git.ozlabs.org Git - ccan/blob - ccan/tdb2/tdb.c

   1 #include "private.h"
   2 #include <ccan/tdb2/tdb2.h>
   3 #include <ccan/hash/hash.h>
   4 #include <ccan/build_assert/build_assert.h>
   5 #include <ccan/likely/likely.h>
   6 #include <assert.h>
   7
   8 /* The null return. */
   9 struct tdb_data tdb_null = { .dptr = NULL, .dsize = 0 };
  10
  11 /* all contexts, to ensure no double-opens (fcntl locks don't nest!) */
  12 static struct tdb_context *tdbs = NULL;
  13
  14 PRINTF_ATTRIBUTE(4, 5) static void
  15 null_log_fn(struct tdb_context *tdb,
  16             enum tdb_debug_level level, void *priv,
  17             const char *fmt, ...)
  18 {
  19 }
  20
  21 /* We do a lot of work assuming our copy of the header volatile area
  22  * is uptodate, and usually it is.  However, once we grab a lock, we have to
  23  * re-check it. */
  24 bool header_changed(struct tdb_context *tdb)
  25 {
  26         uint64_t gen;
  27
  28         if (!(tdb->flags & TDB_NOLOCK) && tdb->header_uptodate) {
  29                 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
  30                          "warning: header uptodate already\n");
  31         }
  32
  33         /* We could get a partial update if we're not holding any locks. */
  34         assert((tdb->flags & TDB_NOLOCK) || tdb_has_locks(tdb));
  35
  36         tdb->header_uptodate = true;
  37         gen = tdb_read_off(tdb, offsetof(struct tdb_header, v.generation));
  38         if (unlikely(gen != tdb->header.v.generation)) {
  39                 tdb_read_convert(tdb, offsetof(struct tdb_header, v),
  40                                  &tdb->header.v, sizeof(tdb->header.v));
  41                 return true;
  42         }
  43         return false;
  44 }
  45
  46 int write_header(struct tdb_context *tdb)
  47 {
  48         assert(tdb_read_off(tdb, offsetof(struct tdb_header, v.generation))
  49                == tdb->header.v.generation);
  50         tdb->header.v.generation++;
  51         return tdb_write_convert(tdb, offsetof(struct tdb_header, v),
  52                                  &tdb->header.v, sizeof(tdb->header.v));
  53 }
  54
  55 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
  56                              void *arg)
  57 {
  58         return hash64_stable((const unsigned char *)key, length, seed);
  59 }
  60
  61 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
  62 {
  63         return tdb->khash(ptr, len, tdb->header.hash_seed, tdb->hash_priv);
  64 }
  65
  66 static bool tdb_already_open(dev_t device, ino_t ino)
  67 {
  68         struct tdb_context *i;
  69
  70         for (i = tdbs; i; i = i->next) {
  71                 if (i->device == device && i->inode == ino) {
  72                         return true;
  73                 }
  74         }
  75
  76         return false;
  77 }
  78
  79 static uint64_t random_number(struct tdb_context *tdb)
  80 {
  81         int fd;
  82         uint64_t ret = 0;
  83         struct timeval now;
  84
  85         fd = open("/dev/urandom", O_RDONLY);
  86         if (fd >= 0) {
  87                 if (tdb_read_all(fd, &ret, sizeof(ret))) {
  88                         tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
  89                                  "tdb_open: random from /dev/urandom\n");
  90                         close(fd);
  91                         return ret;
  92                 }
  93                 close(fd);
  94         }
  95         /* FIXME: Untested!  Based on Wikipedia protocol description! */
  96         fd = open("/dev/egd-pool", O_RDWR);
  97         if (fd >= 0) {
  98                 /* Command is 1, next byte is size we want to read. */
  99                 char cmd[2] = { 1, sizeof(uint64_t) };
 100                 if (write(fd, cmd, sizeof(cmd)) == sizeof(cmd)) {
 101                         char reply[1 + sizeof(uint64_t)];
 102                         int r = read(fd, reply, sizeof(reply));
 103                         if (r > 1) {
 104                                 tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 105                                          "tdb_open: %u random bytes from"
 106                                          " /dev/egd-pool\n", r-1);
 107                                 /* Copy at least some bytes. */
 108                                 memcpy(&ret, reply+1, r - 1);
 109                                 if (reply[0] == sizeof(uint64_t)
 110                                     && r == sizeof(reply)) {
 111                                         close(fd);
 112                                         return ret;
 113                                 }
 114                         }
 115                 }
 116                 close(fd);
 117         }
 118
 119         /* Fallback: pid and time. */
 120         gettimeofday(&now, NULL);
 121         ret = getpid() * 100132289ULL + now.tv_sec * 1000000ULL + now.tv_usec;
 122         tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 123                  "tdb_open: random from getpid and time\n");
 124         return ret;
 125 }
 126
 127 struct new_db_head {
 128         struct tdb_header hdr;
 129         struct free_zone_header zhdr;
 130         tdb_off_t free[BUCKETS_FOR_ZONE(INITIAL_ZONE_BITS) + 1];
 131         struct tdb_used_record hrec;
 132         tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
 133         struct tdb_free_record frec;
 134 };
 135
 136 struct new_database {
 137         struct new_db_head h;
 138         /* Rest up to 1 << INITIAL_ZONE_BITS is empty. */
 139         char space[(1 << INITIAL_ZONE_BITS)
 140                    - (sizeof(struct new_db_head) - sizeof(struct tdb_header))];
 141         uint8_t tailer;
 142         /* Don't count final padding! */
 143 };
 144
 145 /* initialise a new database */
 146 static int tdb_new_database(struct tdb_context *tdb)
 147 {
 148         /* We make it up in memory, then write it out if not internal */
