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/kernel.h>
18 #include <linux/init.h>
19 #include <linux/log2.h>
20 #include <linux/sched.h>
21 #include <linux/slab.h>
22 #include <linux/vmalloc.h>
24 #include <linux/jhash.h>
25 #include <linux/random.h>
26 #include <linux/rhashtable.h>
27 #include <linux/err.h>
29 #define HASH_DEFAULT_SIZE 64UL
30 #define HASH_MIN_SIZE 4U
31 #define BUCKET_LOCKS_PER_CPU 128UL
33 static u32 head_hashfn(struct rhashtable *ht,
34 const struct bucket_table *tbl,
35 const struct rhash_head *he)
37 return rht_head_hashfn(ht, tbl, he, ht->p);
40 #ifdef CONFIG_PROVE_LOCKING
41 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
43 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
45 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
47 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
49 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
51 spinlock_t *lock = rht_bucket_lock(tbl, hash);
53 return (debug_locks) ? lockdep_is_held(lock) : 1;
55 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
57 #define ASSERT_RHT_MUTEX(HT)
61 static int alloc_bucket_locks(struct rhashtable *ht, struct bucket_table *tbl,
65 #if defined(CONFIG_PROVE_LOCKING)
66 unsigned int nr_pcpus = 2;
68 unsigned int nr_pcpus = num_possible_cpus();
71 nr_pcpus = min_t(unsigned int, nr_pcpus, 32UL);
72 size = roundup_pow_of_two(nr_pcpus * ht->p.locks_mul);
74 /* Never allocate more than 0.5 locks per bucket */
75 size = min_t(unsigned int, size, tbl->size >> 1);
77 if (sizeof(spinlock_t) != 0) {
79 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
81 tbl->locks = vmalloc(size * sizeof(spinlock_t));
84 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
88 for (i = 0; i < size; i++)
89 spin_lock_init(&tbl->locks[i]);
91 tbl->locks_mask = size - 1;
96 static void bucket_table_free(const struct bucket_table *tbl)
104 static void bucket_table_free_rcu(struct rcu_head *head)
106 bucket_table_free(container_of(head, struct bucket_table, rcu));
109 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
113 struct bucket_table *tbl = NULL;
117 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
118 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
120 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
121 if (tbl == NULL && gfp == GFP_KERNEL)
126 tbl->size = nbuckets;
128 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
129 bucket_table_free(tbl);
133 INIT_LIST_HEAD(&tbl->walkers);
135 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
137 for (i = 0; i < nbuckets; i++)
138 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
143 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
144 struct bucket_table *tbl)
146 struct bucket_table *new_tbl;
150 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
156 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned old_hash)
158 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
159 struct bucket_table *new_tbl = rhashtable_last_table(ht,
160 rht_dereference_rcu(old_tbl->future_tbl, ht));
161 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
163 struct rhash_head *head, *next, *entry;
164 spinlock_t *new_bucket_lock;
167 rht_for_each(entry, old_tbl, old_hash) {
169 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
171 if (rht_is_a_nulls(next))
174 pprev = &entry->next;
180 new_hash = head_hashfn(ht, new_tbl, entry);
182 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
184 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
185 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
188 if (rht_is_a_nulls(head))
189 INIT_RHT_NULLS_HEAD(entry->next, ht, 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, unsigned old_hash)
204 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
205 spinlock_t *old_bucket_lock;
207 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
209 spin_lock_bh(old_bucket_lock);
210 while (!rhashtable_rehash_one(ht, old_hash))
213 spin_unlock_bh(old_bucket_lock);
216 static int rhashtable_rehash_attach(struct rhashtable *ht,
217 struct bucket_table *old_tbl,
218 struct bucket_table *new_tbl)
220 /* Protect future_tbl using the first bucket lock. */
221 spin_lock_bh(old_tbl->locks);
223 /* Did somebody beat us to it? */
224 if (rcu_access_pointer(old_tbl->future_tbl)) {
225 spin_unlock_bh(old_tbl->locks);
229 /* Make insertions go into the new, empty table right away. Deletions
230 * and lookups will be attempted in both tables until we synchronize.
