xref: /openbmc/linux/drivers/md/dm-bufio.c (revision 5e012745)
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_PERCENT	75
37 
38 /*
39  * Check buffer ages in this interval (seconds)
40  */
41 #define DM_BUFIO_WORK_TIMER_SECS	30
42 
43 /*
44  * Free buffers when they are older than this (seconds)
45  */
46 #define DM_BUFIO_DEFAULT_AGE_SECS	300
47 
48 /*
49  * The nr of bytes of cached data to keep around.
50  */
51 #define DM_BUFIO_DEFAULT_RETAIN_BYTES   (256 * 1024)
52 
53 /*
54  * Align buffer writes to this boundary.
55  * Tests show that SSDs have the highest IOPS when using 4k writes.
56  */
57 #define DM_BUFIO_WRITE_ALIGN		4096
58 
59 /*
60  * dm_buffer->list_mode
61  */
62 #define LIST_CLEAN	0
63 #define LIST_DIRTY	1
64 #define LIST_SIZE	2
65 
66 /*
67  * Linking of buffers:
68  *	All buffers are linked to buffer_tree with their node field.
69  *
70  *	Clean buffers that are not being written (B_WRITING not set)
71  *	are linked to lru[LIST_CLEAN] with their lru_list field.
72  *
73  *	Dirty and clean buffers that are being written are linked to
74  *	lru[LIST_DIRTY] with their lru_list field. When the write
75  *	finishes, the buffer cannot be relinked immediately (because we
76  *	are in an interrupt context and relinking requires process
77  *	context), so some clean-not-writing buffers can be held on
78  *	dirty_lru too.  They are later added to lru in the process
79  *	context.
80  */
81 struct dm_bufio_client {
82 	struct mutex lock;
83 
84 	struct list_head lru[LIST_SIZE];
85 	unsigned long n_buffers[LIST_SIZE];
86 
87 	struct block_device *bdev;
88 	unsigned block_size;
89 	s8 sectors_per_block_bits;
90 	void (*alloc_callback)(struct dm_buffer *);
91 	void (*write_callback)(struct dm_buffer *);
92 
93 	struct kmem_cache *slab_buffer;
94 	struct kmem_cache *slab_cache;
95 	struct dm_io_client *dm_io;
96 
97 	struct list_head reserved_buffers;
98 	unsigned need_reserved_buffers;
99 
100 	unsigned minimum_buffers;
101 
102 	struct rb_root buffer_tree;
103 	wait_queue_head_t free_buffer_wait;
104 
105 	sector_t start;
106 
107 	int async_write_error;
108 
109 	struct list_head client_list;
110 	struct shrinker shrinker;
111 };
112 
113 /*
114  * Buffer state bits.
115  */
116 #define B_READING	0
117 #define B_WRITING	1
118 #define B_DIRTY		2
119 
120 /*
121  * Describes how the block was allocated:
122  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
123  * See the comment at alloc_buffer_data.
124  */
125 enum data_mode {
126 	DATA_MODE_SLAB = 0,
127 	DATA_MODE_GET_FREE_PAGES = 1,
128 	DATA_MODE_VMALLOC = 2,
129 	DATA_MODE_LIMIT = 3
130 };
131 
132 struct dm_buffer {
133 	struct rb_node node;
134 	struct list_head lru_list;
135 	sector_t block;
136 	void *data;
137 	unsigned char data_mode;		/* DATA_MODE_* */
138 	unsigned char list_mode;		/* LIST_* */
139 	blk_status_t read_error;
140 	blk_status_t write_error;
141 	unsigned hold_count;
142 	unsigned long state;
143 	unsigned long last_accessed;
144 	unsigned dirty_start;
145 	unsigned dirty_end;
146 	unsigned write_start;
147 	unsigned write_end;
148 	struct dm_bufio_client *c;
149 	struct list_head write_list;
150 	void (*end_io)(struct dm_buffer *, blk_status_t);
151 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
152 #define MAX_STACK 10
153 	unsigned int stack_len;
154 	unsigned long stack_entries[MAX_STACK];
155 #endif
156 };
157 
158 /*----------------------------------------------------------------*/
159 
160 #define dm_bufio_in_request()	(!!current->bio_list)
161 
162 static void dm_bufio_lock(struct dm_bufio_client *c)
163 {
164 	mutex_lock_nested(&c->lock, dm_bufio_in_request());
165 }
166 
167 static int dm_bufio_trylock(struct dm_bufio_client *c)
168 {
169 	return mutex_trylock(&c->lock);
170 }
171 
172 static void dm_bufio_unlock(struct dm_bufio_client *c)
173 {
174 	mutex_unlock(&c->lock);
175 }
176 
177 /*----------------------------------------------------------------*/
178 
179 /*
180  * Default cache size: available memory divided by the ratio.
181  */
182 static unsigned long dm_bufio_default_cache_size;
183 
184 /*
185  * Total cache size set by the user.
186  */
187 static unsigned long dm_bufio_cache_size;
188 
189 /*
190  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
191  * at any time.  If it disagrees, the user has changed cache size.
192  */
193 static unsigned long dm_bufio_cache_size_latch;
194 
195 static DEFINE_SPINLOCK(param_spinlock);
196 
197 /*
198  * Buffers are freed after this timeout
199  */
200 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
201 static unsigned long dm_bufio_retain_bytes = DM_BUFIO_DEFAULT_RETAIN_BYTES;
202 
203 static unsigned long dm_bufio_peak_allocated;
204 static unsigned long dm_bufio_allocated_kmem_cache;
205 static unsigned long dm_bufio_allocated_get_free_pages;
206 static unsigned long dm_bufio_allocated_vmalloc;
207 static unsigned long dm_bufio_current_allocated;
208 
209 /*----------------------------------------------------------------*/
210 
211 /*
212  * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
213  */
214 static unsigned long dm_bufio_cache_size_per_client;
215 
216 /*
217  * The current number of clients.
218  */
219 static int dm_bufio_client_count;
220 
221 /*
222  * The list of all clients.
223  */
224 static LIST_HEAD(dm_bufio_all_clients);
225 
226 /*
227  * This mutex protects dm_bufio_cache_size_latch,
228  * dm_bufio_cache_size_per_client and dm_bufio_client_count
229  */
230 static DEFINE_MUTEX(dm_bufio_clients_lock);
231 
232 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
233 static void buffer_record_stack(struct dm_buffer *b)
234 {
235 	b->stack_len = stack_trace_save(b->stack_entries, MAX_STACK, 2);
236 }
237 #endif
238 
239 /*----------------------------------------------------------------
240  * A red/black tree acts as an index for all the buffers.
241  *--------------------------------------------------------------*/
242 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
243 {
244 	struct rb_node *n = c->buffer_tree.rb_node;
245 	struct dm_buffer *b;
246 
247 	while (n) {
248 		b = container_of(n, struct dm_buffer, node);
249 
250 		if (b->block == block)
251 			return b;
252 
253 		n = (b->block < block) ? n->rb_left : n->rb_right;
254 	}
255 
256 	return NULL;
257 }
258 
259 static void __insert(struct dm_bufio_client *c, struct dm_buffer *b)
260 {
261 	struct rb_node **new = &c->buffer_tree.rb_node, *parent = NULL;
262 	struct dm_buffer *found;
263 
264 	while (*new) {
265 		found = container_of(*new, struct dm_buffer, node);
266 
267 		if (found->block == b->block) {
268 			BUG_ON(found != b);
269 			return;
270 		}
271 
272 		parent = *new;
273 		new = (found->block < b->block) ?
