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);
124 tbl = __vmalloc(size, gfp | __GFP_HIGHMEM | __GFP_ZERO,
129 tbl->size = nbuckets;
131 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
132 bucket_table_free(tbl);
136 INIT_LIST_HEAD(&tbl->walkers);
138 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
140 for (i = 0; i < nbuckets; i++)
141 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
146 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
147 struct bucket_table *tbl)
149 struct bucket_table *new_tbl;
153 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
159 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
161 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
162 struct bucket_table *new_tbl = rhashtable_last_table(ht,
163 rht_dereference_rcu(old_tbl->future_tbl, ht));
164 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
166 struct rhash_head *head, *next, *entry;
167 spinlock_t *new_bucket_lock;
168 unsigned int new_hash;
170 rht_for_each(entry, old_tbl, old_hash) {
172 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
174 if (rht_is_a_nulls(next))
177 pprev = &entry->next;
183 new_hash = head_hashfn(ht, new_tbl, entry);
185 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
187 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
188 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
191 RCU_INIT_POINTER(entry->next, head);
193 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
194 spin_unlock(new_bucket_lock);
196 rcu_assign_pointer(*pprev, next);
202 static void rhashtable_rehash_chain(struct rhashtable *ht,
203 unsigned int old_hash)
205 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
206 spinlock_t *old_bucket_lock;
208 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
210 spin_lock_bh(old_bucket_lock);
211 while (!rhashtable_rehash_one(ht, old_hash))
214 spin_unlock_bh(old_bucket_lock);
217 static int rhashtable_rehash_attach(struct rhashtable *ht,
218 struct bucket_table *old_tbl,
219 struct bucket_table *new_tbl)
221 /* Protect future_tbl using the first bucket lock. */
222 spin_lock_bh(old_tbl->locks);
224 /* Did somebody beat us to it? */
225 if (rcu_access_pointer(old_tbl->future_tbl)) {
226 spin_unlock_bh(old_tbl->locks);
230 /* Make insertions go into the new, empty table right away. Deletions
231 * and lookups will be attempted in both tables until we synchronize.
233 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
235 /* Ensure the new table is visible to readers. */
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);
331 ASSERT_RHT_MUTEX(ht);
333 size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
334 if (size < ht->p.min_size)
335 size = ht->p.min_size;
337 if (old_tbl->size <= size)
340 if (rht_dereference(old_tbl->future_tbl, ht))
343 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
347 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
349 bucket_table_free(new_tbl);
354 static void rht_deferred_worker(struct work_struct *work)
356 struct rhashtable *ht;
357 struct bucket_table *tbl;
360 ht = container_of(work, struct rhashtable, run_work);
361 mutex_lock(&ht->mutex);
363 tbl = rht_dereference(ht->tbl, ht);
364 tbl = rhashtable_last_table(ht, tbl);
366 if (rht_grow_above_75(ht, tbl))
367 rhashtable_expand(ht);
368 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
369 rhashtable_shrink(ht);
371 err = rhashtable_rehash_table(ht);
373 mutex_unlock(&ht->mutex);
376 schedule_work(&ht->run_work);
379 static bool rhashtable_check_elasticity(struct rhashtable *ht,
380 struct bucket_table *tbl,
383 unsigned int elasticity = ht->elasticity;
384 struct rhash_head *head;
386 rht_for_each(head, tbl, hash)
393 int rhashtable_insert_rehash(struct rhashtable *ht,
394 struct bucket_table *tbl)
396 struct bucket_table *old_tbl;
397 struct bucket_table *new_tbl;
401 old_tbl = rht_dereference_rcu(ht->tbl, ht);
407 if (rht_grow_above_75(ht, tbl))
409 /* Do not schedule more than one rehash */
410 else if (old_tbl != tbl)
415 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
419 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
421 bucket_table_free(new_tbl);
425 schedule_work(&ht->run_work);
430 /* Do not fail the insert if someone else did a rehash. */
431 if (likely(rcu_dereference_raw(tbl->future_tbl)))
434 /* Schedule async rehash to retry allocation in process context. */
436 schedule_work(&ht->run_work);
440 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
442 struct bucket_table *rhashtable_insert_slow(struct rhashtable *ht,
444 struct rhash_head *obj,
445 struct bucket_table *tbl)
447 struct rhash_head *head;
451 tbl = rhashtable_last_table(ht, tbl);
452 hash = head_hashfn(ht, tbl, obj);
453 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
456 if (key && rhashtable_lookup_fast(ht, key, ht->p))
460 if (unlikely(rht_grow_above_max(ht, tbl)))
464 if (rhashtable_check_elasticity(ht, tbl, hash) ||
465 rht_grow_above_100(ht, tbl))
470 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
472 RCU_INIT_POINTER(obj->next, head);
474 rcu_assign_pointer(tbl->buckets[hash], obj);
476 atomic_inc(&ht->nelems);
479 spin_unlock(rht_bucket_lock(tbl, hash));
483 else if (err == -EAGAIN)
488 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
491 * rhashtable_walk_init - Initialise an iterator
492 * @ht: Table to walk over
493 * @iter: Hash table Iterator
495 * This function prepares a hash table walk.
