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 static inline int
  90 __mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
  91 {
  92         return !list_empty(&ce->e_block_list);
  93 }
  94
  95
  96 static void
  97 __mb_cache_entry_unhash(struct mb_cache_entry *ce)
  98 {
  99         if (__mb_cache_entry_is_hashed(ce)) {
 100                 list_del_init(&ce->e_block_list);
 101                 list_del(&ce->e_index.o_list);
 102         }
 103 }
 104
 105
 106 static void
 107 __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
 108 {
 109         struct mb_cache *cache = ce->e_cache;
 110
 111         mb_assert(!(ce->e_used || ce->e_queued));
 112         kmem_cache_free(cache->c_entry_cache, ce);
 113         atomic_dec(&cache->c_entry_count);
 114 }
 115
 116
 117 static void
 118 __mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
 119         __releases(mb_cache_spinlock)
 120 {
 121         /* Wake up all processes queuing for this cache entry. */
 122         if (ce->e_queued)
 123                 wake_up_all(&mb_cache_queue);
 124         if (ce->e_used >= MB_CACHE_WRITER)
 125                 ce->e_used -= MB_CACHE_WRITER;
 126         ce->e_used--;
 127         if (!(ce->e_used || ce->e_queued)) {
 128                 if (!__mb_cache_entry_is_hashed(ce))
 129                         goto forget;
 130                 mb_assert(list_empty(&ce->e_lru_list));
 131                 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
 132         }
 133         spin_unlock(&mb_cache_spinlock);
 134         return;
 135 forget:
 136         spin_unlock(&mb_cache_spinlock);
 137         __mb_cache_entry_forget(ce, GFP_KERNEL);
 138 }
 139
 140
 141 /*
 142  * mb_cache_shrink_scan()  memory pressure callback
 143  *
 144  * This function is called by the kernel memory management when memory
 145  * gets low.
 146  *
 147  * @shrink: (ignored)
 148  * @sc: shrink_control passed from reclaim
 149  *
 150  * Returns the number of objects freed.
 151  */
 152 static unsigned long
 153 mb_cache_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
 154 {
 155         LIST_HEAD(free_list);
 156         struct mb_cache_entry *entry, *tmp;
 157         int nr_to_scan = sc->nr_to_scan;
 158         gfp_t gfp_mask = sc->gfp_mask;
 159         unsigned long freed = 0;
 160
 161         mb_debug("trying to free %d entries", nr_to_scan);
 162         spin_lock(&mb_cache_spinlock);
 163         while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
 164                 struct mb_cache_entry *ce =
 165                         list_entry(mb_cache_lru_list.next,
 166                                    struct mb_cache_entry, e_lru_list);
 167                 list_move_tail(&ce->e_lru_list, &free_list);
 168                 __mb_cache_entry_unhash(ce);
 169                 freed++;
 170         }
 171         spin_unlock(&mb_cache_spinlock);
 172         list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
 173                 __mb_cache_entry_forget(entry, gfp_mask);
 174         }
 175         return freed;
 176 }
 177
 178 static unsigned long
 179 mb_cache_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
 180 {
 181         struct mb_cache *cache;
 182         unsigned long count = 0;
 183
 184         spin_lock(&mb_cache_spinlock);
 185         list_for_each_entry(cache, &mb_cache_list, c_cache_list) {
 186                 mb_debug("cache %s (%d)", cache->c_name,
 187                           atomic_read(&cache->c_entry_count));
 188                 count += atomic_read(&cache->c_entry_count);
 189         }
 190         spin_unlock(&mb_cache_spinlock);
 191
 192         return vfs_pressure_ratio(count);
 193 }
 194
 195 static struct shrinker mb_cache_shrinker = {
 196         .count_objects = mb_cache_shrink_count,
 197         .scan_objects = mb_cache_shrink_scan,
 198         .seeks = DEFAULT_SEEKS,
 199 };
 200
 201 /*
 202  * mb_cache_create()  create a new cache
 203  *
 204  * All entries in one cache are equal size. Cache entries may be from
 205  * multiple devices. If this is the first mbcache created, registers
 206  * the cache with kernel memory management. Returns NULL if no more
 207  * memory was available.
