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/sched/mm.h>
15 #include <linux/jiffies.h>
16 #include <linux/vmalloc.h>
17 #include <linux/shrinker.h>
18 #include <linux/module.h>
19 #include <linux/rbtree.h>
20 #include <linux/stacktrace.h>
22 #define DM_MSG_PREFIX "bufio"
25 * Memory management policy:
26 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
27 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
28 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
29 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
32 #define DM_BUFIO_MIN_BUFFERS 8
34 #define DM_BUFIO_MEMORY_PERCENT 2
35 #define DM_BUFIO_VMALLOC_PERCENT 25
36 #define DM_BUFIO_WRITEBACK_PERCENT 75
39 * Check buffer ages in this interval (seconds)
41 #define DM_BUFIO_WORK_TIMER_SECS 30
44 * Free buffers when they are older than this (seconds)
46 #define DM_BUFIO_DEFAULT_AGE_SECS 300
49 * The nr of bytes of cached data to keep around.
51 #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
54 * The number of bvec entries that are embedded directly in the buffer.
55 * If the chunk size is larger, dm-io is used to do the io.
57 #define DM_BUFIO_INLINE_VECS 16
60 * Don't try to use kmem_cache_alloc for blocks larger than this.
61 * For explanation, see alloc_buffer_data below.
63 #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1)
64 #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1))
67 * dm_buffer->list_mode
75 * All buffers are linked to cache_hash with their hash_list field.
77 * Clean buffers that are not being written (B_WRITING not set)
78 * are linked to lru[LIST_CLEAN] with their lru_list field.
80 * Dirty and clean buffers that are being written are linked to
81 * lru[LIST_DIRTY] with their lru_list field. When the write
82 * finishes, the buffer cannot be relinked immediately (because we
83 * are in an interrupt context and relinking requires process
84 * context), so some clean-not-writing buffers can be held on
85 * dirty_lru too. They are later added to lru in the process
88 struct dm_bufio_client {
91 struct list_head lru[LIST_SIZE];
92 unsigned long n_buffers[LIST_SIZE];
94 struct block_device *bdev;
96 unsigned char sectors_per_block_bits;
97 unsigned char pages_per_block_bits;
98 unsigned char blocks_per_page_bits;
100 void (*alloc_callback)(struct dm_buffer *);
101 void (*write_callback)(struct dm_buffer *);
103 struct dm_io_client *dm_io;
105 struct list_head reserved_buffers;
106 unsigned need_reserved_buffers;
108 unsigned minimum_buffers;
110 struct rb_root buffer_tree;
111 wait_queue_head_t free_buffer_wait;
115 int async_write_error;
117 struct list_head client_list;
118 struct shrinker shrinker;
129 * Describes how the block was allocated:
130 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
131 * See the comment at alloc_buffer_data.
135 DATA_MODE_GET_FREE_PAGES = 1,
136 DATA_MODE_VMALLOC = 2,
142 struct list_head lru_list;
145 enum data_mode data_mode;
146 unsigned char list_mode; /* LIST_* */
148 blk_status_t read_error;
149 blk_status_t write_error;
151 unsigned long last_accessed;
152 struct dm_bufio_client *c;
153 struct list_head write_list;
155 struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
156 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
158 struct stack_trace stack_trace;
159 unsigned long stack_entries[MAX_STACK];
163 /*----------------------------------------------------------------*/
165 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
166 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
168 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
170 unsigned ret = c->blocks_per_page_bits - 1;
172 BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
177 #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
178 #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
180 #define dm_bufio_in_request() (!!current->bio_list)
182 static void dm_bufio_lock(struct dm_bufio_client *c)
184 mutex_lock_nested(&c->lock, dm_bufio_in_request());
187 static int dm_bufio_trylock(struct dm_bufio_client *c)
189 return mutex_trylock(&c->lock);
192 static void dm_bufio_unlock(struct dm_bufio_client *c)
194 mutex_unlock(&c->lock);
197 /*----------------------------------------------------------------*/
200 * Default cache size: available memory divided by the ratio.
202 static unsigned long dm_bufio_default_cache_size;
205 * Total cache size set by the user.
207 static unsigned long dm_bufio_cache_size;
210 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
211 * at any time. If it disagrees, the user has changed cache size.
213 static unsigned long dm_bufio_cache_size_latch;
215 static DEFINE_SPINLOCK(param_spinlock);
218 * Buffers are freed after this timeout
220 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
221 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
223 static unsigned long dm_bufio_peak_allocated;
224 static unsigned long dm_bufio_allocated_kmem_cache;
225 static unsigned long dm_bufio_allocated_get_free_pages;
226 static unsigned long dm_bufio_allocated_vmalloc;
227 static unsigned long dm_bufio_current_allocated;
229 /*----------------------------------------------------------------*/
232 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
234 static unsigned long dm_bufio_cache_size_per_client;
237 * The current number of clients.
239 static int dm_bufio_client_count;
242 * The list of all clients.
