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 int __dirty_map_init(struct dirty_map *dmap, unsigned long nr_words,
469 			    unsigned long nr_regions)
470 {
471 	dmap->changed = 0;
472 
473 	dmap->dirty_words = kvzalloc(bitmap_size(nr_words), GFP_KERNEL);
474 	if (!dmap->dirty_words)
475 		return -ENOMEM;
476 
477 	dmap->dirty_regions = kvzalloc(bitmap_size(nr_regions), GFP_KERNEL);
478 	if (!dmap->dirty_regions) {
479 		kvfree(dmap->dirty_words);
480 		return -ENOMEM;
481 	}
482 
483 	return 0;
484 }
485 
486 static void __dirty_map_exit(struct dirty_map *dmap)
487 {
488 	kvfree(dmap->dirty_words);
489 	kvfree(dmap->dirty_regions);
490 }
491 
492 static int dirty_map_init(struct dm_clone_metadata *cmd)
493 {
494 	if (__dirty_map_init(&cmd->dmap[0], cmd->nr_words, cmd->nr_regions)) {
495 		DMERR("Failed to allocate dirty bitmap");
496 		return -ENOMEM;
497 	}
498 
499 	if (__dirty_map_init(&cmd->dmap[1], cmd->nr_words, cmd->nr_regions)) {
500 		DMERR("Failed to allocate dirty bitmap");
501 		__dirty_map_exit(&cmd->dmap[0]);
502 		return -ENOMEM;
503 	}
504 
505 	cmd->current_dmap = &cmd->dmap[0];
506 	cmd->committing_dmap = NULL;
507 
508 	return 0;
509 }
510 
511 static void dirty_map_exit(struct dm_clone_metadata *cmd)
512 {
513 	__dirty_map_exit(&cmd->dmap[0]);
514 	__dirty_map_exit(&cmd->dmap[1]);
515 }
516 
517 static int __load_bitset_in_core(struct dm_clone_metadata *cmd)
518 {
519 	int r;
520 	unsigned long i;
521 	struct dm_bitset_cursor c;
522 
523 	/* Flush bitset cache */
524 	r = dm_bitset_flush(&cmd->bitset_info, cmd->bitset_root, &cmd->bitset_root);
525 	if (r)
526 		return r;
527 
528 	r = dm_bitset_cursor_begin(&cmd->bitset_info, cmd->bitset_root, cmd->nr_regions, &c);
529 	if (r)
530 		return r;
531 
532 	for (i = 0; ; i++) {
533 		if (dm_bitset_cursor_get_value(&c))
534 			__set_bit(i, cmd->region_map);
535 		else
536 			__clear_bit(i, cmd->region_map);
537 
538 		if (i >= (cmd->nr_regions - 1))
539 			break;
540 
541 		r = dm_bitset_cursor_next(&c);
542 
543 		if (r)
544 			break;
545 	}
546 
547 	dm_bitset_cursor_end(&c);
548 
549 	return r;
550 }
551 
552 struct dm_clone_metadata *dm_clone_metadata_open(struct block_device *bdev,
553 						 sector_t target_size,
554 						 sector_t region_size)
555 {
556 	int r;
557 	struct dm_clone_metadata *cmd;
558 
559 	cmd = kzalloc(sizeof(*cmd), GFP_KERNEL);
560 	if (!cmd) {
561 		DMERR("Failed to allocate memory for dm-clone metadata");
562 		return ERR_PTR(-ENOMEM);
563 	}
564 
565 	cmd->bdev = bdev;
566 	cmd->target_size = target_size;
567 	cmd->region_size = region_size;
568 	cmd->nr_regions = dm_sector_div_up(cmd->target_size, cmd->region_size);
569 	cmd->nr_words = BITS_TO_LONGS(cmd->nr_regions);
570 
571 	init_rwsem(&cmd->lock);
572 	spin_lock_init(&cmd->bitmap_lock);
573 	cmd->read_only = 0;
574 	cmd->fail_io = false;
575 	cmd->hydration_done = false;
576 
577 	cmd->region_map = kvmalloc(bitmap_size(cmd->nr_regions), GFP_KERNEL);
578 	if (!cmd->region_map) {
579 		DMERR("Failed to allocate memory for region bitmap");
580 		r = -ENOMEM;
581 		goto out_with_md;
582 	}
583 
584 	r = __create_persistent_data_structures(cmd, true);
585 	if (r)
586 		goto out_with_region_map;
587 
588 	r = __load_bitset_in_core(cmd);
589 	if (r) {
590 		DMERR("Failed to load on-disk region map");
591 		goto out_with_pds;
592 	}
593 
594 	r = dirty_map_init(cmd);
595 	if (r)
596 		goto out_with_pds;
597 
598 	if (bitmap_full(cmd->region_map, cmd->nr_regions))
