2 * Copyright (C) 2009-2011 Red Hat, Inc.
4 * Author: Mikulas Patocka <mpatocka@redhat.com>
6 * This file is released under the GPL.
9 #include <linux/dm-bufio.h>
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>
21 #include <linux/jump_label.h>
23 #define DM_MSG_PREFIX "bufio"
26 * Memory management policy:
27 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
28 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
29 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
30 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
33 #define DM_BUFIO_MIN_BUFFERS 8
35 #define DM_BUFIO_MEMORY_PERCENT 2
36 #define DM_BUFIO_VMALLOC_PERCENT 25
37 #define DM_BUFIO_WRITEBACK_RATIO 3
38 #define DM_BUFIO_LOW_WATERMARK_RATIO 16
41 * Check buffer ages in this interval (seconds)
43 #define DM_BUFIO_WORK_TIMER_SECS 30
46 * Free buffers when they are older than this (seconds)
48 #define DM_BUFIO_DEFAULT_AGE_SECS 300
51 * The nr of bytes of cached data to keep around.
53 #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
56 * Align buffer writes to this boundary.
57 * Tests show that SSDs have the highest IOPS when using 4k writes.
59 #define DM_BUFIO_WRITE_ALIGN 4096
62 * dm_buffer->list_mode
70 * All buffers are linked to buffer_tree with their node field.
72 * Clean buffers that are not being written (B_WRITING not set)
73 * are linked to lru[LIST_CLEAN] with their lru_list field.
75 * Dirty and clean buffers that are being written are linked to
76 * lru[LIST_DIRTY] with their lru_list field. When the write
77 * finishes, the buffer cannot be relinked immediately (because we
78 * are in an interrupt context and relinking requires process
79 * context), so some clean-not-writing buffers can be held on
80 * dirty_lru too. They are later added to lru in the process
83 struct dm_bufio_client {
88 struct list_head lru[LIST_SIZE];
89 unsigned long n_buffers[LIST_SIZE];
91 struct block_device *bdev;
92 unsigned int block_size;
93 s8 sectors_per_block_bits;
94 void (*alloc_callback)(struct dm_buffer *);
95 void (*write_callback)(struct dm_buffer *);
96 struct kmem_cache *slab_buffer;
97 struct kmem_cache *slab_cache;
98 struct dm_io_client *dm_io;
100 struct list_head reserved_buffers;
101 unsigned int need_reserved_buffers;
103 unsigned int minimum_buffers;
105 struct rb_root buffer_tree;
106 wait_queue_head_t free_buffer_wait;
110 int async_write_error;
112 struct list_head client_list;
114 struct shrinker shrinker;
115 struct work_struct shrink_work;
116 atomic_long_t need_shrink;
127 * Describes how the block was allocated:
128 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
129 * See the comment at alloc_buffer_data.
133 DATA_MODE_GET_FREE_PAGES = 1,
134 DATA_MODE_VMALLOC = 2,
140 struct list_head lru_list;
141 struct list_head global_list;
144 unsigned char data_mode; /* DATA_MODE_* */
145 unsigned char list_mode; /* LIST_* */
146 blk_status_t read_error;
147 blk_status_t write_error;
148 unsigned int accessed;
149 unsigned int hold_count;
151 unsigned long last_accessed;
152 unsigned int dirty_start;
153 unsigned int dirty_end;
154 unsigned int write_start;
155 unsigned int write_end;
156 struct dm_bufio_client *c;
157 struct list_head write_list;
158 void (*end_io)(struct dm_buffer *, blk_status_t);
159 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
161 unsigned int stack_len;
162 unsigned long stack_entries[MAX_STACK];
166 static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
168 /*----------------------------------------------------------------*/
170 #define dm_bufio_in_request() (!!current->bio_list)
172 static void dm_bufio_lock(struct dm_bufio_client *c)
174 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
175 spin_lock_bh(&c->spinlock);
177 mutex_lock_nested(&c->lock, dm_bufio_in_request());
180 static int dm_bufio_trylock(struct dm_bufio_client *c)
182 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
183 return spin_trylock_bh(&c->spinlock);
185 return mutex_trylock(&c->lock);
188 static void dm_bufio_unlock(struct dm_bufio_client *c)
190 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
191 spin_unlock_bh(&c->spinlock);
193 mutex_unlock(&c->lock);
196 /*----------------------------------------------------------------*/
199 * Default cache size: available memory divided by the ratio.
201 static unsigned long dm_bufio_default_cache_size;
204 * Total cache size set by the user.
206 static unsigned long dm_bufio_cache_size;
209 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
210 * at any time. If it disagrees, the user has changed cache size.
212 static unsigned long dm_bufio_cache_size_latch;
214 static DEFINE_SPINLOCK(global_spinlock);
216 static LIST_HEAD(global_queue);
218 static unsigned long global_num = 0;
221 * Buffers are freed after this timeout
223 static unsigned int dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
224 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
226 static unsigned long dm_bufio_peak_allocated;
227 static unsigned long dm_bufio_allocated_kmem_cache;
228 static unsigned long dm_bufio_allocated_get_free_pages;
229 static unsigned long dm_bufio_allocated_vmalloc;
230 static unsigned long dm_bufio_current_allocated;
232 /*----------------------------------------------------------------*/
235 * The current number of clients.
237 static int dm_bufio_client_count;
240 * The list of all clients.
242 static LIST_HEAD(dm_bufio_all_clients);
245 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
247 static DEFINE_MUTEX(dm_bufio_clients_lock);
249 static struct workqueue_struct *dm_bufio_wq;
250 static struct delayed_work dm_bufio_cleanup_old_work;
251 static struct work_struct dm_bufio_replacement_work;
254 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
255 static void buffer_record_stack(struct dm_buffer *b)
257 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
261 /*----------------------------------------------------------------
262 * A red/black tree acts as an index for all the buffers.