 149         struct new_database newdb;
 150         unsigned int bucket, magic_off, dbsize;
 151
 152         /* Don't want any extra padding! */
 153         dbsize = offsetof(struct new_database, tailer) + sizeof(newdb.tailer);
 154
 155         /* Fill in the header */
 156         newdb.h.hdr.version = TDB_VERSION;
 157         newdb.h.hdr.hash_seed = random_number(tdb);
 158         newdb.h.hdr.hash_test = TDB_HASH_MAGIC;
 159         newdb.h.hdr.hash_test = tdb->khash(&newdb.h.hdr.hash_test,
 160                                            sizeof(newdb.h.hdr.hash_test),
 161                                            newdb.h.hdr.hash_seed,
 162                                            tdb->hash_priv);
 163         memset(newdb.h.hdr.reserved, 0, sizeof(newdb.h.hdr.reserved));
 164         newdb.h.hdr.v.generation = 0;
 165         /* Initial hashes are empty. */
 166         newdb.h.hdr.v.hash_bits = INITIAL_HASH_BITS;
 167         newdb.h.hdr.v.hash_off = offsetof(struct new_database, h.hash);
 168         set_header(tdb, &newdb.h.hrec, 0,
 169                    sizeof(newdb.h.hash), sizeof(newdb.h.hash), 0,
 170                    INITIAL_ZONE_BITS);
 171         memset(newdb.h.hash, 0, sizeof(newdb.h.hash));
 172
 173         /* Create the single free entry. */
 174         newdb.h.frec.magic_and_meta = TDB_FREE_MAGIC | INITIAL_ZONE_BITS;
 175         newdb.h.frec.data_len = (sizeof(newdb.h.frec)
 176                                  - sizeof(struct tdb_used_record)
 177                                  + sizeof(newdb.space));
 178
 179         /* Free is mostly empty... */
 180         newdb.h.zhdr.zone_bits = INITIAL_ZONE_BITS;
 181         memset(newdb.h.free, 0, sizeof(newdb.h.free));
 182
 183         /* ... except for this one bucket. */
 184         bucket = size_to_bucket(INITIAL_ZONE_BITS, newdb.h.frec.data_len);
 185         newdb.h.free[bucket] = offsetof(struct new_database, h.frec);
 186         newdb.h.frec.next = newdb.h.frec.prev = 0;
 187
 188         /* Clear free space to keep valgrind happy, and avoid leaking stack. */
 189         memset(newdb.space, 0, sizeof(newdb.space));
 190
 191         /* Tailer contains maximum number of free_zone bits. */
 192         newdb.tailer = INITIAL_ZONE_BITS;
 193
 194         /* Magic food */
 195         memset(newdb.h.hdr.magic_food, 0, sizeof(newdb.h.hdr.magic_food));
 196         strcpy(newdb.h.hdr.magic_food, TDB_MAGIC_FOOD);
 197
 198         /* This creates an endian-converted database, as if read from disk */
 199         magic_off = offsetof(struct tdb_header, magic_food);
 200         tdb_convert(tdb,
 201                     (char *)&newdb.h.hdr + magic_off,
 202                     dbsize - 1 - magic_off);
 203
 204         tdb->header = newdb.h.hdr;
 205
 206         if (tdb->flags & TDB_INTERNAL) {
 207                 tdb->map_size = dbsize;
 208                 tdb->map_ptr = malloc(tdb->map_size);
 209                 if (!tdb->map_ptr) {
 210                         tdb->ecode = TDB_ERR_OOM;
 211                         return -1;
 212                 }
 213                 memcpy(tdb->map_ptr, &newdb, tdb->map_size);
 214                 return 0;
 215         }
 216         if (lseek(tdb->fd, 0, SEEK_SET) == -1)
 217                 return -1;
 218
 219         if (ftruncate(tdb->fd, 0) == -1)
 220                 return -1;
 221
 222         if (!tdb_pwrite_all(tdb->fd, &newdb, dbsize, 0)) {
 223                 tdb->ecode = TDB_ERR_IO;
 224                 return -1;
 225         }
 226         return 0;
 227 }
 228
 229 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 230                              int open_flags, mode_t mode,
 231                              union tdb_attribute *attr)
 232 {
 233         struct tdb_context *tdb;
 234         struct stat st;
 235         int save_errno;
 236         uint64_t hash_test;
 237         unsigned v;
 238
 239         tdb = malloc(sizeof(*tdb));
 240         if (!tdb) {
 241                 /* Can't log this */
 242                 errno = ENOMEM;
 243                 goto fail;
 244         }
 245         tdb->name = NULL;
 246         tdb->map_ptr = NULL;
 247         tdb->fd = -1;
 248         tdb->map_size = sizeof(struct tdb_header);
 249         tdb->ecode = TDB_SUCCESS;
 250         /* header will be read in below. */
 251         tdb->header_uptodate = false;
 252         tdb->flags = tdb_flags;
 253         tdb->log = null_log_fn;
 254         tdb->log_priv = NULL;
 255         tdb->khash = jenkins_hash;
 256         tdb->hash_priv = NULL;
 257         tdb->transaction = NULL;
 258         /* last_zone will be set below. */
 259         tdb_io_init(tdb);
 260         tdb_lock_init(tdb);
 261
 262         while (attr) {
 263                 switch (attr->base.attr) {
 264                 case TDB_ATTRIBUTE_LOG:
 265                         tdb->log = attr->log.log_fn;
 266                         tdb->log_priv = attr->log.log_private;
 267                         break;
 268                 case TDB_ATTRIBUTE_HASH:
 269                         tdb->khash = attr->hash.hash_fn;
 270                         tdb->hash_priv = attr->hash.hash_private;
 271                         break;
 272                 default:
 273                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 274                                  "tdb_open: unknown attribute type %u\n",
 275                                  attr->base.attr);
 276                         errno = EINVAL;
 277                         goto fail;
 278                 }
 279                 attr = attr->base.next;
 280         }
 281
 282         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 283                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 284                          "tdb_open: can't open tdb %s write-only\n", name);
 285                 errno = EINVAL;
 286                 goto fail;
