2 * Resizable, Scalable, Concurrent Hash Table
4 * Copyright (c) 2014 Thomas Graf <tgraf@suug.ch>
5 * Copyright (c) 2008-2014 Patrick McHardy <kaber@trash.net>
7 * Based on the following paper by Josh Triplett, Paul E. McKenney
8 * and Jonathan Walpole:
9 * https://www.usenix.org/legacy/event/atc11/tech/final_files/Triplett.pdf
11 * Code partially derived from nft_hash
13 * This program is free software; you can redistribute it and/or modify
14 * it under the terms of the GNU General Public License version 2 as
15 * published by the Free Software Foundation.
18 #ifndef _LINUX_RHASHTABLE_H
19 #define _LINUX_RHASHTABLE_H
21 #include <linux/compiler.h>
22 #include <linux/list_nulls.h>
23 #include <linux/workqueue.h>
24 #include <linux/mutex.h>
25 #include <linux/rcupdate.h>
28 * The end of the chain is marked with a special nulls marks which has
29 * the following format:
31 * +-------+-----------------------------------------------------+-+
33 * +-------+-----------------------------------------------------+-+
35 * Base (4 bits) : Reserved to distinguish between multiple tables.
36 * Specified via &struct rhashtable_params.nulls_base.
37 * Hash (27 bits): Full hash (unmasked) of first element added to bucket
38 * 1 (1 bit) : Nulls marker (always set)
40 * The remaining bits of the next pointer remain unused for now.
42 #define RHT_BASE_BITS 4
43 #define RHT_HASH_BITS 27
44 #define RHT_BASE_SHIFT RHT_HASH_BITS
46 /* Base bits plus 1 bit for nulls marker */
47 #define RHT_HASH_RESERVED_SPACE (RHT_BASE_BITS + 1)
50 struct rhash_head __rcu *next;
54 * struct bucket_table - Table of hash buckets
55 * @size: Number of hash buckets
56 * @rehash: Current bucket being rehashed
57 * @hash_rnd: Random seed to fold into hash
58 * @locks_mask: Mask to apply before accessing locks[]
59 * @locks: Array of spinlocks protecting individual buckets
60 * @walkers: List of active walkers
61 * @rcu: RCU structure for freeing the table
62 * @future_tbl: Table under construction during rehashing
63 * @buckets: size * hash buckets
69 unsigned int locks_mask;
71 struct list_head walkers;
74 struct bucket_table __rcu *future_tbl;
76 struct rhash_head __rcu *buckets[] ____cacheline_aligned_in_smp;
80 * struct rhashtable_compare_arg - Key for the function rhashtable_compare
82 * @key: Key to compare against
84 struct rhashtable_compare_arg {
85 struct rhashtable *ht;
89 typedef u32 (*rht_hashfn_t)(const void *data, u32 len, u32 seed);
90 typedef u32 (*rht_obj_hashfn_t)(const void *data, u32 seed);
91 typedef int (*rht_obj_cmpfn_t)(struct rhashtable_compare_arg *arg,
97 * struct rhashtable_params - Hash table construction parameters
98 * @nelem_hint: Hint on number of elements, should be 75% of desired size
99 * @key_len: Length of key
100 * @key_offset: Offset of key in struct to be hashed
101 * @head_offset: Offset of rhash_head in struct to be hashed
102 * @max_size: Maximum size while expanding
103 * @min_size: Minimum size while shrinking
104 * @nulls_base: Base value to generate nulls marker
105 * @locks_mul: Number of bucket locks to allocate per cpu (default: 128)
106 * @hashfn: Function to hash key
107 * @obj_hashfn: Function to hash object
108 * @obj_cmpfn: Function to compare key with object
110 struct rhashtable_params {
115 unsigned int max_size;
116 unsigned int min_size;
120 rht_obj_hashfn_t obj_hashfn;
121 rht_obj_cmpfn_t obj_cmpfn;
125 * struct rhashtable - Hash table handle
127 * @nelems: Number of elements in table
