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/jiffies.h>
15 #include <linux/vmalloc.h>
16 #include <linux/shrinker.h>
17 #include <linux/module.h>
18 #include <linux/rbtree.h>
19 #include <linux/stacktrace.h>
21 #define DM_MSG_PREFIX "bufio"
24 * Memory management policy:
25 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
26 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
27 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
28 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
31 #define DM_BUFIO_MIN_BUFFERS 8
33 #define DM_BUFIO_MEMORY_PERCENT 2
34 #define DM_BUFIO_VMALLOC_PERCENT 25
35 #define DM_BUFIO_WRITEBACK_PERCENT 75
38 * Check buffer ages in this interval (seconds)
40 #define DM_BUFIO_WORK_TIMER_SECS 30
43 * Free buffers when they are older than this (seconds)
45 #define DM_BUFIO_DEFAULT_AGE_SECS 300
48 * The nr of bytes of cached data to keep around.
50 #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
53 * The number of bvec entries that are embedded directly in the buffer.
54 * If the chunk size is larger, dm-io is used to do the io.
56 #define DM_BUFIO_INLINE_VECS 16
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 rb_root buffer_tree
;
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,
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
];
153 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
155 struct stack_trace stack_trace
;
156 unsigned long stack_entries
[MAX_STACK
];
160 /*----------------------------------------------------------------*/
162 static struct kmem_cache
*dm_bufio_caches
[PAGE_SHIFT
- SECTOR_SHIFT
];
163 static char *dm_bufio_cache_names
[PAGE_SHIFT
- SECTOR_SHIFT
];
165 static inline int dm_bufio_cache_index(struct dm_bufio_client
*c
)
167 unsigned ret
= c
->blocks_per_page_bits
- 1;
169 BUG_ON(ret
>= ARRAY_SIZE(dm_bufio_caches
));
174 #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
175 #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
177 #define dm_bufio_in_request() (!!current->bio_list)
179 static void dm_bufio_lock(struct dm_bufio_client
*c
)
181 mutex_lock_nested(&c
->lock
, dm_bufio_in_request());
184 static int dm_bufio_trylock(struct dm_bufio_client
*c
)
186 return mutex_trylock(&c
->lock
);
189 static void dm_bufio_unlock(struct dm_bufio_client
*c
)
191 mutex_unlock(&c
->lock
);
194 /*----------------------------------------------------------------*/
197 * Default cache size: available memory divided by the ratio.
199 static unsigned long dm_bufio_default_cache_size
;
202 * Total cache size set by the user.
204 static unsigned long dm_bufio_cache_size
;
207 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
208 * at any time. If it disagrees, the user has changed cache size.
210 static unsigned long dm_bufio_cache_size_latch
;
212 static DEFINE_SPINLOCK(param_spinlock
);
215 * Buffers are freed after this timeout
217 static unsigned dm_bufio_max_age
= DM_BUFIO_DEFAULT_AGE_SECS
;
218 static unsigned dm_bufio_retain_bytes
= DM_BUFIO_DEFAULT_RETAIN_BYTES
;
220 static unsigned long dm_bufio_peak_allocated
;
221 static unsigned long dm_bufio_allocated_kmem_cache
;
222 static unsigned long dm_bufio_allocated_get_free_pages
;
223 static unsigned long dm_bufio_allocated_vmalloc
;
224 static unsigned long dm_bufio_current_allocated
;
226 /*----------------------------------------------------------------*/
229 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
231 static unsigned long dm_bufio_cache_size_per_client
;
234 * The current number of clients.
236 static int dm_bufio_client_count
;
239 * The list of all clients.
241 static LIST_HEAD(dm_bufio_all_clients
);
244 * This mutex protects dm_bufio_cache_size_latch,
245 * dm_bufio_cache_size_per_client and dm_bufio_client_count
247 static DEFINE_MUTEX(dm_bufio_clients_lock
);
249 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
250 static void buffer_record_stack(struct dm_buffer
*b
)
252 b
->stack_trace
.nr_entries
= 0;
253 b
->stack_trace
.max_entries
= MAX_STACK
;
254 b
->stack_trace
.entries
= b
->stack_entries
;
255 b
->stack_trace
.skip
= 2;
256 save_stack_trace(&b
->stack_trace
);
260 /*----------------------------------------------------------------
261 * A red/black tree acts as an index for all the buffers.
262 *--------------------------------------------------------------*/
263 static struct dm_buffer
*__find(struct dm_bufio_client
*c
, sector_t block
)
265 struct rb_node
*n
= c
->buffer_tree
.rb_node
;
269 b
= container_of(n
, struct dm_buffer
, node
);
271 if (b
->block
== block
)
274 n
= (b
->block
< block
) ? n
->rb_left
: n
->rb_right
;
280 static void __insert(struct dm_bufio_client
*c
, struct dm_buffer
*b
)
282 struct rb_node
**new = &c
->buffer_tree
.rb_node
, *parent
= NULL
;
283 struct dm_buffer
*found
;
286 found
= container_of(*new, struct dm_buffer
, node
);
288 if (found
->block
== b
->block
) {
294 new = (found
->block
< b
->block
) ?
295 &((*new)->rb_left
) : &((*new)->rb_right
);
298 rb_link_node(&b
->node
, parent
, new);
299 rb_insert_color(&b
->node
, &c
->buffer_tree
);
302 static void __remove(struct dm_bufio_client
*c
, struct dm_buffer
*b
)
304 rb_erase(&b
->node
, &c
->buffer_tree
);
307 /*----------------------------------------------------------------*/
309 static void adjust_total_allocated(enum data_mode data_mode
, long diff
)
311 static unsigned long * const class_ptr
[DATA_MODE_LIMIT
] = {
312 &dm_bufio_allocated_kmem_cache
,
313 &dm_bufio_allocated_get_free_pages
,
314 &dm_bufio_allocated_vmalloc
,
317 spin_lock(¶m_spinlock
);
319 *class_ptr
[data_mode
] += diff
;
321 dm_bufio_current_allocated
+= diff
;
323 if (dm_bufio_current_allocated
> dm_bufio_peak_allocated
)
324 dm_bufio_peak_allocated
= dm_bufio_current_allocated
;
326 spin_unlock(¶m_spinlock
);
330 * Change the number of clients and recalculate per-client limit.
332 static void __cache_size_refresh(void)
334 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock
));
335 BUG_ON(dm_bufio_client_count
< 0);
337 dm_bufio_cache_size_latch
= ACCESS_ONCE(dm_bufio_cache_size
);
340 * Use default if set to 0 and report the actual cache size used.
