lib/tsocket/tsocket_bsd.c

   1 /*
   2    Unix SMB/CIFS implementation.
   3
   4    Copyright (C) Stefan Metzmacher 2009
   5
   6      ** NOTE! The following LGPL license applies to the tsocket
   7      ** library. This does NOT imply that all of Samba is released
   8      ** under the LGPL
   9
  10    This library is free software; you can redistribute it and/or
  11    modify it under the terms of the GNU Lesser General Public
  12    License as published by the Free Software Foundation; either
  13    version 3 of the License, or (at your option) any later version.
  14
  15    This library is distributed in the hope that it will be useful,
  16    but WITHOUT ANY WARRANTY; without even the implied warranty of
  17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  18    Lesser General Public License for more details.
  19
  20    You should have received a copy of the GNU Lesser General Public
  21    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  22 */
  23
  24 #include "replace.h"
  25 #include "system/filesys.h"
  26 #include "system/network.h"
  27 #include "tsocket.h"
  28 #include "tsocket_internal.h"
  29
  30 static int tsocket_bsd_error_from_errno(int ret,
  31                                         int sys_errno,
  32                                         bool *retry)
  33 {
  34         *retry = false;
  35
  36         if (ret >= 0) {
  37                 return 0;
  38         }
  39
  40         if (ret != -1) {
  41                 return EIO;
  42         }
  43
  44         if (sys_errno == 0) {
  45                 return EIO;
  46         }
  47
  48         if (sys_errno == EINTR) {
  49                 *retry = true;
  50                 return sys_errno;
  51         }
  52
  53         if (sys_errno == EINPROGRESS) {
  54                 *retry = true;
  55                 return sys_errno;
  56         }
  57
  58         if (sys_errno == EAGAIN) {
  59                 *retry = true;
  60                 return sys_errno;
  61         }
  62
  63         /* ENOMEM is retryable on Solaris/illumos, and possibly other systems. */
  64         if (sys_errno == ENOMEM) {
  65                 *retry = true;
  66                 return sys_errno;
  67         }
  68
  69 #ifdef EWOULDBLOCK
  70         if (sys_errno == EWOULDBLOCK) {
  71                 *retry = true;
  72                 return sys_errno;
  73         }
  74 #endif
  75
  76         return sys_errno;
  77 }
  78
  79 static int tsocket_bsd_common_prepare_fd(int fd, bool high_fd)
  80 {
  81         int i;
  82         int sys_errno = 0;
  83         int fds[3];
  84         int num_fds = 0;
  85
  86         int result, flags;
  87
  88         if (fd == -1) {
  89                 return -1;
  90         }
  91
  92         /* first make a fd >= 3 */
  93         if (high_fd) {
  94                 while (fd < 3) {
  95                         fds[num_fds++] = fd;
  96                         fd = dup(fd);
  97                         if (fd == -1) {
  98                                 sys_errno = errno;
  99                                 break;
 100                         }
 101                 }
 102                 for (i=0; i<num_fds; i++) {
 103                         close(fds[i]);
 104                 }
 105                 if (fd == -1) {
 106                         errno = sys_errno;
 107                         return fd;
 108                 }
 109         }
 110
 111         /* fd should be nonblocking. */
 112
 113 #ifdef O_NONBLOCK
 114 #define FLAG_TO_SET O_NONBLOCK
 115 #else
 116 #ifdef SYSV
 117 #define FLAG_TO_SET O_NDELAY
 118 #else /* BSD */
 119 #define FLAG_TO_SET FNDELAY
 120 #endif
 121 #endif
 122
 123         if ((flags = fcntl(fd, F_GETFL)) == -1) {
 124                 goto fail;
 125         }
 126
 127         flags |= FLAG_TO_SET;
 128         if (fcntl(fd, F_SETFL, flags) == -1) {
 129                 goto fail;
 130         }
 131
 132 #undef FLAG_TO_SET
 133
 134         /* fd should be closed on exec() */
 135 #ifdef FD_CLOEXEC
 136         result = flags = fcntl(fd, F_GETFD, 0);
 137         if (flags >= 0) {
 138                 flags |= FD_CLOEXEC;
 139                 result = fcntl(fd, F_SETFD, flags);
 140         }
 141         if (result < 0) {
 142                 goto fail;
 143         }
 144 #endif
 145         return fd;
 146
 147  fail:
 148         if (fd != -1) {
 149                 sys_errno = errno;
 150                 close(fd);
 151                 errno = sys_errno;
 152         }
 153         return -1;
 154 }
 155
 156 static ssize_t tsocket_bsd_pending(int fd)
 157 {
 158         int ret, error;
 159         int value = 0;
 160         socklen_t len;
 161
 162         ret = ioctl(fd, FIONREAD, &value);
 163         if (ret == -1) {
 164                 return ret;
 165         }
 166
 167         if (ret != 0) {
 168                 /* this should not be reached */
 169                 errno = EIO;
 170                 return -1;
 171         }
 172
 173         if (value != 0) {
 174                 return value;
 175         }
 176
 177         error = 0;
 178         len = sizeof(error);
 179
 180         /*
 181          * if no data is available check if the socket is in error state. For
 182          * dgram sockets it's the way to return ICMP error messages of
 183          * connected sockets to the caller.
 184          */
 185         ret = getsockopt(fd, SOL_SOCKET, SO_ERROR, &error, &len);
 186         if (ret == -1) {
 187                 return ret;
 188         }
 189         if (error != 0) {
 190                 errno = error;
 191                 return -1;
 192         }
 193         return 0;
 194 }
 195
 196 static const struct tsocket_address_ops tsocket_address_bsd_ops;
 197
 198 struct tsocket_address_bsd {
 199         socklen_t sa_socklen;
 200         union {
 201                 struct sockaddr sa;
 202                 struct sockaddr_in in;
 203 #ifdef HAVE_IPV6
 204                 struct sockaddr_in6 in6;
 205 #endif
 206                 struct sockaddr_un un;
 207                 struct sockaddr_storage ss;
 208         } u;
 209 };
 210
 211 int _tsocket_address_bsd_from_sockaddr(TALLOC_CTX *mem_ctx,
 212                                        const struct sockaddr *sa,
 213                                        size_t sa_socklen,
 214                                        struct tsocket_address **_addr,
 215                                        const char *location)
 216 {
 217         struct tsocket_address *addr;
 218         struct tsocket_address_bsd *bsda;
 219
 220         if (sa_socklen < sizeof(sa->sa_family)) {
 221                 errno = EINVAL;
 222                 return -1;
 223         }
 224
 225         switch (sa->sa_family) {
 226         case AF_UNIX:
 227                 if (sa_socklen > sizeof(struct sockaddr_un)) {
 228                         sa_socklen = sizeof(struct sockaddr_un);
 229                 }
 230                 break;
 231         case AF_INET:
 232                 if (sa_socklen < sizeof(struct sockaddr_in)) {
 233                         errno = EINVAL;
 234                         return -1;
 235                 }
 236                 sa_socklen = sizeof(struct sockaddr_in);
 237                 break;
 238 #ifdef HAVE_IPV6
 239         case AF_INET6:
 240                 if (sa_socklen < sizeof(struct sockaddr_in6)) {
 241                         errno = EINVAL;
 242                         return -1;
 243                 }
 244                 sa_socklen = sizeof(struct sockaddr_in6);
 245                 break;
 246 #endif
 247         default:
 248                 errno = EAFNOSUPPORT;
 249                 return -1;
 250         }
 251
 252         if (sa_socklen > sizeof(struct sockaddr_storage)) {
 253                 errno = EINVAL;
 254                 return -1;
 255         }
 256
 257         addr = tsocket_address_create(mem_ctx,
 258                                       &tsocket_address_bsd_ops,
 259                                       &bsda,
 260                                       struct tsocket_address_bsd,
 261                                       location);
 262         if (!addr) {
 263                 errno = ENOMEM;
 264                 return -1;
 265         }
 266
 267         ZERO_STRUCTP(bsda);
 268
 269         memcpy(&bsda->u.ss, sa, sa_socklen);
 270
 271         bsda->sa_socklen = sa_socklen;
 272 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 273         bsda->u.sa.sa_len = bsda->sa_socklen;
 274 #endif
 275
 276         *_addr = addr;
 277         return 0;
 278 }
 279
 280 ssize_t tsocket_address_bsd_sockaddr(const struct tsocket_address *addr,
 281                                      struct sockaddr *sa,
 282                                      size_t sa_socklen)
 283 {
 284         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 285                                            struct tsocket_address_bsd);
 286
 287         if (!bsda) {
 288                 errno = EINVAL;
 289                 return -1;
 290         }
 291
 292         if (sa_socklen < bsda->sa_socklen) {
 293                 errno = EINVAL;
 294                 return -1;
 295         }
 296
 297         if (sa_socklen > bsda->sa_socklen) {
 298                 memset(sa, 0, sa_socklen);
 299                 sa_socklen = bsda->sa_socklen;
 300         }
 301
 302         memcpy(sa, &bsda->u.ss, sa_socklen);
 303 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 304         sa->sa_len = sa_socklen;
 305 #endif
 306         return sa_socklen;
 307 }
 308
 309 bool tsocket_address_is_inet(const struct tsocket_address *addr, const char *fam)
 310 {
 311         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 312                                            struct tsocket_address_bsd);
 313
 314         if (!bsda) {
 315                 return false;
 316         }
 317
 318         switch (bsda->u.sa.sa_family) {
 319         case AF_INET:
 320                 if (strcasecmp(fam, "ip") == 0) {
 321                         return true;
 322                 }
 323
 324                 if (strcasecmp(fam, "ipv4") == 0) {
 325                         return true;
 326                 }
 327
 328                 return false;
 329 #ifdef HAVE_IPV6
 330         case AF_INET6:
 331                 if (strcasecmp(fam, "ip") == 0) {
 332                         return true;
 333                 }
 334
 335                 if (strcasecmp(fam, "ipv6") == 0) {
 336                         return true;
 337                 }
 338
 339                 return false;
 340 #endif
 341         }
 342
 343         return false;
 344 }
 345
 346 int _tsocket_address_inet_from_strings(TALLOC_CTX *mem_ctx,
 347                                        const char *fam,
 348                                        const char *addr,
 349                                        uint16_t port,
 350                                        struct tsocket_address **_addr,
 351                                        const char *location)
 352 {
 353         struct addrinfo hints;
 354         struct addrinfo *result = NULL;
 355         char port_str[6];
 356         int ret;
 357
 358         ZERO_STRUCT(hints);
 359         /*
 360          * we use SOCKET_STREAM here to get just one result
 361          * back from getaddrinfo().
