2 Trivial Database 2: hash handling
3 Copyright (C) Rusty Russell 2010
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.
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.
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/>.
20 #include <ccan/hash/hash.h>
22 static uint64_t jenkins_hash(const void *key, size_t length, uint64_t seed,
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);
32 void tdb_hash_init(struct tdb_context *tdb)
34 tdb->khash = jenkins_hash;
35 tdb->hash_priv = NULL;
38 uint64_t tdb_hash(struct tdb_context *tdb, const void *ptr, size_t len)
40 return tdb->khash(ptr, len, tdb->hash_seed, tdb->hash_priv);
43 uint64_t hash_record(struct tdb_context *tdb, tdb_off_t off)
45 struct tdb_used_record pad, *r;
49 r = tdb_get(tdb, off, &pad, sizeof(pad));
54 klen = rec_key_length(r);
55 key = tdb_access_read(tdb, off + sizeof(pad), klen, false);
59 hash = tdb_hash(tdb, key, klen);
60 tdb_access_release(tdb, key);
64 /* Get bits from a value. */
65 static uint32_t bits(uint64_t val, unsigned start, unsigned num)
68 return (val >> start) & ((1U << num) - 1);
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)
76 return bits(h->h, 64 - h->hash_used, num);
79 /* Does entry match? */
80 static bool match(struct tdb_context *tdb,
82 const struct tdb_data *key,
84 struct tdb_used_record *rec)
87 const unsigned char *rkey;
90 /* FIXME: Handle hash value truncated. */
91 if (bits(val, TDB_OFF_HASH_TRUNCATED_BIT, 1))
94 /* Desired bucket must match. */
95 if (h->home_bucket != (val & TDB_OFF_HASH_GROUP_MASK))
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))
104 off = val & TDB_OFF_MASK;
105 if (tdb_read_convert(tdb, off, rec, sizeof(*rec)) == -1)
108 /* FIXME: check extra bits in header? */
109 if (rec_key_length(rec) != key->dsize)
112 rkey = tdb_access_read(tdb, off + sizeof(*rec), key->dsize, false);
115 ret = (memcmp(rkey, key->dptr, key->dsize) == 0);
116 tdb_access_release(tdb, rkey);
120 static tdb_off_t hbucket_off(tdb_off_t group_start, unsigned bucket)
123 + (bucket % (1 << TDB_HASH_GROUP_BITS)) * sizeof(tdb_off_t);
126 /* Truncated hashes can't be all 1: that's how we spot a sub-hash */
127 bool is_subhash(tdb_off_t val)
129 return val >> (64-TDB_OFF_UPPER_STEAL) == (1<<TDB_OFF_UPPER_STEAL) - 1;
132 /* This is the core routine which searches the hashtable for an entry.
133 * On error, no locks are held and TDB_OFF_ERR is returned.
134 * Otherwise, hinfo is filled in (and the optional tinfo).
135 * If not found, the return value is 0.
136 * If found, the return value is the offset, and *rec is the record. */
137 tdb_off_t find_and_lock(struct tdb_context *tdb,
141 struct tdb_used_record *rec,
142 struct traverse_info *tinfo)
147 h->h = tdb_hash(tdb, key.dptr, key.dsize);
149 group = use_bits(h, TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
150 h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
152 /* FIXME: Guess the depth, don't over-lock! */
153 h->hlock_start = (tdb_off_t)group
154 << (64 - (TDB_TOPLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS));
155 h->hlock_range = 1ULL << (64 - (TDB_TOPLEVEL_HASH_BITS
156 - TDB_HASH_GROUP_BITS));
157 if (tdb_lock_hashes(tdb, h->hlock_start, h->hlock_range, ltype,
161 hashtable = offsetof(struct tdb_header, hashtable);
163 tinfo->toplevel_group = group;
164 tinfo->num_levels = 1;
165 tinfo->levels[0].entry = 0;
166 tinfo->levels[0].hashtable = hashtable
167 + (group << TDB_HASH_GROUP_BITS) * sizeof(tdb_off_t);
168 tinfo->levels[0].total_buckets = 1 << TDB_HASH_GROUP_BITS;
171 while (likely(h->hash_used < 64)) {
