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 <ccan/hash/hash.h>
  20 #include <assert.h>
  21
  22 /* all lock info, to detect double-opens (fcntl file don't nest!) */
  23 static struct tdb_file *files = NULL;
  24
  25 static struct tdb_file *find_file(dev_t device, ino_t ino)
  26 {
  27         struct tdb_file *i;
  28
  29         for (i = files; i; i = i->next) {
  30                 if (i->device == device && i->inode == ino) {
  31                         i->refcnt++;
  32                         break;
  33                 }
  34         }
  35         return i;
  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 struct new_database {
 100         struct tdb_header hdr;
 101         struct tdb_freetable ftable;
 102 };
 103
 104 /* initialise a new database */
 105 static enum TDB_ERROR tdb_new_database(struct tdb_context *tdb,
 106                                        struct tdb_attribute_seed *seed,
 107                                        struct tdb_header *hdr)
 108 {
 109         /* We make it up in memory, then write it out if not internal */
 110         struct new_database newdb;
 111         unsigned int magic_len;
 112         ssize_t rlen;
 113         enum TDB_ERROR ecode;
 114
 115         /* Fill in the header */
 116         newdb.hdr.version = TDB_VERSION;
 117         if (seed)
 118                 newdb.hdr.hash_seed = seed->seed;
 119         else
 120                 newdb.hdr.hash_seed = random_number(tdb);
 121         newdb.hdr.hash_test = TDB_HASH_MAGIC;
 122         newdb.hdr.hash_test = tdb->hash_fn(&newdb.hdr.hash_test,
 123                                            sizeof(newdb.hdr.hash_test),
 124                                            newdb.hdr.hash_seed,
 125                                            tdb->hash_data);
 126         newdb.hdr.recovery = 0;
 127         newdb.hdr.features_used = newdb.hdr.features_offered = TDB_FEATURE_MASK;
 128         newdb.hdr.seqnum = 0;
 129         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
 130         /* Initial hashes are empty. */
 131         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 132
 133         /* Free is empty. */
 134         newdb.hdr.free_table = offsetof(struct new_database, ftable);
 135         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
 136         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
 137                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 138                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 139                            0);
 140         if (ecode != TDB_SUCCESS) {
 141                 return ecode;
 142         }
 143
 144         /* Magic food */
 145         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 146         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 147
 148         /* This creates an endian-converted database, as if read from disk */
 149         magic_len = sizeof(newdb.hdr.magic_food);
 150         tdb_convert(tdb,
 151                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 152
 153         *hdr = newdb.hdr;
 154
 155         if (tdb->flags & TDB_INTERNAL) {
 156                 tdb->file->map_size = sizeof(newdb);
 157                 tdb->file->map_ptr = malloc(tdb->file->map_size);
 158                 if (!tdb->file->map_ptr) {
 159                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 160                                           "tdb_new_database:"
 161                                           " failed to allocate");
 162                 }
 163                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
 164                 return TDB_SUCCESS;
 165         }
 166         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
 167                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 168                                   "tdb_new_database:"
 169                                   " failed to seek: %s", strerror(errno));
 170         }
 171
 172         if (ftruncate(tdb->file->fd, 0) == -1) {
 173                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 174                                   "tdb_new_database:"
 175                                   " failed to truncate: %s", strerror(errno));
 176         }
 177
 178         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
 179         if (rlen != sizeof(newdb)) {
 180                 if (rlen >= 0)
 181                         errno = ENOSPC;
 182                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 183                                   "tdb_new_database: %zi writing header: %s",
 184                                   rlen, strerror(errno));
 185         }
 186         return TDB_SUCCESS;
 187 }
 188
 189 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
 190 {
 191         tdb->file = malloc(sizeof(*tdb->file));
 192         if (!tdb->file)
 193                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 194                                   "tdb_open: cannot alloc tdb_file structure");
 195         tdb->file->num_lockrecs = 0;
 196         tdb->file->lockrecs = NULL;
