blob: f9e4d2700cd8148d89e9975fda791377e859a98e [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
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
60DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
61
62MODULE_AUTHOR("Andreas Gruenbacher <a.gruenbacher@computer.org>");
63MODULE_DESCRIPTION("Meta block cache (for extended attributes)");
64MODULE_LICENSE("GPL");
65
66EXPORT_SYMBOL(mb_cache_create);
67EXPORT_SYMBOL(mb_cache_shrink);
68EXPORT_SYMBOL(mb_cache_destroy);
69EXPORT_SYMBOL(mb_cache_entry_alloc);
70EXPORT_SYMBOL(mb_cache_entry_insert);
71EXPORT_SYMBOL(mb_cache_entry_release);
72EXPORT_SYMBOL(mb_cache_entry_free);
73EXPORT_SYMBOL(mb_cache_entry_get);
74#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
75EXPORT_SYMBOL(mb_cache_entry_find_first);
76EXPORT_SYMBOL(mb_cache_entry_find_next);
77#endif
78
79struct mb_cache {
80 struct list_head c_cache_list;
81 const char *c_name;
82 struct mb_cache_op c_op;
83 atomic_t c_entry_count;
84 int c_bucket_bits;
85#ifndef MB_CACHE_INDEXES_COUNT
86 int c_indexes_count;
87#endif
88 kmem_cache_t *c_entry_cache;
89 struct list_head *c_block_hash;
90 struct list_head *c_indexes_hash[0];
91};
92
93
94/*
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.
98 */
99
100static LIST_HEAD(mb_cache_list);
101static LIST_HEAD(mb_cache_lru_list);
102static DEFINE_SPINLOCK(mb_cache_spinlock);
103static struct shrinker *mb_shrinker;
104
105static inline int
106mb_cache_indexes(struct mb_cache *cache)
107{
108#ifdef MB_CACHE_INDEXES_COUNT
109 return MB_CACHE_INDEXES_COUNT;
110#else
111 return cache->c_indexes_count;
112#endif
113}
114
115/*
116 * What the mbcache registers as to get shrunk dynamically.
117 */
118
119static int mb_cache_shrink_fn(int nr_to_scan, unsigned int gfp_mask);
120
121
122static inline int
123__mb_cache_entry_is_hashed(struct mb_cache_entry *ce)
124{
125 return !list_empty(&ce->e_block_list);
126}
127
128
129static inline void
130__mb_cache_entry_unhash(struct mb_cache_entry *ce)
131{
132 int n;
133
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);
138 }
139}
140
141
142static inline void
143__mb_cache_entry_forget(struct mb_cache_entry *ce, int gfp_mask)
144{
145 struct mb_cache *cache = ce->e_cache;
146
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);
154 } else {
155 kmem_cache_free(cache->c_entry_cache, ce);
156 atomic_dec(&cache->c_entry_count);
157 }
158}
159
160
161static inline void
162__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
163{
164 /* Wake up all processes queuing for this cache entry. */
165 if (ce->e_queued)
166 wake_up_all(&mb_cache_queue);
167 if (ce->e_used >= MB_CACHE_WRITER)
168 ce->e_used -= MB_CACHE_WRITER;
169 ce->e_used--;
170 if (!(ce->e_used || ce->e_queued)) {
171 if (!__mb_cache_entry_is_hashed(ce))
172 goto forget;
173 mb_assert(list_empty(&ce->e_lru_list));
174 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
175 }
176 spin_unlock(&mb_cache_spinlock);
177 return;
178forget:
179 spin_unlock(&mb_cache_spinlock);
180 __mb_cache_entry_forget(ce, GFP_KERNEL);
181}
182
183
184/*
185 * mb_cache_shrink_fn() memory pressure callback
186 *
187 * This function is called by the kernel memory management when memory
188 * gets low.
189 *
190 * @nr_to_scan: Number of objects to scan
191 * @gfp_mask: (ignored)
192 *
193 * Returns the number of objects which are present in the cache.
194 */
195static int
196mb_cache_shrink_fn(int nr_to_scan, unsigned int gfp_mask)
197{
198 LIST_HEAD(free_list);
199 struct list_head *l, *ltmp;
200 int count = 0;
201
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);
209 }
210 mb_debug("trying to free %d entries", nr_to_scan);
211 if (nr_to_scan == 0) {
212 spin_unlock(&mb_cache_spinlock);
213 goto out;
214 }
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);
221 }
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);
226 }
227out:
228 return (count / 100) * sysctl_vfs_cache_pressure;
229}
230
231
232/*
233 * mb_cache_create() create a new cache
234 *
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.
239 *
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
246 * hardwired.
