lib/tdb2/open.c

   1  /*
   2    Trivial Database 2: opening and closing TDBs
   3    Copyright (C) Rusty Russell 2010
   4
   5    This library is free software; you can redistribute it and/or
   6    modify it under the terms of the GNU Lesser General Public
   7    License as published by the Free Software Foundation; either
   8    version 3 of the License, or (at your option) any later version.
   9
  10    This library is distributed in the hope that it will be useful,
  11    but WITHOUT ANY WARRANTY; without even the implied warranty of
  12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  13    Lesser General Public License for more details.
  14
  15    You should have received a copy of the GNU Lesser General Public
  16    License along with this library; if not, see <http://www.gnu.org/licenses/>.
  17 */
  18 #include "private.h"
  19 #include <ccan/build_assert/build_assert.h>
  20 #include <assert.h>
  21
  22 /* all tdbs, to detect double-opens (fcntl file don't nest!) */
  23 static struct tdb_context *tdbs = NULL;
  24
  25 static struct tdb_file *find_file(dev_t device, ino_t ino)
  26 {
  27         struct tdb_context *i;
  28
  29         for (i = tdbs; i; i = i->next) {
  30                 if (i->file->device == device && i->file->inode == ino) {
  31                         i->file->refcnt++;
  32                         return i->file;
  33                 }
  34         }
  35         return NULL;
  36 }
  37
  38 static bool read_all(int fd, void *buf, size_t len)
  39 {
  40         while (len) {
  41                 ssize_t ret;
  42                 ret = read(fd, buf, len);
  43                 if (ret < 0)
  44                         return false;
  45                 if (ret == 0) {
  46                         /* ETOOSHORT? */
  47                         errno = EWOULDBLOCK;
  48                         return false;
  49                 }
  50                 buf = (char *)buf + ret;
  51                 len -= ret;
  52         }
  53         return true;
  54 }
  55
  56 static uint64_t random_number(struct tdb_context *tdb)
  57 {
  58         int fd;
  59         uint64_t ret = 0;
  60         struct timeval now;
  61
  62         fd = open("/dev/urandom", O_RDONLY);
  63         if (fd >= 0) {
  64                 if (read_all(fd, &ret, sizeof(ret))) {
  65                         close(fd);
  66                         return ret;
  67                 }
  68                 close(fd);
  69         }
  70         /* FIXME: Untested!  Based on Wikipedia protocol description! */
  71         fd = open("/dev/egd-pool", O_RDWR);
  72         if (fd >= 0) {
  73                 /* Command is 1, next byte is size we want to read. */
  74                 char cmd[2] = { 1, sizeof(uint64_t) };
  75                 if (write(fd, cmd, sizeof(cmd)) == sizeof(cmd)) {
  76                         char reply[1 + sizeof(uint64_t)];
  77                         int r = read(fd, reply, sizeof(reply));
  78                         if (r > 1) {
  79                                 /* Copy at least some bytes. */
  80                                 memcpy(&ret, reply+1, r - 1);
  81                                 if (reply[0] == sizeof(uint64_t)
  82                                     && r == sizeof(reply)) {
  83                                         close(fd);
  84                                         return ret;
  85                                 }
  86                         }
  87                 }
  88                 close(fd);
  89         }
  90
  91         /* Fallback: pid and time. */
  92         gettimeofday(&now, NULL);
  93         ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
  94         tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
  95                    "tdb_open: random from getpid and time");
  96         return ret;
  97 }
  98
  99 static void tdb2_context_init(struct tdb_context *tdb)
 100 {
 101         /* Initialize the TDB2 fields here */
 102         tdb_io_init(tdb);
 103         tdb->tdb2.direct_access = 0;
 104         tdb->tdb2.transaction = NULL;
 105         tdb->tdb2.access = NULL;
 106 }
 107
 108 struct new_database {
 109         struct tdb_header hdr;
 110         struct tdb_freetable ftable;
 111 };
 112
 113 /* initialise a new database */
 114 static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 115                                        struct tdb_attribute_seed *seed,
 116                                        struct tdb_header *hdr)
 117 {
 118         /* We make it up in memory, then write it out if not internal */
 119         struct new_database newdb;
