xref: /freebsd/contrib/jemalloc/src/ckh.c (revision d6b92ffa)
1 /*
2  *******************************************************************************
3  * Implementation of (2^1+,2) cuckoo hashing, where 2^1+ indicates that each
4  * hash bucket contains 2^n cells, for n >= 1, and 2 indicates that two hash
5  * functions are employed.  The original cuckoo hashing algorithm was described
6  * in:
7  *
8  *   Pagh, R., F.F. Rodler (2004) Cuckoo Hashing.  Journal of Algorithms
9  *     51(2):122-144.
10  *
11  * Generalization of cuckoo hashing was discussed in:
12  *
13  *   Erlingsson, U., M. Manasse, F. McSherry (2006) A cool and practical
14  *     alternative to traditional hash tables.  In Proceedings of the 7th
15  *     Workshop on Distributed Data and Structures (WDAS'06), Santa Clara, CA,
16  *     January 2006.
17  *
18  * This implementation uses precisely two hash functions because that is the
19  * fewest that can work, and supporting multiple hashes is an implementation
20  * burden.  Here is a reproduction of Figure 1 from Erlingsson et al. (2006)
21  * that shows approximate expected maximum load factors for various
22  * configurations:
23  *
24  *           |         #cells/bucket         |
25  *   #hashes |   1   |   2   |   4   |   8   |
26  *   --------+-------+-------+-------+-------+
27  *         1 | 0.006 | 0.006 | 0.03  | 0.12  |
28  *         2 | 0.49  | 0.86  |>0.93< |>0.96< |
29  *         3 | 0.91  | 0.97  | 0.98  | 0.999 |
30  *         4 | 0.97  | 0.99  | 0.999 |       |
31  *
32  * The number of cells per bucket is chosen such that a bucket fits in one cache
33  * line.  So, on 32- and 64-bit systems, we use (8,2) and (4,2) cuckoo hashing,
34  * respectively.
35  *
36  ******************************************************************************/
37 #define JEMALLOC_CKH_C_
38 #include "jemalloc/internal/jemalloc_preamble.h"
39 
40 #include "jemalloc/internal/ckh.h"
41 
42 #include "jemalloc/internal/jemalloc_internal_includes.h"
43 
44 #include "jemalloc/internal/assert.h"
45 #include "jemalloc/internal/hash.h"
46 #include "jemalloc/internal/malloc_io.h"
47 #include "jemalloc/internal/prng.h"
48 #include "jemalloc/internal/util.h"
49 
50 /******************************************************************************/
51 /* Function prototypes for non-inline static functions. */
52 
53 static bool	ckh_grow(tsd_t *tsd, ckh_t *ckh);
54 static void	ckh_shrink(tsd_t *tsd, ckh_t *ckh);
55 
56 /******************************************************************************/
57 
58 /*
59  * Search bucket for key and return the cell number if found; SIZE_T_MAX
60  * otherwise.
61  */
62 static size_t
63 ckh_bucket_search(ckh_t *ckh, size_t bucket, const void *key) {
64 	ckhc_t *cell;
65 	unsigned i;
66 
67 	for (i = 0; i < (ZU(1) << LG_CKH_BUCKET_CELLS); i++) {
68 		cell = &ckh->tab[(bucket << LG_CKH_BUCKET_CELLS) + i];
69 		if (cell->key != NULL && ckh->keycomp(key, cell->key)) {
70 			return (bucket << LG_CKH_BUCKET_CELLS) + i;
71 		}
72 	}
73 
74 	return SIZE_T_MAX;
75 }
76 
77 /*
78  * Search table for key and return cell number if found; SIZE_T_MAX otherwise.
79  */
80 static size_t
81 ckh_isearch(ckh_t *ckh, const void *key) {
82 	size_t hashes[2], bucket, cell;
83 
84 	assert(ckh != NULL);
85 
86 	ckh->hash(key, hashes);
87 
88 	/* Search primary bucket. */
89 	bucket = hashes[0] & ((ZU(1) << ckh->lg_curbuckets) - 1);
90 	cell = ckh_bucket_search(ckh, bucket, key);
91 	if (cell != SIZE_T_MAX) {
92 		return cell;
93 	}
94 
95 	/* Search secondary bucket. */
96 	bucket = hashes[1] & ((ZU(1) << ckh->lg_curbuckets) - 1);
97 	cell = ckh_bucket_search(ckh, bucket, key);
98 	return cell;
99 }
100 
101 static bool
102 ckh_try_bucket_insert(ckh_t *ckh, size_t bucket, const void *key,
103     const void *data) {
104 	ckhc_t *cell;
105 	unsigned offset, i;
106 
107 	/*
108 	 * Cycle through the cells in the bucket, starting at a random position.
