNSENT_BYTES(tcp) <= connp->conn_sndlowat) {
1959                 tcp_clrqfull(tcp);
1960         }
1961         mutex_exit(&tcp->tcp_non_sq_lock);
1962         tcp_close_mpp(&tcp->tcp_reass_head);
1963         tcp->tcp_reass_tail = NULL;
1964         if (tcp->tcp_rcv_list != NULL) {
1965                 /* Free b_next chain */
1966                 tcp_close_mpp(&tcp->tcp_rcv_list);
1967                 tcp->tcp_rcv_last_head = NULL;
1968                 tcp->tcp_rcv_last_tail = NULL;
1969                 tcp->tcp_rcv_cnt = 0;
1970         }
1971         tcp->tcp_rcv_last_tail = NULL;
1972 
1973         if ((mp = tcp->tcp_urp_mp) != NULL) {
1974                 freemsg(mp);
1975                 tcp->tcp_urp_mp = NULL;
1976         }
1977         if ((mp = tcp->tcp_urp_mark_mp) != NULL) {
1978                 freemsg(mp);
1979                 tcp->tcp_urp_mark_mp = NULL;
1980         }
1981         if (tcp->tcp_fused_sigurg_mp != NULL) {
1982                 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1983                 freeb(tcp->tcp_fused_sigurg_mp);
1984                 tcp->tcp_fused_sigurg_mp = NULL;
1985         }
1986         if (tcp->tcp_ordrel_mp != NULL) {
1987                 ASSERT(!IPCL_IS_NONSTR(tcp->tcp_connp));
1988                 freeb(tcp->tcp_ordrel_mp);
1989                 tcp->tcp_ordrel_mp = NULL;
1990         }
1991 
1992         /*
1993          * Following is a union with two members which are
1994          * identical types and size so the following cleanup
1995          * is enough.
1996          */
1997         tcp_close_mpp(&tcp->tcp_conn.tcp_eager_conn_ind);
1998 
1999         CL_INET_DISCONNECT(connp);
2000 
2001         /*
2002          * The connection can't be on the tcp_time_wait_head list
2003          * since it is not detached.
2004          */
2005         ASSERT(tcp->tcp_time_wait_next == NULL);
2006         ASSERT(tcp->tcp_time_wait_prev == NULL);
2007         ASSERT(tcp->tcp_time_wait_expire == 0);
2008 
2009         /*
2010          * Reset/preserve other values
2011          */
2012         tcp_reinit_values(tcp);
2013         ipcl_hash_remove(connp);
2014         /* Note that ixa_cred gets cleared in ixa_cleanup */
2015         ixa_cleanup(connp->conn_ixa);
2016         tcp_ipsec_cleanup(tcp);
2017 
2018         connp->conn_laddr_v6 = connp->conn_bound_addr_v6;
2019         connp->conn_saddr_v6 = connp->conn_bound_addr_v6;
2020         oldstate = tcp->tcp_state;
2021 
2022         if (tcp->tcp_conn_req_max != 0) {
2023                 /*
2024                  * This is the case when a TLI program uses the same
2025                  * transport end point to accept a connection.  This
2026                  * makes the TCP both a listener and acceptor.  When
2027                  * this connection is closed, we need to set the state
2028                  * back to TCPS_LISTEN.  Make sure that the eager list
2029                  * is reinitialized.
2030                  *
2031                  * Note that this stream is still bound to the four
2032                  * tuples of the previous connection in IP.  If a new
2033                  * SYN with different foreign address comes in, IP will
2034                  * not find it and will send it to the global queue.  In
2035                  * the global queue, TCP will do a tcp_lookup_listener()
2036                  * to find this stream.  This works because this stream
2037                  * is only removed from connected hash.
2038                  *
2039                  */
2040                 tcp->tcp_state = TCPS_LISTEN;
2041                 tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
2042                 tcp->tcp_eager_next_drop_q0 = tcp;
2043                 tcp->tcp_eager_prev_drop_q0 = tcp;
2044                 /*
2045                  * Initially set conn_recv to tcp_input_listener_unbound to try
2046                  * to pick a good squeue for the listener when the first SYN
2047                  * arrives. tcp_input_listener_unbound sets it to
2048                  * tcp_input_listener on that first SYN.
2049                  */
2050                 connp->conn_recv = tcp_input_listener_unbound;
2051 
2052                 connp->conn_proto = IPPROTO_TCP;
2053                 connp->conn_faddr_v6 = ipv6_all_zeros;
2054                 connp->conn_fport = 0;
2055 
2056                 (void) ipcl_bind_insert(connp);
2057         } else {
2058                 tcp->tcp_state = TCPS_BOUND;
2059         }
2060 
2061         /*
2062          * Initialize to default values
2063          */
2064         tcp_init_values(tcp, NULL);
2065 
2066         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
2067             connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
2068             int32_t, oldstate);
2069 
2070         ASSERT(tcp->tcp_ptpbhn != NULL);
2071         tcp->tcp_rwnd = connp->conn_rcvbuf;
2072         tcp->tcp_mss = connp->conn_ipversion != IPV4_VERSION ?
2073             tcps->tcps_mss_def_ipv6 : tcps->tcps_mss_def_ipv4;
2074 }
2075 
2076 /*
2077  * Force values to zero that need be zero.
2078  * Do not touch values asociated with the BOUND or LISTEN state
2079  * since the connection will end up in that state after the reinit.
2080  * NOTE: tcp_reinit_values MUST have a line for each field in the tcp_t
2081  * structure!
2082  */
2083 static void
2084 tcp_reinit_values(tcp_t *tcp)
2085 {
2086         tcp_stack_t     *tcps = tcp->tcp_tcps;
2087         conn_t          *connp = tcp->tcp_connp;
2088 
2089 #ifndef lint
2090 #define DONTCARE(x)
2091 #define PRESERVE(x)
2092 #else
2093 #define DONTCARE(x)     ((x) = (x))
2094 #define PRESERVE(x)     ((x) = (x))
2095 #endif  /* lint */
2096 
2097         PRESERVE(tcp->tcp_bind_hash_port);
2098         PRESERVE(tcp->tcp_bind_hash);
2099         PRESERVE(tcp->tcp_ptpbhn);
2100         PRESERVE(tcp->tcp_acceptor_hash);
2101         PRESERVE(tcp->tcp_ptpahn);
2102 
2103         /* Should be ASSERT NULL on these with new code! */
2104         ASSERT(tcp->tcp_time_wait_next == NULL);
2105         ASSERT(tcp->tcp_time_wait_prev == NULL);
2106         ASSERT(tcp->tcp_time_wait_expire == 0);
2107         PRESERVE(tcp->tcp_state);
2108         PRESERVE(connp->conn_rq);
2109         PRESERVE(connp->conn_wq);
2110 
2111         ASSERT(tcp->tcp_xmit_head == NULL);
2112         ASSERT(tcp->tcp_xmit_last == NULL);
2113         ASSERT(tcp->tcp_unsent == 0);
2114         ASSERT(tcp->tcp_xmit_tail == NULL);
2115         ASSERT(tcp->tcp_xmit_tail_unsent == 0);
2116 
2117         tcp->tcp_snxt = 0;                   /* Displayed in mib */
2118         tcp->tcp_suna = 0;                   /* Displayed in mib */
2119         tcp->tcp_swnd = 0;
2120         DONTCARE(tcp->tcp_cwnd);     /* Init in tcp_process_options */
2121 
2122         if (connp->conn_ht_iphc != NULL) {
2123                 kmem_free(connp->conn_ht_iphc, connp->conn_ht_iphc_allocated);
2124                 connp->conn_ht_iphc = NULL;
2125                 connp->conn_ht_iphc_allocated = 0;
2126                 connp->conn_ht_iphc_len = 0;
2127                 connp->conn_ht_ulp = NULL;
2128                 connp->conn_ht_ulp_len = 0;
2129                 tcp->tcp_ipha = NULL;
2130                 tcp->tcp_ip6h = NULL;
2131                 tcp->tcp_tcpha = NULL;
2132         }
2133 
2134         /* We clear any IP_OPTIONS and extension headers */
2135         ip_pkt_free(&connp->conn_xmit_ipp);
2136 
2137         DONTCARE(tcp->tcp_naglim);           /* Init in tcp_init_values */
2138         DONTCARE(tcp->tcp_ipha);
2139         DONTCARE(tcp->tcp_ip6h);
2140         DONTCARE(tcp->tcp_tcpha);
2141         tcp->tcp_valid_bits = 0;
2142 
2143         DONTCARE(tcp->tcp_timer_backoff);    /* Init in tcp_init_values */
2144         DONTCARE(tcp->tcp_last_recv_time);   /* Init in tcp_init_values */
2145         tcp->tcp_last_rcv_lbolt = 0;
2146 
2147         tcp->tcp_init_cwnd = 0;
2148 
2149         tcp->tcp_urp_last_valid = 0;
2150         tcp->tcp_hard_binding = 0;
2151 
2152         tcp->tcp_fin_acked = 0;
2153         tcp->tcp_fin_rcvd = 0;
2154         tcp->tcp_fin_sent = 0;
2155         tcp->tcp_ordrel_done = 0;
2156 
2157         tcp->tcp_detached = 0;
2158 
2159         tcp->tcp_snd_ws_ok = B_FALSE;
2160         tcp->tcp_snd_ts_ok = B_FALSE;
2161         tcp->tcp_zero_win_probe = 0;
2162 
2163         tcp->tcp_loopback = 0;
2164         tcp->tcp_localnet = 0;
2165         tcp->tcp_syn_defense = 0;
2166         tcp->tcp_set_timer = 0;
2167 
2168         tcp->tcp_active_open = 0;
2169         tcp->tcp_rexmit = B_FALSE;
2170         tcp->tcp_xmit_zc_clean = B_FALSE;
2171 
2172         tcp->tcp_snd_sack_ok = B_FALSE;
2173         tcp->tcp_hwcksum = B_FALSE;
2174 
2175         DONTCARE(tcp->tcp_maxpsz_multiplier);        /* Init in tcp_init_values */
2176 
2177         tcp->tcp_conn_def_q0 = 0;
2178         tcp->tcp_ip_forward_progress = B_FALSE;
2179         tcp->tcp_ecn_ok = B_FALSE;
2180 
2181         tcp->tcp_cwr = B_FALSE;
2182         tcp->tcp_ecn_echo_on = B_FALSE;
2183         tcp->tcp_is_wnd_shrnk = B_FALSE;
2184 
2185         TCP_NOTSACK_REMOVE_ALL(tcp->tcp_notsack_list, tcp);
2186         bzero(&tcp->tcp_sack_info, sizeof (tcp_sack_info_t));
2187 
2188         tcp->tcp_rcv_ws = 0;
2189         tcp->tcp_snd_ws = 0;
2190         tcp->tcp_ts_recent = 0;
2191         tcp->tcp_rnxt = 0;                   /* Displayed in mib */
2192         DONTCARE(tcp->tcp_rwnd);             /* Set in tcp_reinit() */
2193         tcp->tcp_initial_pmtu = 0;
2194 
2195         ASSERT(tcp->tcp_reass_head == NULL);
2196         ASSERT(tcp->tcp_reass_tail == NULL);
2197 
2198         tcp->tcp_cwnd_cnt = 0;
2199 
2200         ASSERT(tcp->tcp_rcv_list == NULL);
2201         ASSERT(tcp->tcp_rcv_last_head == NULL);
2202         ASSERT(tcp->tcp_rcv_last_tail == NULL);
2203         ASSERT(tcp->tcp_rcv_cnt == 0);
2204 
2205         DONTCARE(tcp->tcp_cwnd_ssthresh); /* Init in tcp_set_destination */
2206         DONTCARE(tcp->tcp_cwnd_max);         /* Init in tcp_init_values */
2207         tcp->tcp_csuna = 0;
2208 
2209         tcp->tcp_rto = 0;                    /* Displayed in MIB */
2210         DONTCARE(tcp->tcp_rtt_sa);           /* Init in tcp_init_values */
2211         DONTCARE(tcp->tcp_rtt_sd);           /* Init in tcp_init_values */
2212         tcp->tcp_rtt_update = 0;
2213         tcp->tcp_rtt_sum = 0;
2214         tcp->tcp_rtt_cnt = 0;
2215 
2216         DONTCARE(tcp->tcp_swl1); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2217         DONTCARE(tcp->tcp_swl2); /* Init in case TCPS_LISTEN/TCPS_SYN_SENT */
2218 
2219         tcp->tcp_rack = 0;                   /* Displayed in mib */
2220         tcp->tcp_rack_cnt = 0;
2221         tcp->tcp_rack_cur_max = 0;
2222         tcp->tcp_rack_abs_max = 0;
2223 
2224         tcp->tcp_max_swnd = 0;
2225 
2226         ASSERT(tcp->tcp_listener == NULL);
2227 
2228         DONTCARE(tcp->tcp_irs);                      /* tcp_valid_bits cleared */
2229         DONTCARE(tcp->tcp_iss);                      /* tcp_valid_bits cleared */
2230         DONTCARE(tcp->tcp_fss);                      /* tcp_valid_bits cleared */
2231         DONTCARE(tcp->tcp_urg);                      /* tcp_valid_bits cleared */
2232 
2233         ASSERT(tcp->tcp_conn_req_cnt_q == 0);
2234         ASSERT(tcp->tcp_conn_req_cnt_q0 == 0);
2235         PRESERVE(tcp->tcp_conn_req_max);
2236         PRESERVE(tcp->tcp_conn_req_seqnum);
2237 
2238         DONTCARE(tcp->tcp_first_timer_threshold); /* Init in tcp_init_values */
2239         DONTCARE(tcp->tcp_second_timer_threshold); /* Init in tcp_init_values */
2240         DONTCARE(tcp->tcp_first_ctimer_threshold); /* Init in tcp_init_values */
2241         DONTCARE(tcp->tcp_second_ctimer_threshold); /* in tcp_init_values */
2242 
2243         DONTCARE(tcp->tcp_urp_last); /* tcp_urp_last_valid is cleared */
2244         ASSERT(tcp->tcp_urp_mp == NULL);
2245         ASSERT(tcp->tcp_urp_mark_mp == NULL);
