2 * Copyright (C) 2009-2011 Red Hat, Inc.
4 * Author: Mikulas Patocka <mpatocka@redhat.com>
6 * This file is released under the GPL.
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/shrinker.h>
16 #include <linux/module.h>
18 #define DM_MSG_PREFIX "bufio"
21 * Memory management policy:
22 * Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
23 * or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
24 * Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
25 * Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
28 #define DM_BUFIO_MIN_BUFFERS 8
30 #define DM_BUFIO_MEMORY_PERCENT 2
31 #define DM_BUFIO_VMALLOC_PERCENT 25
32 #define DM_BUFIO_WRITEBACK_PERCENT 75
35 * Check buffer ages in this interval (seconds)
37 #define DM_BUFIO_WORK_TIMER_SECS 10
40 * Free buffers when they are older than this (seconds)
42 #define DM_BUFIO_DEFAULT_AGE_SECS 60
45 * The number of bvec entries that are embedded directly in the buffer.
46 * If the chunk size is larger, dm-io is used to do the io.
48 #define DM_BUFIO_INLINE_VECS 16
53 #define DM_BUFIO_HASH_BITS 20
54 #define DM_BUFIO_HASH(block) \
55 ((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
56 ((1 << DM_BUFIO_HASH_BITS) - 1))
59 * Don't try to use kmem_cache_alloc for blocks larger than this.
60 * For explanation, see alloc_buffer_data below.
62 #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT (PAGE_SIZE >> 1)
63 #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT (PAGE_SIZE << (MAX_ORDER - 1))
66 * dm_buffer->list_mode
74 * All buffers are linked to cache_hash with their hash_list field.
76 * Clean buffers that are not being written (B_WRITING not set)
77 * are linked to lru[LIST_CLEAN] with their lru_list field.
79 * Dirty and clean buffers that are being written are linked to
80 * lru[LIST_DIRTY] with their lru_list field. When the write
81 * finishes, the buffer cannot be relinked immediately (because we
82 * are in an interrupt context and relinking requires process
83 * context), so some clean-not-writing buffers can be held on
84 * dirty_lru too. They are later added to lru in the process
87 struct dm_bufio_client {
90 struct list_head lru[LIST_SIZE];
91 unsigned long n_buffers[LIST_SIZE];
93 struct block_device *bdev;
95 unsigned char sectors_per_block_bits;
96 unsigned char pages_per_block_bits;
97 unsigned char blocks_per_page_bits;
99 void (*alloc_callback)(struct dm_buffer *);
100 void (*write_callback)(struct dm_buffer *);
102 struct dm_io_client *dm_io;
104 struct list_head reserved_buffers;
105 unsigned need_reserved_buffers;
107 struct hlist_head *cache_hash;
108 wait_queue_head_t free_buffer_wait;
110 int async_write_error;
112 struct list_head client_list;
113 struct shrinker shrinker;
124 * Describes how the block was allocated:
125 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
126 * See the comment at alloc_buffer_data.
130 DATA_MODE_GET_FREE_PAGES = 1,
131 DATA_MODE_VMALLOC = 2,
136 struct hlist_node hash_list;
137 struct list_head lru_list;
140 enum data_mode data_mode;
141 unsigned char list_mode; /* LIST_* */
146 unsigned long last_accessed;
147 struct dm_bufio_client *c;
149 struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
152 /*----------------------------------------------------------------*/
154 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
155 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
157 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
159 unsigned ret = c->blocks_per_page_bits - 1;
161 BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
166 #define DM_BUFIO_CACHE(c) (dm_bufio_caches[dm_bufio_cache_index(c)])
167 #define DM_BUFIO_CACHE_NAME(c) (dm_bufio_cache_names[dm_bufio_cache_index(c)])
169 #define dm_bufio_in_request() (!!current->bio_list)
171 static void dm_bufio_lock(struct dm_bufio_client *c)
173 mutex_lock_nested(&c->lock, dm_bufio_in_request());
176 static int dm_bufio_trylock(struct dm_bufio_client *c)
178 return mutex_trylock(&c->lock);
181 static void dm_bufio_unlock(struct dm_bufio_client *c)
183 mutex_unlock(&c->lock);
187 * FIXME Move to sched.h?
189 #ifdef CONFIG_PREEMPT_VOLUNTARY
190 # define dm_bufio_cond_resched() \
192 if (unlikely(need_resched())) \
196 # define dm_bufio_cond_resched() do { } while (0)
199 /*----------------------------------------------------------------*/
202 * Default cache size: available memory divided by the ratio.
204 static unsigned long dm_bufio_default_cache_size;
207 * Total cache size set by the user.
209 static unsigned long dm_bufio_cache_size;
212 * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
213 * at any time. If it disagrees, the user has changed cache size.
