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>
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_RATIO 3
37 #define DM_BUFIO_LOW_WATERMARK_RATIO 16
40 * Check buffer ages in this interval (seconds)
42 #define DM_BUFIO_WORK_TIMER_SECS 30
45 * Free buffers when they are older than this (seconds)
47 #define DM_BUFIO_DEFAULT_AGE_SECS 300
50 * The nr of bytes of cached data to keep around.
52 #define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
55 * Align buffer writes to this boundary.
56 * Tests show that SSDs have the highest IOPS when using 4k writes.
58 #define DM_BUFIO_WRITE_ALIGN 4096
61 * dm_buffer->list_mode
69 * All buffers are linked to buffer_tree with their node field.
71 * Clean buffers that are not being written (B_WRITING not set)
72 * are linked to lru[LIST_CLEAN] with their lru_list field.
74 * Dirty and clean buffers that are being written are linked to
75 * lru[LIST_DIRTY] with their lru_list field. When the write
76 * finishes, the buffer cannot be relinked immediately (because we
77 * are in an interrupt context and relinking requires process
78 * context), so some clean-not-writing buffers can be held on
79 * dirty_lru too. They are later added to lru in the process
82 struct dm_bufio_client {
85 struct list_head lru[LIST_SIZE];
86 unsigned long n_buffers[LIST_SIZE];
88 struct block_device *bdev;
90 s8 sectors_per_block_bits;
91 void (*alloc_callback)(struct dm_buffer *);
92 void (*write_callback)(struct dm_buffer *);
94 struct kmem_cache *slab_buffer;
95 struct kmem_cache *slab_cache;
96 struct dm_io_client *dm_io;
98 struct list_head reserved_buffers;
99 unsigned need_reserved_buffers;
101 unsigned minimum_buffers;
103 struct rb_root buffer_tree;
104 wait_queue_head_t free_buffer_wait;
108 int async_write_error;
110 struct list_head client_list;
112 struct shrinker shrinker;
113 struct work_struct shrink_work;
114 atomic_long_t need_shrink;
125 * Describes how the block was allocated:
126 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
127 * See the comment at alloc_buffer_data.
131 DATA_MODE_GET_FREE_PAGES = 1,
132 DATA_MODE_VMALLOC = 2,
138 struct list_head lru_list;
139 struct list_head global_list;
142 unsigned char data_mode; /* DATA_MODE_* */
143 unsigned char list_mode; /* LIST_* */
144 blk_status_t read_error;
145 blk_status_t write_error;
149 unsigned long last_accessed;
150 unsigned dirty_start;
152 unsigned write_start;
154 struct dm_bufio_client *c;
155 struct list_head write_list;
156 void (*end_io)(struct dm_buffer *, blk_status_t);
157 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
159 unsigned int stack_len;
160 unsigned long stack_entries[MAX_STACK];
164 /*----------------------------------------------------------------*/
166 #define dm_bufio_in_request() (!!current->bio_list)
168 static void dm_bufio_lock(struct dm_bufio_client *c)
170 mutex_lock_nested(&c->lock, dm_bufio_in_request());
173 static int dm_bufio_trylock(struct dm_bufio_client *c)
175 return mutex_trylock(&c->lock);
178 static void dm_bufio_unlock(struct dm_bufio_client *c)
180 mutex_unlock(&c->lock);
183 /*----------------------------------------------------------------*/
186 * Default cache size: available memory divided by the ratio.
188 static unsigned long dm_bufio_default_cache_size;
191 * Total cache size set by the user.
193 static unsigned long dm_bufio_cache_size;
196 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
197 * at any time. If it disagrees, the user has changed cache size.
199 static unsigned long dm_bufio_cache_size_latch;
201 static DEFINE_SPINLOCK(global_spinlock);
203 static LIST_HEAD(global_queue);
205 static unsigned long global_num = 0;
208 * Buffers are freed after this timeout
210 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
211 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
213 static unsigned long dm_bufio_peak_allocated;
214 static unsigned long dm_bufio_allocated_kmem_cache;
215 static unsigned long dm_bufio_allocated_get_free_pages;
216 static unsigned long dm_bufio_allocated_vmalloc;
217 static unsigned long dm_bufio_current_allocated;
219 /*----------------------------------------------------------------*/
222 * The current number of clients.
224 static int dm_bufio_client_count;
227 * The list of all clients.
229 static LIST_HEAD(dm_bufio_all_clients);
232 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
234 static DEFINE_MUTEX(dm_bufio_clients_lock);
236 static struct workqueue_struct *dm_bufio_wq;
237 static struct delayed_work dm_bufio_cleanup_old_work;
238 static struct work_struct dm_bufio_replacement_work;
241 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
242 static void buffer_record_stack(struct dm_buffer *b)
244 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
248 /*----------------------------------------------------------------
249 * A red/black tree acts as an index for all the buffers.
250 *--------------------------------------------------------------*/
251 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
253 struct rb_node *n = c->buffer_tree.rb_node;
257 b = container_of(n, struct dm_buffer, node);
259 if (b->block == block)
262 n = block < b->block ? n->rb_left : n->rb_right;
268 static struct dm_buffer *__find_next(struct dm_bufio_client *c, sector_t block)
270 struct rb_node *n = c->buffer_tree.rb_node;
272 struct dm_buffer *best = NULL;
275 b = container_of(n, struct dm_buffer, node);
277 if (b->block == block)
280 if (block <= b->block) {
291 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
293 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
294 struct dm_buffer *found;
297 found = container_of(*new, struct dm_buffer, node);
299 if (found->block == b->block) {
305 new = b->block < found->block ?
