xref: /linux/drivers/md/dm-era-target.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 #include "dm.h"
2 #include "persistent-data/dm-transaction-manager.h"
3 #include "persistent-data/dm-bitset.h"
4 #include "persistent-data/dm-space-map.h"
5 
6 #include <linux/dm-io.h>
7 #include <linux/dm-kcopyd.h>
8 #include <linux/init.h>
9 #include <linux/mempool.h>
10 #include <linux/module.h>
11 #include <linux/slab.h>
12 #include <linux/vmalloc.h>
13 
14 #define DM_MSG_PREFIX "era"
15 
16 #define SUPERBLOCK_LOCATION 0
17 #define SUPERBLOCK_MAGIC 2126579579
18 #define SUPERBLOCK_CSUM_XOR 146538381
19 #define MIN_ERA_VERSION 1
20 #define MAX_ERA_VERSION 1
21 #define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION
22 #define MIN_BLOCK_SIZE 8
23 
24 /*----------------------------------------------------------------
25  * Writeset
26  *--------------------------------------------------------------*/
27 struct writeset_metadata {
28 	uint32_t nr_bits;
29 	dm_block_t root;
30 };
31 
32 struct writeset {
33 	struct writeset_metadata md;
34 
35 	/*
36 	 * An in core copy of the bits to save constantly doing look ups on
37 	 * disk.
38 	 */
39 	unsigned long *bits;
40 };
41 
42 /*
43  * This does not free off the on disk bitset as this will normally be done
44  * after digesting into the era array.
45  */
46 static void writeset_free(struct writeset *ws)
47 {
48 	vfree(ws->bits);
49 }
50 
51 static int setup_on_disk_bitset(struct dm_disk_bitset *info,
52 				unsigned nr_bits, dm_block_t *root)
53 {
54 	int r;
55 
56 	r = dm_bitset_empty(info, root);
57 	if (r)
58 		return r;
59 
60 	return dm_bitset_resize(info, *root, 0, nr_bits, false, root);
61 }
62 
63 static size_t bitset_size(unsigned nr_bits)
64 {
65 	return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG);
66 }
67 
68 /*
69  * Allocates memory for the in core bitset.
70  */
71 static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks)
72 {
73 	ws->md.nr_bits = nr_blocks;
74 	ws->md.root = INVALID_WRITESET_ROOT;
75 	ws->bits = vzalloc(bitset_size(nr_blocks));
76 	if (!ws->bits) {
77 		DMERR("%s: couldn't allocate in memory bitset", __func__);
78 		return -ENOMEM;
79 	}
80 
81 	return 0;
82 }
83 
84 /*
85  * Wipes the in-core bitset, and creates a new on disk bitset.
86  */
87 static int writeset_init(struct dm_disk_bitset *info, struct writeset *ws)
88 {
89 	int r;
90 
91 	memset(ws->bits, 0, bitset_size(ws->md.nr_bits));
92 
93 	r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root);
94 	if (r) {
95 		DMERR("%s: setup_on_disk_bitset failed", __func__);
96 		return r;
97 	}
98 
99 	return 0;
100 }
101 
102 static bool writeset_marked(struct writeset *ws, dm_block_t block)
103 {
104 	return test_bit(block, ws->bits);
105 }
106 
107 static int writeset_marked_on_disk(struct dm_disk_bitset *info,
108 				   struct writeset_metadata *m, dm_block_t block,
109 				   bool *result)
110 {
111 	dm_block_t old = m->root;
112 
113 	/*
114 	 * The bitset was flushed when it was archived, so we know there'll
115 	 * be no change to the root.
116 	 */
117 	int r = dm_bitset_test_bit(info, m->root, block, &m->root, result);
118 	if (r) {
119 		DMERR("%s: dm_bitset_test_bit failed", __func__);
120 		return r;
121 	}
122 
123 	BUG_ON(m->root != old);
124 
125 	return r;
126 }
127 
128 /*
129  * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was.
130  */
131 static int writeset_test_and_set(struct dm_disk_bitset *info,
132 				 struct writeset *ws, uint32_t block)
133 {
134 	int r;
135 
136 	if (!test_and_set_bit(block, ws->bits)) {
137 		r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root);
138 		if (r) {
139 			/* FIXME: fail mode */
140 			return r;
141 		}
142 
143 		return 0;
144 	}
145 
146 	return 1;
147 }
148 
149 /*----------------------------------------------------------------
150  * On disk metadata layout
151  *--------------------------------------------------------------*/
152 #define SPACE_MAP_ROOT_SIZE 128
153 #define UUID_LEN 16
154 
155 struct writeset_disk {
156 	__le32 nr_bits;
157 	__le64 root;
158 } __packed;
159 
160 struct superblock_disk {
161 	__le32 csum;
162 	__le32 flags;
163 	__le64 blocknr;
164 
165 	__u8 uuid[UUID_LEN];
166 	__le64 magic;
167 	__le32 version;
168 
169 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
170 
171 	__le32 data_block_size;
172 	__le32 metadata_block_size;
173 	__le32 nr_blocks;
174 
175 	__le32 current_era;
176 	struct writeset_disk current_writeset;
177 
178 	/*
179 	 * Only these two fields are valid within the metadata snapshot.
