fs/mbcache.c

   1 /*
   2  * linux/fs/mbcache.c
   3  * (C) 2001-2002 Andreas Gruenbacher, <a.gruenbacher@computer.org>
   4  */
   5
   6 /*
   7  * Filesystem Meta Information Block Cache (mbcache)
   8  *
   9  * The mbcache caches blocks of block devices that need to be located
  10  * by their device/block number, as well as by other criteria (such
  11  * as the block's contents).
  12  *
  13  * There can only be one cache entry in a cache per device and block number.
  14  * Additional indexes need not be unique in this sense. The number of
  15  * additional indexes (=other criteria) can be hardwired at compile time
  16  * or specified at cache create time.
  17  *
  18  * Each cache entry is of fixed size. An entry may be `valid' or `invalid'
  19  * in the cache. A valid entry is in the main hash tables of the cache,
  20  * and may also be in the lru list. An invalid entry is not in any hashes
  21  * or lists.
  22  *
  23  * A valid cache entry is only in the lru list if no handles refer to it.
  24  * Invalid cache entries will be freed when the last handle to the cache
  25  * entry is released. Entries that cannot be freed immediately are put
  26  * back on the lru list.
  27  */
  28
  29 #include <linux/kernel.h>
  30 #include <linux/module.h>
  31
  32 #include <linux/hash.h>
  33 #include <linux/fs.h>
  34 #include <linux/mm.h>
  35 #include <linux/slab.h>
  36 #include <linux/sched.h>
  37 #include <linux/init.h>
  38 #include <linux/mbcache.h>
  39
  40
  41 #ifdef MB_CACHE_DEBUG
  42 # define mb_debug(f...) do { \
  43                 printk(KERN_DEBUG f); \
  44                 printk("\n"); \
  45         } while (0)
  46 #define mb_assert(c) do { if (!(c)) \
  47                 printk(KERN_ERR "assertion " #c " failed\n"); \
  48         } while(0)
  49 #else
  50 # define mb_debug(f...) do { } while(0)
  51 # define mb_assert(c) do { } while(0)
  52 #endif
  53 #define mb_error(f...) do { \
  54                 printk(KERN_ERR f); \
  55                 printk("\n"); \
  56         } while(0)
  57
  58 #define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
  59
  60 static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
  61
  62 MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
  63 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
  64 MODULE_LICENSE("GPL");
  65
  66 EXPORT_SYMBOL(mb_cache_create);
  67 EXPORT_SYMBOL(mb_cache_shrink);
  68 EXPORT_SYMBOL(mb_cache_destroy);
  69 EXPORT_SYMBOL(mb_cache_entry_alloc);
  70 EXPORT_SYMBOL(mb_cache_entry_insert);
  71 EXPORT_SYMBOL(mb_cache_entry_release);
  72 EXPORT_SYMBOL(mb_cache_entry_free);
  73 EXPORT_SYMBOL(mb_cache_entry_get);
  74 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
  75 EXPORT_SYMBOL(mb_cache_entry_find_first);
  76 EXPORT_SYMBOL(mb_cache_entry_find_next);
  77 #endif
  78
  79 struct mb_cache {
  80         struct list_head                c_cache_list;
  81         const char                      *c_name;
  82         atomic_t                        c_entry_count;
  83         int                             c_bucket_bits;
  84         struct kmem_cache               *c_entry_cache;
  85         struct list_head                *c_block_hash;
  86         struct list_head                *c_index_hash;
  87 };
  88
  89
  90 /*
  91  * Global data: list of all mbcache's, lru list, and a spinlock for
  92  * accessing cache data structures on SMP machines. The lru list is
  93  * global across all mbcaches.
  94  */
  95
  96 static LIST_HEAD(mb_cache_list);
  97 static LIST_HEAD(mb_cache_lru_list);
  98 static DEFINE_SPINLOCK(mb_cache_spinlock);
  99
 100 /*
 101  * What the mbcache registers as to get shrunk dynamically.
