| 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 | |
| 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 | */ |
| 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 | |
| 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 | */ |
| 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 | |
| 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 | */ |
| 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 | */ |
| 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 | */ |
| 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_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 | */ |
| 254 | struct 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 | } |
| 278 | EXPORT_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 | */ |
| 286 | void 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 | } |
| 291 | EXPORT_SYMBOL(mb_cache_entry_touch); |
| 292 | |
| 293 | static unsigned long mb_cache_count(struct shrinker *shrink, |
| 294 | struct shrink_control *sc) |
| 295 | { |
| 296 | struct mb_cache *cache = container_of(shrink, struct mb_cache, |
| 297 | c_shrink); |
| 298 | |
| 299 | return cache->c_entry_count; |
| 300 | } |
| 301 | |
| 302 | /* Shrink number of entries in cache */ |
| 303 | static unsigned long mb_cache_shrink(struct mb_cache *cache, |
| 304 | unsigned long nr_to_scan) |
| 305 | { |
| 306 | struct mb_cache_entry *entry; |
| 307 | unsigned long shrunk = 0; |
| 308 | |
| 309 | spin_lock(&cache->c_list_lock); |
| 310 | while (nr_to_scan-- && !list_empty(&cache->c_list)) { |
| 311 | entry = list_first_entry(&cache->c_list, |
| 312 | struct mb_cache_entry, e_list); |
| 313 | /* Drop initial hash reference if there is no user */ |
| 314 | if (test_bit(MBE_REFERENCED_B, &entry->e_flags) || |
| 315 | atomic_cmpxchg(&entry->e_refcnt, 1, 0) != 1) { |
| 316 | clear_bit(MBE_REFERENCED_B, &entry->e_flags); |
| 317 | list_move_tail(&entry->e_list, &cache->c_list); |
| 318 | continue; |
| 319 | } |
| 320 | list_del_init(&entry->e_list); |
| 321 | cache->c_entry_count--; |
| 322 | spin_unlock(&cache->c_list_lock); |
| 323 | __mb_cache_entry_free(cache, entry); |
| 324 | shrunk++; |
| 325 | cond_resched(); |
| 326 | spin_lock(&cache->c_list_lock); |
| 327 | } |
| 328 | spin_unlock(&cache->c_list_lock); |
| 329 | |
| 330 | return shrunk; |
| 331 | } |
| 332 | |
| 333 | static unsigned long mb_cache_scan(struct shrinker *shrink, |
| 334 | struct shrink_control *sc) |
| 335 | { |
| 336 | struct mb_cache *cache = container_of(shrink, struct mb_cache, |
| 337 | c_shrink); |
| 338 | return mb_cache_shrink(cache, sc->nr_to_scan); |
| 339 | } |
| 340 | |
| 341 | /* We shrink 1/X of the cache when we have too many entries in it */ |
| 342 | #define SHRINK_DIVISOR 16 |
| 343 | |
| 344 | static void mb_cache_shrink_worker(struct work_struct *work) |
| 345 | { |
| 346 | struct mb_cache *cache = container_of(work, struct mb_cache, |
| 347 | c_shrink_work); |
| 348 | mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR); |
| 349 | } |
| 350 | |
| 351 | /* |
| 352 | * mb_cache_create - create cache |
| 353 | * @bucket_bits: log2 of the hash table size |
| 354 | * |
| 355 | * Create cache for keys with 2^bucket_bits hash entries. |
| 356 | */ |
| 357 | struct mb_cache *mb_cache_create(int bucket_bits) |
| 358 | { |
| 359 | struct mb_cache *cache; |
| 360 | unsigned long bucket_count = 1UL << bucket_bits; |
| 361 | unsigned long i; |
| 362 | |
| 363 | cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL); |
| 364 | if (!cache) |
| 365 | goto err_out; |
| 366 | cache->c_bucket_bits = bucket_bits; |
| 367 | cache->c_max_entries = bucket_count << 4; |
| 368 | INIT_LIST_HEAD(&cache->c_list); |
| 369 | spin_lock_init(&cache->c_list_lock); |
| 370 | cache->c_hash = kmalloc_array(bucket_count, |
| 371 | sizeof(struct hlist_bl_head), |
| 372 | GFP_KERNEL); |
| 373 | if (!cache->c_hash) { |
| 374 | kfree(cache); |
| 375 | goto err_out; |
| 376 | } |
| 377 | for (i = 0; i < bucket_count; i++) |
| 378 | INIT_HLIST_BL_HEAD(&cache->c_hash[i]); |
| 379 | |
| 380 | cache->c_shrink.count_objects = mb_cache_count; |
| 381 | cache->c_shrink.scan_objects = mb_cache_scan; |
| 382 | cache->c_shrink.seeks = DEFAULT_SEEKS; |
| 383 | if (register_shrinker(&cache->c_shrink, "mbcache-shrinker")) { |
| 384 | kfree(cache->c_hash); |
| 385 | kfree(cache); |
| 386 | goto err_out; |
| 387 | } |
| 388 | |
| 389 | INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker); |
| 390 | |
| 391 | return cache; |
| 392 | |
| 393 | err_out: |
| 394 | return NULL; |
| 395 | } |
| 396 | EXPORT_SYMBOL(mb_cache_create); |
| 397 | |
| 398 | /* |
| 399 | * mb_cache_destroy - destroy cache |
| 400 | * @cache: the cache to destroy |
| 401 | * |
| 402 | * Free all entries in cache and cache itself. Caller must make sure nobody |
| 403 | * (except shrinker) can reach @cache when calling this. |
| 404 | */ |
| 405 | void mb_cache_destroy(struct mb_cache *cache) |
| 406 | { |
| 407 | struct mb_cache_entry *entry, *next; |
| 408 | |
| 409 | unregister_shrinker(&cache->c_shrink); |
| 410 | |
| 411 | /* |
| 412 | * We don't bother with any locking. Cache must not be used at this |
| 413 | * point. |
| 414 | */ |
| 415 | list_for_each_entry_safe(entry, next, &cache->c_list, e_list) { |
| 416 | list_del(&entry->e_list); |
| 417 | WARN_ON(atomic_read(&entry->e_refcnt) != 1); |
| 418 | mb_cache_entry_put(cache, entry); |
| 419 | } |
| 420 | kfree(cache->c_hash); |
| 421 | kfree(cache); |
| 422 | } |
| 423 | EXPORT_SYMBOL(mb_cache_destroy); |
| 424 | |
| 425 | static int __init mbcache_init(void) |
| 426 | { |
| 427 | mb_entry_cache = kmem_cache_create("mbcache", |
| 428 | sizeof(struct mb_cache_entry), 0, |
| 429 | SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL); |
| 430 | if (!mb_entry_cache) |
| 431 | return -ENOMEM; |
| 432 | return 0; |
| 433 | } |
| 434 | |
| 435 | static void __exit mbcache_exit(void) |
| 436 | { |
| 437 | kmem_cache_destroy(mb_entry_cache); |
| 438 | } |
| 439 | |
| 440 | module_init(mbcache_init) |
| 441 | module_exit(mbcache_exit) |
| 442 | |
| 443 | MODULE_AUTHOR("Jan Kara <jack@suse.cz>"); |
| 444 | MODULE_DESCRIPTION("Meta block cache (for extended attributes)"); |
| 445 | MODULE_LICENSE("GPL"); |