 287         }
 288
 289         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 290                 tdb->read_only = true;
 291                 /* read only databases don't do locking */
 292                 tdb->flags |= TDB_NOLOCK;
 293         } else
 294                 tdb->read_only = false;
 295
 296         /* internal databases don't need any of the rest. */
 297         if (tdb->flags & TDB_INTERNAL) {
 298                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 299                 if (tdb_new_database(tdb) != 0) {
 300                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 301                                  "tdb_open: tdb_new_database failed!");
 302                         goto fail;
 303                 }
 304                 TEST_IT(tdb->flags & TDB_CONVERT);
 305                 tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 306                 tdb_zone_init(tdb);
 307                 return tdb;
 308         }
 309
 310         if ((tdb->fd = open(name, open_flags, mode)) == -1) {
 311                 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
 312                          "tdb_open: could not open file %s: %s\n",
 313                          name, strerror(errno));
 314                 goto fail;      /* errno set by open(2) */
 315         }
 316
 317         /* on exec, don't inherit the fd */
 318         v = fcntl(tdb->fd, F_GETFD, 0);
 319         fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
 320
 321         /* ensure there is only one process initialising at once */
 322         if (tdb_lock_open(tdb) == -1) {
 323                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 324                          "tdb_open: failed to get open lock on %s: %s\n",
 325                          name, strerror(errno));
 326                 goto fail;      /* errno set by tdb_brlock */
 327         }
 328
 329         if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
 330             || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
 331                 if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
 332                         if (errno == 0) {
 333                                 errno = EIO; /* ie bad format or something */
 334                         }
 335                         goto fail;
 336                 }
 337         } else if (tdb->header.version != TDB_VERSION) {
 338                 if (tdb->header.version == bswap_64(TDB_VERSION))
 339                         tdb->flags |= TDB_CONVERT;
 340                 else {
 341                         /* wrong version */
 342                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 343                                  "tdb_open: %s is unknown version 0x%llx\n",
 344                                  name, (long long)tdb->header.version);
 345                         errno = EIO;
 346                         goto fail;
 347                 }
 348         }
 349
 350         tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 351         hash_test = TDB_HASH_MAGIC;
 352         hash_test = tdb->khash(&hash_test, sizeof(hash_test),
 353                                tdb->header.hash_seed, tdb->hash_priv);
 354         if (tdb->header.hash_test != hash_test) {
 355                 /* wrong hash variant */
 356                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 357                          "tdb_open: %s uses a different hash function\n",
 358                          name);
 359                 errno = EIO;
 360                 goto fail;
 361         }
 362
 363         if (fstat(tdb->fd, &st) == -1)
 364                 goto fail;
 365
 366         /* Is it already in the open list?  If so, fail. */
 367         if (tdb_already_open(st.st_dev, st.st_ino)) {
 368                 /* FIXME */
 369                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 370                          "tdb_open: %s (%d,%d) is already open in this process\n",
 371                          name, (int)st.st_dev, (int)st.st_ino);
 372                 errno = EBUSY;
 373                 goto fail;
 374         }
 375
 376         tdb->name = strdup(name);
 377         if (!tdb->name) {
 378                 errno = ENOMEM;
 379                 goto fail;
 380         }
 381
 382         tdb->device = st.st_dev;
 383         tdb->inode = st.st_ino;
 384         tdb_unlock_open(tdb);
 385
 386         /* This make sure we have current map_size and mmap. */
 387         tdb->methods->oob(tdb, tdb->map_size + 1, true);
 388
 389         /* Now we can pick a random free zone to start from. */
 390         if (tdb_zone_init(tdb) == -1)
 391                 goto fail;
 392
 393         tdb->next = tdbs;
 394         tdbs = tdb;
 395         return tdb;
 396
 397  fail:
 398         save_errno = errno;
 399
 400         if (!tdb)
 401                 return NULL;
 402
 403 #ifdef TDB_TRACE
 404         close(tdb->tracefd);
 405 #endif
 406         if (tdb->map_ptr) {
 407                 if (tdb->flags & TDB_INTERNAL) {
 408                         free(tdb->map_ptr);
 409                 } else
 410                         tdb_munmap(tdb);
 411         }
 412         free((char *)tdb->name);
 413         if (tdb->fd != -1)
 414                 if (close(tdb->fd) != 0)
 415                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 416                                  "tdb_open: failed to close tdb->fd"
 417                                  " on error!\n");
 418         free(tdb);
 419         errno = save_errno;
 420         return NULL;
 421 }
 422
 423 tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
 424 {
 425         return tdb->header.v.hash_off
 426                 + ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
 427                    * sizeof(tdb_off_t));
 428 }
 429
 430 /* Returns 0 if the entry is a zero (definitely not a match).