 362          */
 363         hints.ai_socktype = SOCK_STREAM;
 364         hints.ai_flags = AI_NUMERICHOST | AI_NUMERICSERV;
 365
 366         if (strcasecmp(fam, "ip") == 0) {
 367                 hints.ai_family = AF_UNSPEC;
 368                 if (!addr) {
 369 #ifdef HAVE_IPV6
 370                         addr = "::";
 371 #else
 372                         addr = "0.0.0.0";
 373 #endif
 374                 }
 375         } else if (strcasecmp(fam, "ipv4") == 0) {
 376                 hints.ai_family = AF_INET;
 377                 if (!addr) {
 378                         addr = "0.0.0.0";
 379                 }
 380 #ifdef HAVE_IPV6
 381         } else if (strcasecmp(fam, "ipv6") == 0) {
 382                 hints.ai_family = AF_INET6;
 383                 if (!addr) {
 384                         addr = "::";
 385                 }
 386 #endif
 387         } else {
 388                 errno = EAFNOSUPPORT;
 389                 return -1;
 390         }
 391
 392         snprintf(port_str, sizeof(port_str), "%u", port);
 393
 394         ret = getaddrinfo(addr, port_str, &hints, &result);
 395         if (ret != 0) {
 396                 switch (ret) {
 397                 case EAI_FAIL:
 398                         errno = EINVAL;
 399                         break;
 400                 }
 401                 ret = -1;
 402                 goto done;
 403         }
 404
 405         if (result->ai_socktype != SOCK_STREAM) {
 406                 errno = EINVAL;
 407                 ret = -1;
 408                 goto done;
 409         }
 410
 411         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 412                                                   result->ai_addr,
 413                                                   result->ai_addrlen,
 414                                                   _addr,
 415                                                   location);
 416
 417 done:
 418         if (result) {
 419                 freeaddrinfo(result);
 420         }
 421         return ret;
 422 }
 423
 424 char *tsocket_address_inet_addr_string(const struct tsocket_address *addr,
 425                                        TALLOC_CTX *mem_ctx)
 426 {
 427         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 428                                            struct tsocket_address_bsd);
 429         char addr_str[INET6_ADDRSTRLEN+1];
 430         const char *str;
 431
 432         if (!bsda) {
 433                 errno = EINVAL;
 434                 return NULL;
 435         }
 436
 437         switch (bsda->u.sa.sa_family) {
 438         case AF_INET:
 439                 str = inet_ntop(bsda->u.in.sin_family,
 440                                 &bsda->u.in.sin_addr,
 441                                 addr_str, sizeof(addr_str));
 442                 break;
 443 #ifdef HAVE_IPV6
 444         case AF_INET6:
 445                 str = inet_ntop(bsda->u.in6.sin6_family,
 446                                 &bsda->u.in6.sin6_addr,
 447                                 addr_str, sizeof(addr_str));
 448                 break;
 449 #endif
 450         default:
 451                 errno = EINVAL;
 452                 return NULL;
 453         }
 454
 455         if (!str) {
 456                 return NULL;
 457         }
 458
 459         return talloc_strdup(mem_ctx, str);
 460 }
 461
 462 uint16_t tsocket_address_inet_port(const struct tsocket_address *addr)
 463 {
 464         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 465                                            struct tsocket_address_bsd);
 466         uint16_t port = 0;
 467
 468         if (!bsda) {
 469                 errno = EINVAL;
 470                 return 0;
 471         }
 472
 473         switch (bsda->u.sa.sa_family) {
 474         case AF_INET:
 475                 port = ntohs(bsda->u.in.sin_port);
 476                 break;
 477 #ifdef HAVE_IPV6
 478         case AF_INET6:
 479                 port = ntohs(bsda->u.in6.sin6_port);
 480                 break;
 481 #endif
 482         default:
 483                 errno = EINVAL;
 484                 return 0;
 485         }
 486
 487         return port;
 488 }
 489
 490 int tsocket_address_inet_set_port(struct tsocket_address *addr,
 491                                   uint16_t port)
 492 {
 493         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 494                                            struct tsocket_address_bsd);
 495
 496         if (!bsda) {
 497                 errno = EINVAL;
 498                 return -1;
 499         }
 500
 501         switch (bsda->u.sa.sa_family) {
 502         case AF_INET:
 503                 bsda->u.in.sin_port = htons(port);
 504                 break;
 505 #ifdef HAVE_IPV6
 506         case AF_INET6:
 507                 bsda->u.in6.sin6_port = htons(port);
 508                 break;
 509 #endif
 510         default:
 511                 errno = EINVAL;
 512                 return -1;
 513         }
 514
 515         return 0;
 516 }
 517
 518 bool tsocket_address_is_unix(const struct tsocket_address *addr)
 519 {
 520         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 521                                            struct tsocket_address_bsd);
 522
 523         if (!bsda) {
 524                 return false;
 525         }
 526
 527         switch (bsda->u.sa.sa_family) {
 528         case AF_UNIX:
 529                 return true;
 530         }
 531
 532         return false;
 533 }
 534
 535 int _tsocket_address_unix_from_path(TALLOC_CTX *mem_ctx,
 536                                     const char *path,
 537                                     struct tsocket_address **_addr,
 538                                     const char *location)
 539 {
 540         struct sockaddr_un un;
 541         void *p = &un;
 542         int ret;
 543
 544         if (!path) {
 545                 path = "";
 546         }
 547
 548         if (strlen(path) > sizeof(un.sun_path)-1) {
 549                 errno = ENAMETOOLONG;
 550                 return -1;
 551         }
 552
 553         ZERO_STRUCT(un);
 554         un.sun_family = AF_UNIX;
 555         strncpy(un.sun_path, path, sizeof(un.sun_path)-1);
 556
 557         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 558                                                  (struct sockaddr *)p,
 559                                                  sizeof(un),
 560                                                  _addr,
 561                                                  location);
 562
 563         return ret;
 564 }
 565
 566 char *tsocket_address_unix_path(const struct tsocket_address *addr,
 567                                 TALLOC_CTX *mem_ctx)
 568 {
 569         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 570                                            struct tsocket_address_bsd);
 571         const char *str;
 572
 573         if (!bsda) {
 574                 errno = EINVAL;
 575                 return NULL;
 576         }
 577
 578         switch (bsda->u.sa.sa_family) {
 579         case AF_UNIX:
 580                 str = bsda->u.un.sun_path;
 581                 break;
 582         default:
 583                 errno = EINVAL;
 584                 return NULL;
 585         }
 586
 587         return talloc_strdup(mem_ctx, str);
 588 }
 589
 590 static char *tsocket_address_bsd_string(const struct tsocket_address *addr,
 591                                         TALLOC_CTX *mem_ctx)
 592 {
 593         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 594                                            struct tsocket_address_bsd);
 595         char *str;
 596         char *addr_str;
 597         const char *prefix = NULL;
 598         uint16_t port;
 599
 600         switch (bsda->u.sa.sa_family) {
 601         case AF_UNIX:
 602                 return talloc_asprintf(mem_ctx, "unix:%s",
 603                                        bsda->u.un.sun_path);
 604         case AF_INET:
 605                 prefix = "ipv4";
 606                 break;
 607 #ifdef HAVE_IPV6
 608         case AF_INET6:
 609                 prefix = "ipv6";
 610                 break;
 611 #endif
 612         default:
 613                 errno = EINVAL;
 614                 return NULL;
 615         }
 616
 617         addr_str = tsocket_address_inet_addr_string(addr, mem_ctx);
 618         if (!addr_str) {
 619                 return NULL;
 620         }
 621
 622         port = tsocket_address_inet_port(addr);
 623
 624         str = talloc_asprintf(mem_ctx, "%s:%s:%u",
 625                               prefix, addr_str, port);
 626         talloc_free(addr_str);
 627
 628         return str;
 629 }
 630
 631 static struct tsocket_address *tsocket_address_bsd_copy(const struct tsocket_address *addr,
 632                                                          TALLOC_CTX *mem_ctx,
 633                                                          const char *location)
 634 {
 635         struct tsocket_address_bsd *bsda = talloc_get_type(addr->private_data,
 636                                            struct tsocket_address_bsd);
 637         struct tsocket_address *copy;
 638         int ret;
 639
 640         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 641                                                  &bsda->u.sa,
 642                                                  bsda->sa_socklen,
 643                                                  &copy,
 644                                                  location);
 645         if (ret != 0) {
 646                 return NULL;
 647         }
 648
 649         return copy;
 650 }
 651
 652 static const struct tsocket_address_ops tsocket_address_bsd_ops = {
 653         .name           = "bsd",
 654         .string         = tsocket_address_bsd_string,
 655         .copy           = tsocket_address_bsd_copy,
 656 };
 657
 658 struct tdgram_bsd {
 659         int fd;
 660
 661         void *event_ptr;
 662         struct tevent_fd *fde;
 663         bool optimize_recvfrom;
 664
 665         void *readable_private;
 666         void (*readable_handler)(void *private_data);
 667         void *writeable_private;
 668         void (*writeable_handler)(void *private_data);
 669 };
 670
 671 bool tdgram_bsd_optimize_recvfrom(struct tdgram_context *dgram,
 672                                   bool on)
 673 {
 674         struct tdgram_bsd *bsds =
 675                 talloc_get_type(_tdgram_context_data(dgram),
 676                 struct tdgram_bsd);
 677         bool old;
 678
 679         if (bsds == NULL) {
 680                 /* not a bsd socket */
 681                 return false;
 682         }
 683
 684         old = bsds->optimize_recvfrom;
 685         bsds->optimize_recvfrom = on;
