xref: /openbmc/linux/drivers/md/dm-bufio.c (revision d2168146)
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 "dm-bufio.h"
10 
11 #include <linux/device-mapper.h>
12 #include <linux/dm-io.h>
13 #include <linux/slab.h>
14 #include <linux/vmalloc.h>
15 #include <linux/shrinker.h>
16 #include <linux/module.h>
17 
18 #define DM_MSG_PREFIX "bufio"
19 
20 /*
21  * Memory management policy:
22  *	Limit the number of buffers to DM_BUFIO_MEMORY_PERCENT of main memory
23  *	or DM_BUFIO_VMALLOC_PERCENT of vmalloc memory (whichever is lower).
24  *	Always allocate at least DM_BUFIO_MIN_BUFFERS buffers.
25  *	Start background writeback when there are DM_BUFIO_WRITEBACK_PERCENT
26  *	dirty buffers.
27  */
28 #define DM_BUFIO_MIN_BUFFERS		8
29 
30 #define DM_BUFIO_MEMORY_PERCENT		2
31 #define DM_BUFIO_VMALLOC_PERCENT	25
32 #define DM_BUFIO_WRITEBACK_PERCENT	75
33 
34 /*
35  * Check buffer ages in this interval (seconds)
36  */
37 #define DM_BUFIO_WORK_TIMER_SECS	10
38 
39 /*
40  * Free buffers when they are older than this (seconds)
41  */
42 #define DM_BUFIO_DEFAULT_AGE_SECS	60
43 
44 /*
45  * The number of bvec entries that are embedded directly in the buffer.
46  * If the chunk size is larger, dm-io is used to do the io.
47  */
48 #define DM_BUFIO_INLINE_VECS		16
49 
50 /*
51  * Buffer hash
52  */
53 #define DM_BUFIO_HASH_BITS	20
54 #define DM_BUFIO_HASH(block) \
55 	((((block) >> DM_BUFIO_HASH_BITS) ^ (block)) & \
56 	 ((1 << DM_BUFIO_HASH_BITS) - 1))
57 
58 /*
59  * Don't try to use kmem_cache_alloc for blocks larger than this.
60  * For explanation, see alloc_buffer_data below.
61  */
62 #define DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT	(PAGE_SIZE >> 1)
63 #define DM_BUFIO_BLOCK_SIZE_GFP_LIMIT	(PAGE_SIZE << (MAX_ORDER - 1))
64 
65 /*
66  * dm_buffer->list_mode
67  */
68 #define LIST_CLEAN	0
69 #define LIST_DIRTY	1
70 #define LIST_SIZE	2
71 
72 /*
73  * Linking of buffers:
74  *	All buffers are linked to cache_hash with their hash_list field.
75  *
76  *	Clean buffers that are not being written (B_WRITING not set)
77  *	are linked to lru[LIST_CLEAN] with their lru_list field.
78  *
79  *	Dirty and clean buffers that are being written are linked to
80  *	lru[LIST_DIRTY] with their lru_list field. When the write
81  *	finishes, the buffer cannot be relinked immediately (because we
82  *	are in an interrupt context and relinking requires process
83  *	context), so some clean-not-writing buffers can be held on
84  *	dirty_lru too.  They are later added to lru in the process
85  *	context.
86  */
87 struct dm_bufio_client {
88 	struct mutex lock;
89 
90 	struct list_head lru[LIST_SIZE];
91 	unsigned long n_buffers[LIST_SIZE];
92 
93 	struct block_device *bdev;
94 	unsigned block_size;
95 	unsigned char sectors_per_block_bits;
96 	unsigned char pages_per_block_bits;
97 	unsigned char blocks_per_page_bits;
98 	unsigned aux_size;
99 	void (*alloc_callback)(struct dm_buffer *);
100 	void (*write_callback)(struct dm_buffer *);
101 
102 	struct dm_io_client *dm_io;
103 
104 	struct list_head reserved_buffers;
105 	unsigned need_reserved_buffers;
106 
107 	unsigned minimum_buffers;
108 
109 	struct hlist_head *cache_hash;
110 	wait_queue_head_t free_buffer_wait;
111 
112 	int async_write_error;
113 
114 	struct list_head client_list;
115 	struct shrinker shrinker;
116 };
117 
118 /*
119  * Buffer state bits.
120  */
121 #define B_READING	0
122 #define B_WRITING	1
123 #define B_DIRTY		2
124 
125 /*
126  * Describes how the block was allocated:
127  * kmem_cache_alloc(), __get_free_pages() or vmalloc().
128  * See the comment at alloc_buffer_data.
129  */
130 enum data_mode {
131 	DATA_MODE_SLAB = 0,
132 	DATA_MODE_GET_FREE_PAGES = 1,
133 	DATA_MODE_VMALLOC = 2,
134 	DATA_MODE_LIMIT = 3
135 };
136 
137 struct dm_buffer {
138 	struct hlist_node hash_list;
139 	struct list_head lru_list;
140 	sector_t block;
141 	void *data;
142 	enum data_mode data_mode;
143 	unsigned char list_mode;		/* LIST_* */
144 	unsigned hold_count;
145 	int read_error;
146 	int write_error;
147 	unsigned long state;
148 	unsigned long last_accessed;
149 	struct dm_bufio_client *c;
150 	struct list_head write_list;
151 	struct bio bio;
152 	struct bio_vec bio_vec[DM_BUFIO_INLINE_VECS];
153 };
154 
155 /*----------------------------------------------------------------*/
156 
157 static struct kmem_cache *dm_bufio_caches[PAGE_SHIFT - SECTOR_SHIFT];
158 static char *dm_bufio_cache_names[PAGE_SHIFT - SECTOR_SHIFT];
159 
160 static inline int dm_bufio_cache_index(struct dm_bufio_client *c)
161 {
162 	unsigned ret = c->blocks_per_page_bits - 1;
163 
164 	BUG_ON(ret >= ARRAY_SIZE(dm_bufio_caches));
165 
166 	return ret;
167 }
168 
169 #define DM_BUFIO_CACHE(c)	(dm_bufio_caches[dm_bufio_cache_index(c)])
170 #define DM_BUFIO_CACHE_NAME(c)	(dm_bufio_cache_names[dm_bufio_cache_index(c)])
171 
172 #define dm_bufio_in_request()	(!!current->bio_list)
173 
174 static void dm_bufio_lock(struct dm_bufio_client *c)
175 {
176 	mutex_lock_nested(&c->lock, dm_bufio_in_request());
177 }
178 
179 static int dm_bufio_trylock(struct dm_bufio_client *c)
180 {
181 	return mutex_trylock(&c->lock);
182 }
183 
184 static void dm_bufio_unlock(struct dm_bufio_client *c)
185 {
186 	mutex_unlock(&c->lock);
187 }
188 
189 /*
190  * FIXME Move to sched.h?
191  */
192 #ifdef CONFIG_PREEMPT_VOLUNTARY
193 #  define dm_bufio_cond_resched()		\
194 do {						\
195 	if (unlikely(need_resched()))		\
196 		_cond_resched();		\
197 } while (0)
198 #else
199 #  define dm_bufio_cond_resched()                do { } while (0)
200 #endif
201 
202 /*----------------------------------------------------------------*/
203 
204 /*
205  * Default cache size: available memory divided by the ratio.
206  */
207 static unsigned long dm_bufio_default_cache_size;
208 
209 /*
210  * Total cache size set by the user.
211  */
212 static unsigned long dm_bufio_cache_size;
213 
214 /*
215  * A copy of dm_bufio_cache_size because dm_bufio_cache_size can change
216  * at any time.  If it disagrees, the user has changed cache size.
217  */
218 static unsigned long dm_bufio_cache_size_latch;
219 
220 static DEFINE_SPINLOCK(param_spinlock);
221 
222 /*
223  * Buffers are freed after this timeout
224  */
225 static unsigned dm_bufio_max_age = DM_BUFIO_DEFAULT_AGE_SECS;
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  * Per-client cache: dm_bufio_cache_size / dm_bufio_client_count
237  */
238 static unsigned long dm_bufio_cache_size_per_client;
239 
240 /*
241  * The current number of clients.
242  */
243 static int dm_bufio_client_count;
244 
245 /*
246  * The list of all clients.