 431  * Returns a valid entry offset if it's a match.  Fills in rec.
 432  * Otherwise returns TDB_OFF_ERR: keep searching. */
 433 static tdb_off_t entry_matches(struct tdb_context *tdb,
 434                                uint64_t list,
 435                                uint64_t hash,
 436                                const struct tdb_data *key,
 437                                struct tdb_used_record *rec)
 438 {
 439         tdb_off_t off;
 440         uint64_t keylen;
 441         const unsigned char *rkey;
 442
 443         list &= ((1ULL << tdb->header.v.hash_bits) - 1);
 444
 445         off = tdb_read_off(tdb, tdb->header.v.hash_off
 446                            + list * sizeof(tdb_off_t));
 447         if (off == 0 || off == TDB_OFF_ERR)
 448                 return off;
 449
 450 #if 0 /* FIXME: Check other bits. */
 451         unsigned int bits, bitmask, hoffextra;
 452         /* Bottom three bits show how many extra hash bits. */
 453         bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
 454         bitmask = (1 << bits)-1;
 455         hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
 456         uint64_t hextra = hash >> tdb->header.v.hash_bits;
 457         if ((hextra & bitmask) != hoffextra)
 458                 return TDB_OFF_ERR;
 459         off &= ~...;
 460 #endif
 461
 462         if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
 463                 return TDB_OFF_ERR;
 464
 465         /* FIXME: check extra bits in header! */
 466         keylen = rec_key_length(rec);
 467         if (keylen != key->dsize)
 468                 return TDB_OFF_ERR;
 469
 470         rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
 471         if (!rkey)
 472                 return TDB_OFF_ERR;
 473         if (memcmp(rkey, key->dptr, keylen) != 0)
 474                 off = TDB_OFF_ERR;
 475         tdb_access_release(tdb, rkey);
 476         return off;
 477 }
 478
 479 /* FIXME: Optimize? */
 480 static void unlock_lists(struct tdb_context *tdb,
 481                          tdb_off_t list, tdb_len_t num,
 482                          int ltype)
 483 {
 484         tdb_off_t i;
 485
 486         for (i = list; i < list + num; i++)
 487                 tdb_unlock_list(tdb, i, ltype);
 488 }
 489
 490 /* FIXME: Optimize? */
 491 static int lock_lists(struct tdb_context *tdb,
 492                       tdb_off_t list, tdb_len_t num,
 493                       int ltype)
 494 {
 495         tdb_off_t i;
 496
 497         for (i = list; i < list + num; i++) {
 498                 if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
 499                     == TDB_OFF_ERR) {
 500                         unlock_lists(tdb, list, i - list, ltype);
 501                         return -1;
 502                 }
 503         }
 504         return 0;
 505 }
 506
 507 /* We lock hashes up to the next empty offset.  We already hold the
 508  * lock on the start bucket, but we may need to release and re-grab
 509  * it.  If we fail, we hold no locks at all! */
 510 static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
 511                                      tdb_off_t start, int ltype)
 512 {
 513         tdb_len_t num, len;
 514
 515 again:
 516         num = 1ULL << tdb->header.v.hash_bits;
 517         len = tdb_find_zero_off(tdb, hash_off(tdb, start), num - start);
 518         if (unlikely(len == num - start)) {
 519                 /* We hit the end of the hash range.  Drop lock: we have
 520                    to lock start of hash first. */
 521                 tdb_len_t pre_locks;
 522
 523                 tdb_unlock_list(tdb, start, ltype);
 524
 525                 /* Grab something, so header is stable. */
 526                 if (tdb_lock_list(tdb, 0, ltype, TDB_LOCK_WAIT))
 527                         return TDB_OFF_ERR;
 528                 pre_locks = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
 529                 /* We want to lock the zero entry as well. */
 530                 pre_locks++;
 531                 if (lock_lists(tdb, 1, pre_locks - 1, ltype) == -1) {
 532                         tdb_unlock_list(tdb, 0, ltype);
 533                         return TDB_OFF_ERR;
 534                 }
 535
 536                 /* Now lock later ones. */
 537                 if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
 538                         unlock_lists(tdb, 0, pre_locks, ltype);
 539                         return TDB_OFF_ERR;
 540                 }
 541                 len += pre_locks;
 542         } else {
 543                 /* We want to lock the zero entry as well. */
 544                 len++;
 545                 /* But we already have lock on start. */
 546                 if (unlikely(lock_lists(tdb, start+1, len-1, ltype) == -1)) {
 547                         tdb_unlock_list(tdb, start, ltype);
 548                         return TDB_OFF_ERR;
 549                 }
 550         }
 551
 552         /* Now, did we lose the race, and it's not zero any more? */
 553         if (unlikely(tdb_read_off(tdb, hash_off(tdb, start + len - 1)) != 0)) {
 554                 /* Leave the start locked, as expected. */
