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;
111 struct shrinker shrinker;
122 * Describes how the block was allocated:
123 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
124 * See the comment at alloc_buffer_data.
128 DATA_MODE_GET_FREE_PAGES = 1,
129 DATA_MODE_VMALLOC = 2,
135 struct list_head lru_list;
136 struct list_head global_list;
139 unsigned char data_mode; /* DATA_MODE_* */
140 unsigned char list_mode; /* LIST_* */
141 blk_status_t read_error;
142 blk_status_t write_error;
146 unsigned long last_accessed;
147 unsigned dirty_start;
149 unsigned write_start;
151 struct dm_bufio_client *c;
152 struct list_head write_list;
153 void (*end_io)(struct dm_buffer *, blk_status_t);
154 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
156 unsigned int stack_len;
157 unsigned long stack_entries[MAX_STACK];
161 /*----------------------------------------------------------------*/
163 #define dm_bufio_in_request() (!!current->bio_list)
165 static void dm_bufio_lock(struct dm_bufio_client *c)
167 mutex_lock_nested(&c->lock, dm_bufio_in_request());
170 static int dm_bufio_trylock(struct dm_bufio_client *c)
172 return mutex_trylock(&c->lock);
175 static void dm_bufio_unlock(struct dm_bufio_client *c)
177 mutex_unlock(&c->lock);
180 /*----------------------------------------------------------------*/
183 * Default cache size: available memory divided by the ratio.
185 static unsigned long dm_bufio_default_cache_size;
188 * Total cache size set by the user.
190 static unsigned long dm_bufio_cache_size;
193 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
194 * at any time. If it disagrees, the user has changed cache size.
196 static unsigned long dm_bufio_cache_size_latch;
198 static DEFINE_SPINLOCK(global_spinlock);
200 static LIST_HEAD(global_queue);
202 static unsigned long global_num = 0;
205 * Buffers are freed after this timeout
207 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
208 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
210 static unsigned long dm_bufio_peak_allocated;
211 static unsigned long dm_bufio_allocated_kmem_cache;
212 static unsigned long dm_bufio_allocated_get_free_pages;
213 static unsigned long dm_bufio_allocated_vmalloc;
214 static unsigned long dm_bufio_current_allocated;
216 /*----------------------------------------------------------------*/
219 * The current number of clients.
221 static int dm_bufio_client_count;
224 * The list of all clients.
226 static LIST_HEAD(dm_bufio_all_clients);
229 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
231 static DEFINE_MUTEX(dm_bufio_clients_lock);
233 static struct workqueue_struct *dm_bufio_wq;
234 static struct delayed_work dm_bufio_cleanup_old_work;
235 static struct work_struct dm_bufio_replacement_work;
238 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
239 static void buffer_record_stack(struct dm_buffer *b)
241 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
245 /*----------------------------------------------------------------
246 * A red/black tree acts as an index for all the buffers.
247 *--------------------------------------------------------------*/
248 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
250 struct rb_node *n = c->buffer_tree.rb_node;
254 b = container_of(n, struct dm_buffer, node);
256 if (b->block == block)
259 n = (b->block < block) ? n->rb_left : n->rb_right;
265 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
267 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
268 struct dm_buffer *found;
271 found = container_of(*new, struct dm_buffer, node);
273 if (found->block == b->block) {
279 new = (found->block < b->block) ?
280 &((*new)->rb_left) : &((*new)->rb_right);
283 rb_link_node(&b->node, parent, new);
284 rb_insert_color(&b->node, &c->buffer_tree);
287 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
289 rb_erase(&b->node, &c->buffer_tree);
292 /*----------------------------------------------------------------*/
294 static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
296 unsigned char data_mode;
299 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
300 &dm_bufio_allocated_kmem_cache,
301 &dm_bufio_allocated_get_free_pages,
302 &dm_bufio_allocated_vmalloc,
305 data_mode = b->data_mode;
306 diff = (long)b->c->block_size;
310 spin_lock(&global_spinlock);
312 *class_ptr[data_mode] += diff;
314 dm_bufio_current_allocated += diff;
316 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
317 dm_bufio_peak_allocated = dm_bufio_current_allocated;
322 list_add(&b->global_list, &global_queue);
324 if (dm_bufio_current_allocated > dm_bufio_cache_size)
325 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
327 list_del(&b->global_list);
331 spin_unlock(&global_spinlock);
335 * Change the number of clients and recalculate per-client limit.
337 static void __cache_size_refresh(void)
339 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
340 BUG_ON(dm_bufio_client_count < 0);
342 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
345 * Use default if set to 0 and report the actual cache size used.
347 if (!dm_bufio_cache_size_latch) {
348 (void)cmpxchg(&dm_bufio_cache_size, 0,
349 dm_bufio_default_cache_size);
350 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
355 * Allocating buffer data.
357 * Small buffers are allocated with kmem_cache, to use space optimally.
359 * For large buffers, we choose between get_free_pages and vmalloc.
360 * Each has advantages and disadvantages.
