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 #include "lib/util/iov_buf.h"
  30 #include "lib/util/blocking.h"
  31 #include "lib/util/util_net.h"
  32
  33 static int tsocket_bsd_error_from_errno(int ret,
  34                                         int sys_errno,
  35                                         bool *retry)
  36 {
  37         *retry = false;
  38
  39         if (ret >= 0) {
  40                 return 0;
  41         }
  42
  43         if (ret != -1) {
  44                 return EIO;
  45         }
  46
  47         if (sys_errno == 0) {
  48                 return EIO;
  49         }
  50
  51         if (sys_errno == EINTR) {
  52                 *retry = true;
  53                 return sys_errno;
  54         }
  55
  56         if (sys_errno == EINPROGRESS) {
  57                 *retry = true;
  58                 return sys_errno;
  59         }
  60
  61         if (sys_errno == EAGAIN) {
  62                 *retry = true;
  63                 return sys_errno;
  64         }
  65
  66         /* ENOMEM is retryable on Solaris/illumos, and possibly other systems. */
  67         if (sys_errno == ENOMEM) {
  68                 *retry = true;
  69                 return sys_errno;
  70         }
  71
  72 #ifdef EWOULDBLOCK
  73         if (sys_errno == EWOULDBLOCK) {
  74                 *retry = true;
  75                 return sys_errno;
  76         }
  77 #endif
  78
  79         return sys_errno;
  80 }
  81
  82 static int tsocket_bsd_common_prepare_fd(int fd, bool high_fd)
  83 {
  84         int i;
  85         int sys_errno = 0;
  86         int fds[3];
  87         int num_fds = 0;
  88
  89         int result;
  90         bool ok;
  91
  92         if (fd == -1) {
  93                 return -1;
  94         }
  95
  96         /* first make a fd >= 3 */
  97         if (high_fd) {
  98                 while (fd < 3) {
  99                         fds[num_fds++] = fd;
 100                         fd = dup(fd);
 101                         if (fd == -1) {
 102                                 sys_errno = errno;
 103                                 break;
 104                         }
 105                 }
 106                 for (i=0; i<num_fds; i++) {
 107                         close(fds[i]);
 108                 }
 109                 if (fd == -1) {
 110                         errno = sys_errno;
 111                         return fd;
 112                 }
 113         }
 114
 115         result = set_blocking(fd, false);
 116         if (result == -1) {
 117                 goto fail;
 118         }
 119
 120         ok = smb_set_close_on_exec(fd);
 121         if (!ok) {
 122                 goto fail;
 123         }
 124
 125         return fd;
 126
 127  fail:
 128         if (fd != -1) {
 129                 sys_errno = errno;
 130                 close(fd);
 131                 errno = sys_errno;
 132         }
 133         return -1;
 134 }
 135
 136 #ifdef HAVE_LINUX_RTNETLINK_H
 137 /**
 138  * Get the amount of pending bytes from a netlink socket
 139  *
 140  * For some reason netlink sockets don't support querying the amount of pending
 141  * data via ioctl with FIONREAD, which is what we use in tsocket_bsd_pending()
 142  * below.
 143  *
 144  * We know we are on Linux as we're using netlink, which means we have a working
 145  * MSG_TRUNC flag to recvmsg() as well, so we use that together with MSG_PEEK.
 146  **/
 147 static ssize_t tsocket_bsd_netlink_pending(int fd)
 148 {
 149         struct iovec iov;
 150         struct msghdr msg;
 151         char buf[1];
 152
 153         iov = (struct iovec) {
 154                 .iov_base = buf,
 155                 .iov_len = sizeof(buf)
 156         };
 157
 158         msg = (struct msghdr) {
 159                 .msg_iov = &iov,
 160                 .msg_iovlen = 1
 161         };
 162
 163         return recvmsg(fd, &msg, MSG_PEEK | MSG_TRUNC);
 164 }
 165 #else
 166 static ssize_t tsocket_bsd_netlink_pending(int fd)
 167 {
 168         errno = ENOSYS;
 169         return -1;
 170 }
 171 #endif
 172
 173 static ssize_t tsocket_bsd_pending(int fd)
 174 {
 175         int ret, error;
 176         int value = 0;
 177         socklen_t len;
 178
 179         ret = ioctl(fd, FIONREAD, &value);
 180         if (ret == -1) {
 181                 return ret;
 182         }
 183
 184         if (ret != 0) {
 185                 /* this should not be reached */
 186                 errno = EIO;
 187                 return -1;
 188         }
 189
 190         if (value != 0) {
 191                 return value;
 192         }
 193
 194         error = 0;
 195         len = sizeof(error);
 196
 197         /*
 198          * if no data is available check if the socket is in error state. For
 199          * dgram sockets it's the way to return ICMP error messages of
 200          * connected sockets to the caller.
 201          */
 202         ret = getsockopt(fd, SOL_SOCKET, SO_ERROR, &error, &len);
 203         if (ret == -1) {
 204                 return ret;
 205         }
 206         if (error != 0) {
 207                 errno = error;
 208                 return -1;
 209         }
 210         return 0;
 211 }
 212
 213 static const struct tsocket_address_ops tsocket_address_bsd_ops;
 214
 215 int _tsocket_address_bsd_from_sockaddr(TALLOC_CTX *mem_ctx,
 216                                        const struct sockaddr *sa,
 217                                        size_t sa_socklen,
 218                                        struct tsocket_address **_addr,
 219                                        const char *location)
 220 {
 221         struct tsocket_address *addr;
 222         struct samba_sockaddr *bsda = NULL;
 223
 224         if (sa_socklen < sizeof(sa->sa_family)) {
 225                 errno = EINVAL;
 226                 return -1;
 227         }
 228
 229         switch (sa->sa_family) {
 230         case AF_UNIX:
 231                 if (sa_socklen > sizeof(struct sockaddr_un)) {
 232                         sa_socklen = sizeof(struct sockaddr_un);
 233                 }
 234                 break;
 235         case AF_INET:
 236                 if (sa_socklen < sizeof(struct sockaddr_in)) {
 237                         errno = EINVAL;
 238                         return -1;
 239                 }
 240                 sa_socklen = sizeof(struct sockaddr_in);
 241                 break;
 242 #ifdef HAVE_IPV6
 243         case AF_INET6:
 244                 if (sa_socklen < sizeof(struct sockaddr_in6)) {
 245                         errno = EINVAL;
 246                         return -1;
 247                 }
 248                 sa_socklen = sizeof(struct sockaddr_in6);
 249                 break;
 250 #endif
 251         default:
 252                 errno = EAFNOSUPPORT;
 253                 return -1;
 254         }
 255
 256         if (sa_socklen > sizeof(struct sockaddr_storage)) {
 257                 errno = EINVAL;
 258                 return -1;
 259         }
 260
 261         addr = tsocket_address_create(mem_ctx,
 262                                       &tsocket_address_bsd_ops,
 263                                       &bsda,
 264                                       struct samba_sockaddr,
 265                                       location);
 266         if (!addr) {
 267                 errno = ENOMEM;
 268                 return -1;
 269         }
 270
 271         ZERO_STRUCTP(bsda);
 272
 273         memcpy(&bsda->u.ss, sa, sa_socklen);
 274
 275         bsda->sa_socklen = sa_socklen;
 276 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 277         bsda->u.sa.sa_len = bsda->sa_socklen;
 278 #endif
 279
 280         *_addr = addr;
 281         return 0;
 282 }
 283
 284 int _tsocket_address_bsd_from_samba_sockaddr(TALLOC_CTX *mem_ctx,
 285                                          const struct samba_sockaddr *xs_addr,
 286                                          struct tsocket_address **t_addr,
 287                                          const char *location)
 288 {
 289         return _tsocket_address_bsd_from_sockaddr(mem_ctx,
 290                                                   &xs_addr->u.sa,
 291                                                   xs_addr->sa_socklen,
 292                                                   t_addr,
 293                                                   location);
 294 }
 295
 296 ssize_t tsocket_address_bsd_sockaddr(const struct tsocket_address *addr,
 297                                      struct sockaddr *sa,
 298                                      size_t sa_socklen)
 299 {
 300         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 301                                            struct samba_sockaddr);
 302
 303         if (!bsda) {
 304                 errno = EINVAL;
 305                 return -1;
 306         }
 307
 308         if (sa_socklen < bsda->sa_socklen) {
 309                 errno = EINVAL;
 310                 return -1;
 311         }
 312
 313         if (sa_socklen > bsda->sa_socklen) {
 314                 memset(sa, 0, sa_socklen);
 315                 sa_socklen = bsda->sa_socklen;
 316         }
 317
 318         memcpy(sa, &bsda->u.ss, sa_socklen);
 319 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 320         sa->sa_len = sa_socklen;
 321 #endif
 322         return sa_socklen;
 323 }
 324
 325 bool tsocket_address_is_inet(const struct tsocket_address *addr, const char *fam)
 326 {
 327         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 328                                            struct samba_sockaddr);
 329
 330         if (!bsda) {
 331                 return false;
 332         }
 333
 334         switch (bsda->u.sa.sa_family) {
 335         case AF_INET:
 336                 if (strcasecmp(fam, "ip") == 0) {
 337                         return true;
 338                 }
 339
 340                 if (strcasecmp(fam, "ipv4") == 0) {
 341                         return true;
 342                 }
 343
 344                 return false;
 345 #ifdef HAVE_IPV6
 346         case AF_INET6:
 347                 if (strcasecmp(fam, "ip") == 0) {
 348                         return true;
 349                 }
 350
 351                 if (strcasecmp(fam, "ipv6") == 0) {
 352                         return true;
 353                 }
 354
 355                 return false;
 356 #endif
 357         }
 358
 359         return false;
 360 }
 361
 362 int _tsocket_address_inet_from_strings(TALLOC_CTX *mem_ctx,
 363                                        const char *fam,
 364                                        const char *addr,
 365                                        uint16_t port,
 366                                        struct tsocket_address **_addr,
 367                                        const char *location)
 368 {
 369         struct addrinfo hints;
 370         struct addrinfo *result = NULL;
 371         char port_str[6];
 372         int ret;
 373
 374         ZERO_STRUCT(hints);
 375         /*
 376          * we use SOCKET_STREAM here to get just one result
 377          * back from getaddrinfo().
