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 /*
  80  * Global data: list of all mbcache's, lru list, and a spinlock for
  81  * accessing cache data structures on SMP machines. The lru list is
  82  * global across all mbcaches.
  83  */
  84
  85 static LIST_HEAD(mb_cache_list);
  86 static LIST_HEAD(mb_cache_lru_list);
  87 static DEFINE_SPINLOCK(mb_cache_spinlock);
  88
  89 /*
  90  * What the mbcache registers as to get shrunk dynamically.
  91  */
  92
  93 static int mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask);
  94
  95 static struct shrinker mb_cache_shrinker = {
  96         .shrink = mb_cache_shrink_fn,
  97         .seeks = DEFAULT_SEEKS,
  98 };
  99
 100 static inline int
 101 __mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
 102 {
 103         return !list_empty(&ce->e_block_list);
 104 }
 105
 106
 107 static void
 108 __mb_cache_entry_unhash(struct mb_cache_entry *ce)
 109 {
 110         if (__mb_cache_entry_is_hashed(ce)) {
 111                 list_del_init(&ce->e_block_list);
 112                 list_del(&ce->e_index.o_list);
 113         }
 114 }
 115
 116
 117 static void
 118 __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
 119 {
 120         struct mb_cache *cache = ce->e_cache;
 121
 122         mb_assert(!(ce->e_used || ce->e_queued));
 123         kmem_cache_free(cache->c_entry_cache, ce);
 124         atomic_dec(&cache->c_entry_count);
 125 }
 126
 127
 128 static void
 129 __mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
 130         __releases(mb_cache_spinlock)
 131 {
 132         /* Wake up all processes queuing for this cache entry. */
 133         if (ce->e_queued)
 134                 wake_up_all(&mb_cache_queue);
 135         if (ce->e_used >= MB_CACHE_WRITER)
 136                 ce->e_used -= MB_CACHE_WRITER;
 137         ce->e_used--;
 138         if (!(ce->e_used || ce->e_queued)) {
 139                 if (!__mb_cache_entry_is_hashed(ce))
 140                         goto forget;
 141                 mb_assert(list_empty(&ce->e_lru_list));
 142                 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
 143         }
 144         spin_unlock(&mb_cache_spinlock);
 145         return;
 146 forget:
 147         spin_unlock(&mb_cache_spinlock);
 148         __mb_cache_entry_forget(ce, GFP_KERNEL);
 149 }
 150
 151
 152 /*
 153  * mb_cache_shrink_fn()  memory pressure callback
 154  *
 155  * This function is called by the kernel memory management when memory
 156  * gets low.
 157  *
 158  * @shrink: (ignored)
 159  * @nr_to_scan: Number of objects to scan
 160  * @gfp_mask: (ignored)
 161  *
 162  * Returns the number of objects which are present in the cache.
 163  */
 164 static int
 165 mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
 166 {
 167         LIST_HEAD(free_list);
 168         struct mb_cache *cache;
 169         struct mb_cache_entry *entry, *tmp;
 170         int count = 0;
 171
 172         mb_debug("trying to free %d entries", nr_to_scan);
 173         spin_lock(&mb_cache_spinlock);
 174         while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
 175                 struct mb_cache_entry *ce =
 176                         list_entry(mb_cache_lru_list.next,
 177                                    struct mb_cache_entry, e_lru_list);
 178                 list_move_tail(&ce->e_lru_list, &free_list);
 179                 __mb_cache_entry_unhash(ce);
 180         }
 181         list_for_each_entry(cache, &mb_cache_list, c_cache_list) {
 182                 mb_debug("cache %s (%d)", cache->c_name,
 183                           atomic_read(&cache->c_entry_count));
 184                 count += atomic_read(&cache->c_entry_count);
 185         }
 186         spin_unlock(&mb_cache_spinlock);
 187         list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
 188                 __mb_cache_entry_forget(entry, gfp_mask);
 189         }
 190         return (count / 100) * sysctl_vfs_cache_pressure;
 191 }
 192
 193
 194 /*
 195  * mb_cache_create()  create a new cache
 196  *
 197  * All entries in one cache are equal size. Cache entries may be from
 198  * multiple devices. If this is the first mbcache created, registers
 199  * the cache with kernel memory management. Returns NULL if no more
 200  * memory was available.
