2 * Resizable, Scalable, Concurrent Hash Table
7 * Based on the following paper:
8 * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
10 * Code partially derived from nft_hash
12 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2 as
14 * published by the Free Software Foundation.
17 #include <linux/kernel.h>
18 #include <linux/init.h>
19 #include <linux/log2.h>
20 #include <linux/slab.h>
21 #include <linux/vmalloc.h>
23 #include <linux/hash.h>
24 #include <linux/random.h>
25 #include <linux/rhashtable.h>
27 #define HASH_DEFAULT_SIZE 64UL
28 #define HASH_MIN_SIZE 4UL
30 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
32 #ifdef CONFIG_PROVE_LOCKING
33 int lockdep_rht_mutex_is_held(const struct rhashtable *ht)
35 return ht->p.mutex_is_held(ht->p.parent);
37 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
40 static void *rht_obj(const struct rhashtable *ht, const struct rhash_head *he)
42 return (void *) he - ht->p.head_offset;
45 static u32 __hashfn(const struct rhashtable *ht, const void *key,
50 h = ht->p.hashfn(key, len, ht->p.hash_rnd);
52 return h & (hsize - 1);
56 * rhashtable_hashfn - compute hash for key of given length
57 * @ht: hash table to compute for
58 * @key: pointer to key
61 * Computes the hash value using the hash function provided in the 'hashfn'
62 * of struct rhashtable_params. The returned value is guaranteed to be
63 * smaller than the number of buckets in the hash table.
65 u32 rhashtable_hashfn(const struct rhashtable *ht, const void *key, u32 len)
67 struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
69 return __hashfn(ht, key, len, tbl->size);
71 EXPORT_SYMBOL_GPL(rhashtable_hashfn);
73 static u32 obj_hashfn(const struct rhashtable *ht, const void *ptr, u32 hsize)
75 if (unlikely(!ht->p.key_len)) {
78 h = ht->p.obj_hashfn(ptr, ht->p.hash_rnd);
80 return h & (hsize - 1);
83 return __hashfn(ht, ptr + ht->p.key_offset, ht->p.key_len, hsize);
87 * rhashtable_obj_hashfn - compute hash for hashed object
88 * @ht: hash table to compute for
89 * @ptr: pointer to hashed object
91 * Computes the hash value using the hash function `hashfn` respectively
92 * 'obj_hashfn' depending on whether the hash table is set up to work with
93 * a fixed length key. The returned value is guaranteed to be smaller than
94 * the number of buckets in the hash table.
96 u32 rhashtable_obj_hashfn(const struct rhashtable *ht, void *ptr)
98 struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
100 return obj_hashfn(ht, ptr, tbl->size);
102 EXPORT_SYMBOL_GPL(rhashtable_obj_hashfn);
104 static u32 head_hashfn(const struct rhashtable *ht,
105 const struct rhash_head *he, u32 hsize)
107 return obj_hashfn(ht, rht_obj(ht, he), hsize);
110 static struct bucket_table *bucket_table_alloc(size_t nbuckets)
112 struct bucket_table *tbl;
115 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
116 tbl = kzalloc(size, GFP_KERNEL | __GFP_NOWARN);
123 tbl->size = nbuckets;
128 static void bucket_table_free(const struct bucket_table *tbl)
134 * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
136 * @new_size: new table size
138 bool rht_grow_above_75(const struct rhashtable *ht, size_t new_size)
140 /* Expand table when exceeding 75% load */
141 return ht->nelems > (new_size / 4 * 3);
143 EXPORT_SYMBOL_GPL(rht_grow_above_75);
146 * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
148 * @new_size: new table size
150 bool rht_shrink_below_30(const struct rhashtable *ht, size_t new_size)
152 /* Shrink table beneath 30% load */
153 return ht->nelems < (new_size * 3 / 10);
155 EXPORT_SYMBOL_GPL(rht_shrink_below_30);
157 static void hashtable_chain_unzip(const struct rhashtable *ht,
158 const struct bucket_table *new_tbl,
159 struct bucket_table *old_tbl, size_t n)
161 struct rhash_head *he, *p, *next;
164 /* Old bucket empty, no work needed. */
165 p = rht_dereference(old_tbl->buckets[n], ht);
169 /* Advance the old bucket pointer one or more times until it
170 * reaches a node that doesn't hash to the same bucket as the
171 * previous node p. Call the previous node p;
173 h = head_hashfn(ht, p, new_tbl->size);
174 rht_for_each(he, p->next, ht) {
175 if (head_hashfn(ht, he, new_tbl->size) != h)
179 RCU_INIT_POINTER(old_tbl->buckets[n], p->next);
181 /* Find the subsequent node which does hash to the same
182 * bucket as node P, or NULL if no such node exists.