274 			&((*new)->rb_left) : &((*new)->rb_right);
275 	}
276 
277 	rb_link_node(&b->node, parent, new);
278 	rb_insert_color(&b->node, &c->buffer_tree);
279 }
280 
281 static void __remove(struct dm_bufio_client *c, struct dm_buffer *b)
282 {
283 	rb_erase(&b->node, &c->buffer_tree);
284 }
285 
286 /*----------------------------------------------------------------*/
287 
288 static void adjust_total_allocated(unsigned char data_mode, long diff)
289 {
290 	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
291 		&dm_bufio_allocated_kmem_cache,
292 		&dm_bufio_allocated_get_free_pages,
293 		&dm_bufio_allocated_vmalloc,
294 	};
295 
296 	spin_lock(&param_spinlock);
297 
298 	*class_ptr[data_mode] += diff;
299 
300 	dm_bufio_current_allocated += diff;
301 
302 	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
303 		dm_bufio_peak_allocated = dm_bufio_current_allocated;
304 
305 	spin_unlock(&param_spinlock);
306 }
307 
308 /*
309  * Change the number of clients and recalculate per-client limit.
310  */
311 static void __cache_size_refresh(void)
312 {
313 	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
314 	BUG_ON(dm_bufio_client_count < 0);
315 
316 	dm_bufio_cache_size_latch = READ_ONCE(dm_bufio_cache_size);
317 
318 	/*
319 	 * Use default if set to 0 and report the actual cache size used.
320 	 */
321 	if (!dm_bufio_cache_size_latch) {
322 		(void)cmpxchg(&dm_bufio_cache_size, 0,
323 			      dm_bufio_default_cache_size);
324 		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
325 	}
326 
327 	dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
328 					 (dm_bufio_client_count ? : 1);
329 }
330 
331 /*
332  * Allocating buffer data.
333  *
334  * Small buffers are allocated with kmem_cache, to use space optimally.
335  *
336  * For large buffers, we choose between get_free_pages and vmalloc.
337  * Each has advantages and disadvantages.
338  *
339  * __get_free_pages can randomly fail if the memory is fragmented.
340  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
341  * as low as 128M) so using it for caching is not appropriate.
342  *
343  * If the allocation may fail we use __get_free_pages. Memory fragmentation
344  * won't have a fatal effect here, but it just causes flushes of some other
345  * buffers and more I/O will be performed. Don't use __get_free_pages if it
346  * always fails (i.e. order >= MAX_ORDER).
347  *
348  * If the allocation shouldn't fail we use __vmalloc. This is only for the
349  * initial reserve allocation, so there's no risk of wasting all vmalloc
350  * space.
351  */
352 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
353 			       unsigned char *data_mode)
354 {
355 	if (unlikely(c->slab_cache != NULL)) {
356 		*data_mode = DATA_MODE_SLAB;
357 		return kmem_cache_alloc(c->slab_cache, gfp_mask);
358 	}
359 
360 	if (c->block_size <= KMALLOC_MAX_SIZE &&
361 	    gfp_mask & __GFP_NORETRY) {
362 		*data_mode = DATA_MODE_GET_FREE_PAGES;
363 		return (void *)__get_free_pages(gfp_mask,
364 						c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
365 	}
366 
367 	*data_mode = DATA_MODE_VMALLOC;
368 
369 	/*
370 	 * __vmalloc allocates the data pages and auxiliary structures with
371 	 * gfp_flags that were specified, but pagetables are always allocated
372 	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
373 	 *
374 	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
375 	 * all allocations done by this process (including pagetables) are done
376 	 * as if GFP_NOIO was specified.
377 	 */
378 	if (gfp_mask & __GFP_NORETRY) {
379 		unsigned noio_flag = memalloc_noio_save();
380 		void *ptr = __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
381 
382 		memalloc_noio_restore(noio_flag);
383 		return ptr;
384 	}
385 
386 	return __vmalloc(c->block_size, gfp_mask, PAGE_KERNEL);
387 }
388 
389 /*
390  * Free buffer's data.
391  */
392 static void free_buffer_data(struct dm_bufio_client *c,
393 			     void *data, unsigned char data_mode)
394 {
395 	switch (data_mode) {
396 	case DATA_MODE_SLAB:
397 		kmem_cache_free(c->slab_cache, data);
398 		break;
399 
400 	case DATA_MODE_GET_FREE_PAGES:
401 		free_pages((unsigned long)data,
402 			   c->sectors_per_block_bits - (PAGE_SHIFT - SECTOR_SHIFT));
403 		break;
404 
405 	case DATA_MODE_VMALLOC:
406 		vfree(data);
407 		break;
408 
409 	default:
410 		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
411 		       data_mode);
412 		BUG();
413 	}
414 }
415 
416 /*
417  * Allocate buffer and its data.
418  */
419 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
420 {
421 	struct dm_buffer *b = kmem_cache_alloc(c->slab_buffer, gfp_mask);
422 
423 	if (!b)
424 		return NULL;
425 
426 	b->c = c;
427 
428 	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
429 	if (!b->data) {
430 		kmem_cache_free(c->slab_buffer, b);
431 		return NULL;
432 	}
433 
434 	adjust_total_allocated(b->data_mode, (long)c->block_size);
435 
436 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
437 	b->stack_len = 0;
438 #endif
439 	return b;
440 }
441 
442 /*
443  * Free buffer and its data.
444  */
445 static void free_buffer(struct dm_buffer *b)
446 {
447 	struct dm_bufio_client *c = b->c;
448 
449 	adjust_total_allocated(b->data_mode, -(long)c->block_size);
450 
451 	free_buffer_data(c, b->data, b->data_mode);
452 	kmem_cache_free(c->slab_buffer, b);
453 }
454 
455 /*
456  * Link buffer to the buffer tree and clean or dirty queue.
457  */
458 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
459 {
460 	struct dm_bufio_client *c = b->c;
461 
462 	c->n_buffers[dirty]++;
463 	b->block = block;
464 	b->list_mode = dirty;
465 	list_add(&b->lru_list, &c->lru[dirty]);
466 	__insert(b->c, b);
467 	b->last_accessed = jiffies;
468 }
469 
470 /*
471  * Unlink buffer from the buffer tree and dirty or clean queue.
472  */
473 static void __unlink_buffer(struct dm_buffer *b)
474 {
475 	struct dm_bufio_client *c = b->c;
476 
477 	BUG_ON(!c->n_buffers[b->list_mode]);
478 
479 	c->n_buffers[b->list_mode]--;
480 	__remove(b->c, b);
481 	list_del(&b->lru_list);
482 }
483 
484 /*
485  * Place the buffer to the head of dirty or clean LRU queue.
486  */
487 static void __relink_lru(struct dm_buffer *b, int dirty)
488 {
489 	struct dm_bufio_client *c = b->c;
490 
491 	BUG_ON(!c->n_buffers[b->list_mode]);
492 
493 	c->n_buffers[b->list_mode]--;
494 	c->n_buffers[dirty]++;
495 	b->list_mode = dirty;
496 	list_move(&b->lru_list, &c->lru[dirty]);
497 	b->last_accessed = jiffies;
498 }
499 
500 /*----------------------------------------------------------------
501  * Submit I/O on the buffer.
502  *
503  * Bio interface is faster but it has some problems:
504  *	the vector list is limited (increasing this limit increases
505  *	memory-consumption per buffer, so it is not viable);
506  *
507  *	the memory must be direct-mapped, not vmalloced;
508  *
509  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
510  * it is not vmalloced, try using the bio interface.
511  *
512  * If the buffer is big, if it is vmalloced or if the underlying device
513  * rejects the bio because it is too large, use dm-io layer to do the I/O.
514  * The dm-io layer splits the I/O into multiple requests, avoiding the above
515  * shortcomings.
516  *--------------------------------------------------------------*/
517 
518 /*
519  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
520  * that the request was handled directly with bio interface.
