2 * Resizable, Scalable, Concurrent Hash Table
4 * Copyright (c) 2015 Herbert Xu <herbert@gondor.apana.org.au>
5 * Copyright (c) 2014 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 #ifndef _LINUX_RHASHTABLE_H
18 #define _LINUX_RHASHTABLE_H
20 #include <linux/compiler.h>
21 #include <linux/errno.h>
22 #include <linux/jhash.h>
23 #include <linux/list_nulls.h>
24 #include <linux/workqueue.h>
25 #include <linux/mutex.h>
26 #include <linux/rcupdate.h>
29 * The end of the chain is marked with a special nulls marks which has
30 * the following format:
32 * +-------+-----------------------------------------------------+-+
34 * +-------+-----------------------------------------------------+-+
36 * Base (4 bits) : Reserved to distinguish between multiple tables.
37 * Specified via &struct rhashtable_params.nulls_base.
38 * Hash (27 bits): Full hash (unmasked) of first element added to bucket
39 * 1 (1 bit) : Nulls marker (always set)
41 * The remaining bits of the next pointer remain unused for now.
43 #define RHT_BASE_BITS 4
44 #define RHT_HASH_BITS 27
45 #define RHT_BASE_SHIFT RHT_HASH_BITS
47 /* Base bits plus 1 bit for nulls marker */
48 #define RHT_HASH_RESERVED_SPACE (RHT_BASE_BITS + 1)
51 struct rhash_head __rcu *next;
55 * struct bucket_table - Table of hash buckets
56 * @size: Number of hash buckets
57 * @rehash: Current bucket being rehashed
58 * @hash_rnd: Random seed to fold into hash
59 * @locks_mask: Mask to apply before accessing locks[]
60 * @locks: Array of spinlocks protecting individual buckets
61 * @walkers: List of active walkers
62 * @rcu: RCU structure for freeing the table
63 * @future_tbl: Table under construction during rehashing
64 * @buckets: size * hash buckets
70 unsigned int locks_mask;
72 struct list_head walkers;
75 struct bucket_table __rcu *future_tbl;
77 struct rhash_head __rcu *buckets[] ____cacheline_aligned_in_smp;
81 * struct rhashtable_compare_arg - Key for the function rhashtable_compare
83 * @key: Key to compare against
85 struct rhashtable_compare_arg {
86 struct rhashtable *ht;
90 typedef u32 (*rht_hashfn_t)(const void *data, u32 len, u32 seed);
91 typedef u32 (*rht_obj_hashfn_t)(const void *data, u32 seed);
92 typedef int (*rht_obj_cmpfn_t)(struct rhashtable_compare_arg *arg,
98 * struct rhashtable_params - Hash table construction parameters
99 * @nelem_hint: Hint on number of elements, should be 75% of desired size
100 * @key_len: Length of key
101 * @key_offset: Offset of key in struct to be hashed
102 * @head_offset: Offset of rhash_head in struct to be hashed
103 * @max_size: Maximum size while expanding
104 * @min_size: Minimum size while shrinking
105 * @nulls_base: Base value to generate nulls marker
106 * @locks_mul: Number of bucket locks to allocate per cpu (default: 128)
107 * @hashfn: Hash function (default: jhash2 if !(key_len % 4), or jhash)
108 * @obj_hashfn: Function to hash object
109 * @obj_cmpfn: Function to compare key with object
111 struct rhashtable_params {
116 unsigned int max_size;
117 unsigned int min_size;
121 rht_obj_hashfn_t obj_hashfn;
122 rht_obj_cmpfn_t obj_cmpfn;
126 * struct rhashtable - Hash table handle
128 * @nelems: Number of elements in table
129 * @key_len: Key length for hashfn
130 * @p: Configuration parameters
131 * @run_work: Deferred worker to expand/shrink asynchronously
132 * @mutex: Mutex to protect current/future table swapping
133 * @being_destroyed: True if table is set up for destruction
136 struct bucket_table __rcu *tbl;
138 bool being_destroyed;
139 unsigned int key_len;
140 struct rhashtable_params p;
