Loading...
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 40
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 */
98struct 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
138struct disk_device_details {
139 __le64 mapped_blocks;
140 __le64 transaction_id; /* When created. */
141 __le32 creation_time;
142 __le32 snapshotted_time;
143} __packed;
144
145struct 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 * Pre-commit callback.
193 *
194 * This allows the thin provisioning target to run a callback before
195 * the metadata are committed.
196 */
197 dm_pool_pre_commit_fn pre_commit_fn;
198 void *pre_commit_context;
199
200 /*
201 * We reserve a section of the metadata for commit overhead.
202 * All reported space does *not* include this.
203 */
204 dm_block_t metadata_reserve;
205
206 /*
207 * Set if a transaction has to be aborted but the attempt to roll back
208 * to the previous (good) transaction failed. The only pool metadata
209 * operation possible in this state is the closing of the device.
210 */
211 bool fail_io:1;
212
213 /*
214 * Set once a thin-pool has been accessed through one of the interfaces
215 * that imply the pool is in-service (e.g. thin devices created/deleted,
216 * thin-pool message, metadata snapshots, etc).
217 */
218 bool in_service:1;
219
220 /*
221 * Reading the space map roots can fail, so we read it into these
222 * buffers before the superblock is locked and updated.
223 */
224 __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
225 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
226};
227
228struct dm_thin_device {
229 struct list_head list;
230 struct dm_pool_metadata *pmd;
231 dm_thin_id id;
232
233 int open_count;
234 bool changed:1;
235 bool aborted_with_changes:1;
236 uint64_t mapped_blocks;
237 uint64_t transaction_id;
238 uint32_t creation_time;
239 uint32_t snapshotted_time;
240};
241
242/*----------------------------------------------------------------
243 * superblock validator
244 *--------------------------------------------------------------*/
245
246#define SUPERBLOCK_CSUM_XOR 160774
247
248static void sb_prepare_for_write(struct dm_block_validator *v,
249 struct dm_block *b,
250 size_t block_size)
251{
252 struct thin_disk_superblock *disk_super = dm_block_data(b);
253
254 disk_super->blocknr = cpu_to_le64(dm_block_location(b));
255 disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
256 block_size - sizeof(__le32),
257 SUPERBLOCK_CSUM_XOR));
258}
259
260static int sb_check(struct dm_block_validator *v,
261 struct dm_block *b,
262 size_t block_size)
263{
264 struct thin_disk_superblock *disk_super = dm_block_data(b);
265 __le32 csum_le;
266
267 if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
268 DMERR("sb_check failed: blocknr %llu: "
269 "wanted %llu", le64_to_cpu(disk_super->blocknr),
270 (unsigned long long)dm_block_location(b));
271 return -ENOTBLK;
272 }
273
274 if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
275 DMERR("sb_check failed: magic %llu: "
276 "wanted %llu", le64_to_cpu(disk_super->magic),
277 (unsigned long long)THIN_SUPERBLOCK_MAGIC);
278 return -EILSEQ;
279 }
280
281 csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
282 block_size - sizeof(__le32),
283 SUPERBLOCK_CSUM_XOR));
284 if (csum_le != disk_super->csum) {
285 DMERR("sb_check failed: csum %u: wanted %u",
286 le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
287 return -EILSEQ;
288 }
289
290 return 0;
291}
292
293static struct dm_block_validator sb_validator = {
294 .name = "superblock",
295 .prepare_for_write = sb_prepare_for_write,
296 .check = sb_check
297};
298
299/*----------------------------------------------------------------
300 * Methods for the btree value types
301 *--------------------------------------------------------------*/
302
303static uint64_t pack_block_time(dm_block_t b, uint32_t t)
304{
305 return (b << 24) | t;
306}
307
308static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
309{
310 *b = v >> 24;
311 *t = v & ((1 << 24) - 1);
312}
313
314/*
315 * It's more efficient to call dm_sm_{inc,dec}_blocks as few times as
316 * possible. 'with_runs' reads contiguous runs of blocks, and calls the
317 * given sm function.
318 */
319typedef int (*run_fn)(struct dm_space_map *, dm_block_t, dm_block_t);
320
321static void with_runs(struct dm_space_map *sm, const __le64 *value_le, unsigned count, run_fn fn)
322{
323 uint64_t b, begin, end;
324 uint32_t t;
325 bool in_run = false;
326 unsigned i;
327
328 for (i = 0; i < count; i++, value_le++) {
329 /* We know value_le is 8 byte aligned */
330 unpack_block_time(le64_to_cpu(*value_le), &b, &t);
331
332 if (in_run) {
333 if (b == end) {
334 end++;
335 } else {
336 fn(sm, begin, end);
337 begin = b;
338 end = b + 1;
339 }
340 } else {
341 in_run = true;
342 begin = b;
343 end = b + 1;
344 }
345 }
346
347 if (in_run)
348 fn(sm, begin, end);
349}
350
351static void data_block_inc(void *context, const void *value_le, unsigned count)
352{
353 with_runs((struct dm_space_map *) context,
354 (const __le64 *) value_le, count, dm_sm_inc_blocks);
355}
356
357static void data_block_dec(void *context, const void *value_le, unsigned count)
358{
359 with_runs((struct dm_space_map *) context,
360 (const __le64 *) value_le, count, dm_sm_dec_blocks);
361}
362
363static int data_block_equal(void *context, const void *value1_le, const void *value2_le)
364{
365 __le64 v1_le, v2_le;
366 uint64_t b1, b2;
367 uint32_t t;
368
369 memcpy(&v1_le, value1_le, sizeof(v1_le));
370 memcpy(&v2_le, value2_le, sizeof(v2_le));
371 unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
372 unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
373
374 return b1 == b2;
375}
376
377static void subtree_inc(void *context, const void *value, unsigned count)
378{
379 struct dm_btree_info *info = context;
380 const __le64 *root_le = value;
381 unsigned i;
382
383 for (i = 0; i < count; i++, root_le++)
384 dm_tm_inc(info->tm, le64_to_cpu(*root_le));
385}
386
387static void subtree_dec(void *context, const void *value, unsigned count)
388{
389 struct dm_btree_info *info = context;
390 const __le64 *root_le = value;
391 unsigned i;
392
393 for (i = 0; i < count; i++, root_le++)
394 if (dm_btree_del(info, le64_to_cpu(*root_le)))
395 DMERR("btree delete failed");
396}
397
398static int subtree_equal(void *context, const void *value1_le, const void *value2_le)
399{
400 __le64 v1_le, v2_le;
401 memcpy(&v1_le, value1_le, sizeof(v1_le));
402 memcpy(&v2_le, value2_le, sizeof(v2_le));
403
404 return v1_le == v2_le;
405}
406
407/*----------------------------------------------------------------*/
408
409/*
410 * Variant that is used for in-core only changes or code that
411 * shouldn't put the pool in service on its own (e.g. commit).
412 */
413static inline void pmd_write_lock_in_core(struct dm_pool_metadata *pmd)
414 __acquires(pmd->root_lock)
415{
416 down_write(&pmd->root_lock);
417}
418
419static inline void pmd_write_lock(struct dm_pool_metadata *pmd)
420{
421 pmd_write_lock_in_core(pmd);
422 if (unlikely(!pmd->in_service))
423 pmd->in_service = true;
424}
425
426static inline void pmd_write_unlock(struct dm_pool_metadata *pmd)
427 __releases(pmd->root_lock)
428{
429 up_write(&pmd->root_lock);
430}
431
432/*----------------------------------------------------------------*/
433
434static int superblock_lock_zero(struct dm_pool_metadata *pmd,
435 struct dm_block **sblock)
436{
437 return dm_bm_write_lock_zero(pmd->bm, THIN_SUPERBLOCK_LOCATION,
438 &sb_validator, sblock);
439}
440
441static int superblock_lock(struct dm_pool_metadata *pmd,
442 struct dm_block **sblock)
443{
444 return dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
445 &sb_validator, sblock);
446}
447
448static int __superblock_all_zeroes(struct dm_block_manager *bm, int *result)
449{
450 int r;
451 unsigned i;
452 struct dm_block *b;
453 __le64 *data_le, zero = cpu_to_le64(0);
454 unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
455
456 /*
457 * We can't use a validator here - it may be all zeroes.
458 */
459 r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
460 if (r)
461 return r;
462
463 data_le = dm_block_data(b);
464 *result = 1;
465 for (i = 0; i < block_size; i++) {
466 if (data_le[i] != zero) {
467 *result = 0;
468 break;
469 }
470 }
471
472 dm_bm_unlock(b);
473
474 return 0;
475}
476
477static void __setup_btree_details(struct dm_pool_metadata *pmd)
478{
479 pmd->info.tm = pmd->tm;
480 pmd->info.levels = 2;
481 pmd->info.value_type.context = pmd->data_sm;
482 pmd->info.value_type.size = sizeof(__le64);
483 pmd->info.value_type.inc = data_block_inc;
484 pmd->info.value_type.dec = data_block_dec;
485 pmd->info.value_type.equal = data_block_equal;
486
487 memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
488 pmd->nb_info.tm = pmd->nb_tm;
489
490 pmd->tl_info.tm = pmd->tm;
491 pmd->tl_info.levels = 1;
492 pmd->tl_info.value_type.context = &pmd->bl_info;
493 pmd->tl_info.value_type.size = sizeof(__le64);
494 pmd->tl_info.value_type.inc = subtree_inc;
495 pmd->tl_info.value_type.dec = subtree_dec;
496 pmd->tl_info.value_type.equal = subtree_equal;
497
498 pmd->bl_info.tm = pmd->tm;
499 pmd->bl_info.levels = 1;
500 pmd->bl_info.value_type.context = pmd->data_sm;
501 pmd->bl_info.value_type.size = sizeof(__le64);
502 pmd->bl_info.value_type.inc = data_block_inc;
503 pmd->bl_info.value_type.dec = data_block_dec;
504 pmd->bl_info.value_type.equal = data_block_equal;
505
506 pmd->details_info.tm = pmd->tm;
507 pmd->details_info.levels = 1;
508 pmd->details_info.value_type.context = NULL;
509 pmd->details_info.value_type.size = sizeof(struct disk_device_details);
510 pmd->details_info.value_type.inc = NULL;
511 pmd->details_info.value_type.dec = NULL;
512 pmd->details_info.value_type.equal = NULL;
513}
514
515static int save_sm_roots(struct dm_pool_metadata *pmd)
516{
517 int r;
518 size_t len;
519
520 r = dm_sm_root_size(pmd->metadata_sm, &len);
521 if (r < 0)
522 return r;
523
524 r = dm_sm_copy_root(pmd->metadata_sm, &pmd->metadata_space_map_root, len);
525 if (r < 0)
526 return r;
527
528 r = dm_sm_root_size(pmd->data_sm, &len);
529 if (r < 0)
530 return r;
531
532 return dm_sm_copy_root(pmd->data_sm, &pmd->data_space_map_root, len);
533}
534
535static void copy_sm_roots(struct dm_pool_metadata *pmd,
536 struct thin_disk_superblock *disk)
537{
538 memcpy(&disk->metadata_space_map_root,
539 &pmd->metadata_space_map_root,
540 sizeof(pmd->metadata_space_map_root));
541
542 memcpy(&disk->data_space_map_root,
543 &pmd->data_space_map_root,
544 sizeof(pmd->data_space_map_root));
545}
546
547static int __write_initial_superblock(struct dm_pool_metadata *pmd)
548{
549 int r;
550 struct dm_block *sblock;
551 struct thin_disk_superblock *disk_super;
552 sector_t bdev_size = bdev_nr_sectors(pmd->bdev);
553
554 if (bdev_size > THIN_METADATA_MAX_SECTORS)
555 bdev_size = THIN_METADATA_MAX_SECTORS;
556
557 r = dm_sm_commit(pmd->data_sm);
558 if (r < 0)
559 return r;
560
561 r = dm_tm_pre_commit(pmd->tm);
562 if (r < 0)
563 return r;
564
565 r = save_sm_roots(pmd);
566 if (r < 0)
567 return r;
568
569 r = superblock_lock_zero(pmd, &sblock);
570 if (r)
571 return r;
572
573 disk_super = dm_block_data(sblock);
574 disk_super->flags = 0;
575 memset(disk_super->uuid, 0, sizeof(disk_super->uuid));
576 disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
577 disk_super->version = cpu_to_le32(THIN_VERSION);
578 disk_super->time = 0;
579 disk_super->trans_id = 0;
580 disk_super->held_root = 0;
581
582 copy_sm_roots(pmd, disk_super);
583
584 disk_super->data_mapping_root = cpu_to_le64(pmd->root);
585 disk_super->device_details_root = cpu_to_le64(pmd->details_root);
586 disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE);
587 disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
588 disk_super->data_block_size = cpu_to_le32(pmd->data_block_size);
589
590 return dm_tm_commit(pmd->tm, sblock);
591}
592
593static int __format_metadata(struct dm_pool_metadata *pmd)
594{
595 int r;
596
597 r = dm_tm_create_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
598 &pmd->tm, &pmd->metadata_sm);
599 if (r < 0) {
600 DMERR("tm_create_with_sm failed");
601 return r;
602 }
603
604 pmd->data_sm = dm_sm_disk_create(pmd->tm, 0);
605 if (IS_ERR(pmd->data_sm)) {
606 DMERR("sm_disk_create failed");
607 r = PTR_ERR(pmd->data_sm);
608 goto bad_cleanup_tm;
609 }
610
611 pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
612 if (!pmd->nb_tm) {
613 DMERR("could not create non-blocking clone tm");
614 r = -ENOMEM;
615 goto bad_cleanup_data_sm;
616 }
617
618 __setup_btree_details(pmd);
619
620 r = dm_btree_empty(&pmd->info, &pmd->root);
621 if (r < 0)
622 goto bad_cleanup_nb_tm;
623
624 r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
625 if (r < 0) {
626 DMERR("couldn't create devices root");
627 goto bad_cleanup_nb_tm;
628 }
629
630 r = __write_initial_superblock(pmd);
631 if (r)
632 goto bad_cleanup_nb_tm;
633
634 return 0;
635
636bad_cleanup_nb_tm:
637 dm_tm_destroy(pmd->nb_tm);
638bad_cleanup_data_sm:
639 dm_sm_destroy(pmd->data_sm);
640bad_cleanup_tm:
641 dm_tm_destroy(pmd->tm);
642 dm_sm_destroy(pmd->metadata_sm);
643
644 return r;
645}
646
647static int __check_incompat_features(struct thin_disk_superblock *disk_super,
648 struct dm_pool_metadata *pmd)
649{
650 uint32_t features;
651
652 features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
653 if (features) {
654 DMERR("could not access metadata due to unsupported optional features (%lx).",
655 (unsigned long)features);
656 return -EINVAL;
657 }
658
659 /*
660 * Check for read-only metadata to skip the following RDWR checks.
