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/rculist.h>
23 #include <linux/slab.h>
24 #include <linux/vmalloc.h>
26 #include <linux/jhash.h>
27 #include <linux/random.h>
28 #include <linux/rhashtable.h>
29 #include <linux/err.h>
30 #include <linux/export.h>
32 #define HASH_DEFAULT_SIZE 64UL
33 #define HASH_MIN_SIZE 4U
34 #define BUCKET_LOCKS_PER_CPU 32UL
37 union nested_table __rcu *table;
38 struct rhash_head __rcu *bucket;
41 static u32 head_hashfn(struct rhashtable *ht,
42 const struct bucket_table *tbl,
43 const struct rhash_head *he)
45 return rht_head_hashfn(ht, tbl, he, ht->p);
48 #ifdef CONFIG_PROVE_LOCKING
49 #define ASSERT_RHT_MUTEX(HT) BUG_ON(!lockdep_rht_mutex_is_held(HT))
51 int lockdep_rht_mutex_is_held(struct rhashtable *ht)
53 return (debug_locks) ? lockdep_is_held(&ht->mutex) : 1;
55 EXPORT_SYMBOL_GPL(lockdep_rht_mutex_is_held);
57 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash)
59 spinlock_t *lock = rht_bucket_lock(tbl, hash);
61 return (debug_locks) ? lockdep_is_held(lock) : 1;
63 EXPORT_SYMBOL_GPL(lockdep_rht_bucket_is_held);
65 #define ASSERT_RHT_MUTEX(HT)
68 static void nested_table_free(union nested_table *ntbl, unsigned int size)
70 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
71 const unsigned int len = 1 << shift;
74 ntbl = rcu_dereference_raw(ntbl->table);
80 for (i = 0; i < len; i++)
81 nested_table_free(ntbl + i, size);
87 static void nested_bucket_table_free(const struct bucket_table *tbl)
89 unsigned int size = tbl->size >> tbl->nest;
90 unsigned int len = 1 << tbl->nest;
91 union nested_table *ntbl;
94 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
96 for (i = 0; i < len; i++)
97 nested_table_free(ntbl + i, size);
102 static void bucket_table_free(const struct bucket_table *tbl)
105 nested_bucket_table_free(tbl);
107 free_bucket_spinlocks(tbl->locks);
111 static void bucket_table_free_rcu(struct rcu_head *head)
113 bucket_table_free(container_of(head, struct bucket_table, rcu));
116 static union nested_table *nested_table_alloc(struct rhashtable *ht,
117 union nested_table __rcu **prev,
118 unsigned int shifted,
121 union nested_table *ntbl;
124 ntbl = rcu_dereference(*prev);
128 ntbl = kzalloc(PAGE_SIZE, GFP_ATOMIC);
130 if (ntbl && shifted) {
131 for (i = 0; i < PAGE_SIZE / sizeof(ntbl[0].bucket); i++)
132 INIT_RHT_NULLS_HEAD(ntbl[i].bucket, ht,
133 (i << shifted) | nhash);
136 rcu_assign_pointer(*prev, ntbl);
141 static struct bucket_table *nested_bucket_table_alloc(struct rhashtable *ht,
145 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
146 struct bucket_table *tbl;
149 if (nbuckets < (1 << (shift + 1)))
152 size = sizeof(*tbl) + sizeof(tbl->buckets[0]);
154 tbl = kzalloc(size, gfp);
158 if (!nested_table_alloc(ht, (union nested_table __rcu **)tbl->buckets,
164 tbl->nest = (ilog2(nbuckets) - 1) % shift + 1;