215 static unsigned long dm_bufio_cache_size_latch;
217 static DEFINE_SPINLOCK(param_spinlock);
220 * Buffers are freed after this timeout
222 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
224 static unsigned long dm_bufio_peak_allocated;
225 static unsigned long dm_bufio_allocated_kmem_cache;
226 static unsigned long dm_bufio_allocated_get_free_pages;
227 static unsigned long dm_bufio_allocated_vmalloc;
228 static unsigned long dm_bufio_current_allocated;
230 /*----------------------------------------------------------------*/
233 * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
235 static unsigned long dm_bufio_cache_size_per_client;
238 * The current number of clients.
240 static int dm_bufio_client_count;
243 * The list of all clients.
245 static LIST_HEAD(dm_bufio_all_clients);
248 * This mutex protects dm_bufio_cache_size_latch,
249 * dm_bufio_cache_size_per_client and dm_bufio_client_count
251 static DEFINE_MUTEX(dm_bufio_clients_lock);
253 /*----------------------------------------------------------------*/
255 static void adjust_total_allocated(enum data_mode data_mode, long diff)
257 static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
258 &dm_bufio_allocated_kmem_cache,
259 &dm_bufio_allocated_get_free_pages,
260 &dm_bufio_allocated_vmalloc,
263 spin_lock(¶m_spinlock);
265 *class_ptr[data_mode] += diff;
267 dm_bufio_current_allocated += diff;
269 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
270 dm_bufio_peak_allocated = dm_bufio_current_allocated;
272 spin_unlock(¶m_spinlock);
276 * Change the number of clients and recalculate per-client limit.
278 static void __cache_size_refresh(void)
280 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
281 BUG_ON(dm_bufio_client_count < 0);
283 dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
286 * Use default if set to 0 and report the actual cache size used.
288 if (!dm_bufio_cache_size_latch) {
289 (void)cmpxchg(&dm_bufio_cache_size, 0,
290 dm_bufio_default_cache_size);
291 dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
294 dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
295 (dm_bufio_client_count ? : 1);
299 * Allocating buffer data.
301 * Small buffers are allocated with kmem_cache, to use space optimally.
303 * For large buffers, we choose between get_free_pages and vmalloc.
304 * Each has advantages and disadvantages.
306 * __get_free_pages can randomly fail if the memory is fragmented.
307 * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
308 * as low as 128M) so using it for caching is not appropriate.
310 * If the allocation may fail we use __get_free_pages. Memory fragmentation
311 * won't have a fatal effect here, but it just causes flushes of some other
312 * buffers and more I/O will be performed. Don't use __get_free_pages if it
313 * always fails (i.e. order >= MAX_ORDER).
315 * If the allocation shouldn't fail we use __vmalloc. This is only for the
316 * initial reserve allocation, so there's no risk of wasting all vmalloc
319 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
320 enum data_mode *data_mode)
322 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
323 *data_mode = DATA_MODE_SLAB;
324 return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
327 if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
328 gfp_mask & __GFP_NORETRY) {
329 *data_mode = DATA_MODE_GET_FREE_PAGES;
330 return (void *)__get_free_pages(gfp_mask,
331 c->pages_per_block_bits);
334 *data_mode = DATA_MODE_VMALLOC;
335 return __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
339 * Free buffer's data.
341 static void free_buffer_data(struct dm_bufio_client *c,
342 void *data, enum data_mode data_mode)
346 kmem_cache_free(DM_BUFIO_CACHE(c), data);
349 case DATA_MODE_GET_FREE_PAGES:
350 free_pages((unsigned long)data, c->pages_per_block_bits);
353 case DATA_MODE_VMALLOC:
358 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
365 * Allocate buffer and its data.
367 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
369 struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
377 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
383 adjust_total_allocated(b->data_mode, (long)c->block_size);
389 * Free buffer and its data.
391 static void free_buffer(struct dm_buffer *b)
393 struct dm_bufio_client *c = b->c;
395 adjust_total_allocated(b->data_mode, -(long)c->block_size);
397 free_buffer_data(c, b->data, b->data_mode);
402 * Link buffer to the hash list and clean or dirty queue.
404 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
406 struct dm_bufio_client *c = b->c;
408 c->n_buffers[dirty]++;
410 b->list_mode = dirty;
411 list_add(&b->lru_list, &c->lru[dirty]);
412 hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
413 b->last_accessed = jiffies;
417 * Unlink buffer from the hash list and dirty or clean queue.
419 static void __unlink_buffer(struct dm_buffer *b)
421 struct dm_bufio_client *c = b->c;
423 BUG_ON(!c->n_buffers[b->list_mode]);
425 c->n_buffers[b->list_mode]--;
426 hlist_del(&b->hash_list);
427 list_del(&b->lru_list);
431 * Place the buffer to the head of dirty or clean LRU queue.