 197         tdb->file->allrecord_lock.count = 0;
 198         tdb->file->refcnt = 1;
 199         return TDB_SUCCESS;
 200 }
 201
 202 enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
 203                                  const union tdb_attribute *attr)
 204 {
 205         switch (attr->base.attr) {
 206         case TDB_ATTRIBUTE_LOG:
 207                 tdb->log_fn = attr->log.fn;
 208                 tdb->log_data = attr->log.data;
 209                 break;
 210         case TDB_ATTRIBUTE_HASH:
 211         case TDB_ATTRIBUTE_SEED:
 212         case TDB_ATTRIBUTE_OPENHOOK:
 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                                      : "TDB_ATTRIBUTE_OPENHOOK");
 223         case TDB_ATTRIBUTE_STATS:
 224                 return tdb->last_error
 225                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 226                                      TDB_LOG_USE_ERROR,
 227                                      "tdb_set_attribute:"
 228                                      " cannot set TDB_ATTRIBUTE_STATS");
 229         case TDB_ATTRIBUTE_FLOCK:
 230                 tdb->lock_fn = attr->flock.lock;
 231                 tdb->unlock_fn = attr->flock.unlock;
 232                 tdb->lock_data = attr->flock.data;
 233                 break;
 234         default:
 235                 return tdb->last_error
 236                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 237                                      TDB_LOG_USE_ERROR,
 238                                      "tdb_set_attribute:"
 239                                      " unknown attribute type %u",
 240                                      attr->base.attr);
 241         }
 242         return TDB_SUCCESS;
 243 }
 244
 245 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
 246                              void *unused)
 247 {
 248         uint64_t ret;
 249         /* hash64_stable assumes lower bits are more important; they are a
 250          * slightly better hash.  We use the upper bits first, so swap them. */
 251         ret = hash64_stable((const unsigned char *)key, length, seed);
 252         return (ret >> 32) | (ret << 32);
 253 }
 254
 255 enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
 256                                  union tdb_attribute *attr)
 257 {
 258         switch (attr->base.attr) {
 259         case TDB_ATTRIBUTE_LOG:
 260                 if (!tdb->log_fn)
 261                         return tdb->last_error = TDB_ERR_NOEXIST;
 262                 attr->log.fn = tdb->log_fn;
 263                 attr->log.data = tdb->log_data;
 264                 break;
 265         case TDB_ATTRIBUTE_HASH:
 266                 attr->hash.fn = tdb->hash_fn;
 267                 attr->hash.data = tdb->hash_data;
 268                 break;
 269         case TDB_ATTRIBUTE_SEED:
 270                 attr->seed.seed = tdb->hash_seed;
 271                 break;
 272         case TDB_ATTRIBUTE_OPENHOOK:
 273                 if (!tdb->openhook)
 274                         return tdb->last_error = TDB_ERR_NOEXIST;
 275                 attr->openhook.fn = tdb->openhook;
 276                 attr->openhook.data = tdb->openhook_data;
 277                 break;
 278         case TDB_ATTRIBUTE_STATS: {
 279                 size_t size = attr->stats.size;
 280                 if (size > tdb->stats.size)
 281                         size = tdb->stats.size;
 282                 memcpy(&attr->stats, &tdb->stats, size);
 283                 break;
 284         }
 285         case TDB_ATTRIBUTE_FLOCK:
 286                 attr->flock.lock = tdb->lock_fn;
 287                 attr->flock.unlock = tdb->unlock_fn;
 288                 attr->flock.data = tdb->lock_data;
 289                 break;
 290         default:
 291                 return tdb->last_error
 292                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 293                                      TDB_LOG_USE_ERROR,
 294                                      "tdb_get_attribute:"
 295                                      " unknown attribute type %u",
 296                                      attr->base.attr);
 297         }
 298         attr->base.next = NULL;
 299         return TDB_SUCCESS;
 300 }
 301
 302 void tdb_unset_attribute(struct tdb_context *tdb,
 303                          enum tdb_attribute_type type)
 304 {
 305         switch (type) {
 306         case TDB_ATTRIBUTE_LOG:
 307                 tdb->log_fn = NULL;
 308                 break;
 309         case TDB_ATTRIBUTE_OPENHOOK:
 310                 tdb->openhook = NULL;
 311                 break;
 312         case TDB_ATTRIBUTE_HASH:
 313         case TDB_ATTRIBUTE_SEED:
 314                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 315                            "tdb_unset_attribute: cannot unset %s after opening",
 316                            type == TDB_ATTRIBUTE_HASH
 317                            ? "TDB_ATTRIBUTE_HASH"
 318                            : "TDB_ATTRIBUTE_SEED");
 319                 break;
 320         case TDB_ATTRIBUTE_STATS:
 321                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 322                            TDB_LOG_USE_ERROR,
 323                            "tdb_unset_attribute:"
 324                            "cannot unset TDB_ATTRIBUTE_STATS");
 325                 break;
 326         case TDB_ATTRIBUTE_FLOCK:
 327                 tdb->lock_fn = tdb_fcntl_lock;
 328                 tdb->unlock_fn = tdb_fcntl_unlock;
 329                 break;
 330         default:
 331                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 332                            TDB_LOG_USE_ERROR,
 333                            "tdb_unset_attribute: unknown attribute type %u",
 334                            type);
 335         }
 336 }
 337
 338 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 339                              int open_flags, mode_t mode,
 340                              union tdb_attribute *attr)
 341 {
 342         struct tdb_context *tdb;
 343         struct stat st;
 344         int saved_errno = 0;
 345         uint64_t hash_test;
 346         unsigned v;
 347         ssize_t rlen;
 348         struct tdb_header hdr;
 349         struct tdb_attribute_seed *seed = NULL;
 350         tdb_bool_err berr;
 351         enum TDB_ERROR ecode;
 352         int openlock;
 353
 354         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 355         if (!tdb) {
 356                 /* Can't log this */
 357                 errno = ENOMEM;
 358                 return NULL;
 359         }
 360         /* Set name immediately for logging functions. */
 361         if (name) {
 362                 tdb->name = strcpy((char *)(tdb + 1), name);
 363         } else {
 364                 tdb->name = NULL;
 365         }
 366         tdb->direct_access = 0;
 367         tdb->flags = tdb_flags;
 368         tdb->log_fn = NULL;
 369         tdb->transaction = NULL;
 370         tdb->access = NULL;
 371         tdb->open_flags = open_flags;
 372         tdb->last_error = TDB_SUCCESS;
 373         tdb->file = NULL;
 374         tdb->openhook = NULL;
 375         tdb->lock_fn = tdb_fcntl_lock;
 376         tdb->unlock_fn = tdb_fcntl_unlock;
 377         tdb->hash_fn = jenkins_hash;
 378         memset(&tdb->stats, 0, sizeof(tdb->stats));
 379         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
 380         tdb->stats.size = sizeof(tdb->stats);
 381         tdb_io_init(tdb);
 382
 383         while (attr) {
 384                 switch (attr->base.attr) {
 385                 case TDB_ATTRIBUTE_HASH:
 386                         tdb->hash_fn = attr->hash.fn;
 387                         tdb->hash_data = attr->hash.data;
 388                         break;
 389                 case TDB_ATTRIBUTE_SEED:
 390                         seed = &attr->seed;
 391                         break;
 392                 case TDB_ATTRIBUTE_OPENHOOK:
 393                         tdb->openhook = attr->openhook.fn;
 394                         tdb->openhook_data = attr->openhook.data;
 395                         break;
 396                 default:
 397                         /* These are set as normal. */
 398                         ecode = tdb_set_attribute(tdb, attr);
 399                         if (ecode != TDB_SUCCESS)
 400                                 goto fail;
 401                 }
 402                 attr = attr->base.next;
 403         }
 404
 405         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 406                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
 407                           | TDB_RDONLY)) {
 408                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 409                                    "tdb_open: unknown flags %u", tdb_flags);
 410                 goto fail;
 411         }
 412
 413         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 414                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 415                                    "tdb_open: can't open tdb %s write-only",
 416                                    name);
 417                 goto fail;
 418         }
 419
 420         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 421                 openlock = F_RDLCK;
 422                 tdb->flags |= TDB_RDONLY;
 423         } else {
 424                 if (tdb_flags & TDB_RDONLY) {
 425                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 426                                            TDB_LOG_USE_ERROR,
 427                                            "tdb_open: can't use TDB_RDONLY"
 428                                            " without O_RDONLY");
 429                         goto fail;
 430                 }
 431                 openlock = F_WRLCK;
 432         }
 433
 434         /* internal databases don't need any of the rest. */
 435         if (tdb->flags & TDB_INTERNAL) {
 436                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 437                 ecode = tdb_new_file(tdb);
 438                 if (ecode != TDB_SUCCESS) {
 439                         goto fail;
 440                 }
 441                 tdb->file->fd = -1;