186 rht_for_each(he, he->next, ht) {
187 if (head_hashfn(ht, he, new_tbl->size) == h) {
194 /* Set p's next pointer to that subsequent node pointer,
195 * bypassing the nodes which do not hash to p's bucket
197 RCU_INIT_POINTER(p->next, next);
201 * rhashtable_expand - Expand hash table while allowing concurrent lookups
202 * @ht: the hash table to expand
204 * A secondary bucket array is allocated and the hash entries are migrated
205 * while keeping them on both lists until the end of the RCU grace period.
207 * This function may only be called in a context where it is safe to call
208 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
210 * The caller must ensure that no concurrent table mutations take place.
211 * It is however valid to have concurrent lookups if they are RCU protected.
213 int rhashtable_expand(struct rhashtable *ht)
215 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
216 struct rhash_head *he;
220 ASSERT_RHT_MUTEX(ht);
222 if (ht->p.max_shift && ht->shift >= ht->p.max_shift)
225 new_tbl = bucket_table_alloc(old_tbl->size * 2);
231 /* For each new bucket, search the corresponding old bucket
232 * for the first entry that hashes to the new bucket, and
233 * link the new bucket to that entry. Since all the entries
234 * which will end up in the new bucket appear in the same
235 * old bucket, this constructs an entirely valid new hash
236 * table, but with multiple buckets "zipped" together into a
237 * single imprecise chain.
239 for (i = 0; i < new_tbl->size; i++) {
240 h = i & (old_tbl->size - 1);
241 rht_for_each(he, old_tbl->buckets[h], ht) {
242 if (head_hashfn(ht, he, new_tbl->size) == i) {
243 RCU_INIT_POINTER(new_tbl->buckets[i], he);
249 /* Publish the new table pointer. Lookups may now traverse
250 * the new table, but they will not benefit from any
251 * additional efficiency until later steps unzip the buckets.
253 rcu_assign_pointer(ht->tbl, new_tbl);
255 /* Unzip interleaved hash chains */
257 /* Wait for readers. All new readers will see the new
258 * table, and thus no references to the old table will
263 /* For each bucket in the old table (each of which
264 * contains items from multiple buckets of the new
268 for (i = 0; i < old_tbl->size; i++) {
269 hashtable_chain_unzip(ht, new_tbl, old_tbl, i);
270 if (old_tbl->buckets[i] != NULL)
275 bucket_table_free(old_tbl);
278 EXPORT_SYMBOL_GPL(rhashtable_expand);
281 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
282 * @ht: the hash table to shrink
284 * This function may only be called in a context where it is safe to call
285 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
287 * The caller must ensure that no concurrent table mutations take place.
288 * It is however valid to have concurrent lookups if they are RCU protected.
290 int rhashtable_shrink(struct rhashtable *ht)
292 struct bucket_table *ntbl, *tbl = rht_dereference(ht->tbl, ht);
293 struct rhash_head __rcu **pprev;
296 ASSERT_RHT_MUTEX(ht);
298 if (ht->shift <= ht->p.min_shift)
301 ntbl = bucket_table_alloc(tbl->size / 2);
307 /* Link each bucket in the new table to the first bucket
308 * in the old table that contains entries which will hash
311 for (i = 0; i < ntbl->size; i++) {
312 ntbl->buckets[i] = tbl->buckets[i];
314 /* Link each bucket in the new table to the first bucket
315 * in the old table that contains entries which will hash
318 for (pprev = &ntbl->buckets[i]; *pprev != NULL;
319 pprev = &rht_dereference(*pprev, ht)->next)
321 RCU_INIT_POINTER(*pprev, tbl->buckets[i + ntbl->size]);
324 /* Publish the new, valid hash table */
325 rcu_assign_pointer(ht->tbl, ntbl);
327 /* Wait for readers. No new readers will have references to the
332 bucket_table_free(tbl);
336 EXPORT_SYMBOL_GPL(rhashtable_shrink);
339 * rhashtable_insert - insert object into hash hash table
341 * @obj: pointer to hash head inside object
343 * Will automatically grow the table via rhashtable_expand() if the the
344 * grow_decision function specified at rhashtable_init() returns true.
