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

   1  /*
   2    Trivial Database 2: hash handling
   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 <assert.h>
  20 #include <ccan/hash/hash.h>
  21
  22 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
  23                              void *arg)
  24 {
  25         uint64_t ret;
  26         /* hash64_stable assumes lower bits are more important; they are a
  27          * slightly better hash.  We use the upper bits first, so swap them. */
  28         ret = hash64_stable((const unsigned char *)key, length, seed);
  29         return (ret >> 32) | (ret << 32);
  30 }
  31
  32 void tdb_hash_init(struct tdb_context *tdb)
  33 {
  34         tdb->khash = jenkins_hash;
  35         tdb->hash_priv = NULL;
  36 }
  37
  38 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
  39 {
  40         return tdb->khash(ptr, len, tdb->hash_seed, tdb->hash_priv);
  41 }
  42
  43 uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off)
  44 {
  45         const struct tdb_used_record *r;
  46         const void *key;
  47         uint64_t klen, hash;
  48
  49         r = tdb_access_read(tdb, off, sizeof(*r), true);
  50         if (!r)
  51                 /* FIXME */
  52                 return 0;
  53
  54         klen = rec_key_length(r);
  55         tdb_access_release(tdb, r);
  56
  57         key = tdb_access_read(tdb, off + sizeof(*r), klen, false);
  58         if (!key)
  59                 return 0;
  60
  61         hash = tdb_hash(tdb, key, klen);
  62         tdb_access_release(tdb, key);
  63         return hash;
  64 }
  65
  66 /* Get bits from a value. */
  67 static uint32_t bits(uint64_t val, unsigned start, unsigned num)
  68 {
  69         assert(num <= 32);
  70         return (val >> start) & ((1U << num) - 1);
  71 }
  72
  73 /* We take bits from the top: that way we can lock whole sections of the hash
  74  * by using lock ranges. */
  75 static uint32_t use_bits(struct hash_info *h, unsigned num)
  76 {
  77         h->hash_used += num;
  78         return bits(h->h, 64 - h->hash_used, num);
  79 }
  80
  81 /* Does entry match? */
  82 static bool match(struct tdb_context *tdb,
  83                   struct hash_info *h,
  84                   const struct tdb_data *key,
  85                   tdb_off_t val,
  86                   struct tdb_used_record *rec)
  87 {
  88         bool ret = false;
  89         const unsigned char *rkey;
  90         tdb_off_t off;
  91
  92         add_stat(tdb, compares, 1);
  93         /* Desired bucket must match. */
  94         if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK)) {
  95                 add_stat(tdb, compare_wrong_bucket, 1);
  96                 return ret;
  97         }
  98
  99         /* Top bits of offset == next bits of hash. */
 100         if (bits(val, TDB_OFF_HASH_EXTRA_BIT, TDB_OFF_UPPER_STEAL_EXTRA)
 101             != bits(h->h, 64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
 102                     TDB_OFF_UPPER_STEAL_EXTRA)) {
 103                 add_stat(tdb, compare_wrong_offsetbits, 1);
 104                 return ret;
 105         }
 106
 107         off = val & TDB_OFF_MASK;
 108         if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
 109                 return ret;
 110
 111         if (rec_key_length(rec) != key->dsize) {
 112                 add_stat(tdb, compare_wrong_keylen, 1);
 113                 return ret;
 114         }
 115
 116         if ((h->h & ((1 << 11)-1)) != rec_hash(rec)) {
 117                 add_stat(tdb, compare_wrong_rechash, 1);
 118                 return false;
 119         }
 120
 121         rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false);
 122         if (!rkey)
 123                 return ret;
 124         if (memcmp(rkey, key->dptr, key->dsize) == 0)
 125                 ret = true;
 126         else
 127                 add_stat(tdb, compare_wrong_keycmp, 1);
 128         tdb_access_release(tdb, rkey);
 129         return ret;
 130 }
 131
 132 static tdb_off_t hbucket_off(tdb_off_t group_start, unsigned bucket)
 133 {
 134         return group_start
 135                 + (bucket % (1 << TDB_HASH_GROUP_BITS)) * sizeof(tdb_off_t);
 136 }
 137
 138 bool is_subhash(tdb_off_t val)
 139 {
 140         return (val >> TDB_OFF_UPPER_STEAL_SUBHASH_BIT) & 1;
 141 }
 142
 143 /* FIXME: Guess the depth, don't over-lock! */
 144 static tdb_off_t hlock_range(tdb_off_t group, tdb_off_t *size)
 145 {
 146         *size = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
 147         return group << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
 148 }
 149
 150 /* This is the core routine which searches the hashtable for an entry.