342 if (!dm_bufio_cache_size_latch
) {
343 (void)cmpxchg(&dm_bufio_cache_size
, 0,
344 dm_bufio_default_cache_size
);
345 dm_bufio_cache_size_latch
= dm_bufio_default_cache_size
;
348 dm_bufio_cache_size_per_client
= dm_bufio_cache_size_latch
/
349 (dm_bufio_client_count
? : 1);
353 * Allocating buffer data.
355 * Small buffers are allocated with kmem_cache, to use space optimally.
357 * For large buffers, we choose between get_free_pages and vmalloc.
358 * Each has advantages and disadvantages.
360 * __get_free_pages can randomly fail if the memory is fragmented.
361 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
362 * as low as 128M) so using it for caching is not appropriate.
364 * If the allocation may fail we use __get_free_pages. Memory fragmentation
365 * won't have a fatal effect here, but it just causes flushes of some other
366 * buffers and more I/O will be performed. Don't use __get_free_pages if it
367 * always fails (i.e. order >= MAX_ORDER).
369 * If the allocation shouldn't fail we use __vmalloc. This is only for the
370 * initial reserve allocation, so there's no risk of wasting all vmalloc
373 static void *alloc_buffer_data(struct dm_bufio_client
*c
, gfp_t gfp_mask
,
374 enum data_mode
*data_mode
)
379 if (c
->block_size
<= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT
) {
380 *data_mode
= DATA_MODE_SLAB
;
381 return kmem_cache_alloc(DM_BUFIO_CACHE(c
), gfp_mask
);
384 if (c
->block_size
<= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT
&&
385 gfp_mask
& __GFP_NORETRY
) {
386 *data_mode
= DATA_MODE_GET_FREE_PAGES
;
387 return (void *)__get_free_pages(gfp_mask
,
388 c
->pages_per_block_bits
);
391 *data_mode
= DATA_MODE_VMALLOC
;
394 * __vmalloc allocates the data pages and auxiliary structures with
395 * gfp_flags that were specified, but pagetables are always allocated
396 * with GFP_KERNEL, no matter what was specified as gfp_mask.
398 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
399 * all allocations done by this process (including pagetables) are done
400 * as if GFP_NOIO was specified.
403 if (gfp_mask
& __GFP_NORETRY
)
404 noio_flag
= memalloc_noio_save();
406 ptr
= __vmalloc(c
->block_size
, gfp_mask
| __GFP_HIGHMEM
, PAGE_KERNEL
);
408 if (gfp_mask
& __GFP_NORETRY
)
409 memalloc_noio_restore(noio_flag
);
415 * Free buffer's data.
417 static void free_buffer_data(struct dm_bufio_client
*c
,
418 void *data
, enum data_mode data_mode
)
422 kmem_cache_free(DM_BUFIO_CACHE(c
), data
);
425 case DATA_MODE_GET_FREE_PAGES
:
426 free_pages((unsigned long)data
, c
->pages_per_block_bits
);
429 case DATA_MODE_VMALLOC
:
434 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
441 * Allocate buffer and its data.
443 static struct dm_buffer
*alloc_buffer(struct dm_bufio_client
*c
, gfp_t gfp_mask
)
445 struct dm_buffer
*b
= kmalloc(sizeof(struct dm_buffer
) + c
->aux_size
,
453 b
->data
= alloc_buffer_data(c
, gfp_mask
, &b
->data_mode
);
459 adjust_total_allocated(b
->data_mode
, (long)c
->block_size
);
461 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
462 memset(&b
->stack_trace
, 0, sizeof(b
->stack_trace
));
468 * Free buffer and its data.
470 static void free_buffer(struct dm_buffer
*b
)
472 struct dm_bufio_client
*c
= b
->c
;
474 adjust_total_allocated(b
->data_mode
, -(long)c
->block_size
);
476 free_buffer_data(c
, b
->data
, b
->data_mode
);
481 * Link buffer to the hash list and clean or dirty queue.
483 static void __link_buffer(struct dm_buffer
*b
, sector_t block
, int dirty
)
485 struct dm_bufio_client
*c
= b
->c
;
487 c
->n_buffers
[dirty
]++;
489 b
->list_mode
= dirty
;
490 list_add(&b
->lru_list
, &c
->lru
[dirty
]);
492 b
->last_accessed
= jiffies
;
496 * Unlink buffer from the hash list and dirty or clean queue.
498 static void __unlink_buffer(struct dm_buffer
*b
)
500 struct dm_bufio_client
*c
= b
->c
;
502 BUG_ON(!c
->n_buffers
[b
->list_mode
]);
504 c
->n_buffers
[b
->list_mode
]--;
506 list_del(&b
->lru_list
);
510 * Place the buffer to the head of dirty or clean LRU queue.
512 static void __relink_lru(struct dm_buffer
*b
, int dirty
)
514 struct dm_bufio_client
*c
= b
->c
;
516 BUG_ON(!c
->n_buffers
[b
->list_mode
]);
518 c
->n_buffers
[b
->list_mode
]--;
519 c
->n_buffers
[dirty
]++;
520 b
->list_mode
= dirty
;
521 list_move(&b
->lru_list
, &c
->lru
[dirty
]);
522 b
->last_accessed
= jiffies
;
525 /*----------------------------------------------------------------
526 * Submit I/O on the buffer.
528 * Bio interface is faster but it has some problems:
529 * the vector list is limited (increasing this limit increases
530 * memory-consumption per buffer, so it is not viable);
532 * the memory must be direct-mapped, not vmalloced;
534 * the I/O driver can reject requests spuriously if it thinks that
535 * the requests are too big for the device or if they cross a
536 * controller-defined memory boundary.
538 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
539 * it is not vmalloced, try using the bio interface.
541 * If the buffer is big, if it is vmalloced or if the underlying device
542 * rejects the bio because it is too large, use dm-io layer to do the I/O.
543 * The dm-io layer splits the I/O into multiple requests, avoiding the above
545 *--------------------------------------------------------------*/
548 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
549 * that the request was handled directly with bio interface.
551 static void dmio_complete(unsigned long error
, void *context
)
553 struct dm_buffer
*b
= context
;
555 b
->bio
.bi_error
= error
? -EIO
: 0;
556 b
->bio
.bi_end_io(&b
->bio
);
559 static void use_dmio(struct dm_buffer
*b
, int rw
, sector_t block
,
560 bio_end_io_t
*end_io
)
563 struct dm_io_request io_req
= {
566 .notify
.fn
= dmio_complete
,
568 .client
= b
->c
->dm_io
,
570 struct dm_io_region region
= {
572 .sector
= block
<< b
->c
->sectors_per_block_bits
,
573 .count
= b
->c
->block_size
>> SECTOR_SHIFT
,
576 if (b
->data_mode
!= DATA_MODE_VMALLOC
) {
577 io_req
.mem
.type
= DM_IO_KMEM
;
578 io_req
.mem
.ptr
.addr
= b
->data
;
580 io_req
.mem
.type
= DM_IO_VMA
;
581 io_req
.mem
.ptr
.vma
= b
->data
;
584 b
->bio
.bi_end_io
= end_io
;
586 r
= dm_io(&io_req
, 1, ®ion
, NULL
);
593 static void inline_endio(struct bio
*bio
)
595 bio_end_io_t
*end_fn
= bio
->bi_private
;
596 int error
= bio
->bi_error
;
599 * Reset the bio to free any attached resources
600 * (e.g. bio integrity profiles).
