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                         return -1;
 622                 }
 623
 624                 /* cache the event context we're running on */
 625                 bsds->event_ptr = ev;
 626         } else if (!bsds->readable_handler) {
 627                 TEVENT_FD_READABLE(bsds->fde);
 628         }
 629
 630         bsds->readable_handler = handler;
 631         bsds->readable_private = private_data;
 632
 633         return 0;
 634 }
 635
 636 static int tdgram_bsd_set_writeable_handler(struct tdgram_bsd *bsds,
 637                                             struct tevent_context *ev,
 638                                             void (*handler)(void *private_data),
 639                                             void *private_data)
 640 {
 641         if (ev == NULL) {
 642                 if (handler) {
 643                         errno = EINVAL;
 644                         return -1;
 645                 }
 646                 if (!bsds->writeable_handler) {
 647                         return 0;
 648                 }
 649                 bsds->writeable_handler = NULL;
 650                 bsds->writeable_private = NULL;
 651                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
 652
 653                 return 0;
 654         }
 655
 656         /* read and write must use the same tevent_context */
 657         if (bsds->event_ptr != ev) {
 658                 if (bsds->readable_handler || bsds->writeable_handler) {
 659                         errno = EINVAL;
 660                         return -1;
 661                 }
 662                 bsds->event_ptr = NULL;
 663                 TALLOC_FREE(bsds->fde);
 664         }
 665
 666         if (bsds->fde == NULL) {
 667                 bsds->fde = tevent_add_fd(ev, bsds,
 668                                           bsds->fd, TEVENT_FD_WRITE,
 669                                           tdgram_bsd_fde_handler,
 670                                           bsds);
 671                 if (!bsds->fde) {
 672                         return -1;
 673                 }
 674
 675                 /* cache the event context we're running on */
 676                 bsds->event_ptr = ev;
 677         } else if (!bsds->writeable_handler) {
 678                 TEVENT_FD_WRITEABLE(bsds->fde);
 679         }
 680
 681         bsds->writeable_handler = handler;
 682         bsds->writeable_private = private_data;
 683
 684         return 0;
 685 }
 686
 687 struct tdgram_bsd_recvfrom_state {
 688         struct tdgram_context *dgram;
 689
 690         uint8_t *buf;
 691         size_t len;
 692         struct tsocket_address *src;
 693 };
 694
 695 static int tdgram_bsd_recvfrom_destructor(struct tdgram_bsd_recvfrom_state *state)
 696 {
 697         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
 698                                   struct tdgram_bsd);
 699
 700         tdgram_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
 701
 702         return 0;
 703 }
 704
 705 static void tdgram_bsd_recvfrom_handler(void *private_data);
 706
 707 static struct tevent_req *tdgram_bsd_recvfrom_send(TALLOC_CTX *mem_ctx,
 708                                         struct tevent_context *ev,
 709                                         struct tdgram_context *dgram)
 710 {
 711         struct tevent_req *req;
 712         struct tdgram_bsd_recvfrom_state *state;
 713         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 714         int ret;
 715
 716         req = tevent_req_create(mem_ctx, &state,
 717                                 struct tdgram_bsd_recvfrom_state);
 718         if (!req) {
 719                 return NULL;
 720         }
 721
 722         state->dgram    = dgram;
 723         state->buf      = NULL;
 724         state->len      = 0;
 725         state->src      = NULL;
 726
 727         talloc_set_destructor(state, tdgram_bsd_recvfrom_destructor);
 728
 729         if (bsds->fd == -1) {
 730                 tevent_req_error(req, ENOTCONN);
 731                 goto post;
 732         }
 733
 734         /*
 735          * this is a fast path, not waiting for the
 736          * socket to become explicit readable gains
 737          * about 10%-20% performance in benchmark tests.
 738          */
 739         tdgram_bsd_recvfrom_handler(req);
 740         if (!tevent_req_is_in_progress(req)) {
 741                 goto post;
 742         }
 743
 744         ret = tdgram_bsd_set_readable_handler(bsds, ev,
 745                                               tdgram_bsd_recvfrom_handler,
 746                                               req);
 747         if (ret == -1) {
 748                 tevent_req_error(req, errno);
 749                 goto post;
 750         }
 751
 752         return req;
 753
 754  post:
 755         tevent_req_post(req, ev);
 756         return req;
 757 }
 758
 759 static void tdgram_bsd_recvfrom_handler(void *private_data)
 760 {
 761         struct tevent_req *req = talloc_get_type_abort(private_data,
 762                                  struct tevent_req);
 763         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
 764                                         struct tdgram_bsd_recvfrom_state);
 765         struct tdgram_context *dgram = state->dgram;
 766         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 767         struct tsocket_address_bsd *bsda;
 768         ssize_t ret;
 769         struct sockaddr *sa = NULL;
 770         socklen_t sa_len = 0;
 771         int err;
 772         bool retry;
 773
 774         ret = tsocket_bsd_pending(bsds->fd);
 775         if (ret == 0) {
 776                 /* retry later */
 777                 return;
 778         }
 779         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 780         if (retry) {
 781                 /* retry later */
 782                 return;
 783         }
 784         if (tevent_req_error(req, err)) {
 785                 return;
 786         }
 787
 788         state->buf = talloc_array(state, uint8_t, ret);
 789         if (tevent_req_nomem(state->buf, req)) {
 790                 return;
 791         }
 792         state->len = ret;
 793
 794         state->src = tsocket_address_create(state,
 795                                             &tsocket_address_bsd_ops,
 796                                             &bsda,
 797                                             struct tsocket_address_bsd,
 798                                             __location__ "bsd_recvfrom");
 799         if (tevent_req_nomem(state->src, req)) {
 800                 return;
 801         }
 802
 803         ZERO_STRUCTP(bsda);
 804
 805         sa = &bsda->u.sa;
 806         sa_len = sizeof(bsda->u.ss);
 807         /*
 808          * for unix sockets we can't use the size of sockaddr_storage
 809          * we would get EINVAL
 810          */
 811         if (bsda->u.sa.sa_family == AF_UNIX) {
 812                 sa_len = sizeof(bsda->u.un);
 813         }
 814
 815         ret = recvfrom(bsds->fd, state->buf, state->len, 0, sa, &sa_len);
 816         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 817         if (retry) {
