1 /*
2  * Copyright (C) 2012 Red Hat, Inc.
3  *
4  * This file is released under the GPL.
5  */
6 
7 #include "dm-cache-metadata.h"
8 
9 #include "persistent-data/dm-array.h"
10 #include "persistent-data/dm-bitset.h"
11 #include "persistent-data/dm-space-map.h"
12 #include "persistent-data/dm-space-map-disk.h"
13 #include "persistent-data/dm-transaction-manager.h"
14 
15 #include <linux/device-mapper.h>
16 #include <linux/refcount.h>
17 
18 /*----------------------------------------------------------------*/
19 
20 #define DM_MSG_PREFIX   "cache metadata"
21 
22 #define CACHE_SUPERBLOCK_MAGIC 06142003
23 #define CACHE_SUPERBLOCK_LOCATION 0
24 
25 /*
26  * defines a range of metadata versions that this module can handle.
27  */
28 #define MIN_CACHE_VERSION 1
29 #define MAX_CACHE_VERSION 2
30 
31 /*
32  *  3 for btree insert +
33  *  2 for btree lookup used within space map
34  */
35 #define CACHE_MAX_CONCURRENT_LOCKS 5
36 #define SPACE_MAP_ROOT_SIZE 128
37 
38 enum superblock_flag_bits {
39 	/* for spotting crashes that would invalidate the dirty bitset */
40 	CLEAN_SHUTDOWN,
41 	/* metadata must be checked using the tools */
42 	NEEDS_CHECK,
43 };
44 
45 /*
46  * Each mapping from cache block -> origin block carries a set of flags.
47  */
48 enum mapping_bits {
49 	/*
50 	 * A valid mapping.  Because we're using an array we clear this
51 	 * flag for an non existant mapping.
52 	 */
53 	M_VALID = 1,
54 
55 	/*
56 	 * The data on the cache is different from that on the origin.
57 	 * This flag is only used by metadata format 1.
58 	 */
59 	M_DIRTY = 2
60 };
61 
62 struct cache_disk_superblock {
63 	__le32 csum;
64 	__le32 flags;
65 	__le64 blocknr;
66 
67 	__u8 uuid[16];
68 	__le64 magic;
69 	__le32 version;
70 
71 	__u8 policy_name[CACHE_POLICY_NAME_SIZE];
72 	__le32 policy_hint_size;
73 
74 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
75 	__le64 mapping_root;
76 	__le64 hint_root;
77 
78 	__le64 discard_root;
79 	__le64 discard_block_size;
80 	__le64 discard_nr_blocks;
81 
82 	__le32 data_block_size;
83 	__le32 metadata_block_size;
84 	__le32 cache_blocks;
85 
86 	__le32 compat_flags;
87 	__le32 compat_ro_flags;
88 	__le32 incompat_flags;
89 
90 	__le32 read_hits;
91 	__le32 read_misses;
92 	__le32 write_hits;
93 	__le32 write_misses;
94 
95 	__le32 policy_version[CACHE_POLICY_VERSION_SIZE];
96 
97 	/*
98 	 * Metadata format 2 fields.
99 	 */
100 	__le64 dirty_root;
101 } __packed;
102 
103 struct dm_cache_metadata {
104 	refcount_t ref_count;
105 	struct list_head list;
106 
107 	unsigned version;
108 	struct block_device *bdev;
109 	struct dm_block_manager *bm;
110 	struct dm_space_map *metadata_sm;
111 	struct dm_transaction_manager *tm;
112 
113 	struct dm_array_info info;
114 	struct dm_array_info hint_info;
115 	struct dm_disk_bitset discard_info;
116 
117 	struct rw_semaphore root_lock;
118 	unsigned long flags;
119 	dm_block_t root;
120 	dm_block_t hint_root;
121 	dm_block_t discard_root;
122 
123 	sector_t discard_block_size;
124 	dm_dblock_t discard_nr_blocks;
125 
126 	sector_t data_block_size;
127 	dm_cblock_t cache_blocks;
128 	bool changed:1;
129 	bool clean_when_opened:1;
130 
131 	char policy_name[CACHE_POLICY_NAME_SIZE];
132 	unsigned policy_version[CACHE_POLICY_VERSION_SIZE];
133 	size_t policy_hint_size;
134 	struct dm_cache_statistics stats;
135 
136 	/*
137 	 * Reading the space map root can fail, so we read it into this
138 	 * buffer before the superblock is locked and updated.
139 	 */
140 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
141 
142 	/*
143 	 * Set if a transaction has to be aborted but the attempt to roll
144 	 * back to the previous (good) transaction failed.  The only
145 	 * metadata operation permissible in this state is the closing of
146 	 * the device.
147 	 */
148 	bool fail_io:1;
149 
150 	/*
151 	 * Metadata format 2 fields.
152 	 */
153 	dm_block_t dirty_root;
154 	struct dm_disk_bitset dirty_info;
155 
156 	/*
157 	 * These structures are used when loading metadata.  They're too
158 	 * big to put on the stack.
159 	 */
160 	struct dm_array_cursor mapping_cursor;
161 	struct dm_array_cursor hint_cursor;
162 	struct dm_bitset_cursor dirty_cursor;
163 };
164 
165 /*-------------------------------------------------------------------
166  * superblock validator
167  *-----------------------------------------------------------------*/
168 
169 #define SUPERBLOCK_CSUM_XOR 9031977
170 
171 static void sb_prepare_for_write(struct dm_block_validator *v,
172 				 struct dm_block *b,
173 				 size_t sb_block_size)
174 {
175 	struct cache_disk_superblock *disk_super = dm_block_data(b);
176 
177 	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
178 	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
179 						      sb_block_size - sizeof(__le32),
180 						      SUPERBLOCK_CSUM_XOR));
181 }
182 
183 static int check_metadata_version(struct cache_disk_superblock *disk_super)
184 {
185 	uint32_t metadata_version = le32_to_cpu(disk_super->version);
186 
187 	if (metadata_version < MIN_CACHE_VERSION || metadata_version > MAX_CACHE_VERSION) {
188 		DMERR("Cache metadata version %u found, but only versions between %u and %u supported.",
189 		      metadata_version, MIN_CACHE_VERSION, MAX_CACHE_VERSION);
190 		return -EINVAL;
191 	}
192 
193 	return 0;
194 }
195 
196 static int sb_check(struct dm_block_validator *v,
197 		    struct dm_block *b,
198 		    size_t sb_block_size)
199 {
200 	struct cache_disk_superblock *disk_super = dm_block_data(b);
201 	__le32 csum_le;
202 
203 	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
204 		DMERR("sb_check failed: blocknr %llu: wanted %llu",
205 		      le64_to_cpu(disk_super->blocknr),
206 		      (unsigned long long)dm_block_location(b));
207 		return -ENOTBLK;
208 	}
209 
210 	if (le64_to_cpu(disk_super->magic) != CACHE_SUPERBLOCK_MAGIC) {
211 		DMERR("sb_check failed: magic %llu: wanted %llu",
212 		      le64_to_cpu(disk_super->magic),
213 		      (unsigned long long)CACHE_SUPERBLOCK_MAGIC);
214 		return -EILSEQ;
215 	}
216 
217 	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
218 					     sb_block_size - sizeof(__le32),
219 					     SUPERBLOCK_CSUM_XOR));
220 	if (csum_le != disk_super->csum) {
221 		DMERR("sb_check failed: csum %u: wanted %u",
222 		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
223 		return -EILSEQ;
224 	}
225 
226 	return check_metadata_version(disk_super);
227 }
228 
229 static struct dm_block_validator sb_validator = {
230 	.name = "superblock",
231 	.prepare_for_write = sb_prepare_for_write,
232 	.check = sb_check
233 };
234 
235 /*----------------------------------------------------------------*/
236 
237 static int superblock_read_lock(struct dm_cache_metadata *cmd,
238 				struct dm_block **sblock)
239 {
240 	return dm_bm_read_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
241 			       &sb_validator, sblock);
242 }
243 
244 static int superblock_lock_zero(struct dm_cache_metadata *cmd,
245 				struct dm_block **sblock)
246 {
247 	return dm_bm_write_lock_zero(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
248 				     &sb_validator, sblock);
249 }
250 
251 static int superblock_lock(struct dm_cache_metadata *cmd,
252 			   struct dm_block **sblock)
253 {
254 	return dm_bm_write_lock(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
255 				&sb_validator, sblock);
256 }
257 
258 /*----------------------------------------------------------------*/
259 
260 static int __superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
261 {
262 	int r;
263 	unsigned i;
264 	struct dm_block *b;
265 	__le64 *data_le, zero = cpu_to_le64(0);
266 	unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
267 
268 	/*
269 	 * We can't use a validator here - it may be all zeroes.
