2 * Resizable, Scalable, Concurrent Hash Table
4 * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
5 * Copyright (c) 2014-2015 Thomas Graf <tgraf@suug.ch>
6 * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
8 * Code partially derived from nft_hash
9 * Rewritten with rehash code from br_multicast plus single list
10 * pointer as suggested by Josh Triplett
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/atomic.h>
18 #include <linux/kernel.h>
19 #include <linux/init.h>
20 #include <linux/log2.h>
21 #include <linux/sched.h>
22 #include <linux/slab.h>
23 #include <linux/vmalloc.h>
25 #include <linux/jhash.h>
26 #include <linux/random.h>
27 #include <linux/rhashtable.h>
28 #include <linux/err.h>
29 #include <linux/export.h>
31 #define HASH_DEFAULT_SIZE 64UL
32 #define HASH_MIN_SIZE 4U
33 #define BUCKET_LOCKS_PER_CPU 32UL
35 static u32 head_hashfn(struct rhashtable *ht,
36 const struct bucket_table *tbl,
37 const struct rhash_head *he)
39 return rht_head_hashfn(ht, tbl, he, ht->p);
42 #ifdef CONFIG_PROVE_LOCKING
43 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
45 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
47 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
49 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
51 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
53 spinlock_t *lock = rht_bucket_lock(tbl, hash);
55 return (debug_locks) ? lockdep_is_held(lock) : 1;
57 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
59 #define ASSERT_RHT_MUTEX(HT)
63 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
67 #if defined(CONFIG_PROVE_LOCKING)
68 unsigned int nr_pcpus = 2;
70 unsigned int nr_pcpus = num_possible_cpus();
73 nr_pcpus = min_t(unsigned int, nr_pcpus, 64UL);
74 size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
76 /* Never allocate more than 0.5 locks per bucket */
77 size = min_t(unsigned int, size, tbl->size >> 1);
79 if (sizeof(spinlock_t) != 0) {
82 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
84 tbl->locks = vmalloc(size * sizeof(spinlock_t));
86 if (gfp != GFP_KERNEL)
87 gfp |= __GFP_NOWARN | __GFP_NORETRY;
90 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
94 for (i = 0; i < size; i++)
95 spin_lock_init(&tbl->locks[i]);
97 tbl->locks_mask = size - 1;
102 static void bucket_table_free(const struct bucket_table *tbl)
110 static void bucket_table_free_rcu(struct rcu_head *head)
112 bucket_table_free(container_of(head, struct bucket_table, rcu));
115 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
119 struct bucket_table *tbl = NULL;
123 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
124 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
126 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
127 if (tbl == NULL && gfp == GFP_KERNEL)
132 tbl->size = nbuckets;
134 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
135 bucket_table_free(tbl);
139 INIT_LIST_HEAD(&tbl->walkers);
141 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
143 for (i = 0; i < nbuckets; i++)
144 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
149 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
150 struct bucket_table *tbl)
152 struct bucket_table *new_tbl;
156 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
162 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
164 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
165 struct bucket_table *new_tbl = rhashtable_last_table(ht,
166 rht_dereference_rcu(old_tbl->future_tbl, ht));
167 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
169 struct rhash_head *head, *next, *entry;
170 spinlock_t *new_bucket_lock;
171 unsigned int new_hash;
173 rht_for_each(entry, old_tbl, old_hash) {
175 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
177 if (rht_is_a_nulls(next))
180 pprev = &entry->next;
186 new_hash = head_hashfn(ht, new_tbl, entry);
188 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
190 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
191 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
194 RCU_INIT_POINTER(entry->next, head);
196 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
197 spin_unlock(new_bucket_lock);
199 rcu_assign_pointer(*pprev, next);
205 static void rhashtable_rehash_chain(struct rhashtable *ht,
206 unsigned int old_hash)
208 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
209 spinlock_t *old_bucket_lock;
211 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
213 spin_lock_bh(old_bucket_lock);
214 while (!rhashtable_rehash_one(ht, old_hash))
217 spin_unlock_bh(old_bucket_lock);
220 static int rhashtable_rehash_attach(struct rhashtable *ht,
221 struct bucket_table *old_tbl,
222 struct bucket_table *new_tbl)
224 /* Protect future_tbl using the first bucket lock. */
225 spin_lock_bh(old_tbl->locks);
227 /* Did somebody beat us to it? */
228 if (rcu_access_pointer(old_tbl->future_tbl)) {
229 spin_unlock_bh(old_tbl->locks);
233 /* Make insertions go into the new, empty table right away. Deletions
234 * and lookups will be attempted in both tables until we synchronize.