 442                 ecode = tdb_new_database(tdb, seed, &hdr);
 443                 if (ecode != TDB_SUCCESS) {
 444                         goto fail;
 445                 }
 446                 tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
 447                 tdb->hash_seed = hdr.hash_seed;
 448                 tdb_ftable_init(tdb);
 449                 return tdb;
 450         }
 451
 452         if (stat(name, &st) != -1)
 453                 tdb->file = find_file(st.st_dev, st.st_ino);
 454
 455         if (!tdb->file) {
 456                 int fd;
 457
 458                 if ((fd = open(name, open_flags, mode)) == -1) {
 459                         /* errno set by open(2) */
 460                         saved_errno = errno;
 461                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 462                                    "tdb_open: could not open file %s: %s",
 463                                    name, strerror(errno));
 464                         goto fail_errno;
 465                 }
 466
 467                 /* on exec, don't inherit the fd */
 468                 v = fcntl(fd, F_GETFD, 0);
 469                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 470
 471                 if (fstat(fd, &st) == -1) {
 472                         saved_errno = errno;
 473                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 474                                    "tdb_open: could not stat open %s: %s",
 475                                    name, strerror(errno));
 476                         close(fd);
 477                         goto fail_errno;
 478                 }
 479
 480                 ecode = tdb_new_file(tdb);
 481                 if (ecode != TDB_SUCCESS) {
 482                         close(fd);
 483                         goto fail;
 484                 }
 485
 486                 tdb->file->next = files;
 487                 tdb->file->fd = fd;
 488                 tdb->file->device = st.st_dev;
 489                 tdb->file->inode = st.st_ino;
 490                 tdb->file->map_ptr = NULL;
 491                 tdb->file->map_size = sizeof(struct tdb_header);
 492         }
 493
 494         /* ensure there is only one process initialising at once */
 495         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 496         if (ecode != TDB_SUCCESS) {
 497                 saved_errno = errno;
 498                 goto fail_errno;
 499         }
 500
 501         /* call their open hook if they gave us one. */
 502         if (tdb->openhook) {
 503                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
 504                 if (ecode != TDB_SUCCESS) {
 505                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 506                                    "tdb_open: open hook failed");
 507                         goto fail;
 508                 }
 509                 open_flags |= O_CREAT;
 510         }
 511
 512         /* If they used O_TRUNC, read will return 0. */
 513         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 514         if (rlen == 0 && (open_flags & O_CREAT)) {
 515                 ecode = tdb_new_database(tdb, seed, &hdr);
 516                 if (ecode != TDB_SUCCESS) {
 517                         goto fail;
 518                 }
 519         } else if (rlen < 0) {
 520                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 521                                    "tdb_open: error %s reading %s",
 522                                    strerror(errno), name);
 523                 goto fail;
 524         } else if (rlen < sizeof(hdr)
 525                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 526                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 527                                    "tdb_open: %s is not a tdb file", name);
 528                 goto fail;
 529         }
 530
 531         if (hdr.version != TDB_VERSION) {
 532                 if (hdr.version == bswap_64(TDB_VERSION))
 533                         tdb->flags |= TDB_CONVERT;
 534                 else {
 535                         /* wrong version */
 536                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 537                                            "tdb_open:"
 538                                            " %s is unknown version 0x%llx",
 539                                            name, (long long)hdr.version);
 540                         goto fail;
 541                 }
 542         } else if (tdb->flags & TDB_CONVERT) {
 543                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 544                                    "tdb_open:"
 545                                    " %s does not need TDB_CONVERT",
 546                                    name);
 547                 goto fail;
 548         }
 549
 550         tdb_convert(tdb, &hdr, sizeof(hdr));
 551         tdb->hash_seed = hdr.hash_seed;
 552         hash_test = TDB_HASH_MAGIC;
 553         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 554         if (hdr.hash_test != hash_test) {
 555                 /* wrong hash variant */
