xref: /linux/drivers/md/dm-thin-metadata.c (revision ec63e2a4897075e427c121d863bd89c44578094f)
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 SECTOR_TO_BLOCK_SHIFT 3
81 
82 /*
83  * For btree insert:
84  *  3 for btree insert +
85  *  2 for btree lookup used within space map
86  * For btree remove:
87  *  2 for shadow spine +
88  *  4 for rebalance 3 child node
89  */
90 #define THIN_MAX_CONCURRENT_LOCKS 6
91 
92 /* This should be plenty */
93 #define SPACE_MAP_ROOT_SIZE 128
94 
95 /*
96  * Little endian on-disk superblock and device details.
97  */
98 struct thin_disk_superblock {
99 	__le32 csum;	/* Checksum of superblock except for this field. */
100 	__le32 flags;
101 	__le64 blocknr;	/* This block number, dm_block_t. */
102 
103 	__u8 uuid[16];
104 	__le64 magic;
105 	__le32 version;
106 	__le32 time;
107 
108 	__le64 trans_id;
109 
110 	/*
111 	 * Root held by userspace transactions.
112 	 */
113 	__le64 held_root;
114 
115 	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
116 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
117 
118 	/*
119 	 * 2-level btree mapping (dev_id, (dev block, time)) -> data block
120 	 */
121 	__le64 data_mapping_root;
122 
123 	/*
124 	 * Device detail root mapping dev_id -> device_details
125 	 */
126 	__le64 device_details_root;
127 
128 	__le32 data_block_size;		/* In 512-byte sectors. */
129 
130 	__le32 metadata_block_size;	/* In 512-byte sectors. */
131 	__le64 metadata_nr_blocks;
132 
133 	__le32 compat_flags;
134 	__le32 compat_ro_flags;
135 	__le32 incompat_flags;
136 } __packed;
137 
138 struct disk_device_details {
139 	__le64 mapped_blocks;
140 	__le64 transaction_id;		/* When created. */
141 	__le32 creation_time;
142 	__le32 snapshotted_time;
143 } __packed;
144 
145 struct dm_pool_metadata {
146 	struct hlist_node hash;
147 
148 	struct block_device *bdev;
149 	struct dm_block_manager *bm;
150 	struct dm_space_map *metadata_sm;
151 	struct dm_space_map *data_sm;
152 	struct dm_transaction_manager *tm;
153 	struct dm_transaction_manager *nb_tm;
154 
155 	/*
156 	 * Two-level btree.
157 	 * First level holds thin_dev_t.
158 	 * Second level holds mappings.
159 	 */
160 	struct dm_btree_info info;
161 
162 	/*
163 	 * Non-blocking version of the above.
164 	 */
165 	struct dm_btree_info nb_info;
166 
167 	/*
168 	 * Just the top level for deleting whole devices.
169 	 */
170 	struct dm_btree_info tl_info;
171 
172 	/*
173 	 * Just the bottom level for creating new devices.
174 	 */
175 	struct dm_btree_info bl_info;
176 
177 	/*
178 	 * Describes the device details btree.
179 	 */
180 	struct dm_btree_info details_info;
181 
182 	struct rw_semaphore root_lock;
183 	uint32_t time;
184 	dm_block_t root;
185 	dm_block_t details_root;
186 	struct list_head thin_devices;
187 	uint64_t trans_id;
188 	unsigned long flags;
189 	sector_t data_block_size;
190 
191 	/*
192 	 * We reserve a section of the metadata for commit overhead.
193 	 * All reported space does *not* include this.
194 	 */
195 	dm_block_t metadata_reserve;
196 
197 	/*
198 	 * Set if a transaction has to be aborted but the attempt to roll back
199 	 * to the previous (good) transaction failed.  The only pool metadata
200 	 * operation possible in this state is the closing of the device.
201 	 */
202 	bool fail_io:1;
203 
204 	/*
205 	 * Reading the space map roots can fail, so we read it into these
206 	 * buffers before the superblock is locked and updated.
207 	 */
208 	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
209 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
210 };
211 
212 struct dm_thin_device {
213 	struct list_head list;
214 	struct dm_pool_metadata *pmd;
215 	dm_thin_id id;
216 
217 	int open_count;
218 	bool changed:1;
219 	bool aborted_with_changes:1;
220 	uint64_t mapped_blocks;
221 	uint64_t transaction_id;
222 	uint32_t creation_time;
223 	uint32_t snapshotted_time;
224 };
225 
226 /*----------------------------------------------------------------
227  * superblock validator
228  *--------------------------------------------------------------*/
229 
230 #define SUPERBLOCK_CSUM_XOR 160774
231 
232 static void sb_prepare_for_write(struct dm_block_validator *v,
233 				 struct dm_block *b,
234 				 size_t block_size)
235 {
236 	struct thin_disk_superblock *disk_super = dm_block_data(b);
237 
238 	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
239 	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
240 						      block_size - sizeof(__le32),
241 						      SUPERBLOCK_CSUM_XOR));
242 }
243 
244 static int sb_check(struct dm_block_validator *v,
245 		    struct dm_block *b,
246 		    size_t block_size)
247 {
248 	struct thin_disk_superblock *disk_super = dm_block_data(b);
249 	__le32 csum_le;
250 
251 	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
252 		DMERR("sb_check failed: blocknr %llu: "
253 		      "wanted %llu", le64_to_cpu(disk_super->blocknr),
254 		      (unsigned long long)dm_block_location(b));
255 		return -ENOTBLK;
256 	}
257 
258 	if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
259 		DMERR("sb_check failed: magic %llu: "
260 		      "wanted %llu", le64_to_cpu(disk_super->magic),
261 		      (unsigned long long)THIN_SUPERBLOCK_MAGIC);
262 		return -EILSEQ;
263 	}
264 
265 	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
266 					     block_size - sizeof(__le32),
267 					     SUPERBLOCK_CSUM_XOR));
268 	if (csum_le != disk_super->csum) {
269 		DMERR("sb_check failed: csum %u: wanted %u",
270 		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
271 		return -EILSEQ;
272 	}
273 
274 	return 0;
275 }
276 
277 static struct dm_block_validator sb_validator = {
278 	.name = "superblock",
279 	.prepare_for_write = sb_prepare_for_write,
280 	.check = sb_check
281 };
282 
283 /*----------------------------------------------------------------
284  * Methods for the btree value types
285  *--------------------------------------------------------------*/
286 
287 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
288 {
289 	return (b << 24) | t;
290 }
291 
292 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
293 {
294 	*b = v >> 24;
295 	*t = v & ((1 << 24) - 1);
296 }
297 
298 static void data_block_inc(void *context, const void *value_le)
299 {
300 	struct dm_space_map *sm = context;
301 	__le64 v_le;
302 	uint64_t b;
303 	uint32_t t;
304 
305 	memcpy(&v_le, value_le, sizeof(v_le));
306 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
307 	dm_sm_inc_block(sm, b);
308 }
309 
310 static void data_block_dec(void *context, const void *value_le)
311 {
312 	struct dm_space_map *sm = context;
313 	__le64 v_le;
314 	uint64_t b;
315 	uint32_t t;
316 
317 	memcpy(&v_le, value_le, sizeof(v_le));
318 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
319 	dm_sm_dec_block(sm, b);
320 }
321 
322 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
323 {
324 	__le64 v1_le, v2_le;
325 	uint64_t b1, b2;
326 	uint32_t t;
327 
328 	memcpy(&v1_le, value1_le, sizeof(v1_le));
329 	memcpy(&v2_le, value2_le, sizeof(v2_le));
330 	unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
331 	unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
332 
333 	return b1 == b2;
334 }
335 
336 static void subtree_inc(void *context, const void *value)
337 {
338 	struct dm_btree_info *info = context;