433 static void __relink_lru(struct dm_buffer *b, int dirty)
435 struct dm_bufio_client *c = b->c;
437 BUG_ON(!c->n_buffers[b->list_mode]);
439 c->n_buffers[b->list_mode]--;
440 c->n_buffers[dirty]++;
441 b->list_mode = dirty;
442 list_move(&b->lru_list, &c->lru[dirty]);
445 /*----------------------------------------------------------------
446 * Submit I/O on the buffer.
448 * Bio interface is faster but it has some problems:
449 * the vector list is limited (increasing this limit increases
450 * memory-consumption per buffer, so it is not viable);
452 * the memory must be direct-mapped, not vmalloced;
454 * the I/O driver can reject requests spuriously if it thinks that
455 * the requests are too big for the device or if they cross a
456 * controller-defined memory boundary.
458 * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
459 * it is not vmalloced, try using the bio interface.
461 * If the buffer is big, if it is vmalloced or if the underlying device
462 * rejects the bio because it is too large, use dm-io layer to do the I/O.
463 * The dm-io layer splits the I/O into multiple requests, avoiding the above
465 *--------------------------------------------------------------*/
468 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
469 * that the request was handled directly with bio interface.
471 static void dmio_complete(unsigned long error, void *context)
473 struct dm_buffer *b = context;
475 b->bio.bi_end_io(&b->bio, error ? -EIO : 0, NULL);
478 static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
479 bio_end_io_t *end_io)
482 struct dm_io_request io_req = {
484 .notify.fn = dmio_complete,
486 .client = b->c->dm_io,
488 struct dm_io_region region = {
490 .sector = block << b->c->sectors_per_block_bits,
491 .count = b->c->block_size >> SECTOR_SHIFT,
494 if (b->data_mode != DATA_MODE_VMALLOC) {
495 io_req.mem.type = DM_IO_KMEM;
496 io_req.mem.ptr.addr = b->data;
498 io_req.mem.type = DM_IO_VMA;
499 io_req.mem.ptr.vma = b->data;
502 b->bio.bi_end_io = end_io;
504 r = dm_io(&io_req, 1, ®ion, NULL);
506 end_io(&b->bio, r, NULL);
509 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
510 bio_end_io_t *end_io)
516 b->bio.bi_io_vec = b->bio_vec;
517 b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
518 b->bio.bi_sector = block << b->c->sectors_per_block_bits;
519 b->bio.bi_bdev = b->c->bdev;
520 b->bio.bi_end_io = end_io;
523 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
524 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
527 len = b->c->block_size;
529 if (len >= PAGE_SIZE)
530 BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
532 BUG_ON((unsigned long)ptr & (len - 1));
535 if (!bio_add_page(&b->bio, virt_to_page(ptr),
536 len < PAGE_SIZE ? len : PAGE_SIZE,
537 virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
538 BUG_ON(b->c->block_size <= PAGE_SIZE);
539 use_dmio(b, rw, block, end_io);
547 submit_bio(rw, &b->bio);
550 static void submit_io(struct dm_buffer *b, int rw, sector_t block,
551 bio_end_io_t *end_io)
553 if (rw == WRITE && b->c->write_callback)
554 b->c->write_callback(b);
556 if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
557 b->data_mode != DATA_MODE_VMALLOC)
558 use_inline_bio(b, rw, block, end_io);
560 use_dmio(b, rw, block, end_io);
563 /*----------------------------------------------------------------
564 * Writing dirty buffers
565 *--------------------------------------------------------------*/
568 * The endio routine for write.
570 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
573 static void write_endio(struct bio *bio, int error,
574 struct batch_complete *batch)
576 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
578 b->write_error = error;
579 if (unlikely(error)) {
580 struct dm_bufio_client *c = b->c;
581 (void)cmpxchg(&c->async_write_error, 0, error);
584 BUG_ON(!test_bit(B_WRITING, &b->state));
586 smp_mb__before_clear_bit();
587 clear_bit(B_WRITING, &b->state);
588 smp_mb__after_clear_bit();
590 wake_up_bit(&b->state, B_WRITING);
594 * This function is called when wait_on_bit is actually waiting.
596 static int do_io_schedule(void *word)
604 * Initiate a write on a dirty buffer, but don't wait for it.
606 * - If the buffer is not dirty, exit.
607 * - If there some previous write going on, wait for it to finish (we can't
608 * have two writes on the same buffer simultaneously).
609 * - Submit our write and don't wait on it. We set B_WRITING indicating
610 * that there is a write in progress.