109 	 * The randomness avoids worst-case search overhead as buckets fill up.
110 	 */
111 	offset = (unsigned)prng_lg_range_u64(&ckh->prng_state,
112 	    LG_CKH_BUCKET_CELLS);
113 	for (i = 0; i < (ZU(1) << LG_CKH_BUCKET_CELLS); i++) {
114 		cell = &ckh->tab[(bucket << LG_CKH_BUCKET_CELLS) +
115 		    ((i + offset) & ((ZU(1) << LG_CKH_BUCKET_CELLS) - 1))];
116 		if (cell->key == NULL) {
117 			cell->key = key;
118 			cell->data = data;
119 			ckh->count++;
120 			return false;
121 		}
122 	}
123 
124 	return true;
125 }
126 
127 /*
128  * No space is available in bucket.  Randomly evict an item, then try to find an
129  * alternate location for that item.  Iteratively repeat this
130  * eviction/relocation procedure until either success or detection of an
131  * eviction/relocation bucket cycle.
132  */
133 static bool
134 ckh_evict_reloc_insert(ckh_t *ckh, size_t argbucket, void const **argkey,
135     void const **argdata) {
136 	const void *key, *data, *tkey, *tdata;
137 	ckhc_t *cell;
138 	size_t hashes[2], bucket, tbucket;
139 	unsigned i;
140 
141 	bucket = argbucket;
142 	key = *argkey;
143 	data = *argdata;
144 	while (true) {
145 		/*
146 		 * Choose a random item within the bucket to evict.  This is
147 		 * critical to correct function, because without (eventually)
148 		 * evicting all items within a bucket during iteration, it
149 		 * would be possible to get stuck in an infinite loop if there
150 		 * were an item for which both hashes indicated the same
151 		 * bucket.
152 		 */
153 		i = (unsigned)prng_lg_range_u64(&ckh->prng_state,
154 		    LG_CKH_BUCKET_CELLS);
155 		cell = &ckh->tab[(bucket << LG_CKH_BUCKET_CELLS) + i];
156 		assert(cell->key != NULL);
157 
158 		/* Swap cell->{key,data} and {key,data} (evict). */
159 		tkey = cell->key; tdata = cell->data;
160 		cell->key = key; cell->data = data;
161 		key = tkey; data = tdata;
162 
163 #ifdef CKH_COUNT
164 		ckh->nrelocs++;
165 #endif
166 
167 		/* Find the alternate bucket for the evicted item. */
168 		ckh->hash(key, hashes);
169 		tbucket = hashes[1] & ((ZU(1) << ckh->lg_curbuckets) - 1);
170 		if (tbucket == bucket) {
171 			tbucket = hashes[0] & ((ZU(1) << ckh->lg_curbuckets)
172 			    - 1);
173 			/*
174 			 * It may be that (tbucket == bucket) still, if the
175 			 * item's hashes both indicate this bucket.  However,
176 			 * we are guaranteed to eventually escape this bucket
177 			 * during iteration, assuming pseudo-random item
178 			 * selection (true randomness would make infinite
179 			 * looping a remote possibility).  The reason we can
180 			 * never get trapped forever is that there are two
181 			 * cases:
182 			 *
183 			 * 1) This bucket == argbucket, so we will quickly
184 			 *    detect an eviction cycle and terminate.
185 			 * 2) An item was evicted to this bucket from another,
186 			 *    which means that at least one item in this bucket
187 			 *    has hashes that indicate distinct buckets.
188 			 */
189 		}
190 		/* Check for a cycle. */
191 		if (tbucket == argbucket) {
192 			*argkey = key;
193 			*argdata = data;
194 			return true;
195 		}
196 
197 		bucket = tbucket;
198 		if (!ckh_try_bucket_insert(ckh, bucket, key, data)) {
199 			return false;
200 		}
201 	}
202 }
203 
204 static bool
205 ckh_try_insert(ckh_t *ckh, void const**argkey, void const**argdata) {
206 	size_t hashes[2], bucket;
207 	const void *key = *argkey;
208 	const void *data = *argdata;
209 
210 	ckh->hash(key, hashes);
211 
212 	/* Try to insert in primary bucket. */
213 	bucket = hashes[0] & ((ZU(1) << ckh->lg_curbuckets) - 1);
214 	if (!ckh_try_bucket_insert(ckh, bucket, key, data)) {
215 		return false;
216 	}
217 
218 	/* Try to insert in secondary bucket. */
219 	bucket = hashes[1] & ((ZU(1) << ckh->lg_curbuckets) - 1);
220 	if (!ckh_try_bucket_insert(ckh, bucket, key, data)) {
221 		return false;
222 	}
223 
224 	/*
225 	 * Try to find a place for this item via iterative eviction/relocation.
