projects/linux·blob v2.6.21 · refs

676 lines · 18.2 KiB

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
60static DECLARE_WAIT_QUEUE_HEAD(mb_cache_queue);
61
62MODULE_AUTHOR("Andreas Gruenbacher <[email protected]>");
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 struct kmem_cache *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, gfp_t 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 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 void
143__mb_cache_entry_forget(struct mb_cache_entry *ce, gfp_t 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 void
162__mb_cache_entry_release_unlock(struct mb_cache_entry *ce)
163 __releases(mb_cache_spinlock)
164{
165 /* Wake up all processes queuing for this cache entry. */
166 if (ce->e_queued)
167 wake_up_all(&mb_cache_queue);
168 if (ce->e_used >= MB_CACHE_WRITER)
169 ce->e_used -= MB_CACHE_WRITER;
170 ce->e_used--;
171 if (!(ce->e_used || ce->e_queued)) {
172 if (!__mb_cache_entry_is_hashed(ce))
173 goto forget;
174 mb_assert(list_empty(&ce->e_lru_list));
175 list_add_tail(&ce->e_lru_list, &mb_cache_lru_list);
176 }
177 spin_unlock(&mb_cache_spinlock);
178 return;
179forget:
180 spin_unlock(&mb_cache_spinlock);
181 __mb_cache_entry_forget(ce, GFP_KERNEL);
182}
183
184
185/*
186 * mb_cache_shrink_fn() memory pressure callback
187 *
188 * This function is called by the kernel memory management when memory
189 * gets low.
190 *
191 * @nr_to_scan: Number of objects to scan
192 * @gfp_mask: (ignored)
193 *
194 * Returns the number of objects which are present in the cache.
195 */
196static int
197mb_cache_shrink_fn(int nr_to_scan, gfp_t gfp_mask)
198{
199 LIST_HEAD(free_list);
200 struct list_head *l, *ltmp;
201 int count = 0;
202
203 spin_lock(&mb_cache_spinlock);
204 list_for_each(l, &mb_cache_list) {
205 struct mb_cache *cache =
206 list_entry(l, struct mb_cache, c_cache_list);
207 mb_debug("cache %s (%d)", cache->c_name,
208 atomic_read(&cache->c_entry_count));
209 count += atomic_read(&cache->c_entry_count);
210 }
211 mb_debug("trying to free %d entries", nr_to_scan);
212 if (nr_to_scan == 0) {
213 spin_unlock(&mb_cache_spinlock);
214 goto out;
215 }
216 while (nr_to_scan-- && !list_empty(&mb_cache_lru_list)) {
217 struct mb_cache_entry *ce =
218 list_entry(mb_cache_lru_list.next,
219 struct mb_cache_entry, e_lru_list);
220 list_move_tail(&ce->e_lru_list, &free_list);
221 __mb_cache_entry_unhash(ce);
222 }
223 spin_unlock(&mb_cache_spinlock);
224 list_for_each_safe(l, ltmp, &free_list) {
225 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
226 e_lru_list), gfp_mask);
227 }
228out:
229 return (count / 100) * sysctl_vfs_cache_pressure;
230}
231
232
233/*
234 * mb_cache_create() create a new cache
235 *
236 * All entries in one cache are equal size. Cache entries may be from
237 * multiple devices. If this is the first mbcache created, registers
238 * the cache with kernel memory management. Returns NULL if no more
239 * memory was available.
240 *
241 * @name: name of the cache (informal)
242 * @cache_op: contains the callback called when freeing a cache entry
243 * @entry_size: The size of a cache entry, including
244 * struct mb_cache_entry
245 * @indexes_count: number of additional indexes in the cache. Must equal
246 * MB_CACHE_INDEXES_COUNT if the number of indexes is
247 * hardwired.
