fs/mbcache.c

   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         /* One ref for hash, one ref returned */
  94         atomic_set(&entry->e_refcnt, 1);
  95         entry->e_key = key;
  96         entry->e_value = value;
  97         entry->e_reusable = reusable;
  98         entry->e_referenced = 0;
  99         head = mb_cache_entry_head(cache, key);
 100         hlist_bl_lock(head);
 101         hlist_bl_for_each_entry(dup, dup_node, head, e_hash_list) {
 102                 if (dup->e_key == key && dup->e_value == value) {
 103                         hlist_bl_unlock(head);
 104                         kmem_cache_free(mb_entry_cache, entry);
 105                         return -EBUSY;
 106                 }
 107         }
 108         hlist_bl_add_head(&entry->e_hash_list, head);
 109         hlist_bl_unlock(head);
 110
 111         spin_lock(&cache->c_list_lock);
 112         list_add_tail(&entry->e_list, &cache->c_list);
 113         /* Grab ref for LRU list */
 114         atomic_inc(&entry->e_refcnt);
 115         cache->c_entry_count++;
 116         spin_unlock(&cache->c_list_lock);
 117
 118         return 0;
 119 }
 120 EXPORT_SYMBOL(mb_cache_entry_create);
 121
 122 void __mb_cache_entry_free(struct mb_cache_entry *entry)
 123 {
 124         kmem_cache_free(mb_entry_cache, entry);
 125 }
 126 EXPORT_SYMBOL(__mb_cache_entry_free);
 127
 128 /*
 129  * mb_cache_entry_wait_unused - wait to be the last user of the entry
 130  *
 131  * @entry - entry to work on
 132  *
 133  * Wait to be the last user of the entry.
 134  */
 135 void mb_cache_entry_wait_unused(struct mb_cache_entry *entry)
 136 {
 137         wait_var_event(&entry->e_refcnt, atomic_read(&entry->e_refcnt) <= 3);
 138 }
 139 EXPORT_SYMBOL(mb_cache_entry_wait_unused);
 140
 141 static struct mb_cache_entry *__entry_find(struct mb_cache *cache,
 142                                            struct mb_cache_entry *entry,
 143                                            u32 key)
 144 {
 145         struct mb_cache_entry *old_entry = entry;
 146         struct hlist_bl_node *node;
 147         struct hlist_bl_head *head;
 148
 149         head = mb_cache_entry_head(cache, key);
 150         hlist_bl_lock(head);
 151         if (entry && !hlist_bl_unhashed(&entry->e_hash_list))
 152                 node = entry->e_hash_list.next;
 153         else
 154                 node = hlist_bl_first(head);
 155         while (node) {
 156                 entry = hlist_bl_entry(node, struct mb_cache_entry,
 157                                        e_hash_list);
 158                 if (entry->e_key == key && entry->e_reusable) {
 159                         atomic_inc(&entry->e_refcnt);
 160                         goto out;
 161                 }
 162                 node = node->next;
 163         }
 164         entry = NULL;
 165 out:
 166         hlist_bl_unlock(head);
 167         if (old_entry)
 168                 mb_cache_entry_put(cache, old_entry);
 169
 170         return entry;
 171 }
 172
 173 /*
 174  * mb_cache_entry_find_first - find the first reusable entry with the given key
 175  * @cache: cache where we should search
 176  * @key: key to look for
 177  *
 178  * Search in @cache for a reusable entry with key @key. Grabs reference to the
 179  * first reusable entry found and returns the entry.
 180  */
 181 struct mb_cache_entry *mb_cache_entry_find_first(struct mb_cache *cache,
 182                                                  u32 key)
 183 {
 184         return __entry_find(cache, NULL, key);
 185 }
 186 EXPORT_SYMBOL(mb_cache_entry_find_first);
 187
 188 /*
 189  * mb_cache_entry_find_next - find next reusable entry with the same key
 190  * @cache: cache where we should search
 191  * @entry: entry to start search from
 192  *
 193  * Finds next reusable entry in the hash chain which has the same key as @entry.
 194  * If @entry is unhashed (which can happen when deletion of entry races with the
 195  * search), finds the first reusable entry in the hash chain. The function drops
 196  * reference to @entry and returns with a reference to the found entry.