306 &found->node.rb_left : &found->node.rb_right;
309 rb_link_node(&b->node, parent, new);
310 rb_insert_color(&b->node, &c->buffer_tree);
313 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
315 rb_erase(&b->node, &c->buffer_tree);
318 /*----------------------------------------------------------------*/
320 static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
322 unsigned char data_mode;
325 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
326 &dm_bufio_allocated_kmem_cache,
327 &dm_bufio_allocated_get_free_pages,
328 &dm_bufio_allocated_vmalloc,
331 data_mode = b->data_mode;
332 diff = (long)b->c->block_size;
336 spin_lock(&global_spinlock);
338 *class_ptr[data_mode] += diff;
340 dm_bufio_current_allocated += diff;
342 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
343 dm_bufio_peak_allocated = dm_bufio_current_allocated;
348 list_add(&b->global_list, &global_queue);
350 if (dm_bufio_current_allocated > dm_bufio_cache_size)
351 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
353 list_del(&b->global_list);
357 spin_unlock(&global_spinlock);
361 * Change the number of clients and recalculate per-client limit.
363 static void __cache_size_refresh(void)
365 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
366 BUG_ON(dm_bufio_client_count < 0);
368 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
371 * Use default if set to 0 and report the actual cache size used.
373 if (!dm_bufio_cache_size_latch) {
374 (void)cmpxchg(&dm_bufio_cache_size, 0,
375 dm_bufio_default_cache_size);
376 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
381 * Allocating buffer data.
383 * Small buffers are allocated with kmem_cache, to use space optimally.
385 * For large buffers, we choose between get_free_pages and vmalloc.
386 * Each has advantages and disadvantages.
388 * __get_free_pages can randomly fail if the memory is fragmented.
389 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
390 * as low as 128M) so using it for caching is not appropriate.
392 * If the allocation may fail we use __get_free_pages. Memory fragmentation
393 * won't have a fatal effect here, but it just causes flushes of some other
394 * buffers and more I/O will be performed. Don't use __get_free_pages if it
395 * always fails (i.e. order >= MAX_ORDER).
397 * If the allocation shouldn't fail we use __vmalloc. This is only for the
398 * initial reserve allocation, so there's no risk of wasting all vmalloc
401 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
402 unsigned char *data_mode)
404 if (unlikely(c->slab_cache != NULL)) {
405 *data_mode = DATA_MODE_SLAB;
406 return kmem_cache_alloc(c->slab_cache, gfp_mask);
409 if (c->block_size <= KMALLOC_MAX_SIZE &&
410 gfp_mask & __GFP_NORETRY) {
411 *data_mode = DATA_MODE_GET_FREE_PAGES;
412 return (void *)__get_free_pages(gfp_mask,
413 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
416 *data_mode = DATA_MODE_VMALLOC;
419 * __vmalloc allocates the data pages and auxiliary structures with
420 * gfp_flags that were specified, but pagetables are always allocated
421 * with GFP_KERNEL, no matter what was specified as gfp_mask.
423 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
424 * all allocations done by this process (including pagetables) are done
425 * as if GFP_NOIO was specified.
427 if (gfp_mask & __GFP_NORETRY) {
428 unsigned noio_flag = memalloc_noio_save();
429 void *ptr = __vmalloc(c->block_size, gfp_mask);
431 memalloc_noio_restore(noio_flag);
435 return __vmalloc(c->block_size, gfp_mask);
439 * Free buffer's data.
441 static void free_buffer_data(struct dm_bufio_client *c,
442 void *data, unsigned char data_mode)
446 kmem_cache_free(c->slab_cache, data);
449 case DATA_MODE_GET_FREE_PAGES:
450 free_pages((unsigned long)data,
451 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
454 case DATA_MODE_VMALLOC:
459 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
466 * Allocate buffer and its data.
468 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
470 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
477 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
479 kmem_cache_free(c->slab_buffer, b);
483 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
490 * Free buffer and its data.
492 static void free_buffer(struct dm_buffer *b)
494 struct dm_bufio_client *c = b->c;
496 free_buffer_data(c, b->data, b->data_mode);
497 kmem_cache_free(c->slab_buffer, b);
501 * Link buffer to the buffer tree and clean or dirty queue.
503 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
505 struct dm_bufio_client *c = b->c;
507 c->n_buffers[dirty]++;
509 b->list_mode = dirty;
510 list_add(&b->lru_list, &c->lru[dirty]);
512 b->last_accessed = jiffies;
514 adjust_total_allocated(b, false);
518 * Unlink buffer from the buffer tree and dirty or clean queue.
520 static void __unlink_buffer(struct dm_buffer *b)
522 struct dm_bufio_client *c = b->c;
524 BUG_ON(!c->n_buffers[b->list_mode]);
526 c->n_buffers[b->list_mode]--;
528 list_del(&b->lru_list);
530 adjust_total_allocated(b, true);
534 * Place the buffer to the head of dirty or clean LRU queue.
536 static void __relink_lru(struct dm_buffer *b, int dirty)
538 struct dm_bufio_client *c = b->c;
542 BUG_ON(!c->n_buffers[b->list_mode]);
544 c->n_buffers[b->list_mode]--;
545 c->n_buffers[dirty]++;
546 b->list_mode = dirty;
547 list_move(&b->lru_list, &c->lru[dirty]);
548 b->last_accessed = jiffies;
551 /*----------------------------------------------------------------
552 * Submit I/O on the buffer.
