1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2019 Arrikto, Inc. All Rights Reserved.
4  */
5 
6 #include <linux/mm.h>
7 #include <linux/err.h>
8 #include <linux/slab.h>
9 #include <linux/rwsem.h>
10 #include <linux/bitops.h>
11 #include <linux/bitmap.h>
12 #include <linux/device-mapper.h>
13 
14 #include "persistent-data/dm-bitset.h"
15 #include "persistent-data/dm-space-map.h"
16 #include "persistent-data/dm-block-manager.h"
17 #include "persistent-data/dm-transaction-manager.h"
18 
19 #include "dm-clone-metadata.h"
20 
21 #define DM_MSG_PREFIX "clone metadata"
22 
23 #define SUPERBLOCK_LOCATION 0
24 #define SUPERBLOCK_MAGIC 0x8af27f64
25 #define SUPERBLOCK_CSUM_XOR 257649492
26 
27 #define DM_CLONE_MAX_CONCURRENT_LOCKS 5
28 
29 #define UUID_LEN 16
30 
31 /* Min and max dm-clone metadata versions supported */
32 #define DM_CLONE_MIN_METADATA_VERSION 1
33 #define DM_CLONE_MAX_METADATA_VERSION 1
34 
35 /*
36  * On-disk metadata layout
37  */
38 struct superblock_disk {
39 	__le32 csum;
40 	__le32 flags;
41 	__le64 blocknr;
42 
43 	__u8 uuid[UUID_LEN];
44 	__le64 magic;
45 	__le32 version;
46 
47 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
48 
49 	__le64 region_size;
50 	__le64 target_size;
51 
52 	__le64 bitset_root;
53 } __packed;
54 
55 /*
56  * Region and Dirty bitmaps.
57  *
58  * dm-clone logically splits the source and destination devices in regions of
59  * fixed size. The destination device's regions are gradually hydrated, i.e.,
60  * we copy (clone) the source's regions to the destination device. Eventually,
61  * all regions will get hydrated and all I/O will be served from the
62  * destination device.
63  *
64  * We maintain an on-disk bitmap which tracks the state of each of the
65  * destination device's regions, i.e., whether they are hydrated or not.
66  *
67  * To save constantly doing look ups on disk we keep an in core copy of the
68  * on-disk bitmap, the region_map.
69  *
70  * In order to track which regions are hydrated during a metadata transaction,
71  * we use a second set of bitmaps, the dmap (dirty bitmap), which includes two
72  * bitmaps, namely dirty_regions and dirty_words. The dirty_regions bitmap
73  * tracks the regions that got hydrated during the current metadata
74  * transaction. The dirty_words bitmap tracks the dirty words, i.e. longs, of
75  * the dirty_regions bitmap.
76  *
77  * This allows us to precisely track the regions that were hydrated during the
78  * current metadata transaction and update the metadata accordingly, when we
79  * commit the current transaction. This is important because dm-clone should
80  * only commit the metadata of regions that were properly flushed to the
81  * destination device beforehand. Otherwise, in case of a crash, we could end
82  * up with a corrupted dm-clone device.
83  *
84  * When a region finishes hydrating dm-clone calls
85  * dm_clone_set_region_hydrated(), or for discard requests
86  * dm_clone_cond_set_range(), which sets the corresponding bits in region_map
87  * and dmap.
88  *
89  * During a metadata commit we scan dmap->dirty_words and dmap->dirty_regions
90  * and update the on-disk metadata accordingly. Thus, we don't have to flush to
91  * disk the whole region_map. We can just flush the dirty region_map bits.
92  *
93  * We use the helper dmap->dirty_words bitmap, which is smaller than the
94  * original region_map, to reduce the amount of memory accesses during a
95  * metadata commit. Moreover, as dm-bitset also accesses the on-disk bitmap in
96  * 64-bit word granularity, the dirty_words bitmap helps us avoid useless disk
97  * accesses.
98  *
99  * We could update directly the on-disk bitmap, when dm-clone calls either
100  * dm_clone_set_region_hydrated() or dm_clone_cond_set_range(), buts this
101  * inserts significant metadata I/O overhead in dm-clone's I/O path. Also, as
102  * these two functions don't block, we can call them in interrupt context,
103  * e.g., in a hooked overwrite bio's completion routine, and further reduce the
104  * I/O completion latency.
105  *
106  * We maintain two dirty bitmap sets. During a metadata commit we atomically
107  * swap the currently used dmap with the unused one. This allows the metadata
108  * update functions to run concurrently with an ongoing commit.