 120         unsigned int magic_len;
 121         ssize_t rlen;
 122         enum TDB_ERROR ecode;
 123
 124         /* Fill in the header */
 125         newdb.hdr.version = TDB_VERSION;
 126         if (seed)
 127                 newdb.hdr.hash_seed = seed->seed;
 128         else
 129                 newdb.hdr.hash_seed = random_number(tdb);
 130         newdb.hdr.hash_test = TDB_HASH_MAGIC;
 131         newdb.hdr.hash_test = tdb->hash_fn(&newdb.hdr.hash_test,
 132                                            sizeof(newdb.hdr.hash_test),
 133                                            newdb.hdr.hash_seed,
 134                                            tdb->hash_data);
 135         newdb.hdr.recovery = 0;
 136         newdb.hdr.features_used = newdb.hdr.features_offered = TDB_FEATURE_MASK;
 137         newdb.hdr.seqnum = 0;
 138         newdb.hdr.capabilities = 0;
 139         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
 140         /* Initial hashes are empty. */
 141         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 142
 143         /* Free is empty. */
 144         newdb.hdr.free_table = offsetof(struct new_database, ftable);
 145         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
 146         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
 147                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 148                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 149                            0);
 150         if (ecode != TDB_SUCCESS) {
 151                 return ecode;
 152         }
 153
 154         /* Magic food */
 155         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 156         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 157
 158         /* This creates an endian-converted database, as if read from disk */
 159         magic_len = sizeof(newdb.hdr.magic_food);
 160         tdb_convert(tdb,
 161                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 162
 163         *hdr = newdb.hdr;
 164
 165         if (tdb->flags & TDB_INTERNAL) {
 166                 tdb->file->map_size = sizeof(newdb);
 167                 tdb->file->map_ptr = malloc(tdb->file->map_size);
 168                 if (!tdb->file->map_ptr) {
 169                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 170                                           "tdb_new_database:"
 171                                           " failed to allocate");
 172                 }
 173                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
 174                 return TDB_SUCCESS;
 175         }
 176         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
 177                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 178                                   "tdb_new_database:"
 179                                   " failed to seek: %s", strerror(errno));
 180         }
 181
 182         if (ftruncate(tdb->file->fd, 0) == -1) {
 183                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 184                                   "tdb_new_database:"
 185                                   " failed to truncate: %s", strerror(errno));
 186         }
 187
 188         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
 189         if (rlen != sizeof(newdb)) {
 190                 if (rlen >= 0)
 191                         errno = ENOSPC;
 192                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 193                                   "tdb_new_database: %zi writing header: %s",
 194                                   rlen, strerror(errno));
 195         }
 196         return TDB_SUCCESS;
 197 }
 198
 199 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
 200 {
 201         tdb->file = malloc(sizeof(*tdb->file));
 202         if (!tdb->file)
 203                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 204                                   "tdb_open: cannot alloc tdb_file structure");
 205         tdb->file->num_lockrecs = 0;
 206         tdb->file->lockrecs = NULL;
 207         tdb->file->allrecord_lock.count = 0;
 208         tdb->file->refcnt = 1;
 209         tdb->file->map_ptr = NULL;
 210         return TDB_SUCCESS;
 211 }
 212
 213 _PUBLIC_ enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
 214                                  const union tdb_attribute *attr)
 215 {
 216         switch (attr->base.attr) {
 217         case TDB_ATTRIBUTE_LOG:
 218                 tdb->log_fn = attr->log.fn;
 219                 tdb->log_data = attr->log.data;
 220                 break;
 221         case TDB_ATTRIBUTE_HASH:
 222         case TDB_ATTRIBUTE_SEED:
 223         case TDB_ATTRIBUTE_OPENHOOK:
 224                 return tdb->last_error
 225                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 226                                      TDB_LOG_USE_ERROR,
 227                                      "tdb_set_attribute:"
 228                                      " cannot set %s after opening",
 229                                      attr->base.attr == TDB_ATTRIBUTE_HASH
 230                                      ? "TDB_ATTRIBUTE_HASH"
 231                                      : attr->base.attr == TDB_ATTRIBUTE_SEED
 232                                      ? "TDB_ATTRIBUTE_SEED"
 233                                      : "TDB_ATTRIBUTE_OPENHOOK");
 234         case TDB_ATTRIBUTE_STATS:
 235                 return tdb->last_error
 236                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 237                                      TDB_LOG_USE_ERROR,
 238                                      "tdb_set_attribute:"
 239                                      " cannot set TDB_ATTRIBUTE_STATS");
 240         case TDB_ATTRIBUTE_FLOCK:
 241                 tdb->lock_fn = attr->flock.lock;
 242                 tdb->unlock_fn = attr->flock.unlock;
 243                 tdb->lock_data = attr->flock.data;
 244                 break;
 245         default:
 246                 return tdb->last_error
 247                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 248                                      TDB_LOG_USE_ERROR,
 249                                      "tdb_set_attribute:"
 250                                      " unknown attribute type %u",
 251                                      attr->base.attr);
 252         }
 253         return TDB_SUCCESS;
 254 }
 255
 256 _PUBLIC_ enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
 257                                  union tdb_attribute *attr)
 258 {
 259         switch (attr->base.attr) {
 260         case TDB_ATTRIBUTE_LOG:
 261                 if (!tdb->log_fn)
 262                         return tdb->last_error = TDB_ERR_NOEXIST;
 263                 attr->log.fn = tdb->log_fn;
 264                 attr->log.data = tdb->log_data;
 265                 break;
 266         case TDB_ATTRIBUTE_HASH:
 267                 attr->hash.fn = tdb->hash_fn;
 268                 attr->hash.data = tdb->hash_data;
 269                 break;
 270         case TDB_ATTRIBUTE_SEED:
 271                 attr->seed.seed = tdb->hash_seed;
 272                 break;
 273         case TDB_ATTRIBUTE_OPENHOOK:
 274                 if (!tdb->openhook)
 275                         return tdb->last_error = TDB_ERR_NOEXIST;
 276                 attr->openhook.fn = tdb->openhook;
 277                 attr->openhook.data = tdb->openhook_data;
 278                 break;
 279         case TDB_ATTRIBUTE_STATS: {
 280                 size_t size = attr->stats.size;
 281                 if (size > tdb->stats.size)
 282                         size = tdb->stats.size;
 283                 memcpy(&attr->stats, &tdb->stats, size);
 284                 break;
 285         }
 286         case TDB_ATTRIBUTE_FLOCK:
 287                 attr->flock.lock = tdb->lock_fn;
 288                 attr->flock.unlock = tdb->unlock_fn;
 289                 attr->flock.data = tdb->lock_data;
 290                 break;
 291         default:
 292                 return tdb->last_error
 293                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 294                                      TDB_LOG_USE_ERROR,
 295                                      "tdb_get_attribute:"
 296                                      " unknown attribute type %u",
 297                                      attr->base.attr);
 298         }
 299         attr->base.next = NULL;
 300         return TDB_SUCCESS;
 301 }
 302
 303 _PUBLIC_ void tdb_unset_attribute(struct tdb_context *tdb,
 304                          enum tdb_attribute_type type)
 305 {
 306         switch (type) {
 307         case TDB_ATTRIBUTE_LOG:
 308                 tdb->log_fn = NULL;
 309                 break;
 310         case TDB_ATTRIBUTE_OPENHOOK:
 311                 tdb->openhook = NULL;
 312                 break;
 313         case TDB_ATTRIBUTE_HASH:
 314         case TDB_ATTRIBUTE_SEED:
 315                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 316                            "tdb_unset_attribute: cannot unset %s after opening",
 317                            type == TDB_ATTRIBUTE_HASH
 318                            ? "TDB_ATTRIBUTE_HASH"