 151  * On error, no locks are held and TDB_OFF_ERR is returned.
 152  * Otherwise, hinfo is filled in (and the optional tinfo).
 153  * If not found, the return value is 0.
 154  * If found, the return value is the offset, and *rec is the record. */
 155 tdb_off_t find_and_lock(struct tdb_context *tdb,
 156                         struct tdb_data key,
 157                         int ltype,
 158                         struct hash_info *h,
 159                         struct tdb_used_record *rec,
 160                         struct traverse_info *tinfo)
 161 {
 162         uint32_t i, group;
 163         tdb_off_t hashtable;
 164
 165         h->h = tdb_hash(tdb, key.dptr, key.dsize);
 166         h->hash_used = 0;
 167         group = use_bits(h, TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
 168         h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 169
 170         h->hlock_start = hlock_range(group, &h->hlock_range);
 171         if (tdb_lock_hashes(tdb, h->hlock_start, h->hlock_range, ltype,
 172                             TDB_LOCK_WAIT))
 173                 return TDB_OFF_ERR;
 174
 175         hashtable = offsetof(struct tdb_header, hashtable);
 176         if (tinfo) {
 177                 tinfo->toplevel_group = group;
 178                 tinfo->num_levels = 1;
 179                 tinfo->levels[0].entry = 0;
 180                 tinfo->levels[0].hashtable = hashtable
 181                         + (group << TDB_HASH_GROUP_BITS) * sizeof(tdb_off_t);
 182                 tinfo->levels[0].total_buckets = 1 << TDB_HASH_GROUP_BITS;
 183         }
 184
 185         while (likely(h->hash_used < 64)) {
 186                 /* Read in the hash group. */
 187                 h->group_start = hashtable
 188                         + group * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 189
 190                 if (tdb_read_convert(tdb, h->group_start, &h->group,
 191                                      sizeof(h->group)) == -1)
 192                         goto fail;
 193
 194                 /* Pointer to another hash table?  Go down... */
 195                 if (is_subhash(h->group[h->home_bucket])) {
 196                         hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
 197                                 + sizeof(struct tdb_used_record);
 198                         if (tinfo) {
 199                                 /* When we come back, use *next* bucket */
 200                                 tinfo->levels[tinfo->num_levels-1].entry
 201                                         += h->home_bucket + 1;
 202                         }
 203                         group = use_bits(h, TDB_SUBLEVEL_HASH_BITS
 204                                          - TDB_HASH_GROUP_BITS);
 205                         h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 206                         if (tinfo) {
 207                                 tinfo->levels[tinfo->num_levels].hashtable
 208                                         = hashtable;
 209                                 tinfo->levels[tinfo->num_levels].total_buckets
 210                                         = 1 << TDB_SUBLEVEL_HASH_BITS;
 211                                 tinfo->levels[tinfo->num_levels].entry
 212                                         = group << TDB_HASH_GROUP_BITS;
 213                                 tinfo->num_levels++;
 214                         }
 215                         continue;
 216                 }
 217
 218                 /* It's in this group: search (until 0 or all searched) */
 219                 for (i = 0, h->found_bucket = h->home_bucket;
 220                      i < (1 << TDB_HASH_GROUP_BITS);
 221                      i++, h->found_bucket = ((h->found_bucket+1)
 222                                              % (1 << TDB_HASH_GROUP_BITS))) {
 223                         if (is_subhash(h->group[h->found_bucket]))
 224                                 continue;
 225
 226                         if (!h->group[h->found_bucket])
 227                                 break;
 228
 229                         if (match(tdb, h, &key, h->group[h->found_bucket],
 230                                   rec)) {
 231                                 if (tinfo) {
 232                                         tinfo->levels[tinfo->num_levels-1].entry
 233                                                 += h->found_bucket;
 234                                 }
 235                                 return h->group[h->found_bucket] & TDB_OFF_MASK;
 236                         }
 237                 }
 238                 /* Didn't find it: h indicates where it would go. */
 239                 return 0;
 240         }
 241
 242         /* FIXME: We hit the bottom.  Chain! */
 243         abort();
 244
 245 fail:
 246         tdb_unlock_hashes(tdb, h->hlock_start, h->hlock_range, ltype);
 247         return TDB_OFF_ERR;
 248 }
 249
 250 /* I wrote a simple test, expanding a hash to 2GB, for the following
 251  * cases:
 252  * 1) Expanding all the buckets at once,
 253  * 2) Expanding the bucket we wanted to place the new entry into.
