xref: /linux/drivers/md/dm-thin-metadata.c (revision e9f0878c4b2004ac19581274c1ae4c61ae3ca70e)
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 	 * Set if a transaction has to be aborted but the attempt to roll back
193 	 * to the previous (good) transaction failed.  The only pool metadata
194 	 * operation possible in this state is the closing of the device.
195 	 */
196 	bool fail_io:1;
197 
198 	/*
199 	 * Reading the space map roots can fail, so we read it into these
200 	 * buffers before the superblock is locked and updated.
201 	 */
202 	__u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
203 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
204 };
205 
206 struct dm_thin_device {
207 	struct list_head list;
208 	struct dm_pool_metadata *pmd;
209 	dm_thin_id id;
210 
211 	int open_count;
212 	bool changed:1;
213 	bool aborted_with_changes:1;
214 	uint64_t mapped_blocks;
215 	uint64_t transaction_id;
216 	uint32_t creation_time;
217 	uint32_t snapshotted_time;
218 };
219 
220 /*----------------------------------------------------------------
221  * superblock validator
222  *--------------------------------------------------------------*/
223 
224 #define SUPERBLOCK_CSUM_XOR 160774
225 
226 static void sb_prepare_for_write(struct dm_block_validator *v,
227 				 struct dm_block *b,
228 				 size_t block_size)
229 {
230 	struct thin_disk_superblock *disk_super = dm_block_data(b);
231 
232 	disk_super->blocknr = cpu_to_le64(dm_block_location(b));
233 	disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
234 						      block_size - sizeof(__le32),
235 						      SUPERBLOCK_CSUM_XOR));
236 }
237 
238 static int sb_check(struct dm_block_validator *v,
239 		    struct dm_block *b,
240 		    size_t block_size)
241 {
242 	struct thin_disk_superblock *disk_super = dm_block_data(b);
243 	__le32 csum_le;
244 
245 	if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
246 		DMERR("sb_check failed: blocknr %llu: "
247 		      "wanted %llu", le64_to_cpu(disk_super->blocknr),
248 		      (unsigned long long)dm_block_location(b));
249 		return -ENOTBLK;
250 	}
251 
252 	if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
253 		DMERR("sb_check failed: magic %llu: "
254 		      "wanted %llu", le64_to_cpu(disk_super->magic),
255 		      (unsigned long long)THIN_SUPERBLOCK_MAGIC);
256 		return -EILSEQ;
257 	}
258 
259 	csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
260 					     block_size - sizeof(__le32),
261 					     SUPERBLOCK_CSUM_XOR));
262 	if (csum_le != disk_super->csum) {
263 		DMERR("sb_check failed: csum %u: wanted %u",
264 		      le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
265 		return -EILSEQ;
266 	}
267 
268 	return 0;
269 }
270 
271 static struct dm_block_validator sb_validator = {
272 	.name = "superblock",
273 	.prepare_for_write = sb_prepare_for_write,
274 	.check = sb_check
275 };
276 
277 /*----------------------------------------------------------------
278  * Methods for the btree value types
279  *--------------------------------------------------------------*/
280 
281 static uint64_t pack_block_time(dm_block_t b, uint32_t t)
282 {
283 	return (b << 24) | t;
284 }
285 
286 static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
287 {
288 	*b = v >> 24;
289 	*t = v & ((1 << 24) - 1);
290 }
291 
292 static void data_block_inc(void *context, const void *value_le)
293 {
294 	struct dm_space_map *sm = context;
295 	__le64 v_le;
296 	uint64_t b;
297 	uint32_t t;
298 
299 	memcpy(&v_le, value_le, sizeof(v_le));
300 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
301 	dm_sm_inc_block(sm, b);
302 }
303 
304 static void data_block_dec(void *context, const void *value_le)
305 {
306 	struct dm_space_map *sm = context;
307 	__le64 v_le;
308 	uint64_t b;
309 	uint32_t t;
310 
311 	memcpy(&v_le, value_le, sizeof(v_le));
312 	unpack_block_time(le64_to_cpu(v_le), &b, &t);
313 	dm_sm_dec_block(sm, b);
314 }
315 
316 static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
317 {
318 	__le64 v1_le, v2_le;
319 	uint64_t b1, b2;
320 	uint32_t t;
321 
322 	memcpy(&v1_le, value1_le, sizeof(v1_le));
323 	memcpy(&v2_le, value2_le, sizeof(v2_le));
324 	unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
325 	unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
326 
327 	return b1 == b2;
328 }
329 
330 static void subtree_inc(void *context, const void *value)
331 {
332 	struct dm_btree_info *info = context;
333 	__le64 root_le;
334 	uint64_t root;
335 
336 	memcpy(&root_le, value, sizeof(root_le));
337 	root = le64_to_cpu(root_le);
338 	dm_tm_inc(info->tm, root);
339 }
340 
341 static void subtree_dec(void *context, const void *value)
342 {
343 	struct dm_btree_info *info = context;
344 	__le64 root_le;
345 	uint64_t root;
346 
347 	memcpy(&root_le, value, sizeof(root_le));
348 	root = le64_to_cpu(root_le);
349 	if (dm_btree_del(info, root))
350 		DMERR("btree delete failed");
351 }
352 
353 static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
354 {
355 	__le64 v1_le, v2_le;
356 	memcpy(&v1_le, value1_le, sizeof(v1_le));
357 	memcpy(&v2_le, value2_le, sizeof(v2_le));
358 
359 	return v1_le == v2_le;
360 }
361 
362 /*----------------------------------------------------------------*/
363 
364 static int superblock_lock_zero(struct dm_pool_metadata *pmd,
365 				struct dm_block **sblock)
366 {
367 	return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
368 				     &sb_validator, sblock);
369 }
370 
371 static int superblock_lock(struct dm_pool_metadata *pmd,
372 			   struct dm_block **sblock)
373 {
374 	return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
375 				&sb_validator, sblock);
376 }
377 
378 static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
379 {
380 	int r;
381 	unsigned i;
382 	struct dm_block *b;
383 	__le64 *data_le, zero = cpu_to_le64(0);
384 	unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
385 
386 	/*
387 	 * We can't use a validator here - it may be all zeroes.