 818                 /* retry later */
 819                 return;
 820         }
 821         if (tevent_req_error(req, err)) {
 822                 return;
 823         }
 824
 825         if (ret != state->len) {
 826                 tevent_req_error(req, EIO);
 827                 return;
 828         }
 829
 830         tevent_req_done(req);
 831 }
 832
 833 static ssize_t tdgram_bsd_recvfrom_recv(struct tevent_req *req,
 834                                         int *perrno,
 835                                         TALLOC_CTX *mem_ctx,
 836                                         uint8_t **buf,
 837                                         struct tsocket_address **src)
 838 {
 839         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
 840                                         struct tdgram_bsd_recvfrom_state);
 841         ssize_t ret;
 842
 843         ret = tsocket_simple_int_recv(req, perrno);
 844         if (ret == 0) {
 845                 *buf = talloc_move(mem_ctx, &state->buf);
 846                 ret = state->len;
 847                 if (src) {
 848                         *src = talloc_move(mem_ctx, &state->src);
 849                 }
 850         }
 851
 852         tevent_req_received(req);
 853         return ret;
 854 }
 855
 856 struct tdgram_bsd_sendto_state {
 857         struct tdgram_context *dgram;
 858
 859         const uint8_t *buf;
 860         size_t len;
 861         const struct tsocket_address *dst;
 862
 863         ssize_t ret;
 864 };
 865
 866 static int tdgram_bsd_sendto_destructor(struct tdgram_bsd_sendto_state *state)
 867 {
 868         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
 869                                   struct tdgram_bsd);
 870
 871         tdgram_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
 872
 873         return 0;
 874 }
 875
 876 static void tdgram_bsd_sendto_handler(void *private_data);
 877
 878 static struct tevent_req *tdgram_bsd_sendto_send(TALLOC_CTX *mem_ctx,
 879                                                  struct tevent_context *ev,
 880                                                  struct tdgram_context *dgram,
 881                                                  const uint8_t *buf,
 882                                                  size_t len,
 883                                                  const struct tsocket_address *dst)
 884 {
 885         struct tevent_req *req;
 886         struct tdgram_bsd_sendto_state *state;
 887         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 888         int ret;
 889
 890         req = tevent_req_create(mem_ctx, &state,
 891                                 struct tdgram_bsd_sendto_state);
 892         if (!req) {
 893                 return NULL;
 894         }
 895
 896         state->dgram    = dgram;
 897         state->buf      = buf;
 898         state->len      = len;
 899         state->dst      = dst;
 900         state->ret      = -1;
 901
 902         talloc_set_destructor(state, tdgram_bsd_sendto_destructor);
 903
 904         if (bsds->fd == -1) {
 905                 tevent_req_error(req, ENOTCONN);
 906                 goto post;
 907         }
 908
 909         /*
 910          * this is a fast path, not waiting for the
 911          * socket to become explicit writeable gains
 912          * about 10%-20% performance in benchmark tests.
 913          */
 914         tdgram_bsd_sendto_handler(req);
 915         if (!tevent_req_is_in_progress(req)) {
 916                 goto post;
 917         }
 918
 919         ret = tdgram_bsd_set_writeable_handler(bsds, ev,
 920                                                tdgram_bsd_sendto_handler,
 921                                                req);
 922         if (ret == -1) {
 923                 tevent_req_error(req, errno);
 924                 goto post;
 925         }
 926
 927         return req;
 928
 929  post:
 930         tevent_req_post(req, ev);
 931         return req;
 932 }
 933
 934 static void tdgram_bsd_sendto_handler(void *private_data)
 935 {
 936         struct tevent_req *req = talloc_get_type_abort(private_data,
 937                                  struct tevent_req);
 938         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
 939                                         struct tdgram_bsd_sendto_state);
 940         struct tdgram_context *dgram = state->dgram;
 941         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 942         struct sockaddr *sa = NULL;
 943         socklen_t sa_len = 0;
 944         ssize_t ret;
 945         int err;
 946         bool retry;
 947
 948         if (state->dst) {
 949                 struct tsocket_address_bsd *bsda =
 950                         talloc_get_type(state->dst->private_data,
 951                         struct tsocket_address_bsd);
 952
 953                 sa = &bsda->u.sa;
 954                 sa_len = sizeof(bsda->u.ss);
 955                 /*
 956                  * for unix sockets we can't use the size of sockaddr_storage
 957                  * we would get EINVAL
 958                  */
 959                 if (bsda->u.sa.sa_family == AF_UNIX) {
 960                         sa_len = sizeof(bsda->u.un);
 961                 }
 962         }
 963
 964         ret = sendto(bsds->fd, state->buf, state->len, 0, sa, sa_len);
 965         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 966         if (retry) {
 967                 /* retry later */
 968                 return;
 969         }
 970         if (tevent_req_error(req, err)) {
 971                 return;
 972         }
 973
 974         state->ret = ret;
 975
 976         tevent_req_done(req);
 977 }
 978
 979 static ssize_t tdgram_bsd_sendto_recv(struct tevent_req *req, int *perrno)
 980 {
 981         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
 982                                         struct tdgram_bsd_sendto_state);
 983         ssize_t ret;
 984
 985         ret = tsocket_simple_int_recv(req, perrno);
 986         if (ret == 0) {
 987                 ret = state->ret;
 988         }
 989
 990         tevent_req_received(req);
 991         return ret;
 992 }
 993
 994 struct tdgram_bsd_disconnect_state {
 995         uint8_t __dummy;
 996 };
 997
 998 static struct tevent_req *tdgram_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
 999                                                      struct tevent_context *ev,
1000                                                      struct tdgram_context *dgram)
1001 {
1002         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1003         struct tevent_req *req;
1004         struct tdgram_bsd_disconnect_state *state;
1005         int ret;
1006         int err;
1007         bool dummy;
1008
1009         req = tevent_req_create(mem_ctx, &state,
1010                                 struct tdgram_bsd_disconnect_state);
1011         if (req == NULL) {
1012                 return NULL;
1013         }
1014
1015         if (bsds->fd == -1) {
1016                 tevent_req_error(req, ENOTCONN);
1017                 goto post;
1018         }
1019
1020         ret = close(bsds->fd);
1021         bsds->fd = -1;