362 * __get_free_pages can randomly fail if the memory is fragmented.
363 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
364 * as low as 128M) so using it for caching is not appropriate.
366 * If the allocation may fail we use __get_free_pages. Memory fragmentation
367 * won't have a fatal effect here, but it just causes flushes of some other
368 * buffers and more I/O will be performed. Don't use __get_free_pages if it
369 * always fails (i.e. order >= MAX_ORDER).
371 * If the allocation shouldn't fail we use __vmalloc. This is only for the
372 * initial reserve allocation, so there's no risk of wasting all vmalloc
375 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
376 unsigned char *data_mode)
378 if (unlikely(c->slab_cache != NULL)) {
379 *data_mode = DATA_MODE_SLAB;
380 return kmem_cache_alloc(c->slab_cache, gfp_mask);
383 if (c->block_size <= KMALLOC_MAX_SIZE &&
384 gfp_mask & __GFP_NORETRY) {
385 *data_mode = DATA_MODE_GET_FREE_PAGES;
386 return (void *)__get_free_pages(gfp_mask,
387 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
390 *data_mode = DATA_MODE_VMALLOC;
393 * __vmalloc allocates the data pages and auxiliary structures with
394 * gfp_flags that were specified, but pagetables are always allocated
395 * with GFP_KERNEL, no matter what was specified as gfp_mask.
397 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
398 * all allocations done by this process (including pagetables) are done
399 * as if GFP_NOIO was specified.
401 if (gfp_mask & __GFP_NORETRY) {
402 unsigned noio_flag = memalloc_noio_save();
403 void *ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
405 memalloc_noio_restore(noio_flag);
409 return __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
413 * Free buffer's data.
415 static void free_buffer_data(struct dm_bufio_client *c,
416 void *data, unsigned char data_mode)
420 kmem_cache_free(c->slab_cache, data);
423 case DATA_MODE_GET_FREE_PAGES:
424 free_pages((unsigned long)data,
425 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
428 case DATA_MODE_VMALLOC:
433 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
440 * Allocate buffer and its data.
442 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
444 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
451 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
453 kmem_cache_free(c->slab_buffer, b);
457 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
464 * Free buffer and its data.
466 static void free_buffer(struct dm_buffer *b)
468 struct dm_bufio_client *c = b->c;
470 free_buffer_data(c, b->data, b->data_mode);
471 kmem_cache_free(c->slab_buffer, b);
475 * Link buffer to the buffer tree and clean or dirty queue.
477 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
479 struct dm_bufio_client *c = b->c;
481 c->n_buffers[dirty]++;
483 b->list_mode = dirty;
484 list_add(&b->lru_list, &c->lru[dirty]);
486 b->last_accessed = jiffies;
488 adjust_total_allocated(b, false);
492 * Unlink buffer from the buffer tree and dirty or clean queue.
494 static void __unlink_buffer(struct dm_buffer *b)
496 struct dm_bufio_client *c = b->c;
498 BUG_ON(!c->n_buffers[b->list_mode]);
500 c->n_buffers[b->list_mode]--;
502 list_del(&b->lru_list);
504 adjust_total_allocated(b, true);
508 * Place the buffer to the head of dirty or clean LRU queue.
510 static void __relink_lru(struct dm_buffer *b, int dirty)
512 struct dm_bufio_client *c = b->c;
516 BUG_ON(!c->n_buffers[b->list_mode]);
518 c->n_buffers[b->list_mode]--;
519 c->n_buffers[dirty]++;
520 b->list_mode = dirty;
521 list_move(&b->lru_list, &c->lru[dirty]);
522 b->last_accessed = jiffies;
525 /*----------------------------------------------------------------
526 * Submit I/O on the buffer.
528 * Bio interface is faster but it has some problems:
529 * the vector list is limited (increasing this limit increases
530 * memory-consumption per buffer, so it is not viable);
532 * the memory must be direct-mapped, not vmalloced;
534 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
535 * it is not vmalloced, try using the bio interface.
537 * If the buffer is big, if it is vmalloced or if the underlying device
538 * rejects the bio because it is too large, use dm-io layer to do the I/O.
539 * The dm-io layer splits the I/O into multiple requests, avoiding the above
541 *--------------------------------------------------------------*/
544 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
545 * that the request was handled directly with bio interface.
