xref: /openbmc/linux/drivers/md/dm-thin-metadata.c (revision 0883c2c0)
1 /*
2  * Copyright (C) 2011-2012 Red Hat, Inc.
3  *
4  * This file is released under the GPL.
5  */
6 
7 #include "dm-thin-metadata.h"
8 #include "persistent-data/dm-btree.h"
9 #include "persistent-data/dm-space-map.h"
10 #include "persistent-data/dm-space-map-disk.h"
11 #include "persistent-data/dm-transaction-manager.h"
12 
13 #include <linux/list.h>
14 #include <linux/device-mapper.h>
15 #include <linux/workqueue.h>
16 
17 /*--------------------------------------------------------------------------
18  * As far as the metadata goes, there is:
19  *
20  * - A superblock in block zero, taking up fewer than 512 bytes for
21  *   atomic writes.
22  *
23  * - A space map managing the metadata blocks.
24  *
25  * - A space map managing the data blocks.
26  *
27  * - A btree mapping our internal thin dev ids onto struct disk_device_details.
28  *
29  * - A hierarchical btree, with 2 levels which effectively maps (thin
30  *   dev id, virtual block) -> block_time.  Block time is a 64-bit
31  *   field holding the time in the low 24 bits, and block in the top 48
32  *   bits.
33  *
34  * BTrees consist solely of btree_nodes, that fill a block.  Some are
35  * internal nodes, as such their values are a __le64 pointing to other
36  * nodes.  Leaf nodes can store data of any reasonable size (ie. much
37  * smaller than the block size).  The nodes consist of the header,
38  * followed by an array of keys, followed by an array of values.  We have
39  * to binary search on the keys so they're all held together to help the
40  * cpu cache.
41  *
42  * Space maps have 2 btrees:
43  *
44  * - One maps a uint64_t onto a struct index_entry.  Which points to a
45  *   bitmap block, and has some details about how many free entries there
46  *   are etc.
47  *
48  * - The bitmap blocks have a header (for the checksum).  Then the rest
49  *   of the block is pairs of bits.  With the meaning being:
50  *
51  *   0 - ref count is 0
52  *   1 - ref count is 1
53  *   2 - ref count is 2
54  *   3 - ref count is higher than 2
55  *
56  * - If the count is higher than 2 then the ref count is entered in a
57  *   second btree that directly maps the block_address to a uint32_t ref
58  *   count.
59  *
60  * The space map metadata variant doesn't have a bitmaps btree.  Instead
61  * it has one single blocks worth of index_entries.  This avoids
62  * recursive issues with the bitmap btree needing to allocate space in
63  * order to insert.  With a small data block size such as 64k the
64  * metadata support data devices that are hundreds of terrabytes.
65  *
66  * The space maps allocate space linearly from front to back.  Space that
67  * is freed in a transaction is never recycled within that transaction.
68  * To try and avoid fragmenting _free_ space the allocator always goes
69  * back and fills in gaps.
70  *
71  * All metadata io is in THIN_METADATA_BLOCK_SIZE sized/aligned chunks
72  * from the block manager.
73  *--------------------------------------------------------------------------*/
74 
75 #define DM_MSG_PREFIX   "thin metadata"
76 
77 #define THIN_SUPERBLOCK_MAGIC 27022010
78 #define THIN_SUPERBLOCK_LOCATION 0
79 #define THIN_VERSION 2
80 #define THIN_METADATA_CACHE_SIZE 64
81 #define SECTOR_TO_BLOCK_SHIFT 3
82 
83 /*
84  *  3 for btree insert +
85  *  2 for btree lookup used within space map
86  */
87 #define THIN_MAX_CONCURRENT_LOCKS 5
88 
89 /* This should be plenty */
90 #define SPACE_MAP_ROOT_SIZE 128
91 
92 /*
93  * Little endian on-disk superblock and device details.
94  */
95 struct thin_disk_superblock {
96 	__le32 csum;	/* Checksum of superblock except for this field. */
97 	__le32 flags;
98 	__le64 blocknr;	/* This block number, dm_block_t. */
99 
100 	__u8 uuid[16];
101 	__le64 magic;
102 	__le32 version;
103 	__le32 time;
104 
105 	__le64 trans_id;
106 
107 	/*
108 	 * Root held by userspace transactions.
109 	 */
110 	__le64 held_root;
111 
112 	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
113 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
114 
115 	/*
116 	 * 2-level btree mapping (dev_id, (dev block, time)) -> data block
117 	 */
118 	__le64 data_mapping_root;
119 
120 	/*
121 	 * Device detail root mapping dev_id -> device_details
122 	 */
123 	__le64 device_details_root;
124 
125 	__le32 data_block_size;		/* In 512-byte sectors. */
126 
127 	__le32 metadata_block_size;	/* In 512-byte sectors. */
128 	__le64 metadata_nr_blocks;
129 
130 	__le32 compat_flags;
131 	__le32 compat_ro_flags;
132 	__le32 incompat_flags;
133 } __packed;
134 
135 struct disk_device_details {
136 	__le64 mapped_blocks;
137 	__le64 transaction_id;		/* When created. */
138 	__le32 creation_time;
139 	__le32 snapshotted_time;
140 } __packed;
141 
142 struct dm_pool_metadata {
143 	struct hlist_node hash;
144 
145 	struct block_device *bdev;
146 	struct dm_block_manager *bm;
147 	struct dm_space_map *metadata_sm;
148 	struct dm_space_map *data_sm;
149 	struct dm_transaction_manager *tm;
150 	struct dm_transaction_manager *nb_tm;
151 
152 	/*
153 	 * Two-level btree.
154 	 * First level holds thin_dev_t.
155 	 * Second level holds mappings.
156 	 */
157 	struct dm_btree_info info;
158 
159 	/*
160 	 * Non-blocking version of the above.
161 	 */
162 	struct dm_btree_info nb_info;
163 
164 	/*
165 	 * Just the top level for deleting whole devices.
166 	 */
167 	struct dm_btree_info tl_info;
168 
169 	/*
170 	 * Just the bottom level for creating new devices.
171 	 */
172 	struct dm_btree_info bl_info;
173 
174 	/*
175 	 * Describes the device details btree.
176 	 */
177 	struct dm_btree_info details_info;
178 
179 	struct rw_semaphore root_lock;
180 	uint32_t time;
181 	dm_block_t root;
182 	dm_block_t details_root;
183 	struct list_head thin_devices;
184 	uint64_t trans_id;
185 	unsigned long flags;
186 	sector_t data_block_size;
187 
188 	/*
189 	 * Set if a transaction has to be aborted but the attempt to roll back
190 	 * to the previous (good) transaction failed.  The only pool metadata
191 	 * operation possible in this state is the closing of the device.
192 	 */
193 	bool fail_io:1;
194 
195 	/*
196 	 * Reading the space map roots can fail, so we read it into these
197 	 * buffers before the superblock is locked and updated.
