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/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         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
 139         /* Initial hashes are empty. */
 140         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 141
 142         /* Free is empty. */
 143         newdb.hdr.free_table = offsetof(struct new_database, ftable);
 144         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
 145         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
 146                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 147                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 148                            0);
 149         if (ecode != TDB_SUCCESS) {
 150                 return ecode;
 151         }
 152
 153         /* Magic food */
 154         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 155         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 156
 157         /* This creates an endian-converted database, as if read from disk */
 158         magic_len = sizeof(newdb.hdr.magic_food);
 159         tdb_convert(tdb,
 160                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 161
 162         *hdr = newdb.hdr;
 163
 164         if (tdb->flags & TDB_INTERNAL) {
 165                 tdb->file->map_size = sizeof(newdb);
 166                 tdb->file->map_ptr = malloc(tdb->file->map_size);
 167                 if (!tdb->file->map_ptr) {
 168                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 169                                           "tdb_new_database:"
 170                                           " failed to allocate");
 171                 }
 172                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
 173                 return TDB_SUCCESS;
 174         }
 175         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
 176                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 177                                   "tdb_new_database:"
 178                                   " failed to seek: %s", strerror(errno));
 179         }
 180
 181         if (ftruncate(tdb->file->fd, 0) == -1) {
 182                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 183                                   "tdb_new_database:"
 184                                   " failed to truncate: %s", strerror(errno));
 185         }
 186
 187         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
 188         if (rlen != sizeof(newdb)) {
 189                 if (rlen >= 0)
 190                         errno = ENOSPC;
 191                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 192                                   "tdb_new_database: %zi writing header: %s",
 193                                   rlen, strerror(errno));
 194         }
 195         return TDB_SUCCESS;
 196 }
 197
 198 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
 199 {
 200         tdb->file = malloc(sizeof(*tdb->file));
 201         if (!tdb->file)
 202                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 203                                   "tdb_open: cannot alloc tdb_file structure");
 204         tdb->file->num_lockrecs = 0;
 205         tdb->file->lockrecs = NULL;
 206         tdb->file->allrecord_lock.count = 0;
 207         tdb->file->refcnt = 1;
 208         tdb->file->map_ptr = NULL;
 209         return TDB_SUCCESS;
 210 }
 211
 212 enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
 213                                  const union tdb_attribute *attr)
 214 {
 215         switch (attr->base.attr) {
 216         case TDB_ATTRIBUTE_LOG:
 217                 tdb->log_fn = attr->log.fn;
 218                 tdb->log_data = attr->log.data;
 219                 break;
 220         case TDB_ATTRIBUTE_HASH:
 221         case TDB_ATTRIBUTE_SEED:
 222         case TDB_ATTRIBUTE_OPENHOOK:
 223         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 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                                      : attr->base.attr == TDB_ATTRIBUTE_OPENHOOK
 234                                      ? "TDB_ATTRIBUTE_OPENHOOK"
 235                                      : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 236         case TDB_ATTRIBUTE_STATS:
 237                 return tdb->last_error
 238                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 239                                      TDB_LOG_USE_ERROR,
 240                                      "tdb_set_attribute:"
 241                                      " cannot set TDB_ATTRIBUTE_STATS");
 242         case TDB_ATTRIBUTE_FLOCK:
 243                 tdb->lock_fn = attr->flock.lock;
 244                 tdb->unlock_fn = attr->flock.unlock;
 245                 tdb->lock_data = attr->flock.data;
 246                 break;
 247         default:
 248                 return tdb->last_error