612 static void __write_dirty_buffer(struct dm_buffer *b)
614 if (!test_bit(B_DIRTY, &b->state))
617 clear_bit(B_DIRTY, &b->state);
618 wait_on_bit_lock(&b->state, B_WRITING,
619 do_io_schedule, TASK_UNINTERRUPTIBLE);
621 submit_io(b, WRITE, b->block, write_endio);
625 * Wait until any activity on the buffer finishes. Possibly write the
626 * buffer if it is dirty. When this function finishes, there is no I/O
627 * running on the buffer and the buffer is not dirty.
629 static void __make_buffer_clean(struct dm_buffer *b)
631 BUG_ON(b->hold_count);
633 if (!b->state) /* fast case */
636 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
637 __write_dirty_buffer(b);
638 wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
642 * Find some buffer that is not held by anybody, clean it, unlink it and
645 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
649 list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
650 BUG_ON(test_bit(B_WRITING, &b->state));
651 BUG_ON(test_bit(B_DIRTY, &b->state));
653 if (!b->hold_count) {
654 __make_buffer_clean(b);
658 dm_bufio_cond_resched();
661 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
662 BUG_ON(test_bit(B_READING, &b->state));
664 if (!b->hold_count) {
665 __make_buffer_clean(b);
669 dm_bufio_cond_resched();
676 * Wait until some other threads free some buffer or release hold count on
679 * This function is entered with c->lock held, drops it and regains it
682 static void __wait_for_free_buffer(struct dm_bufio_client *c)
684 DECLARE_WAITQUEUE(wait, current);
686 add_wait_queue(&c->free_buffer_wait, &wait);
687 set_task_state(current, TASK_UNINTERRUPTIBLE);
692 set_task_state(current, TASK_RUNNING);
693 remove_wait_queue(&c->free_buffer_wait, &wait);
706 * Allocate a new buffer. If the allocation is not possible, wait until
707 * some other thread frees a buffer.
709 * May drop the lock and regain it.
711 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
716 * dm-bufio is resistant to allocation failures (it just keeps
717 * one buffer reserved in cases all the allocations fail).
718 * So set flags to not try too hard:
719 * GFP_NOIO: don't recurse into the I/O layer
720 * __GFP_NORETRY: don't retry and rather return failure
721 * __GFP_NOMEMALLOC: don't use emergency reserves
722 * __GFP_NOWARN: don't print a warning in case of failure
724 * For debugging, if we set the cache size to 1, no new buffers will
728 if (dm_bufio_cache_size_latch != 1) {
729 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
734 if (nf == NF_PREFETCH)
737 if (!list_empty(&c->reserved_buffers)) {
738 b = list_entry(c->reserved_buffers.next,
739 struct dm_buffer, lru_list);
740 list_del(&b->lru_list);
741 c->need_reserved_buffers++;
746 b = __get_unclaimed_buffer(c);
750 __wait_for_free_buffer(c);
754 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
756 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
761 if (c->alloc_callback)
762 c->alloc_callback(b);
768 * Free a buffer and wake other threads waiting for free buffers.
770 static void __free_buffer_wake(struct dm_buffer *b)
772 struct dm_bufio_client *c = b->c;
774 if (!c->need_reserved_buffers)
777 list_add(&b->lru_list, &c->reserved_buffers);
778 c->need_reserved_buffers--;
781 wake_up(&c->free_buffer_wait);
784 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait)
786 struct dm_buffer *b, *tmp;
788 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
789 BUG_ON(test_bit(B_READING, &b->state));
791 if (!test_bit(B_DIRTY, &b->state) &&
792 !test_bit(B_WRITING, &b->state)) {
793 __relink_lru(b, LIST_CLEAN);
797 if (no_wait && test_bit(B_WRITING, &b->state))
800 __write_dirty_buffer(b);
801 dm_bufio_cond_resched();
806 * Get writeback threshold and buffer limit for a given client.
808 static void __get_memory_limit(struct dm_bufio_client *c,
809 unsigned long *threshold_buffers,
810 unsigned long *limit_buffers)
812 unsigned long buffers;
814 if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
815 mutex_lock(&dm_bufio_clients_lock);
816 __cache_size_refresh();
817 mutex_unlock(&dm_bufio_clients_lock);
820 buffers = dm_bufio_cache_size_per_client >>
821 (c->sectors_per_block_bits + SECTOR_SHIFT);
823 if (buffers < DM_BUFIO_MIN_BUFFERS)
824 buffers = DM_BUFIO_MIN_BUFFERS;
826 *limit_buffers = buffers;
827 *threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
831 * Check if we're over watermark.
832 * If we are over threshold_buffers, start freeing buffers.
833 * If we're over "limit_buffers", block until we get under the limit.
