Loading...
Note: File does not exist in v3.1.
1/*
2 * Copyright (C) 2009-2011 Red Hat, Inc.
3 *
4 * Author: Mikulas Patocka <mpatocka@redhat.com>
5 *
6 * This file is released under the GPL.
7 */
8
9#include <linux/dm-bufio.h>
10
11#include <linux/device-mapper.h>
12#include <linux/dm-io.h>
13#include <linux/slab.h>
14#include <linux/sched/mm.h>
15#include <linux/jiffies.h>
16#include <linux/vmalloc.h>
17#include <linux/shrinker.h>
18#include <linux/module.h>
19#include <linux/rbtree.h>
20#include <linux/stacktrace.h>
21
22#define DM_MSG_PREFIX "bufio"
23
24/*
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
30 * dirty buffers.
31 */
32#define DM_BUFIO_MIN_BUFFERS 8
33
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
38
39/*
40 * Check buffer ages in this interval (seconds)
41 */
42#define DM_BUFIO_WORK_TIMER_SECS 30
43
44/*
45 * Free buffers when they are older than this (seconds)
46 */
47#define DM_BUFIO_DEFAULT_AGE_SECS 300
48
49/*
50 * The nr of bytes of cached data to keep around.
51 */
52#define DM_BUFIO_DEFAULT_RETAIN_BYTES (256 * 1024)
53
54/*
55 * Align buffer writes to this boundary.
56 * Tests show that SSDs have the highest IOPS when using 4k writes.
57 */
58#define DM_BUFIO_WRITE_ALIGN 4096
59
60/*
61 * dm_buffer->list_mode
62 */
63#define LIST_CLEAN 0
64#define LIST_DIRTY 1
65#define LIST_SIZE 2
66
67/*
68 * Linking of buffers:
69 * All buffers are linked to buffer_tree with their node field.
70 *
71 * Clean buffers that are not being written (B_WRITING not set)
72 * are linked to lru[LIST_CLEAN] with their lru_list field.
73 *
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
80 * context.
81 */
82struct dm_bufio_client {
83 struct mutex lock;
84
85 struct list_head lru[LIST_SIZE];
86 unsigned long n_buffers[LIST_SIZE];
87
88 struct block_device *bdev;
89 unsigned block_size;
90 s8 sectors_per_block_bits;
91 void (*alloc_callback)(struct dm_buffer *);
92 void (*write_callback)(struct dm_buffer *);
93
94 struct kmem_cache *slab_buffer;
95 struct kmem_cache *slab_cache;
96 struct dm_io_client *dm_io;
97
98 struct list_head reserved_buffers;
99 unsigned need_reserved_buffers;
100
101 unsigned minimum_buffers;
102
103 struct rb_root buffer_tree;
104 wait_queue_head_t free_buffer_wait;
105
106 sector_t start;
107
108 int async_write_error;
109
110 struct list_head client_list;
111
112 struct shrinker shrinker;
113 struct work_struct shrink_work;
114 atomic_long_t need_shrink;
115};
116
117/*
118 * Buffer state bits.
119 */
120#define B_READING 0
121#define B_WRITING 1
122#define B_DIRTY 2
123
124/*
125 * Describes how the block was allocated:
126 * kmem_cache_alloc(), __get_free_pages() or vmalloc().
127 * See the comment at alloc_buffer_data.
128 */
129enum data_mode {
130 DATA_MODE_SLAB = 0,
131 DATA_MODE_GET_FREE_PAGES = 1,
132 DATA_MODE_VMALLOC = 2,
133 DATA_MODE_LIMIT = 3
134};
135
136struct dm_buffer {
137 struct rb_node node;
138 struct list_head lru_list;
139 struct list_head global_list;
140 sector_t block;
141 void *data;
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;
146 unsigned accessed;
147 unsigned hold_count;
148 unsigned long state;
149 unsigned long last_accessed;
150 unsigned dirty_start;
151 unsigned dirty_end;
152 unsigned write_start;
153 unsigned write_end;
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
158#define MAX_STACK 10
159 unsigned int stack_len;
160 unsigned long stack_entries[MAX_STACK];
161#endif
162};
163
164/*----------------------------------------------------------------*/
165
166#define dm_bufio_in_request() (!!current->bio_list)
167
168static void dm_bufio_lock(struct dm_bufio_client *c)
169{
170 mutex_lock_nested(&c->lock, dm_bufio_in_request());
171}
172
173static int dm_bufio_trylock(struct dm_bufio_client *c)
174{
175 return mutex_trylock(&c->lock);
176}
177
178static void dm_bufio_unlock(struct dm_bufio_client *c)
179{
180 mutex_unlock(&c->lock);
181}
182
183/*----------------------------------------------------------------*/
184
185/*
186 * Default cache size: available memory divided by the ratio.
187 */
188static unsigned long dm_bufio_default_cache_size;
189
190/*
191 * Total cache size set by the user.
192 */
193static unsigned long dm_bufio_cache_size;
194
195/*
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.
198 */
199static unsigned long dm_bufio_cache_size_latch;
200
201static DEFINE_SPINLOCK(global_spinlock);
202
203static LIST_HEAD(global_queue);
204
205static unsigned long global_num = 0;
206
207/*
208 * Buffers are freed after this timeout
209 */
210static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
211static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
212
213static unsigned long dm_bufio_peak_allocated;
214static unsigned long dm_bufio_allocated_kmem_cache;
215static unsigned long dm_bufio_allocated_get_free_pages;
216static unsigned long dm_bufio_allocated_vmalloc;
217static unsigned long dm_bufio_current_allocated;
218
219/*----------------------------------------------------------------*/
220
221/*
222 * The current number of clients.
223 */
224static int dm_bufio_client_count;
225
226/*
227 * The list of all clients.
