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_database {
 128         struct tdb_header hdr;
 129         struct tdb_used_record hrec;
 130         tdb_off_t hash[1ULL << INITIAL_HASH_BITS];
 131         struct tdb_used_record frec;
 132         tdb_off_t free[INITIAL_FREE_BUCKETS + 1]; /* One overflow bucket */
 133 };
 134
 135 /* initialise a new database */
 136 static int tdb_new_database(struct tdb_context *tdb)
 137 {
 138         /* We make it up in memory, then write it out if not internal */
 139         struct new_database newdb;
 140         unsigned int magic_off = offsetof(struct tdb_header, magic_food);
 141
 142         /* Fill in the header */
 143         newdb.hdr.version = TDB_VERSION;
 144         newdb.hdr.hash_seed = random_number(tdb);
 145         newdb.hdr.hash_test = TDB_HASH_MAGIC;
 146         newdb.hdr.hash_test = tdb->khash(&newdb.hdr.hash_test,
 147                                          sizeof(newdb.hdr.hash_test),
 148                                          newdb.hdr.hash_seed,
 149                                          tdb->hash_priv);
 150
 151         newdb.hdr.v.generation = 0;
 152
 153         /* The initial zone must cover the initial database size! */
 154         BUILD_ASSERT((1ULL << INITIAL_ZONE_BITS) >= sizeof(newdb));
 155
 156         /* Free array has 1 zone, 10 buckets.  All buckets empty. */
 157         newdb.hdr.v.num_zones = 1;
 158         newdb.hdr.v.zone_bits = INITIAL_ZONE_BITS;
 159         newdb.hdr.v.free_buckets = INITIAL_FREE_BUCKETS;
 160         newdb.hdr.v.free_off = offsetof(struct new_database, free);
 161         set_header(tdb, &newdb.frec, 0,
 162                    sizeof(newdb.free), sizeof(newdb.free), 0);
 163         memset(newdb.free, 0, sizeof(newdb.free));
 164
 165         /* Initial hashes are empty. */
 166         newdb.hdr.v.hash_bits = INITIAL_HASH_BITS;
 167         newdb.hdr.v.hash_off = offsetof(struct new_database, hash);
 168         set_header(tdb, &newdb.hrec, 0,
 169                    sizeof(newdb.hash), sizeof(newdb.hash), 0);
 170         memset(newdb.hash, 0, sizeof(newdb.hash));
 171
 172         /* Magic food */
 173         memset(newdb.hdr.magic_food, 0, sizeof(newdb.hdr.magic_food));
 174         strcpy(newdb.hdr.magic_food, TDB_MAGIC_FOOD);
 175
 176         /* This creates an endian-converted database, as if read from disk */
 177         tdb_convert(tdb,
 178                     (char *)&newdb.hdr + magic_off,
 179                     sizeof(newdb) - magic_off);
 180
 181         tdb->header = newdb.hdr;
 182
 183         if (tdb->flags & TDB_INTERNAL) {
 184                 tdb->map_size = sizeof(newdb);
 185                 tdb->map_ptr = malloc(tdb->map_size);
 186                 if (!tdb->map_ptr) {
 187                         tdb->ecode = TDB_ERR_OOM;
 188                         return -1;
 189                 }
 190                 memcpy(tdb->map_ptr, &newdb, tdb->map_size);
 191                 return 0;
 192         }
 193         if (lseek(tdb->fd, 0, SEEK_SET) == -1)
 194                 return -1;
 195
 196         if (ftruncate(tdb->fd, 0) == -1)
 197                 return -1;
 198
 199         if (!tdb_pwrite_all(tdb->fd, &newdb, sizeof(newdb), 0)) {
 200                 tdb->ecode = TDB_ERR_IO;
 201                 return -1;
 202         }
 203         return 0;
 204 }
 205
 206 struct tdb_context *tdb_open(const char *name, int tdb_flags,
 207                              int open_flags, mode_t mode,
 208                              union tdb_attribute *attr)
 209 {
 210         struct tdb_context *tdb;
 211         struct stat st;
 212         int save_errno;
 213         uint64_t hash_test;
 214         unsigned v;
 215
 216         tdb = malloc(sizeof(*tdb));
 217         if (!tdb) {
 218                 /* Can't log this */
 219                 errno = ENOMEM;