2246         ASSERT(tcp->tcp_fused_sigurg_mp == NULL);
2247 
2248         ASSERT(tcp->tcp_eager_next_q == NULL);
2249         ASSERT(tcp->tcp_eager_last_q == NULL);
2250         ASSERT((tcp->tcp_eager_next_q0 == NULL &&
2251             tcp->tcp_eager_prev_q0 == NULL) ||
2252             tcp->tcp_eager_next_q0 == tcp->tcp_eager_prev_q0);
2253         ASSERT(tcp->tcp_conn.tcp_eager_conn_ind == NULL);
2254 
2255         ASSERT((tcp->tcp_eager_next_drop_q0 == NULL &&
2256             tcp->tcp_eager_prev_drop_q0 == NULL) ||
2257             tcp->tcp_eager_next_drop_q0 == tcp->tcp_eager_prev_drop_q0);
2258 
2259         DONTCARE(tcp->tcp_ka_rinterval);     /* Init in tcp_init_values */
2260         DONTCARE(tcp->tcp_ka_abort_thres);   /* Init in tcp_init_values */
2261         DONTCARE(tcp->tcp_ka_cnt);           /* Init in tcp_init_values */
2262 
2263         tcp->tcp_client_errno = 0;
2264 
2265         DONTCARE(connp->conn_sum);           /* Init in tcp_init_values */
2266 
2267         connp->conn_faddr_v6 = ipv6_all_zeros;       /* Displayed in MIB */
2268 
2269         PRESERVE(connp->conn_bound_addr_v6);
2270         tcp->tcp_last_sent_len = 0;
2271         tcp->tcp_dupack_cnt = 0;
2272 
2273         connp->conn_fport = 0;                       /* Displayed in MIB */
2274         PRESERVE(connp->conn_lport);
2275 
2276         PRESERVE(tcp->tcp_acceptor_lockp);
2277 
2278         ASSERT(tcp->tcp_ordrel_mp == NULL);
2279         PRESERVE(tcp->tcp_acceptor_id);
2280         DONTCARE(tcp->tcp_ipsec_overhead);
2281 
2282         PRESERVE(connp->conn_family);
2283         /* Remove any remnants of mapped address binding */
2284         if (connp->conn_family == AF_INET6) {
2285                 connp->conn_ipversion = IPV6_VERSION;
2286                 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2287         } else {
2288                 connp->conn_ipversion = IPV4_VERSION;
2289                 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2290         }
2291 
2292         connp->conn_bound_if = 0;
2293         connp->conn_recv_ancillary.crb_all = 0;
2294         tcp->tcp_recvifindex = 0;
2295         tcp->tcp_recvhops = 0;
2296         tcp->tcp_closed = 0;
2297         if (tcp->tcp_hopopts != NULL) {
2298                 mi_free(tcp->tcp_hopopts);
2299                 tcp->tcp_hopopts = NULL;
2300                 tcp->tcp_hopoptslen = 0;
2301         }
2302         ASSERT(tcp->tcp_hopoptslen == 0);
2303         if (tcp->tcp_dstopts != NULL) {
2304                 mi_free(tcp->tcp_dstopts);
2305                 tcp->tcp_dstopts = NULL;
2306                 tcp->tcp_dstoptslen = 0;
2307         }
2308         ASSERT(tcp->tcp_dstoptslen == 0);
2309         if (tcp->tcp_rthdrdstopts != NULL) {
2310                 mi_free(tcp->tcp_rthdrdstopts);
2311                 tcp->tcp_rthdrdstopts = NULL;
2312                 tcp->tcp_rthdrdstoptslen = 0;
2313         }
2314         ASSERT(tcp->tcp_rthdrdstoptslen == 0);
2315         if (tcp->tcp_rthdr != NULL) {
2316                 mi_free(tcp->tcp_rthdr);
2317                 tcp->tcp_rthdr = NULL;
2318                 tcp->tcp_rthdrlen = 0;
2319         }
2320         ASSERT(tcp->tcp_rthdrlen == 0);
2321 
2322         /* Reset fusion-related fields */
2323         tcp->tcp_fused = B_FALSE;
2324         tcp->tcp_unfusable = B_FALSE;
2325         tcp->tcp_fused_sigurg = B_FALSE;
2326         tcp->tcp_loopback_peer = NULL;
2327 
2328         tcp->tcp_lso = B_FALSE;
2329 
2330         tcp->tcp_in_ack_unsent = 0;
2331         tcp->tcp_cork = B_FALSE;
2332         tcp->tcp_tconnind_started = B_FALSE;
2333 
2334         PRESERVE(tcp->tcp_squeue_bytes);
2335 
2336         tcp->tcp_closemp_used = B_FALSE;
2337 
2338         PRESERVE(tcp->tcp_rsrv_mp);
2339         PRESERVE(tcp->tcp_rsrv_mp_lock);
2340 
2341 #ifdef DEBUG
2342         DONTCARE(tcp->tcmp_stk[0]);
2343 #endif
2344 
2345         PRESERVE(tcp->tcp_connid);
2346 
2347         ASSERT(tcp->tcp_listen_cnt == NULL);
2348         ASSERT(tcp->tcp_reass_tid == 0);
2349 
2350         /* Allow the CC algorithm to clean up after itself. */
2351         if (tcp->tcp_cc_algo->cb_destroy != NULL)
2352                 tcp->tcp_cc_algo->cb_destroy(&tcp->tcp_ccv);
2353         tcp->tcp_cc_algo = NULL;
2354 
2355 #undef  DONTCARE
2356 #undef  PRESERVE
2357 }
2358 
2359 /*
2360  * Initialize the various fields in tcp_t.  If parent (the listener) is non
2361  * NULL, certain values will be inheritted from it.
2362  */
2363 void
2364 tcp_init_values(tcp_t *tcp, tcp_t *parent)
2365 {
2366         tcp_stack_t     *tcps = tcp->tcp_tcps;
2367         conn_t          *connp = tcp->tcp_connp;
2368 
2369         ASSERT((connp->conn_family == AF_INET &&
2370             connp->conn_ipversion == IPV4_VERSION) ||
2371             (connp->conn_family == AF_INET6 &&
2372             (connp->conn_ipversion == IPV4_VERSION ||
2373             connp->conn_ipversion == IPV6_VERSION)));
2374 
2375         tcp->tcp_ccv.type = IPPROTO_TCP;
2376         tcp->tcp_ccv.ccvc.tcp = tcp;
2377 
2378         if (parent == NULL) {
2379                 tcp->tcp_cc_algo = tcps->tcps_default_cc_algo;
2380 
2381                 tcp->tcp_naglim = tcps->tcps_naglim_def;
2382 
2383                 tcp->tcp_rto_initial = tcps->tcps_rexmit_interval_initial;
2384                 tcp->tcp_rto_min = tcps->tcps_rexmit_interval_min;
2385                 tcp->tcp_rto_max = tcps->tcps_rexmit_interval_max;
2386 
2387                 tcp->tcp_first_ctimer_threshold =
2388                     tcps->tcps_ip_notify_cinterval;
2389                 tcp->tcp_second_ctimer_threshold =
2390                     tcps->tcps_ip_abort_cinterval;
2391                 tcp->tcp_first_timer_threshold = tcps->tcps_ip_notify_interval;
2392                 tcp->tcp_second_timer_threshold = tcps->tcps_ip_abort_interval;
2393 
2394                 tcp->tcp_fin_wait_2_flush_interval =
2395                     tcps->tcps_fin_wait_2_flush_interval;
2396 
2397                 tcp->tcp_ka_interval = tcps->tcps_keepalive_interval;
2398                 tcp->tcp_ka_abort_thres = tcps->tcps_keepalive_abort_interval;
2399                 tcp->tcp_ka_cnt = 0;
2400                 tcp->tcp_ka_rinterval = 0;
2401 
2402                 /*
2403                  * Default value of tcp_init_cwnd is 0, so no need to set here
2404                  * if parent is NULL.  But we need to inherit it from parent.
2405                  */
2406         } else {
2407                 /* Inherit various TCP parameters from the parent. */
2408                 tcp->tcp_cc_algo = parent->tcp_cc_algo;
2409 
2410                 tcp->tcp_naglim = parent->tcp_naglim;
2411 
2412                 tcp->tcp_rto_initial = parent->tcp_rto_initial;
2413                 tcp->tcp_rto_min = parent->tcp_rto_min;
2414                 tcp->tcp_rto_max = parent->tcp_rto_max;
2415 
2416                 tcp->tcp_first_ctimer_threshold =
2417                     parent->tcp_first_ctimer_threshold;
2418                 tcp->tcp_second_ctimer_threshold =
2419                     parent->tcp_second_ctimer_threshold;
2420                 tcp->tcp_first_timer_threshold =
2421                     parent->tcp_first_timer_threshold;
2422                 tcp->tcp_second_timer_threshold =
2423                     parent->tcp_second_timer_threshold;
2424 
2425                 tcp->tcp_fin_wait_2_flush_interval =
2426                     parent->tcp_fin_wait_2_flush_interval;
2427 
2428                 tcp->tcp_ka_interval = parent->tcp_ka_interval;
2429                 tcp->tcp_ka_abort_thres = parent->tcp_ka_abort_thres;
2430                 tcp->tcp_ka_cnt = parent->tcp_ka_cnt;
2431                 tcp->tcp_ka_rinterval = parent->tcp_ka_rinterval;
2432 
2433                 tcp->tcp_init_cwnd = parent->tcp_init_cwnd;
2434         }
2435 
2436         if (tcp->tcp_cc_algo->cb_init != NULL)
2437                 VERIFY(tcp->tcp_cc_algo->cb_init(&tcp->tcp_ccv) == 0);
2438 
2439         /*
2440          * Initialize tcp_rtt_sa and tcp_rtt_sd so that the calculated RTO
2441          * will be close to tcp_rexmit_interval_initial.  By doing this, we
2442          * allow the algorithm to adjust slowly to large fluctuations of RTT
2443          * during first few transmissions of a connection as seen in slow
2444          * links.
2445          */
2446         tcp->tcp_rtt_sa = MSEC2NSEC(tcp->tcp_rto_initial) << 2;
2447         tcp->tcp_rtt_sd = MSEC2NSEC(tcp->tcp_rto_initial) >> 1;
2448         tcp->tcp_rto = tcp_calculate_rto(tcp, tcps);
2449 
2450         tcp->tcp_timer_backoff = 0;
2451         tcp->tcp_ms_we_have_waited = 0;
2452         tcp->tcp_last_recv_time = ddi_get_lbolt();
2453         tcp->tcp_cwnd_max = tcps->tcps_cwnd_max_;
2454         tcp->tcp_cwnd_ssthresh = TCP_MAX_LARGEWIN;
2455 
2456         tcp->tcp_maxpsz_multiplier = tcps->tcps_maxpsz_multiplier;
2457 
2458         /* NOTE:  ISS is now set in tcp_set_destination(). */
2459 
2460         /* Reset fusion-related fields */
2461         tcp->tcp_fused = B_FALSE;
2462         tcp->tcp_unfusable = B_FALSE;
2463         tcp->tcp_fused_sigurg = B_FALSE;
2464         tcp->tcp_loopback_peer = NULL;
2465 
2466         /* We rebuild the header template on the next connect/conn_request */
2467 
2468         connp->conn_mlp_type = mlptSingle;
2469 
2470         /*
2471          * Init the window scale to the max so tcp_rwnd_set() won't pare
2472          * down tcp_rwnd. tcp_set_destination() will set the right value later.
2473          */
2474         tcp->tcp_rcv_ws = TCP_MAX_WINSHIFT;
2475         tcp->tcp_rwnd = connp->conn_rcvbuf;
2476 
2477         tcp->tcp_cork = B_FALSE;
2478         /*
2479          * Init the tcp_debug option if it wasn't already set.  This value
2480          * determines whether TCP
2481          * calls strlog() to print out debug messages.  Doing this
2482          * initialization here means that this value is not inherited thru
2483          * tcp_reinit().
2484          */
2485         if (!connp->conn_debug)
2486                 connp->conn_debug = tcps->tcps_dbg;
2487 }
2488 
2489 /*
2490  * Update the TCP connection according to change of PMTU.
2491  *
2492  * Path MTU might have changed by either increase or decrease, so need to
2493  * adjust the MSS based on the value of ixa_pmtu. No need to handle tiny
2494  * or negative MSS, since tcp_mss_set() will do it.
2495  *
2496  * Returns B_TRUE when the connection PMTU changes, otherwise B_FALSE.
2497  */
2498 boolean_t
2499 tcp_update_pmtu(tcp_t *tcp, boolean_t decrease_only)
2500 {
2501         uint32_t        pmtu;
2502         int32_t         mss;
2503         conn_t          *connp = tcp->tcp_connp;
2504         ip_xmit_attr_t  *ixa = connp->conn_ixa;
2505         iaflags_t       ixaflags;
2506 
2507         if (tcp->tcp_tcps->tcps_ignore_path_mtu)
2508                 return (B_FALSE);
2509 
2510         if (tcp->tcp_state < TCPS_ESTABLISHED)
2511                 return (B_FALSE);
2512 
2513         /*
2514          * Always call ip_get_pmtu() to make sure that IP has updated
2515          * ixa_flags properly.
2516          */
2517         pmtu = ip_get_pmtu(ixa);
2518         ixaflags = ixa->ixa_flags;
2519 
2520         /*
2521          * Calculate the MSS by decreasing the PMTU by conn_ht_iphc_len and
2522          * IPsec overhead if applied. Make sure to use the most recent
2523          * IPsec information.
2524          */
2525         mss = pmtu - connp->conn_ht_iphc_len - conn_ipsec_length(connp);
2526 
2527         /*
2528          * Nothing to change, so just return.
2529          */
2530         if (mss == tcp->tcp_mss)
2531                 return (B_FALSE);
2532 
2533         /*
2534          * Currently, for ICMP errors, only PMTU decrease is handled.
2535          */
2536         if (mss > tcp->tcp_mss && decrease_only)
2537                 return (B_FALSE);
2538 
2539         DTRACE_PROBE2(tcp_update_pmtu, int32_t, tcp->tcp_mss, uint32_t, mss);
2540 
2541         /*
2542          * Update ixa_fragsize and ixa_pmtu.