226 	 */
227 	return ckh_evict_reloc_insert(ckh, bucket, argkey, argdata);
228 }
229 
230 /*
231  * Try to rebuild the hash table from scratch by inserting all items from the
232  * old table into the new.
233  */
234 static bool
235 ckh_rebuild(ckh_t *ckh, ckhc_t *aTab) {
236 	size_t count, i, nins;
237 	const void *key, *data;
238 
239 	count = ckh->count;
240 	ckh->count = 0;
241 	for (i = nins = 0; nins < count; i++) {
242 		if (aTab[i].key != NULL) {
243 			key = aTab[i].key;
244 			data = aTab[i].data;
245 			if (ckh_try_insert(ckh, &key, &data)) {
246 				ckh->count = count;
247 				return true;
248 			}
249 			nins++;
250 		}
251 	}
252 
253 	return false;
254 }
255 
256 static bool
257 ckh_grow(tsd_t *tsd, ckh_t *ckh) {
258 	bool ret;
259 	ckhc_t *tab, *ttab;
260 	unsigned lg_prevbuckets, lg_curcells;
261 
262 #ifdef CKH_COUNT
263 	ckh->ngrows++;
264 #endif
265 
266 	/*
267 	 * It is possible (though unlikely, given well behaved hashes) that the
268 	 * table will have to be doubled more than once in order to create a
269 	 * usable table.
270 	 */
271 	lg_prevbuckets = ckh->lg_curbuckets;
272 	lg_curcells = ckh->lg_curbuckets + LG_CKH_BUCKET_CELLS;
273 	while (true) {
274 		size_t usize;
275 
276 		lg_curcells++;
277 		usize = sz_sa2u(sizeof(ckhc_t) << lg_curcells, CACHELINE);
278 		if (unlikely(usize == 0 || usize > LARGE_MAXCLASS)) {
279 			ret = true;
280 			goto label_return;
281 		}
282 		tab = (ckhc_t *)ipallocztm(tsd_tsdn(tsd), usize, CACHELINE,
283 		    true, NULL, true, arena_ichoose(tsd, NULL));
284 		if (tab == NULL) {
285 			ret = true;
286 			goto label_return;
287 		}
288 		/* Swap in new table. */
289 		ttab = ckh->tab;
290 		ckh->tab = tab;
291 		tab = ttab;
292 		ckh->lg_curbuckets = lg_curcells - LG_CKH_BUCKET_CELLS;
293 
294 		if (!ckh_rebuild(ckh, tab)) {
295 			idalloctm(tsd_tsdn(tsd), tab, NULL, NULL, true, true);
296 			break;
297 		}
298 
299 		/* Rebuilding failed, so back out partially rebuilt table. */
300 		idalloctm(tsd_tsdn(tsd), ckh->tab, NULL, NULL, true, true);
301 		ckh->tab = tab;
302 		ckh->lg_curbuckets = lg_prevbuckets;
303 	}
304 
305 	ret = false;
306 label_return:
307 	return ret;
308 }
309 
310 static void
311 ckh_shrink(tsd_t *tsd, ckh_t *ckh) {
312 	ckhc_t *tab, *ttab;
313 	size_t usize;
314 	unsigned lg_prevbuckets, lg_curcells;
315 
316 	/*
317 	 * It is possible (though unlikely, given well behaved hashes) that the
318 	 * table rebuild will fail.
319 	 */
320 	lg_prevbuckets = ckh->lg_curbuckets;
321 	lg_curcells = ckh->lg_curbuckets + LG_CKH_BUCKET_CELLS - 1;
322 	usize = sz_sa2u(sizeof(ckhc_t) << lg_curcells, CACHELINE);
323 	if (unlikely(usize == 0 || usize > LARGE_MAXCLASS)) {
324 		return;
325 	}
326 	tab = (ckhc_t *)ipallocztm(tsd_tsdn(tsd), usize, CACHELINE, true, NULL,
327 	    true, arena_ichoose(tsd, NULL));
328 	if (tab == NULL) {
329 		/*
330 		 * An OOM error isn't worth propagating, since it doesn't
331 		 * prevent this or future operations from proceeding.