236 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
238 spin_unlock_bh(old_tbl->locks);
243 static int rhashtable_rehash_table(struct rhashtable *ht)
245 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
246 struct bucket_table *new_tbl;
247 struct rhashtable_walker *walker;
248 unsigned int old_hash;
250 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
254 for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
255 rhashtable_rehash_chain(ht, old_hash);
257 /* Publish the new table pointer. */
258 rcu_assign_pointer(ht->tbl, new_tbl);
260 spin_lock(&ht->lock);
261 list_for_each_entry(walker, &old_tbl->walkers, list)
263 spin_unlock(&ht->lock);
265 /* Wait for readers. All new readers will see the new
266 * table, and thus no references to the old table will
269 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
271 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
275 * rhashtable_expand - Expand hash table while allowing concurrent lookups
276 * @ht: the hash table to expand
278 * A secondary bucket array is allocated and the hash entries are migrated.
280 * This function may only be called in a context where it is safe to call
281 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
283 * The caller must ensure that no concurrent resizing occurs by holding
286 * It is valid to have concurrent insertions and deletions protected by per
287 * bucket locks or concurrent RCU protected lookups and traversals.
289 static int rhashtable_expand(struct rhashtable *ht)
291 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
294 ASSERT_RHT_MUTEX(ht);
296 old_tbl = rhashtable_last_table(ht, old_tbl);
298 new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
302 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
304 bucket_table_free(new_tbl);
310 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
311 * @ht: the hash table to shrink
313 * This function shrinks the hash table to fit, i.e., the smallest
314 * size would not cause it to expand right away automatically.
316 * The caller must ensure that no concurrent resizing occurs by holding
319 * The caller must ensure that no concurrent table mutations take place.
320 * It is however valid to have concurrent lookups if they are RCU protected.
322 * It is valid to have concurrent insertions and deletions protected by per
323 * bucket locks or concurrent RCU protected lookups and traversals.
325 static int rhashtable_shrink(struct rhashtable *ht)
327 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
328 unsigned int nelems = atomic_read(&ht->nelems);
329 unsigned int size = 0;
332 ASSERT_RHT_MUTEX(ht);
335 size = roundup_pow_of_two(nelems * 3 / 2);
336 if (size < ht->p.min_size)
337 size = ht->p.min_size;
339 if (old_tbl->size <= size)
342 if (rht_dereference(old_tbl->future_tbl, ht))
345 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
349 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
351 bucket_table_free(new_tbl);
356 static void rht_deferred_worker(struct work_struct *work)
358 struct rhashtable *ht;
359 struct bucket_table *tbl;
362 ht = container_of(work, struct rhashtable, run_work);
363 mutex_lock(&ht->mutex);
365 tbl = rht_dereference(ht->tbl, ht);
366 tbl = rhashtable_last_table(ht, tbl);
368 if (rht_grow_above_75(ht, tbl))
369 rhashtable_expand(ht);
370 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
371 rhashtable_shrink(ht);
373 err = rhashtable_rehash_table(ht);
375 mutex_unlock(&ht->mutex);
378 schedule_work(&ht->run_work);
381 static bool rhashtable_check_elasticity(struct rhashtable *ht,
382 struct bucket_table *tbl,
385 unsigned int elasticity = ht->elasticity;
386 struct rhash_head *head;
388 rht_for_each(head, tbl, hash)
395 int rhashtable_insert_rehash(struct rhashtable *ht,
396 struct bucket_table *tbl)
398 struct bucket_table *old_tbl;
399 struct bucket_table *new_tbl;
403 old_tbl = rht_dereference_rcu(ht->tbl, ht);
409 if (rht_grow_above_75(ht, tbl))
411 /* Do not schedule more than one rehash */
412 else if (old_tbl != tbl)
417 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
421 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
423 bucket_table_free(new_tbl);
427 schedule_work(&ht->run_work);
432 /* Do not fail the insert if someone else did a rehash. */
433 if (likely(rcu_dereference_raw(tbl->future_tbl)))
436 /* Schedule async rehash to retry allocation in process context. */
438 schedule_work(&ht->run_work);
442 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
444 struct bucket_table *rhashtable_insert_slow(struct rhashtable *ht,
446 struct rhash_head *obj,
447 struct bucket_table *tbl)
449 struct rhash_head *head;
453 tbl = rhashtable_last_table(ht, tbl);
454 hash = head_hashfn(ht, tbl, obj);
455 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
458 if (key && rhashtable_lookup_fast(ht, key, ht->p))
462 if (unlikely(rht_grow_above_max(ht, tbl)))
466 if (rhashtable_check_elasticity(ht, tbl, hash) ||
467 rht_grow_above_100(ht, tbl))
472 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
474 RCU_INIT_POINTER(obj->next, head);
476 rcu_assign_pointer(tbl->buckets[hash], obj);
478 atomic_inc(&ht->nelems);
481 spin_unlock(rht_bucket_lock(tbl, hash));
485 else if (err == -EAGAIN)
490 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
493 * rhashtable_walk_enter - Initialise an iterator
494 * @ht: Table to walk over
495 * @iter: Hash table Iterator
497 * This function prepares a hash table walk.