 197  */
 198 struct mb_cache_entry *mb_cache_entry_find_next(struct mb_cache *cache,
 199                                                 struct mb_cache_entry *entry)
 200 {
 201         return __entry_find(cache, entry, entry->e_key);
 202 }
 203 EXPORT_SYMBOL(mb_cache_entry_find_next);
 204
 205 /*
 206  * mb_cache_entry_get - get a cache entry by value (and key)
 207  * @cache - cache we work with
 208  * @key - key
 209  * @value - value
 210  */
 211 struct mb_cache_entry *mb_cache_entry_get(struct mb_cache *cache, u32 key,
 212                                           u64 value)
 213 {
 214         struct hlist_bl_node *node;
 215         struct hlist_bl_head *head;
 216         struct mb_cache_entry *entry;
 217
 218         head = mb_cache_entry_head(cache, key);
 219         hlist_bl_lock(head);
 220         hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
 221                 if (entry->e_key == key && entry->e_value == value) {
 222                         atomic_inc(&entry->e_refcnt);
 223                         goto out;
 224                 }
 225         }
 226         entry = NULL;
 227 out:
 228         hlist_bl_unlock(head);
 229         return entry;
 230 }
 231 EXPORT_SYMBOL(mb_cache_entry_get);
 232
 233 /* mb_cache_entry_delete - try to remove a cache entry
 234  * @cache - cache we work with
 235  * @key - key
 236  * @value - value
 237  *
 238  * Remove entry from cache @cache with key @key and value @value.
 239  */
 240 void mb_cache_entry_delete(struct mb_cache *cache, u32 key, u64 value)
 241 {
 242         struct hlist_bl_node *node;
 243         struct hlist_bl_head *head;
 244         struct mb_cache_entry *entry;
 245
 246         head = mb_cache_entry_head(cache, key);
 247         hlist_bl_lock(head);
 248         hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
 249                 if (entry->e_key == key && entry->e_value == value) {
 250                         /* We keep hash list reference to keep entry alive */
 251                         hlist_bl_del_init(&entry->e_hash_list);
 252                         hlist_bl_unlock(head);
 253                         spin_lock(&cache->c_list_lock);
 254                         if (!list_empty(&entry->e_list)) {
 255                                 list_del_init(&entry->e_list);
 256                                 if (!WARN_ONCE(cache->c_entry_count == 0,
 257                 "mbcache: attempt to decrement c_entry_count past zero"))
 258                                         cache->c_entry_count--;
 259                                 atomic_dec(&entry->e_refcnt);
 260                         }
 261                         spin_unlock(&cache->c_list_lock);
 262                         mb_cache_entry_put(cache, entry);
 263                         return;
 264                 }
 265         }
 266         hlist_bl_unlock(head);
 267 }
 268 EXPORT_SYMBOL(mb_cache_entry_delete);
 269
 270 /* mb_cache_entry_delete_or_get - remove a cache entry if it has no users
 271  * @cache - cache we work with
 272  * @key - key
 273  * @value - value
 274  *
 275  * Remove entry from cache @cache with key @key and value @value. The removal
 276  * happens only if the entry is unused. The function returns NULL in case the
 277  * entry was successfully removed or there's no entry in cache. Otherwise the
 278  * function grabs reference of the entry that we failed to delete because it
 279  * still has users and return it.