1022         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
1023         if (tevent_req_error(req, err)) {
1024                 goto post;
1025         }
1026
1027         tevent_req_done(req);
1028 post:
1029         tevent_req_post(req, ev);
1030         return req;
1031 }
1032
1033 static int tdgram_bsd_disconnect_recv(struct tevent_req *req,
1034                                       int *perrno)
1035 {
1036         int ret;
1037
1038         ret = tsocket_simple_int_recv(req, perrno);
1039
1040         tevent_req_received(req);
1041         return ret;
1042 }
1043
1044 static const struct tdgram_context_ops tdgram_bsd_ops = {
1045         .name                   = "bsd",
1046
1047         .recvfrom_send          = tdgram_bsd_recvfrom_send,
1048         .recvfrom_recv          = tdgram_bsd_recvfrom_recv,
1049
1050         .sendto_send            = tdgram_bsd_sendto_send,
1051         .sendto_recv            = tdgram_bsd_sendto_recv,
1052
1053         .disconnect_send        = tdgram_bsd_disconnect_send,
1054         .disconnect_recv        = tdgram_bsd_disconnect_recv,
1055 };
1056
1057 static int tdgram_bsd_destructor(struct tdgram_bsd *bsds)
1058 {
1059         TALLOC_FREE(bsds->fde);
1060         if (bsds->fd != -1) {
1061                 close(bsds->fd);
1062                 bsds->fd = -1;
1063         }
1064         return 0;
1065 }
1066
1067 static int tdgram_bsd_dgram_socket(const struct tsocket_address *local,
1068                                    const struct tsocket_address *remote,
1069                                    bool broadcast,
1070                                    TALLOC_CTX *mem_ctx,
1071                                    struct tdgram_context **_dgram,
1072                                    const char *location)
1073 {
1074         struct tsocket_address_bsd *lbsda =
1075                 talloc_get_type_abort(local->private_data,
1076                 struct tsocket_address_bsd);
1077         struct tsocket_address_bsd *rbsda = NULL;
1078         struct tdgram_context *dgram;
1079         struct tdgram_bsd *bsds;
1080         int fd;
1081         int ret;
1082         bool do_bind = false;
1083         bool do_reuseaddr = false;
1084         socklen_t sa_len = sizeof(lbsda->u.ss);
1085
1086         if (remote) {
1087                 rbsda = talloc_get_type_abort(remote->private_data,
1088                         struct tsocket_address_bsd);
1089         }
1090
1091         switch (lbsda->u.sa.sa_family) {
1092         case AF_UNIX:
1093                 if (broadcast) {
1094                         errno = EINVAL;
1095                         return -1;
1096                 }
1097                 if (lbsda->u.un.sun_path[0] != 0) {
1098                         do_reuseaddr = true;
1099                         do_bind = true;
1100                 }
1101                 /*
1102                  * for unix sockets we can't use the size of sockaddr_storage
1103                  * we would get EINVAL
1104                  */
1105                 sa_len = sizeof(lbsda->u.un);
1106                 break;
1107         case AF_INET:
1108                 if (lbsda->u.in.sin_port != 0) {
1109                         do_reuseaddr = true;
1110                         do_bind = true;
1111                 }
1112                 if (lbsda->u.in.sin_addr.s_addr == INADDR_ANY) {
1113                         do_bind = true;
1114                 }
1115                 break;
1116 #ifdef HAVE_IPV6
1117         case AF_INET6:
1118                 if (lbsda->u.in6.sin6_port != 0) {
1119                         do_reuseaddr = true;
1120                         do_bind = true;
1121                 }
1122                 if (memcmp(&in6addr_any,
1123                            &lbsda->u.in6.sin6_addr,
1124                            sizeof(in6addr_any)) != 0) {
1125                         do_bind = true;
1126                 }
1127                 break;
1128 #endif
1129         default:
1130                 errno = EINVAL;
1131                 return -1;
1132         }
1133
1134         fd = socket(lbsda->u.sa.sa_family, SOCK_DGRAM, 0);
1135         if (fd < 0) {
1136                 return fd;
1137         }
1138
1139         fd = tsocket_bsd_common_prepare_fd(fd, true);
1140         if (fd < 0) {
1141                 return fd;
1142         }
1143
1144         dgram = tdgram_context_create(mem_ctx,
1145                                       &tdgram_bsd_ops,
1146                                       &bsds,
1147                                       struct tdgram_bsd,
1148                                       location);
1149         if (!dgram) {
1150                 int saved_errno = errno;
1151                 close(fd);
1152                 errno = saved_errno;
1153                 return -1;
1154         }
1155         ZERO_STRUCTP(bsds);
1156         bsds->fd = fd;
1157         talloc_set_destructor(bsds, tdgram_bsd_destructor);
1158
1159         if (broadcast) {
1160                 int val = 1;
1161
1162                 ret = setsockopt(fd, SOL_SOCKET, SO_BROADCAST,
1163                                  (const void *)&val, sizeof(val));
1164                 if (ret == -1) {
1165                         int saved_errno = errno;
1166                         talloc_free(dgram);
1167                         errno = saved_errno;
1168                         return ret;
1169                 }
1170         }
1171
1172         if (do_reuseaddr) {
1173                 int val = 1;
1174
1175                 ret = setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
1176                                  (const void *)&val, sizeof(val));
1177                 if (ret == -1) {
1178                         int saved_errno = errno;
1179                         talloc_free(dgram);
1180                         errno = saved_errno;
1181                         return ret;
1182                 }
1183         }
1184
1185         if (do_bind) {
1186                 ret = bind(fd, &lbsda->u.sa, sa_len);
1187                 if (ret == -1) {
1188                         int saved_errno = errno;
1189                         talloc_free(dgram);
1190                         errno = saved_errno;
1191                         return ret;
1192                 }
1193         }
1194
1195         if (rbsda) {
1196                 ret = connect(fd, &rbsda->u.sa, sa_len);
1197                 if (ret == -1) {
1198                         int saved_errno = errno;
1199                         talloc_free(dgram);
1200                         errno = saved_errno;
1201                         return ret;
1202                 }
1203         }
1204
1205         *_dgram = dgram;
1206         return 0;
1207 }
1208
1209 int _tdgram_inet_udp_socket(const struct tsocket_address *local,
1210                             const struct tsocket_address *remote,
1211                             TALLOC_CTX *mem_ctx,
1212                             struct tdgram_context **dgram,
1213                             const char *location)
1214 {
1215         struct tsocket_address_bsd *lbsda =
1216                 talloc_get_type_abort(local->private_data,
1217                 struct tsocket_address_bsd);
1218         int ret;
1219