835 static void __check_watermark(struct dm_bufio_client *c)
837 unsigned long threshold_buffers, limit_buffers;
839 __get_memory_limit(c, &threshold_buffers, &limit_buffers);
841 while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
844 struct dm_buffer *b = __get_unclaimed_buffer(c);
849 __free_buffer_wake(b);
850 dm_bufio_cond_resched();
853 if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
854 __write_dirty_buffers_async(c, 1);
858 * Find a buffer in the hash.
860 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
864 hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
866 dm_bufio_cond_resched();
867 if (b->block == block)
874 /*----------------------------------------------------------------
876 *--------------------------------------------------------------*/
878 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
879 enum new_flag nf, int *need_submit)
881 struct dm_buffer *b, *new_b = NULL;
885 b = __find(c, block);
892 new_b = __alloc_buffer_wait(c, nf);
897 * We've had a period where the mutex was unlocked, so need to
898 * recheck the hash table.
900 b = __find(c, block);
902 __free_buffer_wake(new_b);
906 __check_watermark(c);
912 __link_buffer(b, block, LIST_CLEAN);
914 if (nf == NF_FRESH) {
919 b->state = 1 << B_READING;
925 if (nf == NF_PREFETCH)
928 * Note: it is essential that we don't wait for the buffer to be
929 * read if dm_bufio_get function is used. Both dm_bufio_get and
930 * dm_bufio_prefetch can be used in the driver request routine.
931 * If the user called both dm_bufio_prefetch and dm_bufio_get on
932 * the same buffer, it would deadlock if we waited.
934 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
938 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
939 test_bit(B_WRITING, &b->state));
944 * The endio routine for reading: set the error, clear the bit and wake up
945 * anyone waiting on the buffer.
947 static void read_endio(struct bio *bio, int error, struct batch_complete *batch)
949 struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
951 b->read_error = error;
953 BUG_ON(!test_bit(B_READING, &b->state));
955 smp_mb__before_clear_bit();
956 clear_bit(B_READING, &b->state);
957 smp_mb__after_clear_bit();
959 wake_up_bit(&b->state, B_READING);
963 * A common routine for dm_bufio_new and dm_bufio_read. Operation of these
964 * functions is similar except that dm_bufio_new doesn't read the
965 * buffer from the disk (assuming that the caller overwrites all the data
966 * and uses dm_bufio_mark_buffer_dirty to write new data back).
968 static void *new_read(struct dm_bufio_client *c, sector_t block,
969 enum new_flag nf, struct dm_buffer **bp)
975 b = __bufio_new(c, block, nf, &need_submit);
982 submit_io(b, READ, b->block, read_endio);
984 wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
987 int error = b->read_error;
991 return ERR_PTR(error);
999 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1000 struct dm_buffer **bp)
1002 return new_read(c, block, NF_GET, bp);
1004 EXPORT_SYMBOL_GPL(dm_bufio_get);
1006 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1007 struct dm_buffer **bp)
1009 BUG_ON(dm_bufio_in_request());
1011 return new_read(c, block, NF_READ, bp);
1013 EXPORT_SYMBOL_GPL(dm_bufio_read);
1015 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1016 struct dm_buffer **bp)
1018 BUG_ON(dm_bufio_in_request());
1020 return new_read(c, block, NF_FRESH, bp);
1022 EXPORT_SYMBOL_GPL(dm_bufio_new);
1024 void dm_bufio_prefetch(struct dm_bufio_client *c,
1025 sector_t block, unsigned n_blocks)
1027 struct blk_plug plug;
1029 BUG_ON(dm_bufio_in_request());
1031 blk_start_plug(&plug);
1034 for (; n_blocks--; block++) {
1036 struct dm_buffer *b;
1037 b = __bufio_new(c, block, NF_PREFETCH, &need_submit);
1038 if (unlikely(b != NULL)) {
1042 submit_io(b, READ, b->block, read_endio);
1043 dm_bufio_release(b);
1045 dm_bufio_cond_resched();
1057 blk_finish_plug(&plug);
1059 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1061 void dm_bufio_release(struct dm_buffer *b)
1063 struct dm_bufio_client *c = b->c;
1067 BUG_ON(!b->hold_count);
1070 if (!b->hold_count) {
1071 wake_up(&c->free_buffer_wait);
1074 * If there were errors on the buffer, and the buffer is not
1075 * to be written, free the buffer. There is no point in caching
1078 if ((b->read_error || b->write_error) &&
1079 !test_bit(B_READING, &b->state) &&
1080 !test_bit(B_WRITING, &b->state) &&
1081 !test_bit(B_DIRTY, &b->state)) {
1083 __free_buffer_wake(b);
1089 EXPORT_SYMBOL_GPL(dm_bufio_release);
1091 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1093 struct dm_bufio_client *c = b->c;
1097 BUG_ON(test_bit(B_READING, &b->state));
1099 if (!test_and_set_bit(B_DIRTY, &b->state))
1100 __relink_lru(b, LIST_DIRTY);
1104 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1106 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1108 BUG_ON(dm_bufio_in_request());
1111 __write_dirty_buffers_async(c, 0);
1114 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1117 * For performance, it is essential that the buffers are written asynchronously
1118 * and simultaneously (so that the block layer can merge the writes) and then
1121 * Finally, we flush hardware disk cache.