332 		 */
333 		return;
334 	}
335 	/* Swap in new table. */
336 	ttab = ckh->tab;
337 	ckh->tab = tab;
338 	tab = ttab;
339 	ckh->lg_curbuckets = lg_curcells - LG_CKH_BUCKET_CELLS;
340 
341 	if (!ckh_rebuild(ckh, tab)) {
342 		idalloctm(tsd_tsdn(tsd), tab, NULL, NULL, true, true);
343 #ifdef CKH_COUNT
344 		ckh->nshrinks++;
345 #endif
346 		return;
347 	}
348 
349 	/* Rebuilding failed, so back out partially rebuilt table. */
350 	idalloctm(tsd_tsdn(tsd), ckh->tab, NULL, NULL, true, true);
351 	ckh->tab = tab;
352 	ckh->lg_curbuckets = lg_prevbuckets;
353 #ifdef CKH_COUNT
354 	ckh->nshrinkfails++;
355 #endif
356 }
357 
358 bool
359 ckh_new(tsd_t *tsd, ckh_t *ckh, size_t minitems, ckh_hash_t *hash,
360     ckh_keycomp_t *keycomp) {
361 	bool ret;
362 	size_t mincells, usize;
363 	unsigned lg_mincells;
364 
365 	assert(minitems > 0);
366 	assert(hash != NULL);
367 	assert(keycomp != NULL);
368 
369 #ifdef CKH_COUNT
370 	ckh->ngrows = 0;
371 	ckh->nshrinks = 0;
372 	ckh->nshrinkfails = 0;
373 	ckh->ninserts = 0;
374 	ckh->nrelocs = 0;
375 #endif
376 	ckh->prng_state = 42; /* Value doesn't really matter. */
377 	ckh->count = 0;
378 
379 	/*
380 	 * Find the minimum power of 2 that is large enough to fit minitems
381 	 * entries.  We are using (2+,2) cuckoo hashing, which has an expected
382 	 * maximum load factor of at least ~0.86, so 0.75 is a conservative load
383 	 * factor that will typically allow mincells items to fit without ever
384 	 * growing the table.
385 	 */
386 	assert(LG_CKH_BUCKET_CELLS > 0);
387 	mincells = ((minitems + (3 - (minitems % 3))) / 3) << 2;
388 	for (lg_mincells = LG_CKH_BUCKET_CELLS;
389 	    (ZU(1) << lg_mincells) < mincells;
390 	    lg_mincells++) {
391 		/* Do nothing. */
392 	}
393 	ckh->lg_minbuckets = lg_mincells - LG_CKH_BUCKET_CELLS;
394 	ckh->lg_curbuckets = lg_mincells - LG_CKH_BUCKET_CELLS;
395 	ckh->hash = hash;
396 	ckh->keycomp = keycomp;
397 
398 	usize = sz_sa2u(sizeof(ckhc_t) << lg_mincells, CACHELINE);
399 	if (unlikely(usize == 0 || usize > LARGE_MAXCLASS)) {
400 		ret = true;
401 		goto label_return;
402 	}
403 	ckh->tab = (ckhc_t *)ipallocztm(tsd_tsdn(tsd), usize, CACHELINE, true,
404 	    NULL, true, arena_ichoose(tsd, NULL));
405 	if (ckh->tab == NULL) {
406 		ret = true;
407 		goto label_return;
408 	}
409 
410 	ret = false;
411 label_return:
412 	return ret;
413 }
414 
415 void
416 ckh_delete(tsd_t *tsd, ckh_t *ckh) {
417 	assert(ckh != NULL);
418 
419 #ifdef CKH_VERBOSE
420 	malloc_printf(
421 	    "%s(%p): ngrows: %"FMTu64", nshrinks: %"FMTu64","
422 	    " nshrinkfails: %"FMTu64", ninserts: %"FMTu64","
423 	    " nrelocs: %"FMTu64"\n", __func__, ckh,
424 	    (unsigned long long)ckh->ngrows,
425 	    (unsigned long long)ckh->nshrinks,
426 	    (unsigned long long)ckh->nshrinkfails,
427 	    (unsigned long long)ckh->ninserts,
428 	    (unsigned long long)ckh->nrelocs);
429 #endif
430 
431 	idalloctm(tsd_tsdn(tsd), ckh->tab, NULL, NULL, true, true);
432 	if (config_debug) {
433 		memset(ckh, JEMALLOC_FREE_JUNK, sizeof(ckh_t));