198 	 */
199 	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
200 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
201 };
202 
203 struct dm_thin_device {
204 	struct list_head list;
205 	struct dm_pool_metadata *pmd;
206 	dm_thin_id id;
207 
208 	int open_count;
209 	bool changed:1;
210 	bool aborted_with_changes:1;
211 	uint64_t mapped_blocks;
212 	uint64_t transaction_id;
213 	uint32_t creation_time;
214 	uint32_t snapshotted_time;
215 };
216 
217 /*----------------------------------------------------------------
218  * superblock validator
219  *--------------------------------------------------------------*/
220 
221 #define SUPERBLOCK_CSUM_XOR 160774
222 
223 static void sb_prepare_for_write(struct dm_block_validator *v,
224 				 struct dm_block *b,
225 				 size_t block_size)
226 {
227 	struct thin_disk_superblock *disk_super = dm_block_data(b);
228 
229 	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
230 	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
231 						      block_size - sizeof(__le32),
232 						      SUPERBLOCK_CSUM_XOR));
233 }
234 
235 static int sb_check(struct dm_block_validator *v,
236 		    struct dm_block *b,
237 		    size_t block_size)
238 {
239 	struct thin_disk_superblock *disk_super = dm_block_data(b);
240 	__le32 csum_le;
241 
242 	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
243 		DMERR("sb_check failed: blocknr %llu: "
244 		      "wanted %llu", le64_to_cpu(disk_super->blocknr),
245 		      (unsigned long long)dm_block_location(b));
246 		return -ENOTBLK;
247 	}
248 
249 	if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
250 		DMERR("sb_check failed: magic %llu: "
251 		      "wanted %llu", le64_to_cpu(disk_super->magic),
252 		      (unsigned long long)THIN_SUPERBLOCK_MAGIC);
253 		return -EILSEQ;
254 	}
255 
256 	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
257 					     block_size - sizeof(__le32),
258 					     SUPERBLOCK_CSUM_XOR));
259 	if (csum_le != disk_super->csum) {
260 		DMERR("sb_check failed: csum %u: wanted %u",
261 		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
262 		return -EILSEQ;
263 	}
264 
265 	return 0;
266 }
267 
268 static struct dm_block_validator sb_validator = {
269 	.name = "superblock",
270 	.prepare_for_write = sb_prepare_for_write,
271 	.check = sb_check
272 };
273 
274 /*----------------------------------------------------------------
275  * Methods for the btree value types
276  *--------------------------------------------------------------*/
277 
278 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
279 {
280 	return (b << 24) | t;
281 }
282 
283 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
284 {
285 	*b = v >> 24;
286 	*t = v & ((1 << 24) - 1);
287 }
288 
289 static void data_block_inc(void *context, const void *value_le)
290 {
291 	struct dm_space_map *sm = context;
292 	__le64 v_le;
293 	uint64_t b;
294 	uint32_t t;
295 
296 	memcpy(&v_le, value_le, sizeof(v_le));
297 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
298 	dm_sm_inc_block(sm, b);
299 }
300 
301 static void data_block_dec(void *context, const void *value_le)
302 {
303 	struct dm_space_map *sm = context;
304 	__le64 v_le;
305 	uint64_t b;
306 	uint32_t t;
307 
308 	memcpy(&v_le, value_le, sizeof(v_le));
309 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
310 	dm_sm_dec_block(sm, b);
311 }
312 
313 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
314 {
315 	__le64 v1_le, v2_le;
316 	uint64_t b1, b2;
317 	uint32_t t;
318 
319 	memcpy(&v1_le, value1_le, sizeof(v1_le));
320 	memcpy(&v2_le, value2_le, sizeof(v2_le));
321 	unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
322 	unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
323 
324 	return b1 == b2;
325 }
326 
327 static void subtree_inc(void *context, const void *value)
328 {
329 	struct dm_btree_info *info = context;
330 	__le64 root_le;
331 	uint64_t root;
332 
333 	memcpy(&root_le, value, sizeof(root_le));
334 	root = le64_to_cpu(root_le);
335 	dm_tm_inc(info->tm, root);
336 }
337 
338 static void subtree_dec(void *context, const void *value)
339 {
340 	struct dm_btree_info *info = context;
341 	__le64 root_le;
342 	uint64_t root;
343 
344 	memcpy(&root_le, value, sizeof(root_le));
345 	root = le64_to_cpu(root_le);
346 	if (dm_btree_del(info, root))
347 		DMERR("btree delete failed");
348 }
349 
350 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
351 {
352 	__le64 v1_le, v2_le;
353 	memcpy(&v1_le, value1_le, sizeof(v1_le));
354 	memcpy(&v2_le, value2_le, sizeof(v2_le));
355 
356 	return v1_le == v2_le;
357 }
358 
359 /*----------------------------------------------------------------*/
360 
361 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
362 				struct dm_block **sblock)
363 {
364 	return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
365 				     &sb_validator, sblock);
366 }
367 
368 static int superblock_lock(struct dm_pool_metadata *pmd,
369 			   struct dm_block **sblock)
370 {
371 	return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
372 				&sb_validator, sblock);
373 }
374 
375 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
376 {
377 	int r;
378 	unsigned i;
379 	struct dm_block *b;
380 	__le64 *data_le, zero = cpu_to_le64(0);
381 	unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
382 
383 	/*
384 	 * We can't use a validator here - it may be all zeroes.
385 	 */
386 	r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
387 	if (r)
388 		return r;
389 
390 	data_le = dm_block_data(b);
391 	*result = 1;
392 	for (i = 0; i < block_size; i++) {
393 		if (data_le[i] != zero) {
394 			*result = 0;
395 			break;
396 		}
397 	}
398 
399 	dm_bm_unlock(b);
400 
401 	return 0;
402 }
403 
404 static void __setup_btree_details(struct dm_pool_metadata *pmd)
405 {
406 	pmd->info.tm = pmd->tm;
407 	pmd->info.levels = 2;
408 	pmd->info.value_type.context = pmd->data_sm;
409 	pmd->info.value_type.size = sizeof(__le64);
410 	pmd->info.value_type.inc = data_block_inc;
411 	pmd->info.value_type.dec = data_block_dec;
412 	pmd->info.value_type.equal = data_block_equal;
413 
414 	memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
415 	pmd->nb_info.tm = pmd->nb_tm;
416 
417 	pmd->tl_info.tm = pmd->tm;
418 	pmd->tl_info.levels = 1;
419 	pmd->tl_info.value_type.context = &pmd->bl_info;
420 	pmd->tl_info.value_type.size = sizeof(__le64);
421 	pmd->tl_info.value_type.inc = subtree_inc;
422 	pmd->tl_info.value_type.dec = subtree_dec;
423 	pmd->tl_info.value_type.equal = subtree_equal;
424 
425 	pmd->bl_info.tm = pmd->tm;
426 	pmd->bl_info.levels = 1;
427 	pmd->bl_info.value_type.context = pmd->data_sm;
428 	pmd->bl_info.value_type.size = sizeof(__le64);
429 	pmd->bl_info.value_type.inc = data_block_inc;
430 	pmd->bl_info.value_type.dec = data_block_dec;
431 	pmd->bl_info.value_type.equal = data_block_equal;
432 
433 	pmd->details_info.tm = pmd->tm;
434 	pmd->details_info.levels = 1;
435 	pmd->details_info.value_type.context = NULL;
436 	pmd->details_info.value_type.size = sizeof(struct disk_device_details);
437 	pmd->details_info.value_type.inc = NULL;
438 	pmd->details_info.value_type.dec = NULL;
439 	pmd->details_info.value_type.equal = NULL;
440 }
441 
442 static int save_sm_roots(struct dm_pool_metadata *pmd)
443 {
444 	int r;
445 	size_t len;
446 
447 	r = dm_sm_root_size(pmd->metadata_sm, &len);
448 	if (r < 0)
