xref: /openbmc/linux/drivers/md/dm-bufio.c (revision 02891844)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2009-2011 Red Hat, Inc.
4  *
5  * Author: Mikulas Patocka <mpatocka@redhat.com>
6  *
7  * This file is released under the GPL.
8  */
9 
10 #include <linux/dm-bufio.h>
11 
12 #include <linux/device-mapper.h>
13 #include <linux/dm-io.h>
14 #include <linux/slab.h>
15 #include <linux/sched/mm.h>
16 #include <linux/jiffies.h>
17 #include <linux/vmalloc.h>
18 #include <linux/shrinker.h>
19 #include <linux/module.h>
20 #include <linux/rbtree.h>
21 #include <linux/stacktrace.h>
22 #include <linux/jump_label.h>
23 
24 #define DM_MSG_PREFIX "bufio"
25 
26 /*
27  * Memory management policy:
28  *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
29  *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
30  *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
31  *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
32  *	dirty buffers.
33  */
34 #define DM_BUFIO_MIN_BUFFERS		8
35 
36 #define DM_BUFIO_MEMORY_PERCENT		2
37 #define DM_BUFIO_VMALLOC_PERCENT	25
38 #define DM_BUFIO_WRITEBACK_RATIO	3
39 #define DM_BUFIO_LOW_WATERMARK_RATIO	16
40 
41 /*
42  * Check buffer ages in this interval (seconds)
43  */
44 #define DM_BUFIO_WORK_TIMER_SECS	30
45 
46 /*
47  * Free buffers when they are older than this (seconds)
48  */
49 #define DM_BUFIO_DEFAULT_AGE_SECS	300
50 
51 /*
52  * The nr of bytes of cached data to keep around.
53  */
54 #define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
55 
56 /*
57  * Align buffer writes to this boundary.
58  * Tests show that SSDs have the highest IOPS when using 4k writes.
59  */
60 #define DM_BUFIO_WRITE_ALIGN		4096
61 
62 /*
63  * dm_buffer->list_mode
64  */
65 #define LIST_CLEAN	0
66 #define LIST_DIRTY	1
67 #define LIST_SIZE	2
68 
69 /*
70  * Linking of buffers:
71  *	All buffers are linked to buffer_tree with their node field.
72  *
73  *	Clean buffers that are not being written (B_WRITING not set)
74  *	are linked to lru[LIST_CLEAN] with their lru_list field.
75  *
76  *	Dirty and clean buffers that are being written are linked to
77  *	lru[LIST_DIRTY] with their lru_list field. When the write
78  *	finishes, the buffer cannot be relinked immediately (because we
79  *	are in an interrupt context and relinking requires process
80  *	context), so some clean-not-writing buffers can be held on
81  *	dirty_lru too.  They are later added to lru in the process
82  *	context.
83  */
84 struct dm_bufio_client {
85 	struct mutex lock;
86 	spinlock_t spinlock;
87 	bool no_sleep;
88 
89 	struct list_head lru[LIST_SIZE];
90 	unsigned long n_buffers[LIST_SIZE];
91 
92 	struct block_device *bdev;
93 	unsigned int block_size;
94 	s8 sectors_per_block_bits;
95 	void (*alloc_callback)(struct dm_buffer *buf);
96 	void (*write_callback)(struct dm_buffer *buf);
97 	struct kmem_cache *slab_buffer;
98 	struct kmem_cache *slab_cache;
99 	struct dm_io_client *dm_io;
100 
101 	struct list_head reserved_buffers;
102 	unsigned int need_reserved_buffers;
103 
104 	unsigned int minimum_buffers;
105 
106 	struct rb_root buffer_tree;
107 	wait_queue_head_t free_buffer_wait;
108 
109 	sector_t start;
110 
111 	int async_write_error;
112 
113 	struct list_head client_list;
114 
115 	struct shrinker shrinker;
116 	struct work_struct shrink_work;
117 	atomic_long_t need_shrink;
118 };
119 
120 /*
121  * Buffer state bits.
122  */
123 #define B_READING	0
124 #define B_WRITING	1
125 #define B_DIRTY		2
126 
127 /*
128  * Describes how the block was allocated:
129  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
130  * See the comment at alloc_buffer_data.
131  */
132 enum data_mode {
133 	DATA_MODE_SLAB = 0,
134 	DATA_MODE_GET_FREE_PAGES = 1,
135 	DATA_MODE_VMALLOC = 2,
136 	DATA_MODE_LIMIT = 3
137 };
138 
139 struct dm_buffer {
140 	struct rb_node node;
141 	struct list_head lru_list;
142 	struct list_head global_list;
143 	sector_t block;
144 	void *data;
145 	unsigned char data_mode;		/* DATA_MODE_* */
146 	unsigned char list_mode;		/* LIST_* */
147 	blk_status_t read_error;
148 	blk_status_t write_error;
149 	unsigned int accessed;
150 	unsigned int hold_count;
151 	unsigned long state;
152 	unsigned long last_accessed;
153 	unsigned int dirty_start;
154 	unsigned int dirty_end;
155 	unsigned int write_start;
156 	unsigned int write_end;
157 	struct dm_bufio_client *c;
158 	struct list_head write_list;
159 	void (*end_io)(struct dm_buffer *buf, blk_status_t stat);
160 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
161 #define MAX_STACK 10
162 	unsigned int stack_len;
163 	unsigned long stack_entries[MAX_STACK];
164 #endif
165 };
166 
167 static DEFINE_STATIC_KEY_FALSE(no_sleep_enabled);
168 
169 /*----------------------------------------------------------------*/
170 
171 #define dm_bufio_in_request()	(!!current->bio_list)
172 
173 static void dm_bufio_lock(struct dm_bufio_client *c)
174 {
175 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
176 		spin_lock_bh(&c->spinlock);
177 	else
178 		mutex_lock_nested(&c->lock, dm_bufio_in_request());
179 }
180 
181 static int dm_bufio_trylock(struct dm_bufio_client *c)
182 {
183 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
184 		return spin_trylock_bh(&c->spinlock);
185 	else
186 		return mutex_trylock(&c->lock);
187 }
188 
189 static void dm_bufio_unlock(struct dm_bufio_client *c)
190 {
191 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
192 		spin_unlock_bh(&c->spinlock);
193 	else
194 		mutex_unlock(&c->lock);
195 }
196 
197 /*----------------------------------------------------------------*/
198 
199 /*
200  * Default cache size: available memory divided by the ratio.
201  */
202 static unsigned long dm_bufio_default_cache_size;
203 
204 /*
205  * Total cache size set by the user.
206  */
207 static unsigned long dm_bufio_cache_size;
208 
209 /*
210  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
211  * at any time.  If it disagrees, the user has changed cache size.
212  */
213 static unsigned long dm_bufio_cache_size_latch;
214 
215 static DEFINE_SPINLOCK(global_spinlock);
216 
217 static LIST_HEAD(global_queue);
218 
219 static unsigned long global_num;
220 
221 /*
222  * Buffers are freed after this timeout
223  */
224 static unsigned int dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
225 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
226 
227 static unsigned long dm_bufio_peak_allocated;
228 static unsigned long dm_bufio_allocated_kmem_cache;
229 static unsigned long dm_bufio_allocated_get_free_pages;
230 static unsigned long dm_bufio_allocated_vmalloc;
231 static unsigned long dm_bufio_current_allocated;
232 
233 /*----------------------------------------------------------------*/
234 
235 /*
236  * The current number of clients.
237  */
238 static int dm_bufio_client_count;
239 
240 /*
241  * The list of all clients.
242  */
243 static LIST_HEAD(dm_bufio_all_clients);
244 
245 /*
246  * This mutex protects dm_bufio_cache_size_latch and dm_bufio_client_count
247  */
248 static DEFINE_MUTEX(dm_bufio_clients_lock);
249 
250 static struct workqueue_struct *dm_bufio_wq;
251 static struct delayed_work dm_bufio_cleanup_old_work;
252 static struct work_struct dm_bufio_replacement_work;
253 
254 
255 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
256 static void buffer_record_stack(struct dm_buffer *b)
257 {
258 	b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
259 }
260 #endif
261 
262 /*
263  *----------------------------------------------------------------
264  * A red/black tree acts as an index for all the buffers.