 102  */
 103
 104 static int mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask);
 105
 106 static struct shrinker mb_cache_shrinker = {
 107         .shrink = mb_cache_shrink_fn,
 108         .seeks = DEFAULT_SEEKS,
 109 };
 110
 111 static inline int
 112 __mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
 113 {
 114         return !list_empty(&ce->e_block_list);
 115 }
 116
 117
 118 static void
 119 __mb_cache_entry_unhash(struct mb_cache_entry *ce)
 120 {
 121         if (__mb_cache_entry_is_hashed(ce)) {
 122                 list_del_init(&ce->e_block_list);
 123                 list_del(&ce->e_index.o_list);
 124         }
 125 }
 126
 127
 128 static void
 129 __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
 130 {
 131         struct mb_cache *cache = ce->e_cache;
 132
 133         mb_assert(!(ce->e_used || ce->e_queued));
 134         kmem_cache_free(cache->c_entry_cache, ce);
 135         atomic_dec(&cache->c_entry_count);
 136 }
 137
 138
 139 static void
 140 __mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
 141         __releases(mb_cache_spinlock)
 142 {
 143         /* Wake up all processes queuing for this cache entry. */
 144         if (ce->e_queued)
 145                 wake_up_all(&mb_cache_queue);
 146         if (ce->e_used >= MB_CACHE_WRITER)
 147                 ce->e_used -= MB_CACHE_WRITER;
 148         ce->e_used--;
 149         if (!(ce->e_used || ce->e_queued)) {
 150                 if (!__mb_cache_entry_is_hashed(ce))
 151                         goto forget;
 152                 mb_assert(list_empty(&ce->e_lru_list));
 153                 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
 154         }
 155         spin_unlock(&mb_cache_spinlock);
 156         return;
 157 forget:
 158         spin_unlock(&mb_cache_spinlock);
 159         __mb_cache_entry_forget(ce, GFP_KERNEL);
 160 }
 161
 162
 163 /*
 164  * mb_cache_shrink_fn()  memory pressure callback
 165  *
 166  * This function is called by the kernel memory management when memory
 167  * gets low.
 168  *
 169  * @shrink: (ignored)
 170  * @nr_to_scan: Number of objects to scan
 171  * @gfp_mask: (ignored)
 172  *
 173  * Returns the number of objects which are present in the cache.
 174  */
 175 static int
 176 mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
 177 {
 178         LIST_HEAD(free_list);
 179         struct mb_cache *cache;
 180         struct mb_cache_entry *entry, *tmp;
 181         int count = 0;
 182
 183         mb_debug("trying to free %d entries", nr_to_scan);
 184         spin_lock(&mb_cache_spinlock);
 185         while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
 186                 struct mb_cache_entry *ce =
 187                         list_entry(mb_cache_lru_list.next,
 188                                    struct mb_cache_entry, e_lru_list);
 189                 list_move_tail(&ce->e_lru_list, &free_list);
 190                 __mb_cache_entry_unhash(ce);
 191         }
 192         list_for_each_entry(cache, &mb_cache_list, c_cache_list) {
 193                 mb_debug("cache %s (%d)", cache->c_name,
 194                           atomic_read(&cache->c_entry_count));
 195                 count += atomic_read(&cache->c_entry_count);
 196         }
 197         spin_unlock(&mb_cache_spinlock);
 198         list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
 199                 __mb_cache_entry_forget(entry, gfp_mask);
 200         }
 201         return (count / 100) * sysctl_vfs_cache_pressure;
 202 }
 203
 204
 205 /*
 206  * mb_cache_create()  create a new cache
 207  *
 208  * All entries in one cache are equal size. Cache entries may be from
 209  * multiple devices. If this is the first mbcache created, registers
 210  * the cache with kernel memory management. Returns NULL if no more
 211  * memory was available.
 212  *
 213  * @name: name of the cache (informal)
 214  * @bucket_bits: log2(number of hash buckets)
 215  */
 216 struct mb_cache *
 217 mb_cache_create(const char *name, int bucket_bits)
 218 {
 219         int n, bucket_count = 1 << bucket_bits;
 220         struct mb_cache *cache = NULL;
 221
 222         cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
 223         if (!cache)
 224                 return NULL;
 225         cache->c_name = name;
 226         atomic_set(&cache->c_entry_count, 0);
 227         cache->c_bucket_bits = bucket_bits;
 228         cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
 229                                       GFP_KERNEL);
 230         if (!cache->c_block_hash)
 231                 goto fail;
 232         for (n=0; n<bucket_count; n++)
 233                 INIT_LIST_HEAD(&cache->c_block_hash[n]);
 234         cache->c_index_hash = kmalloc(bucket_count * sizeof(struct list_head),
 235                                       GFP_KERNEL);
 236         if (!cache->c_index_hash)
 237                 goto fail;
 238         for (n=0; n<bucket_count; n++)
 239                 INIT_LIST_HEAD(&cache->c_index_hash[n]);
 240         cache->c_entry_cache = kmem_cache_create(name,
 241                 sizeof(struct mb_cache_entry), 0,
 242                 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
 243         if (!cache->c_entry_cache)
 244                 goto fail2;
 245
 246         spin_lock(&mb_cache_spinlock);
 247         list_add(&cache->c_cache_list, &mb_cache_list);
 248         spin_unlock(&mb_cache_spinlock);
 249         return cache;
 250
 251 fail2:
 252         kfree(cache->c_index_hash);
 253
 254 fail:
 255         kfree(cache->c_block_hash);
 256         kfree(cache);
 257         return NULL;
 258 }
 259
 260
 261 /*
 262  * mb_cache_shrink()
 263  *
 264  * Removes all cache entries of a device from the cache. All cache entries
 265  * currently in use cannot be freed, and thus remain in the cache. All others
 266  * are freed.