388 	 */
389 	r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
390 	if (r)
391 		return r;
392 
393 	data_le = dm_block_data(b);
394 	*result = 1;
395 	for (i = 0; i < block_size; i++) {
396 		if (data_le[i] != zero) {
397 			*result = 0;
398 			break;
399 		}
400 	}
401 
402 	dm_bm_unlock(b);
403 
404 	return 0;
405 }
406 
407 static void __setup_btree_details(struct dm_pool_metadata *pmd)
408 {
409 	pmd->info.tm = pmd->tm;
410 	pmd->info.levels = 2;
411 	pmd->info.value_type.context = pmd->data_sm;
412 	pmd->info.value_type.size = sizeof(__le64);
413 	pmd->info.value_type.inc = data_block_inc;
414 	pmd->info.value_type.dec = data_block_dec;
415 	pmd->info.value_type.equal = data_block_equal;
416 
417 	memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
418 	pmd->nb_info.tm = pmd->nb_tm;
419 
420 	pmd->tl_info.tm = pmd->tm;
421 	pmd->tl_info.levels = 1;
422 	pmd->tl_info.value_type.context = &pmd->bl_info;
423 	pmd->tl_info.value_type.size = sizeof(__le64);
424 	pmd->tl_info.value_type.inc = subtree_inc;
425 	pmd->tl_info.value_type.dec = subtree_dec;
426 	pmd->tl_info.value_type.equal = subtree_equal;
427 
428 	pmd->bl_info.tm = pmd->tm;
429 	pmd->bl_info.levels = 1;
430 	pmd->bl_info.value_type.context = pmd->data_sm;
431 	pmd->bl_info.value_type.size = sizeof(__le64);
432 	pmd->bl_info.value_type.inc = data_block_inc;
433 	pmd->bl_info.value_type.dec = data_block_dec;
434 	pmd->bl_info.value_type.equal = data_block_equal;
435 
436 	pmd->details_info.tm = pmd->tm;
437 	pmd->details_info.levels = 1;
438 	pmd->details_info.value_type.context = NULL;
439 	pmd->details_info.value_type.size = sizeof(struct disk_device_details);
440 	pmd->details_info.value_type.inc = NULL;
441 	pmd->details_info.value_type.dec = NULL;
442 	pmd->details_info.value_type.equal = NULL;
443 }
444 
445 static int save_sm_roots(struct dm_pool_metadata *pmd)
446 {
447 	int r;
448 	size_t len;
449 
450 	r = dm_sm_root_size(pmd->metadata_sm, &len);
451 	if (r < 0)
452 		return r;
453 
454 	r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
455 	if (r < 0)
456 		return r;
457 
458 	r = dm_sm_root_size(pmd->data_sm, &len);
459 	if (r < 0)
460 		return r;
461 
462 	return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
463 }
464 
465 static void copy_sm_roots(struct dm_pool_metadata *pmd,
466 			  struct thin_disk_superblock *disk)
467 {
468 	memcpy(&disk->metadata_space_map_root,
469 	       &pmd->metadata_space_map_root,
470 	       sizeof(pmd->metadata_space_map_root));
471 
472 	memcpy(&disk->data_space_map_root,
473 	       &pmd->data_space_map_root,
474 	       sizeof(pmd->data_space_map_root));
475 }
476 
477 static int __write_initial_superblock(struct dm_pool_metadata *pmd)
478 {
479 	int r;
480 	struct dm_block *sblock;
481 	struct thin_disk_superblock *disk_super;
482 	sector_t bdev_size = i_size_read(pmd->bdev->bd_inode) >> SECTOR_SHIFT;
483 
484 	if (bdev_size > THIN_METADATA_MAX_SECTORS)
485 		bdev_size = THIN_METADATA_MAX_SECTORS;
486 
487 	r = dm_sm_commit(pmd->data_sm);
488 	if (r < 0)
489 		return r;
490 
491 	r = dm_tm_pre_commit(pmd->tm);
492 	if (r < 0)
493 		return r;
494 
495 	r = save_sm_roots(pmd);
496 	if (r < 0)
497 		return r;
498 
499 	r = superblock_lock_zero(pmd, &sblock);
500 	if (r)
501 		return r;
502 
503 	disk_super = dm_block_data(sblock);
504 	disk_super->flags = 0;
505 	memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
506 	disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
507 	disk_super->version = cpu_to_le32(THIN_VERSION);
508 	disk_super->time = 0;
509 	disk_super->trans_id = 0;
510 	disk_super->held_root = 0;
511 
512 	copy_sm_roots(pmd, disk_super);
513 
514 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
515 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
516 	disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
517 	disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
518 	disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
519 
520 	return dm_tm_commit(pmd->tm, sblock);
521 }
522 
523 static int __format_metadata(struct dm_pool_metadata *pmd)
524 {
525 	int r;
526 
527 	r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
528 				 &pmd->tm, &pmd->metadata_sm);
529 	if (r < 0) {
530 		DMERR("tm_create_with_sm failed");
531 		return r;
532 	}
533 
534 	pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
535 	if (IS_ERR(pmd->data_sm)) {
536 		DMERR("sm_disk_create failed");
537 		r = PTR_ERR(pmd->data_sm);
538 		goto bad_cleanup_tm;
539 	}
540 
541 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
542 	if (!pmd->nb_tm) {
543 		DMERR("could not create non-blocking clone tm");
544 		r = -ENOMEM;
545 		goto bad_cleanup_data_sm;
546 	}
547 
548 	__setup_btree_details(pmd);
549 
550 	r = dm_btree_empty(&pmd->info, &pmd->root);
551 	if (r < 0)
552 		goto bad_cleanup_nb_tm;
553 
554 	r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
555 	if (r < 0) {
556 		DMERR("couldn't create devices root");
557 		goto bad_cleanup_nb_tm;
558 	}
559 
560 	r = __write_initial_superblock(pmd);
561 	if (r)
562 		goto bad_cleanup_nb_tm;
563 
564 	return 0;
565 
566 bad_cleanup_nb_tm:
567 	dm_tm_destroy(pmd->nb_tm);
568 bad_cleanup_data_sm:
569 	dm_sm_destroy(pmd->data_sm);
570 bad_cleanup_tm:
571 	dm_tm_destroy(pmd->tm);
572 	dm_sm_destroy(pmd->metadata_sm);
573 
574 	return r;
575 }
576 
577 static int __check_incompat_features(struct thin_disk_superblock *disk_super,
578 				     struct dm_pool_metadata *pmd)
579 {
580 	uint32_t features;
581 
582 	features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
583 	if (features) {
584 		DMERR("could not access metadata due to unsupported optional features (%lx).",
585 		      (unsigned long)features);
586 		return -EINVAL;
587 	}
588 
589 	/*
590 	 * Check for read-only metadata to skip the following RDWR checks.