547 static void dmio_complete(unsigned long error, void *context)
549 struct dm_buffer *b = context;
551 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
554 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
555 unsigned n_sectors, unsigned offset)
558 struct dm_io_request io_req = {
561 .notify.fn = dmio_complete,
563 .client = b->c->dm_io,
565 struct dm_io_region region = {
571 if (b->data_mode != DATA_MODE_VMALLOC) {
572 io_req.mem.type = DM_IO_KMEM;
573 io_req.mem.ptr.addr = (char *)b->data + offset;
575 io_req.mem.type = DM_IO_VMA;
576 io_req.mem.ptr.vma = (char *)b->data + offset;
579 r = dm_io(&io_req, 1, ®ion, NULL);
581 b->end_io(b, errno_to_blk_status(r));
584 static void bio_complete(struct bio *bio)
586 struct dm_buffer *b = bio->bi_private;
587 blk_status_t status = bio->bi_status;
589 b->end_io(b, status);
592 static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
593 unsigned n_sectors, unsigned offset)
597 unsigned vec_size, len;
599 vec_size = b->c->block_size >> PAGE_SHIFT;
600 if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
603 bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
606 use_dmio(b, rw, sector, n_sectors, offset);
610 bio->bi_iter.bi_sector = sector;
611 bio_set_dev(bio, b->c->bdev);
612 bio_set_op_attrs(bio, rw, 0);
613 bio->bi_end_io = bio_complete;
616 ptr = (char *)b->data + offset;
617 len = n_sectors << SECTOR_SHIFT;
620 unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
621 if (!bio_add_page(bio, virt_to_page(ptr), this_step,
622 offset_in_page(ptr))) {
634 static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
638 unsigned offset, end;
642 if (likely(b->c->sectors_per_block_bits >= 0))
643 sector = b->block << b->c->sectors_per_block_bits;
645 sector = b->block * (b->c->block_size >> SECTOR_SHIFT);
646 sector += b->c->start;
648 if (rw != REQ_OP_WRITE) {
649 n_sectors = b->c->block_size >> SECTOR_SHIFT;
652 if (b->c->write_callback)
653 b->c->write_callback(b);
654 offset = b->write_start;
656 offset &= -DM_BUFIO_WRITE_ALIGN;
657 end += DM_BUFIO_WRITE_ALIGN - 1;
658 end &= -DM_BUFIO_WRITE_ALIGN;
659 if (unlikely(end > b->c->block_size))
660 end = b->c->block_size;
662 sector += offset >> SECTOR_SHIFT;
663 n_sectors = (end - offset) >> SECTOR_SHIFT;
666 if (b->data_mode != DATA_MODE_VMALLOC)
667 use_bio(b, rw, sector, n_sectors, offset);
669 use_dmio(b, rw, sector, n_sectors, offset);
672 /*----------------------------------------------------------------
673 * Writing dirty buffers
674 *--------------------------------------------------------------*/
677 * The endio routine for write.
679 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
682 static void write_endio(struct dm_buffer *b, blk_status_t status)
684 b->write_error = status;
685 if (unlikely(status)) {
686 struct dm_bufio_client *c = b->c;
688 (void)cmpxchg(&c->async_write_error, 0,
689 blk_status_to_errno(status));
692 BUG_ON(!test_bit(B_WRITING, &b->state));
694 smp_mb__before_atomic();
695 clear_bit(B_WRITING, &b->state);
696 smp_mb__after_atomic();
698 wake_up_bit(&b->state, B_WRITING);
702 * Initiate a write on a dirty buffer, but don't wait for it.
704 * - If the buffer is not dirty, exit.
705 * - If there some previous write going on, wait for it to finish (we can't
706 * have two writes on the same buffer simultaneously).
707 * - Submit our write and don't wait on it. We set B_WRITING indicating
708 * that there is a write in progress.
710 static void __write_dirty_buffer(struct dm_buffer *b,
711 struct list_head *write_list)
713 if (!test_bit(B_DIRTY, &b->state))
716 clear_bit(B_DIRTY, &b->state);
717 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
719 b->write_start = b->dirty_start;
720 b->write_end = b->dirty_end;
723 submit_io(b, REQ_OP_WRITE, write_endio);
725 list_add_tail(&b->write_list, write_list);
728 static void __flush_write_list(struct list_head *write_list)
730 struct blk_plug plug;
731 blk_start_plug(&plug);
732 while (!list_empty(write_list)) {
733 struct dm_buffer *b =
734 list_entry(write_list->next, struct dm_buffer, write_list);
735 list_del(&b->write_list);
736 submit_io(b, REQ_OP_WRITE, write_endio);
739 blk_finish_plug(&plug);
743 * Wait until any activity on the buffer finishes. Possibly write the
744 * buffer if it is dirty. When this function finishes, there is no I/O
745 * running on the buffer and the buffer is not dirty.
747 static void __make_buffer_clean(struct dm_buffer *b)
749 BUG_ON(b->hold_count);
751 if (!b->state) /* fast case */
754 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
755 __write_dirty_buffer(b, NULL);
756 wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
760 * Find some buffer that is not held by anybody, clean it, unlink it and
763 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
767 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
768 BUG_ON(test_bit(B_WRITING, &b->state));
769 BUG_ON(test_bit(B_DIRTY, &b->state));
771 if (!b->hold_count) {
772 __make_buffer_clean(b);
779 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
780 BUG_ON(test_bit(B_READING, &b->state));
782 if (!b->hold_count) {
783 __make_buffer_clean(b);
794 * Wait until some other threads free some buffer or release hold count on
797 * This function is entered with c->lock held, drops it and regains it
800 static void __wait_for_free_buffer(struct dm_bufio_client *c)
802 DECLARE_WAITQUEUE(wait, current);
804 add_wait_queue(&c->free_buffer_wait, &wait);
805 set_current_state(TASK_UNINTERRUPTIBLE);
810 remove_wait_queue(&c->free_buffer_wait, &wait);
823 * Allocate a new buffer. If the allocation is not possible, wait until
824 * some other thread frees a buffer.