141 struct work_struct run_work;
146 * struct rhashtable_walker - Hash table walker
147 * @list: List entry on list of walkers
148 * @tbl: The table that we were walking over
150 struct rhashtable_walker {
151 struct list_head list;
152 struct bucket_table *tbl;
156 * struct rhashtable_iter - Hash table iterator, fits into netlink cb
157 * @ht: Table to iterate through
158 * @p: Current pointer
159 * @walker: Associated rhashtable walker
160 * @slot: Current slot
161 * @skip: Number of entries to skip in slot
163 struct rhashtable_iter {
164 struct rhashtable *ht;
165 struct rhash_head *p;
166 struct rhashtable_walker *walker;
171 static inline unsigned long rht_marker(const struct rhashtable *ht, u32 hash)
173 return NULLS_MARKER(ht->p.nulls_base + hash);
176 #define INIT_RHT_NULLS_HEAD(ptr, ht, hash) \
177 ((ptr) = (typeof(ptr)) rht_marker(ht, hash))
179 static inline bool rht_is_a_nulls(const struct rhash_head *ptr)
181 return ((unsigned long) ptr & 1);
184 static inline unsigned long rht_get_nulls_value(const struct rhash_head *ptr)
186 return ((unsigned long) ptr) >> 1;
189 static inline void *rht_obj(const struct rhashtable *ht,
190 const struct rhash_head *he)
192 return (char *)he - ht->p.head_offset;
195 static inline unsigned int rht_bucket_index(const struct bucket_table *tbl,
198 return (hash >> RHT_HASH_RESERVED_SPACE) & (tbl->size - 1);
201 static inline unsigned int rht_key_hashfn(
202 struct rhashtable *ht, const struct bucket_table *tbl,
203 const void *key, const struct rhashtable_params params)
207 /* params must be equal to ht->p if it isn't constant. */
208 if (!__builtin_constant_p(params.key_len))
209 hash = ht->p.hashfn(key, ht->key_len, tbl->hash_rnd);
210 else if (params.key_len) {
211 unsigned key_len = params.key_len;
214 hash = params.hashfn(key, key_len, tbl->hash_rnd);
215 else if (key_len & (sizeof(u32) - 1))
216 hash = jhash(key, key_len, tbl->hash_rnd);
218 hash = jhash2(key, key_len / sizeof(u32),
221 unsigned key_len = ht->p.key_len;
224 hash = params.hashfn(key, key_len, tbl->hash_rnd);
226 hash = jhash(key, key_len, tbl->hash_rnd);
229 return rht_bucket_index(tbl, hash);
232 static inline unsigned int rht_head_hashfn(
233 struct rhashtable *ht, const struct bucket_table *tbl,
234 const struct rhash_head *he, const struct rhashtable_params params)
236 const char *ptr = rht_obj(ht, he);
238 return likely(params.obj_hashfn) ?
239 rht_bucket_index(tbl, params.obj_hashfn(ptr, tbl->hash_rnd)) :
240 rht_key_hashfn(ht, tbl, ptr + params.key_offset, params);
244 * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
246 * @tbl: current table
248 static inline bool rht_grow_above_75(const struct rhashtable *ht,
249 const struct bucket_table *tbl)
251 /* Expand table when exceeding 75% load */
252 return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) &&
253 (!ht->p.max_size || tbl->size < ht->p.max_size);
257 * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
259 * @tbl: current table
261 static inline bool rht_shrink_below_30(const struct rhashtable *ht,
262 const struct bucket_table *tbl)
264 /* Shrink table beneath 30% load */
265 return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) &&
266 tbl->size > ht->p.min_size;
269 /* The bucket lock is selected based on the hash and protects mutations
270 * on a group of hash buckets.
272 * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
273 * a single lock always covers both buckets which may both contains
274 * entries which link to the same bucket of the old table during resizing.