 220                 goto fail;
 221         }
 222         tdb->name = NULL;
 223         tdb->map_ptr = NULL;
 224         tdb->fd = -1;
 225         /* map_size will be set below. */
 226         tdb->ecode = TDB_SUCCESS;
 227         /* header will be read in below. */
 228         tdb->header_uptodate = false;
 229         tdb->flags = tdb_flags;
 230         tdb->log = null_log_fn;
 231         tdb->log_priv = NULL;
 232         tdb->khash = jenkins_hash;
 233         tdb->hash_priv = NULL;
 234         tdb->transaction = NULL;
 235         /* last_zone will be set below. */
 236         tdb_io_init(tdb);
 237         tdb_lock_init(tdb);
 238
 239         while (attr) {
 240                 switch (attr->base.attr) {
 241                 case TDB_ATTRIBUTE_LOG:
 242                         tdb->log = attr->log.log_fn;
 243                         tdb->log_priv = attr->log.log_private;
 244                         break;
 245                 case TDB_ATTRIBUTE_HASH:
 246                         tdb->khash = attr->hash.hash_fn;
 247                         tdb->hash_priv = attr->hash.hash_private;
 248                         break;
 249                 default:
 250                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 251                                  "tdb_open: unknown attribute type %u\n",
 252                                  attr->base.attr);
 253                         errno = EINVAL;
 254                         goto fail;
 255                 }
 256                 attr = attr->base.next;
 257         }
 258
 259         if ((open_flags & O_ACCMODE) == O_WRONLY) {
 260                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 261                          "tdb_open: can't open tdb %s write-only\n", name);
 262                 errno = EINVAL;
 263                 goto fail;
 264         }
 265
 266         if ((open_flags & O_ACCMODE) == O_RDONLY) {
 267                 tdb->read_only = true;
 268                 /* read only databases don't do locking */
 269                 tdb->flags |= TDB_NOLOCK;
 270         } else
 271                 tdb->read_only = false;
 272
 273         /* internal databases don't need any of the rest. */
 274         if (tdb->flags & TDB_INTERNAL) {
 275                 tdb->flags |= (TDB_NOLOCK | TDB_NOMMAP);
 276                 if (tdb_new_database(tdb) != 0) {
 277                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 278                                  "tdb_open: tdb_new_database failed!");
 279                         goto fail;
 280                 }
 281                 TEST_IT(tdb->flags & TDB_CONVERT);
 282                 tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 283                 /* Zones don't matter for internal db. */
 284                 tdb->last_zone = 0;
 285                 return tdb;
 286         }
 287
 288         if ((tdb->fd = open(name, open_flags, mode)) == -1) {
 289                 tdb->log(tdb, TDB_DEBUG_WARNING, tdb->log_priv,
 290                          "tdb_open: could not open file %s: %s\n",
 291                          name, strerror(errno));
 292                 goto fail;      /* errno set by open(2) */
 293         }
 294
 295         /* on exec, don't inherit the fd */
 296         v = fcntl(tdb->fd, F_GETFD, 0);
 297         fcntl(tdb->fd, F_SETFD, v | FD_CLOEXEC);
 298
 299         /* ensure there is only one process initialising at once */
 300         if (tdb_lock_open(tdb) == -1) {
 301                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 302                          "tdb_open: failed to get open lock on %s: %s\n",
 303                          name, strerror(errno));
 304                 goto fail;      /* errno set by tdb_brlock */
 305         }
 306
 307         if (!tdb_pread_all(tdb->fd, &tdb->header, sizeof(tdb->header), 0)