 556                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 557                                    "tdb_open:"
 558                                    " %s uses a different hash function",
 559                                    name);
 560                 goto fail;
 561         }
 562
 563         /* Clear any features we don't understand. */
 564         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 565                 hdr.features_used &= TDB_FEATURE_MASK;
 566                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 567                                                         features_used),
 568                                           &hdr.features_used,
 569                                           sizeof(hdr.features_used));
 570                 if (ecode != TDB_SUCCESS)
 571                         goto fail;
 572         }
 573
 574         tdb_unlock_open(tdb, openlock);
 575
 576         /* This make sure we have current map_size and mmap. */
 577         ecode = tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
 578         if (unlikely(ecode != TDB_SUCCESS))
 579                 goto fail;
 580
 581         /* Now it's fully formed, recover if necessary. */
 582         berr = tdb_needs_recovery(tdb);
 583         if (unlikely(berr != false)) {
 584                 if (berr < 0) {
 585                         ecode = berr;
 586                         goto fail;
 587                 }
 588                 ecode = tdb_lock_and_recover(tdb);
 589                 if (ecode != TDB_SUCCESS) {
 590                         goto fail;
 591                 }
 592         }
 593
 594         ecode = tdb_ftable_init(tdb);
 595         if (ecode != TDB_SUCCESS) {
 596                 goto fail;
 597         }
 598
 599         /* Add to linked list if we're new. */
 600         if (tdb->file->refcnt == 1)
 601                 files = tdb->file;
 602         return tdb;
 603
 604  fail:
 605         /* Map ecode to some logical errno. */
 606         switch (ecode) {
 607         case TDB_ERR_CORRUPT:
 608         case TDB_ERR_IO:
 609                 saved_errno = EIO;
 610                 break;
 611         case TDB_ERR_LOCK:
 612                 saved_errno = EWOULDBLOCK;
 613                 break;
 614         case TDB_ERR_OOM:
 615                 saved_errno = ENOMEM;
 616                 break;
 617         case TDB_ERR_EINVAL:
 618                 saved_errno = EINVAL;
 619                 break;
 620         default:
 621                 saved_errno = EINVAL;
 622                 break;
 623         }
 624
 625 fail_errno:
 626 #ifdef TDB_TRACE
 627         close(tdb->tracefd);
 628 #endif
 629         if (tdb->file) {
 630                 tdb_lock_cleanup(tdb);
 631                 if (--tdb->file->refcnt == 0) {
 632                         assert(tdb->file->num_lockrecs == 0);
 633                         if (tdb->file->map_ptr) {
 634                                 if (tdb->flags & TDB_INTERNAL) {
 635                                         free(tdb->file->map_ptr);
 636                                 } else
 637                                         tdb_munmap(tdb->file);
 638                         }
 639                         if (close(tdb->file->fd) != 0)
 640                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 641                                            "tdb_open: failed to close tdb fd"
 642                                            " on error: %s", strerror(errno));
 643                         free(tdb->file->lockrecs);
 644                         free(tdb->file);
 645                 }
 646         }
 647
 648         free(tdb);
 649         errno = saved_errno;
 650         return NULL;
 651 }
 652
 653 int tdb_close(struct tdb_context *tdb)
 654 {
 655         int ret = 0;
 656
 657         tdb_trace(tdb, "tdb_close");
 658
 659         if (tdb->transaction) {
 660                 tdb_transaction_cancel(tdb);
 661         }
 662
 663         if (tdb->file->map_ptr) {
 664                 if (tdb->flags & TDB_INTERNAL)
 665                         free(tdb->file->map_ptr);
 666                 else
 667                         tdb_munmap(tdb->file);
 668         }
 669         if (tdb->file) {
 670                 struct tdb_file **i;
 671
 672                 tdb_lock_cleanup(tdb);
 673                 if (--tdb->file->refcnt == 0) {
 674                         ret = close(tdb->file->fd);
 675
 676                         /* Remove from files list */
 677                         for (i = &files; *i; i = &(*i)->next) {
 678                                 if (*i == tdb->file) {
 679                                         *i = tdb->file->next;
 680                                         break;
 681                                 }
 682                         }
 683                         free(tdb->file->lockrecs);
 684                         free(tdb->file);
 685                 }
 686         }
 687
 688 #ifdef TDB_TRACE
 689         close(tdb->tracefd);
 690 #endif
 691         free(tdb);
 692
 693         return ret;
 694 }