275 * This allows to simplify the locking as locking the bucket in both
276 * tables during resize always guarantee protection.
278 * IMPORTANT: When holding the bucket lock of both the old and new table
279 * during expansions and shrinking, the old bucket lock must always be
282 static inline spinlock_t *rht_bucket_lock(const struct bucket_table *tbl,
285 return &tbl->locks[hash & tbl->locks_mask];
288 #ifdef CONFIG_PROVE_LOCKING
289 int lockdep_rht_mutex_is_held(struct rhashtable *ht);
290 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash);
292 static inline int lockdep_rht_mutex_is_held(struct rhashtable *ht)
297 static inline int lockdep_rht_bucket_is_held(const struct bucket_table *tbl,
302 #endif /* CONFIG_PROVE_LOCKING */
304 int rhashtable_init(struct rhashtable *ht,
305 const struct rhashtable_params *params);
307 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
308 struct rhash_head *obj,
309 struct bucket_table *old_tbl);
311 int rhashtable_expand(struct rhashtable *ht);
312 int rhashtable_shrink(struct rhashtable *ht);
314 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter);
315 void rhashtable_walk_exit(struct rhashtable_iter *iter);
316 int rhashtable_walk_start(struct rhashtable_iter *iter) __acquires(RCU);
317 void *rhashtable_walk_next(struct rhashtable_iter *iter);
318 void rhashtable_walk_stop(struct rhashtable_iter *iter) __releases(RCU);
320 void rhashtable_destroy(struct rhashtable *ht);
322 #define rht_dereference(p, ht) \
323 rcu_dereference_protected(p, lockdep_rht_mutex_is_held(ht))
325 #define rht_dereference_rcu(p, ht) \
326 rcu_dereference_check(p, lockdep_rht_mutex_is_held(ht))
328 #define rht_dereference_bucket(p, tbl, hash) \
329 rcu_dereference_protected(p, lockdep_rht_bucket_is_held(tbl, hash))
331 #define rht_dereference_bucket_rcu(p, tbl, hash) \
332 rcu_dereference_check(p, lockdep_rht_bucket_is_held(tbl, hash))
334 #define rht_entry(tpos, pos, member) \
335 ({ tpos = container_of(pos, typeof(*tpos), member); 1; })
338 * rht_for_each_continue - continue iterating over hash chain
339 * @pos: the &struct rhash_head to use as a loop cursor.
340 * @head: the previous &struct rhash_head to continue from
341 * @tbl: the &struct bucket_table
342 * @hash: the hash value / bucket index
344 #define rht_for_each_continue(pos, head, tbl, hash) \
345 for (pos = rht_dereference_bucket(head, tbl, hash); \
346 !rht_is_a_nulls(pos); \
347 pos = rht_dereference_bucket((pos)->next, tbl, hash))
350 * rht_for_each - iterate over hash chain
351 * @pos: the &struct rhash_head to use as a loop cursor.
352 * @tbl: the &struct bucket_table
353 * @hash: the hash value / bucket index
355 #define rht_for_each(pos, tbl, hash) \
356 rht_for_each_continue(pos, (tbl)->buckets[hash], tbl, hash)
359 * rht_for_each_entry_continue - continue iterating over hash chain
360 * @tpos: the type * to use as a loop cursor.
361 * @pos: the &struct rhash_head to use as a loop cursor.
362 * @head: the previous &struct rhash_head to continue from
363 * @tbl: the &struct bucket_table
364 * @hash: the hash value / bucket index
365 * @member: name of the &struct rhash_head within the hashable struct.
367 #define rht_for_each_entry_continue(tpos, pos, head, tbl, hash, member) \
368 for (pos = rht_dereference_bucket(head, tbl, hash); \
369 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
370 pos = rht_dereference_bucket((pos)->next, tbl, hash))
373 * rht_for_each_entry - iterate over hash chain of given type
374 * @tpos: the type * to use as a loop cursor.
375 * @pos: the &struct rhash_head to use as a loop cursor.