2543          */
2544         ixa->ixa_fragsize = ixa->ixa_pmtu = pmtu;
2545 
2546         /*
2547          * Adjust MSS and all relevant variables.
2548          */
2549         tcp_mss_set(tcp, mss);
2550 
2551         /*
2552          * If the PMTU is below the min size maintained by IP, then ip_get_pmtu
2553          * has set IXAF_PMTU_TOO_SMALL and cleared IXAF_PMTU_IPV4_DF. Since TCP
2554          * has a (potentially different) min size we do the same. Make sure to
2555          * clear IXAF_DONTFRAG, which is used by IP to decide whether to
2556          * fragment the packet.
2557          *
2558          * LSO over IPv6 can not be fragmented. So need to disable LSO
2559          * when IPv6 fragmentation is needed.
2560          */
2561         if (mss < tcp->tcp_tcps->tcps_mss_min)
2562                 ixaflags |= IXAF_PMTU_TOO_SMALL;
2563 
2564         if (ixaflags & IXAF_PMTU_TOO_SMALL)
2565                 ixaflags &= ~(IXAF_DONTFRAG | IXAF_PMTU_IPV4_DF);
2566 
2567         if ((connp->conn_ipversion == IPV4_VERSION) &&
2568             !(ixaflags & IXAF_PMTU_IPV4_DF)) {
2569                 tcp->tcp_ipha->ipha_fragment_offset_and_flags = 0;
2570         }
2571         ixa->ixa_flags = ixaflags;
2572         return (B_TRUE);
2573 }
2574 
2575 int
2576 tcp_maxpsz_set(tcp_t *tcp, boolean_t set_maxblk)
2577 {
2578         conn_t  *connp = tcp->tcp_connp;
2579         queue_t *q = connp->conn_rq;
2580         int32_t mss = tcp->tcp_mss;
2581         int     maxpsz;
2582 
2583         if (TCP_IS_DETACHED(tcp))
2584                 return (mss);
2585         if (tcp->tcp_fused) {
2586                 maxpsz = tcp_fuse_maxpsz(tcp);
2587                 mss = INFPSZ;
2588         } else if (tcp->tcp_maxpsz_multiplier == 0) {
2589                 /*
2590                  * Set the sd_qn_maxpsz according to the socket send buffer
2591                  * size, and sd_maxblk to INFPSZ (-1).  This will essentially
2592                  * instruct the stream head to copyin user data into contiguous
2593                  * kernel-allocated buffers without breaking it up into smaller
2594                  * chunks.  We round up the buffer size to the nearest SMSS.
2595                  */
2596                 maxpsz = MSS_ROUNDUP(connp->conn_sndbuf, mss);
2597                 mss = INFPSZ;
2598         } else {
2599                 /*
2600                  * Set sd_qn_maxpsz to approx half the (receivers) buffer
2601                  * (and a multiple of the mss).  This instructs the stream
2602                  * head to break down larger than SMSS writes into SMSS-
2603                  * size mblks, up to tcp_maxpsz_multiplier mblks at a time.
2604                  */
2605                 maxpsz = tcp->tcp_maxpsz_multiplier * mss;
2606                 if (maxpsz > connp->conn_sndbuf / 2) {
2607                         maxpsz = connp->conn_sndbuf / 2;
2608                         /* Round up to nearest mss */
2609                         maxpsz = MSS_ROUNDUP(maxpsz, mss);
2610                 }
2611         }
2612 
2613         (void) proto_set_maxpsz(q, connp, maxpsz);
2614         if (!(IPCL_IS_NONSTR(connp)))
2615                 connp->conn_wq->q_maxpsz = maxpsz;
2616         if (set_maxblk)
2617                 (void) proto_set_tx_maxblk(q, connp, mss);
2618         return (mss);
2619 }
2620 
2621 /* For /dev/tcp aka AF_INET open */
2622 static int
2623 tcp_openv4(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2624 {
2625         return (tcp_open(q, devp, flag, sflag, credp, B_FALSE));
2626 }
2627 
2628 /* For /dev/tcp6 aka AF_INET6 open */
2629 static int
2630 tcp_openv6(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp)
2631 {
2632         return (tcp_open(q, devp, flag, sflag, credp, B_TRUE));
2633 }
2634 
2635 conn_t *
2636 tcp_create_common(cred_t *credp, boolean_t isv6, boolean_t issocket,
2637     int *errorp)
2638 {
2639         tcp_t           *tcp = NULL;
2640         conn_t          *connp;
2641         zoneid_t        zoneid;
2642         tcp_stack_t     *tcps;
2643         squeue_t        *sqp;
2644 
2645         ASSERT(errorp != NULL);
2646         /*
2647          * Find the proper zoneid and netstack.
2648          */
2649         /*
2650          * Special case for install: miniroot needs to be able to
2651          * access files via NFS as though it were always in the
2652          * global zone.
2653          */
2654         if (credp == kcred && nfs_global_client_only != 0) {
2655                 zoneid = GLOBAL_ZONEID;
2656                 tcps = netstack_find_by_stackid(GLOBAL_NETSTACKID)->
2657                     netstack_tcp;
2658                 ASSERT(tcps != NULL);
2659         } else {
2660                 netstack_t *ns;
2661                 int err;
2662 
2663                 if ((err = secpolicy_basic_net_access(credp)) != 0) {
2664                         *errorp = err;
2665                         return (NULL);
2666                 }
2667 
2668                 ns = netstack_find_by_cred(credp);
2669                 ASSERT(ns != NULL);
2670                 tcps = ns->netstack_tcp;
2671                 ASSERT(tcps != NULL);
2672 
2673                 /*
2674                  * For exclusive stacks we set the zoneid to zero
2675                  * to make TCP operate as if in the global zone.
2676                  */
2677                 if (tcps->tcps_netstack->netstack_stackid !=
2678                     GLOBAL_NETSTACKID)
2679                         zoneid = GLOBAL_ZONEID;
2680                 else
2681                         zoneid = crgetzoneid(credp);
2682         }
2683 
2684         sqp = IP_SQUEUE_GET((uint_t)gethrtime());
2685         connp = tcp_get_conn(sqp, tcps);
2686         /*
2687          * Both tcp_get_conn and netstack_find_by_cred incremented refcnt,
2688          * so we drop it by one.
2689          */
2690         netstack_rele(tcps->tcps_netstack);
2691         if (connp == NULL) {
2692                 *errorp = ENOSR;
2693                 return (NULL);
2694         }
2695         ASSERT(connp->conn_ixa->ixa_protocol == connp->conn_proto);
2696 
2697         connp->conn_sqp = sqp;
2698         connp->conn_initial_sqp = connp->conn_sqp;
2699         connp->conn_ixa->ixa_sqp = connp->conn_sqp;
2700         tcp = connp->conn_tcp;
2701 
2702         /*
2703          * Besides asking IP to set the checksum for us, have conn_ip_output
2704          * to do the following checks when necessary:
2705          *
2706          * IXAF_VERIFY_SOURCE: drop packets when our outer source goes invalid
2707          * IXAF_VERIFY_PMTU: verify PMTU changes
2708          * IXAF_VERIFY_LSO: verify LSO capability changes
2709          */
2710         connp->conn_ixa->ixa_flags |= IXAF_SET_ULP_CKSUM | IXAF_VERIFY_SOURCE |
2711             IXAF_VERIFY_PMTU | IXAF_VERIFY_LSO;
2712 
2713         if (!tcps->tcps_dev_flow_ctl)
2714                 connp->conn_ixa->ixa_flags |= IXAF_NO_DEV_FLOW_CTL;
2715 
2716         if (isv6) {
2717                 connp->conn_ixa->ixa_src_preferences = IPV6_PREFER_SRC_DEFAULT;
2718                 connp->conn_ipversion = IPV6_VERSION;
2719                 connp->conn_family = AF_INET6;
2720                 tcp->tcp_mss = tcps->tcps_mss_def_ipv6;
2721                 connp->conn_default_ttl = tcps->tcps_ipv6_hoplimit;
2722         } else {
2723                 connp->conn_ipversion = IPV4_VERSION;
2724                 connp->conn_family = AF_INET;
2725                 tcp->tcp_mss = tcps->tcps_mss_def_ipv4;
2726                 connp->conn_default_ttl = tcps->tcps_ipv4_ttl;
2727         }
2728         connp->conn_xmit_ipp.ipp_unicast_hops = connp->conn_default_ttl;
2729 
2730         crhold(credp);
2731         connp->conn_cred = credp;
2732         connp->conn_cpid = curproc->p_pid;
2733         connp->conn_open_time = ddi_get_lbolt64();
2734 
2735         /* Cache things in the ixa without any refhold */
2736         ASSERT(!(connp->conn_ixa->ixa_free_flags & IXA_FREE_CRED));
2737         connp->conn_ixa->ixa_cred = credp;
2738         connp->conn_ixa->ixa_cpid = connp->conn_cpid;
2739 
2740         connp->conn_zoneid = zoneid;
2741         /* conn_allzones can not be set this early, hence no IPCL_ZONEID */
2742         connp->conn_ixa->ixa_zoneid = zoneid;
2743         connp->conn_mlp_type = mlptSingle;
2744         ASSERT(connp->conn_netstack == tcps->tcps_netstack);
2745         ASSERT(tcp->tcp_tcps == tcps);
2746 
2747         /*
2748          * If the caller has the process-wide flag set, then default to MAC
2749          * exempt mode.  This allows read-down to unlabeled hosts.
2750          */
2751         if (getpflags(NET_MAC_AWARE, credp) != 0)
2752                 connp->conn_mac_mode = CONN_MAC_AWARE;
2753 
2754         connp->conn_zone_is_global = (crgetzoneid(credp) == GLOBAL_ZONEID);
2755 
2756         if (issocket) {
2757                 tcp->tcp_issocket = 1;
2758         }
2759 
2760         connp->conn_rcvbuf = tcps->tcps_recv_hiwat;
2761         connp->conn_sndbuf = tcps->tcps_xmit_hiwat;
2762         if (tcps->tcps_snd_lowat_fraction != 0) {
2763                 connp->conn_sndlowat = connp->conn_sndbuf /
2764                     tcps->tcps_snd_lowat_fraction;
2765         } else {
2766                 connp->conn_sndlowat = tcps->tcps_xmit_lowat;
2767         }
2768         connp->conn_so_type = SOCK_STREAM;
2769         connp->conn_wroff = connp->conn_ht_iphc_allocated +
2770             tcps->tcps_wroff_xtra;
2771 
2772         SOCK_CONNID_INIT(tcp->tcp_connid);
2773         /* DTrace ignores this - it isn't a tcp:::state-change */
2774         tcp->tcp_state = TCPS_IDLE;
2775         tcp_init_values(tcp, NULL);
2776         return (connp);
2777 }
2778 
2779 static int
2780 tcp_open(queue_t *q, dev_t *devp, int flag, int sflag, cred_t *credp,
2781     boolean_t isv6)
2782 {
2783         tcp_t           *tcp = NULL;
2784         conn_t          *connp = NULL;
2785         int             err;
2786         vmem_t          *minor_arena = NULL;
2787         dev_t           conn_dev;
2788         boolean_t       issocket;
2789 
2790         if (q->q_ptr != NULL)
2791                 return (0);
2792 
2793         if (sflag == MODOPEN)
2794                 return (EINVAL);
2795 
2796         if ((ip_minor_arena_la != NULL) && (flag & SO_SOCKSTR) &&
2797             ((conn_dev = inet_minor_alloc(ip_minor_arena_la)) != 0)) {
2798                 minor_arena = ip_minor_arena_la;
2799         } else {
2800                 /*
2801                  * Either minor numbers in the large arena were exhausted
2802                  * or a non socket application is doing the open.
2803                  * Try to allocate from the small arena.
2804                  */
2805                 if ((conn_dev = inet_minor_alloc(ip_minor_arena_sa)) == 0) {
2806                         return (EBUSY);
2807                 }
2808                 minor_arena = ip_minor_arena_sa;
2809         }
2810 
2811         ASSERT(minor_arena != NULL);
2812 
2813         *devp = makedevice(getmajor(*devp), (minor_t)conn_dev);
2814 
2815         if (flag & SO_FALLBACK) {
2816                 /*
2817                  * Non streams socket needs a stream to fallback to
2818                  */
2819                 RD(q)->q_ptr = (void *)conn_dev;
2820                 WR(q)->q_qinfo = &tcp_fallback_sock_winit;
2821                 WR(q)->q_ptr = (void *)minor_arena;
2822                 qprocson(q);
2823                 return (0);
2824         } else if (flag & SO_ACCEPTOR) {
2825                 q->q_qinfo = &tcp_acceptor_rinit;
2826                 /*
2827                  * the conn_dev and minor_arena will be subsequently used by
2828                  * tcp_tli_accept() and tcp_tpi_close_accept() to figure out
2829                  * the minor device number for this connection from the q_ptr.
2830                  */
2831                 RD(q)->q_ptr = (void *)conn_dev;
2832                 WR(q)->q_qinfo = &tcp_acceptor_winit;
2833                 WR(q)->q_ptr = (void *)minor_arena;
2834                 qprocson(q);
2835                 return (0);
2836         }
2837 
2838         issocket = flag & SO_SOCKSTR;
2839         connp = tcp_create_common(credp, isv6, issocket, &err);
2840 
2841         if (connp == NULL) {
2842                 inet_minor_free(minor_arena, conn_dev);
2843                 q->q_ptr = WR(q)->q_ptr = NULL;
2844                 return (err);
2845         }
2846 
2847         connp->conn_rq = q;
2848         connp->conn_wq = WR(q);
2849         q->q_ptr = WR(q)->q_ptr = connp;
2850 
2851         connp->conn_dev = conn_dev;
2852         connp->conn_minor_arena = minor_arena;
2853 
2854         ASSERT(q->q_qinfo == &tcp_rinitv4 || q->q_qinfo == &tcp_rinitv6);
2855         ASSERT(WR(q)->q_qinfo == &tcp_winit);
2856 
2857         tcp = connp->conn_tcp;
2858 
2859         if (issocket) {
2860                 WR(q)->q_qinfo = &tcp_sock_winit;
2861         } else {
2862 #ifdef  _ILP32
2863                 tcp->tcp_acceptor_id = (t_uscalar_t)RD(q);
2864 #else
2865                 tcp->tcp_acceptor_id = conn_dev;
2866 #endif  /* _ILP32 */
2867                 tcp_acceptor_hash_insert(tcp->tcp_acceptor_id, tcp);
2868         }
2869 
2870         /*
2871          * Put the ref for TCP. Ref for IP was already put
2872          * by ipcl_conn_create. Also Make the conn_t globally
2873          * visible to walkers
2874          */
2875         mutex_enter(&connp->conn_lock);
2876         CONN_INC_REF_LOCKED(connp);
2877         ASSERT(connp->conn_ref == 2);
2878         connp->conn_state_flags &= ~CONN_INCIPIENT;
2879         mutex_exit(&connp->conn_lock);
2880 
2881         qprocson(q);
2882         return (0);
2883 }
2884 
2885 /*
2886  * Build/update the tcp header template (in conn_ht_iphc) based on
2887  * conn_xmit_ipp. The headers include ip6_t, any extension
2888  * headers, and the maximum size tcp header (to avoid reallocation
2889  * on the fly for additional tcp options).