339 	__le64 root_le;
340 	uint64_t root;
341 
342 	memcpy(&root_le, value, sizeof(root_le));
343 	root = le64_to_cpu(root_le);
344 	dm_tm_inc(info->tm, root);
345 }
346 
347 static void subtree_dec(void *context, const void *value)
348 {
349 	struct dm_btree_info *info = context;
350 	__le64 root_le;
351 	uint64_t root;
352 
353 	memcpy(&root_le, value, sizeof(root_le));
354 	root = le64_to_cpu(root_le);
355 	if (dm_btree_del(info, root))
356 		DMERR("btree delete failed");
357 }
358 
359 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
360 {
361 	__le64 v1_le, v2_le;
362 	memcpy(&v1_le, value1_le, sizeof(v1_le));
363 	memcpy(&v2_le, value2_le, sizeof(v2_le));
364 
365 	return v1_le == v2_le;
366 }
367 
368 /*----------------------------------------------------------------*/
369 
370 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
371 				struct dm_block **sblock)
372 {
373 	return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
374 				     &sb_validator, sblock);
375 }
376 
377 static int superblock_lock(struct dm_pool_metadata *pmd,
378 			   struct dm_block **sblock)
379 {
380 	return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
381 				&sb_validator, sblock);
382 }
383 
384 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
385 {
386 	int r;
387 	unsigned i;
388 	struct dm_block *b;
389 	__le64 *data_le, zero = cpu_to_le64(0);
390 	unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
391 
392 	/*
393 	 * We can't use a validator here - it may be all zeroes.
394 	 */
395 	r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
396 	if (r)
397 		return r;
398 
399 	data_le = dm_block_data(b);
400 	*result = 1;
401 	for (i = 0; i < block_size; i++) {
402 		if (data_le[i] != zero) {
403 			*result = 0;
404 			break;
405 		}
406 	}
407 
408 	dm_bm_unlock(b);
409 
410 	return 0;
411 }
412 
413 static void __setup_btree_details(struct dm_pool_metadata *pmd)
414 {
415 	pmd->info.tm = pmd->tm;
416 	pmd->info.levels = 2;
417 	pmd->info.value_type.context = pmd->data_sm;
418 	pmd->info.value_type.size = sizeof(__le64);
419 	pmd->info.value_type.inc = data_block_inc;
420 	pmd->info.value_type.dec = data_block_dec;
421 	pmd->info.value_type.equal = data_block_equal;
422 
423 	memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
424 	pmd->nb_info.tm = pmd->nb_tm;
425 
426 	pmd->tl_info.tm = pmd->tm;
427 	pmd->tl_info.levels = 1;
428 	pmd->tl_info.value_type.context = &pmd->bl_info;
429 	pmd->tl_info.value_type.size = sizeof(__le64);
430 	pmd->tl_info.value_type.inc = subtree_inc;
431 	pmd->tl_info.value_type.dec = subtree_dec;
432 	pmd->tl_info.value_type.equal = subtree_equal;
433 
434 	pmd->bl_info.tm = pmd->tm;
435 	pmd->bl_info.levels = 1;
436 	pmd->bl_info.value_type.context = pmd->data_sm;
437 	pmd->bl_info.value_type.size = sizeof(__le64);
438 	pmd->bl_info.value_type.inc = data_block_inc;
439 	pmd->bl_info.value_type.dec = data_block_dec;
440 	pmd->bl_info.value_type.equal = data_block_equal;
441 
442 	pmd->details_info.tm = pmd->tm;
443 	pmd->details_info.levels = 1;
444 	pmd->details_info.value_type.context = NULL;
445 	pmd->details_info.value_type.size = sizeof(struct disk_device_details);
446 	pmd->details_info.value_type.inc = NULL;
447 	pmd->details_info.value_type.dec = NULL;
448 	pmd->details_info.value_type.equal = NULL;
449 }
450 
451 static int save_sm_roots(struct dm_pool_metadata *pmd)
452 {
453 	int r;
454 	size_t len;
455 
456 	r = dm_sm_root_size(pmd->metadata_sm, &len);
457 	if (r < 0)
458 		return r;
459 
460 	r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
461 	if (r < 0)
462 		return r;
463 
464 	r = dm_sm_root_size(pmd->data_sm, &len);
465 	if (r < 0)
466 		return r;
467 
468 	return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
469 }
470 
471 static void copy_sm_roots(struct dm_pool_metadata *pmd,
472 			  struct thin_disk_superblock *disk)
473 {
474 	memcpy(&disk->metadata_space_map_root,
475 	       &pmd->metadata_space_map_root,
476 	       sizeof(pmd->metadata_space_map_root));
477 
478 	memcpy(&disk->data_space_map_root,
479 	       &pmd->data_space_map_root,
480 	       sizeof(pmd->data_space_map_root));
481 }
482 
483 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
484 {
485 	int r;
486 	struct dm_block *sblock;
487 	struct thin_disk_superblock *disk_super;
488 	sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
489 
490 	if (bdev_size > THIN_METADATA_MAX_SECTORS)
491 		bdev_size = THIN_METADATA_MAX_SECTORS;
492 
493 	r = dm_sm_commit(pmd->data_sm);
494 	if (r < 0)
495 		return r;
496 
497 	r = dm_tm_pre_commit(pmd->tm);
498 	if (r < 0)
499 		return r;
500 
501 	r = save_sm_roots(pmd);
502 	if (r < 0)
503 		return r;
504 
505 	r = superblock_lock_zero(pmd, &sblock);
506 	if (r)
507 		return r;
508 
509 	disk_super = dm_block_data(sblock);
510 	disk_super->flags = 0;
511 	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
512 	disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
513 	disk_super->version = cpu_to_le32(THIN_VERSION);
514 	disk_super->time = 0;
515 	disk_super->trans_id = 0;
516 	disk_super->held_root = 0;
517 
518 	copy_sm_roots(pmd, disk_super);
519 
520 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
521 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
522 	disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
523 	disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
524 	disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
525 
526 	return dm_tm_commit(pmd->tm, sblock);
527 }
528 
529 static int __format_metadata(struct dm_pool_metadata *pmd)
530 {
531 	int r;
532 
533 	r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
534 				 &pmd->tm, &pmd->metadata_sm);
535 	if (r < 0) {
536 		DMERR("tm_create_with_sm failed");
537 		return r;
538 	}
539 
540 	pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
541 	if (IS_ERR(pmd->data_sm)) {
542 		DMERR("sm_disk_create failed");
543 		r = PTR_ERR(pmd->data_sm);
544 		goto bad_cleanup_tm;
545 	}
546 
547 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
548 	if (!pmd->nb_tm) {
549 		DMERR("could not create non-blocking clone tm");
550 		r = -ENOMEM;
551 		goto bad_cleanup_data_sm;
552 	}
553 
554 	__setup_btree_details(pmd);
555 
556 	r = dm_btree_empty(&pmd->info, &pmd->root);
557 	if (r < 0)
558 		goto bad_cleanup_nb_tm;
559 
560 	r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
561 	if (r < 0) {
562 		DMERR("couldn't create devices root");
563 		goto bad_cleanup_nb_tm;
564 	}
565 
566 	r = __write_initial_superblock(pmd);
567 	if (r)
568 		goto bad_cleanup_nb_tm;
569 
570 	return 0;
571 
572 bad_cleanup_nb_tm:
573 	dm_tm_destroy(pmd->nb_tm);
574 bad_cleanup_data_sm:
575 	dm_sm_destroy(pmd->data_sm);
576 bad_cleanup_tm:
577 	dm_tm_destroy(pmd->tm);
578 	dm_sm_destroy(pmd->metadata_sm);
579 
580 	return r;
581 }
582 
583 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
584 				     struct dm_pool_metadata *pmd)
585 {
586 	uint32_t features;
587 
588 	features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
589 	if (features) {
590 		DMERR("could not access metadata due to unsupported optional features (%lx).",
591 		      (unsigned long)features);
592 		return -EINVAL;
593 	}
594 
595 	/*
596 	 * Check for read-only metadata to skip the following RDWR checks.