 249                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 250                                      TDB_LOG_USE_ERROR,
 251                                      "tdb_set_attribute:"
 252                                      " unknown attribute type %u",
 253                                      attr->base.attr);
 254         }
 255         return TDB_SUCCESS;
 256 }
 257
 258 enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
 259                                  union tdb_attribute *attr)
 260 {
 261         switch (attr->base.attr) {
 262         case TDB_ATTRIBUTE_LOG:
 263                 if (!tdb->log_fn)
 264                         return tdb->last_error = TDB_ERR_NOEXIST;
 265                 attr->log.fn = tdb->log_fn;
 266                 attr->log.data = tdb->log_data;
 267                 break;
 268         case TDB_ATTRIBUTE_HASH:
 269                 attr->hash.fn = tdb->hash_fn;
 270                 attr->hash.data = tdb->hash_data;
 271                 break;
 272         case TDB_ATTRIBUTE_SEED:
 273                 if (tdb->flags & TDB_VERSION1)
 274                         return tdb->last_error
 275                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 276                                              TDB_LOG_USE_ERROR,
 277                                      "tdb_get_attribute:"
 278                                      " cannot get TDB_ATTRIBUTE_SEED"
 279                                      " on TDB1 tdb.");
 280                 attr->seed.seed = tdb->hash_seed;
 281                 break;
 282         case TDB_ATTRIBUTE_OPENHOOK:
 283                 if (!tdb->openhook)
 284                         return tdb->last_error = TDB_ERR_NOEXIST;
 285                 attr->openhook.fn = tdb->openhook;
 286                 attr->openhook.data = tdb->openhook_data;
 287                 break;
 288         case TDB_ATTRIBUTE_STATS: {
 289                 size_t size = attr->stats.size;
 290                 if (size > tdb->stats.size)
 291                         size = tdb->stats.size;
 292                 memcpy(&attr->stats, &tdb->stats, size);
 293                 break;
 294         }
 295         case TDB_ATTRIBUTE_FLOCK:
 296                 attr->flock.lock = tdb->lock_fn;
 297                 attr->flock.unlock = tdb->unlock_fn;
 298                 attr->flock.data = tdb->lock_data;
 299                 break;
 300         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 301                 if (!(tdb->flags & TDB_VERSION1))
 302                         return tdb->last_error
 303                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 304                                              TDB_LOG_USE_ERROR,
 305                                      "tdb_get_attribute:"
 306                                      " cannot get TDB_ATTRIBUTE_TDB1_HASHSIZE"
 307                                      " on TDB2 tdb.");
 308                 attr->tdb1_hashsize.hsize = tdb->tdb1.header.hash_size;
 309                 break;
 310         default:
 311                 return tdb->last_error
 312                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 313                                      TDB_LOG_USE_ERROR,
 314                                      "tdb_get_attribute:"
 315                                      " unknown attribute type %u",
 316                                      attr->base.attr);
 317         }
 318         attr->base.next = NULL;
 319         return TDB_SUCCESS;
 320 }
 321
 322 void tdb_unset_attribute(struct tdb_context *tdb,
 323                          enum tdb_attribute_type type)
 324 {
 325         switch (type) {
 326         case TDB_ATTRIBUTE_LOG:
 327                 tdb->log_fn = NULL;
 328                 break;
 329         case TDB_ATTRIBUTE_OPENHOOK:
 330                 tdb->openhook = NULL;
 331                 break;
 332         case TDB_ATTRIBUTE_HASH:
 333         case TDB_ATTRIBUTE_SEED:
 334         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 335                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 336                            "tdb_unset_attribute: cannot unset %s after opening",
 337                            type == TDB_ATTRIBUTE_HASH
 338                            ? "TDB_ATTRIBUTE_HASH"
 339                            : type == TDB_ATTRIBUTE_SEED
 340                            ? "TDB_ATTRIBUTE_SEED"
 341                            : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 342                 break;
 343         case TDB_ATTRIBUTE_STATS:
 344                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 345                            TDB_LOG_USE_ERROR,
 346                            "tdb_unset_attribute:"
 347                            "cannot unset TDB_ATTRIBUTE_STATS");
 348                 break;
 349         case TDB_ATTRIBUTE_FLOCK:
 350                 tdb->lock_fn = tdb_fcntl_lock;
 351                 tdb->unlock_fn = tdb_fcntl_unlock;
 352                 break;
