]>
Commit | Line | Data |
---|---|---|
1 | #include <linux/spinlock.h> | |
2 | #include <linux/slab.h> | |
3 | #include <linux/list.h> | |
4 | #include <linux/list_bl.h> | |
5 | #include <linux/module.h> | |
6 | #include <linux/sched.h> | |
7 | #include <linux/workqueue.h> | |
8 | #include <linux/mbcache.h> | |
9 | ||
10 | /* | |
11 | * Mbcache is a simple key-value store. Keys need not be unique, however | |
12 | * key-value pairs are expected to be unique (we use this fact in | |
13 | * mb_cache_entry_delete_block()). | |
14 | * | |
15 | * Ext2 and ext4 use this cache for deduplication of extended attribute blocks. | |
16 | * They use hash of a block contents as a key and block number as a value. | |
17 | * That's why keys need not be unique (different xattr blocks may end up having | |
18 | * the same hash). However block number always uniquely identifies a cache | |
19 | * entry. | |
20 | * | |
21 | * We provide functions for creation and removal of entries, search by key, | |
22 | * and a special "delete entry with given key-value pair" operation. Fixed | |
23 | * size hash table is used for fast key lookups. | |
24 | */ | |
25 | ||
26 | struct mb_cache { | |
27 | /* Hash table of entries */ | |
28 | struct hlist_bl_head *c_hash; | |
29 | /* log2 of hash table size */ | |
30 | int c_bucket_bits; | |
31 | /* Maximum entries in cache to avoid degrading hash too much */ | |
32 | unsigned long c_max_entries; | |
33 | /* Protects c_list, c_entry_count */ | |
34 | spinlock_t c_list_lock; | |
35 | struct list_head c_list; | |
36 | /* Number of entries in cache */ | |
37 | unsigned long c_entry_count; | |
38 | struct shrinker c_shrink; | |
39 | /* Work for shrinking when the cache has too many entries */ | |
40 | struct work_struct c_shrink_work; | |
41 | }; | |
42 | ||
43 | static struct kmem_cache *mb_entry_cache; | |
44 | ||
45 | static unsigned long mb_cache_shrink(struct mb_cache *cache, | |
46 | unsigned long nr_to_scan); | |
47 | ||
48 | static inline struct hlist_bl_head *mb_cache_entry_head(struct mb_cache *cache, | |
49 | u32 key) | |
50 | { | |
51 | return &cache->c_hash[hash_32(key, cache->c_bucket_bits)]; | |
52 | } | |
53 | ||
54 | /* | |
55 | * Number of entries to reclaim synchronously when there are too many entries | |
56 | * in cache | |
57 | */ | |
58 | #define SYNC_SHRINK_BATCH 64 | |
59 | ||
60 | /* | |
61 | * mb_cache_entry_create - create entry in cache | |
62 | * @cache - cache where the entry should be created | |
63 | * @mask - gfp mask with which the entry should be allocated | |
64 | * @key - key of the entry | |
65 | * @block - block that contains data | |
66 | * @reusable - is the block reusable by other inodes? | |
67 | * | |
68 | * Creates entry in @cache with key @key and records that data is stored in | |
69 | * block @block. The function returns -EBUSY if entry with the same key | |
70 | * and for the same block already exists in cache. Otherwise 0 is returned. | |
71 | */ | |
72 | int mb_cache_entry_create(struct mb_cache *cache, gfp_t mask, u32 key, | |
73 | sector_t block, bool reusable) | |
74 | { | |
75 | struct mb_cache_entry *entry, *dup; | |
76 | struct hlist_bl_node *dup_node; | |
77 | struct hlist_bl_head *head; | |
78 | ||
79 | /* Schedule background reclaim if there are too many entries */ | |
80 | if (cache->c_entry_count >= cache->c_max_entries) | |
81 | schedule_work(&cache->c_shrink_work); | |
82 | /* Do some sync reclaim if background reclaim cannot keep up */ | |
83 | if (cache->c_entry_count >= 2*cache->c_max_entries) | |
84 | mb_cache_shrink(cache, SYNC_SHRINK_BATCH); | |
85 | ||
86 | entry = kmem_cache_alloc(mb_entry_cache, mask); | |
87 | if (!entry) | |
88 | return -ENOMEM; | |
89 | ||
90 | INIT_LIST_HEAD(&entry->e_list); | |
91 | /* One ref for hash, one ref returned */ | |
92 | atomic_set(&entry->e_refcnt, 1); | |
93 | entry->e_key = key; | |
94 | entry->e_block = block; | |
95 | entry->e_reusable = reusable; | |