228 */
229static LIST_HEAD(dm_bufio_all_clients);
230
231/*
232 * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
233 */
234static DEFINE_MUTEX(dm_bufio_clients_lock);
235
236static struct workqueue_struct *dm_bufio_wq;
237static struct delayed_work dm_bufio_cleanup_old_work;
238static struct work_struct dm_bufio_replacement_work;
239
240
241#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
242static void buffer_record_stack(struct dm_buffer *b)
243{
244 b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
245}
246#endif
247
248/*----------------------------------------------------------------
249 * A red/black tree acts as an index for all the buffers.
250 *--------------------------------------------------------------*/
251static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
252{
253 struct rb_node *n = c->buffer_tree.rb_node;
254 struct dm_buffer *b;
255
256 while (n) {
257 b = container_of(n, struct dm_buffer, node);
258
259 if (b->block == block)
260 return b;
261
262 n = block < b->block ? n->rb_left : n->rb_right;
263 }
264
265 return NULL;
266}
267
268static struct dm_buffer *__find_next(struct dm_bufio_client *c, sector_t block)
269{
270 struct rb_node *n = c->buffer_tree.rb_node;
271 struct dm_buffer *b;
272 struct dm_buffer *best = NULL;
273
274 while (n) {
275 b = container_of(n, struct dm_buffer, node);
276
277 if (b->block == block)
278 return b;
279
280 if (block <= b->block) {
281 n = n->rb_left;
282 best = b;
283 } else {
284 n = n->rb_right;
285 }
286 }
287
288 return best;
289}
290
291static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
292{
293 struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
294 struct dm_buffer *found;
295
296 while (*new) {
297 found = container_of(*new, struct dm_buffer, node);
298
299 if (found->block == b->block) {
300 BUG_ON(found != b);
301 return;
302 }
303
304 parent = *new;
305 new = b->block < found->block ?
306 &found->node.rb_left : &found->node.rb_right;
307 }
308
309 rb_link_node(&b->node, parent, new);
310 rb_insert_color(&b->node, &c->buffer_tree);
311}
312
313static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
314{
315 rb_erase(&b->node, &c->buffer_tree);
316}
317
318/*----------------------------------------------------------------*/
319
320static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
321{
322 unsigned char data_mode;
323 long diff;
324
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,
329 };
330
331 data_mode = b->data_mode;
332 diff = (long)b->c->block_size;
333 if (unlink)
334 diff = -diff;
335
336 spin_lock(&global_spinlock);
337
338 *class_ptr[data_mode] += diff;
339
340 dm_bufio_current_allocated += diff;
341
342 if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
343 dm_bufio_peak_allocated = dm_bufio_current_allocated;
344
345 b->accessed = 1;
346
347 if (!unlink) {
348 list_add(&b->global_list, &global_queue);
349 global_num++;
350 if (dm_bufio_current_allocated > dm_bufio_cache_size)
351 queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
352 } else {
353 list_del(&b->global_list);
354 global_num--;
355 }
356
357 spin_unlock(&global_spinlock);
358}
359
360/*
361 * Change the number of clients and recalculate per-client limit.
362 */
363static void __cache_size_refresh(void)
364{
365 BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
366 BUG_ON(dm_bufio_client_count < 0);
367
368 dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
369
370 /*
371 * Use default if set to 0 and report the actual cache size used.
372 */
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;
377 }
378}
379
380/*
381 * Allocating buffer data.
382 *
383 * Small buffers are allocated with kmem_cache, to use space optimally.
384 *
385 * For large buffers, we choose between get_free_pages and vmalloc.
386 * Each has advantages and disadvantages.
387 *
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.
391 *
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).
396 *
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
399 * space.
400 */
401static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
402 unsigned char *data_mode)
403{
404 if (unlikely(c->slab_cache != NULL)) {
405 *data_mode = DATA_MODE_SLAB;
406 return kmem_cache_alloc(c->slab_cache, gfp_mask);
407 }
408
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));
414 }
415
416 *data_mode = DATA_MODE_VMALLOC;
417
418 /*
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.
422 *
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.
426 */
427 if (gfp_mask & __GFP_NORETRY) {
428 unsigned noio_flag = memalloc_noio_save();
429 void *ptr = __vmalloc(c->block_size, gfp_mask);
430
431 memalloc_noio_restore(noio_flag);
432 return ptr;
433 }
434
435 return __vmalloc(c->block_size, gfp_mask);
436}
437
438/*
439 * Free buffer's data.
440 */
441static void free_buffer_data(struct dm_bufio_client *c,
442 void *data, unsigned char data_mode)
443{
444 switch (data_mode) {
445 case DATA_MODE_SLAB:
446 kmem_cache_free(c->slab_cache, data);
447 break;
448
449 case DATA_MODE_GET_FREE_PAGES:
450 free_pages((unsigned long)data,
451 c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
452 break;
453
454 case DATA_MODE_VMALLOC:
455 vfree(data);
456 break;
457
458 default:
459 DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
460 data_mode);
461 BUG();
462 }
463}
464
465/*
466 * Allocate buffer and its data.
467 */
468static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
469{
470 struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
471
472 if (!b)
473 return NULL;
474
475 b->c = c;
476
477 b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
478 if (!b->data) {
479 kmem_cache_free(c->slab_buffer, b);
480 return NULL;
481 }
482
483#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
484 b->stack_len = 0;
485#endif
486 return b;
487}
488
489/*
490 * Free buffer and its data.
491 */
492static void free_buffer(struct dm_buffer *b)
493{
494 struct dm_bufio_client *c = b->c;
495
496 free_buffer_data(c, b->data, b->data_mode);
497 kmem_cache_free(c->slab_buffer, b);
498}
499
500/*
501 * Link buffer to the buffer tree and clean or dirty queue.
502 */
503static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
504{
505 struct dm_bufio_client *c = b->c;
506
507 c->n_buffers[dirty]++;
508 b->block = block;
509 b->list_mode = dirty;
510 list_add(&b->lru_list, &c->lru[dirty]);
511 __insert(b->c, b);
512 b->last_accessed = jiffies;
513
514 adjust_total_allocated(b, false);
515}
516
517/*
518 * Unlink buffer from the buffer tree and dirty or clean queue.