449 		return r;
450 
451 	r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
452 	if (r < 0)
453 		return r;
454 
455 	r = dm_sm_root_size(pmd->data_sm, &len);
456 	if (r < 0)
457 		return r;
458 
459 	return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
460 }
461 
462 static void copy_sm_roots(struct dm_pool_metadata *pmd,
463 			  struct thin_disk_superblock *disk)
464 {
465 	memcpy(&disk->metadata_space_map_root,
466 	       &pmd->metadata_space_map_root,
467 	       sizeof(pmd->metadata_space_map_root));
468 
469 	memcpy(&disk->data_space_map_root,
470 	       &pmd->data_space_map_root,
471 	       sizeof(pmd->data_space_map_root));
472 }
473 
474 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
475 {
476 	int r;
477 	struct dm_block *sblock;
478 	struct thin_disk_superblock *disk_super;
479 	sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
480 
481 	if (bdev_size > THIN_METADATA_MAX_SECTORS)
482 		bdev_size = THIN_METADATA_MAX_SECTORS;
483 
484 	r = dm_sm_commit(pmd->data_sm);
485 	if (r < 0)
486 		return r;
487 
488 	r = save_sm_roots(pmd);
489 	if (r < 0)
490 		return r;
491 
492 	r = dm_tm_pre_commit(pmd->tm);
493 	if (r < 0)
494 		return r;
495 
496 	r = superblock_lock_zero(pmd, &sblock);
497 	if (r)
498 		return r;
499 
500 	disk_super = dm_block_data(sblock);
501 	disk_super->flags = 0;
502 	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
503 	disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
504 	disk_super->version = cpu_to_le32(THIN_VERSION);
505 	disk_super->time = 0;
506 	disk_super->trans_id = 0;
507 	disk_super->held_root = 0;
508 
509 	copy_sm_roots(pmd, disk_super);
510 
511 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
512 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
513 	disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
514 	disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
515 	disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
516 
517 	return dm_tm_commit(pmd->tm, sblock);
518 }
519 
520 static int __format_metadata(struct dm_pool_metadata *pmd)
521 {
522 	int r;
523 
524 	r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
525 				 &pmd->tm, &pmd->metadata_sm);
526 	if (r < 0) {
527 		DMERR("tm_create_with_sm failed");
528 		return r;
529 	}
530 
531 	pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
532 	if (IS_ERR(pmd->data_sm)) {
533 		DMERR("sm_disk_create failed");
534 		r = PTR_ERR(pmd->data_sm);
535 		goto bad_cleanup_tm;
536 	}
537 
538 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
539 	if (!pmd->nb_tm) {
540 		DMERR("could not create non-blocking clone tm");
541 		r = -ENOMEM;
542 		goto bad_cleanup_data_sm;
543 	}
544 
545 	__setup_btree_details(pmd);
546 
547 	r = dm_btree_empty(&pmd->info, &pmd->root);
548 	if (r < 0)
549 		goto bad_cleanup_nb_tm;
550 
551 	r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
552 	if (r < 0) {
553 		DMERR("couldn't create devices root");
554 		goto bad_cleanup_nb_tm;
555 	}
556 
557 	r = __write_initial_superblock(pmd);
558 	if (r)
559 		goto bad_cleanup_nb_tm;
560 
561 	return 0;
562 
563 bad_cleanup_nb_tm:
564 	dm_tm_destroy(pmd->nb_tm);
565 bad_cleanup_data_sm:
566 	dm_sm_destroy(pmd->data_sm);
567 bad_cleanup_tm:
568 	dm_tm_destroy(pmd->tm);
569 	dm_sm_destroy(pmd->metadata_sm);
570 
571 	return r;
572 }
573 
574 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
575 				     struct dm_pool_metadata *pmd)
576 {
577 	uint32_t features;
578 
579 	features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
580 	if (features) {
581 		DMERR("could not access metadata due to unsupported optional features (%lx).",
582 		      (unsigned long)features);
583 		return -EINVAL;
584 	}
585 
586 	/*
587 	 * Check for read-only metadata to skip the following RDWR checks.
588 	 */
589 	if (get_disk_ro(pmd->bdev->bd_disk))
590 		return 0;
591 
592 	features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
593 	if (features) {
594 		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
595 		      (unsigned long)features);
596 		return -EINVAL;
597 	}
598 
599 	return 0;
600 }
601 
602 static int __open_metadata(struct dm_pool_metadata *pmd)
603 {
604 	int r;
605 	struct dm_block *sblock;
606 	struct thin_disk_superblock *disk_super;
607 
608 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
609 			    &sb_validator, &sblock);
610 	if (r < 0) {
611 		DMERR("couldn't read superblock");
612 		return r;
613 	}
614 
615 	disk_super = dm_block_data(sblock);
616 
617 	/* Verify the data block size hasn't changed */
618 	if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
619 		DMERR("changing the data block size (from %u to %llu) is not supported",
620 		      le32_to_cpu(disk_super->data_block_size),
621 		      (unsigned long long)pmd->data_block_size);
622 		r = -EINVAL;
623 		goto bad_unlock_sblock;
624 	}
625 
626 	r = __check_incompat_features(disk_super, pmd);
627 	if (r < 0)
628 		goto bad_unlock_sblock;
629 
630 	r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
631 			       disk_super->metadata_space_map_root,
632 			       sizeof(disk_super->metadata_space_map_root),
633 			       &pmd->tm, &pmd->metadata_sm);
634 	if (r < 0) {
635 		DMERR("tm_open_with_sm failed");
636 		goto bad_unlock_sblock;
637 	}
638 
639 	pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
640 				       sizeof(disk_super->data_space_map_root));
641 	if (IS_ERR(pmd->data_sm)) {
642 		DMERR("sm_disk_open failed");
643 		r = PTR_ERR(pmd->data_sm);
644 		goto bad_cleanup_tm;
645 	}
646 
647 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
648 	if (!pmd->nb_tm) {
649 		DMERR("could not create non-blocking clone tm");
650 		r = -ENOMEM;
651 		goto bad_cleanup_data_sm;
652 	}
653 
654 	__setup_btree_details(pmd);
655 	dm_bm_unlock(sblock);
656 
657 	return 0;
658 
659 bad_cleanup_data_sm:
660 	dm_sm_destroy(pmd->data_sm);
661 bad_cleanup_tm:
662 	dm_tm_destroy(pmd->tm);
663 	dm_sm_destroy(pmd->metadata_sm);
664 bad_unlock_sblock:
665 	dm_bm_unlock(sblock);
666 
667 	return r;
668 }
669 
670 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
671 {
672 	int r, unformatted;
673 
674 	r = __superblock_all_zeroes(pmd->bm, &unformatted);
675 	if (r)
676 		return r;
677 
678 	if (unformatted)
679 		return format_device ? __format_metadata(pmd) : -EPERM;
680 
681 	return __open_metadata(pmd);
682 }
683 
684 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
685 {
686 	int r;
687 
688 	pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
689 					  THIN_METADATA_CACHE_SIZE,
690 					  THIN_MAX_CONCURRENT_LOCKS);
691 	if (IS_ERR(pmd->bm)) {
692 		DMERR("could not create block manager");
693 		return PTR_ERR(pmd->bm);
694 	}
695 
696 	r = __open_or_format_metadata(pmd, format_device);
697 	if (r)
698 		dm_block_manager_destroy(pmd->bm);
699 
700 	return r;
701 }
702 
703 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
704 {
705 	dm_sm_destroy(pmd->data_sm);
706 	dm_sm_destroy(pmd->metadata_sm);
707 	dm_tm_destroy(pmd->nb_tm);
708 	dm_tm_destroy(pmd->tm);
709 	dm_block_manager_destroy(pmd->bm);
710 }
711 
712 static int __begin_transaction(struct dm_pool_metadata *pmd)
713 {
714 	int r;
715 	struct thin_disk_superblock *disk_super;
716 	struct dm_block *sblock;
717 
718 	/*
719 	 * We re-read the superblock every time.  Shouldn't need to do this
720 	 * really.