247  */
248 static LIST_HEAD(dm_bufio_all_clients);
249 
250 /*
251  * This mutex protects dm_bufio_cache_size_latch,
252  * dm_bufio_cache_size_per_client and dm_bufio_client_count
253  */
254 static DEFINE_MUTEX(dm_bufio_clients_lock);
255 
256 /*----------------------------------------------------------------*/
257 
258 static void adjust_total_allocated(enum data_mode data_mode, long diff)
259 {
260 	static unsigned long * const class_ptr[DATA_MODE_LIMIT] = {
261 		&dm_bufio_allocated_kmem_cache,
262 		&dm_bufio_allocated_get_free_pages,
263 		&dm_bufio_allocated_vmalloc,
264 	};
265 
266 	spin_lock(&param_spinlock);
267 
268 	*class_ptr[data_mode] += diff;
269 
270 	dm_bufio_current_allocated += diff;
271 
272 	if (dm_bufio_current_allocated > dm_bufio_peak_allocated)
273 		dm_bufio_peak_allocated = dm_bufio_current_allocated;
274 
275 	spin_unlock(&param_spinlock);
276 }
277 
278 /*
279  * Change the number of clients and recalculate per-client limit.
280  */
281 static void __cache_size_refresh(void)
282 {
283 	BUG_ON(!mutex_is_locked(&dm_bufio_clients_lock));
284 	BUG_ON(dm_bufio_client_count < 0);
285 
286 	dm_bufio_cache_size_latch = ACCESS_ONCE(dm_bufio_cache_size);
287 
288 	/*
289 	 * Use default if set to 0 and report the actual cache size used.
290 	 */
291 	if (!dm_bufio_cache_size_latch) {
292 		(void)cmpxchg(&dm_bufio_cache_size, 0,
293 			      dm_bufio_default_cache_size);
294 		dm_bufio_cache_size_latch = dm_bufio_default_cache_size;
295 	}
296 
297 	dm_bufio_cache_size_per_client = dm_bufio_cache_size_latch /
298 					 (dm_bufio_client_count ? : 1);
299 }
300 
301 /*
302  * Allocating buffer data.
303  *
304  * Small buffers are allocated with kmem_cache, to use space optimally.
305  *
306  * For large buffers, we choose between get_free_pages and vmalloc.
307  * Each has advantages and disadvantages.
308  *
309  * __get_free_pages can randomly fail if the memory is fragmented.
310  * __vmalloc won't randomly fail, but vmalloc space is limited (it may be
311  * as low as 128M) so using it for caching is not appropriate.
312  *
313  * If the allocation may fail we use __get_free_pages. Memory fragmentation
314  * won't have a fatal effect here, but it just causes flushes of some other
315  * buffers and more I/O will be performed. Don't use __get_free_pages if it
316  * always fails (i.e. order >= MAX_ORDER).
317  *
318  * If the allocation shouldn't fail we use __vmalloc. This is only for the
319  * initial reserve allocation, so there's no risk of wasting all vmalloc
320  * space.
321  */
322 static void *alloc_buffer_data(struct dm_bufio_client *c, gfp_t gfp_mask,
323 			       enum data_mode *data_mode)
324 {
325 	unsigned noio_flag;
326 	void *ptr;
327 
328 	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_SLAB_LIMIT) {
329 		*data_mode = DATA_MODE_SLAB;
330 		return kmem_cache_alloc(DM_BUFIO_CACHE(c), gfp_mask);
331 	}
332 
333 	if (c->block_size <= DM_BUFIO_BLOCK_SIZE_GFP_LIMIT &&
334 	    gfp_mask & __GFP_NORETRY) {
335 		*data_mode = DATA_MODE_GET_FREE_PAGES;
336 		return (void *)__get_free_pages(gfp_mask,
337 						c->pages_per_block_bits);
338 	}
339 
340 	*data_mode = DATA_MODE_VMALLOC;
341 
342 	/*
343 	 * __vmalloc allocates the data pages and auxiliary structures with
344 	 * gfp_flags that were specified, but pagetables are always allocated
345 	 * with GFP_KERNEL, no matter what was specified as gfp_mask.
346 	 *
347 	 * Consequently, we must set per-process flag PF_MEMALLOC_NOIO so that
348 	 * all allocations done by this process (including pagetables) are done
349 	 * as if GFP_NOIO was specified.
350 	 */
351 
352 	if (gfp_mask & __GFP_NORETRY)
353 		noio_flag = memalloc_noio_save();
354 
355 	ptr = __vmalloc(c->block_size, gfp_mask | __GFP_HIGHMEM, PAGE_KERNEL);
356 
357 	if (gfp_mask & __GFP_NORETRY)
358 		memalloc_noio_restore(noio_flag);
359 
360 	return ptr;
361 }
362 
363 /*
364  * Free buffer's data.
365  */
366 static void free_buffer_data(struct dm_bufio_client *c,
367 			     void *data, enum data_mode data_mode)
368 {
369 	switch (data_mode) {
370 	case DATA_MODE_SLAB:
371 		kmem_cache_free(DM_BUFIO_CACHE(c), data);
372 		break;
373 
374 	case DATA_MODE_GET_FREE_PAGES:
375 		free_pages((unsigned long)data, c->pages_per_block_bits);
376 		break;
377 
378 	case DATA_MODE_VMALLOC:
379 		vfree(data);
380 		break;
381 
382 	default:
383 		DMCRIT("dm_bufio_free_buffer_data: bad data mode: %d",
384 		       data_mode);
385 		BUG();
386 	}
387 }
388 
389 /*
390  * Allocate buffer and its data.
391  */
392 static struct dm_buffer *alloc_buffer(struct dm_bufio_client *c, gfp_t gfp_mask)
393 {
394 	struct dm_buffer *b = kmalloc(sizeof(struct dm_buffer) + c->aux_size,
395 				      gfp_mask);
396 
397 	if (!b)
398 		return NULL;
399 
400 	b->c = c;
401 
402 	b->data = alloc_buffer_data(c, gfp_mask, &b->data_mode);
403 	if (!b->data) {
404 		kfree(b);
405 		return NULL;
406 	}
407 
408 	adjust_total_allocated(b->data_mode, (long)c->block_size);
409 
410 	return b;
411 }
412 
413 /*
414  * Free buffer and its data.
415  */
416 static void free_buffer(struct dm_buffer *b)
417 {
418 	struct dm_bufio_client *c = b->c;
419 
420 	adjust_total_allocated(b->data_mode, -(long)c->block_size);
421 
422 	free_buffer_data(c, b->data, b->data_mode);
423 	kfree(b);
424 }
425 
426 /*
427  * Link buffer to the hash list and clean or dirty queue.
428  */
429 static void __link_buffer(struct dm_buffer *b, sector_t block, int dirty)
430 {
431 	struct dm_bufio_client *c = b->c;
432 
433 	c->n_buffers[dirty]++;
434 	b->block = block;
435 	b->list_mode = dirty;
436 	list_add(&b->lru_list, &c->lru[dirty]);
437 	hlist_add_head(&b->hash_list, &c->cache_hash[DM_BUFIO_HASH(block)]);
438 	b->last_accessed = jiffies;
439 }
440 
441 /*
442  * Unlink buffer from the hash list and dirty or clean queue.
443  */
444 static void __unlink_buffer(struct dm_buffer *b)
445 {
446 	struct dm_bufio_client *c = b->c;
447 
448 	BUG_ON(!c->n_buffers[b->list_mode]);
449 
450 	c->n_buffers[b->list_mode]--;
451 	hlist_del(&b->hash_list);
452 	list_del(&b->lru_list);
453 }
454 
455 /*
456  * Place the buffer to the head of dirty or clean LRU queue.
457  */
458 static void __relink_lru(struct dm_buffer *b, int dirty)
459 {
460 	struct dm_bufio_client *c = b->c;
461 
462 	BUG_ON(!c->n_buffers[b->list_mode]);
463 
464 	c->n_buffers[b->list_mode]--;
465 	c->n_buffers[dirty]++;
466 	b->list_mode = dirty;
467 	list_move(&b->lru_list, &c->lru[dirty]);
468 }
469 
470 /*----------------------------------------------------------------
471  * Submit I/O on the buffer.
472  *
473  * Bio interface is faster but it has some problems:
474  *	the vector list is limited (increasing this limit increases
475  *	memory-consumption per buffer, so it is not viable);
476  *
477  *	the memory must be direct-mapped, not vmalloced;
478  *
479  *	the I/O driver can reject requests spuriously if it thinks that
480  *	the requests are too big for the device or if they cross a
481  *	controller-defined memory boundary.
482  *
483  * If the buffer is small enough (up to DM_BUFIO_INLINE_VECS pages) and
484  * it is not vmalloced, try using the bio interface.
485  *
486  * If the buffer is big, if it is vmalloced or if the underlying device
487  * rejects the bio because it is too large, use dm-io layer to do the I/O.
488  * The dm-io layer splits the I/O into multiple requests, avoiding the above
489  * shortcomings.