265  *----------------------------------------------------------------
266  */
267 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
268 {
269 	struct rb_node *n = c->buffer_tree.rb_node;
270 	struct dm_buffer *b;
271 
272 	while (n) {
273 		b = container_of(n, struct dm_buffer, node);
274 
275 		if (b->block == block)
276 			return b;
277 
278 		n = block < b->block ? n->rb_left : n->rb_right;
279 	}
280 
281 	return NULL;
282 }
283 
284 static struct dm_buffer *__find_next(struct dm_bufio_client *c, sector_t block)
285 {
286 	struct rb_node *n = c->buffer_tree.rb_node;
287 	struct dm_buffer *b;
288 	struct dm_buffer *best = NULL;
289 
290 	while (n) {
291 		b = container_of(n, struct dm_buffer, node);
292 
293 		if (b->block == block)
294 			return b;
295 
296 		if (block <= b->block) {
297 			n = n->rb_left;
298 			best = b;
299 		} else {
300 			n = n->rb_right;
301 		}
302 	}
303 
304 	return best;
305 }
306 
307 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
308 {
309 	struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
310 	struct dm_buffer *found;
311 
312 	while (*new) {
313 		found = container_of(*new, struct dm_buffer, node);
314 
315 		if (found->block == b->block) {
316 			BUG_ON(found != b);
317 			return;
318 		}
319 
320 		parent = *new;
321 		new = b->block < found->block ?
322 			&found->node.rb_left : &found->node.rb_right;
323 	}
324 
325 	rb_link_node(&b->node, parent, new);
326 	rb_insert_color(&b->node, &c->buffer_tree);
327 }
328 
329 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
330 {
331 	rb_erase(&b->node, &c->buffer_tree);
332 }
333 
334 /*----------------------------------------------------------------*/
335 
336 static void adjust_total_allocated(struct dm_buffer *b, bool unlink)
337 {
338 	unsigned char data_mode;
339 	long diff;
340 
341 	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
342 		&dm_bufio_allocated_kmem_cache,
343 		&dm_bufio_allocated_get_free_pages,
344 		&dm_bufio_allocated_vmalloc,
345 	};
346 
347 	data_mode = b->data_mode;
348 	diff = (long)b->c->block_size;
349 	if (unlink)
350 		diff = -diff;
351 
352 	spin_lock(&global_spinlock);
353 
354 	*class_ptr[data_mode] += diff;
355 
356 	dm_bufio_current_allocated += diff;
357 
358 	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
359 		dm_bufio_peak_allocated = dm_bufio_current_allocated;
360 
361 	b->accessed = 1;
362 
363 	if (!unlink) {
364 		list_add(&b->global_list, &global_queue);
365 		global_num++;
366 		if (dm_bufio_current_allocated > dm_bufio_cache_size)
367 			queue_work(dm_bufio_wq, &dm_bufio_replacement_work);
368 	} else {
369 		list_del(&b->global_list);
370 		global_num--;
371 	}
372 
373 	spin_unlock(&global_spinlock);
374 }
375 
376 /*
377  * Change the number of clients and recalculate per-client limit.
378  */
379 static void __cache_size_refresh(void)
380 {
381 	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
382 	BUG_ON(dm_bufio_client_count < 0);
383 
384 	dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
385 
386 	/*
387 	 * Use default if set to 0 and report the actual cache size used.
388 	 */
389 	if (!dm_bufio_cache_size_latch) {
390 		(void)cmpxchg(&dm_bufio_cache_size, 0,
391 			      dm_bufio_default_cache_size);
392 		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
393 	}
394 }
395 
396 /*
397  * Allocating buffer data.
398  *
399  * Small buffers are allocated with kmem_cache, to use space optimally.
400  *
401  * For large buffers, we choose between get_free_pages and vmalloc.
402  * Each has advantages and disadvantages.
403  *
404  * __get_free_pages can randomly fail if the memory is fragmented.
405  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
406  * as low as 128M) so using it for caching is not appropriate.
407  *
408  * If the allocation may fail we use __get_free_pages. Memory fragmentation
409  * won't have a fatal effect here, but it just causes flushes of some other
410  * buffers and more I/O will be performed. Don't use __get_free_pages if it
411  * always fails (i.e. order > MAX_ORDER).
412  *
413  * If the allocation shouldn't fail we use __vmalloc. This is only for the
414  * initial reserve allocation, so there's no risk of wasting all vmalloc
415  * space.
416  */
417 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
418 			       unsigned char *data_mode)
419 {
420 	if (unlikely(c->slab_cache != NULL)) {
421 		*data_mode = DATA_MODE_SLAB;
422 		return kmem_cache_alloc(c->slab_cache, gfp_mask);
423 	}
424 
425 	if (c->block_size <= KMALLOC_MAX_SIZE &&
426 	    gfp_mask & __GFP_NORETRY) {
427 		*data_mode = DATA_MODE_GET_FREE_PAGES;
428 		return (void *)__get_free_pages(gfp_mask,
429 						c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
430 	}
431 
432 	*data_mode = DATA_MODE_VMALLOC;
433 
434 	/*
435 	 * __vmalloc allocates the data pages and auxiliary structures with
436 	 * gfp_flags that were specified, but pagetables are always allocated
437 	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
438 	 *
439 	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
440 	 * all allocations done by this process (including pagetables) are done
441 	 * as if GFP_NOIO was specified.
442 	 */
443 	if (gfp_mask & __GFP_NORETRY) {
444 		unsigned int noio_flag = memalloc_noio_save();
445 		void *ptr = __vmalloc(c->block_size, gfp_mask);
446 
447 		memalloc_noio_restore(noio_flag);
448 		return ptr;
449 	}
450 
451 	return __vmalloc(c->block_size, gfp_mask);
452 }
453 
454 /*
455  * Free buffer's data.
456  */
457 static void free_buffer_data(struct dm_bufio_client *c,
458 			     void *data, unsigned char data_mode)
459 {
460 	switch (data_mode) {
461 	case DATA_MODE_SLAB:
462 		kmem_cache_free(c->slab_cache, data);
463 		break;
464 
465 	case DATA_MODE_GET_FREE_PAGES:
466 		free_pages((unsigned long)data,
467 			   c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
468 		break;
469 
470 	case DATA_MODE_VMALLOC:
471 		vfree(data);
472 		break;
473 
474 	default:
475 		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
476 		       data_mode);
477 		BUG();
478 	}
479 }
480 
481 /*
482  * Allocate buffer and its data.
483  */
484 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
485 {
486 	struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
487 
488 	if (!b)
489 		return NULL;
490 
491 	b->c = c;
492 
493 	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
494 	if (!b->data) {
495 		kmem_cache_free(c->slab_buffer, b);
496 		return NULL;
497 	}
498 
499 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
500 	b->stack_len = 0;
501 #endif
502 	return b;
503 }
504 
505 /*
506  * Free buffer and its data.
507  */
508 static void free_buffer(struct dm_buffer *b)
509 {
510 	struct dm_bufio_client *c = b->c;
511 
512 	free_buffer_data(c, b->data, b->data_mode);
513 	kmem_cache_free(c->slab_buffer, b);
514 }
515 
516 /*
517  * Link buffer to the buffer tree and clean or dirty queue.
518  */
519 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
520 {
521 	struct dm_bufio_client *c = b->c;
522 
523 	c->n_buffers[dirty]++;
524 	b->block = block;
525 	b->list_mode = dirty;
526 	list_add(&b->lru_list, &c->lru[dirty]);
527 	__insert(b->c, b);
528 	b->last_accessed = jiffies;
529 
530 	adjust_total_allocated(b, false);
531 }
532 
533 /*
534  * Unlink buffer from the buffer tree and dirty or clean queue.
535  */
536 static void __unlink_buffer(struct dm_buffer *b)
537 {
538 	struct dm_bufio_client *c = b->c;
539 
540 	BUG_ON(!c->n_buffers[b->list_mode]);
541 
542 	c->n_buffers[b->list_mode]--;
543 	__remove(b->c, b);
544 	list_del(&b->lru_list);
545 
546 	adjust_total_allocated(b, true);
547 }
548 
549 /*
550  * Place the buffer to the head of dirty or clean LRU queue.
551  */
552 static void __relink_lru(struct dm_buffer *b, int dirty)
553 {
554 	struct dm_bufio_client *c = b->c;
555 
556 	b->accessed = 1;
557 
558 	BUG_ON(!c->n_buffers[b->list_mode]);
559 
560 	c->n_buffers[b->list_mode]--;
561 	c->n_buffers[dirty]++;
562 	b->list_mode = dirty;
563 	list_move(&b->lru_list, &c->lru[dirty]);
564 	b->last_accessed = jiffies;
565 }
566 
567 /*
568  *--------------------------------------------------------------------------
569  * Submit I/O on the buffer.
570  *
571  * Bio interface is faster but it has some problems:
572  *	the vector list is limited (increasing this limit increases
573  *	memory-consumption per buffer, so it is not viable);
574  *
575  *	the memory must be direct-mapped, not vmalloced;
576  *
577  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
578  * it is not vmalloced, try using the bio interface.
