git.ozlabs.org Git - ccan/blob - ccan/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 <assert.h>
  20
  21 /* all tdbs, to detect double-opens (fcntl file don't nest!) */
  22 static struct tdb_context *tdbs = NULL;
  23
  24 static struct tdb_file *find_file(dev_t device, ino_t ino)
  25 {
  26         struct tdb_context *i;
  27
  28         for (i = tdbs; i; i = i->next) {
  29                 if (i->file->device == device && i->file->inode == ino) {
  30                         i->file->refcnt++;
  31                         return i->file;
  32                 }
  33         }
  34         return NULL;
  35 }
  36
  37 static bool read_all(int fd, void *buf, size_t len)
  38 {
  39         while (len) {
  40                 ssize_t ret;
  41                 ret = read(fd, buf, len);
  42                 if (ret < 0)
  43                         return false;
  44                 if (ret == 0) {
  45                         /* ETOOSHORT? */
  46                         errno = EWOULDBLOCK;
  47                         return false;
  48                 }
  49                 buf = (char *)buf + ret;
  50                 len -= ret;
  51         }
  52         return true;
  53 }
  54
  55 static uint64_t random_number(struct tdb_context *tdb)
  56 {
  57         int fd;
  58         uint64_t ret = 0;
  59         struct timeval now;
  60
  61         fd = open("/dev/urandom", O_RDONLY);
  62         if (fd >= 0) {
  63                 if (read_all(fd, &ret, sizeof(ret))) {
  64                         close(fd);
  65                         return ret;
  66                 }
  67                 close(fd);
  68         }
  69         /* FIXME: Untested!  Based on Wikipedia protocol description! */
  70         fd = open("/dev/egd-pool", O_RDWR);
  71         if (fd >= 0) {
  72                 /* Command is 1, next byte is size we want to read. */
  73                 char cmd[2] = { 1, sizeof(uint64_t) };
  74                 if (write(fd, cmd, sizeof(cmd)) == sizeof(cmd)) {
  75                         char reply[1 + sizeof(uint64_t)];
  76                         int r = read(fd, reply, sizeof(reply));
  77                         if (r > 1) {
  78                                 /* Copy at least some bytes. */
  79                                 memcpy(&ret, reply+1, r - 1);
  80                                 if (reply[0] == sizeof(uint64_t)
  81                                     && r == sizeof(reply)) {
  82                                         close(fd);
  83                                         return ret;
  84                                 }
  85                         }
  86                 }
  87                 close(fd);
  88         }
  89
  90         /* Fallback: pid and time. */
  91         gettimeofday(&now, NULL);
  92         ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
  93         tdb_logerr(tdb, TDB_SUCCESS, TDB_LOG_WARNING,
  94                    "tdb_open: random from getpid and time");
  95         return ret;
  96 }
  97
  98 struct new_database {
  99         struct tdb_header hdr;
 100         struct tdb_freetable ftable;
 101 };
 102
 103 /* initialise a new database */
 104 static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 105                                        struct tdb_attribute_seed *seed,
 106                                        struct tdb_header *hdr)
 107 {
 108         /* We make it up in memory, then write it out if not internal */
 109         struct new_database newdb;
 110         unsigned int magic_len;
 111         ssize_t rlen;
 112         enum TDB_ERROR ecode;
 113
 114         /* Fill in the header */
 115         newdb.hdr.version = TDB_VERSION;
 116         if (seed)
 117                 newdb.hdr.hash_seed = seed->seed;
 118         else
 119                 newdb.hdr.hash_seed = random_number(tdb);
 120         newdb.hdr.hash_test = TDB_HASH_MAGIC;
 121         newdb.hdr.hash_test = tdb->hash_fn(&newdb.hdr.hash_test,
 122                                            sizeof(newdb.hdr.hash_test),
 123                                            newdb.hdr.hash_seed,
 124                                            tdb->hash_data);
 125         newdb.hdr.recovery = 0;
 126         newdb.hdr.features_used = newdb.hdr.features_offered = TDB_FEATURE_MASK;
 127         newdb.hdr.seqnum = 0;
 128         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
 129         /* Initial hashes are empty. */
 130         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 131
 132         /* Free is empty. */
 133         newdb.hdr.free_table = offsetof(struct new_database, ftable);
 134         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
 135         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
 136                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 137                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 138                            0);
 139         if (ecode != TDB_SUCCESS) {
 140                 return ecode;
 141         }
 142
 143         /* Magic food */
 144         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 145         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 146
 147         /* This creates an endian-converted database, as if read from disk */