521  */
522 static void dmio_complete(unsigned long error, void *context)
523 {
524 	struct dm_buffer *b = context;
525 
526 	b->end_io(b, unlikely(error != 0) ? BLK_STS_IOERR : 0);
527 }
528 
529 static void use_dmio(struct dm_buffer *b, int rw, sector_t sector,
530 		     unsigned n_sectors, unsigned offset)
531 {
532 	int r;
533 	struct dm_io_request io_req = {
534 		.bi_op = rw,
535 		.bi_op_flags = 0,
536 		.notify.fn = dmio_complete,
537 		.notify.context = b,
538 		.client = b->c->dm_io,
539 	};
540 	struct dm_io_region region = {
541 		.bdev = b->c->bdev,
542 		.sector = sector,
543 		.count = n_sectors,
544 	};
545 
546 	if (b->data_mode != DATA_MODE_VMALLOC) {
547 		io_req.mem.type = DM_IO_KMEM;
548 		io_req.mem.ptr.addr = (char *)b->data + offset;
549 	} else {
550 		io_req.mem.type = DM_IO_VMA;
551 		io_req.mem.ptr.vma = (char *)b->data + offset;
552 	}
553 
554 	r = dm_io(&io_req, 1, &region, NULL);
555 	if (unlikely(r))
556 		b->end_io(b, errno_to_blk_status(r));
557 }
558 
559 static void bio_complete(struct bio *bio)
560 {
561 	struct dm_buffer *b = bio->bi_private;
562 	blk_status_t status = bio->bi_status;
563 	bio_put(bio);
564 	b->end_io(b, status);
565 }
566 
567 static void use_bio(struct dm_buffer *b, int rw, sector_t sector,
568 		    unsigned n_sectors, unsigned offset)
569 {
570 	struct bio *bio;
571 	char *ptr;
572 	unsigned vec_size, len;
573 
574 	vec_size = b->c->block_size >> PAGE_SHIFT;
575 	if (unlikely(b->c->sectors_per_block_bits < PAGE_SHIFT - SECTOR_SHIFT))
576 		vec_size += 2;
577 
578 	bio = bio_kmalloc(GFP_NOWAIT | __GFP_NORETRY | __GFP_NOWARN, vec_size);
579 	if (!bio) {
580 dmio:
581 		use_dmio(b, rw, sector, n_sectors, offset);
582 		return;
583 	}
584 
585 	bio->bi_iter.bi_sector = sector;
586 	bio_set_dev(bio, b->c->bdev);
587 	bio_set_op_attrs(bio, rw, 0);
588 	bio->bi_end_io = bio_complete;
589 	bio->bi_private = b;
590 
591 	ptr = (char *)b->data + offset;
592 	len = n_sectors << SECTOR_SHIFT;
593 
594 	do {
595 		unsigned this_step = min((unsigned)(PAGE_SIZE - offset_in_page(ptr)), len);
596 		if (!bio_add_page(bio, virt_to_page(ptr), this_step,
597 				  offset_in_page(ptr))) {
598 			bio_put(bio);
599 			goto dmio;
600 		}
601 
602 		len -= this_step;
603 		ptr += this_step;
604 	} while (len > 0);
605 
606 	submit_bio(bio);
607 }
608 
609 static void submit_io(struct dm_buffer *b, int rw, void (*end_io)(struct dm_buffer *, blk_status_t))
610 {
611 	unsigned n_sectors;
612 	sector_t sector;
613 	unsigned offset, end;
614 
615 	b->end_io = end_io;
616 
617 	if (likely(b->c->sectors_per_block_bits >= 0))
618 		sector = b->block << b->c->sectors_per_block_bits;
619 	else
620 		sector = b->block * (b->c->block_size >> SECTOR_SHIFT);
621 	sector += b->c->start;
622 
623 	if (rw != REQ_OP_WRITE) {
624 		n_sectors = b->c->block_size >> SECTOR_SHIFT;
625 		offset = 0;
626 	} else {
627 		if (b->c->write_callback)
628 			b->c->write_callback(b);
629 		offset = b->write_start;
630 		end = b->write_end;
631 		offset &= -DM_BUFIO_WRITE_ALIGN;
632 		end += DM_BUFIO_WRITE_ALIGN - 1;
633 		end &= -DM_BUFIO_WRITE_ALIGN;
634 		if (unlikely(end > b->c->block_size))
635 			end = b->c->block_size;
636 
637 		sector += offset >> SECTOR_SHIFT;
638 		n_sectors = (end - offset) >> SECTOR_SHIFT;
639 	}
640 
641 	if (b->data_mode != DATA_MODE_VMALLOC)
642 		use_bio(b, rw, sector, n_sectors, offset);
643 	else
644 		use_dmio(b, rw, sector, n_sectors, offset);
645 }
646 
647 /*----------------------------------------------------------------
648  * Writing dirty buffers
649  *--------------------------------------------------------------*/
650 
651 /*
652  * The endio routine for write.
653  *
654  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
655  * it.
656  */
657 static void write_endio(struct dm_buffer *b, blk_status_t status)
658 {
659 	b->write_error = status;
660 	if (unlikely(status)) {
661 		struct dm_bufio_client *c = b->c;
662 
663 		(void)cmpxchg(&c->async_write_error, 0,
664 				blk_status_to_errno(status));
665 	}
666 
667 	BUG_ON(!test_bit(B_WRITING, &b->state));
668 
669 	smp_mb__before_atomic();
670 	clear_bit(B_WRITING, &b->state);
671 	smp_mb__after_atomic();
672 
673 	wake_up_bit(&b->state, B_WRITING);
674 }
675 
676 /*
677  * Initiate a write on a dirty buffer, but don't wait for it.
678  *
679  * - If the buffer is not dirty, exit.
680  * - If there some previous write going on, wait for it to finish (we can't
681  *   have two writes on the same buffer simultaneously).
682  * - Submit our write and don't wait on it. We set B_WRITING indicating
683  *   that there is a write in progress.
684  */
685 static void __write_dirty_buffer(struct dm_buffer *b,
686 				 struct list_head *write_list)
687 {
688 	if (!test_bit(B_DIRTY, &b->state))
689 		return;
690 
691 	clear_bit(B_DIRTY, &b->state);
692 	wait_on_bit_lock_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
693 
694 	b->write_start = b->dirty_start;
695 	b->write_end = b->dirty_end;
696 
697 	if (!write_list)
698 		submit_io(b, REQ_OP_WRITE, write_endio);
699 	else
700 		list_add_tail(&b->write_list, write_list);
701 }
702 
703 static void __flush_write_list(struct list_head *write_list)
704 {
705 	struct blk_plug plug;
706 	blk_start_plug(&plug);
707 	while (!list_empty(write_list)) {
708 		struct dm_buffer *b =
709 			list_entry(write_list->next, struct dm_buffer, write_list);
710 		list_del(&b->write_list);
711 		submit_io(b, REQ_OP_WRITE, write_endio);
712 		cond_resched();
713 	}
714 	blk_finish_plug(&plug);
715 }
716 
717 /*
718  * Wait until any activity on the buffer finishes.  Possibly write the
719  * buffer if it is dirty.  When this function finishes, there is no I/O
720  * running on the buffer and the buffer is not dirty.
721  */
722 static void __make_buffer_clean(struct dm_buffer *b)
723 {
724 	BUG_ON(b->hold_count);
725 
726 	if (!b->state)	/* fast case */
727 		return;
728 
729 	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
730 	__write_dirty_buffer(b, NULL);
731 	wait_on_bit_io(&b->state, B_WRITING, TASK_UNINTERRUPTIBLE);
732 }
733 
734 /*
735  * Find some buffer that is not held by anybody, clean it, unlink it and
736  * return it.
737  */
738 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
739 {
740 	struct dm_buffer *b;
741 
742 	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
743 		BUG_ON(test_bit(B_WRITING, &b->state));
744 		BUG_ON(test_bit(B_DIRTY, &b->state));
745 
746 		if (!b->hold_count) {
747 			__make_buffer_clean(b);
748 			__unlink_buffer(b);
749 			return b;
750 		}
751 		cond_resched();
752 	}
753 
754 	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
755 		BUG_ON(test_bit(B_READING, &b->state));
756 
757 		if (!b->hold_count) {
758 			__make_buffer_clean(b);
759 			__unlink_buffer(b);
760 			return b;
761 		}
762 		cond_resched();
763 	}
764 
765 	return NULL;
766 }
767 
768 /*
769  * Wait until some other threads free some buffer or release hold count on
770  * some buffer.
771  *
772  * This function is entered with c->lock held, drops it and regains it
773  * before exiting.