 267  *
 268  * @bdev: which device's cache entries to shrink
 269  */
 270 void
 271 mb_cache_shrink(struct block_device *bdev)
 272 {
 273         LIST_HEAD(free_list);
 274         struct list_head *l, *ltmp;
 275
 276         spin_lock(&mb_cache_spinlock);
 277         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 278                 struct mb_cache_entry *ce =
 279                         list_entry(l, struct mb_cache_entry, e_lru_list);
 280                 if (ce->e_bdev == bdev) {
 281                         list_move_tail(&ce->e_lru_list, &free_list);
 282                         __mb_cache_entry_unhash(ce);
 283                 }
 284         }
 285         spin_unlock(&mb_cache_spinlock);
 286         list_for_each_safe(l, ltmp, &free_list) {
 287                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 288                                                    e_lru_list), GFP_KERNEL);
 289         }
 290 }
 291
 292
 293 /*
 294  * mb_cache_destroy()
 295  *
 296  * Shrinks the cache to its minimum possible size (hopefully 0 entries),
 297  * and then destroys it. If this was the last mbcache, un-registers the
 298  * mbcache from kernel memory management.
 299  */
 300 void
 301 mb_cache_destroy(struct mb_cache *cache)
 302 {
 303         LIST_HEAD(free_list);
 304         struct list_head *l, *ltmp;
 305
 306         spin_lock(&mb_cache_spinlock);
 307         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 308                 struct mb_cache_entry *ce =
 309                         list_entry(l, struct mb_cache_entry, e_lru_list);
 310                 if (ce->e_cache == cache) {
 311                         list_move_tail(&ce->e_lru_list, &free_list);
 312                         __mb_cache_entry_unhash(ce);
 313                 }
 314         }
 315         list_del(&cache->c_cache_list);
 316         spin_unlock(&mb_cache_spinlock);
 317
 318         list_for_each_safe(l, ltmp, &free_list) {
 319                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 320                                                    e_lru_list), GFP_KERNEL);
 321         }
 322
 323         if (atomic_read(&cache->c_entry_count) > 0) {
 324                 mb_error("cache %s: %d orphaned entries",
 325                           cache->c_name,
 326                           atomic_read(&cache->c_entry_count));
 327         }
 328
 329         kmem_cache_destroy(cache->c_entry_cache);
 330
 331         kfree(cache->c_index_hash);
 332         kfree(cache->c_block_hash);
 333         kfree(cache);
 334 }
 335
 336
 337 /*
 338  * mb_cache_entry_alloc()
 339  *
 340  * Allocates a new cache entry. The new entry will not be valid initially,
 341  * and thus cannot be looked up yet. It should be filled with data, and
 342  * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
 343  * if no more memory was available.
 344  */
 345 struct mb_cache_entry *
 346 mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
 347 {
 348         struct mb_cache_entry *ce;
 349
 350         ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
 351         if (ce) {
 352                 atomic_inc(&cache->c_entry_count);
 353                 INIT_LIST_HEAD(&ce->e_lru_list);
 354                 INIT_LIST_HEAD(&ce->e_block_list);
 355                 ce->e_cache = cache;
 356                 ce->e_used = 1 + MB_CACHE_WRITER;
 357                 ce->e_queued = 0;
 358         }
 359         return ce;
 360 }
 361
 362
 363 /*
 364  * mb_cache_entry_insert()
 365  *
 366  * Inserts an entry that was allocated using mb_cache_entry_alloc() into
 367  * the cache. After this, the cache entry can be looked up, but is not yet
 368  * in the lru list as the caller still holds a handle to it. Returns 0 on
 369  * success, or -EBUSY if a cache entry for that device + inode exists
 370  * already (this may happen after a failed lookup, but when another process
 371  * has inserted the same cache entry in the meantime).