597 	 */
598 	if (get_disk_ro(pmd->bdev->bd_disk))
599 		return 0;
600 
601 	features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
602 	if (features) {
603 		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
604 		      (unsigned long)features);
605 		return -EINVAL;
606 	}
607 
608 	return 0;
609 }
610 
611 static int __open_metadata(struct dm_pool_metadata *pmd)
612 {
613 	int r;
614 	struct dm_block *sblock;
615 	struct thin_disk_superblock *disk_super;
616 
617 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
618 			    &sb_validator, &sblock);
619 	if (r < 0) {
620 		DMERR("couldn't read superblock");
621 		return r;
622 	}
623 
624 	disk_super = dm_block_data(sblock);
625 
626 	/* Verify the data block size hasn't changed */
627 	if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
628 		DMERR("changing the data block size (from %u to %llu) is not supported",
629 		      le32_to_cpu(disk_super->data_block_size),
630 		      (unsigned long long)pmd->data_block_size);
631 		r = -EINVAL;
632 		goto bad_unlock_sblock;
633 	}
634 
635 	r = __check_incompat_features(disk_super, pmd);
636 	if (r < 0)
637 		goto bad_unlock_sblock;
638 
639 	r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
640 			       disk_super->metadata_space_map_root,
641 			       sizeof(disk_super->metadata_space_map_root),
642 			       &pmd->tm, &pmd->metadata_sm);
643 	if (r < 0) {
644 		DMERR("tm_open_with_sm failed");
645 		goto bad_unlock_sblock;
646 	}
647 
648 	pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
649 				       sizeof(disk_super->data_space_map_root));
650 	if (IS_ERR(pmd->data_sm)) {
651 		DMERR("sm_disk_open failed");
652 		r = PTR_ERR(pmd->data_sm);
653 		goto bad_cleanup_tm;
654 	}
655 
656 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
657 	if (!pmd->nb_tm) {
658 		DMERR("could not create non-blocking clone tm");
659 		r = -ENOMEM;
660 		goto bad_cleanup_data_sm;
661 	}
662 
663 	__setup_btree_details(pmd);
664 	dm_bm_unlock(sblock);
665 
666 	return 0;
667 
668 bad_cleanup_data_sm:
669 	dm_sm_destroy(pmd->data_sm);
670 bad_cleanup_tm:
671 	dm_tm_destroy(pmd->tm);
672 	dm_sm_destroy(pmd->metadata_sm);
673 bad_unlock_sblock:
674 	dm_bm_unlock(sblock);
675 
676 	return r;
677 }
678 
679 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
680 {
681 	int r, unformatted;
682 
683 	r = __superblock_all_zeroes(pmd->bm, &unformatted);
684 	if (r)
685 		return r;
686 
687 	if (unformatted)
688 		return format_device ? __format_metadata(pmd) : -EPERM;
689 
690 	return __open_metadata(pmd);
691 }
692 
693 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
694 {
695 	int r;
696 
697 	pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
698 					  THIN_MAX_CONCURRENT_LOCKS);
699 	if (IS_ERR(pmd->bm)) {
700 		DMERR("could not create block manager");
701 		return PTR_ERR(pmd->bm);
702 	}
703 
704 	r = __open_or_format_metadata(pmd, format_device);
705 	if (r)
706 		dm_block_manager_destroy(pmd->bm);
707 
708 	return r;
709 }
710 
711 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
712 {
713 	dm_sm_destroy(pmd->data_sm);
714 	dm_sm_destroy(pmd->metadata_sm);
715 	dm_tm_destroy(pmd->nb_tm);
716 	dm_tm_destroy(pmd->tm);
717 	dm_block_manager_destroy(pmd->bm);
718 }
719 
720 static int __begin_transaction(struct dm_pool_metadata *pmd)
721 {
722 	int r;
723 	struct thin_disk_superblock *disk_super;
724 	struct dm_block *sblock;
725 
726 	/*
727 	 * We re-read the superblock every time.  Shouldn't need to do this
728 	 * really.