232 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
234 /* Ensure the new table is visible to readers. */
237 spin_unlock_bh(old_tbl->locks);
242 static int rhashtable_rehash_table(struct rhashtable *ht)
244 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
245 struct bucket_table *new_tbl;
246 struct rhashtable_walker *walker;
249 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
253 for (old_hash = 0; old_hash < old_tbl->size; old_hash++)
254 rhashtable_rehash_chain(ht, old_hash);
256 /* Publish the new table pointer. */
257 rcu_assign_pointer(ht->tbl, new_tbl);
259 spin_lock(&ht->lock);
260 list_for_each_entry(walker, &old_tbl->walkers, list)
262 spin_unlock(&ht->lock);
264 /* Wait for readers. All new readers will see the new
265 * table, and thus no references to the old table will
268 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
270 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
274 * rhashtable_expand - Expand hash table while allowing concurrent lookups
275 * @ht: the hash table to expand
277 * A secondary bucket array is allocated and the hash entries are migrated.
279 * This function may only be called in a context where it is safe to call
280 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
282 * The caller must ensure that no concurrent resizing occurs by holding
285 * It is valid to have concurrent insertions and deletions protected by per
286 * bucket locks or concurrent RCU protected lookups and traversals.
288 static int rhashtable_expand(struct rhashtable *ht)
290 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
293 ASSERT_RHT_MUTEX(ht);
295 old_tbl = rhashtable_last_table(ht, old_tbl);
297 new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
301 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
303 bucket_table_free(new_tbl);
309 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
310 * @ht: the hash table to shrink
312 * This function shrinks the hash table to fit, i.e., the smallest
313 * size would not cause it to expand right away automatically.
315 * The caller must ensure that no concurrent resizing occurs by holding
318 * The caller must ensure that no concurrent table mutations take place.
319 * It is however valid to have concurrent lookups if they are RCU protected.
321 * It is valid to have concurrent insertions and deletions protected by per
322 * bucket locks or concurrent RCU protected lookups and traversals.
324 static int rhashtable_shrink(struct rhashtable *ht)
326 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
327 unsigned size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
330 ASSERT_RHT_MUTEX(ht);
332 if (size < ht->p.min_size)
333 size = ht->p.min_size;
335 if (old_tbl->size <= size)
338 if (rht_dereference(old_tbl->future_tbl, ht))
341 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
345 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
347 bucket_table_free(new_tbl);
352 static void rht_deferred_worker(struct work_struct *work)
354 struct rhashtable *ht;
355 struct bucket_table *tbl;
358 ht = container_of(work, struct rhashtable, run_work);
359 mutex_lock(&ht->mutex);
360 if (ht->being_destroyed)
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 (rht_shrink_below_30(ht, tbl))
369 rhashtable_shrink(ht);
371 err = rhashtable_rehash_table(ht);
374 mutex_unlock(&ht->mutex);
377 schedule_work(&ht->run_work);
380 static bool rhashtable_check_elasticity(struct rhashtable *ht,
381 struct bucket_table *tbl,
384 unsigned elasticity = ht->elasticity;
385 struct rhash_head *head;
387 rht_for_each(head, tbl, hash)
394 int rhashtable_insert_rehash(struct rhashtable *ht)
396 struct bucket_table *old_tbl;
397 struct bucket_table *new_tbl;
398 struct bucket_table *tbl;
402 old_tbl = rht_dereference_rcu(ht->tbl, ht);
403 tbl = rhashtable_last_table(ht, old_tbl);
407 if (rht_grow_above_75(ht, tbl))
409 /* More than two rehashes (not resizes) detected. */
410 else if (WARN_ON(old_tbl != tbl && old_tbl->size == size))
413 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
417 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
419 bucket_table_free(new_tbl);
423 schedule_work(&ht->run_work);
427 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
429 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
430 struct rhash_head *obj,
431 struct bucket_table *tbl)
433 struct rhash_head *head;
437 tbl = rhashtable_last_table(ht, tbl);
438 hash = head_hashfn(ht, tbl, obj);
439 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
442 if (key && rhashtable_lookup_fast(ht, key, ht->p))
446 if (rhashtable_check_elasticity(ht, tbl, hash) ||
447 rht_grow_above_100(ht, tbl))
452 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
454 RCU_INIT_POINTER(obj->next, head);
456 rcu_assign_pointer(tbl->buckets[hash], obj);
458 atomic_inc(&ht->nelems);
461 spin_unlock(rht_bucket_lock(tbl, hash));
465 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
468 * rhashtable_walk_init - Initialise an iterator
469 * @ht: Table to walk over
470 * @iter: Hash table Iterator
472 * This function prepares a hash table walk.