2890  *
2891  * Assumes the caller has already set conn_{faddr,laddr,fport,lport,flowinfo}.
2892  * Returns failure if can't allocate memory.
2893  */
2894 int
2895 tcp_build_hdrs(tcp_t *tcp)
2896 {
2897         tcp_stack_t     *tcps = tcp->tcp_tcps;
2898         conn_t          *connp = tcp->tcp_connp;
2899         char            buf[TCP_MAX_HDR_LENGTH];
2900         uint_t          buflen;
2901         uint_t          ulplen = TCP_MIN_HEADER_LENGTH;
2902         uint_t          extralen = TCP_MAX_TCP_OPTIONS_LENGTH;
2903         tcpha_t         *tcpha;
2904         uint32_t        cksum;
2905         int             error;
2906 
2907         /*
2908          * We might be called after the connection is set up, and we might
2909          * have TS options already in the TCP header. Thus we  save any
2910          * existing tcp header.
2911          */
2912         buflen = connp->conn_ht_ulp_len;
2913         if (buflen != 0) {
2914                 bcopy(connp->conn_ht_ulp, buf, buflen);
2915                 extralen -= buflen - ulplen;
2916                 ulplen = buflen;
2917         }
2918 
2919         /* Grab lock to satisfy ASSERT; TCP is serialized using squeue */
2920         mutex_enter(&connp->conn_lock);
2921         error = conn_build_hdr_template(connp, ulplen, extralen,
2922             &connp->conn_laddr_v6, &connp->conn_faddr_v6, connp->conn_flowinfo);
2923         mutex_exit(&connp->conn_lock);
2924         if (error != 0)
2925                 return (error);
2926 
2927         /*
2928          * Any routing header/option has been massaged. The checksum difference
2929          * is stored in conn_sum for later use.
2930          */
2931         tcpha = (tcpha_t *)connp->conn_ht_ulp;
2932         tcp->tcp_tcpha = tcpha;
2933 
2934         /* restore any old tcp header */
2935         if (buflen != 0) {
2936                 bcopy(buf, connp->conn_ht_ulp, buflen);
2937         } else {
2938                 tcpha->tha_sum = 0;
2939                 tcpha->tha_urp = 0;
2940                 tcpha->tha_ack = 0;
2941                 tcpha->tha_offset_and_reserved = (5 << 4);
2942                 tcpha->tha_lport = connp->conn_lport;
2943                 tcpha->tha_fport = connp->conn_fport;
2944         }
2945 
2946         /*
2947          * IP wants our header length in the checksum field to
2948          * allow it to perform a single pseudo-header+checksum
2949          * calculation on behalf of TCP.
2950          * Include the adjustment for a source route once IP_OPTIONS is set.
2951          */
2952         cksum = sizeof (tcpha_t) + connp->conn_sum;
2953         cksum = (cksum >> 16) + (cksum & 0xFFFF);
2954         ASSERT(cksum < 0x10000);
2955         tcpha->tha_sum = htons(cksum);
2956 
2957         if (connp->conn_ipversion == IPV4_VERSION)
2958                 tcp->tcp_ipha = (ipha_t *)connp->conn_ht_iphc;
2959         else
2960                 tcp->tcp_ip6h = (ip6_t *)connp->conn_ht_iphc;
2961 
2962         if (connp->conn_ht_iphc_allocated + tcps->tcps_wroff_xtra >
2963             connp->conn_wroff) {
2964                 connp->conn_wroff = connp->conn_ht_iphc_allocated +
2965                     tcps->tcps_wroff_xtra;
2966                 (void) proto_set_tx_wroff(connp->conn_rq, connp,
2967                     connp->conn_wroff);
2968         }
2969         return (0);
2970 }
2971 
2972 /*
2973  * tcp_rwnd_set() is called to adjust the receive window to a desired value.
2974  * We do not allow the receive window to shrink.  After setting rwnd,
2975  * set the flow control hiwat of the stream.
2976  *
2977  * This function is called in 2 cases:
2978  *
2979  * 1) Before data transfer begins, in tcp_input_listener() for accepting a
2980  *    connection (passive open) and in tcp_input_data() for active connect.
2981  *    This is called after tcp_mss_set() when the desired MSS value is known.
2982  *    This makes sure that our window size is a mutiple of the other side's
2983  *    MSS.
2984  * 2) Handling SO_RCVBUF option.
2985  *
2986  * It is ASSUMED that the requested size is a multiple of the current MSS.
2987  *
2988  * XXX - Should allow a lower rwnd than tcp_recv_hiwat_minmss * mss if the
2989  * user requests so.
2990  */
2991 int
2992 tcp_rwnd_set(tcp_t *tcp, uint32_t rwnd)
2993 {
2994         uint32_t        mss = tcp->tcp_mss;
2995         uint32_t        old_max_rwnd;
2996         uint32_t        max_transmittable_rwnd;
2997         boolean_t       tcp_detached = TCP_IS_DETACHED(tcp);
2998         tcp_stack_t     *tcps = tcp->tcp_tcps;
2999         conn_t          *connp = tcp->tcp_connp;
3000 
3001         /*
3002          * Insist on a receive window that is at least
3003          * tcp_recv_hiwat_minmss * MSS (default 4 * MSS) to avoid
3004          * funny TCP interactions of Nagle algorithm, SWS avoidance
3005          * and delayed acknowledgement.
3006          */
3007         rwnd = MAX(rwnd, tcps->tcps_recv_hiwat_minmss * mss);
3008 
3009         if (tcp->tcp_fused) {
3010                 size_t sth_hiwat;
3011                 tcp_t *peer_tcp = tcp->tcp_loopback_peer;
3012 
3013                 ASSERT(peer_tcp != NULL);
3014                 sth_hiwat = tcp_fuse_set_rcv_hiwat(tcp, rwnd);
3015                 if (!tcp_detached) {
3016                         (void) proto_set_rx_hiwat(connp->conn_rq, connp,
3017                             sth_hiwat);
3018                         tcp_set_recv_threshold(tcp, sth_hiwat >> 3);
3019                 }
3020 
3021                 /* Caller could have changed tcp_rwnd; update tha_win */
3022                 if (tcp->tcp_tcpha != NULL) {
3023                         tcp->tcp_tcpha->tha_win =
3024                             htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3025                 }
3026                 if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3027                         tcp->tcp_cwnd_max = rwnd;
3028 
3029                 /*
3030                  * In the fusion case, the maxpsz stream head value of
3031                  * our peer is set according to its send buffer size
3032                  * and our receive buffer size; since the latter may
3033                  * have changed we need to update the peer's maxpsz.
3034                  */
3035                 (void) tcp_maxpsz_set(peer_tcp, B_TRUE);
3036                 return (sth_hiwat);
3037         }
3038 
3039         if (tcp_detached)
3040                 old_max_rwnd = tcp->tcp_rwnd;
3041         else
3042                 old_max_rwnd = connp->conn_rcvbuf;
3043 
3044 
3045         /*
3046          * If window size info has already been exchanged, TCP should not
3047          * shrink the window.  Shrinking window is doable if done carefully.
3048          * We may add that support later.  But so far there is not a real
3049          * need to do that.
3050          */
3051         if (rwnd < old_max_rwnd && tcp->tcp_state > TCPS_SYN_SENT) {
3052                 /* MSS may have changed, do a round up again. */
3053                 rwnd = MSS_ROUNDUP(old_max_rwnd, mss);
3054         }
3055 
3056         /*
3057          * tcp_rcv_ws starts with TCP_MAX_WINSHIFT so the following check
3058          * can be applied even before the window scale option is decided.
3059          */
3060         max_transmittable_rwnd = TCP_MAXWIN << tcp->tcp_rcv_ws;
3061         if (rwnd > max_transmittable_rwnd) {
3062                 rwnd = max_transmittable_rwnd -
3063                     (max_transmittable_rwnd % mss);
3064                 if (rwnd < mss)
3065                         rwnd = max_transmittable_rwnd;
3066                 /*
3067                  * If we're over the limit we may have to back down tcp_rwnd.
3068                  * The increment below won't work for us. So we set all three
3069                  * here and the increment below will have no effect.
3070                  */
3071                 tcp->tcp_rwnd = old_max_rwnd = rwnd;
3072         }
3073         if (tcp->tcp_localnet) {
3074                 tcp->tcp_rack_abs_max =
3075                     MIN(tcps->tcps_local_dacks_max, rwnd / mss / 2);
3076         } else {
3077                 /*
3078                  * For a remote host on a different subnet (through a router),
3079                  * we ack every other packet to be conforming to RFC1122.
3080                  * tcp_deferred_acks_max is default to 2.
3081                  */
3082                 tcp->tcp_rack_abs_max =
3083                     MIN(tcps->tcps_deferred_acks_max, rwnd / mss / 2);
3084         }
3085         if (tcp->tcp_rack_cur_max > tcp->tcp_rack_abs_max)
3086                 tcp->tcp_rack_cur_max = tcp->tcp_rack_abs_max;
3087         else
3088                 tcp->tcp_rack_cur_max = 0;
3089         /*
3090          * Increment the current rwnd by the amount the maximum grew (we
3091          * can not overwrite it since we might be in the middle of a
3092          * connection.)
3093          */
3094         tcp->tcp_rwnd += rwnd - old_max_rwnd;
3095         connp->conn_rcvbuf = rwnd;
3096 
3097         /* Are we already connected? */
3098         if (tcp->tcp_tcpha != NULL) {
3099                 tcp->tcp_tcpha->tha_win =
3100                     htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
3101         }
3102 
3103         if ((tcp->tcp_rcv_ws > 0) && rwnd > tcp->tcp_cwnd_max)
3104                 tcp->tcp_cwnd_max = rwnd;
3105 
3106         if (tcp_detached)
3107                 return (rwnd);
3108 
3109         tcp_set_recv_threshold(tcp, rwnd >> 3);
3110 
3111         (void) proto_set_rx_hiwat(connp->conn_rq, connp, rwnd);
3112         return (rwnd);
3113 }
3114 
3115 int
3116 tcp_do_unbind(conn_t *connp)
3117 {
3118         tcp_t *tcp = connp->conn_tcp;
3119         int32_t oldstate;
3120 
3121         switch (tcp->tcp_state) {
3122         case TCPS_BOUND:
3123         case TCPS_LISTEN:
3124                 break;
3125         default:
3126                 return (-TOUTSTATE);
3127         }
3128 
3129         /*
3130          * Need to clean up all the eagers since after the unbind, segments
3131          * will no longer be delivered to this listener stream.
3132          */
3133         mutex_enter(&tcp->tcp_eager_lock);
3134         if (tcp->tcp_conn_req_cnt_q0 != 0 || tcp->tcp_conn_req_cnt_q != 0) {
3135                 tcp_eager_cleanup(tcp, 0);
3136         }
3137         mutex_exit(&tcp->tcp_eager_lock);
3138 
3139         /* Clean up the listener connection counter if necessary. */
3140         if (tcp->tcp_listen_cnt != NULL)
3141                 TCP_DECR_LISTEN_CNT(tcp);
3142         connp->conn_laddr_v6 = ipv6_all_zeros;
3143         connp->conn_saddr_v6 = ipv6_all_zeros;
3144         tcp_bind_hash_remove(tcp);
3145         oldstate = tcp->tcp_state;
3146         tcp->tcp_state = TCPS_IDLE;
3147         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
3148             connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
3149             int32_t, oldstate);
3150 
3151         ip_unbind(connp);
3152         bzero(&connp->conn_ports, sizeof (connp->conn_ports));
3153 
3154         return (0);
3155 }
3156 
3157 /*
3158  * Collect protocol properties to send to the upper handle.
3159  */
3160 void
3161 tcp_get_proto_props(tcp_t *tcp, struct sock_proto_props *sopp)
3162 {
3163         conn_t *connp = tcp->tcp_connp;
3164 
3165         sopp->sopp_flags = SOCKOPT_RCVHIWAT | SOCKOPT_MAXBLK | SOCKOPT_WROFF;
3166         sopp->sopp_maxblk = tcp_maxpsz_set(tcp, B_FALSE);
3167 
3168         sopp->sopp_rxhiwat = tcp->tcp_fused ?
3169             tcp_fuse_set_rcv_hiwat(tcp, connp->conn_rcvbuf) :
3170             connp->conn_rcvbuf;
3171         /*
3172          * Determine what write offset value to use depending on SACK and
3173          * whether the endpoint is fused or not.
3174          */
3175         if (tcp->tcp_fused) {
3176                 ASSERT(tcp->tcp_loopback);
3177                 ASSERT(tcp->tcp_loopback_peer != NULL);
3178                 /*
3179                  * For fused tcp loopback, set the stream head's write
3180                  * offset value to zero since we won't be needing any room
3181                  * for TCP/IP headers.  This would also improve performance
3182                  * since it would reduce the amount of work done by kmem.
3183                  * Non-fused tcp loopback case is handled separately below.
3184                  */
3185                 sopp->sopp_wroff = 0;
3186                 /*
3187                  * Update the peer's transmit parameters according to
3188                  * our recently calculated high water mark value.