579  *
580  * If the buffer is big, if it is vmalloced or if the underlying device
581  * rejects the bio because it is too large, use dm-io layer to do the I/O.
582  * The dm-io layer splits the I/O into multiple requests, avoiding the above
583  * shortcomings.
584  *--------------------------------------------------------------------------
585  */
586 
587 /*
588  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
589  * that the request was handled directly with bio interface.
590  */
591 static void dmio_complete(unsigned long error, void *context)
592 {
593 	struct dm_buffer *b = context;
594 
595 	b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
596 }
597 
598 static void use_dmio(struct dm_buffer *b, enum req_op op, sector_t sector,
599 		     unsigned int n_sectors, unsigned int offset)
600 {
601 	int r;
602 	struct dm_io_request io_req = {
603 		.bi_opf = op,
604 		.notify.fn = dmio_complete,
605 		.notify.context = b,
606 		.client = b->c->dm_io,
607 	};
608 	struct dm_io_region region = {
609 		.bdev = b->c->bdev,
610 		.sector = sector,
611 		.count = n_sectors,
612 	};
613 
614 	if (b->data_mode != DATA_MODE_VMALLOC) {
615 		io_req.mem.type = DM_IO_KMEM;
616 		io_req.mem.ptr.addr = (char *)b->data + offset;
617 	} else {
618 		io_req.mem.type = DM_IO_VMA;
619 		io_req.mem.ptr.vma = (char *)b->data + offset;
620 	}
621 
622 	r = dm_io(&io_req, 1, &region, NULL);
623 	if (unlikely(r))
624 		b->end_io(b, errno_to_blk_status(r));
625 }
626 
627 static void bio_complete(struct bio *bio)
628 {
629 	struct dm_buffer *b = bio->bi_private;
630 	blk_status_t status = bio->bi_status;
631 
632 	bio_uninit(bio);
633 	kfree(bio);
634 	b->end_io(b, status);
635 }
636 
637 static void use_bio(struct dm_buffer *b, enum req_op op, sector_t sector,
638 		    unsigned int n_sectors, unsigned int offset)
639 {
640 	struct bio *bio;
641 	char *ptr;
642 	unsigned int vec_size, len;
643 
644 	vec_size = b->c->block_size >> PAGE_SHIFT;
645 	if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
646 		vec_size += 2;
647 
648 	bio = bio_kmalloc(vec_size, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN);
649 	if (!bio) {
650 dmio:
651 		use_dmio(b, op, sector, n_sectors, offset);
652 		return;
653 	}
654 	bio_init(bio, b->c->bdev, bio->bi_inline_vecs, vec_size, op);
655 	bio->bi_iter.bi_sector = sector;
656 	bio->bi_end_io = bio_complete;
657 	bio->bi_private = b;
658 
659 	ptr = (char *)b->data + offset;
660 	len = n_sectors << SECTOR_SHIFT;
661 
662 	do {
663 		unsigned int this_step = min((unsigned int)(PAGE_SIZE - offset_in_page(ptr)), len);
664 
665 		if (!bio_add_page(bio, virt_to_page(ptr), this_step,
666 				  offset_in_page(ptr))) {
667 			bio_put(bio);
668 			goto dmio;
669 		}
670 
671 		len -= this_step;
672 		ptr += this_step;
673 	} while (len > 0);
674 
675 	submit_bio(bio);
676 }
677 
678 static inline sector_t block_to_sector(struct dm_bufio_client *c, sector_t block)
679 {
680 	sector_t sector;
681 
682 	if (likely(c->sectors_per_block_bits >= 0))
683 		sector = block << c->sectors_per_block_bits;
684 	else
685 		sector = block * (c->block_size >> SECTOR_SHIFT);
686 	sector += c->start;
687 
688 	return sector;
689 }
690 
691 static void submit_io(struct dm_buffer *b, enum req_op op,
692 		      void (*end_io)(struct dm_buffer *, blk_status_t))
693 {
694 	unsigned int n_sectors;
695 	sector_t sector;
696 	unsigned int offset, end;
697 
698 	b->end_io = end_io;
699 
700 	sector = block_to_sector(b->c, b->block);
701 
702 	if (op != REQ_OP_WRITE) {
703 		n_sectors = b->c->block_size >> SECTOR_SHIFT;
704 		offset = 0;
705 	} else {
706 		if (b->c->write_callback)
707 			b->c->write_callback(b);
708 		offset = b->write_start;
709 		end = b->write_end;
710 		offset &= -DM_BUFIO_WRITE_ALIGN;
711 		end += DM_BUFIO_WRITE_ALIGN - 1;
712 		end &= -DM_BUFIO_WRITE_ALIGN;
713 		if (unlikely(end > b->c->block_size))
714 			end = b->c->block_size;
715 
716 		sector += offset >> SECTOR_SHIFT;
717 		n_sectors = (end - offset) >> SECTOR_SHIFT;
718 	}
719 
720 	if (b->data_mode != DATA_MODE_VMALLOC)
721 		use_bio(b, op, sector, n_sectors, offset);
722 	else
723 		use_dmio(b, op, sector, n_sectors, offset);
724 }
725 
726 /*
727  *--------------------------------------------------------------
728  * Writing dirty buffers
729  *--------------------------------------------------------------
730  */
731 
732 /*
733  * The endio routine for write.
734  *
735  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
736  * it.
737  */
738 static void write_endio(struct dm_buffer *b, blk_status_t status)
739 {
740 	b->write_error = status;
741 	if (unlikely(status)) {
742 		struct dm_bufio_client *c = b->c;
743 
744 		(void)cmpxchg(&c->async_write_error, 0,
745 				blk_status_to_errno(status));
746 	}
747 
748 	BUG_ON(!test_bit(B_WRITING, &b->state));
749 
750 	smp_mb__before_atomic();
751 	clear_bit(B_WRITING, &b->state);
752 	smp_mb__after_atomic();
753 
754 	wake_up_bit(&b->state, B_WRITING);
755 }
756 
757 /*
758  * Initiate a write on a dirty buffer, but don't wait for it.
759  *
760  * - If the buffer is not dirty, exit.
761  * - If there some previous write going on, wait for it to finish (we can't
762  *   have two writes on the same buffer simultaneously).
763  * - Submit our write and don't wait on it. We set B_WRITING indicating
764  *   that there is a write in progress.
765  */
766 static void __write_dirty_buffer(struct dm_buffer *b,
767 				 struct list_head *write_list)
768 {
769 	if (!test_bit(B_DIRTY, &b->state))
770 		return;
771 
772 	clear_bit(B_DIRTY, &b->state);
773 	wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
774 
775 	b->write_start = b->dirty_start;
776 	b->write_end = b->dirty_end;
777 
778 	if (!write_list)
779 		submit_io(b, REQ_OP_WRITE, write_endio);
780 	else
781 		list_add_tail(&b->write_list, write_list);
782 }
783 
784 static void __flush_write_list(struct list_head *write_list)
785 {
786 	struct blk_plug plug;
787 
788 	blk_start_plug(&plug);
789 	while (!list_empty(write_list)) {
790 		struct dm_buffer *b =
791 			list_entry(write_list->next, struct dm_buffer, write_list);
792 		list_del(&b->write_list);
793 		submit_io(b, REQ_OP_WRITE, write_endio);
794 		cond_resched();
795 	}
796 	blk_finish_plug(&plug);
797 }
798 
799 /*
800  * Wait until any activity on the buffer finishes.  Possibly write the
801  * buffer if it is dirty.  When this function finishes, there is no I/O
802  * running on the buffer and the buffer is not dirty.
803  */
804 static void __make_buffer_clean(struct dm_buffer *b)
805 {
806 	BUG_ON(b->hold_count);
807 
808 	/* smp_load_acquire() pairs with read_endio()'s smp_mb__before_atomic() */
809 	if (!smp_load_acquire(&b->state))	/* fast case */
810 		return;
811 
812 	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
813 	__write_dirty_buffer(b, NULL);
814 	wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
815 }
816 
817 /*
818  * Find some buffer that is not held by anybody, clean it, unlink it and
819  * return it.
820  */
821 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
822 {
823 	struct dm_buffer *b;
824 
825 	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
826 		BUG_ON(test_bit(B_WRITING, &b->state));
827 		BUG_ON(test_bit(B_DIRTY, &b->state));
828 
829 		if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep &&
830 		    unlikely(test_bit_acquire(B_READING, &b->state)))
831 			continue;
832 
833 		if (!b->hold_count) {
834 			__make_buffer_clean(b);
835 			__unlink_buffer(b);
836 			return b;
837 		}
838 		cond_resched();
839 	}
840 
841 	if (static_branch_unlikely(&no_sleep_enabled) && c->no_sleep)
842 		return NULL;
843 
844 	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
845 		BUG_ON(test_bit(B_READING, &b->state));
846 
847 		if (!b->hold_count) {
848 			__make_buffer_clean(b);
849 			__unlink_buffer(b);
850 			return b;
851 		}
852 		cond_resched();
853 	}
854 
855 	return NULL;
856 }
857 
858 /*
859  * Wait until some other threads free some buffer or release hold count on
860  * some buffer.