248 * @bucket_bits: log2(number of hash buckets)
249 */
250struct mb_cache *
251mb_cache_create(const char *name, struct mb_cache_op *cache_op,
252 size_t entry_size, int indexes_count, int bucket_bits)
253{
254 int m=0, n, bucket_count = 1 << bucket_bits;
255 struct mb_cache *cache = NULL;
256
257 if(entry_size < sizeof(struct mb_cache_entry) +
258 indexes_count * sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]))
259 return NULL;
260
261 cache = kmalloc(sizeof(struct mb_cache) +
262 indexes_count * sizeof(struct list_head), GFP_KERNEL);
263 if (!cache)
264 goto fail;
265 cache->c_name = name;
266 cache->c_op.free = NULL;
267 if (cache_op)
268 cache->c_op.free = cache_op->free;
269 atomic_set(&cache->c_entry_count, 0);
270 cache->c_bucket_bits = bucket_bits;
271#ifdef MB_CACHE_INDEXES_COUNT
272 mb_assert(indexes_count == MB_CACHE_INDEXES_COUNT);
273#else
274 cache->c_indexes_count = indexes_count;
275#endif
276 cache->c_block_hash = kmalloc(bucket_count * sizeof(struct list_head),
277 GFP_KERNEL);
278 if (!cache->c_block_hash)
279 goto fail;
280 for (n=0; n<bucket_count; n++)
281 INIT_LIST_HEAD(&cache->c_block_hash[n]);
282 for (m=0; m<indexes_count; m++) {
283 cache->c_indexes_hash[m] = kmalloc(bucket_count *
284 sizeof(struct list_head),
285 GFP_KERNEL);
286 if (!cache->c_indexes_hash[m])
287 goto fail;
288 for (n=0; n<bucket_count; n++)
289 INIT_LIST_HEAD(&cache->c_indexes_hash[m][n]);
290 }
291 cache->c_entry_cache = kmem_cache_create(name, entry_size, 0,
292 SLAB_RECLAIM_ACCOUNT|SLAB_MEM_SPREAD, NULL, NULL);
293 if (!cache->c_entry_cache)
294 goto fail;
295
296 spin_lock(&mb_cache_spinlock);
297 list_add(&cache->c_cache_list, &mb_cache_list);
298 spin_unlock(&mb_cache_spinlock);
299 return cache;
300
301fail:
302 if (cache) {
303 while (--m >= 0)
304 kfree(cache->c_indexes_hash[m]);
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 entries 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 * @bdev: which device's cache entries to shrink
320 */
321void
322mb_cache_shrink(struct block_device *bdev)
323{
324 LIST_HEAD(free_list);
325 struct list_head *l, *ltmp;
326
327 spin_lock(&mb_cache_spinlock);
328 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
329 struct mb_cache_entry *ce =
330 list_entry(l, struct mb_cache_entry, e_lru_list);
331 if (ce->e_bdev == bdev) {
332 list_move_tail(&ce->e_lru_list, &free_list);
333 __mb_cache_entry_unhash(ce);
334 }
335 }
336 spin_unlock(&mb_cache_spinlock);
337 list_for_each_safe(l, ltmp, &free_list) {
338 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
339 e_lru_list), GFP_KERNEL);
340 }
341}
342
343
344/*
345 * mb_cache_destroy()
346 *
347 * Shrinks the cache to its minimum possible size (hopefully 0 entries),
348 * and then destroys it. If this was the last mbcache, un-registers the
349 * mbcache from kernel memory management.
350 */
351void
352mb_cache_destroy(struct mb_cache *cache)
353{
354 LIST_HEAD(free_list);
355 struct list_head *l, *ltmp;
356 int n;
357
358 spin_lock(&mb_cache_spinlock);
359 list_for_each_safe(l, ltmp, &mb_cache_lru_list) {
360 struct mb_cache_entry *ce =
361 list_entry(l, struct mb_cache_entry, e_lru_list);
362 if (ce->e_cache == cache) {
363 list_move_tail(&ce->e_lru_list, &free_list);
364 __mb_cache_entry_unhash(ce);
365 }
366 }
367 list_del(&cache->c_cache_list);
368 spin_unlock(&mb_cache_spinlock);
369
370 list_for_each_safe(l, ltmp, &free_list) {
371 __mb_cache_entry_forget(list_entry(l, struct mb_cache_entry,
372 e_lru_list), GFP_KERNEL);
373 }
374
375 if (atomic_read(&cache->c_entry_count) > 0) {
376 mb_error("cache %s: %d orphaned entries",
377 cache->c_name,
378 atomic_read(&cache->c_entry_count));
379 }
380
381 kmem_cache_destroy(cache->c_entry_cache);
382
383 for (n=0; n < mb_cache_indexes(cache); n++)
384 kfree(cache->c_indexes_hash[n]);
385 kfree(cache->c_block_hash);
386 kfree(cache);
387}
388
389
390/*
391 * mb_cache_entry_alloc()
392 *
393 * Allocates a new cache entry. The new entry will not be valid initially,
394 * and thus cannot be looked up yet. It should be filled with data, and
395 * then inserted into the cache using mb_cache_entry_insert(). Returns NULL
396 * if no more memory was available.
