2 * Copyright (C) 2009-2011 Red Hat, Inc.
4 * Author: Mikulas Patocka <mpatocka@redhat.com>
6 * This file is released under the GPL.
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/shrinker.h>
16 #include <linux/module.h>
18 #define DM_MSG_PREFIX "bufio"
21 * Memory management policy:
22 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
23 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
24 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
25 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
28 #define DM_BUFIO_MIN_BUFFERS 8
30 #define DM_BUFIO_MEMORY_PERCENT 2
31 #define DM_BUFIO_VMALLOC_PERCENT 25
32 #define DM_BUFIO_WRITEBACK_PERCENT 75
35 * Check buffer ages in this interval (seconds)
37 #define DM_BUFIO_WORK_TIMER_SECS 10
40 * Free buffers when they are older than this (seconds)
42 #define DM_BUFIO_DEFAULT_AGE_SECS 60
45 * The number of bvec entries that are embedded directly in the buffer.
46 * If the chunk size is larger, dm-io is used to do the io.
48 #define DM_BUFIO_INLINE_VECS 16
53 #define DM_BUFIO_HASH_BITS 20
54 #define DM_BUFIO_HASH(block) \
55 ((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
56 ((1 << DM_BUFIO_HASH_BITS) - 1))
59 * Don't try to use kmem_cache_alloc for blocks larger than this.
60 * For explanation, see alloc_buffer_data below.
62 #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1)
63 #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1))
66 * dm_buffer->list_mode
74 * All buffers are linked to cache_hash with their hash_list field.
76 * Clean buffers that are not being written (B_WRITING not set)
77 * are linked to lru[LIST_CLEAN] with their lru_list field.
79 * Dirty and clean buffers that are being written are linked to
80 * lru[LIST_DIRTY] with their lru_list field. When the write
81 * finishes, the buffer cannot be relinked immediately (because we
82 * are in an interrupt context and relinking requires process
83 * context), so some clean-not-writing buffers can be held on
84 * dirty_lru too. They are later added to lru in the process
87 struct dm_bufio_client {
90 struct list_head lru[LIST_SIZE];
91 unsigned long n_buffers[LIST_SIZE];
93 struct block_device *bdev;
95 unsigned char sectors_per_block_bits;
96 unsigned char pages_per_block_bits;
97 unsigned char blocks_per_page_bits;
99 void (*alloc_callback)(struct dm_buffer *);
100 void (*write_callback)(struct dm_buffer *);
102 struct dm_io_client *dm_io;
104 struct list_head reserved_buffers;
105 unsigned need_reserved_buffers;
107 unsigned minimum_buffers;
109 struct hlist_head *cache_hash;
110 wait_queue_head_t free_buffer_wait;
112 int async_write_error;
114 struct list_head client_list;
115 struct shrinker shrinker;
126 * Describes how the block was allocated:
127 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
128 * See the comment at alloc_buffer_data.
132 DATA_MODE_GET_FREE_PAGES = 1,
133 DATA_MODE_VMALLOC = 2,
138 struct hlist_node hash_list;
139 struct list_head lru_list;
142 enum data_mode data_mode;
143 unsigned char list_mode; /* LIST_* */
148 unsigned long last_accessed;
149 struct dm_bufio_client *c;
150 struct list_head write_list;
152 struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
155 /*----------------------------------------------------------------*/
157 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
158 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
160 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
162 unsigned ret = c->blocks_per_page_bits - 1;
164 BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
169 #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
170 #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
172 #define dm_bufio_in_request() (!!current->bio_list)
174 static void dm_bufio_lock(struct dm_bufio_client *c)
176 mutex_lock_nested(&c->lock, dm_bufio_in_request());
179 static int dm_bufio_trylock(struct dm_bufio_client *c)
181 return mutex_trylock(&c->lock);
184 static void dm_bufio_unlock(struct dm_bufio_client *c)
186 mutex_unlock(&c->lock);
190 * FIXME Move to sched.h?
192 #ifdef CONFIG_PREEMPT_VOLUNTARY
193 # define dm_bufio_cond_resched() \
195 if (unlikely(need_resched())) \
199 # define dm_bufio_cond_resched() do { } while (0)
202 /*----------------------------------------------------------------*/
205 * Default cache size: available memory divided by the ratio.
207 static unsigned long dm_bufio_default_cache_size;
210 * Total cache size set by the user.
212 static unsigned long dm_bufio_cache_size;
215 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
216 * at any time. If it disagrees, the user has changed cache size.
218 static unsigned long dm_bufio_cache_size_latch;
220 static DEFINE_SPINLOCK(param_spinlock);
223 * Buffers are freed after this timeout
225 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
227 static unsigned long dm_bufio_peak_allocated;
228 static unsigned long dm_bufio_allocated_kmem_cache;
229 static unsigned long dm_bufio_allocated_get_free_pages;
230 static unsigned long dm_bufio_allocated_vmalloc;
231 static unsigned long dm_bufio_current_allocated;
233 /*----------------------------------------------------------------*/
236 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
238 static unsigned long dm_bufio_cache_size_per_client;
241 * The current number of clients.