 308             || strcmp(tdb->header.magic_food, TDB_MAGIC_FOOD) != 0) {
 309                 if (!(open_flags & O_CREAT) || tdb_new_database(tdb) == -1) {
 310                         if (errno == 0) {
 311                                 errno = EIO; /* ie bad format or something */
 312                         }
 313                         goto fail;
 314                 }
 315         } else if (tdb->header.version != TDB_VERSION) {
 316                 if (tdb->header.version == bswap_64(TDB_VERSION))
 317                         tdb->flags |= TDB_CONVERT;
 318                 else {
 319                         /* wrong version */
 320                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 321                                  "tdb_open: %s is unknown version 0x%llx\n",
 322                                  name, (long long)tdb->header.version);
 323                         errno = EIO;
 324                         goto fail;
 325                 }
 326         }
 327
 328         tdb_convert(tdb, &tdb->header, sizeof(tdb->header));
 329         hash_test = TDB_HASH_MAGIC;
 330         hash_test = tdb->khash(&hash_test, sizeof(hash_test),
 331                                tdb->header.hash_seed, tdb->hash_priv);
 332         if (tdb->header.hash_test != hash_test) {
 333                 /* wrong hash variant */
 334                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 335                          "tdb_open: %s uses a different hash function\n",
 336                          name);
 337                 errno = EIO;
 338                 goto fail;
 339         }
 340
 341         if (fstat(tdb->fd, &st) == -1)
 342                 goto fail;
 343
 344         /* Is it already in the open list?  If so, fail. */
 345         if (tdb_already_open(st.st_dev, st.st_ino)) {
 346                 /* FIXME */
 347                 tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 348                          "tdb_open: %s (%d,%d) is already open in this process\n",
 349                          name, (int)st.st_dev, (int)st.st_ino);
 350                 errno = EBUSY;
 351                 goto fail;
 352         }
 353
 354         tdb->name = strdup(name);
 355         if (!tdb->name) {
 356                 errno = ENOMEM;
 357                 goto fail;
 358         }
 359
 360         tdb->map_size = st.st_size;
 361         tdb->device = st.st_dev;
 362         tdb->inode = st.st_ino;
 363         tdb_mmap(tdb);
 364         tdb_unlock_open(tdb);
 365         tdb_zone_init(tdb);
 366
 367         tdb->next = tdbs;
 368         tdbs = tdb;
 369         return tdb;
 370
 371  fail:
 372         save_errno = errno;
 373
 374         if (!tdb)
 375                 return NULL;
 376
 377 #ifdef TDB_TRACE
 378         close(tdb->tracefd);
 379 #endif
 380         if (tdb->map_ptr) {
 381                 if (tdb->flags & TDB_INTERNAL) {
 382                         free(tdb->map_ptr);
 383                 } else
 384                         tdb_munmap(tdb);
 385         }
 386         free((char *)tdb->name);
 387         if (tdb->fd != -1)
 388                 if (close(tdb->fd) != 0)
 389                         tdb->log(tdb, TDB_DEBUG_ERROR, tdb->log_priv,
 390                                  "tdb_open: failed to close tdb->fd"
 391                                  " on error!\n");
 392         free(tdb);
 393         errno = save_errno;
 394         return NULL;
 395 }
 396
 397 static tdb_off_t hash_off(struct tdb_context *tdb, uint64_t list)
 398 {
 399         return tdb->header.v.hash_off
 400                 + ((list & ((1ULL << tdb->header.v.hash_bits) - 1))
 401                    * sizeof(tdb_off_t));
 402 }
 403
 404 /* Returns 0 if the entry is a zero (definitely not a match).
 405  * Returns a valid entry offset if it's a match.  Fills in rec.