490  *--------------------------------------------------------------*/
491 
492 /*
493  * dm-io completion routine. It just calls b->bio.bi_end_io, pretending
494  * that the request was handled directly with bio interface.
495  */
496 static void dmio_complete(unsigned long error, void *context)
497 {
498 	struct dm_buffer *b = context;
499 
500 	b->bio.bi_end_io(&b->bio, error ? -EIO : 0);
501 }
502 
503 static void use_dmio(struct dm_buffer *b, int rw, sector_t block,
504 		     bio_end_io_t *end_io)
505 {
506 	int r;
507 	struct dm_io_request io_req = {
508 		.bi_rw = rw,
509 		.notify.fn = dmio_complete,
510 		.notify.context = b,
511 		.client = b->c->dm_io,
512 	};
513 	struct dm_io_region region = {
514 		.bdev = b->c->bdev,
515 		.sector = block << b->c->sectors_per_block_bits,
516 		.count = b->c->block_size >> SECTOR_SHIFT,
517 	};
518 
519 	if (b->data_mode != DATA_MODE_VMALLOC) {
520 		io_req.mem.type = DM_IO_KMEM;
521 		io_req.mem.ptr.addr = b->data;
522 	} else {
523 		io_req.mem.type = DM_IO_VMA;
524 		io_req.mem.ptr.vma = b->data;
525 	}
526 
527 	b->bio.bi_end_io = end_io;
528 
529 	r = dm_io(&io_req, 1, &region, NULL);
530 	if (r)
531 		end_io(&b->bio, r);
532 }
533 
534 static void use_inline_bio(struct dm_buffer *b, int rw, sector_t block,
535 			   bio_end_io_t *end_io)
536 {
537 	char *ptr;
538 	int len;
539 
540 	bio_init(&b->bio);
541 	b->bio.bi_io_vec = b->bio_vec;
542 	b->bio.bi_max_vecs = DM_BUFIO_INLINE_VECS;
543 	b->bio.bi_iter.bi_sector = block << b->c->sectors_per_block_bits;
544 	b->bio.bi_bdev = b->c->bdev;
545 	b->bio.bi_end_io = end_io;
546 
547 	/*
548 	 * We assume that if len >= PAGE_SIZE ptr is page-aligned.
549 	 * If len < PAGE_SIZE the buffer doesn't cross page boundary.
550 	 */
551 	ptr = b->data;
552 	len = b->c->block_size;
553 
554 	if (len >= PAGE_SIZE)
555 		BUG_ON((unsigned long)ptr & (PAGE_SIZE - 1));
556 	else
557 		BUG_ON((unsigned long)ptr & (len - 1));
558 
559 	do {
560 		if (!bio_add_page(&b->bio, virt_to_page(ptr),
561 				  len < PAGE_SIZE ? len : PAGE_SIZE,
562 				  virt_to_phys(ptr) & (PAGE_SIZE - 1))) {
563 			BUG_ON(b->c->block_size <= PAGE_SIZE);
564 			use_dmio(b, rw, block, end_io);
565 			return;
566 		}
567 
568 		len -= PAGE_SIZE;
569 		ptr += PAGE_SIZE;
570 	} while (len > 0);
571 
572 	submit_bio(rw, &b->bio);
573 }
574 
575 static void submit_io(struct dm_buffer *b, int rw, sector_t block,
576 		      bio_end_io_t *end_io)
577 {
578 	if (rw == WRITE && b->c->write_callback)
579 		b->c->write_callback(b);
580 
581 	if (b->c->block_size <= DM_BUFIO_INLINE_VECS * PAGE_SIZE &&
582 	    b->data_mode != DATA_MODE_VMALLOC)
583 		use_inline_bio(b, rw, block, end_io);
584 	else
585 		use_dmio(b, rw, block, end_io);
586 }
587 
588 /*----------------------------------------------------------------
589  * Writing dirty buffers
590  *--------------------------------------------------------------*/
591 
592 /*
593  * The endio routine for write.
594  *
595  * Set the error, clear B_WRITING bit and wake anyone who was waiting on
596  * it.
597  */
598 static void write_endio(struct bio *bio, int error)
599 {
600 	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
601 
602 	b->write_error = error;
603 	if (unlikely(error)) {
604 		struct dm_bufio_client *c = b->c;
605 		(void)cmpxchg(&c->async_write_error, 0, error);
606 	}
607 
608 	BUG_ON(!test_bit(B_WRITING, &b->state));
609 
610 	smp_mb__before_atomic();
611 	clear_bit(B_WRITING, &b->state);
612 	smp_mb__after_atomic();
613 
614 	wake_up_bit(&b->state, B_WRITING);
615 }
616 
617 /*
618  * This function is called when wait_on_bit is actually waiting.
619  */
620 static int do_io_schedule(void *word)
621 {
622 	io_schedule();
623 
624 	return 0;
625 }
626 
627 /*
628  * Initiate a write on a dirty buffer, but don't wait for it.
629  *
630  * - If the buffer is not dirty, exit.
631  * - If there some previous write going on, wait for it to finish (we can't
632  *   have two writes on the same buffer simultaneously).
633  * - Submit our write and don't wait on it. We set B_WRITING indicating
634  *   that there is a write in progress.
635  */
636 static void __write_dirty_buffer(struct dm_buffer *b,
637 				 struct list_head *write_list)
638 {
639 	if (!test_bit(B_DIRTY, &b->state))
640 		return;
641 
642 	clear_bit(B_DIRTY, &b->state);
643 	wait_on_bit_lock(&b->state, B_WRITING,
644 			 do_io_schedule, TASK_UNINTERRUPTIBLE);
645 
646 	if (!write_list)
647 		submit_io(b, WRITE, b->block, write_endio);
648 	else
649 		list_add_tail(&b->write_list, write_list);
650 }
651 
652 static void __flush_write_list(struct list_head *write_list)
653 {
654 	struct blk_plug plug;
655 	blk_start_plug(&plug);
656 	while (!list_empty(write_list)) {
657 		struct dm_buffer *b =
658 			list_entry(write_list->next, struct dm_buffer, write_list);
659 		list_del(&b->write_list);
660 		submit_io(b, WRITE, b->block, write_endio);
661 		dm_bufio_cond_resched();
662 	}
663 	blk_finish_plug(&plug);
664 }
665 
666 /*
667  * Wait until any activity on the buffer finishes.  Possibly write the
668  * buffer if it is dirty.  When this function finishes, there is no I/O
669  * running on the buffer and the buffer is not dirty.
670  */
671 static void __make_buffer_clean(struct dm_buffer *b)
672 {
673 	BUG_ON(b->hold_count);
674 
675 	if (!b->state)	/* fast case */
676 		return;
677 
678 	wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
679 	__write_dirty_buffer(b, NULL);
680 	wait_on_bit(&b->state, B_WRITING, do_io_schedule, TASK_UNINTERRUPTIBLE);
681 }
682 
683 /*
684  * Find some buffer that is not held by anybody, clean it, unlink it and
685  * return it.
686  */
687 static struct dm_buffer *__get_unclaimed_buffer(struct dm_bufio_client *c)
688 {
689 	struct dm_buffer *b;
690 
691 	list_for_each_entry_reverse(b, &c->lru[LIST_CLEAN], lru_list) {
692 		BUG_ON(test_bit(B_WRITING, &b->state));
693 		BUG_ON(test_bit(B_DIRTY, &b->state));
694 
695 		if (!b->hold_count) {
696 			__make_buffer_clean(b);
697 			__unlink_buffer(b);
698 			return b;
699 		}
700 		dm_bufio_cond_resched();
701 	}
702 
703 	list_for_each_entry_reverse(b, &c->lru[LIST_DIRTY], lru_list) {
704 		BUG_ON(test_bit(B_READING, &b->state));
705 
706 		if (!b->hold_count) {
707 			__make_buffer_clean(b);
708 			__unlink_buffer(b);
709 			return b;
710 		}
711 		dm_bufio_cond_resched();
712 	}
713 
714 	return NULL;
715 }
716 
717 /*
718  * Wait until some other threads free some buffer or release hold count on
719  * some buffer.
720  *
721  * This function is entered with c->lock held, drops it and regains it
722  * before exiting.
723  */
724 static void __wait_for_free_buffer(struct dm_bufio_client *c)
725 {
726 	DECLARE_WAITQUEUE(wait, current);
727 
728 	add_wait_queue(&c->free_buffer_wait, &wait);
729 	set_task_state(current, TASK_UNINTERRUPTIBLE);
730 	dm_bufio_unlock(c);
731 
732 	io_schedule();
733 
734 	set_task_state(current, TASK_RUNNING);
735 	remove_wait_queue(&c->free_buffer_wait, &wait);
736 
737 	dm_bufio_lock(c);
738 }
739 
740 enum new_flag {
741 	NF_FRESH = 0,
742 	NF_READ = 1,
743 	NF_GET = 2,
744 	NF_PREFETCH = 3
745 };
746 
747 /*
748  * Allocate a new buffer. If the allocation is not possible, wait until
749  * some other thread frees a buffer.
