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>
10
11 /*
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()).
15 *
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.
22 *
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.
26 */
27
28 struct mb_cache {
29 /* Hash table of entries */
30 struct hlist_bl_head *c_hash;
31 /* log2 of hash table size */
32 int c_bucket_bits;
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;
43 };
44
45 static struct kmem_cache *mb_entry_cache;
46
47 static unsigned long mb_cache_shrink(struct mb_cache *cache,
48 unsigned long nr_to_scan);
49
mb_cache_entry_head(struct mb_cache * cache,u32 key)50 static inline struct hlist_bl_head *mb_cache_entry_head(struct mb_cache *cache,
51 u32 key)
52 {
53 return &cache->c_hash[hash_32(key, cache->c_bucket_bits)];
54 }
55
56 /*
57 * Number of entries to reclaim synchronously when there are too many entries
58 * in cache
59 */
60 #define SYNC_SHRINK_BATCH 64
61
62 /*
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?
69 *
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.
73 */
mb_cache_entry_create(struct mb_cache * cache,gfp_t mask,u32 key,u64 value,bool reusable)74 int mb_cache_entry_create(struct mb_cache *cache, gfp_t mask, u32 key,
75 u64 value, bool reusable)
76 {
77 struct mb_cache_entry *entry, *dup;
78 struct hlist_bl_node *dup_node;
79 struct hlist_bl_head *head;
80
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);
87
88 entry = kmem_cache_alloc(mb_entry_cache, mask);
89 if (!entry)
90 return -ENOMEM;
91
92 INIT_LIST_HEAD(&entry->e_list);
93 /*
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.
99 */
100 atomic_set(&entry->e_refcnt, 2);
101 entry->e_key = key;
102 entry->e_value = value;
103 entry->e_reusable = reusable;
104 entry->e_referenced = 0;
105 head = mb_cache_entry_head(cache, key);
106 hlist_bl_lock(head);
107 hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) {
108 if (dup->e_key == key && dup->e_value == value) {
109 hlist_bl_unlock(head);
110 kmem_cache_free(mb_entry_cache, entry);
111 return -EBUSY;
112 }
113 }
114 hlist_bl_add_head(&entry->e_hash_list, head);
115 hlist_bl_unlock(head);
116 spin_lock(&cache->c_list_lock);
117 list_add_tail(&entry->e_list, &cache->c_list);
118 cache->c_entry_count++;
119 spin_unlock(&cache->c_list_lock);
120 mb_cache_entry_put(cache, entry);
121
122 return 0;
123 }
124 EXPORT_SYMBOL(mb_cache_entry_create);
125
__mb_cache_entry_free(struct mb_cache * cache,struct mb_cache_entry * entry)126 void __mb_cache_entry_free(struct mb_cache *cache, struct mb_cache_entry *entry)
127 {
128 struct hlist_bl_head *head;
129
130 head = mb_cache_entry_head(cache, entry->e_key);
131 hlist_bl_lock(head);
132 hlist_bl_del(&entry->e_hash_list);
133 hlist_bl_unlock(head);
134 kmem_cache_free(mb_entry_cache, entry);
135 }
136 EXPORT_SYMBOL(__mb_cache_entry_free);
137
138 /*
139 * mb_cache_entry_wait_unused - wait to be the last user of the entry
140 *
141 * @entry - entry to work on
142 *
143 * Wait to be the last user of the entry.
144 */
mb_cache_entry_wait_unused(struct mb_cache_entry * entry)145 void mb_cache_entry_wait_unused(struct mb_cache_entry *entry)
146 {
147 wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 2);
148 }
149 EXPORT_SYMBOL(mb_cache_entry_wait_unused);
150
__entry_find(struct mb_cache * cache,struct mb_cache_entry * entry,u32 key)151 static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
152 struct mb_cache_entry *entry,
153 u32 key)
154 {
155 struct mb_cache_entry *old_entry = entry;
156 struct hlist_bl_node *node;
157 struct hlist_bl_head *head;
158
159 head = mb_cache_entry_head(cache, key);
160 hlist_bl_lock(head);
161 if (entry && !hlist_bl_unhashed(&entry->e_hash_list))
162 node = entry->e_hash_list.next;
163 else
164 node = hlist_bl_first(head);
165 while (node) {
166 entry = hlist_bl_entry(node, struct mb_cache_entry,
167 e_hash_list);
168 if (entry->e_key == key && entry->e_reusable &&
169 atomic_inc_not_zero(&entry->e_refcnt))
170 goto out;
171 node = node->next;
172 }
173 entry = NULL;
174 out:
175 hlist_bl_unlock(head);
176 if (old_entry)
177 mb_cache_entry_put(cache, old_entry);
178
179 return entry;
180 }
181
182 /*
183 * mb_cache_entry_find_first - find the first reusable entry with the given key
184 * @cache: cache where we should search
185 * @key: key to look for
186 *
187 * Search in @cache for a reusable entry with key @key. Grabs reference to the
188 * first reusable entry found and returns the entry.