263 *--------------------------------------------------------------*/
264 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
266 struct rb_node *n = c->buffer_tree.rb_node;
270 b = container_of(n, struct dm_buffer, node);
272 if (b->block == block)
275 n = block < b->block ? n->rb_left : n->rb_right;
281 static struct dm_buffer *__find_next(struct dm_bufio_client *c, sector_t block)
283 struct rb_node *n = c->buffer_tree.rb_node;
285 struct dm_buffer *best = NULL;
288 b = container_of(n, struct dm_buffer, node);
290 if (b->block == block)
293 if (block <= b->block) {
304 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
306 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
307 struct dm_buffer *found;
310 found = container_of(*new, struct dm_buffer, node);
312 if (found->block == b->block) {
318 new = b->block < found->block ?
319 &found->node.rb_left : &found->node.rb_right;
322 rb_link_node(&b->node, parent, new);
323 rb_insert_color(&b->node, &c->buffer_tree);
326 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
328 rb_erase(&b->node, &c->buffer_tree);
331 /*----------------------------------------------------------------*/
333 static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
335 unsigned char data_mode;
338 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
339 &dm_bufio_allocated_kmem_cache,
340 &dm_bufio_allocated_get_free_pages,
341 &dm_bufio_allocated_vmalloc,
344 data_mode = b->data_mode;
345 diff = (long)b->c->block_size;
349 spin_lock(&global_spinlock);
351 *class_ptr[data_mode] += diff;
353 dm_bufio_current_allocated += diff;
355 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
356 dm_bufio_peak_allocated = dm_bufio_current_allocated;
361 list_add(&b->global_list, &global_queue);
363 if (dm_bufio_current_allocated > dm_bufio_cache_size)
364 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
366 list_del(&b->global_list);
370 spin_unlock(&global_spinlock);
374 * Change the number of clients and recalculate per-client limit.
376 static void __cache_size_refresh(void)
378 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
379 BUG_ON(dm_bufio_client_count < 0);
381 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
384 * Use default if set to 0 and report the actual cache size used.
386 if (!dm_bufio_cache_size_latch) {
387 (void)cmpxchg(&dm_bufio_cache_size, 0,
388 dm_bufio_default_cache_size);
389 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
394 * Allocating buffer data.
396 * Small buffers are allocated with kmem_cache, to use space optimally.
398 * For large buffers, we choose between get_free_pages and vmalloc.
399 * Each has advantages and disadvantages.
401 * __get_free_pages can randomly fail if the memory is fragmented.
402 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
403 * as low as 128M) so using it for caching is not appropriate.
405 * If the allocation may fail we use __get_free_pages. Memory fragmentation
406 * won't have a fatal effect here, but it just causes flushes of some other
407 * buffers and more I/O will be performed. Don't use __get_free_pages if it
408 * always fails (i.e. order >= MAX_ORDER).
410 * If the allocation shouldn't fail we use __vmalloc. This is only for the
411 * initial reserve allocation, so there's no risk of wasting all vmalloc
414 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
415 unsigned char *data_mode)
417 if (unlikely(c->slab_cache != NULL)) {
418 *data_mode = DATA_MODE_SLAB;
419 return kmem_cache_alloc(c->slab_cache, gfp_mask);
422 if (c->block_size <= KMALLOC_MAX_SIZE &&
423 gfp_mask & __GFP_NORETRY) {
424 *data_mode = DATA_MODE_GET_FREE_PAGES;
425 return (void *)__get_free_pages(gfp_mask,
426 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
429 *data_mode = DATA_MODE_VMALLOC;
432 * __vmalloc allocates the data pages and auxiliary structures with
433 * gfp_flags that were specified, but pagetables are always allocated
434 * with GFP_KERNEL, no matter what was specified as gfp_mask.
436 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
437 * all allocations done by this process (including pagetables) are done
438 * as if GFP_NOIO was specified.
440 if (gfp_mask & __GFP_NORETRY) {
441 unsigned int noio_flag = memalloc_noio_save();
442 void *ptr = __vmalloc(c->block_size, gfp_mask);
444 memalloc_noio_restore(noio_flag);
448 return __vmalloc(c->block_size, gfp_mask);
452 * Free buffer's data.
454 static void free_buffer_data(struct dm_bufio_client *c,
455 void *data, unsigned char data_mode)
459 kmem_cache_free(c->slab_cache, data);
462 case DATA_MODE_GET_FREE_PAGES:
463 free_pages((unsigned long)data,
464 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
467 case DATA_MODE_VMALLOC:
472 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
479 * Allocate buffer and its data.
481 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
483 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
490 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
492 kmem_cache_free(c->slab_buffer, b);
496 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
503 * Free buffer and its data.
505 static void free_buffer(struct dm_buffer *b)
507 struct dm_bufio_client *c = b->c;
509 free_buffer_data(c, b->data, b->data_mode);
510 kmem_cache_free(c->slab_buffer, b);
514 * Link buffer to the buffer tree and clean or dirty queue.
516 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
518 struct dm_bufio_client *c = b->c;
520 c->n_buffers[dirty]++;
522 b->list_mode = dirty;
523 list_add(&b->lru_list, &c->lru[dirty]);
525 b->last_accessed = jiffies;
527 adjust_total_allocated(b, false);
531 * Unlink buffer from the buffer tree and dirty or clean queue.