180 	 */
181 	__le64 writeset_tree_root;
182 	__le64 era_array_root;
183 
184 	__le64 metadata_snap;
185 } __packed;
186 
187 /*----------------------------------------------------------------
188  * Superblock validation
189  *--------------------------------------------------------------*/
190 static void sb_prepare_for_write(struct dm_block_validator *v,
191 				 struct dm_block *b,
192 				 size_t sb_block_size)
193 {
194 	struct superblock_disk *disk = dm_block_data(b);
195 
196 	disk->blocknr = cpu_to_le64(dm_block_location(b));
197 	disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags,
198 						sb_block_size - sizeof(__le32),
199 						SUPERBLOCK_CSUM_XOR));
200 }
201 
202 static int check_metadata_version(struct superblock_disk *disk)
203 {
204 	uint32_t metadata_version = le32_to_cpu(disk->version);
205 	if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) {
206 		DMERR("Era metadata version %u found, but only versions between %u and %u supported.",
207 		      metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION);
208 		return -EINVAL;
209 	}
210 
211 	return 0;
212 }
213 
214 static int sb_check(struct dm_block_validator *v,
215 		    struct dm_block *b,
216 		    size_t sb_block_size)
217 {
218 	struct superblock_disk *disk = dm_block_data(b);
219 	__le32 csum_le;
220 
221 	if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) {
222 		DMERR("sb_check failed: blocknr %llu: wanted %llu",
223 		      le64_to_cpu(disk->blocknr),
224 		      (unsigned long long)dm_block_location(b));
225 		return -ENOTBLK;
226 	}
227 
228 	if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) {
229 		DMERR("sb_check failed: magic %llu: wanted %llu",
230 		      le64_to_cpu(disk->magic),
231 		      (unsigned long long) SUPERBLOCK_MAGIC);
232 		return -EILSEQ;
233 	}
234 
235 	csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags,
236 					     sb_block_size - sizeof(__le32),
237 					     SUPERBLOCK_CSUM_XOR));
238 	if (csum_le != disk->csum) {
239 		DMERR("sb_check failed: csum %u: wanted %u",
240 		      le32_to_cpu(csum_le), le32_to_cpu(disk->csum));
241 		return -EILSEQ;
242 	}
243 
244 	return check_metadata_version(disk);
245 }
246 
247 static struct dm_block_validator sb_validator = {
248 	.name = "superblock",
249 	.prepare_for_write = sb_prepare_for_write,
250 	.check = sb_check
251 };
252 
253 /*----------------------------------------------------------------
254  * Low level metadata handling
255  *--------------------------------------------------------------*/
256 #define DM_ERA_METADATA_BLOCK_SIZE 4096
257 #define DM_ERA_METADATA_CACHE_SIZE 64
258 #define ERA_MAX_CONCURRENT_LOCKS 5
259 
260 struct era_metadata {
261 	struct block_device *bdev;
262 	struct dm_block_manager *bm;
263 	struct dm_space_map *sm;
264 	struct dm_transaction_manager *tm;
265 
266 	dm_block_t block_size;
267 	uint32_t nr_blocks;
268 
269 	uint32_t current_era;
270 
271 	/*
272 	 * We preallocate 2 writesets.  When an era rolls over we
273 	 * switch between them. This means the allocation is done at
274 	 * preresume time, rather than on the io path.
275 	 */
276 	struct writeset writesets[2];
277 	struct writeset *current_writeset;
278 
279 	dm_block_t writeset_tree_root;
280 	dm_block_t era_array_root;
281 
282 	struct dm_disk_bitset bitset_info;
283 	struct dm_btree_info writeset_tree_info;
284 	struct dm_array_info era_array_info;
285 
286 	dm_block_t metadata_snap;
287 
288 	/*
289 	 * A flag that is set whenever a writeset has been archived.
290 	 */
291 	bool archived_writesets;
292 
293 	/*
294 	 * Reading the space map root can fail, so we read it into this
295 	 * buffer before the superblock is locked and updated.
296 	 */
297 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
298 };
299 
300 static int superblock_read_lock(struct era_metadata *md,
301 				struct dm_block **sblock)
302 {
303 	return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION,
304 			       &sb_validator, sblock);
305 }
306 
307 static int superblock_lock_zero(struct era_metadata *md,
308 				struct dm_block **sblock)
309 {
310 	return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION,
311 				     &sb_validator, sblock);
312 }
313 
314 static int superblock_lock(struct era_metadata *md,
315 			   struct dm_block **sblock)
316 {
317 	return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION,
318 				&sb_validator, sblock);
319 }
320 
321 /* FIXME: duplication with cache and thin */
322 static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
323 {
324 	int r;
325 	unsigned i;
326 	struct dm_block *b;
327 	__le64 *data_le, zero = cpu_to_le64(0);
328 	unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
329 
330 	/*
331 	 * We can't use a validator here - it may be all zeroes.
332 	 */
333 	r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b);
334 	if (r)
335 		return r;
336 
337 	data_le = dm_block_data(b);
338 	*result = true;
339 	for (i = 0; i < sb_block_size; i++) {
340 		if (data_le[i] != zero) {
341 			*result = false;
342 			break;
343 		}
344 	}
345 
346 	return dm_bm_unlock(b);
347 }
348 
349 /*----------------------------------------------------------------*/
350 
351 static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk)
352 {
353 	disk->nr_bits = cpu_to_le32(core->nr_bits);
354 	disk->root = cpu_to_le64(core->root);
355 }
356 
357 static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core)
358 {
359 	core->nr_bits = le32_to_cpu(disk->nr_bits);
360 	core->root = le64_to_cpu(disk->root);
361 }
362 
363 static void ws_inc(void *context, const void *value)
364 {
365 	struct era_metadata *md = context;
366 	struct writeset_disk ws_d;
367 	dm_block_t b;
368 
369 	memcpy(&ws_d, value, sizeof(ws_d));
370 	b = le64_to_cpu(ws_d.root);
371 
372 	dm_tm_inc(md->tm, b);
373 }
374 
375 static void ws_dec(void *context, const void *value)
376 {
377 	struct era_metadata *md = context;
378 	struct writeset_disk ws_d;
379 	dm_block_t b;
380 
381 	memcpy(&ws_d, value, sizeof(ws_d));
382 	b = le64_to_cpu(ws_d.root);
383 
384 	dm_bitset_del(&md->bitset_info, b);
385 }
386 
387 static int ws_eq(void *context, const void *value1, const void *value2)
388 {
389 	return !memcmp(value1, value2, sizeof(struct writeset_metadata));
390 }
391 
392 /*----------------------------------------------------------------*/
393 
394 static void setup_writeset_tree_info(struct era_metadata *md)
395 {
396 	struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type;
397 	md->writeset_tree_info.tm = md->tm;
398 	md->writeset_tree_info.levels = 1;
399 	vt->context = md;
400 	vt->size = sizeof(struct writeset_disk);
401 	vt->inc = ws_inc;
402 	vt->dec = ws_dec;
403 	vt->equal = ws_eq;
404 }
405 
406 static void setup_era_array_info(struct era_metadata *md)
407 
408 {
409 	struct dm_btree_value_type vt;
410 	vt.context = NULL;
411 	vt.size = sizeof(__le32);
412 	vt.inc = NULL;
413 	vt.dec = NULL;
414 	vt.equal = NULL;
415 
416 	dm_array_info_init(&md->era_array_info, md->tm, &vt);
417 }
418 
419 static void setup_infos(struct era_metadata *md)
420 {
421 	dm_disk_bitset_init(md->tm, &md->bitset_info);
422 	setup_writeset_tree_info(md);
423 	setup_era_array_info(md);
424 }
425 
426 /*----------------------------------------------------------------*/
427 
428 static int create_fresh_metadata(struct era_metadata *md)
429 {
430 	int r;
431 
432 	r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION,
433 				 &md->tm, &md->sm);
434 	if (r < 0) {
435 		DMERR("dm_tm_create_with_sm failed");
436 		return r;
437 	}
438 
439 	setup_infos(md);
440 
441 	r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root);
442 	if (r) {
443 		DMERR("couldn't create new writeset tree");
444 		goto bad;
445 	}
446 
447 	r = dm_array_empty(&md->era_array_info, &md->era_array_root);
448 	if (r) {
449 		DMERR("couldn't create era array");
450 		goto bad;
451 	}
452 
453 	return 0;
454 
455 bad:
456 	dm_sm_destroy(md->sm);
457 	dm_tm_destroy(md->tm);
458 
459 	return r;
460 }
461 
462 static int save_sm_root(struct era_metadata *md)
463 {
464 	int r;
465 	size_t metadata_len;
466 
467 	r = dm_sm_root_size(md->sm, &metadata_len);
468 	if (r < 0)
469 		return r;
470 
471 	return dm_sm_copy_root(md->sm, &md->metadata_space_map_root,
472 			       metadata_len);
473 }
474 
475 static void copy_sm_root(struct era_metadata *md, struct superblock_disk *disk)
476 {
477 	memcpy(&disk->metadata_space_map_root,
478 	       &md->metadata_space_map_root,
479 	       sizeof(md->metadata_space_map_root));
480 }
481 
482 /*
483  * Writes a superblock, including the static fields that don't get updated
484  * with every commit (possible optimisation here).  'md' should be fully
485  * constructed when this is called.