591 	 */
592 	if (get_disk_ro(pmd->bdev->bd_disk))
593 		return 0;
594 
595 	features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
596 	if (features) {
597 		DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
598 		      (unsigned long)features);
599 		return -EINVAL;
600 	}
601 
602 	return 0;
603 }
604 
605 static int __open_metadata(struct dm_pool_metadata *pmd)
606 {
607 	int r;
608 	struct dm_block *sblock;
609 	struct thin_disk_superblock *disk_super;
610 
611 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
612 			    &sb_validator, &sblock);
613 	if (r < 0) {
614 		DMERR("couldn't read superblock");
615 		return r;
616 	}
617 
618 	disk_super = dm_block_data(sblock);
619 
620 	/* Verify the data block size hasn't changed */
621 	if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
622 		DMERR("changing the data block size (from %u to %llu) is not supported",
623 		      le32_to_cpu(disk_super->data_block_size),
624 		      (unsigned long long)pmd->data_block_size);
625 		r = -EINVAL;
626 		goto bad_unlock_sblock;
627 	}
628 
629 	r = __check_incompat_features(disk_super, pmd);
630 	if (r < 0)
631 		goto bad_unlock_sblock;
632 
633 	r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
634 			       disk_super->metadata_space_map_root,
635 			       sizeof(disk_super->metadata_space_map_root),
636 			       &pmd->tm, &pmd->metadata_sm);
637 	if (r < 0) {
638 		DMERR("tm_open_with_sm failed");
639 		goto bad_unlock_sblock;
640 	}
641 
642 	pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
643 				       sizeof(disk_super->data_space_map_root));
644 	if (IS_ERR(pmd->data_sm)) {
645 		DMERR("sm_disk_open failed");
646 		r = PTR_ERR(pmd->data_sm);
647 		goto bad_cleanup_tm;
648 	}
649 
650 	pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
651 	if (!pmd->nb_tm) {
652 		DMERR("could not create non-blocking clone tm");
653 		r = -ENOMEM;
654 		goto bad_cleanup_data_sm;
655 	}
656 
657 	__setup_btree_details(pmd);
658 	dm_bm_unlock(sblock);
659 
660 	return 0;
661 
662 bad_cleanup_data_sm:
663 	dm_sm_destroy(pmd->data_sm);
664 bad_cleanup_tm:
665 	dm_tm_destroy(pmd->tm);
666 	dm_sm_destroy(pmd->metadata_sm);
667 bad_unlock_sblock:
668 	dm_bm_unlock(sblock);
669 
670 	return r;
671 }
672 
673 static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
674 {
675 	int r, unformatted;
676 
677 	r = __superblock_all_zeroes(pmd->bm, &unformatted);
678 	if (r)
679 		return r;
680 
681 	if (unformatted)
682 		return format_device ? __format_metadata(pmd) : -EPERM;
683 
684 	return __open_metadata(pmd);
685 }
686 
687 static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
688 {
689 	int r;
690 
691 	pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
692 					  THIN_MAX_CONCURRENT_LOCKS);
693 	if (IS_ERR(pmd->bm)) {
694 		DMERR("could not create block manager");
695 		return PTR_ERR(pmd->bm);
696 	}
697 
698 	r = __open_or_format_metadata(pmd, format_device);
699 	if (r)
700 		dm_block_manager_destroy(pmd->bm);
701 
702 	return r;
703 }
704 
705 static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd)
706 {
707 	dm_sm_destroy(pmd->data_sm);
708 	dm_sm_destroy(pmd->metadata_sm);
709 	dm_tm_destroy(pmd->nb_tm);
710 	dm_tm_destroy(pmd->tm);
711 	dm_block_manager_destroy(pmd->bm);
712 }
713 
714 static int __begin_transaction(struct dm_pool_metadata *pmd)
715 {
716 	int r;
717 	struct thin_disk_superblock *disk_super;
718 	struct dm_block *sblock;
719 
720 	/*
721 	 * We re-read the superblock every time.  Shouldn't need to do this
722 	 * really.
723 	 */
724 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
725 			    &sb_validator, &sblock);
726 	if (r)
727 		return r;
728 
729 	disk_super = dm_block_data(sblock);
730 	pmd->time = le32_to_cpu(disk_super->time);
731 	pmd->root = le64_to_cpu(disk_super->data_mapping_root);
732 	pmd->details_root = le64_to_cpu(disk_super->device_details_root);
733 	pmd->trans_id = le64_to_cpu(disk_super->trans_id);
734 	pmd->flags = le32_to_cpu(disk_super->flags);
735 	pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
736 
737 	dm_bm_unlock(sblock);
738 	return 0;
739 }
740 
741 static int __write_changed_details(struct dm_pool_metadata *pmd)
742 {
743 	int r;
744 	struct dm_thin_device *td, *tmp;
745 	struct disk_device_details details;
746 	uint64_t key;
747 
748 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
749 		if (!td->changed)
750 			continue;
751 
752 		key = td->id;
753 
754 		details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
755 		details.transaction_id = cpu_to_le64(td->transaction_id);
756 		details.creation_time = cpu_to_le32(td->creation_time);
757 		details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
758 		__dm_bless_for_disk(&details);
759 
760 		r = dm_btree_insert(&pmd->details_info, pmd->details_root,
761 				    &key, &details, &pmd->details_root);
762 		if (r)
763 			return r;
764 
765 		if (td->open_count)
766 			td->changed = 0;
767 		else {
768 			list_del(&td->list);
769 			kfree(td);
770 		}
771 	}
772 
773 	return 0;
774 }
775 
776 static int __commit_transaction(struct dm_pool_metadata *pmd)
777 {
778 	int r;
779 	struct thin_disk_superblock *disk_super;
780 	struct dm_block *sblock;
781 
782 	/*
783 	 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