 201  *
 202  * @name: name of the cache (informal)
 203  * @bucket_bits: log2(number of hash buckets)
 204  */
 205 struct mb_cache *
 206 mb_cache_create(const char *name, int bucket_bits)
 207 {
 208         int n, bucket_count = 1 << bucket_bits;
 209         struct mb_cache *cache = NULL;
 210
 211         cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
 212         if (!cache)
 213                 return NULL;
 214         cache->c_name = name;
 215         atomic_set(&cache->c_entry_count, 0);
 216         cache->c_bucket_bits = bucket_bits;
 217         cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
 218                                       GFP_KERNEL);
 219         if (!cache->c_block_hash)
 220                 goto fail;
 221         for (n=0; n<bucket_count; n++)
 222                 INIT_LIST_HEAD(&cache->c_block_hash[n]);
 223         cache->c_index_hash = kmalloc(bucket_count * sizeof(struct list_head),
 224                                       GFP_KERNEL);
 225         if (!cache->c_index_hash)
 226                 goto fail;
 227         for (n=0; n<bucket_count; n++)
 228                 INIT_LIST_HEAD(&cache->c_index_hash[n]);
 229         cache->c_entry_cache = kmem_cache_create(name,
 230                 sizeof(struct mb_cache_entry), 0,
 231                 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
 232         if (!cache->c_entry_cache)
 233                 goto fail2;
 234
 235         /*
 236          * Set an upper limit on the number of cache entries so that the hash
 237          * chains won't grow too long.
 238          */
 239         cache->c_max_entries = bucket_count << 4;
 240
 241         spin_lock(&mb_cache_spinlock);
 242         list_add(&cache->c_cache_list, &mb_cache_list);
 243         spin_unlock(&mb_cache_spinlock);
 244         return cache;
 245
 246 fail2:
 247         kfree(cache->c_index_hash);
 248
 249 fail:
 250         kfree(cache->c_block_hash);
 251         kfree(cache);
 252         return NULL;
 253 }
 254
 255
 256 /*
 257  * mb_cache_shrink()
 258  *
 259  * Removes all cache entries of a device from the cache. All cache entries
 260  * currently in use cannot be freed, and thus remain in the cache. All others
 261  * are freed.
 262  *
 263  * @bdev: which device's cache entries to shrink
 264  */
 265 void
 266 mb_cache_shrink(struct block_device *bdev)
 267 {
 268         LIST_HEAD(free_list);
 269         struct list_head *l, *ltmp;
 270
 271         spin_lock(&mb_cache_spinlock);
 272         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 273                 struct mb_cache_entry *ce =
 274                         list_entry(l, struct mb_cache_entry, e_lru_list);
 275                 if (ce->e_bdev == bdev) {
 276                         list_move_tail(&ce->e_lru_list, &free_list);
 277                         __mb_cache_entry_unhash(ce);
 278                 }
 279         }
 280         spin_unlock(&mb_cache_spinlock);
 281         list_for_each_safe(l, ltmp, &free_list) {
 282                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 283                                                    e_lru_list), GFP_KERNEL);
 284         }
 285 }
 286
 287
 288 /*
 289  * mb_cache_destroy()
 290  *
 291  * Shrinks the cache to its minimum possible size (hopefully 0 entries),
 292  * and then destroys it. If this was the last mbcache, un-registers the
 293  * mbcache from kernel memory management.
 294  */
 295 void
 296 mb_cache_destroy(struct mb_cache *cache)
 297 {
 298         LIST_HEAD(free_list);
 299         struct list_head *l, *ltmp;
 300
 301         spin_lock(&mb_cache_spinlock);
 302         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 303                 struct mb_cache_entry *ce =
 304                         list_entry(l, struct mb_cache_entry, e_lru_list);
 305                 if (ce->e_cache == cache) {
 306                         list_move_tail(&ce->e_lru_list, &free_list);
 307                         __mb_cache_entry_unhash(ce);
 308                 }
 309         }
 310         list_del(&cache->c_cache_list);
 311         spin_unlock(&mb_cache_spinlock);
 312
 313         list_for_each_safe(l, ltmp, &free_list) {
 314                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 315                                                    e_lru_list), GFP_KERNEL);
 316         }
 317
 318         if (atomic_read(&cache->c_entry_count) > 0) {
 319                 mb_error("cache %s: %d orphaned entries",
 320                           cache->c_name,
 321                           atomic_read(&cache->c_entry_count));
 322         }
 323
 324         kmem_cache_destroy(cache->c_entry_cache);
 325
 326         kfree(cache->c_index_hash);
 327         kfree(cache->c_block_hash);
 328         kfree(cache);
 329 }
 330
 331 /*
 332  * mb_cache_entry_alloc()
 333  *
 334  * Allocates a new cache entry. The new entry will not be valid initially,
 335  * and thus cannot be looked up yet. It should be filled with data, and
 336  * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
 337  * if no more memory was available.