1123 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1126 unsigned long buffers_processed = 0;
1127 struct dm_buffer *b, *tmp;
1130 __write_dirty_buffers_async(c, 0);
1133 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1134 int dropped_lock = 0;
1136 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1137 buffers_processed++;
1139 BUG_ON(test_bit(B_READING, &b->state));
1141 if (test_bit(B_WRITING, &b->state)) {
1142 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1146 wait_on_bit(&b->state, B_WRITING,
1148 TASK_UNINTERRUPTIBLE);
1152 wait_on_bit(&b->state, B_WRITING,
1154 TASK_UNINTERRUPTIBLE);
1157 if (!test_bit(B_DIRTY, &b->state) &&
1158 !test_bit(B_WRITING, &b->state))
1159 __relink_lru(b, LIST_CLEAN);
1161 dm_bufio_cond_resched();
1164 * If we dropped the lock, the list is no longer consistent,
1165 * so we must restart the search.
1167 * In the most common case, the buffer just processed is
1168 * relinked to the clean list, so we won't loop scanning the
1169 * same buffer again and again.
1171 * This may livelock if there is another thread simultaneously
1172 * dirtying buffers, so we count the number of buffers walked
1173 * and if it exceeds the total number of buffers, it means that
1174 * someone is doing some writes simultaneously with us. In
1175 * this case, stop, dropping the lock.
1180 wake_up(&c->free_buffer_wait);
1183 a = xchg(&c->async_write_error, 0);
1184 f = dm_bufio_issue_flush(c);
1190 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1193 * Use dm-io to send and empty barrier flush the device.
1195 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1197 struct dm_io_request io_req = {
1198 .bi_rw = WRITE_FLUSH,
1199 .mem.type = DM_IO_KMEM,
1200 .mem.ptr.addr = NULL,
1203 struct dm_io_region io_reg = {
1209 BUG_ON(dm_bufio_in_request());
1211 return dm_io(&io_req, 1, &io_reg, NULL);
1213 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1216 * We first delete any other buffer that may be at that new location.
1218 * Then, we write the buffer to the original location if it was dirty.
1220 * Then, if we are the only one who is holding the buffer, relink the buffer
1221 * in the hash queue for the new location.
1223 * If there was someone else holding the buffer, we write it to the new
1224 * location but not relink it, because that other user needs to have the buffer
1225 * at the same place.
1227 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1229 struct dm_bufio_client *c = b->c;
1230 struct dm_buffer *new;
1232 BUG_ON(dm_bufio_in_request());
1237 new = __find(c, new_block);
1239 if (new->hold_count) {
1240 __wait_for_free_buffer(c);
1245 * FIXME: Is there any point waiting for a write that's going
1246 * to be overwritten in a bit?
1248 __make_buffer_clean(new);
1249 __unlink_buffer(new);
1250 __free_buffer_wake(new);
1253 BUG_ON(!b->hold_count);
1254 BUG_ON(test_bit(B_READING, &b->state));
1256 __write_dirty_buffer(b);
1257 if (b->hold_count == 1) {
1258 wait_on_bit(&b->state, B_WRITING,
1259 do_io_schedule, TASK_UNINTERRUPTIBLE);
1260 set_bit(B_DIRTY, &b->state);
1262 __link_buffer(b, new_block, LIST_DIRTY);
1265 wait_on_bit_lock(&b->state, B_WRITING,
1266 do_io_schedule, TASK_UNINTERRUPTIBLE);
1268 * Relink buffer to "new_block" so that write_callback
1269 * sees "new_block" as a block number.
1270 * After the write, link the buffer back to old_block.
1271 * All this must be done in bufio lock, so that block number
1272 * change isn't visible to other threads.
1274 old_block = b->block;
1276 __link_buffer(b, new_block, b->list_mode);
1277 submit_io(b, WRITE, new_block, write_endio);
1278 wait_on_bit(&b->state, B_WRITING,
1279 do_io_schedule, TASK_UNINTERRUPTIBLE);
1281 __link_buffer(b, old_block, b->list_mode);
1285 dm_bufio_release(b);
1287 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1289 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1291 return c->block_size;
1293 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1295 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1297 return i_size_read(c->bdev->bd_inode) >>
1298 (SECTOR_SHIFT + c->sectors_per_block_bits);
1300 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1302 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1306 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1308 void *dm_bufio_get_block_data(struct dm_buffer *b)
1312 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1314 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1318 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1320 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1324 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1326 static void drop_buffers(struct dm_bufio_client *c)
1328 struct dm_buffer *b;
1331 BUG_ON(dm_bufio_in_request());
1334 * An optimization so that the buffers are not written one-by-one.