 378          */
 379         hints.ai_socktype = SOCK_STREAM;
 380         hints.ai_flags = AI_NUMERICHOST | AI_NUMERICSERV;
 381
 382         if (strcasecmp(fam, "ip") == 0) {
 383                 hints.ai_family = AF_UNSPEC;
 384                 if (!addr) {
 385 #ifdef HAVE_IPV6
 386                         addr = "::";
 387 #else
 388                         addr = "0.0.0.0";
 389 #endif
 390                 }
 391         } else if (strcasecmp(fam, "ipv4") == 0) {
 392                 hints.ai_family = AF_INET;
 393                 if (!addr) {
 394                         addr = "0.0.0.0";
 395                 }
 396 #ifdef HAVE_IPV6
 397         } else if (strcasecmp(fam, "ipv6") == 0) {
 398                 hints.ai_family = AF_INET6;
 399                 if (!addr) {
 400                         addr = "::";
 401                 }
 402 #endif
 403         } else {
 404                 errno = EAFNOSUPPORT;
 405                 return -1;
 406         }
 407
 408         snprintf(port_str, sizeof(port_str), "%u", port);
 409
 410         ret = getaddrinfo(addr, port_str, &hints, &result);
 411         if (ret != 0) {
 412                 switch (ret) {
 413                 case EAI_FAIL:
 414                         errno = EINVAL;
 415                         break;
 416                 }
 417                 ret = -1;
 418                 goto done;
 419         }
 420
 421         if (result->ai_socktype != SOCK_STREAM) {
 422                 errno = EINVAL;
 423                 ret = -1;
 424                 goto done;
 425         }
 426
 427         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 428                                                   result->ai_addr,
 429                                                   result->ai_addrlen,
 430                                                   _addr,
 431                                                   location);
 432
 433 done:
 434         if (result) {
 435                 freeaddrinfo(result);
 436         }
 437         return ret;
 438 }
 439
 440 char *tsocket_address_inet_addr_string(const struct tsocket_address *addr,
 441                                        TALLOC_CTX *mem_ctx)
 442 {
 443         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 444                                            struct samba_sockaddr);
 445         char addr_str[INET6_ADDRSTRLEN+1];
 446         const char *str;
 447
 448         if (!bsda) {
 449                 errno = EINVAL;
 450                 return NULL;
 451         }
 452
 453         switch (bsda->u.sa.sa_family) {
 454         case AF_INET:
 455                 str = inet_ntop(bsda->u.in.sin_family,
 456                                 &bsda->u.in.sin_addr,
 457                                 addr_str, sizeof(addr_str));
 458                 break;
 459 #ifdef HAVE_IPV6
 460         case AF_INET6:
 461                 str = inet_ntop(bsda->u.in6.sin6_family,
 462                                 &bsda->u.in6.sin6_addr,
 463                                 addr_str, sizeof(addr_str));
 464                 break;
 465 #endif
 466         default:
 467                 errno = EINVAL;
 468                 return NULL;
 469         }
 470
 471         if (!str) {
 472                 return NULL;
 473         }
 474
 475         return talloc_strdup(mem_ctx, str);
 476 }
 477
 478 uint16_t tsocket_address_inet_port(const struct tsocket_address *addr)
 479 {
 480         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 481                                            struct samba_sockaddr);
 482         uint16_t port = 0;
 483
 484         if (!bsda) {
 485                 errno = EINVAL;
 486                 return 0;
 487         }
 488
 489         switch (bsda->u.sa.sa_family) {
 490         case AF_INET:
 491                 port = ntohs(bsda->u.in.sin_port);
 492                 break;
 493 #ifdef HAVE_IPV6
 494         case AF_INET6:
 495                 port = ntohs(bsda->u.in6.sin6_port);
 496                 break;
 497 #endif
 498         default:
 499                 errno = EINVAL;
 500                 return 0;
 501         }
 502
 503         return port;
 504 }
 505
 506 int tsocket_address_inet_set_port(struct tsocket_address *addr,
 507                                   uint16_t port)
 508 {
 509         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 510                                            struct samba_sockaddr);
 511
 512         if (!bsda) {
 513                 errno = EINVAL;
 514                 return -1;
 515         }
 516
 517         switch (bsda->u.sa.sa_family) {
 518         case AF_INET:
 519                 bsda->u.in.sin_port = htons(port);
 520                 break;
 521 #ifdef HAVE_IPV6
 522         case AF_INET6:
 523                 bsda->u.in6.sin6_port = htons(port);
 524                 break;
 525 #endif
 526         default:
 527                 errno = EINVAL;
 528                 return -1;
 529         }
 530
 531         return 0;
 532 }
 533
 534 bool tsocket_address_is_unix(const struct tsocket_address *addr)
 535 {
 536         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 537                                            struct samba_sockaddr);
 538
 539         if (!bsda) {
 540                 return false;
 541         }
 542
 543         switch (bsda->u.sa.sa_family) {
 544         case AF_UNIX:
 545                 return true;
 546         }
 547
 548         return false;
 549 }
 550
 551 int _tsocket_address_unix_from_path(TALLOC_CTX *mem_ctx,
 552                                     const char *path,
 553                                     struct tsocket_address **_addr,
 554                                     const char *location)
 555 {
 556         struct sockaddr_un un;
 557         void *p = &un;
 558         int ret;
 559
 560         if (!path) {
 561                 path = "";
 562         }
 563
 564         if (strlen(path) > sizeof(un.sun_path)-1) {
 565                 errno = ENAMETOOLONG;
 566                 return -1;
 567         }
 568
 569         ZERO_STRUCT(un);
 570         un.sun_family = AF_UNIX;
 571         strncpy(un.sun_path, path, sizeof(un.sun_path)-1);
 572
 573         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 574                                                  (struct sockaddr *)p,
 575                                                  sizeof(un),
 576                                                  _addr,
 577                                                  location);
 578
 579         return ret;
 580 }
 581
 582 char *tsocket_address_unix_path(const struct tsocket_address *addr,
 583                                 TALLOC_CTX *mem_ctx)
 584 {
 585         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 586                                            struct samba_sockaddr);
 587         const char *str;
 588
 589         if (!bsda) {
 590                 errno = EINVAL;
 591                 return NULL;
 592         }
 593
 594         switch (bsda->u.sa.sa_family) {
 595         case AF_UNIX:
 596                 str = bsda->u.un.sun_path;
 597                 break;
 598         default:
 599                 errno = EINVAL;
 600                 return NULL;
 601         }
 602
 603         return talloc_strdup(mem_ctx, str);
 604 }
 605
 606 static char *tsocket_address_bsd_string(const struct tsocket_address *addr,
 607                                         TALLOC_CTX *mem_ctx)
 608 {
 609         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 610                                            struct samba_sockaddr);
 611         char *str;
 612         char *addr_str;
 613         const char *prefix = NULL;
 614         uint16_t port;
 615
 616         switch (bsda->u.sa.sa_family) {
 617         case AF_UNIX:
 618                 return talloc_asprintf(mem_ctx, "unix:%s",
 619                                        bsda->u.un.sun_path);
 620         case AF_INET:
 621                 prefix = "ipv4";
 622                 break;
 623 #ifdef HAVE_IPV6
 624         case AF_INET6:
 625                 prefix = "ipv6";
 626                 break;
 627 #endif
 628         default:
 629                 errno = EINVAL;
 630                 return NULL;
 631         }
 632
 633         addr_str = tsocket_address_inet_addr_string(addr, mem_ctx);
 634         if (!addr_str) {
 635                 return NULL;
 636         }
 637
 638         port = tsocket_address_inet_port(addr);
 639
 640         str = talloc_asprintf(mem_ctx, "%s:%s:%u",
 641                               prefix, addr_str, port);
 642         talloc_free(addr_str);
 643
 644         return str;
 645 }
 646
 647 static struct tsocket_address *tsocket_address_bsd_copy(const struct tsocket_address *addr,
 648                                                          TALLOC_CTX *mem_ctx,
 649                                                          const char *location)
 650 {
 651         struct samba_sockaddr *bsda = talloc_get_type(addr->private_data,
 652                                            struct samba_sockaddr);
 653         struct tsocket_address *copy;
 654         int ret;
 655
 656         ret = _tsocket_address_bsd_from_sockaddr(mem_ctx,
 657                                                  &bsda->u.sa,
 658                                                  bsda->sa_socklen,
 659                                                  &copy,
 660                                                  location);
 661         if (ret != 0) {
 662                 return NULL;
 663         }
 664
 665         return copy;
 666 }
 667
 668 static const struct tsocket_address_ops tsocket_address_bsd_ops = {
 669         .name           = "bsd",
 670         .string         = tsocket_address_bsd_string,
 671         .copy           = tsocket_address_bsd_copy,
 672 };
 673
 674 struct tdgram_bsd {
 675         int fd;
 676
 677         void *event_ptr;
 678         struct tevent_fd *fde;
 679         bool optimize_recvfrom;
 680         bool netlink;
 681
 682         void *readable_private;
 683         void (*readable_handler)(void *private_data);
 684         void *writeable_private;
 685         void (*writeable_handler)(void *private_data);
 686 };
 687
 688 bool tdgram_bsd_optimize_recvfrom(struct tdgram_context *dgram,
 689                                   bool on)
 690 {
 691         struct tdgram_bsd *bsds =
 692                 talloc_get_type(_tdgram_context_data(dgram),
 693                 struct tdgram_bsd);
 694         bool old;
 695
 696         if (bsds == NULL) {
 697                 /* not a bsd socket */
 698                 return false;
 699         }
 700
 701         old = bsds->optimize_recvfrom;
 702         bsds->optimize_recvfrom = on;
 703
 704         return old;
 705 }
 706