 254  * 3) Expanding the most-populated bucket,
 255  *
 256  * I measured the worst/average/best density during this process.
 257  * 1) 3%/16%/30%
 258  * 2) 4%/20%/38%
 259  * 3) 6%/22%/41%
 260  *
 261  * So we figure out the busiest bucket for the moment.
 262  */
 263 static unsigned fullest_bucket(struct tdb_context *tdb,
 264                                const tdb_off_t *group,
 265                                unsigned new_bucket)
 266 {
 267         unsigned counts[1 << TDB_HASH_GROUP_BITS] = { 0 };
 268         unsigned int i, best_bucket;
 269
 270         /* Count the new entry. */
 271         counts[new_bucket]++;
 272         best_bucket = new_bucket;
 273
 274         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 275                 unsigned this_bucket;
 276
 277                 if (is_subhash(group[i]))
 278                         continue;
 279                 this_bucket = group[i] & TDB_OFF_HASH_GROUP_MASK;
 280                 if (++counts[this_bucket] > counts[best_bucket])
 281                         best_bucket = this_bucket;
 282         }
 283
 284         return best_bucket;
 285 }
 286
 287 static bool put_into_group(tdb_off_t *group,
 288                            unsigned bucket, tdb_off_t encoded)
 289 {
 290         unsigned int i;
 291
 292         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 293                 unsigned b = (bucket + i) % (1 << TDB_HASH_GROUP_BITS);
 294
 295                 if (group[b] == 0) {
 296                         group[b] = encoded;
 297                         return true;
 298                 }
 299         }
 300         return false;
 301 }
 302
 303 static void force_into_group(tdb_off_t *group,
 304                              unsigned bucket, tdb_off_t encoded)
 305 {
 306         if (!put_into_group(group, bucket, encoded))
 307                 abort();
 308 }
 309
 310 static tdb_off_t encode_offset(tdb_off_t new_off, struct hash_info *h)
 311 {
 312         return h->home_bucket
 313                 | new_off
 314                 | ((uint64_t)bits(h->h,
 315                                   64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
 316                                   TDB_OFF_UPPER_STEAL_EXTRA)
 317                    << TDB_OFF_HASH_EXTRA_BIT);
 318 }
 319
 320 /* Simply overwrite the hash entry we found before. */
 321 int replace_in_hash(struct tdb_context *tdb,
 322                     struct hash_info *h,
 323                     tdb_off_t new_off)
 324 {
 325         return tdb_write_off(tdb, hbucket_off(h->group_start, h->found_bucket),
 326                              encode_offset(new_off, h));
 327 }
 328
 329 /* Add into a newly created subhash. */
 330 static int add_to_subhash(struct tdb_context *tdb, tdb_off_t subhash,
 331                           unsigned hash_used, tdb_off_t val)
 332 {
 333         tdb_off_t off = (val & TDB_OFF_MASK), *group;
 334         struct hash_info h;
 335         unsigned int gnum;
 336
 337         h.hash_used = hash_used;
 338
 339         /* FIXME chain if hash_used == 64 */
 340         if (hash_used + TDB_SUBLEVEL_HASH_BITS > 64)
 341                 abort();
 342
 343         h.h = hash_record(tdb, off);
 344         gnum = use_bits(&h, TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS);
 345         h.group_start = subhash + sizeof(struct tdb_used_record)
 346                 + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 347         h.home_bucket = use_bits(&h, TDB_HASH_GROUP_BITS);
 348
 349         group = tdb_access_write(tdb, h.group_start,
 350                                  sizeof(*group) << TDB_HASH_GROUP_BITS, true);
 351         if (!group)
 352                 return -1;
 353         force_into_group(group, h.home_bucket, encode_offset(off, &h));
 354         return tdb_access_commit(tdb, group);
 355 }
 356
 357 static int expand_group(struct tdb_context *tdb, struct hash_info *h)
 358 {
 359         unsigned bucket, num_vals, i;
 360         tdb_off_t subhash;
 361         tdb_off_t vals[1 << TDB_HASH_GROUP_BITS];
 362
 363         /* Attach new empty subhash under fullest bucket. */
 364         bucket = fullest_bucket(tdb, h->group, h->home_bucket);
 365
 366         subhash = alloc(tdb, 0, sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS,
 367                         0, false);
 368         if (subhash == TDB_OFF_ERR)
 369                 return -1;