599 		cmd->hydration_done = true;
600 
601 	return cmd;
602 
603 out_with_pds:
604 	__destroy_persistent_data_structures(cmd);
605 
606 out_with_region_map:
607 	kvfree(cmd->region_map);
608 
609 out_with_md:
610 	kfree(cmd);
611 
612 	return ERR_PTR(r);
613 }
614 
615 void dm_clone_metadata_close(struct dm_clone_metadata *cmd)
616 {
617 	if (!cmd->fail_io)
618 		__destroy_persistent_data_structures(cmd);
619 
620 	dirty_map_exit(cmd);
621 	kvfree(cmd->region_map);
622 	kfree(cmd);
623 }
624 
625 bool dm_clone_is_hydration_done(struct dm_clone_metadata *cmd)
626 {
627 	return cmd->hydration_done;
628 }
629 
630 bool dm_clone_is_region_hydrated(struct dm_clone_metadata *cmd, unsigned long region_nr)
631 {
632 	return dm_clone_is_hydration_done(cmd) || test_bit(region_nr, cmd->region_map);
633 }
634 
635 bool dm_clone_is_range_hydrated(struct dm_clone_metadata *cmd,
636 				unsigned long start, unsigned long nr_regions)
637 {
638 	unsigned long bit;
639 
640 	if (dm_clone_is_hydration_done(cmd))
641 		return true;
642 
643 	bit = find_next_zero_bit(cmd->region_map, cmd->nr_regions, start);
644 
645 	return (bit >= (start + nr_regions));
646 }
647 
648 unsigned int dm_clone_nr_of_hydrated_regions(struct dm_clone_metadata *cmd)
649 {
650 	return bitmap_weight(cmd->region_map, cmd->nr_regions);
651 }
652 
653 unsigned long dm_clone_find_next_unhydrated_region(struct dm_clone_metadata *cmd,
654 						   unsigned long start)
655 {
656 	return find_next_zero_bit(cmd->region_map, cmd->nr_regions, start);
657 }
658 
659 static int __update_metadata_word(struct dm_clone_metadata *cmd,
660 				  unsigned long *dirty_regions,
661 				  unsigned long word)
662 {
663 	int r;
664 	unsigned long index = word * BITS_PER_LONG;
665 	unsigned long max_index = min(cmd->nr_regions, (word + 1) * BITS_PER_LONG);
666 
667 	while (index < max_index) {
668 		if (test_bit(index, dirty_regions)) {
669 			r = dm_bitset_set_bit(&cmd->bitset_info, cmd->bitset_root,
670 					      index, &cmd->bitset_root);
671 			if (r) {
672 				DMERR("dm_bitset_set_bit failed");
673 				return r;
674 			}
675 			__clear_bit(index, dirty_regions);
676 		}
677 		index++;
678 	}
679 
680 	return 0;
681 }
682 
683 static int __metadata_commit(struct dm_clone_metadata *cmd)
684 {
685 	int r;
686 	struct dm_block *sblock;
687 	struct superblock_disk *sb;
688 
689 	/* Flush bitset cache */
690 	r = dm_bitset_flush(&cmd->bitset_info, cmd->bitset_root, &cmd->bitset_root);
691 	if (r) {
692 		DMERR("dm_bitset_flush failed");
693 		return r;
694 	}
695 
696 	/* Flush to disk all blocks, except the superblock */
697 	r = dm_tm_pre_commit(cmd->tm);
698 	if (r) {
699 		DMERR("dm_tm_pre_commit failed");
700 		return r;
701 	}
702 
703 	/* Save the space map root in cmd->metadata_space_map_root */
704 	r = __copy_sm_root(cmd);
705 	if (r) {
706 		DMERR("__copy_sm_root failed");
707 		return r;
708 	}
709 
710 	/* Lock the superblock */
711 	r = superblock_write_lock_zero(cmd, &sblock);
712 	if (r) {
713 		DMERR("Failed to write_lock superblock");
714 		return r;
715 	}
716 
717 	/* Save the metadata in superblock */
718 	sb = dm_block_data(sblock);
719 	__prepare_superblock(cmd, sb);
720 
721 	/* Unlock superblock and commit it to disk */
722 	r = dm_tm_commit(cmd->tm, sblock);
723 	if (r) {
724 		DMERR("Failed to commit superblock");
725 		return r;
726 	}
727 
728 	/*
729 	 * FIXME: Find a more efficient way to check if the hydration is done.