519 */
520static void __unlink_buffer(struct dm_buffer *b)
521{
522 struct dm_bufio_client *c = b->c;
523
524 BUG_ON(!c->n_buffers[b->list_mode]);
525
526 c->n_buffers[b->list_mode]--;
527 __remove(b->c, b);
528 list_del(&b->lru_list);
529
530 adjust_total_allocated(b, true);
531}
532
533/*
534 * Place the buffer to the head of dirty or clean LRU queue.
535 */
536static void __relink_lru(struct dm_buffer *b, int dirty)
537{
538 struct dm_bufio_client *c = b->c;
539
540 b->accessed = 1;
541
542 BUG_ON(!c->n_buffers[b->list_mode]);
543
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;
549}
550
551/*----------------------------------------------------------------
552 * Submit I/O on the buffer.
553 *
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);
557 *
558 * the memory must be direct-mapped, not vmalloced;
559 *
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.
562 *
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
566 * shortcomings.
567 *--------------------------------------------------------------*/
568
569/*
570 * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
571 * that the request was handled directly with bio interface.
572 */
573static void dmio_complete(unsigned long error, void *context)
574{
575 struct dm_buffer *b = context;
576
577 b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
578}
579
580static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
581 unsigned n_sectors, unsigned offset)
582{
583 int r;
584 struct dm_io_request io_req = {
585 .bi_op = rw,
586 .bi_op_flags = 0,
587 .notify.fn = dmio_complete,
588 .notify.context = b,
589 .client = b->c->dm_io,
590 };
591 struct dm_io_region region = {
592 .bdev = b->c->bdev,
593 .sector = sector,
594 .count = n_sectors,
595 };
596
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;
600 } else {
601 io_req.mem.type = DM_IO_VMA;
602 io_req.mem.ptr.vma = (char *)b->data + offset;
603 }
604
605 r = dm_io(&io_req, 1, ®ion, NULL);
606 if (unlikely(r))
607 b->end_io(b, errno_to_blk_status(r));
608}
609
610static void bio_complete(struct bio *bio)
611{
612 struct dm_buffer *b = bio->bi_private;
613 blk_status_t status = bio->bi_status;
614 bio_put(bio);
615 b->end_io(b, status);
616}
617
618static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
619 unsigned n_sectors, unsigned offset)
620{
621 struct bio *bio;
622 char *ptr;
623 unsigned vec_size, len;
624
625 vec_size = b->c->block_size >> PAGE_SHIFT;
626 if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
627 vec_size += 2;
628
629 bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
630 if (!bio) {
631dmio:
632 use_dmio(b, rw, sector, n_sectors, offset);
633 return;
634 }
635
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;
640 bio->bi_private = b;
641
642 ptr = (char *)b->data + offset;
643 len = n_sectors << SECTOR_SHIFT;
644
645 do {
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))) {
649 bio_put(bio);
650 goto dmio;
651 }
652
653 len -= this_step;
654 ptr += this_step;
655 } while (len > 0);
656
657 submit_bio(bio);
658}
659
660static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
661{
662 sector_t sector;
663
664 if (likely(c->sectors_per_block_bits >= 0))
665 sector = block << c->sectors_per_block_bits;
666 else
667 sector = block * (c->block_size >> SECTOR_SHIFT);
668 sector += c->start;
669
670 return sector;
671}
672
673static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
674{
675 unsigned n_sectors;
676 sector_t sector;
677 unsigned offset, end;
678
679 b->end_io = end_io;
680
681 sector = block_to_sector(b->c, b->block);
682
683 if (rw != REQ_OP_WRITE) {
684 n_sectors = b->c->block_size >> SECTOR_SHIFT;
685 offset = 0;
686 } else {
687 if (b->c->write_callback)
688 b->c->write_callback(b);
689 offset = b->write_start;
690 end = b->write_end;
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;
696
697 sector += offset >> SECTOR_SHIFT;
698 n_sectors = (end - offset) >> SECTOR_SHIFT;
699 }
700
701 if (b->data_mode != DATA_MODE_VMALLOC)
702 use_bio(b, rw, sector, n_sectors, offset);
703 else
704 use_dmio(b, rw, sector, n_sectors, offset);
705}
706
707/*----------------------------------------------------------------
708 * Writing dirty buffers
709 *--------------------------------------------------------------*/
710
711/*
712 * The endio routine for write.
713 *
714 * Set the error, clear B_WRITING bit and wake anyone who was waiting on
715 * it.
716 */
717static void write_endio(struct dm_buffer *b, blk_status_t status)
718{
719 b->write_error = status;
720 if (unlikely(status)) {
721 struct dm_bufio_client *c = b->c;
722
723 (void)cmpxchg(&c->async_write_error, 0,
724 blk_status_to_errno(status));
725 }
726
727 BUG_ON(!test_bit(B_WRITING, &b->state));
728
729 smp_mb__before_atomic();
730 clear_bit(B_WRITING, &b->state);
731 smp_mb__after_atomic();
732
733 wake_up_bit(&b->state, B_WRITING);
734}
735
736/*
737 * Initiate a write on a dirty buffer, but don't wait for it.
738 *
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.
744 */
745static void __write_dirty_buffer(struct dm_buffer *b,
746 struct list_head *write_list)
747{
748 if (!test_bit(B_DIRTY, &b->state))
749 return;
750
751 clear_bit(B_DIRTY, &b->state);
752 wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
753
754 b->write_start = b->dirty_start;
755 b->write_end = b->dirty_end;
756
757 if (!write_list)
758 submit_io(b, REQ_OP_WRITE, write_endio);
759 else
760 list_add_tail(&b->write_list, write_list);
761}
762
763static void __flush_write_list(struct list_head *write_list)
764{
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);
772 cond_resched();
773 }
774 blk_finish_plug(&plug);
775}
776
777/*
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.
781 */
782static void __make_buffer_clean(struct dm_buffer *b)
783{
784 BUG_ON(b->hold_count);
785
786 if (!b->state) /* fast case */
787 return;
788
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);
792}
793
794/*
795 * Find some buffer that is not held by anybody, clean it, unlink it and
796 * return it.