128 * @p: Configuration parameters
129 * @run_work: Deferred worker to expand/shrink asynchronously
130 * @mutex: Mutex to protect current/future table swapping
131 * @being_destroyed: True if table is set up for destruction
134 struct bucket_table __rcu *tbl;
136 bool being_destroyed;
137 struct rhashtable_params p;
138 struct work_struct run_work;
143 * struct rhashtable_walker - Hash table walker
144 * @list: List entry on list of walkers
145 * @tbl: The table that we were walking over
147 struct rhashtable_walker {
148 struct list_head list;
149 struct bucket_table *tbl;
153 * struct rhashtable_iter - Hash table iterator, fits into netlink cb
154 * @ht: Table to iterate through
155 * @p: Current pointer
156 * @walker: Associated rhashtable walker
157 * @slot: Current slot
158 * @skip: Number of entries to skip in slot
160 struct rhashtable_iter {
161 struct rhashtable *ht;
162 struct rhash_head *p;
163 struct rhashtable_walker *walker;
168 static inline unsigned long rht_marker(const struct rhashtable *ht, u32 hash)
170 return NULLS_MARKER(ht->p.nulls_base + hash);
173 #define INIT_RHT_NULLS_HEAD(ptr, ht, hash) \
174 ((ptr) = (typeof(ptr)) rht_marker(ht, hash))
176 static inline bool rht_is_a_nulls(const struct rhash_head *ptr)
178 return ((unsigned long) ptr & 1);
181 static inline unsigned long rht_get_nulls_value(const struct rhash_head *ptr)
183 return ((unsigned long) ptr) >> 1;
186 static inline void *rht_obj(const struct rhashtable *ht,
187 const struct rhash_head *he)
189 return (char *)he - ht->p.head_offset;
192 static inline unsigned int rht_bucket_index(const struct bucket_table *tbl,
195 return (hash >> RHT_HASH_RESERVED_SPACE) & (tbl->size - 1);
198 static inline unsigned int rht_key_hashfn(
199 struct rhashtable *ht, const struct bucket_table *tbl,
200 const void *key, const struct rhashtable_params params)
202 return rht_bucket_index(tbl, params.hashfn(key, params.key_len ?:
207 static inline unsigned int rht_head_hashfn(
208 struct rhashtable *ht, const struct bucket_table *tbl,
209 const struct rhash_head *he, const struct rhashtable_params params)
211 const char *ptr = rht_obj(ht, he);
213 return likely(params.obj_hashfn) ?
214 rht_bucket_index(tbl, params.obj_hashfn(ptr, tbl->hash_rnd)) :
215 rht_key_hashfn(ht, tbl, ptr + params.key_offset, params);
219 * rht_grow_above_75 - returns true if nelems > 0.75 * table-size
221 * @tbl: current table
223 static inline bool rht_grow_above_75(const struct rhashtable *ht,
224 const struct bucket_table *tbl)
226 /* Expand table when exceeding 75% load */
227 return atomic_read(&ht->nelems) > (tbl->size / 4 * 3) &&
228 (!ht->p.max_size || tbl->size < ht->p.max_size);
232 * rht_shrink_below_30 - returns true if nelems < 0.3 * table-size
234 * @tbl: current table
236 static inline bool rht_shrink_below_30(const struct rhashtable *ht,
237 const struct bucket_table *tbl)
239 /* Shrink table beneath 30% load */
240 return atomic_read(&ht->nelems) < (tbl->size * 3 / 10) &&
241 tbl->size > ht->p.min_size;
244 /* The bucket lock is selected based on the hash and protects mutations
245 * on a group of hash buckets.
247 * A maximum of tbl->size/2 bucket locks is allocated. This ensures that
248 * a single lock always covers both buckets which may both contains
249 * entries which link to the same bucket of the old table during resizing.
250 * This allows to simplify the locking as locking the bucket in both
251 * tables during resize always guarantee protection.