169 static struct bucket_table *bucket_table_alloc(struct rhashtable *ht,
173 struct bucket_table *tbl = NULL;
174 size_t size, max_locks;
177 size = sizeof(*tbl) + nbuckets * sizeof(tbl->buckets[0]);
178 if (gfp != GFP_KERNEL)
179 tbl = kzalloc(size, gfp | __GFP_NOWARN | __GFP_NORETRY);
181 tbl = kvzalloc(size, gfp);
185 if (tbl == NULL && gfp != GFP_KERNEL) {
186 tbl = nested_bucket_table_alloc(ht, nbuckets, gfp);
194 max_locks = size >> 1;
196 max_locks = min_t(size_t, max_locks, 1U << tbl->nest);
198 if (alloc_bucket_spinlocks(&tbl->locks, &tbl->locks_mask, max_locks,
199 ht->p.locks_mul, gfp) < 0) {
200 bucket_table_free(tbl);
204 INIT_LIST_HEAD(&tbl->walkers);
206 tbl->hash_rnd = get_random_u32();
208 for (i = 0; i < nbuckets; i++)
209 INIT_RHT_NULLS_HEAD(tbl->buckets[i], ht, i);
214 static struct bucket_table *rhashtable_last_table(struct rhashtable *ht,
215 struct bucket_table *tbl)
217 struct bucket_table *new_tbl;
221 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
227 static int rhashtable_rehash_one(struct rhashtable *ht, unsigned int old_hash)
229 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
230 struct bucket_table *new_tbl = rhashtable_last_table(ht,
231 rht_dereference_rcu(old_tbl->future_tbl, ht));
232 struct rhash_head __rcu **pprev = rht_bucket_var(old_tbl, old_hash);
234 struct rhash_head *head, *next, *entry;
235 spinlock_t *new_bucket_lock;
236 unsigned int new_hash;
243 rht_for_each(entry, old_tbl, old_hash) {
245 next = rht_dereference_bucket(entry->next, old_tbl, old_hash);
247 if (rht_is_a_nulls(next))
250 pprev = &entry->next;
256 new_hash = head_hashfn(ht, new_tbl, entry);
258 new_bucket_lock = rht_bucket_lock(new_tbl, new_hash);
260 spin_lock_nested(new_bucket_lock, SINGLE_DEPTH_NESTING);
261 head = rht_dereference_bucket(new_tbl->buckets[new_hash],
264 RCU_INIT_POINTER(entry->next, head);
266 rcu_assign_pointer(new_tbl->buckets[new_hash], entry);
267 spin_unlock(new_bucket_lock);
269 rcu_assign_pointer(*pprev, next);
275 static int rhashtable_rehash_chain(struct rhashtable *ht,
276 unsigned int old_hash)
278 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
279 spinlock_t *old_bucket_lock;
282 old_bucket_lock = rht_bucket_lock(old_tbl, old_hash);
284 spin_lock_bh(old_bucket_lock);
285 while (!(err = rhashtable_rehash_one(ht, old_hash)))
288 if (err == -ENOENT) {
292 spin_unlock_bh(old_bucket_lock);
297 static int rhashtable_rehash_attach(struct rhashtable *ht,
298 struct bucket_table *old_tbl,
299 struct bucket_table *new_tbl)
301 /* Protect future_tbl using the first bucket lock. */
302 spin_lock_bh(old_tbl->locks);
304 /* Did somebody beat us to it? */
305 if (rcu_access_pointer(old_tbl->future_tbl)) {
306 spin_unlock_bh(old_tbl->locks);
310 /* Make insertions go into the new, empty table right away. Deletions
311 * and lookups will be attempted in both tables until we synchronize.
