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 inline_endio(struct bio *bio, int error)
537 bio_end_io_t *end_fn = bio->bi_private;
540 * Reset the bio to free any attached resources
541 * (e.g. bio integrity profiles).
548 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
549 bio_end_io_t *end_io)
555 b->bio.bi_io_vec = b->bio_vec;
556 b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
557 b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
558 b->bio.bi_bdev = b->c->bdev;
559 b->bio.bi_end_io = inline_endio;
561 * Use of .bi_private isn't a problem here because
562 * the dm_buffer's inline bio is local to bufio.
564 b->bio.bi_private = end_io;
567 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
568 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
571 len = b->c->block_size;
573 if (len >= PAGE_SIZE)
574 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
576 BUG_ON((unsigned long)ptr & (len - 1));
579 if (!bio_add_page(&b->bio, virt_to_page(ptr),
580 len < PAGE_SIZE ? len : PAGE_SIZE,
581 virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
582 BUG_ON(b->c->block_size <= PAGE_SIZE);
583 use_dmio(b, rw, block, end_io);
591 submit_bio(rw, &b->bio);
594 static void submit_io(struct dm_buffer *b, int rw, sector_t block,
595 bio_end_io_t *end_io)
597 if (rw == WRITE && b->c->write_callback)
598 b->c->write_callback(b);
600 if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
601 b->data_mode != DATA_MODE_VMALLOC)
602 use_inline_bio(b, rw, block, end_io);
604 use_dmio(b, rw, block, end_io);
607 /*----------------------------------------------------------------
608 * Writing dirty buffers
609 *--------------------------------------------------------------*/
612 * The endio routine for write.
614 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
617 static void write_endio(struct bio *bio, int error)
619 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
621 b->write_error = error;
622 if (unlikely(error)) {
623 struct dm_bufio_client *c = b->c;
624 (void)cmpxchg(&c->async_write_error, 0, error);
627 BUG_ON(!test_bit(B_WRITING, &b->state));
629 smp_mb__before_clear_bit();
630 clear_bit(B_WRITING, &b->state);
631 smp_mb__after_clear_bit();
633 wake_up_bit(&b->state, B_WRITING);
637 * This function is called when wait_on_bit is actually waiting.
639 static int do_io_schedule(void *word)
647 * Initiate a write on a dirty buffer, but don't wait for it.
649 * - If the buffer is not dirty, exit.
650 * - If there some previous write going on, wait for it to finish (we can't
651 * have two writes on the same buffer simultaneously).
652 * - Submit our write and don't wait on it. We set B_WRITING indicating
653 * that there is a write in progress.
655 static void __write_dirty_buffer(struct dm_buffer *b,
656 struct list_head *write_list)
658 if (!test_bit(B_DIRTY, &b->state))
661 clear_bit(B_DIRTY, &b->state);
662 wait_on_bit_lock(&b->state, B_WRITING,
663 do_io_schedule, TASK_UNINTERRUPTIBLE);
666 submit_io(b, WRITE, b->block, write_endio);
668 list_add_tail(&b->write_list, write_list);
671 static void __flush_write_list(struct list_head *write_list)
673 struct blk_plug plug;
674 blk_start_plug(&plug);
675 while (!list_empty(write_list)) {
676 struct dm_buffer *b =
677 list_entry(write_list->next, struct dm_buffer, write_list);
678 list_del(&b->write_list);
679 submit_io(b, WRITE, b->block, write_endio);
680 dm_bufio_cond_resched();
682 blk_finish_plug(&plug);
686 * Wait until any activity on the buffer finishes. Possibly write the
687 * buffer if it is dirty. When this function finishes, there is no I/O
688 * running on the buffer and the buffer is not dirty.