3189                  */
3190                 (void) tcp_maxpsz_set(tcp->tcp_loopback_peer, B_TRUE);
3191         } else if (tcp->tcp_snd_sack_ok) {
3192                 sopp->sopp_wroff = connp->conn_ht_iphc_allocated +
3193                     (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3194         } else {
3195                 sopp->sopp_wroff = connp->conn_ht_iphc_len +
3196                     (tcp->tcp_loopback ? 0 : tcp->tcp_tcps->tcps_wroff_xtra);
3197         }
3198 
3199         if (tcp->tcp_loopback) {
3200                 sopp->sopp_flags |= SOCKOPT_LOOPBACK;
3201                 sopp->sopp_loopback = B_TRUE;
3202         }
3203 }
3204 
3205 /*
3206  * Check the usability of ZEROCOPY. It's instead checking the flag set by IP.
3207  */
3208 boolean_t
3209 tcp_zcopy_check(tcp_t *tcp)
3210 {
3211         conn_t          *connp = tcp->tcp_connp;
3212         ip_xmit_attr_t  *ixa = connp->conn_ixa;
3213         boolean_t       zc_enabled = B_FALSE;
3214         tcp_stack_t     *tcps = tcp->tcp_tcps;
3215 
3216         if (do_tcpzcopy == 2)
3217                 zc_enabled = B_TRUE;
3218         else if ((do_tcpzcopy == 1) && (ixa->ixa_flags & IXAF_ZCOPY_CAPAB))
3219                 zc_enabled = B_TRUE;
3220 
3221         tcp->tcp_snd_zcopy_on = zc_enabled;
3222         if (!TCP_IS_DETACHED(tcp)) {
3223                 if (zc_enabled) {
3224                         ixa->ixa_flags |= IXAF_VERIFY_ZCOPY;
3225                         (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3226                             ZCVMSAFE);
3227                         TCP_STAT(tcps, tcp_zcopy_on);
3228                 } else {
3229                         ixa->ixa_flags &= ~IXAF_VERIFY_ZCOPY;
3230                         (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3231                             ZCVMUNSAFE);
3232                         TCP_STAT(tcps, tcp_zcopy_off);
3233                 }
3234         }
3235         return (zc_enabled);
3236 }
3237 
3238 /*
3239  * Backoff from a zero-copy message by copying data to a new allocated
3240  * message and freeing the original desballoca'ed segmapped message.
3241  *
3242  * This function is called by following two callers:
3243  * 1. tcp_timer: fix_xmitlist is set to B_TRUE, because it's safe to free
3244  *    the origial desballoca'ed message and notify sockfs. This is in re-
3245  *    transmit state.
3246  * 2. tcp_output: fix_xmitlist is set to B_FALSE. Flag STRUIO_ZCNOTIFY need
3247  *    to be copied to new message.
3248  */
3249 mblk_t *
3250 tcp_zcopy_backoff(tcp_t *tcp, mblk_t *bp, boolean_t fix_xmitlist)
3251 {
3252         mblk_t          *nbp;
3253         mblk_t          *head = NULL;
3254         mblk_t          *tail = NULL;
3255         tcp_stack_t     *tcps = tcp->tcp_tcps;
3256 
3257         ASSERT(bp != NULL);
3258         while (bp != NULL) {
3259                 if (IS_VMLOANED_MBLK(bp)) {
3260                         TCP_STAT(tcps, tcp_zcopy_backoff);
3261                         if ((nbp = copyb(bp)) == NULL) {
3262                                 tcp->tcp_xmit_zc_clean = B_FALSE;
3263                                 if (tail != NULL)
3264                                         tail->b_cont = bp;
3265                                 return ((head == NULL) ? bp : head);
3266                         }
3267 
3268                         if (bp->b_datap->db_struioflag & STRUIO_ZCNOTIFY) {
3269                                 if (fix_xmitlist)
3270                                         tcp_zcopy_notify(tcp);
3271                                 else
3272                                         nbp->b_datap->db_struioflag |=
3273                                             STRUIO_ZCNOTIFY;
3274                         }
3275                         nbp->b_cont = bp->b_cont;
3276 
3277                         /*
3278                          * Copy saved information and adjust tcp_xmit_tail
3279                          * if needed.
3280                          */
3281                         if (fix_xmitlist) {
3282                                 nbp->b_prev = bp->b_prev;
3283                                 nbp->b_next = bp->b_next;
3284 
3285                                 if (tcp->tcp_xmit_tail == bp)
3286                                         tcp->tcp_xmit_tail = nbp;
3287                         }
3288 
3289                         /* Free the original message. */
3290                         bp->b_prev = NULL;
3291                         bp->b_next = NULL;
3292                         freeb(bp);
3293 
3294                         bp = nbp;
3295                 }
3296 
3297                 if (head == NULL) {
3298                         head = bp;
3299                 }
3300                 if (tail == NULL) {
3301                         tail = bp;
3302                 } else {
3303                         tail->b_cont = bp;
3304                         tail = bp;
3305                 }
3306 
3307                 /* Move forward. */
3308                 bp = bp->b_cont;
3309         }
3310 
3311         if (fix_xmitlist) {
3312                 tcp->tcp_xmit_last = tail;
3313                 tcp->tcp_xmit_zc_clean = B_TRUE;
3314         }
3315 
3316         return (head);
3317 }
3318 
3319 void
3320 tcp_zcopy_notify(tcp_t *tcp)
3321 {
3322         struct stdata   *stp;
3323         conn_t          *connp;
3324 
3325         if (tcp->tcp_detached)
3326                 return;
3327         connp = tcp->tcp_connp;
3328         if (IPCL_IS_NONSTR(connp)) {
3329                 (*connp->conn_upcalls->su_zcopy_notify)
3330                     (connp->conn_upper_handle);
3331                 return;
3332         }
3333         stp = STREAM(connp->conn_rq);
3334         mutex_enter(&stp->sd_lock);
3335         stp->sd_flag |= STZCNOTIFY;
3336         cv_broadcast(&stp->sd_zcopy_wait);
3337         mutex_exit(&stp->sd_lock);
3338 }
3339 
3340 /*
3341  * Update the TCP connection according to change of LSO capability.
3342  */
3343 static void
3344 tcp_update_lso(tcp_t *tcp, ip_xmit_attr_t *ixa)
3345 {
3346         /*
3347          * We check against IPv4 header length to preserve the old behavior
3348          * of only enabling LSO when there are no IP options.
3349          * But this restriction might not be necessary at all. Before removing
3350          * it, need to verify how LSO is handled for source routing case, with
3351          * which IP does software checksum.
3352          *
3353          * For IPv6, whenever any extension header is needed, LSO is supressed.
3354          */
3355         if (ixa->ixa_ip_hdr_length != ((ixa->ixa_flags & IXAF_IS_IPV4) ?
3356             IP_SIMPLE_HDR_LENGTH : IPV6_HDR_LEN))
3357                 return;
3358 
3359         /*
3360          * Either the LSO capability newly became usable, or it has changed.
3361          */
3362         if (ixa->ixa_flags & IXAF_LSO_CAPAB) {
3363                 ill_lso_capab_t *lsoc = &ixa->ixa_lso_capab;
3364 
3365                 ASSERT(lsoc->ill_lso_max > 0);
3366                 tcp->tcp_lso_max = MIN(TCP_MAX_LSO_LENGTH, lsoc->ill_lso_max);
3367 
3368                 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3369                     boolean_t, B_TRUE, uint32_t, tcp->tcp_lso_max);
3370 
3371                 /*
3372                  * If LSO to be enabled, notify the STREAM header with larger
3373                  * data block.
3374                  */
3375                 if (!tcp->tcp_lso)
3376                         tcp->tcp_maxpsz_multiplier = 0;
3377 
3378                 tcp->tcp_lso = B_TRUE;
3379                 TCP_STAT(tcp->tcp_tcps, tcp_lso_enabled);
3380         } else { /* LSO capability is not usable any more. */
3381                 DTRACE_PROBE3(tcp_update_lso, boolean_t, tcp->tcp_lso,
3382                     boolean_t, B_FALSE, uint32_t, tcp->tcp_lso_max);
3383 
3384                 /*
3385                  * If LSO to be disabled, notify the STREAM header with smaller
3386                  * data block. And need to restore fragsize to PMTU.
3387                  */
3388                 if (tcp->tcp_lso) {
3389                         tcp->tcp_maxpsz_multiplier =
3390                             tcp->tcp_tcps->tcps_maxpsz_multiplier;
3391                         ixa->ixa_fragsize = ixa->ixa_pmtu;
3392                         tcp->tcp_lso = B_FALSE;
3393                         TCP_STAT(tcp->tcp_tcps, tcp_lso_disabled);
3394                 }
3395         }
3396 
3397         (void) tcp_maxpsz_set(tcp, B_TRUE);
3398 }
3399 
3400 /*
3401  * Update the TCP connection according to change of ZEROCOPY capability.
3402  */
3403 static void
3404 tcp_update_zcopy(tcp_t *tcp)
3405 {
3406         conn_t          *connp = tcp->tcp_connp;
3407         tcp_stack_t     *tcps = tcp->tcp_tcps;
3408 
3409         if (tcp->tcp_snd_zcopy_on) {
3410                 tcp->tcp_snd_zcopy_on = B_FALSE;
3411                 if (!TCP_IS_DETACHED(tcp)) {
3412                         (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3413                             ZCVMUNSAFE);
3414                         TCP_STAT(tcps, tcp_zcopy_off);
3415                 }
3416         } else {
3417                 tcp->tcp_snd_zcopy_on = B_TRUE;
3418                 if (!TCP_IS_DETACHED(tcp)) {
3419                         (void) proto_set_tx_copyopt(connp->conn_rq, connp,
3420                             ZCVMSAFE);
3421                         TCP_STAT(tcps, tcp_zcopy_on);
3422                 }
3423         }
3424 }
3425 
3426 /*
3427  * Notify function registered with ip_xmit_attr_t. It's called in the squeue
3428  * so it's safe to update the TCP connection.
3429  */
3430 /* ARGSUSED1 */
3431 static void
3432 tcp_notify(void *arg, ip_xmit_attr_t *ixa, ixa_notify_type_t ntype,
3433     ixa_notify_arg_t narg)
3434 {
3435         tcp_t           *tcp = (tcp_t *)arg;
3436         conn_t          *connp = tcp->tcp_connp;
3437 
3438         switch (ntype) {
3439         case IXAN_LSO:
3440                 tcp_update_lso(tcp, connp->conn_ixa);
3441                 break;
3442         case IXAN_PMTU:
3443                 (void) tcp_update_pmtu(tcp, B_FALSE);
3444                 break;
3445         case IXAN_ZCOPY:
3446                 tcp_update_zcopy(tcp);
3447                 break;
3448         default:
3449                 break;
3450         }
3451 }
3452 
3453 /*
3454  * The TCP write service routine should never be called...
3455  */
3456 /* ARGSUSED */
3457 static void
3458 tcp_wsrv(queue_t *q)
3459 {
3460         tcp_stack_t     *tcps = Q_TO_TCP(q)->tcp_tcps;
3461 
3462         TCP_STAT(tcps, tcp_wsrv_called);
3463 }
3464 
3465 /*
3466  * Hash list lookup routine for tcp_t structures.
3467  * Returns with a CONN_INC_REF tcp structure. Caller must do a CONN_DEC_REF.
3468  */
3469 tcp_t *
3470 tcp_acceptor_hash_lookup(t_uscalar_t id, tcp_stack_t *tcps)
3471 {
3472         tf_t    *tf;
3473         tcp_t   *tcp;
3474 
3475         tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3476         mutex_enter(&tf->tf_lock);
3477         for (tcp = tf->tf_tcp; tcp != NULL;
3478             tcp = tcp->tcp_acceptor_hash) {
3479                 if (tcp->tcp_acceptor_id == id) {
3480                         CONN_INC_REF(tcp->tcp_connp);
3481                         mutex_exit(&tf->tf_lock);
3482                         return (tcp);
3483                 }
3484         }
3485         mutex_exit(&tf->tf_lock);
3486         return (NULL);
3487 }
3488 
3489 /*
3490  * Hash list insertion routine for tcp_t structures.
3491  */
3492 void
3493 tcp_acceptor_hash_insert(t_uscalar_t id, tcp_t *tcp)
3494 {
3495         tf_t    *tf;
3496         tcp_t   **tcpp;
3497         tcp_t   *tcpnext;
3498         tcp_stack_t     *tcps = tcp->tcp_tcps;
3499 
3500         tf = &tcps->tcps_acceptor_fanout[TCP_ACCEPTOR_HASH(id)];
3501 
3502         if (tcp->tcp_ptpahn != NULL)
3503                 tcp_acceptor_hash_remove(tcp);
3504         tcpp = &tf->tf_tcp;
3505         mutex_enter(&tf->tf_lock);
3506         tcpnext = tcpp[0];
3507         if (tcpnext)
3508                 tcpnext->tcp_ptpahn = &tcp->tcp_acceptor_hash;
3509         tcp->tcp_acceptor_hash = tcpnext;
3510         tcp->tcp_ptpahn = tcpp;
3511         tcpp[0] = tcp;
3512         tcp->tcp_acceptor_lockp = &tf->tf_lock;       /* For tcp_*_hash_remove */
3513         mutex_exit(&tf->tf_lock);
3514 }
3515 
3516 /*
3517  * Hash list removal routine for tcp_t structures.
3518  */
3519 void
3520 tcp_acceptor_hash_remove(tcp_t *tcp)
3521 {
3522         tcp_t   *tcpnext;
3523         kmutex_t *lockp;
3524 
3525         /*
3526          * Extract the lock pointer in case there are concurrent
3527          * hash_remove's for this instance.
3528          */
3529         lockp = tcp->tcp_acceptor_lockp;
3530 
3531         if (tcp->tcp_ptpahn == NULL)
3532                 return;
3533 
3534         ASSERT(lockp != NULL);
3535         mutex_enter(lockp);
3536         if (tcp->tcp_ptpahn) {
3537                 tcpnext = tcp->tcp_acceptor_hash;
3538                 if (tcpnext) {
3539                         tcpnext->tcp_ptpahn = tcp->tcp_ptpahn;
3540                         tcp->tcp_acceptor_hash = NULL;
3541                 }
3542                 *tcp->tcp_ptpahn = tcpnext;
3543                 tcp->tcp_ptpahn = NULL;
3544         }
3545         mutex_exit(lockp);
3546         tcp->tcp_acceptor_lockp = NULL;
3547 }
3548 
3549 /*
3550  * Type three generator adapted from the random() function in 4.4 BSD:
3551  */
3552 
3553 /*
3554  * Copyright (c) 1983, 1993
3555  *      The Regents of the University of California.  All rights reserved.