172 /* Read in the hash group. */
173 h->group_start = hashtable
174 + group * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
176 if (tdb_read_convert(tdb, h->group_start, &h->group,
177 sizeof(h->group)) == -1)
180 /* Pointer to another hash table? Go down... */
181 if (is_subhash(h->group[h->home_bucket])) {
182 hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
183 + sizeof(struct tdb_used_record);
185 /* When we come back, use *next* bucket */
186 tinfo->levels[tinfo->num_levels-1].entry
187 += h->home_bucket + 1;
189 group = use_bits(h, TDB_SUBLEVEL_HASH_BITS
190 - TDB_HASH_GROUP_BITS);
191 h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
193 tinfo->levels[tinfo->num_levels].hashtable
195 tinfo->levels[tinfo->num_levels].total_buckets
196 = 1 << TDB_SUBLEVEL_HASH_BITS;
197 tinfo->levels[tinfo->num_levels].entry
198 = group << TDB_HASH_GROUP_BITS;
204 /* It's in this group: search (until 0 or all searched) */
205 for (i = 0, h->found_bucket = h->home_bucket;
206 i < (1 << TDB_HASH_GROUP_BITS);
207 i++, h->found_bucket = ((h->found_bucket+1)
208 % (1 << TDB_HASH_GROUP_BITS))) {
209 if (is_subhash(h->group[h->found_bucket]))
212 if (!h->group[h->found_bucket])
215 if (match(tdb, h, &key, h->group[h->found_bucket],
218 tinfo->levels[tinfo->num_levels-1].entry
221 return h->group[h->found_bucket] & TDB_OFF_MASK;
224 /* Didn't find it: h indicates where it would go. */
228 /* FIXME: We hit the bottom. Chain! */
232 tdb_unlock_hashes(tdb, h->hlock_start, h->hlock_range, ltype);
236 /* I wrote a simple test, expanding a hash to 2GB, for the following
238 * 1) Expanding all the buckets at once,
239 * 2) Expanding the most-populated bucket,
240 * 3) Expanding the bucket we wanted to place the new entry ito.
242 * I measured the worst/average/best density during this process.
247 * So we figure out the busiest bucket for the moment.
249 static unsigned fullest_bucket(struct tdb_context *tdb,
250 const tdb_off_t *group,
253 unsigned counts[1 << TDB_HASH_GROUP_BITS] = { 0 };
254 unsigned int i, best_bucket;
256 /* Count the new entry. */
257 counts[new_bucket]++;
258 best_bucket = new_bucket;
260 for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
261 unsigned this_bucket;
263 if (is_subhash(group[i]))
265 this_bucket = group[i] & TDB_OFF_HASH_GROUP_MASK;
266 if (++counts[this_bucket] > counts[best_bucket])
267 best_bucket = this_bucket;
273 static bool put_into_group(tdb_off_t *group,
274 unsigned bucket, tdb_off_t encoded)
278 for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
279 unsigned b = (bucket + i) % (1 << TDB_HASH_GROUP_BITS);
289 static void force_into_group(tdb_off_t *group,
290 unsigned bucket, tdb_off_t encoded)
292 if (!put_into_group(group, bucket, encoded))
296 static tdb_off_t encode_offset(tdb_off_t new_off, struct hash_info *h)
298 return h->home_bucket
300 | ((uint64_t)bits(h->h,
301 64 - h->hash_used - TDB_OFF_UPPER_STEAL_EXTRA,
302 TDB_OFF_UPPER_STEAL_EXTRA)
303 << TDB_OFF_HASH_EXTRA_BIT);
306 /* Simply overwrite the hash entry we found before. */
307 int replace_in_hash(struct tdb_context *tdb,
311 return tdb_write_off(tdb, hbucket_off(h->group_start, h->found_bucket),
312 encode_offset(new_off, h));
315 /* Add into a newly created subhash. */
316 static int add_to_subhash(struct tdb_context *tdb, tdb_off_t subhash,
317 unsigned hash_used, tdb_off_t val)
319 tdb_off_t off = (val & TDB_OFF_MASK), *group;
323 h.hash_used = hash_used;
325 /* FIXME chain if hash_used == 64 */
326 if (hash_used + TDB_SUBLEVEL_HASH_BITS > 64)
329 /* FIXME: Do truncated hash bits if we can! */