861  *
862  * This function is entered with c->lock held, drops it and regains it
863  * before exiting.
864  */
865 static void __wait_for_free_buffer(struct dm_bufio_client *c)
866 {
867 	DECLARE_WAITQUEUE(wait, current);
868 
869 	add_wait_queue(&c->free_buffer_wait, &wait);
870 	set_current_state(TASK_UNINTERRUPTIBLE);
871 	dm_bufio_unlock(c);
872 
873 	io_schedule();
874 
875 	remove_wait_queue(&c->free_buffer_wait, &wait);
876 
877 	dm_bufio_lock(c);
878 }
879 
880 enum new_flag {
881 	NF_FRESH = 0,
882 	NF_READ = 1,
883 	NF_GET = 2,
884 	NF_PREFETCH = 3
885 };
886 
887 /*
888  * Allocate a new buffer. If the allocation is not possible, wait until
889  * some other thread frees a buffer.
890  *
891  * May drop the lock and regain it.
892  */
893 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
894 {
895 	struct dm_buffer *b;
896 	bool tried_noio_alloc = false;
897 
898 	/*
899 	 * dm-bufio is resistant to allocation failures (it just keeps
900 	 * one buffer reserved in cases all the allocations fail).
901 	 * So set flags to not try too hard:
902 	 *	GFP_NOWAIT: don't wait; if we need to sleep we'll release our
903 	 *		    mutex and wait ourselves.
904 	 *	__GFP_NORETRY: don't retry and rather return failure
905 	 *	__GFP_NOMEMALLOC: don't use emergency reserves
906 	 *	__GFP_NOWARN: don't print a warning in case of failure
907 	 *
908 	 * For debugging, if we set the cache size to 1, no new buffers will
909 	 * be allocated.
910 	 */
911 	while (1) {
912 		if (dm_bufio_cache_size_latch != 1) {
913 			b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
914 			if (b)
915 				return b;
916 		}
917 
918 		if (nf == NF_PREFETCH)
919 			return NULL;
920 
921 		if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
922 			dm_bufio_unlock(c);
923 			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
924 			dm_bufio_lock(c);
925 			if (b)
926 				return b;
927 			tried_noio_alloc = true;
928 		}
929 
930 		if (!list_empty(&c->reserved_buffers)) {
931 			b = list_entry(c->reserved_buffers.next,
932 				       struct dm_buffer, lru_list);
933 			list_del(&b->lru_list);
934 			c->need_reserved_buffers++;
935 
936 			return b;
937 		}
938 
939 		b = __get_unclaimed_buffer(c);
940 		if (b)
941 			return b;
942 
943 		__wait_for_free_buffer(c);
944 	}
945 }
946 
947 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
948 {
949 	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
950 
951 	if (!b)
952 		return NULL;
953 
954 	if (c->alloc_callback)
955 		c->alloc_callback(b);
956 
957 	return b;
958 }
959 
960 /*
961  * Free a buffer and wake other threads waiting for free buffers.
962  */
963 static void __free_buffer_wake(struct dm_buffer *b)
964 {
965 	struct dm_bufio_client *c = b->c;
966 
967 	if (!c->need_reserved_buffers)
968 		free_buffer(b);
969 	else {
970 		list_add(&b->lru_list, &c->reserved_buffers);
971 		c->need_reserved_buffers--;
972 	}
973 
974 	wake_up(&c->free_buffer_wait);
975 }
976 
977 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
978 					struct list_head *write_list)
979 {
980 	struct dm_buffer *b, *tmp;
981 
982 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
983 		BUG_ON(test_bit(B_READING, &b->state));
984 
985 		if (!test_bit(B_DIRTY, &b->state) &&
986 		    !test_bit(B_WRITING, &b->state)) {
987 			__relink_lru(b, LIST_CLEAN);
988 			continue;
989 		}
990 
991 		if (no_wait && test_bit(B_WRITING, &b->state))
992 			return;
993 
994 		__write_dirty_buffer(b, write_list);
995 		cond_resched();
996 	}
997 }
998 
999 /*
1000  * Check if we're over watermark.
1001  * If we are over threshold_buffers, start freeing buffers.
1002  * If we're over "limit_buffers", block until we get under the limit.
1003  */
1004 static void __check_watermark(struct dm_bufio_client *c,
1005 			      struct list_head *write_list)
1006 {
1007 	if (c->n_buffers[LIST_DIRTY] > c->n_buffers[LIST_CLEAN] * DM_BUFIO_WRITEBACK_RATIO)
1008 		__write_dirty_buffers_async(c, 1, write_list);
1009 }
1010 
1011 /*
1012  *--------------------------------------------------------------
1013  * Getting a buffer
1014  *--------------------------------------------------------------
1015  */
1016 
1017 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
1018 				     enum new_flag nf, int *need_submit,
1019 				     struct list_head *write_list)
1020 {
1021 	struct dm_buffer *b, *new_b = NULL;
1022 
1023 	*need_submit = 0;
1024 
1025 	b = __find(c, block);
1026 	if (b)
1027 		goto found_buffer;
1028 
1029 	if (nf == NF_GET)
1030 		return NULL;
1031 
1032 	new_b = __alloc_buffer_wait(c, nf);
1033 	if (!new_b)
1034 		return NULL;
1035 
1036 	/*
1037 	 * We've had a period where the mutex was unlocked, so need to
1038 	 * recheck the buffer tree.
1039 	 */
1040 	b = __find(c, block);
1041 	if (b) {
1042 		__free_buffer_wake(new_b);
1043 		goto found_buffer;
1044 	}
1045 
1046 	__check_watermark(c, write_list);
1047 
1048 	b = new_b;
1049 	b->hold_count = 1;
1050 	b->read_error = 0;
1051 	b->write_error = 0;
1052 	__link_buffer(b, block, LIST_CLEAN);
1053 
1054 	if (nf == NF_FRESH) {
1055 		b->state = 0;
1056 		return b;
1057 	}
1058 
1059 	b->state = 1 << B_READING;
1060 	*need_submit = 1;
1061 
1062 	return b;
1063 
1064 found_buffer:
1065 	if (nf == NF_PREFETCH)
1066 		return NULL;
1067 	/*
1068 	 * Note: it is essential that we don't wait for the buffer to be
1069 	 * read if dm_bufio_get function is used. Both dm_bufio_get and
1070 	 * dm_bufio_prefetch can be used in the driver request routine.
1071 	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1072 	 * the same buffer, it would deadlock if we waited.
1073 	 */
1074 	if (nf == NF_GET && unlikely(test_bit_acquire(B_READING, &b->state)))
1075 		return NULL;
1076 
1077 	b->hold_count++;
1078 	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1079 		     test_bit(B_WRITING, &b->state));
1080 	return b;
1081 }
1082 
1083 /*
1084  * The endio routine for reading: set the error, clear the bit and wake up
1085  * anyone waiting on the buffer.