499 * Note that if you restart a walk after rhashtable_walk_stop you
500 * may see the same object twice. Also, you may miss objects if
501 * there are removals in between rhashtable_walk_stop and the next
502 * call to rhashtable_walk_start.
504 * For a completely stable walk you should construct your own data
505 * structure outside the hash table.
507 * This function may sleep so you must not call it from interrupt
508 * context or with spin locks held.
510 * You must call rhashtable_walk_exit after this function returns.
512 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
519 spin_lock(&ht->lock);
521 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
522 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
523 spin_unlock(&ht->lock);
525 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
528 * rhashtable_walk_exit - Free an iterator
529 * @iter: Hash table Iterator
531 * This function frees resources allocated by rhashtable_walk_init.
533 void rhashtable_walk_exit(struct rhashtable_iter *iter)
535 spin_lock(&iter->ht->lock);
536 if (iter->walker.tbl)
537 list_del(&iter->walker.list);
538 spin_unlock(&iter->ht->lock);
540 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
543 * rhashtable_walk_start - Start a hash table walk
544 * @iter: Hash table iterator
546 * Start a hash table walk. Note that we take the RCU lock in all
547 * cases including when we return an error. So you must always call
548 * rhashtable_walk_stop to clean up.
550 * Returns zero if successful.
552 * Returns -EAGAIN if resize event occured. Note that the iterator
553 * will rewind back to the beginning and you may use it immediately
554 * by calling rhashtable_walk_next.
556 int rhashtable_walk_start(struct rhashtable_iter *iter)
559 struct rhashtable *ht = iter->ht;
563 spin_lock(&ht->lock);
564 if (iter->walker.tbl)
565 list_del(&iter->walker.list);
566 spin_unlock(&ht->lock);
568 if (!iter->walker.tbl) {
569 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
575 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
578 * rhashtable_walk_next - Return the next object and advance the iterator
579 * @iter: Hash table iterator
581 * Note that you must call rhashtable_walk_stop when you are finished
584 * Returns the next object or NULL when the end of the table is reached.
586 * Returns -EAGAIN if resize event occured. Note that the iterator
587 * will rewind back to the beginning and you may continue to use it.
589 void *rhashtable_walk_next(struct rhashtable_iter *iter)
591 struct bucket_table *tbl = iter->walker.tbl;
592 struct rhashtable *ht = iter->ht;
593 struct rhash_head *p = iter->p;
596 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
600 for (; iter->slot < tbl->size; iter->slot++) {
601 int skip = iter->skip;
603 rht_for_each_rcu(p, tbl, iter->slot) {
610 if (!rht_is_a_nulls(p)) {
613 return rht_obj(ht, p);
621 /* Ensure we see any new tables. */
624 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
625 if (iter->walker.tbl) {
628 return ERR_PTR(-EAGAIN);
633 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
636 * rhashtable_walk_stop - Finish a hash table walk
637 * @iter: Hash table iterator
639 * Finish a hash table walk.
