libatalk/tsocket/tsocket_bsd.c

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