533 static void __unlink_buffer(struct dm_buffer *b)
535 struct dm_bufio_client *c = b->c;
537 BUG_ON(!c->n_buffers[b->list_mode]);
539 c->n_buffers[b->list_mode]--;
541 list_del(&b->lru_list);
543 adjust_total_allocated(b, true);
547 * Place the buffer to the head of dirty or clean LRU queue.
549 static void __relink_lru(struct dm_buffer *b, int dirty)
551 struct dm_bufio_client *c = b->c;
555 BUG_ON(!c->n_buffers[b->list_mode]);
557 c->n_buffers[b->list_mode]--;
558 c->n_buffers[dirty]++;
559 b->list_mode = dirty;
560 list_move(&b->lru_list, &c->lru[dirty]);
561 b->last_accessed = jiffies;
564 /*----------------------------------------------------------------
565 * Submit I/O on the buffer.
567 * Bio interface is faster but it has some problems:
568 * the vector list is limited (increasing this limit increases
569 * memory-consumption per buffer, so it is not viable);
571 * the memory must be direct-mapped, not vmalloced;
573 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
574 * it is not vmalloced, try using the bio interface.
576 * If the buffer is big, if it is vmalloced or if the underlying device
577 * rejects the bio because it is too large, use dm-io layer to do the I/O.
578 * The dm-io layer splits the I/O into multiple requests, avoiding the above
580 *--------------------------------------------------------------*/
583 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
584 * that the request was handled directly with bio interface.
586 static void dmio_complete(unsigned long error, void *context)
588 struct dm_buffer *b = context;
590 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
593 static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
594 unsigned int n_sectors, unsigned int offset)
597 struct dm_io_request io_req = {
599 .notify.fn = dmio_complete,
601 .client = b->c->dm_io,
603 struct dm_io_region region = {
609 if (b->data_mode != DATA_MODE_VMALLOC) {
610 io_req.mem.type = DM_IO_KMEM;
611 io_req.mem.ptr.addr = (char *)b->data + offset;
613 io_req.mem.type = DM_IO_VMA;
614 io_req.mem.ptr.vma = (char *)b->data + offset;
617 r = dm_io(&io_req, 1, ®ion, NULL);
619 b->end_io(b, errno_to_blk_status(r));
622 static void bio_complete(struct bio *bio)
624 struct dm_buffer *b = bio->bi_private;
625 blk_status_t status = bio->bi_status;
628 b->end_io(b, status);
631 static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
632 unsigned int n_sectors, unsigned int offset)
636 unsigned int vec_size, len;
638 vec_size = b->c->block_size >> PAGE_SHIFT;
639 if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
642 bio = bio_kmalloc(vec_size, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
645 use_dmio(b, op, sector, n_sectors, offset);
648 bio_init(bio, b->c->bdev, bio->bi_inline_vecs, vec_size, op);
649 bio->bi_iter.bi_sector = sector;
650 bio->bi_end_io = bio_complete;
653 ptr = (char *)b->data + offset;
654 len = n_sectors << SECTOR_SHIFT;
657 unsigned int this_step = min((unsigned int)(PAGE_SIZE - offset_in_page(ptr)), len);
658 if (!bio_add_page(bio, virt_to_page(ptr), this_step,
659 offset_in_page(ptr))) {
671 static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
675 if (likely(c->sectors_per_block_bits >= 0))
676 sector = block << c->sectors_per_block_bits;
678 sector = block * (c->block_size >> SECTOR_SHIFT);
684 static void submit_io(struct dm_buffer *b, enum req_op op,
685 void (*end_io)(struct dm_buffer *, blk_status_t))
687 unsigned int n_sectors;
689 unsigned int offset, end;
693 sector = block_to_sector(b->c, b->block);
695 if (op != REQ_OP_WRITE) {
696 n_sectors = b->c->block_size >> SECTOR_SHIFT;
699 if (b->c->write_callback)
700 b->c->write_callback(b);
701 offset = b->write_start;
703 offset &= -DM_BUFIO_WRITE_ALIGN;
704 end += DM_BUFIO_WRITE_ALIGN - 1;
705 end &= -DM_BUFIO_WRITE_ALIGN;
706 if (unlikely(end > b->c->block_size))
707 end = b->c->block_size;
709 sector += offset >> SECTOR_SHIFT;
710 n_sectors = (end - offset) >> SECTOR_SHIFT;
713 if (b->data_mode != DATA_MODE_VMALLOC)
714 use_bio(b, op, sector, n_sectors, offset);
716 use_dmio(b, op, sector, n_sectors, offset);
719 /*----------------------------------------------------------------
720 * Writing dirty buffers
721 *--------------------------------------------------------------*/
724 * The endio routine for write.
726 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
729 static void write_endio(struct dm_buffer *b, blk_status_t status)
731 b->write_error = status;
732 if (unlikely(status)) {
733 struct dm_bufio_client *c = b->c;
735 (void)cmpxchg(&c->async_write_error, 0,
736 blk_status_to_errno(status));
739 BUG_ON(!test_bit(B_WRITING, &b->state));
741 smp_mb__before_atomic();
742 clear_bit(B_WRITING, &b->state);
743 smp_mb__after_atomic();
745 wake_up_bit(&b->state, B_WRITING);
749 * Initiate a write on a dirty buffer, but don't wait for it.
751 * - If the buffer is not dirty, exit.
752 * - If there some previous write going on, wait for it to finish (we can't
753 * have two writes on the same buffer simultaneously).
754 * - Submit our write and don't wait on it. We set B_WRITING indicating
755 * that there is a write in progress.