604 bio
->bi_error
= error
;
608 static void use_inline_bio(struct dm_buffer
*b
, int rw
, sector_t block
,
609 bio_end_io_t
*end_io
)
614 bio_init(&b
->bio
, b
->bio_vec
, DM_BUFIO_INLINE_VECS
);
615 b
->bio
.bi_iter
.bi_sector
= block
<< b
->c
->sectors_per_block_bits
;
616 b
->bio
.bi_bdev
= b
->c
->bdev
;
617 b
->bio
.bi_end_io
= inline_endio
;
619 * Use of .bi_private isn't a problem here because
620 * the dm_buffer's inline bio is local to bufio.
622 b
->bio
.bi_private
= end_io
;
623 bio_set_op_attrs(&b
->bio
, rw
, 0);
626 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
627 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
630 len
= b
->c
->block_size
;
632 if (len
>= PAGE_SIZE
)
633 BUG_ON((unsigned long)ptr
& (PAGE_SIZE
- 1));
635 BUG_ON((unsigned long)ptr
& (len
- 1));
638 if (!bio_add_page(&b
->bio
, virt_to_page(ptr
),
639 len
< PAGE_SIZE
? len
: PAGE_SIZE
,
640 offset_in_page(ptr
))) {
641 BUG_ON(b
->c
->block_size
<= PAGE_SIZE
);
642 use_dmio(b
, rw
, block
, end_io
);
653 static void submit_io(struct dm_buffer
*b
, int rw
, sector_t block
,
654 bio_end_io_t
*end_io
)
656 if (rw
== WRITE
&& b
->c
->write_callback
)
657 b
->c
->write_callback(b
);
659 if (b
->c
->block_size
<= DM_BUFIO_INLINE_VECS
* PAGE_SIZE
&&
660 b
->data_mode
!= DATA_MODE_VMALLOC
)
661 use_inline_bio(b
, rw
, block
, end_io
);
663 use_dmio(b
, rw
, block
, end_io
);
666 /*----------------------------------------------------------------
667 * Writing dirty buffers
668 *--------------------------------------------------------------*/
671 * The endio routine for write.
673 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
676 static void write_endio(struct bio
*bio
)
678 struct dm_buffer
*b
= container_of(bio
, struct dm_buffer
, bio
);
680 b
->write_error
= bio
->bi_error
;
681 if (unlikely(bio
->bi_error
)) {
682 struct dm_bufio_client
*c
= b
->c
;
683 int error
= bio
->bi_error
;
684 (void)cmpxchg(&c
->async_write_error
, 0, error
);
687 BUG_ON(!test_bit(B_WRITING
, &b
->state
));
689 smp_mb__before_atomic();
690 clear_bit(B_WRITING
, &b
->state
);
691 smp_mb__after_atomic();
693 wake_up_bit(&b
->state
, B_WRITING
);
697 * Initiate a write on a dirty buffer, but don't wait for it.
699 * - If the buffer is not dirty, exit.
700 * - If there some previous write going on, wait for it to finish (we can't
701 * have two writes on the same buffer simultaneously).
702 * - Submit our write and don't wait on it. We set B_WRITING indicating
703 * that there is a write in progress.
705 static void __write_dirty_buffer(struct dm_buffer
*b
,
706 struct list_head
*write_list
)
708 if (!test_bit(B_DIRTY
, &b
->state
))
711 clear_bit(B_DIRTY
, &b
->state
);
712 wait_on_bit_lock_io(&b
->state
, B_WRITING
, TASK_UNINTERRUPTIBLE
);
715 submit_io(b
, WRITE
, b
->block
, write_endio
);
717 list_add_tail(&b
->write_list
, write_list
);
720 static void __flush_write_list(struct list_head
*write_list
)
722 struct blk_plug plug
;
723 blk_start_plug(&plug
);
724 while (!list_empty(write_list
)) {
725 struct dm_buffer
*b
=
726 list_entry(write_list
->next
, struct dm_buffer
, write_list
);
727 list_del(&b
->write_list
);
728 submit_io(b
, WRITE
, b
->block
, write_endio
);
731 blk_finish_plug(&plug
);
735 * Wait until any activity on the buffer finishes. Possibly write the
736 * buffer if it is dirty. When this function finishes, there is no I/O
737 * running on the buffer and the buffer is not dirty.
739 static void __make_buffer_clean(struct dm_buffer
*b
)
741 BUG_ON(b
->hold_count
);
743 if (!b
->state
) /* fast case */
746 wait_on_bit_io(&b
->state
, B_READING
, TASK_UNINTERRUPTIBLE
);
747 __write_dirty_buffer(b
, NULL
);
748 wait_on_bit_io(&b
->state
, B_WRITING
, TASK_UNINTERRUPTIBLE
);
752 * Find some buffer that is not held by anybody, clean it, unlink it and
755 static struct dm_buffer
*__get_unclaimed_buffer(struct dm_bufio_client
*c
)
759 list_for_each_entry_reverse(b
, &c
->lru
[LIST_CLEAN
], lru_list
) {
760 BUG_ON(test_bit(B_WRITING
, &b
->state
));
761 BUG_ON(test_bit(B_DIRTY
, &b
->state
));
763 if (!b
->hold_count
) {
764 __make_buffer_clean(b
);
771 list_for_each_entry_reverse(b
, &c
->lru
[LIST_DIRTY
], lru_list
) {
772 BUG_ON(test_bit(B_READING
, &b
->state
));
774 if (!b
->hold_count
) {
775 __make_buffer_clean(b
);
786 * Wait until some other threads free some buffer or release hold count on
789 * This function is entered with c->lock held, drops it and regains it
792 static void __wait_for_free_buffer(struct dm_bufio_client
*c
)
794 DECLARE_WAITQUEUE(wait
, current
);
796 add_wait_queue(&c
->free_buffer_wait
, &wait
);
797 set_task_state(current
, TASK_UNINTERRUPTIBLE
);
802 remove_wait_queue(&c
->free_buffer_wait
, &wait
);
815 * Allocate a new buffer. If the allocation is not possible, wait until
816 * some other thread frees a buffer.
818 * May drop the lock and regain it.