244 static LIST_HEAD(dm_bufio_all_clients);
247 * This mutex protects dm_bufio_cache_size_latch,
248 * dm_bufio_cache_size_per_client and dm_bufio_client_count
250 static DEFINE_MUTEX(dm_bufio_clients_lock);
252 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
253 static void buffer_record_stack(struct dm_buffer *b)
255 b->stack_trace.nr_entries = 0;
256 b->stack_trace.max_entries = MAX_STACK;
257 b->stack_trace.entries = b->stack_entries;
258 b->stack_trace.skip = 2;
259 save_stack_trace(&b->stack_trace);
263 /*----------------------------------------------------------------
264 * A red/black tree acts as an index for all the buffers.
265 *--------------------------------------------------------------*/
266 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
268 struct rb_node *n = c->buffer_tree.rb_node;
272 b = container_of(n, struct dm_buffer, node);
274 if (b->block == block)
277 n = (b->block < block) ? n->rb_left : n->rb_right;
283 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
285 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
286 struct dm_buffer *found;
289 found = container_of(*new, struct dm_buffer, node);
291 if (found->block == b->block) {
297 new = (found->block < b->block) ?
298 &((*new)->rb_left) : &((*new)->rb_right);
301 rb_link_node(&b->node, parent, new);
302 rb_insert_color(&b->node, &c->buffer_tree);
305 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
307 rb_erase(&b->node, &c->buffer_tree);
310 /*----------------------------------------------------------------*/
312 static void adjust_total_allocated(enum data_mode data_mode, long diff)
314 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
315 &dm_bufio_allocated_kmem_cache,
316 &dm_bufio_allocated_get_free_pages,
317 &dm_bufio_allocated_vmalloc,
320 spin_lock(¶m_spinlock);
322 *class_ptr[data_mode] += diff;
324 dm_bufio_current_allocated += diff;
326 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
327 dm_bufio_peak_allocated = dm_bufio_current_allocated;
329 spin_unlock(¶m_spinlock);
333 * Change the number of clients and recalculate per-client limit.
335 static void __cache_size_refresh(void)
337 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
338 BUG_ON(dm_bufio_client_count < 0);
340 dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
343 * Use default if set to 0 and report the actual cache size used.
345 if (!dm_bufio_cache_size_latch) {
346 (void)cmpxchg(&dm_bufio_cache_size, 0,
347 dm_bufio_default_cache_size);
348 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
351 dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
352 (dm_bufio_client_count ? : 1);
356 * Allocating buffer data.
358 * Small buffers are allocated with kmem_cache, to use space optimally.
360 * For large buffers, we choose between get_free_pages and vmalloc.
361 * Each has advantages and disadvantages.
363 * __get_free_pages can randomly fail if the memory is fragmented.
364 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
365 * as low as 128M) so using it for caching is not appropriate.
367 * If the allocation may fail we use __get_free_pages. Memory fragmentation
368 * won't have a fatal effect here, but it just causes flushes of some other
369 * buffers and more I/O will be performed. Don't use __get_free_pages if it
370 * always fails (i.e. order >= MAX_ORDER).
372 * If the allocation shouldn't fail we use __vmalloc. This is only for the
373 * initial reserve allocation, so there's no risk of wasting all vmalloc
376 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
377 enum data_mode *data_mode)
382 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
383 *data_mode = DATA_MODE_SLAB;
384 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
387 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
388 gfp_mask & __GFP_NORETRY) {
389 *data_mode = DATA_MODE_GET_FREE_PAGES;
390 return (void *)__get_free_pages(gfp_mask,
391 c->pages_per_block_bits);
394 *data_mode = DATA_MODE_VMALLOC;
397 * __vmalloc allocates the data pages and auxiliary structures with
398 * gfp_flags that were specified, but pagetables are always allocated
399 * with GFP_KERNEL, no matter what was specified as gfp_mask.
401 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
402 * all allocations done by this process (including pagetables) are done
403 * as if GFP_NOIO was specified.
406 if (gfp_mask & __GFP_NORETRY)
407 noio_flag = memalloc_noio_save();
409 ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
411 if (gfp_mask & __GFP_NORETRY)
412 memalloc_noio_restore(noio_flag);
418 * Free buffer's data.
420 static void free_buffer_data(struct dm_bufio_client *c,
421 void *data, enum data_mode data_mode)
425 kmem_cache_free(DM_BUFIO_CACHE(c), data);
428 case DATA_MODE_GET_FREE_PAGES:
429 free_pages((unsigned long)data, c->pages_per_block_bits);
432 case DATA_MODE_VMALLOC:
437 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
444 * Allocate buffer and its data.
446 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
448 struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
456 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
462 adjust_total_allocated(b->data_mode, (long)c->block_size);
464 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
465 memset(&b->stack_trace, 0, sizeof(b->stack_trace));
471 * Free buffer and its data.
473 static void free_buffer(struct dm_buffer *b)
475 struct dm_bufio_client *c = b->c;
477 adjust_total_allocated(b->data_mode, -(long)c->block_size);
479 free_buffer_data(c, b->data, b->data_mode);
484 * Link buffer to the hash list and clean or dirty queue.
486 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
488 struct dm_bufio_client *c = b->c;
490 c->n_buffers[dirty]++;
492 b->list_mode = dirty;
493 list_add(&b->lru_list, &c->lru[dirty]);
495 b->last_accessed = jiffies;
499 * Unlink buffer from the hash list and dirty or clean queue.