243 static int dm_bufio_client_count;
246 * The list of all clients.
248 static LIST_HEAD(dm_bufio_all_clients);
251 * This mutex protects dm_bufio_cache_size_latch,
252 * dm_bufio_cache_size_per_client and dm_bufio_client_count
254 static DEFINE_MUTEX(dm_bufio_clients_lock);
256 /*----------------------------------------------------------------*/
258 static void adjust_total_allocated(enum data_mode data_mode, long diff)
260 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
261 &dm_bufio_allocated_kmem_cache,
262 &dm_bufio_allocated_get_free_pages,
263 &dm_bufio_allocated_vmalloc,
266 spin_lock(¶m_spinlock);
268 *class_ptr[data_mode] += diff;
270 dm_bufio_current_allocated += diff;
272 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
273 dm_bufio_peak_allocated = dm_bufio_current_allocated;
275 spin_unlock(¶m_spinlock);
279 * Change the number of clients and recalculate per-client limit.
281 static void __cache_size_refresh(void)
283 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
284 BUG_ON(dm_bufio_client_count < 0);
286 dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
289 * Use default if set to 0 and report the actual cache size used.
291 if (!dm_bufio_cache_size_latch) {
292 (void)cmpxchg(&dm_bufio_cache_size, 0,
293 dm_bufio_default_cache_size);
294 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
297 dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
298 (dm_bufio_client_count ? : 1);
302 * Allocating buffer data.
304 * Small buffers are allocated with kmem_cache, to use space optimally.
306 * For large buffers, we choose between get_free_pages and vmalloc.
307 * Each has advantages and disadvantages.
309 * __get_free_pages can randomly fail if the memory is fragmented.
310 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
311 * as low as 128M) so using it for caching is not appropriate.
313 * If the allocation may fail we use __get_free_pages. Memory fragmentation
314 * won't have a fatal effect here, but it just causes flushes of some other
315 * buffers and more I/O will be performed. Don't use __get_free_pages if it
316 * always fails (i.e. order >= MAX_ORDER).
318 * If the allocation shouldn't fail we use __vmalloc. This is only for the
319 * initial reserve allocation, so there's no risk of wasting all vmalloc
322 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
323 enum data_mode *data_mode)
328 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
329 *data_mode = DATA_MODE_SLAB;
330 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
333 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
334 gfp_mask & __GFP_NORETRY) {
335 *data_mode = DATA_MODE_GET_FREE_PAGES;
336 return (void *)__get_free_pages(gfp_mask,
337 c->pages_per_block_bits);
340 *data_mode = DATA_MODE_VMALLOC;
343 * __vmalloc allocates the data pages and auxiliary structures with
344 * gfp_flags that were specified, but pagetables are always allocated
345 * with GFP_KERNEL, no matter what was specified as gfp_mask.
347 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
348 * all allocations done by this process (including pagetables) are done
349 * as if GFP_NOIO was specified.
352 if (gfp_mask & __GFP_NORETRY)
353 noio_flag = memalloc_noio_save();
355 ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
357 if (gfp_mask & __GFP_NORETRY)
358 memalloc_noio_restore(noio_flag);
364 * Free buffer's data.
366 static void free_buffer_data(struct dm_bufio_client *c,
367 void *data, enum data_mode data_mode)
371 kmem_cache_free(DM_BUFIO_CACHE(c), data);
374 case DATA_MODE_GET_FREE_PAGES:
375 free_pages((unsigned long)data, c->pages_per_block_bits);
378 case DATA_MODE_VMALLOC:
383 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
390 * Allocate buffer and its data.
392 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
394 struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
402 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
408 adjust_total_allocated(b->data_mode, (long)c->block_size);
414 * Free buffer and its data.
416 static void free_buffer(struct dm_buffer *b)
418 struct dm_bufio_client *c = b->c;
420 adjust_total_allocated(b->data_mode, -(long)c->block_size);
422 free_buffer_data(c, b->data, b->data_mode);
427 * Link buffer to the hash list and clean or dirty queue.
429 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
431 struct dm_bufio_client *c = b->c;
433 c->n_buffers[dirty]++;
435 b->list_mode = dirty;
436 list_add(&b->lru_list, &c->lru[dirty]);
437 hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
438 b->last_accessed = jiffies;
442 * Unlink buffer from the hash list and dirty or clean queue.
444 static void __unlink_buffer(struct dm_buffer *b)
446 struct dm_bufio_client *c = b->c;
448 BUG_ON(!c->n_buffers[b->list_mode]);
450 c->n_buffers[b->list_mode]--;
451 hlist_del(&b->hash_list);
452 list_del(&b->lru_list);
456 * Place the buffer to the head of dirty or clean LRU queue.