486  */
487 static void prepare_superblock(struct era_metadata *md, struct superblock_disk *disk)
488 {
489 	disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
490 	disk->flags = cpu_to_le32(0ul);
491 
492 	/* FIXME: can't keep blanking the uuid (uuid is currently unused though) */
493 	memset(disk->uuid, 0, sizeof(disk->uuid));
494 	disk->version = cpu_to_le32(MAX_ERA_VERSION);
495 
496 	copy_sm_root(md, disk);
497 
498 	disk->data_block_size = cpu_to_le32(md->block_size);
499 	disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
500 	disk->nr_blocks = cpu_to_le32(md->nr_blocks);
501 	disk->current_era = cpu_to_le32(md->current_era);
502 
503 	ws_pack(&md->current_writeset->md, &disk->current_writeset);
504 	disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root);
505 	disk->era_array_root = cpu_to_le64(md->era_array_root);
506 	disk->metadata_snap = cpu_to_le64(md->metadata_snap);
507 }
508 
509 static int write_superblock(struct era_metadata *md)
510 {
511 	int r;
512 	struct dm_block *sblock;
513 	struct superblock_disk *disk;
514 
515 	r = save_sm_root(md);
516 	if (r) {
517 		DMERR("%s: save_sm_root failed", __func__);
518 		return r;
519 	}
520 
521 	r = superblock_lock_zero(md, &sblock);
522 	if (r)
523 		return r;
524 
525 	disk = dm_block_data(sblock);
526 	prepare_superblock(md, disk);
527 
528 	return dm_tm_commit(md->tm, sblock);
529 }
530 
531 /*
532  * Assumes block_size and the infos are set.
533  */
534 static int format_metadata(struct era_metadata *md)
535 {
536 	int r;
537 
538 	r = create_fresh_metadata(md);
539 	if (r)
540 		return r;
541 
542 	r = write_superblock(md);
543 	if (r) {
544 		dm_sm_destroy(md->sm);
545 		dm_tm_destroy(md->tm);
546 		return r;
547 	}
548 
549 	return 0;
550 }
551 
552 static int open_metadata(struct era_metadata *md)
553 {
554 	int r;
555 	struct dm_block *sblock;
556 	struct superblock_disk *disk;
557 
558 	r = superblock_read_lock(md, &sblock);
559 	if (r) {
560 		DMERR("couldn't read_lock superblock");
561 		return r;
562 	}
563 
564 	disk = dm_block_data(sblock);
565 	r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION,
566 			       disk->metadata_space_map_root,
567 			       sizeof(disk->metadata_space_map_root),
568 			       &md->tm, &md->sm);
569 	if (r) {
570 		DMERR("dm_tm_open_with_sm failed");
571 		goto bad;
572 	}
573 
574 	setup_infos(md);
575 
576 	md->block_size = le32_to_cpu(disk->data_block_size);
577 	md->nr_blocks = le32_to_cpu(disk->nr_blocks);
578 	md->current_era = le32_to_cpu(disk->current_era);
579 
580 	md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root);
581 	md->era_array_root = le64_to_cpu(disk->era_array_root);
582 	md->metadata_snap = le64_to_cpu(disk->metadata_snap);
583 	md->archived_writesets = true;
584 
585 	return dm_bm_unlock(sblock);
586 
587 bad:
588 	dm_bm_unlock(sblock);
589 	return r;
590 }
591 
592 static int open_or_format_metadata(struct era_metadata *md,
593 				   bool may_format)
594 {
595 	int r;
596 	bool unformatted = false;
597 
598 	r = superblock_all_zeroes(md->bm, &unformatted);
599 	if (r)
600 		return r;
601 
602 	if (unformatted)
603 		return may_format ? format_metadata(md) : -EPERM;
604 
605 	return open_metadata(md);
606 }
607 
608 static int create_persistent_data_objects(struct era_metadata *md,
609 					  bool may_format)
610 {
611 	int r;
612 
613 	md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE,
614 					 DM_ERA_METADATA_CACHE_SIZE,
615 					 ERA_MAX_CONCURRENT_LOCKS);
616 	if (IS_ERR(md->bm)) {
617 		DMERR("could not create block manager");
618 		return PTR_ERR(md->bm);
619 	}
620 
621 	r = open_or_format_metadata(md, may_format);
622 	if (r)
623 		dm_block_manager_destroy(md->bm);
624 
625 	return r;
626 }
627 
628 static void destroy_persistent_data_objects(struct era_metadata *md)
629 {
630 	dm_sm_destroy(md->sm);
631 	dm_tm_destroy(md->tm);
632 	dm_block_manager_destroy(md->bm);
633 }
634 
635 /*
636  * This waits until all era_map threads have picked up the new filter.
637  */
638 static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset)
639 {
640 	rcu_assign_pointer(md->current_writeset, new_writeset);
641 	synchronize_rcu();
642 }
643 
644 /*----------------------------------------------------------------
645  * Writesets get 'digested' into the main era array.
646  *
647  * We're using a coroutine here so the worker thread can do the digestion,
648  * thus avoiding synchronisation of the metadata.  Digesting a whole
649  * writeset in one go would cause too much latency.