 555                 unlock_lists(tdb, start + 1, len - 1, ltype);
 556                 goto again;
 557         }
 558
 559         return len;
 560 }
 561
 562 /* FIXME: modify, don't rewrite! */
 563 static int update_rec_hdr(struct tdb_context *tdb,
 564                           tdb_off_t off,
 565                           tdb_len_t keylen,
 566                           tdb_len_t datalen,
 567                           struct tdb_used_record *rec,
 568                           uint64_t h)
 569 {
 570         uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
 571
 572         if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h,
 573                        rec_zone_bits(rec)))
 574                 return -1;
 575
 576         return tdb_write_convert(tdb, off, rec, sizeof(*rec));
 577 }
 578
 579 static int hash_add(struct tdb_context *tdb,
 580                     uint64_t hash, tdb_off_t off)
 581 {
 582         tdb_off_t i, hoff, len, num;
 583
 584         /* Look for next space. */
 585         i = (hash & ((1ULL << tdb->header.v.hash_bits) - 1));
 586         len = (1ULL << tdb->header.v.hash_bits) - i;
 587         num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
 588
 589         if (unlikely(num == len)) {
 590                 /* We wrapped.  Look through start of hash table. */
 591                 i = 0;
 592                 hoff = hash_off(tdb, 0);
 593                 len = (1ULL << tdb->header.v.hash_bits);
 594                 num = tdb_find_zero_off(tdb, hoff, len);
 595                 if (num == len) {
 596                         tdb->ecode = TDB_ERR_CORRUPT;
 597                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 598                                  "hash_add: full hash table!\n");
 599                         return -1;
 600                 }
 601         }
 602         if (tdb_read_off(tdb, hash_off(tdb, i + num)) != 0) {
 603                 tdb->ecode = TDB_ERR_CORRUPT;
 604                 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 605                          "hash_add: overwriting hash table?\n");
 606                 return -1;
 607         }
 608
 609         /* FIXME: Encode extra hash bits! */
 610         return tdb_write_off(tdb, hash_off(tdb, i + num), off);
 611 }
 612
 613 /* If we fail, others will try after us. */
 614 static void enlarge_hash(struct tdb_context *tdb)
 615 {
 616         tdb_off_t newoff, oldoff, i;
 617         tdb_len_t hlen;
 618         uint64_t num = 1ULL << tdb->header.v.hash_bits;
 619         struct tdb_used_record pad, *r;
 620         unsigned int records = 0;
 621
 622         /* FIXME: We should do this without holding locks throughout. */
 623         if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
 624                 return;
 625
 626         /* Someone else enlarged for us?  Nothing to do. */
 627         if ((1ULL << tdb->header.v.hash_bits) != num)
 628                 goto unlock;
 629
 630         /* Allocate our new array. */
 631         hlen = num * sizeof(tdb_off_t) * 2;
 632         newoff = alloc(tdb, 0, hlen, 0, false);
 633         if (unlikely(newoff == TDB_OFF_ERR))
 634                 goto unlock;
 635         /* Step over record header! */
 636         newoff += sizeof(struct tdb_used_record);
 637
 638         /* Starts all zero. */
 639         if (zero_out(tdb, newoff, hlen) == -1)
 640                 goto unlock;
 641
 642         /* Update header now so we can use normal routines. */
 643         oldoff = tdb->header.v.hash_off;
 644
 645         tdb->header.v.hash_bits++;
 646         tdb->header.v.hash_off = newoff;
 647
 648         /* FIXME: If the space before is empty, we know this is in its ideal
 649          * location.  Or steal a bit from the pointer to avoid rehash. */
 650         for (i = 0; i < num; i++) {
 651                 tdb_off_t off;
 652                 off = tdb_read_off(tdb, oldoff + i * sizeof(tdb_off_t));
 653                 if (unlikely(off == TDB_OFF_ERR))
 654                         goto oldheader;
 655                 if (off && hash_add(tdb, hash_record(tdb, off), off) == -1)
 656                         goto oldheader;
 657                 if (off)
 658                         records++;
 659         }
 660
 661         tdb->log(tdb, TDB_DEBUG_TRACE, tdb->log_priv,
 662                  "enlarge_hash: moved %u records from %llu buckets.\n",
 663                  records, (long long)num);
 664
 665         /* Free up old hash. */
 666         r = tdb_get(tdb, oldoff - sizeof(*r), &pad, sizeof(*r));
 667         if (!r)
 668                 goto oldheader;
 669         add_free_record(tdb, rec_zone_bits(r), oldoff - sizeof(*r),
 670                         sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
 671
 672         /* Now we write the modified header. */
 673         write_header(tdb);
 674 unlock:
 675         tdb_allrecord_unlock(tdb, F_WRLCK);
 676         return;
 677
 678 oldheader:
 679         tdb->header.v.hash_bits--;
 680         tdb->header.v.hash_off = oldoff;
 681         goto unlock;
 682 }
 683
 684
 685 /* This is the slow version of the routine which searches the
 686  * hashtable for an entry.