96 | head = mb_cache_entry_head(cache, key); | |
97 | hlist_bl_lock(head); | |
98 | hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) { | |
99 | if (dup->e_key == key && dup->e_block == block) { | |
100 | hlist_bl_unlock(head); | |
101 | kmem_cache_free(mb_entry_cache, entry); | |
102 | return -EBUSY; | |
103 | } | |
104 | } | |
105 | hlist_bl_add_head(&entry->e_hash_list, head); | |
106 | hlist_bl_unlock(head); | |
107 | ||
108 | spin_lock(&cache->c_list_lock); | |
109 | list_add_tail(&entry->e_list, &cache->c_list); | |
110 | /* Grab ref for LRU list */ | |
111 | atomic_inc(&entry->e_refcnt); | |
112 | cache->c_entry_count++; | |
113 | spin_unlock(&cache->c_list_lock); | |
114 | ||
115 | return 0; | |
116 | } | |
117 | EXPORT_SYMBOL(mb_cache_entry_create); | |
118 | ||
119 | void __mb_cache_entry_free(struct mb_cache_entry *entry) | |
120 | { | |
121 | kmem_cache_free(mb_entry_cache, entry); | |
122 | } | |
123 | EXPORT_SYMBOL(__mb_cache_entry_free); | |
124 | ||
125 | static struct mb_cache_entry *__entry_find(struct mb_cache *cache, | |
126 | struct mb_cache_entry *entry, | |
127 | u32 key) | |
128 | { | |
129 | struct mb_cache_entry *old_entry = entry; | |
130 | struct hlist_bl_node *node; | |
131 | struct hlist_bl_head *head; | |
132 | ||
133 | head = mb_cache_entry_head(cache, key); | |
134 | hlist_bl_lock(head); | |
135 | if (entry && !hlist_bl_unhashed(&entry->e_hash_list)) | |
136 | node = entry->e_hash_list.next; | |
137 | else | |
138 | node = hlist_bl_first(head); | |
139 | while (node) { | |
140 | entry = hlist_bl_entry(node, struct mb_cache_entry, | |
141 | e_hash_list); | |
142 | if (entry->e_key == key && entry->e_reusable) { | |
143 | atomic_inc(&entry->e_refcnt); | |
144 | goto out; | |
145 | } | |
146 | node = node->next; | |
147 | } | |
148 | entry = NULL; | |
149 | out: | |
150 | hlist_bl_unlock(head); | |
151 | if (old_entry) | |
152 | mb_cache_entry_put(cache, old_entry); | |
153 | ||
154 | return entry; | |
155 | } | |
156 | ||
157 | /* | |
158 | * mb_cache_entry_find_first - find the first reusable entry with the given key | |
159 | * @cache: cache where we should search | |
160 | * @key: key to look for | |
161 | * | |
162 | * Search in @cache for a reusable entry with key @key. Grabs reference to the | |
163 | * first reusable entry found and returns the entry. | |
164 | */ | |
165 | struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache, | |
166 | u32 key) | |
167 | { | |
168 | return __entry_find(cache, NULL, key); | |
169 | } | |
170 | EXPORT_SYMBOL(mb_cache_entry_find_first); | |
171 | ||
172 | /* | |
173 | * mb_cache_entry_find_next - find next reusable entry with the same key | |
174 | * @cache: cache where we should search | |
175 | * @entry: entry to start search from | |
176 | * | |
177 | * Finds next reusable entry in the hash chain which has the same key as @entry. | |
178 | * If @entry is unhashed (which can happen when deletion of entry races with the | |
179 | * search), finds the first reusable entry in the hash chain. The function drops | |
180 | * reference to @entry and returns with a reference to the found entry. | |
181 | */ | |
182 | struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache, | |
183 | struct mb_cache_entry *entry) | |
184 | { | |
185 | return __entry_find(cache, entry, entry->e_key); | |
186 | } | |
187 | EXPORT_SYMBOL(mb_cache_entry_find_next); | |
188 | ||
189 | /* | |
190 | * mb_cache_entry_get - get a cache entry by block number (and key) | |
191 | * @cache - cache we work with | |
192 | * @key - key of block number @block | |
193 | * @block - block number | |
194 | */ | |
195 | struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key, | |
196 | sector_t block) | |
197 | { | |
198 | struct hlist_bl_node *node; | |
199 | struct hlist_bl_head *head; | |
200 | struct mb_cache_entry *entry; | |
201 | ||
202 | head = mb_cache_entry_head(cache, key); | |
203 | hlist_bl_lock(head); | |