690 static void __make_buffer_clean(struct dm_buffer *b)
692 BUG_ON(b->hold_count);
694 if (!b->state) /* fast case */
697 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
698 __write_dirty_buffer(b, NULL);
699 wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
703 * Find some buffer that is not held by anybody, clean it, unlink it and
706 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
710 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
711 BUG_ON(test_bit(B_WRITING, &b->state));
712 BUG_ON(test_bit(B_DIRTY, &b->state));
714 if (!b->hold_count) {
715 __make_buffer_clean(b);
719 dm_bufio_cond_resched();
722 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
723 BUG_ON(test_bit(B_READING, &b->state));
725 if (!b->hold_count) {
726 __make_buffer_clean(b);
730 dm_bufio_cond_resched();
737 * Wait until some other threads free some buffer or release hold count on
740 * This function is entered with c->lock held, drops it and regains it
743 static void __wait_for_free_buffer(struct dm_bufio_client *c)
745 DECLARE_WAITQUEUE(wait, current);
747 add_wait_queue(&c->free_buffer_wait, &wait);
748 set_task_state(current, TASK_UNINTERRUPTIBLE);
753 set_task_state(current, TASK_RUNNING);
754 remove_wait_queue(&c->free_buffer_wait, &wait);
767 * Allocate a new buffer. If the allocation is not possible, wait until
768 * some other thread frees a buffer.
770 * May drop the lock and regain it.
772 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
777 * dm-bufio is resistant to allocation failures (it just keeps
778 * one buffer reserved in cases all the allocations fail).
779 * So set flags to not try too hard:
780 * GFP_NOIO: don't recurse into the I/O layer
781 * __GFP_NORETRY: don't retry and rather return failure
782 * __GFP_NOMEMALLOC: don't use emergency reserves
783 * __GFP_NOWARN: don't print a warning in case of failure
785 * For debugging, if we set the cache size to 1, no new buffers will
789 if (dm_bufio_cache_size_latch != 1) {
790 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
795 if (nf == NF_PREFETCH)
798 if (!list_empty(&c->reserved_buffers)) {
799 b = list_entry(c->reserved_buffers.next,
800 struct dm_buffer, lru_list);
801 list_del(&b->lru_list);
802 c->need_reserved_buffers++;
807 b = __get_unclaimed_buffer(c);
811 __wait_for_free_buffer(c);
815 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
817 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
822 if (c->alloc_callback)
823 c->alloc_callback(b);
829 * Free a buffer and wake other threads waiting for free buffers.
831 static void __free_buffer_wake(struct dm_buffer *b)
833 struct dm_bufio_client *c = b->c;
835 if (!c->need_reserved_buffers)
838 list_add(&b->lru_list, &c->reserved_buffers);
839 c->need_reserved_buffers--;
842 wake_up(&c->free_buffer_wait);
845 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
846 struct list_head *write_list)
848 struct dm_buffer *b, *tmp;
850 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
851 BUG_ON(test_bit(B_READING, &b->state));
853 if (!test_bit(B_DIRTY, &b->state) &&
854 !test_bit(B_WRITING, &b->state)) {
855 __relink_lru(b, LIST_CLEAN);
859 if (no_wait && test_bit(B_WRITING, &b->state))
862 __write_dirty_buffer(b, write_list);
863 dm_bufio_cond_resched();
868 * Get writeback threshold and buffer limit for a given client.
870 static void __get_memory_limit(struct dm_bufio_client *c,
871 unsigned long *threshold_buffers,
872 unsigned long *limit_buffers)
874 unsigned long buffers;
876 if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
877 mutex_lock(&dm_bufio_clients_lock);
878 __cache_size_refresh();
879 mutex_unlock(&dm_bufio_clients_lock);
882 buffers = dm_bufio_cache_size_per_client >>
883 (c->sectors_per_block_bits + SECTOR_SHIFT);
885 if (buffers < c->minimum_buffers)
886 buffers = c->minimum_buffers;
888 *limit_buffers = buffers;
889 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
893 * Check if we're over watermark.
894 * If we are over threshold_buffers, start freeing buffers.
895 * If we're over "limit_buffers", block until we get under the limit.