721 	 */
722 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
723 			    &sb_validator, &sblock);
724 	if (r)
725 		return r;
726 
727 	disk_super = dm_block_data(sblock);
728 	pmd->time = le32_to_cpu(disk_super->time);
729 	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
730 	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
731 	pmd->trans_id = le64_to_cpu(disk_super->trans_id);
732 	pmd->flags = le32_to_cpu(disk_super->flags);
733 	pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
734 
735 	dm_bm_unlock(sblock);
736 	return 0;
737 }
738 
739 static int __write_changed_details(struct dm_pool_metadata *pmd)
740 {
741 	int r;
742 	struct dm_thin_device *td, *tmp;
743 	struct disk_device_details details;
744 	uint64_t key;
745 
746 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
747 		if (!td->changed)
748 			continue;
749 
750 		key = td->id;
751 
752 		details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
753 		details.transaction_id = cpu_to_le64(td->transaction_id);
754 		details.creation_time = cpu_to_le32(td->creation_time);
755 		details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
756 		__dm_bless_for_disk(&details);
757 
758 		r = dm_btree_insert(&pmd->details_info, pmd->details_root,
759 				    &key, &details, &pmd->details_root);
760 		if (r)
761 			return r;
762 
763 		if (td->open_count)
764 			td->changed = 0;
765 		else {
766 			list_del(&td->list);
767 			kfree(td);
768 		}
769 	}
770 
771 	return 0;
772 }
773 
774 static int __commit_transaction(struct dm_pool_metadata *pmd)
775 {
776 	int r;
777 	size_t metadata_len, data_len;
778 	struct thin_disk_superblock *disk_super;
779 	struct dm_block *sblock;
780 
781 	/*
782 	 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
783 	 */
784 	BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
785 
786 	r = __write_changed_details(pmd);
787 	if (r < 0)
788 		return r;
789 
790 	r = dm_sm_commit(pmd->data_sm);
791 	if (r < 0)
792 		return r;
793 
794 	r = dm_tm_pre_commit(pmd->tm);
795 	if (r < 0)
796 		return r;
797 
798 	r = dm_sm_root_size(pmd->metadata_sm, &metadata_len);
799 	if (r < 0)
800 		return r;
801 
802 	r = dm_sm_root_size(pmd->data_sm, &data_len);
803 	if (r < 0)
804 		return r;
805 
806 	r = save_sm_roots(pmd);
807 	if (r < 0)
808 		return r;
809 
810 	r = superblock_lock(pmd, &sblock);
811 	if (r)
812 		return r;
813 
814 	disk_super = dm_block_data(sblock);
815 	disk_super->time = cpu_to_le32(pmd->time);
816 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
817 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
818 	disk_super->trans_id = cpu_to_le64(pmd->trans_id);
819 	disk_super->flags = cpu_to_le32(pmd->flags);
820 
821 	copy_sm_roots(pmd, disk_super);
822 
823 	return dm_tm_commit(pmd->tm, sblock);
824 }
825 
826 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
827 					       sector_t data_block_size,
828 					       bool format_device)
829 {
830 	int r;
831 	struct dm_pool_metadata *pmd;
832 
833 	pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
834 	if (!pmd) {
835 		DMERR("could not allocate metadata struct");
836 		return ERR_PTR(-ENOMEM);
837 	}
838 
839 	init_rwsem(&pmd->root_lock);
840 	pmd->time = 0;
841 	INIT_LIST_HEAD(&pmd->thin_devices);
842 	pmd->fail_io = false;
843 	pmd->bdev = bdev;
844 	pmd->data_block_size = data_block_size;
845 
846 	r = __create_persistent_data_objects(pmd, format_device);
847 	if (r) {
848 		kfree(pmd);
849 		return ERR_PTR(r);
850 	}
851 
852 	r = __begin_transaction(pmd);
853 	if (r < 0) {
854 		if (dm_pool_metadata_close(pmd) < 0)
855 			DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
856 		return ERR_PTR(r);
857 	}
858 
859 	return pmd;
860 }
861 
862 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
863 {
864 	int r;
865 	unsigned open_devices = 0;
866 	struct dm_thin_device *td, *tmp;
867 
868 	down_read(&pmd->root_lock);
869 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
870 		if (td->open_count)
871 			open_devices++;
872 		else {
873 			list_del(&td->list);
874 			kfree(td);
875 		}
876 	}
877 	up_read(&pmd->root_lock);
878 
879 	if (open_devices) {
880 		DMERR("attempt to close pmd when %u device(s) are still open",
881 		       open_devices);
882 		return -EBUSY;
883 	}
884 
885 	if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
886 		r = __commit_transaction(pmd);
887 		if (r < 0)
888 			DMWARN("%s: __commit_transaction() failed, error = %d",
889 			       __func__, r);
890 	}
891 
892 	if (!pmd->fail_io)
893 		__destroy_persistent_data_objects(pmd);
894 
895 	kfree(pmd);
896 	return 0;
897 }
898 
899 /*
900  * __open_device: Returns @td corresponding to device with id @dev,
901  * creating it if @create is set and incrementing @td->open_count.
902  * On failure, @td is undefined.
903  */
904 static int __open_device(struct dm_pool_metadata *pmd,
905 			 dm_thin_id dev, int create,
906 			 struct dm_thin_device **td)
907 {
908 	int r, changed = 0;
909 	struct dm_thin_device *td2;
910 	uint64_t key = dev;
911 	struct disk_device_details details_le;
912 
913 	/*
914 	 * If the device is already open, return it.
915 	 */
916 	list_for_each_entry(td2, &pmd->thin_devices, list)
917 		if (td2->id == dev) {
918 			/*
919 			 * May not create an already-open device.
920 			 */
921 			if (create)
922 				return -EEXIST;
923 
924 			td2->open_count++;
925 			*td = td2;
926 			return 0;
927 		}
928 
929 	/*
930 	 * Check the device exists.
931 	 */
932 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
933 			    &key, &details_le);
934 	if (r) {
935 		if (r != -ENODATA || !create)
936 			return r;
937 
938 		/*
939 		 * Create new device.
