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                 return tdb->last_error
 274                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 275                                      TDB_LOG_USE_ERROR,
 276                                      "tdb_get_attribute:"
 277                                      " cannot get TDB_ATTRIBUTE_OPENHOOK");
 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_HASH:
 310         case TDB_ATTRIBUTE_SEED:
 311         case TDB_ATTRIBUTE_OPENHOOK:
 312                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 313                            "tdb_unset_attribute: cannot unset %s after opening",
 314                            type == TDB_ATTRIBUTE_HASH
 315                            ? "TDB_ATTRIBUTE_HASH"
 316                            : type == TDB_ATTRIBUTE_SEED
 317                            ? "TDB_ATTRIBUTE_SEED"
 318                            : "TDB_ATTRIBUTE_OPENHOOK");
 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         struct tdb_attribute_openhook *openhook = NULL;
 351         tdb_bool_err berr;
 352         enum TDB_ERROR ecode;
 353         int openlock;
 354
 355         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 356         if (!tdb) {
 357                 /* Can't log this */
 358                 errno = ENOMEM;
 359                 return NULL;
 360         }
 361         /* Set name immediately for logging functions. */
 362         if (name) {
 363                 tdb->name = strcpy((char *)(tdb + 1), name);
 364         } else {
 365                 tdb->name = NULL;
 366         }
 367         tdb->direct_access = 0;
 368         tdb->flags = tdb_flags;
 369         tdb->log_fn = NULL;
 370         tdb->transaction = NULL;
 371         tdb->access = NULL;
 372         tdb->open_flags = open_flags;
 373         tdb->last_error = TDB_SUCCESS;
 374         tdb->file = 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                         openhook = &attr->openhook;
 394                         break;
 395                 default:
 396                         /* These are set as normal. */
 397                         ecode = tdb_set_attribute(tdb, attr);
 398                         if (ecode != TDB_SUCCESS)
 399                                 goto fail;
 400                 }
 401                 attr = attr->base.next;
 402         }
 403
 404         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 405                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING)) {
 406                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 407                                    "tdb_open: unknown flags %u", tdb_flags);
 408                 goto fail;
 409         }
 410
 411         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 412                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 413                                    "tdb_open: can't open tdb %s write-only",
 414                                    name);
 415                 goto fail;
 416         }
 417
 418         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 419                 tdb->read_only = true;
 420                 openlock = F_RDLCK;
 421         } else {
 422                 tdb->read_only = false;
 423                 openlock = F_WRLCK;
 424         }
 425
 426         /* internal databases don't need any of the rest. */
 427         if (tdb->flags & TDB_INTERNAL) {
 428                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 429                 ecode = tdb_new_file(tdb);
 430                 if (ecode != TDB_SUCCESS) {
 431                         goto fail;
 432                 }
 433                 tdb->file->fd = -1;
 434                 ecode = tdb_new_database(tdb, seed, &hdr);
 435                 if (ecode != TDB_SUCCESS) {
 436                         goto fail;
 437                 }
 438                 tdb_convert(tdb, &hdr.hash_seed, sizeof(hdr.hash_seed));
 439                 tdb->hash_seed = hdr.hash_seed;
 440                 tdb_ftable_init(tdb);
 441                 return tdb;
 442         }
 443
 444         if (stat(name, &st) != -1)
 445                 tdb->file = find_file(st.st_dev, st.st_ino);
 446
 447         if (!tdb->file) {
 448                 int fd;
 449
 450                 if ((fd = open(name, open_flags, mode)) == -1) {
 451                         /* errno set by open(2) */
 452                         saved_errno = errno;
 453                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 454                                    "tdb_open: could not open file %s: %s",
 455                                    name, strerror(errno));
 456                         goto fail_errno;
 457                 }
 458
 459                 /* on exec, don't inherit the fd */
 460                 v = fcntl(fd, F_GETFD, 0);
 461                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 462
 463                 if (fstat(fd, &st) == -1) {
 464                         saved_errno = errno;
 465                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 466                                    "tdb_open: could not stat open %s: %s",
 467                                    name, strerror(errno));
 468                         close(fd);
 469                         goto fail_errno;
 470                 }
 471
 472                 ecode = tdb_new_file(tdb);
 473                 if (ecode != TDB_SUCCESS) {
 474                         close(fd);
 475                         goto fail;
 476                 }
 477
 478                 tdb->file->next = files;
 479                 tdb->file->fd = fd;
 480                 tdb->file->device = st.st_dev;
 481                 tdb->file->inode = st.st_ino;
 482                 tdb->file->map_ptr = NULL;
 483                 tdb->file->map_size = sizeof(struct tdb_header);
 484         }
 485
 486         /* ensure there is only one process initialising at once */
 487         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 488         if (ecode != TDB_SUCCESS) {
 489                 saved_errno = errno;
 490                 goto fail_errno;
 491         }
 492
 493         /* call their open hook if they gave us one. */
 494         if (openhook) {
 495                 ecode = openhook->fn(tdb->file->fd, openhook->data);
 496                 if (ecode != TDB_SUCCESS) {
 497                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 498                                    "tdb_open: open hook failed");
 499                         goto fail;
 500                 }
 501                 open_flags |= O_CREAT;
 502         }
 503
 504         /* If they used O_TRUNC, read will return 0. */
 505         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 506         if (rlen == 0 && (open_flags & O_CREAT)) {
 507                 ecode = tdb_new_database(tdb, seed, &hdr);
 508                 if (ecode != TDB_SUCCESS) {