730 	 */
731 	if (bitmap_full(cmd->region_map, cmd->nr_regions))
732 		cmd->hydration_done = true;
733 
734 	return 0;
735 }
736 
737 static int __flush_dmap(struct dm_clone_metadata *cmd, struct dirty_map *dmap)
738 {
739 	int r;
740 	unsigned long word;
741 
742 	word = 0;
743 	do {
744 		word = find_next_bit(dmap->dirty_words, cmd->nr_words, word);
745 
746 		if (word == cmd->nr_words)
747 			break;
748 
749 		r = __update_metadata_word(cmd, dmap->dirty_regions, word);
750 
751 		if (r)
752 			return r;
753 
754 		__clear_bit(word, dmap->dirty_words);
755 		word++;
756 	} while (word < cmd->nr_words);
757 
758 	r = __metadata_commit(cmd);
759 
760 	if (r)
761 		return r;
762 
763 	/* Update the changed flag */
764 	spin_lock_irq(&cmd->bitmap_lock);
765 	dmap->changed = 0;
766 	spin_unlock_irq(&cmd->bitmap_lock);
767 
768 	return 0;
769 }
770 
771 int dm_clone_metadata_pre_commit(struct dm_clone_metadata *cmd)
772 {
773 	int r = 0;
774 	struct dirty_map *dmap, *next_dmap;
775 
776 	down_write(&cmd->lock);
777 
778 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm)) {
779 		r = -EPERM;
780 		goto out;
781 	}
782 
783 	/* Get current dirty bitmap */
784 	dmap = cmd->current_dmap;
785 
786 	/* Get next dirty bitmap */
787 	next_dmap = (dmap == &cmd->dmap[0]) ? &cmd->dmap[1] : &cmd->dmap[0];
788 
789 	/*
790 	 * The last commit failed, so we don't have a clean dirty-bitmap to
791 	 * use.
792 	 */
793 	if (WARN_ON(next_dmap->changed || cmd->committing_dmap)) {
794 		r = -EINVAL;
795 		goto out;
796 	}
797 
798 	/* Swap dirty bitmaps */
799 	spin_lock_irq(&cmd->bitmap_lock);
800 	cmd->current_dmap = next_dmap;
801 	spin_unlock_irq(&cmd->bitmap_lock);
802 
803 	/* Set old dirty bitmap as currently committing */
804 	cmd->committing_dmap = dmap;
805 out:
806 	up_write(&cmd->lock);
807 
808 	return r;
809 }
810 
811 int dm_clone_metadata_commit(struct dm_clone_metadata *cmd)
812 {
813 	int r = -EPERM;
814 
815 	down_write(&cmd->lock);
816 
817 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm))
818 		goto out;
819 
820 	if (WARN_ON(!cmd->committing_dmap)) {
821 		r = -EINVAL;
822 		goto out;
823 	}
824 
825 	r = __flush_dmap(cmd, cmd->committing_dmap);
826 	if (!r) {
827 		/* Clear committing dmap */
828 		cmd->committing_dmap = NULL;
829 	}
830 out:
831 	up_write(&cmd->lock);
832 
833 	return r;
834 }
835 
836 int dm_clone_set_region_hydrated(struct dm_clone_metadata *cmd, unsigned long region_nr)
837 {
838 	int r = 0;
839 	struct dirty_map *dmap;
840 	unsigned long word, flags;
841 
842 	if (unlikely(region_nr >= cmd->nr_regions)) {
843 		DMERR("Region %lu out of range (total number of regions %lu)",
844 		      region_nr, cmd->nr_regions);
845 		return -ERANGE;
846 	}
847 
848 	word = region_nr / BITS_PER_LONG;
849 
850 	spin_lock_irqsave(&cmd->bitmap_lock, flags);
851 
852 	if (cmd->read_only) {
853 		r = -EPERM;
854 		goto out;
855 	}
856 
857 	dmap = cmd->current_dmap;
858 
859 	__set_bit(word, dmap->dirty_words);
860 	__set_bit(region_nr, dmap->dirty_regions);
861 	__set_bit(region_nr, cmd->region_map);
862 	dmap->changed = 1;
863 
864 out:
865 	spin_unlock_irqrestore(&cmd->bitmap_lock, flags);
866 
867 	return r;
868 }
869 
870 int dm_clone_cond_set_range(struct dm_clone_metadata *cmd, unsigned long start,
871 			    unsigned long nr_regions)
872 {
873 	int r = 0;
874 	struct dirty_map *dmap;
875 	unsigned long word, region_nr;
876 
877 	if (unlikely(start >= cmd->nr_regions || (start + nr_regions) < start ||
878 		     (start + nr_regions) > cmd->nr_regions)) {
879 		DMERR("Invalid region range: start %lu, nr_regions %lu (total number of regions %lu)",
880 		      start, nr_regions, cmd->nr_regions);
881 		return -ERANGE;
882 	}
883 
884 	spin_lock_irq(&cmd->bitmap_lock);
885 
886 	if (cmd->read_only) {
887 		r = -EPERM;
888 		goto out;
889 	}
890 
891 	dmap = cmd->current_dmap;
892 	for (region_nr = start; region_nr < (start + nr_regions); region_nr++) {
893 		if (!test_bit(region_nr, cmd->region_map)) {
894 			word = region_nr / BITS_PER_LONG;
895 			__set_bit(word, dmap->dirty_words);
896 			__set_bit(region_nr, dmap->dirty_regions);
897 			__set_bit(region_nr, cmd->region_map);
898 			dmap->changed = 1;
899 		}
900 	}
901 out:
902 	spin_unlock_irq(&cmd->bitmap_lock);
903 
904 	return r;
905 }
906 
907 /*
908  * WARNING: This must not be called concurrently with either
909  * dm_clone_set_region_hydrated() or dm_clone_cond_set_range(), as it changes
910  * cmd->region_map without taking the cmd->bitmap_lock spinlock. The only
911  * exception is after setting the metadata to read-only mode, using
912  * dm_clone_metadata_set_read_only().