253 * IMPORTANT: When holding the bucket lock of both the old and new table
254 * during expansions and shrinking, the old bucket lock must always be
257 static inline spinlock_t *rht_bucket_lock(const struct bucket_table *tbl,
260 return &tbl->locks[hash & tbl->locks_mask];
263 #ifdef CONFIG_PROVE_LOCKING
264 int lockdep_rht_mutex_is_held(struct rhashtable *ht);
265 int lockdep_rht_bucket_is_held(const struct bucket_table *tbl, u32 hash);
267 static inline int lockdep_rht_mutex_is_held(struct rhashtable *ht)
272 static inline int lockdep_rht_bucket_is_held(const struct bucket_table *tbl,
277 #endif /* CONFIG_PROVE_LOCKING */
279 int rhashtable_init(struct rhashtable *ht,
280 const struct rhashtable_params *params);
282 int rhashtable_insert_slow(struct rhashtable *ht, const void *key,
283 struct rhash_head *obj,
284 struct bucket_table *old_tbl);
285 void rhashtable_insert(struct rhashtable *ht, struct rhash_head *node);
286 bool rhashtable_remove(struct rhashtable *ht, struct rhash_head *node);
288 int rhashtable_expand(struct rhashtable *ht);
289 int rhashtable_shrink(struct rhashtable *ht);
291 void *rhashtable_lookup(struct rhashtable *ht, const void *key);
292 void *rhashtable_lookup_compare(struct rhashtable *ht, const void *key,
293 bool (*compare)(void *, void *), void *arg);
295 bool rhashtable_lookup_insert(struct rhashtable *ht, struct rhash_head *obj);
296 bool rhashtable_lookup_compare_insert(struct rhashtable *ht,
297 struct rhash_head *obj,
298 bool (*compare)(void *, void *),
301 int rhashtable_walk_init(struct rhashtable *ht, struct rhashtable_iter *iter);
302 void rhashtable_walk_exit(struct rhashtable_iter *iter);
303 int rhashtable_walk_start(struct rhashtable_iter *iter) __acquires(RCU);
304 void *rhashtable_walk_next(struct rhashtable_iter *iter);
305 void rhashtable_walk_stop(struct rhashtable_iter *iter) __releases(RCU);
307 void rhashtable_destroy(struct rhashtable *ht);
309 #define rht_dereference(p, ht) \
310 rcu_dereference_protected(p, lockdep_rht_mutex_is_held(ht))
312 #define rht_dereference_rcu(p, ht) \
313 rcu_dereference_check(p, lockdep_rht_mutex_is_held(ht))
315 #define rht_dereference_bucket(p, tbl, hash) \
316 rcu_dereference_protected(p, lockdep_rht_bucket_is_held(tbl, hash))
318 #define rht_dereference_bucket_rcu(p, tbl, hash) \
319 rcu_dereference_check(p, lockdep_rht_bucket_is_held(tbl, hash))
321 #define rht_entry(tpos, pos, member) \
322 ({ tpos = container_of(pos, typeof(*tpos), member); 1; })
325 * rht_for_each_continue - continue iterating over hash chain
326 * @pos: the &struct rhash_head to use as a loop cursor.
327 * @head: the previous &struct rhash_head to continue from
328 * @tbl: the &struct bucket_table
329 * @hash: the hash value / bucket index
331 #define rht_for_each_continue(pos, head, tbl, hash) \
332 for (pos = rht_dereference_bucket(head, tbl, hash); \
333 !rht_is_a_nulls(pos); \
334 pos = rht_dereference_bucket((pos)->next, tbl, hash))
337 * rht_for_each - iterate over hash chain
338 * @pos: the &struct rhash_head to use as a loop cursor.
339 * @tbl: the &struct bucket_table
340 * @hash: the hash value / bucket index
342 #define rht_for_each(pos, tbl, hash) \
343 rht_for_each_continue(pos, (tbl)->buckets[hash], tbl, hash)
346 * rht_for_each_entry_continue - continue iterating over hash chain
347 * @tpos: the type * to use as a loop cursor.
348 * @pos: the &struct rhash_head to use as a loop cursor.
349 * @head: the previous &struct rhash_head to continue from
350 * @tbl: the &struct bucket_table
351 * @hash: the hash value / bucket index
352 * @member: name of the &struct rhash_head within the hashable struct.