 148         magic_len = sizeof(newdb.hdr.magic_food);
 149         tdb_convert(tdb,
 150                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 151
 152         *hdr = newdb.hdr;
 153
 154         if (tdb->flags & TDB_INTERNAL) {
 155                 tdb->file->map_size = sizeof(newdb);
 156                 tdb->file->map_ptr = malloc(tdb->file->map_size);
 157                 if (!tdb->file->map_ptr) {
 158                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 159                                           "tdb_new_database:"
 160                                           " failed to allocate");
 161                 }
 162                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
 163                 return TDB_SUCCESS;
 164         }
 165         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
 166                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 167                                   "tdb_new_database:"
 168                                   " failed to seek: %s", strerror(errno));
 169         }
 170
 171         if (ftruncate(tdb->file->fd, 0) == -1) {
 172                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 173                                   "tdb_new_database:"
 174                                   " failed to truncate: %s", strerror(errno));
 175         }
 176
 177         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
 178         if (rlen != sizeof(newdb)) {
 179                 if (rlen >= 0)
 180                         errno = ENOSPC;
 181                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 182                                   "tdb_new_database: %zi writing header: %s",
 183                                   rlen, strerror(errno));
 184         }
 185         return TDB_SUCCESS;
 186 }
 187
 188 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
 189 {
 190         tdb->file = malloc(sizeof(*tdb->file));
 191         if (!tdb->file)
 192                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 193                                   "tdb_open: cannot alloc tdb_file structure");
 194         tdb->file->num_lockrecs = 0;
 195         tdb->file->lockrecs = NULL;
 196         tdb->file->allrecord_lock.count = 0;
 197         tdb->file->refcnt = 1;
 198         return TDB_SUCCESS;
 199 }
 200
 201 enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
 202                                  const union tdb_attribute *attr)
 203 {
 204         switch (attr->base.attr) {
 205         case TDB_ATTRIBUTE_LOG:
 206                 tdb->log_fn = attr->log.fn;
 207                 tdb->log_data = attr->log.data;
 208                 break;
 209         case TDB_ATTRIBUTE_HASH:
 210         case TDB_ATTRIBUTE_SEED:
 211         case TDB_ATTRIBUTE_OPENHOOK:
 212         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 213                 return tdb->last_error
 214                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 215                                      TDB_LOG_USE_ERROR,
 216                                      "tdb_set_attribute:"
 217                                      " cannot set %s after opening",
 218                                      attr->base.attr == TDB_ATTRIBUTE_HASH
 219                                      ? "TDB_ATTRIBUTE_HASH"
 220                                      : attr->base.attr == TDB_ATTRIBUTE_SEED
 221                                      ? "TDB_ATTRIBUTE_SEED"
 222                                      : attr->base.attr == TDB_ATTRIBUTE_OPENHOOK
 223                                      ? "TDB_ATTRIBUTE_OPENHOOK"
 224                                      : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 225         case TDB_ATTRIBUTE_STATS:
 226                 return tdb->last_error
 227                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 228                                      TDB_LOG_USE_ERROR,
 229                                      "tdb_set_attribute:"
 230                                      " cannot set TDB_ATTRIBUTE_STATS");
 231         case TDB_ATTRIBUTE_FLOCK:
 232                 tdb->lock_fn = attr->flock.lock;
 233                 tdb->unlock_fn = attr->flock.unlock;
 234                 tdb->lock_data = attr->flock.data;
 235                 break;
 236         default:
 237                 return tdb->last_error
 238                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 239                                      TDB_LOG_USE_ERROR,
 240                                      "tdb_set_attribute:"
 241                                      " unknown attribute type %u",
 242                                      attr->base.attr);
 243         }
 244         return TDB_SUCCESS;
 245 }
 246
 247 enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
 248                                  union tdb_attribute *attr)
 249 {
 250         switch (attr->base.attr) {
 251         case TDB_ATTRIBUTE_LOG:
 252                 if (!tdb->log_fn)
 253                         return tdb->last_error = TDB_ERR_NOEXIST;
 254                 attr->log.fn = tdb->log_fn;
 255                 attr->log.data = tdb->log_data;
 256                 break;
