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