897 static void __check_watermark(struct dm_bufio_client *c,
898 struct list_head *write_list)
900 unsigned long threshold_buffers, limit_buffers;
902 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
904 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
907 struct dm_buffer *b = __get_unclaimed_buffer(c);
912 __free_buffer_wake(b);
913 dm_bufio_cond_resched();
916 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
917 __write_dirty_buffers_async(c, 1, write_list);
921 * Find a buffer in the hash.
923 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
927 hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
929 dm_bufio_cond_resched();
930 if (b->block == block)
937 /*----------------------------------------------------------------
939 *--------------------------------------------------------------*/
941 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
942 enum new_flag nf, int *need_submit,
943 struct list_head *write_list)
945 struct dm_buffer *b, *new_b = NULL;
949 b = __find(c, block);
956 new_b = __alloc_buffer_wait(c, nf);
961 * We've had a period where the mutex was unlocked, so need to
962 * recheck the hash table.
964 b = __find(c, block);
966 __free_buffer_wake(new_b);
970 __check_watermark(c, write_list);
976 __link_buffer(b, block, LIST_CLEAN);
978 if (nf == NF_FRESH) {
983 b->state = 1 << B_READING;
989 if (nf == NF_PREFETCH)
992 * Note: it is essential that we don't wait for the buffer to be
993 * read if dm_bufio_get function is used. Both dm_bufio_get and
994 * dm_bufio_prefetch can be used in the driver request routine.
995 * If the user called both dm_bufio_prefetch and dm_bufio_get on
996 * the same buffer, it would deadlock if we waited.
998 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1002 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1003 test_bit(B_WRITING, &b->state));
1008 * The endio routine for reading: set the error, clear the bit and wake up
1009 * anyone waiting on the buffer.
1011 static void read_endio(struct bio *bio, int error)
1013 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1015 b->read_error = error;
1017 BUG_ON(!test_bit(B_READING, &b->state));
1019 smp_mb__before_clear_bit();
1020 clear_bit(B_READING, &b->state);
1021 smp_mb__after_clear_bit();
1023 wake_up_bit(&b->state, B_READING);
1027 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1028 * functions is similar except that dm_bufio_new doesn't read the
1029 * buffer from the disk (assuming that the caller overwrites all the data
1030 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1032 static void *new_read(struct dm_bufio_client *c, sector_t block,
1033 enum new_flag nf, struct dm_buffer **bp)
1036 struct dm_buffer *b;
1038 LIST_HEAD(write_list);
1041 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1044 __flush_write_list(&write_list);
1050 submit_io(b, READ, b->block, read_endio);
1052 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
1054 if (b->read_error) {
1055 int error = b->read_error;
1057 dm_bufio_release(b);
1059 return ERR_PTR(error);
1067 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1068 struct dm_buffer **bp)
1070 return new_read(c, block, NF_GET, bp);
1072 EXPORT_SYMBOL_GPL(dm_bufio_get);
1074 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1075 struct dm_buffer **bp)
1077 BUG_ON(dm_bufio_in_request());
1079 return new_read(c, block, NF_READ, bp);
1081 EXPORT_SYMBOL_GPL(dm_bufio_read);
1083 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1084 struct dm_buffer **bp)
1086 BUG_ON(dm_bufio_in_request());
1088 return new_read(c, block, NF_FRESH, bp);
1090 EXPORT_SYMBOL_GPL(dm_bufio_new);
1092 void dm_bufio_prefetch(struct dm_bufio_client *c,
1093 sector_t block, unsigned n_blocks)
1095 struct blk_plug plug;
1097 LIST_HEAD(write_list);
1099 BUG_ON(dm_bufio_in_request());
1101 blk_start_plug(&plug);
1104 for (; n_blocks--; block++) {
1106 struct dm_buffer *b;
1107 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1109 if (unlikely(!list_empty(&write_list))) {
1111 blk_finish_plug(&plug);