 319                            : "TDB_ATTRIBUTE_SEED");
 320                 break;
 321         case TDB_ATTRIBUTE_STATS:
 322                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 323                            TDB_LOG_USE_ERROR,
 324                            "tdb_unset_attribute:"
 325                            "cannot unset TDB_ATTRIBUTE_STATS");
 326                 break;
 327         case TDB_ATTRIBUTE_FLOCK:
 328                 tdb->lock_fn = tdb_fcntl_lock;
 329                 tdb->unlock_fn = tdb_fcntl_unlock;
 330                 break;
 331         default:
 332                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 333                            TDB_LOG_USE_ERROR,
 334                            "tdb_unset_attribute: unknown attribute type %u",
 335                            type);
 336         }
 337 }
 338
 339 /* The top three bits of the capability tell us whether it matters. */
 340 enum TDB_ERROR unknown_capability(struct tdb_context *tdb, const char *caller,
 341                                   tdb_off_t type)
 342 {
 343         if (type & TDB_CAP_NOOPEN) {
 344                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 345                                   "%s: file has unknown capability %llu",
 346                                   caller, type & TDB_CAP_NOOPEN);
 347         }
 348
 349         if ((type & TDB_CAP_NOWRITE) && !(tdb->flags & TDB_RDONLY)) {
 350                 return tdb_logerr(tdb, TDB_ERR_RDONLY, TDB_LOG_ERROR,
 351                                   "%s: file has unknown capability %llu"
 352                                   " (cannot write to it)",
 353                                   caller, type & TDB_CAP_NOOPEN);
 354         }
 355
 356         if (type & TDB_CAP_NOCHECK) {
 357                 tdb->flags |= TDB_CANT_CHECK;
 358         }
 359         return TDB_SUCCESS;
 360 }
 361
 362 static enum TDB_ERROR capabilities_ok(struct tdb_context *tdb,
 363                                       tdb_off_t capabilities)
 364 {
 365         tdb_off_t off, next;
 366         enum TDB_ERROR ecode = TDB_SUCCESS;
 367         const struct tdb_capability *cap;
 368
 369         /* Check capability list. */
 370         for (off = capabilities; off && ecode == TDB_SUCCESS; off = next) {
 371                 cap = tdb_access_read(tdb, off, sizeof(*cap), true);
 372                 if (TDB_PTR_IS_ERR(cap)) {
 373                         return TDB_PTR_ERR(cap);
 374                 }
 375
 376                 switch (cap->type & TDB_CAP_TYPE_MASK) {
 377                 /* We don't understand any capabilities (yet). */
 378                 default:
 379                         ecode = unknown_capability(tdb, "tdb_open", cap->type);
 380                 }
 381                 next = cap->next;
 382                 tdb_access_release(tdb, cap);
 383         }
 384         return ecode;
 385 }
 386
 387 _PUBLIC_ struct tdb_context *tdb_open(const char *name, int tdb_flags,
 388                              int open_flags, mode_t mode,
 389                              union tdb_attribute *attr)
 390 {
 391         struct tdb_context *tdb;
 392         struct stat st;
 393         int saved_errno = 0;
 394         uint64_t hash_test;
 395         unsigned v;
 396         ssize_t rlen;
 397         struct tdb_header hdr;
 398         struct tdb_attribute_seed *seed = NULL;
 399         tdb_bool_err berr;
 400         enum TDB_ERROR ecode;
 401         int openlock;
 402
 403         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 404         if (!tdb) {
 405                 /* Can't log this */
 406                 errno = ENOMEM;
 407                 return NULL;
 408         }
 409         /* Set name immediately for logging functions. */
 410         if (name) {
 411                 tdb->name = strcpy((char *)(tdb + 1), name);
 412         } else {
 413                 tdb->name = NULL;
 414         }
 415         tdb->flags = tdb_flags;
 416         tdb->log_fn = NULL;
 417         tdb->open_flags = open_flags;
 418         tdb->last_error = TDB_SUCCESS;
 419         tdb->file = NULL;
 420         tdb->openhook = NULL;
 421         tdb->lock_fn = tdb_fcntl_lock;
 422         tdb->unlock_fn = tdb_fcntl_unlock;
 423         tdb->hash_fn = tdb_jenkins_hash;
 424         memset(&tdb->stats, 0, sizeof(tdb->stats));
 425         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
 426         tdb->stats.size = sizeof(tdb->stats);
 427
 428         while (attr) {
 429                 switch (attr->base.attr) {
 430                 case TDB_ATTRIBUTE_HASH:
 431                         tdb->hash_fn = attr->hash.fn;
 432                         tdb->hash_data = attr->hash.data;
 433                         break;
 434                 case TDB_ATTRIBUTE_SEED:
 435                         seed = &attr->seed;
 436                         break;
 437                 case TDB_ATTRIBUTE_OPENHOOK:
 438                         tdb->openhook = attr->openhook.fn;
 439                         tdb->openhook_data = attr->openhook.data;
 440                         break;
 441                 default:
 442                         /* These are set as normal. */
 443                         ecode = tdb_set_attribute(tdb, attr);
 444                         if (ecode != TDB_SUCCESS)
 445                                 goto fail;
 446                 }
 447                 attr = attr->base.next;
 448         }
 449
 450         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 451                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
 452                           | TDB_RDONLY)) {
 453                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 454                                    "tdb_open: unknown flags %u", tdb_flags);
 455                 goto fail;
 456         }
 457
 458         if (seed) {
 459                 if (!(tdb_flags & TDB_INTERNAL) && !(open_flags & O_CREAT)) {
 460                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 461                                            TDB_LOG_USE_ERROR,
 462                                            "tdb_open:"
 463                                            " cannot set TDB_ATTRIBUTE_SEED"
 464                                            " without O_CREAT.");
 465                         goto fail;
 466                 }
 467         }
 468
 469         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 470                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 471                                    "tdb_open: can't open tdb %s write-only",
 472                                    name);
 473                 goto fail;
 474         }
 475
 476         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 477                 openlock = F_RDLCK;
 478                 tdb->flags |= TDB_RDONLY;
 479         } else {
 480                 if (tdb_flags & TDB_RDONLY) {
 481                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 482                                            TDB_LOG_USE_ERROR,
 483                                            "tdb_open: can't use TDB_RDONLY"
 484                                            " without O_RDONLY");
 485                         goto fail;
 486                 }
 487                 openlock = F_WRLCK;
 488         }
 489
 490         /* internal databases don't need any of the rest. */
 491         if (tdb->flags & TDB_INTERNAL) {
 492                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 493                 ecode = tdb_new_file(tdb);
 494                 if (ecode != TDB_SUCCESS) {
 495                         goto fail;
 496                 }
 497                 tdb->file->fd = -1;
 498                 ecode = tdb_new_database(tdb, seed, &hdr);
 499                 if (ecode == TDB_SUCCESS) {
 500                         tdb_convert(tdb, &hdr.hash_seed,
 501                                     sizeof(hdr.hash_seed));
 502                         tdb->hash_seed = hdr.hash_seed;
 503                         tdb2_context_init(tdb);
 504                         tdb_ftable_init(tdb);
 505                 }
 506                 if (ecode != TDB_SUCCESS) {
 507                         goto fail;
 508                 }
 509                 return tdb;
 510         }
 511
 512         if (stat(name, &st) != -1)
 513                 tdb->file = find_file(st.st_dev, st.st_ino);
 514
 515         if (!tdb->file) {
 516                 int fd;
 517
 518                 if ((fd = open(name, open_flags, mode)) == -1) {
 519                         /* errno set by open(2) */
 520                         saved_errno = errno;
 521                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 522                                    "tdb_open: could not open file %s: %s",
 523                                    name, strerror(errno));
 524                         goto fail_errno;
 525                 }
 526
 527                 /* on exec, don't inherit the fd */
 528                 v = fcntl(fd, F_GETFD, 0);
 529                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 530
 531                 if (fstat(fd, &st) == -1) {
 532                         saved_errno = errno;