 406  * Otherwise returns TDB_OFF_ERR: keep searching. */
 407 static tdb_off_t entry_matches(struct tdb_context *tdb,
 408                                uint64_t list,
 409                                uint64_t hash,
 410                                const struct tdb_data *key,
 411                                struct tdb_used_record *rec)
 412 {
 413         tdb_off_t off;
 414         uint64_t keylen;
 415         const unsigned char *rkey;
 416
 417         list &= ((1ULL << tdb->header.v.hash_bits) - 1);
 418
 419         off = tdb_read_off(tdb, tdb->header.v.hash_off
 420                            + list * sizeof(tdb_off_t));
 421         if (off == 0 || off == TDB_OFF_ERR)
 422                 return off;
 423
 424 #if 0 /* FIXME: Check other bits. */
 425         unsigned int bits, bitmask, hoffextra;
 426         /* Bottom three bits show how many extra hash bits. */
 427         bits = (off & ((1 << TDB_EXTRA_HASHBITS_NUM) - 1)) + 1;
 428         bitmask = (1 << bits)-1;
 429         hoffextra = ((off >> TDB_EXTRA_HASHBITS_NUM) & bitmask);
 430         uint64_t hextra = hash >> tdb->header.v.hash_bits;
 431         if ((hextra & bitmask) != hoffextra)
 432                 return TDB_OFF_ERR;
 433         off &= ~...;
 434 #endif
 435
 436         if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
 437                 return TDB_OFF_ERR;
 438
 439         /* FIXME: check extra bits in header! */
 440         keylen = rec_key_length(rec);
 441         if (keylen != key->dsize)
 442                 return TDB_OFF_ERR;
 443
 444         rkey = tdb_access_read(tdb, off + sizeof(*rec), keylen, false);
 445         if (!rkey)
 446                 return TDB_OFF_ERR;
 447         if (memcmp(rkey, key->dptr, keylen) != 0)
 448                 off = TDB_OFF_ERR;
 449         tdb_access_release(tdb, rkey);
 450         return off;
 451 }
 452
 453 /* FIXME: Optimize? */
 454 static void unlock_lists(struct tdb_context *tdb,
 455                          tdb_off_t list, tdb_len_t num,
 456                          int ltype)
 457 {
 458         tdb_off_t i;
 459
 460         for (i = list; i < list + num; i++)
 461                 tdb_unlock_list(tdb, i, ltype);
 462 }
 463
 464 /* FIXME: Optimize? */
 465 static int lock_lists(struct tdb_context *tdb,
 466                       tdb_off_t list, tdb_len_t num,
 467                       int ltype)
 468 {
 469         tdb_off_t i;
 470
 471         for (i = list; i < list + num; i++) {
 472                 if (tdb_lock_list(tdb, i, ltype, TDB_LOCK_WAIT)
 473                     == TDB_OFF_ERR) {
 474                         unlock_lists(tdb, list, i - list, ltype);
 475                         return -1;
 476                 }
 477         }
 478         return 0;
 479 }
 480
 481 /* We lock hashes up to the next empty offset.  We already hold the
 482  * lock on the start bucket, but we may need to release and re-grab
 483  * it.  If we fail, we hold no locks at all! */
 484 static tdb_len_t relock_hash_to_zero(struct tdb_context *tdb,
 485                                      tdb_off_t start, int ltype)
 486 {
 487         tdb_len_t num, len;
 488
 489 again:
 490         num = 1ULL << tdb->header.v.hash_bits;
 491         len = tdb_find_zero_off(tdb, hash_off(tdb, start), num - start);
 492         if (unlikely(len == num - start)) {
 493                 /* We hit the end of the hash range.  Drop lock: we have
 494                    to lock start of hash first. */
 495                 tdb_len_t pre_locks;
 496
 497                 tdb_unlock_list(tdb, start, ltype);
 498
 499                 /* Grab something, so header is stable. */
 500                 if (tdb_lock_list(tdb, 0, ltype, TDB_LOCK_WAIT))
 501                         return TDB_OFF_ERR;
 502                 pre_locks = tdb_find_zero_off(tdb, hash_off(tdb, 0), num);
 503                 /* We want to lock the zero entry as well. */
 504                 pre_locks++;
 505                 if (lock_lists(tdb, 1, pre_locks - 1, ltype) == -1) {
 506                         tdb_unlock_list(tdb, 0, ltype);
 507                         return TDB_OFF_ERR;
 508                 }
 509
 510                 /* Now lock later ones. */
 511                 if (unlikely(lock_lists(tdb, start, len, ltype) == -1)) {
 512                         unlock_lists(tdb, 0, pre_locks, ltype);
 513                         return TDB_OFF_ERR;
 514                 }
 515                 len += pre_locks;
 516         } else {
 517                 /* We want to lock the zero entry as well. */
 518                 len++;
 519                 /* But we already have lock on start. */
 520                 if (unlikely(lock_lists(tdb, start+1, len-1, ltype) == -1)) {