313 rcu_assign_pointer(old_tbl->future_tbl, new_tbl);
315 spin_unlock_bh(old_tbl->locks);
320 static int rhashtable_rehash_table(struct rhashtable *ht)
322 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
323 struct bucket_table *new_tbl;
324 struct rhashtable_walker *walker;
325 unsigned int old_hash;
328 new_tbl = rht_dereference(old_tbl->future_tbl, ht);
332 for (old_hash = 0; old_hash < old_tbl->size; old_hash++) {
333 err = rhashtable_rehash_chain(ht, old_hash);
338 /* Publish the new table pointer. */
339 rcu_assign_pointer(ht->tbl, new_tbl);
341 spin_lock(&ht->lock);
342 list_for_each_entry(walker, &old_tbl->walkers, list)
344 spin_unlock(&ht->lock);
346 /* Wait for readers. All new readers will see the new
347 * table, and thus no references to the old table will
350 call_rcu(&old_tbl->rcu, bucket_table_free_rcu);
352 return rht_dereference(new_tbl->future_tbl, ht) ? -EAGAIN : 0;
355 static int rhashtable_rehash_alloc(struct rhashtable *ht,
356 struct bucket_table *old_tbl,
359 struct bucket_table *new_tbl;
362 ASSERT_RHT_MUTEX(ht);
364 new_tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
368 err = rhashtable_rehash_attach(ht, old_tbl, new_tbl);
370 bucket_table_free(new_tbl);
376 * rhashtable_shrink - Shrink hash table while allowing concurrent lookups
377 * @ht: the hash table to shrink
379 * This function shrinks the hash table to fit, i.e., the smallest
380 * size would not cause it to expand right away automatically.
382 * The caller must ensure that no concurrent resizing occurs by holding
385 * The caller must ensure that no concurrent table mutations take place.
386 * It is however valid to have concurrent lookups if they are RCU protected.
388 * It is valid to have concurrent insertions and deletions protected by per
389 * bucket locks or concurrent RCU protected lookups and traversals.
391 static int rhashtable_shrink(struct rhashtable *ht)
393 struct bucket_table *old_tbl = rht_dereference(ht->tbl, ht);
394 unsigned int nelems = atomic_read(&ht->nelems);
395 unsigned int size = 0;
398 size = roundup_pow_of_two(nelems * 3 / 2);
399 if (size < ht->p.min_size)
400 size = ht->p.min_size;
402 if (old_tbl->size <= size)
405 if (rht_dereference(old_tbl->future_tbl, ht))
408 return rhashtable_rehash_alloc(ht, old_tbl, size);
411 static void rht_deferred_worker(struct work_struct *work)
413 struct rhashtable *ht;
414 struct bucket_table *tbl;
417 ht = container_of(work, struct rhashtable, run_work);
418 mutex_lock(&ht->mutex);
420 tbl = rht_dereference(ht->tbl, ht);
421 tbl = rhashtable_last_table(ht, tbl);
423 if (rht_grow_above_75(ht, tbl))
424 err = rhashtable_rehash_alloc(ht, tbl, tbl->size * 2);
425 else if (ht->p.automatic_shrinking && rht_shrink_below_30(ht, tbl))
426 err = rhashtable_shrink(ht);
428 err = rhashtable_rehash_alloc(ht, tbl, tbl->size);
431 err = rhashtable_rehash_table(ht);
433 mutex_unlock(&ht->mutex);
436 schedule_work(&ht->run_work);
439 static int rhashtable_insert_rehash(struct rhashtable *ht,
440 struct bucket_table *tbl)
442 struct bucket_table *old_tbl;
443 struct bucket_table *new_tbl;
447 old_tbl = rht_dereference_rcu(ht->tbl, ht);
453 if (rht_grow_above_75(ht, tbl))
455 /* Do not schedule more than one rehash */
456 else if (old_tbl != tbl)
461 new_tbl = bucket_table_alloc(ht, size, GFP_ATOMIC);
465 err = rhashtable_rehash_attach(ht, tbl, new_tbl);
467 bucket_table_free(new_tbl);
471 schedule_work(&ht->run_work);
476 /* Do not fail the insert if someone else did a rehash. */
477 if (likely(rcu_dereference_raw(tbl->future_tbl)))
480 /* Schedule async rehash to retry allocation in process context. */
482 schedule_work(&ht->run_work);