797 */
798static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
799{
800 struct dm_buffer *b;
801
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));
805
806 if (!b->hold_count) {
807 __make_buffer_clean(b);
808 __unlink_buffer(b);
809 return b;
810 }
811 cond_resched();
812 }
813
814 list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
815 BUG_ON(test_bit(B_READING, &b->state));
816
817 if (!b->hold_count) {
818 __make_buffer_clean(b);
819 __unlink_buffer(b);
820 return b;
821 }
822 cond_resched();
823 }
824
825 return NULL;
826}
827
828/*
829 * Wait until some other threads free some buffer or release hold count on
830 * some buffer.
831 *
832 * This function is entered with c->lock held, drops it and regains it
833 * before exiting.
834 */
835static void __wait_for_free_buffer(struct dm_bufio_client *c)
836{
837 DECLARE_WAITQUEUE(wait, current);
838
839 add_wait_queue(&c->free_buffer_wait, &wait);
840 set_current_state(TASK_UNINTERRUPTIBLE);
841 dm_bufio_unlock(c);
842
843 io_schedule();
844
845 remove_wait_queue(&c->free_buffer_wait, &wait);
846
847 dm_bufio_lock(c);
848}
849
850enum new_flag {
851 NF_FRESH = 0,
852 NF_READ = 1,
853 NF_GET = 2,
854 NF_PREFETCH = 3
855};
856
857/*
858 * Allocate a new buffer. If the allocation is not possible, wait until
859 * some other thread frees a buffer.
860 *
861 * May drop the lock and regain it.
862 */
863static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
864{
865 struct dm_buffer *b;
866 bool tried_noio_alloc = false;
867
868 /*
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
877 *
878 * For debugging, if we set the cache size to 1, no new buffers will
879 * be allocated.
880 */
881 while (1) {
882 if (dm_bufio_cache_size_latch != 1) {
883 b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
884 if (b)
885 return b;
886 }
887
888 if (nf == NF_PREFETCH)
889 return NULL;
890
891 if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
892 dm_bufio_unlock(c);
893 b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
894 dm_bufio_lock(c);
895 if (b)
896 return b;
897 tried_noio_alloc = true;
898 }
899
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++;
905
906 return b;
907 }
908
909 b = __get_unclaimed_buffer(c);
910 if (b)
911 return b;
912
913 __wait_for_free_buffer(c);
914 }
915}
916
917static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
918{
919 struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
920
921 if (!b)
922 return NULL;
923
924 if (c->alloc_callback)
925 c->alloc_callback(b);
926
927 return b;
928}
929
930/*
931 * Free a buffer and wake other threads waiting for free buffers.
932 */
933static void __free_buffer_wake(struct dm_buffer *b)
934{
935 struct dm_bufio_client *c = b->c;
936
937 if (!c->need_reserved_buffers)
938 free_buffer(b);
939 else {
940 list_add(&b->lru_list, &c->reserved_buffers);
941 c->need_reserved_buffers--;
942 }
943
944 wake_up(&c->free_buffer_wait);
945}
946
947static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
948 struct list_head *write_list)
949{
950 struct dm_buffer *b, *tmp;
951
952 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
953 BUG_ON(test_bit(B_READING, &b->state));
954
955 if (!test_bit(B_DIRTY, &b->state) &&
956 !test_bit(B_WRITING, &b->state)) {
957 __relink_lru(b, LIST_CLEAN);
958 continue;
959 }
960
961 if (no_wait && test_bit(B_WRITING, &b->state))
962 return;
963
964 __write_dirty_buffer(b, write_list);
965 cond_resched();
966 }
967}
968
969/*
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.
973 */
974static void __check_watermark(struct dm_bufio_client *c,
975 struct list_head *write_list)
976{
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);
979}
980
981/*----------------------------------------------------------------
982 * Getting a buffer
983 *--------------------------------------------------------------*/
984
985static 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)
988{
989 struct dm_buffer *b, *new_b = NULL;
990
991 *need_submit = 0;
992
993 b = __find(c, block);
994 if (b)
995 goto found_buffer;
996
997 if (nf == NF_GET)
998 return NULL;
999
1000 new_b = __alloc_buffer_wait(c, nf);
1001 if (!new_b)
1002 return NULL;
1003
1004 /*
1005 * We've had a period where the mutex was unlocked, so need to
1006 * recheck the buffer tree.
1007 */
1008 b = __find(c, block);
1009 if (b) {
1010 __free_buffer_wake(new_b);
1011 goto found_buffer;
1012 }
1013
1014 __check_watermark(c, write_list);
1015
1016 b = new_b;
1017 b->hold_count = 1;
1018 b->read_error = 0;
1019 b->write_error = 0;
1020 __link_buffer(b, block, LIST_CLEAN);
1021
1022 if (nf == NF_FRESH) {
1023 b->state = 0;
1024 return b;
1025 }
1026
1027 b->state = 1 << B_READING;
1028 *need_submit = 1;
1029
1030 return b;
1031
1032found_buffer:
1033 if (nf == NF_PREFETCH)
1034 return NULL;
1035 /*
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.
1041 */
1042 if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1043 return NULL;
1044
1045 b->hold_count++;
1046 __relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1047 test_bit(B_WRITING, &b->state));
1048 return b;
1049}
1050
1051/*
1052 * The endio routine for reading: set the error, clear the bit and wake up
1053 * anyone waiting on the buffer.
1054 */
1055static void read_endio(struct dm_buffer *b, blk_status_t status)
1056{
1057 b->read_error = status;
1058
1059 BUG_ON(!test_bit(B_READING, &b->state));
1060
1061 smp_mb__before_atomic();
1062 clear_bit(B_READING, &b->state);
1063 smp_mb__after_atomic();
1064
1065 wake_up_bit(&b->state, B_READING);
1066}
1067
1068/*
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).