354 #define rht_for_each_entry_continue(tpos, pos, head, tbl, hash, member) \
355 for (pos = rht_dereference_bucket(head, tbl, hash); \
356 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
357 pos = rht_dereference_bucket((pos)->next, tbl, hash))
360 * rht_for_each_entry - iterate over hash chain of given type
361 * @tpos: the type * to use as a loop cursor.
362 * @pos: the &struct rhash_head to use as a loop cursor.
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(tpos, pos, tbl, hash, member) \
368 rht_for_each_entry_continue(tpos, pos, (tbl)->buckets[hash], \
372 * rht_for_each_entry_safe - safely iterate over hash chain of given type
373 * @tpos: the type * to use as a loop cursor.
374 * @pos: the &struct rhash_head to use as a loop cursor.
375 * @next: the &struct rhash_head to use as next in 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 * This hash chain list-traversal primitive allows for the looped code to
381 * remove the loop cursor from the list.
383 #define rht_for_each_entry_safe(tpos, pos, next, tbl, hash, member) \
384 for (pos = rht_dereference_bucket((tbl)->buckets[hash], tbl, hash), \
385 next = !rht_is_a_nulls(pos) ? \
386 rht_dereference_bucket(pos->next, tbl, hash) : NULL; \
387 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
389 next = !rht_is_a_nulls(pos) ? \
390 rht_dereference_bucket(pos->next, tbl, hash) : NULL)
393 * rht_for_each_rcu_continue - continue iterating over rcu hash chain
394 * @pos: the &struct rhash_head to use as a loop cursor.
395 * @head: the previous &struct rhash_head to continue from
396 * @tbl: the &struct bucket_table
397 * @hash: the hash value / bucket index
399 * This hash chain list-traversal primitive may safely run concurrently with
400 * the _rcu mutation primitives such as rhashtable_insert() as long as the
401 * traversal is guarded by rcu_read_lock().
403 #define rht_for_each_rcu_continue(pos, head, tbl, hash) \
404 for (({barrier(); }), \
405 pos = rht_dereference_bucket_rcu(head, tbl, hash); \
406 !rht_is_a_nulls(pos); \
407 pos = rcu_dereference_raw(pos->next))
410 * rht_for_each_rcu - iterate over rcu hash chain
411 * @pos: the &struct rhash_head to use as a loop cursor.
412 * @tbl: the &struct bucket_table
413 * @hash: the hash value / bucket index
415 * This hash chain list-traversal primitive may safely run concurrently with
416 * the _rcu mutation primitives such as rhashtable_insert() as long as the
417 * traversal is guarded by rcu_read_lock().
419 #define rht_for_each_rcu(pos, tbl, hash) \
420 rht_for_each_rcu_continue(pos, (tbl)->buckets[hash], tbl, hash)
423 * rht_for_each_entry_rcu_continue - continue iterating over rcu hash chain
424 * @tpos: the type * to use as a loop cursor.
425 * @pos: the &struct rhash_head to use as a loop cursor.
426 * @head: the previous &struct rhash_head to continue from
427 * @tbl: the &struct bucket_table
428 * @hash: the hash value / bucket index
429 * @member: name of the &struct rhash_head within the hashable struct.
431 * This hash chain list-traversal primitive may safely run concurrently with
432 * the _rcu mutation primitives such as rhashtable_insert() as long as the
433 * traversal is guarded by rcu_read_lock().
435 #define rht_for_each_entry_rcu_continue(tpos, pos, head, tbl, hash, member) \
436 for (({barrier(); }), \
437 pos = rht_dereference_bucket_rcu(head, tbl, hash); \
438 (!rht_is_a_nulls(pos)) && rht_entry(tpos, pos, member); \
439 pos = rht_dereference_bucket_rcu(pos->next, tbl, hash))
442 * rht_for_each_entry_rcu - iterate over rcu hash chain of given type
443 * @tpos: the type * to use as a loop cursor.
444 * @pos: the &struct rhash_head to use as a loop cursor.
445 * @tbl: the &struct bucket_table
446 * @hash: the hash value / bucket index
447 * @member: name of the &struct rhash_head within the hashable struct.