774  */
775 static void __wait_for_free_buffer(struct dm_bufio_client *c)
776 {
777 	DECLARE_WAITQUEUE(wait, current);
778 
779 	add_wait_queue(&c->free_buffer_wait, &wait);
780 	set_current_state(TASK_UNINTERRUPTIBLE);
781 	dm_bufio_unlock(c);
782 
783 	io_schedule();
784 
785 	remove_wait_queue(&c->free_buffer_wait, &wait);
786 
787 	dm_bufio_lock(c);
788 }
789 
790 enum new_flag {
791 	NF_FRESH = 0,
792 	NF_READ = 1,
793 	NF_GET = 2,
794 	NF_PREFETCH = 3
795 };
796 
797 /*
798  * Allocate a new buffer. If the allocation is not possible, wait until
799  * some other thread frees a buffer.
800  *
801  * May drop the lock and regain it.
802  */
803 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
804 {
805 	struct dm_buffer *b;
806 	bool tried_noio_alloc = false;
807 
808 	/*
809 	 * dm-bufio is resistant to allocation failures (it just keeps
810 	 * one buffer reserved in cases all the allocations fail).
811 	 * So set flags to not try too hard:
812 	 *	GFP_NOWAIT: don't wait; if we need to sleep we'll release our
813 	 *		    mutex and wait ourselves.
814 	 *	__GFP_NORETRY: don't retry and rather return failure
815 	 *	__GFP_NOMEMALLOC: don't use emergency reserves
816 	 *	__GFP_NOWARN: don't print a warning in case of failure
817 	 *
818 	 * For debugging, if we set the cache size to 1, no new buffers will
819 	 * be allocated.
820 	 */
821 	while (1) {
822 		if (dm_bufio_cache_size_latch != 1) {
823 			b = alloc_buffer(c, GFP_NOWAIT | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
824 			if (b)
825 				return b;
826 		}
827 
828 		if (nf == NF_PREFETCH)
829 			return NULL;
830 
831 		if (dm_bufio_cache_size_latch != 1 && !tried_noio_alloc) {
832 			dm_bufio_unlock(c);
833 			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
834 			dm_bufio_lock(c);
835 			if (b)
836 				return b;
837 			tried_noio_alloc = true;
838 		}
839 
840 		if (!list_empty(&c->reserved_buffers)) {
841 			b = list_entry(c->reserved_buffers.next,
842 				       struct dm_buffer, lru_list);
843 			list_del(&b->lru_list);
844 			c->need_reserved_buffers++;
845 
846 			return b;
847 		}
848 
849 		b = __get_unclaimed_buffer(c);
850 		if (b)
851 			return b;
852 
853 		__wait_for_free_buffer(c);
854 	}
855 }
856 
857 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
858 {
859 	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
860 
861 	if (!b)
862 		return NULL;
863 
864 	if (c->alloc_callback)
865 		c->alloc_callback(b);
866 
867 	return b;
868 }
869 
870 /*
871  * Free a buffer and wake other threads waiting for free buffers.
872  */
873 static void __free_buffer_wake(struct dm_buffer *b)
874 {
875 	struct dm_bufio_client *c = b->c;
876 
877 	if (!c->need_reserved_buffers)
878 		free_buffer(b);
879 	else {
880 		list_add(&b->lru_list, &c->reserved_buffers);
881 		c->need_reserved_buffers--;
882 	}
883 
884 	wake_up(&c->free_buffer_wait);
885 }
886 
887 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
888 					struct list_head *write_list)
889 {
890 	struct dm_buffer *b, *tmp;
891 
892 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
893 		BUG_ON(test_bit(B_READING, &b->state));
894 
895 		if (!test_bit(B_DIRTY, &b->state) &&
896 		    !test_bit(B_WRITING, &b->state)) {
897 			__relink_lru(b, LIST_CLEAN);
898 			continue;
899 		}
900 
901 		if (no_wait && test_bit(B_WRITING, &b->state))
902 			return;
903 
904 		__write_dirty_buffer(b, write_list);
905 		cond_resched();
906 	}
907 }
908 
909 /*
910  * Get writeback threshold and buffer limit for a given client.
911  */
912 static void __get_memory_limit(struct dm_bufio_client *c,
913 			       unsigned long *threshold_buffers,
914 			       unsigned long *limit_buffers)
915 {
916 	unsigned long buffers;
917 
918 	if (unlikely(READ_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch)) {
919 		if (mutex_trylock(&dm_bufio_clients_lock)) {
920 			__cache_size_refresh();
921 			mutex_unlock(&dm_bufio_clients_lock);
922 		}
923 	}
924 
925 	buffers = dm_bufio_cache_size_per_client;
926 	if (likely(c->sectors_per_block_bits >= 0))
927 		buffers >>= c->sectors_per_block_bits + SECTOR_SHIFT;
928 	else
929 		buffers /= c->block_size;
930 
931 	if (buffers < c->minimum_buffers)
932 		buffers = c->minimum_buffers;
933 
934 	*limit_buffers = buffers;
935 	*threshold_buffers = mult_frac(buffers,
936 				       DM_BUFIO_WRITEBACK_PERCENT, 100);
937 }
938 
939 /*
940  * Check if we're over watermark.
941  * If we are over threshold_buffers, start freeing buffers.
942  * If we're over "limit_buffers", block until we get under the limit.
943  */
944 static void __check_watermark(struct dm_bufio_client *c,
945 			      struct list_head *write_list)
946 {
947 	unsigned long threshold_buffers, limit_buffers;
948 
949 	__get_memory_limit(c, &threshold_buffers, &limit_buffers);
950 
951 	while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
952 	       limit_buffers) {
953 
954 		struct dm_buffer *b = __get_unclaimed_buffer(c);
955 
956 		if (!b)
957 			return;
958 
959 		__free_buffer_wake(b);
960 		cond_resched();
961 	}
962 
963 	if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
964 		__write_dirty_buffers_async(c, 1, write_list);
965 }
966 
967 /*----------------------------------------------------------------
968  * Getting a buffer
969  *--------------------------------------------------------------*/
970 
971 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
972 				     enum new_flag nf, int *need_submit,
973 				     struct list_head *write_list)
974 {
975 	struct dm_buffer *b, *new_b = NULL;
976 
977 	*need_submit = 0;
978 
979 	b = __find(c, block);
980 	if (b)
981 		goto found_buffer;
982 
983 	if (nf == NF_GET)
984 		return NULL;
985 
986 	new_b = __alloc_buffer_wait(c, nf);
987 	if (!new_b)
988 		return NULL;
989 
990 	/*
991 	 * We've had a period where the mutex was unlocked, so need to
992 	 * recheck the buffer tree.
993 	 */
994 	b = __find(c, block);
995 	if (b) {
996 		__free_buffer_wake(new_b);
997 		goto found_buffer;
998 	}
999 
1000 	__check_watermark(c, write_list);
1001 
1002 	b = new_b;
1003 	b->hold_count = 1;
1004 	b->read_error = 0;
1005 	b->write_error = 0;
1006 	__link_buffer(b, block, LIST_CLEAN);
1007 
1008 	if (nf == NF_FRESH) {
1009 		b->state = 0;
1010 		return b;
1011 	}
1012 
1013 	b->state = 1 << B_READING;
1014 	*need_submit = 1;
1015 
1016 	return b;
1017 
1018 found_buffer:
1019 	if (nf == NF_PREFETCH)
1020 		return NULL;
1021 	/*
1022 	 * Note: it is essential that we don't wait for the buffer to be
1023 	 * read if dm_bufio_get function is used. Both dm_bufio_get and
1024 	 * dm_bufio_prefetch can be used in the driver request routine.
1025 	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
1026 	 * the same buffer, it would deadlock if we waited.
1027 	 */
1028 	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
1029 		return NULL;
1030 
1031 	b->hold_count++;
1032 	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
1033 		     test_bit(B_WRITING, &b->state));
1034 	return b;
1035 }
1036 
1037 /*
1038  * The endio routine for reading: set the error, clear the bit and wake up
1039  * anyone waiting on the buffer.