 257         case TDB_ATTRIBUTE_HASH:
 258                 attr->hash.fn = tdb->hash_fn;
 259                 attr->hash.data = tdb->hash_data;
 260                 break;
 261         case TDB_ATTRIBUTE_SEED:
 262                 if (tdb->flags & TDB_VERSION1)
 263                         return tdb->last_error
 264                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 265                                              TDB_LOG_USE_ERROR,
 266                                      "tdb_get_attribute:"
 267                                      " cannot get TDB_ATTRIBUTE_SEED"
 268                                      " on TDB1 tdb.");
 269                 attr->seed.seed = tdb->hash_seed;
 270                 break;
 271         case TDB_ATTRIBUTE_OPENHOOK:
 272                 if (!tdb->openhook)
 273                         return tdb->last_error = TDB_ERR_NOEXIST;
 274                 attr->openhook.fn = tdb->openhook;
 275                 attr->openhook.data = tdb->openhook_data;
 276                 break;
 277         case TDB_ATTRIBUTE_STATS: {
 278                 size_t size = attr->stats.size;
 279                 if (size > tdb->stats.size)
 280                         size = tdb->stats.size;
 281                 memcpy(&attr->stats, &tdb->stats, size);
 282                 break;
 283         }
 284         case TDB_ATTRIBUTE_FLOCK:
 285                 attr->flock.lock = tdb->lock_fn;
 286                 attr->flock.unlock = tdb->unlock_fn;
 287                 attr->flock.data = tdb->lock_data;
 288                 break;
 289         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 290                 if (!(tdb->flags & TDB_VERSION1))
 291                         return tdb->last_error
 292                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 293                                              TDB_LOG_USE_ERROR,
 294                                      "tdb_get_attribute:"
 295                                      " cannot get TDB_ATTRIBUTE_TDB1_HASHSIZE"
 296                                      " on TDB2 tdb.");
 297                 attr->tdb1_hashsize.hsize = tdb->tdb1.header.hash_size;
 298                 break;
 299         default:
 300                 return tdb->last_error
 301                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 302                                      TDB_LOG_USE_ERROR,
 303                                      "tdb_get_attribute:"
 304                                      " unknown attribute type %u",
 305                                      attr->base.attr);
 306         }
 307         attr->base.next = NULL;
 308         return TDB_SUCCESS;
 309 }
 310
 311 void tdb_unset_attribute(struct tdb_context *tdb,
 312                          enum tdb_attribute_type type)
 313 {
 314         switch (type) {
 315         case TDB_ATTRIBUTE_LOG:
 316                 tdb->log_fn = NULL;
 317                 break;
 318         case TDB_ATTRIBUTE_OPENHOOK:
 319                 tdb->openhook = NULL;
 320                 break;
 321         case TDB_ATTRIBUTE_HASH:
 322         case TDB_ATTRIBUTE_SEED:
 323         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 324                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 325                            "tdb_unset_attribute: cannot unset %s after opening",
 326                            type == TDB_ATTRIBUTE_HASH
 327                            ? "TDB_ATTRIBUTE_HASH"
 328                            : type == TDB_ATTRIBUTE_SEED
 329                            ? "TDB_ATTRIBUTE_SEED"
 330                            : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 331                 break;
 332         case TDB_ATTRIBUTE_STATS:
 333                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 334                            TDB_LOG_USE_ERROR,
 335                            "tdb_unset_attribute:"
 336                            "cannot unset TDB_ATTRIBUTE_STATS");
 337                 break;
 338         case TDB_ATTRIBUTE_FLOCK:
 339                 tdb->lock_fn = tdb_fcntl_lock;
 340                 tdb->unlock_fn = tdb_fcntl_unlock;
 341                 break;
 342         default:
 343                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 344                            TDB_LOG_USE_ERROR,
 345                            "tdb_unset_attribute: unknown attribute type %u",
 346                            type);
 347         }
 348 }
 349
 350 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 351                              int open_flags, mode_t mode,
 352                              union tdb_attribute *attr)
 353 {
 354         struct tdb_context *tdb;
 355         struct stat st;
 356         int saved_errno = 0;
 357         uint64_t hash_test;
 358         unsigned v;
 359         ssize_t rlen;
 360         struct tdb_header hdr;
 361         struct tdb_attribute_seed *seed = NULL;
 362         struct tdb_attribute_tdb1_hashsize *hsize_attr = NULL;
 363         tdb_bool_err berr;
 364         enum TDB_ERROR ecode;
 365         int openlock;
 366
 367         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 368         if (!tdb) {
 369                 /* Can't log this */
 370                 errno = ENOMEM;