 687  * We lock every hash bucket up to and including the next zero one.
 688  */
 689 static tdb_off_t find_and_lock_slow(struct tdb_context *tdb,
 690                                     struct tdb_data key,
 691                                     uint64_t h,
 692                                     int ltype,
 693                                     tdb_off_t *start_lock,
 694                                     tdb_len_t *num_locks,
 695                                     tdb_off_t *bucket,
 696                                     struct tdb_used_record *rec)
 697 {
 698         /* Warning: this may drop the lock on *bucket! */
 699         *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
 700         if (*num_locks == TDB_OFF_ERR)
 701                 return TDB_OFF_ERR;
 702
 703         for (*bucket = *start_lock;
 704              *bucket < *start_lock + *num_locks;
 705              (*bucket)++) {
 706                 tdb_off_t off = entry_matches(tdb, *bucket, h, &key, rec);
 707                 /* Empty entry or we found it? */
 708                 if (off == 0 || off != TDB_OFF_ERR)
 709                         return off;
 710         }
 711
 712         /* We didn't find a zero entry?  Something went badly wrong... */
 713         unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
 714         tdb->ecode = TDB_ERR_CORRUPT;
 715         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 716                  "find_and_lock: expected to find an empty hash bucket!\n");
 717         return TDB_OFF_ERR;
 718 }
 719
 720 /* This is the core routine which searches the hashtable for an entry.
 721  * On error, no locks are held and TDB_OFF_ERR is returned.
 722  * Otherwise, *num_locks locks of type ltype from *start_lock are held.
 723  * The bucket where the entry is (or would be) is in *bucket.
 724  * If not found, the return value is 0.
 725  * If found, the return value is the offset, and *rec is the record. */
 726 static tdb_off_t find_and_lock(struct tdb_context *tdb,
 727                                struct tdb_data key,
 728                                uint64_t h,
 729                                int ltype,
 730                                tdb_off_t *start_lock,
 731                                tdb_len_t *num_locks,
 732                                tdb_off_t *bucket,
 733                                struct tdb_used_record *rec)
 734 {
 735         tdb_off_t off;
 736
 737         /* FIXME: can we avoid locks for some fast paths? */
 738         *start_lock = tdb_lock_list(tdb, h, ltype, TDB_LOCK_WAIT);
 739         if (*start_lock == TDB_OFF_ERR)
 740                 return TDB_OFF_ERR;
 741
 742         /* Fast path. */
 743         off = entry_matches(tdb, *start_lock, h, &key, rec);
 744         if (likely(off != TDB_OFF_ERR)) {
 745                 *bucket = *start_lock;
 746                 *num_locks = 1;
 747                 return off;
 748         }
 749
 750         /* Slow path, need to grab more locks and search. */
 751         return find_and_lock_slow(tdb, key, h, ltype, start_lock, num_locks,
 752                                   bucket, rec);
 753 }
 754
 755 /* Returns -1 on error, 0 on OK" */
 756 static int replace_data(struct tdb_context *tdb,
 757                         uint64_t h, struct tdb_data key, struct tdb_data dbuf,
 758                         tdb_off_t bucket,
 759                         tdb_off_t old_off, tdb_len_t old_room,
 760                         unsigned old_zone,
 761                         bool growing)
 762 {
 763         tdb_off_t new_off;
 764
 765         /* Allocate a new record. */
 766         new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
 767         if (unlikely(new_off == TDB_OFF_ERR))
 768                 return -1;
 769
 770         /* We didn't like the existing one: remove it. */
 771         if (old_off)
 772                 add_free_record(tdb, old_zone, old_off,
 773                                 sizeof(struct tdb_used_record)
 774                                 + key.dsize + old_room);
 775
 776         /* FIXME: Encode extra hash bits! */
 777         if (tdb_write_off(tdb, hash_off(tdb, bucket), new_off) == -1)
 778                 return -1;
 779
 780         new_off += sizeof(struct tdb_used_record);
 781         if (tdb->methods->write(tdb, new_off, key.dptr, key.dsize) == -1)
 782                 return -1;
 783
 784         new_off += key.dsize;
 785         if (tdb->methods->write(tdb, new_off, dbuf.dptr, dbuf.dsize) == -1)
 786                 return -1;
 787
 788         /* FIXME: tdb_increment_seqnum(tdb); */
 789         return 0;
 790 }
 791
 792 int tdb_store(struct tdb_context *tdb,
 793               struct tdb_data key, struct tdb_data dbuf, int flag)
 794 {
 795         tdb_off_t off, bucket, start, num;
 796         tdb_len_t old_room = 0;
 797         struct tdb_used_record rec;
 798         uint64_t h;
 799         int ret;
 800
 801         h = tdb_hash(tdb, key.dptr, key.dsize);
 802         off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
 803         if (unlikely(off == TDB_OFF_ERR))
 804                 return -1;
 805
 806         /* Now we have lock on this hash bucket. */
 807         if (flag == TDB_INSERT) {
 808                 if (off) {
 809                         tdb->ecode = TDB_ERR_EXISTS;