 707 static void tdgram_bsd_fde_handler(struct tevent_context *ev,
 708                                    struct tevent_fd *fde,
 709                                    uint16_t flags,
 710                                    void *private_data)
 711 {
 712         struct tdgram_bsd *bsds = talloc_get_type_abort(private_data,
 713                                   struct tdgram_bsd);
 714
 715         if (flags & TEVENT_FD_WRITE) {
 716                 bsds->writeable_handler(bsds->writeable_private);
 717                 return;
 718         }
 719         if (flags & TEVENT_FD_READ) {
 720                 if (!bsds->readable_handler) {
 721                         TEVENT_FD_NOT_READABLE(bsds->fde);
 722                         return;
 723                 }
 724                 bsds->readable_handler(bsds->readable_private);
 725                 return;
 726         }
 727 }
 728
 729 static int tdgram_bsd_set_readable_handler(struct tdgram_bsd *bsds,
 730                                            struct tevent_context *ev,
 731                                            void (*handler)(void *private_data),
 732                                            void *private_data)
 733 {
 734         if (ev == NULL) {
 735                 if (handler) {
 736                         errno = EINVAL;
 737                         return -1;
 738                 }
 739                 if (!bsds->readable_handler) {
 740                         return 0;
 741                 }
 742                 bsds->readable_handler = NULL;
 743                 bsds->readable_private = NULL;
 744
 745                 return 0;
 746         }
 747
 748         /* read and write must use the same tevent_context */
 749         if (bsds->event_ptr != ev) {
 750                 if (bsds->readable_handler || bsds->writeable_handler) {
 751                         errno = EINVAL;
 752                         return -1;
 753                 }
 754                 bsds->event_ptr = NULL;
 755                 TALLOC_FREE(bsds->fde);
 756         }
 757
 758         if (tevent_fd_get_flags(bsds->fde) == 0) {
 759                 TALLOC_FREE(bsds->fde);
 760
 761                 bsds->fde = tevent_add_fd(ev, bsds,
 762                                           bsds->fd, TEVENT_FD_READ,
 763                                           tdgram_bsd_fde_handler,
 764                                           bsds);
 765                 if (!bsds->fde) {
 766                         errno = ENOMEM;
 767                         return -1;
 768                 }
 769
 770                 /* cache the event context we're running on */
 771                 bsds->event_ptr = ev;
 772         } else if (!bsds->readable_handler) {
 773                 TEVENT_FD_READABLE(bsds->fde);
 774         }
 775
 776         bsds->readable_handler = handler;
 777         bsds->readable_private = private_data;
 778
 779         return 0;
 780 }
 781
 782 static int tdgram_bsd_set_writeable_handler(struct tdgram_bsd *bsds,
 783                                             struct tevent_context *ev,
 784                                             void (*handler)(void *private_data),
 785                                             void *private_data)
 786 {
 787         if (ev == NULL) {
 788                 if (handler) {
 789                         errno = EINVAL;
 790                         return -1;
 791                 }
 792                 if (!bsds->writeable_handler) {
 793                         return 0;
 794                 }
 795                 bsds->writeable_handler = NULL;
 796                 bsds->writeable_private = NULL;
 797                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
 798
 799                 return 0;
 800         }
 801
 802         /* read and write must use the same tevent_context */
 803         if (bsds->event_ptr != ev) {
 804                 if (bsds->readable_handler || bsds->writeable_handler) {
 805                         errno = EINVAL;
 806                         return -1;
 807                 }
 808                 bsds->event_ptr = NULL;
 809                 TALLOC_FREE(bsds->fde);
 810         }
 811
 812         if (tevent_fd_get_flags(bsds->fde) == 0) {
 813                 TALLOC_FREE(bsds->fde);
 814
 815                 bsds->fde = tevent_add_fd(ev, bsds,
 816                                           bsds->fd, TEVENT_FD_WRITE,
 817                                           tdgram_bsd_fde_handler,
 818                                           bsds);
 819                 if (!bsds->fde) {
 820                         errno = ENOMEM;
 821                         return -1;
 822                 }
 823
 824                 /* cache the event context we're running on */
 825                 bsds->event_ptr = ev;
 826         } else if (!bsds->writeable_handler) {
 827                 TEVENT_FD_WRITEABLE(bsds->fde);
 828         }
 829
 830         bsds->writeable_handler = handler;
 831         bsds->writeable_private = private_data;
 832
 833         return 0;
 834 }
 835
 836 struct tdgram_bsd_recvfrom_state {
 837         struct tdgram_context *dgram;
 838         bool first_try;
 839         uint8_t *buf;
 840         size_t len;
 841         struct tsocket_address *src;
 842 };
 843
 844 static int tdgram_bsd_recvfrom_destructor(struct tdgram_bsd_recvfrom_state *state)
 845 {
 846         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
 847                                   struct tdgram_bsd);
 848
 849         tdgram_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
 850
 851         return 0;
 852 }
 853
 854 static void tdgram_bsd_recvfrom_handler(void *private_data);
 855
 856 static struct tevent_req *tdgram_bsd_recvfrom_send(TALLOC_CTX *mem_ctx,
 857                                         struct tevent_context *ev,
 858                                         struct tdgram_context *dgram)
 859 {
 860         struct tevent_req *req;
 861         struct tdgram_bsd_recvfrom_state *state;
 862         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 863         int ret;
 864
 865         req = tevent_req_create(mem_ctx, &state,
 866                                 struct tdgram_bsd_recvfrom_state);
 867         if (!req) {
 868                 return NULL;
 869         }
 870
 871         state->dgram    = dgram;
 872         state->first_try= true;
 873         state->buf      = NULL;
 874         state->len      = 0;
 875         state->src      = NULL;
 876
 877         talloc_set_destructor(state, tdgram_bsd_recvfrom_destructor);
 878
 879         if (bsds->fd == -1) {
 880                 tevent_req_error(req, ENOTCONN);
 881                 goto post;
 882         }
 883
 884
 885         /*
 886          * this is a fast path, not waiting for the
 887          * socket to become explicit readable gains
 888          * about 10%-20% performance in benchmark tests.
 889          */
 890         if (bsds->optimize_recvfrom) {
 891                 /*
 892                  * We only do the optimization on
 893                  * recvfrom if the caller asked for it.
 894                  *
 895                  * This is needed because in most cases
 896                  * we prefer to flush send buffers before
 897                  * receiving incoming requests.
 898                  */
 899                 tdgram_bsd_recvfrom_handler(req);
 900                 if (!tevent_req_is_in_progress(req)) {
 901                         goto post;
 902                 }
 903         }
 904
 905         ret = tdgram_bsd_set_readable_handler(bsds, ev,
 906                                               tdgram_bsd_recvfrom_handler,
 907                                               req);
 908         if (ret == -1) {
 909                 tevent_req_error(req, errno);
 910                 goto post;
 911         }
 912
 913         return req;
 914
 915  post:
 916         tevent_req_post(req, ev);
 917         return req;
 918 }
 919
 920 static void tdgram_bsd_recvfrom_handler(void *private_data)
 921 {
 922         struct tevent_req *req = talloc_get_type_abort(private_data,
 923                                  struct tevent_req);
 924         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
 925                                         struct tdgram_bsd_recvfrom_state);
 926         struct tdgram_context *dgram = state->dgram;
 927         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
 928         struct samba_sockaddr *bsda = NULL;
 929         ssize_t ret;
 930         int err;
 931         bool retry;
 932
 933         if (bsds->netlink) {
 934                 ret = tsocket_bsd_netlink_pending(bsds->fd);
 935         } else {
 936                 ret = tsocket_bsd_pending(bsds->fd);
 937         }
 938
 939         if (state->first_try && ret == 0) {
 940                 state->first_try = false;
 941                 /* retry later */
 942                 return;
 943         }
 944         state->first_try = false;
 945
 946         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 947         if (retry) {
 948                 /* retry later */
 949                 return;
 950         }
 951         if (tevent_req_error(req, err)) {
 952                 return;
 953         }
 954
 955         /* note that 'ret' can be 0 here */
 956         state->buf = talloc_array(state, uint8_t, ret);
 957         if (tevent_req_nomem(state->buf, req)) {
 958                 return;
 959         }
 960         state->len = ret;
 961
 962         state->src = tsocket_address_create(state,
 963                                             &tsocket_address_bsd_ops,
 964                                             &bsda,
 965                                             struct samba_sockaddr,
 966                                             __location__ "bsd_recvfrom");
 967         if (tevent_req_nomem(state->src, req)) {
 968                 return;
 969         }
 970
 971         ZERO_STRUCTP(bsda);
 972         bsda->sa_socklen = sizeof(bsda->u.ss);
 973 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
 974         bsda->u.sa.sa_len = bsda->sa_socklen;
 975 #endif
 976
 977         ret = recvfrom(bsds->fd, state->buf, state->len, 0,
 978                        &bsda->u.sa, &bsda->sa_socklen);
 979         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
 980         if (retry) {
 981                 /* retry later */
 982                 return;
 983         }
 984         if (tevent_req_error(req, err)) {
 985                 return;
 986         }
 987
 988         /*
 989          * Some systems (FreeBSD, see bug #7115) return too much
 990          * bytes in tsocket_bsd_pending()/ioctl(fd, FIONREAD, ...),
 991          * the return value includes some IP/UDP header bytes,
 992          * while recvfrom() just returns the payload.