641 void rhashtable_walk_stop(struct rhashtable_iter *iter)
644 struct rhashtable *ht;
645 struct bucket_table *tbl = iter->walker.tbl;
652 spin_lock(&ht->lock);
653 if (tbl->rehash < tbl->size)
654 list_add(&iter->walker.list, &tbl->walkers);
656 iter->walker.tbl = NULL;
657 spin_unlock(&ht->lock);
664 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
666 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
668 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
669 (unsigned long)params->min_size);
672 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
674 return jhash2(key, length, seed);
678 * rhashtable_init - initialize a new hash table
679 * @ht: hash table to be initialized
680 * @params: configuration parameters
682 * Initializes a new hash table based on the provided configuration
683 * parameters. A table can be configured either with a variable or
686 * Configuration Example 1: Fixed length keys
690 * struct rhash_head node;
693 * struct rhashtable_params params = {
694 * .head_offset = offsetof(struct test_obj, node),
695 * .key_offset = offsetof(struct test_obj, key),
696 * .key_len = sizeof(int),
698 * .nulls_base = (1U << RHT_BASE_SHIFT),
701 * Configuration Example 2: Variable length keys
704 * struct rhash_head node;
707 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
709 * struct test_obj *obj = data;
711 * return [... hash ...];
714 * struct rhashtable_params params = {
715 * .head_offset = offsetof(struct test_obj, node),
717 * .obj_hashfn = my_hash_fn,
720 int rhashtable_init(struct rhashtable *ht,
721 const struct rhashtable_params *params)
723 struct bucket_table *tbl;
726 size = HASH_DEFAULT_SIZE;
728 if ((!params->key_len && !params->obj_hashfn) ||
729 (params->obj_hashfn && !params->obj_cmpfn))
732 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
735 memset(ht, 0, sizeof(*ht));
736 mutex_init(&ht->mutex);
737 spin_lock_init(&ht->lock);
738 memcpy(&ht->p, params, sizeof(*params));
740 if (params->min_size)
741 ht->p.min_size = roundup_pow_of_two(params->min_size);
743 if (params->max_size)
744 ht->p.max_size = rounddown_pow_of_two(params->max_size);
746 if (params->insecure_max_entries)
747 ht->p.insecure_max_entries =
748 rounddown_pow_of_two(params->insecure_max_entries);
750 ht->p.insecure_max_entries = ht->p.max_size * 2;
752 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
754 if (params->nelem_hint)
755 size = rounded_hashtable_size(&ht->p);
757 /* The maximum (not average) chain length grows with the
758 * size of the hash table, at a rate of (log N)/(log log N).
759 * The value of 16 is selected so that even if the hash
760 * table grew to 2^32 you would not expect the maximum
761 * chain length to exceed it unless we are under attack
762 * (or extremely unlucky).
764 * As this limit is only to detect attacks, we don't need
765 * to set it to a lower value as you'd need the chain
766 * length to vastly exceed 16 to have any real effect
769 if (!params->insecure_elasticity)
772 if (params->locks_mul)
773 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
775 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
777 ht->key_len = ht->p.key_len;
778 if (!params->hashfn) {
779 ht->p.hashfn = jhash;
781 if (!(ht->key_len & (sizeof(u32) - 1))) {
782 ht->key_len /= sizeof(u32);
783 ht->p.hashfn = rhashtable_jhash2;
787 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
791 atomic_set(&ht->nelems, 0);
793 RCU_INIT_POINTER(ht->tbl, tbl);
795 INIT_WORK(&ht->run_work, rht_deferred_worker);
799 EXPORT_SYMBOL_GPL(rhashtable_init);
802 * rhashtable_free_and_destroy - free elements and destroy hash table
803 * @ht: the hash table to destroy
804 * @free_fn: callback to release resources of element
805 * @arg: pointer passed to free_fn
807 * Stops an eventual async resize. If defined, invokes free_fn for each
808 * element to releasal resources. Please note that RCU protected
809 * readers may still be accessing the elements. Releasing of resources
810 * must occur in a compatible manner. Then frees the bucket array.
812 * This function will eventually sleep to wait for an async resize
813 * to complete. The caller is responsible that no further write operations
814 * occurs in parallel.
816 void rhashtable_free_and_destroy(struct rhashtable *ht,
817 void (*free_fn)(void *ptr, void *arg),
820 const struct bucket_table *tbl;
823 cancel_work_sync(&ht->run_work);
825 mutex_lock(&ht->mutex);
826 tbl = rht_dereference(ht->tbl, ht);
828 for (i = 0; i < tbl->size; i++) {
829 struct rhash_head *pos, *next;
831 for (pos = rht_dereference(tbl->buckets[i], ht),
832 next = !rht_is_a_nulls(pos) ?
833 rht_dereference(pos->next, ht) : NULL;
834 !rht_is_a_nulls(pos);
836 next = !rht_is_a_nulls(pos) ?
837 rht_dereference(pos->next, ht) : NULL)
838 free_fn(rht_obj(ht, pos), arg);
842 bucket_table_free(tbl);
843 mutex_unlock(&ht->mutex);
845 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
847 void rhashtable_destroy(struct rhashtable *ht)
849 return rhashtable_free_and_destroy(ht, NULL, NULL);
851 EXPORT_SYMBOL_GPL(rhashtable_destroy);