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 128UL
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, 32UL);
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) {
81 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
83 tbl->locks = vmalloc(size * sizeof(spinlock_t));
86 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
90 for (i = 0; i < size; i++)
91 spin_lock_init(&tbl->locks[i]);
93 tbl->locks_mask = size - 1;
98 static void bucket_table_free(const struct bucket_table *tbl)
106 static void bucket_table_free_rcu(struct rcu_head *head)
108 bucket_table_free(container_of(head, struct bucket_table, rcu));
111 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
115 struct bucket_table *tbl = NULL;
119 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
120 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
122 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
123 if (tbl == NULL && gfp == GFP_KERNEL)
128 tbl->size = nbuckets;
130 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
131 bucket_table_free(tbl);
135 INIT_LIST_HEAD(&tbl->walkers);
137 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
139 for (i = 0; i < nbuckets; i++)
140 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
145 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
146 struct bucket_table *tbl)
148 struct bucket_table *new_tbl;
152 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
158 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
160 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
161 struct bucket_table *new_tbl = rhashtable_last_table(ht,
162 rht_dereference_rcu(old_tbl->future_tbl, ht));
163 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
165 struct rhash_head *head, *next, *entry;
166 spinlock_t *new_bucket_lock;
167 unsigned int new_hash;
169 rht_for_each(entry, old_tbl, old_hash) {
171 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
173 if (rht_is_a_nulls(next))
176 pprev = &entry->next;
182 new_hash = head_hashfn(ht, new_tbl, entry);
184 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
186 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
187 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
190 if (rht_is_a_nulls(head))
191 INIT_RHT_NULLS_HEAD(entry->next, ht, new_hash);
193 RCU_INIT_POINTER(entry->next, head);
195 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
196 spin_unlock(new_bucket_lock);
198 rcu_assign_pointer(*pprev, next);
204 static void rhashtable_rehash_chain(struct rhashtable *ht,
205 unsigned int old_hash)
207 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
208 spinlock_t *old_bucket_lock;
210 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
212 spin_lock_bh(old_bucket_lock);
213 while (!rhashtable_rehash_one(ht, old_hash))
216 spin_unlock_bh(old_bucket_lock);
219 static int rhashtable_rehash_attach(struct rhashtable *ht,
220 struct bucket_table *old_tbl,
221 struct bucket_table *new_tbl)
223 /* Protect future_tbl using the first bucket lock. */
224 spin_lock_bh(old_tbl->locks);
226 /* Did somebody beat us to it? */
227 if (rcu_access_pointer(old_tbl->future_tbl)) {
228 spin_unlock_bh(old_tbl->locks);
232 /* Make insertions go into the new, empty table right away. Deletions
233 * and lookups will be attempted in both tables until we synchronize.
235 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
237 /* Ensure the new table is visible to readers. */
240 spin_unlock_bh(old_tbl->locks);
245 static int rhashtable_rehash_table(struct rhashtable *ht)
247 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
248 struct bucket_table *new_tbl;
249 struct rhashtable_walker *walker;
250 unsigned int old_hash;
252 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
256 for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
257 rhashtable_rehash_chain(ht, old_hash);
259 /* Publish the new table pointer. */
260 rcu_assign_pointer(ht->tbl, new_tbl);
262 spin_lock(&ht->lock);
263 list_for_each_entry(walker, &old_tbl->walkers, list)
265 spin_unlock(&ht->lock);
267 /* Wait for readers. All new readers will see the new
268 * table, and thus no references to the old table will
271 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
273 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
277 * rhashtable_expand - Expand hash table while allowing concurrent lookups
278 * @ht: the hash table to expand
280 * A secondary bucket array is allocated and the hash entries are migrated.
282 * This function may only be called in a context where it is safe to call
283 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
285 * The caller must ensure that no concurrent resizing occurs by holding
288 * It is valid to have concurrent insertions and deletions protected by per
289 * bucket locks or concurrent RCU protected lookups and traversals.
291 static int rhashtable_expand(struct rhashtable *ht)
293 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
296 ASSERT_RHT_MUTEX(ht);
298 old_tbl = rhashtable_last_table(ht, old_tbl);
300 new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
304 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
306 bucket_table_free(new_tbl);
312 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
313 * @ht: the hash table to shrink
315 * This function shrinks the hash table to fit, i.e., the smallest
316 * size would not cause it to expand right away automatically.
318 * The caller must ensure that no concurrent resizing occurs by holding
321 * The caller must ensure that no concurrent table mutations take place.
322 * It is however valid to have concurrent lookups if they are RCU protected.
324 * It is valid to have concurrent insertions and deletions protected by per
325 * bucket locks or concurrent RCU protected lookups and traversals.