376 * @tbl: the &struct bucket_table
377 * @hash: the hash value / bucket index
378 * @member: name of the &struct rhash_head within the hashable struct.
380 #define rht_for_each_entry(tpos, pos, tbl, hash, member) \
381 rht_for_each_entry_continue(tpos, pos, (tbl)->buckets[hash], \
385 * rht_for_each_entry_safe - safely iterate over hash chain of given type
386 * @tpos: the type * to use as a loop cursor.
387 * @pos: the &struct rhash_head to use as a loop cursor.
388 * @next: the &struct rhash_head to use as next in loop cursor.
389 * @tbl: the &struct bucket_table
390 * @hash: the hash value / bucket index
391 * @member: name of the &struct rhash_head within the hashable struct.
393 * This hash chain list-traversal primitive allows for the looped code to
394 * remove the loop cursor from the list.
396 #define rht_for_each_entry_safe(tpos, pos, next, tbl, hash, member) \
397 for (pos = rht_dereference_bucket((tbl)->buckets[hash], tbl, hash), \
398 next = !rht_is_a_nulls(pos) ? \
399 rht_dereference_bucket(pos->next, tbl, hash) : NULL; \
400 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
402 next = !rht_is_a_nulls(pos) ? \
403 rht_dereference_bucket(pos->next, tbl, hash) : NULL)
406 * rht_for_each_rcu_continue - continue iterating over rcu hash chain
407 * @pos: the &struct rhash_head to use as a loop cursor.
408 * @head: the previous &struct rhash_head to continue from
409 * @tbl: the &struct bucket_table
410 * @hash: the hash value / bucket index
412 * This hash chain list-traversal primitive may safely run concurrently with
413 * the _rcu mutation primitives such as rhashtable_insert() as long as the
414 * traversal is guarded by rcu_read_lock().
416 #define rht_for_each_rcu_continue(pos, head, tbl, hash) \
417 for (({barrier(); }), \
418 pos = rht_dereference_bucket_rcu(head, tbl, hash); \
419 !rht_is_a_nulls(pos); \
420 pos = rcu_dereference_raw(pos->next))
423 * rht_for_each_rcu - iterate over rcu hash chain
424 * @pos: the &struct rhash_head to use as a loop cursor.
425 * @tbl: the &struct bucket_table
426 * @hash: the hash value / bucket index
428 * This hash chain list-traversal primitive may safely run concurrently with
429 * the _rcu mutation primitives such as rhashtable_insert() as long as the
430 * traversal is guarded by rcu_read_lock().
432 #define rht_for_each_rcu(pos, tbl, hash) \
433 rht_for_each_rcu_continue(pos, (tbl)->buckets[hash], tbl, hash)
436 * rht_for_each_entry_rcu_continue - continue iterating over rcu hash chain
437 * @tpos: the type * to use as a loop cursor.
438 * @pos: the &struct rhash_head to use as a loop cursor.
439 * @head: the previous &struct rhash_head to continue from
440 * @tbl: the &struct bucket_table
441 * @hash: the hash value / bucket index
442 * @member: name of the &struct rhash_head within the hashable struct.
444 * This hash chain list-traversal primitive may safely run concurrently with
445 * the _rcu mutation primitives such as rhashtable_insert() as long as the
446 * traversal is guarded by rcu_read_lock().
448 #define rht_for_each_entry_rcu_continue(tpos, pos, head, tbl, hash, member) \
449 for (({barrier(); }), \
450 pos = rht_dereference_bucket_rcu(head, tbl, hash); \
451 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
452 pos = rht_dereference_bucket_rcu(pos->next, tbl, hash))
455 * rht_for_each_entry_rcu - iterate over rcu hash chain of given type
456 * @tpos: the type * to use as a loop cursor.
457 * @pos: the &struct rhash_head to use as a loop cursor.
458 * @tbl: the &struct bucket_table
459 * @hash: the hash value / bucket index
460 * @member: name of the &struct rhash_head within the hashable struct.