1040  */
1041 static void read_endio(struct dm_buffer *b, blk_status_t status)
1042 {
1043 	b->read_error = status;
1044 
1045 	BUG_ON(!test_bit(B_READING, &b->state));
1046 
1047 	smp_mb__before_atomic();
1048 	clear_bit(B_READING, &b->state);
1049 	smp_mb__after_atomic();
1050 
1051 	wake_up_bit(&b->state, B_READING);
1052 }
1053 
1054 /*
1055  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1056  * functions is similar except that dm_bufio_new doesn't read the
1057  * buffer from the disk (assuming that the caller overwrites all the data
1058  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1059  */
1060 static void *new_read(struct dm_bufio_client *c, sector_t block,
1061 		      enum new_flag nf, struct dm_buffer **bp)
1062 {
1063 	int need_submit;
1064 	struct dm_buffer *b;
1065 
1066 	LIST_HEAD(write_list);
1067 
1068 	dm_bufio_lock(c);
1069 	b = __bufio_new(c, block, nf, &need_submit, &write_list);
1070 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1071 	if (b && b->hold_count == 1)
1072 		buffer_record_stack(b);
1073 #endif
1074 	dm_bufio_unlock(c);
1075 
1076 	__flush_write_list(&write_list);
1077 
1078 	if (!b)
1079 		return NULL;
1080 
1081 	if (need_submit)
1082 		submit_io(b, REQ_OP_READ, read_endio);
1083 
1084 	wait_on_bit_io(&b->state, B_READING, TASK_UNINTERRUPTIBLE);
1085 
1086 	if (b->read_error) {
1087 		int error = blk_status_to_errno(b->read_error);
1088 
1089 		dm_bufio_release(b);
1090 
1091 		return ERR_PTR(error);
1092 	}
1093 
1094 	*bp = b;
1095 
1096 	return b->data;
1097 }
1098 
1099 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1100 		   struct dm_buffer **bp)
1101 {
1102 	return new_read(c, block, NF_GET, bp);
1103 }
1104 EXPORT_SYMBOL_GPL(dm_bufio_get);
1105 
1106 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1107 		    struct dm_buffer **bp)
1108 {
1109 	BUG_ON(dm_bufio_in_request());
1110 
1111 	return new_read(c, block, NF_READ, bp);
1112 }
1113 EXPORT_SYMBOL_GPL(dm_bufio_read);
1114 
1115 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1116 		   struct dm_buffer **bp)
1117 {
1118 	BUG_ON(dm_bufio_in_request());
1119 
1120 	return new_read(c, block, NF_FRESH, bp);
1121 }
1122 EXPORT_SYMBOL_GPL(dm_bufio_new);
1123 
1124 void dm_bufio_prefetch(struct dm_bufio_client *c,
1125 		       sector_t block, unsigned n_blocks)
1126 {
1127 	struct blk_plug plug;
1128 
1129 	LIST_HEAD(write_list);
1130 
1131 	BUG_ON(dm_bufio_in_request());
1132 
1133 	blk_start_plug(&plug);
1134 	dm_bufio_lock(c);
1135 
1136 	for (; n_blocks--; block++) {
1137 		int need_submit;
1138 		struct dm_buffer *b;
1139 		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1140 				&write_list);
1141 		if (unlikely(!list_empty(&write_list))) {
1142 			dm_bufio_unlock(c);
1143 			blk_finish_plug(&plug);
1144 			__flush_write_list(&write_list);
1145 			blk_start_plug(&plug);
1146 			dm_bufio_lock(c);
1147 		}
1148 		if (unlikely(b != NULL)) {
1149 			dm_bufio_unlock(c);
1150 
1151 			if (need_submit)
1152 				submit_io(b, REQ_OP_READ, read_endio);
1153 			dm_bufio_release(b);
1154 
1155 			cond_resched();
1156 
1157 			if (!n_blocks)
1158 				goto flush_plug;
1159 			dm_bufio_lock(c);
1160 		}
1161 	}
1162 
1163 	dm_bufio_unlock(c);
1164 
1165 flush_plug:
1166 	blk_finish_plug(&plug);
1167 }
1168 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1169 
1170 void dm_bufio_release(struct dm_buffer *b)
1171 {
1172 	struct dm_bufio_client *c = b->c;
1173 
1174 	dm_bufio_lock(c);
1175 
1176 	BUG_ON(!b->hold_count);
1177 
1178 	b->hold_count--;
1179 	if (!b->hold_count) {
1180 		wake_up(&c->free_buffer_wait);
1181 
1182 		/*
1183 		 * If there were errors on the buffer, and the buffer is not
1184 		 * to be written, free the buffer. There is no point in caching
1185 		 * invalid buffer.
1186 		 */
1187 		if ((b->read_error || b->write_error) &&
1188 		    !test_bit(B_READING, &b->state) &&
1189 		    !test_bit(B_WRITING, &b->state) &&
1190 		    !test_bit(B_DIRTY, &b->state)) {
1191 			__unlink_buffer(b);
1192 			__free_buffer_wake(b);
1193 		}
1194 	}
1195 
1196 	dm_bufio_unlock(c);
1197 }
1198 EXPORT_SYMBOL_GPL(dm_bufio_release);
1199 
1200 void dm_bufio_mark_partial_buffer_dirty(struct dm_buffer *b,
1201 					unsigned start, unsigned end)
1202 {
1203 	struct dm_bufio_client *c = b->c;
1204 
1205 	BUG_ON(start >= end);
1206 	BUG_ON(end > b->c->block_size);
1207 
1208 	dm_bufio_lock(c);
1209 
1210 	BUG_ON(test_bit(B_READING, &b->state));
1211 
1212 	if (!test_and_set_bit(B_DIRTY, &b->state)) {
1213 		b->dirty_start = start;
1214 		b->dirty_end = end;
1215 		__relink_lru(b, LIST_DIRTY);
1216 	} else {
1217 		if (start < b->dirty_start)
1218 			b->dirty_start = start;
1219 		if (end > b->dirty_end)
1220 			b->dirty_end = end;
1221 	}
1222 
1223 	dm_bufio_unlock(c);
1224 }
1225 EXPORT_SYMBOL_GPL(dm_bufio_mark_partial_buffer_dirty);
1226 
1227 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1228 {
1229 	dm_bufio_mark_partial_buffer_dirty(b, 0, b->c->block_size);
1230 }
1231 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1232 
1233 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1234 {
1235 	LIST_HEAD(write_list);
1236 
1237 	BUG_ON(dm_bufio_in_request());
1238 
1239 	dm_bufio_lock(c);
1240 	__write_dirty_buffers_async(c, 0, &write_list);
1241 	dm_bufio_unlock(c);
1242 	__flush_write_list(&write_list);
1243 }
1244 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1245 
1246 /*
1247  * For performance, it is essential that the buffers are written asynchronously
1248  * and simultaneously (so that the block layer can merge the writes) and then
1249  * waited upon.
1250  *
1251  * Finally, we flush hardware disk cache.
1252  */
1253 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1254 {
1255 	int a, f;
1256 	unsigned long buffers_processed = 0;
1257 	struct dm_buffer *b, *tmp;
1258 
1259 	LIST_HEAD(write_list);
1260 
1261 	dm_bufio_lock(c);
1262 	__write_dirty_buffers_async(c, 0, &write_list);
1263 	dm_bufio_unlock(c);
1264 	__flush_write_list(&write_list);
1265 	dm_bufio_lock(c);
1266 
1267 again:
1268 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1269 		int dropped_lock = 0;
1270 
1271 		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1272 			buffers_processed++;
1273 
1274 		BUG_ON(test_bit(B_READING, &b->state));
1275 
1276 		if (test_bit(B_WRITING, &b->state)) {
1277 			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1278 				dropped_lock = 1;
1279 				b->hold_count++;
1280 				dm_bufio_unlock(c);
1281 				wait_on_bit_io(&b->state, B_WRITING,
1282 					       TASK_UNINTERRUPTIBLE);
1283 				dm_bufio_lock(c);
1284 				b->hold_count--;
1285 			} else
1286 				wait_on_bit_io(&b->state, B_WRITING,
1287 					       TASK_UNINTERRUPTIBLE);
1288 		}
1289 
1290 		if (!test_bit(B_DIRTY, &b->state) &&
1291 		    !test_bit(B_WRITING, &b->state))
1292 			__relink_lru(b, LIST_CLEAN);
1293 
1294 		cond_resched();
1295 
1296 		/*
1297 		 * If we dropped the lock, the list is no longer consistent,
1298 		 * so we must restart the search.