458 static void __relink_lru(struct dm_buffer *b, int dirty)
460 struct dm_bufio_client *c = b->c;
462 BUG_ON(!c->n_buffers[b->list_mode]);
464 c->n_buffers[b->list_mode]--;
465 c->n_buffers[dirty]++;
466 b->list_mode = dirty;
467 list_move(&b->lru_list, &c->lru[dirty]);
468 b->last_accessed = jiffies;
471 /*----------------------------------------------------------------
472 * Submit I/O on the buffer.
474 * Bio interface is faster but it has some problems:
475 * the vector list is limited (increasing this limit increases
476 * memory-consumption per buffer, so it is not viable);
478 * the memory must be direct-mapped, not vmalloced;
480 * the I/O driver can reject requests spuriously if it thinks that
481 * the requests are too big for the device or if they cross a
482 * controller-defined memory boundary.
484 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
485 * it is not vmalloced, try using the bio interface.
487 * If the buffer is big, if it is vmalloced or if the underlying device
488 * rejects the bio because it is too large, use dm-io layer to do the I/O.
489 * The dm-io layer splits the I/O into multiple requests, avoiding the above
491 *--------------------------------------------------------------*/
494 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
495 * that the request was handled directly with bio interface.
497 static void dmio_complete(unsigned long error, void *context)
499 struct dm_buffer *b = context;
501 b->bio.bi_end_io(&b->bio, error ? -EIO : 0);
504 static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
505 bio_end_io_t *end_io)
508 struct dm_io_request io_req = {
510 .notify.fn = dmio_complete,
512 .client = b->c->dm_io,
514 struct dm_io_region region = {
516 .sector = block << b->c->sectors_per_block_bits,
517 .count = b->c->block_size >> SECTOR_SHIFT,
520 if (b->data_mode != DATA_MODE_VMALLOC) {
521 io_req.mem.type = DM_IO_KMEM;
522 io_req.mem.ptr.addr = b->data;
524 io_req.mem.type = DM_IO_VMA;
525 io_req.mem.ptr.vma = b->data;
528 b->bio.bi_end_io = end_io;
530 r = dm_io(&io_req, 1, ®ion, NULL);
535 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
536 bio_end_io_t *end_io)
542 b->bio.bi_io_vec = b->bio_vec;
543 b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
544 b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
545 b->bio.bi_bdev = b->c->bdev;
546 b->bio.bi_end_io = end_io;
549 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
550 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
553 len = b->c->block_size;
555 if (len >= PAGE_SIZE)
556 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
558 BUG_ON((unsigned long)ptr & (len - 1));
561 if (!bio_add_page(&b->bio, virt_to_page(ptr),
562 len < PAGE_SIZE ? len : PAGE_SIZE,
563 virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
564 BUG_ON(b->c->block_size <= PAGE_SIZE);
565 use_dmio(b, rw, block, end_io);
573 submit_bio(rw, &b->bio);
576 static void submit_io(struct dm_buffer *b, int rw, sector_t block,
577 bio_end_io_t *end_io)
579 if (rw == WRITE && b->c->write_callback)
580 b->c->write_callback(b);
582 if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
583 b->data_mode != DATA_MODE_VMALLOC)
584 use_inline_bio(b, rw, block, end_io);
586 use_dmio(b, rw, block, end_io);
589 /*----------------------------------------------------------------
590 * Writing dirty buffers
591 *--------------------------------------------------------------*/
594 * The endio routine for write.
596 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
599 static void write_endio(struct bio *bio, int error)
601 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
603 b->write_error = error;
604 if (unlikely(error)) {
605 struct dm_bufio_client *c = b->c;
606 (void)cmpxchg(&c->async_write_error, 0, error);
609 BUG_ON(!test_bit(B_WRITING, &b->state));
611 smp_mb__before_clear_bit();
612 clear_bit(B_WRITING, &b->state);
613 smp_mb__after_clear_bit();
615 wake_up_bit(&b->state, B_WRITING);
619 * This function is called when wait_on_bit is actually waiting.
621 static int do_io_schedule(void *word)
629 * Initiate a write on a dirty buffer, but don't wait for it.
631 * - If the buffer is not dirty, exit.
632 * - If there some previous write going on, wait for it to finish (we can't
633 * have two writes on the same buffer simultaneously).
634 * - Submit our write and don't wait on it. We set B_WRITING indicating
635 * that there is a write in progress.
637 static void __write_dirty_buffer(struct dm_buffer *b,
638 struct list_head *write_list)
640 if (!test_bit(B_DIRTY, &b->state))
643 clear_bit(B_DIRTY, &b->state);
644 wait_on_bit_lock(&b->state, B_WRITING,
645 do_io_schedule, TASK_UNINTERRUPTIBLE);
648 submit_io(b, WRITE, b->block, write_endio);
650 list_add_tail(&b->write_list, write_list);
653 static void __flush_write_list(struct list_head *write_list)
655 struct blk_plug plug;
656 blk_start_plug(&plug);
657 while (!list_empty(write_list)) {
658 struct dm_buffer *b =
659 list_entry(write_list->next, struct dm_buffer, write_list);
660 list_del(&b->write_list);
661 submit_io(b, WRITE, b->block, write_endio);
662 dm_bufio_cond_resched();
664 blk_finish_plug(&plug);
668 * Wait until any activity on the buffer finishes. Possibly write the
669 * buffer if it is dirty. When this function finishes, there is no I/O
670 * running on the buffer and the buffer is not dirty.