757 static void __write_dirty_buffer(struct dm_buffer *b,
758 struct list_head *write_list)
760 if (!test_bit(B_DIRTY, &b->state))
763 clear_bit(B_DIRTY, &b->state);
764 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
766 b->write_start = b->dirty_start;
767 b->write_end = b->dirty_end;
770 submit_io(b, REQ_OP_WRITE, write_endio);
772 list_add_tail(&b->write_list, write_list);
775 static void __flush_write_list(struct list_head *write_list)
777 struct blk_plug plug;
778 blk_start_plug(&plug);
779 while (!list_empty(write_list)) {
780 struct dm_buffer *b =
781 list_entry(write_list->next, struct dm_buffer, write_list);
782 list_del(&b->write_list);
783 submit_io(b, REQ_OP_WRITE, write_endio);
786 blk_finish_plug(&plug);
790 * Wait until any activity on the buffer finishes. Possibly write the
791 * buffer if it is dirty. When this function finishes, there is no I/O
792 * running on the buffer and the buffer is not dirty.
794 static void __make_buffer_clean(struct dm_buffer *b)
796 BUG_ON(b->hold_count);
798 /* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
799 if (!smp_load_acquire(&b->state)) /* fast case */
802 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
803 __write_dirty_buffer(b, NULL);
804 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
808 * Find some buffer that is not held by anybody, clean it, unlink it and
811 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
815 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
816 BUG_ON(test_bit(B_WRITING, &b->state));
817 BUG_ON(test_bit(B_DIRTY, &b->state));
819 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
820 unlikely(test_bit_acquire(B_READING, &b->state)))
823 if (!b->hold_count) {
824 __make_buffer_clean(b);
831 if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
834 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
835 BUG_ON(test_bit(B_READING, &b->state));
837 if (!b->hold_count) {
838 __make_buffer_clean(b);
849 * Wait until some other threads free some buffer or release hold count on
852 * This function is entered with c->lock held, drops it and regains it
855 static void __wait_for_free_buffer(struct dm_bufio_client *c)
857 DECLARE_WAITQUEUE(wait, current);
859 add_wait_queue(&c->free_buffer_wait, &wait);
860 set_current_state(TASK_UNINTERRUPTIBLE);
865 remove_wait_queue(&c->free_buffer_wait, &wait);
878 * Allocate a new buffer. If the allocation is not possible, wait until
879 * some other thread frees a buffer.
881 * May drop the lock and regain it.
883 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
886 bool tried_noio_alloc = false;
889 * dm-bufio is resistant to allocation failures (it just keeps
890 * one buffer reserved in cases all the allocations fail).
891 * So set flags to not try too hard:
892 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
893 * mutex and wait ourselves.
894 * __GFP_NORETRY: don't retry and rather return failure
895 * __GFP_NOMEMALLOC: don't use emergency reserves
896 * __GFP_NOWARN: don't print a warning in case of failure
898 * For debugging, if we set the cache size to 1, no new buffers will
902 if (dm_bufio_cache_size_latch != 1) {
903 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
908 if (nf == NF_PREFETCH)
911 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
913 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
917 tried_noio_alloc = true;
920 if (!list_empty(&c->reserved_buffers)) {
921 b = list_entry(c->reserved_buffers.next,
922 struct dm_buffer, lru_list);
923 list_del(&b->lru_list);
924 c->need_reserved_buffers++;
929 b = __get_unclaimed_buffer(c);
933 __wait_for_free_buffer(c);
937 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
939 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
944 if (c->alloc_callback)
945 c->alloc_callback(b);
951 * Free a buffer and wake other threads waiting for free buffers.
953 static void __free_buffer_wake(struct dm_buffer *b)
955 struct dm_bufio_client *c = b->c;
957 if (!c->need_reserved_buffers)
960 list_add(&b->lru_list, &c->reserved_buffers);
961 c->need_reserved_buffers--;
964 wake_up(&c->free_buffer_wait);
967 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
968 struct list_head *write_list)
970 struct dm_buffer *b, *tmp;
972 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
973 BUG_ON(test_bit(B_READING, &b->state));
975 if (!test_bit(B_DIRTY, &b->state) &&
976 !test_bit(B_WRITING, &b->state)) {
977 __relink_lru(b, LIST_CLEAN);
981 if (no_wait && test_bit(B_WRITING, &b->state))
984 __write_dirty_buffer(b, write_list);
990 * Check if we're over watermark.
991 * If we are over threshold_buffers, start freeing buffers.
992 * If we're over "limit_buffers", block until we get under the limit.
994 static void __check_watermark(struct dm_bufio_client *c,
995 struct list_head *write_list)
997 if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
998 __write_dirty_buffers_async(c, 1, write_list);
1001 /*----------------------------------------------------------------
1003 *--------------------------------------------------------------*/
1005 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1006 enum new_flag nf, int *need_submit,
1007 struct list_head *write_list)
1009 struct dm_buffer *b, *new_b = NULL;
1013 b = __find(c, block);
1020 new_b = __alloc_buffer_wait(c, nf);
1025 * We've had a period where the mutex was unlocked, so need to
1026 * recheck the buffer tree.
1028 b = __find(c, block);
1030 __free_buffer_wake(new_b);
1034 __check_watermark(c, write_list);
1040 __link_buffer(b, block, LIST_CLEAN);
1042 if (nf == NF_FRESH) {
1047 b->state = 1 << B_READING;
1053 if (nf == NF_PREFETCH)
1056 * Note: it is essential that we don't wait for the buffer to be
1057 * read if dm_bufio_get function is used. Both dm_bufio_get and
1058 * dm_bufio_prefetch can be used in the driver request routine.
1059 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1060 * the same buffer, it would deadlock if we waited.
1062 if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state)))
1066 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1067 test_bit(B_WRITING, &b->state));
1072 * The endio routine for reading: set the error, clear the bit and wake up
1073 * anyone waiting on the buffer.