940 		 */
941 		changed = 1;
942 		details_le.mapped_blocks = 0;
943 		details_le.transaction_id = cpu_to_le64(pmd->trans_id);
944 		details_le.creation_time = cpu_to_le32(pmd->time);
945 		details_le.snapshotted_time = cpu_to_le32(pmd->time);
946 	}
947 
948 	*td = kmalloc(sizeof(**td), GFP_NOIO);
949 	if (!*td)
950 		return -ENOMEM;
951 
952 	(*td)->pmd = pmd;
953 	(*td)->id = dev;
954 	(*td)->open_count = 1;
955 	(*td)->changed = changed;
956 	(*td)->aborted_with_changes = false;
957 	(*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
958 	(*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
959 	(*td)->creation_time = le32_to_cpu(details_le.creation_time);
960 	(*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
961 
962 	list_add(&(*td)->list, &pmd->thin_devices);
963 
964 	return 0;
965 }
966 
967 static void __close_device(struct dm_thin_device *td)
968 {
969 	--td->open_count;
970 }
971 
972 static int __create_thin(struct dm_pool_metadata *pmd,
973 			 dm_thin_id dev)
974 {
975 	int r;
976 	dm_block_t dev_root;
977 	uint64_t key = dev;
978 	struct disk_device_details details_le;
979 	struct dm_thin_device *td;
980 	__le64 value;
981 
982 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
983 			    &key, &details_le);
984 	if (!r)
985 		return -EEXIST;
986 
987 	/*
988 	 * Create an empty btree for the mappings.
989 	 */
990 	r = dm_btree_empty(&pmd->bl_info, &dev_root);
991 	if (r)
992 		return r;
993 
994 	/*
995 	 * Insert it into the main mapping tree.
996 	 */
997 	value = cpu_to_le64(dev_root);
998 	__dm_bless_for_disk(&value);
999 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1000 	if (r) {
1001 		dm_btree_del(&pmd->bl_info, dev_root);
1002 		return r;
1003 	}
1004 
1005 	r = __open_device(pmd, dev, 1, &td);
1006 	if (r) {
1007 		dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1008 		dm_btree_del(&pmd->bl_info, dev_root);
1009 		return r;
1010 	}
1011 	__close_device(td);
1012 
1013 	return r;
1014 }
1015 
1016 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1017 {
1018 	int r = -EINVAL;
1019 
1020 	down_write(&pmd->root_lock);
1021 	if (!pmd->fail_io)
1022 		r = __create_thin(pmd, dev);
1023 	up_write(&pmd->root_lock);
1024 
1025 	return r;
1026 }
1027 
1028 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1029 				  struct dm_thin_device *snap,
1030 				  dm_thin_id origin, uint32_t time)
1031 {
1032 	int r;
1033 	struct dm_thin_device *td;
1034 
1035 	r = __open_device(pmd, origin, 0, &td);
1036 	if (r)
1037 		return r;
1038 
1039 	td->changed = 1;
1040 	td->snapshotted_time = time;
1041 
1042 	snap->mapped_blocks = td->mapped_blocks;
1043 	snap->snapshotted_time = time;
1044 	__close_device(td);
1045 
1046 	return 0;
1047 }
1048 
1049 static int __create_snap(struct dm_pool_metadata *pmd,
1050 			 dm_thin_id dev, dm_thin_id origin)
1051 {
1052 	int r;
1053 	dm_block_t origin_root;
1054 	uint64_t key = origin, dev_key = dev;
1055 	struct dm_thin_device *td;
1056 	struct disk_device_details details_le;
1057 	__le64 value;
1058 
1059 	/* check this device is unused */
1060 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1061 			    &dev_key, &details_le);
1062 	if (!r)
1063 		return -EEXIST;
1064 
1065 	/* find the mapping tree for the origin */
1066 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1067 	if (r)
1068 		return r;
1069 	origin_root = le64_to_cpu(value);
1070 
1071 	/* clone the origin, an inc will do */
1072 	dm_tm_inc(pmd->tm, origin_root);
1073 
1074 	/* insert into the main mapping tree */
1075 	value = cpu_to_le64(origin_root);
1076 	__dm_bless_for_disk(&value);
1077 	key = dev;
1078 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1079 	if (r) {
1080 		dm_tm_dec(pmd->tm, origin_root);
1081 		return r;
1082 	}
1083 
1084 	pmd->time++;
1085 
1086 	r = __open_device(pmd, dev, 1, &td);
1087 	if (r)
1088 		goto bad;
1089 
1090 	r = __set_snapshot_details(pmd, td, origin, pmd->time);
1091 	__close_device(td);
1092 
1093 	if (r)
1094 		goto bad;
1095 
1096 	return 0;
1097 
1098 bad:
1099 	dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1100 	dm_btree_remove(&pmd->details_info, pmd->details_root,
1101 			&key, &pmd->details_root);
1102 	return r;
1103 }
1104 
1105 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1106 				 dm_thin_id dev,
1107 				 dm_thin_id origin)
1108 {
1109 	int r = -EINVAL;
1110 
1111 	down_write(&pmd->root_lock);
1112 	if (!pmd->fail_io)
1113 		r = __create_snap(pmd, dev, origin);
1114 	up_write(&pmd->root_lock);
1115 
1116 	return r;
1117 }
1118 
1119 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1120 {
1121 	int r;
1122 	uint64_t key = dev;
1123 	struct dm_thin_device *td;
1124 
1125 	/* TODO: failure should mark the transaction invalid */
1126 	r = __open_device(pmd, dev, 0, &td);
1127 	if (r)
1128 		return r;
1129 
1130 	if (td->open_count > 1) {
1131 		__close_device(td);
1132 		return -EBUSY;
1133 	}
1134 
1135 	list_del(&td->list);
1136 	kfree(td);
1137 	r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1138 			    &key, &pmd->details_root);
1139 	if (r)
1140 		return r;
1141 
1142 	r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1143 	if (r)
1144 		return r;
1145 
1146 	return 0;
1147 }
1148 
1149 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1150 			       dm_thin_id dev)
1151 {
1152 	int r = -EINVAL;
1153 
1154 	down_write(&pmd->root_lock);
1155 	if (!pmd->fail_io)
1156 		r = __delete_device(pmd, dev);
1157 	up_write(&pmd->root_lock);
1158 
1159 	return r;
1160 }
1161 
1162 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1163 					uint64_t current_id,
1164 					uint64_t new_id)
1165 {
1166 	int r = -EINVAL;
1167 
1168 	down_write(&pmd->root_lock);
1169 
1170 	if (pmd->fail_io)
1171 		goto out;
1172 
1173 	if (pmd->trans_id != current_id) {
1174 		DMERR("mismatched transaction id");
1175 		goto out;
1176 	}
1177 
1178 	pmd->trans_id = new_id;
1179 	r = 0;
1180 
1181 out:
1182 	up_write(&pmd->root_lock);
1183 
1184 	return r;
1185 }
1186 
1187 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1188 					uint64_t *result)
1189 {
1190 	int r = -EINVAL;
1191 
1192 	down_read(&pmd->root_lock);
1193 	if (!pmd->fail_io) {
1194 		*result = pmd->trans_id;
1195 		r = 0;
1196 	}
1197 	up_read(&pmd->root_lock);
1198 
1199 	return r;
1200 }
1201 
1202 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1203 {
1204 	int r, inc;
1205 	struct thin_disk_superblock *disk_super;
1206 	struct dm_block *copy, *sblock;
1207 	dm_block_t held_root;
1208 
1209 	/*
1210 	 * We commit to ensure the btree roots which we increment in a
1211 	 * moment are up to date.
