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         newdb.hdr.capabilities = 0;
 139         memset(newdb.hdr.reserved, 0, sizeof(newdb.hdr.reserved));
 140         /* Initial hashes are empty. */
 141         memset(newdb.hdr.hashtable, 0, sizeof(newdb.hdr.hashtable));
 142
 143         /* Free is empty. */
 144         newdb.hdr.free_table = offsetof(struct new_database, ftable);
 145         memset(&newdb.ftable, 0, sizeof(newdb.ftable));
 146         ecode = set_header(NULL, &newdb.ftable.hdr, TDB_FTABLE_MAGIC, 0,
 147                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 148                            sizeof(newdb.ftable) - sizeof(newdb.ftable.hdr),
 149                            0);
 150         if (ecode != TDB_SUCCESS) {
 151                 return ecode;
 152         }
 153
 154         /* Magic food */
 155         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 156         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 157
 158         /* This creates an endian-converted database, as if read from disk */
 159         magic_len = sizeof(newdb.hdr.magic_food);
 160         tdb_convert(tdb,
 161                     (char *)&newdb.hdr + magic_len, sizeof(newdb) - magic_len);
 162
 163         *hdr = newdb.hdr;
 164
 165         if (tdb->flags & TDB_INTERNAL) {
 166                 tdb->file->map_size = sizeof(newdb);
 167                 tdb->file->map_ptr = malloc(tdb->file->map_size);
 168                 if (!tdb->file->map_ptr) {
 169                         return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 170                                           "tdb_new_database:"
 171                                           " failed to allocate");
 172                 }
 173                 memcpy(tdb->file->map_ptr, &newdb, tdb->file->map_size);
 174                 return TDB_SUCCESS;
 175         }
 176         if (lseek(tdb->file->fd, 0, SEEK_SET) == -1) {
 177                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 178                                   "tdb_new_database:"
 179                                   " failed to seek: %s", strerror(errno));
 180         }
 181
 182         if (ftruncate(tdb->file->fd, 0) == -1) {
 183                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 184                                   "tdb_new_database:"
 185                                   " failed to truncate: %s", strerror(errno));
 186         }
 187
 188         rlen = write(tdb->file->fd, &newdb, sizeof(newdb));
 189         if (rlen != sizeof(newdb)) {
 190                 if (rlen >= 0)
 191                         errno = ENOSPC;
 192                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 193                                   "tdb_new_database: %zi writing header: %s",
 194                                   rlen, strerror(errno));
 195         }
 196         return TDB_SUCCESS;
 197 }
 198
 199 static enum TDB_ERROR tdb_new_file(struct tdb_context *tdb)
 200 {
 201         tdb->file = malloc(sizeof(*tdb->file));
 202         if (!tdb->file)
 203                 return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
 204                                   "tdb_open: cannot alloc tdb_file structure");
 205         tdb->file->num_lockrecs = 0;
 206         tdb->file->lockrecs = NULL;
 207         tdb->file->allrecord_lock.count = 0;
 208         tdb->file->refcnt = 1;
 209         tdb->file->map_ptr = NULL;
 210         return TDB_SUCCESS;
 211 }
 212
 213 enum TDB_ERROR tdb_set_attribute(struct tdb_context *tdb,
 214                                  const union tdb_attribute *attr)
 215 {
 216         switch (attr->base.attr) {
 217         case TDB_ATTRIBUTE_LOG:
 218                 tdb->log_fn = attr->log.fn;
 219                 tdb->log_data = attr->log.data;
 220                 break;
 221         case TDB_ATTRIBUTE_HASH:
 222         case TDB_ATTRIBUTE_SEED:
 223         case TDB_ATTRIBUTE_OPENHOOK:
 224         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 225                 return tdb->last_error
 226                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 227                                      TDB_LOG_USE_ERROR,
 228                                      "tdb_set_attribute:"
 229                                      " cannot set %s after opening",
 230                                      attr->base.attr == TDB_ATTRIBUTE_HASH
 231                                      ? "TDB_ATTRIBUTE_HASH"
 232                                      : attr->base.attr == TDB_ATTRIBUTE_SEED
 233                                      ? "TDB_ATTRIBUTE_SEED"
 234                                      : attr->base.attr == TDB_ATTRIBUTE_OPENHOOK