784 	 */
785 	BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
786 
787 	r = __write_changed_details(pmd);
788 	if (r < 0)
789 		return r;
790 
791 	r = dm_sm_commit(pmd->data_sm);
792 	if (r < 0)
793 		return r;
794 
795 	r = dm_tm_pre_commit(pmd->tm);
796 	if (r < 0)
797 		return r;
798 
799 	r = save_sm_roots(pmd);
800 	if (r < 0)
801 		return r;
802 
803 	r = superblock_lock(pmd, &sblock);
804 	if (r)
805 		return r;
806 
807 	disk_super = dm_block_data(sblock);
808 	disk_super->time = cpu_to_le32(pmd->time);
809 	disk_super->data_mapping_root = cpu_to_le64(pmd->root);
810 	disk_super->device_details_root = cpu_to_le64(pmd->details_root);
811 	disk_super->trans_id = cpu_to_le64(pmd->trans_id);
812 	disk_super->flags = cpu_to_le32(pmd->flags);
813 
814 	copy_sm_roots(pmd, disk_super);
815 
816 	return dm_tm_commit(pmd->tm, sblock);
817 }
818 
819 struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
820 					       sector_t data_block_size,
821 					       bool format_device)
822 {
823 	int r;
824 	struct dm_pool_metadata *pmd;
825 
826 	pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
827 	if (!pmd) {
828 		DMERR("could not allocate metadata struct");
829 		return ERR_PTR(-ENOMEM);
830 	}
831 
832 	init_rwsem(&pmd->root_lock);
833 	pmd->time = 0;
834 	INIT_LIST_HEAD(&pmd->thin_devices);
835 	pmd->fail_io = false;
836 	pmd->bdev = bdev;
837 	pmd->data_block_size = data_block_size;
838 
839 	r = __create_persistent_data_objects(pmd, format_device);
840 	if (r) {
841 		kfree(pmd);
842 		return ERR_PTR(r);
843 	}
844 
845 	r = __begin_transaction(pmd);
846 	if (r < 0) {
847 		if (dm_pool_metadata_close(pmd) < 0)
848 			DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
849 		return ERR_PTR(r);
850 	}
851 
852 	return pmd;
853 }
854 
855 int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
856 {
857 	int r;
858 	unsigned open_devices = 0;
859 	struct dm_thin_device *td, *tmp;
860 
861 	down_read(&pmd->root_lock);
862 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
863 		if (td->open_count)
864 			open_devices++;
865 		else {
866 			list_del(&td->list);
867 			kfree(td);
868 		}
869 	}
870 	up_read(&pmd->root_lock);
871 
872 	if (open_devices) {
873 		DMERR("attempt to close pmd when %u device(s) are still open",
874 		       open_devices);
875 		return -EBUSY;
876 	}
877 
878 	if (!dm_bm_is_read_only(pmd->bm) && !pmd->fail_io) {
879 		r = __commit_transaction(pmd);
880 		if (r < 0)
881 			DMWARN("%s: __commit_transaction() failed, error = %d",
882 			       __func__, r);
883 	}
884 
885 	if (!pmd->fail_io)
886 		__destroy_persistent_data_objects(pmd);
887 
888 	kfree(pmd);
889 	return 0;
890 }
891 
892 /*
893  * __open_device: Returns @td corresponding to device with id @dev,
894  * creating it if @create is set and incrementing @td->open_count.
895  * On failure, @td is undefined.
896  */
897 static int __open_device(struct dm_pool_metadata *pmd,
898 			 dm_thin_id dev, int create,
899 			 struct dm_thin_device **td)
900 {
901 	int r, changed = 0;
902 	struct dm_thin_device *td2;
903 	uint64_t key = dev;
904 	struct disk_device_details details_le;
905 
906 	/*
907 	 * If the device is already open, return it.
908 	 */
909 	list_for_each_entry(td2, &pmd->thin_devices, list)
910 		if (td2->id == dev) {
911 			/*
912 			 * May not create an already-open device.
913 			 */
914 			if (create)
915 				return -EEXIST;
916 
917 			td2->open_count++;
918 			*td = td2;
919 			return 0;
920 		}
921 
922 	/*
923 	 * Check the device exists.
924 	 */
925 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
926 			    &key, &details_le);
927 	if (r) {
928 		if (r != -ENODATA || !create)
929 			return r;
930 
931 		/*
932 		 * Create new device.
933 		 */
934 		changed = 1;
935 		details_le.mapped_blocks = 0;
936 		details_le.transaction_id = cpu_to_le64(pmd->trans_id);
937 		details_le.creation_time = cpu_to_le32(pmd->time);
938 		details_le.snapshotted_time = cpu_to_le32(pmd->time);
939 	}
940 
941 	*td = kmalloc(sizeof(**td), GFP_NOIO);
942 	if (!*td)
943 		return -ENOMEM;
944 
945 	(*td)->pmd = pmd;
946 	(*td)->id = dev;
947 	(*td)->open_count = 1;
948 	(*td)->changed = changed;
949 	(*td)->aborted_with_changes = false;
950 	(*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
951 	(*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
952 	(*td)->creation_time = le32_to_cpu(details_le.creation_time);
953 	(*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
954 
955 	list_add(&(*td)->list, &pmd->thin_devices);
956 
957 	return 0;
958 }
959 
960 static void __close_device(struct dm_thin_device *td)
961 {
962 	--td->open_count;
963 }
964 
965 static int __create_thin(struct dm_pool_metadata *pmd,
966 			 dm_thin_id dev)
967 {
968 	int r;
969 	dm_block_t dev_root;
970 	uint64_t key = dev;
971 	struct disk_device_details details_le;
972 	struct dm_thin_device *td;
973 	__le64 value;
974 
975 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
976 			    &key, &details_le);
977 	if (!r)
978 		return -EEXIST;
979 
980 	/*
981 	 * Create an empty btree for the mappings.
982 	 */
983 	r = dm_btree_empty(&pmd->bl_info, &dev_root);
984 	if (r)
985 		return r;
986 
987 	/*
988 	 * Insert it into the main mapping tree.