1299 		 *
1300 		 * In the most common case, the buffer just processed is
1301 		 * relinked to the clean list, so we won't loop scanning the
1302 		 * same buffer again and again.
1303 		 *
1304 		 * This may livelock if there is another thread simultaneously
1305 		 * dirtying buffers, so we count the number of buffers walked
1306 		 * and if it exceeds the total number of buffers, it means that
1307 		 * someone is doing some writes simultaneously with us.  In
1308 		 * this case, stop, dropping the lock.
1309 		 */
1310 		if (dropped_lock)
1311 			goto again;
1312 	}
1313 	wake_up(&c->free_buffer_wait);
1314 	dm_bufio_unlock(c);
1315 
1316 	a = xchg(&c->async_write_error, 0);
1317 	f = dm_bufio_issue_flush(c);
1318 	if (a)
1319 		return a;
1320 
1321 	return f;
1322 }
1323 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1324 
1325 /*
1326  * Use dm-io to send an empty barrier to flush the device.
1327  */
1328 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1329 {
1330 	struct dm_io_request io_req = {
1331 		.bi_op = REQ_OP_WRITE,
1332 		.bi_op_flags = REQ_PREFLUSH | REQ_SYNC,
1333 		.mem.type = DM_IO_KMEM,
1334 		.mem.ptr.addr = NULL,
1335 		.client = c->dm_io,
1336 	};
1337 	struct dm_io_region io_reg = {
1338 		.bdev = c->bdev,
1339 		.sector = 0,
1340 		.count = 0,
1341 	};
1342 
1343 	BUG_ON(dm_bufio_in_request());
1344 
1345 	return dm_io(&io_req, 1, &io_reg, NULL);
1346 }
1347 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1348 
1349 /*
1350  * We first delete any other buffer that may be at that new location.
1351  *
1352  * Then, we write the buffer to the original location if it was dirty.
1353  *
1354  * Then, if we are the only one who is holding the buffer, relink the buffer
1355  * in the buffer tree for the new location.
1356  *
1357  * If there was someone else holding the buffer, we write it to the new
1358  * location but not relink it, because that other user needs to have the buffer
1359  * at the same place.
1360  */
1361 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1362 {
1363 	struct dm_bufio_client *c = b->c;
1364 	struct dm_buffer *new;
1365 
1366 	BUG_ON(dm_bufio_in_request());
1367 
1368 	dm_bufio_lock(c);
1369 
1370 retry:
1371 	new = __find(c, new_block);
1372 	if (new) {
1373 		if (new->hold_count) {
1374 			__wait_for_free_buffer(c);
1375 			goto retry;
1376 		}
1377 
1378 		/*
1379 		 * FIXME: Is there any point waiting for a write that's going
1380 		 * to be overwritten in a bit?
1381 		 */
1382 		__make_buffer_clean(new);
1383 		__unlink_buffer(new);
1384 		__free_buffer_wake(new);
1385 	}
1386 
1387 	BUG_ON(!b->hold_count);
1388 	BUG_ON(test_bit(B_READING, &b->state));
1389 
1390 	__write_dirty_buffer(b, NULL);
1391 	if (b->hold_count == 1) {
1392 		wait_on_bit_io(&b->state, B_WRITING,
1393 			       TASK_UNINTERRUPTIBLE);
1394 		set_bit(B_DIRTY, &b->state);
1395 		b->dirty_start = 0;
1396 		b->dirty_end = c->block_size;
1397 		__unlink_buffer(b);
1398 		__link_buffer(b, new_block, LIST_DIRTY);
1399 	} else {
1400 		sector_t old_block;
1401 		wait_on_bit_lock_io(&b->state, B_WRITING,
1402 				    TASK_UNINTERRUPTIBLE);
1403 		/*
1404 		 * Relink buffer to "new_block" so that write_callback
1405 		 * sees "new_block" as a block number.
1406 		 * After the write, link the buffer back to old_block.
1407 		 * All this must be done in bufio lock, so that block number
1408 		 * change isn't visible to other threads.
1409 		 */
1410 		old_block = b->block;
1411 		__unlink_buffer(b);
1412 		__link_buffer(b, new_block, b->list_mode);
1413 		submit_io(b, REQ_OP_WRITE, write_endio);
1414 		wait_on_bit_io(&b->state, B_WRITING,
1415 			       TASK_UNINTERRUPTIBLE);
1416 		__unlink_buffer(b);
1417 		__link_buffer(b, old_block, b->list_mode);
1418 	}
1419 
1420 	dm_bufio_unlock(c);
1421 	dm_bufio_release(b);
1422 }
1423 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1424 
1425 /*
1426  * Free the given buffer.
1427  *
1428  * This is just a hint, if the buffer is in use or dirty, this function
1429  * does nothing.
1430  */
1431 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1432 {
1433 	struct dm_buffer *b;
1434 
1435 	dm_bufio_lock(c);
1436 
1437 	b = __find(c, block);
1438 	if (b && likely(!b->hold_count) && likely(!b->state)) {
1439 		__unlink_buffer(b);
1440 		__free_buffer_wake(b);
1441 	}
1442 
1443 	dm_bufio_unlock(c);
1444 }
1445 EXPORT_SYMBOL_GPL(dm_bufio_forget);
1446 
1447 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1448 {
1449 	c->minimum_buffers = n;
1450 }
1451 EXPORT_SYMBOL_GPL(dm_bufio_set_minimum_buffers);
1452 
1453 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1454 {
1455 	return c->block_size;
1456 }
1457 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1458 
1459 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1460 {
1461 	sector_t s = i_size_read(c->bdev->bd_inode) >> SECTOR_SHIFT;
1462 	if (likely(c->sectors_per_block_bits >= 0))
1463 		s >>= c->sectors_per_block_bits;
1464 	else
1465 		sector_div(s, c->block_size >> SECTOR_SHIFT);
1466 	return s;
1467 }
1468 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1469 
1470 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1471 {
1472 	return b->block;
1473 }
1474 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1475 
1476 void *dm_bufio_get_block_data(struct dm_buffer *b)
1477 {
1478 	return b->data;
1479 }
1480 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1481 
1482 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1483 {
1484 	return b + 1;
1485 }
1486 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1487 
1488 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1489 {
1490 	return b->c;
1491 }
1492 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1493 
1494 static void drop_buffers(struct dm_bufio_client *c)
1495 {
1496 	struct dm_buffer *b;
1497 	int i;
1498 	bool warned = false;
1499 
1500 	BUG_ON(dm_bufio_in_request());
1501 
1502 	/*
1503 	 * An optimization so that the buffers are not written one-by-one.
1504 	 */
1505 	dm_bufio_write_dirty_buffers_async(c);
1506 
1507 	dm_bufio_lock(c);
1508 
1509 	while ((b = __get_unclaimed_buffer(c)))
1510 		__free_buffer_wake(b);
1511 
1512 	for (i = 0; i < LIST_SIZE; i++)
1513 		list_for_each_entry(b, &c->lru[i], lru_list) {
1514 			WARN_ON(!warned);
1515 			warned = true;
1516 			DMERR("leaked buffer %llx, hold count %u, list %d",
1517 			      (unsigned long long)b->block, b->hold_count, i);
1518 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1519 			stack_trace_print(b->stack_entries, b->stack_len, 1);
1520 			/* mark unclaimed to avoid BUG_ON below */
1521 			b->hold_count = 0;
1522 #endif
1523 		}
1524 
1525 #ifdef CONFIG_DM_DEBUG_BLOCK_STACK_TRACING
1526 	while ((b = __get_unclaimed_buffer(c)))
1527 		__free_buffer_wake(b);
1528 #endif
1529 
1530 	for (i = 0; i < LIST_SIZE; i++)
1531 		BUG_ON(!list_empty(&c->lru[i]));
1532 
1533 	dm_bufio_unlock(c);
1534 }
1535 
1536 /*
1537  * We may not be able to evict this buffer if IO pending or the client
1538  * is still using it.  Caller is expected to know buffer is too old.