554 * Bio interface is faster but it has some problems:
555 * the vector list is limited (increasing this limit increases
556 * memory-consumption per buffer, so it is not viable);
558 * the memory must be direct-mapped, not vmalloced;
560 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
561 * it is not vmalloced, try using the bio interface.
563 * If the buffer is big, if it is vmalloced or if the underlying device
564 * rejects the bio because it is too large, use dm-io layer to do the I/O.
565 * The dm-io layer splits the I/O into multiple requests, avoiding the above
567 *--------------------------------------------------------------*/
570 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
571 * that the request was handled directly with bio interface.
573 static void dmio_complete(unsigned long error, void *context)
575 struct dm_buffer *b = context;
577 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
580 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
581 unsigned n_sectors, unsigned offset)
584 struct dm_io_request io_req = {
587 .notify.fn = dmio_complete,
589 .client = b->c->dm_io,
591 struct dm_io_region region = {
597 if (b->data_mode != DATA_MODE_VMALLOC) {
598 io_req.mem.type = DM_IO_KMEM;
599 io_req.mem.ptr.addr = (char *)b->data + offset;
601 io_req.mem.type = DM_IO_VMA;
602 io_req.mem.ptr.vma = (char *)b->data + offset;
605 r = dm_io(&io_req, 1, ®ion, NULL);
607 b->end_io(b, errno_to_blk_status(r));
610 static void bio_complete(struct bio *bio)
612 struct dm_buffer *b = bio->bi_private;
613 blk_status_t status = bio->bi_status;
615 b->end_io(b, status);
618 static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
619 unsigned n_sectors, unsigned offset)
623 unsigned vec_size, len;
625 vec_size = b->c->block_size >> PAGE_SHIFT;
626 if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
629 bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
632 use_dmio(b, rw, sector, n_sectors, offset);
636 bio->bi_iter.bi_sector = sector;
637 bio_set_dev(bio, b->c->bdev);
638 bio_set_op_attrs(bio, rw, 0);
639 bio->bi_end_io = bio_complete;
642 ptr = (char *)b->data + offset;
643 len = n_sectors << SECTOR_SHIFT;
646 unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
647 if (!bio_add_page(bio, virt_to_page(ptr), this_step,
648 offset_in_page(ptr))) {
660 static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
664 if (likely(c->sectors_per_block_bits >= 0))
665 sector = block << c->sectors_per_block_bits;
667 sector = block * (c->block_size >> SECTOR_SHIFT);
673 static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
677 unsigned offset, end;
681 sector = block_to_sector(b->c, b->block);
683 if (rw != REQ_OP_WRITE) {
684 n_sectors = b->c->block_size >> SECTOR_SHIFT;
687 if (b->c->write_callback)
688 b->c->write_callback(b);
689 offset = b->write_start;
691 offset &= -DM_BUFIO_WRITE_ALIGN;
692 end += DM_BUFIO_WRITE_ALIGN - 1;
693 end &= -DM_BUFIO_WRITE_ALIGN;
694 if (unlikely(end > b->c->block_size))
695 end = b->c->block_size;
697 sector += offset >> SECTOR_SHIFT;
698 n_sectors = (end - offset) >> SECTOR_SHIFT;
701 if (b->data_mode != DATA_MODE_VMALLOC)
702 use_bio(b, rw, sector, n_sectors, offset);
704 use_dmio(b, rw, sector, n_sectors, offset);
707 /*----------------------------------------------------------------
708 * Writing dirty buffers
709 *--------------------------------------------------------------*/
712 * The endio routine for write.
714 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
717 static void write_endio(struct dm_buffer *b, blk_status_t status)
719 b->write_error = status;
720 if (unlikely(status)) {
721 struct dm_bufio_client *c = b->c;
723 (void)cmpxchg(&c->async_write_error, 0,
724 blk_status_to_errno(status));
727 BUG_ON(!test_bit(B_WRITING, &b->state));
729 smp_mb__before_atomic();
730 clear_bit(B_WRITING, &b->state);
731 smp_mb__after_atomic();
733 wake_up_bit(&b->state, B_WRITING);
737 * Initiate a write on a dirty buffer, but don't wait for it.
739 * - If the buffer is not dirty, exit.
740 * - If there some previous write going on, wait for it to finish (we can't
741 * have two writes on the same buffer simultaneously).
742 * - Submit our write and don't wait on it. We set B_WRITING indicating
743 * that there is a write in progress.
745 static void __write_dirty_buffer(struct dm_buffer *b,
746 struct list_head *write_list)
748 if (!test_bit(B_DIRTY, &b->state))
751 clear_bit(B_DIRTY, &b->state);
752 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
754 b->write_start = b->dirty_start;
755 b->write_end = b->dirty_end;
758 submit_io(b, REQ_OP_WRITE, write_endio);
760 list_add_tail(&b->write_list, write_list);
763 static void __flush_write_list(struct list_head *write_list)
765 struct blk_plug plug;
766 blk_start_plug(&plug);
767 while (!list_empty(write_list)) {
768 struct dm_buffer *b =
769 list_entry(write_list->next, struct dm_buffer, write_list);
770 list_del(&b->write_list);
771 submit_io(b, REQ_OP_WRITE, write_endio);
774 blk_finish_plug(&plug);
778 * Wait until any activity on the buffer finishes. Possibly write the
779 * buffer if it is dirty. When this function finishes, there is no I/O
780 * running on the buffer and the buffer is not dirty.