1212 	 */
1213 	__commit_transaction(pmd);
1214 
1215 	/*
1216 	 * Copy the superblock.
1217 	 */
1218 	dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1219 	r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1220 			       &sb_validator, &copy, &inc);
1221 	if (r)
1222 		return r;
1223 
1224 	BUG_ON(!inc);
1225 
1226 	held_root = dm_block_location(copy);
1227 	disk_super = dm_block_data(copy);
1228 
1229 	if (le64_to_cpu(disk_super->held_root)) {
1230 		DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1231 
1232 		dm_tm_dec(pmd->tm, held_root);
1233 		dm_tm_unlock(pmd->tm, copy);
1234 		return -EBUSY;
1235 	}
1236 
1237 	/*
1238 	 * Wipe the spacemap since we're not publishing this.
1239 	 */
1240 	memset(&disk_super->data_space_map_root, 0,
1241 	       sizeof(disk_super->data_space_map_root));
1242 	memset(&disk_super->metadata_space_map_root, 0,
1243 	       sizeof(disk_super->metadata_space_map_root));
1244 
1245 	/*
1246 	 * Increment the data structures that need to be preserved.
1247 	 */
1248 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1249 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1250 	dm_tm_unlock(pmd->tm, copy);
1251 
1252 	/*
1253 	 * Write the held root into the superblock.
1254 	 */
1255 	r = superblock_lock(pmd, &sblock);
1256 	if (r) {
1257 		dm_tm_dec(pmd->tm, held_root);
1258 		return r;
1259 	}
1260 
1261 	disk_super = dm_block_data(sblock);
1262 	disk_super->held_root = cpu_to_le64(held_root);
1263 	dm_bm_unlock(sblock);
1264 	return 0;
1265 }
1266 
1267 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1268 {
1269 	int r = -EINVAL;
1270 
1271 	down_write(&pmd->root_lock);
1272 	if (!pmd->fail_io)
1273 		r = __reserve_metadata_snap(pmd);
1274 	up_write(&pmd->root_lock);
1275 
1276 	return r;
1277 }
1278 
1279 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1280 {
1281 	int r;
1282 	struct thin_disk_superblock *disk_super;
1283 	struct dm_block *sblock, *copy;
1284 	dm_block_t held_root;
1285 
1286 	r = superblock_lock(pmd, &sblock);
1287 	if (r)
1288 		return r;
1289 
1290 	disk_super = dm_block_data(sblock);
1291 	held_root = le64_to_cpu(disk_super->held_root);
1292 	disk_super->held_root = cpu_to_le64(0);
1293 
1294 	dm_bm_unlock(sblock);
1295 
1296 	if (!held_root) {
1297 		DMWARN("No pool metadata snapshot found: nothing to release.");
1298 		return -EINVAL;
1299 	}
1300 
1301 	r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1302 	if (r)
1303 		return r;
1304 
1305 	disk_super = dm_block_data(copy);
1306 	dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1307 	dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1308 	dm_sm_dec_block(pmd->metadata_sm, held_root);
1309 
1310 	dm_tm_unlock(pmd->tm, copy);
1311 
1312 	return 0;
1313 }
1314 
1315 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1316 {
1317 	int r = -EINVAL;
1318 
1319 	down_write(&pmd->root_lock);
1320 	if (!pmd->fail_io)
1321 		r = __release_metadata_snap(pmd);
1322 	up_write(&pmd->root_lock);
1323 
1324 	return r;
1325 }
1326 
1327 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1328 			       dm_block_t *result)
1329 {
1330 	int r;
1331 	struct thin_disk_superblock *disk_super;
1332 	struct dm_block *sblock;
1333 
1334 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1335 			    &sb_validator, &sblock);
1336 	if (r)
1337 		return r;
1338 
1339 	disk_super = dm_block_data(sblock);
1340 	*result = le64_to_cpu(disk_super->held_root);
1341 
1342 	dm_bm_unlock(sblock);
1343 
1344 	return 0;
1345 }
1346 
1347 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1348 			      dm_block_t *result)
1349 {
1350 	int r = -EINVAL;
1351 
1352 	down_read(&pmd->root_lock);
1353 	if (!pmd->fail_io)
1354 		r = __get_metadata_snap(pmd, result);
1355 	up_read(&pmd->root_lock);
1356 
1357 	return r;
1358 }
1359 
1360 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1361 			     struct dm_thin_device **td)
1362 {
1363 	int r = -EINVAL;
1364 
1365 	down_write(&pmd->root_lock);
1366 	if (!pmd->fail_io)
1367 		r = __open_device(pmd, dev, 0, td);
1368 	up_write(&pmd->root_lock);
1369 
1370 	return r;
1371 }
1372 
1373 int dm_pool_close_thin_device(struct dm_thin_device *td)
1374 {
1375 	down_write(&td->pmd->root_lock);
1376 	__close_device(td);
1377 	up_write(&td->pmd->root_lock);
1378 
1379 	return 0;
1380 }
1381 
1382 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1383 {
1384 	return td->id;
1385 }
1386 
1387 /*
1388  * Check whether @time (of block creation) is older than @td's last snapshot.
1389  * If so then the associated block is shared with the last snapshot device.
1390  * Any block on a device created *after* the device last got snapshotted is
1391  * necessarily not shared.