989 	 */
990 	value = cpu_to_le64(dev_root);
991 	__dm_bless_for_disk(&value);
992 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
993 	if (r) {
994 		dm_btree_del(&pmd->bl_info, dev_root);
995 		return r;
996 	}
997 
998 	r = __open_device(pmd, dev, 1, &td);
999 	if (r) {
1000 		dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1001 		dm_btree_del(&pmd->bl_info, dev_root);
1002 		return r;
1003 	}
1004 	__close_device(td);
1005 
1006 	return r;
1007 }
1008 
1009 int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1010 {
1011 	int r = -EINVAL;
1012 
1013 	down_write(&pmd->root_lock);
1014 	if (!pmd->fail_io)
1015 		r = __create_thin(pmd, dev);
1016 	up_write(&pmd->root_lock);
1017 
1018 	return r;
1019 }
1020 
1021 static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1022 				  struct dm_thin_device *snap,
1023 				  dm_thin_id origin, uint32_t time)
1024 {
1025 	int r;
1026 	struct dm_thin_device *td;
1027 
1028 	r = __open_device(pmd, origin, 0, &td);
1029 	if (r)
1030 		return r;
1031 
1032 	td->changed = 1;
1033 	td->snapshotted_time = time;
1034 
1035 	snap->mapped_blocks = td->mapped_blocks;
1036 	snap->snapshotted_time = time;
1037 	__close_device(td);
1038 
1039 	return 0;
1040 }
1041 
1042 static int __create_snap(struct dm_pool_metadata *pmd,
1043 			 dm_thin_id dev, dm_thin_id origin)
1044 {
1045 	int r;
1046 	dm_block_t origin_root;
1047 	uint64_t key = origin, dev_key = dev;
1048 	struct dm_thin_device *td;
1049 	struct disk_device_details details_le;
1050 	__le64 value;
1051 
1052 	/* check this device is unused */
1053 	r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1054 			    &dev_key, &details_le);
1055 	if (!r)
1056 		return -EEXIST;
1057 
1058 	/* find the mapping tree for the origin */
1059 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1060 	if (r)
1061 		return r;
1062 	origin_root = le64_to_cpu(value);
1063 
1064 	/* clone the origin, an inc will do */
1065 	dm_tm_inc(pmd->tm, origin_root);
1066 
1067 	/* insert into the main mapping tree */
1068 	value = cpu_to_le64(origin_root);
1069 	__dm_bless_for_disk(&value);
1070 	key = dev;
1071 	r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1072 	if (r) {
1073 		dm_tm_dec(pmd->tm, origin_root);
1074 		return r;
1075 	}
1076 
1077 	pmd->time++;
1078 
1079 	r = __open_device(pmd, dev, 1, &td);
1080 	if (r)
1081 		goto bad;
1082 
1083 	r = __set_snapshot_details(pmd, td, origin, pmd->time);
1084 	__close_device(td);
1085 
1086 	if (r)
1087 		goto bad;
1088 
1089 	return 0;
1090 
1091 bad:
1092 	dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1093 	dm_btree_remove(&pmd->details_info, pmd->details_root,
1094 			&key, &pmd->details_root);
1095 	return r;
1096 }
1097 
1098 int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1099 				 dm_thin_id dev,
1100 				 dm_thin_id origin)
1101 {
1102 	int r = -EINVAL;
1103 
1104 	down_write(&pmd->root_lock);
1105 	if (!pmd->fail_io)
1106 		r = __create_snap(pmd, dev, origin);
1107 	up_write(&pmd->root_lock);
1108 
1109 	return r;
1110 }
1111 
1112 static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1113 {
1114 	int r;
1115 	uint64_t key = dev;
1116 	struct dm_thin_device *td;
1117 
1118 	/* TODO: failure should mark the transaction invalid */
1119 	r = __open_device(pmd, dev, 0, &td);
1120 	if (r)
1121 		return r;
1122 
1123 	if (td->open_count > 1) {
1124 		__close_device(td);
1125 		return -EBUSY;
1126 	}
1127 
1128 	list_del(&td->list);
1129 	kfree(td);
1130 	r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1131 			    &key, &pmd->details_root);
1132 	if (r)
1133 		return r;
1134 
1135 	r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1136 	if (r)
1137 		return r;
1138 
1139 	return 0;
1140 }
1141 
1142 int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1143 			       dm_thin_id dev)
1144 {
1145 	int r = -EINVAL;
1146 
1147 	down_write(&pmd->root_lock);
1148 	if (!pmd->fail_io)
1149 		r = __delete_device(pmd, dev);
1150 	up_write(&pmd->root_lock);
1151 
1152 	return r;
1153 }
1154 
1155 int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1156 					uint64_t current_id,
1157 					uint64_t new_id)
1158 {
1159 	int r = -EINVAL;
1160 
1161 	down_write(&pmd->root_lock);
1162 
1163 	if (pmd->fail_io)
1164 		goto out;
1165 
1166 	if (pmd->trans_id != current_id) {
1167 		DMERR("mismatched transaction id");
1168 		goto out;
1169 	}
1170 
1171 	pmd->trans_id = new_id;
1172 	r = 0;
1173 
1174 out:
1175 	up_write(&pmd->root_lock);
1176 
1177 	return r;
1178 }
1179 
1180 int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1181 					uint64_t *result)
1182 {
1183 	int r = -EINVAL;
1184 
1185 	down_read(&pmd->root_lock);
1186 	if (!pmd->fail_io) {
1187 		*result = pmd->trans_id;
1188 		r = 0;
1189 	}
1190 	up_read(&pmd->root_lock);
1191 
1192 	return r;
1193 }
1194 
1195 static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1196 {
1197 	int r, inc;
1198 	struct thin_disk_superblock *disk_super;
1199 	struct dm_block *copy, *sblock;
1200 	dm_block_t held_root;
1201 
1202 	/*
1203 	 * We commit to ensure the btree roots which we increment in a
1204 	 * moment are up to date.
1205 	 */
1206 	__commit_transaction(pmd);
1207 
1208 	/*
1209 	 * Copy the superblock.
1210 	 */
1211 	dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1212 	r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1213 			       &sb_validator, &copy, &inc);
1214 	if (r)
1215 		return r;
1216 
1217 	BUG_ON(!inc);
1218 
1219 	held_root = dm_block_location(copy);
1220 	disk_super = dm_block_data(copy);
1221 
1222 	if (le64_to_cpu(disk_super->held_root)) {
1223 		DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1224 
1225 		dm_tm_dec(pmd->tm, held_root);
1226 		dm_tm_unlock(pmd->tm, copy);
1227 		return -EBUSY;
1228 	}
1229 
1230 	/*
1231 	 * Wipe the spacemap since we're not publishing this.
1232 	 */
1233 	memset(&disk_super->data_space_map_root, 0,
1234 	       sizeof(disk_super->data_space_map_root));
1235 	memset(&disk_super->metadata_space_map_root, 0,
1236 	       sizeof(disk_super->metadata_space_map_root));
1237 
1238 	/*
1239 	 * Increment the data structures that need to be preserved.
1240 	 */
1241 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1242 	dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1243 	dm_tm_unlock(pmd->tm, copy);
1244 
1245 	/*
1246 	 * Write the held root into the superblock.