729 	 */
730 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
731 			    &sb_validator, &sblock);
732 	if (r)
733 		return r;
734 
735 	disk_super = dm_block_data(sblock);
736 	pmd->time = le32_to_cpu(disk_super->time);
737 	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
738 	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
739 	pmd->trans_id = le64_to_cpu(disk_super->trans_id);
740 	pmd->flags = le32_to_cpu(disk_super->flags);
741 	pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
742 
743 	dm_bm_unlock(sblock);
744 	return 0;
745 }
746 
747 static int __write_changed_details(struct dm_pool_metadata *pmd)
748 {
749 	int r;
750 	struct dm_thin_device *td, *tmp;
751 	struct disk_device_details details;
752 	uint64_t key;
753 
754 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
755 		if (!td->changed)
756 			continue;
757 
758 		key = td->id;
759 
760 		details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
761 		details.transaction_id = cpu_to_le64(td->transaction_id);
762 		details.creation_time = cpu_to_le32(td->creation_time);
763 		details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
764 		__dm_bless_for_disk(&details);
765 
766 		r = dm_btree_insert(&pmd->details_info, pmd->details_root,
767 				    &key, &details, &pmd->details_root);
768 		if (r)
769 			return r;
770 
771 		if (td->open_count)
772 			td->changed = 0;
773 		else {
774 			list_del(&td->list);
775 			kfree(td);
776 		}
777 	}
778 
779 	return 0;
780 }
781 
782 static int __commit_transaction(struct dm_pool_metadata *pmd)
783 {
784 	int r;
785 	struct thin_disk_superblock *disk_super;
786 	struct dm_block *sblock;
787 
788 	/*
789 	 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
790 	 */
791 	BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
792 
793 	r = __write_changed_details(pmd);
794 	if (r < 0)
795 		return r;
796 
797 	r = dm_sm_commit(pmd->data_sm);
798 	if (r < 0)
799 		return r;
800 
801 	r = dm_tm_pre_commit(pmd->tm);
802 	if (r < 0)
803 		return r;
804 
805 	r = save_sm_roots(pmd);
806 	if (r < 0)
807 		return r;
808 
809 	r = superblock_lock(pmd, &sblock);
810 	if (r)
811 		return r;
812 
813 	disk_super = dm_block_data(sblock);
814 	disk_super->time = cpu_to_le32(pmd->time);
815 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
816 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
817 	disk_super->trans_id = cpu_to_le64(pmd->trans_id);
818 	disk_super->flags = cpu_to_le32(pmd->flags);
819 
820 	copy_sm_roots(pmd, disk_super);
821 
822 	return dm_tm_commit(pmd->tm, sblock);
823 }
824 
825 static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
826 {
827 	int r;
828 	dm_block_t total;
829 	dm_block_t max_blocks = 4096; /* 16M */
830 
831 	r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
832 	if (r) {
833 		DMERR("could not get size of metadata device");
834 		pmd->metadata_reserve = max_blocks;
835 	} else
836 		pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
837 }
838 
839 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
840 					       sector_t data_block_size,
841 					       bool format_device)
842 {
843 	int r;
844 	struct dm_pool_metadata *pmd;
845 
846 	pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
847 	if (!pmd) {
848 		DMERR("could not allocate metadata struct");
849 		return ERR_PTR(-ENOMEM);
850 	}
851 
852 	init_rwsem(&pmd->root_lock);
853 	pmd->time = 0;
854 	INIT_LIST_HEAD(&pmd->thin_devices);
855 	pmd->fail_io = false;
856 	pmd->bdev = bdev;
857 	pmd->data_block_size = data_block_size;
858 
859 	r = __create_persistent_data_objects(pmd, format_device);
860 	if (r) {
861 		kfree(pmd);
862 		return ERR_PTR(r);
863 	}
864 
865 	r = __begin_transaction(pmd);
866 	if (r < 0) {
867 		if (dm_pool_metadata_close(pmd) < 0)
868 			DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
869 		return ERR_PTR(r);
870 	}
871 
872 	__set_metadata_reserve(pmd);
873 
874 	return pmd;
875 }
876 
877 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
878 {
879 	int r;
880 	unsigned open_devices = 0;
881 	struct dm_thin_device *td, *tmp;
882 
883 	down_read(&pmd->root_lock);
884 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
885 		if (td->open_count)
886 			open_devices++;
887 		else {
888 			list_del(&td->list);
889 			kfree(td);
890 		}
891 	}
892 	up_read(&pmd->root_lock);
893 
894 	if (open_devices) {
895 		DMERR("attempt to close pmd when %u device(s) are still open",
896 		       open_devices);
897 		return -EBUSY;
898 	}
899 
900 	if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
901 		r = __commit_transaction(pmd);
902 		if (r < 0)
903 			DMWARN("%s: __commit_transaction() failed, error = %d",
904 			       __func__, r);
905 	}
906 
907 	if (!pmd->fail_io)
908 		__destroy_persistent_data_objects(pmd);
909 
910 	kfree(pmd);
911 	return 0;
912 }
913 
914 /*
915  * __open_device: Returns @td corresponding to device with id @dev,
916  * creating it if @create is set and incrementing @td->open_count.
917  * On failure, @td is undefined.
918  */
919 static int __open_device(struct dm_pool_metadata *pmd,
920 			 dm_thin_id dev, int create,
921 			 struct dm_thin_device **td)
922 {
923 	int r, changed = 0;
924 	struct dm_thin_device *td2;
925 	uint64_t key = dev;
926 	struct disk_device_details details_le;
927 
928 	/*
929 	 * If the device is already open, return it.
930 	 */
931 	list_for_each_entry(td2, &pmd->thin_devices, list)
932 		if (td2->id == dev) {
933 			/*
934 			 * May not create an already-open device.
935 			 */
936 			if (create)
937 				return -EEXIST;
938 
939 			td2->open_count++;
940 			*td = td2;
941 			return 0;
942 		}
943 
944 	/*
945 	 * Check the device exists.
946 	 */
947 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
948 			    &key, &details_le);
949 	if (r) {
950 		if (r != -ENODATA || !create)
951 			return r;
952 
953 		/*
954 		 * Create new device.
955 		 */
956 		changed = 1;
957 		details_le.mapped_blocks = 0;
958 		details_le.transaction_id = cpu_to_le64(pmd->trans_id);
959 		details_le.creation_time = cpu_to_le32(pmd->time);
960 		details_le.snapshotted_time = cpu_to_le32(pmd->time);
961 	}
962 
963 	*td = kmalloc(sizeof(**td), GFP_NOIO);
964 	if (!*td)
965 		return -ENOMEM;
966 
967 	(*td)->pmd = pmd;
968 	(*td)->id = dev;
969 	(*td)->open_count = 1;
970 	(*td)->changed = changed;
971 	(*td)->aborted_with_changes = false;
972 	(*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
973 	(*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
974 	(*td)->creation_time = le32_to_cpu(details_le.creation_time);
975 	(*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
976 
977 	list_add(&(*td)->list, &pmd->thin_devices);
978 
979 	return 0;
980 }
981 
982 static void __close_device(struct dm_thin_device *td)
983 {
984 	--td->open_count;
985 }
986 
987 static int __create_thin(struct dm_pool_metadata *pmd,
988 			 dm_thin_id dev)
989 {
990 	int r;
991 	dm_block_t dev_root;
992 	uint64_t key = dev;
993 	struct disk_device_details details_le;
994 	struct dm_thin_device *td;
995 	__le64 value;
996 
997 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
998 			    &key, &details_le);
999 	if (!r)
1000 		return -EEXIST;
1001 
1002 	/*
1003 	 * Create an empty btree for the mappings.
1004 	 */
1005 	r = dm_btree_empty(&pmd->bl_info, &dev_root);
1006 	if (r)
1007 		return r;
1008 
1009 	/*
1010 	 * Insert it into the main mapping tree.