1392  */
1393 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1394 {
1395 	return td->snapshotted_time > time;
1396 }
1397 
1398 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1399 				 struct dm_thin_lookup_result *result)
1400 {
1401 	uint64_t block_time = 0;
1402 	dm_block_t exception_block;
1403 	uint32_t exception_time;
1404 
1405 	block_time = le64_to_cpu(value);
1406 	unpack_block_time(block_time, &exception_block, &exception_time);
1407 	result->block = exception_block;
1408 	result->shared = __snapshotted_since(td, exception_time);
1409 }
1410 
1411 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1412 			int can_issue_io, struct dm_thin_lookup_result *result)
1413 {
1414 	int r;
1415 	__le64 value;
1416 	struct dm_pool_metadata *pmd = td->pmd;
1417 	dm_block_t keys[2] = { td->id, block };
1418 	struct dm_btree_info *info;
1419 
1420 	if (can_issue_io) {
1421 		info = &pmd->info;
1422 	} else
1423 		info = &pmd->nb_info;
1424 
1425 	r = dm_btree_lookup(info, pmd->root, keys, &value);
1426 	if (!r)
1427 		unpack_lookup_result(td, value, result);
1428 
1429 	return r;
1430 }
1431 
1432 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1433 		       int can_issue_io, struct dm_thin_lookup_result *result)
1434 {
1435 	int r;
1436 	struct dm_pool_metadata *pmd = td->pmd;
1437 
1438 	down_read(&pmd->root_lock);
1439 	if (pmd->fail_io) {
1440 		up_read(&pmd->root_lock);
1441 		return -EINVAL;
1442 	}
1443 
1444 	r = __find_block(td, block, can_issue_io, result);
1445 
1446 	up_read(&pmd->root_lock);
1447 	return r;
1448 }
1449 
1450 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1451 					  dm_block_t *vblock,
1452 					  struct dm_thin_lookup_result *result)
1453 {
1454 	int r;
1455 	__le64 value;
1456 	struct dm_pool_metadata *pmd = td->pmd;
1457 	dm_block_t keys[2] = { td->id, block };
1458 
1459 	r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1460 	if (!r)
1461 		unpack_lookup_result(td, value, result);
1462 
1463 	return r;
1464 }
1465 
1466 static int __find_mapped_range(struct dm_thin_device *td,
1467 			       dm_block_t begin, dm_block_t end,
1468 			       dm_block_t *thin_begin, dm_block_t *thin_end,
1469 			       dm_block_t *pool_begin, bool *maybe_shared)
1470 {
1471 	int r;
1472 	dm_block_t pool_end;
1473 	struct dm_thin_lookup_result lookup;
1474 
1475 	if (end < begin)
1476 		return -ENODATA;
1477 
1478 	r = __find_next_mapped_block(td, begin, &begin, &lookup);
1479 	if (r)
1480 		return r;
1481 
1482 	if (begin >= end)
1483 		return -ENODATA;
1484 
1485 	*thin_begin = begin;
1486 	*pool_begin = lookup.block;
1487 	*maybe_shared = lookup.shared;
1488 
1489 	begin++;
1490 	pool_end = *pool_begin + 1;
1491 	while (begin != end) {
1492 		r = __find_block(td, begin, true, &lookup);
1493 		if (r) {
1494 			if (r == -ENODATA)
1495 				break;
1496 			else
1497 				return r;
1498 		}
1499 
1500 		if ((lookup.block != pool_end) ||
1501 		    (lookup.shared != *maybe_shared))
1502 			break;
1503 
1504 		pool_end++;
1505 		begin++;
1506 	}
1507 
1508 	*thin_end = begin;
1509 	return 0;
1510 }
1511 
1512 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1513 			      dm_block_t begin, dm_block_t end,
1514 			      dm_block_t *thin_begin, dm_block_t *thin_end,
1515 			      dm_block_t *pool_begin, bool *maybe_shared)
1516 {
1517 	int r = -EINVAL;
1518 	struct dm_pool_metadata *pmd = td->pmd;
1519 
1520 	down_read(&pmd->root_lock);
1521 	if (!pmd->fail_io) {
1522 		r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1523 					pool_begin, maybe_shared);
1524 	}
1525 	up_read(&pmd->root_lock);
1526 
1527 	return r;
1528 }
1529 
1530 static int __insert(struct dm_thin_device *td, dm_block_t block,
1531 		    dm_block_t data_block)
1532 {
1533 	int r, inserted;
1534 	__le64 value;
1535 	struct dm_pool_metadata *pmd = td->pmd;
1536 	dm_block_t keys[2] = { td->id, block };
1537 
1538 	value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1539 	__dm_bless_for_disk(&value);
1540 
1541 	r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1542 				   &pmd->root, &inserted);
1543 	if (r)
1544 		return r;
1545 
1546 	td->changed = 1;
1547 	if (inserted)
1548 		td->mapped_blocks++;
1549 
1550 	return 0;
1551 }
1552 
1553 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1554 			 dm_block_t data_block)
1555 {
1556 	int r = -EINVAL;
1557 
1558 	down_write(&td->pmd->root_lock);
1559 	if (!td->pmd->fail_io)
1560 		r = __insert(td, block, data_block);
1561 	up_write(&td->pmd->root_lock);
1562 
1563 	return r;
1564 }
1565 
1566 static int __remove(struct dm_thin_device *td, dm_block_t block)
1567 {
1568 	int r;
1569 	struct dm_pool_metadata *pmd = td->pmd;
1570 	dm_block_t keys[2] = { td->id, block };
1571 
1572 	r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1573 	if (r)
1574 		return r;
1575 
1576 	td->mapped_blocks--;
1577 	td->changed = 1;
1578 
1579 	return 0;
1580 }
1581 
1582 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1583 {
1584 	int r;
1585 	unsigned count, total_count = 0;
1586 	struct dm_pool_metadata *pmd = td->pmd;
1587 	dm_block_t keys[1] = { td->id };
1588 	__le64 value;
1589 	dm_block_t mapping_root;
1590 
1591 	/*
1592 	 * Find the mapping tree
1593 	 */
1594 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1595 	if (r)
1596 		return r;
1597 
1598 	/*
1599 	 * Remove from the mapping tree, taking care to inc the
1600 	 * ref count so it doesn't get deleted.
1601 	 */
1602 	mapping_root = le64_to_cpu(value);
1603 	dm_tm_inc(pmd->tm, mapping_root);
1604 	r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1605 	if (r)
1606 		return r;
1607 
1608 	/*
1609 	 * Remove leaves stops at the first unmapped entry, so we have to
1610 	 * loop round finding mapped ranges.
1611 	 */
1612 	while (begin < end) {
1613 		r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1614 		if (r == -ENODATA)
1615 			break;
1616 
1617 		if (r)
1618 			return r;
1619 
1620 		if (begin >= end)
1621 			break;
1622 
1623 		r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1624 		if (r)
1625 			return r;
1626 
1627 		total_count += count;
1628 	}
1629 
1630 	td->mapped_blocks -= total_count;
1631 	td->changed = 1;
1632 
1633 	/*
1634 	 * Reinsert the mapping tree.
1635 	 */
1636 	value = cpu_to_le64(mapping_root);
1637 	__dm_bless_for_disk(&value);
1638 	return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1639 }
1640 
1641 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1642 {
1643 	int r = -EINVAL;
1644 
1645 	down_write(&td->pmd->root_lock);
1646 	if (!td->pmd->fail_io)
1647 		r = __remove(td, block);
1648 	up_write(&td->pmd->root_lock);
1649 
1650 	return r;
1651 }
1652 
1653 int dm_thin_remove_range(struct dm_thin_device *td,
1654 			 dm_block_t begin, dm_block_t end)
1655 {
1656 	int r = -EINVAL;
1657 
1658 	down_write(&td->pmd->root_lock);
1659 	if (!td->pmd->fail_io)
1660 		r = __remove_range(td, begin, end);
1661 	up_write(&td->pmd->root_lock);
1662 
1663 	return r;
1664 }
1665 
1666 int dm_pool_block_is_used(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1667 {
1668 	int r;
1669 	uint32_t ref_count;
1670 
1671 	down_read(&pmd->root_lock);
1672 	r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1673 	if (!r)
1674 		*result = (ref_count != 0);
1675 	up_read(&pmd->root_lock);
1676 
1677 	return r;
1678 }
1679 
1680 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1681 {
1682 	int r;
1683 
1684 	down_read(&td->pmd->root_lock);
1685 	r = td->changed;
1686 	up_read(&td->pmd->root_lock);
1687 
1688 	return r;
1689 }
1690 
1691 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1692 {
1693 	bool r = false;
1694 	struct dm_thin_device *td, *tmp;
1695 
1696 	down_read(&pmd->root_lock);
1697 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1698 		if (td->changed) {
1699 			r = td->changed;
1700 			break;
1701 		}
1702 	}
1703 	up_read(&pmd->root_lock);
1704 
1705 	return r;
1706 }
1707 
1708 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1709 {
1710 	bool r;
1711 
1712 	down_read(&td->pmd->root_lock);
1713 	r = td->aborted_with_changes;
1714 	up_read(&td->pmd->root_lock);
1715 
1716 	return r;
1717 }
1718 
1719 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1720 {
1721 	int r = -EINVAL;
1722 
1723 	down_write(&pmd->root_lock);
1724 	if (!pmd->fail_io)
1725 		r = dm_sm_new_block(pmd->data_sm, result);
1726 	up_write(&pmd->root_lock);
1727 
1728 	return r;
1729 }
1730 
1731 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1732 {
1733 	int r = -EINVAL;
1734 
1735 	down_write(&pmd->root_lock);
1736 	if (pmd->fail_io)
1737 		goto out;
1738 
1739 	r = __commit_transaction(pmd);
1740 	if (r <= 0)
1741 		goto out;
1742 
1743 	/*
1744 	 * Open the next transaction.