 371                 return NULL;
 372         }
 373         /* Set name immediately for logging functions. */
 374         if (name) {
 375                 tdb->name = strcpy((char *)(tdb + 1), name);
 376         } else {
 377                 tdb->name = NULL;
 378         }
 379         tdb->flags = tdb_flags;
 380         tdb->log_fn = NULL;
 381         tdb->open_flags = open_flags;
 382         tdb->last_error = TDB_SUCCESS;
 383         tdb->file = NULL;
 384         tdb->openhook = NULL;
 385         tdb->lock_fn = tdb_fcntl_lock;
 386         tdb->unlock_fn = tdb_fcntl_unlock;
 387         tdb->hash_fn = tdb_jenkins_hash;
 388         memset(&tdb->stats, 0, sizeof(tdb->stats));
 389         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
 390         tdb->stats.size = sizeof(tdb->stats);
 391         tdb_io_init(tdb);
 392         tdb->tdb2.direct_access = 0;
 393         tdb->tdb2.transaction = NULL;
 394         tdb->tdb2.access = NULL;
 395
 396         while (attr) {
 397                 switch (attr->base.attr) {
 398                 case TDB_ATTRIBUTE_HASH:
 399                         tdb->hash_fn = attr->hash.fn;
 400                         tdb->hash_data = attr->hash.data;
 401                         break;
 402                 case TDB_ATTRIBUTE_SEED:
 403                         seed = &attr->seed;
 404                         break;
 405                 case TDB_ATTRIBUTE_OPENHOOK:
 406                         tdb->openhook = attr->openhook.fn;
 407                         tdb->openhook_data = attr->openhook.data;
 408                         break;
 409                 case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 410                         hsize_attr = &attr->tdb1_hashsize;
 411                         break;
 412                 default:
 413                         /* These are set as normal. */
 414                         ecode = tdb_set_attribute(tdb, attr);
 415                         if (ecode != TDB_SUCCESS)
 416                                 goto fail;
 417                 }
 418                 attr = attr->base.next;
 419         }
 420
 421         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 422                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
 423                           | TDB_RDONLY)) {
 424                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 425                                    "tdb_open: unknown flags %u", tdb_flags);
 426                 goto fail;
 427         }
 428
 429         if (hsize_attr) {
 430                 if (!(tdb_flags & TDB_VERSION1) ||
 431                     (!(tdb_flags & TDB_INTERNAL) && !(open_flags & O_CREAT))) {
 432                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 433                                            TDB_LOG_USE_ERROR,
 434                                            "tdb_open: can only use"
 435                                            " TDB_ATTRIBUTE_TDB1_HASHSIZE when"
 436                                            " creating a TDB_VERSION1 tdb");
 437                         goto fail;
 438                 }
 439         }
 440
 441         if (seed) {
 442                 if (tdb_flags & TDB_VERSION1) {
 443                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 444                                            TDB_LOG_USE_ERROR,
 445                                            "tdb_open:"
 446                                            " cannot set TDB_ATTRIBUTE_SEED"
 447                                            " on TDB1 tdb.");
 448                         goto fail;
 449                 } else if (!(tdb_flags & TDB_INTERNAL)
 450                            && !(open_flags & O_CREAT)) {
 451                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 452                                            TDB_LOG_USE_ERROR,
 453                                            "tdb_open:"
 454                                            " cannot set TDB_ATTRIBUTE_SEED"
 455                                            " without O_CREAT.");
 456                         goto fail;
 457                 }
 458         }
 459
 460         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 461                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 462                                    "tdb_open: can't open tdb %s write-only",
 463                                    name);
 464                 goto fail;
 465         }
 466
 467         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 468                 openlock = F_RDLCK;
 469                 tdb->flags |= TDB_RDONLY;
 470         } else {
 471                 if (tdb_flags & TDB_RDONLY) {
 472                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 473                                            TDB_LOG_USE_ERROR,
 474                                            "tdb_open: can't use TDB_RDONLY"
 475                                            " without O_RDONLY");
 476                         goto fail;
 477                 }
 478                 openlock = F_WRLCK;
 479         }
 480
 481         /* internal databases don't need any of the rest. */
 482         if (tdb->flags & TDB_INTERNAL) {