474 * Note that if you restart a walk after rhashtable_walk_stop you
475 * may see the same object twice. Also, you may miss objects if
476 * there are removals in between rhashtable_walk_stop and the next
477 * call to rhashtable_walk_start.
479 * For a completely stable walk you should construct your own data
480 * structure outside the hash table.
482 * This function may sleep so you must not call it from interrupt
483 * context or with spin locks held.
485 * You must call rhashtable_walk_exit if this function returns
488 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
495 iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
499 mutex_lock(&ht->mutex);
500 iter->walker->tbl = rht_dereference(ht->tbl, ht);
501 list_add(&iter->walker->list, &iter->walker->tbl->walkers);
502 mutex_unlock(&ht->mutex);
506 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
509 * rhashtable_walk_exit - Free an iterator
510 * @iter: Hash table Iterator
512 * This function frees resources allocated by rhashtable_walk_init.
514 void rhashtable_walk_exit(struct rhashtable_iter *iter)
516 mutex_lock(&iter->ht->mutex);
517 if (iter->walker->tbl)
518 list_del(&iter->walker->list);
519 mutex_unlock(&iter->ht->mutex);
522 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
525 * rhashtable_walk_start - Start a hash table walk
526 * @iter: Hash table iterator
528 * Start a hash table walk. Note that we take the RCU lock in all
529 * cases including when we return an error. So you must always call
530 * rhashtable_walk_stop to clean up.
532 * Returns zero if successful.
534 * Returns -EAGAIN if resize event occured. Note that the iterator
535 * will rewind back to the beginning and you may use it immediately
536 * by calling rhashtable_walk_next.
538 int rhashtable_walk_start(struct rhashtable_iter *iter)
541 struct rhashtable *ht = iter->ht;
543 mutex_lock(&ht->mutex);
545 if (iter->walker->tbl)
546 list_del(&iter->walker->list);
550 mutex_unlock(&ht->mutex);
552 if (!iter->walker->tbl) {
553 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
559 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
562 * rhashtable_walk_next - Return the next object and advance the iterator
563 * @iter: Hash table iterator
565 * Note that you must call rhashtable_walk_stop when you are finished
568 * Returns the next object or NULL when the end of the table is reached.
570 * Returns -EAGAIN if resize event occured. Note that the iterator
571 * will rewind back to the beginning and you may continue to use it.
573 void *rhashtable_walk_next(struct rhashtable_iter *iter)
575 struct bucket_table *tbl = iter->walker->tbl;
576 struct rhashtable *ht = iter->ht;
577 struct rhash_head *p = iter->p;
581 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
585 for (; iter->slot < tbl->size; iter->slot++) {
586 int skip = iter->skip;
588 rht_for_each_rcu(p, tbl, iter->slot) {
595 if (!rht_is_a_nulls(p)) {
598 obj = rht_obj(ht, p);
605 /* Ensure we see any new tables. */
608 iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
609 if (iter->walker->tbl) {
612 return ERR_PTR(-EAGAIN);
621 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
624 * rhashtable_walk_stop - Finish a hash table walk
625 * @iter: Hash table iterator
627 * Finish a hash table walk.