 686
 687         return old;
 688 }
 689
 690 static void tdgram_bsd_fde_handler(struct tevent_context *ev,
 691                                    struct tevent_fd *fde,
 692                                    uint16_t flags,
 693                                    void *private_data)
 694 {
 695         struct tdgram_bsd *bsds = talloc_get_type_abort(private_data,
 696                                   struct tdgram_bsd);
 697
 698         if (flags & TEVENT_FD_WRITE) {
 699                 bsds->writeable_handler(bsds->writeable_private);
 700                 return;
 701         }
 702         if (flags & TEVENT_FD_READ) {
 703                 if (!bsds->readable_handler) {
 704                         TEVENT_FD_NOT_READABLE(bsds->fde);
 705                         return;
 706                 }
 707                 bsds->readable_handler(bsds->readable_private);
 708                 return;
 709         }
 710 }
 711
 712 static int tdgram_bsd_set_readable_handler(struct tdgram_bsd *bsds,
 713                                            struct tevent_context *ev,
 714                                            void (*handler)(void *private_data),
 715                                            void *private_data)
 716 {
 717         if (ev == NULL) {
 718                 if (handler) {
 719                         errno = EINVAL;
 720                         return -1;
 721                 }
 722                 if (!bsds->readable_handler) {
 723                         return 0;
 724                 }
 725                 bsds->readable_handler = NULL;
 726                 bsds->readable_private = NULL;
 727
 728                 return 0;
 729         }
 730
 731         /* read and write must use the same tevent_context */
 732         if (bsds->event_ptr != ev) {
 733                 if (bsds->readable_handler || bsds->writeable_handler) {
 734                         errno = EINVAL;
 735                         return -1;
 736                 }
 737                 bsds->event_ptr = NULL;
 738                 TALLOC_FREE(bsds->fde);
 739         }
 740
 741         if (tevent_fd_get_flags(bsds->fde) == 0) {
 742                 TALLOC_FREE(bsds->fde);
 743
 744                 bsds->fde = tevent_add_fd(ev, bsds,
 745                                           bsds->fd, TEVENT_FD_READ,
 746                                           tdgram_bsd_fde_handler,
 747                                           bsds);
 748                 if (!bsds->fde) {
 749                         errno = ENOMEM;
 750                         return -1;
 751                 }
 752
 753                 /* cache the event context we're running on */
 754                 bsds->event_ptr = ev;
 755         } else if (!bsds->readable_handler) {
 756                 TEVENT_FD_READABLE(bsds->fde);
 757         }
 758
 759         bsds->readable_handler = handler;
 760         bsds->readable_private = private_data;
 761
 762         return 0;
 763 }
 764
 765 static int tdgram_bsd_set_writeable_handler(struct tdgram_bsd *bsds,
 766                                             struct tevent_context *ev,
 767                                             void (*handler)(void *private_data),
 768                                             void *private_data)
 769 {
 770         if (ev == NULL) {
 771                 if (handler) {
 772                         errno = EINVAL;
 773                         return -1;
 774                 }
 775                 if (!bsds->writeable_handler) {
 776                         return 0;
 777                 }
 778                 bsds->writeable_handler = NULL;
 779                 bsds->writeable_private = NULL;
 780                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
 781
 782                 return 0;
 783         }
 784
 785         /* read and write must use the same tevent_context */
 786         if (bsds->event_ptr != ev) {
 787                 if (bsds->readable_handler || bsds->writeable_handler) {
 788                         errno = EINVAL;
 789                         return -1;
 790                 }
 791                 bsds->event_ptr = NULL;
 792                 TALLOC_FREE(bsds->fde);
 793         }
 794
 795         if (tevent_fd_get_flags(bsds->fde) == 0) {
 796                 TALLOC_FREE(bsds->fde);
 797
 798                 bsds->fde = tevent_add_fd(ev, bsds,
 799                                           bsds->fd, TEVENT_FD_WRITE,
 800                                           tdgram_bsd_fde_handler,
 801                                           bsds);
 802                 if (!bsds->fde) {
 803                         errno = ENOMEM;
 804                         return -1;
 805                 }
 806
 807                 /* cache the event context we're running on */
 808                 bsds->event_ptr = ev;
 809         } else if (!bsds->writeable_handler) {
 810                 TEVENT_FD_WRITEABLE(bsds->fde);
 811         }
 812
 813         bsds->writeable_handler = handler;
 814         bsds->writeable_private = private_data;
 815
 816         return 0;
 817 }
 818
 819 struct tdgram_bsd_recvfrom_state {
 820         struct tdgram_context *dgram;
 821         bool first_try;
 822         uint8_t *buf;
 823         size_t len;
 824         struct tsocket_address *src;
 825 };
 826
 827 static int tdgram_bsd_recvfrom_destructor(struct tdgram_bsd_recvfrom_state *state)
 828 {
 829         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
 830                                   struct tdgram_bsd);
 831
 832         tdgram_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
 833
 834         return 0;
 835 }
 836
 837 static void tdgram_bsd_recvfrom_handler(void *private_data);
 838
 839 static struct tevent_req *tdgram_bsd_recvfrom_send(TALLOC_CTX *mem_ctx,
 840                                         struct tevent_context *ev,
 841                                         struct tdgram_context *dgram)
 842 {
 843         struct tevent_req *req;
 844         struct tdgram_bsd_recvfrom_state *state;
 845         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 846         int ret;
 847
 848         req = tevent_req_create(mem_ctx, &state,
 849                                 struct tdgram_bsd_recvfrom_state);
 850         if (!req) {
 851                 return NULL;
 852         }
 853
 854         state->dgram    = dgram;
 855         state->first_try= true;
 856         state->buf      = NULL;
 857         state->len      = 0;
 858         state->src      = NULL;
 859
 860         talloc_set_destructor(state, tdgram_bsd_recvfrom_destructor);
 861
 862         if (bsds->fd == -1) {
 863                 tevent_req_error(req, ENOTCONN);
 864                 goto post;
 865         }
 866
 867
 868         /*
 869          * this is a fast path, not waiting for the
 870          * socket to become explicit readable gains
 871          * about 10%-20% performance in benchmark tests.
 872          */
 873         if (bsds->optimize_recvfrom) {
 874                 /*
 875                  * We only do the optimization on
 876                  * recvfrom if the caller asked for it.
 877                  *
 878                  * This is needed because in most cases
 879                  * we preferr to flush send buffers before
 880                  * receiving incoming requests.
 881                  */
 882                 tdgram_bsd_recvfrom_handler(req);
 883                 if (!tevent_req_is_in_progress(req)) {
 884                         goto post;
 885                 }
 886         }
 887
 888         ret = tdgram_bsd_set_readable_handler(bsds, ev,
 889                                               tdgram_bsd_recvfrom_handler,
 890                                               req);
 891         if (ret == -1) {
 892                 tevent_req_error(req, errno);
 893                 goto post;
 894         }
 895
 896         return req;
 897
 898  post:
 899         tevent_req_post(req, ev);
 900         return req;
 901 }
 902
 903 static void tdgram_bsd_recvfrom_handler(void *private_data)
 904 {
 905         struct tevent_req *req = talloc_get_type_abort(private_data,
 906                                  struct tevent_req);
 907         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
 908                                         struct tdgram_bsd_recvfrom_state);
 909         struct tdgram_context *dgram = state->dgram;
 910         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 911         struct tsocket_address_bsd *bsda;
 912         ssize_t ret;
 913         int err;
 914         bool retry;
 915
 916         ret = tsocket_bsd_pending(bsds->fd);
 917         if (state->first_try && ret == 0) {
 918                 state->first_try = false;
 919                 /* retry later */
 920                 return;
 921         }
 922         state->first_try = false;
 923
 924         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 925         if (retry) {
 926                 /* retry later */
 927                 return;
 928         }
 929         if (tevent_req_error(req, err)) {
 930                 return;
 931         }
 932
 933         /* note that 'ret' can be 0 here */
 934         state->buf = talloc_array(state, uint8_t, ret);
 935         if (tevent_req_nomem(state->buf, req)) {
 936                 return;
 937         }
 938         state->len = ret;
 939
 940         state->src = tsocket_address_create(state,
 941                                             &tsocket_address_bsd_ops,
 942                                             &bsda,
 943                                             struct tsocket_address_bsd,
 944                                             __location__ "bsd_recvfrom");
 945         if (tevent_req_nomem(state->src, req)) {
 946                 return;
 947         }
 948
 949         ZERO_STRUCTP(bsda);
 950         bsda->sa_socklen = sizeof(bsda->u.ss);
 951 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 952         bsda->u.sa.sa_len = bsda->sa_socklen;
 953 #endif
 954
 955         ret = recvfrom(bsds->fd, state->buf, state->len, 0,
 956                        &bsda->u.sa, &bsda->sa_socklen);
 957         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 958         if (retry) {
 959                 /* retry later */
 960                 return;
 961         }
 962         if (tevent_req_error(req, err)) {
 963                 return;
 964         }
 965
 966         /*
 967          * Some systems (FreeBSD, see bug #7115) return too much
 968          * bytes in tsocket_bsd_pending()/ioctl(fd, FIONREAD, ...),
 969          * the return value includes some IP/UDP header bytes,
 970          * while recvfrom() just returns the payload.