1539  *
1540  * And if GFP_NOFS is used, we must not do any I/O because we hold
1541  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets
1542  * rerouted to different bufio client.
1543  */
1544 static bool __try_evict_buffer(struct dm_buffer *b, gfp_t gfp)
1545 {
1546 	if (!(gfp & __GFP_FS)) {
1547 		if (test_bit(B_READING, &b->state) ||
1548 		    test_bit(B_WRITING, &b->state) ||
1549 		    test_bit(B_DIRTY, &b->state))
1550 			return false;
1551 	}
1552 
1553 	if (b->hold_count)
1554 		return false;
1555 
1556 	__make_buffer_clean(b);
1557 	__unlink_buffer(b);
1558 	__free_buffer_wake(b);
1559 
1560 	return true;
1561 }
1562 
1563 static unsigned long get_retain_buffers(struct dm_bufio_client *c)
1564 {
1565 	unsigned long retain_bytes = READ_ONCE(dm_bufio_retain_bytes);
1566 	if (likely(c->sectors_per_block_bits >= 0))
1567 		retain_bytes >>= c->sectors_per_block_bits + SECTOR_SHIFT;
1568 	else
1569 		retain_bytes /= c->block_size;
1570 	return retain_bytes;
1571 }
1572 
1573 static unsigned long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1574 			    gfp_t gfp_mask)
1575 {
1576 	int l;
1577 	struct dm_buffer *b, *tmp;
1578 	unsigned long freed = 0;
1579 	unsigned long count = c->n_buffers[LIST_CLEAN] +
1580 			      c->n_buffers[LIST_DIRTY];
1581 	unsigned long retain_target = get_retain_buffers(c);
1582 
1583 	for (l = 0; l < LIST_SIZE; l++) {
1584 		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1585 			if (__try_evict_buffer(b, gfp_mask))
1586 				freed++;
1587 			if (!--nr_to_scan || ((count - freed) <= retain_target))
1588 				return freed;
1589 			cond_resched();
1590 		}
1591 	}
1592 	return freed;
1593 }
1594 
1595 static unsigned long
1596 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1597 {
1598 	struct dm_bufio_client *c;
1599 	unsigned long freed;
1600 
1601 	c = container_of(shrink, struct dm_bufio_client, shrinker);
1602 	if (!dm_bufio_trylock(c))
1603 		return SHRINK_STOP;
1604 
1605 	freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1606 	dm_bufio_unlock(c);
1607 	return freed;
1608 }
1609 
1610 static unsigned long
1611 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1612 {
1613 	struct dm_bufio_client *c = container_of(shrink, struct dm_bufio_client, shrinker);
1614 	unsigned long count = READ_ONCE(c->n_buffers[LIST_CLEAN]) +
1615 			      READ_ONCE(c->n_buffers[LIST_DIRTY]);
1616 	unsigned long retain_target = get_retain_buffers(c);
1617 
1618 	return (count < retain_target) ? 0 : (count - retain_target);
1619 }
1620 
1621 /*
1622  * Create the buffering interface
1623  */
1624 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1625 					       unsigned reserved_buffers, unsigned aux_size,
1626 					       void (*alloc_callback)(struct dm_buffer *),
1627 					       void (*write_callback)(struct dm_buffer *))
1628 {
1629 	int r;
1630 	struct dm_bufio_client *c;
1631 	unsigned i;
1632 	char slab_name[27];
1633 
1634 	if (!block_size || block_size & ((1 << SECTOR_SHIFT) - 1)) {
1635 		DMERR("%s: block size not specified or is not multiple of 512b", __func__);
1636 		r = -EINVAL;
1637 		goto bad_client;
1638 	}
1639 
1640 	c = kzalloc(sizeof(*c), GFP_KERNEL);
1641 	if (!c) {
1642 		r = -ENOMEM;
1643 		goto bad_client;
1644 	}
1645 	c->buffer_tree = RB_ROOT;
1646 
1647 	c->bdev = bdev;
1648 	c->block_size = block_size;
1649 	if (is_power_of_2(block_size))
1650 		c->sectors_per_block_bits = __ffs(block_size) - SECTOR_SHIFT;
1651 	else
1652 		c->sectors_per_block_bits = -1;
1653 
1654 	c->alloc_callback = alloc_callback;
1655 	c->write_callback = write_callback;
1656 
1657 	for (i = 0; i < LIST_SIZE; i++) {
1658 		INIT_LIST_HEAD(&c->lru[i]);
1659 		c->n_buffers[i] = 0;
1660 	}
1661 
1662 	mutex_init(&c->lock);
1663 	INIT_LIST_HEAD(&c->reserved_buffers);
1664 	c->need_reserved_buffers = reserved_buffers;
1665 
1666 	dm_bufio_set_minimum_buffers(c, DM_BUFIO_MIN_BUFFERS);
1667 
1668 	init_waitqueue_head(&c->free_buffer_wait);
1669 	c->async_write_error = 0;
1670 
1671 	c->dm_io = dm_io_client_create();
1672 	if (IS_ERR(c->dm_io)) {
1673 		r = PTR_ERR(c->dm_io);
1674 		goto bad_dm_io;
1675 	}
1676 
1677 	if (block_size <= KMALLOC_MAX_SIZE &&
1678 	    (block_size < PAGE_SIZE || !is_power_of_2(block_size))) {
1679 		unsigned align = min(1U << __ffs(block_size), (unsigned)PAGE_SIZE);
1680 		snprintf(slab_name, sizeof slab_name, "dm_bufio_cache-%u", block_size);
1681 		c->slab_cache = kmem_cache_create(slab_name, block_size, align,
1682 						  SLAB_RECLAIM_ACCOUNT, NULL);
1683 		if (!c->slab_cache) {
1684 			r = -ENOMEM;
1685 			goto bad;
1686 		}
1687 	}
1688 	if (aux_size)
1689 		snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer-%u", aux_size);
1690 	else
1691 		snprintf(slab_name, sizeof slab_name, "dm_bufio_buffer");
1692 	c->slab_buffer = kmem_cache_create(slab_name, sizeof(struct dm_buffer) + aux_size,
1693 					   0, SLAB_RECLAIM_ACCOUNT, NULL);
1694 	if (!c->slab_buffer) {
1695 		r = -ENOMEM;
1696 		goto bad;
1697 	}
1698 
1699 	while (c->need_reserved_buffers) {
1700 		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1701 
1702 		if (!b) {
1703 			r = -ENOMEM;
1704 			goto bad;
1705 		}
1706 		__free_buffer_wake(b);
1707 	}
1708 
1709 	c->shrinker.count_objects = dm_bufio_shrink_count;
1710 	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1711 	c->shrinker.seeks = 1;
1712 	c->shrinker.batch = 0;
1713 	r = register_shrinker(&c->shrinker);
1714 	if (r)
1715 		goto bad;
1716 
1717 	mutex_lock(&dm_bufio_clients_lock);
1718 	dm_bufio_client_count++;
1719 	list_add(&c->client_list, &dm_bufio_all_clients);
1720 	__cache_size_refresh();
1721 	mutex_unlock(&dm_bufio_clients_lock);
1722 
1723 	return c;
1724 
1725 bad:
1726 	while (!list_empty(&c->reserved_buffers)) {
1727 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1728 						 struct dm_buffer, lru_list);
1729 		list_del(&b->lru_list);
1730 		free_buffer(b);
1731 	}
1732 	kmem_cache_destroy(c->slab_cache);
1733 	kmem_cache_destroy(c->slab_buffer);
1734 	dm_io_client_destroy(c->dm_io);
1735 bad_dm_io:
1736 	mutex_destroy(&c->lock);
1737 	kfree(c);
1738 bad_client:
1739 	return ERR_PTR(r);
1740 }
1741 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1742 
1743 /*
1744  * Free the buffering interface.