1073 */
1074static void *new_read(struct dm_bufio_client *c, sector_t block,
1075 enum new_flag nf, struct dm_buffer **bp)
1076{
1077 int need_submit;
1078 struct dm_buffer *b;
1079
1080 LIST_HEAD(write_list);
1081
1082 dm_bufio_lock(c);
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);
1087#endif
1088 dm_bufio_unlock(c);
1089
1090 __flush_write_list(&write_list);
1091
1092 if (!b)
1093 return NULL;
1094
1095 if (need_submit)
1096 submit_io(b, REQ_OP_READ, read_endio);
1097
1098 wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1099
1100 if (b->read_error) {
1101 int error = blk_status_to_errno(b->read_error);
1102
1103 dm_bufio_release(b);
1104
1105 return ERR_PTR(error);
1106 }
1107
1108 *bp = b;
1109
1110 return b->data;
1111}
1112
1113void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1114 struct dm_buffer **bp)
1115{
1116 return new_read(c, block, NF_GET, bp);
1117}
1118EXPORT_SYMBOL_GPL(dm_bufio_get);
1119
1120void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1121 struct dm_buffer **bp)
1122{
1123 BUG_ON(dm_bufio_in_request());
1124
1125 return new_read(c, block, NF_READ, bp);
1126}
1127EXPORT_SYMBOL_GPL(dm_bufio_read);
1128
1129void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1130 struct dm_buffer **bp)
1131{
1132 BUG_ON(dm_bufio_in_request());
1133
1134 return new_read(c, block, NF_FRESH, bp);
1135}
1136EXPORT_SYMBOL_GPL(dm_bufio_new);
1137
1138void dm_bufio_prefetch(struct dm_bufio_client *c,
1139 sector_t block, unsigned n_blocks)
1140{
1141 struct blk_plug plug;
1142
1143 LIST_HEAD(write_list);
1144
1145 BUG_ON(dm_bufio_in_request());
1146
1147 blk_start_plug(&plug);
1148 dm_bufio_lock(c);
1149
1150 for (; n_blocks--; block++) {
1151 int need_submit;
1152 struct dm_buffer *b;
1153 b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1154 &write_list);
1155 if (unlikely(!list_empty(&write_list))) {
1156 dm_bufio_unlock(c);
1157 blk_finish_plug(&plug);
1158 __flush_write_list(&write_list);
1159 blk_start_plug(&plug);
1160 dm_bufio_lock(c);
1161 }
1162 if (unlikely(b != NULL)) {
1163 dm_bufio_unlock(c);
1164
1165 if (need_submit)
1166 submit_io(b, REQ_OP_READ, read_endio);
1167 dm_bufio_release(b);
1168
1169 cond_resched();
1170
1171 if (!n_blocks)
1172 goto flush_plug;
1173 dm_bufio_lock(c);
1174 }
1175 }
1176
1177 dm_bufio_unlock(c);
1178
1179flush_plug:
1180 blk_finish_plug(&plug);
1181}
1182EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1183
1184void dm_bufio_release(struct dm_buffer *b)
1185{
1186 struct dm_bufio_client *c = b->c;
1187
1188 dm_bufio_lock(c);
1189
1190 BUG_ON(!b->hold_count);
1191
1192 b->hold_count--;
1193 if (!b->hold_count) {
1194 wake_up(&c->free_buffer_wait);
1195
1196 /*
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
1199 * invalid buffer.
1200 */
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)) {
1205 __unlink_buffer(b);
1206 __free_buffer_wake(b);
1207 }
1208 }
1209
1210 dm_bufio_unlock(c);
1211}
1212EXPORT_SYMBOL_GPL(dm_bufio_release);
1213
1214void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1215 unsigned start, unsigned end)
1216{
1217 struct dm_bufio_client *c = b->c;
1218
1219 BUG_ON(start >= end);
1220 BUG_ON(end > b->c->block_size);
1221
1222 dm_bufio_lock(c);
1223
1224 BUG_ON(test_bit(B_READING, &b->state));
1225
1226 if (!test_and_set_bit(B_DIRTY, &b->state)) {
1227 b->dirty_start = start;
1228 b->dirty_end = end;
1229 __relink_lru(b, LIST_DIRTY);
1230 } else {
1231 if (start < b->dirty_start)
1232 b->dirty_start = start;
1233 if (end > b->dirty_end)
1234 b->dirty_end = end;
1235 }
1236
1237 dm_bufio_unlock(c);
1238}
1239EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1240
1241void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1242{
1243 dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1244}
1245EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1246
1247void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1248{
1249 LIST_HEAD(write_list);
1250
1251 BUG_ON(dm_bufio_in_request());
1252
1253 dm_bufio_lock(c);
1254 __write_dirty_buffers_async(c, 0, &write_list);
1255 dm_bufio_unlock(c);
1256 __flush_write_list(&write_list);
1257}
1258EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1259
1260/*
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
1263 * waited upon.
1264 *
1265 * Finally, we flush hardware disk cache.
1266 */
1267int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1268{
1269 int a, f;
1270 unsigned long buffers_processed = 0;
1271 struct dm_buffer *b, *tmp;
1272
1273 LIST_HEAD(write_list);
1274
1275 dm_bufio_lock(c);
1276 __write_dirty_buffers_async(c, 0, &write_list);
1277 dm_bufio_unlock(c);
1278 __flush_write_list(&write_list);
1279 dm_bufio_lock(c);
1280
1281again:
1282 list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1283 int dropped_lock = 0;
1284
1285 if (buffers_processed < c->n_buffers[LIST_DIRTY])
1286 buffers_processed++;
1287
1288 BUG_ON(test_bit(B_READING, &b->state));
1289
1290 if (test_bit(B_WRITING, &b->state)) {
1291 if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1292 dropped_lock = 1;
1293 b->hold_count++;
1294 dm_bufio_unlock(c);
1295 wait_on_bit_io(&b->state, B_WRITING,
1296 TASK_UNINTERRUPTIBLE);
1297 dm_bufio_lock(c);
1298 b->hold_count--;
1299 } else
1300 wait_on_bit_io(&b->state, B_WRITING,
1301 TASK_UNINTERRUPTIBLE);
1302 }
1303
1304 if (!test_bit(B_DIRTY, &b->state) &&
1305 !test_bit(B_WRITING, &b->state))
1306 __relink_lru(b, LIST_CLEAN);
1307
1308 cond_resched();
1309
1310 /*
1311 * If we dropped the lock, the list is no longer consistent,
1312 * so we must restart the search.