 993          */
 994         state->buf = talloc_realloc(state, state->buf, uint8_t, ret);
 995         if (tevent_req_nomem(state->buf, req)) {
 996                 return;
 997         }
 998         state->len = ret;
 999
1000         tevent_req_done(req);
1001 }
1002
1003 static ssize_t tdgram_bsd_recvfrom_recv(struct tevent_req *req,
1004                                         int *perrno,
1005                                         TALLOC_CTX *mem_ctx,
1006                                         uint8_t **buf,
1007                                         struct tsocket_address **src)
1008 {
1009         struct tdgram_bsd_recvfrom_state *state = tevent_req_data(req,
1010                                         struct tdgram_bsd_recvfrom_state);
1011         ssize_t ret;
1012
1013         ret = tsocket_simple_int_recv(req, perrno);
1014         if (ret == 0) {
1015                 *buf = talloc_move(mem_ctx, &state->buf);
1016                 ret = state->len;
1017                 if (src) {
1018                         *src = talloc_move(mem_ctx, &state->src);
1019                 }
1020         }
1021
1022         tevent_req_received(req);
1023         return ret;
1024 }
1025
1026 struct tdgram_bsd_sendto_state {
1027         struct tdgram_context *dgram;
1028
1029         const uint8_t *buf;
1030         size_t len;
1031         const struct tsocket_address *dst;
1032
1033         ssize_t ret;
1034 };
1035
1036 static int tdgram_bsd_sendto_destructor(struct tdgram_bsd_sendto_state *state)
1037 {
1038         struct tdgram_bsd *bsds = tdgram_context_data(state->dgram,
1039                                   struct tdgram_bsd);
1040
1041         tdgram_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
1042
1043         return 0;
1044 }
1045
1046 static void tdgram_bsd_sendto_handler(void *private_data);
1047
1048 static struct tevent_req *tdgram_bsd_sendto_send(TALLOC_CTX *mem_ctx,
1049                                                  struct tevent_context *ev,
1050                                                  struct tdgram_context *dgram,
1051                                                  const uint8_t *buf,
1052                                                  size_t len,
1053                                                  const struct tsocket_address *dst)
1054 {
1055         struct tevent_req *req;
1056         struct tdgram_bsd_sendto_state *state;
1057         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1058         int ret;
1059
1060         req = tevent_req_create(mem_ctx, &state,
1061                                 struct tdgram_bsd_sendto_state);
1062         if (!req) {
1063                 return NULL;
1064         }
1065
1066         state->dgram    = dgram;
1067         state->buf      = buf;
1068         state->len      = len;
1069         state->dst      = dst;
1070         state->ret      = -1;
1071
1072         talloc_set_destructor(state, tdgram_bsd_sendto_destructor);
1073
1074         if (bsds->fd == -1) {
1075                 tevent_req_error(req, ENOTCONN);
1076                 goto post;
1077         }
1078
1079         /*
1080          * this is a fast path, not waiting for the
1081          * socket to become explicit writeable gains
1082          * about 10%-20% performance in benchmark tests.
1083          */
1084         tdgram_bsd_sendto_handler(req);
1085         if (!tevent_req_is_in_progress(req)) {
1086                 goto post;
1087         }
1088
1089         ret = tdgram_bsd_set_writeable_handler(bsds, ev,
1090                                                tdgram_bsd_sendto_handler,
1091                                                req);
1092         if (ret == -1) {
1093                 tevent_req_error(req, errno);
1094                 goto post;
1095         }
1096
1097         return req;
1098
1099  post:
1100         tevent_req_post(req, ev);
1101         return req;
1102 }
1103
1104 static void tdgram_bsd_sendto_handler(void *private_data)
1105 {
1106         struct tevent_req *req = talloc_get_type_abort(private_data,
1107                                  struct tevent_req);
1108         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
1109                                         struct tdgram_bsd_sendto_state);
1110         struct tdgram_context *dgram = state->dgram;
1111         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1112         struct sockaddr *sa = NULL;
1113         socklen_t sa_socklen = 0;
1114         ssize_t ret;
1115         int err;
1116         bool retry;
1117
1118         if (state->dst) {
1119                 struct samba_sockaddr *bsda =
1120                         talloc_get_type(state->dst->private_data,
1121                         struct samba_sockaddr);
1122
1123                 sa = &bsda->u.sa;
1124                 sa_socklen = bsda->sa_socklen;
1125         }
1126
1127         ret = sendto(bsds->fd, state->buf, state->len, 0, sa, sa_socklen);
1128         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1129         if (retry) {
1130                 /* retry later */
1131                 return;
1132         }
1133
1134         if (err == EMSGSIZE) {
1135                 /* round up in 1K increments */
1136                 int bufsize = ((state->len + 1023) & (~1023));
1137
1138                 ret = setsockopt(bsds->fd, SOL_SOCKET, SO_SNDBUF, &bufsize,
1139                                  sizeof(bufsize));
1140                 if (ret == 0) {
1141                         /*
1142                          * We do the retry here, rather then via the
1143                          * handler, as we only want to retry once for
1144                          * this condition, so if there is a mismatch
1145                          * between what setsockopt() accepts and what can
1146                          * actually be sent, we do not end up in a
1147                          * loop.
1148                          */
1149
1150                         ret = sendto(bsds->fd, state->buf, state->len,
1151                                      0, sa, sa_socklen);
1152                         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1153                         if (retry) { /* retry later */
1154                                 return;
1155                         }
1156                 }
1157         }
1158
1159         if (tevent_req_error(req, err)) {
1160                 return;
1161         }
1162
1163         state->ret = ret;
1164
1165         tevent_req_done(req);
1166 }
1167
1168 static ssize_t tdgram_bsd_sendto_recv(struct tevent_req *req, int *perrno)
1169 {
1170         struct tdgram_bsd_sendto_state *state = tevent_req_data(req,
1171                                         struct tdgram_bsd_sendto_state);
1172         ssize_t ret;
1173
1174         ret = tsocket_simple_int_recv(req, perrno);
1175         if (ret == 0) {
1176                 ret = state->ret;
1177         }
1178
1179         tevent_req_received(req);
1180         return ret;
1181 }
1182
1183 struct tdgram_bsd_disconnect_state {
1184         uint8_t __dummy;
1185 };
1186
1187 static struct tevent_req *tdgram_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
1188                                                      struct tevent_context *ev,
1189                                                      struct tdgram_context *dgram)
1190 {
1191         struct tdgram_bsd *bsds = tdgram_context_data(dgram, struct tdgram_bsd);
1192         struct tevent_req *req;
1193         struct tdgram_bsd_disconnect_state *state;
1194         int ret;
1195         int err;
1196         bool dummy;
1197
1198         req = tevent_req_create(mem_ctx, &state,
1199                                 struct tdgram_bsd_disconnect_state);
1200         if (req == NULL) {
1201                 return NULL;
1202         }
1203
1204         if (bsds->fd == -1) {
1205                 tevent_req_error(req, ENOTCONN);
1206                 goto post;
1207         }
1208
1209         TALLOC_FREE(bsds->fde);
1210         ret = close(bsds->fd);
1211         bsds->fd = -1;
1212         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
1213         if (tevent_req_error(req, err)) {
1214                 goto post;
1215         }
1216
1217         tevent_req_done(req);
1218 post:
1219         tevent_req_post(req, ev);
1220         return req;
1221 }
1222
1223 static int tdgram_bsd_disconnect_recv(struct tevent_req *req,
1224                                       int *perrno)
1225 {
1226         int ret;
1227
1228         ret = tsocket_simple_int_recv(req, perrno);
1229
1230         tevent_req_received(req);
1231         return ret;
1232 }
1233
1234 static const struct tdgram_context_ops tdgram_bsd_ops = {
1235         .name                   = "bsd",
1236
1237         .recvfrom_send          = tdgram_bsd_recvfrom_send,
1238         .recvfrom_recv          = tdgram_bsd_recvfrom_recv,
1239
1240         .sendto_send            = tdgram_bsd_sendto_send,
1241         .sendto_recv            = tdgram_bsd_sendto_recv,
1242
1243         .disconnect_send        = tdgram_bsd_disconnect_send,
1244         .disconnect_recv        = tdgram_bsd_disconnect_recv,
1245 };
1246
1247 static int tdgram_bsd_destructor(struct tdgram_bsd *bsds)
1248 {
1249         TALLOC_FREE(bsds->fde);
1250         if (bsds->fd != -1) {
1251                 close(bsds->fd);
1252                 bsds->fd = -1;
1253         }
1254         return 0;
1255 }
1256
1257 static int tdgram_bsd_dgram_socket(const struct tsocket_address *local,
1258                                    const struct tsocket_address *remote,
1259                                    bool broadcast,
1260                                    TALLOC_CTX *mem_ctx,
1261                                    struct tdgram_context **_dgram,
1262                                    const char *location)
1263 {
1264         struct samba_sockaddr *lbsda =
1265                 talloc_get_type_abort(local->private_data,
1266                 struct samba_sockaddr);
1267         struct samba_sockaddr *rbsda = NULL;
1268         struct tdgram_context *dgram;
1269         struct tdgram_bsd *bsds;
1270         int fd;
1271         int ret;
1272         bool do_bind = false;
1273         bool do_reuseaddr = false;
1274         bool do_ipv6only = false;
1275         bool is_inet = false;
1276         int sa_fam = lbsda->u.sa.sa_family;
1277
1278         if (remote) {
1279                 rbsda = talloc_get_type_abort(remote->private_data,
1280                         struct samba_sockaddr);
1281         }
1282
1283         switch (lbsda->u.sa.sa_family) {
1284         case AF_UNIX:
1285                 if (broadcast) {
1286                         errno = EINVAL;
1287                         return -1;
1288                 }
1289                 if (lbsda->u.un.sun_path[0] != 0) {
1290                         do_reuseaddr = true;
1291                         do_bind = true;
1292                 }
1293                 break;
1294         case AF_INET:
1295                 if (lbsda->u.in.sin_port != 0) {
1296                         do_reuseaddr = true;
1297                         do_bind = true;
1298                 }
1299                 if (lbsda->u.in.sin_addr.s_addr != INADDR_ANY) {
1300                         do_bind = true;
1301                 }
1302                 is_inet = true;
1303                 break;
1304 #ifdef HAVE_IPV6
1305         case AF_INET6:
1306                 if (lbsda->u.in6.sin6_port != 0) {
1307                         do_reuseaddr = true;