820 static struct dm_buffer
*__alloc_buffer_wait_no_callback(struct dm_bufio_client
*c
, enum new_flag nf
)
825 * dm-bufio is resistant to allocation failures (it just keeps
826 * one buffer reserved in cases all the allocations fail).
827 * So set flags to not try too hard:
828 * GFP_NOIO: don't recurse into the I/O layer
829 * __GFP_NORETRY: don't retry and rather return failure
830 * __GFP_NOMEMALLOC: don't use emergency reserves
831 * __GFP_NOWARN: don't print a warning in case of failure
833 * For debugging, if we set the cache size to 1, no new buffers will
837 if (dm_bufio_cache_size_latch
!= 1) {
838 b
= alloc_buffer(c
, GFP_NOIO
| __GFP_NORETRY
| __GFP_NOMEMALLOC
| __GFP_NOWARN
);
843 if (nf
== NF_PREFETCH
)
846 if (!list_empty(&c
->reserved_buffers
)) {
847 b
= list_entry(c
->reserved_buffers
.next
,
848 struct dm_buffer
, lru_list
);
849 list_del(&b
->lru_list
);
850 c
->need_reserved_buffers
++;
855 b
= __get_unclaimed_buffer(c
);
859 __wait_for_free_buffer(c
);
863 static struct dm_buffer
*__alloc_buffer_wait(struct dm_bufio_client
*c
, enum new_flag nf
)
865 struct dm_buffer
*b
= __alloc_buffer_wait_no_callback(c
, nf
);
870 if (c
->alloc_callback
)
871 c
->alloc_callback(b
);
877 * Free a buffer and wake other threads waiting for free buffers.
879 static void __free_buffer_wake(struct dm_buffer
*b
)
881 struct dm_bufio_client
*c
= b
->c
;
883 if (!c
->need_reserved_buffers
)
886 list_add(&b
->lru_list
, &c
->reserved_buffers
);
887 c
->need_reserved_buffers
--;
890 wake_up(&c
->free_buffer_wait
);
893 static void __write_dirty_buffers_async(struct dm_bufio_client
*c
, int no_wait
,
894 struct list_head
*write_list
)
896 struct dm_buffer
*b
, *tmp
;
898 list_for_each_entry_safe_reverse(b
, tmp
, &c
->lru
[LIST_DIRTY
], lru_list
) {
899 BUG_ON(test_bit(B_READING
, &b
->state
));
901 if (!test_bit(B_DIRTY
, &b
->state
) &&
902 !test_bit(B_WRITING
, &b
->state
)) {
903 __relink_lru(b
, LIST_CLEAN
);
907 if (no_wait
&& test_bit(B_WRITING
, &b
->state
))
910 __write_dirty_buffer(b
, write_list
);
916 * Get writeback threshold and buffer limit for a given client.
918 static void __get_memory_limit(struct dm_bufio_client
*c
,
919 unsigned long *threshold_buffers
,
920 unsigned long *limit_buffers
)
922 unsigned long buffers
;
924 if (ACCESS_ONCE(dm_bufio_cache_size
) != dm_bufio_cache_size_latch
) {
925 mutex_lock(&dm_bufio_clients_lock
);
926 __cache_size_refresh();
927 mutex_unlock(&dm_bufio_clients_lock
);
930 buffers
= dm_bufio_cache_size_per_client
>>
931 (c
->sectors_per_block_bits
+ SECTOR_SHIFT
);
933 if (buffers
< c
->minimum_buffers
)
934 buffers
= c
->minimum_buffers
;
936 *limit_buffers
= buffers
;
937 *threshold_buffers
= buffers
* DM_BUFIO_WRITEBACK_PERCENT
/ 100;
941 * Check if we're over watermark.
942 * If we are over threshold_buffers, start freeing buffers.
943 * If we're over "limit_buffers", block until we get under the limit.
945 static void __check_watermark(struct dm_bufio_client
*c
,
946 struct list_head
*write_list
)
948 unsigned long threshold_buffers
, limit_buffers
;
950 __get_memory_limit(c
, &threshold_buffers
, &limit_buffers
);
952 while (c
->n_buffers
[LIST_CLEAN
] + c
->n_buffers
[LIST_DIRTY
] >
955 struct dm_buffer
*b
= __get_unclaimed_buffer(c
);
960 __free_buffer_wake(b
);
964 if (c
->n_buffers
[LIST_DIRTY
] > threshold_buffers
)
965 __write_dirty_buffers_async(c
, 1, write_list
);
968 /*----------------------------------------------------------------
970 *--------------------------------------------------------------*/
972 static struct dm_buffer
*__bufio_new(struct dm_bufio_client
*c
, sector_t block
,
973 enum new_flag nf
, int *need_submit
,
974 struct list_head
*write_list
)
976 struct dm_buffer
*b
, *new_b
= NULL
;
980 b
= __find(c
, block
);
987 new_b
= __alloc_buffer_wait(c
, nf
);
992 * We've had a period where the mutex was unlocked, so need to
993 * recheck the hash table.
995 b
= __find(c
, block
);
997 __free_buffer_wake(new_b
);
1001 __check_watermark(c
, write_list
);
1007 __link_buffer(b
, block
, LIST_CLEAN
);
1009 if (nf
== NF_FRESH
) {
1014 b
->state
= 1 << B_READING
;
1020 if (nf
== NF_PREFETCH
)
1023 * Note: it is essential that we don't wait for the buffer to be
1024 * read if dm_bufio_get function is used. Both dm_bufio_get and
1025 * dm_bufio_prefetch can be used in the driver request routine.
1026 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1027 * the same buffer, it would deadlock if we waited.
1029 if (nf
== NF_GET
&& unlikely(test_bit(B_READING
, &b
->state
)))
1033 __relink_lru(b
, test_bit(B_DIRTY
, &b
->state
) ||
1034 test_bit(B_WRITING
, &b
->state
));
1039 * The endio routine for reading: set the error, clear the bit and wake up
1040 * anyone waiting on the buffer.