270 	 */
271 	r = dm_bm_read_lock(bm, CACHE_SUPERBLOCK_LOCATION, NULL, &b);
272 	if (r)
273 		return r;
274 
275 	data_le = dm_block_data(b);
276 	*result = true;
277 	for (i = 0; i < sb_block_size; i++) {
278 		if (data_le[i] != zero) {
279 			*result = false;
280 			break;
281 		}
282 	}
283 
284 	dm_bm_unlock(b);
285 
286 	return 0;
287 }
288 
289 static void __setup_mapping_info(struct dm_cache_metadata *cmd)
290 {
291 	struct dm_btree_value_type vt;
292 
293 	vt.context = NULL;
294 	vt.size = sizeof(__le64);
295 	vt.inc = NULL;
296 	vt.dec = NULL;
297 	vt.equal = NULL;
298 	dm_array_info_init(&cmd->info, cmd->tm, &vt);
299 
300 	if (cmd->policy_hint_size) {
301 		vt.size = sizeof(__le32);
302 		dm_array_info_init(&cmd->hint_info, cmd->tm, &vt);
303 	}
304 }
305 
306 static int __save_sm_root(struct dm_cache_metadata *cmd)
307 {
308 	int r;
309 	size_t metadata_len;
310 
311 	r = dm_sm_root_size(cmd->metadata_sm, &metadata_len);
312 	if (r < 0)
313 		return r;
314 
315 	return dm_sm_copy_root(cmd->metadata_sm, &cmd->metadata_space_map_root,
316 			       metadata_len);
317 }
318 
319 static void __copy_sm_root(struct dm_cache_metadata *cmd,
320 			   struct cache_disk_superblock *disk_super)
321 {
322 	memcpy(&disk_super->metadata_space_map_root,
323 	       &cmd->metadata_space_map_root,
324 	       sizeof(cmd->metadata_space_map_root));
325 }
326 
327 static bool separate_dirty_bits(struct dm_cache_metadata *cmd)
328 {
329 	return cmd->version >= 2;
330 }
331 
332 static int __write_initial_superblock(struct dm_cache_metadata *cmd)
333 {
334 	int r;
335 	struct dm_block *sblock;
336 	struct cache_disk_superblock *disk_super;
337 	sector_t bdev_size = i_size_read(cmd->bdev->bd_inode) >> SECTOR_SHIFT;
338 
339 	/* FIXME: see if we can lose the max sectors limit */
340 	if (bdev_size > DM_CACHE_METADATA_MAX_SECTORS)
341 		bdev_size = DM_CACHE_METADATA_MAX_SECTORS;
342 
343 	r = dm_tm_pre_commit(cmd->tm);
344 	if (r < 0)
345 		return r;
346 
347 	/*
348 	 * dm_sm_copy_root() can fail.  So we need to do it before we start
349 	 * updating the superblock.
350 	 */
351 	r = __save_sm_root(cmd);
352 	if (r)
353 		return r;
354 
355 	r = superblock_lock_zero(cmd, &sblock);
356 	if (r)
357 		return r;
358 
359 	disk_super = dm_block_data(sblock);
360 	disk_super->flags = 0;
361 	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
362 	disk_super->magic = cpu_to_le64(CACHE_SUPERBLOCK_MAGIC);
363 	disk_super->version = cpu_to_le32(cmd->version);
364 	memset(disk_super->policy_name, 0, sizeof(disk_super->policy_name));
365 	memset(disk_super->policy_version, 0, sizeof(disk_super->policy_version));
366 	disk_super->policy_hint_size = cpu_to_le32(0);
367 
368 	__copy_sm_root(cmd, disk_super);
369 
370 	disk_super->mapping_root = cpu_to_le64(cmd->root);
371 	disk_super->hint_root = cpu_to_le64(cmd->hint_root);
372 	disk_super->discard_root = cpu_to_le64(cmd->discard_root);
373 	disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
374 	disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
375 	disk_super->metadata_block_size = cpu_to_le32(DM_CACHE_METADATA_BLOCK_SIZE);
376 	disk_super->data_block_size = cpu_to_le32(cmd->data_block_size);
377 	disk_super->cache_blocks = cpu_to_le32(0);
378 
379 	disk_super->read_hits = cpu_to_le32(0);
380 	disk_super->read_misses = cpu_to_le32(0);
381 	disk_super->write_hits = cpu_to_le32(0);
382 	disk_super->write_misses = cpu_to_le32(0);
383 
384 	if (separate_dirty_bits(cmd))
385 		disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
386 
387 	return dm_tm_commit(cmd->tm, sblock);
388 }
389 
390 static int __format_metadata(struct dm_cache_metadata *cmd)
391 {
392 	int r;
393 
394 	r = dm_tm_create_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
395 				 &cmd->tm, &cmd->metadata_sm);
396 	if (r < 0) {
397 		DMERR("tm_create_with_sm failed");
398 		return r;
399 	}
400 
401 	__setup_mapping_info(cmd);
402 
403 	r = dm_array_empty(&cmd->info, &cmd->root);
404 	if (r < 0)
405 		goto bad;
406 
407 	if (separate_dirty_bits(cmd)) {
408 		dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
409 		r = dm_bitset_empty(&cmd->dirty_info, &cmd->dirty_root);
410 		if (r < 0)
411 			goto bad;
412 	}
413 
414 	dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
415 	r = dm_bitset_empty(&cmd->discard_info, &cmd->discard_root);
416 	if (r < 0)
417 		goto bad;
418 
419 	cmd->discard_block_size = 0;
420 	cmd->discard_nr_blocks = 0;
421 
422 	r = __write_initial_superblock(cmd);
423 	if (r)
424 		goto bad;
425 
426 	cmd->clean_when_opened = true;
427 	return 0;
428 
429 bad:
430 	dm_tm_destroy(cmd->tm);
431 	dm_sm_destroy(cmd->metadata_sm);
432 
433 	return r;
434 }
435 
436 static int __check_incompat_features(struct cache_disk_superblock *disk_super,
437 				     struct dm_cache_metadata *cmd)
438 {
439 	uint32_t incompat_flags, features;
440 
441 	incompat_flags = le32_to_cpu(disk_super->incompat_flags);
442 	features = incompat_flags & ~DM_CACHE_FEATURE_INCOMPAT_SUPP;
443 	if (features) {
444 		DMERR("could not access metadata due to unsupported optional features (%lx).",
445 		      (unsigned long)features);
446 		return -EINVAL;
447 	}
448 
449 	/*
450 	 * Check for read-only metadata to skip the following RDWR checks.