330 h.h = hash_record(tdb, off);
331 gnum = use_bits(&h, TDB_SUBLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS);
332 h.group_start = subhash + sizeof(struct tdb_used_record)
333 + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
334 h.home_bucket = use_bits(&h, TDB_HASH_GROUP_BITS);
336 group = tdb_access_write(tdb, h.group_start,
337 sizeof(*group) << TDB_HASH_GROUP_BITS, true);
340 force_into_group(group, h.home_bucket, encode_offset(off, &h));
341 return tdb_access_commit(tdb, group);
344 static int expand_group(struct tdb_context *tdb, struct hash_info *h)
346 unsigned bucket, num_vals, i;
348 tdb_off_t vals[1 << TDB_HASH_GROUP_BITS];
350 /* Attach new empty subhash under fullest bucket. */
351 bucket = fullest_bucket(tdb, h->group, h->home_bucket);
353 subhash = alloc(tdb, 0, sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS,
355 if (subhash == TDB_OFF_ERR)
358 if (zero_out(tdb, subhash + sizeof(struct tdb_used_record),
359 sizeof(tdb_off_t) << TDB_SUBLEVEL_HASH_BITS) == -1)
362 /* Remove any which are destined for bucket or are in wrong place. */
364 for (i = 0; i < (1 << TDB_HASH_GROUP_BITS); i++) {
365 unsigned home_bucket = h->group[i] & TDB_OFF_HASH_GROUP_MASK;
366 if (!h->group[i] || is_subhash(h->group[i]))
368 if (home_bucket == bucket || home_bucket != i) {
369 vals[num_vals++] = h->group[i];
373 /* FIXME: This assert is valid, but we do this during unit test :( */
374 /* assert(num_vals); */
376 /* Overwrite expanded bucket with subhash pointer. */
377 h->group[bucket] = subhash | ~((1ULL << (64 - TDB_OFF_UPPER_STEAL))-1);
379 /* Put values back. */
380 for (i = 0; i < num_vals; i++) {
381 unsigned this_bucket = vals[i] & TDB_OFF_HASH_GROUP_MASK;
383 if (this_bucket == bucket) {
384 if (add_to_subhash(tdb, subhash, h->hash_used, vals[i]))
387 /* There should be room to put this back. */
388 force_into_group(h->group, this_bucket, vals[i]);
394 int delete_from_hash(struct tdb_context *tdb, struct hash_info *h)
396 unsigned int i, num_movers = 0;
397 tdb_off_t movers[1 << TDB_HASH_GROUP_BITS];
399 h->group[h->found_bucket] = 0;
400 for (i = 1; i < (1 << TDB_HASH_GROUP_BITS); i++) {
401 unsigned this_bucket;
403 this_bucket = (h->found_bucket+i) % (1 << TDB_HASH_GROUP_BITS);
404 /* Empty bucket? We're done. */
405 if (!h->group[this_bucket])
408 /* Ignore subhashes. */
409 if (is_subhash(h->group[this_bucket]))
412 /* If this one is not happy where it is, we'll move it. */
413 if ((h->group[this_bucket] & TDB_OFF_HASH_GROUP_MASK)
415 movers[num_movers++] = h->group[this_bucket];
416 h->group[this_bucket] = 0;
420 /* Put back the ones we erased. */
421 for (i = 0; i < num_movers; i++) {
422 force_into_group(h->group, movers[i] & TDB_OFF_HASH_GROUP_MASK,
426 /* Now we write back the hash group */
427 return tdb_write_convert(tdb, h->group_start,
428 h->group, sizeof(h->group));
431 int add_to_hash(struct tdb_context *tdb, struct hash_info *h, tdb_off_t new_off)
434 if (h->hash_used >= 64)
437 /* We hit an empty bucket during search? That's where it goes. */
438 if (!h->group[h->found_bucket]) {
439 h->group[h->found_bucket] = encode_offset(new_off, h);
440 /* Write back the modified group. */
441 return tdb_write_convert(tdb, h->group_start,
442 h->group, sizeof(h->group));
445 /* We're full. Expand. */
446 if (expand_group(tdb, h) == -1)
449 if (is_subhash(h->group[h->home_bucket])) {
450 /* We were expanded! */
454 /* Write back the modified group. */
455 if (tdb_write_convert(tdb, h->group_start, h->group,
459 /* Move hashinfo down a level. */