109  */
110 struct dirty_map {
111 	unsigned long *dirty_words;
112 	unsigned long *dirty_regions;
113 	unsigned int changed;
114 };
115 
116 struct dm_clone_metadata {
117 	/* The metadata block device */
118 	struct block_device *bdev;
119 
120 	sector_t target_size;
121 	sector_t region_size;
122 	unsigned long nr_regions;
123 	unsigned long nr_words;
124 
125 	/* Spinlock protecting the region and dirty bitmaps. */
126 	spinlock_t bitmap_lock;
127 	struct dirty_map dmap[2];
128 	struct dirty_map *current_dmap;
129 
130 	/* Protected by lock */
131 	struct dirty_map *committing_dmap;
132 
133 	/*
134 	 * In core copy of the on-disk bitmap to save constantly doing look ups
135 	 * on disk.
136 	 */
137 	unsigned long *region_map;
138 
139 	/* Protected by bitmap_lock */
140 	unsigned int read_only;
141 
142 	struct dm_block_manager *bm;
143 	struct dm_space_map *sm;
144 	struct dm_transaction_manager *tm;
145 
146 	struct rw_semaphore lock;
147 
148 	struct dm_disk_bitset bitset_info;
149 	dm_block_t bitset_root;
150 
151 	/*
152 	 * Reading the space map root can fail, so we read it into this
153 	 * buffer before the superblock is locked and updated.
154 	 */
155 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
156 
157 	bool hydration_done:1;
158 	bool fail_io:1;
159 };
160 
161 /*---------------------------------------------------------------------------*/
162 
163 /*
164  * Superblock validation.
165  */
166 static void sb_prepare_for_write(struct dm_block_validator *v,
167 				 struct dm_block *b, size_t sb_block_size)
168 {
169 	struct superblock_disk *sb;
170 	u32 csum;
171 
172 	sb = dm_block_data(b);
173 	sb->blocknr = cpu_to_le64(dm_block_location(b));
174 
175 	csum = dm_bm_checksum(&sb->flags, sb_block_size - sizeof(__le32),
176 			      SUPERBLOCK_CSUM_XOR);
177 	sb->csum = cpu_to_le32(csum);
178 }
179 
180 static int sb_check(struct dm_block_validator *v, struct dm_block *b,
181 		    size_t sb_block_size)
182 {
183 	struct superblock_disk *sb;
184 	u32 csum, metadata_version;
185 
186 	sb = dm_block_data(b);
187 
188 	if (dm_block_location(b) != le64_to_cpu(sb->blocknr)) {
189 		DMERR("Superblock check failed: blocknr %llu, expected %llu",
190 		      le64_to_cpu(sb->blocknr),
191 		      (unsigned long long)dm_block_location(b));
192 		return -ENOTBLK;
193 	}
194 
195 	if (le64_to_cpu(sb->magic) != SUPERBLOCK_MAGIC) {
196 		DMERR("Superblock check failed: magic %llu, expected %llu",
197 		      le64_to_cpu(sb->magic),
198 		      (unsigned long long)SUPERBLOCK_MAGIC);
199 		return -EILSEQ;
200 	}
201 
202 	csum = dm_bm_checksum(&sb->flags, sb_block_size - sizeof(__le32),
203 			      SUPERBLOCK_CSUM_XOR);
204 	if (sb->csum != cpu_to_le32(csum)) {
205 		DMERR("Superblock check failed: checksum %u, expected %u",
206 		      csum, le32_to_cpu(sb->csum));
207 		return -EILSEQ;
208 	}
209 
210 	/* Check metadata version */
211 	metadata_version = le32_to_cpu(sb->version);
212 	if (metadata_version < DM_CLONE_MIN_METADATA_VERSION ||
213 	    metadata_version > DM_CLONE_MAX_METADATA_VERSION) {
214 		DMERR("Clone metadata version %u found, but only versions between %u and %u supported.",
215 		      metadata_version, DM_CLONE_MIN_METADATA_VERSION,
216 		      DM_CLONE_MAX_METADATA_VERSION);
217 		return -EINVAL;
218 	}
219 
220 	return 0;
221 }
222 
223 static struct dm_block_validator sb_validator = {
224 	.name = "superblock",
225 	.prepare_for_write = sb_prepare_for_write,
226 	.check = sb_check
227 };
228 
229 /*
230  * Check if the superblock is formatted or not. We consider the superblock to
231  * be formatted in case we find non-zero bytes in it.