650  *--------------------------------------------------------------*/
651 struct digest {
652 	uint32_t era;
653 	unsigned nr_bits, current_bit;
654 	struct writeset_metadata writeset;
655 	__le32 value;
656 	struct dm_disk_bitset info;
657 
658 	int (*step)(struct era_metadata *, struct digest *);
659 };
660 
661 static int metadata_digest_lookup_writeset(struct era_metadata *md,
662 					   struct digest *d);
663 
664 static int metadata_digest_remove_writeset(struct era_metadata *md,
665 					   struct digest *d)
666 {
667 	int r;
668 	uint64_t key = d->era;
669 
670 	r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root,
671 			    &key, &md->writeset_tree_root);
672 	if (r) {
673 		DMERR("%s: dm_btree_remove failed", __func__);
674 		return r;
675 	}
676 
677 	d->step = metadata_digest_lookup_writeset;
678 	return 0;
679 }
680 
681 #define INSERTS_PER_STEP 100
682 
683 static int metadata_digest_transcribe_writeset(struct era_metadata *md,
684 					       struct digest *d)
685 {
686 	int r;
687 	bool marked;
688 	unsigned b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits);
689 
690 	for (b = d->current_bit; b < e; b++) {
691 		r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked);
692 		if (r) {
693 			DMERR("%s: writeset_marked_on_disk failed", __func__);
694 			return r;
695 		}
696 
697 		if (!marked)
698 			continue;
699 
700 		__dm_bless_for_disk(&d->value);
701 		r = dm_array_set_value(&md->era_array_info, md->era_array_root,
702 				       b, &d->value, &md->era_array_root);
703 		if (r) {
704 			DMERR("%s: dm_array_set_value failed", __func__);
705 			return r;
706 		}
707 	}
708 
709 	if (b == d->nr_bits)
710 		d->step = metadata_digest_remove_writeset;
711 	else
712 		d->current_bit = b;
713 
714 	return 0;
715 }
716 
717 static int metadata_digest_lookup_writeset(struct era_metadata *md,
718 					   struct digest *d)
719 {
720 	int r;
721 	uint64_t key;
722 	struct writeset_disk disk;
723 
724 	r = dm_btree_find_lowest_key(&md->writeset_tree_info,
725 				     md->writeset_tree_root, &key);
726 	if (r < 0)
727 		return r;
728 
729 	d->era = key;
730 
731 	r = dm_btree_lookup(&md->writeset_tree_info,
732 			    md->writeset_tree_root, &key, &disk);
733 	if (r) {
734 		if (r == -ENODATA) {
735 			d->step = NULL;
736 			return 0;
737 		}
738 
739 		DMERR("%s: dm_btree_lookup failed", __func__);
740 		return r;
741 	}
742 
743 	ws_unpack(&disk, &d->writeset);
744 	d->value = cpu_to_le32(key);
745 
746 	d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks);
747 	d->current_bit = 0;
748 	d->step = metadata_digest_transcribe_writeset;
749 
750 	return 0;
751 }
752 
753 static int metadata_digest_start(struct era_metadata *md, struct digest *d)
754 {
755 	if (d->step)
756 		return 0;
757 
758 	memset(d, 0, sizeof(*d));
759 
760 	/*
761 	 * We initialise another bitset info to avoid any caching side
762 	 * effects with the previous one.
763 	 */
764 	dm_disk_bitset_init(md->tm, &d->info);
765 	d->step = metadata_digest_lookup_writeset;
766 
767 	return 0;
768 }
769 
770 /*----------------------------------------------------------------
771  * High level metadata interface.  Target methods should use these, and not
772  * the lower level ones.
773  *--------------------------------------------------------------*/
774 static struct era_metadata *metadata_open(struct block_device *bdev,
775 					  sector_t block_size,
776 					  bool may_format)
777 {
778 	int r;
779 	struct era_metadata *md = kzalloc(sizeof(*md), GFP_KERNEL);
780 
781 	if (!md)
782 		return NULL;
783 
784 	md->bdev = bdev;
785 	md->block_size = block_size;
786 
787 	md->writesets[0].md.root = INVALID_WRITESET_ROOT;
788 	md->writesets[1].md.root = INVALID_WRITESET_ROOT;
789 	md->current_writeset = &md->writesets[0];
790 
791 	r = create_persistent_data_objects(md, may_format);
792 	if (r) {
793 		kfree(md);
794 		return ERR_PTR(r);
795 	}
796 
797 	return md;
798 }
799 
800 static void metadata_close(struct era_metadata *md)
801 {
802 	destroy_persistent_data_objects(md);
803 	kfree(md);
804 }
805 
806 static bool valid_nr_blocks(dm_block_t n)
807 {
808 	/*
809 	 * dm_bitset restricts us to 2^32.  test_bit & co. restrict us
810 	 * further to 2^31 - 1
811 	 */
812 	return n < (1ull << 31);
813 }
814 
815 static int metadata_resize(struct era_metadata *md, void *arg)
816 {
817 	int r;
818 	dm_block_t *new_size = arg;
819 	__le32 value;
820 
821 	if (!valid_nr_blocks(*new_size)) {
822 		DMERR("Invalid number of origin blocks %llu",
823 		      (unsigned long long) *new_size);
824 		return -EINVAL;
825 	}
826 
827 	writeset_free(&md->writesets[0]);
828 	writeset_free(&md->writesets[1]);
829 
830 	r = writeset_alloc(&md->writesets[0], *new_size);
831 	if (r) {
832 		DMERR("%s: writeset_alloc failed for writeset 0", __func__);
833 		return r;
834 	}
835 
836 	r = writeset_alloc(&md->writesets[1], *new_size);
837 	if (r) {
838 		DMERR("%s: writeset_alloc failed for writeset 1", __func__);
839 		return r;
840 	}
841 
842 	value = cpu_to_le32(0u);
843 	__dm_bless_for_disk(&value);
844 	r = dm_array_resize(&md->era_array_info, md->era_array_root,
845 			    md->nr_blocks, *new_size,
846 			    &value, &md->era_array_root);
847 	if (r) {
848 		DMERR("%s: dm_array_resize failed", __func__);
849 		return r;
850 	}
851 
852 	md->nr_blocks = *new_size;
853 	return 0;
854 }
855 
856 static int metadata_era_archive(struct era_metadata *md)
857 {
858 	int r;
859 	uint64_t keys[1];
860 	struct writeset_disk value;
861 
862 	r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
863 			    &md->current_writeset->md.root);
864 	if (r) {
865 		DMERR("%s: dm_bitset_flush failed", __func__);
866 		return r;
867 	}
868 
869 	ws_pack(&md->current_writeset->md, &value);
870 	md->current_writeset->md.root = INVALID_WRITESET_ROOT;
871 
872 	keys[0] = md->current_era;
873 	__dm_bless_for_disk(&value);
874 	r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root,
875 			    keys, &value, &md->writeset_tree_root);
876 	if (r) {
877 		DMERR("%s: couldn't insert writeset into btree", __func__);
878 		/* FIXME: fail mode */
879 		return r;
880 	}
881 
882 	md->archived_writesets = true;
883 
884 	return 0;
885 }
886 
887 static struct writeset *next_writeset(struct era_metadata *md)
888 {
889 	return (md->current_writeset == &md->writesets[0]) ?