 353         default:
 354                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 355                            TDB_LOG_USE_ERROR,
 356                            "tdb_unset_attribute: unknown attribute type %u",
 357                            type);
 358         }
 359 }
 360
 361 static bool is_tdb1(struct tdb1_header *hdr, const void *buf, ssize_t rlen)
 362 {
 363         /* This code assumes we've tried to read entire tdb1 header. */
 364         BUILD_ASSERT(sizeof(*hdr) <= sizeof(struct tdb_header));
 365
 366         if (rlen < (ssize_t)sizeof(*hdr)) {
 367                 return false;
 368         }
 369
 370         memcpy(hdr, buf, sizeof(*hdr));
 371         if (strcmp(hdr->magic_food, TDB_MAGIC_FOOD) != 0)
 372                 return false;
 373
 374         return hdr->version == TDB1_VERSION
 375                 || hdr->version == TDB1_BYTEREV(TDB1_VERSION);
 376 }
 377
 378 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 379                              int open_flags, mode_t mode,
 380                              union tdb_attribute *attr)
 381 {
 382         struct tdb_context *tdb;
 383         struct stat st;
 384         int saved_errno = 0;
 385         uint64_t hash_test;
 386         unsigned v;
 387         ssize_t rlen;
 388         struct tdb_header hdr;
 389         struct tdb_attribute_seed *seed = NULL;
 390         struct tdb_attribute_tdb1_hashsize *hsize_attr = NULL;
 391         tdb_bool_err berr;
 392         enum TDB_ERROR ecode;
 393         int openlock;
 394
 395         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 396         if (!tdb) {
 397                 /* Can't log this */
 398                 errno = ENOMEM;
 399                 return NULL;
 400         }
 401         /* Set name immediately for logging functions. */
 402         if (name) {
 403                 tdb->name = strcpy((char *)(tdb + 1), name);
 404         } else {
 405                 tdb->name = NULL;
 406         }
 407         tdb->flags = tdb_flags;
 408         tdb->log_fn = NULL;
 409         tdb->open_flags = open_flags;
 410         tdb->last_error = TDB_SUCCESS;
 411         tdb->file = NULL;
 412         tdb->openhook = NULL;
 413         tdb->lock_fn = tdb_fcntl_lock;
 414         tdb->unlock_fn = tdb_fcntl_unlock;
 415         tdb->hash_fn = tdb_jenkins_hash;
 416         memset(&tdb->stats, 0, sizeof(tdb->stats));
 417         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
 418         tdb->stats.size = sizeof(tdb->stats);
 419
 420         while (attr) {
 421                 switch (attr->base.attr) {
 422                 case TDB_ATTRIBUTE_HASH:
 423                         tdb->hash_fn = attr->hash.fn;
 424                         tdb->hash_data = attr->hash.data;
 425                         break;
 426                 case TDB_ATTRIBUTE_SEED:
 427                         seed = &attr->seed;
 428                         break;
 429                 case TDB_ATTRIBUTE_OPENHOOK:
 430                         tdb->openhook = attr->openhook.fn;
 431                         tdb->openhook_data = attr->openhook.data;
 432                         break;
 433                 case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 434                         hsize_attr = &attr->tdb1_hashsize;
 435                         break;
 436                 default:
 437                         /* These are set as normal. */
 438                         ecode = tdb_set_attribute(tdb, attr);
 439                         if (ecode != TDB_SUCCESS)
 440                                 goto fail;
 441                 }
 442                 attr = attr->base.next;
 443         }
 444
 445         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 446                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
 447                           | TDB_RDONLY | TDB_VERSION1)) {
 448                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 449                                    "tdb_open: unknown flags %u", tdb_flags);
 450                 goto fail;
 451         }
 452
 453         if (hsize_attr) {
 454                 if (!(tdb_flags & TDB_VERSION1) ||
 455                     (!(tdb_flags & TDB_INTERNAL) && !(open_flags & O_CREAT))) {
 456                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 457                                            TDB_LOG_USE_ERROR,
 458                                            "tdb_open: can only use"
 459                                            " TDB_ATTRIBUTE_TDB1_HASHSIZE when"
 460                                            " creating a TDB_VERSION1 tdb");
 461                         goto fail;
 462                 }
 463         }
 464
 465         if (seed) {
 466                 if (tdb_flags & TDB_VERSION1) {
 467                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 468                                            TDB_LOG_USE_ERROR,
 469                                            "tdb_open:"