661 */
662 if (bdev_read_only(pmd->bdev))
663 return 0;
664
665 features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
666 if (features) {
667 DMERR("could not access metadata RDWR due to unsupported optional features (%lx).",
668 (unsigned long)features);
669 return -EINVAL;
670 }
671
672 return 0;
673}
674
675static int __open_metadata(struct dm_pool_metadata *pmd)
676{
677 int r;
678 struct dm_block *sblock;
679 struct thin_disk_superblock *disk_super;
680
681 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
682 &sb_validator, &sblock);
683 if (r < 0) {
684 DMERR("couldn't read superblock");
685 return r;
686 }
687
688 disk_super = dm_block_data(sblock);
689
690 /* Verify the data block size hasn't changed */
691 if (le32_to_cpu(disk_super->data_block_size) != pmd->data_block_size) {
692 DMERR("changing the data block size (from %u to %llu) is not supported",
693 le32_to_cpu(disk_super->data_block_size),
694 (unsigned long long)pmd->data_block_size);
695 r = -EINVAL;
696 goto bad_unlock_sblock;
697 }
698
699 r = __check_incompat_features(disk_super, pmd);
700 if (r < 0)
701 goto bad_unlock_sblock;
702
703 r = dm_tm_open_with_sm(pmd->bm, THIN_SUPERBLOCK_LOCATION,
704 disk_super->metadata_space_map_root,
705 sizeof(disk_super->metadata_space_map_root),
706 &pmd->tm, &pmd->metadata_sm);
707 if (r < 0) {
708 DMERR("tm_open_with_sm failed");
709 goto bad_unlock_sblock;
710 }
711
712 pmd->data_sm = dm_sm_disk_open(pmd->tm, disk_super->data_space_map_root,
713 sizeof(disk_super->data_space_map_root));
714 if (IS_ERR(pmd->data_sm)) {
715 DMERR("sm_disk_open failed");
716 r = PTR_ERR(pmd->data_sm);
717 goto bad_cleanup_tm;
718 }
719
720 pmd->nb_tm = dm_tm_create_non_blocking_clone(pmd->tm);
721 if (!pmd->nb_tm) {
722 DMERR("could not create non-blocking clone tm");
723 r = -ENOMEM;
724 goto bad_cleanup_data_sm;
725 }
726
727 /*
728 * For pool metadata opening process, root setting is redundant
729 * because it will be set again in __begin_transaction(). But dm
730 * pool aborting process really needs to get last transaction's
731 * root to avoid accessing broken btree.
732 */
733 pmd->root = le64_to_cpu(disk_super->data_mapping_root);
734 pmd->details_root = le64_to_cpu(disk_super->device_details_root);
735
736 __setup_btree_details(pmd);
737 dm_bm_unlock(sblock);
738
739 return 0;
740
741bad_cleanup_data_sm:
742 dm_sm_destroy(pmd->data_sm);
743bad_cleanup_tm:
744 dm_tm_destroy(pmd->tm);
745 dm_sm_destroy(pmd->metadata_sm);
746bad_unlock_sblock:
747 dm_bm_unlock(sblock);
748
749 return r;
750}
751
752static int __open_or_format_metadata(struct dm_pool_metadata *pmd, bool format_device)
753{
754 int r, unformatted;
755
756 r = __superblock_all_zeroes(pmd->bm, &unformatted);
757 if (r)
758 return r;
759
760 if (unformatted)
761 return format_device ? __format_metadata(pmd) : -EPERM;
762
763 return __open_metadata(pmd);
764}
765
766static int __create_persistent_data_objects(struct dm_pool_metadata *pmd, bool format_device)
767{
768 int r;
769
770 pmd->bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
771 THIN_MAX_CONCURRENT_LOCKS);
772 if (IS_ERR(pmd->bm)) {
773 DMERR("could not create block manager");
774 r = PTR_ERR(pmd->bm);
775 pmd->bm = NULL;
776 return r;
777 }
778
779 r = __open_or_format_metadata(pmd, format_device);
780 if (r) {
781 dm_block_manager_destroy(pmd->bm);
782 pmd->bm = NULL;
783 }
784
785 return r;
786}
787
788static void __destroy_persistent_data_objects(struct dm_pool_metadata *pmd,
789 bool destroy_bm)
790{
791 dm_sm_destroy(pmd->data_sm);
792 dm_sm_destroy(pmd->metadata_sm);
793 dm_tm_destroy(pmd->nb_tm);
794 dm_tm_destroy(pmd->tm);
795 if (destroy_bm)
796 dm_block_manager_destroy(pmd->bm);
797}
798
799static int __begin_transaction(struct dm_pool_metadata *pmd)
800{
801 int r;
802 struct thin_disk_superblock *disk_super;
803 struct dm_block *sblock;
804
805 /*
806 * We re-read the superblock every time. Shouldn't need to do this
807 * really.
808 */
809 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
810 &sb_validator, &sblock);
811 if (r)
812 return r;
813
814 disk_super = dm_block_data(sblock);
815 pmd->time = le32_to_cpu(disk_super->time);
816 pmd->root = le64_to_cpu(disk_super->data_mapping_root);
817 pmd->details_root = le64_to_cpu(disk_super->device_details_root);
818 pmd->trans_id = le64_to_cpu(disk_super->trans_id);
819 pmd->flags = le32_to_cpu(disk_super->flags);
820 pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
821
822 dm_bm_unlock(sblock);
823 return 0;
824}
825
826static int __write_changed_details(struct dm_pool_metadata *pmd)
827{
828 int r;
829 struct dm_thin_device *td, *tmp;
830 struct disk_device_details details;
831 uint64_t key;
832
833 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
834 if (!td->changed)
835 continue;
836
837 key = td->id;
838
839 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
840 details.transaction_id = cpu_to_le64(td->transaction_id);
841 details.creation_time = cpu_to_le32(td->creation_time);
842 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
843 __dm_bless_for_disk(&details);
844
845 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
846 &key, &details, &pmd->details_root);
847 if (r)
848 return r;
849
850 if (td->open_count)
851 td->changed = false;
852 else {
853 list_del(&td->list);
854 kfree(td);
855 }
856 }
857
858 return 0;
859}
860
861static int __commit_transaction(struct dm_pool_metadata *pmd)
862{
863 int r;
864 struct thin_disk_superblock *disk_super;
865 struct dm_block *sblock;
866
867 /*
868 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
869 */
870 BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
871 BUG_ON(!rwsem_is_locked(&pmd->root_lock));
872
873 if (unlikely(!pmd->in_service))
874 return 0;
875
876 if (pmd->pre_commit_fn) {
877 r = pmd->pre_commit_fn(pmd->pre_commit_context);
878 if (r < 0) {
879 DMERR("pre-commit callback failed");
880 return r;
881 }
882 }
883
884 r = __write_changed_details(pmd);
885 if (r < 0)
886 return r;
887
888 r = dm_sm_commit(pmd->data_sm);
889 if (r < 0)
890 return r;
891
892 r = dm_tm_pre_commit(pmd->tm);
893 if (r < 0)
894 return r;
895
896 r = save_sm_roots(pmd);
897 if (r < 0)
898 return r;
899
900 r = superblock_lock(pmd, &sblock);
901 if (r)
902 return r;
903
904 disk_super = dm_block_data(sblock);
905 disk_super->time = cpu_to_le32(pmd->time);
906 disk_super->data_mapping_root = cpu_to_le64(pmd->root);
907 disk_super->device_details_root = cpu_to_le64(pmd->details_root);
908 disk_super->trans_id = cpu_to_le64(pmd->trans_id);
909 disk_super->flags = cpu_to_le32(pmd->flags);
910
911 copy_sm_roots(pmd, disk_super);
912
913 return dm_tm_commit(pmd->tm, sblock);
914}
915
916static void __set_metadata_reserve(struct dm_pool_metadata *pmd)
917{
918 int r;
919 dm_block_t total;
920 dm_block_t max_blocks = 4096; /* 16M */
921
922 r = dm_sm_get_nr_blocks(pmd->metadata_sm, &total);
923 if (r) {
924 DMERR("could not get size of metadata device");
925 pmd->metadata_reserve = max_blocks;
926 } else
927 pmd->metadata_reserve = min(max_blocks, div_u64(total, 10));
928}
929
930struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
931 sector_t data_block_size,
932 bool format_device)
933{
934 int r;
935 struct dm_pool_metadata *pmd;
936
937 pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
938 if (!pmd) {
939 DMERR("could not allocate metadata struct");
940 return ERR_PTR(-ENOMEM);
941 }
942
943 init_rwsem(&pmd->root_lock);
944 pmd->time = 0;
945 INIT_LIST_HEAD(&pmd->thin_devices);
946 pmd->fail_io = false;
947 pmd->in_service = false;
948 pmd->bdev = bdev;
949 pmd->data_block_size = data_block_size;
950 pmd->pre_commit_fn = NULL;
951 pmd->pre_commit_context = NULL;
952
953 r = __create_persistent_data_objects(pmd, format_device);
954 if (r) {
955 kfree(pmd);
956 return ERR_PTR(r);
957 }
958
959 r = __begin_transaction(pmd);
960 if (r < 0) {
961 if (dm_pool_metadata_close(pmd) < 0)
962 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
963 return ERR_PTR(r);
964 }
965
966 __set_metadata_reserve(pmd);
967
968 return pmd;
969}
970
971int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
972{
973 int r;
974 unsigned open_devices = 0;
975 struct dm_thin_device *td, *tmp;
976
977 down_read(&pmd->root_lock);
978 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
979 if (td->open_count)
980 open_devices++;
981 else {
982 list_del(&td->list);
983 kfree(td);
984 }
985 }
986 up_read(&pmd->root_lock);
987
988 if (open_devices) {
989 DMERR("attempt to close pmd when %u device(s) are still open",
990 open_devices);
991 return -EBUSY;
992 }
993
994 pmd_write_lock_in_core(pmd);
995 if (!pmd->fail_io && !dm_bm_is_read_only(pmd->bm)) {
996 r = __commit_transaction(pmd);
997 if (r < 0)
998 DMWARN("%s: __commit_transaction() failed, error = %d",
999 __func__, r);
1000 }
1001 pmd_write_unlock(pmd);
1002 if (!pmd->fail_io)
1003 __destroy_persistent_data_objects(pmd, true);
1004
1005 kfree(pmd);
1006 return 0;
1007}
1008
1009/*
1010 * __open_device: Returns @td corresponding to device with id @dev,
1011 * creating it if @create is set and incrementing @td->open_count.
1012 * On failure, @td is undefined.
1013 */
1014static int __open_device(struct dm_pool_metadata *pmd,
1015 dm_thin_id dev, int create,
1016 struct dm_thin_device **td)
1017{
1018 int r, changed = 0;
1019 struct dm_thin_device *td2;
1020 uint64_t key = dev;
1021 struct disk_device_details details_le;
1022
1023 /*
1024 * If the device is already open, return it.