 235                                      ? "TDB_ATTRIBUTE_OPENHOOK"
 236                                      : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 237         case TDB_ATTRIBUTE_STATS:
 238                 return tdb->last_error
 239                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 240                                      TDB_LOG_USE_ERROR,
 241                                      "tdb_set_attribute:"
 242                                      " cannot set TDB_ATTRIBUTE_STATS");
 243         case TDB_ATTRIBUTE_FLOCK:
 244                 tdb->lock_fn = attr->flock.lock;
 245                 tdb->unlock_fn = attr->flock.unlock;
 246                 tdb->lock_data = attr->flock.data;
 247                 break;
 248         default:
 249                 return tdb->last_error
 250                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 251                                      TDB_LOG_USE_ERROR,
 252                                      "tdb_set_attribute:"
 253                                      " unknown attribute type %u",
 254                                      attr->base.attr);
 255         }
 256         return TDB_SUCCESS;
 257 }
 258
 259 enum TDB_ERROR tdb_get_attribute(struct tdb_context *tdb,
 260                                  union tdb_attribute *attr)
 261 {
 262         switch (attr->base.attr) {
 263         case TDB_ATTRIBUTE_LOG:
 264                 if (!tdb->log_fn)
 265                         return tdb->last_error = TDB_ERR_NOEXIST;
 266                 attr->log.fn = tdb->log_fn;
 267                 attr->log.data = tdb->log_data;
 268                 break;
 269         case TDB_ATTRIBUTE_HASH:
 270                 attr->hash.fn = tdb->hash_fn;
 271                 attr->hash.data = tdb->hash_data;
 272                 break;
 273         case TDB_ATTRIBUTE_SEED:
 274                 if (tdb->flags & TDB_VERSION1)
 275                         return tdb->last_error
 276                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 277                                              TDB_LOG_USE_ERROR,
 278                                      "tdb_get_attribute:"
 279                                      " cannot get TDB_ATTRIBUTE_SEED"
 280                                      " on TDB1 tdb.");
 281                 attr->seed.seed = tdb->hash_seed;
 282                 break;
 283         case TDB_ATTRIBUTE_OPENHOOK:
 284                 if (!tdb->openhook)
 285                         return tdb->last_error = TDB_ERR_NOEXIST;
 286                 attr->openhook.fn = tdb->openhook;
 287                 attr->openhook.data = tdb->openhook_data;
 288                 break;
 289         case TDB_ATTRIBUTE_STATS: {
 290                 size_t size = attr->stats.size;
 291                 if (size > tdb->stats.size)
 292                         size = tdb->stats.size;
 293                 memcpy(&attr->stats, &tdb->stats, size);
 294                 break;
 295         }
 296         case TDB_ATTRIBUTE_FLOCK:
 297                 attr->flock.lock = tdb->lock_fn;
 298                 attr->flock.unlock = tdb->unlock_fn;
 299                 attr->flock.data = tdb->lock_data;
 300                 break;
 301         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 302                 if (!(tdb->flags & TDB_VERSION1))
 303                         return tdb->last_error
 304                                 = tdb_logerr(tdb, TDB_ERR_EINVAL,
 305                                              TDB_LOG_USE_ERROR,
 306                                      "tdb_get_attribute:"
 307                                      " cannot get TDB_ATTRIBUTE_TDB1_HASHSIZE"
 308                                      " on TDB2 tdb.");
 309                 attr->tdb1_hashsize.hsize = tdb->tdb1.header.hash_size;
 310                 break;
 311         default:
 312                 return tdb->last_error
 313                         = tdb_logerr(tdb, TDB_ERR_EINVAL,
 314                                      TDB_LOG_USE_ERROR,
 315                                      "tdb_get_attribute:"
 316                                      " unknown attribute type %u",
 317                                      attr->base.attr);
 318         }
 319         attr->base.next = NULL;
 320         return TDB_SUCCESS;
 321 }
 322
 323 void tdb_unset_attribute(struct tdb_context *tdb,
 324                          enum tdb_attribute_type type)
 325 {
 326         switch (type) {
 327         case TDB_ATTRIBUTE_LOG:
 328                 tdb->log_fn = NULL;
 329                 break;
 330         case TDB_ATTRIBUTE_OPENHOOK:
 331                 tdb->openhook = NULL;
 332                 break;