232  */
233 static int __superblock_all_zeroes(struct dm_block_manager *bm, bool *formatted)
234 {
235 	int r;
236 	unsigned int i, nr_words;
237 	struct dm_block *sblock;
238 	__le64 *data_le, zero = cpu_to_le64(0);
239 
240 	/*
241 	 * We don't use a validator here because the superblock could be all
242 	 * zeroes.
243 	 */
244 	r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &sblock);
245 	if (r) {
246 		DMERR("Failed to read_lock superblock");
247 		return r;
248 	}
249 
250 	data_le = dm_block_data(sblock);
251 	*formatted = false;
252 
253 	/* This assumes that the block size is a multiple of 8 bytes */
254 	BUG_ON(dm_bm_block_size(bm) % sizeof(__le64));
255 	nr_words = dm_bm_block_size(bm) / sizeof(__le64);
256 	for (i = 0; i < nr_words; i++) {
257 		if (data_le[i] != zero) {
258 			*formatted = true;
259 			break;
260 		}
261 	}
262 
263 	dm_bm_unlock(sblock);
264 
265 	return 0;
266 }
267 
268 /*---------------------------------------------------------------------------*/
269 
270 /*
271  * Low-level metadata handling.
272  */
273 static inline int superblock_read_lock(struct dm_clone_metadata *cmd,
274 				       struct dm_block **sblock)
275 {
276 	return dm_bm_read_lock(cmd->bm, SUPERBLOCK_LOCATION, &sb_validator, sblock);
277 }
278 
279 static inline int superblock_write_lock_zero(struct dm_clone_metadata *cmd,
280 					     struct dm_block **sblock)
281 {
282 	return dm_bm_write_lock_zero(cmd->bm, SUPERBLOCK_LOCATION, &sb_validator, sblock);
283 }
284 
285 static int __copy_sm_root(struct dm_clone_metadata *cmd)
286 {
287 	int r;
288 	size_t root_size;
289 
290 	r = dm_sm_root_size(cmd->sm, &root_size);
291 	if (r)
292 		return r;
293 
294 	return dm_sm_copy_root(cmd->sm, &cmd->metadata_space_map_root, root_size);
295 }
296 
297 /* Save dm-clone metadata in superblock */
298 static void __prepare_superblock(struct dm_clone_metadata *cmd,
299 				 struct superblock_disk *sb)
300 {
301 	sb->flags = cpu_to_le32(0UL);
302 
303 	/* FIXME: UUID is currently unused */
304 	memset(sb->uuid, 0, sizeof(sb->uuid));
305 
306 	sb->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
307 	sb->version = cpu_to_le32(DM_CLONE_MAX_METADATA_VERSION);
308 
309 	/* Save the metadata space_map root */
310 	memcpy(&sb->metadata_space_map_root, &cmd->metadata_space_map_root,
311 	       sizeof(cmd->metadata_space_map_root));
312 
313 	sb->region_size = cpu_to_le64(cmd->region_size);
314 	sb->target_size = cpu_to_le64(cmd->target_size);
315 	sb->bitset_root = cpu_to_le64(cmd->bitset_root);
316 }
317 
318 static int __open_metadata(struct dm_clone_metadata *cmd)
319 {
320 	int r;
321 	struct dm_block *sblock;
322 	struct superblock_disk *sb;
323 
324 	r = superblock_read_lock(cmd, &sblock);
325 
326 	if (r) {
327 		DMERR("Failed to read_lock superblock");
328 		return r;
329 	}
330 
331 	sb = dm_block_data(sblock);
332 
333 	/* Verify that target_size and region_size haven't changed. */
334 	if (cmd->region_size != le64_to_cpu(sb->region_size) ||
335 	    cmd->target_size != le64_to_cpu(sb->target_size)) {
336 		DMERR("Region and/or target size don't match the ones in metadata");
337 		r = -EINVAL;
338 		goto out_with_lock;
339 	}
340 
341 	r = dm_tm_open_with_sm(cmd->bm, SUPERBLOCK_LOCATION,
342 			       sb->metadata_space_map_root,
343 			       sizeof(sb->metadata_space_map_root),
344 			       &cmd->tm, &cmd->sm);