497 * Note that if you restart a walk after rhashtable_walk_stop you
498 * may see the same object twice. Also, you may miss objects if
499 * there are removals in between rhashtable_walk_stop and the next
500 * call to rhashtable_walk_start.
502 * For a completely stable walk you should construct your own data
503 * structure outside the hash table.
505 * This function may sleep so you must not call it from interrupt
506 * context or with spin locks held.
508 * You must call rhashtable_walk_exit if this function returns
511 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
518 iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
522 mutex_lock(&ht->mutex);
523 iter->walker->tbl = rht_dereference(ht->tbl, ht);
524 list_add(&iter->walker->list, &iter->walker->tbl->walkers);
525 mutex_unlock(&ht->mutex);
529 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
532 * rhashtable_walk_exit - Free an iterator
533 * @iter: Hash table Iterator
535 * This function frees resources allocated by rhashtable_walk_init.
537 void rhashtable_walk_exit(struct rhashtable_iter *iter)
539 mutex_lock(&iter->ht->mutex);
540 if (iter->walker->tbl)
541 list_del(&iter->walker->list);
542 mutex_unlock(&iter->ht->mutex);
545 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
548 * rhashtable_walk_start - Start a hash table walk
549 * @iter: Hash table iterator
551 * Start a hash table walk. Note that we take the RCU lock in all
552 * cases including when we return an error. So you must always call
553 * rhashtable_walk_stop to clean up.
555 * Returns zero if successful.
557 * Returns -EAGAIN if resize event occured. Note that the iterator
558 * will rewind back to the beginning and you may use it immediately
559 * by calling rhashtable_walk_next.
561 int rhashtable_walk_start(struct rhashtable_iter *iter)
564 struct rhashtable *ht = iter->ht;
566 mutex_lock(&ht->mutex);
568 if (iter->walker->tbl)
569 list_del(&iter->walker->list);
573 mutex_unlock(&ht->mutex);
575 if (!iter->walker->tbl) {
576 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
582 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
585 * rhashtable_walk_next - Return the next object and advance the iterator
586 * @iter: Hash table iterator
588 * Note that you must call rhashtable_walk_stop when you are finished
591 * Returns the next object or NULL when the end of the table is reached.
593 * Returns -EAGAIN if resize event occured. Note that the iterator
594 * will rewind back to the beginning and you may continue to use it.
596 void *rhashtable_walk_next(struct rhashtable_iter *iter)
598 struct bucket_table *tbl = iter->walker->tbl;
599 struct rhashtable *ht = iter->ht;
600 struct rhash_head *p = iter->p;
603 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
607 for (; iter->slot < tbl->size; iter->slot++) {
608 int skip = iter->skip;
610 rht_for_each_rcu(p, tbl, iter->slot) {
617 if (!rht_is_a_nulls(p)) {
620 return rht_obj(ht, p);
628 /* Ensure we see any new tables. */
631 iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
632 if (iter->walker->tbl) {
635 return ERR_PTR(-EAGAIN);
640 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
643 * rhashtable_walk_stop - Finish a hash table walk
644 * @iter: Hash table iterator
646 * Finish a hash table walk.