1745  * It is required that there are no references on any buffers.
1746  */
1747 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1748 {
1749 	unsigned i;
1750 
1751 	drop_buffers(c);
1752 
1753 	unregister_shrinker(&c->shrinker);
1754 
1755 	mutex_lock(&dm_bufio_clients_lock);
1756 
1757 	list_del(&c->client_list);
1758 	dm_bufio_client_count--;
1759 	__cache_size_refresh();
1760 
1761 	mutex_unlock(&dm_bufio_clients_lock);
1762 
1763 	BUG_ON(!RB_EMPTY_ROOT(&c->buffer_tree));
1764 	BUG_ON(c->need_reserved_buffers);
1765 
1766 	while (!list_empty(&c->reserved_buffers)) {
1767 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1768 						 struct dm_buffer, lru_list);
1769 		list_del(&b->lru_list);
1770 		free_buffer(b);
1771 	}
1772 
1773 	for (i = 0; i < LIST_SIZE; i++)
1774 		if (c->n_buffers[i])
1775 			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1776 
1777 	for (i = 0; i < LIST_SIZE; i++)
1778 		BUG_ON(c->n_buffers[i]);
1779 
1780 	kmem_cache_destroy(c->slab_cache);
1781 	kmem_cache_destroy(c->slab_buffer);
1782 	dm_io_client_destroy(c->dm_io);
1783 	mutex_destroy(&c->lock);
1784 	kfree(c);
1785 }
1786 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1787 
1788 void dm_bufio_set_sector_offset(struct dm_bufio_client *c, sector_t start)
1789 {
1790 	c->start = start;
1791 }
1792 EXPORT_SYMBOL_GPL(dm_bufio_set_sector_offset);
1793 
1794 static unsigned get_max_age_hz(void)
1795 {
1796 	unsigned max_age = READ_ONCE(dm_bufio_max_age);
1797 
1798 	if (max_age > UINT_MAX / HZ)
1799 		max_age = UINT_MAX / HZ;
1800 
1801 	return max_age * HZ;
1802 }
1803 
1804 static bool older_than(struct dm_buffer *b, unsigned long age_hz)
1805 {
1806 	return time_after_eq(jiffies, b->last_accessed + age_hz);
1807 }
1808 
1809 static void __evict_old_buffers(struct dm_bufio_client *c, unsigned long age_hz)
1810 {
1811 	struct dm_buffer *b, *tmp;
1812 	unsigned long retain_target = get_retain_buffers(c);
1813 	unsigned long count;
1814 	LIST_HEAD(write_list);
1815 
1816 	dm_bufio_lock(c);
1817 
1818 	__check_watermark(c, &write_list);
1819 	if (unlikely(!list_empty(&write_list))) {
1820 		dm_bufio_unlock(c);
1821 		__flush_write_list(&write_list);
1822 		dm_bufio_lock(c);
1823 	}
1824 
1825 	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1826 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_CLEAN], lru_list) {
1827 		if (count <= retain_target)
1828 			break;
1829 
1830 		if (!older_than(b, age_hz))
1831 			break;
1832 
1833 		if (__try_evict_buffer(b, 0))
1834 			count--;
1835 
1836 		cond_resched();
1837 	}
1838 
1839 	dm_bufio_unlock(c);
1840 }
1841 
1842 static void cleanup_old_buffers(void)
1843 {
1844 	unsigned long max_age_hz = get_max_age_hz();
1845 	struct dm_bufio_client *c;
1846 
1847 	mutex_lock(&dm_bufio_clients_lock);
1848 
1849 	__cache_size_refresh();
1850 
1851 	list_for_each_entry(c, &dm_bufio_all_clients, client_list)
1852 		__evict_old_buffers(c, max_age_hz);
1853 
1854 	mutex_unlock(&dm_bufio_clients_lock);
1855 }
1856 
1857 static struct workqueue_struct *dm_bufio_wq;
1858 static struct delayed_work dm_bufio_work;
1859 
1860 static void work_fn(struct work_struct *w)
1861 {
1862 	cleanup_old_buffers();
1863 
1864 	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1865 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1866 }
1867 
1868 /*----------------------------------------------------------------
1869  * Module setup
1870  *--------------------------------------------------------------*/
1871 
1872 /*
1873  * This is called only once for the whole dm_bufio module.
1874  * It initializes memory limit.
1875  */
1876 static int __init dm_bufio_init(void)
1877 {
1878 	__u64 mem;
1879 
1880 	dm_bufio_allocated_kmem_cache = 0;
1881 	dm_bufio_allocated_get_free_pages = 0;
1882 	dm_bufio_allocated_vmalloc = 0;
1883 	dm_bufio_current_allocated = 0;
1884 
1885 	mem = (__u64)mult_frac(totalram_pages() - totalhigh_pages(),
1886 			       DM_BUFIO_MEMORY_PERCENT, 100) << PAGE_SHIFT;
1887 
1888 	if (mem > ULONG_MAX)
1889 		mem = ULONG_MAX;
1890 
1891 #ifdef CONFIG_MMU
1892 	if (mem > mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100))
1893 		mem = mult_frac(VMALLOC_TOTAL, DM_BUFIO_VMALLOC_PERCENT, 100);
1894 #endif
1895 
1896 	dm_bufio_default_cache_size = mem;
1897 
1898 	mutex_lock(&dm_bufio_clients_lock);
1899 	__cache_size_refresh();
1900 	mutex_unlock(&dm_bufio_clients_lock);
1901 
1902 	dm_bufio_wq = alloc_workqueue("dm_bufio_cache", WQ_MEM_RECLAIM, 0);
1903 	if (!dm_bufio_wq)
1904 		return -ENOMEM;
1905 
1906 	INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1907 	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1908 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1909 
1910 	return 0;
1911 }
1912 
1913 /*
1914  * This is called once when unloading the dm_bufio module.
1915  */
1916 static void __exit dm_bufio_exit(void)
1917 {
1918 	int bug = 0;
1919 
1920 	cancel_delayed_work_sync(&dm_bufio_work);
1921 	destroy_workqueue(dm_bufio_wq);
1922 
1923 	if (dm_bufio_client_count) {
1924 		DMCRIT("%s: dm_bufio_client_count leaked: %d",
1925 			__func__, dm_bufio_client_count);
1926 		bug = 1;
1927 	}
1928 
1929 	if (dm_bufio_current_allocated) {
1930 		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1931 			__func__, dm_bufio_current_allocated);
1932 		bug = 1;
1933 	}
1934 
1935 	if (dm_bufio_allocated_get_free_pages) {
1936 		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1937 		       __func__, dm_bufio_allocated_get_free_pages);
1938 		bug = 1;
1939 	}
1940 
1941 	if (dm_bufio_allocated_vmalloc) {
1942 		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1943 		       __func__, dm_bufio_allocated_vmalloc);
1944 		bug = 1;
1945 	}
1946 
1947 	BUG_ON(bug);
1948 }
1949 
1950 module_init(dm_bufio_init)
1951 module_exit(dm_bufio_exit)
1952 
1953 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1954 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1955 
1956 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1957 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1958 
1959 module_param_named(retain_bytes, dm_bufio_retain_bytes, ulong, S_IRUGO | S_IWUSR);
1960 MODULE_PARM_DESC(retain_bytes, "Try to keep at least this many bytes cached in memory");
1961 
1962 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1963 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1964 
1965 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1966 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1967 
1968 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1969 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1970 
1971 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1972 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1973 
1974 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1975 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1976 
1977 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1978 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1979 MODULE_LICENSE("GPL");
1980