672 static void __make_buffer_clean(struct dm_buffer *b)
674 BUG_ON(b->hold_count);
676 if (!b->state) /* fast case */
679 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
680 __write_dirty_buffer(b, NULL);
681 wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
685 * Find some buffer that is not held by anybody, clean it, unlink it and
688 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
692 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
693 BUG_ON(test_bit(B_WRITING, &b->state));
694 BUG_ON(test_bit(B_DIRTY, &b->state));
696 if (!b->hold_count) {
697 __make_buffer_clean(b);
701 dm_bufio_cond_resched();
704 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
705 BUG_ON(test_bit(B_READING, &b->state));
707 if (!b->hold_count) {
708 __make_buffer_clean(b);
712 dm_bufio_cond_resched();
719 * Wait until some other threads free some buffer or release hold count on
722 * This function is entered with c->lock held, drops it and regains it
725 static void __wait_for_free_buffer(struct dm_bufio_client *c)
727 DECLARE_WAITQUEUE(wait, current);
729 add_wait_queue(&c->free_buffer_wait, &wait);
730 set_task_state(current, TASK_UNINTERRUPTIBLE);
735 set_task_state(current, TASK_RUNNING);
736 remove_wait_queue(&c->free_buffer_wait, &wait);
749 * Allocate a new buffer. If the allocation is not possible, wait until
750 * some other thread frees a buffer.
752 * May drop the lock and regain it.
754 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
759 * dm-bufio is resistant to allocation failures (it just keeps
760 * one buffer reserved in cases all the allocations fail).
761 * So set flags to not try too hard:
762 * GFP_NOIO: don't recurse into the I/O layer
763 * __GFP_NORETRY: don't retry and rather return failure
764 * __GFP_NOMEMALLOC: don't use emergency reserves
765 * __GFP_NOWARN: don't print a warning in case of failure
767 * For debugging, if we set the cache size to 1, no new buffers will
771 if (dm_bufio_cache_size_latch != 1) {
772 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
777 if (nf == NF_PREFETCH)
780 if (!list_empty(&c->reserved_buffers)) {
781 b = list_entry(c->reserved_buffers.next,
782 struct dm_buffer, lru_list);
783 list_del(&b->lru_list);
784 c->need_reserved_buffers++;
789 b = __get_unclaimed_buffer(c);
793 __wait_for_free_buffer(c);
797 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
799 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
804 if (c->alloc_callback)
805 c->alloc_callback(b);
811 * Free a buffer and wake other threads waiting for free buffers.
813 static void __free_buffer_wake(struct dm_buffer *b)
815 struct dm_bufio_client *c = b->c;
817 if (!c->need_reserved_buffers)
820 list_add(&b->lru_list, &c->reserved_buffers);
821 c->need_reserved_buffers--;
824 wake_up(&c->free_buffer_wait);
827 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
828 struct list_head *write_list)
830 struct dm_buffer *b, *tmp;
832 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
833 BUG_ON(test_bit(B_READING, &b->state));
835 if (!test_bit(B_DIRTY, &b->state) &&
836 !test_bit(B_WRITING, &b->state)) {
837 __relink_lru(b, LIST_CLEAN);
841 if (no_wait && test_bit(B_WRITING, &b->state))
844 __write_dirty_buffer(b, write_list);
845 dm_bufio_cond_resched();
850 * Get writeback threshold and buffer limit for a given client.
852 static void __get_memory_limit(struct dm_bufio_client *c,
853 unsigned long *threshold_buffers,
854 unsigned long *limit_buffers)
856 unsigned long buffers;
858 if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
859 mutex_lock(&dm_bufio_clients_lock);
860 __cache_size_refresh();
861 mutex_unlock(&dm_bufio_clients_lock);
864 buffers = dm_bufio_cache_size_per_client >>
865 (c->sectors_per_block_bits + SECTOR_SHIFT);
867 if (buffers < c->minimum_buffers)
868 buffers = c->minimum_buffers;
870 *limit_buffers = buffers;
871 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
875 * Check if we're over watermark.
876 * If we are over threshold_buffers, start freeing buffers.
877 * If we're over "limit_buffers", block until we get under the limit.
879 static void __check_watermark(struct dm_bufio_client *c,
880 struct list_head *write_list)
882 unsigned long threshold_buffers, limit_buffers;
884 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
886 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
889 struct dm_buffer *b = __get_unclaimed_buffer(c);
894 __free_buffer_wake(b);
895 dm_bufio_cond_resched();
898 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
899 __write_dirty_buffers_async(c, 1, write_list);
903 * Find a buffer in the hash.
