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