1075 static void read_endio(struct dm_buffer *b, blk_status_t status)
1077 b->read_error = status;
1079 BUG_ON(!test_bit(B_READING, &b->state));
1081 smp_mb__before_atomic();
1082 clear_bit(B_READING, &b->state);
1083 smp_mb__after_atomic();
1085 wake_up_bit(&b->state, B_READING);
1089 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1090 * functions is similar except that dm_bufio_new doesn't read the
1091 * buffer from the disk (assuming that the caller overwrites all the data
1092 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1094 static void *new_read(struct dm_bufio_client *c, sector_t block,
1095 enum new_flag nf, struct dm_buffer **bp)
1098 struct dm_buffer *b;
1100 LIST_HEAD(write_list);
1103 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1104 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1105 if (b && b->hold_count == 1)
1106 buffer_record_stack(b);
1110 __flush_write_list(&write_list);
1116 submit_io(b, REQ_OP_READ, read_endio);
1118 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1120 if (b->read_error) {
1121 int error = blk_status_to_errno(b->read_error);
1123 dm_bufio_release(b);
1125 return ERR_PTR(error);
1133 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1134 struct dm_buffer **bp)
1136 return new_read(c, block, NF_GET, bp);
1138 EXPORT_SYMBOL_GPL(dm_bufio_get);
1140 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1141 struct dm_buffer **bp)
1143 BUG_ON(dm_bufio_in_request());
1145 return new_read(c, block, NF_READ, bp);
1147 EXPORT_SYMBOL_GPL(dm_bufio_read);
1149 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1150 struct dm_buffer **bp)
1152 BUG_ON(dm_bufio_in_request());
1154 return new_read(c, block, NF_FRESH, bp);
1156 EXPORT_SYMBOL_GPL(dm_bufio_new);
1158 void dm_bufio_prefetch(struct dm_bufio_client *c,
1159 sector_t block, unsigned int n_blocks)
1161 struct blk_plug plug;
1163 LIST_HEAD(write_list);
1165 BUG_ON(dm_bufio_in_request());
1167 blk_start_plug(&plug);
1170 for (; n_blocks--; block++) {
1172 struct dm_buffer *b;
1173 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1175 if (unlikely(!list_empty(&write_list))) {
1177 blk_finish_plug(&plug);
1178 __flush_write_list(&write_list);
1179 blk_start_plug(&plug);
1182 if (unlikely(b != NULL)) {
1186 submit_io(b, REQ_OP_READ, read_endio);
1187 dm_bufio_release(b);
1200 blk_finish_plug(&plug);
1202 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1204 void dm_bufio_release(struct dm_buffer *b)
1206 struct dm_bufio_client *c = b->c;
1210 BUG_ON(!b->hold_count);
1213 if (!b->hold_count) {
1214 wake_up(&c->free_buffer_wait);
1217 * If there were errors on the buffer, and the buffer is not
1218 * to be written, free the buffer. There is no point in caching
1221 if ((b->read_error || b->write_error) &&
1222 !test_bit_acquire(B_READING, &b->state) &&
1223 !test_bit(B_WRITING, &b->state) &&
1224 !test_bit(B_DIRTY, &b->state)) {
1226 __free_buffer_wake(b);
1232 EXPORT_SYMBOL_GPL(dm_bufio_release);
1234 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1235 unsigned int start, unsigned int end)
1237 struct dm_bufio_client *c = b->c;
1239 BUG_ON(start >= end);
1240 BUG_ON(end > b->c->block_size);
1244 BUG_ON(test_bit(B_READING, &b->state));
1246 if (!test_and_set_bit(B_DIRTY, &b->state)) {
1247 b->dirty_start = start;
1249 __relink_lru(b, LIST_DIRTY);
1251 if (start < b->dirty_start)
1252 b->dirty_start = start;
1253 if (end > b->dirty_end)
1259 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1261 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1263 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1265 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1267 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1269 LIST_HEAD(write_list);
1271 BUG_ON(dm_bufio_in_request());
1274 __write_dirty_buffers_async(c, 0, &write_list);
1276 __flush_write_list(&write_list);
1278 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1281 * For performance, it is essential that the buffers are written asynchronously
1282 * and simultaneously (so that the block layer can merge the writes) and then
1285 * Finally, we flush hardware disk cache.
1287 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1290 unsigned long buffers_processed = 0;
1291 struct dm_buffer *b, *tmp;
1293 LIST_HEAD(write_list);
1296 __write_dirty_buffers_async(c, 0, &write_list);
1298 __flush_write_list(&write_list);
1302 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1303 int dropped_lock = 0;
1305 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1306 buffers_processed++;
1308 BUG_ON(test_bit(B_READING, &b->state));
1310 if (test_bit(B_WRITING, &b->state)) {
1311 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1315 wait_on_bit_io(&b->state, B_WRITING,
1316 TASK_UNINTERRUPTIBLE);
1320 wait_on_bit_io(&b->state, B_WRITING,
1321 TASK_UNINTERRUPTIBLE);
1324 if (!test_bit(B_DIRTY, &b->state) &&
1325 !test_bit(B_WRITING, &b->state))
1326 __relink_lru(b, LIST_CLEAN);
1331 * If we dropped the lock, the list is no longer consistent,
1332 * so we must restart the search.
1334 * In the most common case, the buffer just processed is
1335 * relinked to the clean list, so we won't loop scanning the
1336 * same buffer again and again.
1338 * This may livelock if there is another thread simultaneously
1339 * dirtying buffers, so we count the number of buffers walked
1340 * and if it exceeds the total number of buffers, it means that
1341 * someone is doing some writes simultaneously with us. In
1342 * this case, stop, dropping the lock.