 810                         goto fail;
 811                 }
 812         } else {
 813                 if (off) {
 814                         old_room = rec_data_length(&rec)
 815                                 + rec_extra_padding(&rec);
 816                         if (old_room >= dbuf.dsize) {
 817                                 /* Can modify in-place.  Easy! */
 818                                 if (update_rec_hdr(tdb, off,
 819                                                    key.dsize, dbuf.dsize,
 820                                                    &rec, h))
 821                                         goto fail;
 822                                 if (tdb->methods->write(tdb, off + sizeof(rec)
 823                                                         + key.dsize,
 824                                                         dbuf.dptr, dbuf.dsize))
 825                                         goto fail;
 826                                 unlock_lists(tdb, start, num, F_WRLCK);
 827                                 return 0;
 828                         }
 829                         /* FIXME: See if right record is free? */
 830                 } else {
 831                         if (flag == TDB_MODIFY) {
 832                                 /* if the record doesn't exist and we
 833                                    are in TDB_MODIFY mode then we should fail
 834                                    the store */
 835                                 tdb->ecode = TDB_ERR_NOEXIST;
 836                                 goto fail;
 837                         }
 838                 }
 839         }
 840
 841         /* If we didn't use the old record, this implies we're growing. */
 842         ret = replace_data(tdb, h, key, dbuf, bucket, off, old_room,
 843                            rec_zone_bits(&rec), off != 0);
 844         unlock_lists(tdb, start, num, F_WRLCK);
 845
 846         /* FIXME: by simple simulation, this approximated 60% full.
 847          * Check in real case! */
 848         if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
 849                 enlarge_hash(tdb);
 850
 851         return ret;
 852
 853 fail:
 854         unlock_lists(tdb, start, num, F_WRLCK);
 855         return -1;
 856 }
 857
 858 int tdb_append(struct tdb_context *tdb,
 859                struct tdb_data key, struct tdb_data dbuf)
 860 {
 861         tdb_off_t off, bucket, start, num;
 862         struct tdb_used_record rec;
 863         tdb_len_t old_room = 0, old_dlen;
 864         uint64_t h;
 865         unsigned char *newdata;
 866         struct tdb_data new_dbuf;
 867         int ret;
 868
 869         h = tdb_hash(tdb, key.dptr, key.dsize);
 870         off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
 871         if (unlikely(off == TDB_OFF_ERR))
 872                 return -1;
 873
 874         if (off) {
 875                 old_dlen = rec_data_length(&rec);
 876                 old_room = old_dlen + rec_extra_padding(&rec);
 877
 878                 /* Fast path: can append in place. */
 879                 if (rec_extra_padding(&rec) >= dbuf.dsize) {
 880                         if (update_rec_hdr(tdb, off, key.dsize,
 881                                            old_dlen + dbuf.dsize, &rec, h))
 882                                 goto fail;
 883
 884                         off += sizeof(rec) + key.dsize + old_dlen;
 885                         if (tdb->methods->write(tdb, off, dbuf.dptr,
 886                                                 dbuf.dsize) == -1)
 887                                 goto fail;
 888
 889                         /* FIXME: tdb_increment_seqnum(tdb); */
 890                         unlock_lists(tdb, start, num, F_WRLCK);
 891                         return 0;
 892                 }
 893                 /* FIXME: Check right record free? */
 894
 895                 /* Slow path. */
 896                 newdata = malloc(key.dsize + old_dlen + dbuf.dsize);
 897                 if (!newdata) {
 898                         tdb->ecode = TDB_ERR_OOM;
 899                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 900                                  "tdb_append: cannot allocate %llu bytes!\n",
 901                                  (long long)key.dsize + old_dlen + dbuf.dsize);
 902                         goto fail;
 903                 }
 904                 if (tdb->methods->read(tdb, off + sizeof(rec) + key.dsize,
 905                                        newdata, old_dlen) != 0) {
 906                         free(newdata);
 907                         goto fail;
 908                 }
 909                 memcpy(newdata + old_dlen, dbuf.dptr, dbuf.dsize);
 910                 new_dbuf.dptr = newdata;
 911                 new_dbuf.dsize = old_dlen + dbuf.dsize;
 912         } else {
 913                 newdata = NULL;
 914                 new_dbuf = dbuf;
 915         }
 916
 917         /* If they're using tdb_append(), it implies they're growing record. */
 918         ret = replace_data(tdb, h, key, new_dbuf, bucket, off, old_room,
 919                            rec_zone_bits(&rec), true);
 920         unlock_lists(tdb, start, num, F_WRLCK);
 921         free(newdata);
 922
 923         /* FIXME: by simple simulation, this approximated 60% full.