 333         case TDB_ATTRIBUTE_HASH:
 334         case TDB_ATTRIBUTE_SEED:
 335         case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 336                 tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 337                            "tdb_unset_attribute: cannot unset %s after opening",
 338                            type == TDB_ATTRIBUTE_HASH
 339                            ? "TDB_ATTRIBUTE_HASH"
 340                            : type == TDB_ATTRIBUTE_SEED
 341                            ? "TDB_ATTRIBUTE_SEED"
 342                            : "TDB_ATTRIBUTE_TDB1_HASHSIZE");
 343                 break;
 344         case TDB_ATTRIBUTE_STATS:
 345                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 346                            TDB_LOG_USE_ERROR,
 347                            "tdb_unset_attribute:"
 348                            "cannot unset TDB_ATTRIBUTE_STATS");
 349                 break;
 350         case TDB_ATTRIBUTE_FLOCK:
 351                 tdb->lock_fn = tdb_fcntl_lock;
 352                 tdb->unlock_fn = tdb_fcntl_unlock;
 353                 break;
 354         default:
 355                 tdb_logerr(tdb, TDB_ERR_EINVAL,
 356                            TDB_LOG_USE_ERROR,
 357                            "tdb_unset_attribute: unknown attribute type %u",
 358                            type);
 359         }
 360 }
 361
 362 static bool is_tdb1(struct tdb1_header *hdr, const void *buf, ssize_t rlen)
 363 {
 364         /* This code assumes we've tried to read entire tdb1 header. */
 365         BUILD_ASSERT(sizeof(*hdr) <= sizeof(struct tdb_header));
 366
 367         if (rlen < (ssize_t)sizeof(*hdr)) {
 368                 return false;
 369         }
 370
 371         memcpy(hdr, buf, sizeof(*hdr));
 372         if (strcmp(hdr->magic_food, TDB_MAGIC_FOOD) != 0)
 373                 return false;
 374
 375         return hdr->version == TDB1_VERSION
 376                 || hdr->version == TDB1_BYTEREV(TDB1_VERSION);
 377 }
 378
 379 /* The top three bits of the capability tell us whether it matters. */
 380 enum TDB_ERROR unknown_capability(struct tdb_context *tdb, const char *caller,
 381                                   tdb_off_t type)
 382 {
 383         if (type & TDB_CAP_NOOPEN) {
 384                 return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 385                                   "%s: file has unknown capability %llu",
 386                                   caller, type & TDB_CAP_NOOPEN);
 387         }
 388
 389         if ((type & TDB_CAP_NOWRITE) && !(tdb->flags & TDB_RDONLY)) {
 390                 return tdb_logerr(tdb, TDB_ERR_RDONLY, TDB_LOG_ERROR,
 391                                   "%s: file has unknown capability %llu"
 392                                   " (cannot write to it)",
 393                                   caller, type & TDB_CAP_NOOPEN);
 394         }
 395
 396         if (type & TDB_CAP_NOCHECK) {
 397                 tdb->flags |= TDB_CANT_CHECK;
 398         }
 399         return TDB_SUCCESS;
 400 }
 401
 402 static enum TDB_ERROR capabilities_ok(struct tdb_context *tdb,
 403                                       tdb_off_t capabilities)
 404 {
 405         tdb_off_t off, next;
 406         enum TDB_ERROR ecode = TDB_SUCCESS;
 407         const struct tdb_capability *cap;
 408
 409         /* Check capability list. */
 410         for (off = capabilities; off && ecode == TDB_SUCCESS; off = next) {
 411                 cap = tdb_access_read(tdb, off, sizeof(*cap), true);
 412                 if (TDB_PTR_IS_ERR(cap)) {
 413                         return TDB_PTR_ERR(cap);
 414                 }
 415
 416                 switch (cap->type & TDB_CAP_TYPE_MASK) {
 417                 /* We don't understand any capabilities (yet). */
 418                 default:
 419                         ecode = unknown_capability(tdb, "tdb_open", cap->type);
 420                 }
 421                 next = cap->next;
 422                 tdb_access_release(tdb, cap);
 423         }
 424         return ecode;
 425 }
 426
 427 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 428                              int open_flags, mode_t mode,
 429                              union tdb_attribute *attr)
 430 {
 431         struct tdb_context *tdb;
 432         struct stat st;
 433         int saved_errno = 0;
 434         uint64_t hash_test;
 435         unsigned v;
 436         ssize_t rlen;
 437         struct tdb_header hdr;
 438         struct tdb_attribute_seed *seed = NULL;
 439         struct tdb_attribute_tdb1_hashsize *hsize_attr = NULL;
 440         struct tdb_attribute_tdb1_max_dead *maxsize_attr = NULL;
 441         tdb_bool_err berr;