 483                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 484                 ecode = tdb_new_file(tdb);
 485                 if (ecode != TDB_SUCCESS) {
 486                         goto fail;
 487                 }
 488                 tdb->file->fd = -1;
 489                 ecode = tdb_new_database(tdb, seed, &hdr);
 490                 if (ecode != TDB_SUCCESS) {
 491                         goto fail;
 492                 }
 493                 tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
 494                 tdb->hash_seed = hdr.hash_seed;
 495                 tdb_ftable_init(tdb);
 496                 return tdb;
 497         }
 498
 499         if (stat(name, &st) != -1)
 500                 tdb->file = find_file(st.st_dev, st.st_ino);
 501
 502         if (!tdb->file) {
 503                 int fd;
 504
 505                 if ((fd = open(name, open_flags, mode)) == -1) {
 506                         /* errno set by open(2) */
 507                         saved_errno = errno;
 508                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 509                                    "tdb_open: could not open file %s: %s",
 510                                    name, strerror(errno));
 511                         goto fail_errno;
 512                 }
 513
 514                 /* on exec, don't inherit the fd */
 515                 v = fcntl(fd, F_GETFD, 0);
 516                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 517
 518                 if (fstat(fd, &st) == -1) {
 519                         saved_errno = errno;
 520                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 521                                    "tdb_open: could not stat open %s: %s",
 522                                    name, strerror(errno));
 523                         close(fd);
 524                         goto fail_errno;
 525                 }
 526
 527                 ecode = tdb_new_file(tdb);
 528                 if (ecode != TDB_SUCCESS) {
 529                         close(fd);
 530                         goto fail;
 531                 }
 532
 533                 tdb->file->fd = fd;
 534                 tdb->file->device = st.st_dev;
 535                 tdb->file->inode = st.st_ino;
 536                 tdb->file->map_ptr = NULL;
 537                 tdb->file->map_size = sizeof(struct tdb_header);
 538         }
 539
 540         /* ensure there is only one process initialising at once */
 541         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 542         if (ecode != TDB_SUCCESS) {
 543                 saved_errno = errno;
 544                 goto fail_errno;
 545         }
 546
 547         /* call their open hook if they gave us one. */
 548         if (tdb->openhook) {
 549                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
 550                 if (ecode != TDB_SUCCESS) {
 551                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 552                                    "tdb_open: open hook failed");
 553                         goto fail;
 554                 }
 555                 open_flags |= O_CREAT;
 556         }
 557
 558         /* If they used O_TRUNC, read will return 0. */
 559         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 560         if (rlen == 0 && (open_flags & O_CREAT)) {
 561                 ecode = tdb_new_database(tdb, seed, &hdr);
 562                 if (ecode != TDB_SUCCESS) {
 563                         goto fail;
 564                 }
 565         } else if (rlen < 0) {
 566                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 567                                    "tdb_open: error %s reading %s",
 568                                    strerror(errno), name);
 569                 goto fail;
 570         } else if (rlen < sizeof(hdr)
 571                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 572                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 573                                    "tdb_open: %s is not a tdb file", name);
 574                 goto fail;
 575         }
 576
 577         if (hdr.version != TDB_VERSION) {
 578                 if (hdr.version == bswap_64(TDB_VERSION))
 579                         tdb->flags |= TDB_CONVERT;
 580                 else {
 581                         /* wrong version */
 582                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 583                                            "tdb_open:"
 584                                            " %s is unknown version 0x%llx",
 585                                            name, (long long)hdr.version);
 586                         goto fail;
 587                 }
 588         } else if (tdb->flags & TDB_CONVERT) {
 589                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 590                                    "tdb_open:"
 591                                    " %s does not need TDB_CONVERT",
 592                                    name);
 593                 goto fail;
 594         }
 595
 596         tdb_convert(tdb, &hdr, sizeof(hdr));
 597         tdb->hash_seed = hdr.hash_seed;
 598         hash_test = TDB_HASH_MAGIC;
 599         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 600         if (hdr.hash_test != hash_test) {