1336 dm_bufio_write_dirty_buffers_async(c);
1340 while ((b = __get_unclaimed_buffer(c)))
1341 __free_buffer_wake(b);
1343 for (i = 0; i < LIST_SIZE; i++)
1344 list_for_each_entry(b, &c->lru[i], lru_list)
1345 DMERR("leaked buffer %llx, hold count %u, list %d",
1346 (unsigned long long)b->block, b->hold_count, i);
1348 for (i = 0; i < LIST_SIZE; i++)
1349 BUG_ON(!list_empty(&c->lru[i]));
1355 * Test if the buffer is unused and too old, and commit it.
1356 * At if noio is set, we must not do any I/O because we hold
1357 * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1358 * different bufio client.
1360 static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1361 unsigned long max_jiffies)
1363 if (jiffies - b->last_accessed < max_jiffies)
1366 if (!(gfp & __GFP_IO)) {
1367 if (test_bit(B_READING, &b->state) ||
1368 test_bit(B_WRITING, &b->state) ||
1369 test_bit(B_DIRTY, &b->state))
1376 __make_buffer_clean(b);
1378 __free_buffer_wake(b);
1383 static void __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1384 struct shrink_control *sc)
1387 struct dm_buffer *b, *tmp;
1389 for (l = 0; l < LIST_SIZE; l++) {
1390 list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list)
1391 if (!__cleanup_old_buffer(b, sc->gfp_mask, 0) &&
1394 dm_bufio_cond_resched();
1398 static int shrink(struct shrinker *shrinker, struct shrink_control *sc)
1400 struct dm_bufio_client *c =
1401 container_of(shrinker, struct dm_bufio_client, shrinker);
1403 unsigned long nr_to_scan = sc->nr_to_scan;
1405 if (sc->gfp_mask & __GFP_IO)
1407 else if (!dm_bufio_trylock(c))
1408 return !nr_to_scan ? 0 : -1;
1411 __scan(c, nr_to_scan, sc);
1413 r = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1423 * Create the buffering interface
1425 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1426 unsigned reserved_buffers, unsigned aux_size,
1427 void (*alloc_callback)(struct dm_buffer *),
1428 void (*write_callback)(struct dm_buffer *))
1431 struct dm_bufio_client *c;
1434 BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1435 (block_size & (block_size - 1)));
1437 c = kmalloc(sizeof(*c), GFP_KERNEL);
1442 c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1443 if (!c->cache_hash) {
1449 c->block_size = block_size;
1450 c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1451 c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1452 ffs(block_size) - 1 - PAGE_SHIFT : 0;
1453 c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1454 PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1456 c->aux_size = aux_size;
1457 c->alloc_callback = alloc_callback;
1458 c->write_callback = write_callback;
1460 for (i = 0; i < LIST_SIZE; i++) {
1461 INIT_LIST_HEAD(&c->lru[i]);
1462 c->n_buffers[i] = 0;
1465 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1466 INIT_HLIST_HEAD(&c->cache_hash[i]);
1468 mutex_init(&c->lock);
1469 INIT_LIST_HEAD(&c->reserved_buffers);
1470 c->need_reserved_buffers = reserved_buffers;
1472 init_waitqueue_head(&c->free_buffer_wait);
1473 c->async_write_error = 0;
1475 c->dm_io = dm_io_client_create();
1476 if (IS_ERR(c->dm_io)) {
1477 r = PTR_ERR(c->dm_io);
1481 mutex_lock(&dm_bufio_clients_lock);
1482 if (c->blocks_per_page_bits) {
1483 if (!DM_BUFIO_CACHE_NAME(c)) {
1484 DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1485 if (!DM_BUFIO_CACHE_NAME(c)) {
1487 mutex_unlock(&dm_bufio_clients_lock);
1492 if (!DM_BUFIO_CACHE(c)) {
1493 DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1495 c->block_size, 0, NULL);
1496 if (!DM_BUFIO_CACHE(c)) {
1498 mutex_unlock(&dm_bufio_clients_lock);
1503 mutex_unlock(&dm_bufio_clients_lock);
1505 while (c->need_reserved_buffers) {
1506 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1512 __free_buffer_wake(b);
1515 mutex_lock(&dm_bufio_clients_lock);
1516 dm_bufio_client_count++;
1517 list_add(&c->client_list, &dm_bufio_all_clients);
1518 __cache_size_refresh();
1519 mutex_unlock(&dm_bufio_clients_lock);
1521 c->shrinker.shrink = shrink;
1522 c->shrinker.seeks = 1;
1523 c->shrinker.batch = 0;
1524 register_shrinker(&c->shrinker);
1530 while (!list_empty(&c->reserved_buffers)) {
1531 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1532 struct dm_buffer, lru_list);
1533 list_del(&b->lru_list);
1536 dm_io_client_destroy(c->dm_io);
1538 vfree(c->cache_hash);
1544 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1547 * Free the buffering interface.