 338  */
 339 struct mb_cache_entry *
 340 mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
 341 {
 342         struct mb_cache_entry *ce = NULL;
 343
 344         if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
 345                 spin_lock(&mb_cache_spinlock);
 346                 if (!list_empty(&mb_cache_lru_list)) {
 347                         ce = list_entry(mb_cache_lru_list.next,
 348                                         struct mb_cache_entry, e_lru_list);
 349                         list_del_init(&ce->e_lru_list);
 350                         __mb_cache_entry_unhash(ce);
 351                 }
 352                 spin_unlock(&mb_cache_spinlock);
 353         }
 354         if (!ce) {
 355                 ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
 356                 if (!ce)
 357                         return NULL;
 358                 atomic_inc(&cache->c_entry_count);
 359                 INIT_LIST_HEAD(&ce->e_lru_list);
 360                 INIT_LIST_HEAD(&ce->e_block_list);
 361                 ce->e_cache = cache;
 362                 ce->e_queued = 0;
 363         }
 364         ce->e_used = 1 + MB_CACHE_WRITER;
 365         return ce;
 366 }
 367
 368
 369 /*
 370  * mb_cache_entry_insert()
 371  *
 372  * Inserts an entry that was allocated using mb_cache_entry_alloc() into
 373  * the cache. After this, the cache entry can be looked up, but is not yet
 374  * in the lru list as the caller still holds a handle to it. Returns 0 on
 375  * success, or -EBUSY if a cache entry for that device + inode exists
 376  * already (this may happen after a failed lookup, but when another process
 377  * has inserted the same cache entry in the meantime).
 378  *
 379  * @bdev: device the cache entry belongs to
 380  * @block: block number
 381  * @key: lookup key
 382  */
 383 int
 384 mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
 385                       sector_t block, unsigned int key)
 386 {
 387         struct mb_cache *cache = ce->e_cache;
 388         unsigned int bucket;
 389         struct list_head *l;
 390         int error = -EBUSY;
 391
 392         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 393                            cache->c_bucket_bits);
 394         spin_lock(&mb_cache_spinlock);
 395         list_for_each_prev(l, &cache->c_block_hash[bucket]) {
 396                 struct mb_cache_entry *ce =
 397                         list_entry(l, struct mb_cache_entry, e_block_list);
 398                 if (ce->e_bdev == bdev && ce->e_block == block)
 399                         goto out;
 400         }
 401         __mb_cache_entry_unhash(ce);
 402         ce->e_bdev = bdev;
 403         ce->e_block = block;
 404         list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
 405         ce->e_index.o_key = key;
 406         bucket = hash_long(key, cache->c_bucket_bits);
 407         list_add(&ce->e_index.o_list, &cache->c_index_hash[bucket]);
 408         error = 0;
 409 out:
 410         spin_unlock(&mb_cache_spinlock);
 411         return error;
 412 }
 413
 414
 415 /*
 416  * mb_cache_entry_release()
 417  *
 418  * Release a handle to a cache entry. When the last handle to a cache entry
 419  * is released it is either freed (if it is invalid) or otherwise inserted
 420  * in to the lru list.
 421  */
 422 void
 423 mb_cache_entry_release(struct mb_cache_entry *ce)
 424 {
 425         spin_lock(&mb_cache_spinlock);
 426         __mb_cache_entry_release_unlock(ce);
 427 }
 428
 429
 430 /*
 431  * mb_cache_entry_free()
 432  *
 433  * This is equivalent to the sequence mb_cache_entry_takeout() --
 434  * mb_cache_entry_release().
 435  */
 436 void
 437 mb_cache_entry_free(struct mb_cache_entry *ce)
 438 {
 439         spin_lock(&mb_cache_spinlock);
 440         mb_assert(list_empty(&ce->e_lru_list));
 441         __mb_cache_entry_unhash(ce);
 442         __mb_cache_entry_release_unlock(ce);
 443 }
 444
 445
 446 /*
 447  * mb_cache_entry_get()
 448  *
 449  * Get a cache entry  by device / block number. (There can only be one entry
 450  * in the cache per device and block.) Returns NULL if no such cache entry
 451  * exists. The returned cache entry is locked for exclusive access ("single
 452  * writer").