451 	 */
452 	if (bdev_read_only(cmd->bdev))
453 		return 0;
454 
455 	features = le32_to_cpu(disk_super->compat_ro_flags) & ~DM_CACHE_FEATURE_COMPAT_RO_SUPP;
456 	if (features) {
457 		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
458 		      (unsigned long)features);
459 		return -EINVAL;
460 	}
461 
462 	return 0;
463 }
464 
465 static int __open_metadata(struct dm_cache_metadata *cmd)
466 {
467 	int r;
468 	struct dm_block *sblock;
469 	struct cache_disk_superblock *disk_super;
470 	unsigned long sb_flags;
471 
472 	r = superblock_read_lock(cmd, &sblock);
473 	if (r < 0) {
474 		DMERR("couldn't read lock superblock");
475 		return r;
476 	}
477 
478 	disk_super = dm_block_data(sblock);
479 
480 	/* Verify the data block size hasn't changed */
481 	if (le32_to_cpu(disk_super->data_block_size) != cmd->data_block_size) {
482 		DMERR("changing the data block size (from %u to %llu) is not supported",
483 		      le32_to_cpu(disk_super->data_block_size),
484 		      (unsigned long long)cmd->data_block_size);
485 		r = -EINVAL;
486 		goto bad;
487 	}
488 
489 	r = __check_incompat_features(disk_super, cmd);
490 	if (r < 0)
491 		goto bad;
492 
493 	r = dm_tm_open_with_sm(cmd->bm, CACHE_SUPERBLOCK_LOCATION,
494 			       disk_super->metadata_space_map_root,
495 			       sizeof(disk_super->metadata_space_map_root),
496 			       &cmd->tm, &cmd->metadata_sm);
497 	if (r < 0) {
498 		DMERR("tm_open_with_sm failed");
499 		goto bad;
500 	}
501 
502 	__setup_mapping_info(cmd);
503 	dm_disk_bitset_init(cmd->tm, &cmd->dirty_info);
504 	dm_disk_bitset_init(cmd->tm, &cmd->discard_info);
505 	sb_flags = le32_to_cpu(disk_super->flags);
506 	cmd->clean_when_opened = test_bit(CLEAN_SHUTDOWN, &sb_flags);
507 	dm_bm_unlock(sblock);
508 
509 	return 0;
510 
511 bad:
512 	dm_bm_unlock(sblock);
513 	return r;
514 }
515 
516 static int __open_or_format_metadata(struct dm_cache_metadata *cmd,
517 				     bool format_device)
518 {
519 	int r;
520 	bool unformatted = false;
521 
522 	r = __superblock_all_zeroes(cmd->bm, &unformatted);
523 	if (r)
524 		return r;
525 
526 	if (unformatted)
527 		return format_device ? __format_metadata(cmd) : -EPERM;
528 
529 	return __open_metadata(cmd);
530 }
531 
532 static int __create_persistent_data_objects(struct dm_cache_metadata *cmd,
533 					    bool may_format_device)
534 {
535 	int r;
536 	cmd->bm = dm_block_manager_create(cmd->bdev, DM_CACHE_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
537 					  CACHE_MAX_CONCURRENT_LOCKS);
538 	if (IS_ERR(cmd->bm)) {
539 		DMERR("could not create block manager");
540 		r = PTR_ERR(cmd->bm);
541 		cmd->bm = NULL;
542 		return r;
543 	}
544 
545 	r = __open_or_format_metadata(cmd, may_format_device);
546 	if (r) {
547 		dm_block_manager_destroy(cmd->bm);
548 		cmd->bm = NULL;
549 	}
550 
551 	return r;
552 }
553 
554 static void __destroy_persistent_data_objects(struct dm_cache_metadata *cmd)
555 {
556 	dm_sm_destroy(cmd->metadata_sm);
557 	dm_tm_destroy(cmd->tm);
558 	dm_block_manager_destroy(cmd->bm);
559 }
560 
561 typedef unsigned long (*flags_mutator)(unsigned long);
562 
563 static void update_flags(struct cache_disk_superblock *disk_super,
564 			 flags_mutator mutator)
565 {
566 	uint32_t sb_flags = mutator(le32_to_cpu(disk_super->flags));
567 	disk_super->flags = cpu_to_le32(sb_flags);
568 }
569 
570 static unsigned long set_clean_shutdown(unsigned long flags)
571 {
572 	set_bit(CLEAN_SHUTDOWN, &flags);
573 	return flags;
574 }
575 
576 static unsigned long clear_clean_shutdown(unsigned long flags)
577 {
578 	clear_bit(CLEAN_SHUTDOWN, &flags);
579 	return flags;
580 }
581 
582 static void read_superblock_fields(struct dm_cache_metadata *cmd,
583 				   struct cache_disk_superblock *disk_super)
584 {
585 	cmd->version = le32_to_cpu(disk_super->version);
586 	cmd->flags = le32_to_cpu(disk_super->flags);
587 	cmd->root = le64_to_cpu(disk_super->mapping_root);
588 	cmd->hint_root = le64_to_cpu(disk_super->hint_root);
589 	cmd->discard_root = le64_to_cpu(disk_super->discard_root);
590 	cmd->discard_block_size = le64_to_cpu(disk_super->discard_block_size);
591 	cmd->discard_nr_blocks = to_dblock(le64_to_cpu(disk_super->discard_nr_blocks));
592 	cmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
593 	cmd->cache_blocks = to_cblock(le32_to_cpu(disk_super->cache_blocks));
594 	strncpy(cmd->policy_name, disk_super->policy_name, sizeof(cmd->policy_name));
595 	cmd->policy_version[0] = le32_to_cpu(disk_super->policy_version[0]);
596 	cmd->policy_version[1] = le32_to_cpu(disk_super->policy_version[1]);
597 	cmd->policy_version[2] = le32_to_cpu(disk_super->policy_version[2]);
598 	cmd->policy_hint_size = le32_to_cpu(disk_super->policy_hint_size);
599 
600 	cmd->stats.read_hits = le32_to_cpu(disk_super->read_hits);
601 	cmd->stats.read_misses = le32_to_cpu(disk_super->read_misses);
602 	cmd->stats.write_hits = le32_to_cpu(disk_super->write_hits);
603 	cmd->stats.write_misses = le32_to_cpu(disk_super->write_misses);
604 
605 	if (separate_dirty_bits(cmd))
606 		cmd->dirty_root = le64_to_cpu(disk_super->dirty_root);
607 
608 	cmd->changed = false;
609 }
610 
611 /*
612  * The mutator updates the superblock flags.
613  */
614 static int __begin_transaction_flags(struct dm_cache_metadata *cmd,
615 				     flags_mutator mutator)
616 {
617 	int r;
618 	struct cache_disk_superblock *disk_super;
619 	struct dm_block *sblock;
620 
621 	r = superblock_lock(cmd, &sblock);
622 	if (r)
623 		return r;
624 
625 	disk_super = dm_block_data(sblock);
626 	update_flags(disk_super, mutator);
627 	read_superblock_fields(cmd, disk_super);
628 	dm_bm_unlock(sblock);
629 
630 	return dm_bm_flush(cmd->bm);
631 }
632 
633 static int __begin_transaction(struct dm_cache_metadata *cmd)
634 {
635 	int r;
636 	struct cache_disk_superblock *disk_super;
637 	struct dm_block *sblock;
638 
639 	/*
640 	 * We re-read the superblock every time.  Shouldn't need to do this
641 	 * really.
642 	 */
643 	r = superblock_read_lock(cmd, &sblock);
644 	if (r)
645 		return r;
646 
647 	disk_super = dm_block_data(sblock);
648 	read_superblock_fields(cmd, disk_super);
649 	dm_bm_unlock(sblock);
650 
651 	return 0;
652 }
653 
654 static int __commit_transaction(struct dm_cache_metadata *cmd,
655 				flags_mutator mutator)
656 {
657 	int r;
658 	struct cache_disk_superblock *disk_super;
659 	struct dm_block *sblock;
660 
661 	/*
662 	 * We need to know if the cache_disk_superblock exceeds a 512-byte sector.