1025 */
1026 list_for_each_entry(td2, &pmd->thin_devices, list)
1027 if (td2->id == dev) {
1028 /*
1029 * May not create an already-open device.
1030 */
1031 if (create)
1032 return -EEXIST;
1033
1034 td2->open_count++;
1035 *td = td2;
1036 return 0;
1037 }
1038
1039 /*
1040 * Check the device exists.
1041 */
1042 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1043 &key, &details_le);
1044 if (r) {
1045 if (r != -ENODATA || !create)
1046 return r;
1047
1048 /*
1049 * Create new device.
1050 */
1051 changed = 1;
1052 details_le.mapped_blocks = 0;
1053 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
1054 details_le.creation_time = cpu_to_le32(pmd->time);
1055 details_le.snapshotted_time = cpu_to_le32(pmd->time);
1056 }
1057
1058 *td = kmalloc(sizeof(**td), GFP_NOIO);
1059 if (!*td)
1060 return -ENOMEM;
1061
1062 (*td)->pmd = pmd;
1063 (*td)->id = dev;
1064 (*td)->open_count = 1;
1065 (*td)->changed = changed;
1066 (*td)->aborted_with_changes = false;
1067 (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
1068 (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
1069 (*td)->creation_time = le32_to_cpu(details_le.creation_time);
1070 (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
1071
1072 list_add(&(*td)->list, &pmd->thin_devices);
1073
1074 return 0;
1075}
1076
1077static void __close_device(struct dm_thin_device *td)
1078{
1079 --td->open_count;
1080}
1081
1082static int __create_thin(struct dm_pool_metadata *pmd,
1083 dm_thin_id dev)
1084{
1085 int r;
1086 dm_block_t dev_root;
1087 uint64_t key = dev;
1088 struct dm_thin_device *td;
1089 __le64 value;
1090
1091 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1092 &key, NULL);
1093 if (!r)
1094 return -EEXIST;
1095
1096 /*
1097 * Create an empty btree for the mappings.
1098 */
1099 r = dm_btree_empty(&pmd->bl_info, &dev_root);
1100 if (r)
1101 return r;
1102
1103 /*
1104 * Insert it into the main mapping tree.
1105 */
1106 value = cpu_to_le64(dev_root);
1107 __dm_bless_for_disk(&value);
1108 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1109 if (r) {
1110 dm_btree_del(&pmd->bl_info, dev_root);
1111 return r;
1112 }
1113
1114 r = __open_device(pmd, dev, 1, &td);
1115 if (r) {
1116 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1117 dm_btree_del(&pmd->bl_info, dev_root);
1118 return r;
1119 }
1120 __close_device(td);
1121
1122 return r;
1123}
1124
1125int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
1126{
1127 int r = -EINVAL;
1128
1129 pmd_write_lock(pmd);
1130 if (!pmd->fail_io)
1131 r = __create_thin(pmd, dev);
1132 pmd_write_unlock(pmd);
1133
1134 return r;
1135}
1136
1137static int __set_snapshot_details(struct dm_pool_metadata *pmd,
1138 struct dm_thin_device *snap,
1139 dm_thin_id origin, uint32_t time)
1140{
1141 int r;
1142 struct dm_thin_device *td;
1143
1144 r = __open_device(pmd, origin, 0, &td);
1145 if (r)
1146 return r;
1147
1148 td->changed = true;
1149 td->snapshotted_time = time;
1150
1151 snap->mapped_blocks = td->mapped_blocks;
1152 snap->snapshotted_time = time;
1153 __close_device(td);
1154
1155 return 0;
1156}
1157
1158static int __create_snap(struct dm_pool_metadata *pmd,
1159 dm_thin_id dev, dm_thin_id origin)
1160{
1161 int r;
1162 dm_block_t origin_root;
1163 uint64_t key = origin, dev_key = dev;
1164 struct dm_thin_device *td;
1165 __le64 value;
1166
1167 /* check this device is unused */
1168 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
1169 &dev_key, NULL);
1170 if (!r)
1171 return -EEXIST;
1172
1173 /* find the mapping tree for the origin */
1174 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
1175 if (r)
1176 return r;
1177 origin_root = le64_to_cpu(value);
1178
1179 /* clone the origin, an inc will do */
1180 dm_tm_inc(pmd->tm, origin_root);
1181
1182 /* insert into the main mapping tree */
1183 value = cpu_to_le64(origin_root);
1184 __dm_bless_for_disk(&value);
1185 key = dev;
1186 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
1187 if (r) {
1188 dm_tm_dec(pmd->tm, origin_root);
1189 return r;
1190 }
1191
1192 pmd->time++;
1193
1194 r = __open_device(pmd, dev, 1, &td);
1195 if (r)
1196 goto bad;
1197
1198 r = __set_snapshot_details(pmd, td, origin, pmd->time);
1199 __close_device(td);
1200
1201 if (r)
1202 goto bad;
1203
1204 return 0;
1205
1206bad:
1207 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1208 dm_btree_remove(&pmd->details_info, pmd->details_root,
1209 &key, &pmd->details_root);
1210 return r;
1211}
1212
1213int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1214 dm_thin_id dev,
1215 dm_thin_id origin)
1216{
1217 int r = -EINVAL;
1218
1219 pmd_write_lock(pmd);
1220 if (!pmd->fail_io)
1221 r = __create_snap(pmd, dev, origin);
1222 pmd_write_unlock(pmd);
1223
1224 return r;
1225}
1226
1227static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1228{
1229 int r;
1230 uint64_t key = dev;
1231 struct dm_thin_device *td;
1232
1233 /* TODO: failure should mark the transaction invalid */
1234 r = __open_device(pmd, dev, 0, &td);
1235 if (r)
1236 return r;
1237
1238 if (td->open_count > 1) {
1239 __close_device(td);
1240 return -EBUSY;
1241 }
1242
1243 list_del(&td->list);
1244 kfree(td);
1245 r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1246 &key, &pmd->details_root);
1247 if (r)
1248 return r;
1249
1250 r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1251 if (r)
1252 return r;
1253
1254 return 0;
1255}
1256
1257int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1258 dm_thin_id dev)
1259{
1260 int r = -EINVAL;
1261
1262 pmd_write_lock(pmd);
1263 if (!pmd->fail_io)
1264 r = __delete_device(pmd, dev);
1265 pmd_write_unlock(pmd);
1266
1267 return r;
1268}
1269
1270int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1271 uint64_t current_id,
1272 uint64_t new_id)
1273{
1274 int r = -EINVAL;
1275
1276 pmd_write_lock(pmd);
1277
1278 if (pmd->fail_io)
1279 goto out;
1280
1281 if (pmd->trans_id != current_id) {
1282 DMERR("mismatched transaction id");
1283 goto out;
1284 }
1285
1286 pmd->trans_id = new_id;
1287 r = 0;
1288
1289out:
1290 pmd_write_unlock(pmd);
1291
1292 return r;
1293}
1294
1295int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1296 uint64_t *result)
1297{
1298 int r = -EINVAL;
1299
1300 down_read(&pmd->root_lock);
1301 if (!pmd->fail_io) {
1302 *result = pmd->trans_id;
1303 r = 0;
1304 }
1305 up_read(&pmd->root_lock);
1306
1307 return r;
1308}
1309
1310static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1311{
1312 int r, inc;
1313 struct thin_disk_superblock *disk_super;
1314 struct dm_block *copy, *sblock;
1315 dm_block_t held_root;
1316
1317 /*
1318 * We commit to ensure the btree roots which we increment in a
1319 * moment are up to date.
1320 */
1321 r = __commit_transaction(pmd);
1322 if (r < 0) {
1323 DMWARN("%s: __commit_transaction() failed, error = %d",
1324 __func__, r);
1325 return r;
1326 }
1327
1328 /*
1329 * Copy the superblock.
1330 */
1331 dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1332 r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1333 &sb_validator, ©, &inc);
1334 if (r)
1335 return r;
1336
1337 BUG_ON(!inc);
1338
1339 held_root = dm_block_location(copy);
1340 disk_super = dm_block_data(copy);
1341
1342 if (le64_to_cpu(disk_super->held_root)) {
1343 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1344
1345 dm_tm_dec(pmd->tm, held_root);
1346 dm_tm_unlock(pmd->tm, copy);
1347 return -EBUSY;
1348 }
1349
1350 /*
1351 * Wipe the spacemap since we're not publishing this.
1352 */
1353 memset(&disk_super->data_space_map_root, 0,
1354 sizeof(disk_super->data_space_map_root));
1355 memset(&disk_super->metadata_space_map_root, 0,
1356 sizeof(disk_super->metadata_space_map_root));
1357
1358 /*
1359 * Increment the data structures that need to be preserved.
1360 */
1361 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1362 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1363 dm_tm_unlock(pmd->tm, copy);
1364
1365 /*
1366 * Write the held root into the superblock.
1367 */
1368 r = superblock_lock(pmd, &sblock);
1369 if (r) {
1370 dm_tm_dec(pmd->tm, held_root);
1371 return r;
1372 }
1373
1374 disk_super = dm_block_data(sblock);
1375 disk_super->held_root = cpu_to_le64(held_root);
1376 dm_bm_unlock(sblock);
1377 return 0;
1378}
1379
1380int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1381{
1382 int r = -EINVAL;
1383
1384 pmd_write_lock(pmd);
1385 if (!pmd->fail_io)
1386 r = __reserve_metadata_snap(pmd);
1387 pmd_write_unlock(pmd);
1388
1389 return r;
1390}
1391
1392static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1393{
1394 int r;
1395 struct thin_disk_superblock *disk_super;
1396 struct dm_block *sblock, *copy;
1397 dm_block_t held_root;
1398
1399 r = superblock_lock(pmd, &sblock);
1400 if (r)
1401 return r;
1402
1403 disk_super = dm_block_data(sblock);
1404 held_root = le64_to_cpu(disk_super->held_root);
1405 disk_super->held_root = cpu_to_le64(0);
1406
1407 dm_bm_unlock(sblock);
1408
1409 if (!held_root) {
1410 DMWARN("No pool metadata snapshot found: nothing to release.");
1411 return -EINVAL;
1412 }
1413
1414 r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, ©);
1415 if (r)
1416 return r;
1417
1418 disk_super = dm_block_data(copy);
1419 dm_btree_del(&pmd->info, le64_to_cpu(disk_super->data_mapping_root));
1420 dm_btree_del(&pmd->details_info, le64_to_cpu(disk_super->device_details_root));
1421 dm_sm_dec_block(pmd->metadata_sm, held_root);
1422
1423 dm_tm_unlock(pmd->tm, copy);
1424
1425 return 0;
1426}
1427
1428int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1429{
1430 int r = -EINVAL;
1431
1432 pmd_write_lock(pmd);
1433 if (!pmd->fail_io)
1434 r = __release_metadata_snap(pmd);
1435 pmd_write_unlock(pmd);
1436
1437 return r;
1438}
1439
1440static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1441 dm_block_t *result)
1442{
1443 int r;
1444 struct thin_disk_superblock *disk_super;
1445 struct dm_block *sblock;
1446
1447 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1448 &sb_validator, &sblock);
1449 if (r)
1450 return r;
1451
1452 disk_super = dm_block_data(sblock);
1453 *result = le64_to_cpu(disk_super->held_root);
1454
1455 dm_bm_unlock(sblock);
1456
1457 return 0;
1458}
1459
1460int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1461 dm_block_t *result)
1462{
1463 int r = -EINVAL;
1464
1465 down_read(&pmd->root_lock);
1466 if (!pmd->fail_io)
1467 r = __get_metadata_snap(pmd, result);
1468 up_read(&pmd->root_lock);
1469
1470 return r;
1471}
1472
1473int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1474 struct dm_thin_device **td)
1475{
1476 int r = -EINVAL;
1477
1478 pmd_write_lock_in_core(pmd);
1479 if (!pmd->fail_io)
1480 r = __open_device(pmd, dev, 0, td);
1481 pmd_write_unlock(pmd);
1482
1483 return r;
1484}
1485
1486int dm_pool_close_thin_device(struct dm_thin_device *td)
1487{
1488 pmd_write_lock_in_core(td->pmd);
1489 __close_device(td);
1490 pmd_write_unlock(td->pmd);
1491
1492 return 0;
1493}
1494
1495dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1496{
1497 return td->id;
1498}
1499
1500/*
1501 * Check whether @time (of block creation) is older than @td's last snapshot.
1502 * If so then the associated block is shared with the last snapshot device.
1503 * Any block on a device created *after* the device last got snapshotted is
1504 * necessarily not shared.