826 * May drop the lock and regain it.
828 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
831 bool tried_noio_alloc = false;
834 * dm-bufio is resistant to allocation failures (it just keeps
835 * one buffer reserved in cases all the allocations fail).
836 * So set flags to not try too hard:
837 * GFP_NOWAIT: don't wait; if we need to sleep we'll release our
838 * mutex and wait ourselves.
839 * __GFP_NORETRY: don't retry and rather return failure
840 * __GFP_NOMEMALLOC: don't use emergency reserves
841 * __GFP_NOWARN: don't print a warning in case of failure
843 * For debugging, if we set the cache size to 1, no new buffers will
847 if (dm_bufio_cache_size_latch != 1) {
848 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
853 if (nf == NF_PREFETCH)
856 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
858 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
862 tried_noio_alloc = true;
865 if (!list_empty(&c->reserved_buffers)) {
866 b = list_entry(c->reserved_buffers.next,
867 struct dm_buffer, lru_list);
868 list_del(&b->lru_list);
869 c->need_reserved_buffers++;
874 b = __get_unclaimed_buffer(c);
878 __wait_for_free_buffer(c);
882 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
884 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
889 if (c->alloc_callback)
890 c->alloc_callback(b);
896 * Free a buffer and wake other threads waiting for free buffers.
898 static void __free_buffer_wake(struct dm_buffer *b)
900 struct dm_bufio_client *c = b->c;
902 if (!c->need_reserved_buffers)
905 list_add(&b->lru_list, &c->reserved_buffers);
906 c->need_reserved_buffers--;
909 wake_up(&c->free_buffer_wait);
912 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
913 struct list_head *write_list)
915 struct dm_buffer *b, *tmp;
917 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
918 BUG_ON(test_bit(B_READING, &b->state));
920 if (!test_bit(B_DIRTY, &b->state) &&
921 !test_bit(B_WRITING, &b->state)) {
922 __relink_lru(b, LIST_CLEAN);
926 if (no_wait && test_bit(B_WRITING, &b->state))
929 __write_dirty_buffer(b, write_list);
935 * Check if we're over watermark.
936 * If we are over threshold_buffers, start freeing buffers.
937 * If we're over "limit_buffers", block until we get under the limit.
939 static void __check_watermark(struct dm_bufio_client *c,
940 struct list_head *write_list)
942 if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
943 __write_dirty_buffers_async(c, 1, write_list);
946 /*----------------------------------------------------------------
948 *--------------------------------------------------------------*/
950 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
951 enum new_flag nf, int *need_submit,
952 struct list_head *write_list)
954 struct dm_buffer *b, *new_b = NULL;
958 b = __find(c, block);
965 new_b = __alloc_buffer_wait(c, nf);
970 * We've had a period where the mutex was unlocked, so need to
971 * recheck the buffer tree.
973 b = __find(c, block);
975 __free_buffer_wake(new_b);
979 __check_watermark(c, write_list);
985 __link_buffer(b, block, LIST_CLEAN);
987 if (nf == NF_FRESH) {
992 b->state = 1 << B_READING;
998 if (nf == NF_PREFETCH)
1001 * Note: it is essential that we don't wait for the buffer to be
1002 * read if dm_bufio_get function is used. Both dm_bufio_get and
1003 * dm_bufio_prefetch can be used in the driver request routine.
1004 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1005 * the same buffer, it would deadlock if we waited.
1007 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1011 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1012 test_bit(B_WRITING, &b->state));
1017 * The endio routine for reading: set the error, clear the bit and wake up
1018 * anyone waiting on the buffer.
1020 static void read_endio(struct dm_buffer *b, blk_status_t status)
1022 b->read_error = status;
1024 BUG_ON(!test_bit(B_READING, &b->state));
1026 smp_mb__before_atomic();
1027 clear_bit(B_READING, &b->state);
1028 smp_mb__after_atomic();
1030 wake_up_bit(&b->state, B_READING);
1034 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
1035 * functions is similar except that dm_bufio_new doesn't read the
1036 * buffer from the disk (assuming that the caller overwrites all the data
1037 * and uses dm_bufio_mark_buffer_dirty to write new data back).