890 		&md->writesets[1] : &md->writesets[0];
891 }
892 
893 static int metadata_new_era(struct era_metadata *md)
894 {
895 	int r;
896 	struct writeset *new_writeset = next_writeset(md);
897 
898 	r = writeset_init(&md->bitset_info, new_writeset);
899 	if (r) {
900 		DMERR("%s: writeset_init failed", __func__);
901 		return r;
902 	}
903 
904 	swap_writeset(md, new_writeset);
905 	md->current_era++;
906 
907 	return 0;
908 }
909 
910 static int metadata_era_rollover(struct era_metadata *md)
911 {
912 	int r;
913 
914 	if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
915 		r = metadata_era_archive(md);
916 		if (r) {
917 			DMERR("%s: metadata_archive_era failed", __func__);
918 			/* FIXME: fail mode? */
919 			return r;
920 		}
921 	}
922 
923 	r = metadata_new_era(md);
924 	if (r) {
925 		DMERR("%s: new era failed", __func__);
926 		/* FIXME: fail mode */
927 		return r;
928 	}
929 
930 	return 0;
931 }
932 
933 static bool metadata_current_marked(struct era_metadata *md, dm_block_t block)
934 {
935 	bool r;
936 	struct writeset *ws;
937 
938 	rcu_read_lock();
939 	ws = rcu_dereference(md->current_writeset);
940 	r = writeset_marked(ws, block);
941 	rcu_read_unlock();
942 
943 	return r;
944 }
945 
946 static int metadata_commit(struct era_metadata *md)
947 {
948 	int r;
949 	struct dm_block *sblock;
950 
951 	if (md->current_writeset->md.root != SUPERBLOCK_LOCATION) {
952 		r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
953 				    &md->current_writeset->md.root);
954 		if (r) {
955 			DMERR("%s: bitset flush failed", __func__);
956 			return r;
957 		}
958 	}
959 
960 	r = save_sm_root(md);
961 	if (r) {
962 		DMERR("%s: save_sm_root failed", __func__);
963 		return r;
964 	}
965 
966 	r = dm_tm_pre_commit(md->tm);
967 	if (r) {
968 		DMERR("%s: pre commit failed", __func__);
969 		return r;
970 	}
971 
972 	r = superblock_lock(md, &sblock);
973 	if (r) {
974 		DMERR("%s: superblock lock failed", __func__);
975 		return r;
976 	}
977 
978 	prepare_superblock(md, dm_block_data(sblock));
979 
980 	return dm_tm_commit(md->tm, sblock);
981 }
982 
983 static int metadata_checkpoint(struct era_metadata *md)
984 {
985 	/*
986 	 * For now we just rollover, but later I want to put a check in to
987 	 * avoid this if the filter is still pretty fresh.
988 	 */
989 	return metadata_era_rollover(md);
990 }
991 
992 /*
993  * Metadata snapshots allow userland to access era data.
994  */
995 static int metadata_take_snap(struct era_metadata *md)
996 {
997 	int r, inc;
998 	struct dm_block *clone;
999 
1000 	if (md->metadata_snap != SUPERBLOCK_LOCATION) {
1001 		DMERR("%s: metadata snapshot already exists", __func__);
1002 		return -EINVAL;
1003 	}
1004 
1005 	r = metadata_era_rollover(md);
1006 	if (r) {
1007 		DMERR("%s: era rollover failed", __func__);
1008 		return r;
1009 	}
1010 
1011 	r = metadata_commit(md);
1012 	if (r) {
1013 		DMERR("%s: pre commit failed", __func__);
1014 		return r;
1015 	}
1016 
1017 	r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION);
1018 	if (r) {
1019 		DMERR("%s: couldn't increment superblock", __func__);
1020 		return r;
1021 	}
1022 
1023 	r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION,
1024 			       &sb_validator, &clone, &inc);
1025 	if (r) {
1026 		DMERR("%s: couldn't shadow superblock", __func__);
1027 		dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION);
1028 		return r;
1029 	}
1030 	BUG_ON(!inc);
1031 
1032 	r = dm_sm_inc_block(md->sm, md->writeset_tree_root);
1033 	if (r) {
1034 		DMERR("%s: couldn't inc writeset tree root", __func__);
1035 		dm_tm_unlock(md->tm, clone);
1036 		return r;
1037 	}
1038 
1039 	r = dm_sm_inc_block(md->sm, md->era_array_root);
1040 	if (r) {
1041 		DMERR("%s: couldn't inc era tree root", __func__);
1042 		dm_sm_dec_block(md->sm, md->writeset_tree_root);
1043 		dm_tm_unlock(md->tm, clone);
1044 		return r;
1045 	}
1046 
1047 	md->metadata_snap = dm_block_location(clone);
1048 
1049 	r = dm_tm_unlock(md->tm, clone);
1050 	if (r) {
1051 		DMERR("%s: couldn't unlock clone", __func__);
1052 		md->metadata_snap = SUPERBLOCK_LOCATION;
1053 		return r;
1054 	}
1055 
1056 	return 0;
1057 }
1058 
1059 static int metadata_drop_snap(struct era_metadata *md)
1060 {
1061 	int r;
1062 	dm_block_t location;
1063 	struct dm_block *clone;
1064 	struct superblock_disk *disk;
1065 
1066 	if (md->metadata_snap == SUPERBLOCK_LOCATION) {
1067 		DMERR("%s: no snap to drop", __func__);
1068 		return -EINVAL;
1069 	}
1070 
1071 	r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone);
1072 	if (r) {
1073 		DMERR("%s: couldn't read lock superblock clone", __func__);
1074 		return r;
1075 	}
1076 
1077 	/*
1078 	 * Whatever happens now we'll commit with no record of the metadata
1079 	 * snap.