1313 *
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.
1317 *
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.
1323 */
1324 if (dropped_lock)
1325 goto again;
1326 }
1327 wake_up(&c->free_buffer_wait);
1328 dm_bufio_unlock(c);
1329
1330 a = xchg(&c->async_write_error, 0);
1331 f = dm_bufio_issue_flush(c);
1332 if (a)
1333 return a;
1334
1335 return f;
1336}
1337EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1338
1339/*
1340 * Use dm-io to send an empty barrier to flush the device.
1341 */
1342int dm_bufio_issue_flush(struct dm_bufio_client *c)
1343{
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,
1349 .client = c->dm_io,
1350 };
1351 struct dm_io_region io_reg = {
1352 .bdev = c->bdev,
1353 .sector = 0,
1354 .count = 0,
1355 };
1356
1357 BUG_ON(dm_bufio_in_request());
1358
1359 return dm_io(&io_req, 1, &io_reg, NULL);
1360}
1361EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1362
1363/*
1364 * Use dm-io to send a discard request to flush the device.
1365 */
1366int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
1367{
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,
1373 .client = c->dm_io,
1374 };
1375 struct dm_io_region io_reg = {
1376 .bdev = c->bdev,
1377 .sector = block_to_sector(c, block),
1378 .count = block_to_sector(c, count),
1379 };
1380
1381 BUG_ON(dm_bufio_in_request());
1382
1383 return dm_io(&io_req, 1, &io_reg, NULL);
1384}
1385EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
1386
1387/*
1388 * We first delete any other buffer that may be at that new location.
1389 *
1390 * Then, we write the buffer to the original location if it was dirty.
1391 *
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.
1394 *
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.
1398 */
1399void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1400{
1401 struct dm_bufio_client *c = b->c;
1402 struct dm_buffer *new;
1403
1404 BUG_ON(dm_bufio_in_request());
1405
1406 dm_bufio_lock(c);
1407
1408retry:
1409 new = __find(c, new_block);
1410 if (new) {
1411 if (new->hold_count) {
1412 __wait_for_free_buffer(c);
1413 goto retry;
1414 }
1415
1416 /*
1417 * FIXME: Is there any point waiting for a write that's going
1418 * to be overwritten in a bit?
1419 */
1420 __make_buffer_clean(new);
1421 __unlink_buffer(new);
1422 __free_buffer_wake(new);
1423 }
1424
1425 BUG_ON(!b->hold_count);
1426 BUG_ON(test_bit(B_READING, &b->state));
1427
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);
1433 b->dirty_start = 0;
1434 b->dirty_end = c->block_size;
1435 __unlink_buffer(b);
1436 __link_buffer(b, new_block, LIST_DIRTY);
1437 } else {
1438 sector_t old_block;
1439 wait_on_bit_lock_io(&b->state, B_WRITING,
1440 TASK_UNINTERRUPTIBLE);
1441 /*
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.
1447 */
1448 old_block = b->block;
1449 __unlink_buffer(b);
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);
1454 __unlink_buffer(b);
1455 __link_buffer(b, old_block, b->list_mode);
1456 }
1457
1458 dm_bufio_unlock(c);
1459 dm_bufio_release(b);
1460}
1461EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1462
1463static void forget_buffer_locked(struct dm_buffer *b)
1464{
1465 if (likely(!b->hold_count) && likely(!b->state)) {
1466 __unlink_buffer(b);
1467 __free_buffer_wake(b);
1468 }
1469}
1470
1471/*
1472 * Free the given buffer.
1473 *
1474 * This is just a hint, if the buffer is in use or dirty, this function
1475 * does nothing.
1476 */
1477void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1478{
1479 struct dm_buffer *b;
1480
1481 dm_bufio_lock(c);
1482
1483 b = __find(c, block);
1484 if (b)
1485 forget_buffer_locked(b);
1486
1487 dm_bufio_unlock(c);
1488}
1489EXPORT_SYMBOL_GPL(dm_bufio_forget);
1490
1491void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
1492{
1493 struct dm_buffer *b;
1494 sector_t end_block = block + n_blocks;
1495
1496 while (block < end_block) {
1497 dm_bufio_lock(c);
1498
1499 b = __find_next(c, block);
1500 if (b) {
1501 block = b->block + 1;
1502 forget_buffer_locked(b);
1503 }
1504
1505 dm_bufio_unlock(c);
1506
1507 if (!b)
1508 break;
1509 }
1510
1511}
1512EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
1513
1514void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1515{
1516 c->minimum_buffers = n;
1517}
1518EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1519
1520unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1521{
1522 return c->block_size;
1523}
1524EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1525
1526sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1527{
1528 sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1529 if (s >= c->start)
1530 s -= c->start;
1531 else
1532 s = 0;
1533 if (likely(c->sectors_per_block_bits >= 0))
1534 s >>= c->sectors_per_block_bits;
1535 else
1536 sector_div(s, c->block_size >> SECTOR_SHIFT);
1537 return s;
1538}
1539EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1540
1541struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1542{
1543 return c->dm_io;
1544}
1545EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1546
1547sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1548{
1549 return b->block;
1550}
1551EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1552
1553void *dm_bufio_get_block_data(struct dm_buffer *b)
1554{
1555 return b->data;
1556}
1557EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1558
1559void *dm_bufio_get_aux_data(struct dm_buffer *b)
1560{
1561 return b + 1;
1562}
1563EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1564
1565struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1566{
1567 return b->c;
1568}
1569EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1570
1571static void drop_buffers(struct dm_bufio_client *c)
1572{
1573 struct dm_buffer *b;
1574 int i;
1575 bool warned = false;
1576
1577 BUG_ON(dm_bufio_in_request());
1578
1579 /*
1580 * An optimization so that the buffers are not written one-by-one.