905 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
909 hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
911 dm_bufio_cond_resched();
912 if (b->block == block)
919 /*----------------------------------------------------------------
921 *--------------------------------------------------------------*/
923 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
924 enum new_flag nf, int *need_submit,
925 struct list_head *write_list)
927 struct dm_buffer *b, *new_b = NULL;
931 b = __find(c, block);
938 new_b = __alloc_buffer_wait(c, nf);
943 * We've had a period where the mutex was unlocked, so need to
944 * recheck the hash table.
946 b = __find(c, block);
948 __free_buffer_wake(new_b);
952 __check_watermark(c, write_list);
958 __link_buffer(b, block, LIST_CLEAN);
960 if (nf == NF_FRESH) {
965 b->state = 1 << B_READING;
971 if (nf == NF_PREFETCH)
974 * Note: it is essential that we don't wait for the buffer to be
975 * read if dm_bufio_get function is used. Both dm_bufio_get and
976 * dm_bufio_prefetch can be used in the driver request routine.
977 * If the user called both dm_bufio_prefetch and dm_bufio_get on
978 * the same buffer, it would deadlock if we waited.
980 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
984 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
985 test_bit(B_WRITING, &b->state));
990 * The endio routine for reading: set the error, clear the bit and wake up
991 * anyone waiting on the buffer.
993 static void read_endio(struct bio *bio, int error)
995 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
997 b->read_error = error;
999 BUG_ON(!test_bit(B_READING, &b->state));
1001 smp_mb__before_clear_bit();
1002 clear_bit(B_READING, &b->state);
1003 smp_mb__after_clear_bit();
1005 wake_up_bit(&b->state, B_READING);
1009 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1010 * functions is similar except that dm_bufio_new doesn't read the
1011 * buffer from the disk (assuming that the caller overwrites all the data
1012 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1014 static void *new_read(struct dm_bufio_client *c, sector_t block,
1015 enum new_flag nf, struct dm_buffer **bp)
1018 struct dm_buffer *b;
1020 LIST_HEAD(write_list);
1023 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1026 __flush_write_list(&write_list);
1032 submit_io(b, READ, b->block, read_endio);
1034 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
1036 if (b->read_error) {
1037 int error = b->read_error;
1039 dm_bufio_release(b);
1041 return ERR_PTR(error);
1049 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1050 struct dm_buffer **bp)
1052 return new_read(c, block, NF_GET, bp);
1054 EXPORT_SYMBOL_GPL(dm_bufio_get);
1056 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1057 struct dm_buffer **bp)
1059 BUG_ON(dm_bufio_in_request());
1061 return new_read(c, block, NF_READ, bp);
1063 EXPORT_SYMBOL_GPL(dm_bufio_read);
1065 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1066 struct dm_buffer **bp)
1068 BUG_ON(dm_bufio_in_request());
1070 return new_read(c, block, NF_FRESH, bp);
1072 EXPORT_SYMBOL_GPL(dm_bufio_new);
1074 void dm_bufio_prefetch(struct dm_bufio_client *c,
1075 sector_t block, unsigned n_blocks)
1077 struct blk_plug plug;
1079 LIST_HEAD(write_list);
1081 BUG_ON(dm_bufio_in_request());
1083 blk_start_plug(&plug);
1086 for (; n_blocks--; block++) {
1088 struct dm_buffer *b;
1089 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1091 if (unlikely(!list_empty(&write_list))) {
1093 blk_finish_plug(&plug);
1094 __flush_write_list(&write_list);
1095 blk_start_plug(&plug);
1098 if (unlikely(b != NULL)) {
1102 submit_io(b, READ, b->block, read_endio);
1103 dm_bufio_release(b);
1105 dm_bufio_cond_resched();
1116 blk_finish_plug(&plug);
1118 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1120 void dm_bufio_release(struct dm_buffer *b)
1122 struct dm_bufio_client *c = b->c;
1126 BUG_ON(!b->hold_count);
1129 if (!b->hold_count) {
1130 wake_up(&c->free_buffer_wait);
1133 * If there were errors on the buffer, and the buffer is not
1134 * to be written, free the buffer. There is no point in caching
1137 if ((b->read_error || b->write_error) &&
1138 !test_bit(B_READING, &b->state) &&
1139 !test_bit(B_WRITING, &b->state) &&
1140 !test_bit(B_DIRTY, &b->state)) {
1142 __free_buffer_wake(b);
1148 EXPORT_SYMBOL_GPL(dm_bufio_release);
1150 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1152 struct dm_bufio_client *c = b->c;
1156 BUG_ON(test_bit(B_READING, &b->state));
1158 if (!test_and_set_bit(B_DIRTY, &b->state))
1159 __relink_lru(b, LIST_DIRTY);
1163 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1165 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1167 LIST_HEAD(write_list);
1169 BUG_ON(dm_bufio_in_request());
1172 __write_dirty_buffers_async(c, 0, &write_list);
1174 __flush_write_list(&write_list);
1176 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1179 * For performance, it is essential that the buffers are written asynchronously
1180 * and simultaneously (so that the block layer can merge the writes) and then
1183 * Finally, we flush hardware disk cache.