1505 */
1506static bool __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1507{
1508 return td->snapshotted_time > time;
1509}
1510
1511static void unpack_lookup_result(struct dm_thin_device *td, __le64 value,
1512 struct dm_thin_lookup_result *result)
1513{
1514 uint64_t block_time = 0;
1515 dm_block_t exception_block;
1516 uint32_t exception_time;
1517
1518 block_time = le64_to_cpu(value);
1519 unpack_block_time(block_time, &exception_block, &exception_time);
1520 result->block = exception_block;
1521 result->shared = __snapshotted_since(td, exception_time);
1522}
1523
1524static int __find_block(struct dm_thin_device *td, dm_block_t block,
1525 int can_issue_io, struct dm_thin_lookup_result *result)
1526{
1527 int r;
1528 __le64 value;
1529 struct dm_pool_metadata *pmd = td->pmd;
1530 dm_block_t keys[2] = { td->id, block };
1531 struct dm_btree_info *info;
1532
1533 if (can_issue_io) {
1534 info = &pmd->info;
1535 } else
1536 info = &pmd->nb_info;
1537
1538 r = dm_btree_lookup(info, pmd->root, keys, &value);
1539 if (!r)
1540 unpack_lookup_result(td, value, result);
1541
1542 return r;
1543}
1544
1545int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1546 int can_issue_io, struct dm_thin_lookup_result *result)
1547{
1548 int r;
1549 struct dm_pool_metadata *pmd = td->pmd;
1550
1551 down_read(&pmd->root_lock);
1552 if (pmd->fail_io) {
1553 up_read(&pmd->root_lock);
1554 return -EINVAL;
1555 }
1556
1557 r = __find_block(td, block, can_issue_io, result);
1558
1559 up_read(&pmd->root_lock);
1560 return r;
1561}
1562
1563static int __find_next_mapped_block(struct dm_thin_device *td, dm_block_t block,
1564 dm_block_t *vblock,
1565 struct dm_thin_lookup_result *result)
1566{
1567 int r;
1568 __le64 value;
1569 struct dm_pool_metadata *pmd = td->pmd;
1570 dm_block_t keys[2] = { td->id, block };
1571
1572 r = dm_btree_lookup_next(&pmd->info, pmd->root, keys, vblock, &value);
1573 if (!r)
1574 unpack_lookup_result(td, value, result);
1575
1576 return r;
1577}
1578
1579static int __find_mapped_range(struct dm_thin_device *td,
1580 dm_block_t begin, dm_block_t end,
1581 dm_block_t *thin_begin, dm_block_t *thin_end,
1582 dm_block_t *pool_begin, bool *maybe_shared)
1583{
1584 int r;
1585 dm_block_t pool_end;
1586 struct dm_thin_lookup_result lookup;
1587
1588 if (end < begin)
1589 return -ENODATA;
1590
1591 r = __find_next_mapped_block(td, begin, &begin, &lookup);
1592 if (r)
1593 return r;
1594
1595 if (begin >= end)
1596 return -ENODATA;
1597
1598 *thin_begin = begin;
1599 *pool_begin = lookup.block;
1600 *maybe_shared = lookup.shared;
1601
1602 begin++;
1603 pool_end = *pool_begin + 1;
1604 while (begin != end) {
1605 r = __find_block(td, begin, true, &lookup);
1606 if (r) {
1607 if (r == -ENODATA)
1608 break;
1609 else
1610 return r;
1611 }
1612
1613 if ((lookup.block != pool_end) ||
1614 (lookup.shared != *maybe_shared))
1615 break;
1616
1617 pool_end++;
1618 begin++;
1619 }
1620
1621 *thin_end = begin;
1622 return 0;
1623}
1624
1625int dm_thin_find_mapped_range(struct dm_thin_device *td,
1626 dm_block_t begin, dm_block_t end,
1627 dm_block_t *thin_begin, dm_block_t *thin_end,
1628 dm_block_t *pool_begin, bool *maybe_shared)
1629{
1630 int r = -EINVAL;
1631 struct dm_pool_metadata *pmd = td->pmd;
1632
1633 down_read(&pmd->root_lock);
1634 if (!pmd->fail_io) {
1635 r = __find_mapped_range(td, begin, end, thin_begin, thin_end,
1636 pool_begin, maybe_shared);
1637 }
1638 up_read(&pmd->root_lock);
1639
1640 return r;
1641}
1642
1643static int __insert(struct dm_thin_device *td, dm_block_t block,
1644 dm_block_t data_block)
1645{
1646 int r, inserted;
1647 __le64 value;
1648 struct dm_pool_metadata *pmd = td->pmd;
1649 dm_block_t keys[2] = { td->id, block };
1650
1651 value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1652 __dm_bless_for_disk(&value);
1653
1654 r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1655 &pmd->root, &inserted);
1656 if (r)
1657 return r;
1658
1659 td->changed = true;
1660 if (inserted)
1661 td->mapped_blocks++;
1662
1663 return 0;
1664}
1665
1666int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1667 dm_block_t data_block)
1668{
1669 int r = -EINVAL;
1670
1671 pmd_write_lock(td->pmd);
1672 if (!td->pmd->fail_io)
1673 r = __insert(td, block, data_block);
1674 pmd_write_unlock(td->pmd);
1675
1676 return r;
1677}
1678
1679static int __remove_range(struct dm_thin_device *td, dm_block_t begin, dm_block_t end)
1680{
1681 int r;
1682 unsigned count, total_count = 0;
1683 struct dm_pool_metadata *pmd = td->pmd;
1684 dm_block_t keys[1] = { td->id };
1685 __le64 value;
1686 dm_block_t mapping_root;
1687
1688 /*
1689 * Find the mapping tree
1690 */
1691 r = dm_btree_lookup(&pmd->tl_info, pmd->root, keys, &value);
1692 if (r)
1693 return r;
1694
1695 /*
1696 * Remove from the mapping tree, taking care to inc the
1697 * ref count so it doesn't get deleted.
1698 */
1699 mapping_root = le64_to_cpu(value);
1700 dm_tm_inc(pmd->tm, mapping_root);
1701 r = dm_btree_remove(&pmd->tl_info, pmd->root, keys, &pmd->root);
1702 if (r)
1703 return r;
1704
1705 /*
1706 * Remove leaves stops at the first unmapped entry, so we have to
1707 * loop round finding mapped ranges.
1708 */
1709 while (begin < end) {
1710 r = dm_btree_lookup_next(&pmd->bl_info, mapping_root, &begin, &begin, &value);
1711 if (r == -ENODATA)
1712 break;
1713
1714 if (r)
1715 return r;
1716
1717 if (begin >= end)
1718 break;
1719
1720 r = dm_btree_remove_leaves(&pmd->bl_info, mapping_root, &begin, end, &mapping_root, &count);
1721 if (r)
1722 return r;
1723
1724 total_count += count;
1725 }
1726
1727 td->mapped_blocks -= total_count;
1728 td->changed = true;
1729
1730 /*
1731 * Reinsert the mapping tree.
1732 */
1733 value = cpu_to_le64(mapping_root);
1734 __dm_bless_for_disk(&value);
1735 return dm_btree_insert(&pmd->tl_info, pmd->root, keys, &value, &pmd->root);
1736}
1737
1738int dm_thin_remove_range(struct dm_thin_device *td,
1739 dm_block_t begin, dm_block_t end)
1740{
1741 int r = -EINVAL;
1742
1743 pmd_write_lock(td->pmd);
1744 if (!td->pmd->fail_io)
1745 r = __remove_range(td, begin, end);
1746 pmd_write_unlock(td->pmd);
1747
1748 return r;
1749}
1750
1751int dm_pool_block_is_shared(struct dm_pool_metadata *pmd, dm_block_t b, bool *result)
1752{
1753 int r;
1754 uint32_t ref_count;
1755
1756 down_read(&pmd->root_lock);
1757 r = dm_sm_get_count(pmd->data_sm, b, &ref_count);
1758 if (!r)
1759 *result = (ref_count > 1);
1760 up_read(&pmd->root_lock);
1761
1762 return r;
1763}
1764
1765int dm_pool_inc_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1766{
1767 int r = 0;
1768
1769 pmd_write_lock(pmd);
1770 r = dm_sm_inc_blocks(pmd->data_sm, b, e);
1771 pmd_write_unlock(pmd);
1772
1773 return r;
1774}
1775
1776int dm_pool_dec_data_range(struct dm_pool_metadata *pmd, dm_block_t b, dm_block_t e)
1777{
1778 int r = 0;
1779
1780 pmd_write_lock(pmd);
1781 r = dm_sm_dec_blocks(pmd->data_sm, b, e);
1782 pmd_write_unlock(pmd);
1783
1784 return r;
1785}
1786
1787bool dm_thin_changed_this_transaction(struct dm_thin_device *td)
1788{
1789 int r;
1790
1791 down_read(&td->pmd->root_lock);
1792 r = td->changed;
1793 up_read(&td->pmd->root_lock);
1794
1795 return r;
1796}
1797
1798bool dm_pool_changed_this_transaction(struct dm_pool_metadata *pmd)
1799{
1800 bool r = false;
1801 struct dm_thin_device *td, *tmp;
1802
1803 down_read(&pmd->root_lock);
1804 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
1805 if (td->changed) {
1806 r = td->changed;
1807 break;
1808 }
1809 }
1810 up_read(&pmd->root_lock);
1811
1812 return r;
1813}
1814
1815bool dm_thin_aborted_changes(struct dm_thin_device *td)
1816{
1817 bool r;
1818
1819 down_read(&td->pmd->root_lock);
1820 r = td->aborted_with_changes;
1821 up_read(&td->pmd->root_lock);
1822
1823 return r;
1824}
1825
1826int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1827{
1828 int r = -EINVAL;
1829
1830 pmd_write_lock(pmd);
1831 if (!pmd->fail_io)
1832 r = dm_sm_new_block(pmd->data_sm, result);
1833 pmd_write_unlock(pmd);
1834
1835 return r;
1836}
1837
1838int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1839{
1840 int r = -EINVAL;
1841
1842 /*
1843 * Care is taken to not have commit be what
1844 * triggers putting the thin-pool in-service.
1845 */
1846 pmd_write_lock_in_core(pmd);
1847 if (pmd->fail_io)
1848 goto out;
1849
1850 r = __commit_transaction(pmd);
1851 if (r < 0)
1852 goto out;
1853
1854 /*
1855 * Open the next transaction.
1856 */
1857 r = __begin_transaction(pmd);
1858out:
1859 pmd_write_unlock(pmd);
1860 return r;
1861}
1862
1863static void __set_abort_with_changes_flags(struct dm_pool_metadata *pmd)
1864{
1865 struct dm_thin_device *td;
1866
1867 list_for_each_entry(td, &pmd->thin_devices, list)
1868 td->aborted_with_changes = td->changed;
1869}
1870
1871int dm_pool_abort_metadata(struct dm_pool_metadata *pmd)
1872{
1873 int r = -EINVAL;
1874 struct dm_block_manager *old_bm = NULL, *new_bm = NULL;
1875
1876 /* fail_io is double-checked with pmd->root_lock held below */
1877 if (unlikely(pmd->fail_io))
1878 return r;
1879
1880 /*
1881 * Replacement block manager (new_bm) is created and old_bm destroyed outside of
1882 * pmd root_lock to avoid ABBA deadlock that would result (due to life-cycle of
1883 * shrinker associated with the block manager's bufio client vs pmd root_lock).