345 
346 	if (r) {
347 		DMERR("dm_tm_open_with_sm failed");
348 		goto out_with_lock;
349 	}
350 
351 	dm_disk_bitset_init(cmd->tm, &cmd->bitset_info);
352 	cmd->bitset_root = le64_to_cpu(sb->bitset_root);
353 
354 out_with_lock:
355 	dm_bm_unlock(sblock);
356 
357 	return r;
358 }
359 
360 static int __format_metadata(struct dm_clone_metadata *cmd)
361 {
362 	int r;
363 	struct dm_block *sblock;
364 	struct superblock_disk *sb;
365 
366 	r = dm_tm_create_with_sm(cmd->bm, SUPERBLOCK_LOCATION, &cmd->tm, &cmd->sm);
367 	if (r) {
368 		DMERR("Failed to create transaction manager");
369 		return r;
370 	}
371 
372 	dm_disk_bitset_init(cmd->tm, &cmd->bitset_info);
373 
374 	r = dm_bitset_empty(&cmd->bitset_info, &cmd->bitset_root);
375 	if (r) {
376 		DMERR("Failed to create empty on-disk bitset");
377 		goto err_with_tm;
378 	}
379 
380 	r = dm_bitset_resize(&cmd->bitset_info, cmd->bitset_root, 0,
381 			     cmd->nr_regions, false, &cmd->bitset_root);
382 	if (r) {
383 		DMERR("Failed to resize on-disk bitset to %lu entries", cmd->nr_regions);
384 		goto err_with_tm;
385 	}
386 
387 	/* Flush to disk all blocks, except the superblock */
388 	r = dm_tm_pre_commit(cmd->tm);
389 	if (r) {
390 		DMERR("dm_tm_pre_commit failed");
391 		goto err_with_tm;
392 	}
393 
394 	r = __copy_sm_root(cmd);
395 	if (r) {
396 		DMERR("__copy_sm_root failed");
397 		goto err_with_tm;
398 	}
399 
400 	r = superblock_write_lock_zero(cmd, &sblock);
401 	if (r) {
402 		DMERR("Failed to write_lock superblock");
403 		goto err_with_tm;
404 	}
405 
406 	sb = dm_block_data(sblock);
407 	__prepare_superblock(cmd, sb);
408 	r = dm_tm_commit(cmd->tm, sblock);
409 	if (r) {
410 		DMERR("Failed to commit superblock");
411 		goto err_with_tm;
412 	}
413 
414 	return 0;
415 
416 err_with_tm:
417 	dm_sm_destroy(cmd->sm);
418 	dm_tm_destroy(cmd->tm);
419 
420 	return r;
421 }
422 
423 static int __open_or_format_metadata(struct dm_clone_metadata *cmd, bool may_format_device)
424 {
425 	int r;
426 	bool formatted = false;
427 
428 	r = __superblock_all_zeroes(cmd->bm, &formatted);
429 	if (r)
430 		return r;
431 
432 	if (!formatted)
433 		return may_format_device ? __format_metadata(cmd) : -EPERM;
434 
435 	return __open_metadata(cmd);
436 }
437 
438 static int __create_persistent_data_structures(struct dm_clone_metadata *cmd,
439 					       bool may_format_device)
440 {
441 	int r;
442 
443 	/* Create block manager */
444 	cmd->bm = dm_block_manager_create(cmd->bdev,
445 					 DM_CLONE_METADATA_BLOCK_SIZE << SECTOR_SHIFT,
446 					 DM_CLONE_MAX_CONCURRENT_LOCKS);
447 	if (IS_ERR(cmd->bm)) {
448 		DMERR("Failed to create block manager");
449 		return PTR_ERR(cmd->bm);
450 	}
451 
452 	r = __open_or_format_metadata(cmd, may_format_device);
453 	if (r)
454 		dm_block_manager_destroy(cmd->bm);
455 
456 	return r;
457 }
458 
459 static void __destroy_persistent_data_structures(struct dm_clone_metadata *cmd)
460 {
461 	dm_sm_destroy(cmd->sm);
462 	dm_tm_destroy(cmd->tm);
463 	dm_block_manager_destroy(cmd->bm);
464 }
465 
466 /*---------------------------------------------------------------------------*/
467 
468 static size_t bitmap_size(unsigned long nr_bits)
469 {
470 	return BITS_TO_LONGS(nr_bits) * sizeof(long);
471 }
472 
473 static int __dirty_map_init(struct dirty_map *dmap, unsigned long nr_words,
474 			    unsigned long nr_regions)
475 {
476 	dmap->changed = 0;