1185 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1188 unsigned long buffers_processed = 0;
1189 struct dm_buffer *b, *tmp;
1191 LIST_HEAD(write_list);
1194 __write_dirty_buffers_async(c, 0, &write_list);
1196 __flush_write_list(&write_list);
1200 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1201 int dropped_lock = 0;
1203 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1204 buffers_processed++;
1206 BUG_ON(test_bit(B_READING, &b->state));
1208 if (test_bit(B_WRITING, &b->state)) {
1209 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1213 wait_on_bit(&b->state, B_WRITING,
1215 TASK_UNINTERRUPTIBLE);
1219 wait_on_bit(&b->state, B_WRITING,
1221 TASK_UNINTERRUPTIBLE);
1224 if (!test_bit(B_DIRTY, &b->state) &&
1225 !test_bit(B_WRITING, &b->state))
1226 __relink_lru(b, LIST_CLEAN);
1228 dm_bufio_cond_resched();
1231 * If we dropped the lock, the list is no longer consistent,
1232 * so we must restart the search.
1234 * In the most common case, the buffer just processed is
1235 * relinked to the clean list, so we won't loop scanning the
1236 * same buffer again and again.
1238 * This may livelock if there is another thread simultaneously
1239 * dirtying buffers, so we count the number of buffers walked
1240 * and if it exceeds the total number of buffers, it means that
1241 * someone is doing some writes simultaneously with us. In
1242 * this case, stop, dropping the lock.
1247 wake_up(&c->free_buffer_wait);
1250 a = xchg(&c->async_write_error, 0);
1251 f = dm_bufio_issue_flush(c);
1257 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1260 * Use dm-io to send and empty barrier flush the device.
1262 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1264 struct dm_io_request io_req = {
1265 .bi_rw = WRITE_FLUSH,
1266 .mem.type = DM_IO_KMEM,
1267 .mem.ptr.addr = NULL,
1270 struct dm_io_region io_reg = {
1276 BUG_ON(dm_bufio_in_request());
1278 return dm_io(&io_req, 1, &io_reg, NULL);
1280 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1283 * We first delete any other buffer that may be at that new location.
1285 * Then, we write the buffer to the original location if it was dirty.
1287 * Then, if we are the only one who is holding the buffer, relink the buffer
1288 * in the hash queue for the new location.
1290 * If there was someone else holding the buffer, we write it to the new
1291 * location but not relink it, because that other user needs to have the buffer
1292 * at the same place.
1294 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1296 struct dm_bufio_client *c = b->c;
1297 struct dm_buffer *new;
1299 BUG_ON(dm_bufio_in_request());
1304 new = __find(c, new_block);
1306 if (new->hold_count) {
1307 __wait_for_free_buffer(c);
1312 * FIXME: Is there any point waiting for a write that's going
1313 * to be overwritten in a bit?
1315 __make_buffer_clean(new);
1316 __unlink_buffer(new);
1317 __free_buffer_wake(new);
1320 BUG_ON(!b->hold_count);
1321 BUG_ON(test_bit(B_READING, &b->state));
1323 __write_dirty_buffer(b, NULL);
1324 if (b->hold_count == 1) {
1325 wait_on_bit(&b->state, B_WRITING,
1326 do_io_schedule, TASK_UNINTERRUPTIBLE);
1327 set_bit(B_DIRTY, &b->state);
1329 __link_buffer(b, new_block, LIST_DIRTY);
1332 wait_on_bit_lock(&b->state, B_WRITING,
1333 do_io_schedule, TASK_UNINTERRUPTIBLE);
1335 * Relink buffer to "new_block" so that write_callback
1336 * sees "new_block" as a block number.
1337 * After the write, link the buffer back to old_block.
1338 * All this must be done in bufio lock, so that block number
1339 * change isn't visible to other threads.
1341 old_block = b->block;
1343 __link_buffer(b, new_block, b->list_mode);
1344 submit_io(b, WRITE, new_block, write_endio);
1345 wait_on_bit(&b->state, B_WRITING,
1346 do_io_schedule, TASK_UNINTERRUPTIBLE);
1348 __link_buffer(b, old_block, b->list_mode);
1352 dm_bufio_release(b);
1354 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1357 * Free the given buffer.