204 | hlist_bl_for_each_entry(entry, node, head, e_hash_list) { | |
205 | if (entry->e_key == key && entry->e_block == block) { | |
206 | atomic_inc(&entry->e_refcnt); | |
207 | goto out; | |
208 | } | |
209 | } | |
210 | entry = NULL; | |
211 | out: | |
212 | hlist_bl_unlock(head); | |
213 | return entry; | |
214 | } | |
215 | EXPORT_SYMBOL(mb_cache_entry_get); | |
216 | ||
217 | /* mb_cache_entry_delete_block - remove information about block from cache | |
218 | * @cache - cache we work with | |
219 | * @key - key of block @block | |
220 | * @block - block number | |
221 | * | |
222 | * Remove entry from cache @cache with key @key with data stored in @block. | |
223 | */ | |
224 | void mb_cache_entry_delete_block(struct mb_cache *cache, u32 key, | |
225 | sector_t block) | |
226 | { | |
227 | struct hlist_bl_node *node; | |
228 | struct hlist_bl_head *head; | |
229 | struct mb_cache_entry *entry; | |
230 | ||
231 | head = mb_cache_entry_head(cache, key); | |
232 | hlist_bl_lock(head); | |
233 | hlist_bl_for_each_entry(entry, node, head, e_hash_list) { | |
234 | if (entry->e_key == key && entry->e_block == block) { | |
235 | /* We keep hash list reference to keep entry alive */ | |
236 | hlist_bl_del_init(&entry->e_hash_list); | |
237 | hlist_bl_unlock(head); | |
238 | spin_lock(&cache->c_list_lock); | |
239 | if (!list_empty(&entry->e_list)) { | |
240 | list_del_init(&entry->e_list); | |
241 | cache->c_entry_count--; | |
242 | atomic_dec(&entry->e_refcnt); | |
243 | } | |
244 | spin_unlock(&cache->c_list_lock); | |
245 | mb_cache_entry_put(cache, entry); | |
246 | return; | |
247 | } | |
248 | } | |
249 | hlist_bl_unlock(head); | |
250 | } | |
251 | EXPORT_SYMBOL(mb_cache_entry_delete_block); | |
252 | ||
253 | /* mb_cache_entry_touch - cache entry got used | |
254 | * @cache - cache the entry belongs to | |
255 | * @entry - entry that got used | |
256 | * | |
257 | * Marks entry as used to give hit higher chances of surviving in cache. | |
258 | */ | |
259 | void mb_cache_entry_touch(struct mb_cache *cache, | |
260 | struct mb_cache_entry *entry) | |
261 | { | |
262 | entry->e_referenced = 1; | |
263 | } | |
264 | EXPORT_SYMBOL(mb_cache_entry_touch); | |
265 | ||
266 | static unsigned long mb_cache_count(struct shrinker *shrink, | |
267 | struct shrink_control *sc) | |
268 | { | |
269 | struct mb_cache *cache = container_of(shrink, struct mb_cache, | |
270 | c_shrink); | |
271 | ||
272 | return cache->c_entry_count; | |
273 | } | |
274 | ||
275 | /* Shrink number of entries in cache */ | |
276 | static unsigned long mb_cache_shrink(struct mb_cache *cache, | |
277 | unsigned long nr_to_scan) | |
278 | { | |
279 | struct mb_cache_entry *entry; | |
280 | struct hlist_bl_head *head; | |
281 | unsigned long shrunk = 0; | |
282 | ||
283 | spin_lock(&cache->c_list_lock); | |
284 | while (nr_to_scan-- && !list_empty(&cache->c_list)) { | |
285 | entry = list_first_entry(&cache->c_list, | |
286 | struct mb_cache_entry, e_list); | |
287 | if (entry->e_referenced) { | |
288 | entry->e_referenced = 0; | |
289 | list_move_tail(&entry->e_list, &cache->c_list); | |
290 | continue; | |
291 | } | |
292 | list_del_init(&entry->e_list); | |
293 | cache->c_entry_count--; | |
294 | /* | |
295 | * We keep LRU list reference so that entry doesn't go away | |
296 | * from under us. | |
297 | */ | |
298 | spin_unlock(&cache->c_list_lock); | |
299 | head = mb_cache_entry_head(cache, entry->e_key); | |
300 | hlist_bl_lock(head); | |
301 | if (!hlist_bl_unhashed(&entry->e_hash_list)) { | |
302 | hlist_bl_del_init(&entry->e_hash_list); | |
303 | atomic_dec(&entry->e_refcnt); | |
304 | } | |
305 | hlist_bl_unlock(head); | |
306 | if (mb_cache_entry_put(cache, entry)) | |
307 | shrunk++; | |
308 | cond_resched(); | |
309 | spin_lock(&cache->c_list_lock); | |
310 | } | |
311 | spin_unlock(&cache->c_list_lock); | |
312 | ||
313 | return shrunk; | |
314 | } | |
315 | ||
316 | static unsigned long mb_cache_scan(struct shrinker *shrink, | |