782 static void __make_buffer_clean(struct dm_buffer *b)
784 BUG_ON(b->hold_count);
786 if (!b->state) /* fast case */
789 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
790 __write_dirty_buffer(b, NULL);
791 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
795 * Find some buffer that is not held by anybody, clean it, unlink it and
798 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
802 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
803 BUG_ON(test_bit(B_WRITING, &b->state));
804 BUG_ON(test_bit(B_DIRTY, &b->state));
806 if (!b->hold_count) {
807 __make_buffer_clean(b);
814 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
815 BUG_ON(test_bit(B_READING, &b->state));
817 if (!b->hold_count) {
818 __make_buffer_clean(b);
829 * Wait until some other threads free some buffer or release hold count on
832 * This function is entered with c->lock held, drops it and regains it
835 static void __wait_for_free_buffer(struct dm_bufio_client *c)
837 DECLARE_WAITQUEUE(wait, current);
839 add_wait_queue(&c->free_buffer_wait, &wait);
840 set_current_state(TASK_UNINTERRUPTIBLE);
845 remove_wait_queue(&c->free_buffer_wait, &wait);
858 * Allocate a new buffer. If the allocation is not possible, wait until
859 * some other thread frees a buffer.
861 * May drop the lock and regain it.
863 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
866 bool tried_noio_alloc = false;
869 * dm-bufio is resistant to allocation failures (it just keeps
870 * one buffer reserved in cases all the allocations fail).
871 * So set flags to not try too hard:
872 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
873 * mutex and wait ourselves.
874 * __GFP_NORETRY: don't retry and rather return failure
875 * __GFP_NOMEMALLOC: don't use emergency reserves
876 * __GFP_NOWARN: don't print a warning in case of failure
878 * For debugging, if we set the cache size to 1, no new buffers will
882 if (dm_bufio_cache_size_latch != 1) {
883 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
888 if (nf == NF_PREFETCH)
891 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
893 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
897 tried_noio_alloc = true;
900 if (!list_empty(&c->reserved_buffers)) {
901 b = list_entry(c->reserved_buffers.next,
902 struct dm_buffer, lru_list);
903 list_del(&b->lru_list);
904 c->need_reserved_buffers++;
909 b = __get_unclaimed_buffer(c);
913 __wait_for_free_buffer(c);
917 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
919 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
924 if (c->alloc_callback)
925 c->alloc_callback(b);
931 * Free a buffer and wake other threads waiting for free buffers.
933 static void __free_buffer_wake(struct dm_buffer *b)
935 struct dm_bufio_client *c = b->c;
937 if (!c->need_reserved_buffers)
940 list_add(&b->lru_list, &c->reserved_buffers);
941 c->need_reserved_buffers--;
944 wake_up(&c->free_buffer_wait);
947 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
948 struct list_head *write_list)
950 struct dm_buffer *b, *tmp;
952 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
953 BUG_ON(test_bit(B_READING, &b->state));
955 if (!test_bit(B_DIRTY, &b->state) &&
956 !test_bit(B_WRITING, &b->state)) {
957 __relink_lru(b, LIST_CLEAN);
961 if (no_wait && test_bit(B_WRITING, &b->state))
964 __write_dirty_buffer(b, write_list);
970 * Check if we're over watermark.
971 * If we are over threshold_buffers, start freeing buffers.
972 * If we're over "limit_buffers", block until we get under the limit.
974 static void __check_watermark(struct dm_bufio_client *c,
975 struct list_head *write_list)
977 if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
978 __write_dirty_buffers_async(c, 1, write_list);
981 /*----------------------------------------------------------------
983 *--------------------------------------------------------------*/
985 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
986 enum new_flag nf, int *need_submit,
987 struct list_head *write_list)
989 struct dm_buffer *b, *new_b = NULL;
993 b = __find(c, block);
1000 new_b = __alloc_buffer_wait(c, nf);
1005 * We've had a period where the mutex was unlocked, so need to
1006 * recheck the buffer tree.
1008 b = __find(c, block);
1010 __free_buffer_wake(new_b);
1014 __check_watermark(c, write_list);
1020 __link_buffer(b, block, LIST_CLEAN);
1022 if (nf == NF_FRESH) {
1027 b->state = 1 << B_READING;
1033 if (nf == NF_PREFETCH)
1036 * Note: it is essential that we don't wait for the buffer to be
1037 * read if dm_bufio_get function is used. Both dm_bufio_get and
1038 * dm_bufio_prefetch can be used in the driver request routine.
1039 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1040 * the same buffer, it would deadlock if we waited.
1042 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1046 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1047 test_bit(B_WRITING, &b->state));
1052 * The endio routine for reading: set the error, clear the bit and wake up
1053 * anyone waiting on the buffer.