1308                         do_bind = true;
1309                 }
1310                 if (memcmp(&in6addr_any,
1311                            &lbsda->u.in6.sin6_addr,
1312                            sizeof(in6addr_any)) != 0) {
1313                         do_bind = true;
1314                 }
1315                 is_inet = true;
1316                 do_ipv6only = true;
1317                 break;
1318 #endif
1319         default:
1320                 errno = EINVAL;
1321                 return -1;
1322         }
1323
1324         if (!do_bind && is_inet && rbsda) {
1325                 sa_fam = rbsda->u.sa.sa_family;
1326                 switch (sa_fam) {
1327                 case AF_INET:
1328                         do_ipv6only = false;
1329                         break;
1330 #ifdef HAVE_IPV6
1331                 case AF_INET6:
1332                         do_ipv6only = true;
1333                         break;
1334 #endif
1335                 }
1336         }
1337
1338         fd = socket(sa_fam, SOCK_DGRAM, 0);
1339         if (fd < 0) {
1340                 return -1;
1341         }
1342
1343         fd = tsocket_bsd_common_prepare_fd(fd, true);
1344         if (fd < 0) {
1345                 return -1;
1346         }
1347
1348         dgram = tdgram_context_create(mem_ctx,
1349                                       &tdgram_bsd_ops,
1350                                       &bsds,
1351                                       struct tdgram_bsd,
1352                                       location);
1353         if (!dgram) {
1354                 int saved_errno = errno;
1355                 close(fd);
1356                 errno = saved_errno;
1357                 return -1;
1358         }
1359         ZERO_STRUCTP(bsds);
1360         bsds->fd = fd;
1361         talloc_set_destructor(bsds, tdgram_bsd_destructor);
1362
1363 #ifdef HAVE_IPV6
1364         if (do_ipv6only) {
1365                 int val = 1;
1366
1367                 ret = setsockopt(fd, IPPROTO_IPV6, IPV6_V6ONLY,
1368                                  (const void *)&val, sizeof(val));
1369                 if (ret == -1) {
1370                         int saved_errno = errno;
1371                         talloc_free(dgram);
1372                         errno = saved_errno;
1373                         return -1;
1374                 }
1375         }
1376 #endif
1377
1378         if (broadcast) {
1379                 int val = 1;
1380
1381                 ret = setsockopt(fd, SOL_SOCKET, SO_BROADCAST,
1382                                  (const void *)&val, sizeof(val));
1383                 if (ret == -1) {
1384                         int saved_errno = errno;
1385                         talloc_free(dgram);
1386                         errno = saved_errno;
1387                         return -1;
1388                 }
1389         }
1390
1391         if (do_reuseaddr) {
1392                 int val = 1;
1393
1394                 ret = setsockopt(fd, SOL_SOCKET, SO_REUSEADDR,
1395                                  (const void *)&val, sizeof(val));
1396                 if (ret == -1) {
1397                         int saved_errno = errno;
1398                         talloc_free(dgram);
1399                         errno = saved_errno;
1400                         return -1;
1401                 }
1402         }
1403
1404         if (do_bind) {
1405                 ret = bind(fd, &lbsda->u.sa, lbsda->sa_socklen);
1406                 if (ret == -1) {
1407                         int saved_errno = errno;
1408                         talloc_free(dgram);
1409                         errno = saved_errno;
1410                         return -1;
1411                 }
1412         }
1413
1414         if (rbsda) {
1415                 if (rbsda->u.sa.sa_family != sa_fam) {
1416                         talloc_free(dgram);
1417                         errno = EINVAL;
1418                         return -1;
1419                 }
1420
1421                 ret = connect(fd, &rbsda->u.sa, rbsda->sa_socklen);
1422                 if (ret == -1) {
1423                         int saved_errno = errno;
1424                         talloc_free(dgram);
1425                         errno = saved_errno;
1426                         return -1;
1427                 }
1428         }
1429
1430         *_dgram = dgram;
1431         return 0;
1432 }
1433
1434 int _tdgram_bsd_existing_socket(TALLOC_CTX *mem_ctx,
1435                                 int fd,
1436                                 struct tdgram_context **_dgram,
1437                                 const char *location)
1438 {
1439         struct tdgram_context *dgram;
1440         struct tdgram_bsd *bsds;
1441 #ifdef HAVE_LINUX_RTNETLINK_H
1442         int result;
1443         struct sockaddr sa;
1444         socklen_t sa_len = sizeof(struct sockaddr);
1445 #endif
1446
1447         dgram = tdgram_context_create(mem_ctx,
1448                                       &tdgram_bsd_ops,
1449                                       &bsds,
1450                                       struct tdgram_bsd,
1451                                       location);
1452         if (!dgram) {
1453                 return -1;
1454         }
1455         ZERO_STRUCTP(bsds);
1456         bsds->fd = fd;
1457         talloc_set_destructor(bsds, tdgram_bsd_destructor);
1458
1459         *_dgram = dgram;
1460
1461 #ifdef HAVE_LINUX_RTNETLINK_H
1462         /*
1463          * Try to determine the protocol family and remember if it's
1464          * AF_NETLINK. We don't care if this fails.
1465          */
1466         result = getsockname(fd, &sa, &sa_len);
1467         if (result == 0 && sa.sa_family == AF_NETLINK) {
1468                 bsds->netlink = true;
1469         }
1470 #endif
1471
1472         return 0;
1473 }
1474
1475 int _tdgram_inet_udp_socket(const struct tsocket_address *local,
1476                             const struct tsocket_address *remote,
1477                             TALLOC_CTX *mem_ctx,
1478                             struct tdgram_context **dgram,
1479                             const char *location)
1480 {
1481         struct samba_sockaddr *lbsda =
1482                 talloc_get_type_abort(local->private_data,
1483                 struct samba_sockaddr);
1484         int ret;
1485
1486         switch (lbsda->u.sa.sa_family) {
1487         case AF_INET:
1488                 break;
1489 #ifdef HAVE_IPV6
1490         case AF_INET6:
1491                 break;
1492 #endif
1493         default:
1494                 errno = EINVAL;
1495                 return -1;
1496         }
1497
1498         ret = tdgram_bsd_dgram_socket(local, remote, false,
1499                                       mem_ctx, dgram, location);
1500
1501         return ret;
1502 }
1503
1504 int _tdgram_inet_udp_broadcast_socket(const struct tsocket_address *local,
1505                                       TALLOC_CTX *mem_ctx,
1506                                       struct tdgram_context **dgram,
1507                                       const char *location)
1508 {
1509         struct samba_sockaddr *lbsda =
1510                 talloc_get_type_abort(local->private_data,
1511                 struct samba_sockaddr);
1512         int ret;
1513
1514         switch (lbsda->u.sa.sa_family) {
1515         case AF_INET:
1516                 break;
1517 #ifdef HAVE_IPV6
1518         case AF_INET6:
1519                 /* only ipv4 */
1520                 errno = EINVAL;
1521                 return -1;
1522 #endif
1523         default:
1524                 errno = EINVAL;
1525                 return -1;
1526         }
1527
1528         ret = tdgram_bsd_dgram_socket(local, NULL, true,
1529                                       mem_ctx, dgram, location);
1530
1531         return ret;
1532 }
1533
1534 int _tdgram_unix_socket(const struct tsocket_address *local,
1535                         const struct tsocket_address *remote,
1536                         TALLOC_CTX *mem_ctx,
1537                         struct tdgram_context **dgram,
1538                         const char *location)
1539 {
1540         struct samba_sockaddr *lbsda =
1541                 talloc_get_type_abort(local->private_data,
1542                 struct samba_sockaddr);
1543         int ret;
1544
1545         switch (lbsda->u.sa.sa_family) {
1546         case AF_UNIX:
1547                 break;
1548         default:
1549                 errno = EINVAL;
1550                 return -1;
1551         }
1552
1553         ret = tdgram_bsd_dgram_socket(local, remote, false,
1554                                       mem_ctx, dgram, location);
1555
1556         return ret;
1557 }
1558
1559 struct tstream_bsd {
1560         int fd;
1561
1562         void *event_ptr;
1563         struct tevent_fd *fde;
1564         bool optimize_readv;
1565
1566         void *readable_private;
1567         void (*readable_handler)(void *private_data);
1568         void *writeable_private;
1569         void (*writeable_handler)(void *private_data);
1570 };
1571
1572 bool tstream_bsd_optimize_readv(struct tstream_context *stream,
1573                                 bool on)
1574 {
1575         struct tstream_bsd *bsds =
1576                 talloc_get_type(_tstream_context_data(stream),
1577                 struct tstream_bsd);
1578         bool old;
1579
1580         if (bsds == NULL) {
1581                 /* not a bsd socket */
1582                 return false;
1583         }
1584
1585         old = bsds->optimize_readv;
1586         bsds->optimize_readv = on;
1587
1588         return old;
1589 }
1590
1591 static void tstream_bsd_fde_handler(struct tevent_context *ev,
1592                                     struct tevent_fd *fde,
1593                                     uint16_t flags,
1594                                     void *private_data)
1595 {
1596         struct tstream_bsd *bsds = talloc_get_type_abort(private_data,
1597                                    struct tstream_bsd);
1598
1599         if (flags & TEVENT_FD_WRITE) {
1600                 bsds->writeable_handler(bsds->writeable_private);
1601                 return;
1602         }
1603         if (flags & TEVENT_FD_READ) {
1604                 if (!bsds->readable_handler) {
1605                         if (bsds->writeable_handler) {
1606                                 bsds->writeable_handler(bsds->writeable_private);
1607                                 return;
1608                         }
1609                         TEVENT_FD_NOT_READABLE(bsds->fde);
1610                         return;
1611                 }
1612                 bsds->readable_handler(bsds->readable_private);
1613                 return;
1614         }
1615 }
1616
1617 static int tstream_bsd_set_readable_handler(struct tstream_bsd *bsds,
1618                                             struct tevent_context *ev,
1619                                             void (*handler)(void *private_data),
1620                                             void *private_data)
1621 {
1622         if (ev == NULL) {
1623                 if (handler) {
1624                         errno = EINVAL;
1625                         return -1;
1626                 }
1627                 if (!bsds->readable_handler) {
1628                         return 0;
1629                 }
1630                 bsds->readable_handler = NULL;
1631                 bsds->readable_private = NULL;
1632
1633                 return 0;
1634         }
1635
1636         /* read and write must use the same tevent_context */
1637         if (bsds->event_ptr != ev) {
