[linux.git] / fs / mbcache.c

/*
 * linux/fs/mbcache.c
 * (C) 2001-2002 Andreas Gruenbacher, <[email protected]>
 */

/*
 * Filesystem Meta Information Block Cache (mbcache)
 *
 * The mbcache caches blocks of block devices that need to be located
 * by their device/block number, as well as by other criteria (such
 * as the block's contents).
 *
 * There can only be one cache entry in a cache per device and block number.
 * Additional indexes need not be unique in this sense. The number of
 * additional indexes (=other criteria) can be hardwired at compile time
 * or specified at cache create time.
 *
 * Each cache entry is of fixed size. An entry may be `valid' or `invalid'
 * in the cache. A valid entry is in the main hash tables of the cache,
 * and may also be in the lru list. An invalid entry is not in any hashes
 * or lists.
 *
 * A valid cache entry is only in the lru list if no handles refer to it.
 * Invalid cache entries will be freed when the last handle to the cache
 * entry is released. Entries that cannot be freed immediately are put
 * back on the lru list.
 */

#include <linux/kernel.h>
#include <linux/module.h>

#include <linux/hash.h>
#include <linux/fs.h>
#include <linux/mm.h>
#include <linux/slab.h>
#include <linux/sched.h>
#include <linux/init.h>
#include <linux/mbcache.h>


#ifdef MB_CACHE_DEBUG
# define mb_debug(f...) do { \
		printk(KERN_DEBUG f); \
		printk("\n"); \
	} while (0)
#define mb_assert(c) do { if (!(c)) \
		printk(KERN_ERR "assertion " #c " failed\n"); \
	} while(0)
#else
# define mb_debug(f...) do { } while(0)
# define mb_assert(c) do { } while(0)
#endif
#define mb_error(f...) do { \
		printk(KERN_ERR f); \
		printk("\n"); \
	} while(0)

#define MB_CACHE_WRITER ((unsigned short)~0U >> 1)

static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
		
MODULE_AUTHOR("Andreas Gruenbacher <[email protected]>");
MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
MODULE_LICENSE("GPL");

EXPORT_SYMBOL(mb_cache_create);
EXPORT_SYMBOL(mb_cache_shrink);
EXPORT_SYMBOL(mb_cache_destroy);
EXPORT_SYMBOL(mb_cache_entry_alloc);
EXPORT_SYMBOL(mb_cache_entry_insert);
EXPORT_SYMBOL(mb_cache_entry_release);
EXPORT_SYMBOL(mb_cache_entry_free);
EXPORT_SYMBOL(mb_cache_entry_get);
#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
EXPORT_SYMBOL(mb_cache_entry_find_first);
EXPORT_SYMBOL(mb_cache_entry_find_next);
#endif

/*
 * Global data: list of all mbcache's, lru list, and a spinlock for
 * accessing cache data structures on SMP machines. The lru list is
 * global across all mbcaches.
 */

static LIST_HEAD(mb_cache_list);
static LIST_HEAD(mb_cache_lru_list);
static DEFINE_SPINLOCK(mb_cache_spinlock);

/*
 * What the mbcache registers as to get shrunk dynamically.
 */

static int mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask);

static struct shrinker mb_cache_shrinker = {
	.shrink = mb_cache_shrink_fn,
	.seeks = DEFAULT_SEEKS,
};

static inline int
__mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
{
	return !list_empty(&ce->e_block_list);
}


static void
__mb_cache_entry_unhash(struct mb_cache_entry *ce)
{
	if (__mb_cache_entry_is_hashed(ce)) {
		list_del_init(&ce->e_block_list);
		list_del(&ce->e_index.o_list);
	}
}


static void
__mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
{
	struct mb_cache *cache = ce->e_cache;

	mb_assert(!(ce->e_used || ce->e_queued));
	kmem_cache_free(cache->c_entry_cache, ce);
	atomic_dec(&cache->c_entry_count);
}