501 static void __unlink_buffer(struct dm_buffer *b)
503 struct dm_bufio_client *c = b->c;
505 BUG_ON(!c->n_buffers[b->list_mode]);
507 c->n_buffers[b->list_mode]--;
509 list_del(&b->lru_list);
513 * Place the buffer to the head of dirty or clean LRU queue.
515 static void __relink_lru(struct dm_buffer *b, int dirty)
517 struct dm_bufio_client *c = b->c;
519 BUG_ON(!c->n_buffers[b->list_mode]);
521 c->n_buffers[b->list_mode]--;
522 c->n_buffers[dirty]++;
523 b->list_mode = dirty;
524 list_move(&b->lru_list, &c->lru[dirty]);
525 b->last_accessed = jiffies;
528 /*----------------------------------------------------------------
529 * Submit I/O on the buffer.
531 * Bio interface is faster but it has some problems:
532 * the vector list is limited (increasing this limit increases
533 * memory-consumption per buffer, so it is not viable);
535 * the memory must be direct-mapped, not vmalloced;
537 * the I/O driver can reject requests spuriously if it thinks that
538 * the requests are too big for the device or if they cross a
539 * controller-defined memory boundary.
541 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
542 * it is not vmalloced, try using the bio interface.
544 * If the buffer is big, if it is vmalloced or if the underlying device
545 * rejects the bio because it is too large, use dm-io layer to do the I/O.
546 * The dm-io layer splits the I/O into multiple requests, avoiding the above
548 *--------------------------------------------------------------*/
551 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
552 * that the request was handled directly with bio interface.
554 static void dmio_complete(unsigned long error, void *context)
556 struct dm_buffer *b = context;
558 b->bio.bi_status = error ? BLK_STS_IOERR : 0;
559 b->bio.bi_end_io(&b->bio);
562 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
563 unsigned n_sectors, bio_end_io_t *end_io)
566 struct dm_io_request io_req = {
569 .notify.fn = dmio_complete,
571 .client = b->c->dm_io,
573 struct dm_io_region region = {
579 if (b->data_mode != DATA_MODE_VMALLOC) {
580 io_req.mem.type = DM_IO_KMEM;
581 io_req.mem.ptr.addr = b->data;
583 io_req.mem.type = DM_IO_VMA;
584 io_req.mem.ptr.vma = b->data;
587 b->bio.bi_end_io = end_io;
589 r = dm_io(&io_req, 1, ®ion, NULL);
591 b->bio.bi_status = errno_to_blk_status(r);
596 static void inline_endio(struct bio *bio)
598 bio_end_io_t *end_fn = bio->bi_private;
599 blk_status_t status = bio->bi_status;
602 * Reset the bio to free any attached resources
603 * (e.g. bio integrity profiles).
607 bio->bi_status = status;
611 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t sector,
612 unsigned n_sectors, bio_end_io_t *end_io)
617 bio_init(&b->bio, b->bio_vec, DM_BUFIO_INLINE_VECS);
618 b->bio.bi_iter.bi_sector = sector;
619 bio_set_dev(&b->bio, b->c->bdev);
620 b->bio.bi_end_io = inline_endio;
622 * Use of .bi_private isn't a problem here because
623 * the dm_buffer's inline bio is local to bufio.
625 b->bio.bi_private = end_io;
626 bio_set_op_attrs(&b->bio, rw, 0);
629 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
630 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
633 len = n_sectors << SECTOR_SHIFT;
635 if (len >= PAGE_SIZE)
636 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
638 BUG_ON((unsigned long)ptr & (len - 1));
641 if (!bio_add_page(&b->bio, virt_to_page(ptr),
642 len < PAGE_SIZE ? len : PAGE_SIZE,
643 offset_in_page(ptr))) {
644 BUG_ON(b->c->block_size <= PAGE_SIZE);
645 use_dmio(b, rw, sector, n_sectors, end_io);
656 static void submit_io(struct dm_buffer *b, int rw, bio_end_io_t *end_io)
661 if (rw == WRITE && b->c->write_callback)
662 b->c->write_callback(b);
664 sector = (b->block << b->c->sectors_per_block_bits) + b->c->start;
665 n_sectors = 1 << b->c->sectors_per_block_bits;
667 if (n_sectors <= ((DM_BUFIO_INLINE_VECS * PAGE_SIZE) >> SECTOR_SHIFT) &&
668 b->data_mode != DATA_MODE_VMALLOC)
669 use_inline_bio(b, rw, sector, n_sectors, end_io);
671 use_dmio(b, rw, sector, n_sectors, end_io);
674 /*----------------------------------------------------------------
675 * Writing dirty buffers
676 *--------------------------------------------------------------*/
679 * The endio routine for write.
681 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
684 static void write_endio(struct bio *bio)
686 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
688 b->write_error = bio->bi_status;
689 if (unlikely(bio->bi_status)) {
690 struct dm_bufio_client *c = b->c;
692 (void)cmpxchg(&c->async_write_error, 0,
693 blk_status_to_errno(bio->bi_status));
696 BUG_ON(!test_bit(B_WRITING, &b->state));
698 smp_mb__before_atomic();
699 clear_bit(B_WRITING, &b->state);
700 smp_mb__after_atomic();
702 wake_up_bit(&b->state, B_WRITING);
706 * Initiate a write on a dirty buffer, but don't wait for it.