1042 static void read_endio(struct bio
*bio
)
1044 struct dm_buffer
*b
= container_of(bio
, struct dm_buffer
, bio
);
1046 b
->read_error
= bio
->bi_error
;
1048 BUG_ON(!test_bit(B_READING
, &b
->state
));
1050 smp_mb__before_atomic();
1051 clear_bit(B_READING
, &b
->state
);
1052 smp_mb__after_atomic();
1054 wake_up_bit(&b
->state
, B_READING
);
1058 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1059 * functions is similar except that dm_bufio_new doesn't read the
1060 * buffer from the disk (assuming that the caller overwrites all the data
1061 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1063 static void *new_read(struct dm_bufio_client
*c
, sector_t block
,
1064 enum new_flag nf
, struct dm_buffer
**bp
)
1067 struct dm_buffer
*b
;
1069 LIST_HEAD(write_list
);
1072 b
= __bufio_new(c
, block
, nf
, &need_submit
, &write_list
);
1073 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1074 if (b
&& b
->hold_count
== 1)
1075 buffer_record_stack(b
);
1079 __flush_write_list(&write_list
);
1085 submit_io(b
, READ
, b
->block
, read_endio
);
1087 wait_on_bit_io(&b
->state
, B_READING
, TASK_UNINTERRUPTIBLE
);
1089 if (b
->read_error
) {
1090 int error
= b
->read_error
;
1092 dm_bufio_release(b
);
1094 return ERR_PTR(error
);
1102 void *dm_bufio_get(struct dm_bufio_client
*c
, sector_t block
,
1103 struct dm_buffer
**bp
)
1105 return new_read(c
, block
, NF_GET
, bp
);
1107 EXPORT_SYMBOL_GPL(dm_bufio_get
);
1109 void *dm_bufio_read(struct dm_bufio_client
*c
, sector_t block
,
1110 struct dm_buffer
**bp
)
1112 BUG_ON(dm_bufio_in_request());
1114 return new_read(c
, block
, NF_READ
, bp
);
1116 EXPORT_SYMBOL_GPL(dm_bufio_read
);
1118 void *dm_bufio_new(struct dm_bufio_client
*c
, sector_t block
,
1119 struct dm_buffer
**bp
)
1121 BUG_ON(dm_bufio_in_request());
1123 return new_read(c
, block
, NF_FRESH
, bp
);
1125 EXPORT_SYMBOL_GPL(dm_bufio_new
);
1127 void dm_bufio_prefetch(struct dm_bufio_client
*c
,
1128 sector_t block
, unsigned n_blocks
)
1130 struct blk_plug plug
;
1132 LIST_HEAD(write_list
);
1134 BUG_ON(dm_bufio_in_request());
1136 blk_start_plug(&plug
);
1139 for (; n_blocks
--; block
++) {
1141 struct dm_buffer
*b
;
1142 b
= __bufio_new(c
, block
, NF_PREFETCH
, &need_submit
,
1144 if (unlikely(!list_empty(&write_list
))) {
1146 blk_finish_plug(&plug
);
1147 __flush_write_list(&write_list
);
1148 blk_start_plug(&plug
);
1151 if (unlikely(b
!= NULL
)) {
1155 submit_io(b
, READ
, b
->block
, read_endio
);
1156 dm_bufio_release(b
);
1169 blk_finish_plug(&plug
);
1171 EXPORT_SYMBOL_GPL(dm_bufio_prefetch
);
1173 void dm_bufio_release(struct dm_buffer
*b
)
1175 struct dm_bufio_client
*c
= b
->c
;
1179 BUG_ON(!b
->hold_count
);
1182 if (!b
->hold_count
) {
1183 wake_up(&c
->free_buffer_wait
);
1186 * If there were errors on the buffer, and the buffer is not
1187 * to be written, free the buffer. There is no point in caching
1190 if ((b
->read_error
|| b
->write_error
) &&
1191 !test_bit(B_READING
, &b
->state
) &&
1192 !test_bit(B_WRITING
, &b
->state
) &&
1193 !test_bit(B_DIRTY
, &b
->state
)) {
1195 __free_buffer_wake(b
);
1201 EXPORT_SYMBOL_GPL(dm_bufio_release
);
1203 void dm_bufio_mark_buffer_dirty(struct dm_buffer
*b
)
1205 struct dm_bufio_client
*c
= b
->c
;
1209 BUG_ON(test_bit(B_READING
, &b
->state
));
1211 if (!test_and_set_bit(B_DIRTY
, &b
->state
))
1212 __relink_lru(b
, LIST_DIRTY
);
1216 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty
);
1218 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client
*c
)
1220 LIST_HEAD(write_list
);
1222 BUG_ON(dm_bufio_in_request());
1225 __write_dirty_buffers_async(c
, 0, &write_list
);
1227 __flush_write_list(&write_list
);
1229 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async
);
1232 * For performance, it is essential that the buffers are written asynchronously
1233 * and simultaneously (so that the block layer can merge the writes) and then
1236 * Finally, we flush hardware disk cache.
1238 int dm_bufio_write_dirty_buffers(struct dm_bufio_client
*c
)
1241 unsigned long buffers_processed
= 0;
1242 struct dm_buffer
*b
, *tmp
;
1244 LIST_HEAD(write_list
);
1247 __write_dirty_buffers_async(c
, 0, &write_list
);
1249 __flush_write_list(&write_list
);
1253 list_for_each_entry_safe_reverse(b
, tmp
, &c
->lru
[LIST_DIRTY
], lru_list
) {
1254 int dropped_lock
= 0;
1256 if (buffers_processed
< c
->n_buffers
[LIST_DIRTY
])
1257 buffers_processed
++;
1259 BUG_ON(test_bit(B_READING
, &b
->state
));
1261 if (test_bit(B_WRITING
, &b
->state
)) {
1262 if (buffers_processed
< c
->n_buffers
[LIST_DIRTY
]) {
1266 wait_on_bit_io(&b
->state
, B_WRITING
,
1267 TASK_UNINTERRUPTIBLE
);
1271 wait_on_bit_io(&b
->state
, B_WRITING
,
1272 TASK_UNINTERRUPTIBLE
);
1275 if (!test_bit(B_DIRTY
, &b
->state
) &&
1276 !test_bit(B_WRITING
, &b
->state
))
1277 __relink_lru(b
, LIST_CLEAN
);
1282 * If we dropped the lock, the list is no longer consistent,
1283 * so we must restart the search.
1285 * In the most common case, the buffer just processed is
1286 * relinked to the clean list, so we won't loop scanning the
1287 * same buffer again and again.
1289 * This may livelock if there is another thread simultaneously
1290 * dirtying buffers, so we count the number of buffers walked
1291 * and if it exceeds the total number of buffers, it means that
1292 * someone is doing some writes simultaneously with us. In
1293 * this case, stop, dropping the lock.
1298 wake_up(&c
->free_buffer_wait
);
1301 a
= xchg(&c
->async_write_error
, 0);
1302 f
= dm_bufio_issue_flush(c
);
1308 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers
);
1311 * Use dm-io to send and empty barrier flush the device.
1313 int dm_bufio_issue_flush(struct dm_bufio_client
*c
)
1315 struct dm_io_request io_req
= {
1316 .bi_op
= REQ_OP_WRITE
,
1317 .bi_op_flags
= REQ_PREFLUSH
,
1318 .mem
.type
= DM_IO_KMEM
,
1319 .mem
.ptr
.addr
= NULL
,
1322 struct dm_io_region io_reg
= {
1328 BUG_ON(dm_bufio_in_request());
1330 return dm_io(&io_req
, 1, &io_reg
, NULL
);
1332 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush
);
1335 * We first delete any other buffer that may be at that new location.
