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) {
81 if (size * sizeof(spinlock_t) > PAGE_SIZE &&
83 tbl->locks = vmalloc(size * sizeof(spinlock_t));
86 if (gfp != GFP_KERNEL)
87 gfp |= __GFP_NOWARN | __GFP_NORETRY;
89 tbl->locks = kmalloc_array(size, sizeof(spinlock_t),
93 for (i = 0; i < size; i++)
94 spin_lock_init(&tbl->locks[i]);
96 tbl->locks_mask = size - 1;
101 static void bucket_table_free(const struct bucket_table *tbl)
109 static void bucket_table_free_rcu(struct rcu_head *head)
111 bucket_table_free(container_of(head, struct bucket_table, rcu));
114 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
118 struct bucket_table *tbl = NULL;
122 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
123 if (size <= (PAGE_SIZE << PAGE_ALLOC_COSTLY_ORDER) ||
125 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
126 if (tbl == NULL && gfp == GFP_KERNEL)
131 tbl->size = nbuckets;
133 if (alloc_bucket_locks(ht, tbl, gfp) < 0) {
134 bucket_table_free(tbl);
138 INIT_LIST_HEAD(&tbl->walkers);
140 get_random_bytes(&tbl->hash_rnd, sizeof(tbl->hash_rnd));
142 for (i = 0; i < nbuckets; i++)
143 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
148 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
149 struct bucket_table *tbl)
151 struct bucket_table *new_tbl;
155 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
161 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
163 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
164 struct bucket_table *new_tbl = rhashtable_last_table(ht,
165 rht_dereference_rcu(old_tbl->future_tbl, ht));
166 struct rhash_head __rcu **pprev = &old_tbl->buckets[old_hash];
168 struct rhash_head *head, *next, *entry;
169 spinlock_t *new_bucket_lock;
170 unsigned int new_hash;
172 rht_for_each(entry, old_tbl, old_hash) {
174 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
176 if (rht_is_a_nulls(next))
179 pprev = &entry->next;
185 new_hash = head_hashfn(ht, new_tbl, entry);
187 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
189 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
190 head = rht_dereference_bucket(new_tbl->buckets[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 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;
247 unsigned int old_hash;
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 int nelems = atomic_read(&ht->nelems);
328 unsigned int size = 0;
331 ASSERT_RHT_MUTEX(ht);
334 size = roundup_pow_of_two(nelems * 3 / 2);
335 if (size < ht->p.min_size)
336 size = ht->p.min_size;
338 if (old_tbl->size <= size)
341 if (rht_dereference(old_tbl->future_tbl, ht))
344 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
348 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
350 bucket_table_free(new_tbl);
355 static void rht_deferred_worker(struct work_struct *work)
357 struct rhashtable *ht;
358 struct bucket_table *tbl;
361 ht = container_of(work, struct rhashtable, run_work);
362 mutex_lock(&ht->mutex);
364 tbl = rht_dereference(ht->tbl, ht);
365 tbl = rhashtable_last_table(ht, tbl);
367 if (rht_grow_above_75(ht, tbl))
368 rhashtable_expand(ht);
369 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
370 rhashtable_shrink(ht);
372 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 int elasticity = ht->elasticity;
385 struct rhash_head *head;
387 rht_for_each(head, tbl, hash)
394 int rhashtable_insert_rehash(struct rhashtable *ht,
395 struct bucket_table *tbl)
397 struct bucket_table *old_tbl;
398 struct bucket_table *new_tbl;
402 old_tbl = rht_dereference_rcu(ht->tbl, ht);
408 if (rht_grow_above_75(ht, tbl))
410 /* Do not schedule more than one rehash */
411 else if (old_tbl != tbl)
416 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
420 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
422 bucket_table_free(new_tbl);
426 schedule_work(&ht->run_work);
431 /* Do not fail the insert if someone else did a rehash. */
432 if (likely(rcu_dereference_raw(tbl->future_tbl)))
435 /* Schedule async rehash to retry allocation in process context. */
437 schedule_work(&ht->run_work);
441 EXPORT_SYMBOL_GPL(rhashtable_insert_rehash);
443 struct bucket_table *rhashtable_insert_slow(struct rhashtable *ht,
445 struct rhash_head *obj,
446 struct bucket_table *tbl)
448 struct rhash_head *head;
452 tbl = rhashtable_last_table(ht, tbl);
453 hash = head_hashfn(ht, tbl, obj);
454 spin_lock_nested(rht_bucket_lock(tbl, hash), SINGLE_DEPTH_NESTING);
457 if (key && rhashtable_lookup_fast(ht, key, ht->p))
461 if (unlikely(rht_grow_above_max(ht, tbl)))
465 if (rhashtable_check_elasticity(ht, tbl, hash) ||
466 rht_grow_above_100(ht, tbl))
471 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
473 RCU_INIT_POINTER(obj->next, head);
475 rcu_assign_pointer(tbl->buckets[hash], obj);
477 atomic_inc(&ht->nelems);
480 spin_unlock(rht_bucket_lock(tbl, hash));
484 else if (err == -EAGAIN)
489 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
492 * rhashtable_walk_init - Initialise an iterator
493 * @ht: Table to walk over
494 * @iter: Hash table Iterator
495 * @gfp: GFP flags for allocations
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 if this function returns
513 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter,
521 iter->walker = kmalloc(sizeof(*iter->walker), gfp);
525 spin_lock(&ht->lock);
527 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
528 list_add(&iter->walker->list, &iter->walker->tbl->walkers);
529 spin_unlock(&ht->lock);
533 EXPORT_SYMBOL_GPL(rhashtable_walk_init);
536 * rhashtable_walk_exit - Free an iterator
537 * @iter: Hash table Iterator
539 * This function frees resources allocated by rhashtable_walk_init.