247 * @bucket_bits: log2(number of hash buckets)
248 */
249struct mb_cache *
250mb_cache_create(const char *name, struct mb_cache_op *cache_op,
251 size_t entry_size, int indexes_count, int bucket_bits)
252{
253 int m=0, n, bucket_count = 1 << bucket_bits;
254 struct mb_cache *cache = NULL;
255
256 if(entry_size < sizeof(struct mb_cache_entry) +
257 indexes_count * sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]))
258 return NULL;
259
260 cache = kmalloc(sizeof(struct mb_cache) +
261 indexes_count * sizeof(struct list_head), GFP_KERNEL);
262 if (!cache)
263 goto fail;
264 cache->c_name = name;
265 cache->c_op.free = NULL;
266 if (cache_op)
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);
272#else
273 cache->c_indexes_count = indexes_count;
274#endif
275 cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
276 GFP_KERNEL);
277 if (!cache->c_block_hash)
278 goto fail;
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),
284 GFP_KERNEL);
285 if (!cache->c_indexes_hash[m])
286 goto fail;
287 for (n=0; n<bucket_count; n++)
288 INIT_LIST_HEAD(&cache->c_indexes_hash[m][n]);
289 }
290 cache->c_entry_cache = kmem_cache_create(name, entry_size, 0,
291 SLAB_RECLAIM_ACCOUNT, NULL, NULL);
292 if (!cache->c_entry_cache)
293 goto fail;
294
295 spin_lock(&mb_cache_spinlock);
296 list_add(&cache->c_cache_list, &mb_cache_list);
297 spin_unlock(&mb_cache_spinlock);
298 return cache;
299
300fail:
301 if (cache) {
302 while (--m >= 0)
303 kfree(cache->c_indexes_hash[m]);
304 if (cache->c_block_hash)
305 kfree(cache->c_block_hash);
306 kfree(cache);
307 }
308 return NULL;
309}
310
311
312/*
313 * mb_cache_shrink()
314 *
315 * Removes all cache entires of a device from the cache. All cache entries
316 * currently in use cannot be freed, and thus remain in the cache. All others
317 * are freed.
318 *
319 * @cache: which cache to shrink
320 * @bdev: which device's cache entries to shrink
321 */
322void
323mb_cache_shrink(struct mb_cache *cache, struct block_device *bdev)
324{
325 LIST_HEAD(free_list);
326 struct list_head *l, *ltmp;
327
328 spin_lock(&mb_cache_spinlock);
329 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
330 struct mb_cache_entry *ce =
331 list_entry(l, struct mb_cache_entry, e_lru_list);
332 if (ce->e_bdev == bdev) {
333 list_move_tail(&ce->e_lru_list, &free_list);
334 __mb_cache_entry_unhash(ce);
335 }
336 }
337 spin_unlock(&mb_cache_spinlock);
338 list_for_each_safe(l, ltmp, &free_list) {
339 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
340 e_lru_list), GFP_KERNEL);
341 }
342}
343
344
345/*
346 * mb_cache_destroy()
347 *
348 * Shrinks the cache to its minimum possible size (hopefully 0 entries),
349 * and then destroys it. If this was the last mbcache, un-registers the
350 * mbcache from kernel memory management.
351 */
352void
353mb_cache_destroy(struct mb_cache *cache)
354{
355 LIST_HEAD(free_list);
356 struct list_head *l, *ltmp;
357 int n;
358
359 spin_lock(&mb_cache_spinlock);
360 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
361 struct mb_cache_entry *ce =
362 list_entry(l, struct mb_cache_entry, e_lru_list);
363 if (ce->e_cache == cache) {
364 list_move_tail(&ce->e_lru_list, &free_list);
365 __mb_cache_entry_unhash(ce);
366 }
367 }
368 list_del(&cache->c_cache_list);
369 spin_unlock(&mb_cache_spinlock);
370
371 list_for_each_safe(l, ltmp, &free_list) {
372 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
373 e_lru_list), GFP_KERNEL);
374 }
375
376 if (atomic_read(&cache->c_entry_count) > 0) {
377 mb_error("cache %s: %d orphaned entries",
378 cache->c_name,
379 atomic_read(&cache->c_entry_count));
380 }
381
382 kmem_cache_destroy(cache->c_entry_cache);
383
384 for (n=0; n < mb_cache_indexes(cache); n++)
385 kfree(cache->c_indexes_hash[n]);
386 kfree(cache->c_block_hash);
387 kfree(cache);
388}
389
390
391/*
392 * mb_cache_entry_alloc()
393 *
394 * Allocates a new cache entry. The new entry will not be valid initially,
395 * and thus cannot be looked up yet. It should be filled with data, and
396 * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
397 * if no more memory was available.