487 static void *rhashtable_lookup_one(struct rhashtable *ht,
488 struct bucket_table *tbl, unsigned int hash,
489 const void *key, struct rhash_head *obj)
491 struct rhashtable_compare_arg arg = {
495 struct rhash_head __rcu **pprev;
496 struct rhash_head *head;
499 elasticity = RHT_ELASTICITY;
500 pprev = rht_bucket_var(tbl, hash);
501 rht_for_each_continue(head, *pprev, tbl, hash) {
502 struct rhlist_head *list;
503 struct rhlist_head *plist;
508 ht->p.obj_cmpfn(&arg, rht_obj(ht, head)) :
509 rhashtable_compare(&arg, rht_obj(ht, head)))) {
515 return rht_obj(ht, head);
517 list = container_of(obj, struct rhlist_head, rhead);
518 plist = container_of(head, struct rhlist_head, rhead);
520 RCU_INIT_POINTER(list->next, plist);
521 head = rht_dereference_bucket(head->next, tbl, hash);
522 RCU_INIT_POINTER(list->rhead.next, head);
523 rcu_assign_pointer(*pprev, obj);
529 return ERR_PTR(-EAGAIN);
531 return ERR_PTR(-ENOENT);
534 static struct bucket_table *rhashtable_insert_one(struct rhashtable *ht,
535 struct bucket_table *tbl,
537 struct rhash_head *obj,
540 struct rhash_head __rcu **pprev;
541 struct bucket_table *new_tbl;
542 struct rhash_head *head;
544 if (!IS_ERR_OR_NULL(data))
545 return ERR_PTR(-EEXIST);
547 if (PTR_ERR(data) != -EAGAIN && PTR_ERR(data) != -ENOENT)
548 return ERR_CAST(data);
550 new_tbl = rcu_dereference(tbl->future_tbl);
554 if (PTR_ERR(data) != -ENOENT)
555 return ERR_CAST(data);
557 if (unlikely(rht_grow_above_max(ht, tbl)))
558 return ERR_PTR(-E2BIG);
560 if (unlikely(rht_grow_above_100(ht, tbl)))
561 return ERR_PTR(-EAGAIN);
563 pprev = rht_bucket_insert(ht, tbl, hash);
565 return ERR_PTR(-ENOMEM);
567 head = rht_dereference_bucket(*pprev, tbl, hash);
569 RCU_INIT_POINTER(obj->next, head);
571 struct rhlist_head *list;
573 list = container_of(obj, struct rhlist_head, rhead);
574 RCU_INIT_POINTER(list->next, NULL);
577 rcu_assign_pointer(*pprev, obj);
579 atomic_inc(&ht->nelems);
580 if (rht_grow_above_75(ht, tbl))
581 schedule_work(&ht->run_work);
586 static void *rhashtable_try_insert(struct rhashtable *ht, const void *key,
587 struct rhash_head *obj)
589 struct bucket_table *new_tbl;
590 struct bucket_table *tbl;
595 tbl = rcu_dereference(ht->tbl);
597 /* All insertions must grab the oldest table containing
598 * the hashed bucket that is yet to be rehashed.
601 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
602 lock = rht_bucket_lock(tbl, hash);
605 if (tbl->rehash <= hash)
608 spin_unlock_bh(lock);
609 tbl = rcu_dereference(tbl->future_tbl);
612 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
613 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
614 if (PTR_ERR(new_tbl) != -EEXIST)
615 data = ERR_CAST(new_tbl);
617 while (!IS_ERR_OR_NULL(new_tbl)) {
619 hash = rht_head_hashfn(ht, tbl, obj, ht->p);
620 spin_lock_nested(rht_bucket_lock(tbl, hash),
621 SINGLE_DEPTH_NESTING);
623 data = rhashtable_lookup_one(ht, tbl, hash, key, obj);
624 new_tbl = rhashtable_insert_one(ht, tbl, hash, obj, data);
625 if (PTR_ERR(new_tbl) != -EEXIST)
626 data = ERR_CAST(new_tbl);
628 spin_unlock(rht_bucket_lock(tbl, hash));
631 spin_unlock_bh(lock);
633 if (PTR_ERR(data) == -EAGAIN)
634 data = ERR_PTR(rhashtable_insert_rehash(ht, tbl) ?:
640 void *rhashtable_insert_slow(struct rhashtable *ht, const void *key,
641 struct rhash_head *obj)
647 data = rhashtable_try_insert(ht, key, obj);
649 } while (PTR_ERR(data) == -EAGAIN);
653 EXPORT_SYMBOL_GPL(rhashtable_insert_slow);
656 * rhashtable_walk_enter - Initialise an iterator
657 * @ht: Table to walk over
658 * @iter: Hash table Iterator
660 * This function prepares a hash table walk.