 370
 371         add_stat(tdb, alloc_subhash, 1);
 372         if (zero_out(tdb, subhash + sizeof(struct tdb_used_record),
 373                      sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS) == -1)
 374                 return -1;
 375
 376         /* Remove any which are destined for bucket or are in wrong place. */
 377         num_vals = 0;
 378         for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 379                 unsigned home_bucket = h->group[i] & TDB_OFF_HASH_GROUP_MASK;
 380                 if (!h->group[i] || is_subhash(h->group[i]))
 381                         continue;
 382                 if (home_bucket == bucket || home_bucket != i) {
 383                         vals[num_vals++] = h->group[i];
 384                         h->group[i] = 0;
 385                 }
 386         }
 387         /* FIXME: This assert is valid, but we do this during unit test :( */
 388         /* assert(num_vals); */
 389
 390         /* Overwrite expanded bucket with subhash pointer. */
 391         h->group[bucket] = subhash | (1ULL << TDB_OFF_UPPER_STEAL_SUBHASH_BIT);
 392
 393         /* Put values back. */
 394         for (i = 0; i < num_vals; i++) {
 395                 unsigned this_bucket = vals[i] & TDB_OFF_HASH_GROUP_MASK;
 396
 397                 if (this_bucket == bucket) {
 398                         if (add_to_subhash(tdb, subhash, h->hash_used, vals[i]))
 399                                 return -1;
 400                 } else {
 401                         /* There should be room to put this back. */
 402                         force_into_group(h->group, this_bucket, vals[i]);
 403                 }
 404         }
 405         return 0;
 406 }
 407
 408 int delete_from_hash(struct tdb_context *tdb, struct hash_info *h)
 409 {
 410         unsigned int i, num_movers = 0;
 411         tdb_off_t movers[1 << TDB_HASH_GROUP_BITS];
 412
 413         h->group[h->found_bucket] = 0;
 414         for (i = 1; i < (1 << TDB_HASH_GROUP_BITS); i++) {
 415                 unsigned this_bucket;
 416
 417                 this_bucket = (h->found_bucket+i) % (1 << TDB_HASH_GROUP_BITS);
 418                 /* Empty bucket?  We're done. */
 419                 if (!h->group[this_bucket])
 420                         break;
 421
 422                 /* Ignore subhashes. */
 423                 if (is_subhash(h->group[this_bucket]))
 424                         continue;
 425
 426                 /* If this one is not happy where it is, we'll move it. */
 427                 if ((h->group[this_bucket] & TDB_OFF_HASH_GROUP_MASK)
 428                     != this_bucket) {
 429                         movers[num_movers++] = h->group[this_bucket];
 430                         h->group[this_bucket] = 0;
 431                 }
 432         }
 433
 434         /* Put back the ones we erased. */
 435         for (i = 0; i < num_movers; i++) {
 436                 force_into_group(h->group, movers[i] & TDB_OFF_HASH_GROUP_MASK,
 437                                  movers[i]);
 438         }
 439
 440         /* Now we write back the hash group */
 441         return tdb_write_convert(tdb, h->group_start,
 442                                  h->group, sizeof(h->group));
 443 }
 444
 445 int add_to_hash(struct tdb_context *tdb, struct hash_info *h, tdb_off_t new_off)
 446 {
 447         /* FIXME: chain! */
 448         if (h->hash_used >= 64)
 449                 abort();
 450
 451         /* We hit an empty bucket during search?  That's where it goes. */
 452         if (!h->group[h->found_bucket]) {
 453                 h->group[h->found_bucket] = encode_offset(new_off, h);
 454                 /* Write back the modified group. */
 455                 return tdb_write_convert(tdb, h->group_start,
 456                                          h->group, sizeof(h->group));
 457         }
 458
 459         /* We're full.  Expand. */
 460         if (expand_group(tdb, h) == -1)
 461                 return -1;
 462
 463         if (is_subhash(h->group[h->home_bucket])) {
 464                 /* We were expanded! */
 465                 tdb_off_t hashtable;
 466                 unsigned int gnum;
 467
 468                 /* Write back the modified group. */
 469                 if (tdb_write_convert(tdb, h->group_start, h->group,
 470                                       sizeof(h->group)))
 471                         return -1;
 472