346 * The caller must ensure that no concurrent table mutations occur. It is
347 * however valid to have concurrent lookups if they are RCU protected.
349 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *obj)
351 struct bucket_table *tbl = rht_dereference(ht->tbl, ht);
354 ASSERT_RHT_MUTEX(ht);
356 hash = head_hashfn(ht, obj, tbl->size);
357 RCU_INIT_POINTER(obj->next, tbl->buckets[hash]);
358 rcu_assign_pointer(tbl->buckets[hash], obj);
361 if (ht->p.grow_decision && ht->p.grow_decision(ht, tbl->size))
362 rhashtable_expand(ht);
364 EXPORT_SYMBOL_GPL(rhashtable_insert);
367 * rhashtable_remove_pprev - remove object from hash table given previous element
369 * @obj: pointer to hash head inside object
370 * @pprev: pointer to previous element
372 * Identical to rhashtable_remove() but caller is alreayd aware of the element
373 * in front of the element to be deleted. This is in particular useful for
374 * deletion when combined with walking or lookup.
376 void rhashtable_remove_pprev(struct rhashtable *ht, struct rhash_head *obj,
377 struct rhash_head __rcu **pprev)
379 struct bucket_table *tbl = rht_dereference(ht->tbl, ht);
381 ASSERT_RHT_MUTEX(ht);
383 RCU_INIT_POINTER(*pprev, obj->next);
386 if (ht->p.shrink_decision &&
387 ht->p.shrink_decision(ht, tbl->size))
388 rhashtable_shrink(ht);
390 EXPORT_SYMBOL_GPL(rhashtable_remove_pprev);
393 * rhashtable_remove - remove object from hash table
395 * @obj: pointer to hash head inside object
397 * Since the hash chain is single linked, the removal operation needs to
398 * walk the bucket chain upon removal. The removal operation is thus
399 * considerable slow if the hash table is not correctly sized.
401 * Will automatically shrink the table via rhashtable_expand() if the the
402 * shrink_decision function specified at rhashtable_init() returns true.
404 * The caller must ensure that no concurrent table mutations occur. It is
405 * however valid to have concurrent lookups if they are RCU protected.
407 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *obj)
409 struct bucket_table *tbl = rht_dereference(ht->tbl, ht);
410 struct rhash_head __rcu **pprev;
411 struct rhash_head *he;
414 ASSERT_RHT_MUTEX(ht);
416 h = head_hashfn(ht, obj, tbl->size);
418 pprev = &tbl->buckets[h];
419 rht_for_each(he, tbl->buckets[h], ht) {
425 rhashtable_remove_pprev(ht, he, pprev);
431 EXPORT_SYMBOL_GPL(rhashtable_remove);
434 * rhashtable_lookup - lookup key in hash table
436 * @key: pointer to key
438 * Computes the hash value for the key and traverses the bucket chain looking
439 * for a entry with an identical key. The first matching entry is returned.
441 * This lookup function may only be used for fixed key hash table (key_len
442 * paramter set). It will BUG() if used inappropriately.
444 * Lookups may occur in parallel with hash mutations as long as the lookup is
445 * guarded by rcu_read_lock(). The caller must take care of this.
447 void *rhashtable_lookup(const struct rhashtable *ht, const void *key)
449 const struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
450 struct rhash_head *he;
453 BUG_ON(!ht->p.key_len);
455 h = __hashfn(ht, key, ht->p.key_len, tbl->size);
456 rht_for_each_rcu(he, tbl->buckets[h], ht) {
457 if (memcmp(rht_obj(ht, he) + ht->p.key_offset, key,
460 return (void *) he - ht->p.head_offset;
465 EXPORT_SYMBOL_GPL(rhashtable_lookup);
468 * rhashtable_lookup_compare - search hash table with compare function
470 * @hash: hash value of desired entry
471 * @compare: compare function, must return true on match
472 * @arg: argument passed on to compare function
474 * Traverses the bucket chain behind the provided hash value and calls the
475 * specified compare function for each entry.