1039 static void *new_read(struct dm_bufio_client *c, sector_t block,
1040 enum new_flag nf, struct dm_buffer **bp)
1043 struct dm_buffer *b;
1045 LIST_HEAD(write_list);
1048 b = __bufio_new(c, block, nf, &need_submit, &write_list);
1049 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1050 if (b && b->hold_count == 1)
1051 buffer_record_stack(b);
1055 __flush_write_list(&write_list);
1061 submit_io(b, REQ_OP_READ, read_endio);
1063 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1065 if (b->read_error) {
1066 int error = blk_status_to_errno(b->read_error);
1068 dm_bufio_release(b);
1070 return ERR_PTR(error);
1078 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1079 struct dm_buffer **bp)
1081 return new_read(c, block, NF_GET, bp);
1083 EXPORT_SYMBOL_GPL(dm_bufio_get);
1085 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1086 struct dm_buffer **bp)
1088 BUG_ON(dm_bufio_in_request());
1090 return new_read(c, block, NF_READ, bp);
1092 EXPORT_SYMBOL_GPL(dm_bufio_read);
1094 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1095 struct dm_buffer **bp)
1097 BUG_ON(dm_bufio_in_request());
1099 return new_read(c, block, NF_FRESH, bp);
1101 EXPORT_SYMBOL_GPL(dm_bufio_new);
1103 void dm_bufio_prefetch(struct dm_bufio_client *c,
1104 sector_t block, unsigned n_blocks)
1106 struct blk_plug plug;
1108 LIST_HEAD(write_list);
1110 BUG_ON(dm_bufio_in_request());
1112 blk_start_plug(&plug);
1115 for (; n_blocks--; block++) {
1117 struct dm_buffer *b;
1118 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1120 if (unlikely(!list_empty(&write_list))) {
1122 blk_finish_plug(&plug);
1123 __flush_write_list(&write_list);
1124 blk_start_plug(&plug);
1127 if (unlikely(b != NULL)) {
1131 submit_io(b, REQ_OP_READ, read_endio);
1132 dm_bufio_release(b);
1145 blk_finish_plug(&plug);
1147 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1149 void dm_bufio_release(struct dm_buffer *b)
1151 struct dm_bufio_client *c = b->c;
1155 BUG_ON(!b->hold_count);
1158 if (!b->hold_count) {
1159 wake_up(&c->free_buffer_wait);
1162 * If there were errors on the buffer, and the buffer is not
1163 * to be written, free the buffer. There is no point in caching
1166 if ((b->read_error || b->write_error) &&
1167 !test_bit(B_READING, &b->state) &&
1168 !test_bit(B_WRITING, &b->state) &&
1169 !test_bit(B_DIRTY, &b->state)) {
1171 __free_buffer_wake(b);
1177 EXPORT_SYMBOL_GPL(dm_bufio_release);
1179 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1180 unsigned start, unsigned end)
1182 struct dm_bufio_client *c = b->c;
1184 BUG_ON(start >= end);
1185 BUG_ON(end > b->c->block_size);
1189 BUG_ON(test_bit(B_READING, &b->state));
1191 if (!test_and_set_bit(B_DIRTY, &b->state)) {
1192 b->dirty_start = start;
1194 __relink_lru(b, LIST_DIRTY);
1196 if (start < b->dirty_start)
1197 b->dirty_start = start;
1198 if (end > b->dirty_end)
1204 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1206 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1208 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1210 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1212 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1214 LIST_HEAD(write_list);
1216 BUG_ON(dm_bufio_in_request());
1219 __write_dirty_buffers_async(c, 0, &write_list);
1221 __flush_write_list(&write_list);
1223 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1226 * For performance, it is essential that the buffers are written asynchronously
1227 * and simultaneously (so that the block layer can merge the writes) and then
1230 * Finally, we flush hardware disk cache.
1232 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1235 unsigned long buffers_processed = 0;
1236 struct dm_buffer *b, *tmp;
1238 LIST_HEAD(write_list);
1241 __write_dirty_buffers_async(c, 0, &write_list);
1243 __flush_write_list(&write_list);
1247 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1248 int dropped_lock = 0;
1250 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1251 buffers_processed++;
1253 BUG_ON(test_bit(B_READING, &b->state));
1255 if (test_bit(B_WRITING, &b->state)) {
1256 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1260 wait_on_bit_io(&b->state, B_WRITING,
1261 TASK_UNINTERRUPTIBLE);
1265 wait_on_bit_io(&b->state, B_WRITING,
1266 TASK_UNINTERRUPTIBLE);
1269 if (!test_bit(B_DIRTY, &b->state) &&
1270 !test_bit(B_WRITING, &b->state))
1271 __relink_lru(b, LIST_CLEAN);
1276 * If we dropped the lock, the list is no longer consistent,
1277 * so we must restart the search.
1279 * In the most common case, the buffer just processed is
1280 * relinked to the clean list, so we won't loop scanning the
1281 * same buffer again and again.
1283 * This may livelock if there is another thread simultaneously
1284 * dirtying buffers, so we count the number of buffers walked
1285 * and if it exceeds the total number of buffers, it means that
1286 * someone is doing some writes simultaneously with us. In
1287 * this case, stop, dropping the lock.
1292 wake_up(&c->free_buffer_wait);
1295 a = xchg(&c->async_write_error, 0);
1296 f = dm_bufio_issue_flush(c);
1302 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1305 * Use dm-io to send an empty barrier to flush the device.
