1 // SPDX-License-Identifier: GPL-2.0-only
2 #include <linux/spinlock.h>
3 #include <linux/slab.h>
4 #include <linux/list.h>
5 #include <linux/list_bl.h>
6 #include <linux/module.h>
7 #include <linux/sched.h>
8 #include <linux/workqueue.h>
9 #include <linux/mbcache.h>
12 * Mbcache is a simple key-value store. Keys need not be unique, however
13 * key-value pairs are expected to be unique (we use this fact in
14 * mb_cache_entry_delete_or_get()).
16 * Ext2 and ext4 use this cache for deduplication of extended attribute blocks.
17 * Ext4 also uses it for deduplication of xattr values stored in inodes.
18 * They use hash of data as a key and provide a value that may represent a
19 * block or inode number. That's why keys need not be unique (hash of different
20 * data may be the same). However user provided value always uniquely
21 * identifies a cache entry.
23 * We provide functions for creation and removal of entries, search by key,
24 * and a special "delete entry with given key-value pair" operation. Fixed
25 * size hash table is used for fast key lookups.
29 /* Hash table of entries */
30 struct hlist_bl_head *c_hash;
31 /* log2 of hash table size */
33 /* Maximum entries in cache to avoid degrading hash too much */
34 unsigned long c_max_entries;
35 /* Protects c_list, c_entry_count */
36 spinlock_t c_list_lock;
37 struct list_head c_list;
38 /* Number of entries in cache */
39 unsigned long c_entry_count;
40 struct shrinker *c_shrink;
41 /* Work for shrinking when the cache has too many entries */
42 struct work_struct c_shrink_work;
45 static struct kmem_cache *mb_entry_cache;
47 static unsigned long mb_cache_shrink(struct mb_cache *cache,
48 unsigned long nr_to_scan);
50 static inline struct hlist_bl_head *mb_cache_entry_head(struct mb_cache *cache,
53 return &cache->c_hash[hash_32(key, cache->c_bucket_bits)];
57 * Number of entries to reclaim synchronously when there are too many entries
60 #define SYNC_SHRINK_BATCH 64
63 * mb_cache_entry_create - create entry in cache
64 * @cache - cache where the entry should be created
65 * @mask - gfp mask with which the entry should be allocated
66 * @key - key of the entry
67 * @value - value of the entry
68 * @reusable - is the entry reusable by others?
70 * Creates entry in @cache with key @key and value @value. The function returns
71 * -EBUSY if entry with the same key and value already exists in cache.
72 * Otherwise 0 is returned.
74 int mb_cache_entry_create(struct mb_cache *cache, gfp_t mask, u32 key,
75 u64 value, bool reusable)
77 struct mb_cache_entry *entry, *dup;
78 struct hlist_bl_node *dup_node;
79 struct hlist_bl_head *head;
81 /* Schedule background reclaim if there are too many entries */
82 if (cache->c_entry_count >= cache->c_max_entries)
83 schedule_work(&cache->c_shrink_work);
84 /* Do some sync reclaim if background reclaim cannot keep up */
85 if (cache->c_entry_count >= 2*cache->c_max_entries)
86 mb_cache_shrink(cache, SYNC_SHRINK_BATCH);
88 entry = kmem_cache_alloc(mb_entry_cache, mask);
92 INIT_LIST_HEAD(&entry->e_list);
94 * We create entry with two references. One reference is kept by the
95 * hash table, the other reference is used to protect us from
96 * mb_cache_entry_delete_or_get() until the entry is fully setup. This
97 * avoids nesting of cache->c_list_lock into hash table bit locks which
98 * is problematic for RT.
100 atomic_set(&entry->e_refcnt, 2);
102 entry->e_value = value;
105 set_bit(MBE_REUSABLE_B, &entry->e_flags);
106 head = mb_cache_entry_head(cache, key);
108 hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) {
109 if (dup->e_key == key && dup->e_value == value) {
110 hlist_bl_unlock(head);
111 kmem_cache_free(mb_entry_cache, entry);
115 hlist_bl_add_head(&entry->e_hash_list, head);
116 hlist_bl_unlock(head);
117 spin_lock(&cache->c_list_lock);
118 list_add_tail(&entry->e_list, &cache->c_list);
119 cache->c_entry_count++;
120 spin_unlock(&cache->c_list_lock);
121 mb_cache_entry_put(cache, entry);
125 EXPORT_SYMBOL(mb_cache_entry_create);
127 void __mb_cache_entry_free(struct mb_cache *cache, struct mb_cache_entry *entry)
129 struct hlist_bl_head *head;
131 head = mb_cache_entry_head(cache, entry->e_key);
133 hlist_bl_del(&entry->e_hash_list);
134 hlist_bl_unlock(head);
135 kmem_cache_free(mb_entry_cache, entry);
137 EXPORT_SYMBOL(__mb_cache_entry_free);
140 * mb_cache_entry_wait_unused - wait to be the last user of the entry
142 * @entry - entry to work on
144 * Wait to be the last user of the entry.