462 * This hash chain list-traversal primitive may safely run concurrently with
463 * the _rcu mutation primitives such as rhashtable_insert() as long as the
464 * traversal is guarded by rcu_read_lock().
466 #define rht_for_each_entry_rcu(tpos, pos, tbl, hash, member) \
467 rht_for_each_entry_rcu_continue(tpos, pos, (tbl)->buckets[hash],\
470 static inline int rhashtable_compare(struct rhashtable_compare_arg *arg,
473 struct rhashtable *ht = arg->ht;
474 const char *ptr = obj;
476 return memcmp(ptr + ht->p.key_offset, arg->key, ht->p.key_len);
480 * rhashtable_lookup_fast - search hash table, inlined version
482 * @key: the pointer to the key
483 * @params: hash table parameters
485 * Computes the hash value for the key and traverses the bucket chain looking
486 * for a entry with an identical key. The first matching entry is returned.
488 * Returns the first entry on which the compare function returned true.
490 static inline void *rhashtable_lookup_fast(
491 struct rhashtable *ht, const void *key,
492 const struct rhashtable_params params)
494 struct rhashtable_compare_arg arg = {
498 const struct bucket_table *tbl;
499 struct rhash_head *he;
504 tbl = rht_dereference_rcu(ht->tbl, ht);
506 hash = rht_key_hashfn(ht, tbl, key, params);
507 rht_for_each_rcu(he, tbl, hash) {
508 if (params.obj_cmpfn ?
509 params.obj_cmpfn(&arg, rht_obj(ht, he)) :
510 rhashtable_compare(&arg, rht_obj(ht, he)))
513 return rht_obj(ht, he);
516 /* Ensure we see any new tables. */
519 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
527 static inline int __rhashtable_insert_fast(
528 struct rhashtable *ht, const void *key, struct rhash_head *obj,
529 const struct rhashtable_params params)
531 struct rhashtable_compare_arg arg = {
536 struct bucket_table *tbl, *new_tbl;
537 struct rhash_head *head;
543 tbl = rht_dereference_rcu(ht->tbl, ht);
544 hash = rht_head_hashfn(ht, tbl, obj, params);
545 lock = rht_bucket_lock(tbl, hash);
549 /* Because we have already taken the bucket lock in tbl,
550 * if we find that future_tbl is not yet visible then
551 * that guarantees all other insertions of the same entry
552 * will also grab the bucket lock in tbl because until
553 * the rehash completes ht->tbl won't be changed.
555 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
556 if (unlikely(new_tbl)) {
557 err = rhashtable_insert_slow(ht, key, obj, new_tbl);
564 rht_for_each(head, tbl, hash) {
565 if (unlikely(!(params.obj_cmpfn ?
566 params.obj_cmpfn(&arg, rht_obj(ht, head)) :
567 rhashtable_compare(&arg, rht_obj(ht, head)))))
574 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
576 RCU_INIT_POINTER(obj->next, head);
578 rcu_assign_pointer(tbl->buckets[hash], obj);
580 atomic_inc(&ht->nelems);
581 if (rht_grow_above_75(ht, tbl))
582 schedule_work(&ht->run_work);
585 spin_unlock_bh(lock);
592 * rhashtable_insert_fast - insert object into hash table
594 * @obj: pointer to hash head inside object
595 * @params: hash table parameters
597 * Will take a per bucket spinlock to protect against mutual mutations
598 * on the same bucket. Multiple insertions may occur in parallel unless
599 * they map to the same bucket lock.
601 * It is safe to call this function from atomic context.
603 * Will trigger an automatic deferred table resizing if the size grows
604 * beyond the watermark indicated by grow_decision() which can be passed
605 * to rhashtable_init().