317 | struct shrink_control *sc) | |
318 | { | |
319 | struct mb_cache *cache = container_of(shrink, struct mb_cache, | |
320 | c_shrink); | |
321 | return mb_cache_shrink(cache, sc->nr_to_scan); | |
322 | } | |
323 | ||
324 | /* We shrink 1/X of the cache when we have too many entries in it */ | |
325 | #define SHRINK_DIVISOR 16 | |
326 | ||
327 | static void mb_cache_shrink_worker(struct work_struct *work) | |
328 | { | |
329 | struct mb_cache *cache = container_of(work, struct mb_cache, | |
330 | c_shrink_work); | |
331 | mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR); | |
332 | } | |
333 | ||
334 | /* | |
335 | * mb_cache_create - create cache | |
336 | * @bucket_bits: log2 of the hash table size | |
337 | * | |
338 | * Create cache for keys with 2^bucket_bits hash entries. | |
339 | */ | |
340 | struct mb_cache *mb_cache_create(int bucket_bits) | |
341 | { | |
342 | struct mb_cache *cache; | |
343 | unsigned long bucket_count = 1UL << bucket_bits; | |
344 | unsigned long i; | |
345 | ||
346 | cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL); | |
347 | if (!cache) | |
348 | goto err_out; | |
349 | cache->c_bucket_bits = bucket_bits; | |
350 | cache->c_max_entries = bucket_count << 4; | |
351 | INIT_LIST_HEAD(&cache->c_list); | |
352 | spin_lock_init(&cache->c_list_lock); | |
353 | cache->c_hash = kmalloc(bucket_count * sizeof(struct hlist_bl_head), | |
354 | GFP_KERNEL); | |
355 | if (!cache->c_hash) { | |
356 | kfree(cache); | |
357 | goto err_out; | |
358 | } | |
359 | for (i = 0; i < bucket_count; i++) | |
360 | INIT_HLIST_BL_HEAD(&cache->c_hash[i]); | |
361 | ||
362 | cache->c_shrink.count_objects = mb_cache_count; | |
363 | cache->c_shrink.scan_objects = mb_cache_scan; | |
364 | cache->c_shrink.seeks = DEFAULT_SEEKS; | |
365 | if (register_shrinker(&cache->c_shrink)) { | |
366 | kfree(cache->c_hash); | |
367 | kfree(cache); | |
368 | goto err_out; | |
369 | } | |
370 | ||
371 | INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker); | |
372 | ||
373 | return cache; | |
374 | ||
375 | err_out: | |
376 | return NULL; | |
377 | } | |
378 | EXPORT_SYMBOL(mb_cache_create); | |
379 | ||
380 | /* | |
381 | * mb_cache_destroy - destroy cache | |
382 | * @cache: the cache to destroy | |
383 | * | |
384 | * Free all entries in cache and cache itself. Caller must make sure nobody | |
385 | * (except shrinker) can reach @cache when calling this. | |
386 | */ | |
387 | void mb_cache_destroy(struct mb_cache *cache) | |
388 | { | |
389 | struct mb_cache_entry *entry, *next; | |
390 | ||
391 | unregister_shrinker(&cache->c_shrink); | |
392 | ||
393 | /* | |
394 | * We don't bother with any locking. Cache must not be used at this | |
395 | * point. | |
396 | */ | |
397 | list_for_each_entry_safe(entry, next, &cache->c_list, e_list) { | |
398 | if (!hlist_bl_unhashed(&entry->e_hash_list)) { | |
399 | hlist_bl_del_init(&entry->e_hash_list); | |
400 | atomic_dec(&entry->e_refcnt); | |
401 | } else | |
402 | WARN_ON(1); | |
403 | list_del(&entry->e_list); | |
404 | WARN_ON(atomic_read(&entry->e_refcnt) != 1); | |
405 | mb_cache_entry_put(cache, entry); | |
406 | } | |
407 | kfree(cache->c_hash); | |
408 | kfree(cache); | |
409 | } | |
410 | EXPORT_SYMBOL(mb_cache_destroy); | |
411 | ||
412 | static int __init mbcache_init(void) | |
413 | { | |
414 | mb_entry_cache = kmem_cache_create("mbcache", | |
415 | sizeof(struct mb_cache_entry), 0, | |
416 | SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL); | |
417 | if (!mb_entry_cache) | |
418 | return -ENOMEM; | |
419 | return 0; | |
420 | } | |
421 | ||
422 | static void __exit mbcache_exit(void) | |
423 | { | |
424 | kmem_cache_destroy(mb_entry_cache); | |
425 | } | |
426 | ||
427 | module_init(mbcache_init) | |
428 | module_exit(mbcache_exit) | |
429 | ||
430 | MODULE_AUTHOR("Jan Kara <jack@suse.cz>"); | |
431 | MODULE_DESCRIPTION("Meta block cache (for extended attributes)"); | |
432 | MODULE_LICENSE("GPL"); |