 533                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 534                                    "tdb_open: could not stat open %s: %s",
 535                                    name, strerror(errno));
 536                         close(fd);
 537                         goto fail_errno;
 538                 }
 539
 540                 ecode = tdb_new_file(tdb);
 541                 if (ecode != TDB_SUCCESS) {
 542                         close(fd);
 543                         goto fail;
 544                 }
 545
 546                 tdb->file->fd = fd;
 547                 tdb->file->device = st.st_dev;
 548                 tdb->file->inode = st.st_ino;
 549                 tdb->file->map_ptr = NULL;
 550                 tdb->file->map_size = 0;
 551         }
 552
 553         /* ensure there is only one process initialising at once */
 554         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 555         if (ecode != TDB_SUCCESS) {
 556                 saved_errno = errno;
 557                 goto fail_errno;
 558         }
 559
 560         /* call their open hook if they gave us one. */
 561         if (tdb->openhook) {
 562                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
 563                 if (ecode != TDB_SUCCESS) {
 564                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 565                                    "tdb_open: open hook failed");
 566                         goto fail;
 567                 }
 568                 open_flags |= O_CREAT;
 569         }
 570
 571         /* If they used O_TRUNC, read will return 0. */
 572         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 573         if (rlen == 0 && (open_flags & O_CREAT)) {
 574                 ecode = tdb_new_database(tdb, seed, &hdr);
 575                 if (ecode != TDB_SUCCESS) {
 576                         goto fail;
 577                 }
 578         } else if (rlen < 0) {
 579                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 580                                    "tdb_open: error %s reading %s",
 581                                    strerror(errno), name);
 582                 goto fail;
 583         } else if (rlen < sizeof(hdr)
 584                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 585                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 586                                    "tdb_open: %s is not a tdb2 file", name);
 587                 goto fail;
 588         }
 589
 590         if (hdr.version != TDB_VERSION) {
 591                 if (hdr.version == bswap_64(TDB_VERSION))
 592                         tdb->flags |= TDB_CONVERT;
 593                 else {
 594                         /* wrong version */
 595                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 596                                            "tdb_open:"
 597                                            " %s is unknown version 0x%llx",
 598                                            name, (long long)hdr.version);
 599                         goto fail;
 600                 }
 601         } else if (tdb->flags & TDB_CONVERT) {
 602                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 603                                    "tdb_open:"
 604                                    " %s does not need TDB_CONVERT",
 605                                    name);
 606                 goto fail;
 607         }
 608
 609         tdb2_context_init(tdb);
 610
 611         tdb_convert(tdb, &hdr, sizeof(hdr));
 612         tdb->hash_seed = hdr.hash_seed;
 613         hash_test = TDB_HASH_MAGIC;
 614         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 615         if (hdr.hash_test != hash_test) {
 616                 /* wrong hash variant */
 617                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 618                                    "tdb_open:"
 619                                    " %s uses a different hash function",
 620                                    name);
 621                 goto fail;
 622         }
 623
 624         ecode = capabilities_ok(tdb, hdr.capabilities);
 625         if (ecode != TDB_SUCCESS) {
 626                 goto fail;
 627         }
 628
 629         /* Clear any features we don't understand. */
 630         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 631                 hdr.features_used &= TDB_FEATURE_MASK;
 632                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 633                                                         features_used),
 634                                           &hdr.features_used,