 971          */
 972         state->buf = talloc_realloc(state, state->buf, uint8_t, ret);
 973         if (tevent_req_nomem(state->buf, req)) {
 974                 return;
 975         }
 976         state->len = ret;
 977
 978         tevent_req_done(req);
 979 }
 980
 981 static ssize_t tdgram_bsd_recvfrom_recv(struct tevent_req *req,
 982                                         int *perrno,
 983                                         TALLOC_CTX *mem_ctx,
 984                                         uint8_t **buf,
 985                                         struct tsocket_address **src)
 986 {
 987         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
 988                                         struct tdgram_bsd_recvfrom_state);
 989         ssize_t ret;
 990
 991         ret = tsocket_simple_int_recv(req, perrno);
 992         if (ret == 0) {
 993                 *buf = talloc_move(mem_ctx, &state->buf);
 994                 ret = state->len;
 995                 if (src) {
 996                         *src = talloc_move(mem_ctx, &state->src);
 997                 }
 998         }
 999
1000         tevent_req_received(req);
1001         return ret;
1002 }
1003
1004 struct tdgram_bsd_sendto_state {
1005         struct tdgram_context *dgram;
1006
1007         const uint8_t *buf;
1008         size_t len;
1009         const struct tsocket_address *dst;
1010
1011         ssize_t ret;
1012 };
1013
1014 static int tdgram_bsd_sendto_destructor(struct tdgram_bsd_sendto_state *state)
1015 {
1016         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
1017                                   struct tdgram_bsd);
1018
1019         tdgram_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
1020
1021         return 0;
1022 }
1023
1024 static void tdgram_bsd_sendto_handler(void *private_data);
1025
1026 static struct tevent_req *tdgram_bsd_sendto_send(TALLOC_CTX *mem_ctx,
1027                                                  struct tevent_context *ev,
1028                                                  struct tdgram_context *dgram,
1029                                                  const uint8_t *buf,
1030                                                  size_t len,
1031                                                  const struct tsocket_address *dst)
1032 {
1033         struct tevent_req *req;
1034         struct tdgram_bsd_sendto_state *state;
1035         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1036         int ret;
1037
1038         req = tevent_req_create(mem_ctx, &state,
1039                                 struct tdgram_bsd_sendto_state);
1040         if (!req) {
1041                 return NULL;
1042         }
1043
1044         state->dgram    = dgram;
1045         state->buf      = buf;
1046         state->len      = len;
1047         state->dst      = dst;
1048         state->ret      = -1;
1049
1050         talloc_set_destructor(state, tdgram_bsd_sendto_destructor);
1051
1052         if (bsds->fd == -1) {
1053                 tevent_req_error(req, ENOTCONN);
1054                 goto post;
1055         }
1056
1057         /*
1058          * this is a fast path, not waiting for the
1059          * socket to become explicit writeable gains
1060          * about 10%-20% performance in benchmark tests.
1061          */
1062         tdgram_bsd_sendto_handler(req);
1063         if (!tevent_req_is_in_progress(req)) {
1064                 goto post;
1065         }
1066
1067         ret = tdgram_bsd_set_writeable_handler(bsds, ev,
1068                                                tdgram_bsd_sendto_handler,
1069                                                req);
1070         if (ret == -1) {
1071                 tevent_req_error(req, errno);
1072                 goto post;
1073         }
1074
1075         return req;
1076
1077  post:
1078         tevent_req_post(req, ev);
1079         return req;
1080 }
1081
1082 static void tdgram_bsd_sendto_handler(void *private_data)
1083 {
1084         struct tevent_req *req = talloc_get_type_abort(private_data,
1085                                  struct tevent_req);
1086         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
1087                                         struct tdgram_bsd_sendto_state);
1088         struct tdgram_context *dgram = state->dgram;
1089         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1090         struct sockaddr *sa = NULL;
1091         socklen_t sa_socklen = 0;
1092         ssize_t ret;
1093         int err;
1094         bool retry;
1095
1096         if (state->dst) {
1097                 struct tsocket_address_bsd *bsda =
1098                         talloc_get_type(state->dst->private_data,
1099                         struct tsocket_address_bsd);
1100
1101                 sa = &bsda->u.sa;
1102                 sa_socklen = bsda->sa_socklen;
1103         }
1104
1105         ret = sendto(bsds->fd, state->buf, state->len, 0, sa, sa_socklen);
1106         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1107         if (retry) {
1108                 /* retry later */
1109                 return;
1110         }
1111
1112         if (err == EMSGSIZE) {
1113                 /* round up in 1K increments */
1114                 int bufsize = ((state->len + 1023) & (~1023));
1115
1116                 ret = setsockopt(bsds->fd, SOL_SOCKET, SO_SNDBUF, &bufsize,
1117                                  sizeof(bufsize));
1118                 if (ret == 0) {
1119                         /*
1120                          * We do the rety here, rather then via the
1121                          * handler, as we only want to retry once for
1122                          * this condition, so if there is a mismatch
1123                          * between what setsockopt() accepts and what can
1124                          * actually be sent, we do not end up in a
1125                          * loop.
1126                          */
1127
1128                         ret = sendto(bsds->fd, state->buf, state->len,
1129                                      0, sa, sa_socklen);
1130                         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1131                         if (retry) { /* retry later */
1132                                 return;
1133                         }
1134                 }
1135         }
1136
1137         if (tevent_req_error(req, err)) {
1138                 return;
1139         }
1140
1141         state->ret = ret;
1142
1143         tevent_req_done(req);
1144 }
1145
1146 static ssize_t tdgram_bsd_sendto_recv(struct tevent_req *req, int *perrno)
1147 {
1148         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
1149                                         struct tdgram_bsd_sendto_state);
1150         ssize_t ret;
1151
1152         ret = tsocket_simple_int_recv(req, perrno);
1153         if (ret == 0) {
1154                 ret = state->ret;
1155         }
1156
1157         tevent_req_received(req);
1158         return ret;
1159 }
1160
1161 struct tdgram_bsd_disconnect_state {
1162         uint8_t __dummy;
1163 };
1164
1165 static struct tevent_req *tdgram_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
1166                                                      struct tevent_context *ev,
1167                                                      struct tdgram_context *dgram)
1168 {
1169         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1170         struct tevent_req *req;
1171         struct tdgram_bsd_disconnect_state *state;
1172         int ret;
1173         int err;
1174         bool dummy;
1175
1176         req = tevent_req_create(mem_ctx, &state,
1177                                 struct tdgram_bsd_disconnect_state);
1178         if (req == NULL) {
1179                 return NULL;
1180         }
1181
1182         if (bsds->fd == -1) {
1183                 tevent_req_error(req, ENOTCONN);
1184                 goto post;
1185         }
1186
1187         TALLOC_FREE(bsds->fde);
1188         ret = close(bsds->fd);
1189         bsds->fd = -1;
1190         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
1191         if (tevent_req_error(req, err)) {
1192                 goto post;
1193         }
1194
1195         tevent_req_done(req);
1196 post:
1197         tevent_req_post(req, ev);
1198         return req;
1199 }
1200
1201 static int tdgram_bsd_disconnect_recv(struct tevent_req *req,
1202                                       int *perrno)
1203 {
1204         int ret;
1205
1206         ret = tsocket_simple_int_recv(req, perrno);
1207
1208         tevent_req_received(req);
1209         return ret;
1210 }
1211
1212 static const struct tdgram_context_ops tdgram_bsd_ops = {
1213         .name                   = "bsd",
1214
1215         .recvfrom_send          = tdgram_bsd_recvfrom_send,
1216         .recvfrom_recv          = tdgram_bsd_recvfrom_recv,
1217
1218         .sendto_send            = tdgram_bsd_sendto_send,
1219         .sendto_recv            = tdgram_bsd_sendto_recv,
1220
1221         .disconnect_send        = tdgram_bsd_disconnect_send,
1222         .disconnect_recv        = tdgram_bsd_disconnect_recv,
1223 };
1224
1225 static int tdgram_bsd_destructor(struct tdgram_bsd *bsds)
1226 {
1227         TALLOC_FREE(bsds->fde);
1228         if (bsds->fd != -1) {
1229                 close(bsds->fd);
1230                 bsds->fd = -1;
1231         }
1232         return 0;
1233 }
1234
1235 static int tdgram_bsd_dgram_socket(const struct tsocket_address *local,
1236                                    const struct tsocket_address *remote,
1237                                    bool broadcast,
1238                                    TALLOC_CTX *mem_ctx,
1239                                    struct tdgram_context **_dgram,
1240                                    const char *location)
1241 {
1242         struct tsocket_address_bsd *lbsda =
1243                 talloc_get_type_abort(local->private_data,
1244                 struct tsocket_address_bsd);
1245         struct tsocket_address_bsd *rbsda = NULL;
1246         struct tdgram_context *dgram;
1247         struct tdgram_bsd *bsds;
1248         int fd;
1249         int ret;
1250         bool do_bind = false;
1251         bool do_reuseaddr = false;
1252         bool do_ipv6only = false;
1253         bool is_inet = false;
1254         int sa_fam = lbsda->u.sa.sa_family;
1255
1256         if (remote) {
1257                 rbsda = talloc_get_type_abort(remote->private_data,
1258                         struct tsocket_address_bsd);
1259         }
1260
1261         switch (lbsda->u.sa.sa_family) {
1262         case AF_UNIX:
1263                 if (broadcast) {
1264                         errno = EINVAL;
1265                         return -1;
1266                 }
1267                 if (lbsda->u.un.sun_path[0] != 0) {
1268                         do_reuseaddr = true;
1269                         do_bind = true;
1270                 }
1271                 break;
1272         case AF_INET:
1273                 if (lbsda->u.in.sin_port != 0) {
1274                         do_reuseaddr = true;
1275                         do_bind = true;
1276                 }
1277                 if (lbsda->u.in.sin_addr.s_addr != INADDR_ANY) {
1278                         do_bind = true;
1279                 }
1280                 is_inet = true;
1281                 break;
1282 #ifdef HAVE_IPV6
1283         case AF_INET6:
1284                 if (lbsda->u.in6.sin6_port != 0) {
1285                         do_reuseaddr = true;
1286                         do_bind = true;
1287                 }
1288                 if (memcmp(&in6addr_any,
1289                            &lbsda->u.in6.sin6_addr,
1290                            sizeof(in6addr_any)) != 0) {
1291                         do_bind = true;
1292                 }
1293                 is_inet = true;
1294                 do_ipv6only = true;
1295                 break;
1296 #endif
1297         default:
1298                 errno = EINVAL;
1299                 return -1;
1300         }
1301
1302         if (!do_bind && is_inet && rbsda) {
1303                 sa_fam = rbsda->u.sa.sa_family;
1304                 switch (sa_fam) {
1305                 case AF_INET:
1306                         do_ipv6only = false;
1307                         break;
1308 #ifdef HAVE_IPV6
1309                 case AF_INET6:
1310                         do_ipv6only = true;
1311                         break;
1312 #endif
1313                 }
1314         }
1315
1316         fd = socket(sa_fam, SOCK_DGRAM, 0);
1317         if (fd < 0) {
1318                 return -1;
1319         }
1320
1321         fd = tsocket_bsd_common_prepare_fd(fd, true);
1322         if (fd < 0) {
1323                 return -1;
1324         }
1325
1326         dgram = tdgram_context_create(mem_ctx,
1327                                       &tdgram_bsd_ops,
1328                                       &bsds,
1329                                       struct tdgram_bsd,
1330                                       location);
1331         if (!dgram) {
1332                 int saved_errno = errno;
1333                 close(fd);
1334                 errno = saved_errno;
1335                 return -1;
1336         }
1337         ZERO_STRUCTP(bsds);
1338         bsds->fd = fd;
1339         talloc_set_destructor(bsds, tdgram_bsd_destructor);
1340
1341 #ifdef HAVE_IPV6
1342         if (do_ipv6only) {
1343                 int val = 1;
1344
1345                 ret = setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY,
1346                                  (const void *)&val, sizeof(val));
1347                 if (ret == -1) {
1348                         int saved_errno = errno;
1349                         talloc_free(dgram);
1350                         errno = saved_errno;
1351                         return -1;
1352                 }
1353         }
1354 #endif
1355
1356         if (broadcast) {
1357                 int val = 1;
1358
1359                 ret = setsockopt(fd, SOL_SOCKET, SO_BROADCAST,
1360                                  (const void *)&val, sizeof(val));
1361                 if (ret == -1) {
1362                         int saved_errno = errno;
1363                         talloc_free(dgram);
1364                         errno = saved_errno;
1365                         return -1;
1366                 }
1367         }
1368
1369         if (do_reuseaddr) {
1370                 int val = 1;
1371
1372                 ret = setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
1373                                  (const void *)&val, sizeof(val));
1374                 if (ret == -1) {
1375                         int saved_errno = errno;
1376                         talloc_free(dgram);
1377                         errno = saved_errno;
1378                         return -1;
1379                 }
1380         }
1381
1382         if (do_bind) {
1383                 ret = bind(fd, &lbsda->u.sa, lbsda->sa_socklen);
1384                 if (ret == -1) {
1385                         int saved_errno = errno;
1386                         talloc_free(dgram);
1387                         errno = saved_errno;
1388                         return -1;
1389                 }
1390         }
1391
1392         if (rbsda) {
1393                 if (rbsda->u.sa.sa_family != sa_fam) {
1394                         talloc_free(dgram);
1395                         errno = EINVAL;
1396                         return -1;
1397                 }
1398
1399                 ret = connect(fd, &rbsda->u.sa, rbsda->sa_socklen);
1400                 if (ret == -1) {
1401                         int saved_errno = errno;
1402                         talloc_free(dgram);
1403                         errno = saved_errno;
1404                         return -1;
1405                 }
1406         }
1407
1408         *_dgram = dgram;
1409         return 0;
1410 }
1411
1412 int _tdgram_inet_udp_socket(const struct tsocket_address *local,
1413                             const struct tsocket_address *remote,
1414                             TALLOC_CTX *mem_ctx,
1415                             struct tdgram_context **dgram,
1416                             const char *location)
1417 {
1418         struct tsocket_address_bsd *lbsda =
1419                 talloc_get_type_abort(local->private_data,
1420                 struct tsocket_address_bsd);
1421         int ret;
1422
1423         switch (lbsda->u.sa.sa_family) {
1424         case AF_INET:
1425                 break;
1426 #ifdef HAVE_IPV6
1427         case AF_INET6:
1428                 break;
1429 #endif
1430         default:
1431                 errno = EINVAL;
1432                 return -1;
1433         }
1434
1435         ret = tdgram_bsd_dgram_socket(local, remote, false,
1436                                       mem_ctx, dgram, location);
1437
1438         return ret;
1439 }
1440
1441 int _tdgram_unix_socket(const struct tsocket_address *local,
1442                         const struct tsocket_address *remote,
1443                         TALLOC_CTX *mem_ctx,
1444                         struct tdgram_context **dgram,
1445                         const char *location)
1446 {
1447         struct tsocket_address_bsd *lbsda =
1448                 talloc_get_type_abort(local->private_data,
1449                 struct tsocket_address_bsd);
1450         int ret;
1451
1452         switch (lbsda->u.sa.sa_family) {
1453         case AF_UNIX:
1454                 break;
1455         default:
1456                 errno = EINVAL;
1457                 return -1;
1458         }
1459
1460         ret = tdgram_bsd_dgram_socket(local, remote, false,
1461                                       mem_ctx, dgram, location);
1462
1463         return ret;
1464 }
1465
1466 struct tstream_bsd {
1467         int fd;
1468
1469         void *event_ptr;
1470         struct tevent_fd *fde;
1471         bool optimize_readv;
1472
1473         void *readable_private;
1474         void (*readable_handler)(void *private_data);
1475         void *writeable_private;
1476         void (*writeable_handler)(void *private_data);
1477 };
1478
1479 bool tstream_bsd_optimize_readv(struct tstream_context *stream,
1480                                 bool on)
1481 {
1482         struct tstream_bsd *bsds =
1483                 talloc_get_type(_tstream_context_data(stream),
1484                 struct tstream_bsd);
1485         bool old;
1486
1487         if (bsds == NULL) {
1488                 /* not a bsd socket */
1489                 return false;
1490         }
1491
1492         old = bsds->optimize_readv;
1493         bsds->optimize_readv = on;
1494
1495         return old;
1496 }
1497
1498 static void tstream_bsd_fde_handler(struct tevent_context *ev,
1499                                     struct tevent_fd *fde,
1500                                     uint16_t flags,
1501                                     void *private_data)
1502 {
1503         struct tstream_bsd *bsds = talloc_get_type_abort(private_data,
1504                                    struct tstream_bsd);
1505
1506         if (flags & TEVENT_FD_WRITE) {
1507                 bsds->writeable_handler(bsds->writeable_private);
1508                 return;
1509         }
1510         if (flags & TEVENT_FD_READ) {
1511                 if (!bsds->readable_handler) {
1512                         if (bsds->writeable_handler) {
1513                                 bsds->writeable_handler(bsds->writeable_private);
1514                                 return;
1515                         }
1516                         TEVENT_FD_NOT_READABLE(bsds->fde);
1517                         return;
1518                 }
1519                 bsds->readable_handler(bsds->readable_private);
1520                 return;
1521         }
1522 }
1523
1524 static int tstream_bsd_set_readable_handler(struct tstream_bsd *bsds,
1525                                             struct tevent_context *ev,
1526                                             void (*handler)(void *private_data),
1527                                             void *private_data)
1528 {
1529         if (ev == NULL) {
1530                 if (handler) {
1531                         errno = EINVAL;
1532                         return -1;
1533                 }
1534                 if (!bsds->readable_handler) {
1535                         return 0;
1536                 }
1537                 bsds->readable_handler = NULL;
1538                 bsds->readable_private = NULL;
1539
1540                 return 0;
1541         }
1542
1543         /* read and write must use the same tevent_context */
1544         if (bsds->event_ptr != ev) {
1545                 if (bsds->readable_handler || bsds->writeable_handler) {
1546                         errno = EINVAL;
1547                         return -1;
1548                 }
1549                 bsds->event_ptr = NULL;
1550                 TALLOC_FREE(bsds->fde);
1551         }
1552
1553         if (tevent_fd_get_flags(bsds->fde) == 0) {
1554                 TALLOC_FREE(bsds->fde);
1555
1556                 bsds->fde = tevent_add_fd(ev, bsds,
1557                                           bsds->fd, TEVENT_FD_READ,
1558                                           tstream_bsd_fde_handler,
1559                                           bsds);
1560                 if (!bsds->fde) {
1561                         errno = ENOMEM;
1562                         return -1;
1563                 }
1564
1565                 /* cache the event context we're running on */
1566                 bsds->event_ptr = ev;
1567         } else if (!bsds->readable_handler) {
1568                 TEVENT_FD_READABLE(bsds->fde);
1569         }
1570
1571         bsds->readable_handler = handler;
1572         bsds->readable_private = private_data;
1573
1574         return 0;
1575 }
1576
1577 static int tstream_bsd_set_writeable_handler(struct tstream_bsd *bsds,
1578                                              struct tevent_context *ev,