1220         switch (lbsda->u.sa.sa_family) {
1221         case AF_INET:
1222                 break;
1223 #ifdef HAVE_IPV6
1224         case AF_INET6:
1225                 break;
1226 #endif
1227         default:
1228                 errno = EINVAL;
1229                 return -1;
1230         }
1231
1232         ret = tdgram_bsd_dgram_socket(local, remote, false,
1233                                       mem_ctx, dgram, location);
1234
1235         return ret;
1236 }
1237
1238 int _tdgram_unix_socket(const struct tsocket_address *local,
1239                         const struct tsocket_address *remote,
1240                         TALLOC_CTX *mem_ctx,
1241                         struct tdgram_context **dgram,
1242                         const char *location)
1243 {
1244         struct tsocket_address_bsd *lbsda =
1245                 talloc_get_type_abort(local->private_data,
1246                 struct tsocket_address_bsd);
1247         int ret;
1248
1249         switch (lbsda->u.sa.sa_family) {
1250         case AF_UNIX:
1251                 break;
1252         default:
1253                 errno = EINVAL;
1254                 return -1;
1255         }
1256
1257         ret = tdgram_bsd_dgram_socket(local, remote, false,
1258                                       mem_ctx, dgram, location);
1259
1260         return ret;
1261 }
1262
1263 struct tstream_bsd {
1264         int fd;
1265
1266         void *event_ptr;
1267         struct tevent_fd *fde;
1268
1269         void *readable_private;
1270         void (*readable_handler)(void *private_data);
1271         void *writeable_private;
1272         void (*writeable_handler)(void *private_data);
1273 };
1274
1275 static void tstream_bsd_fde_handler(struct tevent_context *ev,
1276                                     struct tevent_fd *fde,
1277                                     uint16_t flags,
1278                                     void *private_data)
1279 {
1280         struct tstream_bsd *bsds = talloc_get_type_abort(private_data,
1281                                    struct tstream_bsd);
1282
1283         if (flags & TEVENT_FD_WRITE) {
1284                 bsds->writeable_handler(bsds->writeable_private);
1285                 return;
1286         }
1287         if (flags & TEVENT_FD_READ) {
1288                 if (!bsds->readable_handler) {
1289                         TEVENT_FD_NOT_READABLE(bsds->fde);
1290                         return;
1291                 }
1292                 bsds->readable_handler(bsds->readable_private);
1293                 return;
1294         }
1295 }
1296
1297 static int tstream_bsd_set_readable_handler(struct tstream_bsd *bsds,
1298                                             struct tevent_context *ev,
1299                                             void (*handler)(void *private_data),
1300                                             void *private_data)
1301 {
1302         if (ev == NULL) {
1303                 if (handler) {
1304                         errno = EINVAL;
1305                         return -1;
1306                 }
1307                 if (!bsds->readable_handler) {
1308                         return 0;
1309                 }
1310                 bsds->readable_handler = NULL;
1311                 bsds->readable_private = NULL;
1312
1313                 return 0;
1314         }
1315
1316         /* read and write must use the same tevent_context */
1317         if (bsds->event_ptr != ev) {
1318                 if (bsds->readable_handler || bsds->writeable_handler) {
1319                         errno = EINVAL;
1320                         return -1;
1321                 }
1322                 bsds->event_ptr = NULL;
1323                 TALLOC_FREE(bsds->fde);
1324         }
1325
1326         if (bsds->fde == NULL) {
1327                 bsds->fde = tevent_add_fd(ev, bsds,
1328                                           bsds->fd, TEVENT_FD_READ,
1329                                           tstream_bsd_fde_handler,
1330                                           bsds);
1331                 if (!bsds->fde) {
1332                         return -1;
1333                 }
1334
1335                 /* cache the event context we're running on */
1336                 bsds->event_ptr = ev;
1337         } else if (!bsds->readable_handler) {
1338                 TEVENT_FD_READABLE(bsds->fde);
1339         }
1340
1341         bsds->readable_handler = handler;
1342         bsds->readable_private = private_data;
1343
1344         return 0;
1345 }
1346
1347 static int tstream_bsd_set_writeable_handler(struct tstream_bsd *bsds,
1348                                              struct tevent_context *ev,
1349                                              void (*handler)(void *private_data),
1350                                              void *private_data)
1351 {
1352         if (ev == NULL) {
1353                 if (handler) {
1354                         errno = EINVAL;
1355                         return -1;
1356                 }
1357                 if (!bsds->writeable_handler) {
1358                         return 0;
1359                 }
1360                 bsds->writeable_handler = NULL;
1361                 bsds->writeable_private = NULL;
1362                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
1363
1364                 return 0;
1365         }
1366
1367         /* read and write must use the same tevent_context */
1368         if (bsds->event_ptr != ev) {
1369                 if (bsds->readable_handler || bsds->writeable_handler) {
1370                         errno = EINVAL;
1371                         return -1;
1372                 }
1373                 bsds->event_ptr = NULL;
1374                 TALLOC_FREE(bsds->fde);
1375         }
1376
1377         if (bsds->fde == NULL) {
1378                 bsds->fde = tevent_add_fd(ev, bsds,
1379                                           bsds->fd, TEVENT_FD_WRITE,
1380                                           tstream_bsd_fde_handler,
1381                                           bsds);
1382                 if (!bsds->fde) {
1383                         return -1;
1384                 }
1385
1386                 /* cache the event context we're running on */
1387                 bsds->event_ptr = ev;
1388         } else if (!bsds->writeable_handler) {
1389                 TEVENT_FD_WRITEABLE(bsds->fde);
1390         }
1391
1392         bsds->writeable_handler = handler;
1393         bsds->writeable_private = private_data;
1394
1395         return 0;
1396 }
1397
1398 static ssize_t tstream_bsd_pending_bytes(struct tstream_context *stream)
1399 {
1400         struct tstream_bsd *bsds = tstream_context_data(stream,
1401                                    struct tstream_bsd);
1402         ssize_t ret;
1403
1404         if (bsds->fd == -1) {
1405                 errno = ENOTCONN;
1406                 return -1;
1407         }
1408
1409         ret = tsocket_bsd_pending(bsds->fd);
1410
1411         return ret;
1412 }
1413
1414 struct tstream_bsd_readv_state {
1415         struct tstream_context *stream;
1416
1417         struct iovec *vector;
1418         size_t count;
1419
1420         int ret;
1421 };
1422
