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_flags = 0;
104 if (reusable)
105 set_bit(MBE_REUSABLE_B, &entry->e_flags);
106 head = mb_cache_entry_head(cache, key);
107 hlist_bl_lock(head);
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);
112 return -EBUSY;
113 }
114 }
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);
122
123 return 0;
124 }
125 EXPORT_SYMBOL(mb_cache_entry_create);
126
__mb_cache_entry_free(struct mb_cache * cache,struct mb_cache_entry * entry)127 void __mb_cache_entry_free(struct mb_cache *cache, struct mb_cache_entry *entry)
128 {
129 struct hlist_bl_head *head;
130
131 head = mb_cache_entry_head(cache, entry->e_key);
132 hlist_bl_lock(head);
133 hlist_bl_del(&entry->e_hash_list);
134 hlist_bl_unlock(head);
135 kmem_cache_free(mb_entry_cache, entry);
136 }
137 EXPORT_SYMBOL(__mb_cache_entry_free);
138
139 /*
140 * mb_cache_entry_wait_unused - wait to be the last user of the entry
141 *
142 * @entry - entry to work on
143 *
144 * Wait to be the last user of the entry.
145 */
mb_cache_entry_wait_unused(struct mb_cache_entry * entry)146 void mb_cache_entry_wait_unused(struct mb_cache_entry *entry)
147 {
148 wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 2);
149 }
150 EXPORT_SYMBOL(mb_cache_entry_wait_unused);
151
__entry_find(struct mb_cache * cache,struct mb_cache_entry * entry,u32 key)152 static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
153 struct mb_cache_entry *entry,
154 u32 key)
155 {
156 struct mb_cache_entry *old_entry = entry;
157 struct hlist_bl_node *node;
158 struct hlist_bl_head *head;
159
160 head = mb_cache_entry_head(cache, key);
161 hlist_bl_lock(head);
162 if (entry && !hlist_bl_unhashed(&entry->e_hash_list))
163 node = entry->e_hash_list.next;
164 else
165 node = hlist_bl_first(head);
166 while (node) {
167 entry = hlist_bl_entry(node, struct mb_cache_entry,
168 e_hash_list);
169 if (entry->e_key == key &&
170 test_bit(MBE_REUSABLE_B, &entry->e_flags) &&
171 atomic_inc_not_zero(&entry->e_refcnt))
172 goto out;
173 node = node->next;
174 }
175 entry = NULL;
176 out:
177 hlist_bl_unlock(head);
178 if (old_entry)
179 mb_cache_entry_put(cache, old_entry);
180
181 return entry;
182 }
183
184 /*
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
188 *
189 * Search in @cache for a reusable entry with key @key. Grabs reference to the
190 * first reusable entry found and returns the entry.
191 */
mb_cache_entry_find_first(struct mb_cache * cache,u32 key)192 struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
193 u32 key)
194 {
195 return __entry_find(cache, NULL, key);
196 }
197 EXPORT_SYMBOL(mb_cache_entry_find_first);
198
199 /*
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
203 *
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.
208 */
mb_cache_entry_find_next(struct mb_cache * cache,struct mb_cache_entry * entry)209 struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
210 struct mb_cache_entry *entry)
211 {
212 return __entry_find(cache, entry, entry->e_key);
213 }
214 EXPORT_SYMBOL(mb_cache_entry_find_next);
215
216 /*
217 * mb_cache_entry_get - get a cache entry by value (and key)
218 * @cache - cache we work with
219 * @key - key
220 * @value - value
221 */
mb_cache_entry_get(struct mb_cache * cache,u32 key,u64 value)222 struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
223 u64 value)
224 {
225 struct hlist_bl_node *node;
226 struct hlist_bl_head *head;
227 struct mb_cache_entry *entry;
228
229 head = mb_cache_entry_head(cache, key);
230 hlist_bl_lock(head);
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))
234 goto out;
235 }
236 entry = NULL;
237 out:
238 hlist_bl_unlock(head);
239 return entry;
240 }
241 EXPORT_SYMBOL(mb_cache_entry_get);
242
243 /* mb_cache_entry_delete - try to remove a cache entry
244 * @cache - cache we work with
245 * @key - key
246 * @value - value
247 *
248 * Remove entry from cache @cache with key @key and value @value.