1884 * - must take shrinker_rwsem without holding pmd->root_lock
1885 */
1886 new_bm = dm_block_manager_create(pmd->bdev, THIN_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
1887 THIN_MAX_CONCURRENT_LOCKS);
1888
1889 pmd_write_lock(pmd);
1890 if (pmd->fail_io) {
1891 pmd_write_unlock(pmd);
1892 goto out;
1893 }
1894
1895 __set_abort_with_changes_flags(pmd);
1896 __destroy_persistent_data_objects(pmd, false);
1897 old_bm = pmd->bm;
1898 if (IS_ERR(new_bm)) {
1899 DMERR("could not create block manager during abort");
1900 pmd->bm = NULL;
1901 r = PTR_ERR(new_bm);
1902 goto out_unlock;
1903 }
1904
1905 pmd->bm = new_bm;
1906 r = __open_or_format_metadata(pmd, false);
1907 if (r) {
1908 pmd->bm = NULL;
1909 goto out_unlock;
1910 }
1911 new_bm = NULL;
1912out_unlock:
1913 if (r)
1914 pmd->fail_io = true;
1915 pmd_write_unlock(pmd);
1916 dm_block_manager_destroy(old_bm);
1917out:
1918 if (new_bm && !IS_ERR(new_bm))
1919 dm_block_manager_destroy(new_bm);
1920
1921 return r;
1922}
1923
1924int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1925{
1926 int r = -EINVAL;
1927
1928 down_read(&pmd->root_lock);
1929 if (!pmd->fail_io)
1930 r = dm_sm_get_nr_free(pmd->data_sm, result);
1931 up_read(&pmd->root_lock);
1932
1933 return r;
1934}
1935
1936int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1937 dm_block_t *result)
1938{
1939 int r = -EINVAL;
1940
1941 down_read(&pmd->root_lock);
1942 if (!pmd->fail_io)
1943 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1944
1945 if (!r) {
1946 if (*result < pmd->metadata_reserve)
1947 *result = 0;
1948 else
1949 *result -= pmd->metadata_reserve;
1950 }
1951 up_read(&pmd->root_lock);
1952
1953 return r;
1954}
1955
1956int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1957 dm_block_t *result)
1958{
1959 int r = -EINVAL;
1960
1961 down_read(&pmd->root_lock);
1962 if (!pmd->fail_io)
1963 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1964 up_read(&pmd->root_lock);
1965
1966 return r;
1967}
1968
1969int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1970{
1971 int r = -EINVAL;
1972
1973 down_read(&pmd->root_lock);
1974 if (!pmd->fail_io)
1975 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1976 up_read(&pmd->root_lock);
1977
1978 return r;
1979}
1980
1981int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1982{
1983 int r = -EINVAL;
1984 struct dm_pool_metadata *pmd = td->pmd;
1985
1986 down_read(&pmd->root_lock);
1987 if (!pmd->fail_io) {
1988 *result = td->mapped_blocks;
1989 r = 0;
1990 }
1991 up_read(&pmd->root_lock);
1992
1993 return r;
1994}
1995
1996static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1997{
1998 int r;
1999 __le64 value_le;
2000 dm_block_t thin_root;
2001 struct dm_pool_metadata *pmd = td->pmd;
2002
2003 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
2004 if (r)
2005 return r;
2006
2007 thin_root = le64_to_cpu(value_le);
2008
2009 return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
2010}
2011
2012int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
2013 dm_block_t *result)
2014{
2015 int r = -EINVAL;
2016 struct dm_pool_metadata *pmd = td->pmd;
2017
2018 down_read(&pmd->root_lock);
2019 if (!pmd->fail_io)
2020 r = __highest_block(td, result);
2021 up_read(&pmd->root_lock);
2022
2023 return r;
2024}
2025
2026static int __resize_space_map(struct dm_space_map *sm, dm_block_t new_count)
2027{
2028 int r;
2029 dm_block_t old_count;
2030
2031 r = dm_sm_get_nr_blocks(sm, &old_count);
2032 if (r)
2033 return r;
2034
2035 if (new_count == old_count)
2036 return 0;
2037
2038 if (new_count < old_count) {
2039 DMERR("cannot reduce size of space map");
2040 return -EINVAL;
2041 }
2042
2043 return dm_sm_extend(sm, new_count - old_count);
2044}
2045
2046int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2047{
2048 int r = -EINVAL;
2049
2050 pmd_write_lock(pmd);
2051 if (!pmd->fail_io)
2052 r = __resize_space_map(pmd->data_sm, new_count);
2053 pmd_write_unlock(pmd);
2054
2055 return r;
2056}
2057
2058int dm_pool_resize_metadata_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
2059{
2060 int r = -EINVAL;
2061
2062 pmd_write_lock(pmd);
2063 if (!pmd->fail_io) {
2064 r = __resize_space_map(pmd->metadata_sm, new_count);
2065 if (!r)
2066 __set_metadata_reserve(pmd);
2067 }
2068 pmd_write_unlock(pmd);
2069
2070 return r;
2071}
2072
2073void dm_pool_metadata_read_only(struct dm_pool_metadata *pmd)
2074{
2075 pmd_write_lock_in_core(pmd);
2076 dm_bm_set_read_only(pmd->bm);
2077 pmd_write_unlock(pmd);
2078}
2079
2080void dm_pool_metadata_read_write(struct dm_pool_metadata *pmd)
2081{
2082 pmd_write_lock_in_core(pmd);
2083 dm_bm_set_read_write(pmd->bm);
2084 pmd_write_unlock(pmd);
2085}
2086
2087int dm_pool_register_metadata_threshold(struct dm_pool_metadata *pmd,
2088 dm_block_t threshold,
2089 dm_sm_threshold_fn fn,
2090 void *context)
2091{
2092 int r = -EINVAL;
2093
2094 pmd_write_lock_in_core(pmd);
2095 if (!pmd->fail_io) {
2096 r = dm_sm_register_threshold_callback(pmd->metadata_sm,
2097 threshold, fn, context);
2098 }
2099 pmd_write_unlock(pmd);
2100
2101 return r;
2102}
2103
2104void dm_pool_register_pre_commit_callback(struct dm_pool_metadata *pmd,
2105 dm_pool_pre_commit_fn fn,
2106 void *context)
2107{
2108 pmd_write_lock_in_core(pmd);
2109 pmd->pre_commit_fn = fn;
2110 pmd->pre_commit_context = context;
2111 pmd_write_unlock(pmd);
2112}
2113
2114int dm_pool_metadata_set_needs_check(struct dm_pool_metadata *pmd)
2115{
2116 int r = -EINVAL;
2117 struct dm_block *sblock;
2118 struct thin_disk_superblock *disk_super;
2119
2120 pmd_write_lock(pmd);
2121 if (pmd->fail_io)
2122 goto out;
2123
2124 pmd->flags |= THIN_METADATA_NEEDS_CHECK_FLAG;
2125
2126 r = superblock_lock(pmd, &sblock);
2127 if (r) {
2128 DMERR("couldn't lock superblock");
2129 goto out;
2130 }
2131
2132 disk_super = dm_block_data(sblock);
2133 disk_super->flags = cpu_to_le32(pmd->flags);
2134
2135 dm_bm_unlock(sblock);
2136out:
2137 pmd_write_unlock(pmd);
2138 return r;
2139}
2140
2141bool dm_pool_metadata_needs_check(struct dm_pool_metadata *pmd)
2142{
2143 bool needs_check;
2144
2145 down_read(&pmd->root_lock);
2146 needs_check = pmd->flags & THIN_METADATA_NEEDS_CHECK_FLAG;
2147 up_read(&pmd->root_lock);
2148
2149 return needs_check;
2150}
2151
2152void dm_pool_issue_prefetches(struct dm_pool_metadata *pmd)
2153{
2154 down_read(&pmd->root_lock);
2155 if (!pmd->fail_io)
2156 dm_tm_issue_prefetches(pmd->tm);
2157 up_read(&pmd->root_lock);
2158}
1/*
2 * Copyright (C) 2011 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 1
80#define THIN_METADATA_CACHE_SIZE 64
81#define SECTOR_TO_BLOCK_SHIFT 3
82
83/* This should be plenty */
84#define SPACE_MAP_ROOT_SIZE 128
85
86/*
87 * Little endian on-disk superblock and device details.
88 */
89struct thin_disk_superblock {
90 __le32 csum; /* Checksum of superblock except for this field. */
91 __le32 flags;
92 __le64 blocknr; /* This block number, dm_block_t. */
93
94 __u8 uuid[16];
95 __le64 magic;
96 __le32 version;
97 __le32 time;
98
99 __le64 trans_id;
100
101 /*
102 * Root held by userspace transactions.
103 */
104 __le64 held_root;
105
106 __u8 data_space_map_root[SPACE_MAP_ROOT_SIZE];
107 __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
108
109 /*
110 * 2-level btree mapping (dev_id, (dev block, time)) -> data block
111 */
112 __le64 data_mapping_root;
113
114 /*
115 * Device detail root mapping dev_id -> device_details
116 */
117 __le64 device_details_root;
118
119 __le32 data_block_size; /* In 512-byte sectors. */
120
121 __le32 metadata_block_size; /* In 512-byte sectors. */
122 __le64 metadata_nr_blocks;
123
124 __le32 compat_flags;
125 __le32 compat_ro_flags;
126 __le32 incompat_flags;
127} __packed;
128
129struct disk_device_details {
130 __le64 mapped_blocks;
131 __le64 transaction_id; /* When created. */
132 __le32 creation_time;
133 __le32 snapshotted_time;
134} __packed;
135
136struct dm_pool_metadata {
137 struct hlist_node hash;
138
139 struct block_device *bdev;
140 struct dm_block_manager *bm;
141 struct dm_space_map *metadata_sm;
142 struct dm_space_map *data_sm;
143 struct dm_transaction_manager *tm;
144 struct dm_transaction_manager *nb_tm;
145
146 /*
147 * Two-level btree.
148 * First level holds thin_dev_t.
149 * Second level holds mappings.
150 */
151 struct dm_btree_info info;
152
153 /*
154 * Non-blocking version of the above.
155 */
156 struct dm_btree_info nb_info;
157
158 /*
159 * Just the top level for deleting whole devices.
160 */
161 struct dm_btree_info tl_info;
162
163 /*
164 * Just the bottom level for creating new devices.
165 */
166 struct dm_btree_info bl_info;
167
168 /*
169 * Describes the device details btree.
170 */
171 struct dm_btree_info details_info;
172
173 struct rw_semaphore root_lock;
174 uint32_t time;
175 int need_commit;
176 dm_block_t root;
177 dm_block_t details_root;
178 struct list_head thin_devices;
179 uint64_t trans_id;
180 unsigned long flags;
181 sector_t data_block_size;
182};
183
184struct dm_thin_device {
185 struct list_head list;
186 struct dm_pool_metadata *pmd;
187 dm_thin_id id;
188
189 int open_count;
190 int changed;
191 uint64_t mapped_blocks;
192 uint64_t transaction_id;
193 uint32_t creation_time;
194 uint32_t snapshotted_time;
195};
196
197/*----------------------------------------------------------------
198 * superblock validator
199 *--------------------------------------------------------------*/
200
201#define SUPERBLOCK_CSUM_XOR 160774
202
203static void sb_prepare_for_write(struct dm_block_validator *v,
204 struct dm_block *b,
205 size_t block_size)
206{
207 struct thin_disk_superblock *disk_super = dm_block_data(b);
208
209 disk_super->blocknr = cpu_to_le64(dm_block_location(b));
210 disk_super->csum = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
211 block_size - sizeof(__le32),
212 SUPERBLOCK_CSUM_XOR));
213}
214
215static int sb_check(struct dm_block_validator *v,
216 struct dm_block *b,
217 size_t block_size)
218{
219 struct thin_disk_superblock *disk_super = dm_block_data(b);
220 __le32 csum_le;
221
222 if (dm_block_location(b) != le64_to_cpu(disk_super->blocknr)) {
223 DMERR("sb_check failed: blocknr %llu: "
224 "wanted %llu", le64_to_cpu(disk_super->blocknr),
225 (unsigned long long)dm_block_location(b));
226 return -ENOTBLK;
227 }
228
229 if (le64_to_cpu(disk_super->magic) != THIN_SUPERBLOCK_MAGIC) {
230 DMERR("sb_check failed: magic %llu: "
231 "wanted %llu", le64_to_cpu(disk_super->magic),
232 (unsigned long long)THIN_SUPERBLOCK_MAGIC);
233 return -EILSEQ;
234 }
235
236 csum_le = cpu_to_le32(dm_bm_checksum(&disk_super->flags,
237 block_size - sizeof(__le32),
238 SUPERBLOCK_CSUM_XOR));
239 if (csum_le != disk_super->csum) {
240 DMERR("sb_check failed: csum %u: wanted %u",
241 le32_to_cpu(csum_le), le32_to_cpu(disk_super->csum));
242 return -EILSEQ;
243 }
244
245 return 0;
246}
247
248static struct dm_block_validator sb_validator = {
249 .name = "superblock",
250 .prepare_for_write = sb_prepare_for_write,
251 .check = sb_check
252};
253
254/*----------------------------------------------------------------
255 * Methods for the btree value types
256 *--------------------------------------------------------------*/
257
258static uint64_t pack_block_time(dm_block_t b, uint32_t t)
259{
260 return (b << 24) | t;
261}
262
263static void unpack_block_time(uint64_t v, dm_block_t *b, uint32_t *t)
264{
265 *b = v >> 24;
266 *t = v & ((1 << 24) - 1);
267}
268
269static void data_block_inc(void *context, void *value_le)
270{
271 struct dm_space_map *sm = context;
272 __le64 v_le;
273 uint64_t b;
274 uint32_t t;
275
276 memcpy(&v_le, value_le, sizeof(v_le));
277 unpack_block_time(le64_to_cpu(v_le), &b, &t);
278 dm_sm_inc_block(sm, b);
279}
280
281static void data_block_dec(void *context, void *value_le)
282{
283 struct dm_space_map *sm = context;
284 __le64 v_le;
285 uint64_t b;
286 uint32_t t;
287
288 memcpy(&v_le, value_le, sizeof(v_le));
289 unpack_block_time(le64_to_cpu(v_le), &b, &t);
290 dm_sm_dec_block(sm, b);
291}
292
293static int data_block_equal(void *context, void *value1_le, void *value2_le)
294{
295 __le64 v1_le, v2_le;
296 uint64_t b1, b2;
297 uint32_t t;
298
299 memcpy(&v1_le, value1_le, sizeof(v1_le));
300 memcpy(&v2_le, value2_le, sizeof(v2_le));
301 unpack_block_time(le64_to_cpu(v1_le), &b1, &t);
302 unpack_block_time(le64_to_cpu(v2_le), &b2, &t);
303
304 return b1 == b2;
305}
306
307static void subtree_inc(void *context, void *value)
308{
309 struct dm_btree_info *info = context;
310 __le64 root_le;
311 uint64_t root;
312
313 memcpy(&root_le, value, sizeof(root_le));
314 root = le64_to_cpu(root_le);
315 dm_tm_inc(info->tm, root);
316}
317
318static void subtree_dec(void *context, void *value)
319{
320 struct dm_btree_info *info = context;
321 __le64 root_le;
322 uint64_t root;
323
324 memcpy(&root_le, value, sizeof(root_le));
325 root = le64_to_cpu(root_le);
326 if (dm_btree_del(info, root))
327 DMERR("btree delete failed\n");
328}
329
330static int subtree_equal(void *context, void *value1_le, void *value2_le)
331{
332 __le64 v1_le, v2_le;
333 memcpy(&v1_le, value1_le, sizeof(v1_le));
334 memcpy(&v2_le, value2_le, sizeof(v2_le));
335
336 return v1_le == v2_le;
337}
338
339/*----------------------------------------------------------------*/
340
341static int superblock_all_zeroes(struct dm_block_manager *bm, int *result)
342{
343 int r;
344 unsigned i;
345 struct dm_block *b;
346 __le64 *data_le, zero = cpu_to_le64(0);
347 unsigned block_size = dm_bm_block_size(bm) / sizeof(__le64);
348
349 /*
350 * We can't use a validator here - it may be all zeroes.