1548 * It is required that there are no references on any buffers.
1550 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1556 unregister_shrinker(&c->shrinker);
1558 mutex_lock(&dm_bufio_clients_lock);
1560 list_del(&c->client_list);
1561 dm_bufio_client_count--;
1562 __cache_size_refresh();
1564 mutex_unlock(&dm_bufio_clients_lock);
1566 for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1567 BUG_ON(!hlist_empty(&c->cache_hash[i]));
1569 BUG_ON(c->need_reserved_buffers);
1571 while (!list_empty(&c->reserved_buffers)) {
1572 struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1573 struct dm_buffer, lru_list);
1574 list_del(&b->lru_list);
1578 for (i = 0; i < LIST_SIZE; i++)
1579 if (c->n_buffers[i])
1580 DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1582 for (i = 0; i < LIST_SIZE; i++)
1583 BUG_ON(c->n_buffers[i]);
1585 dm_io_client_destroy(c->dm_io);
1586 vfree(c->cache_hash);
1589 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1591 static void cleanup_old_buffers(void)
1593 unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1594 struct dm_bufio_client *c;
1596 if (max_age > ULONG_MAX / HZ)
1597 max_age = ULONG_MAX / HZ;
1599 mutex_lock(&dm_bufio_clients_lock);
1600 list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1601 if (!dm_bufio_trylock(c))
1604 while (!list_empty(&c->lru[LIST_CLEAN])) {
1605 struct dm_buffer *b;
1606 b = list_entry(c->lru[LIST_CLEAN].prev,
1607 struct dm_buffer, lru_list);
1608 if (__cleanup_old_buffer(b, 0, max_age * HZ))
1610 dm_bufio_cond_resched();
1614 dm_bufio_cond_resched();
1616 mutex_unlock(&dm_bufio_clients_lock);
1619 static struct workqueue_struct *dm_bufio_wq;
1620 static struct delayed_work dm_bufio_work;
1622 static void work_fn(struct work_struct *w)
1624 cleanup_old_buffers();
1626 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1627 DM_BUFIO_WORK_TIMER_SECS * HZ);
1630 /*----------------------------------------------------------------
1632 *--------------------------------------------------------------*/
1635 * This is called only once for the whole dm_bufio module.
1636 * It initializes memory limit.
1638 static int __init dm_bufio_init(void)
1642 memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1643 memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1645 mem = (__u64)((totalram_pages - totalhigh_pages) *
1646 DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1648 if (mem > ULONG_MAX)
1653 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1654 * in fs/proc/internal.h
1656 if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1657 mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1660 dm_bufio_default_cache_size = mem;
1662 mutex_lock(&dm_bufio_clients_lock);
1663 __cache_size_refresh();
1664 mutex_unlock(&dm_bufio_clients_lock);
1666 dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1670 INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1671 queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1672 DM_BUFIO_WORK_TIMER_SECS * HZ);
1678 * This is called once when unloading the dm_bufio module.
1680 static void __exit dm_bufio_exit(void)
1685 cancel_delayed_work_sync(&dm_bufio_work);
1686 destroy_workqueue(dm_bufio_wq);
1688 for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1689 struct kmem_cache *kc = dm_bufio_caches[i];
1692 kmem_cache_destroy(kc);
1695 for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1696 kfree(dm_bufio_cache_names[i]);
1698 if (dm_bufio_client_count) {
1699 DMCRIT("%s: dm_bufio_client_count leaked: %d",
1700 __func__, dm_bufio_client_count);
1704 if (dm_bufio_current_allocated) {
1705 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1706 __func__, dm_bufio_current_allocated);
1710 if (dm_bufio_allocated_get_free_pages) {
1711 DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1712 __func__, dm_bufio_allocated_get_free_pages);
1716 if (dm_bufio_allocated_vmalloc) {
1717 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1718 __func__, dm_bufio_allocated_vmalloc);
1726 module_init(dm_bufio_init)
1727 module_exit(dm_bufio_exit)
1729 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1730 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1732 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1733 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1735 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1736 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1738 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1739 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1741 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1742 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1744 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1745 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1747 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1748 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1750 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1751 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1752 MODULE_LICENSE("GPL");