750  *
751  * May drop the lock and regain it.
752  */
753 static struct dm_buffer *__alloc_buffer_wait_no_callback(struct dm_bufio_client *c, enum new_flag nf)
754 {
755 	struct dm_buffer *b;
756 
757 	/*
758 	 * dm-bufio is resistant to allocation failures (it just keeps
759 	 * one buffer reserved in cases all the allocations fail).
760 	 * So set flags to not try too hard:
761 	 *	GFP_NOIO: don't recurse into the I/O layer
762 	 *	__GFP_NORETRY: don't retry and rather return failure
763 	 *	__GFP_NOMEMALLOC: don't use emergency reserves
764 	 *	__GFP_NOWARN: don't print a warning in case of failure
765 	 *
766 	 * For debugging, if we set the cache size to 1, no new buffers will
767 	 * be allocated.
768 	 */
769 	while (1) {
770 		if (dm_bufio_cache_size_latch != 1) {
771 			b = alloc_buffer(c, GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
772 			if (b)
773 				return b;
774 		}
775 
776 		if (nf == NF_PREFETCH)
777 			return NULL;
778 
779 		if (!list_empty(&c->reserved_buffers)) {
780 			b = list_entry(c->reserved_buffers.next,
781 				       struct dm_buffer, lru_list);
782 			list_del(&b->lru_list);
783 			c->need_reserved_buffers++;
784 
785 			return b;
786 		}
787 
788 		b = __get_unclaimed_buffer(c);
789 		if (b)
790 			return b;
791 
792 		__wait_for_free_buffer(c);
793 	}
794 }
795 
796 static struct dm_buffer *__alloc_buffer_wait(struct dm_bufio_client *c, enum new_flag nf)
797 {
798 	struct dm_buffer *b = __alloc_buffer_wait_no_callback(c, nf);
799 
800 	if (!b)
801 		return NULL;
802 
803 	if (c->alloc_callback)
804 		c->alloc_callback(b);
805 
806 	return b;
807 }
808 
809 /*
810  * Free a buffer and wake other threads waiting for free buffers.
811  */
812 static void __free_buffer_wake(struct dm_buffer *b)
813 {
814 	struct dm_bufio_client *c = b->c;
815 
816 	if (!c->need_reserved_buffers)
817 		free_buffer(b);
818 	else {
819 		list_add(&b->lru_list, &c->reserved_buffers);
820 		c->need_reserved_buffers--;
821 	}
822 
823 	wake_up(&c->free_buffer_wait);
824 }
825 
826 static void __write_dirty_buffers_async(struct dm_bufio_client *c, int no_wait,
827 					struct list_head *write_list)
828 {
829 	struct dm_buffer *b, *tmp;
830 
831 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
832 		BUG_ON(test_bit(B_READING, &b->state));
833 
834 		if (!test_bit(B_DIRTY, &b->state) &&
835 		    !test_bit(B_WRITING, &b->state)) {
836 			__relink_lru(b, LIST_CLEAN);
837 			continue;
838 		}
839 
840 		if (no_wait && test_bit(B_WRITING, &b->state))
841 			return;
842 
843 		__write_dirty_buffer(b, write_list);
844 		dm_bufio_cond_resched();
845 	}
846 }
847 
848 /*
849  * Get writeback threshold and buffer limit for a given client.
850  */
851 static void __get_memory_limit(struct dm_bufio_client *c,
852 			       unsigned long *threshold_buffers,
853 			       unsigned long *limit_buffers)
854 {
855 	unsigned long buffers;
856 
857 	if (ACCESS_ONCE(dm_bufio_cache_size) != dm_bufio_cache_size_latch) {
858 		mutex_lock(&dm_bufio_clients_lock);
859 		__cache_size_refresh();
860 		mutex_unlock(&dm_bufio_clients_lock);
861 	}
862 
863 	buffers = dm_bufio_cache_size_per_client >>
864 		  (c->sectors_per_block_bits + SECTOR_SHIFT);
865 
866 	if (buffers < c->minimum_buffers)
867 		buffers = c->minimum_buffers;
868 
869 	*limit_buffers = buffers;
870 	*threshold_buffers = buffers * DM_BUFIO_WRITEBACK_PERCENT / 100;
871 }
872 
873 /*
874  * Check if we're over watermark.
875  * If we are over threshold_buffers, start freeing buffers.
876  * If we're over "limit_buffers", block until we get under the limit.
877  */
878 static void __check_watermark(struct dm_bufio_client *c,
879 			      struct list_head *write_list)
880 {
881 	unsigned long threshold_buffers, limit_buffers;
882 
883 	__get_memory_limit(c, &threshold_buffers, &limit_buffers);
884 
885 	while (c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY] >
886 	       limit_buffers) {
887 
888 		struct dm_buffer *b = __get_unclaimed_buffer(c);
889 
890 		if (!b)
891 			return;
892 
893 		__free_buffer_wake(b);
894 		dm_bufio_cond_resched();
895 	}
896 
897 	if (c->n_buffers[LIST_DIRTY] > threshold_buffers)
898 		__write_dirty_buffers_async(c, 1, write_list);
899 }
900 
901 /*
902  * Find a buffer in the hash.
903  */
904 static struct dm_buffer *__find(struct dm_bufio_client *c, sector_t block)
905 {
906 	struct dm_buffer *b;
907 
908 	hlist_for_each_entry(b, &c->cache_hash[DM_BUFIO_HASH(block)],
909 			     hash_list) {
910 		dm_bufio_cond_resched();
911 		if (b->block == block)
912 			return b;
913 	}
914 
915 	return NULL;
916 }
917 
918 /*----------------------------------------------------------------
919  * Getting a buffer
920  *--------------------------------------------------------------*/
921 
922 static struct dm_buffer *__bufio_new(struct dm_bufio_client *c, sector_t block,
923 				     enum new_flag nf, int *need_submit,
924 				     struct list_head *write_list)
925 {
926 	struct dm_buffer *b, *new_b = NULL;
927 
928 	*need_submit = 0;
929 
930 	b = __find(c, block);
931 	if (b)
932 		goto found_buffer;
933 
934 	if (nf == NF_GET)
935 		return NULL;
936 
937 	new_b = __alloc_buffer_wait(c, nf);
938 	if (!new_b)
939 		return NULL;
940 
941 	/*
942 	 * We've had a period where the mutex was unlocked, so need to
943 	 * recheck the hash table.
944 	 */
945 	b = __find(c, block);
946 	if (b) {
947 		__free_buffer_wake(new_b);
948 		goto found_buffer;
949 	}
950 
951 	__check_watermark(c, write_list);
952 
953 	b = new_b;
954 	b->hold_count = 1;
955 	b->read_error = 0;
956 	b->write_error = 0;
957 	__link_buffer(b, block, LIST_CLEAN);
958 
959 	if (nf == NF_FRESH) {
960 		b->state = 0;
961 		return b;
962 	}
963 
964 	b->state = 1 << B_READING;
965 	*need_submit = 1;
966 
967 	return b;
968 
969 found_buffer:
970 	if (nf == NF_PREFETCH)
971 		return NULL;
972 	/*
973 	 * Note: it is essential that we don't wait for the buffer to be
974 	 * read if dm_bufio_get function is used. Both dm_bufio_get and
975 	 * dm_bufio_prefetch can be used in the driver request routine.
976 	 * If the user called both dm_bufio_prefetch and dm_bufio_get on
977 	 * the same buffer, it would deadlock if we waited.
978 	 */
979 	if (nf == NF_GET && unlikely(test_bit(B_READING, &b->state)))
980 		return NULL;
981 
982 	b->hold_count++;
983 	__relink_lru(b, test_bit(B_DIRTY, &b->state) ||
984 		     test_bit(B_WRITING, &b->state));
985 	return b;
986 }
987 
988 /*
989  * The endio routine for reading: set the error, clear the bit and wake up
990  * anyone waiting on the buffer.