662 * Note that if you restart a walk after rhashtable_walk_stop you
663 * may see the same object twice. Also, you may miss objects if
664 * there are removals in between rhashtable_walk_stop and the next
665 * call to rhashtable_walk_start.
667 * For a completely stable walk you should construct your own data
668 * structure outside the hash table.
670 * This function may sleep so you must not call it from interrupt
671 * context or with spin locks held.
673 * You must call rhashtable_walk_exit after this function returns.
675 void rhashtable_walk_enter(struct rhashtable *ht, struct rhashtable_iter *iter)
681 iter->end_of_table = 0;
683 spin_lock(&ht->lock);
685 rcu_dereference_protected(ht->tbl, lockdep_is_held(&ht->lock));
686 list_add(&iter->walker.list, &iter->walker.tbl->walkers);
687 spin_unlock(&ht->lock);
689 EXPORT_SYMBOL_GPL(rhashtable_walk_enter);
692 * rhashtable_walk_exit - Free an iterator
693 * @iter: Hash table Iterator
695 * This function frees resources allocated by rhashtable_walk_init.
697 void rhashtable_walk_exit(struct rhashtable_iter *iter)
699 spin_lock(&iter->ht->lock);
700 if (iter->walker.tbl)
701 list_del(&iter->walker.list);
702 spin_unlock(&iter->ht->lock);
704 EXPORT_SYMBOL_GPL(rhashtable_walk_exit);
707 * rhashtable_walk_start_check - Start a hash table walk
708 * @iter: Hash table iterator
710 * Start a hash table walk at the current iterator position. Note that we take
711 * the RCU lock in all cases including when we return an error. So you must
712 * always call rhashtable_walk_stop to clean up.
714 * Returns zero if successful.
716 * Returns -EAGAIN if resize event occured. Note that the iterator
717 * will rewind back to the beginning and you may use it immediately
718 * by calling rhashtable_walk_next.
720 * rhashtable_walk_start is defined as an inline variant that returns
721 * void. This is preferred in cases where the caller would ignore
722 * resize events and always continue.
724 int rhashtable_walk_start_check(struct rhashtable_iter *iter)
727 struct rhashtable *ht = iter->ht;
731 spin_lock(&ht->lock);
732 if (iter->walker.tbl)
733 list_del(&iter->walker.list);
734 spin_unlock(&ht->lock);
736 if (!iter->walker.tbl && !iter->end_of_table) {
737 iter->walker.tbl = rht_dereference_rcu(ht->tbl, ht);
743 EXPORT_SYMBOL_GPL(rhashtable_walk_start_check);
746 * __rhashtable_walk_find_next - Find the next element in a table (or the first
747 * one in case of a new walk).
749 * @iter: Hash table iterator
751 * Returns the found object or NULL when the end of the table is reached.
753 * Returns -EAGAIN if resize event occurred.