477 
478 	dmap->dirty_words = kvzalloc(bitmap_size(nr_words), GFP_KERNEL);
479 	if (!dmap->dirty_words)
480 		return -ENOMEM;
481 
482 	dmap->dirty_regions = kvzalloc(bitmap_size(nr_regions), GFP_KERNEL);
483 	if (!dmap->dirty_regions) {
484 		kvfree(dmap->dirty_words);
485 		return -ENOMEM;
486 	}
487 
488 	return 0;
489 }
490 
491 static void __dirty_map_exit(struct dirty_map *dmap)
492 {
493 	kvfree(dmap->dirty_words);
494 	kvfree(dmap->dirty_regions);
495 }
496 
497 static int dirty_map_init(struct dm_clone_metadata *cmd)
498 {
499 	if (__dirty_map_init(&cmd->dmap[0], cmd->nr_words, cmd->nr_regions)) {
500 		DMERR("Failed to allocate dirty bitmap");
501 		return -ENOMEM;
502 	}
503 
504 	if (__dirty_map_init(&cmd->dmap[1], cmd->nr_words, cmd->nr_regions)) {
505 		DMERR("Failed to allocate dirty bitmap");
506 		__dirty_map_exit(&cmd->dmap[0]);
507 		return -ENOMEM;
508 	}
509 
510 	cmd->current_dmap = &cmd->dmap[0];
511 	cmd->committing_dmap = NULL;
512 
513 	return 0;
514 }
515 
516 static void dirty_map_exit(struct dm_clone_metadata *cmd)
517 {
518 	__dirty_map_exit(&cmd->dmap[0]);
519 	__dirty_map_exit(&cmd->dmap[1]);
520 }
521 
522 static int __load_bitset_in_core(struct dm_clone_metadata *cmd)
523 {
524 	int r;
525 	unsigned long i;
526 	struct dm_bitset_cursor c;
527 
528 	/* Flush bitset cache */
529 	r = dm_bitset_flush(&cmd->bitset_info, cmd->bitset_root, &cmd->bitset_root);
530 	if (r)
531 		return r;
532 
533 	r = dm_bitset_cursor_begin(&cmd->bitset_info, cmd->bitset_root, cmd->nr_regions, &c);
534 	if (r)
535 		return r;
536 
537 	for (i = 0; ; i++) {
538 		if (dm_bitset_cursor_get_value(&c))
539 			__set_bit(i, cmd->region_map);
540 		else
541 			__clear_bit(i, cmd->region_map);
542 
543 		if (i >= (cmd->nr_regions - 1))
544 			break;
545 
546 		r = dm_bitset_cursor_next(&c);
547 
548 		if (r)
549 			break;
550 	}
551 
552 	dm_bitset_cursor_end(&c);
553 
554 	return r;
555 }
556 
557 struct dm_clone_metadata *dm_clone_metadata_open(struct block_device *bdev,
558 						 sector_t target_size,
559 						 sector_t region_size)
560 {
561 	int r;
562 	struct dm_clone_metadata *cmd;
563 
564 	cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
565 	if (!cmd) {
566 		DMERR("Failed to allocate memory for dm-clone metadata");
567 		return ERR_PTR(-ENOMEM);
568 	}
569 
570 	cmd->bdev = bdev;
571 	cmd->target_size = target_size;
572 	cmd->region_size = region_size;
573 	cmd->nr_regions = dm_sector_div_up(cmd->target_size, cmd->region_size);
574 	cmd->nr_words = BITS_TO_LONGS(cmd->nr_regions);
575 
576 	init_rwsem(&cmd->lock);
577 	spin_lock_init(&cmd->bitmap_lock);
578 	cmd->read_only = 0;
579 	cmd->fail_io = false;
580 	cmd->hydration_done = false;
581 
582 	cmd->region_map = kvmalloc(bitmap_size(cmd->nr_regions), GFP_KERNEL);
583 	if (!cmd->region_map) {
584 		DMERR("Failed to allocate memory for region bitmap");
585 		r = -ENOMEM;
586 		goto out_with_md;
587 	}
588 
589 	r = __create_persistent_data_structures(cmd, true);
590 	if (r)
591 		goto out_with_region_map;
592 
593 	r = __load_bitset_in_core(cmd);
594 	if (r) {
595 		DMERR("Failed to load on-disk region map");
596 		goto out_with_pds;
597 	}
598 
599 	r = dirty_map_init(cmd);
600 	if (r)
601 		goto out_with_pds;
602 
603 	if (bitmap_full(cmd->region_map, cmd->nr_regions))
604 		cmd->hydration_done = true;
605 
606 	return cmd;
607 