1347 wake_up(&c->free_buffer_wait);
1350 a = xchg(&c->async_write_error, 0);
1351 f = dm_bufio_issue_flush(c);
1357 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1360 * Use dm-io to send an empty barrier to flush the device.
1362 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1364 struct dm_io_request io_req = {
1365 .bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
1366 .mem.type = DM_IO_KMEM,
1367 .mem.ptr.addr = NULL,
1370 struct dm_io_region io_reg = {
1376 BUG_ON(dm_bufio_in_request());
1378 return dm_io(&io_req, 1, &io_reg, NULL);
1380 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1383 * Use dm-io to send a discard request to flush the device.
1385 int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
1387 struct dm_io_request io_req = {
1388 .bi_opf = REQ_OP_DISCARD | REQ_SYNC,
1389 .mem.type = DM_IO_KMEM,
1390 .mem.ptr.addr = NULL,
1393 struct dm_io_region io_reg = {
1395 .sector = block_to_sector(c, block),
1396 .count = block_to_sector(c, count),
1399 BUG_ON(dm_bufio_in_request());
1401 return dm_io(&io_req, 1, &io_reg, NULL);
1403 EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
1406 * We first delete any other buffer that may be at that new location.
1408 * Then, we write the buffer to the original location if it was dirty.
1410 * Then, if we are the only one who is holding the buffer, relink the buffer
1411 * in the buffer tree for the new location.
1413 * If there was someone else holding the buffer, we write it to the new
1414 * location but not relink it, because that other user needs to have the buffer
1415 * at the same place.
1417 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1419 struct dm_bufio_client *c = b->c;
1420 struct dm_buffer *new;
1422 BUG_ON(dm_bufio_in_request());
1427 new = __find(c, new_block);
1429 if (new->hold_count) {
1430 __wait_for_free_buffer(c);
1435 * FIXME: Is there any point waiting for a write that's going
1436 * to be overwritten in a bit?
1438 __make_buffer_clean(new);
1439 __unlink_buffer(new);
1440 __free_buffer_wake(new);
1443 BUG_ON(!b->hold_count);
1444 BUG_ON(test_bit(B_READING, &b->state));
1446 __write_dirty_buffer(b, NULL);
1447 if (b->hold_count == 1) {
1448 wait_on_bit_io(&b->state, B_WRITING,
1449 TASK_UNINTERRUPTIBLE);
1450 set_bit(B_DIRTY, &b->state);
1452 b->dirty_end = c->block_size;
1454 __link_buffer(b, new_block, LIST_DIRTY);
1457 wait_on_bit_lock_io(&b->state, B_WRITING,
1458 TASK_UNINTERRUPTIBLE);
1460 * Relink buffer to "new_block" so that write_callback
1461 * sees "new_block" as a block number.
1462 * After the write, link the buffer back to old_block.
1463 * All this must be done in bufio lock, so that block number
1464 * change isn't visible to other threads.
1466 old_block = b->block;
1468 __link_buffer(b, new_block, b->list_mode);
1469 submit_io(b, REQ_OP_WRITE, write_endio);
1470 wait_on_bit_io(&b->state, B_WRITING,
1471 TASK_UNINTERRUPTIBLE);
1473 __link_buffer(b, old_block, b->list_mode);
1477 dm_bufio_release(b);
1479 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1481 static void forget_buffer_locked(struct dm_buffer *b)
1483 if (likely(!b->hold_count) && likely(!smp_load_acquire(&b->state))) {
1485 __free_buffer_wake(b);
1490 * Free the given buffer.
1492 * This is just a hint, if the buffer is in use or dirty, this function
1495 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1497 struct dm_buffer *b;
1501 b = __find(c, block);
1503 forget_buffer_locked(b);
1507 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1509 void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
1511 struct dm_buffer *b;
1512 sector_t end_block = block + n_blocks;
1514 while (block < end_block) {
1517 b = __find_next(c, block);
1519 block = b->block + 1;
1520 forget_buffer_locked(b);
1530 EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
1532 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned int n)
1534 c->minimum_buffers = n;
1536 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1538 unsigned int dm_bufio_get_block_size(struct dm_bufio_client *c)
1540 return c->block_size;
1542 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1544 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1546 sector_t s = bdev_nr_sectors(c->bdev);
1551 if (likely(c->sectors_per_block_bits >= 0))
1552 s >>= c->sectors_per_block_bits;
1554 sector_div(s, c->block_size >> SECTOR_SHIFT);
1557 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1559 struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1563 EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1565 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1569 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1571 void *dm_bufio_get_block_data(struct dm_buffer *b)
1575 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1577 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1581 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1583 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1587 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1589 static void drop_buffers(struct dm_bufio_client *c)
1591 struct dm_buffer *b;
1593 bool warned = false;
1595 BUG_ON(dm_bufio_in_request());
1598 * An optimization so that the buffers are not written one-by-one.
1600 dm_bufio_write_dirty_buffers_async(c);
1604 while ((b = __get_unclaimed_buffer(c)))
1605 __free_buffer_wake(b);
1607 for (i = 0; i < LIST_SIZE; i++)
1608 list_for_each_entry(b, &c->lru[i], lru_list) {
1611 DMERR("leaked buffer %llx, hold count %u, list %d",
1612 (unsigned long long)b->block, b->hold_count, i);
1613 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1614 stack_trace_print(b->stack_entries, b->stack_len, 1);
1615 /* mark unclaimed to avoid BUG_ON below */
1620 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1621 while ((b = __get_unclaimed_buffer(c)))
1622 __free_buffer_wake(b);
1625 for (i = 0; i < LIST_SIZE; i++)
1626 BUG_ON(!list_empty(&c->lru[i]));
1632 * We may not be able to evict this buffer if IO pending or the client
1633 * is still using it. Caller is expected to know buffer is too old.