3556  *
3557  * Redistribution and use in source and binary forms, with or without
3558  * modification, are permitted provided that the following conditions
3559  * are met:
3560  * 1. Redistributions of source code must retain the above copyright
3561  *    notice, this list of conditions and the following disclaimer.
3562  * 2. Redistributions in binary form must reproduce the above copyright
3563  *    notice, this list of conditions and the following disclaimer in the
3564  *    documentation and/or other materials provided with the distribution.
3565  * 3. All advertising materials mentioning features or use of this software
3566  *    must display the following acknowledgement:
3567  *      This product includes software developed by the University of
3568  *      California, Berkeley and its contributors.
3569  * 4. Neither the name of the University nor the names of its contributors
3570  *    may be used to endorse or promote products derived from this software
3571  *    without specific prior written permission.
3572  *
3573  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
3574  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
3575  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
3576  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
3577  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
3578  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
3579  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
3580  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
3581  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
3582  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
3583  * SUCH DAMAGE.
3584  */
3585 
3586 /* Type 3 -- x**31 + x**3 + 1 */
3587 #define DEG_3           31
3588 #define SEP_3           3
3589 
3590 
3591 /* Protected by tcp_random_lock */
3592 static int tcp_randtbl[DEG_3 + 1];
3593 
3594 static int *tcp_random_fptr = &tcp_randtbl[SEP_3 + 1];
3595 static int *tcp_random_rptr = &tcp_randtbl[1];
3596 
3597 static int *tcp_random_state = &tcp_randtbl[1];
3598 static int *tcp_random_end_ptr = &tcp_randtbl[DEG_3 + 1];
3599 
3600 kmutex_t tcp_random_lock;
3601 
3602 void
3603 tcp_random_init(void)
3604 {
3605         int i;
3606         hrtime_t hrt;
3607         time_t wallclock;
3608         uint64_t result;
3609 
3610         /*
3611          * Use high-res timer and current time for seed.  Gethrtime() returns
3612          * a longlong, which may contain resolution down to nanoseconds.
3613          * The current time will either be a 32-bit or a 64-bit quantity.
3614          * XOR the two together in a 64-bit result variable.
3615          * Convert the result to a 32-bit value by multiplying the high-order
3616          * 32-bits by the low-order 32-bits.
3617          */
3618 
3619         hrt = gethrtime();
3620         (void) drv_getparm(TIME, &wallclock);
3621         result = (uint64_t)wallclock ^ (uint64_t)hrt;
3622         mutex_enter(&tcp_random_lock);
3623         tcp_random_state[0] = ((result >> 32) & 0xffffffff) *
3624             (result & 0xffffffff);
3625 
3626         for (i = 1; i < DEG_3; i++)
3627                 tcp_random_state[i] = 1103515245 * tcp_random_state[i - 1]
3628                     + 12345;
3629         tcp_random_fptr = &tcp_random_state[SEP_3];
3630         tcp_random_rptr = &tcp_random_state[0];
3631         mutex_exit(&tcp_random_lock);
3632         for (i = 0; i < 10 * DEG_3; i++)
3633                 (void) tcp_random();
3634 }
3635 
3636 /*
3637  * tcp_random: Return a random number in the range [1 - (128K + 1)].
3638  * This range is selected to be approximately centered on TCP_ISS / 2,
3639  * and easy to compute. We get this value by generating a 32-bit random
3640  * number, selecting out the high-order 17 bits, and then adding one so
3641  * that we never return zero.
3642  */
3643 int
3644 tcp_random(void)
3645 {
3646         int i;
3647 
3648         mutex_enter(&tcp_random_lock);
3649         *tcp_random_fptr += *tcp_random_rptr;
3650 
3651         /*
3652          * The high-order bits are more random than the low-order bits,
3653          * so we select out the high-order 17 bits and add one so that
3654          * we never return zero.
3655          */
3656         i = ((*tcp_random_fptr >> 15) & 0x1ffff) + 1;
3657         if (++tcp_random_fptr >= tcp_random_end_ptr) {
3658                 tcp_random_fptr = tcp_random_state;
3659                 ++tcp_random_rptr;
3660         } else if (++tcp_random_rptr >= tcp_random_end_ptr)
3661                 tcp_random_rptr = tcp_random_state;
3662 
3663         mutex_exit(&tcp_random_lock);
3664         return (i);
3665 }
3666 
3667 /*
3668  * Split this function out so that if the secret changes, I'm okay.
3669  *
3670  * Initialize the tcp_iss_cookie and tcp_iss_key.
3671  */
3672 
3673 #define PASSWD_SIZE 16  /* MUST be multiple of 4 */
3674 
3675 void
3676 tcp_iss_key_init(uint8_t *phrase, int len, tcp_stack_t *tcps)
3677 {
3678         struct {
3679                 int32_t current_time;
3680                 uint32_t randnum;
3681                 uint16_t pad;
3682                 uint8_t ether[6];
3683                 uint8_t passwd[PASSWD_SIZE];
3684         } tcp_iss_cookie;
3685         time_t t;
3686 
3687         /*
3688          * Start with the current absolute time.
3689          */
3690         (void) drv_getparm(TIME, &t);
3691         tcp_iss_cookie.current_time = t;
3692 
3693         /*
3694          * XXX - Need a more random number per RFC 1750, not this crap.
3695          * OTOH, if what follows is pretty random, then I'm in better shape.
3696          */
3697         tcp_iss_cookie.randnum = (uint32_t)(gethrtime() + tcp_random());
3698         tcp_iss_cookie.pad = 0x365c;  /* Picked from HMAC pad values. */
3699 
3700         /*
3701          * The cpu_type_info is pretty non-random.  Ugggh.  It does serve
3702          * as a good template.
3703          */
3704         bcopy(&cpu_list->cpu_type_info, &tcp_iss_cookie.passwd,
3705             min(PASSWD_SIZE, sizeof (cpu_list->cpu_type_info)));
3706 
3707         /*
3708          * The pass-phrase.  Normally this is supplied by user-called NDD.
3709          */
3710         bcopy(phrase, &tcp_iss_cookie.passwd, min(PASSWD_SIZE, len));
3711 
3712         /*
3713          * See 4010593 if this section becomes a problem again,
3714          * but the local ethernet address is useful here.
3715          */
3716         (void) localetheraddr(NULL,
3717             (struct ether_addr *)&tcp_iss_cookie.ether);
3718 
3719         /*
3720          * Hash 'em all together.  The MD5Final is called per-connection.
3721          */
3722         mutex_enter(&tcps->tcps_iss_key_lock);
3723         MD5Init(&tcps->tcps_iss_key);
3724         MD5Update(&tcps->tcps_iss_key, (uchar_t *)&tcp_iss_cookie,
3725             sizeof (tcp_iss_cookie));
3726         mutex_exit(&tcps->tcps_iss_key_lock);
3727 }
3728 
3729 /*
3730  * Called by IP when IP is loaded into the kernel
3731  */
3732 void
3733 tcp_ddi_g_init(void)
3734 {
3735         tcp_timercache = kmem_cache_create("tcp_timercache",
3736             sizeof (tcp_timer_t) + sizeof (mblk_t), 0,
3737             NULL, NULL, NULL, NULL, NULL, 0);
3738 
3739         tcp_notsack_blk_cache = kmem_cache_create("tcp_notsack_blk_cache",
3740             sizeof (notsack_blk_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3741 
3742         mutex_init(&tcp_random_lock, NULL, MUTEX_DEFAULT, NULL);
3743 
3744         /* Initialize the random number generator */
3745         tcp_random_init();
3746 
3747         /* A single callback independently of how many netstacks we have */
3748         ip_squeue_init(tcp_squeue_add);
3749 
3750         tcp_g_kstat = tcp_g_kstat_init(&tcp_g_statistics);
3751 
3752         tcp_squeue_flag = tcp_squeue_switch(tcp_squeue_wput);
3753 
3754         /*
3755          * We want to be informed each time a stack is created or
3756          * destroyed in the kernel, so we can maintain the
3757          * set of tcp_stack_t's.
3758          */
3759         netstack_register(NS_TCP, tcp_stack_init, NULL, tcp_stack_fini);
3760 }
3761 
3762 
3763 #define INET_NAME       "ip"
3764 
3765 /*
3766  * Initialize the TCP stack instance.
3767  */
3768 static void *
3769 tcp_stack_init(netstackid_t stackid, netstack_t *ns)
3770 {
3771         tcp_stack_t     *tcps;
3772         int             i;
3773         major_t         major;
3774         size_t          arrsz;
3775 
3776         tcps = (tcp_stack_t *)kmem_zalloc(sizeof (*tcps), KM_SLEEP);
3777         tcps->tcps_netstack = ns;
3778 
3779         /* Initialize locks */
3780         mutex_init(&tcps->tcps_iss_key_lock, NULL, MUTEX_DEFAULT, NULL);
3781         mutex_init(&tcps->tcps_epriv_port_lock, NULL, MUTEX_DEFAULT, NULL);
3782 
3783         tcps->tcps_g_num_epriv_ports = TCP_NUM_EPRIV_PORTS;
3784         tcps->tcps_g_epriv_ports[0] = ULP_DEF_EPRIV_PORT1;
3785         tcps->tcps_g_epriv_ports[1] = ULP_DEF_EPRIV_PORT2;
3786         tcps->tcps_min_anonpriv_port = 512;
3787 
3788         tcps->tcps_bind_fanout = kmem_zalloc(sizeof (tf_t) *
3789             TCP_BIND_FANOUT_SIZE, KM_SLEEP);
3790         tcps->tcps_acceptor_fanout = kmem_zalloc(sizeof (tf_t) *
3791             TCP_ACCEPTOR_FANOUT_SIZE, KM_SLEEP);
3792 
3793         for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3794                 mutex_init(&tcps->tcps_bind_fanout[i].tf_lock, NULL,
3795                     MUTEX_DEFAULT, NULL);
3796         }
3797 
3798         for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3799                 mutex_init(&tcps->tcps_acceptor_fanout[i].tf_lock, NULL,
3800                     MUTEX_DEFAULT, NULL);
3801         }
3802 
3803         /* TCP's IPsec code calls the packet dropper. */
3804         ip_drop_register(&tcps->tcps_dropper, "TCP IPsec policy enforcement");
3805 
3806         arrsz = tcp_propinfo_count * sizeof (mod_prop_info_t);
3807         tcps->tcps_propinfo_tbl = (mod_prop_info_t *)kmem_alloc(arrsz,
3808             KM_SLEEP);
3809         bcopy(tcp_propinfo_tbl, tcps->tcps_propinfo_tbl, arrsz);
3810 
3811         /*
3812          * Note: To really walk the device tree you need the devinfo
3813          * pointer to your device which is only available after probe/attach.
3814          * The following is safe only because it uses ddi_root_node()
3815          */
3816         tcp_max_optsize = optcom_max_optsize(tcp_opt_obj.odb_opt_des_arr,
3817             tcp_opt_obj.odb_opt_arr_cnt);
3818 
3819         /*
3820          * Initialize RFC 1948 secret values.  This will probably be reset once
3821          * by the boot scripts.
3822          *
3823          * Use NULL name, as the name is caught by the new lockstats.
3824          *
3825          * Initialize with some random, non-guessable string, like the global
3826          * T_INFO_ACK.
3827          */
3828 
3829         tcp_iss_key_init((uint8_t *)&tcp_g_t_info_ack,
3830             sizeof (tcp_g_t_info_ack), tcps);
3831 
3832         tcps->tcps_kstat = tcp_kstat2_init(stackid);
3833         tcps->tcps_mibkp = tcp_kstat_init(stackid);
3834 
3835         major = mod_name_to_major(INET_NAME);
3836         VERIFY0(ldi_ident_from_major(major, &tcps->tcps_ldi_ident));
3837         tcps->tcps_ixa_cleanup_mp = allocb_wait(0, BPRI_MED, STR_NOSIG, NULL);
3838         ASSERT(tcps->tcps_ixa_cleanup_mp != NULL);
3839         cv_init(&tcps->tcps_ixa_cleanup_ready_cv, NULL, CV_DEFAULT, NULL);
3840         cv_init(&tcps->tcps_ixa_cleanup_done_cv, NULL, CV_DEFAULT, NULL);
3841         mutex_init(&tcps->tcps_ixa_cleanup_lock, NULL, MUTEX_DEFAULT, NULL);
3842 
3843         mutex_init(&tcps->tcps_reclaim_lock, NULL, MUTEX_DEFAULT, NULL);
3844         tcps->tcps_reclaim = B_FALSE;
3845         tcps->tcps_reclaim_tid = 0;
3846         tcps->tcps_reclaim_period = tcps->tcps_rexmit_interval_max;
3847 
3848         /*
3849          * ncpus is the current number of CPUs, which can be bigger than
3850          * boot_ncpus.  But we don't want to use ncpus to allocate all the
3851          * tcp_stats_cpu_t at system boot up time since it will be 1.  While
3852          * we handle adding CPU in tcp_cpu_update(), it will be slow if
3853          * there are many CPUs as we will be adding them 1 by 1.
3854          *
3855          * Note that tcps_sc_cnt never decreases and the tcps_sc[x] pointers
3856          * are not freed until the stack is going away.  So there is no need
3857          * to grab a lock to access the per CPU tcps_sc[x] pointer.