608 out_with_pds:
609 	__destroy_persistent_data_structures(cmd);
610 
611 out_with_region_map:
612 	kvfree(cmd->region_map);
613 
614 out_with_md:
615 	kfree(cmd);
616 
617 	return ERR_PTR(r);
618 }
619 
620 void dm_clone_metadata_close(struct dm_clone_metadata *cmd)
621 {
622 	if (!cmd->fail_io)
623 		__destroy_persistent_data_structures(cmd);
624 
625 	dirty_map_exit(cmd);
626 	kvfree(cmd->region_map);
627 	kfree(cmd);
628 }
629 
630 bool dm_clone_is_hydration_done(struct dm_clone_metadata *cmd)
631 {
632 	return cmd->hydration_done;
633 }
634 
635 bool dm_clone_is_region_hydrated(struct dm_clone_metadata *cmd, unsigned long region_nr)
636 {
637 	return dm_clone_is_hydration_done(cmd) || test_bit(region_nr, cmd->region_map);
638 }
639 
640 bool dm_clone_is_range_hydrated(struct dm_clone_metadata *cmd,
641 				unsigned long start, unsigned long nr_regions)
642 {
643 	unsigned long bit;
644 
645 	if (dm_clone_is_hydration_done(cmd))
646 		return true;
647 
648 	bit = find_next_zero_bit(cmd->region_map, cmd->nr_regions, start);
649 
650 	return (bit >= (start + nr_regions));
651 }
652 
653 unsigned int dm_clone_nr_of_hydrated_regions(struct dm_clone_metadata *cmd)
654 {
655 	return bitmap_weight(cmd->region_map, cmd->nr_regions);
656 }
657 
658 unsigned long dm_clone_find_next_unhydrated_region(struct dm_clone_metadata *cmd,
659 						   unsigned long start)
660 {
661 	return find_next_zero_bit(cmd->region_map, cmd->nr_regions, start);
662 }
663 
664 static int __update_metadata_word(struct dm_clone_metadata *cmd,
665 				  unsigned long *dirty_regions,
666 				  unsigned long word)
667 {
668 	int r;
669 	unsigned long index = word * BITS_PER_LONG;
670 	unsigned long max_index = min(cmd->nr_regions, (word + 1) * BITS_PER_LONG);
671 
672 	while (index < max_index) {
673 		if (test_bit(index, dirty_regions)) {
674 			r = dm_bitset_set_bit(&cmd->bitset_info, cmd->bitset_root,
675 					      index, &cmd->bitset_root);
676 			if (r) {
677 				DMERR("dm_bitset_set_bit failed");
678 				return r;
679 			}
680 			__clear_bit(index, dirty_regions);
681 		}
682 		index++;
683 	}
684 
685 	return 0;
686 }
687 
688 static int __metadata_commit(struct dm_clone_metadata *cmd)
689 {
690 	int r;
691 	struct dm_block *sblock;
692 	struct superblock_disk *sb;
693 
694 	/* Flush bitset cache */
695 	r = dm_bitset_flush(&cmd->bitset_info, cmd->bitset_root, &cmd->bitset_root);
696 	if (r) {
697 		DMERR("dm_bitset_flush failed");
698 		return r;
699 	}
700 
701 	/* Flush to disk all blocks, except the superblock */
702 	r = dm_tm_pre_commit(cmd->tm);
703 	if (r) {
704 		DMERR("dm_tm_pre_commit failed");
705 		return r;
706 	}
707 
708 	/* Save the space map root in cmd->metadata_space_map_root */
709 	r = __copy_sm_root(cmd);
710 	if (r) {
711 		DMERR("__copy_sm_root failed");
712 		return r;
713 	}
714 
715 	/* Lock the superblock */
716 	r = superblock_write_lock_zero(cmd, &sblock);
717 	if (r) {
718 		DMERR("Failed to write_lock superblock");
719 		return r;
720 	}
721 
722 	/* Save the metadata in superblock */
723 	sb = dm_block_data(sblock);
724 	__prepare_superblock(cmd, sb);
725 
726 	/* Unlock superblock and commit it to disk */
727 	r = dm_tm_commit(cmd->tm, sblock);
728 	if (r) {
729 		DMERR("Failed to commit superblock");
730 		return r;
731 	}
732 
733 	/*
734 	 * FIXME: Find a more efficient way to check if the hydration is done.