1581 */
1582 dm_bufio_write_dirty_buffers_async(c);
1583
1584 dm_bufio_lock(c);
1585
1586 while ((b = __get_unclaimed_buffer(c)))
1587 __free_buffer_wake(b);
1588
1589 for (i = 0; i < LIST_SIZE; i++)
1590 list_for_each_entry(b, &c->lru[i], lru_list) {
1591 WARN_ON(!warned);
1592 warned = true;
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 */
1598 b->hold_count = 0;
1599#endif
1600 }
1601
1602#ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1603 while ((b = __get_unclaimed_buffer(c)))
1604 __free_buffer_wake(b);
1605#endif
1606
1607 for (i = 0; i < LIST_SIZE; i++)
1608 BUG_ON(!list_empty(&c->lru[i]));
1609
1610 dm_bufio_unlock(c);
1611}
1612
1613/*
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.
1616 *
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.
1620 */
1621static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1622{
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))
1627 return false;
1628 }
1629
1630 if (b->hold_count)
1631 return false;
1632
1633 __make_buffer_clean(b);
1634 __unlink_buffer(b);
1635 __free_buffer_wake(b);
1636
1637 return true;
1638}
1639
1640static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1641{
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;
1645 else
1646 retain_bytes /= c->block_size;
1647 return retain_bytes;
1648}
1649
1650static void __scan(struct dm_bufio_client *c)
1651{
1652 int l;
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);
1658
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))
1664 return;
1665 if (__try_evict_buffer(b, GFP_KERNEL)) {
1666 atomic_long_dec(&c->need_shrink);
1667 freed++;
1668 }
1669 cond_resched();
1670 }
1671 }
1672}
1673
1674static void shrink_work(struct work_struct *w)
1675{
1676 struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
1677
1678 dm_bufio_lock(c);
1679 __scan(c);
1680 dm_bufio_unlock(c);
1681}
1682
1683static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1684{
1685 struct dm_bufio_client *c;
1686
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);
1690
1691 return sc->nr_to_scan;
1692}
1693
1694static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1695{
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);
1701
1702 if (unlikely(count < retain_target))
1703 count = 0;
1704 else
1705 count -= retain_target;
1706
1707 if (unlikely(count < queued_for_cleanup))
1708 count = 0;
1709 else
1710 count -= queued_for_cleanup;
1711
1712 return count;
1713}
1714
1715/*
1716 * Create the buffering interface
1717 */
1718struct 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 *))
1722{
1723 int r;
1724 struct dm_bufio_client *c;
1725 unsigned i;
1726 char slab_name[27];
1727
1728 if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1729 DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1730 r = -EINVAL;
1731 goto bad_client;
1732 }
1733
1734 c = kzalloc(sizeof(*c), GFP_KERNEL);
1735 if (!c) {
1736 r = -ENOMEM;
1737 goto bad_client;
1738 }
1739 c->buffer_tree = RB_ROOT;
1740
1741 c->bdev = bdev;
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;
1745 else
1746 c->sectors_per_block_bits = -1;
1747
1748 c->alloc_callback = alloc_callback;
1749 c->write_callback = write_callback;
1750
1751 for (i = 0; i < LIST_SIZE; i++) {
1752 INIT_LIST_HEAD(&c->lru[i]);
1753 c->n_buffers[i] = 0;
1754 }
1755
1756 mutex_init(&c->lock);
1757 INIT_LIST_HEAD(&c->reserved_buffers);
1758 c->need_reserved_buffers = reserved_buffers;
1759
1760 dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1761
1762 init_waitqueue_head(&c->free_buffer_wait);
1763 c->async_write_error = 0;
1764
1765 c->dm_io = dm_io_client_create();
1766 if (IS_ERR(c->dm_io)) {
1767 r = PTR_ERR(c->dm_io);
1768 goto bad_dm_io;
1769 }
1770
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) {
1778 r = -ENOMEM;
1779 goto bad;
1780 }
1781 }
1782 if (aux_size)
1783 snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1784 else
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) {
1789 r = -ENOMEM;
1790 goto bad;
1791 }
1792
1793 while (c->need_reserved_buffers) {
1794 struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1795
1796 if (!b) {
1797 r = -ENOMEM;
1798 goto bad;
1799 }
1800 __free_buffer_wake(b);
1801 }
1802
1803 INIT_WORK(&c->shrink_work, shrink_work);
1804 atomic_long_set(&c->need_shrink, 0);
1805
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);
1811 if (r)
1812 goto bad;
1813
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);
1819
1820 return c;
1821
1822bad:
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);
1827 free_buffer(b);
1828 }
1829 kmem_cache_destroy(c->slab_cache);
1830 kmem_cache_destroy(c->slab_buffer);
1831 dm_io_client_destroy(c->dm_io);
1832bad_dm_io:
1833 mutex_destroy(&c->lock);
1834 kfree(c);
1835bad_client:
1836 return ERR_PTR(r);
1837}
1838EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1839
1840/*
1841 * Free the buffering interface.
1842 * It is required that there are no references on any buffers.