991  */
992 static void read_endio(struct bio *bio, int error)
993 {
994 	struct dm_buffer *b = container_of(bio, struct dm_buffer, bio);
995 
996 	b->read_error = error;
997 
998 	BUG_ON(!test_bit(B_READING, &b->state));
999 
1000 	smp_mb__before_atomic();
1001 	clear_bit(B_READING, &b->state);
1002 	smp_mb__after_atomic();
1003 
1004 	wake_up_bit(&b->state, B_READING);
1005 }
1006 
1007 /*
1008  * A common routine for dm_bufio_new and dm_bufio_read.  Operation of these
1009  * functions is similar except that dm_bufio_new doesn't read the
1010  * buffer from the disk (assuming that the caller overwrites all the data
1011  * and uses dm_bufio_mark_buffer_dirty to write new data back).
1012  */
1013 static void *new_read(struct dm_bufio_client *c, sector_t block,
1014 		      enum new_flag nf, struct dm_buffer **bp)
1015 {
1016 	int need_submit;
1017 	struct dm_buffer *b;
1018 
1019 	LIST_HEAD(write_list);
1020 
1021 	dm_bufio_lock(c);
1022 	b = __bufio_new(c, block, nf, &need_submit, &write_list);
1023 	dm_bufio_unlock(c);
1024 
1025 	__flush_write_list(&write_list);
1026 
1027 	if (!b)
1028 		return b;
1029 
1030 	if (need_submit)
1031 		submit_io(b, READ, b->block, read_endio);
1032 
1033 	wait_on_bit(&b->state, B_READING, do_io_schedule, TASK_UNINTERRUPTIBLE);
1034 
1035 	if (b->read_error) {
1036 		int error = b->read_error;
1037 
1038 		dm_bufio_release(b);
1039 
1040 		return ERR_PTR(error);
1041 	}
1042 
1043 	*bp = b;
1044 
1045 	return b->data;
1046 }
1047 
1048 void *dm_bufio_get(struct dm_bufio_client *c, sector_t block,
1049 		   struct dm_buffer **bp)
1050 {
1051 	return new_read(c, block, NF_GET, bp);
1052 }
1053 EXPORT_SYMBOL_GPL(dm_bufio_get);
1054 
1055 void *dm_bufio_read(struct dm_bufio_client *c, sector_t block,
1056 		    struct dm_buffer **bp)
1057 {
1058 	BUG_ON(dm_bufio_in_request());
1059 
1060 	return new_read(c, block, NF_READ, bp);
1061 }
1062 EXPORT_SYMBOL_GPL(dm_bufio_read);
1063 
1064 void *dm_bufio_new(struct dm_bufio_client *c, sector_t block,
1065 		   struct dm_buffer **bp)
1066 {
1067 	BUG_ON(dm_bufio_in_request());
1068 
1069 	return new_read(c, block, NF_FRESH, bp);
1070 }
1071 EXPORT_SYMBOL_GPL(dm_bufio_new);
1072 
1073 void dm_bufio_prefetch(struct dm_bufio_client *c,
1074 		       sector_t block, unsigned n_blocks)
1075 {
1076 	struct blk_plug plug;
1077 
1078 	LIST_HEAD(write_list);
1079 
1080 	BUG_ON(dm_bufio_in_request());
1081 
1082 	blk_start_plug(&plug);
1083 	dm_bufio_lock(c);
1084 
1085 	for (; n_blocks--; block++) {
1086 		int need_submit;
1087 		struct dm_buffer *b;
1088 		b = __bufio_new(c, block, NF_PREFETCH, &need_submit,
1089 				&write_list);
1090 		if (unlikely(!list_empty(&write_list))) {
1091 			dm_bufio_unlock(c);
1092 			blk_finish_plug(&plug);
1093 			__flush_write_list(&write_list);
1094 			blk_start_plug(&plug);
1095 			dm_bufio_lock(c);
1096 		}
1097 		if (unlikely(b != NULL)) {
1098 			dm_bufio_unlock(c);
1099 
1100 			if (need_submit)
1101 				submit_io(b, READ, b->block, read_endio);
1102 			dm_bufio_release(b);
1103 
1104 			dm_bufio_cond_resched();
1105 
1106 			if (!n_blocks)
1107 				goto flush_plug;
1108 			dm_bufio_lock(c);
1109 		}
1110 	}
1111 
1112 	dm_bufio_unlock(c);
1113 
1114 flush_plug:
1115 	blk_finish_plug(&plug);
1116 }
1117 EXPORT_SYMBOL_GPL(dm_bufio_prefetch);
1118 
1119 void dm_bufio_release(struct dm_buffer *b)
1120 {
1121 	struct dm_bufio_client *c = b->c;
1122 
1123 	dm_bufio_lock(c);
1124 
1125 	BUG_ON(!b->hold_count);
1126 
1127 	b->hold_count--;
1128 	if (!b->hold_count) {
1129 		wake_up(&c->free_buffer_wait);
1130 
1131 		/*
1132 		 * If there were errors on the buffer, and the buffer is not
1133 		 * to be written, free the buffer. There is no point in caching
1134 		 * invalid buffer.
1135 		 */
1136 		if ((b->read_error || b->write_error) &&
1137 		    !test_bit(B_READING, &b->state) &&
1138 		    !test_bit(B_WRITING, &b->state) &&
1139 		    !test_bit(B_DIRTY, &b->state)) {
1140 			__unlink_buffer(b);
1141 			__free_buffer_wake(b);
1142 		}
1143 	}
1144 
1145 	dm_bufio_unlock(c);
1146 }
1147 EXPORT_SYMBOL_GPL(dm_bufio_release);
1148 
1149 void dm_bufio_mark_buffer_dirty(struct dm_buffer *b)
1150 {
1151 	struct dm_bufio_client *c = b->c;
1152 
1153 	dm_bufio_lock(c);
1154 
1155 	BUG_ON(test_bit(B_READING, &b->state));
1156 
1157 	if (!test_and_set_bit(B_DIRTY, &b->state))
1158 		__relink_lru(b, LIST_DIRTY);
1159 
1160 	dm_bufio_unlock(c);
1161 }
1162 EXPORT_SYMBOL_GPL(dm_bufio_mark_buffer_dirty);
1163 
1164 void dm_bufio_write_dirty_buffers_async(struct dm_bufio_client *c)
1165 {
1166 	LIST_HEAD(write_list);
1167 
1168 	BUG_ON(dm_bufio_in_request());
1169 
1170 	dm_bufio_lock(c);
1171 	__write_dirty_buffers_async(c, 0, &write_list);
1172 	dm_bufio_unlock(c);
1173 	__flush_write_list(&write_list);
1174 }
1175 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers_async);
1176 
1177 /*
1178  * For performance, it is essential that the buffers are written asynchronously
1179  * and simultaneously (so that the block layer can merge the writes) and then
1180  * waited upon.
1181  *
1182  * Finally, we flush hardware disk cache.
1183  */
1184 int dm_bufio_write_dirty_buffers(struct dm_bufio_client *c)
1185 {
1186 	int a, f;
1187 	unsigned long buffers_processed = 0;
1188 	struct dm_buffer *b, *tmp;
1189 
1190 	LIST_HEAD(write_list);
1191 
1192 	dm_bufio_lock(c);
1193 	__write_dirty_buffers_async(c, 0, &write_list);
1194 	dm_bufio_unlock(c);
1195 	__flush_write_list(&write_list);
1196 	dm_bufio_lock(c);
1197 
1198 again:
1199 	list_for_each_entry_safe_reverse(b, tmp, &c->lru[LIST_DIRTY], lru_list) {
1200 		int dropped_lock = 0;
1201 
1202 		if (buffers_processed < c->n_buffers[LIST_DIRTY])
1203 			buffers_processed++;
1204 
1205 		BUG_ON(test_bit(B_READING, &b->state));
1206 
1207 		if (test_bit(B_WRITING, &b->state)) {
1208 			if (buffers_processed < c->n_buffers[LIST_DIRTY]) {
1209 				dropped_lock = 1;
1210 				b->hold_count++;
1211 				dm_bufio_unlock(c);
1212 				wait_on_bit(&b->state, B_WRITING,
1213 					    do_io_schedule,
1214 					    TASK_UNINTERRUPTIBLE);
1215 				dm_bufio_lock(c);
1216 				b->hold_count--;
1217 			} else
1218 				wait_on_bit(&b->state, B_WRITING,
1219 					    do_io_schedule,
1220 					    TASK_UNINTERRUPTIBLE);
1221 		}
1222 
1223 		if (!test_bit(B_DIRTY, &b->state) &&
1224 		    !test_bit(B_WRITING, &b->state))
1225 			__relink_lru(b, LIST_CLEAN);
1226 
1227 		dm_bufio_cond_resched();
1228 
1229 		/*
1230 		 * If we dropped the lock, the list is no longer consistent,
1231 		 * so we must restart the search.
1232 		 *
1233 		 * In the most common case, the buffer just processed is
1234 		 * relinked to the clean list, so we won't loop scanning the
1235 		 * same buffer again and again.