351 */
352 r = dm_bm_read_lock(bm, THIN_SUPERBLOCK_LOCATION, NULL, &b);
353 if (r)
354 return r;
355
356 data_le = dm_block_data(b);
357 *result = 1;
358 for (i = 0; i < block_size; i++) {
359 if (data_le[i] != zero) {
360 *result = 0;
361 break;
362 }
363 }
364
365 return dm_bm_unlock(b);
366}
367
368static int init_pmd(struct dm_pool_metadata *pmd,
369 struct dm_block_manager *bm,
370 dm_block_t nr_blocks, int create)
371{
372 int r;
373 struct dm_space_map *sm, *data_sm;
374 struct dm_transaction_manager *tm;
375 struct dm_block *sblock;
376
377 if (create) {
378 r = dm_tm_create_with_sm(bm, THIN_SUPERBLOCK_LOCATION,
379 &sb_validator, &tm, &sm, &sblock);
380 if (r < 0) {
381 DMERR("tm_create_with_sm failed");
382 return r;
383 }
384
385 data_sm = dm_sm_disk_create(tm, nr_blocks);
386 if (IS_ERR(data_sm)) {
387 DMERR("sm_disk_create failed");
388 dm_tm_unlock(tm, sblock);
389 r = PTR_ERR(data_sm);
390 goto bad;
391 }
392 } else {
393 struct thin_disk_superblock *disk_super = NULL;
394 size_t space_map_root_offset =
395 offsetof(struct thin_disk_superblock, metadata_space_map_root);
396
397 r = dm_tm_open_with_sm(bm, THIN_SUPERBLOCK_LOCATION,
398 &sb_validator, space_map_root_offset,
399 SPACE_MAP_ROOT_SIZE, &tm, &sm, &sblock);
400 if (r < 0) {
401 DMERR("tm_open_with_sm failed");
402 return r;
403 }
404
405 disk_super = dm_block_data(sblock);
406 data_sm = dm_sm_disk_open(tm, disk_super->data_space_map_root,
407 sizeof(disk_super->data_space_map_root));
408 if (IS_ERR(data_sm)) {
409 DMERR("sm_disk_open failed");
410 r = PTR_ERR(data_sm);
411 goto bad;
412 }
413 }
414
415
416 r = dm_tm_unlock(tm, sblock);
417 if (r < 0) {
418 DMERR("couldn't unlock superblock");
419 goto bad_data_sm;
420 }
421
422 pmd->bm = bm;
423 pmd->metadata_sm = sm;
424 pmd->data_sm = data_sm;
425 pmd->tm = tm;
426 pmd->nb_tm = dm_tm_create_non_blocking_clone(tm);
427 if (!pmd->nb_tm) {
428 DMERR("could not create clone tm");
429 r = -ENOMEM;
430 goto bad_data_sm;
431 }
432
433 pmd->info.tm = tm;
434 pmd->info.levels = 2;
435 pmd->info.value_type.context = pmd->data_sm;
436 pmd->info.value_type.size = sizeof(__le64);
437 pmd->info.value_type.inc = data_block_inc;
438 pmd->info.value_type.dec = data_block_dec;
439 pmd->info.value_type.equal = data_block_equal;
440
441 memcpy(&pmd->nb_info, &pmd->info, sizeof(pmd->nb_info));
442 pmd->nb_info.tm = pmd->nb_tm;
443
444 pmd->tl_info.tm = tm;
445 pmd->tl_info.levels = 1;
446 pmd->tl_info.value_type.context = &pmd->info;
447 pmd->tl_info.value_type.size = sizeof(__le64);
448 pmd->tl_info.value_type.inc = subtree_inc;
449 pmd->tl_info.value_type.dec = subtree_dec;
450 pmd->tl_info.value_type.equal = subtree_equal;
451
452 pmd->bl_info.tm = tm;
453 pmd->bl_info.levels = 1;
454 pmd->bl_info.value_type.context = pmd->data_sm;
455 pmd->bl_info.value_type.size = sizeof(__le64);
456 pmd->bl_info.value_type.inc = data_block_inc;
457 pmd->bl_info.value_type.dec = data_block_dec;
458 pmd->bl_info.value_type.equal = data_block_equal;
459
460 pmd->details_info.tm = tm;
461 pmd->details_info.levels = 1;
462 pmd->details_info.value_type.context = NULL;
463 pmd->details_info.value_type.size = sizeof(struct disk_device_details);
464 pmd->details_info.value_type.inc = NULL;
465 pmd->details_info.value_type.dec = NULL;
466 pmd->details_info.value_type.equal = NULL;
467
468 pmd->root = 0;
469
470 init_rwsem(&pmd->root_lock);
471 pmd->time = 0;
472 pmd->need_commit = 0;
473 pmd->details_root = 0;
474 pmd->trans_id = 0;
475 pmd->flags = 0;
476 INIT_LIST_HEAD(&pmd->thin_devices);
477
478 return 0;
479
480bad_data_sm:
481 dm_sm_destroy(data_sm);
482bad:
483 dm_tm_destroy(tm);
484 dm_sm_destroy(sm);
485
486 return r;
487}
488
489static int __begin_transaction(struct dm_pool_metadata *pmd)
490{
491 int r;
492 u32 features;
493 struct thin_disk_superblock *disk_super;
494 struct dm_block *sblock;
495
496 /*
497 * __maybe_commit_transaction() resets these
498 */
499 WARN_ON(pmd->need_commit);
500
501 /*
502 * We re-read the superblock every time. Shouldn't need to do this
503 * really.
504 */
505 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
506 &sb_validator, &sblock);
507 if (r)
508 return r;
509
510 disk_super = dm_block_data(sblock);
511 pmd->time = le32_to_cpu(disk_super->time);
512 pmd->root = le64_to_cpu(disk_super->data_mapping_root);
513 pmd->details_root = le64_to_cpu(disk_super->device_details_root);
514 pmd->trans_id = le64_to_cpu(disk_super->trans_id);
515 pmd->flags = le32_to_cpu(disk_super->flags);
516 pmd->data_block_size = le32_to_cpu(disk_super->data_block_size);
517
518 features = le32_to_cpu(disk_super->incompat_flags) & ~THIN_FEATURE_INCOMPAT_SUPP;
519 if (features) {
520 DMERR("could not access metadata due to "
521 "unsupported optional features (%lx).",
522 (unsigned long)features);
523 r = -EINVAL;
524 goto out;
525 }
526
527 /*
528 * Check for read-only metadata to skip the following RDWR checks.
529 */
530 if (get_disk_ro(pmd->bdev->bd_disk))
531 goto out;
532
533 features = le32_to_cpu(disk_super->compat_ro_flags) & ~THIN_FEATURE_COMPAT_RO_SUPP;
534 if (features) {
535 DMERR("could not access metadata RDWR due to "
536 "unsupported optional features (%lx).",
537 (unsigned long)features);
538 r = -EINVAL;
539 }
540
541out:
542 dm_bm_unlock(sblock);
543 return r;
544}
545
546static int __write_changed_details(struct dm_pool_metadata *pmd)
547{
548 int r;
549 struct dm_thin_device *td, *tmp;
550 struct disk_device_details details;
551 uint64_t key;
552
553 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
554 if (!td->changed)
555 continue;
556
557 key = td->id;
558
559 details.mapped_blocks = cpu_to_le64(td->mapped_blocks);
560 details.transaction_id = cpu_to_le64(td->transaction_id);
561 details.creation_time = cpu_to_le32(td->creation_time);
562 details.snapshotted_time = cpu_to_le32(td->snapshotted_time);
563 __dm_bless_for_disk(&details);
564
565 r = dm_btree_insert(&pmd->details_info, pmd->details_root,
566 &key, &details, &pmd->details_root);
567 if (r)
568 return r;
569
570 if (td->open_count)
571 td->changed = 0;
572 else {
573 list_del(&td->list);
574 kfree(td);
575 }
576
577 pmd->need_commit = 1;
578 }
579
580 return 0;
581}
582
583static int __commit_transaction(struct dm_pool_metadata *pmd)
584{
585 /*
586 * FIXME: Associated pool should be made read-only on failure.
587 */
588 int r;
589 size_t metadata_len, data_len;
590 struct thin_disk_superblock *disk_super;
591 struct dm_block *sblock;
592
593 /*
594 * We need to know if the thin_disk_superblock exceeds a 512-byte sector.
595 */
596 BUILD_BUG_ON(sizeof(struct thin_disk_superblock) > 512);
597
598 r = __write_changed_details(pmd);
599 if (r < 0)
600 goto out;
601
602 if (!pmd->need_commit)
603 goto out;
604
605 r = dm_sm_commit(pmd->data_sm);
606 if (r < 0)
607 goto out;
608
609 r = dm_tm_pre_commit(pmd->tm);
610 if (r < 0)
611 goto out;
612
613 r = dm_sm_root_size(pmd->metadata_sm, &metadata_len);
614 if (r < 0)
615 goto out;
616
617 r = dm_sm_root_size(pmd->data_sm, &data_len);
618 if (r < 0)
619 goto out;
620
621 r = dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
622 &sb_validator, &sblock);
623 if (r)
624 goto out;
625
626 disk_super = dm_block_data(sblock);
627 disk_super->time = cpu_to_le32(pmd->time);
628 disk_super->data_mapping_root = cpu_to_le64(pmd->root);
629 disk_super->device_details_root = cpu_to_le64(pmd->details_root);
630 disk_super->trans_id = cpu_to_le64(pmd->trans_id);
631 disk_super->flags = cpu_to_le32(pmd->flags);
632
633 r = dm_sm_copy_root(pmd->metadata_sm, &disk_super->metadata_space_map_root,
634 metadata_len);
635 if (r < 0)
636 goto out_locked;
637
638 r = dm_sm_copy_root(pmd->data_sm, &disk_super->data_space_map_root,
639 data_len);
640 if (r < 0)
641 goto out_locked;
642
643 r = dm_tm_commit(pmd->tm, sblock);
644 if (!r)
645 pmd->need_commit = 0;
646
647out:
648 return r;
649
650out_locked:
651 dm_bm_unlock(sblock);
652 return r;
653}
654
655struct dm_pool_metadata *dm_pool_metadata_open(struct block_device *bdev,
656 sector_t data_block_size)
657{
658 int r;
659 struct thin_disk_superblock *disk_super;
660 struct dm_pool_metadata *pmd;
661 sector_t bdev_size = i_size_read(bdev->bd_inode) >> SECTOR_SHIFT;
662 struct dm_block_manager *bm;
663 int create;
664 struct dm_block *sblock;
665
666 pmd = kmalloc(sizeof(*pmd), GFP_KERNEL);
667 if (!pmd) {
668 DMERR("could not allocate metadata struct");
669 return ERR_PTR(-ENOMEM);
670 }
671
672 /*
673 * Max hex locks:
674 * 3 for btree insert +
675 * 2 for btree lookup used within space map
676 */
677 bm = dm_block_manager_create(bdev, THIN_METADATA_BLOCK_SIZE,
678 THIN_METADATA_CACHE_SIZE, 5);
679 if (!bm) {
680 DMERR("could not create block manager");
681 kfree(pmd);
682 return ERR_PTR(-ENOMEM);
683 }
684
685 r = superblock_all_zeroes(bm, &create);
686 if (r) {
687 dm_block_manager_destroy(bm);
688 kfree(pmd);
689 return ERR_PTR(r);
690 }
691
692
693 r = init_pmd(pmd, bm, 0, create);
694 if (r) {
695 dm_block_manager_destroy(bm);
696 kfree(pmd);
697 return ERR_PTR(r);
698 }
699 pmd->bdev = bdev;
700
701 if (!create) {
702 r = __begin_transaction(pmd);
703 if (r < 0)
704 goto bad;
705 return pmd;
706 }
707
708 /*
709 * Create.