397 */
398struct mb_cache_entry *
399mb_cache_entry_alloc(struct mb_cache *cache)
400{
401 struct mb_cache_entry *ce;
402
403 atomic_inc(&cache->c_entry_count);
404 ce = kmem_cache_alloc(cache->c_entry_cache, GFP_KERNEL);
405 if (ce) {
406 INIT_LIST_HEAD(&ce->e_lru_list);
407 INIT_LIST_HEAD(&ce->e_block_list);
408 ce->e_cache = cache;
409 ce->e_used = 1 + MB_CACHE_WRITER;
410 ce->e_queued = 0;
411 }
412 return ce;
413}
414
415
416/*
417 * mb_cache_entry_insert()
418 *
419 * Inserts an entry that was allocated using mb_cache_entry_alloc() into
420 * the cache. After this, the cache entry can be looked up, but is not yet
421 * in the lru list as the caller still holds a handle to it. Returns 0 on
422 * success, or -EBUSY if a cache entry for that device + inode exists
423 * already (this may happen after a failed lookup, but when another process
424 * has inserted the same cache entry in the meantime).
425 *
426 * @bdev: device the cache entry belongs to
427 * @block: block number
428 * @keys: array of additional keys. There must be indexes_count entries
429 * in the array (as specified when creating the cache).
430 */
431int
432mb_cache_entry_insert(struct mb_cache_entry *ce, struct block_device *bdev,
433 sector_t block, unsigned int keys[])
434{
435 struct mb_cache *cache = ce->e_cache;
436 unsigned int bucket;
437 struct list_head *l;
438 int error = -EBUSY, n;
439
440 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
441 cache->c_bucket_bits);
442 spin_lock(&mb_cache_spinlock);
443 list_for_each_prev(l, &cache->c_block_hash[bucket]) {
444 struct mb_cache_entry *ce =
445 list_entry(l, struct mb_cache_entry, e_block_list);
446 if (ce->e_bdev == bdev && ce->e_block == block)
447 goto out;
448 }
449 __mb_cache_entry_unhash(ce);
450 ce->e_bdev = bdev;
451 ce->e_block = block;
452 list_add(&ce->e_block_list, &cache->c_block_hash[bucket]);
453 for (n=0; n<mb_cache_indexes(cache); n++) {
454 ce->e_indexes[n].o_key = keys[n];
455 bucket = hash_long(keys[n], cache->c_bucket_bits);
456 list_add(&ce->e_indexes[n].o_list,
457 &cache->c_indexes_hash[n][bucket]);
458 }
459 error = 0;
460out:
461 spin_unlock(&mb_cache_spinlock);
462 return error;
463}
464
465
466/*
467 * mb_cache_entry_release()
468 *
469 * Release a handle to a cache entry. When the last handle to a cache entry
470 * is released it is either freed (if it is invalid) or otherwise inserted
471 * in to the lru list.
472 */
473void
474mb_cache_entry_release(struct mb_cache_entry *ce)
475{
476 spin_lock(&mb_cache_spinlock);
477 __mb_cache_entry_release_unlock(ce);
478}
479
480
481/*
482 * mb_cache_entry_free()
483 *
484 * This is equivalent to the sequence mb_cache_entry_takeout() --
485 * mb_cache_entry_release().
486 */
487void
488mb_cache_entry_free(struct mb_cache_entry *ce)
489{
490 spin_lock(&mb_cache_spinlock);
491 mb_assert(list_empty(&ce->e_lru_list));
492 __mb_cache_entry_unhash(ce);
493 __mb_cache_entry_release_unlock(ce);
494}
495
496
497/*
498 * mb_cache_entry_get()
499 *
500 * Get a cache entry by device / block number. (There can only be one entry
501 * in the cache per device and block.) Returns NULL if no such cache entry
502 * exists. The returned cache entry is locked for exclusive access ("single
503 * writer").