249 */
mb_cache_entry_delete(struct mb_cache * cache,u32 key,u64 value)250 void mb_cache_entry_delete(struct mb_cache *cache, u32 key, u64 value)
251 {
252 struct hlist_bl_node *node;
253 struct hlist_bl_head *head;
254 struct mb_cache_entry *entry;
255
256 head = mb_cache_entry_head(cache, key);
257 hlist_bl_lock(head);
258 hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
259 if (entry->e_key == key && entry->e_value == value) {
260 /* We keep hash list reference to keep entry alive */
261 hlist_bl_del_init(&entry->e_hash_list);
262 hlist_bl_unlock(head);
263 spin_lock(&cache->c_list_lock);
264 if (!list_empty(&entry->e_list)) {
265 list_del_init(&entry->e_list);
266 if (!WARN_ONCE(cache->c_entry_count == 0,
267 "mbcache: attempt to decrement c_entry_count past zero"))
268 cache->c_entry_count--;
269 atomic_dec(&entry->e_refcnt);
270 }
271 spin_unlock(&cache->c_list_lock);
272 mb_cache_entry_put(cache, entry);
273 return;
274 }
275 }
276 hlist_bl_unlock(head);
277 }
278 EXPORT_SYMBOL(mb_cache_entry_delete);
279
280 /* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
281 * @cache - cache we work with
282 * @key - key
283 * @value - value
284 *
285 * Remove entry from cache @cache with key @key and value @value. The removal
286 * happens only if the entry is unused. The function returns NULL in case the
287 * entry was successfully removed or there's no entry in cache. Otherwise the
288 * function grabs reference of the entry that we failed to delete because it
289 * still has users and return it.
290 */
mb_cache_entry_delete_or_get(struct mb_cache * cache,u32 key,u64 value)291 struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
292 u32 key, u64 value)
293 {
294 struct mb_cache_entry *entry;
295
296 entry = mb_cache_entry_get(cache, key, value);
297 if (!entry)
298 return NULL;
299
300 /*
301 * Drop the ref we got from mb_cache_entry_get() and the initial hash
302 * ref if we are the last user
303 */
304 if (atomic_cmpxchg(&entry->e_refcnt, 2, 0) != 2)
305 return entry;
306
307 spin_lock(&cache->c_list_lock);
308 if (!list_empty(&entry->e_list))
309 list_del_init(&entry->e_list);
310 cache->c_entry_count--;
311 spin_unlock(&cache->c_list_lock);
312 __mb_cache_entry_free(cache, entry);
313 return NULL;
314 }
315 EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
316
317 /* mb_cache_entry_touch - cache entry got used
318 * @cache - cache the entry belongs to
319 * @entry - entry that got used
320 *
321 * Marks entry as used to give hit higher chances of surviving in cache.
322 */
mb_cache_entry_touch(struct mb_cache * cache,struct mb_cache_entry * entry)323 void mb_cache_entry_touch(struct mb_cache *cache,
324 struct mb_cache_entry *entry)
325 {
326 set_bit(MBE_REFERENCED_B, &entry->e_flags);
327 }
328 EXPORT_SYMBOL(mb_cache_entry_touch);
329
mb_cache_count(struct shrinker * shrink,struct shrink_control * sc)330 static unsigned long mb_cache_count(struct shrinker *shrink,
331 struct shrink_control *sc)
332 {
333 struct mb_cache *cache = container_of(shrink, struct mb_cache,
334 c_shrink);
335
336 return cache->c_entry_count;
337 }
338
339 /* Shrink number of entries in cache */
mb_cache_shrink(struct mb_cache * cache,unsigned long nr_to_scan)340 static unsigned long mb_cache_shrink(struct mb_cache *cache,
341 unsigned long nr_to_scan)
342 {
343 struct mb_cache_entry *entry;
344 unsigned long shrunk = 0;
345
346 spin_lock(&cache->c_list_lock);
347 while (nr_to_scan-- && !list_empty(&cache->c_list)) {
348 entry = list_first_entry(&cache->c_list,
349 struct mb_cache_entry, e_list);
350 /* Drop initial hash reference if there is no user */
351 if (test_bit(MBE_REFERENCED_B, &entry->e_flags) ||
352 atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) {
353 clear_bit(MBE_REFERENCED_B, &entry->e_flags);