1011 	 */
1012 	value = cpu_to_le64(dev_root);
1013 	__dm_bless_for_disk(&value);
1014 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1015 	if (r) {
1016 		dm_btree_del(&pmd->bl_info, dev_root);
1017 		return r;
1018 	}
1019 
1020 	r = __open_device(pmd, dev, 1, &td);
1021 	if (r) {
1022 		dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1023 		dm_btree_del(&pmd->bl_info, dev_root);
1024 		return r;
1025 	}
1026 	__close_device(td);
1027 
1028 	return r;
1029 }
1030 
1031 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1032 {
1033 	int r = -EINVAL;
1034 
1035 	down_write(&pmd->root_lock);
1036 	if (!pmd->fail_io)
1037 		r = __create_thin(pmd, dev);
1038 	up_write(&pmd->root_lock);
1039 
1040 	return r;
1041 }
1042 
1043 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1044 				  struct dm_thin_device *snap,
1045 				  dm_thin_id origin, uint32_t time)
1046 {
1047 	int r;
1048 	struct dm_thin_device *td;
1049 
1050 	r = __open_device(pmd, origin, 0, &td);
1051 	if (r)
1052 		return r;
1053 
1054 	td->changed = 1;
1055 	td->snapshotted_time = time;
1056 
1057 	snap->mapped_blocks = td->mapped_blocks;
1058 	snap->snapshotted_time = time;
1059 	__close_device(td);
1060 
1061 	return 0;
1062 }
1063 
1064 static int __create_snap(struct dm_pool_metadata *pmd,
1065 			 dm_thin_id dev, dm_thin_id origin)
1066 {
1067 	int r;
1068 	dm_block_t origin_root;
1069 	uint64_t key = origin, dev_key = dev;
1070 	struct dm_thin_device *td;
1071 	struct disk_device_details details_le;
1072 	__le64 value;
1073 
1074 	/* check this device is unused */
1075 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1076 			    &dev_key, &details_le);
1077 	if (!r)
1078 		return -EEXIST;
1079 
1080 	/* find the mapping tree for the origin */
1081 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1082 	if (r)
1083 		return r;
1084 	origin_root = le64_to_cpu(value);
1085 
1086 	/* clone the origin, an inc will do */
1087 	dm_tm_inc(pmd->tm, origin_root);
1088 
1089 	/* insert into the main mapping tree */
1090 	value = cpu_to_le64(origin_root);
1091 	__dm_bless_for_disk(&value);
1092 	key = dev;
1093 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1094 	if (r) {
1095 		dm_tm_dec(pmd->tm, origin_root);
1096 		return r;
1097 	}
1098 
1099 	pmd->time++;
1100 
1101 	r = __open_device(pmd, dev, 1, &td);
1102 	if (r)
1103 		goto bad;
1104 
1105 	r = __set_snapshot_details(pmd, td, origin, pmd->time);
1106 	__close_device(td);
1107 
1108 	if (r)
1109 		goto bad;
1110 
1111 	return 0;
1112 
1113 bad:
1114 	dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1115 	dm_btree_remove(&pmd->details_info, pmd->details_root,
1116 			&key, &pmd->details_root);
1117 	return r;
1118 }
1119 
1120 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1121 				 dm_thin_id dev,
1122 				 dm_thin_id origin)
1123 {
1124 	int r = -EINVAL;
1125 
1126 	down_write(&pmd->root_lock);
1127 	if (!pmd->fail_io)
1128 		r = __create_snap(pmd, dev, origin);
1129 	up_write(&pmd->root_lock);
1130 
1131 	return r;
1132 }
1133 
1134 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1135 {
1136 	int r;
1137 	uint64_t key = dev;
1138 	struct dm_thin_device *td;
1139 
1140 	/* TODO: failure should mark the transaction invalid */
1141 	r = __open_device(pmd, dev, 0, &td);
1142 	if (r)
1143 		return r;
1144 
1145 	if (td->open_count > 1) {
1146 		__close_device(td);
1147 		return -EBUSY;
1148 	}
1149 
1150 	list_del(&td->list);
1151 	kfree(td);
1152 	r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1153 			    &key, &pmd->details_root);
1154 	if (r)
1155 		return r;
1156 
1157 	r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1158 	if (r)
1159 		return r;
1160 
1161 	return 0;
1162 }
1163 
1164 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1165 			       dm_thin_id dev)
1166 {
1167 	int r = -EINVAL;
1168 
1169 	down_write(&pmd->root_lock);
1170 	if (!pmd->fail_io)
1171 		r = __delete_device(pmd, dev);
1172 	up_write(&pmd->root_lock);
1173 
1174 	return r;
1175 }
1176 
1177 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1178 					uint64_t current_id,
1179 					uint64_t new_id)
1180 {
1181 	int r = -EINVAL;
1182 
1183 	down_write(&pmd->root_lock);
1184 
1185 	if (pmd->fail_io)
1186 		goto out;
1187 
1188 	if (pmd->trans_id != current_id) {
1189 		DMERR("mismatched transaction id");
1190 		goto out;
1191 	}
1192 
1193 	pmd->trans_id = new_id;
1194 	r = 0;
1195 
1196 out:
1197 	up_write(&pmd->root_lock);
1198 
1199 	return r;
1200 }
1201 
1202 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1203 					uint64_t *result)
1204 {
1205 	int r = -EINVAL;
1206 
1207 	down_read(&pmd->root_lock);
1208 	if (!pmd->fail_io) {
1209 		*result = pmd->trans_id;
1210 		r = 0;
1211 	}
1212 	up_read(&pmd->root_lock);
1213 
1214 	return r;
1215 }
1216 
1217 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1218 {
1219 	int r, inc;
1220 	struct thin_disk_superblock *disk_super;
1221 	struct dm_block *copy, *sblock;
1222 	dm_block_t held_root;
1223 
1224 	/*
1225 	 * We commit to ensure the btree roots which we increment in a
1226 	 * moment are up to date.
1227 	 */
1228 	__commit_transaction(pmd);
1229 
1230 	/*
1231 	 * Copy the superblock.
1232 	 */
1233 	dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1234 	r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1235 			       &sb_validator, &copy, &inc);
1236 	if (r)
1237 		return r;
1238 
1239 	BUG_ON(!inc);
1240 
1241 	held_root = dm_block_location(copy);
1242 	disk_super = dm_block_data(copy);
1243 
1244 	if (le64_to_cpu(disk_super->held_root)) {
1245 		DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1246 
1247 		dm_tm_dec(pmd->tm, held_root);
1248 		dm_tm_unlock(pmd->tm, copy);
1249 		return -EBUSY;
1250 	}
1251 
1252 	/*
1253 	 * Wipe the spacemap since we're not publishing this.
1254 	 */
1255 	memset(&disk_super->data_space_map_root, 0,
1256 	       sizeof(disk_super->data_space_map_root));
1257 	memset(&disk_super->metadata_space_map_root, 0,
1258 	       sizeof(disk_super->metadata_space_map_root));
1259 
1260 	/*
1261 	 * Increment the data structures that need to be preserved.
1262 	 */
1263 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1264 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1265 	dm_tm_unlock(pmd->tm, copy);
1266 
1267 	/*
1268 	 * Write the held root into the superblock.