449 * This hash chain list-traversal primitive may safely run concurrently with
450 * the _rcu mutation primitives such as rhashtable_insert() as long as the
451 * traversal is guarded by rcu_read_lock().
453 #define rht_for_each_entry_rcu(tpos, pos, tbl, hash, member) \
454 rht_for_each_entry_rcu_continue(tpos, pos, (tbl)->buckets[hash],\
457 static inline int rhashtable_compare(struct rhashtable_compare_arg *arg,
460 struct rhashtable *ht = arg->ht;
461 const char *ptr = obj;
463 return memcmp(ptr + ht->p.key_offset, arg->key, ht->p.key_len);
467 * rhashtable_lookup_fast - search hash table, inlined version
469 * @key: the pointer to the key
470 * @params: hash table parameters
472 * Computes the hash value for the key and traverses the bucket chain looking
473 * for a entry with an identical key. The first matching entry is returned.
475 * Returns the first entry on which the compare function returned true.
477 static inline void *rhashtable_lookup_fast(
478 struct rhashtable *ht, const void *key,
479 const struct rhashtable_params params)
481 struct rhashtable_compare_arg arg = {
485 const struct bucket_table *tbl;
486 struct rhash_head *he;
491 tbl = rht_dereference_rcu(ht->tbl, ht);
493 hash = rht_key_hashfn(ht, tbl, key, params);
494 rht_for_each_rcu(he, tbl, hash) {
495 if (params.obj_cmpfn ?
496 params.obj_cmpfn(&arg, rht_obj(ht, he)) :
497 rhashtable_compare(&arg, rht_obj(ht, he)))
500 return rht_obj(ht, he);
503 /* Ensure we see any new tables. */
506 tbl = rht_dereference_rcu(tbl->future_tbl, ht);
514 static inline int __rhashtable_insert_fast(
515 struct rhashtable *ht, const void *key, struct rhash_head *obj,
516 const struct rhashtable_params params)
518 struct rhashtable_compare_arg arg = {
523 struct bucket_table *tbl, *new_tbl;
524 struct rhash_head *head;
530 tbl = rht_dereference_rcu(ht->tbl, ht);
531 hash = rht_head_hashfn(ht, tbl, obj, params);
532 lock = rht_bucket_lock(tbl, hash);
536 /* Because we have already taken the bucket lock in tbl,
537 * if we find that future_tbl is not yet visible then
538 * that guarantees all other insertions of the same entry
539 * will also grab the bucket lock in tbl because until
540 * the rehash completes ht->tbl won't be changed.
542 new_tbl = rht_dereference_rcu(tbl->future_tbl, ht);
543 if (unlikely(new_tbl)) {
544 err = rhashtable_insert_slow(ht, key, obj, new_tbl);
551 rht_for_each(head, tbl, hash) {
552 if (unlikely(!(params.obj_cmpfn ?
553 params.obj_cmpfn(&arg, rht_obj(ht, head)) :
554 rhashtable_compare(&arg, rht_obj(ht, head)))))
561 head = rht_dereference_bucket(tbl->buckets[hash], tbl, hash);
563 RCU_INIT_POINTER(obj->next, head);
565 rcu_assign_pointer(tbl->buckets[hash], obj);
567 atomic_inc(&ht->nelems);
568 if (rht_grow_above_75(ht, tbl))
569 schedule_work(&ht->run_work);
572 spin_unlock_bh(lock);
579 * rhashtable_insert_fast - insert object into hash table
581 * @obj: pointer to hash head inside object
582 * @params: hash table parameters
584 * Will take a per bucket spinlock to protect against mutual mutations
585 * on the same bucket. Multiple insertions may occur in parallel unless
586 * they map to the same bucket lock.
588 * It is safe to call this function from atomic context.
590 * Will trigger an automatic deferred table resizing if the size grows
591 * beyond the watermark indicated by grow_decision() which can be passed
592 * to rhashtable_init().