1359 * This is just a hint, if the buffer is in use or dirty, this function
1362 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1364 struct dm_buffer *b;
1368 b = __find(c, block);
1369 if (b && likely(!b->hold_count) && likely(!b->state)) {
1371 __free_buffer_wake(b);
1376 EXPORT_SYMBOL(dm_bufio_forget);
1378 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1380 c->minimum_buffers = n;
1382 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1384 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1386 return c->block_size;
1388 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1390 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1392 return i_size_read(c->bdev->bd_inode) >>
1393 (SECTOR_SHIFT + c->sectors_per_block_bits);
1395 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1397 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1401 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1403 void *dm_bufio_get_block_data(struct dm_buffer *b)
1407 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1409 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1413 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1415 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1419 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1421 static void drop_buffers(struct dm_bufio_client *c)
1423 struct dm_buffer *b;
1426 BUG_ON(dm_bufio_in_request());
1429 * An optimization so that the buffers are not written one-by-one.
1431 dm_bufio_write_dirty_buffers_async(c);
1435 while ((b = __get_unclaimed_buffer(c)))
1436 __free_buffer_wake(b);
1438 for (i = 0; i < LIST_SIZE; i++)
1439 list_for_each_entry(b, &c->lru[i], lru_list)
1440 DMERR("leaked buffer %llx, hold count %u, list %d",
1441 (unsigned long long)b->block, b->hold_count, i);
1443 for (i = 0; i < LIST_SIZE; i++)
1444 BUG_ON(!list_empty(&c->lru[i]));
1450 * Test if the buffer is unused and too old, and commit it.
1451 * At if noio is set, we must not do any I/O because we hold
1452 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1453 * different bufio client.
1455 static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1456 unsigned long max_jiffies)
1458 if (jiffies - b->last_accessed < max_jiffies)
1461 if (!(gfp & __GFP_IO)) {
1462 if (test_bit(B_READING, &b->state) ||
1463 test_bit(B_WRITING, &b->state) ||
1464 test_bit(B_DIRTY, &b->state))
1471 __make_buffer_clean(b);
1473 __free_buffer_wake(b);
1478 static long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1482 struct dm_buffer *b, *tmp;
1485 for (l = 0; l < LIST_SIZE; l++) {
1486 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1487 freed += __cleanup_old_buffer(b, gfp_mask, 0);
1491 dm_bufio_cond_resched();
1496 static unsigned long
1497 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1499 struct dm_bufio_client *c;
1500 unsigned long freed;
1502 c = container_of(shrink, struct dm_bufio_client, shrinker);
1503 if (sc->gfp_mask & __GFP_IO)
1505 else if (!dm_bufio_trylock(c))
1508 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1513 static unsigned long
1514 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1516 struct dm_bufio_client *c;
1517 unsigned long count;
1519 c = container_of(shrink, struct dm_bufio_client, shrinker);
1520 if (sc->gfp_mask & __GFP_IO)
1522 else if (!dm_bufio_trylock(c))
1525 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1531 * Create the buffering interface
1533 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1534 unsigned reserved_buffers, unsigned aux_size,
1535 void (*alloc_callback)(struct dm_buffer *),
1536 void (*write_callback)(struct dm_buffer *))
1539 struct dm_bufio_client *c;
1542 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1543 (block_size & (block_size - 1)));
1545 c = kzalloc(sizeof(*c), GFP_KERNEL);
1550 c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1551 if (!c->cache_hash) {
1557 c->block_size = block_size;
1558 c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1559 c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1560 ffs(block_size) - 1 - PAGE_SHIFT : 0;
1561 c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1562 PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1564 c->aux_size = aux_size;
1565 c->alloc_callback = alloc_callback;
1566 c->write_callback = write_callback;
1568 for (i = 0; i < LIST_SIZE; i++) {
1569 INIT_LIST_HEAD(&c->lru[i]);
1570 c->n_buffers[i] = 0;
1573 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1574 INIT_HLIST_HEAD(&c->cache_hash[i]);
1576 mutex_init(&c->lock);
1577 INIT_LIST_HEAD(&c->reserved_buffers);
1578 c->need_reserved_buffers = reserved_buffers;
1580 c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1582 init_waitqueue_head(&c->free_buffer_wait);
1583 c->async_write_error = 0;
1585 c->dm_io = dm_io_client_create();
1586 if (IS_ERR(c->dm_io)) {
1587 r = PTR_ERR(c->dm_io);
1591 mutex_lock(&dm_bufio_clients_lock);
1592 if (c->blocks_per_page_bits) {
1593 if (!DM_BUFIO_CACHE_NAME(c)) {
1594 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1595 if (!DM_BUFIO_CACHE_NAME(c)) {
1597 mutex_unlock(&dm_bufio_clients_lock);
1602 if (!DM_BUFIO_CACHE(c)) {
1603 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1605 c->block_size, 0, NULL);
1606 if (!DM_BUFIO_CACHE(c)) {
1608 mutex_unlock(&dm_bufio_clients_lock);
1613 mutex_unlock(&dm_bufio_clients_lock);
1615 while (c->need_reserved_buffers) {
1616 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1622 __free_buffer_wake(b);
1625 mutex_lock(&dm_bufio_clients_lock);
1626 dm_bufio_client_count++;
1627 list_add(&c->client_list, &dm_bufio_all_clients);
1628 __cache_size_refresh();
1629 mutex_unlock(&dm_bufio_clients_lock);
1631 c->shrinker.count_objects = dm_bufio_shrink_count;
1632 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1633 c->shrinker.seeks = 1;
1634 c->shrinker.batch = 0;
1635 register_shrinker(&c->shrinker);
1641 while (!list_empty(&c->reserved_buffers)) {
1642 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1643 struct dm_buffer, lru_list);
1644 list_del(&b->lru_list);
1647 dm_io_client_destroy(c->dm_io);
1649 vfree(c->cache_hash);
1655 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1658 * Free the buffering interface.