708 * - If the buffer is not dirty, exit.
709 * - If there some previous write going on, wait for it to finish (we can't
710 * have two writes on the same buffer simultaneously).
711 * - Submit our write and don't wait on it. We set B_WRITING indicating
712 * that there is a write in progress.
714 static void __write_dirty_buffer(struct dm_buffer *b,
715 struct list_head *write_list)
717 if (!test_bit(B_DIRTY, &b->state))
720 clear_bit(B_DIRTY, &b->state);
721 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
724 submit_io(b, WRITE, write_endio);
726 list_add_tail(&b->write_list, write_list);
729 static void __flush_write_list(struct list_head *write_list)
731 struct blk_plug plug;
732 blk_start_plug(&plug);
733 while (!list_empty(write_list)) {
734 struct dm_buffer *b =
735 list_entry(write_list->next, struct dm_buffer, write_list);
736 list_del(&b->write_list);
737 submit_io(b, WRITE, write_endio);
740 blk_finish_plug(&plug);
744 * Wait until any activity on the buffer finishes. Possibly write the
745 * buffer if it is dirty. When this function finishes, there is no I/O
746 * running on the buffer and the buffer is not dirty.
748 static void __make_buffer_clean(struct dm_buffer *b)
750 BUG_ON(b->hold_count);
752 if (!b->state) /* fast case */
755 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
756 __write_dirty_buffer(b, NULL);
757 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
761 * Find some buffer that is not held by anybody, clean it, unlink it and
764 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
768 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
769 BUG_ON(test_bit(B_WRITING, &b->state));
770 BUG_ON(test_bit(B_DIRTY, &b->state));
772 if (!b->hold_count) {
773 __make_buffer_clean(b);
780 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
781 BUG_ON(test_bit(B_READING, &b->state));
783 if (!b->hold_count) {
784 __make_buffer_clean(b);
795 * Wait until some other threads free some buffer or release hold count on
798 * This function is entered with c->lock held, drops it and regains it
801 static void __wait_for_free_buffer(struct dm_bufio_client *c)
803 DECLARE_WAITQUEUE(wait, current);
805 add_wait_queue(&c->free_buffer_wait, &wait);
806 set_current_state(TASK_UNINTERRUPTIBLE);
811 remove_wait_queue(&c->free_buffer_wait, &wait);
824 * Allocate a new buffer. If the allocation is not possible, wait until
825 * some other thread frees a buffer.
827 * May drop the lock and regain it.
829 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
832 bool tried_noio_alloc = false;
835 * dm-bufio is resistant to allocation failures (it just keeps
836 * one buffer reserved in cases all the allocations fail).
837 * So set flags to not try too hard:
838 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
839 * mutex and wait ourselves.
840 * __GFP_NORETRY: don't retry and rather return failure
841 * __GFP_NOMEMALLOC: don't use emergency reserves
842 * __GFP_NOWARN: don't print a warning in case of failure
844 * For debugging, if we set the cache size to 1, no new buffers will
848 if (dm_bufio_cache_size_latch != 1) {
849 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
854 if (nf == NF_PREFETCH)
857 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
859 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
863 tried_noio_alloc = true;
866 if (!list_empty(&c->reserved_buffers)) {
867 b = list_entry(c->reserved_buffers.next,
868 struct dm_buffer, lru_list);
869 list_del(&b->lru_list);
870 c->need_reserved_buffers++;
875 b = __get_unclaimed_buffer(c);
879 __wait_for_free_buffer(c);
883 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
885 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
890 if (c->alloc_callback)
891 c->alloc_callback(b);
897 * Free a buffer and wake other threads waiting for free buffers.
899 static void __free_buffer_wake(struct dm_buffer *b)
901 struct dm_bufio_client *c = b->c;
903 if (!c->need_reserved_buffers)
906 list_add(&b->lru_list, &c->reserved_buffers);
907 c->need_reserved_buffers--;
910 wake_up(&c->free_buffer_wait);
913 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
914 struct list_head *write_list)
916 struct dm_buffer *b, *tmp;
918 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
919 BUG_ON(test_bit(B_READING, &b->state));
921 if (!test_bit(B_DIRTY, &b->state) &&
922 !test_bit(B_WRITING, &b->state)) {
923 __relink_lru(b, LIST_CLEAN);
927 if (no_wait && test_bit(B_WRITING, &b->state))
930 __write_dirty_buffer(b, write_list);
936 * Get writeback threshold and buffer limit for a given client.
938 static void __get_memory_limit(struct dm_bufio_client *c,
939 unsigned long *threshold_buffers,
940 unsigned long *limit_buffers)
942 unsigned long buffers;
944 if (unlikely(ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch)) {
945 if (mutex_trylock(&dm_bufio_clients_lock)) {
946 __cache_size_refresh();
947 mutex_unlock(&dm_bufio_clients_lock);
951 buffers = dm_bufio_cache_size_per_client >>
952 (c->sectors_per_block_bits + SECTOR_SHIFT);
954 if (buffers < c->minimum_buffers)
955 buffers = c->minimum_buffers;
957 *limit_buffers = buffers;
958 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
962 * Check if we're over watermark.