541 void rhashtable_walk_exit(struct rhashtable_iter *iter)
543 spin_lock(&iter->ht->lock);
544 if (iter->walker->tbl)
545 list_del(&iter->walker->list);
546 spin_unlock(&iter->ht->lock);
549 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
552 * rhashtable_walk_start - Start a hash table walk
553 * @iter: Hash table iterator
555 * Start a hash table walk. Note that we take the RCU lock in all
556 * cases including when we return an error. So you must always call
557 * rhashtable_walk_stop to clean up.
559 * Returns zero if successful.
561 * Returns -EAGAIN if resize event occured. Note that the iterator
562 * will rewind back to the beginning and you may use it immediately
563 * by calling rhashtable_walk_next.
565 int rhashtable_walk_start(struct rhashtable_iter *iter)
568 struct rhashtable *ht = iter->ht;
572 spin_lock(&ht->lock);
573 if (iter->walker->tbl)
574 list_del(&iter->walker->list);
575 spin_unlock(&ht->lock);
577 if (!iter->walker->tbl) {
578 iter->walker->tbl = rht_dereference_rcu(ht->tbl, ht);
584 EXPORT_SYMBOL_GPL(rhashtable_walk_start);
587 * rhashtable_walk_next - Return the next object and advance the iterator
588 * @iter: Hash table iterator
590 * Note that you must call rhashtable_walk_stop when you are finished
593 * Returns the next object or NULL when the end of the table is reached.
595 * Returns -EAGAIN if resize event occured. Note that the iterator
596 * will rewind back to the beginning and you may continue to use it.
598 void *rhashtable_walk_next(struct rhashtable_iter *iter)
600 struct bucket_table *tbl = iter->walker->tbl;
601 struct rhashtable *ht = iter->ht;
602 struct rhash_head *p = iter->p;
605 p = rht_dereference_bucket_rcu(p->next, tbl, iter->slot);
609 for (; iter->slot < tbl->size; iter->slot++) {
610 int skip = iter->skip;
612 rht_for_each_rcu(p, tbl, iter->slot) {
619 if (!rht_is_a_nulls(p)) {
622 return rht_obj(ht, p);
630 /* Ensure we see any new tables. */
633 iter->walker->tbl = rht_dereference_rcu(tbl->future_tbl, ht);
634 if (iter->walker->tbl) {
637 return ERR_PTR(-EAGAIN);
642 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
645 * rhashtable_walk_stop - Finish a hash table walk
646 * @iter: Hash table iterator
648 * Finish a hash table walk.