1055 static void read_endio(struct dm_buffer *b, blk_status_t status)
1057 b->read_error = status;
1059 BUG_ON(!test_bit(B_READING, &b->state));
1061 smp_mb__before_atomic();
1062 clear_bit(B_READING, &b->state);
1063 smp_mb__after_atomic();
1065 wake_up_bit(&b->state, B_READING);
1069 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1070 * functions is similar except that dm_bufio_new doesn't read the
1071 * buffer from the disk (assuming that the caller overwrites all the data
1072 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1074 static void *new_read(struct dm_bufio_client *c, sector_t block,
1075 enum new_flag nf, struct dm_buffer **bp)
1078 struct dm_buffer *b;
1080 LIST_HEAD(write_list);
1083 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1084 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1085 if (b && b->hold_count == 1)
1086 buffer_record_stack(b);
1090 __flush_write_list(&write_list);
1096 submit_io(b, REQ_OP_READ, read_endio);
1098 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1100 if (b->read_error) {
1101 int error = blk_status_to_errno(b->read_error);
1103 dm_bufio_release(b);
1105 return ERR_PTR(error);
1113 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1114 struct dm_buffer **bp)
1116 return new_read(c, block, NF_GET, bp);
1118 EXPORT_SYMBOL_GPL(dm_bufio_get);
1120 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1121 struct dm_buffer **bp)
1123 BUG_ON(dm_bufio_in_request());
1125 return new_read(c, block, NF_READ, bp);
1127 EXPORT_SYMBOL_GPL(dm_bufio_read);
1129 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1130 struct dm_buffer **bp)
1132 BUG_ON(dm_bufio_in_request());
1134 return new_read(c, block, NF_FRESH, bp);
1136 EXPORT_SYMBOL_GPL(dm_bufio_new);
1138 void dm_bufio_prefetch(struct dm_bufio_client *c,
1139 sector_t block, unsigned n_blocks)
1141 struct blk_plug plug;
1143 LIST_HEAD(write_list);
1145 BUG_ON(dm_bufio_in_request());
1147 blk_start_plug(&plug);
1150 for (; n_blocks--; block++) {
1152 struct dm_buffer *b;
1153 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1155 if (unlikely(!list_empty(&write_list))) {
1157 blk_finish_plug(&plug);
1158 __flush_write_list(&write_list);
1159 blk_start_plug(&plug);
1162 if (unlikely(b != NULL)) {
1166 submit_io(b, REQ_OP_READ, read_endio);
1167 dm_bufio_release(b);
1180 blk_finish_plug(&plug);
1182 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1184 void dm_bufio_release(struct dm_buffer *b)
1186 struct dm_bufio_client *c = b->c;
1190 BUG_ON(!b->hold_count);
1193 if (!b->hold_count) {
1194 wake_up(&c->free_buffer_wait);
1197 * If there were errors on the buffer, and the buffer is not
1198 * to be written, free the buffer. There is no point in caching
1201 if ((b->read_error || b->write_error) &&
1202 !test_bit(B_READING, &b->state) &&
1203 !test_bit(B_WRITING, &b->state) &&
1204 !test_bit(B_DIRTY, &b->state)) {
1206 __free_buffer_wake(b);
1212 EXPORT_SYMBOL_GPL(dm_bufio_release);
1214 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1215 unsigned start, unsigned end)
1217 struct dm_bufio_client *c = b->c;
1219 BUG_ON(start >= end);
1220 BUG_ON(end > b->c->block_size);
1224 BUG_ON(test_bit(B_READING, &b->state));
1226 if (!test_and_set_bit(B_DIRTY, &b->state)) {
1227 b->dirty_start = start;
1229 __relink_lru(b, LIST_DIRTY);
1231 if (start < b->dirty_start)
1232 b->dirty_start = start;
1233 if (end > b->dirty_end)
1239 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1241 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1243 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1245 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1247 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1249 LIST_HEAD(write_list);
1251 BUG_ON(dm_bufio_in_request());
1254 __write_dirty_buffers_async(c, 0, &write_list);
1256 __flush_write_list(&write_list);
1258 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1261 * For performance, it is essential that the buffers are written asynchronously
1262 * and simultaneously (so that the block layer can merge the writes) and then
1265 * Finally, we flush hardware disk cache.
1267 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1270 unsigned long buffers_processed = 0;
1271 struct dm_buffer *b, *tmp;
1273 LIST_HEAD(write_list);
1276 __write_dirty_buffers_async(c, 0, &write_list);
1278 __flush_write_list(&write_list);
1282 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1283 int dropped_lock = 0;
1285 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1286 buffers_processed++;
1288 BUG_ON(test_bit(B_READING, &b->state));
1290 if (test_bit(B_WRITING, &b->state)) {
1291 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1295 wait_on_bit_io(&b->state, B_WRITING,
1296 TASK_UNINTERRUPTIBLE);
1300 wait_on_bit_io(&b->state, B_WRITING,
1301 TASK_UNINTERRUPTIBLE);
1304 if (!test_bit(B_DIRTY, &b->state) &&
1305 !test_bit(B_WRITING, &b->state))
1306 __relink_lru(b, LIST_CLEAN);
1311 * If we dropped the lock, the list is no longer consistent,
1312 * so we must restart the search.
1314 * In the most common case, the buffer just processed is
1315 * relinked to the clean list, so we won't loop scanning the
1316 * same buffer again and again.
1318 * This may livelock if there is another thread simultaneously
1319 * dirtying buffers, so we count the number of buffers walked
1320 * and if it exceeds the total number of buffers, it means that
1321 * someone is doing some writes simultaneously with us. In
1322 * this case, stop, dropping the lock.
1327 wake_up(&c->free_buffer_wait);
1330 a = xchg(&c->async_write_error, 0);
1331 f = dm_bufio_issue_flush(c);
1337 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1340 * Use dm-io to send an empty barrier to flush the device.