1635 * And if GFP_NOFS is used, we must not do any I/O because we hold
1636 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1637 * rerouted to different bufio client.
1639 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1641 if (!(gfp & __GFP_FS) ||
1642 (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
1643 if (test_bit_acquire(B_READING, &b->state) ||
1644 test_bit(B_WRITING, &b->state) ||
1645 test_bit(B_DIRTY, &b->state))
1652 __make_buffer_clean(b);
1654 __free_buffer_wake(b);
1659 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1661 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1662 if (likely(c->sectors_per_block_bits >= 0))
1663 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1665 retain_bytes /= c->block_size;
1666 return retain_bytes;
1669 static void __scan(struct dm_bufio_client *c)
1672 struct dm_buffer *b, *tmp;
1673 unsigned long freed = 0;
1674 unsigned long count = c->n_buffers[LIST_CLEAN] +
1675 c->n_buffers[LIST_DIRTY];
1676 unsigned long retain_target = get_retain_buffers(c);
1678 for (l = 0; l < LIST_SIZE; l++) {
1679 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1680 if (count - freed <= retain_target)
1681 atomic_long_set(&c->need_shrink, 0);
1682 if (!atomic_long_read(&c->need_shrink))
1684 if (__try_evict_buffer(b, GFP_KERNEL)) {
1685 atomic_long_dec(&c->need_shrink);
1693 static void shrink_work(struct work_struct *w)
1695 struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
1702 static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1704 struct dm_bufio_client *c;
1706 c = container_of(shrink, struct dm_bufio_client, shrinker);
1707 atomic_long_add(sc->nr_to_scan, &c->need_shrink);
1708 queue_work(dm_bufio_wq, &c->shrink_work);
1710 return sc->nr_to_scan;
1713 static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1715 struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1716 unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1717 READ_ONCE(c->n_buffers[LIST_DIRTY]);
1718 unsigned long retain_target = get_retain_buffers(c);
1719 unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
1721 if (unlikely(count < retain_target))
1724 count -= retain_target;
1726 if (unlikely(count < queued_for_cleanup))
1729 count -= queued_for_cleanup;
1735 * Create the buffering interface
1737 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned int block_size,
1738 unsigned int reserved_buffers, unsigned int aux_size,
1739 void (*alloc_callback)(struct dm_buffer *),
1740 void (*write_callback)(struct dm_buffer *),
1744 struct dm_bufio_client *c;
1748 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1749 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1754 c = kzalloc(sizeof(*c), GFP_KERNEL);
1759 c->buffer_tree = RB_ROOT;
1762 c->block_size = block_size;
1763 if (is_power_of_2(block_size))
1764 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1766 c->sectors_per_block_bits = -1;
1768 c->alloc_callback = alloc_callback;
1769 c->write_callback = write_callback;
1771 if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
1773 static_branch_inc(&no_sleep_enabled);
1776 for (i = 0; i < LIST_SIZE; i++) {
1777 INIT_LIST_HEAD(&c->lru[i]);
1778 c->n_buffers[i] = 0;
1781 mutex_init(&c->lock);
1782 spin_lock_init(&c->spinlock);
1783 INIT_LIST_HEAD(&c->reserved_buffers);
1784 c->need_reserved_buffers = reserved_buffers;
1786 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1788 init_waitqueue_head(&c->free_buffer_wait);
1789 c->async_write_error = 0;
1791 c->dm_io = dm_io_client_create();
1792 if (IS_ERR(c->dm_io)) {
1793 r = PTR_ERR(c->dm_io);
1797 if (block_size <= KMALLOC_MAX_SIZE &&
1798 (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1799 unsigned int align = min(1U << __ffs(block_size), (unsigned int)PAGE_SIZE);
1800 snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1801 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1802 SLAB_RECLAIM_ACCOUNT, NULL);
1803 if (!c->slab_cache) {
1809 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1811 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1812 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1813 0, SLAB_RECLAIM_ACCOUNT, NULL);
1814 if (!c->slab_buffer) {
1819 while (c->need_reserved_buffers) {
1820 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1826 __free_buffer_wake(b);
1829 INIT_WORK(&c->shrink_work, shrink_work);
1830 atomic_long_set(&c->need_shrink, 0);
1832 c->shrinker.count_objects = dm_bufio_shrink_count;
1833 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1834 c->shrinker.seeks = 1;
1835 c->shrinker.batch = 0;
1836 r = register_shrinker(&c->shrinker, "dm-bufio:(%u:%u)",
1837 MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
1841 mutex_lock(&dm_bufio_clients_lock);
1842 dm_bufio_client_count++;
1843 list_add(&c->client_list, &dm_bufio_all_clients);
1844 __cache_size_refresh();
1845 mutex_unlock(&dm_bufio_clients_lock);
1850 while (!list_empty(&c->reserved_buffers)) {
1851 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1852 struct dm_buffer, lru_list);
1853 list_del(&b->lru_list);
1856 kmem_cache_destroy(c->slab_cache);
1857 kmem_cache_destroy(c->slab_buffer);
1858 dm_io_client_destroy(c->dm_io);
1860 mutex_destroy(&c->lock);
1862 static_branch_dec(&no_sleep_enabled);
1867 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1870 * Free the buffering interface.
1871 * It is required that there are no references on any buffers.