 453  */
 454 struct mb_cache_entry *
 455 mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
 456                    sector_t block)
 457 {
 458         unsigned int bucket;
 459         struct list_head *l;
 460         struct mb_cache_entry *ce;
 461
 462         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 463                            cache->c_bucket_bits);
 464         spin_lock(&mb_cache_spinlock);
 465         list_for_each(l, &cache->c_block_hash[bucket]) {
 466                 ce = list_entry(l, struct mb_cache_entry, e_block_list);
 467                 if (ce->e_bdev == bdev && ce->e_block == block) {
 468                         DEFINE_WAIT(wait);
 469
 470                         if (!list_empty(&ce->e_lru_list))
 471                                 list_del_init(&ce->e_lru_list);
 472
 473                         while (ce->e_used > 0) {
 474                                 ce->e_queued++;
 475                                 prepare_to_wait(&mb_cache_queue, &wait,
 476                                                 TASK_UNINTERRUPTIBLE);
 477                                 spin_unlock(&mb_cache_spinlock);
 478                                 schedule();
 479                                 spin_lock(&mb_cache_spinlock);
 480                                 ce->e_queued--;
 481                         }
 482                         finish_wait(&mb_cache_queue, &wait);
 483                         ce->e_used += 1 + MB_CACHE_WRITER;
 484
 485                         if (!__mb_cache_entry_is_hashed(ce)) {
 486                                 __mb_cache_entry_release_unlock(ce);
 487                                 return NULL;
 488                         }
 489                         goto cleanup;
 490                 }
 491         }
 492         ce = NULL;
 493
 494 cleanup:
 495         spin_unlock(&mb_cache_spinlock);
 496         return ce;
 497 }
 498
 499 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
 500
 501 static struct mb_cache_entry *
 502 __mb_cache_entry_find(struct list_head *l, struct list_head *head,
 503                       struct block_device *bdev, unsigned int key)
 504 {
 505         while (l != head) {
 506                 struct mb_cache_entry *ce =
 507                         list_entry(l, struct mb_cache_entry, e_index.o_list);
 508                 if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
 509                         DEFINE_WAIT(wait);
 510
 511                         if (!list_empty(&ce->e_lru_list))
 512                                 list_del_init(&ce->e_lru_list);
 513
 514                         /* Incrementing before holding the lock gives readers
 515                            priority over writers. */
 516                         ce->e_used++;
 517                         while (ce->e_used >= MB_CACHE_WRITER) {
 518                                 ce->e_queued++;
 519                                 prepare_to_wait(&mb_cache_queue, &wait,
 520                                                 TASK_UNINTERRUPTIBLE);
 521                                 spin_unlock(&mb_cache_spinlock);
 522                                 schedule();
 523                                 spin_lock(&mb_cache_spinlock);
 524                                 ce->e_queued--;
 525                         }
 526                         finish_wait(&mb_cache_queue, &wait);
 527
 528                         if (!__mb_cache_entry_is_hashed(ce)) {
 529                                 __mb_cache_entry_release_unlock(ce);
 530                                 spin_lock(&mb_cache_spinlock);
 531                                 return ERR_PTR(-EAGAIN);
 532                         }
 533                         return ce;
 534                 }
 535                 l = l->next;
 536         }
 537         return NULL;
 538 }
 539
 540
 541 /*
 542  * mb_cache_entry_find_first()
 543  *
 544  * Find the first cache entry on a given device with a certain key in
 545  * an additional index. Additonal matches can be found with
 546  * mb_cache_entry_find_next(). Returns NULL if no match was found. The
 547  * returned cache entry is locked for shared access ("multiple readers").
 548  *
 549  * @cache: the cache to search
 550  * @bdev: the device the cache entry should belong to
 551  * @key: the key in the index
 552  */
 553 struct mb_cache_entry *
 554 mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
 555                           unsigned int key)
 556 {
 557         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 558         struct list_head *l;
 559         struct mb_cache_entry *ce;
 560
 561         spin_lock(&mb_cache_spinlock);
 562         l = cache->c_index_hash[bucket].next;
 563         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 564         spin_unlock(&mb_cache_spinlock);
 565         return ce;
 566 }
 567
 568
 569 /*
 570  * mb_cache_entry_find_next()
 571  *
 572  * Find the next cache entry on a given device with a certain key in an
 573  * additional index. Returns NULL if no match could be found. The previous
 574  * entry is atomatically released, so that mb_cache_entry_find_next() can
 575  * be called like this:
 576  *
 577  * entry = mb_cache_entry_find_first();
 578  * while (entry) {
 579  *      ...
 580  *      entry = mb_cache_entry_find_next(entry, ...);
 581  * }
 582  *
 583  * @prev: The previous match
 584  * @bdev: the device the cache entry should belong to
 585  * @key: the key in the index
 586  */
 587 struct mb_cache_entry *
 588 mb_cache_entry_find_next(struct mb_cache_entry *prev,
 589                          struct block_device *bdev, unsigned int key)
 590 {
 591         struct mb_cache *cache = prev->e_cache;
 592         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 593         struct list_head *l;
 594         struct mb_cache_entry *ce;
 595
 596         spin_lock(&mb_cache_spinlock);
 597         l = prev->e_index.o_list.next;
 598         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 599         __mb_cache_entry_release_unlock(prev);
 600         return ce;
 601 }
 602
 603 #endif  /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
 604
 605 static int __init init_mbcache(void)
 606 {
 607         register_shrinker(&mb_cache_shrinker);
 608         return 0;
 609 }
 610
 611 static void __exit exit_mbcache(void)
 612 {
 613         unregister_shrinker(&mb_cache_shrinker);
 614 }
 615
 616 module_init(init_mbcache)
 617 module_exit(exit_mbcache)
 618