3 * (C) 2001-2002 Andreas Gruenbacher, <a.gruenbacher@computer.org>
7 * Filesystem Meta Information Block Cache (mbcache)
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).
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.
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
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.
29 #include <linux/kernel.h>
30 #include <linux/module.h>
32 #include <linux/hash.h>
35 #include <linux/slab.h>
36 #include <linux/sched.h>
37 #include <linux/init.h>
38 #include <linux/mbcache.h>
42 # define mb_debug(f...) do { \
43 printk(KERN_DEBUG f); \
46 #define mb_assert(c) do { if (!(c)) \
47 printk(KERN_ERR "assertion " #c " failed\n"); \
50 # define mb_debug(f...) do { } while(0)
51 # define mb_assert(c) do { } while(0)
53 #define mb_error(f...) do { \
58 #define MB_CACHE_WRITER ((unsigned short)~0U >> 1)
60 static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
62 MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
63 MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
64 MODULE_LICENSE("GPL");
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);
80 struct list_head c_cache_list;
82 struct mb_cache_op c_op;
83 atomic_t c_entry_count;
85 #ifndef MB_CACHE_INDEXES_COUNT
88 kmem_cache_t *c_entry_cache;
89 struct list_head *c_block_hash;
90 struct list_head *c_indexes_hash[0];
95 * Global data: list of all mbcache's, lru list, and a spinlock for
96 * accessing cache data structures on SMP machines. The lru list is
97 * global across all mbcaches.
100 static LIST_HEAD(mb_cache_list);
101 static LIST_HEAD(mb_cache_lru_list);
102 static DEFINE_SPINLOCK(mb_cache_spinlock);
103 static struct shrinker *mb_shrinker;
106 mb_cache_indexes(struct mb_cache *cache)
108 #ifdef MB_CACHE_INDEXES_COUNT
109 return MB_CACHE_INDEXES_COUNT;
111 return cache->c_indexes_count;
116 * What the mbcache registers as to get shrunk dynamically.
119 static int mb_cache_shrink_fn(int nr_to_scan, gfp_t gfp_mask);
123 __mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
125 return !list_empty(&ce->e_block_list);
130 __mb_cache_entry_unhash(struct mb_cache_entry *ce)
134 if (__mb_cache_entry_is_hashed(ce)) {
135 list_del_init(&ce->e_block_list);
136 for (n=0; n<mb_cache_indexes(ce->e_cache); n++)
137 list_del(&ce->e_indexes[n].o_list);
143 __mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
145 struct mb_cache *cache = ce->e_cache;
147 mb_assert(!(ce->e_used || ce->e_queued));
148 if (cache->c_op.free && cache->c_op.free(ce, gfp_mask)) {
149 /* free failed -- put back on the lru list
150 for freeing later. */
151 spin_lock(&mb_cache_spinlock);
152 list_add(&ce->e_lru_list, &mb_cache_lru_list);
153 spin_unlock(&mb_cache_spinlock);
155 kmem_cache_free(cache->c_entry_cache, ce);
156 atomic_dec(&cache->c_entry_count);
162 __mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
164 /* Wake up all processes queuing for this cache entry. */
166 wake_up_all(&mb_cache_queue);
167 if (ce->e_used >= MB_CACHE_WRITER)
168 ce->e_used -= MB_CACHE_WRITER;
170 if (!(ce->e_used || ce->e_queued)) {
171 if (!__mb_cache_entry_is_hashed(ce))
173 mb_assert(list_empty(&ce->e_lru_list));
174 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
176 spin_unlock(&mb_cache_spinlock);
179 spin_unlock(&mb_cache_spinlock);
180 __mb_cache_entry_forget(ce, GFP_KERNEL);
185 * mb_cache_shrink_fn() memory pressure callback
187 * This function is called by the kernel memory management when memory
190 * @nr_to_scan: Number of objects to scan
191 * @gfp_mask: (ignored)
193 * Returns the number of objects which are present in the cache.
