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
28struct 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
45static struct kmem_cache *mb_entry_cache;
46
47static unsigned long mb_cache_shrink(struct mb_cache *cache,
48				     unsigned long nr_to_scan);
49
50static 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 */
74int 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}
125EXPORT_SYMBOL(mb_cache_entry_create);
126
127void __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}
137EXPORT_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 */
146void 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}
150EXPORT_SYMBOL(mb_cache_entry_wait_unused);
151
152static 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;
176out:
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 */
192struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
193						 u32 key)
194{
195	return __entry_find(cache, NULL, key);
196}
197EXPORT_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 */
209struct 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}
214EXPORT_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 */
222struct 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;
237out:
238	hlist_bl_unlock(head);
239	return entry;
240}
241EXPORT_SYMBOL(mb_cache_entry_get);
242
243/* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
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. 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.
253 */
254struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
255						    u32 key, u64 value)
256{
257	struct mb_cache_entry *entry;
258
259	entry = mb_cache_entry_get(cache, key, value);
260	if (!entry)
261		return NULL;
262
263	/*
264	 * Drop the ref we got from mb_cache_entry_get() and the initial hash
265	 * ref if we are the last user
266	 */
267	if (atomic_cmpxchg(&entry->e_refcnt, 2, 0) != 2)
268		return entry;
269
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);
276	return NULL;
277}
278EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
279
280/* mb_cache_entry_touch - cache entry got used
281 * @cache - cache the entry belongs to
282 * @entry - entry that got used
283 *
284 * Marks entry as used to give hit higher chances of surviving in cache.
285 */
286void mb_cache_entry_touch(struct mb_cache *cache,
287			  struct mb_cache_entry *entry)
288{
289	set_bit(MBE_REFERENCED_B, &entry->e_flags);
290}
291EXPORT_SYMBOL(mb_cache_entry_touch);
292
293static unsigned long mb_cache_count(struct shrinker *shrink,
294				    struct shrink_control *sc)
295{
296	struct mb_cache *cache = shrink->private_data;
297
298	return cache->c_entry_count;
299}
300
301/* Shrink number of entries in cache */
302static unsigned long mb_cache_shrink(struct mb_cache *cache,
303				     unsigned long nr_to_scan)
304{
305	struct mb_cache_entry *entry;
306	unsigned long shrunk = 0;
307
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);
317			continue;
318		}
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);
323		shrunk++;
324		cond_resched();
325		spin_lock(&cache->c_list_lock);
326	}
327	spin_unlock(&cache->c_list_lock);
328
329	return shrunk;
330}
331
332static unsigned long mb_cache_scan(struct shrinker *shrink,
333				   struct shrink_control *sc)
334{
335	struct mb_cache *cache = shrink->private_data;
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
342static 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 */
355struct 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 = shrinker_alloc(0, "mbcache-shrinker");
379	if (!cache->c_shrink) {
380		kfree(cache->c_hash);
381		kfree(cache);
382		goto err_out;
383	}
384
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;
388
389	shrinker_register(cache->c_shrink);
390
391	INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
392
393	return cache;
394
395err_out:
396	return NULL;
397}
398EXPORT_SYMBOL(mb_cache_create);
399
400/*
401 * mb_cache_destroy - destroy cache
402 * @cache: the cache to destroy
403 *
404 * Free all entries in cache and cache itself. Caller must make sure nobody
405 * (except shrinker) can reach @cache when calling this.
406 */
407void mb_cache_destroy(struct mb_cache *cache)
408{
409	struct mb_cache_entry *entry, *next;
410
411	shrinker_free(cache->c_shrink);
412
413	/*
414	 * We don't bother with any locking. Cache must not be used at this
415	 * point.
416	 */
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);
421	}
422	kfree(cache->c_hash);
423	kfree(cache);
424}
425EXPORT_SYMBOL(mb_cache_destroy);
426
427static int __init mbcache_init(void)
428{
429	mb_entry_cache = KMEM_CACHE(mb_cache_entry, SLAB_RECLAIM_ACCOUNT);
430	if (!mb_entry_cache)
431		return -ENOMEM;
432	return 0;
433}
434
435static void __exit mbcache_exit(void)
436{
437	kmem_cache_destroy(mb_entry_cache);
438}
439
440module_init(mbcache_init)
441module_exit(mbcache_exit)
442
443MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
444MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
445MODULE_LICENSE("GPL");
446