460 hashtable = (h->group[h->home_bucket] & TDB_OFF_MASK)
461 + sizeof(struct tdb_used_record);
462 gnum = use_bits(h,TDB_SUBLEVEL_HASH_BITS - TDB_HASH_GROUP_BITS);
463 h->home_bucket = use_bits(h, TDB_HASH_GROUP_BITS);
464 h->group_start = hashtable
465 + gnum * (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
466 if (tdb_read_convert(tdb, h->group_start, &h->group,
467 sizeof(h->group)) == -1)
471 /* Expanding the group must have made room if it didn't choose this
473 if (put_into_group(h->group, h->home_bucket, encode_offset(new_off, h)))
474 return tdb_write_convert(tdb, h->group_start,
475 h->group, sizeof(h->group));
477 /* This can happen if all hashes in group (and us) dropped into same
478 * group in subhash. */
479 return add_to_hash(tdb, h, new_off);
482 /* Traverse support: returns offset of record, or 0 or TDB_OFF_ERR. */
483 static tdb_off_t iterate_hash(struct tdb_context *tdb,
484 struct traverse_info *tinfo)
488 struct traverse_level *tlevel;
490 tlevel = &tinfo->levels[tinfo->num_levels-1];
493 for (i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
494 tlevel->entry, tlevel->total_buckets);
495 i != tlevel->total_buckets;
496 i = tdb_find_nonzero_off(tdb, tlevel->hashtable,
497 i+1, tlevel->total_buckets)) {
498 val = tdb_read_off(tdb, tlevel->hashtable+sizeof(tdb_off_t)*i);
499 if (unlikely(val == TDB_OFF_ERR))
502 off = val & TDB_OFF_MASK;
504 /* This makes the delete-all-in-traverse case work
505 * (and simplifies our logic a little). */
506 if (off == tinfo->prev)
511 if (!is_subhash(val)) {
517 /* When we come back, we want the next one */
521 tlevel->hashtable = off + sizeof(struct tdb_used_record);
523 tlevel->total_buckets = (1 << TDB_SUBLEVEL_HASH_BITS);
528 if (tinfo->num_levels == 1)
531 /* Go back up and keep searching. */
537 /* Return 1 if we find something, 0 if not, -1 on error. */
538 int next_in_hash(struct tdb_context *tdb, int ltype,
539 struct traverse_info *tinfo,
540 TDB_DATA *kbuf, size_t *dlen)
542 const unsigned group_bits = TDB_TOPLEVEL_HASH_BITS-TDB_HASH_GROUP_BITS;
543 tdb_off_t hlock_start, hlock_range, off;
545 while (tinfo->toplevel_group < (1 << group_bits)) {
546 hlock_start = (tdb_off_t)tinfo->toplevel_group
547 << (64 - group_bits);
548 hlock_range = 1ULL << group_bits;
549 if (tdb_lock_hashes(tdb, hlock_start, hlock_range, ltype,
553 off = iterate_hash(tdb, tinfo);
555 struct tdb_used_record rec;
557 if (tdb_read_convert(tdb, off, &rec, sizeof(rec))) {
558 tdb_unlock_hashes(tdb,
559 hlock_start, hlock_range,
563 if (rec_magic(&rec) != TDB_MAGIC) {
564 tdb->log(tdb, TDB_DEBUG_FATAL, tdb->log_priv,
566 " corrupt record at %llu\n",
571 kbuf->dsize = rec_key_length(&rec);
573 /* They want data as well? */
575 *dlen = rec_data_length(&rec);
576 kbuf->dptr = tdb_alloc_read(tdb,
581 kbuf->dptr = tdb_alloc_read(tdb,
585 tdb_unlock_hashes(tdb, hlock_start, hlock_range, ltype);
586 return kbuf->dptr ? 1 : -1;
589 tdb_unlock_hashes(tdb, hlock_start, hlock_range, ltype);
591 tinfo->toplevel_group++;
592 tinfo->levels[0].hashtable
593 += (sizeof(tdb_off_t) << TDB_HASH_GROUP_BITS);
594 tinfo->levels[0].entry = 0;
599 /* Return 1 if we find something, 0 if not, -1 on error. */
600 int first_in_hash(struct tdb_context *tdb, int ltype,
601 struct traverse_info *tinfo,
602 TDB_DATA *kbuf, size_t *dlen)
605 tinfo->toplevel_group = 0;
606 tinfo->num_levels = 1;
607 tinfo->levels[0].hashtable = offsetof(struct tdb_header, hashtable);
608 tinfo->levels[0].entry = 0;
609 tinfo->levels[0].total_buckets = (1 << TDB_HASH_GROUP_BITS);
611 return next_in_hash(tdb, ltype, tinfo, kbuf, dlen);