 470                                            " cannot set TDB_ATTRIBUTE_SEED"
 471                                            " on TDB1 tdb.");
 472                         goto fail;
 473                 } else if (!(tdb_flags & TDB_INTERNAL)
 474                            && !(open_flags & O_CREAT)) {
 475                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 476                                            TDB_LOG_USE_ERROR,
 477                                            "tdb_open:"
 478                                            " cannot set TDB_ATTRIBUTE_SEED"
 479                                            " without O_CREAT.");
 480                         goto fail;
 481                 }
 482         }
 483
 484         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 485                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 486                                    "tdb_open: can't open tdb %s write-only",
 487                                    name);
 488                 goto fail;
 489         }
 490
 491         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 492                 openlock = F_RDLCK;
 493                 tdb->flags |= TDB_RDONLY;
 494         } else {
 495                 if (tdb_flags & TDB_RDONLY) {
 496                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 497                                            TDB_LOG_USE_ERROR,
 498                                            "tdb_open: can't use TDB_RDONLY"
 499                                            " without O_RDONLY");
 500                         goto fail;
 501                 }
 502                 openlock = F_WRLCK;
 503         }
 504
 505         /* internal databases don't need any of the rest. */
 506         if (tdb->flags & TDB_INTERNAL) {
 507                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 508                 ecode = tdb_new_file(tdb);
 509                 if (ecode != TDB_SUCCESS) {
 510                         goto fail;
 511                 }
 512                 tdb->file->fd = -1;
 513                 if (tdb->flags & TDB_VERSION1)
 514                         ecode = tdb1_new_database(tdb, hsize_attr);
 515                 else {
 516                         ecode = tdb_new_database(tdb, seed, &hdr);
 517                         if (ecode == TDB_SUCCESS) {
 518                                 tdb_convert(tdb, &hdr.hash_seed,
 519                                             sizeof(hdr.hash_seed));
 520                                 tdb->hash_seed = hdr.hash_seed;
 521                                 tdb2_context_init(tdb);
 522                                 tdb_ftable_init(tdb);
 523                         }
 524                 }
 525                 if (ecode != TDB_SUCCESS) {
 526                         goto fail;
 527                 }
 528                 return tdb;
 529         }
 530
 531         if (stat(name, &st) != -1)
 532                 tdb->file = find_file(st.st_dev, st.st_ino);
 533
 534         if (!tdb->file) {
 535                 int fd;
 536
 537                 if ((fd = open(name, open_flags, mode)) == -1) {
 538                         /* errno set by open(2) */
 539                         saved_errno = errno;
 540                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 541                                    "tdb_open: could not open file %s: %s",
 542                                    name, strerror(errno));
 543                         goto fail_errno;
 544                 }
 545
 546                 /* on exec, don't inherit the fd */
 547                 v = fcntl(fd, F_GETFD, 0);
 548                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 549
 550                 if (fstat(fd, &st) == -1) {
 551                         saved_errno = errno;
 552                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 553                                    "tdb_open: could not stat open %s: %s",
 554                                    name, strerror(errno));
 555                         close(fd);
 556                         goto fail_errno;
 557                 }
 558
 559                 ecode = tdb_new_file(tdb);
 560                 if (ecode != TDB_SUCCESS) {
 561                         close(fd);
 562                         goto fail;
 563                 }
 564
 565                 tdb->file->fd = fd;
 566                 tdb->file->device = st.st_dev;
 567                 tdb->file->inode = st.st_ino;
 568                 tdb->file->map_ptr = NULL;
 569                 tdb->file->map_size = 0;
 570         }
 571
 572         /* ensure there is only one process initialising at once */
 573         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 574         if (ecode != TDB_SUCCESS) {
 575                 saved_errno = errno;
 576                 goto fail_errno;
 577         }
 578
 579         /* call their open hook if they gave us one. */