 473                 /* Move hashinfo down a level. */
 474                 hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
 475                         + sizeof(struct tdb_used_record);
 476                 gnum = use_bits(h,TDB_SUBLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
 477                 h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
 478                 h->group_start = hashtable
 479                         + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 480                 if (tdb_read_convert(tdb, h->group_start, &h->group,
 481                                      sizeof(h->group)) == -1)
 482                         return -1;
 483         }
 484
 485         /* Expanding the group must have made room if it didn't choose this
 486          * bucket. */
 487         if (put_into_group(h->group, h->home_bucket, encode_offset(new_off, h)))
 488                 return tdb_write_convert(tdb, h->group_start,
 489                                          h->group, sizeof(h->group));
 490
 491         /* This can happen if all hashes in group (and us) dropped into same
 492          * group in subhash. */
 493         return add_to_hash(tdb, h, new_off);
 494 }
 495
 496 /* Traverse support: returns offset of record, or 0 or TDB_OFF_ERR. */
 497 static tdb_off_t iterate_hash(struct tdb_context *tdb,
 498                               struct traverse_info *tinfo)
 499 {
 500         tdb_off_t off, val;
 501         unsigned int i;
 502         struct traverse_level *tlevel;
 503
 504         tlevel = &tinfo->levels[tinfo->num_levels-1];
 505
 506 again:
 507         for (i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
 508                                       tlevel->entry, tlevel->total_buckets);
 509              i != tlevel->total_buckets;
 510              i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
 511                                       i+1, tlevel->total_buckets)) {
 512                 val = tdb_read_off(tdb, tlevel->hashtable+sizeof(tdb_off_t)*i);
 513                 if (unlikely(val == TDB_OFF_ERR))
 514                         return TDB_OFF_ERR;
 515
 516                 off = val & TDB_OFF_MASK;
 517
 518                 /* This makes the delete-all-in-traverse case work
 519                  * (and simplifies our logic a little). */
 520                 if (off == tinfo->prev)
 521                         continue;
 522
 523                 tlevel->entry = i;
 524
 525                 if (!is_subhash(val)) {
 526                         /* Found one. */
 527                         tinfo->prev = off;
 528                         return off;
 529                 }
 530
 531                 /* When we come back, we want the next one */
 532                 tlevel->entry++;
 533                 tinfo->num_levels++;
 534                 tlevel++;
 535                 tlevel->hashtable = off + sizeof(struct tdb_used_record);
 536                 tlevel->entry = 0;
 537                 tlevel->total_buckets = (1 << TDB_SUBLEVEL_HASH_BITS);
 538                 goto again;
 539         }
 540
 541         /* Nothing there? */
 542         if (tinfo->num_levels == 1)
 543                 return 0;
 544
 545         /* Go back up and keep searching. */
 546         tinfo->num_levels--;
 547         tlevel--;
 548         goto again;
 549 }
 550
 551 /* Return 1 if we find something, 0 if not, -1 on error. */
 552 int next_in_hash(struct tdb_context *tdb, int ltype,
 553                  struct traverse_info *tinfo,
 554                  TDB_DATA *kbuf, size_t *dlen)
 555 {
 556         const unsigned group_bits = TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS;
 557         tdb_off_t hlock_start, hlock_range, off;
 558
 559         while (tinfo->toplevel_group < (1 << group_bits)) {
 560                 hlock_start = (tdb_off_t)tinfo->toplevel_group
 561                         << (64 - group_bits);
 562                 hlock_range = 1ULL << group_bits;
 563                 if (tdb_lock_hashes(tdb, hlock_start, hlock_range, ltype,
 564                                     TDB_LOCK_WAIT) != 0)
 565                         return -1;
 566
 567                 off = iterate_hash(tdb, tinfo);
 568                 if (off) {
 569                         struct tdb_used_record rec;
 570
 571                         if (tdb_read_convert(tdb, off, &rec, sizeof(rec))) {