1579                                              void (*handler)(void *private_data),
1580                                              void *private_data)
1581 {
1582         if (ev == NULL) {
1583                 if (handler) {
1584                         errno = EINVAL;
1585                         return -1;
1586                 }
1587                 if (!bsds->writeable_handler) {
1588                         return 0;
1589                 }
1590                 bsds->writeable_handler = NULL;
1591                 bsds->writeable_private = NULL;
1592                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
1593
1594                 return 0;
1595         }
1596
1597         /* read and write must use the same tevent_context */
1598         if (bsds->event_ptr != ev) {
1599                 if (bsds->readable_handler || bsds->writeable_handler) {
1600                         errno = EINVAL;
1601                         return -1;
1602                 }
1603                 bsds->event_ptr = NULL;
1604                 TALLOC_FREE(bsds->fde);
1605         }
1606
1607         if (tevent_fd_get_flags(bsds->fde) == 0) {
1608                 TALLOC_FREE(bsds->fde);
1609
1610                 bsds->fde = tevent_add_fd(ev, bsds,
1611                                           bsds->fd,
1612                                           TEVENT_FD_READ | TEVENT_FD_WRITE,
1613                                           tstream_bsd_fde_handler,
1614                                           bsds);
1615                 if (!bsds->fde) {
1616                         errno = ENOMEM;
1617                         return -1;
1618                 }
1619
1620                 /* cache the event context we're running on */
1621                 bsds->event_ptr = ev;
1622         } else if (!bsds->writeable_handler) {
1623                 uint16_t flags = tevent_fd_get_flags(bsds->fde);
1624                 flags |= TEVENT_FD_READ | TEVENT_FD_WRITE;
1625                 tevent_fd_set_flags(bsds->fde, flags);
1626         }
1627
1628         bsds->writeable_handler = handler;
1629         bsds->writeable_private = private_data;
1630
1631         return 0;
1632 }
1633
1634 static ssize_t tstream_bsd_pending_bytes(struct tstream_context *stream)
1635 {
1636         struct tstream_bsd *bsds = tstream_context_data(stream,
1637                                    struct tstream_bsd);
1638         ssize_t ret;
1639
1640         if (bsds->fd == -1) {
1641                 errno = ENOTCONN;
1642                 return -1;
1643         }
1644
1645         ret = tsocket_bsd_pending(bsds->fd);
1646
1647         return ret;
1648 }
1649
1650 struct tstream_bsd_readv_state {
1651         struct tstream_context *stream;
1652
1653         struct iovec *vector;
1654         size_t count;
1655
1656         int ret;
1657 };
1658
1659 static int tstream_bsd_readv_destructor(struct tstream_bsd_readv_state *state)
1660 {
1661         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1662                                    struct tstream_bsd);
1663
1664         tstream_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
1665
1666         return 0;
1667 }
1668
1669 static void tstream_bsd_readv_handler(void *private_data);
1670
1671 static struct tevent_req *tstream_bsd_readv_send(TALLOC_CTX *mem_ctx,
1672                                         struct tevent_context *ev,
1673                                         struct tstream_context *stream,
1674                                         struct iovec *vector,
1675                                         size_t count)
1676 {
1677         struct tevent_req *req;
1678         struct tstream_bsd_readv_state *state;
1679         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1680         int ret;
1681
1682         req = tevent_req_create(mem_ctx, &state,
1683                                 struct tstream_bsd_readv_state);
1684         if (!req) {
1685                 return NULL;
1686         }
1687
1688         state->stream   = stream;
1689         /* we make a copy of the vector so that we can modify it */
1690         state->vector   = talloc_array(state, struct iovec, count);
1691         if (tevent_req_nomem(state->vector, req)) {
1692                 goto post;
1693         }
1694         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1695         state->count    = count;
1696         state->ret      = 0;
1697
1698         talloc_set_destructor(state, tstream_bsd_readv_destructor);
1699
1700         if (bsds->fd == -1) {
1701                 tevent_req_error(req, ENOTCONN);
1702                 goto post;
1703         }
1704
1705         /*
1706          * this is a fast path, not waiting for the
1707          * socket to become explicit readable gains
1708          * about 10%-20% performance in benchmark tests.
1709          */
1710         if (bsds->optimize_readv) {
1711                 /*
1712                  * We only do the optimization on
1713                  * readv if the caller asked for it.
1714                  *
1715                  * This is needed because in most cases
1716                  * we preferr to flush send buffers before
1717                  * receiving incoming requests.
1718                  */
1719                 tstream_bsd_readv_handler(req);
1720                 if (!tevent_req_is_in_progress(req)) {
1721                         goto post;
1722                 }
1723         }
1724
1725         ret = tstream_bsd_set_readable_handler(bsds, ev,
1726                                               tstream_bsd_readv_handler,
1727                                               req);
1728         if (ret == -1) {
1729                 tevent_req_error(req, errno);
1730                 goto post;
1731         }
1732
1733         return req;
1734
1735  post:
1736         tevent_req_post(req, ev);
1737         return req;
1738 }
1739
1740 static void tstream_bsd_readv_handler(void *private_data)
1741 {
1742         struct tevent_req *req = talloc_get_type_abort(private_data,
1743                                  struct tevent_req);
1744         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1745                                         struct tstream_bsd_readv_state);
1746         struct tstream_context *stream = state->stream;
1747         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1748         int ret;
1749         int err;
1750         bool retry;
1751
1752         ret = readv(bsds->fd, state->vector, state->count);
1753         if (ret == 0) {
1754                 /* propagate end of file */
1755                 tevent_req_error(req, EPIPE);
1756                 return;
1757         }
1758         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1759         if (retry) {
1760                 /* retry later */
1761                 return;
1762         }
1763         if (tevent_req_error(req, err)) {
1764                 return;
1765         }
1766
1767         state->ret += ret;
1768
1769         while (ret > 0) {
1770                 if (ret < state->vector[0].iov_len) {
1771                         uint8_t *base;
1772                         base = (uint8_t *)state->vector[0].iov_base;
1773                         base += ret;
1774                         state->vector[0].iov_base = (void *)base;
1775                         state->vector[0].iov_len -= ret;
1776                         break;
1777                 }
1778                 ret -= state->vector[0].iov_len;
1779                 state->vector += 1;
1780                 state->count -= 1;
1781         }
1782
1783         /*
1784          * there're maybe some empty vectors at the end
1785          * which we need to skip, otherwise we would get
1786          * ret == 0 from the readv() call and return EPIPE
1787          */
1788         while (state->count > 0) {
1789                 if (state->vector[0].iov_len > 0) {
1790                         break;
1791                 }
1792                 state->vector += 1;
1793                 state->count -= 1;
1794         }
1795
1796         if (state->count > 0) {
1797                 /* we have more to read */
1798                 return;
1799         }
1800
1801         tevent_req_done(req);
1802 }
1803
1804 static int tstream_bsd_readv_recv(struct tevent_req *req,
1805                                   int *perrno)
1806 {
1807         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1808                                         struct tstream_bsd_readv_state);
1809         int ret;
1810
1811         ret = tsocket_simple_int_recv(req, perrno);
1812         if (ret == 0) {
1813                 ret = state->ret;
1814         }
1815
1816         tevent_req_received(req);
1817         return ret;
1818 }
1819
1820 struct tstream_bsd_writev_state {
1821         struct tstream_context *stream;
1822
1823         struct iovec *vector;
1824         size_t count;
1825
1826         int ret;
1827 };
1828
1829 static int tstream_bsd_writev_destructor(struct tstream_bsd_writev_state *state)
1830 {
1831         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1832                                   struct tstream_bsd);
1833
1834         tstream_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
1835
1836         return 0;
1837 }
1838
1839 static void tstream_bsd_writev_handler(void *private_data);
1840
1841 static struct tevent_req *tstream_bsd_writev_send(TALLOC_CTX *mem_ctx,
1842                                                  struct tevent_context *ev,
1843                                                  struct tstream_context *stream,
1844                                                  const struct iovec *vector,
1845                                                  size_t count)
1846 {
1847         struct tevent_req *req;
1848         struct tstream_bsd_writev_state *state;
1849         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1850         int ret;
1851
1852         req = tevent_req_create(mem_ctx, &state,
1853                                 struct tstream_bsd_writev_state);
1854         if (!req) {
1855                 return NULL;
1856         }
1857
1858         state->stream   = stream;
1859         /* we make a copy of the vector so that we can modify it */
1860         state->vector   = talloc_array(state, struct iovec, count);
1861         if (tevent_req_nomem(state->vector, req)) {
1862                 goto post;
1863         }
1864         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1865         state->count    = count;
1866         state->ret      = 0;
1867
1868         talloc_set_destructor(state, tstream_bsd_writev_destructor);
1869
1870         if (bsds->fd == -1) {
1871                 tevent_req_error(req, ENOTCONN);
1872                 goto post;
1873         }
1874
1875         /*
1876          * this is a fast path, not waiting for the
1877          * socket to become explicit writeable gains
1878          * about 10%-20% performance in benchmark tests.