 509                         goto fail;
 510                 }
 511         } else if (rlen < 0) {
 512                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 513                                    "tdb_open: error %s reading %s",
 514                                    strerror(errno), name);
 515                 goto fail;
 516         } else if (rlen < sizeof(hdr)
 517                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 518                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 519                                    "tdb_open: %s is not a tdb file", name);
 520                 goto fail;
 521         }
 522
 523         if (hdr.version != TDB_VERSION) {
 524                 if (hdr.version == bswap_64(TDB_VERSION))
 525                         tdb->flags |= TDB_CONVERT;
 526                 else {
 527                         /* wrong version */
 528                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 529                                            "tdb_open:"
 530                                            " %s is unknown version 0x%llx",
 531                                            name, (long long)hdr.version);
 532                         goto fail;
 533                 }
 534         } else if (tdb->flags & TDB_CONVERT) {
 535                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 536                                    "tdb_open:"
 537                                    " %s does not need TDB_CONVERT",
 538                                    name);
 539                 goto fail;
 540         }
 541
 542         tdb_convert(tdb, &hdr, sizeof(hdr));
 543         tdb->hash_seed = hdr.hash_seed;
 544         hash_test = TDB_HASH_MAGIC;
 545         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 546         if (hdr.hash_test != hash_test) {
 547                 /* wrong hash variant */
 548                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 549                                    "tdb_open:"
 550                                    " %s uses a different hash function",
 551                                    name);
 552                 goto fail;
 553         }
 554
 555         /* Clear any features we don't understand. */
 556         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 557                 hdr.features_used &= TDB_FEATURE_MASK;
 558                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 559                                                         features_used),
 560                                           &hdr.features_used,
 561                                           sizeof(hdr.features_used));
 562                 if (ecode != TDB_SUCCESS)
 563                         goto fail;
 564         }
 565
 566         tdb_unlock_open(tdb, openlock);
 567
 568         /* This make sure we have current map_size and mmap. */
 569         ecode = tdb->methods->oob(tdb, tdb->file->map_size + 1, true);
 570         if (unlikely(ecode != TDB_SUCCESS))
 571                 goto fail;
 572
 573         /* Now it's fully formed, recover if necessary. */
 574         berr = tdb_needs_recovery(tdb);
 575         if (unlikely(berr != false)) {
 576                 if (berr < 0) {
 577                         ecode = berr;
 578                         goto fail;
 579                 }
 580                 ecode = tdb_lock_and_recover(tdb);
 581                 if (ecode != TDB_SUCCESS) {
 582                         goto fail;
 583                 }
 584         }
 585
 586         ecode = tdb_ftable_init(tdb);
 587         if (ecode != TDB_SUCCESS) {
 588                 goto fail;
 589         }
 590
 591         /* Add to linked list if we're new. */
 592         if (tdb->file->refcnt == 1)
 593                 files = tdb->file;
 594         return tdb;
 595
 596  fail:
 597         /* Map ecode to some logical errno. */
 598         switch (ecode) {
 599         case TDB_ERR_CORRUPT:
 600         case TDB_ERR_IO:
 601                 saved_errno = EIO;
 602                 break;
 603         case TDB_ERR_LOCK:
 604                 saved_errno = EWOULDBLOCK;
 605                 break;
 606         case TDB_ERR_OOM:
 607                 saved_errno = ENOMEM;
 608                 break;
 609         case TDB_ERR_EINVAL:
 610                 saved_errno = EINVAL;
 611                 break;
 612         default:
 613                 saved_errno = EINVAL;
 614                 break;
 615         }
 616
 617 fail_errno:
 618 #ifdef TDB_TRACE
 619         close(tdb->tracefd);
 620 #endif
 621         if (tdb->file) {
 622                 tdb_lock_cleanup(tdb);
 623                 if (--tdb->file->refcnt == 0) {
 624                         assert(tdb->file->num_lockrecs == 0);
 625                         if (tdb->file->map_ptr) {
 626                                 if (tdb->flags & TDB_INTERNAL) {
 627                                         free(tdb->file->map_ptr);
 628                                 } else
 629                                         tdb_munmap(tdb->file);
 630                         }
 631                         if (close(tdb->file->fd) != 0)
 632                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 633                                            "tdb_open: failed to close tdb fd"
 634                                            " on error: %s", strerror(errno));
 635                         free(tdb->file->lockrecs);
 636                         free(tdb->file);
 637                 }
 638         }
 639
 640         free(tdb);
 641         errno = saved_errno;
 642         return NULL;
 643 }
 644
 645 int tdb_close(struct tdb_context *tdb)
 646 {
 647         int ret = 0;
 648
 649         tdb_trace(tdb, "tdb_close");
 650
 651         if (tdb->transaction) {
 652                 tdb_transaction_cancel(tdb);
 653         }
 654
 655         if (tdb->file->map_ptr) {
 656                 if (tdb->flags & TDB_INTERNAL)
 657                         free(tdb->file->map_ptr);
 658                 else
 659                         tdb_munmap(tdb->file);
 660         }
 661         if (tdb->file) {
 662                 struct tdb_file **i;
 663
 664                 tdb_lock_cleanup(tdb);
 665                 if (--tdb->file->refcnt == 0) {
 666                         ret = close(tdb->file->fd);
 667
 668                         /* Remove from files list */
 669                         for (i = &files; *i; i = &(*i)->next) {
 670                                 if (*i == tdb->file) {
 671                                         *i = tdb->file->next;
 672                                         break;
 673                                 }
 674                         }
 675                         free(tdb->file->lockrecs);
 676                         free(tdb->file);
 677                 }
 678         }
 679
 680 #ifdef TDB_TRACE
 681         close(tdb->tracefd);
 682 #endif
 683         free(tdb);
 684
 685         return ret;
 686 }