 572                                 tdb_unlock_hashes(tdb,
 573                                                   hlock_start, hlock_range,
 574                                                   ltype);
 575                                 return -1;
 576                         }
 577                         if (rec_magic(&rec) != TDB_MAGIC) {
 578                                 tdb_logerr(tdb, TDB_ERR_CORRUPT,
 579                                            TDB_DEBUG_FATAL,
 580                                            "next_in_hash:"
 581                                            " corrupt record at %llu",
 582                                            (long long)off);
 583                                 return -1;
 584                         }
 585
 586                         kbuf->dsize = rec_key_length(&rec);
 587
 588                         /* They want data as well? */
 589                         if (dlen) {
 590                                 *dlen = rec_data_length(&rec);
 591                                 kbuf->dptr = tdb_alloc_read(tdb,
 592                                                             off + sizeof(rec),
 593                                                             kbuf->dsize
 594                                                             + *dlen);
 595                         } else {
 596                                 kbuf->dptr = tdb_alloc_read(tdb,
 597                                                             off + sizeof(rec),
 598                                                             kbuf->dsize);
 599                         }
 600                         tdb_unlock_hashes(tdb, hlock_start, hlock_range, ltype);
 601                         return kbuf->dptr ? 1 : -1;
 602                 }
 603
 604                 tdb_unlock_hashes(tdb, hlock_start, hlock_range, ltype);
 605
 606                 tinfo->toplevel_group++;
 607                 tinfo->levels[0].hashtable
 608                         += (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
 609                 tinfo->levels[0].entry = 0;
 610         }
 611         return 0;
 612 }
 613
 614 /* Return 1 if we find something, 0 if not, -1 on error. */
 615 int first_in_hash(struct tdb_context *tdb, int ltype,
 616                   struct traverse_info *tinfo,
 617                   TDB_DATA *kbuf, size_t *dlen)
 618 {
 619         tinfo->prev = 0;
 620         tinfo->toplevel_group = 0;
 621         tinfo->num_levels = 1;
 622         tinfo->levels[0].hashtable = offsetof(struct tdb_header, hashtable);
 623         tinfo->levels[0].entry = 0;
 624         tinfo->levels[0].total_buckets = (1 << TDB_HASH_GROUP_BITS);
 625
 626         return next_in_hash(tdb, ltype, tinfo, kbuf, dlen);
 627 }
 628
 629 /* Even if the entry isn't in this hash bucket, you'd have to lock this
 630  * bucket to find it. */
 631 static int chainlock(struct tdb_context *tdb, const TDB_DATA *key,
 632                      int ltype, enum tdb_lock_flags waitflag,
 633                      const char *func)
 634 {
 635         int ret;
 636         uint64_t h = tdb_hash(tdb, key->dptr, key->dsize);
 637         tdb_off_t lockstart, locksize;
 638         unsigned int group, gbits;
 639
 640         gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
 641         group = bits(h, 64 - gbits, gbits);
 642
 643         lockstart = hlock_range(group, &locksize);
 644
 645         ret = tdb_lock_hashes(tdb, lockstart, locksize, ltype, waitflag);
 646         tdb_trace_1rec(tdb, func, *key);
 647         return ret;
 648 }
 649
 650 /* lock/unlock one hash chain. This is meant to be used to reduce
 651    contention - it cannot guarantee how many records will be locked */
 652 int tdb_chainlock(struct tdb_context *tdb, TDB_DATA key)
 653 {
 654         return chainlock(tdb, &key, F_WRLCK, TDB_LOCK_WAIT, "tdb_chainlock");
 655 }
 656
 657 int tdb_chainunlock(struct tdb_context *tdb, TDB_DATA key)
 658 {
 659         uint64_t h = tdb_hash(tdb, key.dptr, key.dsize);
 660         tdb_off_t lockstart, locksize;
 661         unsigned int group, gbits;
 662
 663         gbits = TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS;
 664         group = bits(h, 64 - gbits, gbits);
 665
 666         lockstart = hlock_range(group, &locksize);
 667
 668         tdb_trace_1rec(tdb, "tdb_chainunlock", key);
 669         return tdb_unlock_hashes(tdb, lockstart, locksize, F_WRLCK);
 670 }