735 	 */
736 	if (bitmap_full(cmd->region_map, cmd->nr_regions))
737 		cmd->hydration_done = true;
738 
739 	return 0;
740 }
741 
742 static int __flush_dmap(struct dm_clone_metadata *cmd, struct dirty_map *dmap)
743 {
744 	int r;
745 	unsigned long word;
746 
747 	word = 0;
748 	do {
749 		word = find_next_bit(dmap->dirty_words, cmd->nr_words, word);
750 
751 		if (word == cmd->nr_words)
752 			break;
753 
754 		r = __update_metadata_word(cmd, dmap->dirty_regions, word);
755 
756 		if (r)
757 			return r;
758 
759 		__clear_bit(word, dmap->dirty_words);
760 		word++;
761 	} while (word < cmd->nr_words);
762 
763 	r = __metadata_commit(cmd);
764 
765 	if (r)
766 		return r;
767 
768 	/* Update the changed flag */
769 	spin_lock_irq(&cmd->bitmap_lock);
770 	dmap->changed = 0;
771 	spin_unlock_irq(&cmd->bitmap_lock);
772 
773 	return 0;
774 }
775 
776 int dm_clone_metadata_pre_commit(struct dm_clone_metadata *cmd)
777 {
778 	int r = 0;
779 	struct dirty_map *dmap, *next_dmap;
780 
781 	down_write(&cmd->lock);
782 
783 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
784 		r = -EPERM;
785 		goto out;
786 	}
787 
788 	/* Get current dirty bitmap */
789 	dmap = cmd->current_dmap;
790 
791 	/* Get next dirty bitmap */
792 	next_dmap = (dmap == &cmd->dmap[0]) ? &cmd->dmap[1] : &cmd->dmap[0];
793 
794 	/*
795 	 * The last commit failed, so we don't have a clean dirty-bitmap to
796 	 * use.
797 	 */
798 	if (WARN_ON(next_dmap->changed || cmd->committing_dmap)) {
799 		r = -EINVAL;
800 		goto out;
801 	}
802 
803 	/* Swap dirty bitmaps */
804 	spin_lock_irq(&cmd->bitmap_lock);
805 	cmd->current_dmap = next_dmap;
806 	spin_unlock_irq(&cmd->bitmap_lock);
807 
808 	/* Set old dirty bitmap as currently committing */
809 	cmd->committing_dmap = dmap;
810 out:
811 	up_write(&cmd->lock);
812 
813 	return r;
814 }
815 
816 int dm_clone_metadata_commit(struct dm_clone_metadata *cmd)
817 {
818 	int r = -EPERM;
819 
820 	down_write(&cmd->lock);
821 
822 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm))
823 		goto out;
824 
825 	if (WARN_ON(!cmd->committing_dmap)) {
826 		r = -EINVAL;
827 		goto out;
828 	}
829 
830 	r = __flush_dmap(cmd, cmd->committing_dmap);
831 	if (!r) {
832 		/* Clear committing dmap */
833 		cmd->committing_dmap = NULL;
834 	}
835 out:
836 	up_write(&cmd->lock);
837 
838 	return r;
839 }
840 
841 int dm_clone_set_region_hydrated(struct dm_clone_metadata *cmd, unsigned long region_nr)
842 {
843 	int r = 0;
844 	struct dirty_map *dmap;
845 	unsigned long word, flags;
846 
847 	if (unlikely(region_nr >= cmd->nr_regions)) {
848 		DMERR("Region %lu out of range (total number of regions %lu)",
849 		      region_nr, cmd->nr_regions);
850 		return -ERANGE;
851 	}
852 
853 	word = region_nr / BITS_PER_LONG;
854 
855 	spin_lock_irqsave(&cmd->bitmap_lock, flags);
856 
857 	if (cmd->read_only) {
858 		r = -EPERM;
859 		goto out;
860 	}
861 
862 	dmap = cmd->current_dmap;
863 
864 	__set_bit(word, dmap->dirty_words);
865 	__set_bit(region_nr, dmap->dirty_regions);
866 	__set_bit(region_nr, cmd->region_map);
867 	dmap->changed = 1;
868 
869 out:
870 	spin_unlock_irqrestore(&cmd->bitmap_lock, flags);
871 
872 	return r;
873 }
874 
875 int dm_clone_cond_set_range(struct dm_clone_metadata *cmd, unsigned long start,
876 			    unsigned long nr_regions)
877 {
878 	int r = 0;
879 	struct dirty_map *dmap;
880 	unsigned long word, region_nr;
881 
882 	if (unlikely(start >= cmd->nr_regions || (start + nr_regions) < start ||
883 		     (start + nr_regions) > cmd->nr_regions)) {
884 		DMERR("Invalid region range: start %lu, nr_regions %lu (total number of regions %lu)",
885 		      start, nr_regions, cmd->nr_regions);
886 		return -ERANGE;
887 	}
888 
889 	spin_lock_irq(&cmd->bitmap_lock);
890 
891 	if (cmd->read_only) {
892 		r = -EPERM;
893 		goto out;
894 	}
895 
896 	dmap = cmd->current_dmap;
897 	for (region_nr = start; region_nr < (start + nr_regions); region_nr++) {
898 		if (!test_bit(region_nr, cmd->region_map)) {
899 			word = region_nr / BITS_PER_LONG;
900 			__set_bit(word, dmap->dirty_words);
901 			__set_bit(region_nr, dmap->dirty_regions);
902 			__set_bit(region_nr, cmd->region_map);
903 			dmap->changed = 1;
904 		}
905 	}
906 out:
907 	spin_unlock_irq(&cmd->bitmap_lock);
908 
909 	return r;
910 }
911 
912 /*
913  * WARNING: This must not be called concurrently with either
914  * dm_clone_set_region_hydrated() or dm_clone_cond_set_range(), as it changes
915  * cmd->region_map without taking the cmd->bitmap_lock spinlock. The only
916  * exception is after setting the metadata to read-only mode, using
917  * dm_clone_metadata_set_read_only().