1638                 if (bsds->readable_handler || bsds->writeable_handler) {
1639                         errno = EINVAL;
1640                         return -1;
1641                 }
1642                 bsds->event_ptr = NULL;
1643                 TALLOC_FREE(bsds->fde);
1644         }
1645
1646         if (tevent_fd_get_flags(bsds->fde) == 0) {
1647                 TALLOC_FREE(bsds->fde);
1648
1649                 bsds->fde = tevent_add_fd(ev, bsds,
1650                                           bsds->fd, TEVENT_FD_READ,
1651                                           tstream_bsd_fde_handler,
1652                                           bsds);
1653                 if (!bsds->fde) {
1654                         errno = ENOMEM;
1655                         return -1;
1656                 }
1657
1658                 /* cache the event context we're running on */
1659                 bsds->event_ptr = ev;
1660         } else if (!bsds->readable_handler) {
1661                 TEVENT_FD_READABLE(bsds->fde);
1662         }
1663
1664         bsds->readable_handler = handler;
1665         bsds->readable_private = private_data;
1666
1667         return 0;
1668 }
1669
1670 static int tstream_bsd_set_writeable_handler(struct tstream_bsd *bsds,
1671                                              struct tevent_context *ev,
1672                                              void (*handler)(void *private_data),
1673                                              void *private_data)
1674 {
1675         if (ev == NULL) {
1676                 if (handler) {
1677                         errno = EINVAL;
1678                         return -1;
1679                 }
1680                 if (!bsds->writeable_handler) {
1681                         return 0;
1682                 }
1683                 bsds->writeable_handler = NULL;
1684                 bsds->writeable_private = NULL;
1685                 TEVENT_FD_NOT_WRITEABLE(bsds->fde);
1686
1687                 return 0;
1688         }
1689
1690         /* read and write must use the same tevent_context */
1691         if (bsds->event_ptr != ev) {
1692                 if (bsds->readable_handler || bsds->writeable_handler) {
1693                         errno = EINVAL;
1694                         return -1;
1695                 }
1696                 bsds->event_ptr = NULL;
1697                 TALLOC_FREE(bsds->fde);
1698         }
1699
1700         if (tevent_fd_get_flags(bsds->fde) == 0) {
1701                 TALLOC_FREE(bsds->fde);
1702
1703                 bsds->fde = tevent_add_fd(ev, bsds,
1704                                           bsds->fd,
1705                                           TEVENT_FD_READ | TEVENT_FD_WRITE,
1706                                           tstream_bsd_fde_handler,
1707                                           bsds);
1708                 if (!bsds->fde) {
1709                         errno = ENOMEM;
1710                         return -1;
1711                 }
1712
1713                 /* cache the event context we're running on */
1714                 bsds->event_ptr = ev;
1715         } else if (!bsds->writeable_handler) {
1716                 uint16_t flags = tevent_fd_get_flags(bsds->fde);
1717                 flags |= TEVENT_FD_READ | TEVENT_FD_WRITE;
1718                 tevent_fd_set_flags(bsds->fde, flags);
1719         }
1720
1721         bsds->writeable_handler = handler;
1722         bsds->writeable_private = private_data;
1723
1724         return 0;
1725 }
1726
1727 static ssize_t tstream_bsd_pending_bytes(struct tstream_context *stream)
1728 {
1729         struct tstream_bsd *bsds = tstream_context_data(stream,
1730                                    struct tstream_bsd);
1731         ssize_t ret;
1732
1733         if (bsds->fd == -1) {
1734                 errno = ENOTCONN;
1735                 return -1;
1736         }
1737
1738         ret = tsocket_bsd_pending(bsds->fd);
1739
1740         return ret;
1741 }
1742
1743 struct tstream_bsd_readv_state {
1744         struct tstream_context *stream;
1745
1746         struct iovec *vector;
1747         size_t count;
1748
1749         int ret;
1750 };
1751
1752 static int tstream_bsd_readv_destructor(struct tstream_bsd_readv_state *state)
1753 {
1754         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1755                                    struct tstream_bsd);
1756
1757         tstream_bsd_set_readable_handler(bsds, NULL, NULL, NULL);
1758
1759         return 0;
1760 }
1761
1762 static void tstream_bsd_readv_handler(void *private_data);
1763
1764 static struct tevent_req *tstream_bsd_readv_send(TALLOC_CTX *mem_ctx,
1765                                         struct tevent_context *ev,
1766                                         struct tstream_context *stream,
1767                                         struct iovec *vector,
1768                                         size_t count)
1769 {
1770         struct tevent_req *req;
1771         struct tstream_bsd_readv_state *state;
1772         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1773         int ret;
1774
1775         req = tevent_req_create(mem_ctx, &state,
1776                                 struct tstream_bsd_readv_state);
1777         if (!req) {
1778                 return NULL;
1779         }
1780
1781         state->stream   = stream;
1782         /* we make a copy of the vector so that we can modify it */
1783         state->vector   = talloc_array(state, struct iovec, count);
1784         if (tevent_req_nomem(state->vector, req)) {
1785                 goto post;
1786         }
1787         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1788         state->count    = count;
1789         state->ret      = 0;
1790
1791         talloc_set_destructor(state, tstream_bsd_readv_destructor);
1792
1793         if (bsds->fd == -1) {
1794                 tevent_req_error(req, ENOTCONN);
1795                 goto post;
1796         }
1797
1798         /*
1799          * this is a fast path, not waiting for the
1800          * socket to become explicit readable gains
1801          * about 10%-20% performance in benchmark tests.
1802          */
1803         if (bsds->optimize_readv) {
1804                 /*
1805                  * We only do the optimization on
1806                  * readv if the caller asked for it.
1807                  *
1808                  * This is needed because in most cases
1809                  * we prefer to flush send buffers before
1810                  * receiving incoming requests.
1811                  */
1812                 tstream_bsd_readv_handler(req);
1813                 if (!tevent_req_is_in_progress(req)) {
1814                         goto post;
1815                 }
1816         }
1817
1818         ret = tstream_bsd_set_readable_handler(bsds, ev,
1819                                               tstream_bsd_readv_handler,
1820                                               req);
1821         if (ret == -1) {
1822                 tevent_req_error(req, errno);
1823                 goto post;
1824         }
1825
1826         return req;
1827
1828  post:
1829         tevent_req_post(req, ev);
1830         return req;
1831 }
1832
1833 static void tstream_bsd_readv_handler(void *private_data)
1834 {
1835         struct tevent_req *req = talloc_get_type_abort(private_data,
1836                                  struct tevent_req);
1837         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1838                                         struct tstream_bsd_readv_state);
1839         struct tstream_context *stream = state->stream;
1840         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1841         int ret;
1842         int err;
1843         int _count;
1844         bool ok, retry;
1845
1846         ret = readv(bsds->fd, state->vector, state->count);
1847         if (ret == 0) {
1848                 /* propagate end of file */
1849                 tevent_req_error(req, EPIPE);
1850                 return;
1851         }
1852         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1853         if (retry) {
1854                 /* retry later */
1855                 return;
1856         }
1857         if (tevent_req_error(req, err)) {
1858                 return;
1859         }
1860
1861         state->ret += ret;
1862
1863         _count = state->count; /* tstream has size_t count, readv has int */
1864         ok = iov_advance(&state->vector, &_count, ret);
1865         state->count = _count;
1866
1867         if (!ok) {
1868                 tevent_req_error(req, EINVAL);
1869                 return;
1870         }
1871
1872         if (state->count > 0) {
1873                 /* we have more to read */
1874                 return;
1875         }
1876
1877         tevent_req_done(req);
1878 }
1879
1880 static int tstream_bsd_readv_recv(struct tevent_req *req,
1881                                   int *perrno)
1882 {
1883         struct tstream_bsd_readv_state *state = tevent_req_data(req,
1884                                         struct tstream_bsd_readv_state);
1885         int ret;
1886
1887         ret = tsocket_simple_int_recv(req, perrno);
1888         if (ret == 0) {
1889                 ret = state->ret;
1890         }
1891
1892         tevent_req_received(req);
1893         return ret;
1894 }
1895
1896 struct tstream_bsd_writev_state {
1897         struct tstream_context *stream;
1898
1899         struct iovec *vector;
1900         size_t count;
1901
1902         int ret;
1903 };
1904
1905 static int tstream_bsd_writev_destructor(struct tstream_bsd_writev_state *state)
1906 {
1907         struct tstream_bsd *bsds = tstream_context_data(state->stream,
1908                                   struct tstream_bsd);
1909
1910         tstream_bsd_set_writeable_handler(bsds, NULL, NULL, NULL);
1911
1912         return 0;
1913 }
1914
1915 static void tstream_bsd_writev_handler(void *private_data);
1916
1917 static struct tevent_req *tstream_bsd_writev_send(TALLOC_CTX *mem_ctx,
1918                                                  struct tevent_context *ev,
1919                                                  struct tstream_context *stream,
1920                                                  const struct iovec *vector,
1921                                                  size_t count)
1922 {
1923         struct tevent_req *req;
1924         struct tstream_bsd_writev_state *state;
1925         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1926         int ret;
1927
1928         req = tevent_req_create(mem_ctx, &state,
1929                                 struct tstream_bsd_writev_state);
1930         if (!req) {
1931                 return NULL;
1932         }
1933
1934         state->stream   = stream;
1935         /* we make a copy of the vector so that we can modify it */
1936         state->vector   = talloc_array(state, struct iovec, count);
1937         if (tevent_req_nomem(state->vector, req)) {
1938                 goto post;
1939         }
1940         memcpy(state->vector, vector, sizeof(struct iovec)*count);
1941         state->count    = count;
1942         state->ret      = 0;
1943
1944         talloc_set_destructor(state, tstream_bsd_writev_destructor);
1945
1946         if (bsds->fd == -1) {
1947                 tevent_req_error(req, ENOTCONN);
1948                 goto post;
1949         }
1950
1951         /*
1952          * this is a fast path, not waiting for the
1953          * socket to become explicit writeable gains
1954          * about 10%-20% performance in benchmark tests.