static void
__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
	__releases(mb_cache_spinlock)
{
	/* Wake up all processes queuing for this cache entry. */
	if (ce->e_queued)
		wake_up_all(&mb_cache_queue);
	if (ce->e_used >= MB_CACHE_WRITER)
		ce->e_used -= MB_CACHE_WRITER;
	ce->e_used--;
	if (!(ce->e_used || ce->e_queued)) {
		if (!__mb_cache_entry_is_hashed(ce))
			goto forget;
		mb_assert(list_empty(&ce->e_lru_list));
		list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
	}
	spin_unlock(&mb_cache_spinlock);
	return;
forget:
	spin_unlock(&mb_cache_spinlock);
	__mb_cache_entry_forget(ce, GFP_KERNEL);
}


/*
 * mb_cache_shrink_fn()  memory pressure callback
 *
 * This function is called by the kernel memory management when memory
 * gets low.
 *
 * @shrink: (ignored)
 * @nr_to_scan: Number of objects to scan
 * @gfp_mask: (ignored)
 *
 * Returns the number of objects which are present in the cache.
 */
static int
mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
{
	LIST_HEAD(free_list);
	struct mb_cache *cache;
	struct mb_cache_entry *entry, *tmp;
	int count = 0;

	mb_debug("trying to free %d entries", nr_to_scan);
	spin_lock(&mb_cache_spinlock);
	while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
		struct mb_cache_entry *ce =
			list_entry(mb_cache_lru_list.next,
				   struct mb_cache_entry, e_lru_list);
		list_move_tail(&ce->e_lru_list, &free_list);
		__mb_cache_entry_unhash(ce);
	}
	list_for_each_entry(cache, &mb_cache_list, c_cache_list) {
		mb_debug("cache %s (%d)", cache->c_name,
			  atomic_read(&cache->c_entry_count));
		count += atomic_read(&cache->c_entry_count);
	}
	spin_unlock(&mb_cache_spinlock);
	list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
		__mb_cache_entry_forget(entry, gfp_mask);
	}
	return (count / 100) * sysctl_vfs_cache_pressure;
}


/*
 * mb_cache_create()  create a new cache
 *
 * All entries in one cache are equal size. Cache entries may be from
 * multiple devices. If this is the first mbcache created, registers
 * the cache with kernel memory management. Returns NULL if no more
 * memory was available.
 *
 * @name: name of the cache (informal)
 * @bucket_bits: log2(number of hash buckets)
 */
struct mb_cache *
mb_cache_create(const char *name, int bucket_bits)
{
	int n, bucket_count = 1 << bucket_bits;
	struct mb_cache *cache = NULL;

	cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
	if (!cache)
		return NULL;
	cache->c_name = name;
	atomic_set(&cache->c_entry_count, 0);
	cache->c_bucket_bits = bucket_bits;
	cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
	                              GFP_KERNEL);
	if (!cache->c_block_hash)
		goto fail;
	for (n=0; n<bucket_count; n++)
		INIT_LIST_HEAD(&cache->c_block_hash[n]);
	cache->c_index_hash = kmalloc(bucket_count * sizeof(struct list_head),
				      GFP_KERNEL);
	if (!cache->c_index_hash)
		goto fail;
	for (n=0; n<bucket_count; n++)
		INIT_LIST_HEAD(&cache->c_index_hash[n]);
	cache->c_entry_cache = kmem_cache_create(name,
		sizeof(struct mb_cache_entry), 0,
		SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL);
	if (!cache->c_entry_cache)
		goto fail2;