1337 * Then, we write the buffer to the original location if it was dirty.
1339 * Then, if we are the only one who is holding the buffer, relink the buffer
1340 * in the hash queue for the new location.
1342 * If there was someone else holding the buffer, we write it to the new
1343 * location but not relink it, because that other user needs to have the buffer
1344 * at the same place.
1346 void dm_bufio_release_move(struct dm_buffer
*b
, sector_t new_block
)
1348 struct dm_bufio_client
*c
= b
->c
;
1349 struct dm_buffer
*new;
1351 BUG_ON(dm_bufio_in_request());
1356 new = __find(c
, new_block
);
1358 if (new->hold_count
) {
1359 __wait_for_free_buffer(c
);
1364 * FIXME: Is there any point waiting for a write that's going
1365 * to be overwritten in a bit?
1367 __make_buffer_clean(new);
1368 __unlink_buffer(new);
1369 __free_buffer_wake(new);
1372 BUG_ON(!b
->hold_count
);
1373 BUG_ON(test_bit(B_READING
, &b
->state
));
1375 __write_dirty_buffer(b
, NULL
);
1376 if (b
->hold_count
== 1) {
1377 wait_on_bit_io(&b
->state
, B_WRITING
,
1378 TASK_UNINTERRUPTIBLE
);
1379 set_bit(B_DIRTY
, &b
->state
);
1381 __link_buffer(b
, new_block
, LIST_DIRTY
);
1384 wait_on_bit_lock_io(&b
->state
, B_WRITING
,
1385 TASK_UNINTERRUPTIBLE
);
1387 * Relink buffer to "new_block" so that write_callback
1388 * sees "new_block" as a block number.
1389 * After the write, link the buffer back to old_block.
1390 * All this must be done in bufio lock, so that block number
1391 * change isn't visible to other threads.
1393 old_block
= b
->block
;
1395 __link_buffer(b
, new_block
, b
->list_mode
);
1396 submit_io(b
, WRITE
, new_block
, write_endio
);
1397 wait_on_bit_io(&b
->state
, B_WRITING
,
1398 TASK_UNINTERRUPTIBLE
);
1400 __link_buffer(b
, old_block
, b
->list_mode
);
1404 dm_bufio_release(b
);
1406 EXPORT_SYMBOL_GPL(dm_bufio_release_move
);
1409 * Free the given buffer.
1411 * This is just a hint, if the buffer is in use or dirty, this function
1414 void dm_bufio_forget(struct dm_bufio_client
*c
, sector_t block
)
1416 struct dm_buffer
*b
;
1420 b
= __find(c
, block
);
1421 if (b
&& likely(!b
->hold_count
) && likely(!b
->state
)) {
1423 __free_buffer_wake(b
);
1428 EXPORT_SYMBOL(dm_bufio_forget
);
1430 void dm_bufio_set_minimum_buffers(struct dm_bufio_client
*c
, unsigned n
)
1432 c
->minimum_buffers
= n
;
1434 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers
);
1436 unsigned dm_bufio_get_block_size(struct dm_bufio_client
*c
)
1438 return c
->block_size
;
1440 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size
);
1442 sector_t
dm_bufio_get_device_size(struct dm_bufio_client
*c
)
1444 return i_size_read(c
->bdev
->bd_inode
) >>
1445 (SECTOR_SHIFT
+ c
->sectors_per_block_bits
);
1447 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size
);
1449 sector_t
dm_bufio_get_block_number(struct dm_buffer
*b
)
1453 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number
);
1455 void *dm_bufio_get_block_data(struct dm_buffer
*b
)
1459 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data
);
1461 void *dm_bufio_get_aux_data(struct dm_buffer
*b
)
1465 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data
);
1467 struct dm_bufio_client
*dm_bufio_get_client(struct dm_buffer
*b
)
1471 EXPORT_SYMBOL_GPL(dm_bufio_get_client
);
1473 static void drop_buffers(struct dm_bufio_client
*c
)
1475 struct dm_buffer
*b
;
1477 bool warned
= false;
1479 BUG_ON(dm_bufio_in_request());
1482 * An optimization so that the buffers are not written one-by-one.
1484 dm_bufio_write_dirty_buffers_async(c
);
1488 while ((b
= __get_unclaimed_buffer(c
)))
1489 __free_buffer_wake(b
);
1491 for (i
= 0; i
< LIST_SIZE
; i
++)
1492 list_for_each_entry(b
, &c
->lru
[i
], lru_list
) {
1495 DMERR("leaked buffer %llx, hold count %u, list %d",
1496 (unsigned long long)b
->block
, b
->hold_count
, i
);
1497 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1498 print_stack_trace(&b
->stack_trace
, 1);
1499 b
->hold_count
= 0; /* mark unclaimed to avoid BUG_ON below */
1503 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1504 while ((b
= __get_unclaimed_buffer(c
)))
1505 __free_buffer_wake(b
);
1508 for (i
= 0; i
< LIST_SIZE
; i
++)
1509 BUG_ON(!list_empty(&c
->lru
[i
]));
1515 * We may not be able to evict this buffer if IO pending or the client
1516 * is still using it. Caller is expected to know buffer is too old.
1518 * And if GFP_NOFS is used, we must not do any I/O because we hold
1519 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1520 * rerouted to different bufio client.