1086  */
1087 static void read_endio(struct dm_buffer *b, blk_status_t status)
1088 {
1089 	b->read_error = status;
1090 
1091 	BUG_ON(!test_bit(B_READING, &b->state));
1092 
1093 	smp_mb__before_atomic();
1094 	clear_bit(B_READING, &b->state);
1095 	smp_mb__after_atomic();
1096 
1097 	wake_up_bit(&b->state, B_READING);
1098 }
1099 
1100 /*
1101  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1102  * functions is similar except that dm_bufio_new doesn't read the
1103  * buffer from the disk (assuming that the caller overwrites all the data
1104  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1105  */
1106 static void *new_read(struct dm_bufio_client *c, sector_t block,
1107 		      enum new_flag nf, struct dm_buffer **bp)
1108 {
1109 	int need_submit;
1110 	struct dm_buffer *b;
1111 
1112 	LIST_HEAD(write_list);
1113 
1114 	dm_bufio_lock(c);
1115 	b = __bufio_new(c, block, nf, &need_submit, &write_list);
1116 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1117 	if (b && b->hold_count == 1)
1118 		buffer_record_stack(b);
1119 #endif
1120 	dm_bufio_unlock(c);
1121 
1122 	__flush_write_list(&write_list);
1123 
1124 	if (!b)
1125 		return NULL;
1126 
1127 	if (need_submit)
1128 		submit_io(b, REQ_OP_READ, read_endio);
1129 
1130 	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1131 
1132 	if (b->read_error) {
1133 		int error = blk_status_to_errno(b->read_error);
1134 
1135 		dm_bufio_release(b);
1136 
1137 		return ERR_PTR(error);
1138 	}
1139 
1140 	*bp = b;
1141 
1142 	return b->data;
1143 }
1144 
1145 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1146 		   struct dm_buffer **bp)
1147 {
1148 	return new_read(c, block, NF_GET, bp);
1149 }
1150 EXPORT_SYMBOL_GPL(dm_bufio_get);
1151 
1152 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1153 		    struct dm_buffer **bp)
1154 {
1155 	BUG_ON(dm_bufio_in_request());
1156 
1157 	return new_read(c, block, NF_READ, bp);
1158 }
1159 EXPORT_SYMBOL_GPL(dm_bufio_read);
1160 
1161 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1162 		   struct dm_buffer **bp)
1163 {
1164 	BUG_ON(dm_bufio_in_request());
1165 
1166 	return new_read(c, block, NF_FRESH, bp);
1167 }
1168 EXPORT_SYMBOL_GPL(dm_bufio_new);
1169 
1170 void dm_bufio_prefetch(struct dm_bufio_client *c,
1171 		       sector_t block, unsigned int n_blocks)
1172 {
1173 	struct blk_plug plug;
1174 
1175 	LIST_HEAD(write_list);
1176 
1177 	BUG_ON(dm_bufio_in_request());
1178 
1179 	blk_start_plug(&plug);
1180 	dm_bufio_lock(c);
1181 
1182 	for (; n_blocks--; block++) {
1183 		int need_submit;
1184 		struct dm_buffer *b;
1185 
1186 		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1187 				&write_list);
1188 		if (unlikely(!list_empty(&write_list))) {
1189 			dm_bufio_unlock(c);
1190 			blk_finish_plug(&plug);
1191 			__flush_write_list(&write_list);
1192 			blk_start_plug(&plug);
1193 			dm_bufio_lock(c);
1194 		}
1195 		if (unlikely(b != NULL)) {
1196 			dm_bufio_unlock(c);
1197 
1198 			if (need_submit)
1199 				submit_io(b, REQ_OP_READ, read_endio);
1200 			dm_bufio_release(b);
1201 
1202 			cond_resched();
1203 
1204 			if (!n_blocks)
1205 				goto flush_plug;
1206 			dm_bufio_lock(c);
1207 		}
1208 	}
1209 
1210 	dm_bufio_unlock(c);
1211 
1212 flush_plug:
1213 	blk_finish_plug(&plug);
1214 }
1215 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1216 
1217 void dm_bufio_release(struct dm_buffer *b)
1218 {
1219 	struct dm_bufio_client *c = b->c;
1220 
1221 	dm_bufio_lock(c);
1222 
1223 	BUG_ON(!b->hold_count);
1224 
1225 	b->hold_count--;
1226 	if (!b->hold_count) {
1227 		wake_up(&c->free_buffer_wait);
1228 
1229 		/*
1230 		 * If there were errors on the buffer, and the buffer is not
1231 		 * to be written, free the buffer. There is no point in caching
1232 		 * invalid buffer.
1233 		 */
1234 		if ((b->read_error || b->write_error) &&
1235 		    !test_bit_acquire(B_READING, &b->state) &&
1236 		    !test_bit(B_WRITING, &b->state) &&
1237 		    !test_bit(B_DIRTY, &b->state)) {
1238 			__unlink_buffer(b);
1239 			__free_buffer_wake(b);
1240 		}
1241 	}
1242 
1243 	dm_bufio_unlock(c);
1244 }
1245 EXPORT_SYMBOL_GPL(dm_bufio_release);
1246 
1247 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1248 					unsigned int start, unsigned int end)
1249 {
1250 	struct dm_bufio_client *c = b->c;
1251 
1252 	BUG_ON(start >= end);
1253 	BUG_ON(end > b->c->block_size);
1254 
1255 	dm_bufio_lock(c);
1256 
1257 	BUG_ON(test_bit(B_READING, &b->state));
1258 
1259 	if (!test_and_set_bit(B_DIRTY, &b->state)) {
1260 		b->dirty_start = start;
1261 		b->dirty_end = end;
1262 		__relink_lru(b, LIST_DIRTY);
1263 	} else {
1264 		if (start < b->dirty_start)
1265 			b->dirty_start = start;
1266 		if (end > b->dirty_end)
1267 			b->dirty_end = end;
1268 	}
1269 
1270 	dm_bufio_unlock(c);
1271 }
1272 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1273 
1274 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1275 {
1276 	dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1277 }
1278 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1279 
1280 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1281 {
1282 	LIST_HEAD(write_list);
1283 
1284 	BUG_ON(dm_bufio_in_request());
1285 
1286 	dm_bufio_lock(c);
1287 	__write_dirty_buffers_async(c, 0, &write_list);
1288 	dm_bufio_unlock(c);
1289 	__flush_write_list(&write_list);
1290 }
1291 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1292 
1293 /*
1294  * For performance, it is essential that the buffers are written asynchronously
1295  * and simultaneously (so that the block layer can merge the writes) and then
1296  * waited upon.
1297  *
1298  * Finally, we flush hardware disk cache.
1299  */
1300 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1301 {
1302 	int a, f;
1303 	unsigned long buffers_processed = 0;
1304 	struct dm_buffer *b, *tmp;
1305 
1306 	LIST_HEAD(write_list);
1307 
1308 	dm_bufio_lock(c);
1309 	__write_dirty_buffers_async(c, 0, &write_list);
1310 	dm_bufio_unlock(c);
1311 	__flush_write_list(&write_list);
1312 	dm_bufio_lock(c);
1313 
1314 again:
1315 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1316 		int dropped_lock = 0;
1317 
1318 		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1319 			buffers_processed++;
1320 
1321 		BUG_ON(test_bit(B_READING, &b->state));
1322 
1323 		if (test_bit(B_WRITING, &b->state)) {
1324 			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1325 				dropped_lock = 1;
1326 				b->hold_count++;
1327 				dm_bufio_unlock(c);
1328 				wait_on_bit_io(&b->state, B_WRITING,
1329 					       TASK_UNINTERRUPTIBLE);
1330 				dm_bufio_lock(c);
1331 				b->hold_count--;
1332 			} else
1333 				wait_on_bit_io(&b->state, B_WRITING,
1334 					       TASK_UNINTERRUPTIBLE);
1335 		}
1336 
1337 		if (!test_bit(B_DIRTY, &b->state) &&
1338 		    !test_bit(B_WRITING, &b->state))
1339 			__relink_lru(b, LIST_CLEAN);
1340 
1341 		cond_resched();
1342 
1343 		/*
1344 		 * If we dropped the lock, the list is no longer consistent,
1345 		 * so we must restart the search.
1346 		 *
1347 		 * In the most common case, the buffer just processed is
1348 		 * relinked to the clean list, so we won't loop scanning the
1349 		 * same buffer again and again.
1350 		 *
1351 		 * This may livelock if there is another thread simultaneously
1352 		 * dirtying buffers, so we count the number of buffers walked
1353 		 * and if it exceeds the total number of buffers, it means that
1354 		 * someone is doing some writes simultaneously with us.  In
1355 		 * this case, stop, dropping the lock.
1356 		 */
1357 		if (dropped_lock)
1358 			goto again;
1359 	}
1360 	wake_up(&c->free_buffer_wait);
1361 	dm_bufio_unlock(c);
1362 
1363 	a = xchg(&c->async_write_error, 0);
1364 	f = dm_bufio_issue_flush(c);
1365 	if (a)
1366 		return a;
1367 
1368 	return f;
1369 }
1370 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1371 
1372 /*
1373  * Use dm-io to send an empty barrier to flush the device.
1374  */
1375 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1376 {
1377 	struct dm_io_request io_req = {
1378 		.bi_opf = REQ_OP_WRITE | REQ_PREFLUSH | REQ_SYNC,
1379 		.mem.type = DM_IO_KMEM,
1380 		.mem.ptr.addr = NULL,
1381 		.client = c->dm_io,
1382 	};
1383 	struct dm_io_region io_reg = {
1384 		.bdev = c->bdev,
1385 		.sector = 0,
1386 		.count = 0,
1387 	};
1388 
1389 	BUG_ON(dm_bufio_in_request());
1390 
1391 	return dm_io(&io_req, 1, &io_reg, NULL);
1392 }
1393 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1394 
1395 /*
1396  * Use dm-io to send a discard request to flush the device.