 521                         tdb_unlock_list(tdb, start, ltype);
 522                         return TDB_OFF_ERR;
 523                 }
 524         }
 525
 526         /* Now, did we lose the race, and it's not zero any more? */
 527         if (unlikely(tdb_read_off(tdb, hash_off(tdb, start + len - 1)) != 0)) {
 528                 /* Leave the start locked, as expected. */
 529                 unlock_lists(tdb, start + 1, len - 1, ltype);
 530                 goto again;
 531         }
 532
 533         return len;
 534 }
 535
 536 /* FIXME: modify, don't rewrite! */
 537 static int update_rec_hdr(struct tdb_context *tdb,
 538                           tdb_off_t off,
 539                           tdb_len_t keylen,
 540                           tdb_len_t datalen,
 541                           struct tdb_used_record *rec,
 542                           uint64_t h)
 543 {
 544         uint64_t dataroom = rec_data_length(rec) + rec_extra_padding(rec);
 545
 546         if (set_header(tdb, rec, keylen, datalen, keylen + dataroom, h))
 547                 return -1;
 548
 549         return tdb_write_convert(tdb, off, rec, sizeof(*rec));
 550 }
 551
 552 /* If we fail, others will try after us. */
 553 static void enlarge_hash(struct tdb_context *tdb)
 554 {
 555         tdb_off_t newoff, oldoff, i;
 556         tdb_len_t hlen;
 557         uint64_t h, num = 1ULL << tdb->header.v.hash_bits;
 558         struct tdb_used_record pad, *r;
 559
 560         /* FIXME: We should do this without holding locks throughout. */
 561         if (tdb_allrecord_lock(tdb, F_WRLCK, TDB_LOCK_WAIT, false) == -1)
 562                 return;
 563
 564         /* Someone else enlarged for us?  Nothing to do. */
 565         if ((1ULL << tdb->header.v.hash_bits) != num)
 566                 goto unlock;
 567
 568         /* Allocate our new array. */
 569         hlen = num * sizeof(tdb_off_t) * 2;
 570         newoff = alloc(tdb, 0, hlen, 0, false);
 571         if (unlikely(newoff == TDB_OFF_ERR))
 572                 goto unlock;
 573         if (unlikely(newoff == 0)) {
 574                 if (tdb_expand(tdb, 0, hlen, false) == -1)
 575                         goto unlock;
 576                 newoff = alloc(tdb, 0, hlen, 0, false);
 577                 if (newoff == TDB_OFF_ERR || newoff == 0)
 578                         goto unlock;
 579         }
 580         /* Step over record header! */
 581         newoff += sizeof(struct tdb_used_record);
 582
 583         /* Starts all zero. */
 584         if (zero_out(tdb, newoff, hlen) == -1)
 585                 goto unlock;
 586
 587         /* FIXME: If the space before is empty, we know this is in its ideal
 588          * location.  Or steal a bit from the pointer to avoid rehash. */
 589         for (i = tdb_find_nonzero_off(tdb, hash_off(tdb, 0), num);
 590              i < num;
 591              i += tdb_find_nonzero_off(tdb, hash_off(tdb, i), num - i)) {
 592                 tdb_off_t off;
 593                 off = tdb_read_off(tdb, hash_off(tdb, i));
 594                 if (unlikely(off == TDB_OFF_ERR))
 595                         goto unlock;
 596                 if (unlikely(!off)) {
 597                         tdb->ecode = TDB_ERR_CORRUPT;
 598                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 599                                  "enlarge_hash: zero hash bucket!\n");
 600                         goto unlock;
 601                 }
 602
 603                 /* Find next empty hash slot. */
 604                 for (h = hash_record(tdb, off);
 605                      tdb_read_off(tdb, newoff + (h & ((num * 2)-1))
 606                                   * sizeof(tdb_off_t)) != 0;
 607                      h++);
 608
 609                 /* FIXME: Encode extra hash bits! */
 610                 if (tdb_write_off(tdb, newoff + (h & ((num * 2)-1))
 611                                   * sizeof(tdb_off_t), off) == -1)
 612                         goto unlock;
 613                 i++;
 614         }
 615
 616         /* Free up old hash. */
 617         oldoff = tdb->header.v.hash_off - sizeof(*r);
 618         r = tdb_get(tdb, oldoff, &pad, sizeof(*r));
 619         if (!r)
 620                 goto unlock;
 621         add_free_record(tdb, oldoff,
 622                         sizeof(*r)+rec_data_length(r)+rec_extra_padding(r));
 623
 624         /* Now we write the modified header. */
 625         tdb->header.v.hash_bits++;
 626         tdb->header.v.hash_off = newoff;
 627         write_header(tdb);
 628 unlock:
 629         tdb_allrecord_unlock(tdb, F_WRLCK);
 630 }
 631
 632 /* This is the core routine which searches the hashtable for an entry.