 580         if (tdb->openhook) {
 581                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
 582                 if (ecode != TDB_SUCCESS) {
 583                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 584                                    "tdb_open: open hook failed");
 585                         goto fail;
 586                 }
 587                 open_flags |= O_CREAT;
 588         }
 589
 590         /* If they used O_TRUNC, read will return 0. */
 591         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 592         if (rlen == 0 && (open_flags & O_CREAT)) {
 593                 if (tdb->flags & TDB_VERSION1) {
 594                         ecode = tdb1_new_database(tdb, hsize_attr);
 595                         if (ecode != TDB_SUCCESS)
 596                                 goto fail;
 597                         goto finished;
 598                 }
 599                 ecode = tdb_new_database(tdb, seed, &hdr);
 600                 if (ecode != TDB_SUCCESS) {
 601                         goto fail;
 602                 }
 603         } else if (rlen < 0) {
 604                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 605                                    "tdb_open: error %s reading %s",
 606                                    strerror(errno), name);
 607                 goto fail;
 608         } else if (rlen < sizeof(hdr)
 609                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 610                 if (is_tdb1(&tdb->tdb1.header, &hdr, rlen)) {
 611                         ecode = tdb1_open(tdb);
 612                         if (!ecode)
 613                                 goto finished;
 614                         goto fail;
 615                 }
 616                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 617                                    "tdb_open: %s is not a tdb file", name);
 618                 goto fail;
 619         }
 620
 621         if (hdr.version != TDB_VERSION) {
 622                 if (hdr.version == bswap_64(TDB_VERSION))
 623                         tdb->flags |= TDB_CONVERT;
 624                 else {
 625                         if (is_tdb1(&tdb->tdb1.header, &hdr, rlen)) {
 626                                 ecode = tdb1_open(tdb);
 627                                 if (!ecode)
 628                                         goto finished;
 629                                 goto fail;
 630                         }
 631                         /* wrong version */
 632                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 633                                            "tdb_open:"
 634                                            " %s is unknown version 0x%llx",
 635                                            name, (long long)hdr.version);
 636                         goto fail;
 637                 }
 638         } else if (tdb->flags & TDB_CONVERT) {
 639                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 640                                    "tdb_open:"
 641                                    " %s does not need TDB_CONVERT",
 642                                    name);
 643                 goto fail;
 644         }
 645
 646         if (tdb->flags & TDB_VERSION1) {
 647                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 648                                    "tdb_open:"
 649                                    " %s does not need TDB_VERSION1",
 650                                    name);
 651                 goto fail;
 652         }
 653
 654         tdb2_context_init(tdb);
 655
 656         tdb_convert(tdb, &hdr, sizeof(hdr));
 657         tdb->hash_seed = hdr.hash_seed;
 658         hash_test = TDB_HASH_MAGIC;
 659         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 660         if (hdr.hash_test != hash_test) {
 661                 /* wrong hash variant */
 662                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 663                                    "tdb_open:"
 664                                    " %s uses a different hash function",
 665                                    name);
 666                 goto fail;
 667         }
 668
 669         /* Clear any features we don't understand. */
 670         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 671                 hdr.features_used &= TDB_FEATURE_MASK;
 672                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 673                                                         features_used),
 674                                           &hdr.features_used,
 675                                           sizeof(hdr.features_used));
 676                 if (ecode != TDB_SUCCESS)
 677                         goto fail;
 678         }
 679
 680 finished:
 681         if (tdb->flags & TDB_VERSION1) {
 682                 /* if needed, run recovery */
 683                 if (tdb1_transaction_recover(tdb) == -1) {