1397  */
1398 int dm_bufio_issue_discard(struct dm_bufio_client *c, sector_t block, sector_t count)
1399 {
1400 	struct dm_io_request io_req = {
1401 		.bi_opf = REQ_OP_DISCARD | REQ_SYNC,
1402 		.mem.type = DM_IO_KMEM,
1403 		.mem.ptr.addr = NULL,
1404 		.client = c->dm_io,
1405 	};
1406 	struct dm_io_region io_reg = {
1407 		.bdev = c->bdev,
1408 		.sector = block_to_sector(c, block),
1409 		.count = block_to_sector(c, count),
1410 	};
1411 
1412 	BUG_ON(dm_bufio_in_request());
1413 
1414 	return dm_io(&io_req, 1, &io_reg, NULL);
1415 }
1416 EXPORT_SYMBOL_GPL(dm_bufio_issue_discard);
1417 
1418 /*
1419  * We first delete any other buffer that may be at that new location.
1420  *
1421  * Then, we write the buffer to the original location if it was dirty.
1422  *
1423  * Then, if we are the only one who is holding the buffer, relink the buffer
1424  * in the buffer tree for the new location.
1425  *
1426  * If there was someone else holding the buffer, we write it to the new
1427  * location but not relink it, because that other user needs to have the buffer
1428  * at the same place.
1429  */
1430 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1431 {
1432 	struct dm_bufio_client *c = b->c;
1433 	struct dm_buffer *new;
1434 
1435 	BUG_ON(dm_bufio_in_request());
1436 
1437 	dm_bufio_lock(c);
1438 
1439 retry:
1440 	new = __find(c, new_block);
1441 	if (new) {
1442 		if (new->hold_count) {
1443 			__wait_for_free_buffer(c);
1444 			goto retry;
1445 		}
1446 
1447 		/*
1448 		 * FIXME: Is there any point waiting for a write that's going
1449 		 * to be overwritten in a bit?
1450 		 */
1451 		__make_buffer_clean(new);
1452 		__unlink_buffer(new);
1453 		__free_buffer_wake(new);
1454 	}
1455 
1456 	BUG_ON(!b->hold_count);
1457 	BUG_ON(test_bit(B_READING, &b->state));
1458 
1459 	__write_dirty_buffer(b, NULL);
1460 	if (b->hold_count == 1) {
1461 		wait_on_bit_io(&b->state, B_WRITING,
1462 			       TASK_UNINTERRUPTIBLE);
1463 		set_bit(B_DIRTY, &b->state);
1464 		b->dirty_start = 0;
1465 		b->dirty_end = c->block_size;
1466 		__unlink_buffer(b);
1467 		__link_buffer(b, new_block, LIST_DIRTY);
1468 	} else {
1469 		sector_t old_block;
1470 
1471 		wait_on_bit_lock_io(&b->state, B_WRITING,
1472 				    TASK_UNINTERRUPTIBLE);
1473 		/*
1474 		 * Relink buffer to "new_block" so that write_callback
1475 		 * sees "new_block" as a block number.
1476 		 * After the write, link the buffer back to old_block.
1477 		 * All this must be done in bufio lock, so that block number
1478 		 * change isn't visible to other threads.
1479 		 */
1480 		old_block = b->block;
1481 		__unlink_buffer(b);
1482 		__link_buffer(b, new_block, b->list_mode);
1483 		submit_io(b, REQ_OP_WRITE, write_endio);
1484 		wait_on_bit_io(&b->state, B_WRITING,
1485 			       TASK_UNINTERRUPTIBLE);
1486 		__unlink_buffer(b);
1487 		__link_buffer(b, old_block, b->list_mode);
1488 	}
1489 
1490 	dm_bufio_unlock(c);
1491 	dm_bufio_release(b);
1492 }
1493 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1494 
1495 static void forget_buffer_locked(struct dm_buffer *b)
1496 {
1497 	if (likely(!b->hold_count) && likely(!smp_load_acquire(&b->state))) {
1498 		__unlink_buffer(b);
1499 		__free_buffer_wake(b);
1500 	}
1501 }
1502 
1503 /*
1504  * Free the given buffer.
1505  *
1506  * This is just a hint, if the buffer is in use or dirty, this function
1507  * does nothing.
1508  */
1509 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1510 {
1511 	struct dm_buffer *b;
1512 
1513 	dm_bufio_lock(c);
1514 
1515 	b = __find(c, block);
1516 	if (b)
1517 		forget_buffer_locked(b);
1518 
1519 	dm_bufio_unlock(c);
1520 }
1521 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1522 
1523 void dm_bufio_forget_buffers(struct dm_bufio_client *c, sector_t block, sector_t n_blocks)
1524 {
1525 	struct dm_buffer *b;
1526 	sector_t end_block = block + n_blocks;
1527 
1528 	while (block < end_block) {
1529 		dm_bufio_lock(c);
1530 
1531 		b = __find_next(c, block);
1532 		if (b) {
1533 			block = b->block + 1;
1534 			forget_buffer_locked(b);
1535 		}
1536 
1537 		dm_bufio_unlock(c);
1538 
1539 		if (!b)
1540 			break;
1541 	}
1542 
1543 }
1544 EXPORT_SYMBOL_GPL(dm_bufio_forget_buffers);
1545 
1546 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned int n)
1547 {
1548 	c->minimum_buffers = n;
1549 }
1550 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1551 
1552 unsigned int dm_bufio_get_block_size(struct dm_bufio_client *c)
1553 {
1554 	return c->block_size;
1555 }
1556 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1557 
1558 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1559 {
1560 	sector_t s = bdev_nr_sectors(c->bdev);
1561 
1562 	if (s >= c->start)
1563 		s -= c->start;
1564 	else
1565 		s = 0;
1566 	if (likely(c->sectors_per_block_bits >= 0))
1567 		s >>= c->sectors_per_block_bits;
1568 	else
1569 		sector_div(s, c->block_size >> SECTOR_SHIFT);
1570 	return s;
1571 }
1572 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1573 
1574 struct dm_io_client *dm_bufio_get_dm_io_client(struct dm_bufio_client *c)
1575 {
1576 	return c->dm_io;
1577 }
1578 EXPORT_SYMBOL_GPL(dm_bufio_get_dm_io_client);
1579 
1580 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1581 {
1582 	return b->block;
1583 }
1584 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1585 
1586 void *dm_bufio_get_block_data(struct dm_buffer *b)
1587 {
1588 	return b->data;
1589 }
1590 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1591 
1592 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1593 {
1594 	return b + 1;
1595 }
1596 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1597 
1598 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1599 {
1600 	return b->c;
1601 }
1602 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1603 
1604 static void drop_buffers(struct dm_bufio_client *c)
1605 {
1606 	struct dm_buffer *b;
1607 	int i;
1608 	bool warned = false;
1609 
1610 	BUG_ON(dm_bufio_in_request());
1611 
1612 	/*
1613 	 * An optimization so that the buffers are not written one-by-one.
1614 	 */
1615 	dm_bufio_write_dirty_buffers_async(c);
1616 
1617 	dm_bufio_lock(c);
1618 
1619 	while ((b = __get_unclaimed_buffer(c)))
1620 		__free_buffer_wake(b);
1621 
1622 	for (i = 0; i < LIST_SIZE; i++)
1623 		list_for_each_entry(b, &c->lru[i], lru_list) {
1624 			WARN_ON(!warned);
1625 			warned = true;
1626 			DMERR("leaked buffer %llx, hold count %u, list %d",
1627 			      (unsigned long long)b->block, b->hold_count, i);
1628 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1629 			stack_trace_print(b->stack_entries, b->stack_len, 1);
1630 			/* mark unclaimed to avoid BUG_ON below */
1631 			b->hold_count = 0;
1632 #endif
1633 		}
1634 
1635 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1636 	while ((b = __get_unclaimed_buffer(c)))
1637 		__free_buffer_wake(b);
1638 #endif
1639 
1640 	for (i = 0; i < LIST_SIZE; i++)
1641 		BUG_ON(!list_empty(&c->lru[i]));
1642 
1643 	dm_bufio_unlock(c);
1644 }
1645 
1646 /*
1647  * We may not be able to evict this buffer if IO pending or the client
1648  * is still using it.  Caller is expected to know buffer is too old.
1649  *
1650  * And if GFP_NOFS is used, we must not do any I/O because we hold
1651  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1652  * rerouted to different bufio client.