1112 __flush_write_list(&write_list);
1113 blk_start_plug(&plug);
1116 if (unlikely(b != NULL)) {
1120 submit_io(b, READ, b->block, read_endio);
1121 dm_bufio_release(b);
1123 dm_bufio_cond_resched();
1134 blk_finish_plug(&plug);
1136 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1138 void dm_bufio_release(struct dm_buffer *b)
1140 struct dm_bufio_client *c = b->c;
1144 BUG_ON(!b->hold_count);
1147 if (!b->hold_count) {
1148 wake_up(&c->free_buffer_wait);
1151 * If there were errors on the buffer, and the buffer is not
1152 * to be written, free the buffer. There is no point in caching
1155 if ((b->read_error || b->write_error) &&
1156 !test_bit(B_READING, &b->state) &&
1157 !test_bit(B_WRITING, &b->state) &&
1158 !test_bit(B_DIRTY, &b->state)) {
1160 __free_buffer_wake(b);
1166 EXPORT_SYMBOL_GPL(dm_bufio_release);
1168 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1170 struct dm_bufio_client *c = b->c;
1174 BUG_ON(test_bit(B_READING, &b->state));
1176 if (!test_and_set_bit(B_DIRTY, &b->state))
1177 __relink_lru(b, LIST_DIRTY);
1181 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1183 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1185 LIST_HEAD(write_list);
1187 BUG_ON(dm_bufio_in_request());
1190 __write_dirty_buffers_async(c, 0, &write_list);
1192 __flush_write_list(&write_list);
1194 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1197 * For performance, it is essential that the buffers are written asynchronously
1198 * and simultaneously (so that the block layer can merge the writes) and then
1201 * Finally, we flush hardware disk cache.
1203 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1206 unsigned long buffers_processed = 0;
1207 struct dm_buffer *b, *tmp;
1209 LIST_HEAD(write_list);
1212 __write_dirty_buffers_async(c, 0, &write_list);
1214 __flush_write_list(&write_list);
1218 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1219 int dropped_lock = 0;
1221 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1222 buffers_processed++;
1224 BUG_ON(test_bit(B_READING, &b->state));
1226 if (test_bit(B_WRITING, &b->state)) {
1227 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1231 wait_on_bit(&b->state, B_WRITING,
1233 TASK_UNINTERRUPTIBLE);
1237 wait_on_bit(&b->state, B_WRITING,
1239 TASK_UNINTERRUPTIBLE);
1242 if (!test_bit(B_DIRTY, &b->state) &&
1243 !test_bit(B_WRITING, &b->state))
1244 __relink_lru(b, LIST_CLEAN);
1246 dm_bufio_cond_resched();
1249 * If we dropped the lock, the list is no longer consistent,
1250 * so we must restart the search.
1252 * In the most common case, the buffer just processed is
1253 * relinked to the clean list, so we won't loop scanning the
1254 * same buffer again and again.
1256 * This may livelock if there is another thread simultaneously
1257 * dirtying buffers, so we count the number of buffers walked
1258 * and if it exceeds the total number of buffers, it means that
1259 * someone is doing some writes simultaneously with us. In
1260 * this case, stop, dropping the lock.
1265 wake_up(&c->free_buffer_wait);
1268 a = xchg(&c->async_write_error, 0);
1269 f = dm_bufio_issue_flush(c);
1275 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1278 * Use dm-io to send and empty barrier flush the device.
1280 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1282 struct dm_io_request io_req = {
1283 .bi_rw = WRITE_FLUSH,
1284 .mem.type = DM_IO_KMEM,
1285 .mem.ptr.addr = NULL,
1288 struct dm_io_region io_reg = {
1294 BUG_ON(dm_bufio_in_request());
1296 return dm_io(&io_req, 1, &io_reg, NULL);
1298 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1301 * We first delete any other buffer that may be at that new location.
1303 * Then, we write the buffer to the original location if it was dirty.
1305 * Then, if we are the only one who is holding the buffer, relink the buffer
1306 * in the hash queue for the new location.
1308 * If there was someone else holding the buffer, we write it to the new
1309 * location but not relink it, because that other user needs to have the buffer
1310 * at the same place.