1080 	 */
1081 	md->metadata_snap = SUPERBLOCK_LOCATION;
1082 
1083 	disk = dm_block_data(clone);
1084 	r = dm_btree_del(&md->writeset_tree_info,
1085 			 le64_to_cpu(disk->writeset_tree_root));
1086 	if (r) {
1087 		DMERR("%s: error deleting writeset tree clone", __func__);
1088 		dm_tm_unlock(md->tm, clone);
1089 		return r;
1090 	}
1091 
1092 	r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root));
1093 	if (r) {
1094 		DMERR("%s: error deleting era array clone", __func__);
1095 		dm_tm_unlock(md->tm, clone);
1096 		return r;
1097 	}
1098 
1099 	location = dm_block_location(clone);
1100 	dm_tm_unlock(md->tm, clone);
1101 
1102 	return dm_sm_dec_block(md->sm, location);
1103 }
1104 
1105 struct metadata_stats {
1106 	dm_block_t used;
1107 	dm_block_t total;
1108 	dm_block_t snap;
1109 	uint32_t era;
1110 };
1111 
1112 static int metadata_get_stats(struct era_metadata *md, void *ptr)
1113 {
1114 	int r;
1115 	struct metadata_stats *s = ptr;
1116 	dm_block_t nr_free, nr_total;
1117 
1118 	r = dm_sm_get_nr_free(md->sm, &nr_free);
1119 	if (r) {
1120 		DMERR("dm_sm_get_nr_free returned %d", r);
1121 		return r;
1122 	}
1123 
1124 	r = dm_sm_get_nr_blocks(md->sm, &nr_total);
1125 	if (r) {
1126 		DMERR("dm_pool_get_metadata_dev_size returned %d", r);
1127 		return r;
1128 	}
1129 
1130 	s->used = nr_total - nr_free;
1131 	s->total = nr_total;
1132 	s->snap = md->metadata_snap;
1133 	s->era = md->current_era;
1134 
1135 	return 0;
1136 }
1137 
1138 /*----------------------------------------------------------------*/
1139 
1140 struct era {
1141 	struct dm_target *ti;
1142 	struct dm_target_callbacks callbacks;
1143 
1144 	struct dm_dev *metadata_dev;
1145 	struct dm_dev *origin_dev;
1146 
1147 	dm_block_t nr_blocks;
1148 	uint32_t sectors_per_block;
1149 	int sectors_per_block_shift;
1150 	struct era_metadata *md;
1151 
1152 	struct workqueue_struct *wq;
1153 	struct work_struct worker;
1154 
1155 	spinlock_t deferred_lock;
1156 	struct bio_list deferred_bios;
1157 
1158 	spinlock_t rpc_lock;
1159 	struct list_head rpc_calls;
1160 
1161 	struct digest digest;
1162 	atomic_t suspended;
1163 };
1164 
1165 struct rpc {
1166 	struct list_head list;
1167 
1168 	int (*fn0)(struct era_metadata *);
1169 	int (*fn1)(struct era_metadata *, void *);
1170 	void *arg;
1171 	int result;
1172 
1173 	struct completion complete;
1174 };
1175 
1176 /*----------------------------------------------------------------
1177  * Remapping.
1178  *---------------------------------------------------------------*/
1179 static bool block_size_is_power_of_two(struct era *era)
1180 {
1181 	return era->sectors_per_block_shift >= 0;
1182 }
1183 
1184 static dm_block_t get_block(struct era *era, struct bio *bio)
1185 {
1186 	sector_t block_nr = bio->bi_iter.bi_sector;
1187 
1188 	if (!block_size_is_power_of_two(era))
1189 		(void) sector_div(block_nr, era->sectors_per_block);
1190 	else
1191 		block_nr >>= era->sectors_per_block_shift;
1192 
1193 	return block_nr;
1194 }
1195 
1196 static void remap_to_origin(struct era *era, struct bio *bio)
1197 {
1198 	bio->bi_bdev = era->origin_dev->bdev;
1199 }
1200 
1201 /*----------------------------------------------------------------
1202  * Worker thread
1203  *--------------------------------------------------------------*/
1204 static void wake_worker(struct era *era)
1205 {
1206 	if (!atomic_read(&era->suspended))
1207 		queue_work(era->wq, &era->worker);
1208 }
1209 
1210 static void process_old_eras(struct era *era)
1211 {
1212 	int r;
1213 
1214 	if (!era->digest.step)
1215 		return;
1216 
1217 	r = era->digest.step(era->md, &era->digest);
1218 	if (r < 0) {
1219 		DMERR("%s: digest step failed, stopping digestion", __func__);
1220 		era->digest.step = NULL;
1221 
1222 	} else if (era->digest.step)
1223 		wake_worker(era);
1224 }
1225 
1226 static void process_deferred_bios(struct era *era)
1227 {
1228 	int r;
1229 	struct bio_list deferred_bios, marked_bios;
1230 	struct bio *bio;
1231 	bool commit_needed = false;
1232 	bool failed = false;
1233 
1234 	bio_list_init(&deferred_bios);
1235 	bio_list_init(&marked_bios);
1236 
1237 	spin_lock(&era->deferred_lock);
1238 	bio_list_merge(&deferred_bios, &era->deferred_bios);
1239 	bio_list_init(&era->deferred_bios);
1240 	spin_unlock(&era->deferred_lock);
1241 
1242 	while ((bio = bio_list_pop(&deferred_bios))) {
1243 		r = writeset_test_and_set(&era->md->bitset_info,
1244 					  era->md->current_writeset,
1245 					  get_block(era, bio));
1246 		if (r < 0) {
1247 			/*
1248 			 * This is bad news, we need to rollback.
1249 			 * FIXME: finish.