1247 	 */
1248 	r = superblock_lock(pmd, &sblock);
1249 	if (r) {
1250 		dm_tm_dec(pmd->tm, held_root);
1251 		return r;
1252 	}
1253 
1254 	disk_super = dm_block_data(sblock);
1255 	disk_super->held_root = cpu_to_le64(held_root);
1256 	dm_bm_unlock(sblock);
1257 	return 0;
1258 }
1259 
1260 int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1261 {
1262 	int r = -EINVAL;
1263 
1264 	down_write(&pmd->root_lock);
1265 	if (!pmd->fail_io)
1266 		r = __reserve_metadata_snap(pmd);
1267 	up_write(&pmd->root_lock);
1268 
1269 	return r;
1270 }
1271 
1272 static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1273 {
1274 	int r;
1275 	struct thin_disk_superblock *disk_super;
1276 	struct dm_block *sblock, *copy;
1277 	dm_block_t held_root;
1278 
1279 	r = superblock_lock(pmd, &sblock);
1280 	if (r)
1281 		return r;
1282 
1283 	disk_super = dm_block_data(sblock);
1284 	held_root = le64_to_cpu(disk_super->held_root);
1285 	disk_super->held_root = cpu_to_le64(0);
1286 
1287 	dm_bm_unlock(sblock);
1288 
1289 	if (!held_root) {
1290 		DMWARN("No pool metadata snapshot found: nothing to release.");
1291 		return -EINVAL;
1292 	}
1293 
1294 	r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, &copy);
1295 	if (r)
1296 		return r;
1297 
1298 	disk_super = dm_block_data(copy);
1299 	dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1300 	dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1301 	dm_sm_dec_block(pmd->metadata_sm, held_root);
1302 
1303 	dm_tm_unlock(pmd->tm, copy);
1304 
1305 	return 0;
1306 }
1307 
1308 int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1309 {
1310 	int r = -EINVAL;
1311 
1312 	down_write(&pmd->root_lock);
1313 	if (!pmd->fail_io)
1314 		r = __release_metadata_snap(pmd);
1315 	up_write(&pmd->root_lock);
1316 
1317 	return r;
1318 }
1319 
1320 static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1321 			       dm_block_t *result)
1322 {
1323 	int r;
1324 	struct thin_disk_superblock *disk_super;
1325 	struct dm_block *sblock;
1326 
1327 	r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1328 			    &sb_validator, &sblock);
1329 	if (r)
1330 		return r;
1331 
1332 	disk_super = dm_block_data(sblock);
1333 	*result = le64_to_cpu(disk_super->held_root);
1334 
1335 	dm_bm_unlock(sblock);
1336 
1337 	return 0;
1338 }
1339 
1340 int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1341 			      dm_block_t *result)
1342 {
1343 	int r = -EINVAL;
1344 
1345 	down_read(&pmd->root_lock);
1346 	if (!pmd->fail_io)
1347 		r = __get_metadata_snap(pmd, result);
1348 	up_read(&pmd->root_lock);
1349 
1350 	return r;
1351 }
1352 
1353 int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1354 			     struct dm_thin_device **td)
1355 {
1356 	int r = -EINVAL;
1357 
1358 	down_write(&pmd->root_lock);
1359 	if (!pmd->fail_io)
1360 		r = __open_device(pmd, dev, 0, td);
1361 	up_write(&pmd->root_lock);
1362 
1363 	return r;
1364 }
1365 
1366 int dm_pool_close_thin_device(struct dm_thin_device *td)
1367 {
1368 	down_write(&td->pmd->root_lock);
1369 	__close_device(td);
1370 	up_write(&td->pmd->root_lock);
1371 
1372 	return 0;
1373 }
1374 
1375 dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1376 {
1377 	return td->id;
1378 }
1379 
1380 /*
1381  * Check whether @time (of block creation) is older than @td's last snapshot.
1382  * If so then the associated block is shared with the last snapshot device.
1383  * Any block on a device created *after* the device last got snapshotted is
1384  * necessarily not shared.
1385  */
1386 static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1387 {
1388 	return td->snapshotted_time > time;
1389 }
1390 
1391 static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1392 				 struct dm_thin_lookup_result *result)
1393 {
1394 	uint64_t block_time = 0;
1395 	dm_block_t exception_block;
1396 	uint32_t exception_time;
1397 
1398 	block_time = le64_to_cpu(value);
1399 	unpack_block_time(block_time, &exception_block, &exception_time);
1400 	result->block = exception_block;
1401 	result->shared = __snapshotted_since(td, exception_time);
1402 }
1403 
1404 static int __find_block(struct dm_thin_device *td, dm_block_t block,
1405 			int can_issue_io, struct dm_thin_lookup_result *result)
1406 {
1407 	int r;
1408 	__le64 value;
1409 	struct dm_pool_metadata *pmd = td->pmd;
1410 	dm_block_t keys[2] = { td->id, block };
1411 	struct dm_btree_info *info;
1412 
1413 	if (can_issue_io) {
1414 		info = &pmd->info;
1415 	} else
1416 		info = &pmd->nb_info;
1417 
1418 	r = dm_btree_lookup(info, pmd->root, keys, &value);
1419 	if (!r)
1420 		unpack_lookup_result(td, value, result);
1421 
1422 	return r;
1423 }
1424 
1425 int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1426 		       int can_issue_io, struct dm_thin_lookup_result *result)
1427 {
1428 	int r;
1429 	struct dm_pool_metadata *pmd = td->pmd;
1430 
1431 	down_read(&pmd->root_lock);
1432 	if (pmd->fail_io) {
1433 		up_read(&pmd->root_lock);
1434 		return -EINVAL;
1435 	}
1436 
1437 	r = __find_block(td, block, can_issue_io, result);
1438 
1439 	up_read(&pmd->root_lock);
1440 	return r;
1441 }
1442 
1443 static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1444 					  dm_block_t *vblock,
1445 					  struct dm_thin_lookup_result *result)
1446 {
1447 	int r;
1448 	__le64 value;
1449 	struct dm_pool_metadata *pmd = td->pmd;
1450 	dm_block_t keys[2] = { td->id, block };
1451 
1452 	r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1453 	if (!r)
1454 		unpack_lookup_result(td, value, result);
1455 
1456 	return r;
1457 }
1458 
1459 static int __find_mapped_range(struct dm_thin_device *td,
1460 			       dm_block_t begin, dm_block_t end,
1461 			       dm_block_t *thin_begin, dm_block_t *thin_end,
1462 			       dm_block_t *pool_begin, bool *maybe_shared)
1463 {
1464 	int r;
1465 	dm_block_t pool_end;
1466 	struct dm_thin_lookup_result lookup;
1467 
1468 	if (end < begin)
1469 		return -ENODATA;
1470 
1471 	r = __find_next_mapped_block(td, begin, &begin, &lookup);
1472 	if (r)
1473 		return r;
1474 
1475 	if (begin >= end)
1476 		return -ENODATA;
1477 
1478 	*thin_begin = begin;
1479 	*pool_begin = lookup.block;
1480 	*maybe_shared = lookup.shared;
1481 
1482 	begin++;
1483 	pool_end = *pool_begin + 1;
1484 	while (begin != end) {
1485 		r = __find_block(td, begin, true, &lookup);
1486 		if (r) {
1487 			if (r == -ENODATA)
1488 				break;
1489 			else
1490 				return r;
1491 		}
1492 