189 */
mb_cache_entry_find_first(struct mb_cache * cache,u32 key)190 struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
191 u32 key)
192 {
193 return __entry_find(cache, NULL, key);
194 }
195 EXPORT_SYMBOL(mb_cache_entry_find_first);
196
197 /*
198 * mb_cache_entry_find_next - find next reusable entry with the same key
199 * @cache: cache where we should search
200 * @entry: entry to start search from
201 *
202 * Finds next reusable entry in the hash chain which has the same key as @entry.
203 * If @entry is unhashed (which can happen when deletion of entry races with the
204 * search), finds the first reusable entry in the hash chain. The function drops
205 * reference to @entry and returns with a reference to the found entry.
206 */
mb_cache_entry_find_next(struct mb_cache * cache,struct mb_cache_entry * entry)207 struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
208 struct mb_cache_entry *entry)
209 {
210 return __entry_find(cache, entry, entry->e_key);
211 }
212 EXPORT_SYMBOL(mb_cache_entry_find_next);
213
214 /*
215 * mb_cache_entry_get - get a cache entry by value (and key)
216 * @cache - cache we work with
217 * @key - key
218 * @value - value
219 */
mb_cache_entry_get(struct mb_cache * cache,u32 key,u64 value)220 struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
221 u64 value)
222 {
223 struct hlist_bl_node *node;
224 struct hlist_bl_head *head;
225 struct mb_cache_entry *entry;
226
227 head = mb_cache_entry_head(cache, key);
228 hlist_bl_lock(head);
229 hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
230 if (entry->e_key == key && entry->e_value == value &&
231 atomic_inc_not_zero(&entry->e_refcnt))
232 goto out;
233 }
234 entry = NULL;
235 out:
236 hlist_bl_unlock(head);
237 return entry;
238 }
239 EXPORT_SYMBOL(mb_cache_entry_get);
240
241 /* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
242 * @cache - cache we work with
243 * @key - key
244 * @value - value
245 *
246 * Remove entry from cache @cache with key @key and value @value. The removal
247 * happens only if the entry is unused. The function returns NULL in case the
248 * entry was successfully removed or there's no entry in cache. Otherwise the
249 * function grabs reference of the entry that we failed to delete because it
250 * still has users and return it.
251 */
mb_cache_entry_delete_or_get(struct mb_cache * cache,u32 key,u64 value)252 struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
253 u32 key, u64 value)
254 {
255 struct mb_cache_entry *entry;
256
257 entry = mb_cache_entry_get(cache, key, value);
258 if (!entry)
259 return NULL;
260
261 /*
262 * Drop the ref we got from mb_cache_entry_get() and the initial hash
263 * ref if we are the last user
264 */
265 if (atomic_cmpxchg(&entry->e_refcnt, 2, 0) != 2)
266 return entry;
267
268 spin_lock(&cache->c_list_lock);
269 if (!list_empty(&entry->e_list))
270 list_del_init(&entry->e_list);
271 cache->c_entry_count--;
272 spin_unlock(&cache->c_list_lock);
273 __mb_cache_entry_free(cache, entry);
274 return NULL;
275 }
276 EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
277
278 /* mb_cache_entry_touch - cache entry got used
279 * @cache - cache the entry belongs to
280 * @entry - entry that got used
281 *
282 * Marks entry as used to give hit higher chances of surviving in cache.