1522 static bool __try_evict_buffer(struct dm_buffer
*b
, gfp_t gfp
)
1524 if (!(gfp
& __GFP_FS
)) {
1525 if (test_bit(B_READING
, &b
->state
) ||
1526 test_bit(B_WRITING
, &b
->state
) ||
1527 test_bit(B_DIRTY
, &b
->state
))
1534 __make_buffer_clean(b
);
1536 __free_buffer_wake(b
);
1541 static unsigned get_retain_buffers(struct dm_bufio_client
*c
)
1543 unsigned retain_bytes
= ACCESS_ONCE(dm_bufio_retain_bytes
);
1544 return retain_bytes
/ c
->block_size
;
1547 static unsigned long __scan(struct dm_bufio_client
*c
, unsigned long nr_to_scan
,
1551 struct dm_buffer
*b
, *tmp
;
1552 unsigned long freed
= 0;
1553 unsigned long count
= nr_to_scan
;
1554 unsigned retain_target
= get_retain_buffers(c
);
1556 for (l
= 0; l
< LIST_SIZE
; l
++) {
1557 list_for_each_entry_safe_reverse(b
, tmp
, &c
->lru
[l
], lru_list
) {
1558 if (__try_evict_buffer(b
, gfp_mask
))
1560 if (!--nr_to_scan
|| ((count
- freed
) <= retain_target
))
1568 static unsigned long
1569 dm_bufio_shrink_scan(struct shrinker
*shrink
, struct shrink_control
*sc
)
1571 struct dm_bufio_client
*c
;
1572 unsigned long freed
;
1574 c
= container_of(shrink
, struct dm_bufio_client
, shrinker
);
1575 if (sc
->gfp_mask
& __GFP_FS
)
1577 else if (!dm_bufio_trylock(c
))
1580 freed
= __scan(c
, sc
->nr_to_scan
, sc
->gfp_mask
);
1585 static unsigned long
1586 dm_bufio_shrink_count(struct shrinker
*shrink
, struct shrink_control
*sc
)
1588 struct dm_bufio_client
*c
;
1589 unsigned long count
;
1591 c
= container_of(shrink
, struct dm_bufio_client
, shrinker
);
1592 if (sc
->gfp_mask
& __GFP_FS
)
1594 else if (!dm_bufio_trylock(c
))
1597 count
= c
->n_buffers
[LIST_CLEAN
] + c
->n_buffers
[LIST_DIRTY
];
1603 * Create the buffering interface
1605 struct dm_bufio_client
*dm_bufio_client_create(struct block_device
*bdev
, unsigned block_size
,
1606 unsigned reserved_buffers
, unsigned aux_size
,
1607 void (*alloc_callback
)(struct dm_buffer
*),
1608 void (*write_callback
)(struct dm_buffer
*))
1611 struct dm_bufio_client
*c
;
1614 BUG_ON(block_size
< 1 << SECTOR_SHIFT
||
1615 (block_size
& (block_size
- 1)));
1617 c
= kzalloc(sizeof(*c
), GFP_KERNEL
);
1622 c
->buffer_tree
= RB_ROOT
;
1625 c
->block_size
= block_size
;
1626 c
->sectors_per_block_bits
= __ffs(block_size
) - SECTOR_SHIFT
;
1627 c
->pages_per_block_bits
= (__ffs(block_size
) >= PAGE_SHIFT
) ?
1628 __ffs(block_size
) - PAGE_SHIFT
: 0;
1629 c
->blocks_per_page_bits
= (__ffs(block_size
) < PAGE_SHIFT
?
1630 PAGE_SHIFT
- __ffs(block_size
) : 0);
1632 c
->aux_size
= aux_size
;
1633 c
->alloc_callback
= alloc_callback
;
1634 c
->write_callback
= write_callback
;
1636 for (i
= 0; i
< LIST_SIZE
; i
++) {
1637 INIT_LIST_HEAD(&c
->lru
[i
]);
1638 c
->n_buffers
[i
] = 0;
1641 mutex_init(&c
->lock
);
1642 INIT_LIST_HEAD(&c
->reserved_buffers
);
1643 c
->need_reserved_buffers
= reserved_buffers
;
1645 c
->minimum_buffers
= DM_BUFIO_MIN_BUFFERS
;
1647 init_waitqueue_head(&c
->free_buffer_wait
);
1648 c
->async_write_error
= 0;
1650 c
->dm_io
= dm_io_client_create();
1651 if (IS_ERR(c
->dm_io
)) {
1652 r
= PTR_ERR(c
->dm_io
);
1656 mutex_lock(&dm_bufio_clients_lock
);
1657 if (c
->blocks_per_page_bits
) {
1658 if (!DM_BUFIO_CACHE_NAME(c
)) {
1659 DM_BUFIO_CACHE_NAME(c
) = kasprintf(GFP_KERNEL
, "dm_bufio_cache-%u", c
->block_size
);
1660 if (!DM_BUFIO_CACHE_NAME(c
)) {
1662 mutex_unlock(&dm_bufio_clients_lock
);
1667 if (!DM_BUFIO_CACHE(c
)) {
1668 DM_BUFIO_CACHE(c
) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c
),
1670 c
->block_size
, 0, NULL
);
1671 if (!DM_BUFIO_CACHE(c
)) {
1673 mutex_unlock(&dm_bufio_clients_lock
);
1678 mutex_unlock(&dm_bufio_clients_lock
);
1680 while (c
->need_reserved_buffers
) {
1681 struct dm_buffer
*b
= alloc_buffer(c
, GFP_KERNEL
);
1687 __free_buffer_wake(b
);
1690 mutex_lock(&dm_bufio_clients_lock
);
1691 dm_bufio_client_count
++;
1692 list_add(&c
->client_list
, &dm_bufio_all_clients
);
1693 __cache_size_refresh();
1694 mutex_unlock(&dm_bufio_clients_lock
);
1696 c
->shrinker
.count_objects
= dm_bufio_shrink_count
;
1697 c
->shrinker
.scan_objects
= dm_bufio_shrink_scan
;
1698 c
->shrinker
.seeks
= 1;
1699 c
->shrinker
.batch
= 0;
1700 register_shrinker(&c
->shrinker
);
1706 while (!list_empty(&c
->reserved_buffers
)) {
1707 struct dm_buffer
*b
= list_entry(c
->reserved_buffers
.next
,
1708 struct dm_buffer
, lru_list
);
1709 list_del(&b
->lru_list
);
1712 dm_io_client_destroy(c
->dm_io
);
1718 EXPORT_SYMBOL_GPL(dm_bufio_client_create
);
1721 * Free the buffering interface.
1722 * It is required that there are no references on any buffers.