1312 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1314 struct dm_bufio_client *c = b->c;
1315 struct dm_buffer *new;
1317 BUG_ON(dm_bufio_in_request());
1322 new = __find(c, new_block);
1324 if (new->hold_count) {
1325 __wait_for_free_buffer(c);
1330 * FIXME: Is there any point waiting for a write that's going
1331 * to be overwritten in a bit?
1333 __make_buffer_clean(new);
1334 __unlink_buffer(new);
1335 __free_buffer_wake(new);
1338 BUG_ON(!b->hold_count);
1339 BUG_ON(test_bit(B_READING, &b->state));
1341 __write_dirty_buffer(b, NULL);
1342 if (b->hold_count == 1) {
1343 wait_on_bit(&b->state, B_WRITING,
1344 do_io_schedule, TASK_UNINTERRUPTIBLE);
1345 set_bit(B_DIRTY, &b->state);
1347 __link_buffer(b, new_block, LIST_DIRTY);
1350 wait_on_bit_lock(&b->state, B_WRITING,
1351 do_io_schedule, TASK_UNINTERRUPTIBLE);
1353 * Relink buffer to "new_block" so that write_callback
1354 * sees "new_block" as a block number.
1355 * After the write, link the buffer back to old_block.
1356 * All this must be done in bufio lock, so that block number
1357 * change isn't visible to other threads.
1359 old_block = b->block;
1361 __link_buffer(b, new_block, b->list_mode);
1362 submit_io(b, WRITE, new_block, write_endio);
1363 wait_on_bit(&b->state, B_WRITING,
1364 do_io_schedule, TASK_UNINTERRUPTIBLE);
1366 __link_buffer(b, old_block, b->list_mode);
1370 dm_bufio_release(b);
1372 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1375 * Free the given buffer.
1377 * This is just a hint, if the buffer is in use or dirty, this function
1380 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1382 struct dm_buffer *b;
1386 b = __find(c, block);
1387 if (b && likely(!b->hold_count) && likely(!b->state)) {
1389 __free_buffer_wake(b);
1394 EXPORT_SYMBOL(dm_bufio_forget);
1396 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1398 c->minimum_buffers = n;
1400 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1402 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1404 return c->block_size;
1406 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1408 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1410 return i_size_read(c->bdev->bd_inode) >>
1411 (SECTOR_SHIFT + c->sectors_per_block_bits);
1413 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1415 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1419 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1421 void *dm_bufio_get_block_data(struct dm_buffer *b)
1425 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1427 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1431 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1433 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1437 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1439 static void drop_buffers(struct dm_bufio_client *c)
1441 struct dm_buffer *b;
1444 BUG_ON(dm_bufio_in_request());
1447 * An optimization so that the buffers are not written one-by-one.
1449 dm_bufio_write_dirty_buffers_async(c);
1453 while ((b = __get_unclaimed_buffer(c)))
1454 __free_buffer_wake(b);
1456 for (i = 0; i < LIST_SIZE; i++)
1457 list_for_each_entry(b, &c->lru[i], lru_list)
1458 DMERR("leaked buffer %llx, hold count %u, list %d",
1459 (unsigned long long)b->block, b->hold_count, i);
1461 for (i = 0; i < LIST_SIZE; i++)
1462 BUG_ON(!list_empty(&c->lru[i]));
1468 * Test if the buffer is unused and too old, and commit it.
1469 * And if GFP_NOFS is used, we must not do any I/O because we hold
1470 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1471 * rerouted to different bufio client.