1843 */
1844void dm_bufio_client_destroy(struct dm_bufio_client *c)
1845{
1846 unsigned i;
1847
1848 drop_buffers(c);
1849
1850 unregister_shrinker(&c->shrinker);
1851 flush_work(&c->shrink_work);
1852
1853 mutex_lock(&dm_bufio_clients_lock);
1854
1855 list_del(&c->client_list);
1856 dm_bufio_client_count--;
1857 __cache_size_refresh();
1858
1859 mutex_unlock(&dm_bufio_clients_lock);
1860
1861 BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1862 BUG_ON(c->need_reserved_buffers);
1863
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);
1868 free_buffer(b);
1869 }
1870
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]);
1874
1875 for (i = 0; i < LIST_SIZE; i++)
1876 BUG_ON(c->n_buffers[i]);
1877
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);
1882 kfree(c);
1883}
1884EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1885
1886void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1887{
1888 c->start = start;
1889}
1890EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1891
1892static unsigned get_max_age_hz(void)
1893{
1894 unsigned max_age = READ_ONCE(dm_bufio_max_age);
1895
1896 if (max_age > UINT_MAX / HZ)
1897 max_age = UINT_MAX / HZ;
1898
1899 return max_age * HZ;
1900}
1901
1902static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1903{
1904 return time_after_eq(jiffies, b->last_accessed + age_hz);
1905}
1906
1907static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1908{
1909 struct dm_buffer *b, *tmp;
1910 unsigned long retain_target = get_retain_buffers(c);
1911 unsigned long count;
1912 LIST_HEAD(write_list);
1913
1914 dm_bufio_lock(c);
1915
1916 __check_watermark(c, &write_list);
1917 if (unlikely(!list_empty(&write_list))) {
1918 dm_bufio_unlock(c);
1919 __flush_write_list(&write_list);
1920 dm_bufio_lock(c);
1921 }
1922
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)
1926 break;
1927
1928 if (!older_than(b, age_hz))
1929 break;
1930
1931 if (__try_evict_buffer(b, 0))
1932 count--;
1933
1934 cond_resched();
1935 }
1936
1937 dm_bufio_unlock(c);
1938}
1939
1940static void do_global_cleanup(struct work_struct *w)
1941{
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;
1949
1950 mutex_lock(&dm_bufio_clients_lock);
1951
1952 while (1) {
1953 cond_resched();
1954
1955 spin_lock(&global_spinlock);
1956 if (unlikely(dm_bufio_current_allocated <= threshold))
1957 break;
1958
1959 spinlock_hold_count = 0;
1960get_next:
1961 if (!loops--)
1962 break;
1963 if (unlikely(list_empty(&global_queue)))
1964 break;
1965 b = list_entry(global_queue.prev, struct dm_buffer, global_list);
1966
1967 if (b->accessed) {
1968 b->accessed = 0;
1969 list_move(&b->global_list, &global_queue);
1970 if (likely(++spinlock_hold_count < 16))
1971 goto get_next;
1972 spin_unlock(&global_spinlock);
1973 continue;
1974 }
1975
1976 current_client = b->c;
1977 if (unlikely(current_client != locked_client)) {
1978 if (locked_client)
1979 dm_bufio_unlock(locked_client);
1980
1981 if (!dm_bufio_trylock(current_client)) {
1982 spin_unlock(&global_spinlock);
1983 dm_bufio_lock(current_client);
1984 locked_client = current_client;
1985 continue;
1986 }
1987
1988 locked_client = current_client;
1989 }
1990
1991 spin_unlock(&global_spinlock);
1992
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);
1997 }
1998 }
1999
2000 spin_unlock(&global_spinlock);
2001
2002 if (locked_client)
2003 dm_bufio_unlock(locked_client);
2004
2005 mutex_unlock(&dm_bufio_clients_lock);
2006}
2007
2008static void cleanup_old_buffers(void)
2009{
2010 unsigned long max_age_hz = get_max_age_hz();
2011 struct dm_bufio_client *c;
2012
2013 mutex_lock(&dm_bufio_clients_lock);
2014
2015 __cache_size_refresh();
2016
2017 list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2018 __evict_old_buffers(c, max_age_hz);
2019
2020 mutex_unlock(&dm_bufio_clients_lock);
2021}
2022
2023static void work_fn(struct work_struct *w)
2024{
2025 cleanup_old_buffers();
2026
2027 queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2028 DM_BUFIO_WORK_TIMER_SECS * HZ);
2029}
2030
2031/*----------------------------------------------------------------
2032 * Module setup
2033 *--------------------------------------------------------------*/
2034
2035/*
2036 * This is called only once for the whole dm_bufio module.
2037 * It initializes memory limit.
2038 */
2039static int __init dm_bufio_init(void)
2040{
2041 __u64 mem;
2042
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;
2047
2048 mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2049 DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2050
2051 if (mem > ULONG_MAX)
2052 mem = ULONG_MAX;
2053
2054#ifdef CONFIG_MMU
2055 if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2056 mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2057#endif
2058
2059 dm_bufio_default_cache_size = mem;
2060
2061 mutex_lock(&dm_bufio_clients_lock);
2062 __cache_size_refresh();
2063 mutex_unlock(&dm_bufio_clients_lock);
2064
2065 dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2066 if (!dm_bufio_wq)
2067 return -ENOMEM;
2068
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);
2073
2074 return 0;
2075}
2076
2077/*
2078 * This is called once when unloading the dm_bufio module.
2079 */
2080static void __exit dm_bufio_exit(void)
2081{
2082 int bug = 0;
2083
2084 cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2085 flush_workqueue(dm_bufio_wq);
2086 destroy_workqueue(dm_bufio_wq);
2087
2088 if (dm_bufio_client_count) {
2089 DMCRIT("%s: dm_bufio_client_count leaked: %d",
2090 __func__, dm_bufio_client_count);
2091 bug = 1;
2092 }
2093
2094 if (dm_bufio_current_allocated) {
2095 DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2096 __func__, dm_bufio_current_allocated);
2097 bug = 1;
2098 }
2099
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);
2103 bug = 1;
2104 }
2105
2106 if (dm_bufio_allocated_vmalloc) {
2107 DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2108 __func__, dm_bufio_allocated_vmalloc);
2109 bug = 1;
2110 }
2111
2112 BUG_ON(bug);
2113}
2114
2115module_init(dm_bufio_init)
2116module_exit(dm_bufio_exit)
2117
2118module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
2119MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2120
2121module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
2122MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2123
2124module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
2125MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2126
2127module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
2128MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2129
2130module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
2131MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2132
2133module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
2134MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2135
2136module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
2137MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2138
2139module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
2140MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2141
2142MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2143MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2144MODULE_LICENSE("GPL");