 442         enum TDB_ERROR ecode;
 443         int openlock;
 444
 445         tdb = malloc(sizeof(*tdb) + (name ? strlen(name) + 1 : 0));
 446         if (!tdb) {
 447                 /* Can't log this */
 448                 errno = ENOMEM;
 449                 return NULL;
 450         }
 451         /* Set name immediately for logging functions. */
 452         if (name) {
 453                 tdb->name = strcpy((char *)(tdb + 1), name);
 454         } else {
 455                 tdb->name = NULL;
 456         }
 457         tdb->flags = tdb_flags;
 458         tdb->log_fn = NULL;
 459         tdb->open_flags = open_flags;
 460         tdb->last_error = TDB_SUCCESS;
 461         tdb->file = NULL;
 462         tdb->openhook = NULL;
 463         tdb->lock_fn = tdb_fcntl_lock;
 464         tdb->unlock_fn = tdb_fcntl_unlock;
 465         tdb->hash_fn = tdb_jenkins_hash;
 466         memset(&tdb->stats, 0, sizeof(tdb->stats));
 467         tdb->stats.base.attr = TDB_ATTRIBUTE_STATS;
 468         tdb->stats.size = sizeof(tdb->stats);
 469
 470         while (attr) {
 471                 switch (attr->base.attr) {
 472                 case TDB_ATTRIBUTE_HASH:
 473                         tdb->hash_fn = attr->hash.fn;
 474                         tdb->hash_data = attr->hash.data;
 475                         break;
 476                 case TDB_ATTRIBUTE_SEED:
 477                         seed = &attr->seed;
 478                         break;
 479                 case TDB_ATTRIBUTE_OPENHOOK:
 480                         tdb->openhook = attr->openhook.fn;
 481                         tdb->openhook_data = attr->openhook.data;
 482                         break;
 483                 case TDB_ATTRIBUTE_TDB1_HASHSIZE:
 484                         hsize_attr = &attr->tdb1_hashsize;
 485                         break;
 486                 case TDB_ATTRIBUTE_TDB1_MAX_DEAD:
 487                         maxsize_attr = &attr->tdb1_max_dead;
 488                         break;
 489                 default:
 490                         /* These are set as normal. */
 491                         ecode = tdb_set_attribute(tdb, attr);
 492                         if (ecode != TDB_SUCCESS)
 493                                 goto fail;
 494                 }
 495                 attr = attr->base.next;
 496         }
 497
 498         if (tdb_flags & ~(TDB_INTERNAL | TDB_NOLOCK | TDB_NOMMAP | TDB_CONVERT
 499                           | TDB_NOSYNC | TDB_SEQNUM | TDB_ALLOW_NESTING
 500                           | TDB_RDONLY | TDB_VERSION1)) {
 501                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 502                                    "tdb_open: unknown flags %u", tdb_flags);
 503                 goto fail;
 504         }
 505
 506         if (hsize_attr) {
 507                 if (!(tdb_flags & TDB_VERSION1) ||
 508                     (!(tdb_flags & TDB_INTERNAL) && !(open_flags & O_CREAT))) {
 509                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 510                                            TDB_LOG_USE_ERROR,
 511                                            "tdb_open: can only use"
 512                                            " TDB_ATTRIBUTE_TDB1_HASHSIZE when"
 513                                            " creating a TDB_VERSION1 tdb");
 514                         goto fail;
 515                 }
 516         }
 517
 518         if (seed) {
 519                 if (tdb_flags & TDB_VERSION1) {
 520                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 521                                            TDB_LOG_USE_ERROR,
 522                                            "tdb_open:"
 523                                            " cannot set TDB_ATTRIBUTE_SEED"
 524                                            " on TDB1 tdb.");
 525                         goto fail;
 526                 } else if (!(tdb_flags & TDB_INTERNAL)
 527                            && !(open_flags & O_CREAT)) {
 528                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 529                                            TDB_LOG_USE_ERROR,
 530                                            "tdb_open:"
 531                                            " cannot set TDB_ATTRIBUTE_SEED"
 532                                            " without O_CREAT.");
 533                         goto fail;
 534                 }
 535         }
 536
 537         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 538                 ecode = tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
 539                                    "tdb_open: can't open tdb %s write-only",
 540                                    name);