1873 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1879 unregister_shrinker(&c->shrinker);
1880 flush_work(&c->shrink_work);
1882 mutex_lock(&dm_bufio_clients_lock);
1884 list_del(&c->client_list);
1885 dm_bufio_client_count--;
1886 __cache_size_refresh();
1888 mutex_unlock(&dm_bufio_clients_lock);
1890 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1891 BUG_ON(c->need_reserved_buffers);
1893 while (!list_empty(&c->reserved_buffers)) {
1894 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1895 struct dm_buffer, lru_list);
1896 list_del(&b->lru_list);
1900 for (i = 0; i < LIST_SIZE; i++)
1901 if (c->n_buffers[i])
1902 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1904 for (i = 0; i < LIST_SIZE; i++)
1905 BUG_ON(c->n_buffers[i]);
1907 kmem_cache_destroy(c->slab_cache);
1908 kmem_cache_destroy(c->slab_buffer);
1909 dm_io_client_destroy(c->dm_io);
1910 mutex_destroy(&c->lock);
1912 static_branch_dec(&no_sleep_enabled);
1915 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1917 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1921 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1923 static unsigned int get_max_age_hz(void)
1925 unsigned int max_age = READ_ONCE(dm_bufio_max_age);
1927 if (max_age > UINT_MAX / HZ)
1928 max_age = UINT_MAX / HZ;
1930 return max_age * HZ;
1933 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1935 return time_after_eq(jiffies, b->last_accessed + age_hz);
1938 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1940 struct dm_buffer *b, *tmp;
1941 unsigned long retain_target = get_retain_buffers(c);
1942 unsigned long count;
1943 LIST_HEAD(write_list);
1947 __check_watermark(c, &write_list);
1948 if (unlikely(!list_empty(&write_list))) {
1950 __flush_write_list(&write_list);
1954 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1955 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1956 if (count <= retain_target)
1959 if (!older_than(b, age_hz))
1962 if (__try_evict_buffer(b, 0))
1971 static void do_global_cleanup(struct work_struct *w)
1973 struct dm_bufio_client *locked_client = NULL;
1974 struct dm_bufio_client *current_client;
1975 struct dm_buffer *b;
1976 unsigned int spinlock_hold_count;
1977 unsigned long threshold = dm_bufio_cache_size -
1978 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1979 unsigned long loops = global_num * 2;
1981 mutex_lock(&dm_bufio_clients_lock);
1986 spin_lock(&global_spinlock);
1987 if (unlikely(dm_bufio_current_allocated <= threshold))
1990 spinlock_hold_count = 0;
1994 if (unlikely(list_empty(&global_queue)))
1996 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
2000 list_move(&b->global_list, &global_queue);
2001 if (likely(++spinlock_hold_count < 16))
2003 spin_unlock(&global_spinlock);
2007 current_client = b->c;
2008 if (unlikely(current_client != locked_client)) {
2010 dm_bufio_unlock(locked_client);
2012 if (!dm_bufio_trylock(current_client)) {
2013 spin_unlock(&global_spinlock);
2014 dm_bufio_lock(current_client);
2015 locked_client = current_client;
2019 locked_client = current_client;
2022 spin_unlock(&global_spinlock);
2024 if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
2025 spin_lock(&global_spinlock);
2026 list_move(&b->global_list, &global_queue);
2027 spin_unlock(&global_spinlock);
2031 spin_unlock(&global_spinlock);
2034 dm_bufio_unlock(locked_client);
2036 mutex_unlock(&dm_bufio_clients_lock);
2039 static void cleanup_old_buffers(void)
2041 unsigned long max_age_hz = get_max_age_hz();
2042 struct dm_bufio_client *c;
2044 mutex_lock(&dm_bufio_clients_lock);
2046 __cache_size_refresh();
2048 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2049 __evict_old_buffers(c, max_age_hz);
2051 mutex_unlock(&dm_bufio_clients_lock);
2054 static void work_fn(struct work_struct *w)
2056 cleanup_old_buffers();
2058 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2059 DM_BUFIO_WORK_TIMER_SECS * HZ);
2062 /*----------------------------------------------------------------
2064 *--------------------------------------------------------------*/
2067 * This is called only once for the whole dm_bufio module.
2068 * It initializes memory limit.
2070 static int __init dm_bufio_init(void)
2074 dm_bufio_allocated_kmem_cache = 0;
2075 dm_bufio_allocated_get_free_pages = 0;
2076 dm_bufio_allocated_vmalloc = 0;
2077 dm_bufio_current_allocated = 0;
2079 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2080 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2082 if (mem > ULONG_MAX)
2086 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2087 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2090 dm_bufio_default_cache_size = mem;
2092 mutex_lock(&dm_bufio_clients_lock);
2093 __cache_size_refresh();
2094 mutex_unlock(&dm_bufio_clients_lock);
2096 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2100 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2101 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2102 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2103 DM_BUFIO_WORK_TIMER_SECS * HZ);
2109 * This is called once when unloading the dm_bufio module.
2111 static void __exit dm_bufio_exit(void)
2115 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2116 destroy_workqueue(dm_bufio_wq);
2118 if (dm_bufio_client_count) {
2119 DMCRIT("%s: dm_bufio_client_count leaked: %d",
2120 __func__, dm_bufio_client_count);
2124 if (dm_bufio_current_allocated) {
2125 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2126 __func__, dm_bufio_current_allocated);
2130 if (dm_bufio_allocated_get_free_pages) {
2131 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2132 __func__, dm_bufio_allocated_get_free_pages);
2136 if (dm_bufio_allocated_vmalloc) {
2137 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2138 __func__, dm_bufio_allocated_vmalloc);
2145 module_init(dm_bufio_init)
2146 module_exit(dm_bufio_exit)
2148 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2149 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2151 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2152 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2154 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2155 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2157 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2158 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2160 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2161 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2163 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2164 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2166 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2167 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2169 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2170 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2172 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2173 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2174 MODULE_LICENSE("GPL");