963 * If we are over threshold_buffers, start freeing buffers.
964 * If we're over "limit_buffers", block until we get under the limit.
966 static void __check_watermark(struct dm_bufio_client *c,
967 struct list_head *write_list)
969 unsigned long threshold_buffers, limit_buffers;
971 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
973 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
976 struct dm_buffer *b = __get_unclaimed_buffer(c);
981 __free_buffer_wake(b);
985 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
986 __write_dirty_buffers_async(c, 1, write_list);
989 /*----------------------------------------------------------------
991 *--------------------------------------------------------------*/
993 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
994 enum new_flag nf, int *need_submit,
995 struct list_head *write_list)
997 struct dm_buffer *b, *new_b = NULL;
1001 b = __find(c, block);
1008 new_b = __alloc_buffer_wait(c, nf);
1013 * We've had a period where the mutex was unlocked, so need to
1014 * recheck the hash table.
1016 b = __find(c, block);
1018 __free_buffer_wake(new_b);
1022 __check_watermark(c, write_list);
1028 __link_buffer(b, block, LIST_CLEAN);
1030 if (nf == NF_FRESH) {
1035 b->state = 1 << B_READING;
1041 if (nf == NF_PREFETCH)
1044 * Note: it is essential that we don't wait for the buffer to be
1045 * read if dm_bufio_get function is used. Both dm_bufio_get and
1046 * dm_bufio_prefetch can be used in the driver request routine.
1047 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1048 * the same buffer, it would deadlock if we waited.
1050 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1054 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1055 test_bit(B_WRITING, &b->state));
1060 * The endio routine for reading: set the error, clear the bit and wake up
1061 * anyone waiting on the buffer.
1063 static void read_endio(struct bio *bio)
1065 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
1067 b->read_error = bio->bi_status;
1069 BUG_ON(!test_bit(B_READING, &b->state));
1071 smp_mb__before_atomic();
1072 clear_bit(B_READING, &b->state);
1073 smp_mb__after_atomic();
1075 wake_up_bit(&b->state, B_READING);
1079 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1080 * functions is similar except that dm_bufio_new doesn't read the
1081 * buffer from the disk (assuming that the caller overwrites all the data
1082 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1084 static void *new_read(struct dm_bufio_client *c, sector_t block,
1085 enum new_flag nf, struct dm_buffer **bp)
1088 struct dm_buffer *b;
1090 LIST_HEAD(write_list);
1093 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1094 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1095 if (b && b->hold_count == 1)
1096 buffer_record_stack(b);
1100 __flush_write_list(&write_list);
1106 submit_io(b, READ, read_endio);
1108 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1110 if (b->read_error) {
1111 int error = blk_status_to_errno(b->read_error);
1113 dm_bufio_release(b);
1115 return ERR_PTR(error);
1123 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1124 struct dm_buffer **bp)
1126 return new_read(c, block, NF_GET, bp);
1128 EXPORT_SYMBOL_GPL(dm_bufio_get);
1130 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1131 struct dm_buffer **bp)
1133 BUG_ON(dm_bufio_in_request());
1135 return new_read(c, block, NF_READ, bp);
1137 EXPORT_SYMBOL_GPL(dm_bufio_read);
1139 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1140 struct dm_buffer **bp)
1142 BUG_ON(dm_bufio_in_request());
1144 return new_read(c, block, NF_FRESH, bp);
1146 EXPORT_SYMBOL_GPL(dm_bufio_new);
1148 void dm_bufio_prefetch(struct dm_bufio_client *c,
1149 sector_t block, unsigned n_blocks)
1151 struct blk_plug plug;
1153 LIST_HEAD(write_list);
1155 BUG_ON(dm_bufio_in_request());
1157 blk_start_plug(&plug);
1160 for (; n_blocks--; block++) {
1162 struct dm_buffer *b;
1163 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1165 if (unlikely(!list_empty(&write_list))) {
1167 blk_finish_plug(&plug);
1168 __flush_write_list(&write_list);
1169 blk_start_plug(&plug);
1172 if (unlikely(b != NULL)) {
1176 submit_io(b, READ, read_endio);
1177 dm_bufio_release(b);
1190 blk_finish_plug(&plug);
1192 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1194 void dm_bufio_release(struct dm_buffer *b)
1196 struct dm_bufio_client *c = b->c;
1200 BUG_ON(!b->hold_count);
1203 if (!b->hold_count) {
1204 wake_up(&c->free_buffer_wait);
1207 * If there were errors on the buffer, and the buffer is not
1208 * to be written, free the buffer. There is no point in caching
1211 if ((b->read_error || b->write_error) &&
1212 !test_bit(B_READING, &b->state) &&
1213 !test_bit(B_WRITING, &b->state) &&
1214 !test_bit(B_DIRTY, &b->state)) {
1216 __free_buffer_wake(b);
1222 EXPORT_SYMBOL_GPL(dm_bufio_release);
1224 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1226 struct dm_bufio_client *c = b->c;
1230 BUG_ON(test_bit(B_READING, &b->state));
1232 if (!test_and_set_bit(B_DIRTY, &b->state))
1233 __relink_lru(b, LIST_DIRTY);
1237 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1239 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1241 LIST_HEAD(write_list);
1243 BUG_ON(dm_bufio_in_request());
1246 __write_dirty_buffers_async(c, 0, &write_list);
1248 __flush_write_list(&write_list);
1250 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1253 * For performance, it is essential that the buffers are written asynchronously
1254 * and simultaneously (so that the block layer can merge the writes) and then
1257 * Finally, we flush hardware disk cache.