 208  *
 209  * @name: name of the cache (informal)
 210  * @bucket_bits: log2(number of hash buckets)
 211  */
 212 struct mb_cache *
 213 mb_cache_create(const char *name, int bucket_bits)
 214 {
 215         int n, bucket_count = 1 << bucket_bits;
 216         struct mb_cache *cache = NULL;
 217
 218         cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
 219         if (!cache)
 220                 return NULL;
 221         cache->c_name = name;
 222         atomic_set(&cache->c_entry_count, 0);
 223         cache->c_bucket_bits = bucket_bits;
 224         cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
 225                                       GFP_KERNEL);
 226         if (!cache->c_block_hash)
 227                 goto fail;
 228         for (n=0; n<bucket_count; n++)
 229                 INIT_LIST_HEAD(&cache->c_block_hash[n]);
 230         cache->c_index_hash = kmalloc(bucket_count * sizeof(struct list_head),
 231                                       GFP_KERNEL);
 232         if (!cache->c_index_hash)
 233                 goto fail;
 234         for (n=0; n<bucket_count; n++)
 235                 INIT_LIST_HEAD(&cache->c_index_hash[n]);
 236         cache->c_entry_cache = kmem_cache_create(name,
 237                 sizeof(struct mb_cache_entry), 0,
 238                 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
 239         if (!cache->c_entry_cache)
 240                 goto fail2;
 241
 242         /*
 243          * Set an upper limit on the number of cache entries so that the hash
 244          * chains won't grow too long.
 245          */
 246         cache->c_max_entries = bucket_count << 4;
 247
 248         spin_lock(&mb_cache_spinlock);
 249         list_add(&cache->c_cache_list, &mb_cache_list);
 250         spin_unlock(&mb_cache_spinlock);
 251         return cache;
 252
 253 fail2:
 254         kfree(cache->c_index_hash);
 255
 256 fail:
 257         kfree(cache->c_block_hash);
 258         kfree(cache);
 259         return NULL;
 260 }
 261
 262
 263 /*
 264  * mb_cache_shrink()
 265  *
 266  * Removes all cache entries of a device from the cache. All cache entries
 267  * currently in use cannot be freed, and thus remain in the cache. All others
 268  * are freed.
 269  *
 270  * @bdev: which device's cache entries to shrink
 271  */
 272 void
 273 mb_cache_shrink(struct block_device *bdev)
 274 {
 275         LIST_HEAD(free_list);
 276         struct list_head *l, *ltmp;
 277
 278         spin_lock(&mb_cache_spinlock);
 279         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 280                 struct mb_cache_entry *ce =
 281                         list_entry(l, struct mb_cache_entry, e_lru_list);
 282                 if (ce->e_bdev == bdev) {
 283                         list_move_tail(&ce->e_lru_list, &free_list);
 284                         __mb_cache_entry_unhash(ce);
 285                 }
 286         }
 287         spin_unlock(&mb_cache_spinlock);
 288         list_for_each_safe(l, ltmp, &free_list) {
 289                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 290                                                    e_lru_list), GFP_KERNEL);
 291         }
 292 }
 293
 294
 295 /*
 296  * mb_cache_destroy()
 297  *
 298  * Shrinks the cache to its minimum possible size (hopefully 0 entries),
 299  * and then destroys it. If this was the last mbcache, un-registers the
 300  * mbcache from kernel memory management.
 301  */
 302 void
 303 mb_cache_destroy(struct mb_cache *cache)
 304 {
 305         LIST_HEAD(free_list);
 306         struct list_head *l, *ltmp;
 307
 308         spin_lock(&mb_cache_spinlock);
 309         list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
 310                 struct mb_cache_entry *ce =
 311                         list_entry(l, struct mb_cache_entry, e_lru_list);
 312                 if (ce->e_cache == cache) {
 313                         list_move_tail(&ce->e_lru_list, &free_list);
 314                         __mb_cache_entry_unhash(ce);
 315                 }
 316         }
 317         list_del(&cache->c_cache_list);
 318         spin_unlock(&mb_cache_spinlock);
 319
 320         list_for_each_safe(l, ltmp, &free_list) {
 321                 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
 322                                                    e_lru_list), GFP_KERNEL);
 323         }
 324
 325         if (atomic_read(&cache->c_entry_count) > 0) {
 326                 mb_error("cache %s: %d orphaned entries",
 327                           cache->c_name,
 328                           atomic_read(&cache->c_entry_count));
 329         }
 330
 331         kmem_cache_destroy(cache->c_entry_cache);
 332
 333         kfree(cache->c_index_hash);
 334         kfree(cache->c_block_hash);
 335         kfree(cache);
 336 }
 337
 338 /*
 339  * mb_cache_entry_alloc()
 340  *
 341  * Allocates a new cache entry. The new entry will not be valid initially,
 342  * and thus cannot be looked up yet. It should be filled with data, and
 343  * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
 344  * if no more memory was available.