1250 			 */
1251 			failed = true;
1252 
1253 		} else if (r == 0)
1254 			commit_needed = true;
1255 
1256 		bio_list_add(&marked_bios, bio);
1257 	}
1258 
1259 	if (commit_needed) {
1260 		r = metadata_commit(era->md);
1261 		if (r)
1262 			failed = true;
1263 	}
1264 
1265 	if (failed)
1266 		while ((bio = bio_list_pop(&marked_bios)))
1267 			bio_io_error(bio);
1268 	else
1269 		while ((bio = bio_list_pop(&marked_bios)))
1270 			generic_make_request(bio);
1271 }
1272 
1273 static void process_rpc_calls(struct era *era)
1274 {
1275 	int r;
1276 	bool need_commit = false;
1277 	struct list_head calls;
1278 	struct rpc *rpc, *tmp;
1279 
1280 	INIT_LIST_HEAD(&calls);
1281 	spin_lock(&era->rpc_lock);
1282 	list_splice_init(&era->rpc_calls, &calls);
1283 	spin_unlock(&era->rpc_lock);
1284 
1285 	list_for_each_entry_safe(rpc, tmp, &calls, list) {
1286 		rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg);
1287 		need_commit = true;
1288 	}
1289 
1290 	if (need_commit) {
1291 		r = metadata_commit(era->md);
1292 		if (r)
1293 			list_for_each_entry_safe(rpc, tmp, &calls, list)
1294 				rpc->result = r;
1295 	}
1296 
1297 	list_for_each_entry_safe(rpc, tmp, &calls, list)
1298 		complete(&rpc->complete);
1299 }
1300 
1301 static void kick_off_digest(struct era *era)
1302 {
1303 	if (era->md->archived_writesets) {
1304 		era->md->archived_writesets = false;
1305 		metadata_digest_start(era->md, &era->digest);
1306 	}
1307 }
1308 
1309 static void do_work(struct work_struct *ws)
1310 {
1311 	struct era *era = container_of(ws, struct era, worker);
1312 
1313 	kick_off_digest(era);
1314 	process_old_eras(era);
1315 	process_deferred_bios(era);
1316 	process_rpc_calls(era);
1317 }
1318 
1319 static void defer_bio(struct era *era, struct bio *bio)
1320 {
1321 	spin_lock(&era->deferred_lock);
1322 	bio_list_add(&era->deferred_bios, bio);
1323 	spin_unlock(&era->deferred_lock);
1324 
1325 	wake_worker(era);
1326 }
1327 
1328 /*
1329  * Make an rpc call to the worker to change the metadata.
1330  */
1331 static int perform_rpc(struct era *era, struct rpc *rpc)
1332 {
1333 	rpc->result = 0;
1334 	init_completion(&rpc->complete);
1335 
1336 	spin_lock(&era->rpc_lock);
1337 	list_add(&rpc->list, &era->rpc_calls);
1338 	spin_unlock(&era->rpc_lock);
1339 
1340 	wake_worker(era);
1341 	wait_for_completion(&rpc->complete);
1342 
1343 	return rpc->result;
1344 }
1345 
1346 static int in_worker0(struct era *era, int (*fn)(struct era_metadata *))
1347 {
1348 	struct rpc rpc;
1349 	rpc.fn0 = fn;
1350 	rpc.fn1 = NULL;
1351 
1352 	return perform_rpc(era, &rpc);
1353 }
1354 
1355 static int in_worker1(struct era *era,
1356 		      int (*fn)(struct era_metadata *, void *), void *arg)
1357 {
1358 	struct rpc rpc;
1359 	rpc.fn0 = NULL;
1360 	rpc.fn1 = fn;
1361 	rpc.arg = arg;
1362 
1363 	return perform_rpc(era, &rpc);
1364 }
1365 
1366 static void start_worker(struct era *era)
1367 {
1368 	atomic_set(&era->suspended, 0);
1369 }
1370 
1371 static void stop_worker(struct era *era)
1372 {
1373 	atomic_set(&era->suspended, 1);
1374 	flush_workqueue(era->wq);
1375 }
1376 
1377 /*----------------------------------------------------------------
1378  * Target methods
1379  *--------------------------------------------------------------*/
1380 static int dev_is_congested(struct dm_dev *dev, int bdi_bits)
1381 {
1382 	struct request_queue *q = bdev_get_queue(dev->bdev);
1383 	return bdi_congested(&q->backing_dev_info, bdi_bits);
1384 }
1385 
1386 static int era_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
1387 {
1388 	struct era *era = container_of(cb, struct era, callbacks);
1389 	return dev_is_congested(era->origin_dev, bdi_bits);
1390 }
1391 
1392 static void era_destroy(struct era *era)
1393 {
1394 	if (era->md)
1395 		metadata_close(era->md);
1396 
1397 	if (era->wq)
1398 		destroy_workqueue(era->wq);
1399 
1400 	if (era->origin_dev)
1401 		dm_put_device(era->ti, era->origin_dev);
1402 
1403 	if (era->metadata_dev)
1404 		dm_put_device(era->ti, era->metadata_dev);
1405 
1406 	kfree(era);
1407 }
1408 
1409 static dm_block_t calc_nr_blocks(struct era *era)
1410 {
1411 	return dm_sector_div_up(era->ti->len, era->sectors_per_block);
1412 }
1413 
1414 static bool valid_block_size(dm_block_t block_size)
1415 {
1416 	bool greater_than_zero = block_size > 0;
1417 	bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0;
1418 
1419 	return greater_than_zero && multiple_of_min_block_size;
1420 }
1421 
1422 /*
1423  * <metadata dev> <data dev> <data block size (sectors)>
1424  */
1425 static int era_ctr(struct dm_target *ti, unsigned argc, char **argv)
1426 {
1427 	int r;
1428 	char dummy;
1429 	struct era *era;
1430 	struct era_metadata *md;
1431 
1432 	if (argc != 3) {
1433 		ti->error = "Invalid argument count";
1434 		return -EINVAL;
1435 	}
1436 
1437 	era = kzalloc(sizeof(*era), GFP_KERNEL);
1438 	if (!era) {
1439 		ti->error = "Error allocating era structure";
1440 		return -ENOMEM;
1441 	}
1442 
1443 	era->ti = ti;
1444 
1445 	r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &era->metadata_dev);
1446 	if (r) {
1447 		ti->error = "Error opening metadata device";
1448 		era_destroy(era);
1449 		return -EINVAL;
1450 	}
1451 
1452 	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &era->origin_dev);
1453 	if (r) {
1454 		ti->error = "Error opening data device";
1455 		era_destroy(era);
1456 		return -EINVAL;
1457 	}
1458 
1459 	r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy);
1460 	if (r != 1) {
1461 		ti->error = "Error parsing block size";
1462 		era_destroy(era);
1463 		return -EINVAL;
1464 	}
1465 
1466 	r = dm_set_target_max_io_len(ti, era->sectors_per_block);
1467 	if (r) {
1468 		ti->error = "could not set max io len";
1469 		era_destroy(era);
1470 		return -EINVAL;
1471 	}
1472 
1473 	if (!valid_block_size(era->sectors_per_block)) {
1474 		ti->error = "Invalid block size";
1475 		era_destroy(era);
1476 		return -EINVAL;
1477 	}
1478 	if (era->sectors_per_block & (era->sectors_per_block - 1))
1479 		era->sectors_per_block_shift = -1;
1480 	else
1481 		era->sectors_per_block_shift = __ffs(era->sectors_per_block);
1482 
1483 	md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true);
1484 	if (IS_ERR(md)) {
1485 		ti->error = "Error reading metadata";
1486 		era_destroy(era);
1487 		return PTR_ERR(md);
1488 	}
1489 	era->md = md;
1490 
1491 	era->nr_blocks = calc_nr_blocks(era);
1492 
1493 	r = metadata_resize(era->md, &era->nr_blocks);
1494 	if (r) {
1495 		ti->error = "couldn't resize metadata";
1496 		era_destroy(era);
1497 		return -ENOMEM;
1498 	}
1499 
1500 	era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1501 	if (!era->wq) {
1502 		ti->error = "could not create workqueue for metadata object";
1503 		era_destroy(era);
1504 		return -ENOMEM;
1505 	}
1506 	INIT_WORK(&era->worker, do_work);
1507 
1508 	spin_lock_init(&era->deferred_lock);
1509 	bio_list_init(&era->deferred_bios);
1510 
1511 	spin_lock_init(&era->rpc_lock);
1512 	INIT_LIST_HEAD(&era->rpc_calls);
1513 
1514 	ti->private = era;
1515 	ti->num_flush_bios = 1;
1516 	ti->flush_supported = true;
1517 
1518 	ti->num_discard_bios = 1;
1519 	ti->discards_supported = true;
1520 	era->callbacks.congested_fn = era_is_congested;
1521 	dm_table_add_target_callbacks(ti->table, &era->callbacks);
1522 
1523 	return 0;
1524 }
1525 
1526 static void era_dtr(struct dm_target *ti)
1527 {
1528 	era_destroy(ti->private);
1529 }
1530 
1531 static int era_map(struct dm_target *ti, struct bio *bio)
1532 {
1533 	struct era *era = ti->private;
1534 	dm_block_t block = get_block(era, bio);
1535 
1536 	/*
1537 	 * All bios get remapped to the origin device.  We do this now, but
1538 	 * it may not get issued until later.  Depending on whether the
1539 	 * block is marked in this era.