1879          */
1880         tstream_bsd_writev_handler(req);
1881         if (!tevent_req_is_in_progress(req)) {
1882                 goto post;
1883         }
1884
1885         ret = tstream_bsd_set_writeable_handler(bsds, ev,
1886                                                tstream_bsd_writev_handler,
1887                                                req);
1888         if (ret == -1) {
1889                 tevent_req_error(req, errno);
1890                 goto post;
1891         }
1892
1893         return req;
1894
1895  post:
1896         tevent_req_post(req, ev);
1897         return req;
1898 }
1899
1900 static void tstream_bsd_writev_handler(void *private_data)
1901 {
1902         struct tevent_req *req = talloc_get_type_abort(private_data,
1903                                  struct tevent_req);
1904         struct tstream_bsd_writev_state *state = tevent_req_data(req,
1905                                         struct tstream_bsd_writev_state);
1906         struct tstream_context *stream = state->stream;
1907         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1908         ssize_t ret;
1909         int err;
1910         bool retry;
1911
1912         ret = writev(bsds->fd, state->vector, state->count);
1913         if (ret == 0) {
1914                 /* propagate end of file */
1915                 tevent_req_error(req, EPIPE);
1916                 return;
1917         }
1918         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1919         if (retry) {
1920                 /* retry later */
1921                 return;
1922         }
1923         if (tevent_req_error(req, err)) {
1924                 return;
1925         }
1926
1927         state->ret += ret;
1928
1929         while (ret > 0) {
1930                 if (ret < state->vector[0].iov_len) {
1931                         uint8_t *base;
1932                         base = (uint8_t *)state->vector[0].iov_base;
1933                         base += ret;
1934                         state->vector[0].iov_base = (void *)base;
1935                         state->vector[0].iov_len -= ret;
1936                         break;
1937                 }
1938                 ret -= state->vector[0].iov_len;
1939                 state->vector += 1;
1940                 state->count -= 1;
1941         }
1942
1943         /*
1944          * there're maybe some empty vectors at the end
1945          * which we need to skip, otherwise we would get
1946          * ret == 0 from the writev() call and return EPIPE
1947          */
1948         while (state->count > 0) {
1949                 if (state->vector[0].iov_len > 0) {
1950                         break;
1951                 }
1952                 state->vector += 1;
1953                 state->count -= 1;
1954         }
1955
1956         if (state->count > 0) {
1957                 /* we have more to read */
1958                 return;
1959         }
1960
1961         tevent_req_done(req);
1962 }
1963
1964 static int tstream_bsd_writev_recv(struct tevent_req *req, int *perrno)
1965 {
1966         struct tstream_bsd_writev_state *state = tevent_req_data(req,
1967                                         struct tstream_bsd_writev_state);
1968         int ret;
1969
1970         ret = tsocket_simple_int_recv(req, perrno);
1971         if (ret == 0) {
1972                 ret = state->ret;
1973         }
1974
1975         tevent_req_received(req);
1976         return ret;
1977 }
1978
1979 struct tstream_bsd_disconnect_state {
1980         void *__dummy;
1981 };
1982
1983 static struct tevent_req *tstream_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
1984                                                      struct tevent_context *ev,
1985                                                      struct tstream_context *stream)
1986 {
1987         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1988         struct tevent_req *req;
1989         struct tstream_bsd_disconnect_state *state;
1990         int ret;
1991         int err;
1992         bool dummy;
1993
1994         req = tevent_req_create(mem_ctx, &state,
1995                                 struct tstream_bsd_disconnect_state);
1996         if (req == NULL) {
1997                 return NULL;
1998         }
1999
2000         if (bsds->fd == -1) {
2001                 tevent_req_error(req, ENOTCONN);
2002                 goto post;
2003         }
2004
2005         TALLOC_FREE(bsds->fde);
2006         ret = close(bsds->fd);
2007         bsds->fd = -1;
2008         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
2009         if (tevent_req_error(req, err)) {
2010                 goto post;
2011         }
2012
2013         tevent_req_done(req);
2014 post:
2015         tevent_req_post(req, ev);
2016         return req;
2017 }
2018
2019 static int tstream_bsd_disconnect_recv(struct tevent_req *req,
2020                                       int *perrno)
2021 {
2022         int ret;
2023
2024         ret = tsocket_simple_int_recv(req, perrno);
2025
2026         tevent_req_received(req);
2027         return ret;
2028 }
2029
2030 static const struct tstream_context_ops tstream_bsd_ops = {
2031         .name                   = "bsd",
2032
2033         .pending_bytes          = tstream_bsd_pending_bytes,
2034
2035         .readv_send             = tstream_bsd_readv_send,
2036         .readv_recv             = tstream_bsd_readv_recv,
2037
2038         .writev_send            = tstream_bsd_writev_send,
2039         .writev_recv            = tstream_bsd_writev_recv,
2040
2041         .disconnect_send        = tstream_bsd_disconnect_send,
2042         .disconnect_recv        = tstream_bsd_disconnect_recv,
2043 };
2044
2045 static int tstream_bsd_destructor(struct tstream_bsd *bsds)
2046 {
2047         TALLOC_FREE(bsds->fde);
2048         if (bsds->fd != -1) {
2049                 close(bsds->fd);
2050                 bsds->fd = -1;
2051         }
2052         return 0;
2053 }
2054
2055 int _tstream_bsd_existing_socket(TALLOC_CTX *mem_ctx,
2056                                  int fd,
2057                                  struct tstream_context **_stream,
2058                                  const char *location)
2059 {
2060         struct tstream_context *stream;
2061         struct tstream_bsd *bsds;
2062
2063         stream = tstream_context_create(mem_ctx,
2064                                         &tstream_bsd_ops,
2065                                         &bsds,
2066                                         struct tstream_bsd,
2067                                         location);
2068         if (!stream) {
2069                 return -1;
2070         }
2071         ZERO_STRUCTP(bsds);
2072         bsds->fd = fd;
2073         talloc_set_destructor(bsds, tstream_bsd_destructor);
2074
2075         *_stream = stream;
2076         return 0;
2077 }
2078
2079 struct tstream_bsd_connect_state {
2080         int fd;
2081         struct tevent_fd *fde;
2082         struct tstream_conext *stream;
2083         struct tsocket_address *local;
2084 };
2085
2086 static int tstream_bsd_connect_destructor(struct tstream_bsd_connect_state *state)
2087 {
2088         TALLOC_FREE(state->fde);
2089         if (state->fd != -1) {
2090                 close(state->fd);
2091                 state->fd = -1;
2092         }
2093
2094         return 0;
2095 }
2096
2097 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
2098                                             struct tevent_fd *fde,
2099                                             uint16_t flags,
2100                                             void *private_data);
2101
2102 static struct tevent_req *tstream_bsd_connect_send(TALLOC_CTX *mem_ctx,
2103                                         struct tevent_context *ev,
2104                                         int sys_errno,
2105                                         const struct tsocket_address *local,
2106                                         const struct tsocket_address *remote)
2107 {
2108         struct tevent_req *req;
2109         struct tstream_bsd_connect_state *state;
2110         struct tsocket_address_bsd *lbsda =
2111                 talloc_get_type_abort(local->private_data,
2112                 struct tsocket_address_bsd);
2113         struct tsocket_address_bsd *lrbsda = NULL;
2114         struct tsocket_address_bsd *rbsda =
2115                 talloc_get_type_abort(remote->private_data,
2116                 struct tsocket_address_bsd);
2117         int ret;
2118         int err;
2119         bool retry;
2120         bool do_bind = false;
2121         bool do_reuseaddr = false;
2122         bool do_ipv6only = false;
2123         bool is_inet = false;
2124         int sa_fam = lbsda->u.sa.sa_family;
2125
2126         req = tevent_req_create(mem_ctx, &state,
2127                                 struct tstream_bsd_connect_state);
2128         if (!req) {
2129                 return NULL;
2130         }
2131         state->fd = -1;
2132         state->fde = NULL;
2133
2134         talloc_set_destructor(state, tstream_bsd_connect_destructor);
2135
2136         /* give the wrappers a chance to report an error */
2137         if (sys_errno != 0) {
2138                 tevent_req_error(req, sys_errno);
2139                 goto post;
2140         }
2141
2142         switch (lbsda->u.sa.sa_family) {
2143         case AF_UNIX:
2144                 if (lbsda->u.un.sun_path[0] != 0) {
2145                         do_reuseaddr = true;
2146                         do_bind = true;
2147                 }
2148                 break;
2149         case AF_INET:
2150                 if (lbsda->u.in.sin_port != 0) {
2151                         do_reuseaddr = true;
2152                         do_bind = true;
2153                 }
2154                 if (lbsda->u.in.sin_addr.s_addr != INADDR_ANY) {
2155                         do_bind = true;
2156                 }
2157                 is_inet = true;
2158                 break;
2159 #ifdef HAVE_IPV6
2160         case AF_INET6:
2161                 if (lbsda->u.in6.sin6_port != 0) {
2162                         do_reuseaddr = true;
2163                         do_bind = true;
2164                 }
2165                 if (memcmp(&in6addr_any,
2166                            &lbsda->u.in6.sin6_addr,
2167                            sizeof(in6addr_any)) != 0) {
2168                         do_bind = true;
2169                 }
2170                 is_inet = true;
2171                 do_ipv6only = true;
2172                 break;
2173 #endif
2174         default:
2175                 tevent_req_error(req, EINVAL);
2176                 goto post;
2177         }
2178
2179         if (!do_bind && is_inet) {
2180                 sa_fam = rbsda->u.sa.sa_family;
2181                 switch (sa_fam) {
2182                 case AF_INET:
2183                         do_ipv6only = false;
2184                         break;
2185 #ifdef HAVE_IPV6
2186                 case AF_INET6:
2187                         do_ipv6only = true;
2188                         break;
2189 #endif
2190                 }
2191         }
2192
2193         if (is_inet) {
2194                 state->local = tsocket_address_create(state,
2195                                                       &tsocket_address_bsd_ops,
2196                                                       &lrbsda,
2197                                                       struct tsocket_address_bsd,
2198                                                       __location__ "bsd_connect");
2199                 if (tevent_req_nomem(state->local, req)) {
2200                         goto post;
2201                 }