 345  */
 346 struct mb_cache_entry *
 347 mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
 348 {
 349         struct mb_cache_entry *ce = NULL;
 350
 351         if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
 352                 spin_lock(&mb_cache_spinlock);
 353                 if (!list_empty(&mb_cache_lru_list)) {
 354                         ce = list_entry(mb_cache_lru_list.next,
 355                                         struct mb_cache_entry, e_lru_list);
 356                         list_del_init(&ce->e_lru_list);
 357                         __mb_cache_entry_unhash(ce);
 358                 }
 359                 spin_unlock(&mb_cache_spinlock);
 360         }
 361         if (!ce) {
 362                 ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
 363                 if (!ce)
 364                         return NULL;
 365                 atomic_inc(&cache->c_entry_count);
 366                 INIT_LIST_HEAD(&ce->e_lru_list);
 367                 INIT_LIST_HEAD(&ce->e_block_list);
 368                 ce->e_cache = cache;
 369                 ce->e_queued = 0;
 370         }
 371         ce->e_used = 1 + MB_CACHE_WRITER;
 372         return ce;
 373 }
 374
 375
 376 /*
 377  * mb_cache_entry_insert()
 378  *
 379  * Inserts an entry that was allocated using mb_cache_entry_alloc() into
 380  * the cache. After this, the cache entry can be looked up, but is not yet
 381  * in the lru list as the caller still holds a handle to it. Returns 0 on
 382  * success, or -EBUSY if a cache entry for that device + inode exists
 383  * already (this may happen after a failed lookup, but when another process
 384  * has inserted the same cache entry in the meantime).
 385  *
 386  * @bdev: device the cache entry belongs to
 387  * @block: block number
 388  * @key: lookup key
 389  */
 390 int
 391 mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
 392                       sector_t block, unsigned int key)
 393 {
 394         struct mb_cache *cache = ce->e_cache;
 395         unsigned int bucket;
 396         struct list_head *l;
 397         int error = -EBUSY;
 398
 399         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 400                            cache->c_bucket_bits);
 401         spin_lock(&mb_cache_spinlock);
 402         list_for_each_prev(l, &cache->c_block_hash[bucket]) {
 403                 struct mb_cache_entry *ce =
 404                         list_entry(l, struct mb_cache_entry, e_block_list);
 405                 if (ce->e_bdev == bdev && ce->e_block == block)
 406                         goto out;
 407         }
 408         __mb_cache_entry_unhash(ce);
 409         ce->e_bdev = bdev;
 410         ce->e_block = block;
 411         list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
 412         ce->e_index.o_key = key;
 413         bucket = hash_long(key, cache->c_bucket_bits);
 414         list_add(&ce->e_index.o_list, &cache->c_index_hash[bucket]);
 415         error = 0;
 416 out:
 417         spin_unlock(&mb_cache_spinlock);
 418         return error;
 419 }
 420
 421
 422 /*
 423  * mb_cache_entry_release()
 424  *
 425  * Release a handle to a cache entry. When the last handle to a cache entry
 426  * is released it is either freed (if it is invalid) or otherwise inserted
 427  * in to the lru list.
 428  */
 429 void
 430 mb_cache_entry_release(struct mb_cache_entry *ce)
 431 {
 432         spin_lock(&mb_cache_spinlock);
 433         __mb_cache_entry_release_unlock(ce);
 434 }
 435
 436
 437 /*
 438  * mb_cache_entry_free()
 439  *
 440  * This is equivalent to the sequence mb_cache_entry_takeout() --
 441  * mb_cache_entry_release().
 442  */
 443 void
 444 mb_cache_entry_free(struct mb_cache_entry *ce)
 445 {
 446         spin_lock(&mb_cache_spinlock);
 447         mb_assert(list_empty(&ce->e_lru_list));
 448         __mb_cache_entry_unhash(ce);
 449         __mb_cache_entry_release_unlock(ce);
 450 }
 451
 452
 453 /*
 454  * mb_cache_entry_get()
 455  *
 456  * Get a cache entry  by device / block number. (There can only be one entry
 457  * in the cache per device and block.) Returns NULL if no such cache entry
 458  * exists. The returned cache entry is locked for exclusive access ("single
 459  * writer").