913  *
914  * We don't take the spinlock because __load_bitset_in_core() does I/O, so it
915  * may block.
916  */
917 int dm_clone_reload_in_core_bitset(struct dm_clone_metadata *cmd)
918 {
919 	int r = -EINVAL;
920 
921 	down_write(&cmd->lock);
922 
923 	if (cmd->fail_io)
924 		goto out;
925 
926 	r = __load_bitset_in_core(cmd);
927 out:
928 	up_write(&cmd->lock);
929 
930 	return r;
931 }
932 
933 bool dm_clone_changed_this_transaction(struct dm_clone_metadata *cmd)
934 {
935 	bool r;
936 	unsigned long flags;
937 
938 	spin_lock_irqsave(&cmd->bitmap_lock, flags);
939 	r = cmd->dmap[0].changed || cmd->dmap[1].changed;
940 	spin_unlock_irqrestore(&cmd->bitmap_lock, flags);
941 
942 	return r;
943 }
944 
945 int dm_clone_metadata_abort(struct dm_clone_metadata *cmd)
946 {
947 	int r = -EPERM;
948 
949 	down_write(&cmd->lock);
950 
951 	if (cmd->fail_io || dm_bm_is_read_only(cmd->bm))
952 		goto out;
953 
954 	__destroy_persistent_data_structures(cmd);
955 
956 	r = __create_persistent_data_structures(cmd, false);
957 	if (r) {
958 		/* If something went wrong we can neither write nor read the metadata */
959 		cmd->fail_io = true;
960 	}
961 out:
962 	up_write(&cmd->lock);
963 
964 	return r;
965 }
966 
967 void dm_clone_metadata_set_read_only(struct dm_clone_metadata *cmd)
968 {
969 	down_write(&cmd->lock);
970 
971 	spin_lock_irq(&cmd->bitmap_lock);
972 	cmd->read_only = 1;
973 	spin_unlock_irq(&cmd->bitmap_lock);
974 
975 	if (!cmd->fail_io)
976 		dm_bm_set_read_only(cmd->bm);
977 
978 	up_write(&cmd->lock);
979 }
980 
981 void dm_clone_metadata_set_read_write(struct dm_clone_metadata *cmd)
982 {
983 	down_write(&cmd->lock);
984 
985 	spin_lock_irq(&cmd->bitmap_lock);
986 	cmd->read_only = 0;
987 	spin_unlock_irq(&cmd->bitmap_lock);
988 
989 	if (!cmd->fail_io)
990 		dm_bm_set_read_write(cmd->bm);
991 
992 	up_write(&cmd->lock);
993 }
994 
995 int dm_clone_get_free_metadata_block_count(struct dm_clone_metadata *cmd,
996 					   dm_block_t *result)
997 {
998 	int r = -EINVAL;
999 
1000 	down_read(&cmd->lock);
1001 
1002 	if (!cmd->fail_io)
1003 		r = dm_sm_get_nr_free(cmd->sm, result);
1004 
1005 	up_read(&cmd->lock);
1006 
1007 	return r;
1008 }
1009 
1010 int dm_clone_get_metadata_dev_size(struct dm_clone_metadata *cmd,
1011 				   dm_block_t *result)
1012 {
1013 	int r = -EINVAL;
1014 
1015 	down_read(&cmd->lock);
1016 
1017 	if (!cmd->fail_io)
1018 		r = dm_sm_get_nr_blocks(cmd->sm, result);
1019 
1020 	up_read(&cmd->lock);
1021 
1022 	return r;
1023 }
1024