 633  * On error, no locks are held and TDB_OFF_ERR is returned.
 634  * Otherwise, *num_locks locks of type ltype from *start_lock are held.
 635  * The bucket where the entry is (or would be) is in *bucket.
 636  * If not found, the return value is 0.
 637  * If found, the return value is the offset, and *rec is the record. */
 638 static tdb_off_t find_and_lock(struct tdb_context *tdb,
 639                                struct tdb_data key,
 640                                uint64_t h,
 641                                int ltype,
 642                                tdb_off_t *start_lock,
 643                                tdb_len_t *num_locks,
 644                                tdb_off_t *bucket,
 645                                struct tdb_used_record *rec)
 646 {
 647         tdb_off_t off;
 648
 649         /* FIXME: can we avoid locks for some fast paths? */
 650         *start_lock = tdb_lock_list(tdb, h, ltype, TDB_LOCK_WAIT);
 651         if (*start_lock == TDB_OFF_ERR)
 652                 return TDB_OFF_ERR;
 653
 654         /* Fast path. */
 655         off = entry_matches(tdb, *start_lock, h, &key, rec);
 656         if (likely(off != TDB_OFF_ERR)) {
 657                 *bucket = *start_lock;
 658                 *num_locks = 1;
 659                 return off;
 660         }
 661
 662         /* Slow path, need to grab more locks and search. */
 663
 664         /* Warning: this may drop the lock on *bucket! */
 665         *num_locks = relock_hash_to_zero(tdb, *start_lock, ltype);
 666         if (*num_locks == TDB_OFF_ERR)
 667                 return TDB_OFF_ERR;
 668
 669         for (*bucket = *start_lock;
 670              *bucket < *start_lock + *num_locks;
 671              (*bucket)++) {
 672                 off = entry_matches(tdb, *bucket, h, &key, rec);
 673                 /* Empty entry or we found it? */
 674                 if (off == 0 || off != TDB_OFF_ERR)
 675                         return off;
 676         }
 677
 678         /* We didn't find a zero entry?  Something went badly wrong... */
 679         unlock_lists(tdb, *start_lock, *start_lock + *num_locks, ltype);
 680         tdb->ecode = TDB_ERR_CORRUPT;
 681         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 682                  "find_and_lock: expected to find an empty hash bucket!\n");
 683         return TDB_OFF_ERR;
 684 }
 685
 686 int tdb_store(struct tdb_context *tdb,
 687               struct tdb_data key, struct tdb_data dbuf, int flag)
 688 {
 689         tdb_off_t new_off, off, bucket, start, num;
 690         struct tdb_used_record rec;
 691         uint64_t h;
 692         bool growing = false;
 693
 694         h = tdb_hash(tdb, key.dptr, key.dsize);
 695         off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
 696         if (unlikely(off == TDB_OFF_ERR))
 697                 return -1;
 698
 699         /* Now we have lock on this hash bucket. */
 700         if (flag == TDB_INSERT) {
 701                 if (off) {
 702                         tdb->ecode = TDB_ERR_EXISTS;
 703                         goto fail;
 704                 }
 705         } else {
 706                 if (off) {
 707                         if (rec_data_length(&rec) + rec_extra_padding(&rec)
 708                             >= dbuf.dsize) {
 709                                 new_off = off;
 710                                 if (update_rec_hdr(tdb, off,
 711                                                    key.dsize, dbuf.dsize,
 712                                                    &rec, h))
 713                                         goto fail;
 714                                 goto write;
 715                         }
 716                         /* FIXME: See if right record is free? */
 717                         /* Hint to allocator that we've realloced. */
 718                         growing = true;
 719                 } else {