504 */
505struct mb_cache_entry *
506mb_cache_entry_get(struct mb_cache *cache, struct block_device *bdev,
507 sector_t block)
508{
509 unsigned int bucket;
510 struct list_head *l;
511 struct mb_cache_entry *ce;
512
513 bucket = hash_long((unsigned long)bdev + (block & 0xffffffff),
514 cache->c_bucket_bits);
515 spin_lock(&mb_cache_spinlock);
516 list_for_each(l, &cache->c_block_hash[bucket]) {
517 ce = list_entry(l, struct mb_cache_entry, e_block_list);
518 if (ce->e_bdev == bdev && ce->e_block == block) {
519 DEFINE_WAIT(wait);
520
521 if (!list_empty(&ce->e_lru_list))
522 list_del_init(&ce->e_lru_list);
523
524 while (ce->e_used > 0) {
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 ce->e_used += 1 + MB_CACHE_WRITER;
535
536 if (!__mb_cache_entry_is_hashed(ce)) {
537 __mb_cache_entry_release_unlock(ce);
538 return NULL;
539 }
540 goto cleanup;
541 }
542 }
543 ce = NULL;
544
545cleanup:
546 spin_unlock(&mb_cache_spinlock);
547 return ce;
548}
549
550#if !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0)
551
552static struct mb_cache_entry *
553__mb_cache_entry_find(struct list_head *l, struct list_head *head,
554 int index, struct block_device *bdev, unsigned int key)
555{
556 while (l != head) {
557 struct mb_cache_entry *ce =
558 list_entry(l, struct mb_cache_entry,
559 e_indexes[index].o_list);
560 if (ce->e_bdev == bdev && ce->e_indexes[index].o_key == key) {
561 DEFINE_WAIT(wait);
562
563 if (!list_empty(&ce->e_lru_list))
564 list_del_init(&ce->e_lru_list);
565
566 /* Incrementing before holding the lock gives readers
567 priority over writers. */
568 ce->e_used++;
569 while (ce->e_used >= MB_CACHE_WRITER) {
570 ce->e_queued++;
571 prepare_to_wait(&mb_cache_queue, &wait,
572 TASK_UNINTERRUPTIBLE);
573 spin_unlock(&mb_cache_spinlock);
574 schedule();
575 spin_lock(&mb_cache_spinlock);
576 ce->e_queued--;
577 }
578 finish_wait(&mb_cache_queue, &wait);
579
580 if (!__mb_cache_entry_is_hashed(ce)) {
581 __mb_cache_entry_release_unlock(ce);
582 spin_lock(&mb_cache_spinlock);
583 return ERR_PTR(-EAGAIN);
584 }
585 return ce;
586 }
587 l = l->next;
588 }
589 return NULL;
590}
591
592
593/*
594 * mb_cache_entry_find_first()
595 *
596 * Find the first cache entry on a given device with a certain key in
597 * an additional index. Additonal matches can be found with
598 * mb_cache_entry_find_next(). Returns NULL if no match was found. The
599 * returned cache entry is locked for shared access ("multiple readers").
600 *
601 * @cache: the cache to search
602 * @index: the number of the additonal index to search (0<=index<indexes_count)
603 * @bdev: the device the cache entry should belong to
604 * @key: the key in the index
605 */
606struct mb_cache_entry *
607mb_cache_entry_find_first(struct mb_cache *cache, int index,
608 struct block_device *bdev, unsigned int key)
609{
610 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
611 struct list_head *l;
612 struct mb_cache_entry *ce;
613
614 mb_assert(index < mb_cache_indexes(cache));
615 spin_lock(&mb_cache_spinlock);
616 l = cache->c_indexes_hash[index][bucket].next;
617 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
618 index, bdev, key);
619 spin_unlock(&mb_cache_spinlock);
620 return ce;
621}
622
623
624/*
625 * mb_cache_entry_find_next()
626 *
627 * Find the next cache entry on a given device with a certain key in an
628 * additional index. Returns NULL if no match could be found. The previous
629 * entry is atomatically released, so that mb_cache_entry_find_next() can
630 * be called like this:
631 *
632 * entry = mb_cache_entry_find_first();
633 * while (entry) {
634 * ...
635 * entry = mb_cache_entry_find_next(entry, ...);
636 * }
637 *
638 * @prev: The previous match
639 * @index: the number of the additonal index to search (0<=index<indexes_count)
640 * @bdev: the device the cache entry should belong to
641 * @key: the key in the index
642 */
643struct mb_cache_entry *
644mb_cache_entry_find_next(struct mb_cache_entry *prev, int index,
645 struct block_device *bdev, unsigned int key)
646{
647 struct mb_cache *cache = prev->e_cache;
648 unsigned int bucket = hash_long(key, cache->c_bucket_bits);
649 struct list_head *l;
650 struct mb_cache_entry *ce;
651
652 mb_assert(index < mb_cache_indexes(cache));
653 spin_lock(&mb_cache_spinlock);
654 l = prev->e_indexes[index].o_list.next;
655 ce = __mb_cache_entry_find(l, &cache->c_indexes_hash[index][bucket],
656 index, bdev, key);
657 __mb_cache_entry_release_unlock(prev);
658 return ce;
659}
660
661#endif /* !defined(MB_CACHE_INDEXES_COUNT) || (MB_CACHE_INDEXES_COUNT > 0) */
662
663static int __init init_mbcache(void)
664{
665 mb_shrinker = set_shrinker(DEFAULT_SEEKS, mb_cache_shrink_fn);
666 return 0;
667}
668
669static void __exit exit_mbcache(void)
670{
671 remove_shrinker(mb_shrinker);
672}
673
674module_init(init_mbcache)
675module_exit(exit_mbcache)
676