 460  */
 461 struct mb_cache_entry *
 462 mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
 463                    sector_t block)
 464 {
 465         unsigned int bucket;
 466         struct list_head *l;
 467         struct mb_cache_entry *ce;
 468
 469         bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
 470                            cache->c_bucket_bits);
 471         spin_lock(&mb_cache_spinlock);
 472         list_for_each(l, &cache->c_block_hash[bucket]) {
 473                 ce = list_entry(l, struct mb_cache_entry, e_block_list);
 474                 if (ce->e_bdev == bdev && ce->e_block == block) {
 475                         DEFINE_WAIT(wait);
 476
 477                         if (!list_empty(&ce->e_lru_list))
 478                                 list_del_init(&ce->e_lru_list);
 479
 480                         while (ce->e_used > 0) {
 481                                 ce->e_queued++;
 482                                 prepare_to_wait(&mb_cache_queue, &wait,
 483                                                 TASK_UNINTERRUPTIBLE);
 484                                 spin_unlock(&mb_cache_spinlock);
 485                                 schedule();
 486                                 spin_lock(&mb_cache_spinlock);
 487                                 ce->e_queued--;
 488                         }
 489                         finish_wait(&mb_cache_queue, &wait);
 490                         ce->e_used += 1 + MB_CACHE_WRITER;
 491
 492                         if (!__mb_cache_entry_is_hashed(ce)) {
 493                                 __mb_cache_entry_release_unlock(ce);
 494                                 return NULL;
 495                         }
 496                         goto cleanup;
 497                 }
 498         }
 499         ce = NULL;
 500
 501 cleanup:
 502         spin_unlock(&mb_cache_spinlock);
 503         return ce;
 504 }
 505
 506 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
 507
 508 static struct mb_cache_entry *
 509 __mb_cache_entry_find(struct list_head *l, struct list_head *head,
 510                       struct block_device *bdev, unsigned int key)
 511 {
 512         while (l != head) {
 513                 struct mb_cache_entry *ce =
 514                         list_entry(l, struct mb_cache_entry, e_index.o_list);
 515                 if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
 516                         DEFINE_WAIT(wait);
 517
 518                         if (!list_empty(&ce->e_lru_list))
 519                                 list_del_init(&ce->e_lru_list);
 520
 521                         /* Incrementing before holding the lock gives readers
 522                            priority over writers. */
 523                         ce->e_used++;
 524                         while (ce->e_used >= MB_CACHE_WRITER) {
 525                                 ce->e_queued++;
 526                                 prepare_to_wait(&mb_cache_queue, &wait,
 527                                                 TASK_UNINTERRUPTIBLE);
 528                                 spin_unlock(&mb_cache_spinlock);
 529                                 schedule();
 530                                 spin_lock(&mb_cache_spinlock);
 531                                 ce->e_queued--;
 532                         }
 533                         finish_wait(&mb_cache_queue, &wait);
 534
 535                         if (!__mb_cache_entry_is_hashed(ce)) {
 536                                 __mb_cache_entry_release_unlock(ce);
 537                                 spin_lock(&mb_cache_spinlock);
 538                                 return ERR_PTR(-EAGAIN);
 539                         }
 540                         return ce;
 541                 }
 542                 l = l->next;
 543         }
 544         return NULL;
 545 }
 546
 547
 548 /*
 549  * mb_cache_entry_find_first()
 550  *
 551  * Find the first cache entry on a given device with a certain key in
 552  * an additional index. Additional matches can be found with
 553  * mb_cache_entry_find_next(). Returns NULL if no match was found. The
 554  * returned cache entry is locked for shared access ("multiple readers").
 555  *
 556  * @cache: the cache to search
 557  * @bdev: the device the cache entry should belong to
 558  * @key: the key in the index
 559  */
 560 struct mb_cache_entry *
 561 mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
 562                           unsigned int key)
 563 {
 564         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 565         struct list_head *l;
 566         struct mb_cache_entry *ce;
 567
 568         spin_lock(&mb_cache_spinlock);
 569         l = cache->c_index_hash[bucket].next;
 570         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 571         spin_unlock(&mb_cache_spinlock);
 572         return ce;
 573 }
 574
 575
 576 /*
 577  * mb_cache_entry_find_next()
 578  *
 579  * Find the next cache entry on a given device with a certain key in an
 580  * additional index. Returns NULL if no match could be found. The previous
 581  * entry is atomatically released, so that mb_cache_entry_find_next() can
 582  * be called like this:
 583  *
 584  * entry = mb_cache_entry_find_first();
 585  * while (entry) {
 586  *      ...
 587  *      entry = mb_cache_entry_find_next(entry, ...);
 588  * }
 589  *
 590  * @prev: The previous match
 591  * @bdev: the device the cache entry should belong to
 592  * @key: the key in the index
 593  */
 594 struct mb_cache_entry *
 595 mb_cache_entry_find_next(struct mb_cache_entry *prev,
 596                          struct block_device *bdev, unsigned int key)
 597 {
 598         struct mb_cache *cache = prev->e_cache;
 599         unsigned int bucket = hash_long(key, cache->c_bucket_bits);
 600         struct list_head *l;
 601         struct mb_cache_entry *ce;
 602
 603         spin_lock(&mb_cache_spinlock);
 604         l = prev->e_index.o_list.next;
 605         ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
 606         __mb_cache_entry_release_unlock(prev);
 607         return ce;
 608 }
 609
 610 #endif  /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
 611
 612 static int __init init_mbcache(void)
 613 {
 614         register_shrinker(&mb_cache_shrinker);
 615         return 0;
 616 }
 617
 618 static void __exit exit_mbcache(void)
 619 {
 620         unregister_shrinker(&mb_cache_shrinker);
 621 }
 622
 623 module_init(init_mbcache)
 624 module_exit(exit_mbcache)
 625