 372  *
 373  * @bdev: device the cache entry belongs to
 374  * @block: block number
 375  * @key: lookup key
 376  */
 377 int
 378 mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
 379                       sector_t block, unsigned int key)
 380 {
 381         struct mb_cache *cache = ce->e_cache;
 382         unsigned int bucket;
 383         struct list_head *l;
 384         int error = -EBUSY;
 385
 386         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 387                            cache->c_bucket_bits);
 388         spin_lock(&mb_cache_spinlock);
 389         list_for_each_prev(l, &cache->c_block_hash[bucket]) {
 390                 struct mb_cache_entry *ce =
 391                         list_entry(l, struct mb_cache_entry, e_block_list);
 392                 if (ce->e_bdev == bdev && ce->e_block == block)
 393                         goto out;
 394         }
 395         __mb_cache_entry_unhash(ce);
 396         ce->e_bdev = bdev;
 397         ce->e_block = block;
 398         list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
 399         ce->e_index.o_key = key;
 400         bucket = hash_long(key, cache->c_bucket_bits);
 401         list_add(&ce->e_index.o_list, &cache->c_index_hash[bucket]);
 402         error = 0;
 403 out:
 404         spin_unlock(&mb_cache_spinlock);
 405         return error;
 406 }
 407
 408
 409 /*
 410  * mb_cache_entry_release()
 411  *
 412  * Release a handle to a cache entry. When the last handle to a cache entry
 413  * is released it is either freed (if it is invalid) or otherwise inserted
 414  * in to the lru list.
 415  */
 416 void
 417 mb_cache_entry_release(struct mb_cache_entry *ce)
 418 {
 419         spin_lock(&mb_cache_spinlock);
 420         __mb_cache_entry_release_unlock(ce);
 421 }
 422
 423
 424 /*
 425  * mb_cache_entry_free()
 426  *
 427  * This is equivalent to the sequence mb_cache_entry_takeout() --
 428  * mb_cache_entry_release().
 429  */
 430 void
 431 mb_cache_entry_free(struct mb_cache_entry *ce)
 432 {
 433         spin_lock(&mb_cache_spinlock);
 434         mb_assert(list_empty(&ce->e_lru_list));
 435         __mb_cache_entry_unhash(ce);
 436         __mb_cache_entry_release_unlock(ce);
 437 }
 438
 439
 440 /*
 441  * mb_cache_entry_get()
 442  *
 443  * Get a cache entry  by device / block number. (There can only be one entry
 444  * in the cache per device and block.) Returns NULL if no such cache entry
 445  * exists. The returned cache entry is locked for exclusive access ("single
 446  * writer").
 447  */
 448 struct mb_cache_entry *
 449 mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
 450                    sector_t block)
 451 {
 452         unsigned int bucket;
 453         struct list_head *l;
 454         struct mb_cache_entry *ce;
 455
 456         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 457                            cache->c_bucket_bits);
 458         spin_lock(&mb_cache_spinlock);
 459         list_for_each(l, &cache->c_block_hash[bucket]) {
 460                 ce = list_entry(l, struct mb_cache_entry, e_block_list);
 461                 if (ce->e_bdev == bdev && ce->e_block == block) {
 462                         DEFINE_WAIT(wait);
 463
 464                         if (!list_empty(&ce->e_lru_list))
 465                                 list_del_init(&ce->e_lru_list);
 466
 467                         while (ce->e_used > 0) {
 468                                 ce->e_queued++;
 469                                 prepare_to_wait(&mb_cache_queue, &wait,
 470                                                 TASK_UNINTERRUPTIBLE);
 471                                 spin_unlock(&mb_cache_spinlock);
 472                                 schedule();
 473                                 spin_lock(&mb_cache_spinlock);
 474                                 ce->e_queued--;
 475                         }
 476                         finish_wait(&mb_cache_queue, &wait);
 477                         ce->e_used += 1 + MB_CACHE_WRITER;
 478
 479                         if (!__mb_cache_entry_is_hashed(ce)) {
 480                                 __mb_cache_entry_release_unlock(ce);
 481                                 return NULL;
 482                         }
 483                         goto cleanup;
 484                 }
 485         }
 486         ce = NULL;
 487
 488 cleanup:
 489         spin_unlock(&mb_cache_spinlock);
 490         return ce;
 491 }
 492
 493 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
 494