1342 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1344 struct dm_io_request io_req = {
1345 .bi_op = REQ_OP_WRITE,
1346 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1347 .mem.type = DM_IO_KMEM,
1348 .mem.ptr.addr = NULL,
1351 struct dm_io_region io_reg = {
1357 BUG_ON(dm_bufio_in_request());
1359 return dm_io(&io_req, 1, &io_reg, NULL);
1361 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1364 * Use dm-io to send a discard request to flush the device.
1366 int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
1368 struct dm_io_request io_req = {
1369 .bi_op = REQ_OP_DISCARD,
1370 .bi_op_flags = REQ_SYNC,
1371 .mem.type = DM_IO_KMEM,
1372 .mem.ptr.addr = NULL,
1375 struct dm_io_region io_reg = {
1377 .sector = block_to_sector(c, block),
1378 .count = block_to_sector(c, count),
1381 BUG_ON(dm_bufio_in_request());
1383 return dm_io(&io_req, 1, &io_reg, NULL);
1385 EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
1388 * We first delete any other buffer that may be at that new location.
1390 * Then, we write the buffer to the original location if it was dirty.
1392 * Then, if we are the only one who is holding the buffer, relink the buffer
1393 * in the buffer tree for the new location.
1395 * If there was someone else holding the buffer, we write it to the new
1396 * location but not relink it, because that other user needs to have the buffer
1397 * at the same place.
1399 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1401 struct dm_bufio_client *c = b->c;
1402 struct dm_buffer *new;
1404 BUG_ON(dm_bufio_in_request());
1409 new = __find(c, new_block);
1411 if (new->hold_count) {
1412 __wait_for_free_buffer(c);
1417 * FIXME: Is there any point waiting for a write that's going
1418 * to be overwritten in a bit?
1420 __make_buffer_clean(new);
1421 __unlink_buffer(new);
1422 __free_buffer_wake(new);
1425 BUG_ON(!b->hold_count);
1426 BUG_ON(test_bit(B_READING, &b->state));
1428 __write_dirty_buffer(b, NULL);
1429 if (b->hold_count == 1) {
1430 wait_on_bit_io(&b->state, B_WRITING,
1431 TASK_UNINTERRUPTIBLE);
1432 set_bit(B_DIRTY, &b->state);
1434 b->dirty_end = c->block_size;
1436 __link_buffer(b, new_block, LIST_DIRTY);
1439 wait_on_bit_lock_io(&b->state, B_WRITING,
1440 TASK_UNINTERRUPTIBLE);
1442 * Relink buffer to "new_block" so that write_callback
1443 * sees "new_block" as a block number.
1444 * After the write, link the buffer back to old_block.
1445 * All this must be done in bufio lock, so that block number
1446 * change isn't visible to other threads.
1448 old_block = b->block;
1450 __link_buffer(b, new_block, b->list_mode);
1451 submit_io(b, REQ_OP_WRITE, write_endio);
1452 wait_on_bit_io(&b->state, B_WRITING,
1453 TASK_UNINTERRUPTIBLE);
1455 __link_buffer(b, old_block, b->list_mode);
1459 dm_bufio_release(b);
1461 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1463 static void forget_buffer_locked(struct dm_buffer *b)
1465 if (likely(!b->hold_count) && likely(!b->state)) {
1467 __free_buffer_wake(b);
1472 * Free the given buffer.
1474 * This is just a hint, if the buffer is in use or dirty, this function
1477 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1479 struct dm_buffer *b;
1483 b = __find(c, block);
1485 forget_buffer_locked(b);
1489 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1491 void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
1493 struct dm_buffer *b;
1494 sector_t end_block = block + n_blocks;
1496 while (block < end_block) {
1499 b = __find_next(c, block);
1501 block = b->block + 1;
1502 forget_buffer_locked(b);
1512 EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
1514 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1516 c->minimum_buffers = n;
1518 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1520 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1522 return c->block_size;
1524 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1526 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1528 sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1533 if (likely(c->sectors_per_block_bits >= 0))
1534 s >>= c->sectors_per_block_bits;
1536 sector_div(s, c->block_size >> SECTOR_SHIFT);
1539 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1541 struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1545 EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1547 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1551 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1553 void *dm_bufio_get_block_data(struct dm_buffer *b)
1557 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1559 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1563 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1565 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1569 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1571 static void drop_buffers(struct dm_bufio_client *c)
1573 struct dm_buffer *b;
1575 bool warned = false;
1577 BUG_ON(dm_bufio_in_request());
1580 * An optimization so that the buffers are not written one-by-one.
1582 dm_bufio_write_dirty_buffers_async(c);
1586 while ((b = __get_unclaimed_buffer(c)))
1587 __free_buffer_wake(b);
1589 for (i = 0; i < LIST_SIZE; i++)
1590 list_for_each_entry(b, &c->lru[i], lru_list) {
1593 DMERR("leaked buffer %llx, hold count %u, list %d",
1594 (unsigned long long)b->block, b->hold_count, i);
1595 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1596 stack_trace_print(b->stack_entries, b->stack_len, 1);
1597 /* mark unclaimed to avoid BUG_ON below */
1602 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1603 while ((b = __get_unclaimed_buffer(c)))
1604 __free_buffer_wake(b);
1607 for (i = 0; i < LIST_SIZE; i++)
1608 BUG_ON(!list_empty(&c->lru[i]));
1614 * We may not be able to evict this buffer if IO pending or the client
1615 * is still using it. Caller is expected to know buffer is too old.