663 	 */
664 	BUILD_BUG_ON(sizeof(struct cache_disk_superblock) > 512);
665 
666 	if (separate_dirty_bits(cmd)) {
667 		r = dm_bitset_flush(&cmd->dirty_info, cmd->dirty_root,
668 				    &cmd->dirty_root);
669 		if (r)
670 			return r;
671 	}
672 
673 	r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root,
674 			    &cmd->discard_root);
675 	if (r)
676 		return r;
677 
678 	r = dm_tm_pre_commit(cmd->tm);
679 	if (r < 0)
680 		return r;
681 
682 	r = __save_sm_root(cmd);
683 	if (r)
684 		return r;
685 
686 	r = superblock_lock(cmd, &sblock);
687 	if (r)
688 		return r;
689 
690 	disk_super = dm_block_data(sblock);
691 
692 	disk_super->flags = cpu_to_le32(cmd->flags);
693 	if (mutator)
694 		update_flags(disk_super, mutator);
695 
696 	disk_super->mapping_root = cpu_to_le64(cmd->root);
697 	if (separate_dirty_bits(cmd))
698 		disk_super->dirty_root = cpu_to_le64(cmd->dirty_root);
699 	disk_super->hint_root = cpu_to_le64(cmd->hint_root);
700 	disk_super->discard_root = cpu_to_le64(cmd->discard_root);
701 	disk_super->discard_block_size = cpu_to_le64(cmd->discard_block_size);
702 	disk_super->discard_nr_blocks = cpu_to_le64(from_dblock(cmd->discard_nr_blocks));
703 	disk_super->cache_blocks = cpu_to_le32(from_cblock(cmd->cache_blocks));
704 	strncpy(disk_super->policy_name, cmd->policy_name, sizeof(disk_super->policy_name));
705 	disk_super->policy_version[0] = cpu_to_le32(cmd->policy_version[0]);
706 	disk_super->policy_version[1] = cpu_to_le32(cmd->policy_version[1]);
707 	disk_super->policy_version[2] = cpu_to_le32(cmd->policy_version[2]);
708 	disk_super->policy_hint_size = cpu_to_le32(cmd->policy_hint_size);
709 
710 	disk_super->read_hits = cpu_to_le32(cmd->stats.read_hits);
711 	disk_super->read_misses = cpu_to_le32(cmd->stats.read_misses);
712 	disk_super->write_hits = cpu_to_le32(cmd->stats.write_hits);
713 	disk_super->write_misses = cpu_to_le32(cmd->stats.write_misses);
714 	__copy_sm_root(cmd, disk_super);
715 
716 	return dm_tm_commit(cmd->tm, sblock);
717 }
718 
719 /*----------------------------------------------------------------*/
720 
721 /*
722  * The mappings are held in a dm-array that has 64-bit values stored in
723  * little-endian format.  The index is the cblock, the high 48bits of the
724  * value are the oblock and the low 16 bit the flags.
725  */
726 #define FLAGS_MASK ((1 << 16) - 1)
727 
728 static __le64 pack_value(dm_oblock_t block, unsigned flags)
729 {
730 	uint64_t value = from_oblock(block);
731 	value <<= 16;
732 	value = value | (flags & FLAGS_MASK);
733 	return cpu_to_le64(value);
734 }
735 
736 static void unpack_value(__le64 value_le, dm_oblock_t *block, unsigned *flags)
737 {
738 	uint64_t value = le64_to_cpu(value_le);
739 	uint64_t b = value >> 16;
740 	*block = to_oblock(b);
741 	*flags = value & FLAGS_MASK;
742 }
743 
744 /*----------------------------------------------------------------*/
745 
746 static struct dm_cache_metadata *metadata_open(struct block_device *bdev,
747 					       sector_t data_block_size,
748 					       bool may_format_device,
749 					       size_t policy_hint_size,
750 					       unsigned metadata_version)
751 {
752 	int r;
753 	struct dm_cache_metadata *cmd;
754 
755 	cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
756 	if (!cmd) {
757 		DMERR("could not allocate metadata struct");
758 		return ERR_PTR(-ENOMEM);
759 	}
760 
761 	cmd->version = metadata_version;
762 	refcount_set(&cmd->ref_count, 1);
763 	init_rwsem(&cmd->root_lock);
764 	cmd->bdev = bdev;
765 	cmd->data_block_size = data_block_size;
766 	cmd->cache_blocks = 0;
767 	cmd->policy_hint_size = policy_hint_size;
768 	cmd->changed = true;
769 	cmd->fail_io = false;
770 
771 	r = __create_persistent_data_objects(cmd, may_format_device);
772 	if (r) {
773 		kfree(cmd);
774 		return ERR_PTR(r);
775 	}
776 
777 	r = __begin_transaction_flags(cmd, clear_clean_shutdown);
778 	if (r < 0) {
779 		dm_cache_metadata_close(cmd);
780 		return ERR_PTR(r);
781 	}
782 
783 	return cmd;
784 }
785 
786 /*
787  * We keep a little list of ref counted metadata objects to prevent two
788  * different target instances creating separate bufio instances.  This is
789  * an issue if a table is reloaded before the suspend.
790  */
791 static DEFINE_MUTEX(table_lock);
792 static LIST_HEAD(table);
793 
794 static struct dm_cache_metadata *lookup(struct block_device *bdev)
795 {
796 	struct dm_cache_metadata *cmd;
797 
798 	list_for_each_entry(cmd, &table, list)
799 		if (cmd->bdev == bdev) {
800 			refcount_inc(&cmd->ref_count);
801 			return cmd;
802 		}
803 
804 	return NULL;
805 }
806 
807 static struct dm_cache_metadata *lookup_or_open(struct block_device *bdev,
808 						sector_t data_block_size,
809 						bool may_format_device,
810 						size_t policy_hint_size,
811 						unsigned metadata_version)
812 {
813 	struct dm_cache_metadata *cmd, *cmd2;
814 
815 	mutex_lock(&table_lock);
816 	cmd = lookup(bdev);
817 	mutex_unlock(&table_lock);
818 
819 	if (cmd)
820 		return cmd;
821 
822 	cmd = metadata_open(bdev, data_block_size, may_format_device,
823 			    policy_hint_size, metadata_version);
824 	if (!IS_ERR(cmd)) {
825 		mutex_lock(&table_lock);
826 		cmd2 = lookup(bdev);
827 		if (cmd2) {
828 			mutex_unlock(&table_lock);
829 			__destroy_persistent_data_objects(cmd);
830 			kfree(cmd);
831 			return cmd2;
832 		}
833 		list_add(&cmd->list, &table);
834 		mutex_unlock(&table_lock);
835 	}
836 
837 	return cmd;
838 }
839 
840 static bool same_params(struct dm_cache_metadata *cmd, sector_t data_block_size)
841 {
842 	if (cmd->data_block_size != data_block_size) {
843 		DMERR("data_block_size (%llu) different from that in metadata (%llu)",
844 		      (unsigned long long) data_block_size,
845 		      (unsigned long long) cmd->data_block_size);
846 		return false;
847 	}
848 
849 	return true;
850 }
851 
852 struct dm_cache_metadata *dm_cache_metadata_open(struct block_device *bdev,
853 						 sector_t data_block_size,
854 						 bool may_format_device,
855 						 size_t policy_hint_size,
856 						 unsigned metadata_version)
857 {
858 	struct dm_cache_metadata *cmd = lookup_or_open(bdev, data_block_size, may_format_device,
859 						       policy_hint_size, metadata_version);
860 
861 	if (!IS_ERR(cmd) && !same_params(cmd, data_block_size)) {
862 		dm_cache_metadata_close(cmd);
863 		return ERR_PTR(-EINVAL);
864 	}
865 
866 	return cmd;
867 }
868 
869 void dm_cache_metadata_close(struct dm_cache_metadata *cmd)
870 {
871 	if (refcount_dec_and_test(&cmd->ref_count)) {
872 		mutex_lock(&table_lock);
873 		list_del(&cmd->list);
874 		mutex_unlock(&table_lock);
875 
876 		if (!cmd->fail_io)
877 			__destroy_persistent_data_objects(cmd);
878 		kfree(cmd);
879 	}
880 }
881 
882 /*
883  * Checks that the given cache block is either unmapped or clean.