 495 static struct mb_cache_entry *
 496 __mb_cache_entry_find(struct list_head *l, struct list_head *head,
 497                       struct block_device *bdev, unsigned int key)
 498 {
 499         while (l != head) {
 500                 struct mb_cache_entry *ce =
 501                         list_entry(l, struct mb_cache_entry, e_index.o_list);
 502                 if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
 503                         DEFINE_WAIT(wait);
 504
 505                         if (!list_empty(&ce->e_lru_list))
 506                                 list_del_init(&ce->e_lru_list);
 507
 508                         /* Incrementing before holding the lock gives readers
 509                            priority over writers. */
 510                         ce->e_used++;
 511                         while (ce->e_used >= MB_CACHE_WRITER) {
 512                                 ce->e_queued++;
 513                                 prepare_to_wait(&mb_cache_queue, &wait,
 514                                                 TASK_UNINTERRUPTIBLE);
 515                                 spin_unlock(&mb_cache_spinlock);
 516                                 schedule();
 517                                 spin_lock(&mb_cache_spinlock);
 518                                 ce->e_queued--;
 519                         }
 520                         finish_wait(&mb_cache_queue, &wait);
 521
 522                         if (!__mb_cache_entry_is_hashed(ce)) {
 523                                 __mb_cache_entry_release_unlock(ce);
 524                                 spin_lock(&mb_cache_spinlock);
 525                                 return ERR_PTR(-EAGAIN);
 526                         }
 527                         return ce;
 528                 }
 529                 l = l->next;
 530         }
 531         return NULL;
 532 }
 533
 534
 535 /*
 536  * mb_cache_entry_find_first()
 537  *
 538  * Find the first cache entry on a given device with a certain key in
 539  * an additional index. Additonal matches can be found with
 540  * mb_cache_entry_find_next(). Returns NULL if no match was found. The
 541  * returned cache entry is locked for shared access ("multiple readers").
 542  *
 543  * @cache: the cache to search
 544  * @bdev: the device the cache entry should belong to
 545  * @key: the key in the index
 546  */
 547 struct mb_cache_entry *
 548 mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
 549                           unsigned int key)
 550 {
 551         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 552         struct list_head *l;
 553         struct mb_cache_entry *ce;
 554
 555         spin_lock(&mb_cache_spinlock);
 556         l = cache->c_index_hash[bucket].next;
 557         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 558         spin_unlock(&mb_cache_spinlock);
 559         return ce;
 560 }
 561
 562
 563 /*
 564  * mb_cache_entry_find_next()
 565  *
 566  * Find the next cache entry on a given device with a certain key in an
 567  * additional index. Returns NULL if no match could be found. The previous
 568  * entry is atomatically released, so that mb_cache_entry_find_next() can
 569  * be called like this:
 570  *
 571  * entry = mb_cache_entry_find_first();
 572  * while (entry) {
 573  *      ...
 574  *      entry = mb_cache_entry_find_next(entry, ...);
 575  * }
 576  *
 577  * @prev: The previous match
 578  * @bdev: the device the cache entry should belong to
 579  * @key: the key in the index
 580  */
 581 struct mb_cache_entry *
 582 mb_cache_entry_find_next(struct mb_cache_entry *prev,
 583                          struct block_device *bdev, unsigned int key)
 584 {
 585         struct mb_cache *cache = prev->e_cache;
 586         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 587         struct list_head *l;
 588         struct mb_cache_entry *ce;
 589
 590         spin_lock(&mb_cache_spinlock);
 591         l = prev->e_index.o_list.next;
 592         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 593         __mb_cache_entry_release_unlock(prev);
 594         return ce;
 595 }
 596
 597 #endif  /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
 598
 599 static int __init init_mbcache(void)
 600 {
 601         register_shrinker(&mb_cache_shrinker);
 602         return 0;
 603 }
 604
 605 static void __exit exit_mbcache(void)
 606 {
 607         unregister_shrinker(&mb_cache_shrinker);
 608 }
 609
 610 module_init(init_mbcache)
 611 module_exit(exit_mbcache)
 612