327 static int rhashtable_shrink(struct rhashtable *ht)
329 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
333 ASSERT_RHT_MUTEX(ht);
335 size = roundup_pow_of_two(atomic_read(&ht->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)
397 struct bucket_table *old_tbl;
398 struct bucket_table *new_tbl;
399 struct bucket_table *tbl;
403 old_tbl = rht_dereference_rcu(ht->tbl, ht);
404 tbl = rhashtable_last_table(ht, old_tbl);
408 if (rht_grow_above_75(ht, tbl))
410 /* Do not schedule more than one rehash */
411 else if (old_tbl != tbl)
414 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
415 if (new_tbl == NULL) {
416 /* Schedule async resize/rehash to try allocation
417 * non-atomic context.
419 schedule_work(&ht->run_work);
423 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
425 bucket_table_free(new_tbl);
429 schedule_work(&ht->run_work);
433 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
435 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
436 struct rhash_head *obj,
437 struct bucket_table *tbl)
439 struct rhash_head *head;
443 tbl = rhashtable_last_table(ht, tbl);
444 hash = head_hashfn(ht, tbl, obj);
445 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
448 if (key && rhashtable_lookup_fast(ht, key, ht->p))
452 if (unlikely(rht_grow_above_max(ht, tbl)))
456 if (rhashtable_check_elasticity(ht, tbl, hash) ||
457 rht_grow_above_100(ht, tbl))
462 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
464 RCU_INIT_POINTER(obj->next, head);
466 rcu_assign_pointer(tbl->buckets[hash], obj);
468 atomic_inc(&ht->nelems);
471 spin_unlock(rht_bucket_lock(tbl, hash));
475 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
478 * rhashtable_walk_init - Initialise an iterator
479 * @ht: Table to walk over
480 * @iter: Hash table Iterator
482 * This function prepares a hash table walk.
484 * Note that if you restart a walk after rhashtable_walk_stop you
485 * may see the same object twice. Also, you may miss objects if
486 * there are removals in between rhashtable_walk_stop and the next
487 * call to rhashtable_walk_start.
489 * For a completely stable walk you should construct your own data
490 * structure outside the hash table.
492 * This function may sleep so you must not call it from interrupt
493 * context or with spin locks held.
495 * You must call rhashtable_walk_exit if this function returns
498 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
505 iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
509 mutex_lock(&ht->mutex);
510 iter->walker->tbl = rht_dereference(ht->tbl, ht);
511 list_add(&iter->walker->list, &iter->walker->tbl->walkers);
512 mutex_unlock(&ht->mutex);
516 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
519 * rhashtable_walk_exit - Free an iterator
520 * @iter: Hash table Iterator
522 * This function frees resources allocated by rhashtable_walk_init.
524 void rhashtable_walk_exit(struct rhashtable_iter *iter)
526 mutex_lock(&iter->ht->mutex);
527 if (iter->walker->tbl)
528 list_del(&iter->walker->list);
529 mutex_unlock(&iter->ht->mutex);
532 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
535 * rhashtable_walk_start - Start a hash table walk
536 * @iter: Hash table iterator
538 * Start a hash table walk. Note that we take the RCU lock in all
539 * cases including when we return an error. So you must always call
540 * rhashtable_walk_stop to clean up.
542 * Returns zero if successful.
544 * Returns -EAGAIN if resize event occured. Note that the iterator
545 * will rewind back to the beginning and you may use it immediately
546 * by calling rhashtable_walk_next.
548 int rhashtable_walk_start(struct rhashtable_iter *iter)
551 struct rhashtable *ht = iter->ht;
553 mutex_lock(&ht->mutex);
555 if (iter->walker->tbl)
556 list_del(&iter->walker->list);
560 mutex_unlock(&ht->mutex);
562 if (!iter->walker->tbl) {
563 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
569 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
572 * rhashtable_walk_next - Return the next object and advance the iterator
573 * @iter: Hash table iterator
575 * Note that you must call rhashtable_walk_stop when you are finished
578 * Returns the next object or NULL when the end of the table is reached.
580 * Returns -EAGAIN if resize event occured. Note that the iterator
581 * will rewind back to the beginning and you may continue to use it.
583 void *rhashtable_walk_next(struct rhashtable_iter *iter)
585 struct bucket_table *tbl = iter->walker->tbl;
586 struct rhashtable *ht = iter->ht;
587 struct rhash_head *p = iter->p;
590 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
594 for (; iter->slot < tbl->size; iter->slot++) {
595 int skip = iter->skip;
597 rht_for_each_rcu(p, tbl, iter->slot) {
604 if (!rht_is_a_nulls(p)) {
607 return rht_obj(ht, p);
615 /* Ensure we see any new tables. */
618 iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
619 if (iter->walker->tbl) {
622 return ERR_PTR(-EAGAIN);
627 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
630 * rhashtable_walk_stop - Finish a hash table walk
631 * @iter: Hash table iterator
633 * Finish a hash table walk.