146 void mb_cache_entry_wait_unused(struct mb_cache_entry *entry)
148 wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 2);
150 EXPORT_SYMBOL(mb_cache_entry_wait_unused);
152 static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
153 struct mb_cache_entry *entry,
156 struct mb_cache_entry *old_entry = entry;
157 struct hlist_bl_node *node;
158 struct hlist_bl_head *head;
160 head = mb_cache_entry_head(cache, key);
162 if (entry && !hlist_bl_unhashed(&entry->e_hash_list))
163 node = entry->e_hash_list.next;
165 node = hlist_bl_first(head);
167 entry = hlist_bl_entry(node, struct mb_cache_entry,
169 if (entry->e_key == key &&
170 test_bit(MBE_REUSABLE_B, &entry->e_flags) &&
171 atomic_inc_not_zero(&entry->e_refcnt))
177 hlist_bl_unlock(head);
179 mb_cache_entry_put(cache, old_entry);
185 * mb_cache_entry_find_first - find the first reusable entry with the given key
186 * @cache: cache where we should search
187 * @key: key to look for
189 * Search in @cache for a reusable entry with key @key. Grabs reference to the
190 * first reusable entry found and returns the entry.
192 struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
195 return __entry_find(cache, NULL, key);
197 EXPORT_SYMBOL(mb_cache_entry_find_first);
200 * mb_cache_entry_find_next - find next reusable entry with the same key
201 * @cache: cache where we should search
202 * @entry: entry to start search from
204 * Finds next reusable entry in the hash chain which has the same key as @entry.
205 * If @entry is unhashed (which can happen when deletion of entry races with the
206 * search), finds the first reusable entry in the hash chain. The function drops
207 * reference to @entry and returns with a reference to the found entry.
209 struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
210 struct mb_cache_entry *entry)
212 return __entry_find(cache, entry, entry->e_key);
214 EXPORT_SYMBOL(mb_cache_entry_find_next);
217 * mb_cache_entry_get - get a cache entry by value (and key)
218 * @cache - cache we work with
222 struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
225 struct hlist_bl_node *node;
226 struct hlist_bl_head *head;
227 struct mb_cache_entry *entry;
229 head = mb_cache_entry_head(cache, key);
231 hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
232 if (entry->e_key == key && entry->e_value == value &&
233 atomic_inc_not_zero(&entry->e_refcnt))
238 hlist_bl_unlock(head);
241 EXPORT_SYMBOL(mb_cache_entry_get);
243 /* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
244 * @cache - cache we work with
248 * Remove entry from cache @cache with key @key and value @value. The removal
249 * happens only if the entry is unused. The function returns NULL in case the
250 * entry was successfully removed or there's no entry in cache. Otherwise the
251 * function grabs reference of the entry that we failed to delete because it
252 * still has users and return it.
254 struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
257 struct mb_cache_entry *entry;
259 entry = mb_cache_entry_get(cache, key, value);
264 * Drop the ref we got from mb_cache_entry_get() and the initial hash
265 * ref if we are the last user
267 if (atomic_cmpxchg(&entry->e_refcnt, 2, 0) != 2)
270 spin_lock(&cache->c_list_lock);
271 if (!list_empty(&entry->e_list))
272 list_del_init(&entry->e_list);
273 cache->c_entry_count--;
274 spin_unlock(&cache->c_list_lock);
275 __mb_cache_entry_free(cache, entry);
278 EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
280 /* mb_cache_entry_touch - cache entry got used
281 * @cache - cache the entry belongs to
282 * @entry - entry that got used
284 * Marks entry as used to give hit higher chances of surviving in cache.