884  */
885 static int block_clean_combined_dirty(struct dm_cache_metadata *cmd, dm_cblock_t b,
886 				      bool *result)
887 {
888 	int r;
889 	__le64 value;
890 	dm_oblock_t ob;
891 	unsigned flags;
892 
893 	r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(b), &value);
894 	if (r)
895 		return r;
896 
897 	unpack_value(value, &ob, &flags);
898 	*result = !((flags & M_VALID) && (flags & M_DIRTY));
899 
900 	return 0;
901 }
902 
903 static int blocks_are_clean_combined_dirty(struct dm_cache_metadata *cmd,
904 					   dm_cblock_t begin, dm_cblock_t end,
905 					   bool *result)
906 {
907 	int r;
908 	*result = true;
909 
910 	while (begin != end) {
911 		r = block_clean_combined_dirty(cmd, begin, result);
912 		if (r) {
913 			DMERR("block_clean_combined_dirty failed");
914 			return r;
915 		}
916 
917 		if (!*result) {
918 			DMERR("cache block %llu is dirty",
919 			      (unsigned long long) from_cblock(begin));
920 			return 0;
921 		}
922 
923 		begin = to_cblock(from_cblock(begin) + 1);
924 	}
925 
926 	return 0;
927 }
928 
929 static int blocks_are_clean_separate_dirty(struct dm_cache_metadata *cmd,
930 					   dm_cblock_t begin, dm_cblock_t end,
931 					   bool *result)
932 {
933 	int r;
934 	bool dirty_flag;
935 	*result = true;
936 
937 	if (from_cblock(cmd->cache_blocks) == 0)
938 		/* Nothing to do */
939 		return 0;
940 
941 	r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
942 				   from_cblock(cmd->cache_blocks), &cmd->dirty_cursor);
943 	if (r) {
944 		DMERR("%s: dm_bitset_cursor_begin for dirty failed", __func__);
945 		return r;
946 	}
947 
948 	r = dm_bitset_cursor_skip(&cmd->dirty_cursor, from_cblock(begin));
949 	if (r) {
950 		DMERR("%s: dm_bitset_cursor_skip for dirty failed", __func__);
951 		dm_bitset_cursor_end(&cmd->dirty_cursor);
952 		return r;
953 	}
954 
955 	while (begin != end) {
956 		/*
957 		 * We assume that unmapped blocks have their dirty bit
958 		 * cleared.
959 		 */
960 		dirty_flag = dm_bitset_cursor_get_value(&cmd->dirty_cursor);
961 		if (dirty_flag) {
962 			DMERR("%s: cache block %llu is dirty", __func__,
963 			      (unsigned long long) from_cblock(begin));
964 			dm_bitset_cursor_end(&cmd->dirty_cursor);
965 			*result = false;
966 			return 0;
967 		}
968 
969 		begin = to_cblock(from_cblock(begin) + 1);
970 		if (begin == end)
971 			break;
972 
973 		r = dm_bitset_cursor_next(&cmd->dirty_cursor);
974 		if (r) {
975 			DMERR("%s: dm_bitset_cursor_next for dirty failed", __func__);
976 			dm_bitset_cursor_end(&cmd->dirty_cursor);
977 			return r;
978 		}
979 	}
980 
981 	dm_bitset_cursor_end(&cmd->dirty_cursor);
982 
983 	return 0;
984 }
985 
986 static int blocks_are_unmapped_or_clean(struct dm_cache_metadata *cmd,
987 					dm_cblock_t begin, dm_cblock_t end,
988 					bool *result)
989 {
990 	if (separate_dirty_bits(cmd))
991 		return blocks_are_clean_separate_dirty(cmd, begin, end, result);
992 	else
993 		return blocks_are_clean_combined_dirty(cmd, begin, end, result);
994 }
995 
996 static bool cmd_write_lock(struct dm_cache_metadata *cmd)
997 {
998 	down_write(&cmd->root_lock);
999 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
1000 		up_write(&cmd->root_lock);
1001 		return false;
1002 	}
1003 	return true;
1004 }
1005 
1006 #define WRITE_LOCK(cmd)				\
1007 	do {					\
1008 		if (!cmd_write_lock((cmd)))	\
1009 			return -EINVAL;		\
1010 	} while(0)
1011 
1012 #define WRITE_LOCK_VOID(cmd)			\
1013 	do {					\
1014 		if (!cmd_write_lock((cmd)))	\
1015 			return;			\
1016 	} while(0)
1017 
1018 #define WRITE_UNLOCK(cmd) \
1019 	up_write(&(cmd)->root_lock)
1020 
1021 static bool cmd_read_lock(struct dm_cache_metadata *cmd)
1022 {
1023 	down_read(&cmd->root_lock);
1024 	if (cmd->fail_io) {
1025 		up_read(&cmd->root_lock);
1026 		return false;
1027 	}
1028 	return true;
1029 }
1030 
1031 #define READ_LOCK(cmd)				\
1032 	do {					\
1033 		if (!cmd_read_lock((cmd)))	\
1034 			return -EINVAL;		\
1035 	} while(0)
1036 
1037 #define READ_LOCK_VOID(cmd)			\
1038 	do {					\
1039 		if (!cmd_read_lock((cmd)))	\
1040 			return;			\
1041 	} while(0)
1042 
1043 #define READ_UNLOCK(cmd) \
1044 	up_read(&(cmd)->root_lock)
1045 
1046 int dm_cache_resize(struct dm_cache_metadata *cmd, dm_cblock_t new_cache_size)
1047 {
1048 	int r;
1049 	bool clean;
1050 	__le64 null_mapping = pack_value(0, 0);
1051 
1052 	WRITE_LOCK(cmd);
1053 	__dm_bless_for_disk(&null_mapping);
1054 
1055 	if (from_cblock(new_cache_size) < from_cblock(cmd->cache_blocks)) {
1056 		r = blocks_are_unmapped_or_clean(cmd, new_cache_size, cmd->cache_blocks, &clean);
1057 		if (r) {
1058 			__dm_unbless_for_disk(&null_mapping);
1059 			goto out;
1060 		}
1061 
1062 		if (!clean) {
1063 			DMERR("unable to shrink cache due to dirty blocks");
1064 			r = -EINVAL;
1065 			__dm_unbless_for_disk(&null_mapping);
1066 			goto out;
1067 		}
1068 	}
1069 
1070 	r = dm_array_resize(&cmd->info, cmd->root, from_cblock(cmd->cache_blocks),
1071 			    from_cblock(new_cache_size),
1072 			    &null_mapping, &cmd->root);
1073 	if (r)
1074 		goto out;
1075 
1076 	if (separate_dirty_bits(cmd)) {
1077 		r = dm_bitset_resize(&cmd->dirty_info, cmd->dirty_root,
1078 				     from_cblock(cmd->cache_blocks), from_cblock(new_cache_size),
1079 				     false, &cmd->dirty_root);
1080 		if (r)
1081 			goto out;
1082 	}
1083 
1084 	cmd->cache_blocks = new_cache_size;
1085 	cmd->changed = true;
1086 
1087 out:
1088 	WRITE_UNLOCK(cmd);
1089 
1090 	return r;
1091 }
1092 
1093 int dm_cache_discard_bitset_resize(struct dm_cache_metadata *cmd,
1094 				   sector_t discard_block_size,
1095 				   dm_dblock_t new_nr_entries)
1096 {
1097 	int r;
1098 
1099 	WRITE_LOCK(cmd);
1100 	r = dm_bitset_resize(&cmd->discard_info,
1101 			     cmd->discard_root,
1102 			     from_dblock(cmd->discard_nr_blocks),
1103 			     from_dblock(new_nr_entries),
1104 			     false, &cmd->discard_root);
1105 	if (!r) {
1106 		cmd->discard_block_size = discard_block_size;
1107 		cmd->discard_nr_blocks = new_nr_entries;
1108 	}
1109 
1110 	cmd->changed = true;
1111 	WRITE_UNLOCK(cmd);
1112 
1113 	return r;
1114 }
1115 
1116 static int __set_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1117 {