196 mb_cache_shrink_fn(int nr_to_scan, gfp_t gfp_mask)
198 LIST_HEAD(free_list);
199 struct list_head *l, *ltmp;
202 spin_lock(&mb_cache_spinlock);
203 list_for_each(l, &mb_cache_list) {
204 struct mb_cache *cache =
205 list_entry(l, struct mb_cache, c_cache_list);
206 mb_debug("cache %s (%d)", cache->c_name,
207 atomic_read(&cache->c_entry_count));
208 count += atomic_read(&cache->c_entry_count);
210 mb_debug("trying to free %d entries", nr_to_scan);
211 if (nr_to_scan == 0) {
212 spin_unlock(&mb_cache_spinlock);
215 while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
216 struct mb_cache_entry *ce =
217 list_entry(mb_cache_lru_list.next,
218 struct mb_cache_entry, e_lru_list);
219 list_move_tail(&ce->e_lru_list, &free_list);
220 __mb_cache_entry_unhash(ce);
222 spin_unlock(&mb_cache_spinlock);
223 list_for_each_safe(l, ltmp, &free_list) {
224 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
225 e_lru_list), gfp_mask);
228 return (count / 100) * sysctl_vfs_cache_pressure;
233 * mb_cache_create() create a new cache
235 * All entries in one cache are equal size. Cache entries may be from
236 * multiple devices. If this is the first mbcache created, registers
237 * the cache with kernel memory management. Returns NULL if no more
238 * memory was available.
240 * @name: name of the cache (informal)
241 * @cache_op: contains the callback called when freeing a cache entry
242 * @entry_size: The size of a cache entry, including
243 * struct mb_cache_entry
244 * @indexes_count: number of additional indexes in the cache. Must equal
245 * MB_CACHE_INDEXES_COUNT if the number of indexes is
247 * @bucket_bits: log2(number of hash buckets)
250 mb_cache_create(const char *name, struct mb_cache_op *cache_op,
251 size_t entry_size, int indexes_count, int bucket_bits)
253 int m=0, n, bucket_count = 1 << bucket_bits;
254 struct mb_cache *cache = NULL;
256 if(entry_size < sizeof(struct mb_cache_entry) +
257 indexes_count * sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]))
260 cache = kmalloc(sizeof(struct mb_cache) +
261 indexes_count * sizeof(struct list_head), GFP_KERNEL);
264 cache->c_name = name;
265 cache->c_op.free = NULL;
267 cache->c_op.free = cache_op->free;
268 atomic_set(&cache->c_entry_count, 0);
269 cache->c_bucket_bits = bucket_bits;
270 #ifdef MB_CACHE_INDEXES_COUNT
271 mb_assert(indexes_count == MB_CACHE_INDEXES_COUNT);
273 cache->c_indexes_count = indexes_count;
275 cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
277 if (!cache->c_block_hash)
279 for (n=0; n<bucket_count; n++)
280 INIT_LIST_HEAD(&cache->c_block_hash[n]);
281 for (m=0; m<indexes_count; m++) {
282 cache->c_indexes_hash[m] = kmalloc(bucket_count *
283 sizeof(struct list_head),
285 if (!cache->c_indexes_hash[m])
287 for (n=0; n<bucket_count; n++)
288 INIT_LIST_HEAD(&cache->c_indexes_hash[m][n]);
290 cache->c_entry_cache = kmem_cache_create(name, entry_size, 0,
291 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL, NULL);
292 if (!cache->c_entry_cache)
295 spin_lock(&mb_cache_spinlock);
296 list_add(&cache->c_cache_list, &mb_cache_list);
297 spin_unlock(&mb_cache_spinlock);
303 kfree(cache->c_indexes_hash[m]);
304 kfree(cache->c_block_hash);
314 * Removes all cache entires of a device from the cache. All cache entries
315 * currently in use cannot be freed, and thus remain in the cache. All others
318 * @bdev: which device's cache entries to shrink
321 mb_cache_shrink(struct block_device *bdev)
323 LIST_HEAD(free_list);
324 struct list_head *l, *ltmp;
326 spin_lock(&mb_cache_spinlock);
327 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
328 struct mb_cache_entry *ce =
329 list_entry(l, struct mb_cache_entry, e_lru_list);
330 if (ce->e_bdev == bdev) {
331 list_move_tail(&ce->e_lru_list, &free_list);
332 __mb_cache_entry_unhash(ce);
335 spin_unlock(&mb_cache_spinlock);
336 list_for_each_safe(l, ltmp, &free_list) {
337 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
338 e_lru_list), GFP_KERNEL);
346 * Shrinks the cache to its minimum possible size (hopefully 0 entries),
347 * and then destroys it. If this was the last mbcache, un-registers the
348 * mbcache from kernel memory management.