 684                         ecode = tdb->last_error;
 685                         goto fail;
 686                 }
 687         }
 688
 689         tdb_unlock_open(tdb, openlock);
 690
 691         /* This makes sure we have current map_size and mmap. */
 692         if (tdb->flags & TDB_VERSION1) {
 693                 ecode = tdb1_probe_length(tdb);
 694         } else {
 695                 ecode = tdb->tdb2.io->oob(tdb, tdb->file->map_size + 1, true);
 696         }
 697         if (unlikely(ecode != TDB_SUCCESS))
 698                 goto fail;
 699
 700         if (!(tdb->flags & TDB_VERSION1)) {
 701                 /* Now it's fully formed, recover if necessary. */
 702                 berr = tdb_needs_recovery(tdb);
 703                 if (unlikely(berr != false)) {
 704                         if (berr < 0) {
 705                                 ecode = berr;
 706                                 goto fail;
 707                         }
 708                         ecode = tdb_lock_and_recover(tdb);
 709                         if (ecode != TDB_SUCCESS) {
 710                                 goto fail;
 711                         }
 712                 }
 713
 714                 ecode = tdb_ftable_init(tdb);
 715                 if (ecode != TDB_SUCCESS) {
 716                         goto fail;
 717                 }
 718         }
 719
 720         tdb->next = tdbs;
 721         tdbs = tdb;
 722         return tdb;
 723
 724  fail:
 725         /* Map ecode to some logical errno. */
 726         switch (ecode) {
 727         case TDB_ERR_CORRUPT:
 728         case TDB_ERR_IO:
 729                 saved_errno = EIO;
 730                 break;
 731         case TDB_ERR_LOCK:
 732                 saved_errno = EWOULDBLOCK;
 733                 break;
 734         case TDB_ERR_OOM:
 735                 saved_errno = ENOMEM;
 736                 break;
 737         case TDB_ERR_EINVAL:
 738                 saved_errno = EINVAL;
 739                 break;
 740         default:
 741                 saved_errno = EINVAL;
 742                 break;
 743         }
 744
 745 fail_errno:
 746 #ifdef TDB_TRACE
 747         close(tdb->tracefd);
 748 #endif
 749         if (tdb->file) {
 750                 tdb_lock_cleanup(tdb);
 751                 if (--tdb->file->refcnt == 0) {
 752                         assert(tdb->file->num_lockrecs == 0);
 753                         if (tdb->file->map_ptr) {
 754                                 if (tdb->flags & TDB_INTERNAL) {
 755                                         free(tdb->file->map_ptr);
 756                                 } else
 757                                         tdb_munmap(tdb->file);
 758                         }
 759                         if (close(tdb->file->fd) != 0)
 760                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 761                                            "tdb_open: failed to close tdb fd"
 762                                            " on error: %s", strerror(errno));
 763                         free(tdb->file->lockrecs);
 764                         free(tdb->file);
 765                 }
 766         }
 767
 768         free(tdb);
 769         errno = saved_errno;
 770         return NULL;
 771 }
 772
 773 int tdb_close(struct tdb_context *tdb)
 774 {
 775         int ret = 0;
 776         struct tdb_context **i;
 777
 778         tdb_trace(tdb, "tdb_close");
 779
 780         if (tdb->flags & TDB_VERSION1) {
 781                 if (tdb->tdb1.transaction) {
 782                         tdb1_transaction_cancel(tdb);
 783                 }
 784         } else {
 785                 if (tdb->tdb2.transaction) {
 786                         tdb_transaction_cancel(tdb);
 787                 }
 788         }
 789
 790         if (tdb->file->map_ptr) {
 791                 if (tdb->flags & TDB_INTERNAL)
 792                         free(tdb->file->map_ptr);
 793                 else
 794                         tdb_munmap(tdb->file);
 795         }
 796         if (tdb->file) {
 797                 tdb_lock_cleanup(tdb);
 798                 if (--tdb->file->refcnt == 0) {
 799                         ret = close(tdb->file->fd);
 800                         free(tdb->file->lockrecs);
 801                         free(tdb->file);
 802                 }
 803         }
 804
 805         /* Remove from tdbs list */
 806         for (i = &tdbs; *i; i = &(*i)->next) {
 807                 if (*i == tdb) {
 808                         *i = tdb->next;
 809                         break;
 810                 }
 811         }
 812
 813 #ifdef TDB_TRACE
 814         close(tdb->tracefd);
 815 #endif
 816         free(tdb);
 817
 818         return ret;
 819 }
 820
 821 void tdb_foreach_(int (*fn)(struct tdb_context *, void *), void *p)
 822 {
 823         struct tdb_context *i;
 824
 825         for (i = tdbs; i; i = i->next) {
 826                 if (fn(i, p) != 0)
 827                         break;
 828         }
 829 }