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 list_for_each_entry(walker, &old_tbl->walkers, list)
262 /* Wait for readers. All new readers will see the new
263 * table, and thus no references to the old table will
266 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
268 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
272 * rhashtable_expand - Expand hash table while allowing concurrent lookups
273 * @ht: the hash table to expand
275 * A secondary bucket array is allocated and the hash entries are migrated.
277 * This function may only be called in a context where it is safe to call
278 * synchronize_rcu(), e.g. not within a rcu_read_lock() section.
280 * The caller must ensure that no concurrent resizing occurs by holding
283 * It is valid to have concurrent insertions and deletions protected by per
284 * bucket locks or concurrent RCU protected lookups and traversals.
286 static int rhashtable_expand(struct rhashtable *ht)
288 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
291 ASSERT_RHT_MUTEX(ht);
293 old_tbl = rhashtable_last_table(ht, old_tbl);
295 new_tbl = bucket_table_alloc(ht, old_tbl->size * 2, GFP_KERNEL);
299 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
301 bucket_table_free(new_tbl);
307 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
308 * @ht: the hash table to shrink
310 * This function shrinks the hash table to fit, i.e., the smallest
311 * size would not cause it to expand right away automatically.
313 * The caller must ensure that no concurrent resizing occurs by holding
316 * The caller must ensure that no concurrent table mutations take place.
317 * It is however valid to have concurrent lookups if they are RCU protected.
319 * It is valid to have concurrent insertions and deletions protected by per
320 * bucket locks or concurrent RCU protected lookups and traversals.
322 static int rhashtable_shrink(struct rhashtable *ht)
324 struct bucket_table *new_tbl, *old_tbl = rht_dereference(ht->tbl, ht);
325 unsigned size = roundup_pow_of_two(atomic_read(&ht->nelems) * 3 / 2);
328 ASSERT_RHT_MUTEX(ht);
330 if (size < ht->p.min_size)
331 size = ht->p.min_size;
333 if (old_tbl->size <= size)
336 if (rht_dereference(old_tbl->future_tbl, ht))
339 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
343 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
345 bucket_table_free(new_tbl);
350 static void rht_deferred_worker(struct work_struct *work)
352 struct rhashtable *ht;
353 struct bucket_table *tbl;
356 ht = container_of(work, struct rhashtable, run_work);
357 mutex_lock(&ht->mutex);
358 if (ht->being_destroyed)
361 tbl = rht_dereference(ht->tbl, ht);
362 tbl = rhashtable_last_table(ht, tbl);
364 if (rht_grow_above_75(ht, tbl))
365 rhashtable_expand(ht);
366 else if (rht_shrink_below_30(ht, tbl))
367 rhashtable_shrink(ht);
369 err = rhashtable_rehash_table(ht);
372 mutex_unlock(&ht->mutex);
375 schedule_work(&ht->run_work);
378 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
379 struct rhash_head *obj,
380 struct bucket_table *tbl)
382 struct rhash_head *head;
386 tbl = rhashtable_last_table(ht, tbl);
387 hash = head_hashfn(ht, tbl, obj);
388 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
390 if (key && rhashtable_lookup_fast(ht, key, ht->p))
395 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
397 RCU_INIT_POINTER(obj->next, head);
399 rcu_assign_pointer(tbl->buckets[hash], obj);
401 atomic_inc(&ht->nelems);
404 spin_unlock(rht_bucket_lock(tbl, hash));
408 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
411 * rhashtable_walk_init - Initialise an iterator
412 * @ht: Table to walk over
413 * @iter: Hash table Iterator
415 * This function prepares a hash table walk.
417 * Note that if you restart a walk after rhashtable_walk_stop you
418 * may see the same object twice. Also, you may miss objects if
419 * there are removals in between rhashtable_walk_stop and the next
420 * call to rhashtable_walk_start.
422 * For a completely stable walk you should construct your own data
423 * structure outside the hash table.
425 * This function may sleep so you must not call it from interrupt
426 * context or with spin locks held.
428 * You must call rhashtable_walk_exit if this function returns
431 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter)
438 iter->walker = kmalloc(sizeof(*iter->walker), GFP_KERNEL);
442 mutex_lock(&ht->mutex);
443 iter->walker->tbl = rht_dereference(ht->tbl, ht);
444 list_add(&iter->walker->list, &iter->walker->tbl->walkers);
445 mutex_unlock(&ht->mutex);
449 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
452 * rhashtable_walk_exit - Free an iterator
453 * @iter: Hash table Iterator
455 * This function frees resources allocated by rhashtable_walk_init.
457 void rhashtable_walk_exit(struct rhashtable_iter *iter)
459 mutex_lock(&iter->ht->mutex);
460 if (iter->walker->tbl)
461 list_del(&iter->walker->list);
462 mutex_unlock(&iter->ht->mutex);
465 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
468 * rhashtable_walk_start - Start a hash table walk
469 * @iter: Hash table iterator
471 * Start a hash table walk. Note that we take the RCU lock in all
472 * cases including when we return an error. So you must always call
473 * rhashtable_walk_stop to clean up.
475 * Returns zero if successful.
