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