398 */
399struct mb_cache_entry *
400mb_cache_entry_alloc(struct mb_cache *cache)
401{
402 struct mb_cache_entry *ce;
403
404 atomic_inc(&cache->c_entry_count);
405 ce = kmem_cache_alloc(cache->c_entry_cache, GFP_KERNEL);
406 if (ce) {
407 INIT_LIST_HEAD(&ce->e_lru_list);
408 INIT_LIST_HEAD(&ce->e_block_list);
409 ce->e_cache = cache;
410 ce->e_used = 1 + MB_CACHE_WRITER;
411 ce->e_queued = 0;
412 }
413 return ce;
414}
415
416
417/*
418 * mb_cache_entry_insert()
419 *
420 * Inserts an entry that was allocated using mb_cache_entry_alloc() into
421 * the cache. After this, the cache entry can be looked up, but is not yet
422 * in the lru list as the caller still holds a handle to it. Returns 0 on
423 * success, or -EBUSY if a cache entry for that device + inode exists
424 * already (this may happen after a failed lookup, but when another process
425 * has inserted the same cache entry in the meantime).
426 *
427 * @bdev: device the cache entry belongs to
428 * @block: block number
429 * @keys: array of additional keys. There must be indexes_count entries
430 * in the array (as specified when creating the cache).
431 */
432int
433mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
434 sector_t block, unsigned int keys[])
435{
436 struct mb_cache *cache = ce->e_cache;
437 unsigned int bucket;
438 struct list_head *l;
439 int error = -EBUSY, n;
440
441 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
442 cache->c_bucket_bits);
443 spin_lock(&mb_cache_spinlock);
444 list_for_each_prev(l, &cache->c_block_hash[bucket]) {
445 struct mb_cache_entry *ce =
446 list_entry(l, struct mb_cache_entry, e_block_list);
447 if (ce->e_bdev == bdev && ce->e_block == block)
448 goto out;
449 }
450 __mb_cache_entry_unhash(ce);
451 ce->e_bdev = bdev;
452 ce->e_block = block;
453 list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
454 for (n=0; n<mb_cache_indexes(cache); n++) {
455 ce->e_indexes[n].o_key = keys[n];
456 bucket = hash_long(keys[n], cache->c_bucket_bits);
457 list_add(&ce->e_indexes[n].o_list,
458 &cache->c_indexes_hash[n][bucket]);
459 }
460 error = 0;
461out:
462 spin_unlock(&mb_cache_spinlock);
463 return error;
464}
465
466
467/*
468 * mb_cache_entry_release()
469 *
470 * Release a handle to a cache entry. When the last handle to a cache entry
471 * is released it is either freed (if it is invalid) or otherwise inserted
472 * in to the lru list.
473 */
474void
475mb_cache_entry_release(struct mb_cache_entry *ce)
476{
477 spin_lock(&mb_cache_spinlock);
478 __mb_cache_entry_release_unlock(ce);
479}
480
481
482/*
483 * mb_cache_entry_free()
484 *
485 * This is equivalent to the sequence mb_cache_entry_takeout() --
486 * mb_cache_entry_release().
487 */
488void
489mb_cache_entry_free(struct mb_cache_entry *ce)
490{
491 spin_lock(&mb_cache_spinlock);
492 mb_assert(list_empty(&ce->e_lru_list));
493 __mb_cache_entry_unhash(ce);
494 __mb_cache_entry_release_unlock(ce);
495}
496
497
498/*
499 * mb_cache_entry_get()
500 *
501 * Get a cache entry by device / block number. (There can only be one entry
502 * in the cache per device and block.) Returns NULL if no such cache entry
503 * exists. The returned cache entry is locked for exclusive access ("single
504 * writer").