 635                                           sizeof(hdr.features_used));
 636                 if (ecode != TDB_SUCCESS)
 637                         goto fail;
 638         }
 639
 640         tdb_unlock_open(tdb, openlock);
 641
 642         /* This makes sure we have current map_size and mmap. */
 643         ecode = tdb->tdb2.io->oob(tdb, tdb->file->map_size, 1, true);
 644         if (unlikely(ecode != TDB_SUCCESS))
 645                 goto fail;
 646
 647         /* Now it's fully formed, recover if necessary. */
 648         berr = tdb_needs_recovery(tdb);
 649         if (unlikely(berr != false)) {
 650                 if (berr < 0) {
 651                         ecode = TDB_OFF_TO_ERR(berr);
 652                         goto fail;
 653                 }
 654                 ecode = tdb_lock_and_recover(tdb);
 655                 if (ecode != TDB_SUCCESS) {
 656                         goto fail;
 657                 }
 658         }
 659
 660         ecode = tdb_ftable_init(tdb);
 661         if (ecode != TDB_SUCCESS) {
 662                 goto fail;
 663         }
 664
 665         tdb->next = tdbs;
 666         tdbs = tdb;
 667         return tdb;
 668
 669  fail:
 670         /* Map ecode to some logical errno. */
 671         switch (TDB_ERR_TO_OFF(ecode)) {
 672         case TDB_ERR_TO_OFF(TDB_ERR_CORRUPT):
 673         case TDB_ERR_TO_OFF(TDB_ERR_IO):
 674                 saved_errno = EIO;
 675                 break;
 676         case TDB_ERR_TO_OFF(TDB_ERR_LOCK):
 677                 saved_errno = EWOULDBLOCK;
 678                 break;
 679         case TDB_ERR_TO_OFF(TDB_ERR_OOM):
 680                 saved_errno = ENOMEM;
 681                 break;
 682         case TDB_ERR_TO_OFF(TDB_ERR_EINVAL):
 683                 saved_errno = EINVAL;
 684                 break;
 685         default:
 686                 saved_errno = EINVAL;
 687                 break;
 688         }
 689
 690 fail_errno:
 691 #ifdef TDB_TRACE
 692         close(tdb->tracefd);
 693 #endif
 694         if (tdb->file) {
 695                 tdb_lock_cleanup(tdb);
 696                 if (--tdb->file->refcnt == 0) {
 697                         assert(tdb->file->num_lockrecs == 0);
 698                         if (tdb->file->map_ptr) {
 699                                 if (tdb->flags & TDB_INTERNAL) {
 700                                         free(tdb->file->map_ptr);
 701                                 } else
 702                                         tdb_munmap(tdb->file);
 703                         }
 704                         if (close(tdb->file->fd) != 0)
 705                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 706                                            "tdb_open: failed to close tdb fd"
 707                                            " on error: %s", strerror(errno));
 708                         free(tdb->file->lockrecs);
 709                         free(tdb->file);
 710                 }
 711         }
 712
 713         free(tdb);
 714         errno = saved_errno;
 715         return NULL;
 716 }
 717
 718 _PUBLIC_ int tdb_close(struct tdb_context *tdb)
 719 {
 720         int ret = 0;
 721         struct tdb_context **i;
 722
 723         tdb_trace(tdb, "tdb_close");
 724
 725         if (tdb->tdb2.transaction) {
 726                 tdb_transaction_cancel(tdb);
 727         }
 728
 729         if (tdb->file->map_ptr) {
 730                 if (tdb->flags & TDB_INTERNAL)
 731                         free(tdb->file->map_ptr);
 732                 else
 733                         tdb_munmap(tdb->file);
 734         }
 735         if (tdb->file) {
 736                 tdb_lock_cleanup(tdb);
 737                 if (--tdb->file->refcnt == 0) {
 738                         ret = close(tdb->file->fd);
 739                         free(tdb->file->lockrecs);
 740                         free(tdb->file);
 741                 }
 742         }
 743
 744         /* Remove from tdbs list */
 745         for (i = &tdbs; *i; i = &(*i)->next) {
 746                 if (*i == tdb) {
 747                         *i = tdb->next;
 748                         break;
 749                 }
 750         }
 751
 752 #ifdef TDB_TRACE
 753         close(tdb->tracefd);
 754 #endif
 755         free(tdb);
 756
 757         return ret;
 758 }
 759
 760 _PUBLIC_ void tdb_foreach_(int (*fn)(struct tdb_context *, void *), void *p)
 761 {
 762         struct tdb_context *i;
 763
 764         for (i = tdbs; i; i = i->next) {
 765                 if (fn(i, p) != 0)
 766                         break;
 767         }
 768 }