 924          * Check in real case! */
 925         if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
 926                 enlarge_hash(tdb);
 927
 928         return ret;
 929
 930 fail:
 931         unlock_lists(tdb, start, num, F_WRLCK);
 932         return -1;
 933 }
 934
 935 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 936 {
 937         tdb_off_t off, start, num, bucket;
 938         struct tdb_used_record rec;
 939         uint64_t h;
 940         struct tdb_data ret;
 941
 942         h = tdb_hash(tdb, key.dptr, key.dsize);
 943         off = find_and_lock(tdb, key, h, F_RDLCK, &start, &num, &bucket, &rec);
 944         if (unlikely(off == TDB_OFF_ERR))
 945                 return tdb_null;
 946
 947         if (!off) {
 948                 tdb->ecode = TDB_ERR_NOEXIST;
 949                 ret = tdb_null;
 950         } else {
 951                 ret.dsize = rec_data_length(&rec);
 952                 ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
 953                                           ret.dsize);
 954         }
 955
 956         unlock_lists(tdb, start, num, F_RDLCK);
 957         return ret;
 958 }
 959
 960 int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 961 {
 962         tdb_off_t i, bucket, off, start, num;
 963         struct tdb_used_record rec;
 964         uint64_t h;
 965
 966         h = tdb_hash(tdb, key.dptr, key.dsize);
 967         start = tdb_lock_list(tdb, h, F_WRLCK, TDB_LOCK_WAIT);
 968         if (unlikely(start == TDB_OFF_ERR))
 969                 return -1;
 970
 971         /* FIXME: Fastpath: if next is zero, we can delete without lock,
 972          * since this lock protects us. */
 973         off = find_and_lock_slow(tdb, key, h, F_WRLCK,
 974                                  &start, &num, &bucket, &rec);
 975         if (unlikely(off == TDB_OFF_ERR))
 976                 return -1;
 977
 978         if (!off) {
 979                 /* FIXME: We could optimize not found case if it mattered, by
 980                  * reading offset after first lock: if it's zero, goto here. */
 981                 unlock_lists(tdb, start, num, F_WRLCK);
 982                 tdb->ecode = TDB_ERR_NOEXIST;
 983                 return -1;
 984         }
 985         /* Since we found the entry, we must have locked it and a zero. */
 986         assert(num >= 2);
 987
 988         /* This actually unlinks it. */
 989         if (tdb_write_off(tdb, hash_off(tdb, bucket), 0) == -1)
 990                 goto unlock_err;
 991
 992         /* Rehash anything following. */
 993         for (i = bucket+1; i != bucket + num - 1; i++) {
 994                 tdb_off_t hoff, off2;
 995                 uint64_t h2;
 996
 997                 hoff = hash_off(tdb, i);
 998                 off2 = tdb_read_off(tdb, hoff);
 999                 if (unlikely(off2 == TDB_OFF_ERR))
1000                         goto unlock_err;
1001
1002                 /* This can happen if we raced. */
1003                 if (unlikely(off2 == 0))
1004                         break;
1005
1006                 /* Maybe use a bit to indicate it is in ideal place? */
1007                 h2 = hash_record(tdb, off2);
1008                 /* Is it happy where it is? */
1009                 if (hash_off(tdb, h2) == hoff)
1010                         continue;
1011
1012                 /* Remove it. */
1013                 if (tdb_write_off(tdb, hoff, 0) == -1)
1014                         goto unlock_err;
1015
1016                 /* Rehash it. */
1017                 if (hash_add(tdb, h2, off2) == -1)
1018                         goto unlock_err;
1019         }
1020
1021         /* Free the deleted entry. */
1022         if (add_free_record(tdb, rec_zone_bits(&rec), off,
1023                             sizeof(struct tdb_used_record)
1024                             + rec_key_length(&rec)
1025                             + rec_data_length(&rec)
1026                             + rec_extra_padding(&rec)) != 0)
1027                 goto unlock_err;
1028
1029         unlock_lists(tdb, start, num, F_WRLCK);
1030         return 0;
1031
1032 unlock_err:
1033         unlock_lists(tdb, start, num, F_WRLCK);
1034         return -1;
1035 }
1036
1037 int tdb_close(struct tdb_context *tdb)
1038 {
1039         struct tdb_context **i;
1040         int ret = 0;
1041
1042         /* FIXME:
1043         if (tdb->transaction) {
1044                 tdb_transaction_cancel(tdb);
1045         }
1046         */
1047         tdb_trace(tdb, "tdb_close");
1048
1049         if (tdb->map_ptr) {
1050                 if (tdb->flags & TDB_INTERNAL)
1051                         free(tdb->map_ptr);
1052                 else
1053                         tdb_munmap(tdb);
1054         }
1055         free((char *)tdb->name);
1056         if (tdb->fd != -1) {
1057                 ret = close(tdb->fd);
1058                 tdb->fd = -1;
1059         }
1060         free(tdb->lockrecs);
1061
1062         /* Remove from contexts list */
1063         for (i = &tdbs; *i; i = &(*i)->next) {
1064                 if (*i == tdb) {
1065                         *i = tdb->next;
1066                         break;
1067                 }
1068         }
1069
1070 #ifdef TDB_TRACE
1071         close(tdb->tracefd);
1072 #endif
1073         free(tdb);
1074
1075         return ret;
1076 }
1077
1078 enum TDB_ERROR tdb_error(struct tdb_context *tdb)
1079 {
1080         return tdb->ecode;
1081 }