286 void mb_cache_entry_touch(struct mb_cache *cache,
287 struct mb_cache_entry *entry)
289 set_bit(MBE_REFERENCED_B, &entry->e_flags);
291 EXPORT_SYMBOL(mb_cache_entry_touch);
293 static unsigned long mb_cache_count(struct shrinker *shrink,
294 struct shrink_control *sc)
296 struct mb_cache *cache = shrink->private_data;
298 return cache->c_entry_count;
301 /* Shrink number of entries in cache */
302 static unsigned long mb_cache_shrink(struct mb_cache *cache,
303 unsigned long nr_to_scan)
305 struct mb_cache_entry *entry;
306 unsigned long shrunk = 0;
308 spin_lock(&cache->c_list_lock);
309 while (nr_to_scan-- && !list_empty(&cache->c_list)) {
310 entry = list_first_entry(&cache->c_list,
311 struct mb_cache_entry, e_list);
312 /* Drop initial hash reference if there is no user */
313 if (test_bit(MBE_REFERENCED_B, &entry->e_flags) ||
314 atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) {
315 clear_bit(MBE_REFERENCED_B, &entry->e_flags);
316 list_move_tail(&entry->e_list, &cache->c_list);
319 list_del_init(&entry->e_list);
320 cache->c_entry_count--;
321 spin_unlock(&cache->c_list_lock);
322 __mb_cache_entry_free(cache, entry);
325 spin_lock(&cache->c_list_lock);
327 spin_unlock(&cache->c_list_lock);
332 static unsigned long mb_cache_scan(struct shrinker *shrink,
333 struct shrink_control *sc)
335 struct mb_cache *cache = shrink->private_data;
336 return mb_cache_shrink(cache, sc->nr_to_scan);
339 /* We shrink 1/X of the cache when we have too many entries in it */
340 #define SHRINK_DIVISOR 16
342 static void mb_cache_shrink_worker(struct work_struct *work)
344 struct mb_cache *cache = container_of(work, struct mb_cache,
346 mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR);
350 * mb_cache_create - create cache
351 * @bucket_bits: log2 of the hash table size
353 * Create cache for keys with 2^bucket_bits hash entries.
355 struct mb_cache *mb_cache_create(int bucket_bits)
357 struct mb_cache *cache;
358 unsigned long bucket_count = 1UL << bucket_bits;
361 cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
364 cache->c_bucket_bits = bucket_bits;
365 cache->c_max_entries = bucket_count << 4;
366 INIT_LIST_HEAD(&cache->c_list);
367 spin_lock_init(&cache->c_list_lock);
368 cache->c_hash = kmalloc_array(bucket_count,
369 sizeof(struct hlist_bl_head),
371 if (!cache->c_hash) {
375 for (i = 0; i < bucket_count; i++)
376 INIT_HLIST_BL_HEAD(&cache->c_hash[i]);
378 cache->c_shrink = shrinker_alloc(0, "mbcache-shrinker");
379 if (!cache->c_shrink) {
380 kfree(cache->c_hash);
385 cache->c_shrink->count_objects = mb_cache_count;
386 cache->c_shrink->scan_objects = mb_cache_scan;
387 cache->c_shrink->private_data = cache;
389 shrinker_register(cache->c_shrink);
391 INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
398 EXPORT_SYMBOL(mb_cache_create);
401 * mb_cache_destroy - destroy cache
402 * @cache: the cache to destroy
404 * Free all entries in cache and cache itself. Caller must make sure nobody
405 * (except shrinker) can reach @cache when calling this.
407 void mb_cache_destroy(struct mb_cache *cache)
409 struct mb_cache_entry *entry, *next;
411 shrinker_free(cache->c_shrink);
414 * We don't bother with any locking. Cache must not be used at this
417 list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
418 list_del(&entry->e_list);
419 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
420 mb_cache_entry_put(cache, entry);
422 kfree(cache->c_hash);
425 EXPORT_SYMBOL(mb_cache_destroy);
427 static int __init mbcache_init(void)
429 mb_entry_cache = kmem_cache_create("mbcache",
430 sizeof(struct mb_cache_entry), 0,
431 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
437 static void __exit mbcache_exit(void)
439 kmem_cache_destroy(mb_entry_cache);
442 module_init(mbcache_init)
443 module_exit(mbcache_exit)
445 MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
446 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
447 MODULE_LICENSE("GPL");