434 	}
435 }
436 
437 size_t
438 ckh_count(ckh_t *ckh) {
439 	assert(ckh != NULL);
440 
441 	return ckh->count;
442 }
443 
444 bool
445 ckh_iter(ckh_t *ckh, size_t *tabind, void **key, void **data) {
446 	size_t i, ncells;
447 
448 	for (i = *tabind, ncells = (ZU(1) << (ckh->lg_curbuckets +
449 	    LG_CKH_BUCKET_CELLS)); i < ncells; i++) {
450 		if (ckh->tab[i].key != NULL) {
451 			if (key != NULL) {
452 				*key = (void *)ckh->tab[i].key;
453 			}
454 			if (data != NULL) {
455 				*data = (void *)ckh->tab[i].data;
456 			}
457 			*tabind = i + 1;
458 			return false;
459 		}
460 	}
461 
462 	return true;
463 }
464 
465 bool
466 ckh_insert(tsd_t *tsd, ckh_t *ckh, const void *key, const void *data) {
467 	bool ret;
468 
469 	assert(ckh != NULL);
470 	assert(ckh_search(ckh, key, NULL, NULL));
471 
472 #ifdef CKH_COUNT
473 	ckh->ninserts++;
474 #endif
475 
476 	while (ckh_try_insert(ckh, &key, &data)) {
477 		if (ckh_grow(tsd, ckh)) {
478 			ret = true;
479 			goto label_return;
480 		}
481 	}
482 
483 	ret = false;
484 label_return:
485 	return ret;
486 }
487 
488 bool
489 ckh_remove(tsd_t *tsd, ckh_t *ckh, const void *searchkey, void **key,
490     void **data) {
491 	size_t cell;
492 
493 	assert(ckh != NULL);
494 
495 	cell = ckh_isearch(ckh, searchkey);
496 	if (cell != SIZE_T_MAX) {
497 		if (key != NULL) {
498 			*key = (void *)ckh->tab[cell].key;
499 		}
500 		if (data != NULL) {
501 			*data = (void *)ckh->tab[cell].data;
502 		}
503 		ckh->tab[cell].key = NULL;
504 		ckh->tab[cell].data = NULL; /* Not necessary. */
505 
506 		ckh->count--;
507 		/* Try to halve the table if it is less than 1/4 full. */
508 		if (ckh->count < (ZU(1) << (ckh->lg_curbuckets
509 		    + LG_CKH_BUCKET_CELLS - 2)) && ckh->lg_curbuckets
510 		    > ckh->lg_minbuckets) {
511 			/* Ignore error due to OOM. */
512 			ckh_shrink(tsd, ckh);
513 		}
514 
515 		return false;
516 	}
517 
518 	return true;
519 }
520 
521 bool
522 ckh_search(ckh_t *ckh, const void *searchkey, void **key, void **data) {
523 	size_t cell;
524 
525 	assert(ckh != NULL);
526 
527 	cell = ckh_isearch(ckh, searchkey);
528 	if (cell != SIZE_T_MAX) {
529 		if (key != NULL) {
530 			*key = (void *)ckh->tab[cell].key;
531 		}
532 		if (data != NULL) {
533 			*data = (void *)ckh->tab[cell].data;
534 		}
535 		return false;
536 	}
537 
538 	return true;
539 }
540 
541 void
542 ckh_string_hash(const void *key, size_t r_hash[2]) {
543 	hash(key, strlen((const char *)key), 0x94122f33U, r_hash);
544 }
545 
546 bool
547 ckh_string_keycomp(const void *k1, const void *k2) {
548 	assert(k1 != NULL);
549 	assert(k2 != NULL);
550 
551 	return !strcmp((char *)k1, (char *)k2);
552 }
553 
554 void
555 ckh_pointer_hash(const void *key, size_t r_hash[2]) {
556 	union {
557 		const void	*v;
558 		size_t		i;
559 	} u;
560 
561 	assert(sizeof(u.v) == sizeof(u.i));
562 	u.v = key;
563 	hash(&u.i, sizeof(u.i), 0xd983396eU, r_hash);
564 }
565 
566 bool
567 ckh_pointer_keycomp(const void *k1, const void *k2) {
568 	return (k1 == k2);
569 }
570