1236 		 *
1237 		 * This may livelock if there is another thread simultaneously
1238 		 * dirtying buffers, so we count the number of buffers walked
1239 		 * and if it exceeds the total number of buffers, it means that
1240 		 * someone is doing some writes simultaneously with us.  In
1241 		 * this case, stop, dropping the lock.
1242 		 */
1243 		if (dropped_lock)
1244 			goto again;
1245 	}
1246 	wake_up(&c->free_buffer_wait);
1247 	dm_bufio_unlock(c);
1248 
1249 	a = xchg(&c->async_write_error, 0);
1250 	f = dm_bufio_issue_flush(c);
1251 	if (a)
1252 		return a;
1253 
1254 	return f;
1255 }
1256 EXPORT_SYMBOL_GPL(dm_bufio_write_dirty_buffers);
1257 
1258 /*
1259  * Use dm-io to send and empty barrier flush the device.
1260  */
1261 int dm_bufio_issue_flush(struct dm_bufio_client *c)
1262 {
1263 	struct dm_io_request io_req = {
1264 		.bi_rw = WRITE_FLUSH,
1265 		.mem.type = DM_IO_KMEM,
1266 		.mem.ptr.addr = NULL,
1267 		.client = c->dm_io,
1268 	};
1269 	struct dm_io_region io_reg = {
1270 		.bdev = c->bdev,
1271 		.sector = 0,
1272 		.count = 0,
1273 	};
1274 
1275 	BUG_ON(dm_bufio_in_request());
1276 
1277 	return dm_io(&io_req, 1, &io_reg, NULL);
1278 }
1279 EXPORT_SYMBOL_GPL(dm_bufio_issue_flush);
1280 
1281 /*
1282  * We first delete any other buffer that may be at that new location.
1283  *
1284  * Then, we write the buffer to the original location if it was dirty.
1285  *
1286  * Then, if we are the only one who is holding the buffer, relink the buffer
1287  * in the hash queue for the new location.
1288  *
1289  * If there was someone else holding the buffer, we write it to the new
1290  * location but not relink it, because that other user needs to have the buffer
1291  * at the same place.
1292  */
1293 void dm_bufio_release_move(struct dm_buffer *b, sector_t new_block)
1294 {
1295 	struct dm_bufio_client *c = b->c;
1296 	struct dm_buffer *new;
1297 
1298 	BUG_ON(dm_bufio_in_request());
1299 
1300 	dm_bufio_lock(c);
1301 
1302 retry:
1303 	new = __find(c, new_block);
1304 	if (new) {
1305 		if (new->hold_count) {
1306 			__wait_for_free_buffer(c);
1307 			goto retry;
1308 		}
1309 
1310 		/*
1311 		 * FIXME: Is there any point waiting for a write that's going
1312 		 * to be overwritten in a bit?
1313 		 */
1314 		__make_buffer_clean(new);
1315 		__unlink_buffer(new);
1316 		__free_buffer_wake(new);
1317 	}
1318 
1319 	BUG_ON(!b->hold_count);
1320 	BUG_ON(test_bit(B_READING, &b->state));
1321 
1322 	__write_dirty_buffer(b, NULL);
1323 	if (b->hold_count == 1) {
1324 		wait_on_bit(&b->state, B_WRITING,
1325 			    do_io_schedule, TASK_UNINTERRUPTIBLE);
1326 		set_bit(B_DIRTY, &b->state);
1327 		__unlink_buffer(b);
1328 		__link_buffer(b, new_block, LIST_DIRTY);
1329 	} else {
1330 		sector_t old_block;
1331 		wait_on_bit_lock(&b->state, B_WRITING,
1332 				 do_io_schedule, TASK_UNINTERRUPTIBLE);
1333 		/*
1334 		 * Relink buffer to "new_block" so that write_callback
1335 		 * sees "new_block" as a block number.
1336 		 * After the write, link the buffer back to old_block.
1337 		 * All this must be done in bufio lock, so that block number
1338 		 * change isn't visible to other threads.
1339 		 */
1340 		old_block = b->block;
1341 		__unlink_buffer(b);
1342 		__link_buffer(b, new_block, b->list_mode);
1343 		submit_io(b, WRITE, new_block, write_endio);
1344 		wait_on_bit(&b->state, B_WRITING,
1345 			    do_io_schedule, TASK_UNINTERRUPTIBLE);
1346 		__unlink_buffer(b);
1347 		__link_buffer(b, old_block, b->list_mode);
1348 	}
1349 
1350 	dm_bufio_unlock(c);
1351 	dm_bufio_release(b);
1352 }
1353 EXPORT_SYMBOL_GPL(dm_bufio_release_move);
1354 
1355 /*
1356  * Free the given buffer.
1357  *
1358  * This is just a hint, if the buffer is in use or dirty, this function
1359  * does nothing.
1360  */
1361 void dm_bufio_forget(struct dm_bufio_client *c, sector_t block)
1362 {
1363 	struct dm_buffer *b;
1364 
1365 	dm_bufio_lock(c);
1366 
1367 	b = __find(c, block);
1368 	if (b && likely(!b->hold_count) && likely(!b->state)) {
1369 		__unlink_buffer(b);
1370 		__free_buffer_wake(b);
1371 	}
1372 
1373 	dm_bufio_unlock(c);
1374 }
1375 EXPORT_SYMBOL(dm_bufio_forget);
1376 
1377 void dm_bufio_set_minimum_buffers(struct dm_bufio_client *c, unsigned n)
1378 {
1379 	c->minimum_buffers = n;
1380 }
1381 EXPORT_SYMBOL(dm_bufio_set_minimum_buffers);
1382 
1383 unsigned dm_bufio_get_block_size(struct dm_bufio_client *c)
1384 {
1385 	return c->block_size;
1386 }
1387 EXPORT_SYMBOL_GPL(dm_bufio_get_block_size);
1388 
1389 sector_t dm_bufio_get_device_size(struct dm_bufio_client *c)
1390 {
1391 	return i_size_read(c->bdev->bd_inode) >>
1392 			   (SECTOR_SHIFT + c->sectors_per_block_bits);
1393 }
1394 EXPORT_SYMBOL_GPL(dm_bufio_get_device_size);
1395 
1396 sector_t dm_bufio_get_block_number(struct dm_buffer *b)
1397 {
1398 	return b->block;
1399 }
1400 EXPORT_SYMBOL_GPL(dm_bufio_get_block_number);
1401 
1402 void *dm_bufio_get_block_data(struct dm_buffer *b)
1403 {
1404 	return b->data;
1405 }
1406 EXPORT_SYMBOL_GPL(dm_bufio_get_block_data);
1407 
1408 void *dm_bufio_get_aux_data(struct dm_buffer *b)
1409 {
1410 	return b + 1;
1411 }
1412 EXPORT_SYMBOL_GPL(dm_bufio_get_aux_data);
1413 
1414 struct dm_bufio_client *dm_bufio_get_client(struct dm_buffer *b)
1415 {
1416 	return b->c;
1417 }
1418 EXPORT_SYMBOL_GPL(dm_bufio_get_client);
1419 
1420 static void drop_buffers(struct dm_bufio_client *c)
1421 {
1422 	struct dm_buffer *b;
1423 	int i;
1424 
1425 	BUG_ON(dm_bufio_in_request());
1426 
1427 	/*
1428 	 * An optimization so that the buffers are not written one-by-one.
1429 	 */
1430 	dm_bufio_write_dirty_buffers_async(c);
1431 
1432 	dm_bufio_lock(c);
1433 
1434 	while ((b = __get_unclaimed_buffer(c)))
1435 		__free_buffer_wake(b);
1436 
1437 	for (i = 0; i < LIST_SIZE; i++)
1438 		list_for_each_entry(b, &c->lru[i], lru_list)
1439 			DMERR("leaked buffer %llx, hold count %u, list %d",
1440 			      (unsigned long long)b->block, b->hold_count, i);
1441 
1442 	for (i = 0; i < LIST_SIZE; i++)
1443 		BUG_ON(!list_empty(&c->lru[i]));
1444 
1445 	dm_bufio_unlock(c);
1446 }
1447 
1448 /*
1449  * Test if the buffer is unused and too old, and commit it.
1450  * At if noio is set, we must not do any I/O because we hold
1451  * dm_bufio_clients_lock and we would risk deadlock if the I/O gets rerouted to
1452  * different bufio client.