607 static inline int rhashtable_insert_fast(
608 struct rhashtable *ht, struct rhash_head *obj,
609 const struct rhashtable_params params)
611 return __rhashtable_insert_fast(ht, NULL, obj, params);
615 * rhashtable_lookup_insert_fast - lookup and insert object into hash table
617 * @obj: pointer to hash head inside object
618 * @params: hash table parameters
620 * Locks down the bucket chain in both the old and new table if a resize
621 * is in progress to ensure that writers can't remove from the old table
622 * and can't insert to the new table during the atomic operation of search
623 * and insertion. Searches for duplicates in both the old and new table if
624 * a resize is in progress.
626 * This lookup function may only be used for fixed key hash table (key_len
627 * parameter set). It will BUG() if used inappropriately.
629 * It is safe to call this function from atomic context.
631 * Will trigger an automatic deferred table resizing if the size grows
632 * beyond the watermark indicated by grow_decision() which can be passed
633 * to rhashtable_init().
635 static inline int rhashtable_lookup_insert_fast(
636 struct rhashtable *ht, struct rhash_head *obj,
637 const struct rhashtable_params params)
639 const char *key = rht_obj(ht, obj);
641 BUG_ON(ht->p.obj_hashfn);
643 return __rhashtable_insert_fast(ht, key + ht->p.key_offset, obj,
648 * rhashtable_lookup_insert_key - search and insert object to hash table
652 * @obj: pointer to hash head inside object
653 * @params: hash table parameters
655 * Locks down the bucket chain in both the old and new table if a resize
656 * is in progress to ensure that writers can't remove from the old table
657 * and can't insert to the new table during the atomic operation of search
658 * and insertion. Searches for duplicates in both the old and new table if
659 * a resize is in progress.
661 * Lookups may occur in parallel with hashtable mutations and resizing.
663 * Will trigger an automatic deferred table resizing if the size grows
664 * beyond the watermark indicated by grow_decision() which can be passed
665 * to rhashtable_init().
667 * Returns zero on success.
669 static inline int rhashtable_lookup_insert_key(
670 struct rhashtable *ht, const void *key, struct rhash_head *obj,
671 const struct rhashtable_params params)
673 BUG_ON(!ht->p.obj_hashfn || !key);
675 return __rhashtable_insert_fast(ht, key, obj, params);
678 static inline int __rhashtable_remove_fast(
679 struct rhashtable *ht, struct bucket_table *tbl,
680 struct rhash_head *obj, const struct rhashtable_params params)
682 struct rhash_head __rcu **pprev;
683 struct rhash_head *he;
688 hash = rht_head_hashfn(ht, tbl, obj, params);
689 lock = rht_bucket_lock(tbl, hash);
693 pprev = &tbl->buckets[hash];
694 rht_for_each(he, tbl, hash) {
700 rcu_assign_pointer(*pprev, obj->next);
705 spin_unlock_bh(lock);
711 * rhashtable_remove_fast - remove object from hash table
713 * @obj: pointer to hash head inside object
714 * @params: hash table parameters
716 * Since the hash chain is single linked, the removal operation needs to
717 * walk the bucket chain upon removal. The removal operation is thus
718 * considerable slow if the hash table is not correctly sized.
720 * Will automatically shrink the table via rhashtable_expand() if the
721 * shrink_decision function specified at rhashtable_init() returns true.
723 * Returns zero on success, -ENOENT if the entry could not be found.
725 static inline int rhashtable_remove_fast(
726 struct rhashtable *ht, struct rhash_head *obj,
727 const struct rhashtable_params params)
729 struct bucket_table *tbl;
734 tbl = rht_dereference_rcu(ht->tbl, ht);
736 /* Because we have already taken (and released) the bucket
737 * lock in old_tbl, if we find that future_tbl is not yet
738 * visible then that guarantees the entry to still be in
739 * the old tbl if it exists.
741 while ((err = __rhashtable_remove_fast(ht, tbl, obj, params)) &&
742 (tbl = rht_dereference_rcu(tbl->future_tbl, ht)))
748 atomic_dec(&ht->nelems);
749 if (rht_shrink_below_30(ht, tbl))
750 schedule_work(&ht->run_work);
758 #endif /* _LINUX_RHASHTABLE_H */