650 void rhashtable_walk_stop(struct rhashtable_iter *iter)
653 struct rhashtable *ht;
654 struct bucket_table *tbl = iter->walker->tbl;
661 spin_lock(&ht->lock);
662 if (tbl->rehash < tbl->size)
663 list_add(&iter->walker->list, &tbl->walkers);
665 iter->walker->tbl = NULL;
666 spin_unlock(&ht->lock);
673 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
675 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
677 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
678 (unsigned long)params->min_size);
681 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
683 return jhash2(key, length, seed);
687 * rhashtable_init - initialize a new hash table
688 * @ht: hash table to be initialized
689 * @params: configuration parameters
691 * Initializes a new hash table based on the provided configuration
692 * parameters. A table can be configured either with a variable or
695 * Configuration Example 1: Fixed length keys
699 * struct rhash_head node;
702 * struct rhashtable_params params = {
703 * .head_offset = offsetof(struct test_obj, node),
704 * .key_offset = offsetof(struct test_obj, key),
705 * .key_len = sizeof(int),
707 * .nulls_base = (1U << RHT_BASE_SHIFT),
710 * Configuration Example 2: Variable length keys
713 * struct rhash_head node;
716 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
718 * struct test_obj *obj = data;
720 * return [... hash ...];
723 * struct rhashtable_params params = {
724 * .head_offset = offsetof(struct test_obj, node),
726 * .obj_hashfn = my_hash_fn,
729 int rhashtable_init(struct rhashtable *ht,
730 const struct rhashtable_params *params)
732 struct bucket_table *tbl;
735 size = HASH_DEFAULT_SIZE;
737 if ((!params->key_len && !params->obj_hashfn) ||
738 (params->obj_hashfn && !params->obj_cmpfn))
741 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
744 memset(ht, 0, sizeof(*ht));
745 mutex_init(&ht->mutex);
746 spin_lock_init(&ht->lock);
747 memcpy(&ht->p, params, sizeof(*params));
749 if (params->min_size)
750 ht->p.min_size = roundup_pow_of_two(params->min_size);
752 if (params->max_size)
753 ht->p.max_size = rounddown_pow_of_two(params->max_size);
755 if (params->insecure_max_entries)
756 ht->p.insecure_max_entries =
757 rounddown_pow_of_two(params->insecure_max_entries);
759 ht->p.insecure_max_entries = ht->p.max_size * 2;
761 ht->p.min_size = max(ht->p.min_size, HASH_MIN_SIZE);
763 if (params->nelem_hint)
764 size = rounded_hashtable_size(&ht->p);
766 /* The maximum (not average) chain length grows with the
767 * size of the hash table, at a rate of (log N)/(log log N).
768 * The value of 16 is selected so that even if the hash
769 * table grew to 2^32 you would not expect the maximum
770 * chain length to exceed it unless we are under attack
771 * (or extremely unlucky).
773 * As this limit is only to detect attacks, we don't need
774 * to set it to a lower value as you'd need the chain
775 * length to vastly exceed 16 to have any real effect
778 if (!params->insecure_elasticity)
781 if (params->locks_mul)
782 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
784 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
786 ht->key_len = ht->p.key_len;
787 if (!params->hashfn) {
788 ht->p.hashfn = jhash;
790 if (!(ht->key_len & (sizeof(u32) - 1))) {
791 ht->key_len /= sizeof(u32);
792 ht->p.hashfn = rhashtable_jhash2;
796 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
800 atomic_set(&ht->nelems, 0);
802 RCU_INIT_POINTER(ht->tbl, tbl);
804 INIT_WORK(&ht->run_work, rht_deferred_worker);
808 EXPORT_SYMBOL_GPL(rhashtable_init);
811 * rhashtable_free_and_destroy - free elements and destroy hash table
812 * @ht: the hash table to destroy
813 * @free_fn: callback to release resources of element
814 * @arg: pointer passed to free_fn
816 * Stops an eventual async resize. If defined, invokes free_fn for each
817 * element to releasal resources. Please note that RCU protected
818 * readers may still be accessing the elements. Releasing of resources
819 * must occur in a compatible manner. Then frees the bucket array.
821 * This function will eventually sleep to wait for an async resize
822 * to complete. The caller is responsible that no further write operations
823 * occurs in parallel.
825 void rhashtable_free_and_destroy(struct rhashtable *ht,
826 void (*free_fn)(void *ptr, void *arg),
829 const struct bucket_table *tbl;
832 cancel_work_sync(&ht->run_work);
834 mutex_lock(&ht->mutex);
835 tbl = rht_dereference(ht->tbl, ht);
837 for (i = 0; i < tbl->size; i++) {
838 struct rhash_head *pos, *next;
840 for (pos = rht_dereference(tbl->buckets[i], ht),
841 next = !rht_is_a_nulls(pos) ?
842 rht_dereference(pos->next, ht) : NULL;
843 !rht_is_a_nulls(pos);
845 next = !rht_is_a_nulls(pos) ?
846 rht_dereference(pos->next, ht) : NULL)
847 free_fn(rht_obj(ht, pos), arg);
851 bucket_table_free(tbl);
852 mutex_unlock(&ht->mutex);
854 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
856 void rhashtable_destroy(struct rhashtable *ht)
858 return rhashtable_free_and_destroy(ht, NULL, NULL);
860 EXPORT_SYMBOL_GPL(rhashtable_destroy);