755 static void *__rhashtable_walk_find_next(struct rhashtable_iter *iter)
757 struct bucket_table *tbl = iter->walker.tbl;
758 struct rhlist_head *list = iter->list;
759 struct rhashtable *ht = iter->ht;
760 struct rhash_head *p = iter->p;
761 bool rhlist = ht->rhlist;
766 for (; iter->slot < tbl->size; iter->slot++) {
767 int skip = iter->skip;
769 rht_for_each_rcu(p, tbl, iter->slot) {
771 list = container_of(p, struct rhlist_head,
777 list = rcu_dereference(list->next);
788 if (!rht_is_a_nulls(p)) {
792 return rht_obj(ht, rhlist ? &list->rhead : p);
800 /* Ensure we see any new tables. */
803 iter->walker.tbl = rht_dereference_rcu(tbl->future_tbl, ht);
804 if (iter->walker.tbl) {
807 return ERR_PTR(-EAGAIN);
809 iter->end_of_table = true;
816 * rhashtable_walk_next - Return the next object and advance the iterator
817 * @iter: Hash table iterator
819 * Note that you must call rhashtable_walk_stop when you are finished
822 * Returns the next object or NULL when the end of the table is reached.
824 * Returns -EAGAIN if resize event occurred. Note that the iterator
825 * will rewind back to the beginning and you may continue to use it.
827 void *rhashtable_walk_next(struct rhashtable_iter *iter)
829 struct rhlist_head *list = iter->list;
830 struct rhashtable *ht = iter->ht;
831 struct rhash_head *p = iter->p;
832 bool rhlist = ht->rhlist;
835 if (!rhlist || !(list = rcu_dereference(list->next))) {
836 p = rcu_dereference(p->next);
837 list = container_of(p, struct rhlist_head, rhead);
839 if (!rht_is_a_nulls(p)) {
843 return rht_obj(ht, rhlist ? &list->rhead : p);
846 /* At the end of this slot, switch to next one and then find
847 * next entry from that point.
853 return __rhashtable_walk_find_next(iter);
855 EXPORT_SYMBOL_GPL(rhashtable_walk_next);
858 * rhashtable_walk_peek - Return the next object but don't advance the iterator
859 * @iter: Hash table iterator
861 * Returns the next object or NULL when the end of the table is reached.
863 * Returns -EAGAIN if resize event occurred. Note that the iterator
864 * will rewind back to the beginning and you may continue to use it.
866 void *rhashtable_walk_peek(struct rhashtable_iter *iter)
868 struct rhlist_head *list = iter->list;
869 struct rhashtable *ht = iter->ht;
870 struct rhash_head *p = iter->p;
873 return rht_obj(ht, ht->rhlist ? &list->rhead : p);
875 /* No object found in current iter, find next one in the table. */
878 /* A nonzero skip value points to the next entry in the table
879 * beyond that last one that was found. Decrement skip so
880 * we find the current value. __rhashtable_walk_find_next
881 * will restore the original value of skip assuming that
882 * the table hasn't changed.
887 return __rhashtable_walk_find_next(iter);
889 EXPORT_SYMBOL_GPL(rhashtable_walk_peek);
892 * rhashtable_walk_stop - Finish a hash table walk
893 * @iter: Hash table iterator
895 * Finish a hash table walk. Does not reset the iterator to the start of the
898 void rhashtable_walk_stop(struct rhashtable_iter *iter)
901 struct rhashtable *ht;
902 struct bucket_table *tbl = iter->walker.tbl;
909 spin_lock(&ht->lock);
910 if (tbl->rehash < tbl->size)
911 list_add(&iter->walker.list, &tbl->walkers);
913 iter->walker.tbl = NULL;
914 spin_unlock(&ht->lock);
921 EXPORT_SYMBOL_GPL(rhashtable_walk_stop);
923 static size_t rounded_hashtable_size(const struct rhashtable_params *params)
925 return max(roundup_pow_of_two(params->nelem_hint * 4 / 3),
926 (unsigned long)params->min_size);
929 static u32 rhashtable_jhash2(const void *key, u32 length, u32 seed)
931 return jhash2(key, length, seed);
935 * rhashtable_init - initialize a new hash table
936 * @ht: hash table to be initialized