1118 	return dm_bitset_set_bit(&cmd->discard_info, cmd->discard_root,
1119 				 from_dblock(b), &cmd->discard_root);
1120 }
1121 
1122 static int __clear_discard(struct dm_cache_metadata *cmd, dm_dblock_t b)
1123 {
1124 	return dm_bitset_clear_bit(&cmd->discard_info, cmd->discard_root,
1125 				   from_dblock(b), &cmd->discard_root);
1126 }
1127 
1128 static int __discard(struct dm_cache_metadata *cmd,
1129 		     dm_dblock_t dblock, bool discard)
1130 {
1131 	int r;
1132 
1133 	r = (discard ? __set_discard : __clear_discard)(cmd, dblock);
1134 	if (r)
1135 		return r;
1136 
1137 	cmd->changed = true;
1138 	return 0;
1139 }
1140 
1141 int dm_cache_set_discard(struct dm_cache_metadata *cmd,
1142 			 dm_dblock_t dblock, bool discard)
1143 {
1144 	int r;
1145 
1146 	WRITE_LOCK(cmd);
1147 	r = __discard(cmd, dblock, discard);
1148 	WRITE_UNLOCK(cmd);
1149 
1150 	return r;
1151 }
1152 
1153 static int __load_discards(struct dm_cache_metadata *cmd,
1154 			   load_discard_fn fn, void *context)
1155 {
1156 	int r = 0;
1157 	uint32_t b;
1158 	struct dm_bitset_cursor c;
1159 
1160 	if (from_dblock(cmd->discard_nr_blocks) == 0)
1161 		/* nothing to do */
1162 		return 0;
1163 
1164 	if (cmd->clean_when_opened) {
1165 		r = dm_bitset_flush(&cmd->discard_info, cmd->discard_root, &cmd->discard_root);
1166 		if (r)
1167 			return r;
1168 
1169 		r = dm_bitset_cursor_begin(&cmd->discard_info, cmd->discard_root,
1170 					   from_dblock(cmd->discard_nr_blocks), &c);
1171 		if (r)
1172 			return r;
1173 
1174 		for (b = 0; ; b++) {
1175 			r = fn(context, cmd->discard_block_size, to_dblock(b),
1176 			       dm_bitset_cursor_get_value(&c));
1177 			if (r)
1178 				break;
1179 
1180 			if (b >= (from_dblock(cmd->discard_nr_blocks) - 1))
1181 				break;
1182 
1183 			r = dm_bitset_cursor_next(&c);
1184 			if (r)
1185 				break;
1186 		}
1187 
1188 		dm_bitset_cursor_end(&c);
1189 
1190 	} else {
1191 		for (b = 0; b < from_dblock(cmd->discard_nr_blocks); b++) {
1192 			r = fn(context, cmd->discard_block_size, to_dblock(b), false);
1193 			if (r)
1194 				return r;
1195 		}
1196 	}
1197 
1198 	return r;
1199 }
1200 
1201 int dm_cache_load_discards(struct dm_cache_metadata *cmd,
1202 			   load_discard_fn fn, void *context)
1203 {
1204 	int r;
1205 
1206 	READ_LOCK(cmd);
1207 	r = __load_discards(cmd, fn, context);
1208 	READ_UNLOCK(cmd);
1209 
1210 	return r;
1211 }
1212 
1213 int dm_cache_size(struct dm_cache_metadata *cmd, dm_cblock_t *result)
1214 {
1215 	READ_LOCK(cmd);
1216 	*result = cmd->cache_blocks;
1217 	READ_UNLOCK(cmd);
1218 
1219 	return 0;
1220 }
1221 
1222 static int __remove(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1223 {
1224 	int r;
1225 	__le64 value = pack_value(0, 0);
1226 
1227 	__dm_bless_for_disk(&value);
1228 	r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1229 			       &value, &cmd->root);
1230 	if (r)
1231 		return r;
1232 
1233 	cmd->changed = true;
1234 	return 0;
1235 }
1236 
1237 int dm_cache_remove_mapping(struct dm_cache_metadata *cmd, dm_cblock_t cblock)
1238 {
1239 	int r;
1240 
1241 	WRITE_LOCK(cmd);
1242 	r = __remove(cmd, cblock);
1243 	WRITE_UNLOCK(cmd);
1244 
1245 	return r;
1246 }
1247 
1248 static int __insert(struct dm_cache_metadata *cmd,
1249 		    dm_cblock_t cblock, dm_oblock_t oblock)
1250 {
1251 	int r;
1252 	__le64 value = pack_value(oblock, M_VALID);
1253 	__dm_bless_for_disk(&value);
1254 
1255 	r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1256 			       &value, &cmd->root);
1257 	if (r)
1258 		return r;
1259 
1260 	cmd->changed = true;
1261 	return 0;
1262 }
1263 
1264 int dm_cache_insert_mapping(struct dm_cache_metadata *cmd,
1265 			    dm_cblock_t cblock, dm_oblock_t oblock)
1266 {
1267 	int r;
1268 
1269 	WRITE_LOCK(cmd);
1270 	r = __insert(cmd, cblock, oblock);
1271 	WRITE_UNLOCK(cmd);
1272 
1273 	return r;
1274 }
1275 
1276 struct thunk {
1277 	load_mapping_fn fn;
1278 	void *context;
1279 
1280 	struct dm_cache_metadata *cmd;
1281 	bool respect_dirty_flags;
1282 	bool hints_valid;
1283 };
1284 
1285 static bool policy_unchanged(struct dm_cache_metadata *cmd,
1286 			     struct dm_cache_policy *policy)
1287 {
1288 	const char *policy_name = dm_cache_policy_get_name(policy);
1289 	const unsigned *policy_version = dm_cache_policy_get_version(policy);
1290 	size_t policy_hint_size = dm_cache_policy_get_hint_size(policy);
1291 
1292 	/*
1293 	 * Ensure policy names match.
1294 	 */
1295 	if (strncmp(cmd->policy_name, policy_name, sizeof(cmd->policy_name)))
1296 		return false;
1297 
1298 	/*
1299 	 * Ensure policy major versions match.
1300 	 */
1301 	if (cmd->policy_version[0] != policy_version[0])
1302 		return false;
1303 
1304 	/*
1305 	 * Ensure policy hint sizes match.
1306 	 */
1307 	if (cmd->policy_hint_size != policy_hint_size)
1308 		return false;
1309 
1310 	return true;
1311 }
1312 
1313 static bool hints_array_initialized(struct dm_cache_metadata *cmd)
1314 {
1315 	return cmd->hint_root && cmd->policy_hint_size;
1316 }
1317 
1318 static bool hints_array_available(struct dm_cache_metadata *cmd,
1319 				  struct dm_cache_policy *policy)
1320 {
1321 	return cmd->clean_when_opened && policy_unchanged(cmd, policy) &&
1322 		hints_array_initialized(cmd);
1323 }
1324 
1325 static int __load_mapping_v1(struct dm_cache_metadata *cmd,
1326 			     uint64_t cb, bool hints_valid,
1327 			     struct dm_array_cursor *mapping_cursor,
1328 			     struct dm_array_cursor *hint_cursor,
1329 			     load_mapping_fn fn, void *context)
1330 {
1331 	int r = 0;
1332 
1333 	__le64 mapping;
1334 	__le32 hint = 0;
1335 
1336 	__le64 *mapping_value_le;
1337 	__le32 *hint_value_le;
1338 
1339 	dm_oblock_t oblock;
1340 	unsigned flags;
1341 	bool dirty = true;
1342 
1343 	dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1344 	memcpy(&mapping, mapping_value_le, sizeof(mapping));
1345 	unpack_value(mapping, &oblock, &flags);
1346 
1347 	if (flags & M_VALID) {
1348 		if (hints_valid) {
1349 			dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1350 			memcpy(&hint, hint_value_le, sizeof(hint));
1351 		}
1352 		if (cmd->clean_when_opened)
1353 			dirty = flags & M_DIRTY;
1354 
1355 		r = fn(context, oblock, to_cblock(cb), dirty,