 541                 goto fail;
 542         }
 543
 544         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 545                 openlock = F_RDLCK;
 546                 tdb->flags |= TDB_RDONLY;
 547         } else {
 548                 if (tdb_flags & TDB_RDONLY) {
 549                         ecode = tdb_logerr(tdb, TDB_ERR_EINVAL,
 550                                            TDB_LOG_USE_ERROR,
 551                                            "tdb_open: can't use TDB_RDONLY"
 552                                            " without O_RDONLY");
 553                         goto fail;
 554                 }
 555                 openlock = F_WRLCK;
 556         }
 557
 558         /* internal databases don't need any of the rest. */
 559         if (tdb->flags & TDB_INTERNAL) {
 560                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 561                 ecode = tdb_new_file(tdb);
 562                 if (ecode != TDB_SUCCESS) {
 563                         goto fail;
 564                 }
 565                 tdb->file->fd = -1;
 566                 if (tdb->flags & TDB_VERSION1)
 567                         ecode = tdb1_new_database(tdb, hsize_attr, maxsize_attr);
 568                 else {
 569                         ecode = tdb_new_database(tdb, seed, &hdr);
 570                         if (ecode == TDB_SUCCESS) {
 571                                 tdb_convert(tdb, &hdr.hash_seed,
 572                                             sizeof(hdr.hash_seed));
 573                                 tdb->hash_seed = hdr.hash_seed;
 574                                 tdb2_context_init(tdb);
 575                                 tdb_ftable_init(tdb);
 576                         }
 577                 }
 578                 if (ecode != TDB_SUCCESS) {
 579                         goto fail;
 580                 }
 581                 return tdb;
 582         }
 583
 584         if (stat(name, &st) != -1)
 585                 tdb->file = find_file(st.st_dev, st.st_ino);
 586
 587         if (!tdb->file) {
 588                 int fd;
 589
 590                 if ((fd = open(name, open_flags, mode)) == -1) {
 591                         /* errno set by open(2) */
 592                         saved_errno = errno;
 593                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 594                                    "tdb_open: could not open file %s: %s",
 595                                    name, strerror(errno));
 596                         goto fail_errno;
 597                 }
 598
 599                 /* on exec, don't inherit the fd */
 600                 v = fcntl(fd, F_GETFD, 0);
 601                 fcntl(fd, F_SETFD, v | FD_CLOEXEC);
 602
 603                 if (fstat(fd, &st) == -1) {
 604                         saved_errno = errno;
 605                         tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 606                                    "tdb_open: could not stat open %s: %s",
 607                                    name, strerror(errno));
 608                         close(fd);
 609                         goto fail_errno;
 610                 }
 611
 612                 ecode = tdb_new_file(tdb);
 613                 if (ecode != TDB_SUCCESS) {
 614                         close(fd);
 615                         goto fail;
 616                 }
 617
 618                 tdb->file->fd = fd;
 619                 tdb->file->device = st.st_dev;
 620                 tdb->file->inode = st.st_ino;
 621                 tdb->file->map_ptr = NULL;
 622                 tdb->file->map_size = 0;
 623         }
 624
 625         /* ensure there is only one process initialising at once */
 626         ecode = tdb_lock_open(tdb, openlock, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
 627         if (ecode != TDB_SUCCESS) {
 628                 saved_errno = errno;
 629                 goto fail_errno;
 630         }
 631
 632         /* call their open hook if they gave us one. */
 633         if (tdb->openhook) {
 634                 ecode = tdb->openhook(tdb->file->fd, tdb->openhook_data);
 635                 if (ecode != TDB_SUCCESS) {
 636                         tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
 637                                    "tdb_open: open hook failed");
 638                         goto fail;
 639                 }
 640                 open_flags |= O_CREAT;
 641         }
 642
 643         /* If they used O_TRUNC, read will return 0. */
 644         rlen = pread(tdb->file->fd, &hdr, sizeof(hdr), 0);
 645         if (rlen == 0 && (open_flags & O_CREAT)) {
 646                 if (tdb->flags & TDB_VERSION1) {