	/*
	 * Set an upper limit on the number of cache entries so that the hash
	 * chains won't grow too long.
	 */
	cache->c_max_entries = bucket_count << 4;

	spin_lock(&mb_cache_spinlock);
	list_add(&cache->c_cache_list, &mb_cache_list);
	spin_unlock(&mb_cache_spinlock);
	return cache;

fail2:
	kfree(cache->c_index_hash);

fail:
	kfree(cache->c_block_hash);
	kfree(cache);
	return NULL;
}


/*
 * mb_cache_shrink()
 *
 * Removes all cache entries of a device from the cache. All cache entries
 * currently in use cannot be freed, and thus remain in the cache. All others
 * are freed.
 *
 * @bdev: which device's cache entries to shrink
 */
void
mb_cache_shrink(struct block_device *bdev)
{
	LIST_HEAD(free_list);
	struct list_head *l, *ltmp;

	spin_lock(&mb_cache_spinlock);
	list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
		struct mb_cache_entry *ce =
			list_entry(l, struct mb_cache_entry, e_lru_list);
		if (ce->e_bdev == bdev) {
			list_move_tail(&ce->e_lru_list, &free_list);
			__mb_cache_entry_unhash(ce);
		}
	}
	spin_unlock(&mb_cache_spinlock);
	list_for_each_safe(l, ltmp, &free_list) {
		__mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
						   e_lru_list), GFP_KERNEL);
	}
}


/*
 * mb_cache_destroy()
 *
 * Shrinks the cache to its minimum possible size (hopefully 0 entries),
 * and then destroys it. If this was the last mbcache, un-registers the
 * mbcache from kernel memory management.
 */
void
mb_cache_destroy(struct mb_cache *cache)
{
	LIST_HEAD(free_list);
	struct list_head *l, *ltmp;

	spin_lock(&mb_cache_spinlock);
	list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
		struct mb_cache_entry *ce =
			list_entry(l, struct mb_cache_entry, e_lru_list);
		if (ce->e_cache == cache) {
			list_move_tail(&ce->e_lru_list, &free_list);
			__mb_cache_entry_unhash(ce);
		}
	}
	list_del(&cache->c_cache_list);
	spin_unlock(&mb_cache_spinlock);

	list_for_each_safe(l, ltmp, &free_list) {
		__mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
						   e_lru_list), GFP_KERNEL);
	}

	if (atomic_read(&cache->c_entry_count) > 0) {
		mb_error("cache %s: %d orphaned entries",
			  cache->c_name,
			  atomic_read(&cache->c_entry_count));
	}

	kmem_cache_destroy(cache->c_entry_cache);

	kfree(cache->c_index_hash);
	kfree(cache->c_block_hash);
	kfree(cache);
}

/*
 * mb_cache_entry_alloc()
 *
 * Allocates a new cache entry. The new entry will not be valid initially,
 * and thus cannot be looked up yet. It should be filled with data, and
 * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
 * if no more memory was available.
 */
struct mb_cache_entry *
mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
{
	struct mb_cache_entry *ce = NULL;

	if (atomic_read(&cache->c_entry_count) >= cache->c_max_entries) {
		spin_lock(&mb_cache_spinlock);
		if (!list_empty(&mb_cache_lru_list)) {
			ce = list_entry(mb_cache_lru_list.next,
					struct mb_cache_entry, e_lru_list);
			list_del_init(&ce->e_lru_list);
			__mb_cache_entry_unhash(ce);
		}
		spin_unlock(&mb_cache_spinlock);
	}
	if (!ce) {
		ce = kmem_cache_alloc(cache->c_entry_cache, gfp_flags);
		if (!ce)
			return NULL;
		atomic_inc(&cache->c_entry_count);
		INIT_LIST_HEAD(&ce->e_lru_list);
		INIT_LIST_HEAD(&ce->e_block_list);
		ce->e_cache = cache;
		ce->e_queued = 0;
	}
	ce->e_used = 1 + MB_CACHE_WRITER;
	return ce;
}