1453  */
1454 static int __cleanup_old_buffer(struct dm_buffer *b, gfp_t gfp,
1455 				unsigned long max_jiffies)
1456 {
1457 	if (jiffies - b->last_accessed < max_jiffies)
1458 		return 0;
1459 
1460 	if (!(gfp & __GFP_IO)) {
1461 		if (test_bit(B_READING, &b->state) ||
1462 		    test_bit(B_WRITING, &b->state) ||
1463 		    test_bit(B_DIRTY, &b->state))
1464 			return 0;
1465 	}
1466 
1467 	if (b->hold_count)
1468 		return 0;
1469 
1470 	__make_buffer_clean(b);
1471 	__unlink_buffer(b);
1472 	__free_buffer_wake(b);
1473 
1474 	return 1;
1475 }
1476 
1477 static long __scan(struct dm_bufio_client *c, unsigned long nr_to_scan,
1478 		   gfp_t gfp_mask)
1479 {
1480 	int l;
1481 	struct dm_buffer *b, *tmp;
1482 	long freed = 0;
1483 
1484 	for (l = 0; l < LIST_SIZE; l++) {
1485 		list_for_each_entry_safe_reverse(b, tmp, &c->lru[l], lru_list) {
1486 			freed += __cleanup_old_buffer(b, gfp_mask, 0);
1487 			if (!--nr_to_scan)
1488 				break;
1489 		}
1490 		dm_bufio_cond_resched();
1491 	}
1492 	return freed;
1493 }
1494 
1495 static unsigned long
1496 dm_bufio_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
1497 {
1498 	struct dm_bufio_client *c;
1499 	unsigned long freed;
1500 
1501 	c = container_of(shrink, struct dm_bufio_client, shrinker);
1502 	if (sc->gfp_mask & __GFP_IO)
1503 		dm_bufio_lock(c);
1504 	else if (!dm_bufio_trylock(c))
1505 		return SHRINK_STOP;
1506 
1507 	freed  = __scan(c, sc->nr_to_scan, sc->gfp_mask);
1508 	dm_bufio_unlock(c);
1509 	return freed;
1510 }
1511 
1512 static unsigned long
1513 dm_bufio_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
1514 {
1515 	struct dm_bufio_client *c;
1516 	unsigned long count;
1517 
1518 	c = container_of(shrink, struct dm_bufio_client, shrinker);
1519 	if (sc->gfp_mask & __GFP_IO)
1520 		dm_bufio_lock(c);
1521 	else if (!dm_bufio_trylock(c))
1522 		return 0;
1523 
1524 	count = c->n_buffers[LIST_CLEAN] + c->n_buffers[LIST_DIRTY];
1525 	dm_bufio_unlock(c);
1526 	return count;
1527 }
1528 
1529 /*
1530  * Create the buffering interface
1531  */
1532 struct dm_bufio_client *dm_bufio_client_create(struct block_device *bdev, unsigned block_size,
1533 					       unsigned reserved_buffers, unsigned aux_size,
1534 					       void (*alloc_callback)(struct dm_buffer *),
1535 					       void (*write_callback)(struct dm_buffer *))
1536 {
1537 	int r;
1538 	struct dm_bufio_client *c;
1539 	unsigned i;
1540 
1541 	BUG_ON(block_size < 1 << SECTOR_SHIFT ||
1542 	       (block_size & (block_size - 1)));
1543 
1544 	c = kmalloc(sizeof(*c), GFP_KERNEL);
1545 	if (!c) {
1546 		r = -ENOMEM;
1547 		goto bad_client;
1548 	}
1549 	c->cache_hash = vmalloc(sizeof(struct hlist_head) << DM_BUFIO_HASH_BITS);
1550 	if (!c->cache_hash) {
1551 		r = -ENOMEM;
1552 		goto bad_hash;
1553 	}
1554 
1555 	c->bdev = bdev;
1556 	c->block_size = block_size;
1557 	c->sectors_per_block_bits = ffs(block_size) - 1 - SECTOR_SHIFT;
1558 	c->pages_per_block_bits = (ffs(block_size) - 1 >= PAGE_SHIFT) ?
1559 				  ffs(block_size) - 1 - PAGE_SHIFT : 0;
1560 	c->blocks_per_page_bits = (ffs(block_size) - 1 < PAGE_SHIFT ?
1561 				  PAGE_SHIFT - (ffs(block_size) - 1) : 0);
1562 
1563 	c->aux_size = aux_size;
1564 	c->alloc_callback = alloc_callback;
1565 	c->write_callback = write_callback;
1566 
1567 	for (i = 0; i < LIST_SIZE; i++) {
1568 		INIT_LIST_HEAD(&c->lru[i]);
1569 		c->n_buffers[i] = 0;
1570 	}
1571 
1572 	for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1573 		INIT_HLIST_HEAD(&c->cache_hash[i]);
1574 
1575 	mutex_init(&c->lock);
1576 	INIT_LIST_HEAD(&c->reserved_buffers);
1577 	c->need_reserved_buffers = reserved_buffers;
1578 
1579 	c->minimum_buffers = DM_BUFIO_MIN_BUFFERS;
1580 
1581 	init_waitqueue_head(&c->free_buffer_wait);
1582 	c->async_write_error = 0;
1583 
1584 	c->dm_io = dm_io_client_create();
1585 	if (IS_ERR(c->dm_io)) {
1586 		r = PTR_ERR(c->dm_io);
1587 		goto bad_dm_io;
1588 	}
1589 
1590 	mutex_lock(&dm_bufio_clients_lock);
1591 	if (c->blocks_per_page_bits) {
1592 		if (!DM_BUFIO_CACHE_NAME(c)) {
1593 			DM_BUFIO_CACHE_NAME(c) = kasprintf(GFP_KERNEL, "dm_bufio_cache-%u", c->block_size);
1594 			if (!DM_BUFIO_CACHE_NAME(c)) {
1595 				r = -ENOMEM;
1596 				mutex_unlock(&dm_bufio_clients_lock);
1597 				goto bad_cache;
1598 			}
1599 		}
1600 
1601 		if (!DM_BUFIO_CACHE(c)) {
1602 			DM_BUFIO_CACHE(c) = kmem_cache_create(DM_BUFIO_CACHE_NAME(c),
1603 							      c->block_size,
1604 							      c->block_size, 0, NULL);
1605 			if (!DM_BUFIO_CACHE(c)) {
1606 				r = -ENOMEM;
1607 				mutex_unlock(&dm_bufio_clients_lock);
1608 				goto bad_cache;
1609 			}
1610 		}
1611 	}
1612 	mutex_unlock(&dm_bufio_clients_lock);
1613 
1614 	while (c->need_reserved_buffers) {
1615 		struct dm_buffer *b = alloc_buffer(c, GFP_KERNEL);
1616 
1617 		if (!b) {
1618 			r = -ENOMEM;
1619 			goto bad_buffer;
1620 		}
1621 		__free_buffer_wake(b);
1622 	}
1623 
1624 	mutex_lock(&dm_bufio_clients_lock);
1625 	dm_bufio_client_count++;
1626 	list_add(&c->client_list, &dm_bufio_all_clients);
1627 	__cache_size_refresh();
1628 	mutex_unlock(&dm_bufio_clients_lock);
1629 
1630 	c->shrinker.count_objects = dm_bufio_shrink_count;
1631 	c->shrinker.scan_objects = dm_bufio_shrink_scan;
1632 	c->shrinker.seeks = 1;
1633 	c->shrinker.batch = 0;
1634 	register_shrinker(&c->shrinker);
1635 
1636 	return c;
1637 
1638 bad_buffer:
1639 bad_cache:
1640 	while (!list_empty(&c->reserved_buffers)) {
1641 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1642 						 struct dm_buffer, lru_list);
1643 		list_del(&b->lru_list);
1644 		free_buffer(b);
1645 	}
1646 	dm_io_client_destroy(c->dm_io);
1647 bad_dm_io:
1648 	vfree(c->cache_hash);
1649 bad_hash:
1650 	kfree(c);
1651 bad_client:
1652 	return ERR_PTR(r);
1653 }
1654 EXPORT_SYMBOL_GPL(dm_bufio_client_create);
1655 
1656 /*
1657  * Free the buffering interface.
1658  * It is required that there are no references on any buffers.