351 mb_cache_destroy(struct mb_cache *cache)
353 LIST_HEAD(free_list);
354 struct list_head *l, *ltmp;
357 spin_lock(&mb_cache_spinlock);
358 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
359 struct mb_cache_entry *ce =
360 list_entry(l, struct mb_cache_entry, e_lru_list);
361 if (ce->e_cache == cache) {
362 list_move_tail(&ce->e_lru_list, &free_list);
363 __mb_cache_entry_unhash(ce);
366 list_del(&cache->c_cache_list);
367 spin_unlock(&mb_cache_spinlock);
369 list_for_each_safe(l, ltmp, &free_list) {
370 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
371 e_lru_list), GFP_KERNEL);
374 if (atomic_read(&cache->c_entry_count) > 0) {
375 mb_error("cache %s: %d orphaned entries",
377 atomic_read(&cache->c_entry_count));
380 kmem_cache_destroy(cache->c_entry_cache);
382 for (n=0; n < mb_cache_indexes(cache); n++)
383 kfree(cache->c_indexes_hash[n]);
384 kfree(cache->c_block_hash);
390 * mb_cache_entry_alloc()
392 * Allocates a new cache entry. The new entry will not be valid initially,
393 * and thus cannot be looked up yet. It should be filled with data, and
394 * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
395 * if no more memory was available.
397 struct mb_cache_entry *
398 mb_cache_entry_alloc(struct mb_cache *cache)
400 struct mb_cache_entry *ce;
402 atomic_inc(&cache->c_entry_count);
403 ce = kmem_cache_alloc(cache->c_entry_cache, GFP_KERNEL);
405 INIT_LIST_HEAD(&ce->e_lru_list);
406 INIT_LIST_HEAD(&ce->e_block_list);
408 ce->e_used = 1 + MB_CACHE_WRITER;
416 * mb_cache_entry_insert()
418 * Inserts an entry that was allocated using mb_cache_entry_alloc() into
419 * the cache. After this, the cache entry can be looked up, but is not yet
420 * in the lru list as the caller still holds a handle to it. Returns 0 on
421 * success, or -EBUSY if a cache entry for that device + inode exists
422 * already (this may happen after a failed lookup, but when another process
423 * has inserted the same cache entry in the meantime).
425 * @bdev: device the cache entry belongs to
426 * @block: block number
427 * @keys: array of additional keys. There must be indexes_count entries
428 * in the array (as specified when creating the cache).
431 mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
432 sector_t block, unsigned int keys[])
434 struct mb_cache *cache = ce->e_cache;
437 int error = -EBUSY, n;
439 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
440 cache->c_bucket_bits);
441 spin_lock(&mb_cache_spinlock);
442 list_for_each_prev(l, &cache->c_block_hash[bucket]) {
443 struct mb_cache_entry *ce =
444 list_entry(l, struct mb_cache_entry, e_block_list);
445 if (ce->e_bdev == bdev && ce->e_block == block)
448 __mb_cache_entry_unhash(ce);
451 list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
452 for (n=0; n<mb_cache_indexes(cache); n++) {
453 ce->e_indexes[n].o_key = keys[n];
454 bucket = hash_long(keys[n], cache->c_bucket_bits);
455 list_add(&ce->e_indexes[n].o_list,
456 &cache->c_indexes_hash[n][bucket]);
460 spin_unlock(&mb_cache_spinlock);
466 * mb_cache_entry_release()
468 * Release a handle to a cache entry. When the last handle to a cache entry
469 * is released it is either freed (if it is invalid) or otherwise inserted
470 * in to the lru list.
473 mb_cache_entry_release(struct mb_cache_entry *ce)
475 spin_lock(&mb_cache_spinlock);
476 __mb_cache_entry_release_unlock(ce);
481 * mb_cache_entry_free()
483 * This is equivalent to the sequence mb_cache_entry_takeout() --
484 * mb_cache_entry_release().