918  *
919  * We don't take the spinlock because __load_bitset_in_core() does I/O, so it
920  * may block.
921  */
922 int dm_clone_reload_in_core_bitset(struct dm_clone_metadata *cmd)
923 {
924 	int r = -EINVAL;
925 
926 	down_write(&cmd->lock);
927 
928 	if (cmd->fail_io)
929 		goto out;
930 
931 	r = __load_bitset_in_core(cmd);
932 out:
933 	up_write(&cmd->lock);
934 
935 	return r;
936 }
937 
938 bool dm_clone_changed_this_transaction(struct dm_clone_metadata *cmd)
939 {
940 	bool r;
941 	unsigned long flags;
942 
943 	spin_lock_irqsave(&cmd->bitmap_lock, flags);
944 	r = cmd->dmap[0].changed || cmd->dmap[1].changed;
945 	spin_unlock_irqrestore(&cmd->bitmap_lock, flags);
946 
947 	return r;
948 }
949 
950 int dm_clone_metadata_abort(struct dm_clone_metadata *cmd)
951 {
952 	int r = -EPERM;
953 
954 	down_write(&cmd->lock);
955 
956 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm))
957 		goto out;
958 
959 	__destroy_persistent_data_structures(cmd);
960 
961 	r = __create_persistent_data_structures(cmd, false);
962 	if (r) {
963 		/* If something went wrong we can neither write nor read the metadata */
964 		cmd->fail_io = true;
965 	}
966 out:
967 	up_write(&cmd->lock);
968 
969 	return r;
970 }
971 
972 void dm_clone_metadata_set_read_only(struct dm_clone_metadata *cmd)
973 {
974 	down_write(&cmd->lock);
975 
976 	spin_lock_irq(&cmd->bitmap_lock);
977 	cmd->read_only = 1;
978 	spin_unlock_irq(&cmd->bitmap_lock);
979 
980 	if (!cmd->fail_io)
981 		dm_bm_set_read_only(cmd->bm);
982 
983 	up_write(&cmd->lock);
984 }
985 
986 void dm_clone_metadata_set_read_write(struct dm_clone_metadata *cmd)
987 {
988 	down_write(&cmd->lock);
989 
990 	spin_lock_irq(&cmd->bitmap_lock);
991 	cmd->read_only = 0;
992 	spin_unlock_irq(&cmd->bitmap_lock);
993 
994 	if (!cmd->fail_io)
995 		dm_bm_set_read_write(cmd->bm);
996 
997 	up_write(&cmd->lock);
998 }
999 
1000 int dm_clone_get_free_metadata_block_count(struct dm_clone_metadata *cmd,
1001 					   dm_block_t *result)
1002 {
1003 	int r = -EINVAL;
1004 
1005 	down_read(&cmd->lock);
1006 
1007 	if (!cmd->fail_io)
1008 		r = dm_sm_get_nr_free(cmd->sm, result);
1009 
1010 	up_read(&cmd->lock);
1011 
1012 	return r;
1013 }
1014 
1015 int dm_clone_get_metadata_dev_size(struct dm_clone_metadata *cmd,
1016 				   dm_block_t *result)
1017 {
1018 	int r = -EINVAL;
1019 
1020 	down_read(&cmd->lock);
1021 
1022 	if (!cmd->fail_io)
1023 		r = dm_sm_get_nr_blocks(cmd->sm, result);
1024 
1025 	up_read(&cmd->lock);
1026 
1027 	return r;
1028 }
1029