1659  */
1660 void dm_bufio_client_destroy(struct dm_bufio_client *c)
1661 {
1662 	unsigned i;
1663 
1664 	drop_buffers(c);
1665 
1666 	unregister_shrinker(&c->shrinker);
1667 
1668 	mutex_lock(&dm_bufio_clients_lock);
1669 
1670 	list_del(&c->client_list);
1671 	dm_bufio_client_count--;
1672 	__cache_size_refresh();
1673 
1674 	mutex_unlock(&dm_bufio_clients_lock);
1675 
1676 	for (i = 0; i < 1 << DM_BUFIO_HASH_BITS; i++)
1677 		BUG_ON(!hlist_empty(&c->cache_hash[i]));
1678 
1679 	BUG_ON(c->need_reserved_buffers);
1680 
1681 	while (!list_empty(&c->reserved_buffers)) {
1682 		struct dm_buffer *b = list_entry(c->reserved_buffers.next,
1683 						 struct dm_buffer, lru_list);
1684 		list_del(&b->lru_list);
1685 		free_buffer(b);
1686 	}
1687 
1688 	for (i = 0; i < LIST_SIZE; i++)
1689 		if (c->n_buffers[i])
1690 			DMERR("leaked buffer count %d: %ld", i, c->n_buffers[i]);
1691 
1692 	for (i = 0; i < LIST_SIZE; i++)
1693 		BUG_ON(c->n_buffers[i]);
1694 
1695 	dm_io_client_destroy(c->dm_io);
1696 	vfree(c->cache_hash);
1697 	kfree(c);
1698 }
1699 EXPORT_SYMBOL_GPL(dm_bufio_client_destroy);
1700 
1701 static void cleanup_old_buffers(void)
1702 {
1703 	unsigned long max_age = ACCESS_ONCE(dm_bufio_max_age);
1704 	struct dm_bufio_client *c;
1705 
1706 	if (max_age > ULONG_MAX / HZ)
1707 		max_age = ULONG_MAX / HZ;
1708 
1709 	mutex_lock(&dm_bufio_clients_lock);
1710 	list_for_each_entry(c, &dm_bufio_all_clients, client_list) {
1711 		if (!dm_bufio_trylock(c))
1712 			continue;
1713 
1714 		while (!list_empty(&c->lru[LIST_CLEAN])) {
1715 			struct dm_buffer *b;
1716 			b = list_entry(c->lru[LIST_CLEAN].prev,
1717 				       struct dm_buffer, lru_list);
1718 			if (!__cleanup_old_buffer(b, 0, max_age * HZ))
1719 				break;
1720 			dm_bufio_cond_resched();
1721 		}
1722 
1723 		dm_bufio_unlock(c);
1724 		dm_bufio_cond_resched();
1725 	}
1726 	mutex_unlock(&dm_bufio_clients_lock);
1727 }
1728 
1729 static struct workqueue_struct *dm_bufio_wq;
1730 static struct delayed_work dm_bufio_work;
1731 
1732 static void work_fn(struct work_struct *w)
1733 {
1734 	cleanup_old_buffers();
1735 
1736 	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1737 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1738 }
1739 
1740 /*----------------------------------------------------------------
1741  * Module setup
1742  *--------------------------------------------------------------*/
1743 
1744 /*
1745  * This is called only once for the whole dm_bufio module.
1746  * It initializes memory limit.
1747  */
1748 static int __init dm_bufio_init(void)
1749 {
1750 	__u64 mem;
1751 
1752 	dm_bufio_allocated_kmem_cache = 0;
1753 	dm_bufio_allocated_get_free_pages = 0;
1754 	dm_bufio_allocated_vmalloc = 0;
1755 	dm_bufio_current_allocated = 0;
1756 
1757 	memset(&dm_bufio_caches, 0, sizeof dm_bufio_caches);
1758 	memset(&dm_bufio_cache_names, 0, sizeof dm_bufio_cache_names);
1759 
1760 	mem = (__u64)((totalram_pages - totalhigh_pages) *
1761 		      DM_BUFIO_MEMORY_PERCENT / 100) << PAGE_SHIFT;
1762 
1763 	if (mem > ULONG_MAX)
1764 		mem = ULONG_MAX;
1765 
1766 #ifdef CONFIG_MMU
1767 	/*
1768 	 * Get the size of vmalloc space the same way as VMALLOC_TOTAL
1769 	 * in fs/proc/internal.h
1770 	 */
1771 	if (mem > (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100)
1772 		mem = (VMALLOC_END - VMALLOC_START) * DM_BUFIO_VMALLOC_PERCENT / 100;
1773 #endif
1774 
1775 	dm_bufio_default_cache_size = mem;
1776 
1777 	mutex_lock(&dm_bufio_clients_lock);
1778 	__cache_size_refresh();
1779 	mutex_unlock(&dm_bufio_clients_lock);
1780 
1781 	dm_bufio_wq = create_singlethread_workqueue("dm_bufio_cache");
1782 	if (!dm_bufio_wq)
1783 		return -ENOMEM;
1784 
1785 	INIT_DELAYED_WORK(&dm_bufio_work, work_fn);
1786 	queue_delayed_work(dm_bufio_wq, &dm_bufio_work,
1787 			   DM_BUFIO_WORK_TIMER_SECS * HZ);
1788 
1789 	return 0;
1790 }
1791 
1792 /*
1793  * This is called once when unloading the dm_bufio module.
1794  */
1795 static void __exit dm_bufio_exit(void)
1796 {
1797 	int bug = 0;
1798 	int i;
1799 
1800 	cancel_delayed_work_sync(&dm_bufio_work);
1801 	destroy_workqueue(dm_bufio_wq);
1802 
1803 	for (i = 0; i < ARRAY_SIZE(dm_bufio_caches); i++) {
1804 		struct kmem_cache *kc = dm_bufio_caches[i];
1805 
1806 		if (kc)
1807 			kmem_cache_destroy(kc);
1808 	}
1809 
1810 	for (i = 0; i < ARRAY_SIZE(dm_bufio_cache_names); i++)
1811 		kfree(dm_bufio_cache_names[i]);
1812 
1813 	if (dm_bufio_client_count) {
1814 		DMCRIT("%s: dm_bufio_client_count leaked: %d",
1815 			__func__, dm_bufio_client_count);
1816 		bug = 1;
1817 	}
1818 
1819 	if (dm_bufio_current_allocated) {
1820 		DMCRIT("%s: dm_bufio_current_allocated leaked: %lu",
1821 			__func__, dm_bufio_current_allocated);
1822 		bug = 1;
1823 	}
1824 
1825 	if (dm_bufio_allocated_get_free_pages) {
1826 		DMCRIT("%s: dm_bufio_allocated_get_free_pages leaked: %lu",
1827 		       __func__, dm_bufio_allocated_get_free_pages);
1828 		bug = 1;
1829 	}
1830 
1831 	if (dm_bufio_allocated_vmalloc) {
1832 		DMCRIT("%s: dm_bufio_vmalloc leaked: %lu",
1833 		       __func__, dm_bufio_allocated_vmalloc);
1834 		bug = 1;
1835 	}
1836 
1837 	if (bug)
1838 		BUG();
1839 }
1840 
1841 module_init(dm_bufio_init)
1842 module_exit(dm_bufio_exit)
1843 
1844 module_param_named(max_cache_size_bytes, dm_bufio_cache_size, ulong, S_IRUGO | S_IWUSR);
1845 MODULE_PARM_DESC(max_cache_size_bytes, "Size of metadata cache");
1846 
1847 module_param_named(max_age_seconds, dm_bufio_max_age, uint, S_IRUGO | S_IWUSR);
1848 MODULE_PARM_DESC(max_age_seconds, "Max age of a buffer in seconds");
1849 
1850 module_param_named(peak_allocated_bytes, dm_bufio_peak_allocated, ulong, S_IRUGO | S_IWUSR);
1851 MODULE_PARM_DESC(peak_allocated_bytes, "Tracks the maximum allocated memory");
1852 
1853 module_param_named(allocated_kmem_cache_bytes, dm_bufio_allocated_kmem_cache, ulong, S_IRUGO);
1854 MODULE_PARM_DESC(allocated_kmem_cache_bytes, "Memory allocated with kmem_cache_alloc");
1855 
1856 module_param_named(allocated_get_free_pages_bytes, dm_bufio_allocated_get_free_pages, ulong, S_IRUGO);
1857 MODULE_PARM_DESC(allocated_get_free_pages_bytes, "Memory allocated with get_free_pages");
1858 
1859 module_param_named(allocated_vmalloc_bytes, dm_bufio_allocated_vmalloc, ulong, S_IRUGO);
1860 MODULE_PARM_DESC(allocated_vmalloc_bytes, "Memory allocated with vmalloc");
1861 
1862 module_param_named(current_allocated_bytes, dm_bufio_current_allocated, ulong, S_IRUGO);
1863 MODULE_PARM_DESC(current_allocated_bytes, "Memory currently used by the cache");
1864 
1865 MODULE_AUTHOR("Mikulas Patocka <dm-devel@redhat.com>");
1866 MODULE_DESCRIPTION(DM_NAME " buffered I/O library");
1867 MODULE_LICENSE("GPL");
1868