1423 static int tstream_bsd_readv_destructor(struct tstream_bsd_readv_state *state)
1424 {
1425         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1426                                    struct tstream_bsd);
1427
1428         tstream_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
1429
1430         return 0;
1431 }
1432
1433 static void tstream_bsd_readv_handler(void *private_data);
1434
1435 static struct tevent_req *tstream_bsd_readv_send(TALLOC_CTX *mem_ctx,
1436                                         struct tevent_context *ev,
1437                                         struct tstream_context *stream,
1438                                         struct iovec *vector,
1439                                         size_t count)
1440 {
1441         struct tevent_req *req;
1442         struct tstream_bsd_readv_state *state;
1443         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1444         int ret;
1445
1446         req = tevent_req_create(mem_ctx, &state,
1447                                 struct tstream_bsd_readv_state);
1448         if (!req) {
1449                 return NULL;
1450         }
1451
1452         state->stream   = stream;
1453         /* we make a copy of the vector so that we can modify it */
1454         state->vector   = talloc_array(state, struct iovec, count);
1455         if (tevent_req_nomem(state->vector, req)) {
1456                 goto post;
1457         }
1458         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1459         state->count    = count;
1460         state->ret      = 0;
1461
1462         talloc_set_destructor(state, tstream_bsd_readv_destructor);
1463
1464         if (bsds->fd == -1) {
1465                 tevent_req_error(req, ENOTCONN);
1466                 goto post;
1467         }
1468
1469         /*
1470          * this is a fast path, not waiting for the
1471          * socket to become explicit readable gains
1472          * about 10%-20% performance in benchmark tests.
1473          */
1474         tstream_bsd_readv_handler(req);
1475         if (!tevent_req_is_in_progress(req)) {
1476                 goto post;
1477         }
1478
1479         ret = tstream_bsd_set_readable_handler(bsds, ev,
1480                                               tstream_bsd_readv_handler,
1481                                               req);
1482         if (ret == -1) {
1483                 tevent_req_error(req, errno);
1484                 goto post;
1485         }
1486
1487         return req;
1488
1489  post:
1490         tevent_req_post(req, ev);
1491         return req;
1492 }
1493
1494 static void tstream_bsd_readv_handler(void *private_data)
1495 {
1496         struct tevent_req *req = talloc_get_type_abort(private_data,
1497                                  struct tevent_req);
1498         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1499                                         struct tstream_bsd_readv_state);
1500         struct tstream_context *stream = state->stream;
1501         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1502         int ret;
1503         int err;
1504         bool retry;
1505
1506         ret = readv(bsds->fd, state->vector, state->count);
1507         if (ret == 0) {
1508                 /* propagate end of file */
1509                 tevent_req_error(req, EPIPE);
1510                 return;
1511         }
1512         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1513         if (retry) {
1514                 /* retry later */
1515                 return;
1516         }
1517         if (tevent_req_error(req, err)) {
1518                 return;
1519         }
1520
1521         state->ret += ret;
1522
1523         while (ret > 0) {
1524                 if (ret < state->vector[0].iov_len) {
1525                         uint8_t *base;
1526                         base = (uint8_t *)state->vector[0].iov_base;
1527                         base += ret;
1528                         state->vector[0].iov_base = base;
1529                         state->vector[0].iov_len -= ret;
1530                         break;
1531                 }
1532                 ret -= state->vector[0].iov_len;
1533                 state->vector += 1;
1534                 state->count -= 1;
1535         }
1536
1537         if (state->count > 0) {
1538                 /* we have more to read */
1539                 return;
1540         }
1541
1542         tevent_req_done(req);
1543 }
1544
1545 static int tstream_bsd_readv_recv(struct tevent_req *req,
1546                                   int *perrno)
1547 {
1548         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1549                                         struct tstream_bsd_readv_state);
1550         int ret;
1551
1552         ret = tsocket_simple_int_recv(req, perrno);
1553         if (ret == 0) {
1554                 ret = state->ret;
1555         }
1556
1557         tevent_req_received(req);
1558         return ret;
1559 }
1560
1561 struct tstream_bsd_writev_state {
1562         struct tstream_context *stream;
1563
1564         struct iovec *vector;
1565         size_t count;
1566
1567         int ret;
1568 };
1569
1570 static int tstream_bsd_writev_destructor(struct tstream_bsd_writev_state *state)
1571 {
1572         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1573                                   struct tstream_bsd);
1574
1575         tstream_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
1576
1577         return 0;
1578 }
1579
1580 static void tstream_bsd_writev_handler(void *private_data);
1581
1582 static struct tevent_req *tstream_bsd_writev_send(TALLOC_CTX *mem_ctx,
1583                                                  struct tevent_context *ev,
1584                                                  struct tstream_context *stream,
1585                                                  const struct iovec *vector,
1586                                                  size_t count)
1587 {
1588         struct tevent_req *req;
1589         struct tstream_bsd_writev_state *state;
1590         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1591         int ret;
1592
1593         req = tevent_req_create(mem_ctx, &state,
1594                                 struct tstream_bsd_writev_state);
1595         if (!req) {
1596                 return NULL;
1597         }
1598
1599         state->stream   = stream;
1600         /* we make a copy of the vector so that we can modify it */
1601         state->vector   = talloc_array(state, struct iovec, count);
1602         if (tevent_req_nomem(state->vector, req)) {
1603                 goto post;
1604         }
1605         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1606         state->count    = count;
1607         state->ret      = 0;
1608
1609         talloc_set_destructor(state, tstream_bsd_writev_destructor);
1610
1611         if (bsds->fd == -1) {
1612                 tevent_req_error(req, ENOTCONN);
1613                 goto post;
1614         }
1615
1616         /*
1617          * this is a fast path, not waiting for the
1618          * socket to become explicit writeable gains
1619          * about 10%-20% performance in benchmark tests.