594 static inline int rhashtable_insert_fast(
595 struct rhashtable *ht, struct rhash_head *obj,
596 const struct rhashtable_params params)
598 return __rhashtable_insert_fast(ht, NULL, obj, params);
602 * rhashtable_lookup_insert_fast - lookup and insert object into hash table
604 * @obj: pointer to hash head inside object
605 * @params: hash table parameters
607 * Locks down the bucket chain in both the old and new table if a resize
608 * is in progress to ensure that writers can't remove from the old table
609 * and can't insert to the new table during the atomic operation of search
610 * and insertion. Searches for duplicates in both the old and new table if
611 * a resize is in progress.
613 * This lookup function may only be used for fixed key hash table (key_len
614 * parameter set). It will BUG() if used inappropriately.
616 * It is safe to call this function from atomic context.
618 * Will trigger an automatic deferred table resizing if the size grows
619 * beyond the watermark indicated by grow_decision() which can be passed
620 * to rhashtable_init().
622 static inline int rhashtable_lookup_insert_fast(
623 struct rhashtable *ht, struct rhash_head *obj,
624 const struct rhashtable_params params)
626 const char *key = rht_obj(ht, obj);
628 BUG_ON(ht->p.obj_hashfn);
630 return __rhashtable_insert_fast(ht, key + ht->p.key_offset, obj,
635 * rhashtable_lookup_insert_key - search and insert object to hash table
639 * @obj: pointer to hash head inside object
640 * @params: hash table parameters
642 * Locks down the bucket chain in both the old and new table if a resize
643 * is in progress to ensure that writers can't remove from the old table
644 * and can't insert to the new table during the atomic operation of search
645 * and insertion. Searches for duplicates in both the old and new table if
646 * a resize is in progress.
648 * Lookups may occur in parallel with hashtable mutations and resizing.
650 * Will trigger an automatic deferred table resizing if the size grows
651 * beyond the watermark indicated by grow_decision() which can be passed
652 * to rhashtable_init().
654 * Returns zero on success.
656 static inline int rhashtable_lookup_insert_key(
657 struct rhashtable *ht, const void *key, struct rhash_head *obj,
658 const struct rhashtable_params params)
660 BUG_ON(!ht->p.obj_hashfn || !key);
662 return __rhashtable_insert_fast(ht, key, obj, params);
665 static inline int __rhashtable_remove_fast(
666 struct rhashtable *ht, struct bucket_table *tbl,
667 struct rhash_head *obj, const struct rhashtable_params params)
669 struct rhash_head __rcu **pprev;
670 struct rhash_head *he;
675 hash = rht_head_hashfn(ht, tbl, obj, params);
676 lock = rht_bucket_lock(tbl, hash);
680 pprev = &tbl->buckets[hash];
681 rht_for_each(he, tbl, hash) {
687 rcu_assign_pointer(*pprev, obj->next);
692 spin_unlock_bh(lock);
698 * rhashtable_remove_fast - remove object from hash table
700 * @obj: pointer to hash head inside object
701 * @params: hash table parameters
703 * Since the hash chain is single linked, the removal operation needs to
704 * walk the bucket chain upon removal. The removal operation is thus
705 * considerable slow if the hash table is not correctly sized.
707 * Will automatically shrink the table via rhashtable_expand() if the
708 * shrink_decision function specified at rhashtable_init() returns true.
710 * Returns zero on success, -ENOENT if the entry could not be found.
712 static inline int rhashtable_remove_fast(
713 struct rhashtable *ht, struct rhash_head *obj,
714 const struct rhashtable_params params)
716 struct bucket_table *tbl;
721 tbl = rht_dereference_rcu(ht->tbl, ht);
723 /* Because we have already taken (and released) the bucket
724 * lock in old_tbl, if we find that future_tbl is not yet
725 * visible then that guarantees the entry to still be in
726 * the old tbl if it exists.
728 while ((err = __rhashtable_remove_fast(ht, tbl, obj, params)) &&
729 (tbl = rht_dereference_rcu(tbl->future_tbl, ht)))
735 atomic_dec(&ht->nelems);
736 if (rht_shrink_below_30(ht, tbl))
737 schedule_work(&ht->run_work);
745 #endif /* _LINUX_RHASHTABLE_H */