 280  */
 281 struct mb_cache_entry *mb_cache_entry_delete_or_get(struct mb_cache *cache,
 282                                                     u32 key, u64 value)
 283 {
 284         struct hlist_bl_node *node;
 285         struct hlist_bl_head *head;
 286         struct mb_cache_entry *entry;
 287
 288         head = mb_cache_entry_head(cache, key);
 289         hlist_bl_lock(head);
 290         hlist_bl_for_each_entry(entry, node, head, e_hash_list) {
 291                 if (entry->e_key == key && entry->e_value == value) {
 292                         if (atomic_read(&entry->e_refcnt) > 2) {
 293                                 atomic_inc(&entry->e_refcnt);
 294                                 hlist_bl_unlock(head);
 295                                 return entry;
 296                         }
 297                         /* We keep hash list reference to keep entry alive */
 298                         hlist_bl_del_init(&entry->e_hash_list);
 299                         hlist_bl_unlock(head);
 300                         spin_lock(&cache->c_list_lock);
 301                         if (!list_empty(&entry->e_list)) {
 302                                 list_del_init(&entry->e_list);
 303                                 if (!WARN_ONCE(cache->c_entry_count == 0,
 304                 "mbcache: attempt to decrement c_entry_count past zero"))
 305                                         cache->c_entry_count--;
 306                                 atomic_dec(&entry->e_refcnt);
 307                         }
 308                         spin_unlock(&cache->c_list_lock);
 309                         mb_cache_entry_put(cache, entry);
 310                         return NULL;
 311                 }
 312         }
 313         hlist_bl_unlock(head);
 314
 315         return NULL;
 316 }
 317 EXPORT_SYMBOL(mb_cache_entry_delete_or_get);
 318
 319 /* mb_cache_entry_touch - cache entry got used
 320  * @cache - cache the entry belongs to
 321  * @entry - entry that got used
 322  *
 323  * Marks entry as used to give hit higher chances of surviving in cache.
 324  */
 325 void mb_cache_entry_touch(struct mb_cache *cache,
 326                           struct mb_cache_entry *entry)
 327 {
 328         entry->e_referenced = 1;
 329 }
 330 EXPORT_SYMBOL(mb_cache_entry_touch);
 331
 332 static unsigned long mb_cache_count(struct shrinker *shrink,
 333                                     struct shrink_control *sc)
 334 {
 335         struct mb_cache *cache = container_of(shrink, struct mb_cache,
 336                                               c_shrink);
 337
 338         return cache->c_entry_count;
 339 }
 340
 341 /* Shrink number of entries in cache */
 342 static unsigned long mb_cache_shrink(struct mb_cache *cache,
 343                                      unsigned long nr_to_scan)
 344 {
 345         struct mb_cache_entry *entry;
 346         struct hlist_bl_head *head;
 347         unsigned long shrunk = 0;
 348
 349         spin_lock(&cache->c_list_lock);
 350         while (nr_to_scan-- && !list_empty(&cache->c_list)) {
 351                 entry = list_first_entry(&cache->c_list,
 352                                          struct mb_cache_entry, e_list);
 353                 if (entry->e_referenced || atomic_read(&entry->e_refcnt) > 2) {
 354                         entry->e_referenced = 0;
 355                         list_move_tail(&entry->e_list, &cache->c_list);
 356                         continue;
 357                 }
 358                 list_del_init(&entry->e_list);
 359                 cache->c_entry_count--;
 360                 /*
 361                  * We keep LRU list reference so that entry doesn't go away
 362                  * from under us.
 363                  */
 364                 spin_unlock(&cache->c_list_lock);
 365                 head = mb_cache_entry_head(cache, entry->e_key);
 366                 hlist_bl_lock(head);
 367                 /* Now a reliable check if the entry didn't get used... */
 368                 if (atomic_read(&entry->e_refcnt) > 2) {
 369                         hlist_bl_unlock(head);
 370                         spin_lock(&cache->c_list_lock);
 371                         list_add_tail(&entry->e_list, &cache->c_list);
 372                         cache->c_entry_count++;
 373                         continue;
 374                 }
 375                 if (!hlist_bl_unhashed(&entry->e_hash_list)) {
 376                         hlist_bl_del_init(&entry->e_hash_list);
 377                         atomic_dec(&entry->e_refcnt);
 378                 }
 379                 hlist_bl_unlock(head);
 380                 if (mb_cache_entry_put(cache, entry))
 381                         shrunk++;
 382                 cond_resched();
 383                 spin_lock(&cache->c_list_lock);
 384         }
 385         spin_unlock(&cache->c_list_lock);
 386
 387         return shrunk;
 388 }
 389
 390 static unsigned long mb_cache_scan(struct shrinker *shrink,
 391                                    struct shrink_control *sc)
 392 {
 393         struct mb_cache *cache = container_of(shrink, struct mb_cache,
 394                                               c_shrink);
 395         return mb_cache_shrink(cache, sc->nr_to_scan);
 396 }
 397
 398 /* We shrink 1/X of the cache when we have too many entries in it */
 399 #define SHRINK_DIVISOR 16
 400
 401 static void mb_cache_shrink_worker(struct work_struct *work)
 402 {
 403         struct mb_cache *cache = container_of(work, struct mb_cache,
 404                                               c_shrink_work);
 405         mb_cache_shrink(cache, cache->c_max_entries / SHRINK_DIVISOR);
 406 }
 407
 408 /*
 409  * mb_cache_create - create cache
 410  * @bucket_bits: log2 of the hash table size
 411  *
 412  * Create cache for keys with 2^bucket_bits hash entries.