 720                         if (flag == TDB_MODIFY) {
 721                                 /* if the record doesn't exist and we
 722                                    are in TDB_MODIFY mode then we should fail
 723                                    the store */
 724                                 tdb->ecode = TDB_ERR_NOEXIST;
 725                                 goto fail;
 726                         }
 727                 }
 728         }
 729
 730         /* Allocate a new record. */
 731         new_off = alloc(tdb, key.dsize, dbuf.dsize, h, growing);
 732         if (new_off == 0) {
 733                 unlock_lists(tdb, start, num, F_WRLCK);
 734                 /* Expand, then try again... */
 735                 if (tdb_expand(tdb, key.dsize, dbuf.dsize, growing) == -1)
 736                         return -1;
 737                 return tdb_store(tdb, key, dbuf, flag);
 738         }
 739
 740         /* We didn't like the existing one: remove it. */
 741         if (off) {
 742                 add_free_record(tdb, off, sizeof(struct tdb_used_record)
 743                                 + rec_key_length(&rec)
 744                                 + rec_data_length(&rec)
 745                                 + rec_extra_padding(&rec));
 746         }
 747
 748         /* FIXME: Encode extra hash bits! */
 749         if (tdb_write_off(tdb, hash_off(tdb, bucket), new_off) == -1)
 750                 goto fail;
 751
 752 write:
 753         off = new_off + sizeof(struct tdb_used_record);
 754         if (tdb->methods->write(tdb, off, key.dptr, key.dsize) == -1)
 755                 goto fail;
 756         off += key.dsize;
 757         if (tdb->methods->write(tdb, off, dbuf.dptr, dbuf.dsize) == -1)
 758                 goto fail;
 759
 760         /* FIXME: tdb_increment_seqnum(tdb); */
 761         unlock_lists(tdb, start, num, F_WRLCK);
 762
 763         /* FIXME: by simple simulation, this approximated 60% full.
 764          * Check in real case! */
 765         if (unlikely(num > 4 * tdb->header.v.hash_bits - 30))
 766                 enlarge_hash(tdb);
 767
 768         return 0;
 769
 770 fail:
 771         unlock_lists(tdb, start, num, F_WRLCK);
 772         return -1;
 773 }
 774
 775 struct tdb_data tdb_fetch(struct tdb_context *tdb, struct tdb_data key)
 776 {
 777         tdb_off_t off, start, num, bucket;
 778         struct tdb_used_record rec;
 779         uint64_t h;
 780         struct tdb_data ret;
 781
 782         h = tdb_hash(tdb, key.dptr, key.dsize);
 783         off = find_and_lock(tdb, key, h, F_RDLCK, &start, &num, &bucket, &rec);
 784         if (unlikely(off == TDB_OFF_ERR))
 785                 return tdb_null;
 786
 787         if (!off) {
 788                 tdb->ecode = TDB_ERR_NOEXIST;
 789                 ret = tdb_null;
 790         } else {
 791                 ret.dsize = rec_data_length(&rec);
 792                 ret.dptr = tdb_alloc_read(tdb, off + sizeof(rec) + key.dsize,
 793                                           ret.dsize);
 794         }
 795
 796         unlock_lists(tdb, start, num, F_RDLCK);
 797         return ret;
 798 }
 799
 800 static int hash_add(struct tdb_context *tdb, uint64_t h, tdb_off_t off)
 801 {
 802         tdb_off_t i, hoff, len, num;
 803
 804         /* Look for next space. */
 805         i = (h & ((1ULL << tdb->header.v.hash_bits) - 1));
 806         len = (1ULL << tdb->header.v.hash_bits) - i;
 807         num = tdb_find_zero_off(tdb, hash_off(tdb, i), len);
 808
 809         if (unlikely(num == len)) {
 810                 /* We wrapped.  Look through start of hash table. */
 811                 hoff = hash_off(tdb, 0);
 812                 len = (1ULL << tdb->header.v.hash_bits);
 813                 num = tdb_find_zero_off(tdb, hoff, len);
 814                 if (i == len) {
 815                         tdb->ecode = TDB_ERR_CORRUPT;
 816                         tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