354 list_move_tail(&entry->e_list, &cache->c_list);
355 continue;
356 }
357 list_del_init(&entry->e_list);
358 cache->c_entry_count--;
359 spin_unlock(&cache->c_list_lock);
360 __mb_cache_entry_free(cache, entry);
361 shrunk++;
362 cond_resched();
363 spin_lock(&cache->c_list_lock);
364 }
365 spin_unlock(&cache->c_list_lock);
366
367 return shrunk;
368 }
369
mb_cache_scan(struct shrinker * shrink,struct shrink_control * sc)370 static unsigned long mb_cache_scan(struct shrinker *shrink,
371 struct shrink_control *sc)
372 {
373 struct mb_cache *cache = container_of(shrink, struct mb_cache,
374 c_shrink);
375 return mb_cache_shrink(cache, sc->nr_to_scan);
376 }
377
378 /* We shrink 1/X of the cache when we have too many entries in it */
379 #define SHRINK_DIVISOR 16
380
mb_cache_shrink_worker(struct work_struct * work)381 static void mb_cache_shrink_worker(struct work_struct *work)
382 {
383 struct mb_cache *cache = container_of(work, struct mb_cache,
384 c_shrink_work);
385 mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR);
386 }
387
388 /*
389 * mb_cache_create - create cache
390 * @bucket_bits: log2 of the hash table size
391 *
392 * Create cache for keys with 2^bucket_bits hash entries.
393 */
mb_cache_create(int bucket_bits)394 struct mb_cache *mb_cache_create(int bucket_bits)
395 {
396 struct mb_cache *cache;
397 unsigned long bucket_count = 1UL << bucket_bits;
398 unsigned long i;
399
400 cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
401 if (!cache)
402 goto err_out;
403 cache->c_bucket_bits = bucket_bits;
404 cache->c_max_entries = bucket_count << 4;
405 INIT_LIST_HEAD(&cache->c_list);
406 spin_lock_init(&cache->c_list_lock);
407 cache->c_hash = kmalloc_array(bucket_count,
408 sizeof(struct hlist_bl_head),
409 GFP_KERNEL);
410 if (!cache->c_hash) {
411 kfree(cache);
412 goto err_out;
413 }
414 for (i = 0; i < bucket_count; i++)
415 INIT_HLIST_BL_HEAD(&cache->c_hash[i]);
416
417 cache->c_shrink.count_objects = mb_cache_count;
418 cache->c_shrink.scan_objects = mb_cache_scan;
419 cache->c_shrink.seeks = DEFAULT_SEEKS;
420 if (register_shrinker(&cache->c_shrink)) {
421 kfree(cache->c_hash);
422 kfree(cache);
423 goto err_out;
424 }
425
426 INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
427
428 return cache;
429
430 err_out:
431 return NULL;
432 }
433 EXPORT_SYMBOL(mb_cache_create);
434
435 /*
436 * mb_cache_destroy - destroy cache
437 * @cache: the cache to destroy
438 *
439 * Free all entries in cache and cache itself. Caller must make sure nobody
440 * (except shrinker) can reach @cache when calling this.
441 */
mb_cache_destroy(struct mb_cache * cache)442 void mb_cache_destroy(struct mb_cache *cache)
443 {
444 struct mb_cache_entry *entry, *next;
445
446 unregister_shrinker(&cache->c_shrink);
447
448 /*
449 * We don't bother with any locking. Cache must not be used at this
450 * point.
451 */
452 list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
453 list_del(&entry->e_list);
454 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
455 mb_cache_entry_put(cache, entry);
456 }
457 kfree(cache->c_hash);
458 kfree(cache);
459 }
460 EXPORT_SYMBOL(mb_cache_destroy);
461
mbcache_init(void)462 static int __init mbcache_init(void)
463 {
464 mb_entry_cache = kmem_cache_create("mbcache",
465 sizeof(struct mb_cache_entry), 0,
466 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
467 if (!mb_entry_cache)
468 return -ENOMEM;
469 return 0;
470 }
471
mbcache_exit(void)472 static void __exit mbcache_exit(void)
473 {
474 kmem_cache_destroy(mb_entry_cache);
475 }
476
477 module_init(mbcache_init)
478 module_exit(mbcache_exit)
479
480 MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
481 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
482 MODULE_LICENSE("GPL");
483