/*
 * mb_cache_entry_insert()
 *
 * Inserts an entry that was allocated using mb_cache_entry_alloc() into
 * the cache. After this, the cache entry can be looked up, but is not yet
 * in the lru list as the caller still holds a handle to it. Returns 0 on
 * success, or -EBUSY if a cache entry for that device + inode exists
 * already (this may happen after a failed lookup, but when another process
 * has inserted the same cache entry in the meantime).
 *
 * @bdev: device the cache entry belongs to
 * @block: block number
 * @key: lookup key
 */
int
mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
		      sector_t block, unsigned int key)
{
	struct mb_cache *cache = ce->e_cache;
	unsigned int bucket;
	struct list_head *l;
	int error = -EBUSY;

	bucket = hash_long((unsigned long)bdev + (block & 0xffffffff), 
			   cache->c_bucket_bits);
	spin_lock(&mb_cache_spinlock);
	list_for_each_prev(l, &cache->c_block_hash[bucket]) {
		struct mb_cache_entry *ce =
			list_entry(l, struct mb_cache_entry, e_block_list);
		if (ce->e_bdev == bdev && ce->e_block == block)
			goto out;
	}
	__mb_cache_entry_unhash(ce);
	ce->e_bdev = bdev;
	ce->e_block = block;
	list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
	ce->e_index.o_key = key;
	bucket = hash_long(key, cache->c_bucket_bits);
	list_add(&ce->e_index.o_list, &cache->c_index_hash[bucket]);
	error = 0;
out:
	spin_unlock(&mb_cache_spinlock);
	return error;
}


/*
 * mb_cache_entry_release()
 *
 * Release a handle to a cache entry. When the last handle to a cache entry
 * is released it is either freed (if it is invalid) or otherwise inserted
 * in to the lru list.
 */
void
mb_cache_entry_release(struct mb_cache_entry *ce)
{
	spin_lock(&mb_cache_spinlock);
	__mb_cache_entry_release_unlock(ce);
}


/*
 * mb_cache_entry_free()
 *
 * This is equivalent to the sequence mb_cache_entry_takeout() --
 * mb_cache_entry_release().
 */
void
mb_cache_entry_free(struct mb_cache_entry *ce)
{
	spin_lock(&mb_cache_spinlock);
	mb_assert(list_empty(&ce->e_lru_list));
	__mb_cache_entry_unhash(ce);
	__mb_cache_entry_release_unlock(ce);
}


/*
 * mb_cache_entry_get()
 *
 * Get a cache entry  by device / block number. (There can only be one entry
 * in the cache per device and block.) Returns NULL if no such cache entry
 * exists. The returned cache entry is locked for exclusive access ("single
 * writer").
 */
struct mb_cache_entry *
mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
		   sector_t block)
{
	unsigned int bucket;
	struct list_head *l;
	struct mb_cache_entry *ce;

	bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
			   cache->c_bucket_bits);
	spin_lock(&mb_cache_spinlock);
	list_for_each(l, &cache->c_block_hash[bucket]) {
		ce = list_entry(l, struct mb_cache_entry, e_block_list);
		if (ce->e_bdev == bdev && ce->e_block == block) {
			DEFINE_WAIT(wait);

			if (!list_empty(&ce->e_lru_list))
				list_del_init(&ce->e_lru_list);

			while (ce->e_used > 0) {
				ce->e_queued++;
				prepare_to_wait(&mb_cache_queue, &wait,
						TASK_UNINTERRUPTIBLE);
				spin_unlock(&mb_cache_spinlock);
				schedule();
				spin_lock(&mb_cache_spinlock);
				ce->e_queued--;
			}
			finish_wait(&mb_cache_queue, &wait);
			ce->e_used += 1 + MB_CACHE_WRITER;

			if (!__mb_cache_entry_is_hashed(ce)) {
				__mb_cache_entry_release_unlock(ce);
				return NULL;
			}
			goto cleanup;
		}
	}
	ce = NULL;

cleanup:
	spin_unlock(&mb_cache_spinlock);
	return ce;
}

#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)

static struct mb_cache_entry *
__mb_cache_entry_find(struct list_head *l, struct list_head *head,
		      struct block_device *bdev, unsigned int key)
{
	while (l != head) {
		struct mb_cache_entry *ce =
			list_entry(l, struct mb_cache_entry, e_index.o_list);
		if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
			DEFINE_WAIT(wait);

			if (!list_empty(&ce->e_lru_list))
				list_del_init(&ce->e_lru_list);