1620          */
1621         tstream_bsd_writev_handler(req);
1622         if (!tevent_req_is_in_progress(req)) {
1623                 goto post;
1624         }
1625
1626         ret = tstream_bsd_set_writeable_handler(bsds, ev,
1627                                                tstream_bsd_writev_handler,
1628                                                req);
1629         if (ret == -1) {
1630                 tevent_req_error(req, errno);
1631                 goto post;
1632         }
1633
1634         return req;
1635
1636  post:
1637         tevent_req_post(req, ev);
1638         return req;
1639 }
1640
1641 static void tstream_bsd_writev_handler(void *private_data)
1642 {
1643         struct tevent_req *req = talloc_get_type_abort(private_data,
1644                                  struct tevent_req);
1645         struct tstream_bsd_writev_state *state = tevent_req_data(req,
1646                                         struct tstream_bsd_writev_state);
1647         struct tstream_context *stream = state->stream;
1648         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1649         ssize_t ret;
1650         int err;
1651         bool retry;
1652
1653         ret = writev(bsds->fd, state->vector, state->count);
1654         if (ret == 0) {
1655                 /* propagate end of file */
1656                 tevent_req_error(req, EPIPE);
1657                 return;
1658         }
1659         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1660         if (retry) {
1661                 /* retry later */
1662                 return;
1663         }
1664         if (tevent_req_error(req, err)) {
1665                 return;
1666         }
1667
1668         state->ret += ret;
1669
1670         while (ret > 0) {
1671                 if (ret < state->vector[0].iov_len) {
1672                         uint8_t *base;
1673                         base = (uint8_t *)state->vector[0].iov_base;
1674                         base += ret;
1675                         state->vector[0].iov_base = base;
1676                         state->vector[0].iov_len -= ret;
1677                         break;
1678                 }
1679                 ret -= state->vector[0].iov_len;
1680                 state->vector += 1;
1681                 state->count -= 1;
1682         }
1683
1684         if (state->count > 0) {
1685                 /* we have more to read */
1686                 return;
1687         }
1688
1689         tevent_req_done(req);
1690 }
1691
1692 static int tstream_bsd_writev_recv(struct tevent_req *req, int *perrno)
1693 {
1694         struct tstream_bsd_writev_state *state = tevent_req_data(req,
1695                                         struct tstream_bsd_writev_state);
1696         int ret;
1697
1698         ret = tsocket_simple_int_recv(req, perrno);
1699         if (ret == 0) {
1700                 ret = state->ret;
1701         }
1702
1703         tevent_req_received(req);
1704         return ret;
1705 }
1706
1707 struct tstream_bsd_disconnect_state {
1708         void *__dummy;
1709 };
1710
1711 static struct tevent_req *tstream_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
1712                                                      struct tevent_context *ev,
1713                                                      struct tstream_context *stream)
1714 {
1715         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1716         struct tevent_req *req;
1717         struct tstream_bsd_disconnect_state *state;
1718         int ret;
1719         int err;
1720         bool dummy;
1721
1722         req = tevent_req_create(mem_ctx, &state,
1723                                 struct tstream_bsd_disconnect_state);
1724         if (req == NULL) {
1725                 return NULL;
1726         }
1727
1728         if (bsds->fd == -1) {
1729                 tevent_req_error(req, ENOTCONN);
1730                 goto post;
1731         }
1732
1733         ret = close(bsds->fd);
1734         bsds->fd = -1;
1735         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
1736         if (tevent_req_error(req, err)) {
1737                 goto post;
1738         }
1739
1740         tevent_req_done(req);
1741 post:
1742         tevent_req_post(req, ev);
1743         return req;
1744 }
1745
1746 static int tstream_bsd_disconnect_recv(struct tevent_req *req,
1747                                       int *perrno)
1748 {
1749         int ret;
1750
1751         ret = tsocket_simple_int_recv(req, perrno);
1752
1753         tevent_req_received(req);
1754         return ret;
1755 }
1756
1757 static const struct tstream_context_ops tstream_bsd_ops = {
1758         .name                   = "bsd",
1759
1760         .pending_bytes          = tstream_bsd_pending_bytes,
1761
1762         .readv_send             = tstream_bsd_readv_send,
1763         .readv_recv             = tstream_bsd_readv_recv,
1764
1765         .writev_send            = tstream_bsd_writev_send,
1766         .writev_recv            = tstream_bsd_writev_recv,
1767
1768         .disconnect_send        = tstream_bsd_disconnect_send,
1769         .disconnect_recv        = tstream_bsd_disconnect_recv,
1770 };
1771
1772 static int tstream_bsd_destructor(struct tstream_bsd *bsds)
1773 {
1774         TALLOC_FREE(bsds->fde);
1775         if (bsds->fd != -1) {
1776                 close(bsds->fd);
1777                 bsds->fd = -1;
1778         }
1779         return 0;
1780 }
1781
1782 int _tstream_bsd_existing_socket(TALLOC_CTX *mem_ctx,
1783                                  int fd,
1784                                  struct tstream_context **_stream,
1785                                  const char *location)
1786 {
1787         struct tstream_context *stream;
1788         struct tstream_bsd *bsds;
1789
1790         stream = tstream_context_create(mem_ctx,
1791                                         &tstream_bsd_ops,
1792                                         &bsds,
1793                                         struct tstream_bsd,
1794                                         location);
1795         if (!stream) {
1796                 return -1;
1797         }
1798         ZERO_STRUCTP(bsds);
1799         bsds->fd = fd;
1800         talloc_set_destructor(bsds, tstream_bsd_destructor);
1801
1802         *_stream = stream;
1803         return 0;
1804 }
1805
1806 struct tstream_bsd_connect_state {
1807         int fd;
1808         struct tevent_fd *fde;
1809         struct tstream_conext *stream;
1810 };
1811