1653  */
1654 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1655 {
1656 	if (!(gfp & __GFP_FS) ||
1657 	    (static_branch_unlikely(&no_sleep_enabled) && b->c->no_sleep)) {
1658 		if (test_bit_acquire(B_READING, &b->state) ||
1659 		    test_bit(B_WRITING, &b->state) ||
1660 		    test_bit(B_DIRTY, &b->state))
1661 			return false;
1662 	}
1663 
1664 	if (b->hold_count)
1665 		return false;
1666 
1667 	__make_buffer_clean(b);
1668 	__unlink_buffer(b);
1669 	__free_buffer_wake(b);
1670 
1671 	return true;
1672 }
1673 
1674 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1675 {
1676 	unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1677 
1678 	if (likely(c->sectors_per_block_bits >= 0))
1679 		retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1680 	else
1681 		retain_bytes /= c->block_size;
1682 
1683 	return retain_bytes;
1684 }
1685 
1686 static void __scan(struct dm_bufio_client *c)
1687 {
1688 	int l;
1689 	struct dm_buffer *b, *tmp;
1690 	unsigned long freed = 0;
1691 	unsigned long count = c->n_buffers[LIST_CLEAN] +
1692 			      c->n_buffers[LIST_DIRTY];
1693 	unsigned long retain_target = get_retain_buffers(c);
1694 
1695 	for (l = 0; l < LIST_SIZE; l++) {
1696 		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1697 			if (count - freed <= retain_target)
1698 				atomic_long_set(&c->need_shrink, 0);
1699 			if (!atomic_long_read(&c->need_shrink))
1700 				return;
1701 			if (__try_evict_buffer(b, GFP_KERNEL)) {
1702 				atomic_long_dec(&c->need_shrink);
1703 				freed++;
1704 			}
1705 			cond_resched();
1706 		}
1707 	}
1708 }
1709 
1710 static void shrink_work(struct work_struct *w)
1711 {
1712 	struct dm_bufio_client *c = container_of(w, struct dm_bufio_client, shrink_work);
1713 
1714 	dm_bufio_lock(c);
1715 	__scan(c);
1716 	dm_bufio_unlock(c);
1717 }
1718 
1719 static unsigned long dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1720 {
1721 	struct dm_bufio_client *c;
1722 
1723 	c = container_of(shrink, struct dm_bufio_client, shrinker);
1724 	atomic_long_add(sc->nr_to_scan, &c->need_shrink);
1725 	queue_work(dm_bufio_wq, &c->shrink_work);
1726 
1727 	return sc->nr_to_scan;
1728 }
1729 
1730 static unsigned long dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1731 {
1732 	struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1733 	unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1734 			      READ_ONCE(c->n_buffers[LIST_DIRTY]);
1735 	unsigned long retain_target = get_retain_buffers(c);
1736 	unsigned long queued_for_cleanup = atomic_long_read(&c->need_shrink);
1737 
1738 	if (unlikely(count < retain_target))
1739 		count = 0;
1740 	else
1741 		count -= retain_target;
1742 
1743 	if (unlikely(count < queued_for_cleanup))
1744 		count = 0;
1745 	else
1746 		count -= queued_for_cleanup;
1747 
1748 	return count;
1749 }
1750 
1751 /*
1752  * Create the buffering interface
1753  */
1754 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned int block_size,
1755 					       unsigned int reserved_buffers, unsigned int aux_size,
1756 					       void (*alloc_callback)(struct dm_buffer *),
1757 					       void (*write_callback)(struct dm_buffer *),
1758 					       unsigned int flags)
1759 {
1760 	int r;
1761 	struct dm_bufio_client *c;
1762 	unsigned int i;
1763 	char slab_name[27];
1764 
1765 	if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1766 		DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1767 		r = -EINVAL;
1768 		goto bad_client;
1769 	}
1770 
1771 	c = kzalloc(sizeof(*c), GFP_KERNEL);
1772 	if (!c) {
1773 		r = -ENOMEM;
1774 		goto bad_client;
1775 	}
1776 	c->buffer_tree = RB_ROOT;
1777 
1778 	c->bdev = bdev;
1779 	c->block_size = block_size;
1780 	if (is_power_of_2(block_size))
1781 		c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1782 	else
1783 		c->sectors_per_block_bits = -1;
1784 
1785 	c->alloc_callback = alloc_callback;
1786 	c->write_callback = write_callback;
1787 
1788 	if (flags & DM_BUFIO_CLIENT_NO_SLEEP) {
1789 		c->no_sleep = true;
1790 		static_branch_inc(&no_sleep_enabled);
1791 	}
1792 
1793 	for (i = 0; i < LIST_SIZE; i++) {
1794 		INIT_LIST_HEAD(&c->lru[i]);
1795 		c->n_buffers[i] = 0;
1796 	}
1797 
1798 	mutex_init(&c->lock);
1799 	spin_lock_init(&c->spinlock);
1800 	INIT_LIST_HEAD(&c->reserved_buffers);
1801 	c->need_reserved_buffers = reserved_buffers;
1802 
1803 	dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1804 
1805 	init_waitqueue_head(&c->free_buffer_wait);
1806 	c->async_write_error = 0;
1807 
1808 	c->dm_io = dm_io_client_create();
1809 	if (IS_ERR(c->dm_io)) {
1810 		r = PTR_ERR(c->dm_io);
1811 		goto bad_dm_io;
1812 	}
1813 
1814 	if (block_size <= KMALLOC_MAX_SIZE &&
1815 	    (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1816 		unsigned int align = min(1U << __ffs(block_size), (unsigned int)PAGE_SIZE);
1817 
1818 		snprintf(slab_name, sizeof(slab_name), "dm_bufio_cache-%u", block_size);
1819 		c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1820 						  SLAB_RECLAIM_ACCOUNT, NULL);
1821 		if (!c->slab_cache) {
1822 			r = -ENOMEM;
1823 			goto bad;
1824 		}
1825 	}
1826 	if (aux_size)
1827 		snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer-%u", aux_size);
1828 	else
1829 		snprintf(slab_name, sizeof(slab_name), "dm_bufio_buffer");
1830 	c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1831 					   0, SLAB_RECLAIM_ACCOUNT, NULL);
1832 	if (!c->slab_buffer) {
1833 		r = -ENOMEM;
1834 		goto bad;
1835 	}
1836 
1837 	while (c->need_reserved_buffers) {
1838 		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1839 
1840 		if (!b) {
1841 			r = -ENOMEM;
1842 			goto bad;
1843 		}
1844 		__free_buffer_wake(b);
1845 	}
1846 
1847 	INIT_WORK(&c->shrink_work, shrink_work);
1848 	atomic_long_set(&c->need_shrink, 0);
1849 
1850 	c->shrinker.count_objects = dm_bufio_shrink_count;
1851 	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1852 	c->shrinker.seeks = 1;
1853 	c->shrinker.batch = 0;
1854 	r = register_shrinker(&c->shrinker, "dm-bufio:(%u:%u)",
1855 			      MAJOR(bdev->bd_dev), MINOR(bdev->bd_dev));
1856 	if (r)
1857 		goto bad;
1858 
1859 	mutex_lock(&dm_bufio_clients_lock);
1860 	dm_bufio_client_count++;
1861 	list_add(&c->client_list, &dm_bufio_all_clients);
1862 	__cache_size_refresh();
1863 	mutex_unlock(&dm_bufio_clients_lock);
1864 
1865 	return c;
1866 
1867 bad:
1868 	while (!list_empty(&c->reserved_buffers)) {
1869 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1870 						 struct dm_buffer, lru_list);
1871 		list_del(&b->lru_list);
1872 		free_buffer(b);
1873 	}
1874 	kmem_cache_destroy(c->slab_cache);
1875 	kmem_cache_destroy(c->slab_buffer);
1876 	dm_io_client_destroy(c->dm_io);
1877 bad_dm_io:
1878 	mutex_destroy(&c->lock);
1879 	if (c->no_sleep)
1880 		static_branch_dec(&no_sleep_enabled);
1881 	kfree(c);
1882 bad_client:
1883 	return ERR_PTR(r);
1884 }
1885 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1886 
1887 /*
1888  * Free the buffering interface.
1889  * It is required that there are no references on any buffers.