			/* Incrementing before holding the lock gives readers
			   priority over writers. */
			ce->e_used++;
			while (ce->e_used >= MB_CACHE_WRITER) {
				ce->e_queued++;
				prepare_to_wait(&mb_cache_queue, &wait,
						TASK_UNINTERRUPTIBLE);
				spin_unlock(&mb_cache_spinlock);
				schedule();
				spin_lock(&mb_cache_spinlock);
				ce->e_queued--;
			}
			finish_wait(&mb_cache_queue, &wait);

			if (!__mb_cache_entry_is_hashed(ce)) {
				__mb_cache_entry_release_unlock(ce);
				spin_lock(&mb_cache_spinlock);
				return ERR_PTR(-EAGAIN);
			}
			return ce;
		}
		l = l->next;
	}
	return NULL;
}


/*
 * mb_cache_entry_find_first()
 *
 * Find the first cache entry on a given device with a certain key in
 * an additional index. Additonal matches can be found with
 * mb_cache_entry_find_next(). Returns NULL if no match was found. The
 * returned cache entry is locked for shared access ("multiple readers").
 *
 * @cache: the cache to search
 * @bdev: the device the cache entry should belong to
 * @key: the key in the index
 */
struct mb_cache_entry *
mb_cache_entry_find_first(struct mb_cache *cache, struct block_device *bdev,
			  unsigned int key)
{
	unsigned int bucket = hash_long(key, cache->c_bucket_bits);
	struct list_head *l;
	struct mb_cache_entry *ce;

	spin_lock(&mb_cache_spinlock);
	l = cache->c_index_hash[bucket].next;
	ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
	spin_unlock(&mb_cache_spinlock);
	return ce;
}


/*
 * mb_cache_entry_find_next()
 *
 * Find the next cache entry on a given device with a certain key in an
 * additional index. Returns NULL if no match could be found. The previous
 * entry is atomatically released, so that mb_cache_entry_find_next() can
 * be called like this:
 *
 * entry = mb_cache_entry_find_first();
 * while (entry) {
 * 	...
 *	entry = mb_cache_entry_find_next(entry, ...);
 * }
 *
 * @prev: The previous match
 * @bdev: the device the cache entry should belong to
 * @key: the key in the index
 */
struct mb_cache_entry *
mb_cache_entry_find_next(struct mb_cache_entry *prev,
			 struct block_device *bdev, unsigned int key)
{
	struct mb_cache *cache = prev->e_cache;
	unsigned int bucket = hash_long(key, cache->c_bucket_bits);
	struct list_head *l;
	struct mb_cache_entry *ce;

	spin_lock(&mb_cache_spinlock);
	l = prev->e_index.o_list.next;
	ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
	__mb_cache_entry_release_unlock(prev);
	return ce;
}

#endif  /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */

static int __init init_mbcache(void)
{
	register_shrinker(&mb_cache_shrinker);
	return 0;
}

static void __exit exit_mbcache(void)
{
	unregister_shrinker(&mb_cache_shrinker);
}