1493 		if ((lookup.block != pool_end) ||
1494 		    (lookup.shared != *maybe_shared))
1495 			break;
1496 
1497 		pool_end++;
1498 		begin++;
1499 	}
1500 
1501 	*thin_end = begin;
1502 	return 0;
1503 }
1504 
1505 int dm_thin_find_mapped_range(struct dm_thin_device *td,
1506 			      dm_block_t begin, dm_block_t end,
1507 			      dm_block_t *thin_begin, dm_block_t *thin_end,
1508 			      dm_block_t *pool_begin, bool *maybe_shared)
1509 {
1510 	int r = -EINVAL;
1511 	struct dm_pool_metadata *pmd = td->pmd;
1512 
1513 	down_read(&pmd->root_lock);
1514 	if (!pmd->fail_io) {
1515 		r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1516 					pool_begin, maybe_shared);
1517 	}
1518 	up_read(&pmd->root_lock);
1519 
1520 	return r;
1521 }
1522 
1523 static int __insert(struct dm_thin_device *td, dm_block_t block,
1524 		    dm_block_t data_block)
1525 {
1526 	int r, inserted;
1527 	__le64 value;
1528 	struct dm_pool_metadata *pmd = td->pmd;
1529 	dm_block_t keys[2] = { td->id, block };
1530 
1531 	value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1532 	__dm_bless_for_disk(&value);
1533 
1534 	r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1535 				   &pmd->root, &inserted);
1536 	if (r)
1537 		return r;
1538 
1539 	td->changed = 1;
1540 	if (inserted)
1541 		td->mapped_blocks++;
1542 
1543 	return 0;
1544 }
1545 
1546 int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1547 			 dm_block_t data_block)
1548 {
1549 	int r = -EINVAL;
1550 
1551 	down_write(&td->pmd->root_lock);
1552 	if (!td->pmd->fail_io)
1553 		r = __insert(td, block, data_block);
1554 	up_write(&td->pmd->root_lock);
1555 
1556 	return r;
1557 }
1558 
1559 static int __remove(struct dm_thin_device *td, dm_block_t block)
1560 {
1561 	int r;
1562 	struct dm_pool_metadata *pmd = td->pmd;
1563 	dm_block_t keys[2] = { td->id, block };
1564 
1565 	r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1566 	if (r)
1567 		return r;
1568 
1569 	td->mapped_blocks--;
1570 	td->changed = 1;
1571 
1572 	return 0;
1573 }
1574 
1575 static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1576 {
1577 	int r;
1578 	unsigned count, total_count = 0;
1579 	struct dm_pool_metadata *pmd = td->pmd;
1580 	dm_block_t keys[1] = { td->id };
1581 	__le64 value;
1582 	dm_block_t mapping_root;
1583 
1584 	/*
1585 	 * Find the mapping tree
1586 	 */
1587 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1588 	if (r)
1589 		return r;
1590 
1591 	/*
1592 	 * Remove from the mapping tree, taking care to inc the
1593 	 * ref count so it doesn't get deleted.
1594 	 */
1595 	mapping_root = le64_to_cpu(value);
1596 	dm_tm_inc(pmd->tm, mapping_root);
1597 	r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1598 	if (r)
1599 		return r;
1600 
1601 	/*
1602 	 * Remove leaves stops at the first unmapped entry, so we have to
1603 	 * loop round finding mapped ranges.
1604 	 */
1605 	while (begin < end) {
1606 		r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1607 		if (r == -ENODATA)
1608 			break;
1609 
1610 		if (r)
1611 			return r;
1612 
1613 		if (begin >= end)
1614 			break;
1615 
1616 		r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1617 		if (r)
1618 			return r;
1619 
1620 		total_count += count;
1621 	}
1622 
1623 	td->mapped_blocks -= total_count;
1624 	td->changed = 1;
1625 
1626 	/*
1627 	 * Reinsert the mapping tree.
1628 	 */
1629 	value = cpu_to_le64(mapping_root);
1630 	__dm_bless_for_disk(&value);
1631 	return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1632 }
1633 
1634 int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1635 {
1636 	int r = -EINVAL;
1637 
1638 	down_write(&td->pmd->root_lock);
1639 	if (!td->pmd->fail_io)
1640 		r = __remove(td, block);
1641 	up_write(&td->pmd->root_lock);
1642 
1643 	return r;
1644 }
1645 
1646 int dm_thin_remove_range(struct dm_thin_device *td,
1647 			 dm_block_t begin, dm_block_t end)
1648 {
1649 	int r = -EINVAL;
1650 
1651 	down_write(&td->pmd->root_lock);
1652 	if (!td->pmd->fail_io)
1653 		r = __remove_range(td, begin, end);
1654 	up_write(&td->pmd->root_lock);
1655 
1656 	return r;
1657 }
1658 
1659 int dm_pool_block_is_used(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1660 {
1661 	int r;
1662 	uint32_t ref_count;
1663 
1664 	down_read(&pmd->root_lock);
1665 	r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1666 	if (!r)
1667 		*result = (ref_count != 0);
1668 	up_read(&pmd->root_lock);
1669 
1670 	return r;
1671 }
1672 
1673 int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1674 {
1675 	int r = 0;
1676 
1677 	down_write(&pmd->root_lock);
1678 	for (; b != e; b++) {
1679 		r = dm_sm_inc_block(pmd->data_sm, b);
1680 		if (r)
1681 			break;
1682 	}
1683 	up_write(&pmd->root_lock);
1684 
1685 	return r;
1686 }
1687 
1688 int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1689 {
1690 	int r = 0;
1691 
1692 	down_write(&pmd->root_lock);
1693 	for (; b != e; b++) {
1694 		r = dm_sm_dec_block(pmd->data_sm, b);
1695 		if (r)
1696 			break;
1697 	}
1698 	up_write(&pmd->root_lock);
1699 
1700 	return r;
1701 }
1702 
1703 bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1704 {
1705 	int r;
1706 
1707 	down_read(&td->pmd->root_lock);
1708 	r = td->changed;
1709 	up_read(&td->pmd->root_lock);
1710 
1711 	return r;
1712 }
1713 
1714 bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1715 {
1716 	bool r = false;
1717 	struct dm_thin_device *td, *tmp;
1718 
1719 	down_read(&pmd->root_lock);
1720 	list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1721 		if (td->changed) {
1722 			r = td->changed;
1723 			break;
1724 		}
1725 	}
1726 	up_read(&pmd->root_lock);
1727 
1728 	return r;
1729 }
1730 
1731 bool dm_thin_aborted_changes(struct dm_thin_device *td)
1732 {
1733 	bool r;
1734 
1735 	down_read(&td->pmd->root_lock);
1736 	r = td->aborted_with_changes;
1737 	up_read(&td->pmd->root_lock);
1738 
1739 	return r;
1740 }
1741 
1742 int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1743 {
1744 	int r = -EINVAL;
1745 
1746 	down_write(&pmd->root_lock);
1747 	if (!pmd->fail_io)
1748 		r = dm_sm_new_block(pmd->data_sm, result);
1749 	up_write(&pmd->root_lock);
1750 
1751 	return r;
1752 }
1753 
1754 int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1755 {
1756 	int r = -EINVAL;
1757 
1758 	down_write(&pmd->root_lock);
1759 	if (pmd->fail_io)
1760 		goto out;
1761 
1762 	r = __commit_transaction(pmd);
1763 	if (r <= 0)
1764 		goto out;
1765 
1766 	/*
1767 	 * Open the next transaction.