1955          */
1956         tstream_bsd_writev_handler(req);
1957         if (!tevent_req_is_in_progress(req)) {
1958                 goto post;
1959         }
1960
1961         ret = tstream_bsd_set_writeable_handler(bsds, ev,
1962                                                tstream_bsd_writev_handler,
1963                                                req);
1964         if (ret == -1) {
1965                 tevent_req_error(req, errno);
1966                 goto post;
1967         }
1968
1969         return req;
1970
1971  post:
1972         tevent_req_post(req, ev);
1973         return req;
1974 }
1975
1976 static void tstream_bsd_writev_handler(void *private_data)
1977 {
1978         struct tevent_req *req = talloc_get_type_abort(private_data,
1979                                  struct tevent_req);
1980         struct tstream_bsd_writev_state *state = tevent_req_data(req,
1981                                         struct tstream_bsd_writev_state);
1982         struct tstream_context *stream = state->stream;
1983         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
1984         ssize_t ret;
1985         int err;
1986         int _count;
1987         bool ok, retry;
1988
1989         ret = writev(bsds->fd, state->vector, state->count);
1990         if (ret == 0) {
1991                 /* propagate end of file */
1992                 tevent_req_error(req, EPIPE);
1993                 return;
1994         }
1995         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
1996         if (retry) {
1997                 /* retry later */
1998                 return;
1999         }
2000         if (tevent_req_error(req, err)) {
2001                 return;
2002         }
2003
2004         state->ret += ret;
2005
2006         _count = state->count; /* tstream has size_t count, writev has int */
2007         ok = iov_advance(&state->vector, &_count, ret);
2008         state->count = _count;
2009
2010         if (!ok) {
2011                 tevent_req_error(req, EINVAL);
2012                 return;
2013         }
2014
2015         if (state->count > 0) {
2016                 /* we have more to read */
2017                 return;
2018         }
2019
2020         tevent_req_done(req);
2021 }
2022
2023 static int tstream_bsd_writev_recv(struct tevent_req *req, int *perrno)
2024 {
2025         struct tstream_bsd_writev_state *state = tevent_req_data(req,
2026                                         struct tstream_bsd_writev_state);
2027         int ret;
2028
2029         ret = tsocket_simple_int_recv(req, perrno);
2030         if (ret == 0) {
2031                 ret = state->ret;
2032         }
2033
2034         tevent_req_received(req);
2035         return ret;
2036 }
2037
2038 struct tstream_bsd_disconnect_state {
2039         void *__dummy;
2040 };
2041
2042 static struct tevent_req *tstream_bsd_disconnect_send(TALLOC_CTX *mem_ctx,
2043                                                      struct tevent_context *ev,
2044                                                      struct tstream_context *stream)
2045 {
2046         struct tstream_bsd *bsds = tstream_context_data(stream, struct tstream_bsd);
2047         struct tevent_req *req;
2048         struct tstream_bsd_disconnect_state *state;
2049         int ret;
2050         int err;
2051         bool dummy;
2052
2053         req = tevent_req_create(mem_ctx, &state,
2054                                 struct tstream_bsd_disconnect_state);
2055         if (req == NULL) {
2056                 return NULL;
2057         }
2058
2059         if (bsds->fd == -1) {
2060                 tevent_req_error(req, ENOTCONN);
2061                 goto post;
2062         }
2063
2064         TALLOC_FREE(bsds->fde);
2065         ret = close(bsds->fd);
2066         bsds->fd = -1;
2067         err = tsocket_bsd_error_from_errno(ret, errno, &dummy);
2068         if (tevent_req_error(req, err)) {
2069                 goto post;
2070         }
2071
2072         tevent_req_done(req);
2073 post:
2074         tevent_req_post(req, ev);
2075         return req;
2076 }
2077
2078 static int tstream_bsd_disconnect_recv(struct tevent_req *req,
2079                                       int *perrno)
2080 {
2081         int ret;
2082
2083         ret = tsocket_simple_int_recv(req, perrno);
2084
2085         tevent_req_received(req);
2086         return ret;
2087 }
2088
2089 static const struct tstream_context_ops tstream_bsd_ops = {
2090         .name                   = "bsd",
2091
2092         .pending_bytes          = tstream_bsd_pending_bytes,
2093
2094         .readv_send             = tstream_bsd_readv_send,
2095         .readv_recv             = tstream_bsd_readv_recv,
2096
2097         .writev_send            = tstream_bsd_writev_send,
2098         .writev_recv            = tstream_bsd_writev_recv,
2099
2100         .disconnect_send        = tstream_bsd_disconnect_send,
2101         .disconnect_recv        = tstream_bsd_disconnect_recv,
2102 };
2103
2104 static int tstream_bsd_destructor(struct tstream_bsd *bsds)
2105 {
2106         TALLOC_FREE(bsds->fde);
2107         if (bsds->fd != -1) {
2108                 close(bsds->fd);
2109                 bsds->fd = -1;
2110         }
2111         return 0;
2112 }
2113
2114 int _tstream_bsd_existing_socket(TALLOC_CTX *mem_ctx,
2115                                  int fd,
2116                                  struct tstream_context **_stream,
2117                                  const char *location)
2118 {
2119         struct tstream_context *stream;
2120         struct tstream_bsd *bsds;
2121
2122         stream = tstream_context_create(mem_ctx,
2123                                         &tstream_bsd_ops,
2124                                         &bsds,
2125                                         struct tstream_bsd,
2126                                         location);
2127         if (!stream) {
2128                 return -1;
2129         }
2130         ZERO_STRUCTP(bsds);
2131         bsds->fd = fd;
2132         talloc_set_destructor(bsds, tstream_bsd_destructor);
2133
2134         *_stream = stream;
2135         return 0;
2136 }
2137
2138 struct tstream_bsd_connect_state {
2139         int fd;
2140         struct tevent_fd *fde;
2141         struct tstream_conext *stream;
2142         struct tsocket_address *local;
2143 };
2144
2145 static int tstream_bsd_connect_destructor(struct tstream_bsd_connect_state *state)
2146 {
2147         TALLOC_FREE(state->fde);
2148         if (state->fd != -1) {
2149                 close(state->fd);
2150                 state->fd = -1;
2151         }
2152
2153         return 0;
2154 }
2155
2156 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
2157                                             struct tevent_fd *fde,
2158                                             uint16_t flags,
2159                                             void *private_data);
2160
2161 static struct tevent_req *tstream_bsd_connect_send(TALLOC_CTX *mem_ctx,
2162                                         struct tevent_context *ev,
2163                                         int sys_errno,
2164                                         const struct tsocket_address *local,
2165                                         const struct tsocket_address *remote)
2166 {
2167         struct tevent_req *req;
2168         struct tstream_bsd_connect_state *state;
2169         struct samba_sockaddr *lbsda =
2170                 talloc_get_type_abort(local->private_data,
2171                 struct samba_sockaddr);
2172         struct samba_sockaddr *lrbsda = NULL;
2173         struct samba_sockaddr *rbsda =
2174                 talloc_get_type_abort(remote->private_data,
2175                 struct samba_sockaddr);
2176         int ret;
2177         bool do_bind = false;
2178         bool do_reuseaddr = false;
2179         bool do_ipv6only = false;
2180         bool is_inet = false;
2181         int sa_fam = lbsda->u.sa.sa_family;
2182
2183         req = tevent_req_create(mem_ctx, &state,
2184                                 struct tstream_bsd_connect_state);
2185         if (!req) {
2186                 return NULL;
2187         }
2188         state->fd = -1;
2189         state->fde = NULL;
2190
2191         talloc_set_destructor(state, tstream_bsd_connect_destructor);
2192
2193         /* give the wrappers a chance to report an error */
2194         if (sys_errno != 0) {
2195                 tevent_req_error(req, sys_errno);
2196                 goto post;
2197         }
2198
2199         switch (lbsda->u.sa.sa_family) {
2200         case AF_UNIX:
2201                 if (lbsda->u.un.sun_path[0] != 0) {
2202                         do_reuseaddr = true;
2203                         do_bind = true;
2204                 }
2205                 break;
2206         case AF_INET:
2207                 if (lbsda->u.in.sin_port != 0) {
2208                         do_reuseaddr = true;
2209                         do_bind = true;
2210                 }
2211                 if (lbsda->u.in.sin_addr.s_addr != INADDR_ANY) {
2212                         do_bind = true;
2213                 }
2214                 is_inet = true;
2215                 break;
2216 #ifdef HAVE_IPV6
2217         case AF_INET6:
2218                 if (lbsda->u.in6.sin6_port != 0) {
2219                         do_reuseaddr = true;
2220                         do_bind = true;
2221                 }
2222                 if (memcmp(&in6addr_any,
2223                            &lbsda->u.in6.sin6_addr,
2224                            sizeof(in6addr_any)) != 0) {
2225                         do_bind = true;
2226                 }
2227                 is_inet = true;
2228                 do_ipv6only = true;
2229                 break;
2230 #endif
2231         default:
2232                 tevent_req_error(req, EINVAL);
2233                 goto post;
2234         }
2235
2236         if (!do_bind && is_inet) {
2237                 sa_fam = rbsda->u.sa.sa_family;
2238                 switch (sa_fam) {
2239                 case AF_INET:
2240                         do_ipv6only = false;
2241                         break;
2242 #ifdef HAVE_IPV6
2243                 case AF_INET6:
2244                         do_ipv6only = true;
2245                         break;
2246 #endif
2247                 }
2248         }
2249
2250         if (is_inet) {
2251                 state->local = tsocket_address_create(state,
2252                                                       &tsocket_address_bsd_ops,
2253                                                       &lrbsda,
2254                                                       struct samba_sockaddr,
2255                                                       __location__ "bsd_connect");
2256                 if (tevent_req_nomem(state->local, req)) {
2257                         goto post;
2258                 }
2259
2260                 ZERO_STRUCTP(lrbsda);
2261                 lrbsda->sa_socklen = sizeof(lrbsda->u.ss);
2262 #ifdef HAVE_STRUCT_SOCKADDR_SA_LEN
2263                 lrbsda->u.sa.sa_len = lrbsda->sa_socklen;
2264 #endif
2265         }
2266
2267         state->fd = socket(sa_fam, SOCK_STREAM, 0);
2268         if (state->fd == -1) {
2269                 tevent_req_error(req, errno);
2270                 goto post;
2271         }
2272
2273         state->fd = tsocket_bsd_common_prepare_fd(state->fd, true);
2274         if (state->fd == -1) {
2275                 tevent_req_error(req, errno);
2276                 goto post;
2277         }
2278
2279 #ifdef HAVE_IPV6
2280         if (do_ipv6only) {
2281                 int val = 1;
2282
2283                 ret = setsockopt(state->fd, IPPROTO_IPV6, IPV6_V6ONLY,
2284                                  (const void *)&val, sizeof(val));
2285                 if (ret == -1) {
2286                         tevent_req_error(req, errno);
2287                         goto post;
2288                 }
2289         }
2290 #endif
2291
2292         if (do_reuseaddr) {
2293                 int val = 1;
2294
2295                 ret = setsockopt(state->fd, SOL_SOCKET, SO_REUSEADDR,
2296                                  (const void *)&val, sizeof(val));
2297                 if (ret == -1) {
2298                         tevent_req_error(req, errno);
2299                         goto post;
2300                 }
2301         }
2302
2303         if (do_bind) {
2304                 ret = bind(state->fd, &lbsda->u.sa, lbsda->sa_socklen);
2305                 if (ret == -1) {
2306                         tevent_req_error(req, errno);
2307                         goto post;
2308                 }
2309         }
2310
2311         if (rbsda->u.sa.sa_family != sa_fam) {
2312                 tevent_req_error(req, EINVAL);
2313                 goto post;
2314         }
2315
2316         ret = connect(state->fd, &rbsda->u.sa, rbsda->sa_socklen);
2317         if (ret == -1) {
2318                 if (errno == EINPROGRESS) {
2319                         goto async;
2320                 }
2321                 tevent_req_error(req, errno);
2322                 goto post;
2323         }
2324
2325         if (!state->local) {
2326                 tevent_req_done(req);
2327                 goto post;
2328         }
2329
2330         if (lrbsda != NULL) {
2331                 ret = getsockname(state->fd,
2332                                   &lrbsda->u.sa,
2333                                   &lrbsda->sa_socklen);
2334                 if (ret == -1) {
2335                         tevent_req_error(req, errno);
2336                         goto post;
2337                 }
2338         }
2339
2340         tevent_req_done(req);
2341         goto post;
2342
2343  async:
2344
2345         /*
2346          * Note for historic reasons TEVENT_FD_WRITE is not enough
2347          * to get notified for POLLERR or EPOLLHUP even if they
2348          * come together with POLLOUT. That means we need to
2349          * use TEVENT_FD_READ in addition until we have
2350          * TEVENT_FD_ERROR.