1259 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1262 unsigned long buffers_processed = 0;
1263 struct dm_buffer *b, *tmp;
1265 LIST_HEAD(write_list);
1268 __write_dirty_buffers_async(c, 0, &write_list);
1270 __flush_write_list(&write_list);
1274 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1275 int dropped_lock = 0;
1277 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1278 buffers_processed++;
1280 BUG_ON(test_bit(B_READING, &b->state));
1282 if (test_bit(B_WRITING, &b->state)) {
1283 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1287 wait_on_bit_io(&b->state, B_WRITING,
1288 TASK_UNINTERRUPTIBLE);
1292 wait_on_bit_io(&b->state, B_WRITING,
1293 TASK_UNINTERRUPTIBLE);
1296 if (!test_bit(B_DIRTY, &b->state) &&
1297 !test_bit(B_WRITING, &b->state))
1298 __relink_lru(b, LIST_CLEAN);
1303 * If we dropped the lock, the list is no longer consistent,
1304 * so we must restart the search.
1306 * In the most common case, the buffer just processed is
1307 * relinked to the clean list, so we won't loop scanning the
1308 * same buffer again and again.
1310 * This may livelock if there is another thread simultaneously
1311 * dirtying buffers, so we count the number of buffers walked
1312 * and if it exceeds the total number of buffers, it means that
1313 * someone is doing some writes simultaneously with us. In
1314 * this case, stop, dropping the lock.
1319 wake_up(&c->free_buffer_wait);
1322 a = xchg(&c->async_write_error, 0);
1323 f = dm_bufio_issue_flush(c);
1329 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1332 * Use dm-io to send and empty barrier flush the device.
1334 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1336 struct dm_io_request io_req = {
1337 .bi_op = REQ_OP_WRITE,
1338 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1339 .mem.type = DM_IO_KMEM,
1340 .mem.ptr.addr = NULL,
1343 struct dm_io_region io_reg = {
1349 BUG_ON(dm_bufio_in_request());
1351 return dm_io(&io_req, 1, &io_reg, NULL);
1353 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1356 * We first delete any other buffer that may be at that new location.
1358 * Then, we write the buffer to the original location if it was dirty.
1360 * Then, if we are the only one who is holding the buffer, relink the buffer
1361 * in the hash queue for the new location.
1363 * If there was someone else holding the buffer, we write it to the new
1364 * location but not relink it, because that other user needs to have the buffer
1365 * at the same place.
1367 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1369 struct dm_bufio_client *c = b->c;
1370 struct dm_buffer *new;
1372 BUG_ON(dm_bufio_in_request());
1377 new = __find(c, new_block);
1379 if (new->hold_count) {
1380 __wait_for_free_buffer(c);
1385 * FIXME: Is there any point waiting for a write that's going
1386 * to be overwritten in a bit?
1388 __make_buffer_clean(new);
1389 __unlink_buffer(new);
1390 __free_buffer_wake(new);
1393 BUG_ON(!b->hold_count);
1394 BUG_ON(test_bit(B_READING, &b->state));
1396 __write_dirty_buffer(b, NULL);
1397 if (b->hold_count == 1) {
1398 wait_on_bit_io(&b->state, B_WRITING,
1399 TASK_UNINTERRUPTIBLE);
1400 set_bit(B_DIRTY, &b->state);
1402 __link_buffer(b, new_block, LIST_DIRTY);
1405 wait_on_bit_lock_io(&b->state, B_WRITING,
1406 TASK_UNINTERRUPTIBLE);
1408 * Relink buffer to "new_block" so that write_callback
1409 * sees "new_block" as a block number.
1410 * After the write, link the buffer back to old_block.
1411 * All this must be done in bufio lock, so that block number
1412 * change isn't visible to other threads.
1414 old_block = b->block;
1416 __link_buffer(b, new_block, b->list_mode);
1417 submit_io(b, WRITE, write_endio);
1418 wait_on_bit_io(&b->state, B_WRITING,
1419 TASK_UNINTERRUPTIBLE);
1421 __link_buffer(b, old_block, b->list_mode);
1425 dm_bufio_release(b);
1427 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1430 * Free the given buffer.
1432 * This is just a hint, if the buffer is in use or dirty, this function
1435 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1437 struct dm_buffer *b;
1441 b = __find(c, block);
1442 if (b && likely(!b->hold_count) && likely(!b->state)) {
1444 __free_buffer_wake(b);
1449 EXPORT_SYMBOL(dm_bufio_forget);
1451 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1453 c->minimum_buffers = n;
1455 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1457 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1459 return c->block_size;
1461 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1463 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1465 return i_size_read(c->bdev->bd_inode) >>
1466 (SECTOR_SHIFT + c->sectors_per_block_bits);
1468 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1470 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1474 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1476 void *dm_bufio_get_block_data(struct dm_buffer *b)
1480 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1482 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1486 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1488 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1492 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1494 static void drop_buffers(struct dm_bufio_client *c)
1496 struct dm_buffer *b;
1498 bool warned = false;
1500 BUG_ON(dm_bufio_in_request());
1503 * An optimization so that the buffers are not written one-by-one.