1473 static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1474 unsigned long max_jiffies)
1476 if (jiffies - b->last_accessed < max_jiffies)
1479 if (!(gfp & __GFP_FS)) {
1480 if (test_bit(B_READING, &b->state) ||
1481 test_bit(B_WRITING, &b->state) ||
1482 test_bit(B_DIRTY, &b->state))
1489 __make_buffer_clean(b);
1491 __free_buffer_wake(b);
1496 static long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1500 struct dm_buffer *b, *tmp;
1503 for (l = 0; l < LIST_SIZE; l++) {
1504 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1505 freed += __cleanup_old_buffer(b, gfp_mask, 0);
1508 dm_bufio_cond_resched();
1514 static unsigned long
1515 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1517 struct dm_bufio_client *c;
1518 unsigned long freed;
1520 c = container_of(shrink, struct dm_bufio_client, shrinker);
1521 if (sc->gfp_mask & __GFP_FS)
1523 else if (!dm_bufio_trylock(c))
1526 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1531 static unsigned long
1532 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1534 struct dm_bufio_client *c;
1535 unsigned long count;
1537 c = container_of(shrink, struct dm_bufio_client, shrinker);
1538 if (sc->gfp_mask & __GFP_FS)
1540 else if (!dm_bufio_trylock(c))
1543 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1549 * Create the buffering interface
1551 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1552 unsigned reserved_buffers, unsigned aux_size,
1553 void (*alloc_callback)(struct dm_buffer *),
1554 void (*write_callback)(struct dm_buffer *))
1557 struct dm_bufio_client *c;
1560 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1561 (block_size & (block_size - 1)));
1563 c = kzalloc(sizeof(*c), GFP_KERNEL);
1568 c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1569 if (!c->cache_hash) {
1575 c->block_size = block_size;
1576 c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1577 c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1578 ffs(block_size) - 1 - PAGE_SHIFT : 0;
1579 c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1580 PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1582 c->aux_size = aux_size;
1583 c->alloc_callback = alloc_callback;
1584 c->write_callback = write_callback;
1586 for (i = 0; i < LIST_SIZE; i++) {
1587 INIT_LIST_HEAD(&c->lru[i]);
1588 c->n_buffers[i] = 0;
1591 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1592 INIT_HLIST_HEAD(&c->cache_hash[i]);
1594 mutex_init(&c->lock);
1595 INIT_LIST_HEAD(&c->reserved_buffers);
1596 c->need_reserved_buffers = reserved_buffers;
1598 c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1600 init_waitqueue_head(&c->free_buffer_wait);
1601 c->async_write_error = 0;
1603 c->dm_io = dm_io_client_create();
1604 if (IS_ERR(c->dm_io)) {
1605 r = PTR_ERR(c->dm_io);
1609 mutex_lock(&dm_bufio_clients_lock);
1610 if (c->blocks_per_page_bits) {
1611 if (!DM_BUFIO_CACHE_NAME(c)) {
1612 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1613 if (!DM_BUFIO_CACHE_NAME(c)) {
1615 mutex_unlock(&dm_bufio_clients_lock);
1620 if (!DM_BUFIO_CACHE(c)) {
1621 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1623 c->block_size, 0, NULL);
1624 if (!DM_BUFIO_CACHE(c)) {
1626 mutex_unlock(&dm_bufio_clients_lock);
1631 mutex_unlock(&dm_bufio_clients_lock);
1633 while (c->need_reserved_buffers) {
1634 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1640 __free_buffer_wake(b);
1643 mutex_lock(&dm_bufio_clients_lock);
1644 dm_bufio_client_count++;
1645 list_add(&c->client_list, &dm_bufio_all_clients);
1646 __cache_size_refresh();
1647 mutex_unlock(&dm_bufio_clients_lock);
1649 c->shrinker.count_objects = dm_bufio_shrink_count;
1650 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1651 c->shrinker.seeks = 1;
1652 c->shrinker.batch = 0;
1653 register_shrinker(&c->shrinker);
1659 while (!list_empty(&c->reserved_buffers)) {
1660 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1661 struct dm_buffer, lru_list);
1662 list_del(&b->lru_list);
1665 dm_io_client_destroy(c->dm_io);
1667 vfree(c->cache_hash);
1673 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1676 * Free the buffering interface.
1677 * It is required that there are no references on any buffers.