 413  */
 414 struct mb_cache *mb_cache_create(int bucket_bits)
 415 {
 416         struct mb_cache *cache;
 417         unsigned long bucket_count = 1UL << bucket_bits;
 418         unsigned long i;
 419
 420         cache = kzalloc(sizeof(struct mb_cache), GFP_KERNEL);
 421         if (!cache)
 422                 goto err_out;
 423         cache->c_bucket_bits = bucket_bits;
 424         cache->c_max_entries = bucket_count << 4;
 425         INIT_LIST_HEAD(&cache->c_list);
 426         spin_lock_init(&cache->c_list_lock);
 427         cache->c_hash = kmalloc_array(bucket_count,
 428                                       sizeof(struct hlist_bl_head),
 429                                       GFP_KERNEL);
 430         if (!cache->c_hash) {
 431                 kfree(cache);
 432                 goto err_out;
 433         }
 434         for (i = 0; i < bucket_count; i++)
 435                 INIT_HLIST_BL_HEAD(&cache->c_hash[i]);
 436
 437         cache->c_shrink.count_objects = mb_cache_count;
 438         cache->c_shrink.scan_objects = mb_cache_scan;
 439         cache->c_shrink.seeks = DEFAULT_SEEKS;
 440         if (register_shrinker(&cache->c_shrink)) {
 441                 kfree(cache->c_hash);
 442                 kfree(cache);
 443                 goto err_out;
 444         }
 445
 446         INIT_WORK(&cache->c_shrink_work, mb_cache_shrink_worker);
 447
 448         return cache;
 449
 450 err_out:
 451         return NULL;
 452 }
 453 EXPORT_SYMBOL(mb_cache_create);
 454
 455 /*
 456  * mb_cache_destroy - destroy cache
 457  * @cache: the cache to destroy
 458  *
 459  * Free all entries in cache and cache itself. Caller must make sure nobody
 460  * (except shrinker) can reach @cache when calling this.
 461  */
 462 void mb_cache_destroy(struct mb_cache *cache)
 463 {
 464         struct mb_cache_entry *entry, *next;
 465
 466         unregister_shrinker(&cache->c_shrink);
 467
 468         /*
 469          * We don't bother with any locking. Cache must not be used at this
 470          * point.
 471          */
 472         list_for_each_entry_safe(entry, next, &cache->c_list, e_list) {
 473                 if (!hlist_bl_unhashed(&entry->e_hash_list)) {
 474                         hlist_bl_del_init(&entry->e_hash_list);
 475                         atomic_dec(&entry->e_refcnt);
 476                 } else
 477                         WARN_ON(1);
 478                 list_del(&entry->e_list);
 479                 WARN_ON(atomic_read(&entry->e_refcnt) != 1);
 480                 mb_cache_entry_put(cache, entry);
 481         }
 482         kfree(cache->c_hash);
 483         kfree(cache);
 484 }
 485 EXPORT_SYMBOL(mb_cache_destroy);
 486
 487 static int __init mbcache_init(void)
 488 {
 489         mb_entry_cache = kmem_cache_create("mbcache",
 490                                 sizeof(struct mb_cache_entry), 0,
 491                                 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
 492         if (!mb_entry_cache)
 493                 return -ENOMEM;
 494         return 0;
 495 }
 496
 497 static void __exit mbcache_exit(void)
 498 {
 499         kmem_cache_destroy(mb_entry_cache);
 500 }
 501
 502 module_init(mbcache_init)
 503 module_exit(mbcache_exit)
 504
 505 MODULE_AUTHOR("Jan Kara <jack@suse.cz>");
 506 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
 507 MODULE_LICENSE("GPL");