477 * Lookups may occur in parallel with hash mutations as long as the lookup is
478 * guarded by rcu_read_lock(). The caller must take care of this.
480 * Returns the first entry on which the compare function returned true.
482 void *rhashtable_lookup_compare(const struct rhashtable *ht, u32 hash,
483 bool (*compare)(void *, void *), void *arg)
485 const struct bucket_table *tbl = rht_dereference_rcu(ht->tbl, ht);
486 struct rhash_head *he;
488 if (unlikely(hash >= tbl->size))
491 rht_for_each_rcu(he, tbl->buckets[hash], ht) {
492 if (!compare(rht_obj(ht, he), arg))
494 return (void *) he - ht->p.head_offset;
499 EXPORT_SYMBOL_GPL(rhashtable_lookup_compare);
501 static size_t rounded_hashtable_size(struct rhashtable_params *params)
503 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
504 1UL << params->min_shift);
508 * rhashtable_init - initialize a new hash table
509 * @ht: hash table to be initialized
510 * @params: configuration parameters
512 * Initializes a new hash table based on the provided configuration
513 * parameters. A table can be configured either with a variable or
516 * Configuration Example 1: Fixed length keys
520 * struct rhash_head node;
523 * struct rhashtable_params params = {
524 * .head_offset = offsetof(struct test_obj, node),
525 * .key_offset = offsetof(struct test_obj, key),
526 * .key_len = sizeof(int),
527 * .hashfn = arch_fast_hash,
528 * #ifdef CONFIG_PROVE_LOCKING
529 * .mutex_is_held = &my_mutex_is_held,
533 * Configuration Example 2: Variable length keys
536 * struct rhash_head node;
539 * u32 my_hash_fn(const void *data, u32 seed)
541 * struct test_obj *obj = data;
543 * return [... hash ...];
546 * struct rhashtable_params params = {
547 * .head_offset = offsetof(struct test_obj, node),
548 * .hashfn = arch_fast_hash,
549 * .obj_hashfn = my_hash_fn,
550 * #ifdef CONFIG_PROVE_LOCKING
551 * .mutex_is_held = &my_mutex_is_held,
555 int rhashtable_init(struct rhashtable *ht, struct rhashtable_params *params)
557 struct bucket_table *tbl;
560 size = HASH_DEFAULT_SIZE;
562 if ((params->key_len && !params->hashfn) ||
563 (!params->key_len && !params->obj_hashfn))
566 params->min_shift = max_t(size_t, params->min_shift,
567 ilog2(HASH_MIN_SIZE));
569 if (params->nelem_hint)
570 size = rounded_hashtable_size(params);
572 tbl = bucket_table_alloc(size);
576 memset(ht, 0, sizeof(*ht));
577 ht->shift = ilog2(tbl->size);
578 memcpy(&ht->p, params, sizeof(*params));
579 RCU_INIT_POINTER(ht->tbl, tbl);
582 get_random_bytes(&ht->p.hash_rnd, sizeof(ht->p.hash_rnd));
586 EXPORT_SYMBOL_GPL(rhashtable_init);
589 * rhashtable_destroy - destroy hash table
590 * @ht: the hash table to destroy
592 * Frees the bucket array. This function is not rcu safe, therefore the caller
593 * has to make sure that no resizing may happen by unpublishing the hashtable
594 * and waiting for the quiescent cycle before releasing the bucket array.