1679 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1685 unregister_shrinker(&c->shrinker);
1687 mutex_lock(&dm_bufio_clients_lock);
1689 list_del(&c->client_list);
1690 dm_bufio_client_count--;
1691 __cache_size_refresh();
1693 mutex_unlock(&dm_bufio_clients_lock);
1695 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1696 BUG_ON(!hlist_empty(&c->cache_hash[i]));
1698 BUG_ON(c->need_reserved_buffers);
1700 while (!list_empty(&c->reserved_buffers)) {
1701 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1702 struct dm_buffer, lru_list);
1703 list_del(&b->lru_list);
1707 for (i = 0; i < LIST_SIZE; i++)
1708 if (c->n_buffers[i])
1709 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1711 for (i = 0; i < LIST_SIZE; i++)
1712 BUG_ON(c->n_buffers[i]);
1714 dm_io_client_destroy(c->dm_io);
1715 vfree(c->cache_hash);
1718 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1720 static void cleanup_old_buffers(void)
1722 unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1723 struct dm_bufio_client *c;
1725 if (max_age > ULONG_MAX / HZ)
1726 max_age = ULONG_MAX / HZ;
1728 mutex_lock(&dm_bufio_clients_lock);
1729 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1730 if (!dm_bufio_trylock(c))
1733 while (!list_empty(&c->lru[LIST_CLEAN])) {
1734 struct dm_buffer *b;
1735 b = list_entry(c->lru[LIST_CLEAN].prev,
1736 struct dm_buffer, lru_list);
1737 if (!__cleanup_old_buffer(b, 0, max_age * HZ))
1739 dm_bufio_cond_resched();
1743 dm_bufio_cond_resched();
1745 mutex_unlock(&dm_bufio_clients_lock);
1748 static struct workqueue_struct *dm_bufio_wq;
1749 static struct delayed_work dm_bufio_work;
1751 static void work_fn(struct work_struct *w)
1753 cleanup_old_buffers();
1755 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1756 DM_BUFIO_WORK_TIMER_SECS * HZ);
1759 /*----------------------------------------------------------------
1761 *--------------------------------------------------------------*/
1764 * This is called only once for the whole dm_bufio module.
1765 * It initializes memory limit.
1767 static int __init dm_bufio_init(void)
1771 dm_bufio_allocated_kmem_cache = 0;
1772 dm_bufio_allocated_get_free_pages = 0;
1773 dm_bufio_allocated_vmalloc = 0;
1774 dm_bufio_current_allocated = 0;
1776 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1777 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1779 mem = (__u64)((totalram_pages - totalhigh_pages) *
1780 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1782 if (mem > ULONG_MAX)
1787 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1788 * in fs/proc/internal.h
1790 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1791 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1794 dm_bufio_default_cache_size = mem;
1796 mutex_lock(&dm_bufio_clients_lock);
1797 __cache_size_refresh();
1798 mutex_unlock(&dm_bufio_clients_lock);
1800 dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1804 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1805 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1806 DM_BUFIO_WORK_TIMER_SECS * HZ);
1812 * This is called once when unloading the dm_bufio module.
1814 static void __exit dm_bufio_exit(void)
1819 cancel_delayed_work_sync(&dm_bufio_work);
1820 destroy_workqueue(dm_bufio_wq);
1822 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1823 struct kmem_cache *kc = dm_bufio_caches[i];
1826 kmem_cache_destroy(kc);
1829 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1830 kfree(dm_bufio_cache_names[i]);
1832 if (dm_bufio_client_count) {
1833 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1834 __func__, dm_bufio_client_count);
1838 if (dm_bufio_current_allocated) {
1839 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1840 __func__, dm_bufio_current_allocated);
1844 if (dm_bufio_allocated_get_free_pages) {
1845 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1846 __func__, dm_bufio_allocated_get_free_pages);
1850 if (dm_bufio_allocated_vmalloc) {
1851 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1852 __func__, dm_bufio_allocated_vmalloc);
1860 module_init(dm_bufio_init)
1861 module_exit(dm_bufio_exit)
1863 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1864 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1866 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1867 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1869 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1870 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1872 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1873 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1875 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1876 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1878 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1879 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1881 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1882 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1884 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1885 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1886 MODULE_LICENSE("GPL");