1505 dm_bufio_write_dirty_buffers_async(c);
1509 while ((b = __get_unclaimed_buffer(c)))
1510 __free_buffer_wake(b);
1512 for (i = 0; i < LIST_SIZE; i++)
1513 list_for_each_entry(b, &c->lru[i], lru_list) {
1516 DMERR("leaked buffer %llx, hold count %u, list %d",
1517 (unsigned long long)b->block, b->hold_count, i);
1518 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1519 print_stack_trace(&b->stack_trace, 1);
1520 b->hold_count = 0; /* mark unclaimed to avoid BUG_ON below */
1524 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1525 while ((b = __get_unclaimed_buffer(c)))
1526 __free_buffer_wake(b);
1529 for (i = 0; i < LIST_SIZE; i++)
1530 BUG_ON(!list_empty(&c->lru[i]));
1536 * We may not be able to evict this buffer if IO pending or the client
1537 * is still using it. Caller is expected to know buffer is too old.
1539 * And if GFP_NOFS is used, we must not do any I/O because we hold
1540 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1541 * rerouted to different bufio client.
1543 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1545 if (!(gfp & __GFP_FS)) {
1546 if (test_bit(B_READING, &b->state) ||
1547 test_bit(B_WRITING, &b->state) ||
1548 test_bit(B_DIRTY, &b->state))
1555 __make_buffer_clean(b);
1557 __free_buffer_wake(b);
1562 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1564 unsigned long retain_bytes = ACCESS_ONCE(dm_bufio_retain_bytes);
1565 return retain_bytes >> (c->sectors_per_block_bits + SECTOR_SHIFT);
1568 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1572 struct dm_buffer *b, *tmp;
1573 unsigned long freed = 0;
1574 unsigned long count = nr_to_scan;
1575 unsigned long retain_target = get_retain_buffers(c);
1577 for (l = 0; l < LIST_SIZE; l++) {
1578 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1579 if (__try_evict_buffer(b, gfp_mask))
1581 if (!--nr_to_scan || ((count - freed) <= retain_target))
1589 static unsigned long
1590 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1592 struct dm_bufio_client *c;
1593 unsigned long freed;
1595 c = container_of(shrink, struct dm_bufio_client, shrinker);
1596 if (sc->gfp_mask & __GFP_FS)
1598 else if (!dm_bufio_trylock(c))
1601 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1606 static unsigned long
1607 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1609 struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1611 return ACCESS_ONCE(c->n_buffers[LIST_CLEAN]) + ACCESS_ONCE(c->n_buffers[LIST_DIRTY]);
1615 * Create the buffering interface
1617 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1618 unsigned reserved_buffers, unsigned aux_size,
1619 void (*alloc_callback)(struct dm_buffer *),
1620 void (*write_callback)(struct dm_buffer *))
1623 struct dm_bufio_client *c;
1626 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1627 (block_size & (block_size - 1)));
1629 c = kzalloc(sizeof(*c), GFP_KERNEL);
1634 c->buffer_tree = RB_ROOT;
1637 c->block_size = block_size;
1638 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1639 c->pages_per_block_bits = (__ffs(block_size) >= PAGE_SHIFT) ?
1640 __ffs(block_size) - PAGE_SHIFT : 0;
1641 c->blocks_per_page_bits = (__ffs(block_size) < PAGE_SHIFT ?
1642 PAGE_SHIFT - __ffs(block_size) : 0);
1644 c->aux_size = aux_size;
1645 c->alloc_callback = alloc_callback;
1646 c->write_callback = write_callback;
1648 for (i = 0; i < LIST_SIZE; i++) {
1649 INIT_LIST_HEAD(&c->lru[i]);
1650 c->n_buffers[i] = 0;
1653 mutex_init(&c->lock);
1654 INIT_LIST_HEAD(&c->reserved_buffers);
1655 c->need_reserved_buffers = reserved_buffers;
1657 c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1659 init_waitqueue_head(&c->free_buffer_wait);
1660 c->async_write_error = 0;
1662 c->dm_io = dm_io_client_create();
1663 if (IS_ERR(c->dm_io)) {
1664 r = PTR_ERR(c->dm_io);
1668 mutex_lock(&dm_bufio_clients_lock);
1669 if (c->blocks_per_page_bits) {
1670 if (!DM_BUFIO_CACHE_NAME(c)) {
1671 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1672 if (!DM_BUFIO_CACHE_NAME(c)) {
1674 mutex_unlock(&dm_bufio_clients_lock);
1679 if (!DM_BUFIO_CACHE(c)) {
1680 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1682 c->block_size, 0, NULL);
1683 if (!DM_BUFIO_CACHE(c)) {
1685 mutex_unlock(&dm_bufio_clients_lock);
1690 mutex_unlock(&dm_bufio_clients_lock);
1692 while (c->need_reserved_buffers) {
1693 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1699 __free_buffer_wake(b);
1702 mutex_lock(&dm_bufio_clients_lock);
1703 dm_bufio_client_count++;
1704 list_add(&c->client_list, &dm_bufio_all_clients);
1705 __cache_size_refresh();
1706 mutex_unlock(&dm_bufio_clients_lock);
1708 c->shrinker.count_objects = dm_bufio_shrink_count;
1709 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1710 c->shrinker.seeks = 1;
1711 c->shrinker.batch = 0;
1712 register_shrinker(&c->shrinker);
1718 while (!list_empty(&c->reserved_buffers)) {
1719 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1720 struct dm_buffer, lru_list);
1721 list_del(&b->lru_list);
1724 dm_io_client_destroy(c->dm_io);
1730 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1733 * Free the buffering interface.