 647                         ecode = tdb1_new_database(tdb, hsize_attr, maxsize_attr);
 648                         if (ecode != TDB_SUCCESS)
 649                                 goto fail;
 650                         goto finished;
 651                 }
 652                 ecode = tdb_new_database(tdb, seed, &hdr);
 653                 if (ecode != TDB_SUCCESS) {
 654                         goto fail;
 655                 }
 656         } else if (rlen < 0) {
 657                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 658                                    "tdb_open: error %s reading %s",
 659                                    strerror(errno), name);
 660                 goto fail;
 661         } else if (rlen < sizeof(hdr)
 662                    || strcmp(hdr.magic_food, TDB_MAGIC_FOOD) != 0) {
 663                 if (is_tdb1(&tdb->tdb1.header, &hdr, rlen)) {
 664                         ecode = tdb1_open(tdb, maxsize_attr);
 665                         if (!ecode)
 666                                 goto finished;
 667                         goto fail;
 668                 }
 669                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 670                                    "tdb_open: %s is not a tdb file", name);
 671                 goto fail;
 672         }
 673
 674         if (hdr.version != TDB_VERSION) {
 675                 if (hdr.version == bswap_64(TDB_VERSION))
 676                         tdb->flags |= TDB_CONVERT;
 677                 else {
 678                         if (is_tdb1(&tdb->tdb1.header, &hdr, rlen)) {
 679                                 ecode = tdb1_open(tdb, maxsize_attr);
 680                                 if (!ecode)
 681                                         goto finished;
 682                                 goto fail;
 683                         }
 684                         /* wrong version */
 685                         ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 686                                            "tdb_open:"
 687                                            " %s is unknown version 0x%llx",
 688                                            name, (long long)hdr.version);
 689                         goto fail;
 690                 }
 691         } else if (tdb->flags & TDB_CONVERT) {
 692                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 693                                    "tdb_open:"
 694                                    " %s does not need TDB_CONVERT",
 695                                    name);
 696                 goto fail;
 697         }
 698
 699         /* This is a version2 tdb. */
 700         if (tdb->flags & TDB_VERSION1) {
 701                 tdb->flags &= ~TDB_VERSION1;
 702         }
 703
 704         tdb2_context_init(tdb);
 705
 706         tdb_convert(tdb, &hdr, sizeof(hdr));
 707         tdb->hash_seed = hdr.hash_seed;
 708         hash_test = TDB_HASH_MAGIC;
 709         hash_test = tdb_hash(tdb, &hash_test, sizeof(hash_test));
 710         if (hdr.hash_test != hash_test) {
 711                 /* wrong hash variant */
 712                 ecode = tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 713                                    "tdb_open:"
 714                                    " %s uses a different hash function",
 715                                    name);
 716                 goto fail;
 717         }
 718
 719         ecode = capabilities_ok(tdb, hdr.capabilities);
 720         if (ecode != TDB_SUCCESS) {
 721                 goto fail;
 722         }
 723
 724         /* Clear any features we don't understand. */
 725         if ((open_flags & O_ACCMODE) != O_RDONLY) {
 726                 hdr.features_used &= TDB_FEATURE_MASK;
 727                 ecode = tdb_write_convert(tdb, offsetof(struct tdb_header,
 728                                                         features_used),
 729                                           &hdr.features_used,
 730                                           sizeof(hdr.features_used));
 731                 if (ecode != TDB_SUCCESS)
 732                         goto fail;
 733         }
 734
 735 finished:
 736         if (tdb->flags & TDB_VERSION1) {
 737                 /* if needed, run recovery */
 738                 if (tdb1_transaction_recover(tdb) == -1) {
 739                         ecode = tdb->last_error;
 740                         goto fail;
 741                 }
 742         }
 743
 744         tdb_unlock_open(tdb, openlock);
 745
 746         /* This makes sure we have current map_size and mmap. */
 747         if (tdb->flags & TDB_VERSION1) {
 748                 ecode = tdb1_probe_length(tdb);
 749         } else {