1307 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1309 struct dm_io_request io_req = {
1310 .bi_op = REQ_OP_WRITE,
1311 .bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1312 .mem.type = DM_IO_KMEM,
1313 .mem.ptr.addr = NULL,
1316 struct dm_io_region io_reg = {
1322 BUG_ON(dm_bufio_in_request());
1324 return dm_io(&io_req, 1, &io_reg, NULL);
1326 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1329 * We first delete any other buffer that may be at that new location.
1331 * Then, we write the buffer to the original location if it was dirty.
1333 * Then, if we are the only one who is holding the buffer, relink the buffer
1334 * in the buffer tree for the new location.
1336 * If there was someone else holding the buffer, we write it to the new
1337 * location but not relink it, because that other user needs to have the buffer
1338 * at the same place.
1340 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1342 struct dm_bufio_client *c = b->c;
1343 struct dm_buffer *new;
1345 BUG_ON(dm_bufio_in_request());
1350 new = __find(c, new_block);
1352 if (new->hold_count) {
1353 __wait_for_free_buffer(c);
1358 * FIXME: Is there any point waiting for a write that's going
1359 * to be overwritten in a bit?
1361 __make_buffer_clean(new);
1362 __unlink_buffer(new);
1363 __free_buffer_wake(new);
1366 BUG_ON(!b->hold_count);
1367 BUG_ON(test_bit(B_READING, &b->state));
1369 __write_dirty_buffer(b, NULL);
1370 if (b->hold_count == 1) {
1371 wait_on_bit_io(&b->state, B_WRITING,
1372 TASK_UNINTERRUPTIBLE);
1373 set_bit(B_DIRTY, &b->state);
1375 b->dirty_end = c->block_size;
1377 __link_buffer(b, new_block, LIST_DIRTY);
1380 wait_on_bit_lock_io(&b->state, B_WRITING,
1381 TASK_UNINTERRUPTIBLE);
1383 * Relink buffer to "new_block" so that write_callback
1384 * sees "new_block" as a block number.
1385 * After the write, link the buffer back to old_block.
1386 * All this must be done in bufio lock, so that block number
1387 * change isn't visible to other threads.
1389 old_block = b->block;
1391 __link_buffer(b, new_block, b->list_mode);
1392 submit_io(b, REQ_OP_WRITE, write_endio);
1393 wait_on_bit_io(&b->state, B_WRITING,
1394 TASK_UNINTERRUPTIBLE);
1396 __link_buffer(b, old_block, b->list_mode);
1400 dm_bufio_release(b);
1402 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1405 * Free the given buffer.
1407 * This is just a hint, if the buffer is in use or dirty, this function
1410 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1412 struct dm_buffer *b;
1416 b = __find(c, block);
1417 if (b && likely(!b->hold_count) && likely(!b->state)) {
1419 __free_buffer_wake(b);
1424 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1426 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1428 c->minimum_buffers = n;
1430 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1432 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1434 return c->block_size;
1436 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1438 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1440 sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1445 if (likely(c->sectors_per_block_bits >= 0))
1446 s >>= c->sectors_per_block_bits;
1448 sector_div(s, c->block_size >> SECTOR_SHIFT);
1451 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1453 struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1457 EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1459 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1463 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1465 void *dm_bufio_get_block_data(struct dm_buffer *b)
1469 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1471 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1475 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1477 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1481 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1483 static void drop_buffers(struct dm_bufio_client *c)
1485 struct dm_buffer *b;
1487 bool warned = false;
1489 BUG_ON(dm_bufio_in_request());
1492 * An optimization so that the buffers are not written one-by-one.
1494 dm_bufio_write_dirty_buffers_async(c);
1498 while ((b = __get_unclaimed_buffer(c)))
1499 __free_buffer_wake(b);
1501 for (i = 0; i < LIST_SIZE; i++)
1502 list_for_each_entry(b, &c->lru[i], lru_list) {
1505 DMERR("leaked buffer %llx, hold count %u, list %d",
1506 (unsigned long long)b->block, b->hold_count, i);
1507 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1508 stack_trace_print(b->stack_entries, b->stack_len, 1);
1509 /* mark unclaimed to avoid BUG_ON below */
1514 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1515 while ((b = __get_unclaimed_buffer(c)))
1516 __free_buffer_wake(b);
1519 for (i = 0; i < LIST_SIZE; i++)
1520 BUG_ON(!list_empty(&c->lru[i]));
1526 * We may not be able to evict this buffer if IO pending or the client
1527 * is still using it. Caller is expected to know buffer is too old.
1529 * And if GFP_NOFS is used, we must not do any I/O because we hold
1530 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1531 * rerouted to different bufio client.