937 * @params: configuration parameters
939 * Initializes a new hash table based on the provided configuration
940 * parameters. A table can be configured either with a variable or
943 * Configuration Example 1: Fixed length keys
947 * struct rhash_head node;
950 * struct rhashtable_params params = {
951 * .head_offset = offsetof(struct test_obj, node),
952 * .key_offset = offsetof(struct test_obj, key),
953 * .key_len = sizeof(int),
955 * .nulls_base = (1U << RHT_BASE_SHIFT),
958 * Configuration Example 2: Variable length keys
961 * struct rhash_head node;
964 * u32 my_hash_fn(const void *data, u32 len, u32 seed)
966 * struct test_obj *obj = data;
968 * return [... hash ...];
971 * struct rhashtable_params params = {
972 * .head_offset = offsetof(struct test_obj, node),
974 * .obj_hashfn = my_hash_fn,
977 int rhashtable_init(struct rhashtable *ht,
978 const struct rhashtable_params *params)
980 struct bucket_table *tbl;
983 size = HASH_DEFAULT_SIZE;
985 if ((!params->key_len && !params->obj_hashfn) ||
986 (params->obj_hashfn && !params->obj_cmpfn))
989 if (params->nulls_base && params->nulls_base < (1U << RHT_BASE_SHIFT))
992 memset(ht, 0, sizeof(*ht));
993 mutex_init(&ht->mutex);
994 spin_lock_init(&ht->lock);
995 memcpy(&ht->p, params, sizeof(*params));
997 if (params->min_size)
998 ht->p.min_size = roundup_pow_of_two(params->min_size);
1000 /* Cap total entries at 2^31 to avoid nelems overflow. */
1001 ht->max_elems = 1u << 31;
1003 if (params->max_size) {
1004 ht->p.max_size = rounddown_pow_of_two(params->max_size);
1005 if (ht->p.max_size < ht->max_elems / 2)
1006 ht->max_elems = ht->p.max_size * 2;
1009 ht->p.min_size = max_t(u16, ht->p.min_size, HASH_MIN_SIZE);
1011 if (params->nelem_hint)
1012 size = rounded_hashtable_size(&ht->p);
1014 if (params->locks_mul)
1015 ht->p.locks_mul = roundup_pow_of_two(params->locks_mul);
1017 ht->p.locks_mul = BUCKET_LOCKS_PER_CPU;
1019 ht->key_len = ht->p.key_len;
1020 if (!params->hashfn) {
1021 ht->p.hashfn = jhash;
1023 if (!(ht->key_len & (sizeof(u32) - 1))) {
1024 ht->key_len /= sizeof(u32);
1025 ht->p.hashfn = rhashtable_jhash2;
1029 tbl = bucket_table_alloc(ht, size, GFP_KERNEL);
1033 atomic_set(&ht->nelems, 0);
1035 RCU_INIT_POINTER(ht->tbl, tbl);
1037 INIT_WORK(&ht->run_work, rht_deferred_worker);
1041 EXPORT_SYMBOL_GPL(rhashtable_init);
1044 * rhltable_init - initialize a new hash list table
1045 * @hlt: hash list table to be initialized
1046 * @params: configuration parameters
1048 * Initializes a new hash list table.
1050 * See documentation for rhashtable_init.
1052 int rhltable_init(struct rhltable *hlt, const struct rhashtable_params *params)
1056 /* No rhlist NULLs marking for now. */
1057 if (params->nulls_base)
1060 err = rhashtable_init(&hlt->ht, params);
1061 hlt->ht.rhlist = true;
1064 EXPORT_SYMBOL_GPL(rhltable_init);
1066 static void rhashtable_free_one(struct rhashtable *ht, struct rhash_head *obj,
1067 void (*free_fn)(void *ptr, void *arg),
1070 struct rhlist_head *list;
1073 free_fn(rht_obj(ht, obj), arg);
1077 list = container_of(obj, struct rhlist_head, rhead);
1080 list = rht_dereference(list->next, ht);
1081 free_fn(rht_obj(ht, obj), arg);
1086 * rhashtable_free_and_destroy - free elements and destroy hash table
1087 * @ht: the hash table to destroy
1088 * @free_fn: callback to release resources of element
1089 * @arg: pointer passed to free_fn
1091 * Stops an eventual async resize. If defined, invokes free_fn for each
1092 * element to releasal resources. Please note that RCU protected
1093 * readers may still be accessing the elements. Releasing of resources
1094 * must occur in a compatible manner. Then frees the bucket array.
