lib/tsocket/tsocket_bsd.c

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