1533 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1535 if (!(gfp & __GFP_FS)) {
1536 if (test_bit(B_READING, &b->state) ||
1537 test_bit(B_WRITING, &b->state) ||
1538 test_bit(B_DIRTY, &b->state))
1545 __make_buffer_clean(b);
1547 __free_buffer_wake(b);
1552 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1554 unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1555 if (likely(c->sectors_per_block_bits >= 0))
1556 retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1558 retain_bytes /= c->block_size;
1559 return retain_bytes;
1562 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1566 struct dm_buffer *b, *tmp;
1567 unsigned long freed = 0;
1568 unsigned long count = c->n_buffers[LIST_CLEAN] +
1569 c->n_buffers[LIST_DIRTY];
1570 unsigned long retain_target = get_retain_buffers(c);
1572 for (l = 0; l < LIST_SIZE; l++) {
1573 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1574 if (__try_evict_buffer(b, gfp_mask))
1576 if (!--nr_to_scan || ((count - freed) <= retain_target))
1584 static unsigned long
1585 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1587 struct dm_bufio_client *c;
1588 unsigned long freed;
1590 c = container_of(shrink, struct dm_bufio_client, shrinker);
1591 if (sc->gfp_mask & __GFP_FS)
1593 else if (!dm_bufio_trylock(c))
1596 freed = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1601 static unsigned long
1602 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1604 struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1605 unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1606 READ_ONCE(c->n_buffers[LIST_DIRTY]);
1607 unsigned long retain_target = get_retain_buffers(c);
1609 return (count < retain_target) ? 0 : (count - retain_target);
1613 * Create the buffering interface
1615 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1616 unsigned reserved_buffers, unsigned aux_size,
1617 void (*alloc_callback)(struct dm_buffer *),
1618 void (*write_callback)(struct dm_buffer *))
1621 struct dm_bufio_client *c;
1625 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1626 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1631 c = kzalloc(sizeof(*c), GFP_KERNEL);
1636 c->buffer_tree = RB_ROOT;
1639 c->block_size = block_size;
1640 if (is_power_of_2(block_size))
1641 c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1643 c->sectors_per_block_bits = -1;
1645 c->alloc_callback = alloc_callback;
1646 c->write_callback = write_callback;
1648 for (i = 0; i < LIST_SIZE; i++) {
1649 INIT_LIST_HEAD(&c->lru[i]);
1650 c->n_buffers[i] = 0;
1653 mutex_init(&c->lock);
1654 INIT_LIST_HEAD(&c->reserved_buffers);
1655 c->need_reserved_buffers = reserved_buffers;
1657 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1659 init_waitqueue_head(&c->free_buffer_wait);
1660 c->async_write_error = 0;
1662 c->dm_io = dm_io_client_create();
1663 if (IS_ERR(c->dm_io)) {
1664 r = PTR_ERR(c->dm_io);
1668 if (block_size <= KMALLOC_MAX_SIZE &&
1669 (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1670 unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1671 snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1672 c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1673 SLAB_RECLAIM_ACCOUNT, NULL);
1674 if (!c->slab_cache) {
1680 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1682 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1683 c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1684 0, SLAB_RECLAIM_ACCOUNT, NULL);
1685 if (!c->slab_buffer) {
1690 while (c->need_reserved_buffers) {
1691 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1697 __free_buffer_wake(b);
1700 c->shrinker.count_objects = dm_bufio_shrink_count;
1701 c->shrinker.scan_objects = dm_bufio_shrink_scan;
1702 c->shrinker.seeks = 1;
1703 c->shrinker.batch = 0;
1704 r = register_shrinker(&c->shrinker);
1708 mutex_lock(&dm_bufio_clients_lock);
1709 dm_bufio_client_count++;
1710 list_add(&c->client_list, &dm_bufio_all_clients);
1711 __cache_size_refresh();
1712 mutex_unlock(&dm_bufio_clients_lock);
1717 while (!list_empty(&c->reserved_buffers)) {
1718 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1719 struct dm_buffer, lru_list);
1720 list_del(&b->lru_list);
1723 kmem_cache_destroy(c->slab_cache);
1724 kmem_cache_destroy(c->slab_buffer);
1725 dm_io_client_destroy(c->dm_io);
1727 mutex_destroy(&c->lock);
1732 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1735 * Free the buffering interface.
1736 * It is required that there are no references on any buffers.