283 */
mb_cache_entry_touch(struct mb_cache * cache,struct mb_cache_entry * entry)284 void mb_cache_entry_touch(struct mb_cache *cache,
285 struct mb_cache_entry *entry)
286 {
287 entry->e_referenced = 1;
288 }
289 EXPORT_SYMBOL(mb_cache_entry_touch);
290
mb_cache_count(struct shrinker * shrink,struct shrink_control * sc)291 static unsigned long mb_cache_count(struct shrinker *shrink,
292 struct shrink_control *sc)
293 {
294 struct mb_cache *cache = container_of(shrink, struct mb_cache,
295 c_shrink);
296
297 return cache->c_entry_count;
298 }
299
300 /* Shrink number of entries in cache */
mb_cache_shrink(struct mb_cache * cache,unsigned long nr_to_scan)301 static unsigned long mb_cache_shrink(struct mb_cache *cache,
302 unsigned long nr_to_scan)
303 {
304 struct mb_cache_entry *entry;
305 unsigned long shrunk = 0;
306
307 spin_lock(&cache->c_list_lock);
308 while (nr_to_scan-- && !list_empty(&cache->c_list)) {
309 entry = list_first_entry(&cache->c_list,
310 struct mb_cache_entry, e_list);
311 /* Drop initial hash reference if there is no user */
312 if (entry->e_referenced ||
313 atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) {
314 entry->e_referenced = 0;
315 list_move_tail(&entry->e_list, &cache->c_list);
316 continue;
317 }
318 list_del_init(&entry->e_list);
319 cache->c_entry_count--;
320 spin_unlock(&cache->c_list_lock);
321 __mb_cache_entry_free(cache, entry);
322 shrunk++;
323 cond_resched();
324 spin_lock(&cache->c_list_lock);
325 }
326 spin_unlock(&cache->c_list_lock);
327
328 return shrunk;
329 }
330
mb_cache_scan(struct shrinker * shrink,struct shrink_control * sc)331 static unsigned long mb_cache_scan(struct shrinker *shrink,
332 struct shrink_control *sc)
333 {
334 struct mb_cache *cache = container_of(shrink, struct mb_cache,
335 c_shrink);
336 return mb_cache_shrink(cache, sc->nr_to_scan);
337 }
338
339 /* We shrink 1/X of the cache when we have too many entries in it */
340 #define SHRINK_DIVISOR 16
341
mb_cache_shrink_worker(struct work_struct * work)342 static void mb_cache_shrink_worker(struct work_struct *work)
343 {
344 struct mb_cache *cache = container_of(work, struct mb_cache,
345 c_shrink_work);
346 mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR);
347 }
348
349 /*
350 * mb_cache_create - create cache
351 * @bucket_bits: log2 of the hash table size
352 *
353 * Create cache for keys with 2^bucket_bits hash entries.
354 */
mb_cache_create(int bucket_bits)355 struct mb_cache *mb_cache_create(int bucket_bits)
356 {
357 struct mb_cache *cache;
358 unsigned long bucket_count = 1UL << bucket_bits;
359 unsigned long i;
360
361 cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
362 if (!cache)
363 goto err_out;
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),
370 GFP_KERNEL);
371 if (!cache->c_hash) {
372 kfree(cache);
373 goto err_out;
374 }
375 for (i = 0; i < bucket_count; i++)
376 INIT_HLIST_BL_HEAD(&cache->c_hash[i]);
377
378 cache->c_shrink.count_objects = mb_cache_count;
379 cache->c_shrink.scan_objects = mb_cache_scan;
380 cache->c_shrink.seeks = DEFAULT_SEEKS;
381 if (register_shrinker(&cache->c_shrink, "mbcache-shrinker")) {
382 kfree(cache->c_hash);
383 kfree(cache);
384 goto err_out;
385 }
386
387 INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
388
389 return cache;
390
391 err_out:
392 return NULL;
393 }
394 EXPORT_SYMBOL(mb_cache_create);
395
396 /*
397 * mb_cache_destroy - destroy cache
398 * @cache: the cache to destroy
399 *
400 * Free all entries in cache and cache itself. Caller must make sure nobody
401 * (except shrinker) can reach @cache when calling this.
402 */
mb_cache_destroy(struct mb_cache * cache)403 void mb_cache_destroy(struct mb_cache *cache)
404 {
405 struct mb_cache_entry *entry, *next;
406
407 unregister_shrinker(&cache->c_shrink);
408
409 /*
410 * We don't bother with any locking. Cache must not be used at this
411 * point.
412 */
413 list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
414 list_del(&entry->e_list);
415 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
416 mb_cache_entry_put(cache, entry);
417 }
418 kfree(cache->c_hash);
419 kfree(cache);
420 }
421 EXPORT_SYMBOL(mb_cache_destroy);
422
mbcache_init(void)423 static int __init mbcache_init(void)
424 {
425 mb_entry_cache = kmem_cache_create("mbcache",
426 sizeof(struct mb_cache_entry), 0,
427 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
428 if (!mb_entry_cache)
429 return -ENOMEM;
430 return 0;
431 }
432
mbcache_exit(void)433 static void __exit mbcache_exit(void)
434 {
435 kmem_cache_destroy(mb_entry_cache);
436 }
437
438 module_init(mbcache_init)
439 module_exit(mbcache_exit)
440
441 MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
442 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
443 MODULE_LICENSE("GPL");
444