1617 * And if GFP_NOFS is used, we must not do any I/O because we hold
1618 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1619 * rerouted to different bufio client.
1621 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1623 if (!(gfp & __GFP_FS)) {
1624 if (test_bit(B_READING, &b->state) ||
1625 test_bit(B_WRITING, &b->state) ||
1626 test_bit(B_DIRTY, &b->state))
1633 __make_buffer_clean(b);
1635 __free_buffer_wake(b);
1640 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1642 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1643 if (likely(c->sectors_per_block_bits >= 0))
1644 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1646 retain_bytes /= c->block_size;
1647 return retain_bytes;
1650 static void __scan(struct dm_bufio_client *c)
1653 struct dm_buffer *b, *tmp;
1654 unsigned long freed = 0;
1655 unsigned long count = c->n_buffers[LIST_CLEAN] +
1656 c->n_buffers[LIST_DIRTY];
1657 unsigned long retain_target = get_retain_buffers(c);
1659 for (l = 0; l < LIST_SIZE; l++) {
1660 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1661 if (count - freed <= retain_target)
1662 atomic_long_set(&c->need_shrink, 0);
1663 if (!atomic_long_read(&c->need_shrink))
1665 if (__try_evict_buffer(b, GFP_KERNEL)) {
1666 atomic_long_dec(&c->need_shrink);
1674 static void shrink_work(struct work_struct *w)
1676 struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
1683 static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1685 struct dm_bufio_client *c;
1687 c = container_of(shrink, struct dm_bufio_client, shrinker);
1688 atomic_long_add(sc->nr_to_scan, &c->need_shrink);
1689 queue_work(dm_bufio_wq, &c->shrink_work);
1691 return sc->nr_to_scan;
1694 static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1696 struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1697 unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1698 READ_ONCE(c->n_buffers[LIST_DIRTY]);
1699 unsigned long retain_target = get_retain_buffers(c);
1700 unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
1702 if (unlikely(count < retain_target))
1705 count -= retain_target;
1707 if (unlikely(count < queued_for_cleanup))
1710 count -= queued_for_cleanup;
1716 * Create the buffering interface
1718 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1719 unsigned reserved_buffers, unsigned aux_size,
1720 void (*alloc_callback)(struct dm_buffer *),
1721 void (*write_callback)(struct dm_buffer *))
1724 struct dm_bufio_client *c;
1728 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1729 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1734 c = kzalloc(sizeof(*c), GFP_KERNEL);
1739 c->buffer_tree = RB_ROOT;
1742 c->block_size = block_size;
1743 if (is_power_of_2(block_size))
1744 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1746 c->sectors_per_block_bits = -1;
1748 c->alloc_callback = alloc_callback;
1749 c->write_callback = write_callback;
1751 for (i = 0; i < LIST_SIZE; i++) {
1752 INIT_LIST_HEAD(&c->lru[i]);
1753 c->n_buffers[i] = 0;
1756 mutex_init(&c->lock);
1757 INIT_LIST_HEAD(&c->reserved_buffers);
1758 c->need_reserved_buffers = reserved_buffers;
1760 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1762 init_waitqueue_head(&c->free_buffer_wait);
1763 c->async_write_error = 0;
1765 c->dm_io = dm_io_client_create();
1766 if (IS_ERR(c->dm_io)) {
1767 r = PTR_ERR(c->dm_io);
1771 if (block_size <= KMALLOC_MAX_SIZE &&
1772 (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1773 unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1774 snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1775 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1776 SLAB_RECLAIM_ACCOUNT, NULL);
1777 if (!c->slab_cache) {
1783 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1785 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1786 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1787 0, SLAB_RECLAIM_ACCOUNT, NULL);
1788 if (!c->slab_buffer) {
1793 while (c->need_reserved_buffers) {
1794 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1800 __free_buffer_wake(b);
1803 INIT_WORK(&c->shrink_work, shrink_work);
1804 atomic_long_set(&c->need_shrink, 0);
1806 c->shrinker.count_objects = dm_bufio_shrink_count;
1807 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1808 c->shrinker.seeks = 1;
1809 c->shrinker.batch = 0;
1810 r = register_shrinker(&c->shrinker);
1814 mutex_lock(&dm_bufio_clients_lock);
1815 dm_bufio_client_count++;
1816 list_add(&c->client_list, &dm_bufio_all_clients);
1817 __cache_size_refresh();
1818 mutex_unlock(&dm_bufio_clients_lock);
1823 while (!list_empty(&c->reserved_buffers)) {
1824 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1825 struct dm_buffer, lru_list);
1826 list_del(&b->lru_list);
1829 kmem_cache_destroy(c->slab_cache);
1830 kmem_cache_destroy(c->slab_buffer);
1831 dm_io_client_destroy(c->dm_io);
1833 mutex_destroy(&c->lock);
1838 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1841 * Free the buffering interface.
1842 * It is required that there are no references on any buffers.