596 void rhashtable_destroy(const struct rhashtable *ht)
598 bucket_table_free(ht->tbl);
600 EXPORT_SYMBOL_GPL(rhashtable_destroy);
602 /**************************************************************************
604 **************************************************************************/
606 #ifdef CONFIG_TEST_RHASHTABLE
608 #define TEST_HT_SIZE 8
609 #define TEST_ENTRIES 2048
610 #define TEST_PTR ((void *) 0xdeadbeef)
611 #define TEST_NEXPANDS 4
613 #ifdef CONFIG_PROVE_LOCKING
614 static int test_mutex_is_held(void *parent)
623 struct rhash_head node;
626 static int __init test_rht_lookup(struct rhashtable *ht)
630 for (i = 0; i < TEST_ENTRIES * 2; i++) {
631 struct test_obj *obj;
632 bool expected = !(i % 2);
635 obj = rhashtable_lookup(ht, &key);
637 if (expected && !obj) {
638 pr_warn("Test failed: Could not find key %u\n", key);
640 } else if (!expected && obj) {
641 pr_warn("Test failed: Unexpected entry found for key %u\n",
644 } else if (expected && obj) {
645 if (obj->ptr != TEST_PTR || obj->value != i) {
646 pr_warn("Test failed: Lookup value mismatch %p!=%p, %u!=%u\n",
647 obj->ptr, TEST_PTR, obj->value, i);
656 static void test_bucket_stats(struct rhashtable *ht,
657 struct bucket_table *tbl,
660 unsigned int cnt, i, total = 0;
661 struct test_obj *obj;
663 for (i = 0; i < tbl->size; i++) {
667 pr_info(" [%#4x/%zu]", i, tbl->size);
669 rht_for_each_entry_rcu(obj, tbl->buckets[i], node) {
673 pr_cont(" [%p],", obj);
677 pr_cont("\n [%#x] first element: %p, chain length: %u\n",
678 i, tbl->buckets[i], cnt);
681 pr_info(" Traversal complete: counted=%u, nelems=%zu, entries=%d\n",
682 total, ht->nelems, TEST_ENTRIES);
685 static int __init test_rhashtable(struct rhashtable *ht)
687 struct bucket_table *tbl;
688 struct test_obj *obj, *next;
694 * Insert TEST_ENTRIES into table with all keys even numbers
696 pr_info(" Adding %d keys\n", TEST_ENTRIES);
697 for (i = 0; i < TEST_ENTRIES; i++) {
698 struct test_obj *obj;
700 obj = kzalloc(sizeof(*obj), GFP_KERNEL);
709 rhashtable_insert(ht, &obj->node);
713 tbl = rht_dereference_rcu(ht->tbl, ht);
714 test_bucket_stats(ht, tbl, true);
718 for (i = 0; i < TEST_NEXPANDS; i++) {
719 pr_info(" Table expansion iteration %u...\n", i);
720 rhashtable_expand(ht);
723 pr_info(" Verifying lookups...\n");
728 for (i = 0; i < TEST_NEXPANDS; i++) {
729 pr_info(" Table shrinkage iteration %u...\n", i);
730 rhashtable_shrink(ht);
733 pr_info(" Verifying lookups...\n");
738 pr_info(" Deleting %d keys\n", TEST_ENTRIES);
739 for (i = 0; i < TEST_ENTRIES; i++) {
742 obj = rhashtable_lookup(ht, &key);
745 rhashtable_remove(ht, &obj->node);
752 tbl = rht_dereference_rcu(ht->tbl, ht);
753 for (i = 0; i < tbl->size; i++)
754 rht_for_each_entry_safe(obj, next, tbl->buckets[i], ht, node)
760 static int __init test_rht_init(void)
762 struct rhashtable ht;
763 struct rhashtable_params params = {
764 .nelem_hint = TEST_HT_SIZE,
765 .head_offset = offsetof(struct test_obj, node),
766 .key_offset = offsetof(struct test_obj, value),
767 .key_len = sizeof(int),
768 .hashfn = arch_fast_hash,
769 #ifdef CONFIG_PROVE_LOCKING
770 .mutex_is_held = &test_mutex_is_held,
772 .grow_decision = rht_grow_above_75,
773 .shrink_decision = rht_shrink_below_30,
777 pr_info("Running resizable hashtable tests...\n");
779 err = rhashtable_init(&ht, ¶ms);
781 pr_warn("Test failed: Unable to initialize hashtable: %d\n",
786 err = test_rhashtable(&ht);
788 rhashtable_destroy(&ht);
793 subsys_initcall(test_rht_init);
795 #endif /* CONFIG_TEST_RHASHTABLE */