3858          */
3859         mutex_enter(&cpu_lock);
3860         tcps->tcps_sc_cnt = MAX(ncpus, boot_ncpus);
3861         mutex_exit(&cpu_lock);
3862         tcps->tcps_sc = kmem_zalloc(max_ncpus  * sizeof (tcp_stats_cpu_t *),
3863             KM_SLEEP);
3864         for (i = 0; i < tcps->tcps_sc_cnt; i++) {
3865                 tcps->tcps_sc[i] = kmem_zalloc(sizeof (tcp_stats_cpu_t),
3866                     KM_SLEEP);
3867         }
3868 
3869         mutex_init(&tcps->tcps_listener_conf_lock, NULL, MUTEX_DEFAULT, NULL);
3870         list_create(&tcps->tcps_listener_conf, sizeof (tcp_listener_t),
3871             offsetof(tcp_listener_t, tl_link));
3872 
3873         tcps->tcps_default_cc_algo = cc_load_algo(CC_DEFAULT_ALGO_NAME);
3874         ASSERT3P(tcps->tcps_default_cc_algo, !=, NULL);
3875 
3876         return (tcps);
3877 }
3878 
3879 /*
3880  * Called when the IP module is about to be unloaded.
3881  */
3882 void
3883 tcp_ddi_g_destroy(void)
3884 {
3885         tcp_g_kstat_fini(tcp_g_kstat);
3886         tcp_g_kstat = NULL;
3887         bzero(&tcp_g_statistics, sizeof (tcp_g_statistics));
3888 
3889         mutex_destroy(&tcp_random_lock);
3890 
3891         kmem_cache_destroy(tcp_timercache);
3892         kmem_cache_destroy(tcp_notsack_blk_cache);
3893 
3894         netstack_unregister(NS_TCP);
3895 }
3896 
3897 /*
3898  * Free the TCP stack instance.
3899  */
3900 static void
3901 tcp_stack_fini(netstackid_t stackid, void *arg)
3902 {
3903         tcp_stack_t *tcps = (tcp_stack_t *)arg;
3904         int i;
3905 
3906         freeb(tcps->tcps_ixa_cleanup_mp);
3907         tcps->tcps_ixa_cleanup_mp = NULL;
3908         cv_destroy(&tcps->tcps_ixa_cleanup_ready_cv);
3909         cv_destroy(&tcps->tcps_ixa_cleanup_done_cv);
3910         mutex_destroy(&tcps->tcps_ixa_cleanup_lock);
3911 
3912         /*
3913          * Set tcps_reclaim to false tells tcp_reclaim_timer() not to restart
3914          * the timer.
3915          */
3916         mutex_enter(&tcps->tcps_reclaim_lock);
3917         tcps->tcps_reclaim = B_FALSE;
3918         mutex_exit(&tcps->tcps_reclaim_lock);
3919         if (tcps->tcps_reclaim_tid != 0)
3920                 (void) untimeout(tcps->tcps_reclaim_tid);
3921         mutex_destroy(&tcps->tcps_reclaim_lock);
3922 
3923         tcp_listener_conf_cleanup(tcps);
3924 
3925         for (i = 0; i < tcps->tcps_sc_cnt; i++)
3926                 kmem_free(tcps->tcps_sc[i], sizeof (tcp_stats_cpu_t));
3927         kmem_free(tcps->tcps_sc, max_ncpus * sizeof (tcp_stats_cpu_t *));
3928 
3929         kmem_free(tcps->tcps_propinfo_tbl,
3930             tcp_propinfo_count * sizeof (mod_prop_info_t));
3931         tcps->tcps_propinfo_tbl = NULL;
3932 
3933         for (i = 0; i < TCP_BIND_FANOUT_SIZE; i++) {
3934                 ASSERT(tcps->tcps_bind_fanout[i].tf_tcp == NULL);
3935                 mutex_destroy(&tcps->tcps_bind_fanout[i].tf_lock);
3936         }
3937 
3938         for (i = 0; i < TCP_ACCEPTOR_FANOUT_SIZE; i++) {
3939                 ASSERT(tcps->tcps_acceptor_fanout[i].tf_tcp == NULL);
3940                 mutex_destroy(&tcps->tcps_acceptor_fanout[i].tf_lock);
3941         }
3942 
3943         kmem_free(tcps->tcps_bind_fanout, sizeof (tf_t) * TCP_BIND_FANOUT_SIZE);
3944         tcps->tcps_bind_fanout = NULL;
3945 
3946         kmem_free(tcps->tcps_acceptor_fanout, sizeof (tf_t) *
3947             TCP_ACCEPTOR_FANOUT_SIZE);
3948         tcps->tcps_acceptor_fanout = NULL;
3949 
3950         mutex_destroy(&tcps->tcps_iss_key_lock);
3951         mutex_destroy(&tcps->tcps_epriv_port_lock);
3952 
3953         ip_drop_unregister(&tcps->tcps_dropper);
3954 
3955         tcp_kstat2_fini(stackid, tcps->tcps_kstat);
3956         tcps->tcps_kstat = NULL;
3957 
3958         tcp_kstat_fini(stackid, tcps->tcps_mibkp);
3959         tcps->tcps_mibkp = NULL;
3960 
3961         ldi_ident_release(tcps->tcps_ldi_ident);
3962         kmem_free(tcps, sizeof (*tcps));
3963 }
3964 
3965 /*
3966  * Generate ISS, taking into account NDD changes may happen halfway through.
3967  * (If the iss is not zero, set it.)
3968  */
3969 
3970 static void
3971 tcp_iss_init(tcp_t *tcp)
3972 {
3973         MD5_CTX context;
3974         struct { uint32_t ports; in6_addr_t src; in6_addr_t dst; } arg;
3975         uint32_t answer[4];
3976         tcp_stack_t     *tcps = tcp->tcp_tcps;
3977         conn_t          *connp = tcp->tcp_connp;
3978 
3979         tcps->tcps_iss_incr_extra += (tcps->tcps_iss_incr >> 1);
3980         tcp->tcp_iss = tcps->tcps_iss_incr_extra;
3981         switch (tcps->tcps_strong_iss) {
3982         case 2:
3983                 mutex_enter(&tcps->tcps_iss_key_lock);
3984                 context = tcps->tcps_iss_key;
3985                 mutex_exit(&tcps->tcps_iss_key_lock);
3986                 arg.ports = connp->conn_ports;
3987                 arg.src = connp->conn_laddr_v6;
3988                 arg.dst = connp->conn_faddr_v6;
3989                 MD5Update(&context, (uchar_t *)&arg, sizeof (arg));
3990                 MD5Final((uchar_t *)answer, &context);
3991                 tcp->tcp_iss += answer[0] ^ answer[1] ^ answer[2] ^ answer[3];
3992                 /*
3993                  * Now that we've hashed into a unique per-connection sequence
3994                  * space, add a random increment per strong_iss == 1.  So I
3995                  * guess we'll have to...
3996                  */
3997                 /* FALLTHRU */
3998         case 1:
3999                 tcp->tcp_iss += (gethrtime() >> ISS_NSEC_SHT) + tcp_random();
4000                 break;
4001         default:
4002                 tcp->tcp_iss += (uint32_t)gethrestime_sec() *
4003                     tcps->tcps_iss_incr;
4004                 break;
4005         }
4006         tcp->tcp_valid_bits = TCP_ISS_VALID;
4007         tcp->tcp_fss = tcp->tcp_iss - 1;
4008         tcp->tcp_suna = tcp->tcp_iss;
4009         tcp->tcp_snxt = tcp->tcp_iss + 1;
4010         tcp->tcp_rexmit_nxt = tcp->tcp_snxt;
4011         tcp->tcp_csuna = tcp->tcp_snxt;
4012 }
4013 
4014 /*
4015  * tcp_{set,clr}qfull() functions are used to either set or clear QFULL
4016  * on the specified backing STREAMS q. Note, the caller may make the
4017  * decision to call based on the tcp_t.tcp_flow_stopped value which
4018  * when check outside the q's lock is only an advisory check ...
4019  */
4020 void
4021 tcp_setqfull(tcp_t *tcp)
4022 {
4023         tcp_stack_t     *tcps = tcp->tcp_tcps;
4024         conn_t  *connp = tcp->tcp_connp;
4025 
4026         if (tcp->tcp_closed)
4027                 return;
4028 
4029         conn_setqfull(connp, &tcp->tcp_flow_stopped);
4030         if (tcp->tcp_flow_stopped)
4031                 TCP_STAT(tcps, tcp_flwctl_on);
4032 }
4033 
4034 void
4035 tcp_clrqfull(tcp_t *tcp)
4036 {
4037         conn_t  *connp = tcp->tcp_connp;
4038 
4039         if (tcp->tcp_closed)
4040                 return;
4041         conn_clrqfull(connp, &tcp->tcp_flow_stopped);
4042 }
4043 
4044 static int
4045 tcp_squeue_switch(int val)
4046 {
4047         int rval = SQ_FILL;
4048 
4049         switch (val) {
4050         case 1:
4051                 rval = SQ_NODRAIN;
4052                 break;
4053         case 2:
4054                 rval = SQ_PROCESS;
4055                 break;
4056         default:
4057                 break;
4058         }
4059         return (rval);
4060 }
4061 
4062 /*
4063  * This is called once for each squeue - globally for all stack
4064  * instances.
4065  */
4066 static void
4067 tcp_squeue_add(squeue_t *sqp)
4068 {
4069         tcp_squeue_priv_t *tcp_time_wait = kmem_zalloc(
4070             sizeof (tcp_squeue_priv_t), KM_SLEEP);
4071 
4072         *squeue_getprivate(sqp, SQPRIVATE_TCP) = (intptr_t)tcp_time_wait;
4073         if (tcp_free_list_max_cnt == 0) {
4074                 int tcp_ncpus = ((boot_max_ncpus == -1) ?
4075                     max_ncpus : boot_max_ncpus);
4076 
4077                 /*
4078                  * Limit number of entries to 1% of availble memory / tcp_ncpus
4079                  */
4080                 tcp_free_list_max_cnt = (freemem * PAGESIZE) /
4081                     (tcp_ncpus * sizeof (tcp_t) * 100);
4082         }
4083         tcp_time_wait->tcp_free_list_cnt = 0;
4084 }
4085 /*
4086  * Return unix error is tli error is TSYSERR, otherwise return a negative
4087  * tli error.
4088  */
4089 int
4090 tcp_do_bind(conn_t *connp, struct sockaddr *sa, socklen_t len, cred_t *cr,
4091     boolean_t bind_to_req_port_only)
4092 {
4093         int error;
4094         tcp_t *tcp = connp->conn_tcp;
4095 
4096         if (tcp->tcp_state >= TCPS_BOUND) {
4097                 if (connp->conn_debug) {
4098                         (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4099                             "tcp_bind: bad state, %d", tcp->tcp_state);
4100                 }
4101                 return (-TOUTSTATE);
4102         }
4103 
4104         error = tcp_bind_check(connp, sa, len, cr, bind_to_req_port_only);
4105         if (error != 0)
4106                 return (error);
4107 
4108         ASSERT(tcp->tcp_state == TCPS_BOUND);
4109         tcp->tcp_conn_req_max = 0;
4110         return (0);
4111 }
4112 
4113 /*
4114  * If the return value from this function is positive, it's a UNIX error.
4115  * Otherwise, if it's negative, then the absolute value is a TLI error.
4116  * the TPI routine tcp_tpi_connect() is a wrapper function for this.
4117  */
4118 int
4119 tcp_do_connect(conn_t *connp, const struct sockaddr *sa, socklen_t len,
4120     cred_t *cr, pid_t pid)
4121 {
4122         tcp_t           *tcp = connp->conn_tcp;
4123         sin_t           *sin = (sin_t *)sa;
4124         sin6_t          *sin6 = (sin6_t *)sa;
4125         ipaddr_t        *dstaddrp;
4126         in_port_t       dstport;
4127         uint_t          srcid;
4128         int             error;
4129         uint32_t        mss;
4130         mblk_t          *syn_mp;
4131         tcp_stack_t     *tcps = tcp->tcp_tcps;
4132         int32_t         oldstate;
4133         ip_xmit_attr_t  *ixa = connp->conn_ixa;
4134 
4135         oldstate = tcp->tcp_state;
4136 
4137         switch (len) {
4138         default:
4139                 /*
4140                  * Should never happen
4141                  */
4142                 return (EINVAL);
4143 
4144         case sizeof (sin_t):
4145                 sin = (sin_t *)sa;
4146                 if (sin->sin_port == 0) {
4147                         return (-TBADADDR);
4148                 }
4149                 if (connp->conn_ipv6_v6only) {
4150                         return (EAFNOSUPPORT);
4151                 }
4152                 break;
4153 
4154         case sizeof (sin6_t):
4155                 sin6 = (sin6_t *)sa;
4156                 if (sin6->sin6_port == 0) {
4157                         return (-TBADADDR);
4158                 }
4159                 break;
4160         }
4161         /*
4162          * If we're connecting to an IPv4-mapped IPv6 address, we need to
4163          * make sure that the conn_ipversion is IPV4_VERSION.  We
4164          * need to this before we call tcp_bindi() so that the port lookup
4165          * code will look for ports in the correct port space (IPv4 and
4166          * IPv6 have separate port spaces).
4167          */
4168         if (connp->conn_family == AF_INET6 &&
4169             connp->conn_ipversion == IPV6_VERSION &&
4170             IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4171                 if (connp->conn_ipv6_v6only)
4172                         return (EADDRNOTAVAIL);
4173 
4174                 connp->conn_ipversion = IPV4_VERSION;
4175         }
4176 
4177         switch (tcp->tcp_state) {
4178         case TCPS_LISTEN:
4179                 /*
4180                  * Listening sockets are not allowed to issue connect().
4181                  */
4182                 if (IPCL_IS_NONSTR(connp))
4183                         return (EOPNOTSUPP);
4184                 /* FALLTHRU */
4185         case TCPS_IDLE:
4186                 /*
4187                  * We support quick connect, refer to comments in
4188                  * tcp_connect_*()
4189                  */
4190                 /* FALLTHRU */
4191         case TCPS_BOUND:
4192                 break;
4193         default:
4194                 return (-TOUTSTATE);
4195         }
4196 
4197         /*
4198          * We update our cred/cpid based on the caller of connect
4199          */
4200         if (connp->conn_cred != cr) {
4201                 crhold(cr);
4202                 crfree(connp->conn_cred);
4203                 connp->conn_cred = cr;
4204         }
4205         connp->conn_cpid = pid;
4206 
4207         /* Cache things in the ixa without any refhold */
4208         ASSERT(!(ixa->ixa_free_flags & IXA_FREE_CRED));
4209         ixa->ixa_cred = cr;
4210         ixa->ixa_cpid = pid;
4211         if (is_system_labeled()) {
4212                 /* We need to restart with a label based on the cred */
4213                 ip_xmit_attr_restore_tsl(ixa, ixa->ixa_cred);
4214         }
4215 
4216         if (connp->conn_family == AF_INET6) {
4217                 if (!IN6_IS_ADDR_V4MAPPED(&sin6->sin6_addr)) {
4218                         error = tcp_connect_ipv6(tcp, &sin6->sin6_addr,
4219                             sin6->sin6_port, sin6->sin6_flowinfo,
4220                             sin6->__sin6_src_id, sin6->sin6_scope_id);
4221                 } else {
4222                         /*
4223                          * Destination adress is mapped IPv6 address.