1844 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1850 unregister_shrinker(&c->shrinker);
1851 flush_work(&c->shrink_work);
1853 mutex_lock(&dm_bufio_clients_lock);
1855 list_del(&c->client_list);
1856 dm_bufio_client_count--;
1857 __cache_size_refresh();
1859 mutex_unlock(&dm_bufio_clients_lock);
1861 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1862 BUG_ON(c->need_reserved_buffers);
1864 while (!list_empty(&c->reserved_buffers)) {
1865 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1866 struct dm_buffer, lru_list);
1867 list_del(&b->lru_list);
1871 for (i = 0; i < LIST_SIZE; i++)
1872 if (c->n_buffers[i])
1873 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1875 for (i = 0; i < LIST_SIZE; i++)
1876 BUG_ON(c->n_buffers[i]);
1878 kmem_cache_destroy(c->slab_cache);
1879 kmem_cache_destroy(c->slab_buffer);
1880 dm_io_client_destroy(c->dm_io);
1881 mutex_destroy(&c->lock);
1884 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1886 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1890 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1892 static unsigned get_max_age_hz(void)
1894 unsigned max_age = READ_ONCE(dm_bufio_max_age);
1896 if (max_age > UINT_MAX / HZ)
1897 max_age = UINT_MAX / HZ;
1899 return max_age * HZ;
1902 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1904 return time_after_eq(jiffies, b->last_accessed + age_hz);
1907 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1909 struct dm_buffer *b, *tmp;
1910 unsigned long retain_target = get_retain_buffers(c);
1911 unsigned long count;
1912 LIST_HEAD(write_list);
1916 __check_watermark(c, &write_list);
1917 if (unlikely(!list_empty(&write_list))) {
1919 __flush_write_list(&write_list);
1923 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1924 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1925 if (count <= retain_target)
1928 if (!older_than(b, age_hz))
1931 if (__try_evict_buffer(b, 0))
1940 static void do_global_cleanup(struct work_struct *w)
1942 struct dm_bufio_client *locked_client = NULL;
1943 struct dm_bufio_client *current_client;
1944 struct dm_buffer *b;
1945 unsigned spinlock_hold_count;
1946 unsigned long threshold = dm_bufio_cache_size -
1947 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1948 unsigned long loops = global_num * 2;
1950 mutex_lock(&dm_bufio_clients_lock);
1955 spin_lock(&global_spinlock);
1956 if (unlikely(dm_bufio_current_allocated <= threshold))
1959 spinlock_hold_count = 0;
1963 if (unlikely(list_empty(&global_queue)))
1965 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1969 list_move(&b->global_list, &global_queue);
1970 if (likely(++spinlock_hold_count < 16))
1972 spin_unlock(&global_spinlock);
1976 current_client = b->c;
1977 if (unlikely(current_client != locked_client)) {
1979 dm_bufio_unlock(locked_client);
1981 if (!dm_bufio_trylock(current_client)) {
1982 spin_unlock(&global_spinlock);
1983 dm_bufio_lock(current_client);
1984 locked_client = current_client;
1988 locked_client = current_client;
1991 spin_unlock(&global_spinlock);
1993 if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
1994 spin_lock(&global_spinlock);
1995 list_move(&b->global_list, &global_queue);
1996 spin_unlock(&global_spinlock);
2000 spin_unlock(&global_spinlock);
2003 dm_bufio_unlock(locked_client);
2005 mutex_unlock(&dm_bufio_clients_lock);
2008 static void cleanup_old_buffers(void)
2010 unsigned long max_age_hz = get_max_age_hz();
2011 struct dm_bufio_client *c;
2013 mutex_lock(&dm_bufio_clients_lock);
2015 __cache_size_refresh();
2017 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2018 __evict_old_buffers(c, max_age_hz);
2020 mutex_unlock(&dm_bufio_clients_lock);
2023 static void work_fn(struct work_struct *w)
2025 cleanup_old_buffers();
2027 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2028 DM_BUFIO_WORK_TIMER_SECS * HZ);
2031 /*----------------------------------------------------------------
2033 *--------------------------------------------------------------*/
2036 * This is called only once for the whole dm_bufio module.
2037 * It initializes memory limit.
2039 static int __init dm_bufio_init(void)
2043 dm_bufio_allocated_kmem_cache = 0;
2044 dm_bufio_allocated_get_free_pages = 0;
2045 dm_bufio_allocated_vmalloc = 0;
2046 dm_bufio_current_allocated = 0;
2048 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2049 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2051 if (mem > ULONG_MAX)
2055 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2056 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2059 dm_bufio_default_cache_size = mem;
2061 mutex_lock(&dm_bufio_clients_lock);
2062 __cache_size_refresh();
2063 mutex_unlock(&dm_bufio_clients_lock);
2065 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2069 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2070 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2071 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2072 DM_BUFIO_WORK_TIMER_SECS * HZ);
2078 * This is called once when unloading the dm_bufio module.
2080 static void __exit dm_bufio_exit(void)
2084 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2085 flush_workqueue(dm_bufio_wq);
2086 destroy_workqueue(dm_bufio_wq);
2088 if (dm_bufio_client_count) {
2089 DMCRIT("%s: dm_bufio_client_count leaked: %d",
2090 __func__, dm_bufio_client_count);
2094 if (dm_bufio_current_allocated) {
2095 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2096 __func__, dm_bufio_current_allocated);
2100 if (dm_bufio_allocated_get_free_pages) {
2101 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2102 __func__, dm_bufio_allocated_get_free_pages);
2106 if (dm_bufio_allocated_vmalloc) {
2107 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2108 __func__, dm_bufio_allocated_vmalloc);
2115 module_init(dm_bufio_init)
2116 module_exit(dm_bufio_exit)
2118 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2119 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2121 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2122 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2124 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2125 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2127 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2128 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2130 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2131 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2133 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2134 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2136 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2137 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2139 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2140 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2142 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2143 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2144 MODULE_LICENSE("GPL");