module_init(init_mbcache)
module_exit(exit_mbcache)
Commit	Line	Data
1da177e4 LT	1	/*
	2	* linux/fs/mbcache.c
	3	* (C) 2001-2002 Andreas Gruenbacher, <[email protected]>
	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
75c96f85	60	static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
1da177e4 LT	61
	62	MODULE_AUTHOR("Andreas Gruenbacher <[email protected]>");
	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
1da177e4 LT	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);
1da177e4	88
1da177e4 LT	89	/*
	90	* What the mbcache registers as to get shrunk dynamically.
	91	*/
	92
7f8275d0	93	static int mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask);
1da177e4	94
8e1f936b RR	95	static struct shrinker mb_cache_shrinker = {
	96	.shrink = mb_cache_shrink_fn,
	97	.seeks = DEFAULT_SEEKS,
	98	};
1da177e4 LT	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
858119e1	107	static void
1da177e4 LT	108	__mb_cache_entry_unhash(struct mb_cache_entry *ce)
1da177e4 LT	109	{
1da177e4 LT	110	if (__mb_cache_entry_is_hashed(ce)) {
1da177e4 LT	111	list_del_init(&ce->e_block_list);
2aec7c52	112	list_del(&ce->e_index.o_list);
1da177e4 LT	113	}
	114	}
	115
	116
858119e1	117	static void
27496a8c	118	__mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t gfp_mask)
1da177e4 LT	119	{
	120	struct mb_cache *cache = ce->e_cache;
	121
	122	mb_assert(!(ce->e_used \|\| ce->e_queued));
2aec7c52 AG	123	kmem_cache_free(cache->c_entry_cache, ce);
2aec7c52 AG	124	atomic_dec(&cache->c_entry_count);
1da177e4 LT	125	}
	126
	127
858119e1	128	static void
1da177e4	129	__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
58f555e5	130	__releases(mb_cache_spinlock)
1da177e4 LT	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	*
7f8275d0	158	* @shrink: (ignored)
1da177e4 LT	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
7f8275d0	165	mb_cache_shrink_fn(struct shrinker *shrink, int nr_to_scan, gfp_t gfp_mask)
1da177e4 LT	166	{
1da177e4 LT	167	LIST_HEAD(free_list);
e566d48c AG	168	struct mb_cache *cache;
e566d48c AG	169	struct mb_cache_entry entry, tmp;
1da177e4 LT	170	int count = 0;
1da177e4 LT	171
1da177e4	172	mb_debug("trying to free %d entries", nr_to_scan);
e566d48c	173	spin_lock(&mb_cache_spinlock);
1da177e4 LT	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	}
e566d48c AG	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	}
1da177e4	186	spin_unlock(&mb_cache_spinlock);
e566d48c AG	187	list_for_each_entry_safe(entry, tmp, &free_list, e_lru_list) {
e566d48c AG	188	__mb_cache_entry_forget(entry, gfp_mask);
1da177e4	189	}
1da177e4 LT	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)
1da177e4 LT	203	* @bucket_bits: log2(number of hash buckets)
	204	*/
	205	struct mb_cache *
2aec7c52	206	mb_cache_create(const char *name, int bucket_bits)
1da177e4	207	{
2aec7c52	208	int n, bucket_count = 1 << bucket_bits;
1da177e4 LT	209	struct mb_cache *cache = NULL;
1da177e4 LT	210
2aec7c52	211	cache = kmalloc(sizeof(struct mb_cache), GFP_KERNEL);
1da177e4	212	if (!cache)
2aec7c52	213	return NULL;
1da177e4	214	cache->c_name = name;
1da177e4 LT	215	atomic_set(&cache->c_entry_count, 0);
1da177e4 LT	216	cache->c_bucket_bits = bucket_bits;
1da177e4 LT	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]);
2aec7c52 AG	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,
20c2df83	231	SLAB_RECLAIM_ACCOUNT\|SLAB_MEM_SPREAD, NULL);
1da177e4	232	if (!cache->c_entry_cache)
2aec7c52	233	goto fail2;
1da177e4	234
3a48ee8a AG	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
1da177e4 LT	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
2aec7c52 AG	246	fail2:
	247	kfree(cache->c_index_hash);
	248