1890  */
1891 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1892 {
1893 	unsigned int i;
1894 
1895 	drop_buffers(c);
1896 
1897 	unregister_shrinker(&c->shrinker);
1898 	flush_work(&c->shrink_work);
1899 
1900 	mutex_lock(&dm_bufio_clients_lock);
1901 
1902 	list_del(&c->client_list);
1903 	dm_bufio_client_count--;
1904 	__cache_size_refresh();
1905 
1906 	mutex_unlock(&dm_bufio_clients_lock);
1907 
1908 	BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1909 	BUG_ON(c->need_reserved_buffers);
1910 
1911 	while (!list_empty(&c->reserved_buffers)) {
1912 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1913 						 struct dm_buffer, lru_list);
1914 		list_del(&b->lru_list);
1915 		free_buffer(b);
1916 	}
1917 
1918 	for (i = 0; i < LIST_SIZE; i++)
1919 		if (c->n_buffers[i])
1920 			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1921 
1922 	for (i = 0; i < LIST_SIZE; i++)
1923 		BUG_ON(c->n_buffers[i]);
1924 
1925 	kmem_cache_destroy(c->slab_cache);
1926 	kmem_cache_destroy(c->slab_buffer);
1927 	dm_io_client_destroy(c->dm_io);
1928 	mutex_destroy(&c->lock);
1929 	if (c->no_sleep)
1930 		static_branch_dec(&no_sleep_enabled);
1931 	kfree(c);
1932 }
1933 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1934 
1935 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1936 {
1937 	c->start = start;
1938 }
1939 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1940 
1941 static unsigned int get_max_age_hz(void)
1942 {
1943 	unsigned int max_age = READ_ONCE(dm_bufio_max_age);
1944 
1945 	if (max_age > UINT_MAX / HZ)
1946 		max_age = UINT_MAX / HZ;
1947 
1948 	return max_age * HZ;
1949 }
1950 
1951 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1952 {
1953 	return time_after_eq(jiffies, b->last_accessed + age_hz);
1954 }
1955 
1956 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1957 {
1958 	struct dm_buffer *b, *tmp;
1959 	unsigned long retain_target = get_retain_buffers(c);
1960 	unsigned long count;
1961 	LIST_HEAD(write_list);
1962 
1963 	dm_bufio_lock(c);
1964 
1965 	__check_watermark(c, &write_list);
1966 	if (unlikely(!list_empty(&write_list))) {
1967 		dm_bufio_unlock(c);
1968 		__flush_write_list(&write_list);
1969 		dm_bufio_lock(c);
1970 	}
1971 
1972 	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1973 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1974 		if (count <= retain_target)
1975 			break;
1976 
1977 		if (!older_than(b, age_hz))
1978 			break;
1979 
1980 		if (__try_evict_buffer(b, 0))
1981 			count--;
1982 
1983 		cond_resched();
1984 	}
1985 
1986 	dm_bufio_unlock(c);
1987 }
1988 
1989 static void do_global_cleanup(struct work_struct *w)
1990 {
1991 	struct dm_bufio_client *locked_client = NULL;
1992 	struct dm_bufio_client *current_client;
1993 	struct dm_buffer *b;
1994 	unsigned int spinlock_hold_count;
1995 	unsigned long threshold = dm_bufio_cache_size -
1996 		dm_bufio_cache_size / DM_BUFIO_LOW_WATERMARK_RATIO;
1997 	unsigned long loops = global_num * 2;
1998 
1999 	mutex_lock(&dm_bufio_clients_lock);
2000 
2001 	while (1) {
2002 		cond_resched();
2003 
2004 		spin_lock(&global_spinlock);
2005 		if (unlikely(dm_bufio_current_allocated <= threshold))
2006 			break;
2007 
2008 		spinlock_hold_count = 0;
2009 get_next:
2010 		if (!loops--)
2011 			break;
2012 		if (unlikely(list_empty(&global_queue)))
2013 			break;
2014 		b = list_entry(global_queue.prev, struct dm_buffer, global_list);
2015 
2016 		if (b->accessed) {
2017 			b->accessed = 0;
2018 			list_move(&b->global_list, &global_queue);
2019 			if (likely(++spinlock_hold_count < 16))
2020 				goto get_next;
2021 			spin_unlock(&global_spinlock);
2022 			continue;
2023 		}
2024 
2025 		current_client = b->c;
2026 		if (unlikely(current_client != locked_client)) {
2027 			if (locked_client)
2028 				dm_bufio_unlock(locked_client);
2029 
2030 			if (!dm_bufio_trylock(current_client)) {
2031 				spin_unlock(&global_spinlock);
2032 				dm_bufio_lock(current_client);
2033 				locked_client = current_client;
2034 				continue;
2035 			}
2036 
2037 			locked_client = current_client;
2038 		}
2039 
2040 		spin_unlock(&global_spinlock);
2041 
2042 		if (unlikely(!__try_evict_buffer(b, GFP_KERNEL))) {
2043 			spin_lock(&global_spinlock);
2044 			list_move(&b->global_list, &global_queue);
2045 			spin_unlock(&global_spinlock);
2046 		}
2047 	}
2048 
2049 	spin_unlock(&global_spinlock);
2050 
2051 	if (locked_client)
2052 		dm_bufio_unlock(locked_client);
2053 
2054 	mutex_unlock(&dm_bufio_clients_lock);
2055 }
2056 
2057 static void cleanup_old_buffers(void)
2058 {
2059 	unsigned long max_age_hz = get_max_age_hz();
2060 	struct dm_bufio_client *c;
2061 
2062 	mutex_lock(&dm_bufio_clients_lock);
2063 
2064 	__cache_size_refresh();
2065 
2066 	list_for_each_entry(c, &dm_bufio_all_clients, client_list)
2067 		__evict_old_buffers(c, max_age_hz);
2068 
2069 	mutex_unlock(&dm_bufio_clients_lock);
2070 }
2071 
2072 static void work_fn(struct work_struct *w)
2073 {
2074 	cleanup_old_buffers();
2075 
2076 	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2077 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2078 }
2079 
2080 /*
2081  *--------------------------------------------------------------
2082  * Module setup
2083  *--------------------------------------------------------------
2084  */
2085 
2086 /*
2087  * This is called only once for the whole dm_bufio module.
2088  * It initializes memory limit.
2089  */
2090 static int __init dm_bufio_init(void)
2091 {
2092 	__u64 mem;
2093 
2094 	dm_bufio_allocated_kmem_cache = 0;
2095 	dm_bufio_allocated_get_free_pages = 0;
2096 	dm_bufio_allocated_vmalloc = 0;
2097 	dm_bufio_current_allocated = 0;
2098 
2099 	mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
2100 			       DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
2101 
2102 	if (mem > ULONG_MAX)
2103 		mem = ULONG_MAX;
2104 
2105 #ifdef CONFIG_MMU
2106 	if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
2107 		mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
2108 #endif
2109 
2110 	dm_bufio_default_cache_size = mem;
2111 
2112 	mutex_lock(&dm_bufio_clients_lock);
2113 	__cache_size_refresh();
2114 	mutex_unlock(&dm_bufio_clients_lock);
2115 
2116 	dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
2117 	if (!dm_bufio_wq)
2118 		return -ENOMEM;
2119 
2120 	INIT_DELAYED_WORK(&dm_bufio_cleanup_old_work, work_fn);
2121 	INIT_WORK(&dm_bufio_replacement_work, do_global_cleanup);
2122 	queue_delayed_work(dm_bufio_wq, &dm_bufio_cleanup_old_work,
2123 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
2124 
2125 	return 0;
2126 }
2127 
2128 /*
2129  * This is called once when unloading the dm_bufio module.
2130  */
2131 static void __exit dm_bufio_exit(void)
2132 {
2133 	int bug = 0;
2134 
2135 	cancel_delayed_work_sync(&dm_bufio_cleanup_old_work);
2136 	destroy_workqueue(dm_bufio_wq);
2137 
2138 	if (dm_bufio_client_count) {
2139 		DMCRIT("%s: dm_bufio_client_count leaked: %d",
2140 			__func__, dm_bufio_client_count);
2141 		bug = 1;
2142 	}
2143 
2144 	if (dm_bufio_current_allocated) {
2145 		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
2146 			__func__, dm_bufio_current_allocated);
2147 		bug = 1;
2148 	}
2149 
2150 	if (dm_bufio_allocated_get_free_pages) {
2151 		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
2152 		       __func__, dm_bufio_allocated_get_free_pages);
2153 		bug = 1;
2154 	}
2155 
2156 	if (dm_bufio_allocated_vmalloc) {
2157 		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
2158 		       __func__, dm_bufio_allocated_vmalloc);
2159 		bug = 1;
2160 	}
2161 
2162 	BUG_ON(bug);
2163 }
2164 
2165 module_init(dm_bufio_init)
2166 module_exit(dm_bufio_exit)
2167 
2168 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, 0644);
2169 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
2170 
2171 module_param_named(max_age_seconds, dm_bufio_max_age, uint, 0644);
2172 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
2173 
2174 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, 0644);
2175 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
2176 
2177 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, 0644);
2178 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
2179 
2180 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, 0444);
2181 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
2182 
2183 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, 0444);
2184 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
2185 
2186 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, 0444);
2187 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
2188 
2189 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, 0444);
2190 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
2191 
2192 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
2193 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
2194 MODULE_LICENSE("GPL");
2195