1738 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1744 unregister_shrinker(&c->shrinker);
1746 mutex_lock(&dm_bufio_clients_lock);
1748 list_del(&c->client_list);
1749 dm_bufio_client_count--;
1750 __cache_size_refresh();
1752 mutex_unlock(&dm_bufio_clients_lock);
1754 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1755 BUG_ON(c->need_reserved_buffers);
1757 while (!list_empty(&c->reserved_buffers)) {
1758 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1759 struct dm_buffer, lru_list);
1760 list_del(&b->lru_list);
1764 for (i = 0; i < LIST_SIZE; i++)
1765 if (c->n_buffers[i])
1766 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1768 for (i = 0; i < LIST_SIZE; i++)
1769 BUG_ON(c->n_buffers[i]);
1771 kmem_cache_destroy(c->slab_cache);
1772 kmem_cache_destroy(c->slab_buffer);
1773 dm_io_client_destroy(c->dm_io);
1774 mutex_destroy(&c->lock);
1777 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1779 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1783 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1785 static unsigned get_max_age_hz(void)
1787 unsigned max_age = READ_ONCE(dm_bufio_max_age);
1789 if (max_age > UINT_MAX / HZ)
1790 max_age = UINT_MAX / HZ;
1792 return max_age * HZ;
1795 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1797 return time_after_eq(jiffies, b->last_accessed + age_hz);
1800 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1802 struct dm_buffer *b, *tmp;
1803 unsigned long retain_target = get_retain_buffers(c);
1804 unsigned long count;
1805 LIST_HEAD(write_list);
1809 __check_watermark(c, &write_list);
1810 if (unlikely(!list_empty(&write_list))) {
1812 __flush_write_list(&write_list);
1816 count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1817 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1818 if (count <= retain_target)
1821 if (!older_than(b, age_hz))
1824 if (__try_evict_buffer(b, 0))
1833 static void do_global_cleanup(struct work_struct *w)
1835 struct dm_bufio_client *locked_client = NULL;
1836 struct dm_bufio_client *current_client;
1837 struct dm_buffer *b;
1838 unsigned spinlock_hold_count;
1839 unsigned long threshold = dm_bufio_cache_size -
1840 dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1841 unsigned long loops = global_num * 2;
1843 mutex_lock(&dm_bufio_clients_lock);
1848 spin_lock(&global_spinlock);
1849 if (unlikely(dm_bufio_current_allocated <= threshold))
1852 spinlock_hold_count = 0;
1856 if (unlikely(list_empty(&global_queue)))
1858 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1862 list_move(&b->global_list, &global_queue);
1863 if (likely(++spinlock_hold_count < 16))
1865 spin_unlock(&global_spinlock);
1869 current_client = b->c;
1870 if (unlikely(current_client != locked_client)) {
1872 dm_bufio_unlock(locked_client);
1874 if (!dm_bufio_trylock(current_client)) {
1875 spin_unlock(&global_spinlock);
1876 dm_bufio_lock(current_client);
1877 locked_client = current_client;
1881 locked_client = current_client;
1884 spin_unlock(&global_spinlock);
1886 if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
1887 spin_lock(&global_spinlock);
1888 list_move(&b->global_list, &global_queue);
1889 spin_unlock(&global_spinlock);
1893 spin_unlock(&global_spinlock);
1896 dm_bufio_unlock(locked_client);
1898 mutex_unlock(&dm_bufio_clients_lock);
1901 static void cleanup_old_buffers(void)
1903 unsigned long max_age_hz = get_max_age_hz();
1904 struct dm_bufio_client *c;
1906 mutex_lock(&dm_bufio_clients_lock);
1908 __cache_size_refresh();
1910 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1911 __evict_old_buffers(c, max_age_hz);
1913 mutex_unlock(&dm_bufio_clients_lock);
1916 static void work_fn(struct work_struct *w)
1918 cleanup_old_buffers();
1920 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1921 DM_BUFIO_WORK_TIMER_SECS * HZ);
1924 /*----------------------------------------------------------------
1926 *--------------------------------------------------------------*/
1929 * This is called only once for the whole dm_bufio module.
1930 * It initializes memory limit.
1932 static int __init dm_bufio_init(void)
1936 dm_bufio_allocated_kmem_cache = 0;
1937 dm_bufio_allocated_get_free_pages = 0;
1938 dm_bufio_allocated_vmalloc = 0;
1939 dm_bufio_current_allocated = 0;
1941 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
1942 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
1944 if (mem > ULONG_MAX)
1948 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
1949 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
1952 dm_bufio_default_cache_size = mem;
1954 mutex_lock(&dm_bufio_clients_lock);
1955 __cache_size_refresh();
1956 mutex_unlock(&dm_bufio_clients_lock);
1958 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1962 INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
1963 INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
1964 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
1965 DM_BUFIO_WORK_TIMER_SECS * HZ);
1971 * This is called once when unloading the dm_bufio module.
1973 static void __exit dm_bufio_exit(void)
1977 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
1978 flush_workqueue(dm_bufio_wq);
1979 destroy_workqueue(dm_bufio_wq);
1981 if (dm_bufio_client_count) {
1982 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1983 __func__, dm_bufio_client_count);
1987 if (dm_bufio_current_allocated) {
1988 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1989 __func__, dm_bufio_current_allocated);
1993 if (dm_bufio_allocated_get_free_pages) {
1994 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1995 __func__, dm_bufio_allocated_get_free_pages);
1999 if (dm_bufio_allocated_vmalloc) {
2000 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2001 __func__, dm_bufio_allocated_vmalloc);
2008 module_init(dm_bufio_init)
2009 module_exit(dm_bufio_exit)
2011 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2012 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2014 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2015 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2017 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2018 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2020 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2021 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2023 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2024 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2026 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2027 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2029 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2030 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2032 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2033 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2035 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2036 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2037 MODULE_LICENSE("GPL");