1812 static int tstream_bsd_connect_destructor(struct tstream_bsd_connect_state *state)
1813 {
1814         TALLOC_FREE(state->fde);
1815         if (state->fd != -1) {
1816                 close(state->fd);
1817                 state->fd = -1;
1818         }
1819
1820         return 0;
1821 }
1822
1823 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
1824                                             struct tevent_fd *fde,
1825                                             uint16_t flags,
1826                                             void *private_data);
1827
1828 static struct tevent_req * tstream_bsd_connect_send(TALLOC_CTX *mem_ctx,
1829                                         struct tevent_context *ev,
1830                                         int sys_errno,
1831                                         const struct tsocket_address *local,
1832                                         const struct tsocket_address *remote)
1833 {
1834         struct tevent_req *req;
1835         struct tstream_bsd_connect_state *state;
1836         struct tsocket_address_bsd *lbsda =
1837                 talloc_get_type_abort(local->private_data,
1838                 struct tsocket_address_bsd);
1839         struct tsocket_address_bsd *rbsda =
1840                 talloc_get_type_abort(remote->private_data,
1841                 struct tsocket_address_bsd);
1842         int ret;
1843         int err;
1844         bool retry;
1845         bool do_bind = false;
1846         bool do_reuseaddr = false;
1847         socklen_t sa_len = sizeof(rbsda->u.ss);
1848
1849         req = tevent_req_create(mem_ctx, &state,
1850                                 struct tstream_bsd_connect_state);
1851         if (!req) {
1852                 return NULL;
1853         }
1854         state->fd = -1;
1855         state->fde = NULL;
1856
1857         talloc_set_destructor(state, tstream_bsd_connect_destructor);
1858
1859         /* give the wrappers a chance to report an error */
1860         if (sys_errno != 0) {
1861                 tevent_req_error(req, sys_errno);
1862                 goto post;
1863         }
1864
1865         switch (lbsda->u.sa.sa_family) {
1866         case AF_UNIX:
1867                 if (lbsda->u.un.sun_path[0] != 0) {
1868                         do_reuseaddr = true;
1869                         do_bind = true;
1870                 }
1871                 /*
1872                  * for unix sockets we can't use the size of sockaddr_storage
1873                  * we would get EINVAL
1874                  */
1875                 sa_len = sizeof(rbsda->u.un);
1876                 break;
1877         case AF_INET:
1878                 if (lbsda->u.in.sin_port != 0) {
1879                         do_reuseaddr = true;
1880                         do_bind = true;
1881                 }
1882                 if (lbsda->u.in.sin_addr.s_addr == INADDR_ANY) {
1883                         do_bind = true;
1884                 }
1885                 break;
1886 #ifdef HAVE_IPV6
1887         case AF_INET6:
1888                 if (lbsda->u.in6.sin6_port != 0) {
1889                         do_reuseaddr = true;
1890                         do_bind = true;
1891                 }
1892                 if (memcmp(&in6addr_any,
1893                            &lbsda->u.in6.sin6_addr,
1894                            sizeof(in6addr_any)) != 0) {
1895                         do_bind = true;
1896                 }
1897                 break;
1898 #endif
1899         default:
1900                 tevent_req_error(req, EINVAL);
1901                 goto post;
1902         }
1903
1904         state->fd = socket(lbsda->u.sa.sa_family, SOCK_STREAM, 0);
1905         if (state->fd == -1) {
1906                 tevent_req_error(req, errno);
1907                 goto post;
1908         }
1909
1910         state->fd = tsocket_bsd_common_prepare_fd(state->fd, true);
1911         if (state->fd == -1) {
1912                 tevent_req_error(req, errno);
1913                 goto post;
1914         }
1915
1916         if (do_reuseaddr) {
1917                 int val = 1;
1918
1919                 ret = setsockopt(state->fd, SOL_SOCKET, SO_REUSEADDR,
1920                                  (const void *)&val, sizeof(val));
1921                 if (ret == -1) {
1922                         tevent_req_error(req, errno);
1923                         goto post;
1924                 }
1925         }
1926
1927         if (do_bind) {
1928                 ret = bind(state->fd, &lbsda->u.sa, sizeof(lbsda->u.ss));
1929                 if (ret == -1) {
1930                         tevent_req_error(req, errno);
1931                         goto post;
1932                 }
1933         }
1934
1935         ret = connect(state->fd, &rbsda->u.sa, sa_len);
1936         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1937         if (retry) {
1938                 /* retry later */
1939                 goto async;
1940         }
1941         if (tevent_req_error(req, err)) {
1942                 goto post;
1943         }
1944
1945         tevent_req_done(req);
1946         goto post;
1947
1948  async:
1949         state->fde = tevent_add_fd(ev, state,
1950                                    state->fd,
1951                                    TEVENT_FD_READ | TEVENT_FD_WRITE,
1952                                    tstream_bsd_connect_fde_handler,
1953                                    req);
1954         if (tevent_req_nomem(state->fde, req)) {
1955                 goto post;
1956         }
1957
1958         return req;
1959
1960  post:
1961         tevent_req_post(req, ev);
1962         return req;
1963 }
1964
1965 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
1966                                             struct tevent_fd *fde,
1967                                             uint16_t flags,
1968                                             void *private_data)
1969 {
1970         struct tevent_req *req = talloc_get_type_abort(private_data,
1971                                  struct tevent_req);
1972         struct tstream_bsd_connect_state *state = tevent_req_data(req,
1973                                         struct tstream_bsd_connect_state);
1974         int ret;
1975         int error=0;
1976         socklen_t len = sizeof(error);
1977         int err;
1978         bool retry;
1979
1980         ret = getsockopt(state->fd, SOL_SOCKET, SO_ERROR, &error, &len);
1981         if (ret == 0) {
1982                 if (error != 0) {
1983                         errno = error;