1269 	 */
1270 	r = superblock_lock(pmd, &sblock);
1271 	if (r) {
1272 		dm_tm_dec(pmd->tm, held_root);
1273 		return r;
1274 	}
1275 
1276 	disk_super = dm_block_data(sblock);
1277 	disk_super->held_root = cpu_to_le64(held_root);
1278 	dm_bm_unlock(sblock);
1279 	return 0;
1280 }
1281 
1282 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1283 {
1284 	int r = -EINVAL;
1285 
1286 	down_write(&pmd->root_lock);
1287 	if (!pmd->fail_io)
1288 		r = __reserve_metadata_snap(pmd);
1289 	up_write(&pmd->root_lock);
1290 
1291 	return r;
1292 }
1293 
1294 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1295 {
1296 	int r;
1297 	struct thin_disk_superblock *disk_super;
1298 	struct dm_block *sblock, *copy;
1299 	dm_block_t held_root;
1300 
1301 	r = superblock_lock(pmd, &sblock);
1302 	if (r)
1303 		return r;
1304 
1305 	disk_super = dm_block_data(sblock);
1306 	held_root = le64_to_cpu(disk_super->held_root);
1307 	disk_super->held_root = cpu_to_le64(0);
1308 
1309 	dm_bm_unlock(sblock);
1310 
1311 	if (!held_root) {
1312 		DMWARN("No pool metadata snapshot found: nothing to release.");
1313 		return -EINVAL;
1314 	}
1315 
1316 	r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1317 	if (r)
1318 		return r;
1319 
1320 	disk_super = dm_block_data(copy);
1321 	dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1322 	dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1323 	dm_sm_dec_block(pmd->metadata_sm, held_root);
1324 
1325 	dm_tm_unlock(pmd->tm, copy);
1326 
1327 	return 0;
1328 }
1329 
1330 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1331 {
1332 	int r = -EINVAL;
1333 
1334 	down_write(&pmd->root_lock);
1335 	if (!pmd->fail_io)
1336 		r = __release_metadata_snap(pmd);
1337 	up_write(&pmd->root_lock);
1338 
1339 	return r;
1340 }
1341 
1342 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1343 			       dm_block_t *result)
1344 {
1345 	int r;
1346 	struct thin_disk_superblock *disk_super;
1347 	struct dm_block *sblock;
1348 
1349 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1350 			    &sb_validator, &sblock);
1351 	if (r)
1352 		return r;
1353 
1354 	disk_super = dm_block_data(sblock);
1355 	*result = le64_to_cpu(disk_super->held_root);
1356 
1357 	dm_bm_unlock(sblock);
1358 
1359 	return 0;
1360 }
1361 
1362 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1363 			      dm_block_t *result)
1364 {
1365 	int r = -EINVAL;
1366 
1367 	down_read(&pmd->root_lock);
1368 	if (!pmd->fail_io)
1369 		r = __get_metadata_snap(pmd, result);
1370 	up_read(&pmd->root_lock);
1371 
1372 	return r;
1373 }
1374 
1375 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1376 			     struct dm_thin_device **td)
1377 {
1378 	int r = -EINVAL;
1379 
1380 	down_write(&pmd->root_lock);
1381 	if (!pmd->fail_io)
1382 		r = __open_device(pmd, dev, 0, td);
1383 	up_write(&pmd->root_lock);
1384 
1385 	return r;
1386 }
1387 
1388 int dm_pool_close_thin_device(struct dm_thin_device *td)
1389 {
1390 	down_write(&td->pmd->root_lock);
1391 	__close_device(td);
1392 	up_write(&td->pmd->root_lock);
1393 
1394 	return 0;
1395 }
1396 
1397 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1398 {
1399 	return td->id;
1400 }
1401 
1402 /*
1403  * Check whether @time (of block creation) is older than @td's last snapshot.
1404  * If so then the associated block is shared with the last snapshot device.
1405  * Any block on a device created *after* the device last got snapshotted is
1406  * necessarily not shared.
1407  */
1408 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1409 {
1410 	return td->snapshotted_time > time;
1411 }
1412 
1413 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1414 				 struct dm_thin_lookup_result *result)
1415 {
1416 	uint64_t block_time = 0;
1417 	dm_block_t exception_block;
1418 	uint32_t exception_time;
1419 
1420 	block_time = le64_to_cpu(value);
1421 	unpack_block_time(block_time, &exception_block, &exception_time);
1422 	result->block = exception_block;
1423 	result->shared = __snapshotted_since(td, exception_time);
1424 }
1425 
1426 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1427 			int can_issue_io, struct dm_thin_lookup_result *result)
1428 {
1429 	int r;
1430 	__le64 value;
1431 	struct dm_pool_metadata *pmd = td->pmd;
1432 	dm_block_t keys[2] = { td->id, block };
1433 	struct dm_btree_info *info;
1434 
1435 	if (can_issue_io) {
1436 		info = &pmd->info;
1437 	} else
1438 		info = &pmd->nb_info;
1439 
1440 	r = dm_btree_lookup(info, pmd->root, keys, &value);
1441 	if (!r)
1442 		unpack_lookup_result(td, value, result);
1443 
1444 	return r;
1445 }
1446 
1447 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1448 		       int can_issue_io, struct dm_thin_lookup_result *result)
1449 {
1450 	int r;
1451 	struct dm_pool_metadata *pmd = td->pmd;
1452 
1453 	down_read(&pmd->root_lock);
1454 	if (pmd->fail_io) {
1455 		up_read(&pmd->root_lock);
1456 		return -EINVAL;
1457 	}
1458 
1459 	r = __find_block(td, block, can_issue_io, result);
1460 
1461 	up_read(&pmd->root_lock);
1462 	return r;
1463 }
1464 
1465 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1466 					  dm_block_t *vblock,
1467 					  struct dm_thin_lookup_result *result)
1468 {
1469 	int r;
1470 	__le64 value;
1471 	struct dm_pool_metadata *pmd = td->pmd;
1472 	dm_block_t keys[2] = { td->id, block };
1473 
1474 	r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1475 	if (!r)
1476 		unpack_lookup_result(td, value, result);
1477 
1478 	return r;
1479 }
1480 
1481 static int __find_mapped_range(struct dm_thin_device *td,
1482 			       dm_block_t begin, dm_block_t end,
1483 			       dm_block_t *thin_begin, dm_block_t *thin_end,
1484 			       dm_block_t *pool_begin, bool *maybe_shared)
1485 {
1486 	int r;
1487 	dm_block_t pool_end;
1488 	struct dm_thin_lookup_result lookup;
1489 
1490 	if (end < begin)
1491 		return -ENODATA;
1492 
1493 	r = __find_next_mapped_block(td, begin, &begin, &lookup);
1494 	if (r)
1495 		return r;
1496 
1497 	if (begin >= end)
1498 		return -ENODATA;
1499 
1500 	*thin_begin = begin;
1501 	*pool_begin = lookup.block;
1502 	*maybe_shared = lookup.shared;
1503 
1504 	begin++;
1505 	pool_end = *pool_begin + 1;
1506 	while (begin != end) {
1507 		r = __find_block(td, begin, true, &lookup);
1508 		if (r) {
1509 			if (r == -ENODATA)
1510 				break;
1511 			else
1512 				return r;
1513 		}
1514 
1515 		if ((lookup.block != pool_end) ||