1768 	 */
1769 	r = __begin_transaction(pmd);
1770 out:
1771 	up_write(&pmd->root_lock);
1772 	return r;
1773 }
1774 
1775 static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1776 {
1777 	struct dm_thin_device *td;
1778 
1779 	list_for_each_entry(td, &pmd->thin_devices, list)
1780 		td->aborted_with_changes = td->changed;
1781 }
1782 
1783 int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1784 {
1785 	int r = -EINVAL;
1786 
1787 	down_write(&pmd->root_lock);
1788 	if (pmd->fail_io)
1789 		goto out;
1790 
1791 	__set_abort_with_changes_flags(pmd);
1792 	__destroy_persistent_data_objects(pmd);
1793 	r = __create_persistent_data_objects(pmd, false);
1794 	if (r)
1795 		pmd->fail_io = true;
1796 
1797 out:
1798 	up_write(&pmd->root_lock);
1799 
1800 	return r;
1801 }
1802 
1803 int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1804 {
1805 	int r = -EINVAL;
1806 
1807 	down_read(&pmd->root_lock);
1808 	if (!pmd->fail_io)
1809 		r = dm_sm_get_nr_free(pmd->data_sm, result);
1810 	up_read(&pmd->root_lock);
1811 
1812 	return r;
1813 }
1814 
1815 int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1816 					  dm_block_t *result)
1817 {
1818 	int r = -EINVAL;
1819 
1820 	down_read(&pmd->root_lock);
1821 	if (!pmd->fail_io)
1822 		r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1823 	up_read(&pmd->root_lock);
1824 
1825 	return r;
1826 }
1827 
1828 int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1829 				  dm_block_t *result)
1830 {
1831 	int r = -EINVAL;
1832 
1833 	down_read(&pmd->root_lock);
1834 	if (!pmd->fail_io)
1835 		r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1836 	up_read(&pmd->root_lock);
1837 
1838 	return r;
1839 }
1840 
1841 int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1842 {
1843 	int r = -EINVAL;
1844 
1845 	down_read(&pmd->root_lock);
1846 	if (!pmd->fail_io)
1847 		r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1848 	up_read(&pmd->root_lock);
1849 
1850 	return r;
1851 }
1852 
1853 int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1854 {
1855 	int r = -EINVAL;
1856 	struct dm_pool_metadata *pmd = td->pmd;
1857 
1858 	down_read(&pmd->root_lock);
1859 	if (!pmd->fail_io) {
1860 		*result = td->mapped_blocks;
1861 		r = 0;
1862 	}
1863 	up_read(&pmd->root_lock);
1864 
1865 	return r;
1866 }
1867 
1868 static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1869 {
1870 	int r;
1871 	__le64 value_le;
1872 	dm_block_t thin_root;
1873 	struct dm_pool_metadata *pmd = td->pmd;
1874 
1875 	r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1876 	if (r)
1877 		return r;
1878 
1879 	thin_root = le64_to_cpu(value_le);
1880 
1881 	return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1882 }
1883 
1884 int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1885 				     dm_block_t *result)
1886 {
1887 	int r = -EINVAL;
1888 	struct dm_pool_metadata *pmd = td->pmd;
1889 
1890 	down_read(&pmd->root_lock);
1891 	if (!pmd->fail_io)
1892 		r = __highest_block(td, result);
1893 	up_read(&pmd->root_lock);
1894 
1895 	return r;
1896 }
1897 
1898 static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
1899 {
1900 	int r;
1901 	dm_block_t old_count;
1902 
1903 	r = dm_sm_get_nr_blocks(sm, &old_count);
1904 	if (r)
1905 		return r;
1906 
1907 	if (new_count == old_count)
1908 		return 0;
1909 
1910 	if (new_count < old_count) {
1911 		DMERR("cannot reduce size of space map");
1912 		return -EINVAL;
1913 	}
1914 
1915 	return dm_sm_extend(sm, new_count - old_count);
1916 }
1917 
1918 int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1919 {
1920 	int r = -EINVAL;
1921 
1922 	down_write(&pmd->root_lock);
1923 	if (!pmd->fail_io)
1924 		r = __resize_space_map(pmd->data_sm, new_count);
1925 	up_write(&pmd->root_lock);
1926 
1927 	return r;
1928 }
1929 
1930 int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1931 {
1932 	int r = -EINVAL;
1933 
1934 	down_write(&pmd->root_lock);
1935 	if (!pmd->fail_io)
1936 		r = __resize_space_map(pmd->metadata_sm, new_count);
1937 	up_write(&pmd->root_lock);
1938 
1939 	return r;
1940 }
1941 
1942 void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
1943 {
1944 	down_write(&pmd->root_lock);
1945 	dm_bm_set_read_only(pmd->bm);
1946 	up_write(&pmd->root_lock);
1947 }
1948 
1949 void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
1950 {
1951 	down_write(&pmd->root_lock);
1952 	dm_bm_set_read_write(pmd->bm);
1953 	up_write(&pmd->root_lock);
1954 }
1955 
1956 int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
1957 					dm_block_t threshold,
1958 					dm_sm_threshold_fn fn,
1959 					void *context)
1960 {
1961 	int r;
1962 
1963 	down_write(&pmd->root_lock);
1964 	r = dm_sm_register_threshold_callback(pmd->metadata_sm, threshold, fn, context);
1965 	up_write(&pmd->root_lock);
1966 
1967 	return r;
1968 }
1969 
1970 int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
1971 {
1972 	int r;
1973 	struct dm_block *sblock;
1974 	struct thin_disk_superblock *disk_super;
1975 
1976 	down_write(&pmd->root_lock);
1977 	pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
1978 
1979 	r = superblock_lock(pmd, &sblock);
1980 	if (r) {
1981 		DMERR("couldn't read superblock");
1982 		goto out;
1983 	}
1984 
1985 	disk_super = dm_block_data(sblock);
1986 	disk_super->flags = cpu_to_le32(pmd->flags);
1987 
1988 	dm_bm_unlock(sblock);
1989 out:
1990 	up_write(&pmd->root_lock);
1991 	return r;
1992 }
1993 
1994 bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
1995 {
1996 	bool needs_check;
1997 
1998 	down_read(&pmd->root_lock);
1999 	needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2000 	up_read(&pmd->root_lock);
2001 
2002 	return needs_check;
2003 }
2004 
2005 void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2006 {
2007 	down_read(&pmd->root_lock);
2008 	if (!pmd->fail_io)
2009 		dm_tm_issue_prefetches(pmd->tm);
2010 	up_read(&pmd->root_lock);
2011 }
2012