487 mb_cache_entry_free(struct mb_cache_entry *ce)
489 spin_lock(&mb_cache_spinlock);
490 mb_assert(list_empty(&ce->e_lru_list));
491 __mb_cache_entry_unhash(ce);
492 __mb_cache_entry_release_unlock(ce);
497 * mb_cache_entry_get()
499 * Get a cache entry by device / block number. (There can only be one entry
500 * in the cache per device and block.) Returns NULL if no such cache entry
501 * exists. The returned cache entry is locked for exclusive access ("single
504 struct mb_cache_entry *
505 mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
510 struct mb_cache_entry *ce;
512 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
513 cache->c_bucket_bits);
514 spin_lock(&mb_cache_spinlock);
515 list_for_each(l, &cache->c_block_hash[bucket]) {
516 ce = list_entry(l, struct mb_cache_entry, e_block_list);
517 if (ce->e_bdev == bdev && ce->e_block == block) {
520 if (!list_empty(&ce->e_lru_list))
521 list_del_init(&ce->e_lru_list);
523 while (ce->e_used > 0) {
525 prepare_to_wait(&mb_cache_queue, &wait,
526 TASK_UNINTERRUPTIBLE);
527 spin_unlock(&mb_cache_spinlock);
529 spin_lock(&mb_cache_spinlock);
532 finish_wait(&mb_cache_queue, &wait);
533 ce->e_used += 1 + MB_CACHE_WRITER;
535 if (!__mb_cache_entry_is_hashed(ce)) {
536 __mb_cache_entry_release_unlock(ce);
545 spin_unlock(&mb_cache_spinlock);
549 #if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
551 static struct mb_cache_entry *
552 __mb_cache_entry_find(struct list_head *l, struct list_head *head,
553 int index, struct block_device *bdev, unsigned int key)
556 struct mb_cache_entry *ce =
557 list_entry(l, struct mb_cache_entry,
558 e_indexes[index].o_list);
559 if (ce->e_bdev == bdev && ce->e_indexes[index].o_key == key) {
562 if (!list_empty(&ce->e_lru_list))
563 list_del_init(&ce->e_lru_list);
565 /* Incrementing before holding the lock gives readers
566 priority over writers. */
568 while (ce->e_used >= MB_CACHE_WRITER) {
570 prepare_to_wait(&mb_cache_queue, &wait,
571 TASK_UNINTERRUPTIBLE);
572 spin_unlock(&mb_cache_spinlock);
574 spin_lock(&mb_cache_spinlock);
577 finish_wait(&mb_cache_queue, &wait);
579 if (!__mb_cache_entry_is_hashed(ce)) {
580 __mb_cache_entry_release_unlock(ce);
581 spin_lock(&mb_cache_spinlock);
582 return ERR_PTR(-EAGAIN);
593 * mb_cache_entry_find_first()
595 * Find the first cache entry on a given device with a certain key in
596 * an additional index. Additonal matches can be found with
597 * mb_cache_entry_find_next(). Returns NULL if no match was found. The
598 * returned cache entry is locked for shared access ("multiple readers").
600 * @cache: the cache to search
601 * @index: the number of the additonal index to search (0<=index<indexes_count)
602 * @bdev: the device the cache entry should belong to
603 * @key: the key in the index
605 struct mb_cache_entry *
606 mb_cache_entry_find_first(struct mb_cache *cache, int index,
607 struct block_device *bdev, unsigned int key)
609 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
611 struct mb_cache_entry *ce;
613 mb_assert(index < mb_cache_indexes(cache));
614 spin_lock(&mb_cache_spinlock);
615 l = cache->c_indexes_hash[index][bucket].next;
616 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
618 spin_unlock(&mb_cache_spinlock);
624 * mb_cache_entry_find_next()
626 * Find the next cache entry on a given device with a certain key in an
627 * additional index. Returns NULL if no match could be found. The previous
628 * entry is atomatically released, so that mb_cache_entry_find_next() can
629 * be called like this:
631 * entry = mb_cache_entry_find_first();
634 * entry = mb_cache_entry_find_next(entry, ...);
637 * @prev: The previous match
638 * @index: the number of the additonal index to search (0<=index<indexes_count)
639 * @bdev: the device the cache entry should belong to
640 * @key: the key in the index
642 struct mb_cache_entry *
643 mb_cache_entry_find_next(struct mb_cache_entry *prev, int index,
644 struct block_device *bdev, unsigned int key)
646 struct mb_cache *cache = prev->e_cache;
647 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
649 struct mb_cache_entry *ce;
651 mb_assert(index < mb_cache_indexes(cache));
652 spin_lock(&mb_cache_spinlock);
653 l = prev->e_indexes[index].o_list.next;
654 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
656 __mb_cache_entry_release_unlock(prev);
660 #endif /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
662 static int __init init_mbcache(void)
664 mb_shrinker = set_shrinker(DEFAULT_SEEKS, mb_cache_shrink_fn);
668 static void __exit exit_mbcache(void)
670 remove_shrinker(mb_shrinker);
673 module_init(init_mbcache)
674 module_exit(exit_mbcache)