1516 		    (lookup.shared != *maybe_shared))
1517 			break;
1518 
1519 		pool_end++;
1520 		begin++;
1521 	}
1522 
1523 	*thin_end = begin;
1524 	return 0;
1525 }
1526 
1527 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1528 			      dm_block_t begin, dm_block_t end,
1529 			      dm_block_t *thin_begin, dm_block_t *thin_end,
1530 			      dm_block_t *pool_begin, bool *maybe_shared)
1531 {
1532 	int r = -EINVAL;
1533 	struct dm_pool_metadata *pmd = td->pmd;
1534 
1535 	down_read(&pmd->root_lock);
1536 	if (!pmd->fail_io) {
1537 		r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1538 					pool_begin, maybe_shared);
1539 	}
1540 	up_read(&pmd->root_lock);
1541 
1542 	return r;
1543 }
1544 
1545 static int __insert(struct dm_thin_device *td, dm_block_t block,
1546 		    dm_block_t data_block)
1547 {
1548 	int r, inserted;
1549 	__le64 value;
1550 	struct dm_pool_metadata *pmd = td->pmd;
1551 	dm_block_t keys[2] = { td->id, block };
1552 
1553 	value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1554 	__dm_bless_for_disk(&value);
1555 
1556 	r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1557 				   &pmd->root, &inserted);
1558 	if (r)
1559 		return r;
1560 
1561 	td->changed = 1;
1562 	if (inserted)
1563 		td->mapped_blocks++;
1564 
1565 	return 0;
1566 }
1567 
1568 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1569 			 dm_block_t data_block)
1570 {
1571 	int r = -EINVAL;
1572 
1573 	down_write(&td->pmd->root_lock);
1574 	if (!td->pmd->fail_io)
1575 		r = __insert(td, block, data_block);
1576 	up_write(&td->pmd->root_lock);
1577 
1578 	return r;
1579 }
1580 
1581 static int __remove(struct dm_thin_device *td, dm_block_t block)
1582 {
1583 	int r;
1584 	struct dm_pool_metadata *pmd = td->pmd;
1585 	dm_block_t keys[2] = { td->id, block };
1586 
1587 	r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1588 	if (r)
1589 		return r;
1590 
1591 	td->mapped_blocks--;
1592 	td->changed = 1;
1593 
1594 	return 0;
1595 }
1596 
1597 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1598 {
1599 	int r;
1600 	unsigned count, total_count = 0;
1601 	struct dm_pool_metadata *pmd = td->pmd;
1602 	dm_block_t keys[1] = { td->id };
1603 	__le64 value;
1604 	dm_block_t mapping_root;
1605 
1606 	/*
1607 	 * Find the mapping tree
1608 	 */
1609 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1610 	if (r)
1611 		return r;
1612 
1613 	/*
1614 	 * Remove from the mapping tree, taking care to inc the
1615 	 * ref count so it doesn't get deleted.
1616 	 */
1617 	mapping_root = le64_to_cpu(value);
1618 	dm_tm_inc(pmd->tm, mapping_root);
1619 	r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1620 	if (r)
1621 		return r;
1622 
1623 	/*
1624 	 * Remove leaves stops at the first unmapped entry, so we have to
1625 	 * loop round finding mapped ranges.
1626 	 */
1627 	while (begin < end) {
1628 		r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1629 		if (r == -ENODATA)
1630 			break;
1631 
1632 		if (r)
1633 			return r;
1634 
1635 		if (begin >= end)
1636 			break;
1637 
1638 		r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1639 		if (r)
1640 			return r;
1641 
1642 		total_count += count;
1643 	}
1644 
1645 	td->mapped_blocks -= total_count;
1646 	td->changed = 1;
1647 
1648 	/*
1649 	 * Reinsert the mapping tree.
1650 	 */
1651 	value = cpu_to_le64(mapping_root);
1652 	__dm_bless_for_disk(&value);
1653 	return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1654 }
1655 
1656 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1657 {
1658 	int r = -EINVAL;
1659 
1660 	down_write(&td->pmd->root_lock);
1661 	if (!td->pmd->fail_io)
1662 		r = __remove(td, block);
1663 	up_write(&td->pmd->root_lock);
1664 
1665 	return r;
1666 }
1667 
1668 int dm_thin_remove_range(struct dm_thin_device *td,
1669 			 dm_block_t begin, dm_block_t end)
1670 {
1671 	int r = -EINVAL;
1672 
1673 	down_write(&td->pmd->root_lock);
1674 	if (!td->pmd->fail_io)
1675 		r = __remove_range(td, begin, end);
1676 	up_write(&td->pmd->root_lock);
1677 
1678 	return r;
1679 }
1680 
1681 int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1682 {
1683 	int r;
1684 	uint32_t ref_count;
1685 
1686 	down_read(&pmd->root_lock);
1687 	r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1688 	if (!r)
1689 		*result = (ref_count > 1);
1690 	up_read(&pmd->root_lock);
1691 
1692 	return r;
1693 }
1694 
1695 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1696 {
1697 	int r = 0;
1698 
1699 	down_write(&pmd->root_lock);
1700 	for (; b != e; b++) {
1701 		r = dm_sm_inc_block(pmd->data_sm, b);
1702 		if (r)
1703 			break;
1704 	}
1705 	up_write(&pmd->root_lock);
1706 
1707 	return r;
1708 }
1709 
1710 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1711 {
1712 	int r = 0;
1713 
1714 	down_write(&pmd->root_lock);
1715 	for (; b != e; b++) {
1716 		r = dm_sm_dec_block(pmd->data_sm, b);
1717 		if (r)
1718 			break;
1719 	}
1720 	up_write(&pmd->root_lock);
1721 
1722 	return r;
1723 }
1724 
1725 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1726 {
1727 	int r;
1728 
1729 	down_read(&td->pmd->root_lock);
1730 	r = td->changed;
1731 	up_read(&td->pmd->root_lock);
1732 
1733 	return r;
1734 }
1735 
1736 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1737 {
1738 	bool r = false;
1739 	struct dm_thin_device *td, *tmp;
1740 
1741 	down_read(&pmd->root_lock);
1742 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1743 		if (td->changed) {
1744 			r = td->changed;
1745 			break;
1746 		}
1747 	}
1748 	up_read(&pmd->root_lock);
1749 
1750 	return r;
1751 }
1752 
1753 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1754 {
1755 	bool r;
1756 
1757 	down_read(&td->pmd->root_lock);
1758 	r = td->aborted_with_changes;
1759 	up_read(&td->pmd->root_lock);
1760 
1761 	return r;
1762 }
1763 
1764 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1765 {
1766 	int r = -EINVAL;
1767 
1768 	down_write(&pmd->root_lock);
1769 	if (!pmd->fail_io)
1770 		r = dm_sm_new_block(pmd->data_sm, result);
1771 	up_write(&pmd->root_lock);
1772 
1773 	return r;
1774 }
1775 
1776 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1777 {
1778 	int r = -EINVAL;
1779 
1780 	down_write(&pmd->root_lock);
1781 	if (pmd->fail_io)
1782 		goto out;
1783 
1784 	r = __commit_transaction(pmd);
1785 	if (r <= 0)
1786 		goto out;
1787 
1788 	/*
1789 	 * Open the next transaction.