1745 	 */
1746 	r = __begin_transaction(pmd);
1747 out:
1748 	up_write(&pmd->root_lock);
1749 	return r;
1750 }
1751 
1752 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1753 {
1754 	struct dm_thin_device *td;
1755 
1756 	list_for_each_entry(td, &pmd->thin_devices, list)
1757 		td->aborted_with_changes = td->changed;
1758 }
1759 
1760 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1761 {
1762 	int r = -EINVAL;
1763 
1764 	down_write(&pmd->root_lock);
1765 	if (pmd->fail_io)
1766 		goto out;
1767 
1768 	__set_abort_with_changes_flags(pmd);
1769 	__destroy_persistent_data_objects(pmd);
1770 	r = __create_persistent_data_objects(pmd, false);
1771 	if (r)
1772 		pmd->fail_io = true;
1773 
1774 out:
1775 	up_write(&pmd->root_lock);
1776 
1777 	return r;
1778 }
1779 
1780 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1781 {
1782 	int r = -EINVAL;
1783 
1784 	down_read(&pmd->root_lock);
1785 	if (!pmd->fail_io)
1786 		r = dm_sm_get_nr_free(pmd->data_sm, result);
1787 	up_read(&pmd->root_lock);
1788 
1789 	return r;
1790 }
1791 
1792 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1793 					  dm_block_t *result)
1794 {
1795 	int r = -EINVAL;
1796 
1797 	down_read(&pmd->root_lock);
1798 	if (!pmd->fail_io)
1799 		r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1800 	up_read(&pmd->root_lock);
1801 
1802 	return r;
1803 }
1804 
1805 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1806 				  dm_block_t *result)
1807 {
1808 	int r = -EINVAL;
1809 
1810 	down_read(&pmd->root_lock);
1811 	if (!pmd->fail_io)
1812 		r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1813 	up_read(&pmd->root_lock);
1814 
1815 	return r;
1816 }
1817 
1818 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1819 {
1820 	int r = -EINVAL;
1821 
1822 	down_read(&pmd->root_lock);
1823 	if (!pmd->fail_io)
1824 		r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1825 	up_read(&pmd->root_lock);
1826 
1827 	return r;
1828 }
1829 
1830 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1831 {
1832 	int r = -EINVAL;
1833 	struct dm_pool_metadata *pmd = td->pmd;
1834 
1835 	down_read(&pmd->root_lock);
1836 	if (!pmd->fail_io) {
1837 		*result = td->mapped_blocks;
1838 		r = 0;
1839 	}
1840 	up_read(&pmd->root_lock);
1841 
1842 	return r;
1843 }
1844 
1845 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1846 {
1847 	int r;
1848 	__le64 value_le;
1849 	dm_block_t thin_root;
1850 	struct dm_pool_metadata *pmd = td->pmd;
1851 
1852 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1853 	if (r)
1854 		return r;
1855 
1856 	thin_root = le64_to_cpu(value_le);
1857 
1858 	return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1859 }
1860 
1861 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1862 				     dm_block_t *result)
1863 {
1864 	int r = -EINVAL;
1865 	struct dm_pool_metadata *pmd = td->pmd;
1866 
1867 	down_read(&pmd->root_lock);
1868 	if (!pmd->fail_io)
1869 		r = __highest_block(td, result);
1870 	up_read(&pmd->root_lock);
1871 
1872 	return r;
1873 }
1874 
1875 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1876 {
1877 	int r;
1878 	dm_block_t old_count;
1879 
1880 	r = dm_sm_get_nr_blocks(sm, &old_count);
1881 	if (r)
1882 		return r;
1883 
1884 	if (new_count == old_count)
1885 		return 0;
1886 
1887 	if (new_count < old_count) {
1888 		DMERR("cannot reduce size of space map");
1889 		return -EINVAL;
1890 	}
1891 
1892 	return dm_sm_extend(sm, new_count - old_count);
1893 }
1894 
1895 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1896 {
1897 	int r = -EINVAL;
1898 
1899 	down_write(&pmd->root_lock);
1900 	if (!pmd->fail_io)
1901 		r = __resize_space_map(pmd->data_sm, new_count);
1902 	up_write(&pmd->root_lock);
1903 
1904 	return r;
1905 }
1906 
1907 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1908 {
1909 	int r = -EINVAL;
1910 
1911 	down_write(&pmd->root_lock);
1912 	if (!pmd->fail_io)
1913 		r = __resize_space_map(pmd->metadata_sm, new_count);
1914 	up_write(&pmd->root_lock);
1915 
1916 	return r;
1917 }
1918 
1919 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1920 {
1921 	down_write(&pmd->root_lock);
1922 	dm_bm_set_read_only(pmd->bm);
1923 	up_write(&pmd->root_lock);
1924 }
1925 
1926 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
1927 {
1928 	down_write(&pmd->root_lock);
1929 	dm_bm_set_read_write(pmd->bm);
1930 	up_write(&pmd->root_lock);
1931 }
1932 
1933 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
1934 					dm_block_t threshold,
1935 					dm_sm_threshold_fn fn,
1936 					void *context)
1937 {
1938 	int r;
1939 
1940 	down_write(&pmd->root_lock);
1941 	r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
1942 	up_write(&pmd->root_lock);
1943 
1944 	return r;
1945 }
1946 
1947 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
1948 {
1949 	int r;
1950 	struct dm_block *sblock;
1951 	struct thin_disk_superblock *disk_super;
1952 
1953 	down_write(&pmd->root_lock);
1954 	pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
1955 
1956 	r = superblock_lock(pmd, &sblock);
1957 	if (r) {
1958 		DMERR("couldn't read superblock");
1959 		goto out;
1960 	}
1961 
1962 	disk_super = dm_block_data(sblock);
1963 	disk_super->flags = cpu_to_le32(pmd->flags);
1964 
1965 	dm_bm_unlock(sblock);
1966 out:
1967 	up_write(&pmd->root_lock);
1968 	return r;
1969 }
1970 
1971 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
1972 {
1973 	bool needs_check;
1974 
1975 	down_read(&pmd->root_lock);
1976 	needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
1977 	up_read(&pmd->root_lock);
1978 
1979 	return needs_check;
1980 }
1981 
1982 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
1983 {
1984 	down_read(&pmd->root_lock);
1985 	if (!pmd->fail_io)
1986 		dm_tm_issue_prefetches(pmd->tm);
1987 	up_read(&pmd->root_lock);
1988 }
1989