648 void rhashtable_walk_stop(struct rhashtable_iter *iter)
651 struct rhashtable *ht;
652 struct bucket_table *tbl = iter->walker->tbl;
659 spin_lock(&ht->lock);
660 if (tbl->rehash < tbl->size)
661 list_add(&iter->walker->list, &tbl->walkers);
663 iter->walker->tbl = NULL;
664 spin_unlock(&ht->lock);
671 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
673 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
675 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
676 (unsigned long)params->min_size);
679 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
681 return jhash2(key, length, seed);
685 * rhashtable_init - initialize a new hash table
686 * @ht: hash table to be initialized
687 * @params: configuration parameters
689 * Initializes a new hash table based on the provided configuration
690 * parameters. A table can be configured either with a variable or
693 * Configuration Example 1: Fixed length keys
697 * struct rhash_head node;
700 * struct rhashtable_params params = {
701 * .head_offset = offsetof(struct test_obj, node),
702 * .key_offset = offsetof(struct test_obj, key),
703 * .key_len = sizeof(int),
705 * .nulls_base = (1U << RHT_BASE_SHIFT),
708 * Configuration Example 2: Variable length keys
711 * struct rhash_head node;
714 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
716 * struct test_obj *obj = data;
718 * return [... hash ...];
721 * struct rhashtable_params params = {
722 * .head_offset = offsetof(struct test_obj, node),
724 * .obj_hashfn = my_hash_fn,
727 int rhashtable_init(struct rhashtable *ht,
728 const struct rhashtable_params *params)
730 struct bucket_table *tbl;
733 size = HASH_DEFAULT_SIZE;
735 if ((!params->key_len && !params->obj_hashfn) ||
736 (params->obj_hashfn && !params->obj_cmpfn))
739 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
742 if (params->nelem_hint)
743 size = rounded_hashtable_size(params);
745 memset(ht, 0, sizeof(*ht));
746 mutex_init(&ht->mutex);
747 spin_lock_init(&ht->lock);
748 memcpy(&ht->p, params, sizeof(*params));
750 if (params->min_size)
751 ht->p.min_size = roundup_pow_of_two(params->min_size);
753 if (params->max_size)
754 ht->p.max_size = rounddown_pow_of_two(params->max_size);
756 if (params->insecure_max_entries)
757 ht->p.insecure_max_entries =
758 rounddown_pow_of_two(params->insecure_max_entries);
760 ht->p.insecure_max_entries = ht->p.max_size * 2;
762 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
764 /* The maximum (not average) chain length grows with the
765 * size of the hash table, at a rate of (log N)/(log log N).
766 * The value of 16 is selected so that even if the hash
767 * table grew to 2^32 you would not expect the maximum
768 * chain length to exceed it unless we are under attack
769 * (or extremely unlucky).
771 * As this limit is only to detect attacks, we don't need
772 * to set it to a lower value as you'd need the chain
773 * length to vastly exceed 16 to have any real effect
776 if (!params->insecure_elasticity)
779 if (params->locks_mul)
780 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
782 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
784 ht->key_len = ht->p.key_len;
785 if (!params->hashfn) {
786 ht->p.hashfn = jhash;
788 if (!(ht->key_len & (sizeof(u32) - 1))) {
789 ht->key_len /= sizeof(u32);
790 ht->p.hashfn = rhashtable_jhash2;
794 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
798 atomic_set(&ht->nelems, 0);
800 RCU_INIT_POINTER(ht->tbl, tbl);
802 INIT_WORK(&ht->run_work, rht_deferred_worker);
806 EXPORT_SYMBOL_GPL(rhashtable_init);
809 * rhashtable_free_and_destroy - free elements and destroy hash table
810 * @ht: the hash table to destroy
811 * @free_fn: callback to release resources of element
812 * @arg: pointer passed to free_fn
814 * Stops an eventual async resize. If defined, invokes free_fn for each
815 * element to releasal resources. Please note that RCU protected
816 * readers may still be accessing the elements. Releasing of resources
817 * must occur in a compatible manner. Then frees the bucket array.
819 * This function will eventually sleep to wait for an async resize
820 * to complete. The caller is responsible that no further write operations
821 * occurs in parallel.
823 void rhashtable_free_and_destroy(struct rhashtable *ht,
824 void (*free_fn)(void *ptr, void *arg),
827 const struct bucket_table *tbl;
830 cancel_work_sync(&ht->run_work);
832 mutex_lock(&ht->mutex);
833 tbl = rht_dereference(ht->tbl, ht);
835 for (i = 0; i < tbl->size; i++) {
836 struct rhash_head *pos, *next;
838 for (pos = rht_dereference(tbl->buckets[i], ht),
839 next = !rht_is_a_nulls(pos) ?
840 rht_dereference(pos->next, ht) : NULL;
841 !rht_is_a_nulls(pos);
843 next = !rht_is_a_nulls(pos) ?
844 rht_dereference(pos->next, ht) : NULL)
845 free_fn(rht_obj(ht, pos), arg);
849 bucket_table_free(tbl);
850 mutex_unlock(&ht->mutex);
852 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
854 void rhashtable_destroy(struct rhashtable *ht)
856 return rhashtable_free_and_destroy(ht, NULL, NULL);
858 EXPORT_SYMBOL_GPL(rhashtable_destroy);