1984                         ret = -1;
1985                 }
1986         }
1987         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1988         if (retry) {
1989                 /* retry later */
1990                 return;
1991         }
1992         if (tevent_req_error(req, err)) {
1993                 return;
1994         }
1995
1996         tevent_req_done(req);
1997 }
1998
1999 static int tstream_bsd_connect_recv(struct tevent_req *req,
2000                                     int *perrno,
2001                                     TALLOC_CTX *mem_ctx,
2002                                     struct tstream_context **stream,
2003                                     const char *location)
2004 {
2005         struct tstream_bsd_connect_state *state = tevent_req_data(req,
2006                                         struct tstream_bsd_connect_state);
2007         int ret;
2008
2009         ret = tsocket_simple_int_recv(req, perrno);
2010         if (ret == 0) {
2011                 ret = _tstream_bsd_existing_socket(mem_ctx,
2012                                                    state->fd,
2013                                                    stream,
2014                                                    location);
2015                 if (ret == -1) {
2016                         *perrno = errno;
2017                         goto done;
2018                 }
2019                 TALLOC_FREE(state->fde);
2020                 state->fd = -1;
2021         }
2022
2023 done:
2024         tevent_req_received(req);
2025         return ret;
2026 }
2027
2028 struct tevent_req * tstream_inet_tcp_connect_send(TALLOC_CTX *mem_ctx,
2029                                         struct tevent_context *ev,
2030                                         const struct tsocket_address *local,
2031                                         const struct tsocket_address *remote)
2032 {
2033         struct tsocket_address_bsd *lbsda =
2034                 talloc_get_type_abort(local->private_data,
2035                 struct tsocket_address_bsd);
2036         struct tevent_req *req;
2037         int sys_errno = 0;
2038
2039         switch (lbsda->u.sa.sa_family) {
2040         case AF_INET:
2041                 break;
2042 #ifdef HAVE_IPV6
2043         case AF_INET6:
2044                 break;
2045 #endif
2046         default:
2047                 sys_errno = EINVAL;
2048                 break;
2049         }
2050
2051         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2052
2053         return req;
2054 }
2055
2056 int _tstream_inet_tcp_connect_recv(struct tevent_req *req,
2057                                    int *perrno,
2058                                    TALLOC_CTX *mem_ctx,
2059                                    struct tstream_context **stream,
2060                                    const char *location)
2061 {
2062         return tstream_bsd_connect_recv(req, perrno, mem_ctx, stream, location);
2063 }
2064
2065 struct tevent_req * tstream_unix_connect_send(TALLOC_CTX *mem_ctx,
2066                                         struct tevent_context *ev,
2067                                         const struct tsocket_address *local,
2068                                         const struct tsocket_address *remote)
2069 {
2070         struct tsocket_address_bsd *lbsda =
2071                 talloc_get_type_abort(local->private_data,
2072                 struct tsocket_address_bsd);
2073         struct tevent_req *req;
2074         int sys_errno = 0;
2075
2076         switch (lbsda->u.sa.sa_family) {
2077         case AF_UNIX:
2078                 break;
2079         default:
2080                 sys_errno = EINVAL;
2081                 break;
2082         }
2083
2084         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2085
2086         return req;
2087 }
2088
2089 int _tstream_unix_connect_recv(struct tevent_req *req,
2090                                       int *perrno,
2091                                       TALLOC_CTX *mem_ctx,
2092                                       struct tstream_context **stream,
2093                                       const char *location)
2094 {
2095         return tstream_bsd_connect_recv(req, perrno, mem_ctx, stream, location);
2096 }
2097
2098 int _tstream_unix_socketpair(TALLOC_CTX *mem_ctx1,
2099                              struct tstream_context **_stream1,
2100                              TALLOC_CTX *mem_ctx2,
2101                              struct tstream_context **_stream2,
2102                              const char *location)
2103 {
2104         int ret;
2105         int fds[2];
2106         int fd1;
2107         int fd2;
2108         struct tstream_context *stream1 = NULL;
2109         struct tstream_context *stream2 = NULL;
2110
2111         ret = socketpair(AF_UNIX, SOCK_STREAM, 0, fds);
2112         if (ret == -1) {
2113                 return -1;
2114         }
2115         fd1 = fds[0];
2116         fd2 = fds[1];
2117
2118         fd1 = tsocket_bsd_common_prepare_fd(fd1, true);
2119         if (fd1 == -1) {
2120                 int sys_errno = errno;
2121                 close(fd2);
2122                 errno = sys_errno;
2123                 return -1;
2124         }
2125
2126         fd2 = tsocket_bsd_common_prepare_fd(fd2, true);
2127         if (fd2 == -1) {
2128                 int sys_errno = errno;
2129                 close(fd1);
2130                 errno = sys_errno;
2131                 return -1;
2132         }
2133
2134         ret = _tstream_bsd_existing_socket(mem_ctx1,
2135                                            fd1,
2136                                            &stream1,
2137                                            location);
2138         if (ret == -1) {
2139                 int sys_errno = errno;
2140                 close(fd1);
2141                 close(fd2);
2142                 errno = sys_errno;
2143                 return -1;
2144         }
2145
2146         ret = _tstream_bsd_existing_socket(mem_ctx2,
2147                                            fd2,
2148                                            &stream2,
2149                                            location);
2150         if (ret == -1) {
2151                 int sys_errno = errno;
2152                 talloc_free(stream1);
2153                 close(fd2);
2154                 errno = sys_errno;
2155                 return -1;
2156         }
2157
2158         *_stream1 = stream1;
2159         *_stream2 = stream2;
2160         return 0;
2161 }
2162