 601                 /* wrong hash variant */
 602                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 603                                    "tdb_open:"
 604                                    " %s uses a different hash function",
 605                                    name);
 606                 goto fail;
 607         }
 608
 609         /* Clear any features we don't understand. */
 610         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 611                 hdr.features_used &= TDB_FEATURE_MASK;
 612                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 613                                                         features_used),
 614                                           &hdr.features_used,
 615                                           sizeof(hdr.features_used));
 616                 if (ecode != TDB_SUCCESS)
 617                         goto fail;
 618         }
 619
 620         tdb_unlock_open(tdb, openlock);
 621
 622         /* This make sure we have current map_size and mmap. */
 623         ecode = tdb->tdb2.io->oob(tdb, tdb->file->map_size + 1, true);
 624         if (unlikely(ecode != TDB_SUCCESS))
 625                 goto fail;
 626
 627         /* Now it's fully formed, recover if necessary. */
 628         berr = tdb_needs_recovery(tdb);
 629         if (unlikely(berr != false)) {
 630                 if (berr < 0) {
 631                         ecode = berr;
 632                         goto fail;
 633                 }
 634                 ecode = tdb_lock_and_recover(tdb);
 635                 if (ecode != TDB_SUCCESS) {
 636                         goto fail;
 637                 }
 638         }
 639
 640         ecode = tdb_ftable_init(tdb);
 641         if (ecode != TDB_SUCCESS) {
 642                 goto fail;
 643         }
 644
 645         tdb->next = tdbs;
 646         tdbs = tdb;
 647         return tdb;
 648
 649  fail:
 650         /* Map ecode to some logical errno. */
 651         switch (ecode) {
 652         case TDB_ERR_CORRUPT:
 653         case TDB_ERR_IO:
 654                 saved_errno = EIO;
 655                 break;
 656         case TDB_ERR_LOCK:
 657                 saved_errno = EWOULDBLOCK;
 658                 break;
 659         case TDB_ERR_OOM:
 660                 saved_errno = ENOMEM;
 661                 break;
 662         case TDB_ERR_EINVAL:
 663                 saved_errno = EINVAL;
 664                 break;
 665         default:
 666                 saved_errno = EINVAL;
 667                 break;
 668         }
 669
 670 fail_errno:
 671 #ifdef TDB_TRACE
 672         close(tdb->tracefd);
 673 #endif
 674         if (tdb->file) {
 675                 tdb_lock_cleanup(tdb);
 676                 if (--tdb->file->refcnt == 0) {
 677                         assert(tdb->file->num_lockrecs == 0);
 678                         if (tdb->file->map_ptr) {
 679                                 if (tdb->flags & TDB_INTERNAL) {
 680                                         free(tdb->file->map_ptr);
 681                                 } else
 682                                         tdb_munmap(tdb->file);
 683                         }
 684                         if (close(tdb->file->fd) != 0)
 685                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 686                                            "tdb_open: failed to close tdb fd"
 687                                            " on error: %s", strerror(errno));
 688                         free(tdb->file->lockrecs);
 689                         free(tdb->file);
 690                 }
 691         }
 692
 693         free(tdb);
 694         errno = saved_errno;
 695         return NULL;
 696 }
 697
 698 int tdb_close(struct tdb_context *tdb)
 699 {
 700         int ret = 0;
 701         struct tdb_context **i;
 702
 703         tdb_trace(tdb, "tdb_close");
 704
 705         if (tdb->tdb2.transaction) {
 706                 tdb_transaction_cancel(tdb);
 707         }
 708
 709         if (tdb->file->map_ptr) {
 710                 if (tdb->flags & TDB_INTERNAL)
 711                         free(tdb->file->map_ptr);
 712                 else
 713                         tdb_munmap(tdb->file);
 714         }
 715         if (tdb->file) {
 716                 tdb_lock_cleanup(tdb);
 717                 if (--tdb->file->refcnt == 0) {
 718                         ret = close(tdb->file->fd);
 719                         free(tdb->file->lockrecs);
 720                         free(tdb->file);
 721                 }
 722         }
 723
 724         /* Remove from tdbs list */
 725         for (i = &tdbs; *i; i = &(*i)->next) {
 726                 if (*i == tdb) {
 727                         *i = tdb->next;
 728                         break;
 729                 }
 730         }
 731
 732 #ifdef TDB_TRACE
 733         close(tdb->tracefd);
 734 #endif
 735         free(tdb);
 736
 737         return ret;
 738 }
 739
 740 void tdb_foreach_(int (*fn)(struct tdb_context *, void *), void *p)
 741 {
 742         struct tdb_context *i;
 743
 744         for (i = tdbs; i; i = i->next) {
 745                 if (fn(i, p) != 0)
 746                         break;
 747         }
 748 }