1356 		       le32_to_cpu(hint), hints_valid);
1357 		if (r) {
1358 			DMERR("policy couldn't load cache block %llu",
1359 			      (unsigned long long) from_cblock(to_cblock(cb)));
1360 		}
1361 	}
1362 
1363 	return r;
1364 }
1365 
1366 static int __load_mapping_v2(struct dm_cache_metadata *cmd,
1367 			     uint64_t cb, bool hints_valid,
1368 			     struct dm_array_cursor *mapping_cursor,
1369 			     struct dm_array_cursor *hint_cursor,
1370 			     struct dm_bitset_cursor *dirty_cursor,
1371 			     load_mapping_fn fn, void *context)
1372 {
1373 	int r = 0;
1374 
1375 	__le64 mapping;
1376 	__le32 hint = 0;
1377 
1378 	__le64 *mapping_value_le;
1379 	__le32 *hint_value_le;
1380 
1381 	dm_oblock_t oblock;
1382 	unsigned flags;
1383 	bool dirty = true;
1384 
1385 	dm_array_cursor_get_value(mapping_cursor, (void **) &mapping_value_le);
1386 	memcpy(&mapping, mapping_value_le, sizeof(mapping));
1387 	unpack_value(mapping, &oblock, &flags);
1388 
1389 	if (flags & M_VALID) {
1390 		if (hints_valid) {
1391 			dm_array_cursor_get_value(hint_cursor, (void **) &hint_value_le);
1392 			memcpy(&hint, hint_value_le, sizeof(hint));
1393 		}
1394 		if (cmd->clean_when_opened)
1395 			dirty = dm_bitset_cursor_get_value(dirty_cursor);
1396 
1397 		r = fn(context, oblock, to_cblock(cb), dirty,
1398 		       le32_to_cpu(hint), hints_valid);
1399 		if (r) {
1400 			DMERR("policy couldn't load cache block %llu",
1401 			      (unsigned long long) from_cblock(to_cblock(cb)));
1402 		}
1403 	}
1404 
1405 	return r;
1406 }
1407 
1408 static int __load_mappings(struct dm_cache_metadata *cmd,
1409 			   struct dm_cache_policy *policy,
1410 			   load_mapping_fn fn, void *context)
1411 {
1412 	int r;
1413 	uint64_t cb;
1414 
1415 	bool hints_valid = hints_array_available(cmd, policy);
1416 
1417 	if (from_cblock(cmd->cache_blocks) == 0)
1418 		/* Nothing to do */
1419 		return 0;
1420 
1421 	r = dm_array_cursor_begin(&cmd->info, cmd->root, &cmd->mapping_cursor);
1422 	if (r)
1423 		return r;
1424 
1425 	if (hints_valid) {
1426 		r = dm_array_cursor_begin(&cmd->hint_info, cmd->hint_root, &cmd->hint_cursor);
1427 		if (r) {
1428 			dm_array_cursor_end(&cmd->mapping_cursor);
1429 			return r;
1430 		}
1431 	}
1432 
1433 	if (separate_dirty_bits(cmd)) {
1434 		r = dm_bitset_cursor_begin(&cmd->dirty_info, cmd->dirty_root,
1435 					   from_cblock(cmd->cache_blocks),
1436 					   &cmd->dirty_cursor);
1437 		if (r) {
1438 			dm_array_cursor_end(&cmd->hint_cursor);
1439 			dm_array_cursor_end(&cmd->mapping_cursor);
1440 			return r;
1441 		}
1442 	}
1443 
1444 	for (cb = 0; ; cb++) {
1445 		if (separate_dirty_bits(cmd))
1446 			r = __load_mapping_v2(cmd, cb, hints_valid,
1447 					      &cmd->mapping_cursor,
1448 					      &cmd->hint_cursor,
1449 					      &cmd->dirty_cursor,
1450 					      fn, context);
1451 		else
1452 			r = __load_mapping_v1(cmd, cb, hints_valid,
1453 					      &cmd->mapping_cursor, &cmd->hint_cursor,
1454 					      fn, context);
1455 		if (r)
1456 			goto out;
1457 
1458 		/*
1459 		 * We need to break out before we move the cursors.
1460 		 */
1461 		if (cb >= (from_cblock(cmd->cache_blocks) - 1))
1462 			break;
1463 
1464 		r = dm_array_cursor_next(&cmd->mapping_cursor);
1465 		if (r) {
1466 			DMERR("dm_array_cursor_next for mapping failed");
1467 			goto out;
1468 		}
1469 
1470 		if (hints_valid) {
1471 			r = dm_array_cursor_next(&cmd->hint_cursor);
1472 			if (r) {
1473 				dm_array_cursor_end(&cmd->hint_cursor);
1474 				hints_valid = false;
1475 			}
1476 		}
1477 
1478 		if (separate_dirty_bits(cmd)) {
1479 			r = dm_bitset_cursor_next(&cmd->dirty_cursor);
1480 			if (r) {
1481 				DMERR("dm_bitset_cursor_next for dirty failed");
1482 				goto out;
1483 			}
1484 		}
1485 	}
1486 out:
1487 	dm_array_cursor_end(&cmd->mapping_cursor);
1488 	if (hints_valid)
1489 		dm_array_cursor_end(&cmd->hint_cursor);
1490 
1491 	if (separate_dirty_bits(cmd))
1492 		dm_bitset_cursor_end(&cmd->dirty_cursor);
1493 
1494 	return r;
1495 }
1496 
1497 int dm_cache_load_mappings(struct dm_cache_metadata *cmd,
1498 			   struct dm_cache_policy *policy,
1499 			   load_mapping_fn fn, void *context)
1500 {
1501 	int r;
1502 
1503 	READ_LOCK(cmd);
1504 	r = __load_mappings(cmd, policy, fn, context);
1505 	READ_UNLOCK(cmd);
1506 
1507 	return r;
1508 }
1509 
1510 static int __dump_mapping(void *context, uint64_t cblock, void *leaf)
1511 {
1512 	int r = 0;
1513 	__le64 value;
1514 	dm_oblock_t oblock;
1515 	unsigned flags;
1516 
1517 	memcpy(&value, leaf, sizeof(value));
1518 	unpack_value(value, &oblock, &flags);
1519 
1520 	return r;
1521 }
1522 
1523 static int __dump_mappings(struct dm_cache_metadata *cmd)
1524 {
1525 	return dm_array_walk(&cmd->info, cmd->root, __dump_mapping, NULL);
1526 }
1527 
1528 void dm_cache_dump(struct dm_cache_metadata *cmd)
1529 {
1530 	READ_LOCK_VOID(cmd);
1531 	__dump_mappings(cmd);
1532 	READ_UNLOCK(cmd);
1533 }
1534 
1535 int dm_cache_changed_this_transaction(struct dm_cache_metadata *cmd)
1536 {
1537 	int r;
1538 
1539 	READ_LOCK(cmd);
1540 	r = cmd->changed;
1541 	READ_UNLOCK(cmd);
1542 
1543 	return r;
1544 }
1545 
1546 static int __dirty(struct dm_cache_metadata *cmd, dm_cblock_t cblock, bool dirty)
1547 {
1548 	int r;
1549 	unsigned flags;
1550 	dm_oblock_t oblock;
1551 	__le64 value;
1552 
1553 	r = dm_array_get_value(&cmd->info, cmd->root, from_cblock(cblock), &value);
1554 	if (r)
1555 		return r;
1556 
1557 	unpack_value(value, &oblock, &flags);
1558 
1559 	if (((flags & M_DIRTY) && dirty) || (!(flags & M_DIRTY) && !dirty))
1560 		/* nothing to be done */
1561 		return 0;
1562 
1563 	value = pack_value(oblock, (flags & ~M_DIRTY) | (dirty ? M_DIRTY : 0));
1564 	__dm_bless_for_disk(&value);
1565 
1566 	r = dm_array_set_value(&cmd->info, cmd->root, from_cblock(cblock),
1567 			       &value, &cmd->root);
1568 	if (r)
1569 		return r;
1570 
1571 	cmd->changed = true;
1572 	return 0;
1573 
1574 }
1575 
1576 static int __set_dirty_bits_v1(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1577 {
1578 	int r;
1579 	unsigned i;
1580 	for (i = 0; i < nr_bits; i++) {
1581 		r = __dirty(cmd, to_cblock(i), test_bit(i, bits));
1582 		if (r)