 817                                  "hash_add: full hash table!\n");
 818                         return -1;
 819                 }
 820         }
 821         /* FIXME: Encode extra hash bits! */
 822         return tdb_write_off(tdb, hash_off(tdb, i + num), off);
 823 }
 824
 825 int tdb_delete(struct tdb_context *tdb, struct tdb_data key)
 826 {
 827         tdb_off_t i, bucket, off, start, num;
 828         struct tdb_used_record rec;
 829         uint64_t h;
 830
 831         h = tdb_hash(tdb, key.dptr, key.dsize);
 832         off = find_and_lock(tdb, key, h, F_WRLCK, &start, &num, &bucket, &rec);
 833         if (unlikely(off == TDB_OFF_ERR))
 834                 return -1;
 835
 836         if (!off) {
 837                 unlock_lists(tdb, start, num, F_WRLCK);
 838                 tdb->ecode = TDB_ERR_NOEXIST;
 839                 return -1;
 840         }
 841
 842         /* This actually unlinks it. */
 843         if (tdb_write_off(tdb, hash_off(tdb, bucket), 0) == -1)
 844                 goto unlock_err;
 845
 846         /* Rehash anything following. */
 847         for (i = hash_off(tdb, bucket+1);
 848              i != hash_off(tdb, h + num - 1);
 849              i += sizeof(tdb_off_t)) {
 850                 tdb_off_t off2;
 851                 uint64_t h2;
 852
 853                 off2 = tdb_read_off(tdb, i);
 854                 if (unlikely(off2 == TDB_OFF_ERR))
 855                         goto unlock_err;
 856
 857                 /* This can happen if we raced. */
 858                 if (unlikely(off2 == 0))
 859                         break;
 860
 861                 /* Maybe use a bit to indicate it is in ideal place? */
 862                 h2 = hash_record(tdb, off2);
 863                 /* Is it happy where it is? */
 864                 if (hash_off(tdb, h2) == i)
 865                         continue;
 866
 867                 /* Remove it. */
 868                 if (tdb_write_off(tdb, i, 0) == -1)
 869                         goto unlock_err;
 870
 871                 /* Rehash it. */
 872                 if (hash_add(tdb, h2, off2) == -1)
 873                         goto unlock_err;
 874         }
 875
 876         /* Free the deleted entry. */
 877         if (add_free_record(tdb, off,
 878                             sizeof(struct tdb_used_record)
 879                             + rec_key_length(&rec)
 880                             + rec_data_length(&rec)
 881                             + rec_extra_padding(&rec)) != 0)
 882                 goto unlock_err;
 883
 884         unlock_lists(tdb, start, num, F_WRLCK);
 885         return 0;
 886
 887 unlock_err:
 888         unlock_lists(tdb, start, num, F_WRLCK);
 889         return -1;
 890 }
 891
 892 int tdb_close(struct tdb_context *tdb)
 893 {
 894         struct tdb_context **i;
 895         int ret = 0;
 896
 897         /* FIXME:
 898         if (tdb->transaction) {
 899                 tdb_transaction_cancel(tdb);
 900         }
 901         */
 902         tdb_trace(tdb, "tdb_close");
 903
 904         if (tdb->map_ptr) {
 905                 if (tdb->flags & TDB_INTERNAL)
 906                         free(tdb->map_ptr);
 907                 else
 908                         tdb_munmap(tdb);
 909         }
 910         free((char *)tdb->name);
 911         if (tdb->fd != -1) {
 912                 ret = close(tdb->fd);
 913                 tdb->fd = -1;
 914         }
 915         free(tdb->lockrecs);
 916
 917         /* Remove from contexts list */
 918         for (i = &tdbs; *i; i = &(*i)->next) {
 919                 if (*i == tdb) {
 920                         *i = tdb->next;
 921                         break;
 922                 }
 923         }
 924
 925 #ifdef TDB_TRACE
 926         close(tdb->tracefd);
 927 #endif
 928         free(tdb);
 929
 930         return ret;
 931 }
 932
 933 enum TDB_ERROR tdb_error(struct tdb_context *tdb)
 934 {
 935         return tdb->ecode;
 936 }