629 void rhashtable_walk_stop(struct rhashtable_iter *iter)
632 struct rhashtable *ht;
633 struct bucket_table *tbl = iter->walker->tbl;
640 spin_lock(&ht->lock);
641 if (tbl->rehash < tbl->size)
642 list_add(&iter->walker->list, &tbl->walkers);
644 iter->walker->tbl = NULL;
645 spin_unlock(&ht->lock);
652 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
654 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
656 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
657 (unsigned long)params->min_size);
660 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
662 return jhash2(key, length, seed);
666 * rhashtable_init - initialize a new hash table
667 * @ht: hash table to be initialized
668 * @params: configuration parameters
670 * Initializes a new hash table based on the provided configuration
671 * parameters. A table can be configured either with a variable or
674 * Configuration Example 1: Fixed length keys
678 * struct rhash_head node;
681 * struct rhashtable_params params = {
682 * .head_offset = offsetof(struct test_obj, node),
683 * .key_offset = offsetof(struct test_obj, key),
684 * .key_len = sizeof(int),
686 * .nulls_base = (1U << RHT_BASE_SHIFT),
689 * Configuration Example 2: Variable length keys
692 * struct rhash_head node;
695 * u32 my_hash_fn(const void *data, u32 seed)
697 * struct test_obj *obj = data;
699 * return [... hash ...];
702 * struct rhashtable_params params = {
703 * .head_offset = offsetof(struct test_obj, node),
705 * .obj_hashfn = my_hash_fn,
708 int rhashtable_init(struct rhashtable *ht,
709 const struct rhashtable_params *params)
711 struct bucket_table *tbl;
714 size = HASH_DEFAULT_SIZE;
716 if ((!params->key_len && !params->obj_hashfn) ||
717 (params->obj_hashfn && !params->obj_cmpfn))
720 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
723 if (params->nelem_hint)
724 size = rounded_hashtable_size(params);
726 memset(ht, 0, sizeof(*ht));
727 mutex_init(&ht->mutex);
728 spin_lock_init(&ht->lock);
729 memcpy(&ht->p, params, sizeof(*params));
731 if (params->min_size)
732 ht->p.min_size = roundup_pow_of_two(params->min_size);
734 if (params->max_size)
735 ht->p.max_size = rounddown_pow_of_two(params->max_size);
737 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
739 if (!params->insecure_elasticity)
742 if (params->locks_mul)
743 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
745 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
747 ht->key_len = ht->p.key_len;
748 if (!params->hashfn) {
749 ht->p.hashfn = jhash;
751 if (!(ht->key_len & (sizeof(u32) - 1))) {
752 ht->key_len /= sizeof(u32);
753 ht->p.hashfn = rhashtable_jhash2;
757 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
761 atomic_set(&ht->nelems, 0);
763 RCU_INIT_POINTER(ht->tbl, tbl);
765 INIT_WORK(&ht->run_work, rht_deferred_worker);
769 EXPORT_SYMBOL_GPL(rhashtable_init);
772 * rhashtable_destroy - destroy hash table
773 * @ht: the hash table to destroy
775 * Frees the bucket array. This function is not rcu safe, therefore the caller
776 * has to make sure that no resizing may happen by unpublishing the hashtable
777 * and waiting for the quiescent cycle before releasing the bucket array.
779 void rhashtable_destroy(struct rhashtable *ht)
781 ht->being_destroyed = true;
783 cancel_work_sync(&ht->run_work);
785 mutex_lock(&ht->mutex);
786 bucket_table_free(rht_dereference(ht->tbl, ht));
787 mutex_unlock(&ht->mutex);
789 EXPORT_SYMBOL_GPL(rhashtable_destroy);