1734 * It is required that there are no references on any buffers.
1736 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1742 unregister_shrinker(&c->shrinker);
1744 mutex_lock(&dm_bufio_clients_lock);
1746 list_del(&c->client_list);
1747 dm_bufio_client_count--;
1748 __cache_size_refresh();
1750 mutex_unlock(&dm_bufio_clients_lock);
1752 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1753 BUG_ON(c->need_reserved_buffers);
1755 while (!list_empty(&c->reserved_buffers)) {
1756 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1757 struct dm_buffer, lru_list);
1758 list_del(&b->lru_list);
1762 for (i = 0; i < LIST_SIZE; i++)
1763 if (c->n_buffers[i])
1764 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1766 for (i = 0; i < LIST_SIZE; i++)
1767 BUG_ON(c->n_buffers[i]);
1769 dm_io_client_destroy(c->dm_io);
1772 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1774 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1778 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1780 static unsigned get_max_age_hz(void)
1782 unsigned max_age = ACCESS_ONCE(dm_bufio_max_age);
1784 if (max_age > UINT_MAX / HZ)
1785 max_age = UINT_MAX / HZ;
1787 return max_age * HZ;
1790 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1792 return time_after_eq(jiffies, b->last_accessed + age_hz);
1795 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1797 struct dm_buffer *b, *tmp;
1798 unsigned long retain_target = get_retain_buffers(c);
1799 unsigned long count;
1800 LIST_HEAD(write_list);
1804 __check_watermark(c, &write_list);
1805 if (unlikely(!list_empty(&write_list))) {
1807 __flush_write_list(&write_list);
1811 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1812 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1813 if (count <= retain_target)
1816 if (!older_than(b, age_hz))
1819 if (__try_evict_buffer(b, 0))
1828 static void cleanup_old_buffers(void)
1830 unsigned long max_age_hz = get_max_age_hz();
1831 struct dm_bufio_client *c;
1833 mutex_lock(&dm_bufio_clients_lock);
1835 __cache_size_refresh();
1837 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1838 __evict_old_buffers(c, max_age_hz);
1840 mutex_unlock(&dm_bufio_clients_lock);
1843 static struct workqueue_struct *dm_bufio_wq;
1844 static struct delayed_work dm_bufio_work;
1846 static void work_fn(struct work_struct *w)
1848 cleanup_old_buffers();
1850 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1851 DM_BUFIO_WORK_TIMER_SECS * HZ);
1854 /*----------------------------------------------------------------
1856 *--------------------------------------------------------------*/
1859 * This is called only once for the whole dm_bufio module.
1860 * It initializes memory limit.
1862 static int __init dm_bufio_init(void)
1866 dm_bufio_allocated_kmem_cache = 0;
1867 dm_bufio_allocated_get_free_pages = 0;
1868 dm_bufio_allocated_vmalloc = 0;
1869 dm_bufio_current_allocated = 0;
1871 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1872 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1874 mem = (__u64)((totalram_pages - totalhigh_pages) *
1875 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1877 if (mem > ULONG_MAX)
1882 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1883 * in fs/proc/internal.h
1885 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1886 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1889 dm_bufio_default_cache_size = mem;
1891 mutex_lock(&dm_bufio_clients_lock);
1892 __cache_size_refresh();
1893 mutex_unlock(&dm_bufio_clients_lock);
1895 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1899 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1900 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1901 DM_BUFIO_WORK_TIMER_SECS * HZ);
1907 * This is called once when unloading the dm_bufio module.
1909 static void __exit dm_bufio_exit(void)
1914 cancel_delayed_work_sync(&dm_bufio_work);
1915 destroy_workqueue(dm_bufio_wq);
1917 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++)
1918 kmem_cache_destroy(dm_bufio_caches[i]);
1920 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1921 kfree(dm_bufio_cache_names[i]);
1923 if (dm_bufio_client_count) {
1924 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1925 __func__, dm_bufio_client_count);
1929 if (dm_bufio_current_allocated) {
1930 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1931 __func__, dm_bufio_current_allocated);
1935 if (dm_bufio_allocated_get_free_pages) {
1936 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1937 __func__, dm_bufio_allocated_get_free_pages);
1941 if (dm_bufio_allocated_vmalloc) {
1942 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1943 __func__, dm_bufio_allocated_vmalloc);
1950 module_init(dm_bufio_init)
1951 module_exit(dm_bufio_exit)
1953 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1954 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1956 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1957 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1959 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
1960 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1962 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1963 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1965 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1966 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1968 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1969 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1971 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1972 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1974 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1975 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1977 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1978 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1979 MODULE_LICENSE("GPL");