710 */
711 r = dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
712 &sb_validator, &sblock);
713 if (r)
714 goto bad;
715
716 if (bdev_size > THIN_METADATA_MAX_SECTORS)
717 bdev_size = THIN_METADATA_MAX_SECTORS;
718
719 disk_super = dm_block_data(sblock);
720 disk_super->magic = cpu_to_le64(THIN_SUPERBLOCK_MAGIC);
721 disk_super->version = cpu_to_le32(THIN_VERSION);
722 disk_super->time = 0;
723 disk_super->metadata_block_size = cpu_to_le32(THIN_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
724 disk_super->metadata_nr_blocks = cpu_to_le64(bdev_size >> SECTOR_TO_BLOCK_SHIFT);
725 disk_super->data_block_size = cpu_to_le32(data_block_size);
726
727 r = dm_bm_unlock(sblock);
728 if (r < 0)
729 goto bad;
730
731 r = dm_btree_empty(&pmd->info, &pmd->root);
732 if (r < 0)
733 goto bad;
734
735 r = dm_btree_empty(&pmd->details_info, &pmd->details_root);
736 if (r < 0) {
737 DMERR("couldn't create devices root");
738 goto bad;
739 }
740
741 pmd->flags = 0;
742 pmd->need_commit = 1;
743 r = dm_pool_commit_metadata(pmd);
744 if (r < 0) {
745 DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
746 __func__, r);
747 goto bad;
748 }
749
750 return pmd;
751
752bad:
753 if (dm_pool_metadata_close(pmd) < 0)
754 DMWARN("%s: dm_pool_metadata_close() failed.", __func__);
755 return ERR_PTR(r);
756}
757
758int dm_pool_metadata_close(struct dm_pool_metadata *pmd)
759{
760 int r;
761 unsigned open_devices = 0;
762 struct dm_thin_device *td, *tmp;
763
764 down_read(&pmd->root_lock);
765 list_for_each_entry_safe(td, tmp, &pmd->thin_devices, list) {
766 if (td->open_count)
767 open_devices++;
768 else {
769 list_del(&td->list);
770 kfree(td);
771 }
772 }
773 up_read(&pmd->root_lock);
774
775 if (open_devices) {
776 DMERR("attempt to close pmd when %u device(s) are still open",
777 open_devices);
778 return -EBUSY;
779 }
780
781 r = __commit_transaction(pmd);
782 if (r < 0)
783 DMWARN("%s: __commit_transaction() failed, error = %d",
784 __func__, r);
785
786 dm_tm_destroy(pmd->tm);
787 dm_tm_destroy(pmd->nb_tm);
788 dm_block_manager_destroy(pmd->bm);
789 dm_sm_destroy(pmd->metadata_sm);
790 dm_sm_destroy(pmd->data_sm);
791 kfree(pmd);
792
793 return 0;
794}
795
796/*
797 * __open_device: Returns @td corresponding to device with id @dev,
798 * creating it if @create is set and incrementing @td->open_count.
799 * On failure, @td is undefined.
800 */
801static int __open_device(struct dm_pool_metadata *pmd,
802 dm_thin_id dev, int create,
803 struct dm_thin_device **td)
804{
805 int r, changed = 0;
806 struct dm_thin_device *td2;
807 uint64_t key = dev;
808 struct disk_device_details details_le;
809
810 /*
811 * If the device is already open, return it.
812 */
813 list_for_each_entry(td2, &pmd->thin_devices, list)
814 if (td2->id == dev) {
815 /*
816 * May not create an already-open device.
817 */
818 if (create)
819 return -EEXIST;
820
821 td2->open_count++;
822 *td = td2;
823 return 0;
824 }
825
826 /*
827 * Check the device exists.
828 */
829 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
830 &key, &details_le);
831 if (r) {
832 if (r != -ENODATA || !create)
833 return r;
834
835 /*
836 * Create new device.
837 */
838 changed = 1;
839 details_le.mapped_blocks = 0;
840 details_le.transaction_id = cpu_to_le64(pmd->trans_id);
841 details_le.creation_time = cpu_to_le32(pmd->time);
842 details_le.snapshotted_time = cpu_to_le32(pmd->time);
843 }
844
845 *td = kmalloc(sizeof(**td), GFP_NOIO);
846 if (!*td)
847 return -ENOMEM;
848
849 (*td)->pmd = pmd;
850 (*td)->id = dev;
851 (*td)->open_count = 1;
852 (*td)->changed = changed;
853 (*td)->mapped_blocks = le64_to_cpu(details_le.mapped_blocks);
854 (*td)->transaction_id = le64_to_cpu(details_le.transaction_id);
855 (*td)->creation_time = le32_to_cpu(details_le.creation_time);
856 (*td)->snapshotted_time = le32_to_cpu(details_le.snapshotted_time);
857
858 list_add(&(*td)->list, &pmd->thin_devices);
859
860 return 0;
861}
862
863static void __close_device(struct dm_thin_device *td)
864{
865 --td->open_count;
866}
867
868static int __create_thin(struct dm_pool_metadata *pmd,
869 dm_thin_id dev)
870{
871 int r;
872 dm_block_t dev_root;
873 uint64_t key = dev;
874 struct disk_device_details details_le;
875 struct dm_thin_device *td;
876 __le64 value;
877
878 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
879 &key, &details_le);
880 if (!r)
881 return -EEXIST;
882
883 /*
884 * Create an empty btree for the mappings.
885 */
886 r = dm_btree_empty(&pmd->bl_info, &dev_root);
887 if (r)
888 return r;
889
890 /*
891 * Insert it into the main mapping tree.
892 */
893 value = cpu_to_le64(dev_root);
894 __dm_bless_for_disk(&value);
895 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
896 if (r) {
897 dm_btree_del(&pmd->bl_info, dev_root);
898 return r;
899 }
900
901 r = __open_device(pmd, dev, 1, &td);
902 if (r) {
903 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
904 dm_btree_del(&pmd->bl_info, dev_root);
905 return r;
906 }
907 __close_device(td);
908
909 return r;
910}
911
912int dm_pool_create_thin(struct dm_pool_metadata *pmd, dm_thin_id dev)
913{
914 int r;
915
916 down_write(&pmd->root_lock);
917 r = __create_thin(pmd, dev);
918 up_write(&pmd->root_lock);
919
920 return r;
921}
922
923static int __set_snapshot_details(struct dm_pool_metadata *pmd,
924 struct dm_thin_device *snap,
925 dm_thin_id origin, uint32_t time)
926{
927 int r;
928 struct dm_thin_device *td;
929
930 r = __open_device(pmd, origin, 0, &td);
931 if (r)
932 return r;
933
934 td->changed = 1;
935 td->snapshotted_time = time;
936
937 snap->mapped_blocks = td->mapped_blocks;
938 snap->snapshotted_time = time;
939 __close_device(td);
940
941 return 0;
942}
943
944static int __create_snap(struct dm_pool_metadata *pmd,
945 dm_thin_id dev, dm_thin_id origin)
946{
947 int r;
948 dm_block_t origin_root;
949 uint64_t key = origin, dev_key = dev;
950 struct dm_thin_device *td;
951 struct disk_device_details details_le;
952 __le64 value;
953
954 /* check this device is unused */
955 r = dm_btree_lookup(&pmd->details_info, pmd->details_root,
956 &dev_key, &details_le);
957 if (!r)
958 return -EEXIST;
959
960 /* find the mapping tree for the origin */
961 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &key, &value);
962 if (r)
963 return r;
964 origin_root = le64_to_cpu(value);
965
966 /* clone the origin, an inc will do */
967 dm_tm_inc(pmd->tm, origin_root);
968
969 /* insert into the main mapping tree */
970 value = cpu_to_le64(origin_root);
971 __dm_bless_for_disk(&value);
972 key = dev;
973 r = dm_btree_insert(&pmd->tl_info, pmd->root, &key, &value, &pmd->root);
974 if (r) {
975 dm_tm_dec(pmd->tm, origin_root);
976 return r;
977 }
978
979 pmd->time++;
980
981 r = __open_device(pmd, dev, 1, &td);
982 if (r)
983 goto bad;
984
985 r = __set_snapshot_details(pmd, td, origin, pmd->time);
986 __close_device(td);
987
988 if (r)
989 goto bad;
990
991 return 0;
992
993bad:
994 dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
995 dm_btree_remove(&pmd->details_info, pmd->details_root,
996 &key, &pmd->details_root);
997 return r;
998}
999
1000int dm_pool_create_snap(struct dm_pool_metadata *pmd,
1001 dm_thin_id dev,
1002 dm_thin_id origin)
1003{
1004 int r;
1005
1006 down_write(&pmd->root_lock);
1007 r = __create_snap(pmd, dev, origin);
1008 up_write(&pmd->root_lock);
1009
1010 return r;
1011}
1012
1013static int __delete_device(struct dm_pool_metadata *pmd, dm_thin_id dev)
1014{
1015 int r;
1016 uint64_t key = dev;
1017 struct dm_thin_device *td;
1018
1019 /* TODO: failure should mark the transaction invalid */
1020 r = __open_device(pmd, dev, 0, &td);
1021 if (r)
1022 return r;
1023
1024 if (td->open_count > 1) {
1025 __close_device(td);
1026 return -EBUSY;
1027 }
1028
1029 list_del(&td->list);
1030 kfree(td);
1031 r = dm_btree_remove(&pmd->details_info, pmd->details_root,
1032 &key, &pmd->details_root);
1033 if (r)
1034 return r;
1035
1036 r = dm_btree_remove(&pmd->tl_info, pmd->root, &key, &pmd->root);
1037 if (r)
1038 return r;
1039
1040 pmd->need_commit = 1;
1041
1042 return 0;
1043}
1044
1045int dm_pool_delete_thin_device(struct dm_pool_metadata *pmd,
1046 dm_thin_id dev)
1047{
1048 int r;
1049
1050 down_write(&pmd->root_lock);
1051 r = __delete_device(pmd, dev);
1052 up_write(&pmd->root_lock);
1053
1054 return r;
1055}
1056
1057int dm_pool_set_metadata_transaction_id(struct dm_pool_metadata *pmd,
1058 uint64_t current_id,
1059 uint64_t new_id)
1060{
1061 down_write(&pmd->root_lock);
1062 if (pmd->trans_id != current_id) {
1063 up_write(&pmd->root_lock);
1064 DMERR("mismatched transaction id");
1065 return -EINVAL;
1066 }
1067
1068 pmd->trans_id = new_id;
1069 pmd->need_commit = 1;
1070 up_write(&pmd->root_lock);
1071
1072 return 0;
1073}
1074
1075int dm_pool_get_metadata_transaction_id(struct dm_pool_metadata *pmd,
1076 uint64_t *result)
1077{
1078 down_read(&pmd->root_lock);
1079 *result = pmd->trans_id;
1080 up_read(&pmd->root_lock);
1081
1082 return 0;
1083}
1084
1085static int __reserve_metadata_snap(struct dm_pool_metadata *pmd)
1086{
1087 int r, inc;
1088 struct thin_disk_superblock *disk_super;
1089 struct dm_block *copy, *sblock;
1090 dm_block_t held_root;
1091
1092 /*
1093 * Copy the superblock.
1094 */
1095 dm_sm_inc_block(pmd->metadata_sm, THIN_SUPERBLOCK_LOCATION);
1096 r = dm_tm_shadow_block(pmd->tm, THIN_SUPERBLOCK_LOCATION,
1097 &sb_validator, ©, &inc);
1098 if (r)
1099 return r;
1100
1101 BUG_ON(!inc);
1102
1103 held_root = dm_block_location(copy);
1104 disk_super = dm_block_data(copy);
1105
1106 if (le64_to_cpu(disk_super->held_root)) {
1107 DMWARN("Pool metadata snapshot already exists: release this before taking another.");
1108
1109 dm_tm_dec(pmd->tm, held_root);
1110 dm_tm_unlock(pmd->tm, copy);
1111 pmd->need_commit = 1;
1112
1113 return -EBUSY;
1114 }
1115
1116 /*
1117 * Wipe the spacemap since we're not publishing this.