4224                          * Source bound address should be unspecified or
4225                          * IPv6 mapped address as well.
4226                          */
4227                         if (!IN6_IS_ADDR_UNSPECIFIED(
4228                             &connp->conn_bound_addr_v6) &&
4229                             !IN6_IS_ADDR_V4MAPPED(&connp->conn_bound_addr_v6)) {
4230                                 return (EADDRNOTAVAIL);
4231                         }
4232                         dstaddrp = &V4_PART_OF_V6((sin6->sin6_addr));
4233                         dstport = sin6->sin6_port;
4234                         srcid = sin6->__sin6_src_id;
4235                         error = tcp_connect_ipv4(tcp, dstaddrp, dstport,
4236                             srcid);
4237                 }
4238         } else {
4239                 dstaddrp = &sin->sin_addr.s_addr;
4240                 dstport = sin->sin_port;
4241                 srcid = 0;
4242                 error = tcp_connect_ipv4(tcp, dstaddrp, dstport, srcid);
4243         }
4244 
4245         if (error != 0)
4246                 goto connect_failed;
4247 
4248         CL_INET_CONNECT(connp, B_TRUE, error);
4249         if (error != 0)
4250                 goto connect_failed;
4251 
4252         /* connect succeeded */
4253         TCPS_BUMP_MIB(tcps, tcpActiveOpens);
4254         tcp->tcp_active_open = 1;
4255 
4256         /*
4257          * tcp_set_destination() does not adjust for TCP/IP header length.
4258          */
4259         mss = tcp->tcp_mss - connp->conn_ht_iphc_len;
4260 
4261         /*
4262          * Just make sure our rwnd is at least rcvbuf * MSS large, and round up
4263          * to the nearest MSS.
4264          *
4265          * We do the round up here because we need to get the interface MTU
4266          * first before we can do the round up.
4267          */
4268         tcp->tcp_rwnd = connp->conn_rcvbuf;
4269         tcp->tcp_rwnd = MAX(MSS_ROUNDUP(tcp->tcp_rwnd, mss),
4270             tcps->tcps_recv_hiwat_minmss * mss);
4271         connp->conn_rcvbuf = tcp->tcp_rwnd;
4272         tcp_set_ws_value(tcp);
4273         tcp->tcp_tcpha->tha_win = htons(tcp->tcp_rwnd >> tcp->tcp_rcv_ws);
4274         if (tcp->tcp_rcv_ws > 0 || tcps->tcps_wscale_always)
4275                 tcp->tcp_snd_ws_ok = B_TRUE;
4276 
4277         /*
4278          * Set tcp_snd_ts_ok to true
4279          * so that tcp_xmit_mp will
4280          * include the timestamp
4281          * option in the SYN segment.
4282          */
4283         if (tcps->tcps_tstamp_always ||
4284             (tcp->tcp_rcv_ws && tcps->tcps_tstamp_if_wscale)) {
4285                 tcp->tcp_snd_ts_ok = B_TRUE;
4286         }
4287 
4288         /*
4289          * Note that tcp_snd_sack_ok can be set in tcp_set_destination() if
4290          * the SACK metric is set.  So here we just check the per stack SACK
4291          * permitted param.
4292          */
4293         if (tcps->tcps_sack_permitted == 2) {
4294                 ASSERT(tcp->tcp_num_sack_blk == 0);
4295                 ASSERT(tcp->tcp_notsack_list == NULL);
4296                 tcp->tcp_snd_sack_ok = B_TRUE;
4297         }
4298 
4299         /*
4300          * Should we use ECN?  Note that the current
4301          * default value (SunOS 5.9) of tcp_ecn_permitted
4302          * is 1.  The reason for doing this is that there
4303          * are equipments out there that will drop ECN
4304          * enabled IP packets.  Setting it to 1 avoids
4305          * compatibility problems.
4306          */
4307         if (tcps->tcps_ecn_permitted == 2)
4308                 tcp->tcp_ecn_ok = B_TRUE;
4309 
4310         /* Trace change from BOUND -> SYN_SENT here */
4311         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4312             connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4313             int32_t, TCPS_BOUND);
4314 
4315         TCP_TIMER_RESTART(tcp, tcp->tcp_rto);
4316         syn_mp = tcp_xmit_mp(tcp, NULL, 0, NULL, NULL,
4317             tcp->tcp_iss, B_FALSE, NULL, B_FALSE);
4318         if (syn_mp != NULL) {
4319                 /*
4320                  * We must bump the generation before sending the syn
4321                  * to ensure that we use the right generation in case
4322                  * this thread issues a "connected" up call.
4323                  */
4324                 SOCK_CONNID_BUMP(tcp->tcp_connid);
4325                 /*
4326                  * DTrace sending the first SYN as a
4327                  * tcp:::connect-request event.
4328                  */
4329                 DTRACE_TCP5(connect__request, mblk_t *, NULL,
4330                     ip_xmit_attr_t *, connp->conn_ixa,
4331                     void_ip_t *, syn_mp->b_rptr, tcp_t *, tcp,
4332                     tcph_t *,
4333                     &syn_mp->b_rptr[connp->conn_ixa->ixa_ip_hdr_length]);
4334                 tcp_send_data(tcp, syn_mp);
4335         }
4336 
4337         if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4338                 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4339         return (0);
4340 
4341 connect_failed:
4342         connp->conn_faddr_v6 = ipv6_all_zeros;
4343         connp->conn_fport = 0;
4344         tcp->tcp_state = oldstate;
4345         if (tcp->tcp_conn.tcp_opts_conn_req != NULL)
4346                 tcp_close_mpp(&tcp->tcp_conn.tcp_opts_conn_req);
4347         return (error);
4348 }
4349 
4350 int
4351 tcp_do_listen(conn_t *connp, struct sockaddr *sa, socklen_t len,
4352     int backlog, cred_t *cr, boolean_t bind_to_req_port_only)
4353 {
4354         tcp_t           *tcp = connp->conn_tcp;
4355         int             error = 0;
4356         tcp_stack_t     *tcps = tcp->tcp_tcps;
4357         int32_t         oldstate;
4358 
4359         /* All Solaris components should pass a cred for this operation. */
4360         ASSERT(cr != NULL);
4361 
4362         if (tcp->tcp_state >= TCPS_BOUND) {
4363                 if ((tcp->tcp_state == TCPS_BOUND ||
4364                     tcp->tcp_state == TCPS_LISTEN) && backlog > 0) {
4365                         /*
4366                          * Handle listen() increasing backlog.
4367                          * This is more "liberal" then what the TPI spec
4368                          * requires but is needed to avoid a t_unbind
4369                          * when handling listen() since the port number
4370                          * might be "stolen" between the unbind and bind.
4371                          */
4372                         goto do_listen;
4373                 }
4374                 if (connp->conn_debug) {
4375                         (void) strlog(TCP_MOD_ID, 0, 1, SL_ERROR|SL_TRACE,
4376                             "tcp_listen: bad state, %d", tcp->tcp_state);
4377                 }
4378                 return (-TOUTSTATE);
4379         } else {
4380                 if (sa == NULL) {
4381                         sin6_t  addr;
4382                         sin_t *sin;
4383                         sin6_t *sin6;
4384 
4385                         ASSERT(IPCL_IS_NONSTR(connp));
4386                         /* Do an implicit bind: Request for a generic port. */
4387                         if (connp->conn_family == AF_INET) {
4388                                 len = sizeof (sin_t);
4389                                 sin = (sin_t *)&addr;
4390                                 *sin = sin_null;
4391                                 sin->sin_family = AF_INET;
4392                         } else {
4393                                 ASSERT(connp->conn_family == AF_INET6);
4394                                 len = sizeof (sin6_t);
4395                                 sin6 = (sin6_t *)&addr;
4396                                 *sin6 = sin6_null;
4397                                 sin6->sin6_family = AF_INET6;
4398                         }
4399                         sa = (struct sockaddr *)&addr;
4400                 }
4401 
4402                 error = tcp_bind_check(connp, sa, len, cr,
4403                     bind_to_req_port_only);
4404                 if (error)
4405                         return (error);
4406                 /* Fall through and do the fanout insertion */
4407         }
4408 
4409 do_listen:
4410         ASSERT(tcp->tcp_state == TCPS_BOUND || tcp->tcp_state == TCPS_LISTEN);
4411         tcp->tcp_conn_req_max = backlog;
4412         if (tcp->tcp_conn_req_max) {
4413                 if (tcp->tcp_conn_req_max < tcps->tcps_conn_req_min)
4414                         tcp->tcp_conn_req_max = tcps->tcps_conn_req_min;
4415                 if (tcp->tcp_conn_req_max > tcps->tcps_conn_req_max_q)
4416                         tcp->tcp_conn_req_max = tcps->tcps_conn_req_max_q;
4417                 /*
4418                  * If this is a listener, do not reset the eager list
4419                  * and other stuffs.  Note that we don't check if the
4420                  * existing eager list meets the new tcp_conn_req_max
4421                  * requirement.
4422                  */
4423                 if (tcp->tcp_state != TCPS_LISTEN) {
4424                         tcp->tcp_state = TCPS_LISTEN;
4425                         DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4426                             connp->conn_ixa, void, NULL, tcp_t *, tcp,
4427                             void, NULL, int32_t, TCPS_BOUND);
4428                         /* Initialize the chain. Don't need the eager_lock */
4429                         tcp->tcp_eager_next_q0 = tcp->tcp_eager_prev_q0 = tcp;
4430                         tcp->tcp_eager_next_drop_q0 = tcp;
4431                         tcp->tcp_eager_prev_drop_q0 = tcp;
4432                         tcp->tcp_second_ctimer_threshold =
4433                             tcps->tcps_ip_abort_linterval;
4434                 }
4435         }
4436 
4437         /*
4438          * We need to make sure that the conn_recv is set to a non-null
4439          * value before we insert the conn into the classifier table.
4440          * This is to avoid a race with an incoming packet which does an
4441          * ipcl_classify().
4442          * We initially set it to tcp_input_listener_unbound to try to
4443          * pick a good squeue for the listener when the first SYN arrives.
4444          * tcp_input_listener_unbound sets it to tcp_input_listener on that
4445          * first SYN.
4446          */
4447         connp->conn_recv = tcp_input_listener_unbound;
4448 
4449         /* Insert the listener in the classifier table */
4450         error = ip_laddr_fanout_insert(connp);
4451         if (error != 0) {
4452                 /* Undo the bind - release the port number */
4453                 oldstate = tcp->tcp_state;
4454                 tcp->tcp_state = TCPS_IDLE;
4455                 DTRACE_TCP6(state__change, void, NULL, ip_xmit_attr_t *,
4456                     connp->conn_ixa, void, NULL, tcp_t *, tcp, void, NULL,
4457                     int32_t, oldstate);
4458                 connp->conn_bound_addr_v6 = ipv6_all_zeros;
4459 
4460                 connp->conn_laddr_v6 = ipv6_all_zeros;
4461                 connp->conn_saddr_v6 = ipv6_all_zeros;
4462                 connp->conn_ports = 0;
4463 
4464                 if (connp->conn_anon_port) {
4465                         zone_t          *zone;
4466 
4467                         zone = crgetzone(cr);
4468                         connp->conn_anon_port = B_FALSE;
4469                         (void) tsol_mlp_anon(zone, connp->conn_mlp_type,
4470                             connp->conn_proto, connp->conn_lport, B_FALSE);
4471                 }
4472                 connp->conn_mlp_type = mlptSingle;
4473 
4474                 tcp_bind_hash_remove(tcp);
4475                 return (error);
4476         } else {
4477                 /*
4478                  * If there is a connection limit, allocate and initialize
4479                  * the counter struct.  Note that since listen can be called
4480                  * multiple times, the struct may have been allready allocated.
4481                  */
4482                 if (!list_is_empty(&tcps->tcps_listener_conf) &&
4483                     tcp->tcp_listen_cnt == NULL) {
4484                         tcp_listen_cnt_t *tlc;
4485                         uint32_t ratio;
4486 
4487                         ratio = tcp_find_listener_conf(tcps,
4488                             ntohs(connp->conn_lport));
4489                         if (ratio != 0) {
4490                                 uint32_t mem_ratio, tot_buf;
4491 
4492                                 tlc = kmem_alloc(sizeof (tcp_listen_cnt_t),
4493                                     KM_SLEEP);
4494                                 /*
4495                                  * Calculate the connection limit based on
4496                                  * the configured ratio and maxusers.  Maxusers
4497                                  * are calculated based on memory size,
4498                                  * ~ 1 user per MB.  Note that the conn_rcvbuf
4499                                  * and conn_sndbuf may change after a
4500                                  * connection is accepted.  So what we have
4501                                  * is only an approximation.
4502                                  */
4503                                 if ((tot_buf = connp->conn_rcvbuf +
4504                                     connp->conn_sndbuf) < MB) {
4505                                         mem_ratio = MB / tot_buf;
4506                                         tlc->tlc_max = maxusers / ratio *
4507                                             mem_ratio;
4508                                 } else {
4509                                         mem_ratio = tot_buf / MB;
4510                                         tlc->tlc_max = maxusers / ratio /
4511                                             mem_ratio;
4512                                 }
4513                                 /* At least we should allow two connections! */
4514                                 if (tlc->tlc_max <= tcp_min_conn_listener)
4515                                         tlc->tlc_max = tcp_min_conn_listener;
4516                                 tlc->tlc_cnt = 1;
4517                                 tlc->tlc_drop = 0;
4518                                 tcp->tcp_listen_cnt = tlc;
4519                         }
4520                 }
4521         }
4522         return (error);
4523 }