1790 	 */
1791 	r = __begin_transaction(pmd);
1792 out:
1793 	up_write(&pmd->root_lock);
1794 	return r;
1795 }
1796 
1797 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1798 {
1799 	struct dm_thin_device *td;
1800 
1801 	list_for_each_entry(td, &pmd->thin_devices, list)
1802 		td->aborted_with_changes = td->changed;
1803 }
1804 
1805 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1806 {
1807 	int r = -EINVAL;
1808 
1809 	down_write(&pmd->root_lock);
1810 	if (pmd->fail_io)
1811 		goto out;
1812 
1813 	__set_abort_with_changes_flags(pmd);
1814 	__destroy_persistent_data_objects(pmd);
1815 	r = __create_persistent_data_objects(pmd, false);
1816 	if (r)
1817 		pmd->fail_io = true;
1818 
1819 out:
1820 	up_write(&pmd->root_lock);
1821 
1822 	return r;
1823 }
1824 
1825 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1826 {
1827 	int r = -EINVAL;
1828 
1829 	down_read(&pmd->root_lock);
1830 	if (!pmd->fail_io)
1831 		r = dm_sm_get_nr_free(pmd->data_sm, result);
1832 	up_read(&pmd->root_lock);
1833 
1834 	return r;
1835 }
1836 
1837 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1838 					  dm_block_t *result)
1839 {
1840 	int r = -EINVAL;
1841 
1842 	down_read(&pmd->root_lock);
1843 	if (!pmd->fail_io)
1844 		r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1845 
1846 	if (!r) {
1847 		if (*result < pmd->metadata_reserve)
1848 			*result = 0;
1849 		else
1850 			*result -= pmd->metadata_reserve;
1851 	}
1852 	up_read(&pmd->root_lock);
1853 
1854 	return r;
1855 }
1856 
1857 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1858 				  dm_block_t *result)
1859 {
1860 	int r = -EINVAL;
1861 
1862 	down_read(&pmd->root_lock);
1863 	if (!pmd->fail_io)
1864 		r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1865 	up_read(&pmd->root_lock);
1866 
1867 	return r;
1868 }
1869 
1870 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1871 {
1872 	int r = -EINVAL;
1873 
1874 	down_read(&pmd->root_lock);
1875 	if (!pmd->fail_io)
1876 		r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1877 	up_read(&pmd->root_lock);
1878 
1879 	return r;
1880 }
1881 
1882 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1883 {
1884 	int r = -EINVAL;
1885 	struct dm_pool_metadata *pmd = td->pmd;
1886 
1887 	down_read(&pmd->root_lock);
1888 	if (!pmd->fail_io) {
1889 		*result = td->mapped_blocks;
1890 		r = 0;
1891 	}
1892 	up_read(&pmd->root_lock);
1893 
1894 	return r;
1895 }
1896 
1897 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1898 {
1899 	int r;
1900 	__le64 value_le;
1901 	dm_block_t thin_root;
1902 	struct dm_pool_metadata *pmd = td->pmd;
1903 
1904 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1905 	if (r)
1906 		return r;
1907 
1908 	thin_root = le64_to_cpu(value_le);
1909 
1910 	return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1911 }
1912 
1913 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1914 				     dm_block_t *result)
1915 {
1916 	int r = -EINVAL;
1917 	struct dm_pool_metadata *pmd = td->pmd;
1918 
1919 	down_read(&pmd->root_lock);
1920 	if (!pmd->fail_io)
1921 		r = __highest_block(td, result);
1922 	up_read(&pmd->root_lock);
1923 
1924 	return r;
1925 }
1926 
1927 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1928 {
1929 	int r;
1930 	dm_block_t old_count;
1931 
1932 	r = dm_sm_get_nr_blocks(sm, &old_count);
1933 	if (r)
1934 		return r;
1935 
1936 	if (new_count == old_count)
1937 		return 0;
1938 
1939 	if (new_count < old_count) {
1940 		DMERR("cannot reduce size of space map");
1941 		return -EINVAL;
1942 	}
1943 
1944 	return dm_sm_extend(sm, new_count - old_count);
1945 }
1946 
1947 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1948 {
1949 	int r = -EINVAL;
1950 
1951 	down_write(&pmd->root_lock);
1952 	if (!pmd->fail_io)
1953 		r = __resize_space_map(pmd->data_sm, new_count);
1954 	up_write(&pmd->root_lock);
1955 
1956 	return r;
1957 }
1958 
1959 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1960 {
1961 	int r = -EINVAL;
1962 
1963 	down_write(&pmd->root_lock);
1964 	if (!pmd->fail_io) {
1965 		r = __resize_space_map(pmd->metadata_sm, new_count);
1966 		if (!r)
1967 			__set_metadata_reserve(pmd);
1968 	}
1969 	up_write(&pmd->root_lock);
1970 
1971 	return r;
1972 }
1973 
1974 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1975 {
1976 	down_write(&pmd->root_lock);
1977 	dm_bm_set_read_only(pmd->bm);
1978 	up_write(&pmd->root_lock);
1979 }
1980 
1981 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
1982 {
1983 	down_write(&pmd->root_lock);
1984 	dm_bm_set_read_write(pmd->bm);
1985 	up_write(&pmd->root_lock);
1986 }
1987 
1988 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
1989 					dm_block_t threshold,
1990 					dm_sm_threshold_fn fn,
1991 					void *context)
1992 {
1993 	int r;
1994 
1995 	down_write(&pmd->root_lock);
1996 	r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
1997 	up_write(&pmd->root_lock);
1998 
1999 	return r;
2000 }
2001 
2002 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2003 {
2004 	int r;
2005 	struct dm_block *sblock;
2006 	struct thin_disk_superblock *disk_super;
2007 
2008 	down_write(&pmd->root_lock);
2009 	pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2010 
2011 	r = superblock_lock(pmd, &sblock);
2012 	if (r) {
2013 		DMERR("couldn't read superblock");
2014 		goto out;
2015 	}
2016 
2017 	disk_super = dm_block_data(sblock);
2018 	disk_super->flags = cpu_to_le32(pmd->flags);
2019 
2020 	dm_bm_unlock(sblock);
2021 out:
2022 	up_write(&pmd->root_lock);
2023 	return r;
2024 }
2025 
2026 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2027 {
2028 	bool needs_check;
2029 
2030 	down_read(&pmd->root_lock);
2031 	needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2032 	up_read(&pmd->root_lock);
2033 
2034 	return needs_check;
2035 }
2036 
2037 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2038 {
2039 	down_read(&pmd->root_lock);
2040 	if (!pmd->fail_io)
2041 		dm_tm_issue_prefetches(pmd->tm);
2042 	up_read(&pmd->root_lock);
2043 }
2044