1096 * This function will eventually sleep to wait for an async resize
1097 * to complete. The caller is responsible that no further write operations
1098 * occurs in parallel.
1100 void rhashtable_free_and_destroy(struct rhashtable *ht,
1101 void (*free_fn)(void *ptr, void *arg),
1104 struct bucket_table *tbl;
1107 cancel_work_sync(&ht->run_work);
1109 mutex_lock(&ht->mutex);
1110 tbl = rht_dereference(ht->tbl, ht);
1112 for (i = 0; i < tbl->size; i++) {
1113 struct rhash_head *pos, *next;
1115 for (pos = rht_dereference(*rht_bucket(tbl, i), ht),
1116 next = !rht_is_a_nulls(pos) ?
1117 rht_dereference(pos->next, ht) : NULL;
1118 !rht_is_a_nulls(pos);
1120 next = !rht_is_a_nulls(pos) ?
1121 rht_dereference(pos->next, ht) : NULL)
1122 rhashtable_free_one(ht, pos, free_fn, arg);
1126 bucket_table_free(tbl);
1127 mutex_unlock(&ht->mutex);
1129 EXPORT_SYMBOL_GPL(rhashtable_free_and_destroy);
1131 void rhashtable_destroy(struct rhashtable *ht)
1133 return rhashtable_free_and_destroy(ht, NULL, NULL);
1135 EXPORT_SYMBOL_GPL(rhashtable_destroy);
1137 struct rhash_head __rcu **rht_bucket_nested(const struct bucket_table *tbl,
1140 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1141 static struct rhash_head __rcu *rhnull =
1142 (struct rhash_head __rcu *)NULLS_MARKER(0);
1143 unsigned int index = hash & ((1 << tbl->nest) - 1);
1144 unsigned int size = tbl->size >> tbl->nest;
1145 unsigned int subhash = hash;
1146 union nested_table *ntbl;
1148 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1149 ntbl = rht_dereference_bucket_rcu(ntbl[index].table, tbl, hash);
1150 subhash >>= tbl->nest;
1152 while (ntbl && size > (1 << shift)) {
1153 index = subhash & ((1 << shift) - 1);
1154 ntbl = rht_dereference_bucket_rcu(ntbl[index].table,
1163 return &ntbl[subhash].bucket;
1166 EXPORT_SYMBOL_GPL(rht_bucket_nested);
1168 struct rhash_head __rcu **rht_bucket_nested_insert(struct rhashtable *ht,
1169 struct bucket_table *tbl,
1172 const unsigned int shift = PAGE_SHIFT - ilog2(sizeof(void *));
1173 unsigned int index = hash & ((1 << tbl->nest) - 1);
1174 unsigned int size = tbl->size >> tbl->nest;
1175 union nested_table *ntbl;
1176 unsigned int shifted;
1179 ntbl = (union nested_table *)rcu_dereference_raw(tbl->buckets[0]);
1182 shifted = tbl->nest;
1183 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1184 size <= (1 << shift) ? shifted : 0, nhash);
1186 while (ntbl && size > (1 << shift)) {
1187 index = hash & ((1 << shift) - 1);
1190 nhash |= index << shifted;
1192 ntbl = nested_table_alloc(ht, &ntbl[index].table,
1193 size <= (1 << shift) ? shifted : 0,
1200 return &ntbl[hash].bucket;
1203 EXPORT_SYMBOL_GPL(rht_bucket_nested_insert);