2202
2203                 ZERO_STRUCTP(lrbsda);
2204                 lrbsda->sa_socklen = sizeof(lrbsda->u.ss);
2205 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
2206                 lrbsda->u.sa.sa_len = lrbsda->sa_socklen;
2207 #endif
2208         }
2209
2210         state->fd = socket(sa_fam, SOCK_STREAM, 0);
2211         if (state->fd == -1) {
2212                 tevent_req_error(req, errno);
2213                 goto post;
2214         }
2215
2216         state->fd = tsocket_bsd_common_prepare_fd(state->fd, true);
2217         if (state->fd == -1) {
2218                 tevent_req_error(req, errno);
2219                 goto post;
2220         }
2221
2222 #ifdef HAVE_IPV6
2223         if (do_ipv6only) {
2224                 int val = 1;
2225
2226                 ret = setsockopt(state->fd, IPPROTO_IPV6, IPV6_V6ONLY,
2227                                  (const void *)&val, sizeof(val));
2228                 if (ret == -1) {
2229                         tevent_req_error(req, errno);
2230                         goto post;
2231                 }
2232         }
2233 #endif
2234
2235         if (do_reuseaddr) {
2236                 int val = 1;
2237
2238                 ret = setsockopt(state->fd, SOL_SOCKET, SO_REUSEADDR,
2239                                  (const void *)&val, sizeof(val));
2240                 if (ret == -1) {
2241                         tevent_req_error(req, errno);
2242                         goto post;
2243                 }
2244         }
2245
2246         if (do_bind) {
2247                 ret = bind(state->fd, &lbsda->u.sa, lbsda->sa_socklen);
2248                 if (ret == -1) {
2249                         tevent_req_error(req, errno);
2250                         goto post;
2251                 }
2252         }
2253
2254         if (rbsda->u.sa.sa_family != sa_fam) {
2255                 tevent_req_error(req, EINVAL);
2256                 goto post;
2257         }
2258
2259         ret = connect(state->fd, &rbsda->u.sa, rbsda->sa_socklen);
2260         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
2261         if (retry) {
2262                 /* retry later */
2263                 goto async;
2264         }
2265         if (tevent_req_error(req, err)) {
2266                 goto post;
2267         }
2268
2269         if (!state->local) {
2270                 tevent_req_done(req);
2271                 goto post;
2272         }
2273
2274         ret = getsockname(state->fd, &lrbsda->u.sa, &lrbsda->sa_socklen);
2275         if (ret == -1) {
2276                 tevent_req_error(req, errno);
2277                 goto post;
2278         }
2279
2280         tevent_req_done(req);
2281         goto post;
2282
2283  async:
2284         state->fde = tevent_add_fd(ev, state,
2285                                    state->fd,
2286                                    TEVENT_FD_READ | TEVENT_FD_WRITE,
2287                                    tstream_bsd_connect_fde_handler,
2288                                    req);
2289         if (tevent_req_nomem(state->fde, req)) {
2290                 goto post;
2291         }
2292
2293         return req;
2294
2295  post:
2296         tevent_req_post(req, ev);
2297         return req;
2298 }
2299
2300 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
2301                                             struct tevent_fd *fde,
2302                                             uint16_t flags,
2303                                             void *private_data)
2304 {
2305         struct tevent_req *req = talloc_get_type_abort(private_data,
2306                                  struct tevent_req);
2307         struct tstream_bsd_connect_state *state = tevent_req_data(req,
2308                                         struct tstream_bsd_connect_state);
2309         struct tsocket_address_bsd *lrbsda = NULL;
2310         int ret;
2311         int error=0;
2312         socklen_t len = sizeof(error);
2313         int err;
2314         bool retry;
2315
2316         ret = getsockopt(state->fd, SOL_SOCKET, SO_ERROR, &error, &len);
2317         if (ret == 0) {
2318                 if (error != 0) {
2319                         errno = error;
2320                         ret = -1;
2321                 }
2322         }
2323         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
2324         if (retry) {
2325                 /* retry later */
2326                 return;
2327         }
2328         if (tevent_req_error(req, err)) {
2329                 return;
2330         }
2331
2332         if (!state->local) {
2333                 tevent_req_done(req);
2334                 return;
2335         }
2336
2337         lrbsda = talloc_get_type_abort(state->local->private_data,
2338                                        struct tsocket_address_bsd);
2339
2340         ret = getsockname(state->fd, &lrbsda->u.sa, &lrbsda->sa_socklen);
2341         if (ret == -1) {
2342                 tevent_req_error(req, errno);
2343                 return;
2344         }
2345
2346         tevent_req_done(req);
2347 }
2348
2349 static int tstream_bsd_connect_recv(struct tevent_req *req,
2350                                     int *perrno,
2351                                     TALLOC_CTX *mem_ctx,
2352                                     struct tstream_context **stream,
2353                                     struct tsocket_address **local,
2354                                     const char *location)
2355 {
2356         struct tstream_bsd_connect_state *state = tevent_req_data(req,
2357                                         struct tstream_bsd_connect_state);
2358         int ret;
2359
2360         ret = tsocket_simple_int_recv(req, perrno);
2361         if (ret == 0) {
2362                 ret = _tstream_bsd_existing_socket(mem_ctx,
2363                                                    state->fd,
2364                                                    stream,
2365                                                    location);
2366                 if (ret == -1) {
2367                         *perrno = errno;
2368                         goto done;
2369                 }
2370                 TALLOC_FREE(state->fde);
2371                 state->fd = -1;
2372
2373                 if (local) {
2374                         *local = talloc_move(mem_ctx, &state->local);
2375                 }
2376         }
2377
2378 done:
2379         tevent_req_received(req);
2380         return ret;
2381 }
2382
2383 struct tevent_req * tstream_inet_tcp_connect_send(TALLOC_CTX *mem_ctx,
2384                                         struct tevent_context *ev,
2385                                         const struct tsocket_address *local,
2386                                         const struct tsocket_address *remote)
2387 {
2388         struct tsocket_address_bsd *lbsda =
2389                 talloc_get_type_abort(local->private_data,
2390                 struct tsocket_address_bsd);
2391         struct tevent_req *req;
2392         int sys_errno = 0;
2393
2394         switch (lbsda->u.sa.sa_family) {
2395         case AF_INET:
2396                 break;
2397 #ifdef HAVE_IPV6
2398         case AF_INET6:
2399                 break;
2400 #endif
2401         default:
2402                 sys_errno = EINVAL;
2403                 break;
2404         }
2405
2406         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2407
2408         return req;
2409 }
2410
2411 int _tstream_inet_tcp_connect_recv(struct tevent_req *req,
2412                                    int *perrno,
2413                                    TALLOC_CTX *mem_ctx,
2414                                    struct tstream_context **stream,
2415                                    struct tsocket_address **local,
2416                                    const char *location)
2417 {
2418         return tstream_bsd_connect_recv(req, perrno,
2419                                         mem_ctx, stream, local,
2420                                         location);
2421 }
2422
2423 struct tevent_req * tstream_unix_connect_send(TALLOC_CTX *mem_ctx,
2424                                         struct tevent_context *ev,
2425                                         const struct tsocket_address *local,
2426                                         const struct tsocket_address *remote)
2427 {
2428         struct tsocket_address_bsd *lbsda =
2429                 talloc_get_type_abort(local->private_data,
2430                 struct tsocket_address_bsd);
2431         struct tevent_req *req;
2432         int sys_errno = 0;
2433
2434         switch (lbsda->u.sa.sa_family) {
2435         case AF_UNIX:
2436                 break;
2437         default:
2438                 sys_errno = EINVAL;
2439                 break;
2440         }
2441
2442         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2443
2444         return req;
2445 }
2446
2447 int _tstream_unix_connect_recv(struct tevent_req *req,
2448                                       int *perrno,
2449                                       TALLOC_CTX *mem_ctx,
2450                                       struct tstream_context **stream,
2451                                       const char *location)
2452 {
2453         return tstream_bsd_connect_recv(req, perrno,
2454                                         mem_ctx, stream, NULL,
2455                                         location);
2456 }
2457
2458 int _tstream_unix_socketpair(TALLOC_CTX *mem_ctx1,
2459                              struct tstream_context **_stream1,
2460                              TALLOC_CTX *mem_ctx2,
2461                              struct tstream_context **_stream2,
2462                              const char *location)
2463 {
2464         int ret;
2465         int fds[2];
2466         int fd1;
2467         int fd2;
2468         struct tstream_context *stream1 = NULL;
2469         struct tstream_context *stream2 = NULL;
2470
2471         ret = socketpair(AF_UNIX, SOCK_STREAM, 0, fds);
2472         if (ret == -1) {
2473                 return -1;
2474         }
2475         fd1 = fds[0];
2476         fd2 = fds[1];
2477
2478         fd1 = tsocket_bsd_common_prepare_fd(fd1, true);
2479         if (fd1 == -1) {
2480                 int sys_errno = errno;
2481                 close(fd2);
2482                 errno = sys_errno;
2483                 return -1;
2484         }
2485
2486         fd2 = tsocket_bsd_common_prepare_fd(fd2, true);
2487         if (fd2 == -1) {
2488                 int sys_errno = errno;
2489                 close(fd1);
2490                 errno = sys_errno;
2491                 return -1;
2492         }
2493
2494         ret = _tstream_bsd_existing_socket(mem_ctx1,
2495                                            fd1,
2496                                            &stream1,
2497                                            location);
2498         if (ret == -1) {
2499                 int sys_errno = errno;
2500                 close(fd1);
2501                 close(fd2);
2502                 errno = sys_errno;
2503                 return -1;
2504         }
2505
2506         ret = _tstream_bsd_existing_socket(mem_ctx2,
2507                                            fd2,
2508                                            &stream2,
2509                                            location);
2510         if (ret == -1) {
2511                 int sys_errno = errno;
2512                 talloc_free(stream1);
2513                 close(fd2);
2514                 errno = sys_errno;
2515                 return -1;
2516         }
2517
2518         *_stream1 = stream1;
2519         *_stream2 = stream2;
2520         return 0;
2521 }
2522