1da177e4	249	fail:
2aec7c52 AG	250	kfree(cache->c_block_hash);
2aec7c52 AG	251	kfree(cache);
1da177e4 LT	252	return NULL;
	253	}
	254
	255
	256	/*
	257	* mb_cache_shrink()
	258	*
7f927fcc	259	* Removes all cache entries of a device from the cache. All cache entries
1da177e4 LT	260	* currently in use cannot be freed, and thus remain in the cache. All others
	261	* are freed.
	262	*
1da177e4 LT	263	* @bdev: which device's cache entries to shrink
	264	*/
	265	void
8c52ab42	266	mb_cache_shrink(struct block_device *bdev)
1da177e4 LT	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;
1da177e4 LT	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
2aec7c52	326	kfree(cache->c_index_hash);
1da177e4 LT	327	kfree(cache->c_block_hash);
	328	kfree(cache);
	329	}
	330
1da177e4 LT	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 *
335e92e8	340	mb_cache_entry_alloc(struct mb_cache *cache, gfp_t gfp_flags)
1da177e4	341	{
3a48ee8a AG	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;
f9e83489	358	atomic_inc(&cache->c_entry_count);
1da177e4 LT	359	INIT_LIST_HEAD(&ce->e_lru_list);
	360	INIT_LIST_HEAD(&ce->e_block_list);
	361	ce->e_cache = cache;
1da177e4 LT	362	ce->e_queued = 0;
1da177e4 LT	363	}
3a48ee8a	364	ce->e_used = 1 + MB_CACHE_WRITER;
1da177e4 LT	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
2aec7c52	381	* @key: lookup key
1da177e4 LT	382	*/
	383	int
	384	mb_cache_entry_insert(struct mb_cache_entry ce, struct block_device bdev,
2aec7c52	385	sector_t block, unsigned int key)
1da177e4 LT	386	{
	387	struct mb_cache *cache = ce->e_cache;
	388	unsigned int bucket;
	389	struct list_head *l;
2aec7c52	390	int error = -EBUSY;
1da177e4 LT	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]);
2aec7c52 AG	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]);
1da177e4 LT	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,
2aec7c52	503	struct block_device *bdev, unsigned int key)
1da177e4 LT	504	{
	505	while (l != head) {
	506	struct mb_cache_entry *ce =
2aec7c52 AG	507	list_entry(l, struct mb_cache_entry, e_index.o_list);
2aec7c52 AG	508	if (ce->e_bdev == bdev && ce->e_index.o_key == key) {
1da177e4 LT	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
1da177e4 LT	550	* @bdev: the device the cache entry should belong to
	551	* @key: the key in the index
	552	*/
	553	struct mb_cache_entry *
2aec7c52 AG	554	mb_cache_entry_find_first(struct mb_cache cache, struct block_device bdev,
2aec7c52 AG	555	unsigned int key)
1da177e4 LT	556	{
	557	unsigned int bucket = hash_long(key, cache->c_bucket_bits);
	558	struct list_head *l;
	559	struct mb_cache_entry *ce;
	560
1da177e4	561	spin_lock(&mb_cache_spinlock);
2aec7c52 AG	562	l = cache->c_index_hash[bucket].next;
2aec7c52 AG	563	ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
1da177e4 LT	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
1da177e4 LT	584	* @bdev: the device the cache entry should belong to
	585	* @key: the key in the index
	586	*/
	587	struct mb_cache_entry *
2aec7c52	588	mb_cache_entry_find_next(struct mb_cache_entry *prev,
1da177e4 LT	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
1da177e4	596	spin_lock(&mb_cache_spinlock);
2aec7c52 AG	597	l = prev->e_index.o_list.next;
2aec7c52 AG	598	ce = __mb_cache_entry_find(l, &cache->c_index_hash[bucket], bdev, key);
1da177e4 LT	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	{
8e1f936b	607	register_shrinker(&mb_cache_shrinker);
1da177e4 LT	608	return 0;
	609	}
	610
	611	static void __exit exit_mbcache(void)
	612	{
8e1f936b	613	unregister_shrinker(&mb_cache_shrinker);
1da177e4 LT	614	}
	615
	616	module_init(init_mbcache)
	617	module_exit(exit_mbcache)
	618