635 void rhashtable_walk_stop(struct rhashtable_iter *iter)
638 struct rhashtable *ht;
639 struct bucket_table *tbl = iter->walker->tbl;
646 spin_lock(&ht->lock);
647 if (tbl->rehash < tbl->size)
648 list_add(&iter->walker->list, &tbl->walkers);
650 iter->walker->tbl = NULL;
651 spin_unlock(&ht->lock);
658 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
660 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
662 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
663 (unsigned long)params->min_size);
666 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
668 return jhash2(key, length, seed);
672 * rhashtable_init - initialize a new hash table
673 * @ht: hash table to be initialized
674 * @params: configuration parameters
676 * Initializes a new hash table based on the provided configuration
677 * parameters. A table can be configured either with a variable or
680 * Configuration Example 1: Fixed length keys
684 * struct rhash_head node;
687 * struct rhashtable_params params = {
688 * .head_offset = offsetof(struct test_obj, node),
689 * .key_offset = offsetof(struct test_obj, key),
690 * .key_len = sizeof(int),
692 * .nulls_base = (1U << RHT_BASE_SHIFT),
695 * Configuration Example 2: Variable length keys
698 * struct rhash_head node;
701 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
703 * struct test_obj *obj = data;
705 * return [... hash ...];
708 * struct rhashtable_params params = {
709 * .head_offset = offsetof(struct test_obj, node),
711 * .obj_hashfn = my_hash_fn,
714 int rhashtable_init(struct rhashtable *ht,
715 const struct rhashtable_params *params)
717 struct bucket_table *tbl;
720 size = HASH_DEFAULT_SIZE;
722 if ((!params->key_len && !params->obj_hashfn) ||
723 (params->obj_hashfn && !params->obj_cmpfn))
726 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
729 if (params->nelem_hint)
730 size = rounded_hashtable_size(params);
732 memset(ht, 0, sizeof(*ht));
733 mutex_init(&ht->mutex);
734 spin_lock_init(&ht->lock);
735 memcpy(&ht->p, params, sizeof(*params));
737 if (params->min_size)
738 ht->p.min_size = roundup_pow_of_two(params->min_size);
740 if (params->max_size)
741 ht->p.max_size = rounddown_pow_of_two(params->max_size);
743 if (params->insecure_max_entries)
744 ht->p.insecure_max_entries =
745 rounddown_pow_of_two(params->insecure_max_entries);
747 ht->p.insecure_max_entries = ht->p.max_size * 2;
749 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
751 /* The maximum (not average) chain length grows with the
752 * size of the hash table, at a rate of (log N)/(log log N).
753 * The value of 16 is selected so that even if the hash
754 * table grew to 2^32 you would not expect the maximum
755 * chain length to exceed it unless we are under attack
756 * (or extremely unlucky).
758 * As this limit is only to detect attacks, we don't need
759 * to set it to a lower value as you'd need the chain
760 * length to vastly exceed 16 to have any real effect
763 if (!params->insecure_elasticity)
766 if (params->locks_mul)
767 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
769 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
771 ht->key_len = ht->p.key_len;
772 if (!params->hashfn) {
773 ht->p.hashfn = jhash;
775 if (!(ht->key_len & (sizeof(u32) - 1))) {
776 ht->key_len /= sizeof(u32);
777 ht->p.hashfn = rhashtable_jhash2;
781 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
785 atomic_set(&ht->nelems, 0);
787 RCU_INIT_POINTER(ht->tbl, tbl);
789 INIT_WORK(&ht->run_work, rht_deferred_worker);
793 EXPORT_SYMBOL_GPL(rhashtable_init);
796 * rhashtable_free_and_destroy - free elements and destroy hash table
797 * @ht: the hash table to destroy
798 * @free_fn: callback to release resources of element
799 * @arg: pointer passed to free_fn
801 * Stops an eventual async resize. If defined, invokes free_fn for each
802 * element to releasal resources. Please note that RCU protected
803 * readers may still be accessing the elements. Releasing of resources
804 * must occur in a compatible manner. Then frees the bucket array.
806 * This function will eventually sleep to wait for an async resize
807 * to complete. The caller is responsible that no further write operations
808 * occurs in parallel.
810 void rhashtable_free_and_destroy(struct rhashtable *ht,
811 void (*free_fn)(void *ptr, void *arg),
814 const struct bucket_table *tbl;
817 cancel_work_sync(&ht->run_work);
819 mutex_lock(&ht->mutex);
820 tbl = rht_dereference(ht->tbl, ht);
822 for (i = 0; i < tbl->size; i++) {
823 struct rhash_head *pos, *next;
825 for (pos = rht_dereference(tbl->buckets[i], ht),
826 next = !rht_is_a_nulls(pos) ?
827 rht_dereference(pos->next, ht) : NULL;
828 !rht_is_a_nulls(pos);
830 next = !rht_is_a_nulls(pos) ?
831 rht_dereference(pos->next, ht) : NULL)
832 free_fn(rht_obj(ht, pos), arg);
836 bucket_table_free(tbl);
837 mutex_unlock(&ht->mutex);
839 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
841 void rhashtable_destroy(struct rhashtable *ht)
843 return rhashtable_free_and_destroy(ht, NULL, NULL);
845 EXPORT_SYMBOL_GPL(rhashtable_destroy);