 750                 ecode = tdb->tdb2.io->oob(tdb, tdb->file->map_size + 1, true);
 751         }
 752         if (unlikely(ecode != TDB_SUCCESS))
 753                 goto fail;
 754
 755         if (!(tdb->flags & TDB_VERSION1)) {
 756                 /* Now it's fully formed, recover if necessary. */
 757                 berr = tdb_needs_recovery(tdb);
 758                 if (unlikely(berr != false)) {
 759                         if (berr < 0) {
 760                                 ecode = TDB_OFF_TO_ERR(berr);
 761                                 goto fail;
 762                         }
 763                         ecode = tdb_lock_and_recover(tdb);
 764                         if (ecode != TDB_SUCCESS) {
 765                                 goto fail;
 766                         }
 767                 }
 768
 769                 ecode = tdb_ftable_init(tdb);
 770                 if (ecode != TDB_SUCCESS) {
 771                         goto fail;
 772                 }
 773         }
 774
 775         tdb->next = tdbs;
 776         tdbs = tdb;
 777         return tdb;
 778
 779  fail:
 780         /* Map ecode to some logical errno. */
 781         switch (TDB_ERR_TO_OFF(ecode)) {
 782         case TDB_ERR_TO_OFF(TDB_ERR_CORRUPT):
 783         case TDB_ERR_TO_OFF(TDB_ERR_IO):
 784                 saved_errno = EIO;
 785                 break;
 786         case TDB_ERR_TO_OFF(TDB_ERR_LOCK):
 787                 saved_errno = EWOULDBLOCK;
 788                 break;
 789         case TDB_ERR_TO_OFF(TDB_ERR_OOM):
 790                 saved_errno = ENOMEM;
 791                 break;
 792         case TDB_ERR_TO_OFF(TDB_ERR_EINVAL):
 793                 saved_errno = EINVAL;
 794                 break;
 795         default:
 796                 saved_errno = EINVAL;
 797                 break;
 798         }
 799
 800 fail_errno:
 801 #ifdef TDB_TRACE
 802         close(tdb->tracefd);
 803 #endif
 804         if (tdb->file) {
 805                 tdb_lock_cleanup(tdb);
 806                 if (--tdb->file->refcnt == 0) {
 807                         assert(tdb->file->num_lockrecs == 0);
 808                         if (tdb->file->map_ptr) {
 809                                 if (tdb->flags & TDB_INTERNAL) {
 810                                         free(tdb->file->map_ptr);
 811                                 } else
 812                                         tdb_munmap(tdb->file);
 813                         }
 814                         if (close(tdb->file->fd) != 0)
 815                                 tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
 816                                            "tdb_open: failed to close tdb fd"
 817                                            " on error: %s", strerror(errno));
 818                         free(tdb->file->lockrecs);
 819                         free(tdb->file);
 820                 }
 821         }
 822
 823         free(tdb);
 824         errno = saved_errno;
 825         return NULL;
 826 }
 827
 828 int tdb_close(struct tdb_context *tdb)
 829 {
 830         int ret = 0;
 831         struct tdb_context **i;
 832
 833         tdb_trace(tdb, "tdb_close");
 834
 835         if (tdb->flags & TDB_VERSION1) {
 836                 if (tdb->tdb1.transaction) {
 837                         tdb1_transaction_cancel(tdb);
 838                 }
 839         } else {
 840                 if (tdb->tdb2.transaction) {
 841                         tdb_transaction_cancel(tdb);
 842                 }
 843         }
 844
 845         if (tdb->file->map_ptr) {
 846                 if (tdb->flags & TDB_INTERNAL)
 847                         free(tdb->file->map_ptr);
 848                 else
 849                         tdb_munmap(tdb->file);
 850         }
 851         if (tdb->file) {
 852                 tdb_lock_cleanup(tdb);
 853                 if (--tdb->file->refcnt == 0) {
 854                         ret = close(tdb->file->fd);
 855                         free(tdb->file->lockrecs);
 856                         free(tdb->file);
 857                 }
 858         }
 859
 860         /* Remove from tdbs list */
 861         for (i = &tdbs; *i; i = &(*i)->next) {
 862                 if (*i == tdb) {
 863                         *i = tdb->next;
 864                         break;
 865                 }
 866         }
 867
 868 #ifdef TDB_TRACE
 869         close(tdb->tracefd);
 870 #endif
 871         free(tdb);
 872
 873         return ret;
 874 }
 875
 876 void tdb_foreach_(int (*fn)(struct tdb_context *, void *), void *p)
 877 {
 878         struct tdb_context *i;
 879
 880         for (i = tdbs; i; i = i->next) {
 881                 if (fn(i, p) != 0)
 882                         break;
 883         }
 884 }