1118 */
1119 memset(&disk_super->data_space_map_root, 0,
1120 sizeof(disk_super->data_space_map_root));
1121 memset(&disk_super->metadata_space_map_root, 0,
1122 sizeof(disk_super->metadata_space_map_root));
1123
1124 /*
1125 * Increment the data structures that need to be preserved.
1126 */
1127 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->data_mapping_root));
1128 dm_tm_inc(pmd->tm, le64_to_cpu(disk_super->device_details_root));
1129 dm_tm_unlock(pmd->tm, copy);
1130
1131 /*
1132 * Write the held root into the superblock.
1133 */
1134 r = dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1135 &sb_validator, &sblock);
1136 if (r) {
1137 dm_tm_dec(pmd->tm, held_root);
1138 pmd->need_commit = 1;
1139 return r;
1140 }
1141
1142 disk_super = dm_block_data(sblock);
1143 disk_super->held_root = cpu_to_le64(held_root);
1144 dm_bm_unlock(sblock);
1145
1146 pmd->need_commit = 1;
1147
1148 return 0;
1149}
1150
1151int dm_pool_reserve_metadata_snap(struct dm_pool_metadata *pmd)
1152{
1153 int r;
1154
1155 down_write(&pmd->root_lock);
1156 r = __reserve_metadata_snap(pmd);
1157 up_write(&pmd->root_lock);
1158
1159 return r;
1160}
1161
1162static int __release_metadata_snap(struct dm_pool_metadata *pmd)
1163{
1164 int r;
1165 struct thin_disk_superblock *disk_super;
1166 struct dm_block *sblock, *copy;
1167 dm_block_t held_root;
1168
1169 r = dm_bm_write_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1170 &sb_validator, &sblock);
1171 if (r)
1172 return r;
1173
1174 disk_super = dm_block_data(sblock);
1175 held_root = le64_to_cpu(disk_super->held_root);
1176 disk_super->held_root = cpu_to_le64(0);
1177 pmd->need_commit = 1;
1178
1179 dm_bm_unlock(sblock);
1180
1181 if (!held_root) {
1182 DMWARN("No pool metadata snapshot found: nothing to release.");
1183 return -EINVAL;
1184 }
1185
1186 r = dm_tm_read_lock(pmd->tm, held_root, &sb_validator, ©);
1187 if (r)
1188 return r;
1189
1190 disk_super = dm_block_data(copy);
1191 dm_sm_dec_block(pmd->metadata_sm, le64_to_cpu(disk_super->data_mapping_root));
1192 dm_sm_dec_block(pmd->metadata_sm, le64_to_cpu(disk_super->device_details_root));
1193 dm_sm_dec_block(pmd->metadata_sm, held_root);
1194
1195 return dm_tm_unlock(pmd->tm, copy);
1196}
1197
1198int dm_pool_release_metadata_snap(struct dm_pool_metadata *pmd)
1199{
1200 int r;
1201
1202 down_write(&pmd->root_lock);
1203 r = __release_metadata_snap(pmd);
1204 up_write(&pmd->root_lock);
1205
1206 return r;
1207}
1208
1209static int __get_metadata_snap(struct dm_pool_metadata *pmd,
1210 dm_block_t *result)
1211{
1212 int r;
1213 struct thin_disk_superblock *disk_super;
1214 struct dm_block *sblock;
1215
1216 r = dm_bm_read_lock(pmd->bm, THIN_SUPERBLOCK_LOCATION,
1217 &sb_validator, &sblock);
1218 if (r)
1219 return r;
1220
1221 disk_super = dm_block_data(sblock);
1222 *result = le64_to_cpu(disk_super->held_root);
1223
1224 return dm_bm_unlock(sblock);
1225}
1226
1227int dm_pool_get_metadata_snap(struct dm_pool_metadata *pmd,
1228 dm_block_t *result)
1229{
1230 int r;
1231
1232 down_read(&pmd->root_lock);
1233 r = __get_metadata_snap(pmd, result);
1234 up_read(&pmd->root_lock);
1235
1236 return r;
1237}
1238
1239int dm_pool_open_thin_device(struct dm_pool_metadata *pmd, dm_thin_id dev,
1240 struct dm_thin_device **td)
1241{
1242 int r;
1243
1244 down_write(&pmd->root_lock);
1245 r = __open_device(pmd, dev, 0, td);
1246 up_write(&pmd->root_lock);
1247
1248 return r;
1249}
1250
1251int dm_pool_close_thin_device(struct dm_thin_device *td)
1252{
1253 down_write(&td->pmd->root_lock);
1254 __close_device(td);
1255 up_write(&td->pmd->root_lock);
1256
1257 return 0;
1258}
1259
1260dm_thin_id dm_thin_dev_id(struct dm_thin_device *td)
1261{
1262 return td->id;
1263}
1264
1265static int __snapshotted_since(struct dm_thin_device *td, uint32_t time)
1266{
1267 return td->snapshotted_time > time;
1268}
1269
1270int dm_thin_find_block(struct dm_thin_device *td, dm_block_t block,
1271 int can_block, struct dm_thin_lookup_result *result)
1272{
1273 int r;
1274 uint64_t block_time = 0;
1275 __le64 value;
1276 struct dm_pool_metadata *pmd = td->pmd;
1277 dm_block_t keys[2] = { td->id, block };
1278
1279 if (can_block) {
1280 down_read(&pmd->root_lock);
1281 r = dm_btree_lookup(&pmd->info, pmd->root, keys, &value);
1282 if (!r)
1283 block_time = le64_to_cpu(value);
1284 up_read(&pmd->root_lock);
1285
1286 } else if (down_read_trylock(&pmd->root_lock)) {
1287 r = dm_btree_lookup(&pmd->nb_info, pmd->root, keys, &value);
1288 if (!r)
1289 block_time = le64_to_cpu(value);
1290 up_read(&pmd->root_lock);
1291
1292 } else
1293 return -EWOULDBLOCK;
1294
1295 if (!r) {
1296 dm_block_t exception_block;
1297 uint32_t exception_time;
1298 unpack_block_time(block_time, &exception_block,
1299 &exception_time);
1300 result->block = exception_block;
1301 result->shared = __snapshotted_since(td, exception_time);
1302 }
1303
1304 return r;
1305}
1306
1307static int __insert(struct dm_thin_device *td, dm_block_t block,
1308 dm_block_t data_block)
1309{
1310 int r, inserted;
1311 __le64 value;
1312 struct dm_pool_metadata *pmd = td->pmd;
1313 dm_block_t keys[2] = { td->id, block };
1314
1315 pmd->need_commit = 1;
1316 value = cpu_to_le64(pack_block_time(data_block, pmd->time));
1317 __dm_bless_for_disk(&value);
1318
1319 r = dm_btree_insert_notify(&pmd->info, pmd->root, keys, &value,
1320 &pmd->root, &inserted);
1321 if (r)
1322 return r;
1323
1324 if (inserted) {
1325 td->mapped_blocks++;
1326 td->changed = 1;
1327 }
1328
1329 return 0;
1330}
1331
1332int dm_thin_insert_block(struct dm_thin_device *td, dm_block_t block,
1333 dm_block_t data_block)
1334{
1335 int r;
1336
1337 down_write(&td->pmd->root_lock);
1338 r = __insert(td, block, data_block);
1339 up_write(&td->pmd->root_lock);
1340
1341 return r;
1342}
1343
1344static int __remove(struct dm_thin_device *td, dm_block_t block)
1345{
1346 int r;
1347 struct dm_pool_metadata *pmd = td->pmd;
1348 dm_block_t keys[2] = { td->id, block };
1349
1350 r = dm_btree_remove(&pmd->info, pmd->root, keys, &pmd->root);
1351 if (r)
1352 return r;
1353
1354 td->mapped_blocks--;
1355 td->changed = 1;
1356 pmd->need_commit = 1;
1357
1358 return 0;
1359}
1360
1361int dm_thin_remove_block(struct dm_thin_device *td, dm_block_t block)
1362{
1363 int r;
1364
1365 down_write(&td->pmd->root_lock);
1366 r = __remove(td, block);
1367 up_write(&td->pmd->root_lock);
1368
1369 return r;
1370}
1371
1372int dm_pool_alloc_data_block(struct dm_pool_metadata *pmd, dm_block_t *result)
1373{
1374 int r;
1375
1376 down_write(&pmd->root_lock);
1377
1378 r = dm_sm_new_block(pmd->data_sm, result);
1379 pmd->need_commit = 1;
1380
1381 up_write(&pmd->root_lock);
1382
1383 return r;
1384}
1385
1386int dm_pool_commit_metadata(struct dm_pool_metadata *pmd)
1387{
1388 int r;
1389
1390 down_write(&pmd->root_lock);
1391
1392 r = __commit_transaction(pmd);
1393 if (r <= 0)
1394 goto out;
1395
1396 /*
1397 * Open the next transaction.
1398 */
1399 r = __begin_transaction(pmd);
1400out:
1401 up_write(&pmd->root_lock);
1402 return r;
1403}
1404
1405int dm_pool_get_free_block_count(struct dm_pool_metadata *pmd, dm_block_t *result)
1406{
1407 int r;
1408
1409 down_read(&pmd->root_lock);
1410 r = dm_sm_get_nr_free(pmd->data_sm, result);
1411 up_read(&pmd->root_lock);
1412
1413 return r;
1414}
1415
1416int dm_pool_get_free_metadata_block_count(struct dm_pool_metadata *pmd,
1417 dm_block_t *result)
1418{
1419 int r;
1420
1421 down_read(&pmd->root_lock);
1422 r = dm_sm_get_nr_free(pmd->metadata_sm, result);
1423 up_read(&pmd->root_lock);
1424
1425 return r;
1426}
1427
1428int dm_pool_get_metadata_dev_size(struct dm_pool_metadata *pmd,
1429 dm_block_t *result)
1430{
1431 int r;
1432
1433 down_read(&pmd->root_lock);
1434 r = dm_sm_get_nr_blocks(pmd->metadata_sm, result);
1435 up_read(&pmd->root_lock);
1436
1437 return r;
1438}
1439
1440int dm_pool_get_data_block_size(struct dm_pool_metadata *pmd, sector_t *result)
1441{
1442 down_read(&pmd->root_lock);
1443 *result = pmd->data_block_size;
1444 up_read(&pmd->root_lock);
1445
1446 return 0;
1447}
1448
1449int dm_pool_get_data_dev_size(struct dm_pool_metadata *pmd, dm_block_t *result)
1450{
1451 int r;
1452
1453 down_read(&pmd->root_lock);
1454 r = dm_sm_get_nr_blocks(pmd->data_sm, result);
1455 up_read(&pmd->root_lock);
1456
1457 return r;
1458}
1459
1460int dm_thin_get_mapped_count(struct dm_thin_device *td, dm_block_t *result)
1461{
1462 struct dm_pool_metadata *pmd = td->pmd;
1463
1464 down_read(&pmd->root_lock);
1465 *result = td->mapped_blocks;
1466 up_read(&pmd->root_lock);
1467
1468 return 0;
1469}
1470
1471static int __highest_block(struct dm_thin_device *td, dm_block_t *result)
1472{
1473 int r;
1474 __le64 value_le;
1475 dm_block_t thin_root;
1476 struct dm_pool_metadata *pmd = td->pmd;
1477
1478 r = dm_btree_lookup(&pmd->tl_info, pmd->root, &td->id, &value_le);
1479 if (r)
1480 return r;
1481
1482 thin_root = le64_to_cpu(value_le);
1483
1484 return dm_btree_find_highest_key(&pmd->bl_info, thin_root, result);
1485}
1486
1487int dm_thin_get_highest_mapped_block(struct dm_thin_device *td,
1488 dm_block_t *result)
1489{
1490 int r;
1491 struct dm_pool_metadata *pmd = td->pmd;
1492
1493 down_read(&pmd->root_lock);
1494 r = __highest_block(td, result);
1495 up_read(&pmd->root_lock);
1496
1497 return r;
1498}
1499
1500static int __resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1501{
1502 int r;
1503 dm_block_t old_count;
1504
1505 r = dm_sm_get_nr_blocks(pmd->data_sm, &old_count);
1506 if (r)
1507 return r;
1508
1509 if (new_count == old_count)
1510 return 0;
1511
1512 if (new_count < old_count) {
1513 DMERR("cannot reduce size of data device");
1514 return -EINVAL;
1515 }
1516
1517 r = dm_sm_extend(pmd->data_sm, new_count - old_count);
1518 if (!r)
1519 pmd->need_commit = 1;
1520
1521 return r;
1522}
1523
1524int dm_pool_resize_data_dev(struct dm_pool_metadata *pmd, dm_block_t new_count)
1525{
1526 int r;
1527
1528 down_write(&pmd->root_lock);
1529 r = __resize_data_dev(pmd, new_count);
1530 up_write(&pmd->root_lock);
1531
1532 return r;
1533}