1583 			return r;
1584 	}
1585 
1586 	return 0;
1587 }
1588 
1589 static int is_dirty_callback(uint32_t index, bool *value, void *context)
1590 {
1591 	unsigned long *bits = context;
1592 	*value = test_bit(index, bits);
1593 	return 0;
1594 }
1595 
1596 static int __set_dirty_bits_v2(struct dm_cache_metadata *cmd, unsigned nr_bits, unsigned long *bits)
1597 {
1598 	int r = 0;
1599 
1600 	/* nr_bits is really just a sanity check */
1601 	if (nr_bits != from_cblock(cmd->cache_blocks)) {
1602 		DMERR("dirty bitset is wrong size");
1603 		return -EINVAL;
1604 	}
1605 
1606 	r = dm_bitset_del(&cmd->dirty_info, cmd->dirty_root);
1607 	if (r)
1608 		return r;
1609 
1610 	cmd->changed = true;
1611 	return dm_bitset_new(&cmd->dirty_info, &cmd->dirty_root, nr_bits, is_dirty_callback, bits);
1612 }
1613 
1614 int dm_cache_set_dirty_bits(struct dm_cache_metadata *cmd,
1615 			    unsigned nr_bits,
1616 			    unsigned long *bits)
1617 {
1618 	int r;
1619 
1620 	WRITE_LOCK(cmd);
1621 	if (separate_dirty_bits(cmd))
1622 		r = __set_dirty_bits_v2(cmd, nr_bits, bits);
1623 	else
1624 		r = __set_dirty_bits_v1(cmd, nr_bits, bits);
1625 	WRITE_UNLOCK(cmd);
1626 
1627 	return r;
1628 }
1629 
1630 void dm_cache_metadata_get_stats(struct dm_cache_metadata *cmd,
1631 				 struct dm_cache_statistics *stats)
1632 {
1633 	READ_LOCK_VOID(cmd);
1634 	*stats = cmd->stats;
1635 	READ_UNLOCK(cmd);
1636 }
1637 
1638 void dm_cache_metadata_set_stats(struct dm_cache_metadata *cmd,
1639 				 struct dm_cache_statistics *stats)
1640 {
1641 	WRITE_LOCK_VOID(cmd);
1642 	cmd->stats = *stats;
1643 	WRITE_UNLOCK(cmd);
1644 }
1645 
1646 int dm_cache_commit(struct dm_cache_metadata *cmd, bool clean_shutdown)
1647 {
1648 	int r = -EINVAL;
1649 	flags_mutator mutator = (clean_shutdown ? set_clean_shutdown :
1650 				 clear_clean_shutdown);
1651 
1652 	WRITE_LOCK(cmd);
1653 	if (cmd->fail_io)
1654 		goto out;
1655 
1656 	r = __commit_transaction(cmd, mutator);
1657 	if (r)
1658 		goto out;
1659 
1660 	r = __begin_transaction(cmd);
1661 out:
1662 	WRITE_UNLOCK(cmd);
1663 	return r;
1664 }
1665 
1666 int dm_cache_get_free_metadata_block_count(struct dm_cache_metadata *cmd,
1667 					   dm_block_t *result)
1668 {
1669 	int r = -EINVAL;
1670 
1671 	READ_LOCK(cmd);
1672 	if (!cmd->fail_io)
1673 		r = dm_sm_get_nr_free(cmd->metadata_sm, result);
1674 	READ_UNLOCK(cmd);
1675 
1676 	return r;
1677 }
1678 
1679 int dm_cache_get_metadata_dev_size(struct dm_cache_metadata *cmd,
1680 				   dm_block_t *result)
1681 {
1682 	int r = -EINVAL;
1683 
1684 	READ_LOCK(cmd);
1685 	if (!cmd->fail_io)
1686 		r = dm_sm_get_nr_blocks(cmd->metadata_sm, result);
1687 	READ_UNLOCK(cmd);
1688 
1689 	return r;
1690 }
1691 
1692 /*----------------------------------------------------------------*/
1693 
1694 static int get_hint(uint32_t index, void *value_le, void *context)
1695 {
1696 	uint32_t value;
1697 	struct dm_cache_policy *policy = context;
1698 
1699 	value = policy_get_hint(policy, to_cblock(index));
1700 	*((__le32 *) value_le) = cpu_to_le32(value);
1701 
1702 	return 0;
1703 }
1704 
1705 /*
1706  * It's quicker to always delete the hint array, and recreate with
1707  * dm_array_new().
1708  */
1709 static int write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1710 {
1711 	int r;
1712 	size_t hint_size;
1713 	const char *policy_name = dm_cache_policy_get_name(policy);
1714 	const unsigned *policy_version = dm_cache_policy_get_version(policy);
1715 
1716 	if (!policy_name[0] ||
1717 	    (strlen(policy_name) > sizeof(cmd->policy_name) - 1))
1718 		return -EINVAL;
1719 
1720 	strncpy(cmd->policy_name, policy_name, sizeof(cmd->policy_name));
1721 	memcpy(cmd->policy_version, policy_version, sizeof(cmd->policy_version));
1722 
1723 	hint_size = dm_cache_policy_get_hint_size(policy);
1724 	if (!hint_size)
1725 		return 0; /* short-circuit hints initialization */
1726 	cmd->policy_hint_size = hint_size;
1727 
1728 	if (cmd->hint_root) {
1729 		r = dm_array_del(&cmd->hint_info, cmd->hint_root);
1730 		if (r)
1731 			return r;
1732 	}
1733 
1734 	return dm_array_new(&cmd->hint_info, &cmd->hint_root,
1735 			    from_cblock(cmd->cache_blocks),
1736 			    get_hint, policy);
1737 }
1738 
1739 int dm_cache_write_hints(struct dm_cache_metadata *cmd, struct dm_cache_policy *policy)
1740 {
1741 	int r;
1742 
1743 	WRITE_LOCK(cmd);
1744 	r = write_hints(cmd, policy);
1745 	WRITE_UNLOCK(cmd);
1746 
1747 	return r;
1748 }
1749 
1750 int dm_cache_metadata_all_clean(struct dm_cache_metadata *cmd, bool *result)
1751 {
1752 	int r;
1753 
1754 	READ_LOCK(cmd);
1755 	r = blocks_are_unmapped_or_clean(cmd, 0, cmd->cache_blocks, result);
1756 	READ_UNLOCK(cmd);
1757 
1758 	return r;
1759 }
1760 
1761 void dm_cache_metadata_set_read_only(struct dm_cache_metadata *cmd)
1762 {
1763 	WRITE_LOCK_VOID(cmd);
1764 	dm_bm_set_read_only(cmd->bm);
1765 	WRITE_UNLOCK(cmd);
1766 }
1767 
1768 void dm_cache_metadata_set_read_write(struct dm_cache_metadata *cmd)
1769 {
1770 	WRITE_LOCK_VOID(cmd);
1771 	dm_bm_set_read_write(cmd->bm);
1772 	WRITE_UNLOCK(cmd);
1773 }
1774 
1775 int dm_cache_metadata_set_needs_check(struct dm_cache_metadata *cmd)
1776 {
1777 	int r;
1778 	struct dm_block *sblock;
1779 	struct cache_disk_superblock *disk_super;
1780 
1781 	WRITE_LOCK(cmd);
1782 	set_bit(NEEDS_CHECK, &cmd->flags);
1783 
1784 	r = superblock_lock(cmd, &sblock);
1785 	if (r) {
1786 		DMERR("couldn't read superblock");
1787 		goto out;
1788 	}
1789 
1790 	disk_super = dm_block_data(sblock);
1791 	disk_super->flags = cpu_to_le32(cmd->flags);
1792 
1793 	dm_bm_unlock(sblock);
1794 
1795 out:
1796 	WRITE_UNLOCK(cmd);
1797 	return r;
1798 }
1799 
1800 int dm_cache_metadata_needs_check(struct dm_cache_metadata *cmd, bool *result)
1801 {
1802 	READ_LOCK(cmd);
1803 	*result = !!test_bit(NEEDS_CHECK, &cmd->flags);
1804 	READ_UNLOCK(cmd);
1805 
1806 	return 0;
1807 }
1808 
1809 int dm_cache_metadata_abort(struct dm_cache_metadata *cmd)
1810 {
1811 	int r;
1812 
1813 	WRITE_LOCK(cmd);
1814 	__destroy_persistent_data_objects(cmd);
1815 	r = __create_persistent_data_objects(cmd, false);
1816 	if (r)
1817 		cmd->fail_io = true;
1818 	WRITE_UNLOCK(cmd);
1819 
1820 	return r;
1821 }
1822