1659 * It is required that there are no references on any buffers.
1661 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1667 unregister_shrinker(&c->shrinker);
1669 mutex_lock(&dm_bufio_clients_lock);
1671 list_del(&c->client_list);
1672 dm_bufio_client_count--;
1673 __cache_size_refresh();
1675 mutex_unlock(&dm_bufio_clients_lock);
1677 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1678 BUG_ON(!hlist_empty(&c->cache_hash[i]));
1680 BUG_ON(c->need_reserved_buffers);
1682 while (!list_empty(&c->reserved_buffers)) {
1683 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1684 struct dm_buffer, lru_list);
1685 list_del(&b->lru_list);
1689 for (i = 0; i < LIST_SIZE; i++)
1690 if (c->n_buffers[i])
1691 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1693 for (i = 0; i < LIST_SIZE; i++)
1694 BUG_ON(c->n_buffers[i]);
1696 dm_io_client_destroy(c->dm_io);
1697 vfree(c->cache_hash);
1700 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1702 static void cleanup_old_buffers(void)
1704 unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1705 struct dm_bufio_client *c;
1707 if (max_age > ULONG_MAX / HZ)
1708 max_age = ULONG_MAX / HZ;
1710 mutex_lock(&dm_bufio_clients_lock);
1711 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1712 if (!dm_bufio_trylock(c))
1715 while (!list_empty(&c->lru[LIST_CLEAN])) {
1716 struct dm_buffer *b;
1717 b = list_entry(c->lru[LIST_CLEAN].prev,
1718 struct dm_buffer, lru_list);
1719 if (!__cleanup_old_buffer(b, 0, max_age * HZ))
1721 dm_bufio_cond_resched();
1725 dm_bufio_cond_resched();
1727 mutex_unlock(&dm_bufio_clients_lock);
1730 static struct workqueue_struct *dm_bufio_wq;
1731 static struct delayed_work dm_bufio_work;
1733 static void work_fn(struct work_struct *w)
1735 cleanup_old_buffers();
1737 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1738 DM_BUFIO_WORK_TIMER_SECS * HZ);
1741 /*----------------------------------------------------------------
1743 *--------------------------------------------------------------*/
1746 * This is called only once for the whole dm_bufio module.
1747 * It initializes memory limit.
1749 static int __init dm_bufio_init(void)
1753 dm_bufio_allocated_kmem_cache = 0;
1754 dm_bufio_allocated_get_free_pages = 0;
1755 dm_bufio_allocated_vmalloc = 0;
1756 dm_bufio_current_allocated = 0;
1758 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1759 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1761 mem = (__u64)((totalram_pages - totalhigh_pages) *
1762 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1764 if (mem > ULONG_MAX)
1769 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1770 * in fs/proc/internal.h
1772 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1773 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1776 dm_bufio_default_cache_size = mem;
1778 mutex_lock(&dm_bufio_clients_lock);
1779 __cache_size_refresh();
1780 mutex_unlock(&dm_bufio_clients_lock);
1782 dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1786 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1787 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1788 DM_BUFIO_WORK_TIMER_SECS * HZ);
1794 * This is called once when unloading the dm_bufio module.
1796 static void __exit dm_bufio_exit(void)
1801 cancel_delayed_work_sync(&dm_bufio_work);
1802 destroy_workqueue(dm_bufio_wq);
1804 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1805 struct kmem_cache *kc = dm_bufio_caches[i];
1808 kmem_cache_destroy(kc);
1811 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1812 kfree(dm_bufio_cache_names[i]);
1814 if (dm_bufio_client_count) {
1815 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1816 __func__, dm_bufio_client_count);
1820 if (dm_bufio_current_allocated) {
1821 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1822 __func__, dm_bufio_current_allocated);
1826 if (dm_bufio_allocated_get_free_pages) {
1827 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1828 __func__, dm_bufio_allocated_get_free_pages);
1832 if (dm_bufio_allocated_vmalloc) {
1833 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1834 __func__, dm_bufio_allocated_vmalloc);
1842 module_init(dm_bufio_init)
1843 module_exit(dm_bufio_exit)
1845 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1846 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1848 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1849 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1851 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1852 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1854 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1855 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1857 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1858 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1860 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1861 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1863 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1864 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1866 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1867 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1868 MODULE_LICENSE("GPL");