1724 void dm_bufio_client_destroy(struct dm_bufio_client
*c
)
1730 unregister_shrinker(&c
->shrinker
);
1732 mutex_lock(&dm_bufio_clients_lock
);
1734 list_del(&c
->client_list
);
1735 dm_bufio_client_count
--;
1736 __cache_size_refresh();
1738 mutex_unlock(&dm_bufio_clients_lock
);
1740 BUG_ON(!RB_EMPTY_ROOT(&c
->buffer_tree
));
1741 BUG_ON(c
->need_reserved_buffers
);
1743 while (!list_empty(&c
->reserved_buffers
)) {
1744 struct dm_buffer
*b
= list_entry(c
->reserved_buffers
.next
,
1745 struct dm_buffer
, lru_list
);
1746 list_del(&b
->lru_list
);
1750 for (i
= 0; i
< LIST_SIZE
; i
++)
1751 if (c
->n_buffers
[i
])
1752 DMERR("leaked buffer count %d: %ld", i
, c
->n_buffers
[i
]);
1754 for (i
= 0; i
< LIST_SIZE
; i
++)
1755 BUG_ON(c
->n_buffers
[i
]);
1757 dm_io_client_destroy(c
->dm_io
);
1760 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy
);
1762 static unsigned get_max_age_hz(void)
1764 unsigned max_age
= ACCESS_ONCE(dm_bufio_max_age
);
1766 if (max_age
> UINT_MAX
/ HZ
)
1767 max_age
= UINT_MAX
/ HZ
;
1769 return max_age
* HZ
;
1772 static bool older_than(struct dm_buffer
*b
, unsigned long age_hz
)
1774 return time_after_eq(jiffies
, b
->last_accessed
+ age_hz
);
1777 static void __evict_old_buffers(struct dm_bufio_client
*c
, unsigned long age_hz
)
1779 struct dm_buffer
*b
, *tmp
;
1780 unsigned retain_target
= get_retain_buffers(c
);
1785 count
= c
->n_buffers
[LIST_CLEAN
] + c
->n_buffers
[LIST_DIRTY
];
1786 list_for_each_entry_safe_reverse(b
, tmp
, &c
->lru
[LIST_CLEAN
], lru_list
) {
1787 if (count
<= retain_target
)
1790 if (!older_than(b
, age_hz
))
1793 if (__try_evict_buffer(b
, 0))
1802 static void cleanup_old_buffers(void)
1804 unsigned long max_age_hz
= get_max_age_hz();
1805 struct dm_bufio_client
*c
;
1807 mutex_lock(&dm_bufio_clients_lock
);
1809 list_for_each_entry(c
, &dm_bufio_all_clients
, client_list
)
1810 __evict_old_buffers(c
, max_age_hz
);
1812 mutex_unlock(&dm_bufio_clients_lock
);
1815 static struct workqueue_struct
*dm_bufio_wq
;
1816 static struct delayed_work dm_bufio_work
;
1818 static void work_fn(struct work_struct
*w
)
1820 cleanup_old_buffers();
1822 queue_delayed_work(dm_bufio_wq
, &dm_bufio_work
,
1823 DM_BUFIO_WORK_TIMER_SECS
* HZ
);
1826 /*----------------------------------------------------------------
1828 *--------------------------------------------------------------*/
1831 * This is called only once for the whole dm_bufio module.
1832 * It initializes memory limit.
1834 static int __init
dm_bufio_init(void)
1838 dm_bufio_allocated_kmem_cache
= 0;
1839 dm_bufio_allocated_get_free_pages
= 0;
1840 dm_bufio_allocated_vmalloc
= 0;
1841 dm_bufio_current_allocated
= 0;
1843 memset(&dm_bufio_caches
, 0, sizeof dm_bufio_caches
);
1844 memset(&dm_bufio_cache_names
, 0, sizeof dm_bufio_cache_names
);
1846 mem
= (__u64
)((totalram_pages
- totalhigh_pages
) *
1847 DM_BUFIO_MEMORY_PERCENT
/ 100) << PAGE_SHIFT
;
1849 if (mem
> ULONG_MAX
)
1854 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1855 * in fs/proc/internal.h
1857 if (mem
> (VMALLOC_END
- VMALLOC_START
) * DM_BUFIO_VMALLOC_PERCENT
/ 100)
1858 mem
= (VMALLOC_END
- VMALLOC_START
) * DM_BUFIO_VMALLOC_PERCENT
/ 100;
1861 dm_bufio_default_cache_size
= mem
;
1863 mutex_lock(&dm_bufio_clients_lock
);
1864 __cache_size_refresh();
1865 mutex_unlock(&dm_bufio_clients_lock
);
1867 dm_bufio_wq
= alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM
, 0);
1871 INIT_DELAYED_WORK(&dm_bufio_work
, work_fn
);
1872 queue_delayed_work(dm_bufio_wq
, &dm_bufio_work
,
1873 DM_BUFIO_WORK_TIMER_SECS
* HZ
);
1879 * This is called once when unloading the dm_bufio module.
1881 static void __exit
dm_bufio_exit(void)
1886 cancel_delayed_work_sync(&dm_bufio_work
);
1887 destroy_workqueue(dm_bufio_wq
);
1889 for (i
= 0; i
< ARRAY_SIZE(dm_bufio_caches
); i
++)
1890 kmem_cache_destroy(dm_bufio_caches
[i
]);
1892 for (i
= 0; i
< ARRAY_SIZE(dm_bufio_cache_names
); i
++)
1893 kfree(dm_bufio_cache_names
[i
]);
1895 if (dm_bufio_client_count
) {
1896 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1897 __func__
, dm_bufio_client_count
);
1901 if (dm_bufio_current_allocated
) {
1902 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1903 __func__
, dm_bufio_current_allocated
);
1907 if (dm_bufio_allocated_get_free_pages
) {
1908 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1909 __func__
, dm_bufio_allocated_get_free_pages
);
1913 if (dm_bufio_allocated_vmalloc
) {
1914 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1915 __func__
, dm_bufio_allocated_vmalloc
);
1922 module_init(dm_bufio_init
)
1923 module_exit(dm_bufio_exit
)
1925 module_param_named(max_cache_size_bytes
, dm_bufio_cache_size
, ulong
, S_IRUGO
| S_IWUSR
);
1926 MODULE_PARM_DESC(max_cache_size_bytes
, "Size of metadata cache");
1928 module_param_named(max_age_seconds
, dm_bufio_max_age
, uint
, S_IRUGO
| S_IWUSR
);
1929 MODULE_PARM_DESC(max_age_seconds
, "Max age of a buffer in seconds");
1931 module_param_named(retain_bytes
, dm_bufio_retain_bytes
, uint
, S_IRUGO
| S_IWUSR
);
1932 MODULE_PARM_DESC(retain_bytes
, "Try to keep at least this many bytes cached in memory");
1934 module_param_named(peak_allocated_bytes
, dm_bufio_peak_allocated
, ulong
, S_IRUGO
| S_IWUSR
);
1935 MODULE_PARM_DESC(peak_allocated_bytes
, "Tracks the maximum allocated memory");
1937 module_param_named(allocated_kmem_cache_bytes
, dm_bufio_allocated_kmem_cache
, ulong
, S_IRUGO
);
1938 MODULE_PARM_DESC(allocated_kmem_cache_bytes
, "Memory allocated with kmem_cache_alloc");
1940 module_param_named(allocated_get_free_pages_bytes
, dm_bufio_allocated_get_free_pages
, ulong
, S_IRUGO
);
1941 MODULE_PARM_DESC(allocated_get_free_pages_bytes
, "Memory allocated with get_free_pages");
1943 module_param_named(allocated_vmalloc_bytes
, dm_bufio_allocated_vmalloc
, ulong
, S_IRUGO
);
1944 MODULE_PARM_DESC(allocated_vmalloc_bytes
, "Memory allocated with vmalloc");
1946 module_param_named(current_allocated_bytes
, dm_bufio_current_allocated
, ulong
, S_IRUGO
);
1947 MODULE_PARM_DESC(current_allocated_bytes
, "Memory currently used by the cache");
1949 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1950 MODULE_DESCRIPTION(DM_NAME
" buffered I/O library");
1951 MODULE_LICENSE("GPL");