2351          */
2352         state->fde = tevent_add_fd(ev, state,
2353                                    state->fd,
2354                                    TEVENT_FD_READ | TEVENT_FD_WRITE,
2355                                    tstream_bsd_connect_fde_handler,
2356                                    req);
2357         if (tevent_req_nomem(state->fde, req)) {
2358                 goto post;
2359         }
2360
2361         return req;
2362
2363  post:
2364         tevent_req_post(req, ev);
2365         return req;
2366 }
2367
2368 static void tstream_bsd_connect_fde_handler(struct tevent_context *ev,
2369                                             struct tevent_fd *fde,
2370                                             uint16_t flags,
2371                                             void *private_data)
2372 {
2373         struct tevent_req *req = talloc_get_type_abort(private_data,
2374                                  struct tevent_req);
2375         struct tstream_bsd_connect_state *state = tevent_req_data(req,
2376                                         struct tstream_bsd_connect_state);
2377         struct samba_sockaddr *lrbsda = NULL;
2378         int ret;
2379         int error=0;
2380         socklen_t len = sizeof(error);
2381         int err;
2382         bool retry;
2383
2384         ret = getsockopt(state->fd, SOL_SOCKET, SO_ERROR, &error, &len);
2385         if (ret == 0) {
2386                 if (error != 0) {
2387                         errno = error;
2388                         ret = -1;
2389                 }
2390         }
2391         err = tsocket_bsd_error_from_errno(ret, errno, &retry);
2392         if (retry) {
2393                 /* retry later */
2394                 return;
2395         }
2396         if (tevent_req_error(req, err)) {
2397                 return;
2398         }
2399
2400         if (!state->local) {
2401                 tevent_req_done(req);
2402                 return;
2403         }
2404
2405         lrbsda = talloc_get_type_abort(state->local->private_data,
2406                                        struct samba_sockaddr);
2407
2408         ret = getsockname(state->fd, &lrbsda->u.sa, &lrbsda->sa_socklen);
2409         if (ret == -1) {
2410                 tevent_req_error(req, errno);
2411                 return;
2412         }
2413
2414         tevent_req_done(req);
2415 }
2416
2417 static int tstream_bsd_connect_recv(struct tevent_req *req,
2418                                     int *perrno,
2419                                     TALLOC_CTX *mem_ctx,
2420                                     struct tstream_context **stream,
2421                                     struct tsocket_address **local,
2422                                     const char *location)
2423 {
2424         struct tstream_bsd_connect_state *state = tevent_req_data(req,
2425                                         struct tstream_bsd_connect_state);
2426         int ret;
2427
2428         ret = tsocket_simple_int_recv(req, perrno);
2429         if (ret == 0) {
2430                 ret = _tstream_bsd_existing_socket(mem_ctx,
2431                                                    state->fd,
2432                                                    stream,
2433                                                    location);
2434                 if (ret == -1) {
2435                         *perrno = errno;
2436                         goto done;
2437                 }
2438                 TALLOC_FREE(state->fde);
2439                 state->fd = -1;
2440
2441                 if (local) {
2442                         *local = talloc_move(mem_ctx, &state->local);
2443                 }
2444         }
2445
2446 done:
2447         tevent_req_received(req);
2448         return ret;
2449 }
2450
2451 struct tevent_req * tstream_inet_tcp_connect_send(TALLOC_CTX *mem_ctx,
2452                                         struct tevent_context *ev,
2453                                         const struct tsocket_address *local,
2454                                         const struct tsocket_address *remote)
2455 {
2456         struct samba_sockaddr *lbsda =
2457                 talloc_get_type_abort(local->private_data,
2458                 struct samba_sockaddr);
2459         struct tevent_req *req;
2460         int sys_errno = 0;
2461
2462         switch (lbsda->u.sa.sa_family) {
2463         case AF_INET:
2464                 break;
2465 #ifdef HAVE_IPV6
2466         case AF_INET6:
2467                 break;
2468 #endif
2469         default:
2470                 sys_errno = EINVAL;
2471                 break;
2472         }
2473
2474         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2475
2476         return req;
2477 }
2478
2479 int _tstream_inet_tcp_connect_recv(struct tevent_req *req,
2480                                    int *perrno,
2481                                    TALLOC_CTX *mem_ctx,
2482                                    struct tstream_context **stream,
2483                                    struct tsocket_address **local,
2484                                    const char *location)
2485 {
2486         return tstream_bsd_connect_recv(req, perrno,
2487                                         mem_ctx, stream, local,
2488                                         location);
2489 }
2490
2491 struct tevent_req * tstream_unix_connect_send(TALLOC_CTX *mem_ctx,
2492                                         struct tevent_context *ev,
2493                                         const struct tsocket_address *local,
2494                                         const struct tsocket_address *remote)
2495 {
2496         struct samba_sockaddr *lbsda =
2497                 talloc_get_type_abort(local->private_data,
2498                 struct samba_sockaddr);
2499         struct tevent_req *req;
2500         int sys_errno = 0;
2501
2502         switch (lbsda->u.sa.sa_family) {
2503         case AF_UNIX:
2504                 break;
2505         default:
2506                 sys_errno = EINVAL;
2507                 break;
2508         }
2509
2510         req = tstream_bsd_connect_send(mem_ctx, ev, sys_errno, local, remote);
2511
2512         return req;
2513 }
2514
2515 int _tstream_unix_connect_recv(struct tevent_req *req,
2516                                       int *perrno,
2517                                       TALLOC_CTX *mem_ctx,
2518                                       struct tstream_context **stream,
2519                                       const char *location)
2520 {
2521         return tstream_bsd_connect_recv(req, perrno,
2522                                         mem_ctx, stream, NULL,
2523                                         location);
2524 }
2525
2526 int _tstream_unix_socketpair(TALLOC_CTX *mem_ctx1,
2527                              struct tstream_context **_stream1,
2528                              TALLOC_CTX *mem_ctx2,
2529                              struct tstream_context **_stream2,
2530                              const char *location)
2531 {
2532         int ret;
2533         int fds[2];
2534         int fd1;
2535         int fd2;
2536         struct tstream_context *stream1 = NULL;
2537         struct tstream_context *stream2 = NULL;
2538
2539         ret = socketpair(AF_UNIX, SOCK_STREAM, 0, fds);
2540         if (ret == -1) {
2541                 return -1;
2542         }
2543         fd1 = fds[0];
2544         fd2 = fds[1];
2545
2546         fd1 = tsocket_bsd_common_prepare_fd(fd1, true);
2547         if (fd1 == -1) {
2548                 int sys_errno = errno;
2549                 close(fd2);
2550                 errno = sys_errno;
2551                 return -1;
2552         }
2553
2554         fd2 = tsocket_bsd_common_prepare_fd(fd2, true);
2555         if (fd2 == -1) {
2556                 int sys_errno = errno;
2557                 close(fd1);
2558                 errno = sys_errno;
2559                 return -1;
2560         }
2561
2562         ret = _tstream_bsd_existing_socket(mem_ctx1,
2563                                            fd1,
2564                                            &stream1,
2565                                            location);
2566         if (ret == -1) {
2567                 int sys_errno = errno;
2568                 close(fd1);
2569                 close(fd2);
2570                 errno = sys_errno;
2571                 return -1;
2572         }
2573
2574         ret = _tstream_bsd_existing_socket(mem_ctx2,
2575                                            fd2,
2576                                            &stream2,
2577                                            location);
2578         if (ret == -1) {
2579                 int sys_errno = errno;
2580                 talloc_free(stream1);
2581                 close(fd2);
2582                 errno = sys_errno;
2583                 return -1;
2584         }
2585
2586         *_stream1 = stream1;
2587         *_stream2 = stream2;
2588         return 0;
2589 }
2590