505 */
506struct mb_cache_entry *
507mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
508 sector_t block)
509{
510 unsigned int bucket;
511 struct list_head *l;
512 struct mb_cache_entry *ce;
513
514 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
515 cache->c_bucket_bits);
516 spin_lock(&mb_cache_spinlock);
517 list_for_each(l, &cache->c_block_hash[bucket]) {
518 ce = list_entry(l, struct mb_cache_entry, e_block_list);
519 if (ce->e_bdev == bdev && ce->e_block == block) {
520 DEFINE_WAIT(wait);
521
522 if (!list_empty(&ce->e_lru_list))
523 list_del_init(&ce->e_lru_list);
524
525 while (ce->e_used > 0) {
526 ce->e_queued++;
527 prepare_to_wait(&mb_cache_queue, &wait,
528 TASK_UNINTERRUPTIBLE);
529 spin_unlock(&mb_cache_spinlock);
530 schedule();
531 spin_lock(&mb_cache_spinlock);
532 ce->e_queued--;
533 }
534 finish_wait(&mb_cache_queue, &wait);
535 ce->e_used += 1 + MB_CACHE_WRITER;
536
537 if (!__mb_cache_entry_is_hashed(ce)) {
538 __mb_cache_entry_release_unlock(ce);
539 return NULL;
540 }
541 goto cleanup;
542 }
543 }
544 ce = NULL;
545
546cleanup:
547 spin_unlock(&mb_cache_spinlock);
548 return ce;
549}
550
551#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
552
553static struct mb_cache_entry *
554__mb_cache_entry_find(struct list_head *l, struct list_head *head,
555 int index, struct block_device *bdev, unsigned int key)
556{
557 while (l != head) {
558 struct mb_cache_entry *ce =
559 list_entry(l, struct mb_cache_entry,
560 e_indexes[index].o_list);
561 if (ce->e_bdev == bdev && ce->e_indexes[index].o_key == key) {
562 DEFINE_WAIT(wait);
563
564 if (!list_empty(&ce->e_lru_list))
565 list_del_init(&ce->e_lru_list);
566
567 /* Incrementing before holding the lock gives readers
568 priority over writers. */
569 ce->e_used++;
570 while (ce->e_used >= MB_CACHE_WRITER) {
571 ce->e_queued++;
572 prepare_to_wait(&mb_cache_queue, &wait,
573 TASK_UNINTERRUPTIBLE);
574 spin_unlock(&mb_cache_spinlock);
575 schedule();
576 spin_lock(&mb_cache_spinlock);
577 ce->e_queued--;
578 }
579 finish_wait(&mb_cache_queue, &wait);
580
581 if (!__mb_cache_entry_is_hashed(ce)) {
582 __mb_cache_entry_release_unlock(ce);
583 spin_lock(&mb_cache_spinlock);
584 return ERR_PTR(-EAGAIN);
585 }
586 return ce;
587 }
588 l = l->next;
589 }
590 return NULL;
591}
592
593
594/*
595 * mb_cache_entry_find_first()
596 *
597 * Find the first cache entry on a given device with a certain key in
598 * an additional index. Additonal matches can be found with
599 * mb_cache_entry_find_next(). Returns NULL if no match was found. The
600 * returned cache entry is locked for shared access ("multiple readers").
601 *
602 * @cache: the cache to search
603 * @index: the number of the additonal index to search (0<=index<indexes_count)
604 * @bdev: the device the cache entry should belong to
605 * @key: the key in the index
606 */
607struct mb_cache_entry *
608mb_cache_entry_find_first(struct mb_cache *cache, int index,
609 struct block_device *bdev, unsigned int key)
610{
611 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
612 struct list_head *l;
613 struct mb_cache_entry *ce;
614
615 mb_assert(index < mb_cache_indexes(cache));
616 spin_lock(&mb_cache_spinlock);
617 l = cache->c_indexes_hash[index][bucket].next;
618 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
619 index, bdev, key);
620 spin_unlock(&mb_cache_spinlock);
621 return ce;
622}
623
624
625/*
626 * mb_cache_entry_find_next()
627 *
628 * Find the next cache entry on a given device with a certain key in an
629 * additional index. Returns NULL if no match could be found. The previous
630 * entry is atomatically released, so that mb_cache_entry_find_next() can
631 * be called like this:
632 *
633 * entry = mb_cache_entry_find_first();
634 * while (entry) {
635 * ...
636 * entry = mb_cache_entry_find_next(entry, ...);
637 * }
638 *
639 * @prev: The previous match
640 * @index: the number of the additonal index to search (0<=index<indexes_count)
641 * @bdev: the device the cache entry should belong to
642 * @key: the key in the index
643 */
644struct mb_cache_entry *
645mb_cache_entry_find_next(struct mb_cache_entry *prev, int index,
646 struct block_device *bdev, unsigned int key)
647{
648 struct mb_cache *cache = prev->e_cache;
649 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
650 struct list_head *l;
651 struct mb_cache_entry *ce;
652
653 mb_assert(index < mb_cache_indexes(cache));
654 spin_lock(&mb_cache_spinlock);
655 l = prev->e_indexes[index].o_list.next;
656 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
657 index, bdev, key);
658 __mb_cache_entry_release_unlock(prev);
659 return ce;
660}
661
662#endif /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
663
664static int __init init_mbcache(void)
665{
666 mb_shrinker = set_shrinker(DEFAULT_SEEKS, mb_cache_shrink_fn);
667 return 0;
668}
669
670static void __exit exit_mbcache(void)
671{
672 remove_shrinker(mb_shrinker);
673}
674
675module_init(init_mbcache)
676module_exit(exit_mbcache)
677