477 * Returns -EAGAIN if resize event occured. Note that the iterator
478 * will rewind back to the beginning and you may use it immediately
479 * by calling rhashtable_walk_next.
481 int rhashtable_walk_start(struct rhashtable_iter *iter)
484 struct rhashtable *ht = iter->ht;
486 mutex_lock(&ht->mutex);
488 if (iter->walker->tbl)
489 list_del(&iter->walker->list);
493 mutex_unlock(&ht->mutex);
495 if (!iter->walker->tbl) {
496 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
502 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
505 * rhashtable_walk_next - Return the next object and advance the iterator
506 * @iter: Hash table iterator
508 * Note that you must call rhashtable_walk_stop when you are finished
511 * Returns the next object or NULL when the end of the table is reached.
513 * Returns -EAGAIN if resize event occured. Note that the iterator
514 * will rewind back to the beginning and you may continue to use it.
516 void *rhashtable_walk_next(struct rhashtable_iter *iter)
518 struct bucket_table *tbl = iter->walker->tbl;
519 struct rhashtable *ht = iter->ht;
520 struct rhash_head *p = iter->p;
524 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
528 for (; iter->slot < tbl->size; iter->slot++) {
529 int skip = iter->skip;
531 rht_for_each_rcu(p, tbl, iter->slot) {
538 if (!rht_is_a_nulls(p)) {
541 obj = rht_obj(ht, p);
548 /* Ensure we see any new tables. */
551 iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
552 if (iter->walker->tbl) {
555 return ERR_PTR(-EAGAIN);
564 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
567 * rhashtable_walk_stop - Finish a hash table walk
568 * @iter: Hash table iterator
570 * Finish a hash table walk.
572 void rhashtable_walk_stop(struct rhashtable_iter *iter)
575 struct rhashtable *ht;
576 struct bucket_table *tbl = iter->walker->tbl;
583 mutex_lock(&ht->mutex);
584 if (tbl->rehash < tbl->size)
585 list_add(&iter->walker->list, &tbl->walkers);
587 iter->walker->tbl = NULL;
588 mutex_unlock(&ht->mutex);
595 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
597 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
599 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
600 (unsigned long)params->min_size);
603 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
605 return jhash2(key, length, seed);
609 * rhashtable_init - initialize a new hash table
610 * @ht: hash table to be initialized
611 * @params: configuration parameters
613 * Initializes a new hash table based on the provided configuration
614 * parameters. A table can be configured either with a variable or
617 * Configuration Example 1: Fixed length keys
621 * struct rhash_head node;
624 * struct rhashtable_params params = {
625 * .head_offset = offsetof(struct test_obj, node),
626 * .key_offset = offsetof(struct test_obj, key),
627 * .key_len = sizeof(int),
629 * .nulls_base = (1U << RHT_BASE_SHIFT),
632 * Configuration Example 2: Variable length keys
635 * struct rhash_head node;
638 * u32 my_hash_fn(const void *data, u32 seed)
640 * struct test_obj *obj = data;
642 * return [... hash ...];
645 * struct rhashtable_params params = {
646 * .head_offset = offsetof(struct test_obj, node),
648 * .obj_hashfn = my_hash_fn,
651 int rhashtable_init(struct rhashtable *ht,
652 const struct rhashtable_params *params)
654 struct bucket_table *tbl;
657 size = HASH_DEFAULT_SIZE;
659 if ((!params->key_len && !params->obj_hashfn) ||
660 (params->obj_hashfn && !params->obj_cmpfn))
663 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
666 if (params->nelem_hint)
667 size = rounded_hashtable_size(params);
669 memset(ht, 0, sizeof(*ht));
670 mutex_init(&ht->mutex);
671 memcpy(&ht->p, params, sizeof(*params));
673 if (params->min_size)
674 ht->p.min_size = roundup_pow_of_two(params->min_size);
676 if (params->max_size)
677 ht->p.max_size = rounddown_pow_of_two(params->max_size);
679 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
681 if (params->locks_mul)
682 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
684 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
686 ht->key_len = ht->p.key_len;
687 if (!params->hashfn) {
688 ht->p.hashfn = jhash;
690 if (!(ht->key_len & (sizeof(u32) - 1))) {
691 ht->key_len /= sizeof(u32);
692 ht->p.hashfn = rhashtable_jhash2;
696 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
700 atomic_set(&ht->nelems, 0);
702 RCU_INIT_POINTER(ht->tbl, tbl);
704 INIT_WORK(&ht->run_work, rht_deferred_worker);
708 EXPORT_SYMBOL_GPL(rhashtable_init);
711 * rhashtable_destroy - destroy hash table
712 * @ht: the hash table to destroy
714 * Frees the bucket array. This function is not rcu safe, therefore the caller
715 * has to make sure that no resizing may happen by unpublishing the hashtable
716 * and waiting for the quiescent cycle before releasing the bucket array.
718 void rhashtable_destroy(struct rhashtable *ht)
720 ht->being_destroyed = true;
722 cancel_work_sync(&ht->run_work);
724 mutex_lock(&ht->mutex);
725 bucket_table_free(rht_dereference(ht->tbl, ht));
726 mutex_unlock(&ht->mutex);
728 EXPORT_SYMBOL_GPL(rhashtable_destroy);