1540 	 */
1541 	remap_to_origin(era, bio);
1542 
1543 	/*
1544 	 * REQ_FLUSH bios carry no data, so we're not interested in them.
1545 	 */
1546 	if (!(bio->bi_rw & REQ_FLUSH) &&
1547 	    (bio_data_dir(bio) == WRITE) &&
1548 	    !metadata_current_marked(era->md, block)) {
1549 		defer_bio(era, bio);
1550 		return DM_MAPIO_SUBMITTED;
1551 	}
1552 
1553 	return DM_MAPIO_REMAPPED;
1554 }
1555 
1556 static void era_postsuspend(struct dm_target *ti)
1557 {
1558 	int r;
1559 	struct era *era = ti->private;
1560 
1561 	r = in_worker0(era, metadata_era_archive);
1562 	if (r) {
1563 		DMERR("%s: couldn't archive current era", __func__);
1564 		/* FIXME: fail mode */
1565 	}
1566 
1567 	stop_worker(era);
1568 }
1569 
1570 static int era_preresume(struct dm_target *ti)
1571 {
1572 	int r;
1573 	struct era *era = ti->private;
1574 	dm_block_t new_size = calc_nr_blocks(era);
1575 
1576 	if (era->nr_blocks != new_size) {
1577 		r = in_worker1(era, metadata_resize, &new_size);
1578 		if (r)
1579 			return r;
1580 
1581 		era->nr_blocks = new_size;
1582 	}
1583 
1584 	start_worker(era);
1585 
1586 	r = in_worker0(era, metadata_new_era);
1587 	if (r) {
1588 		DMERR("%s: metadata_era_rollover failed", __func__);
1589 		return r;
1590 	}
1591 
1592 	return 0;
1593 }
1594 
1595 /*
1596  * Status format:
1597  *
1598  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
1599  * <current era> <held metadata root | '-'>
1600  */
1601 static void era_status(struct dm_target *ti, status_type_t type,
1602 		       unsigned status_flags, char *result, unsigned maxlen)
1603 {
1604 	int r;
1605 	struct era *era = ti->private;
1606 	ssize_t sz = 0;
1607 	struct metadata_stats stats;
1608 	char buf[BDEVNAME_SIZE];
1609 
1610 	switch (type) {
1611 	case STATUSTYPE_INFO:
1612 		r = in_worker1(era, metadata_get_stats, &stats);
1613 		if (r)
1614 			goto err;
1615 
1616 		DMEMIT("%u %llu/%llu %u",
1617 		       (unsigned) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
1618 		       (unsigned long long) stats.used,
1619 		       (unsigned long long) stats.total,
1620 		       (unsigned) stats.era);
1621 
1622 		if (stats.snap != SUPERBLOCK_LOCATION)
1623 			DMEMIT(" %llu", stats.snap);
1624 		else
1625 			DMEMIT(" -");
1626 		break;
1627 
1628 	case STATUSTYPE_TABLE:
1629 		format_dev_t(buf, era->metadata_dev->bdev->bd_dev);
1630 		DMEMIT("%s ", buf);
1631 		format_dev_t(buf, era->origin_dev->bdev->bd_dev);
1632 		DMEMIT("%s %u", buf, era->sectors_per_block);
1633 		break;
1634 	}
1635 
1636 	return;
1637 
1638 err:
1639 	DMEMIT("Error");
1640 }
1641 
1642 static int era_message(struct dm_target *ti, unsigned argc, char **argv)
1643 {
1644 	struct era *era = ti->private;
1645 
1646 	if (argc != 1) {
1647 		DMERR("incorrect number of message arguments");
1648 		return -EINVAL;
1649 	}
1650 
1651 	if (!strcasecmp(argv[0], "checkpoint"))
1652 		return in_worker0(era, metadata_checkpoint);
1653 
1654 	if (!strcasecmp(argv[0], "take_metadata_snap"))
1655 		return in_worker0(era, metadata_take_snap);
1656 
1657 	if (!strcasecmp(argv[0], "drop_metadata_snap"))
1658 		return in_worker0(era, metadata_drop_snap);
1659 
1660 	DMERR("unsupported message '%s'", argv[0]);
1661 	return -EINVAL;
1662 }
1663 
1664 static sector_t get_dev_size(struct dm_dev *dev)
1665 {
1666 	return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1667 }
1668 
1669 static int era_iterate_devices(struct dm_target *ti,
1670 			       iterate_devices_callout_fn fn, void *data)
1671 {
1672 	struct era *era = ti->private;
1673 	return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data);
1674 }
1675 
1676 static void era_io_hints(struct dm_target *ti, struct queue_limits *limits)
1677 {
1678 	struct era *era = ti->private;
1679 	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
1680 
1681 	/*
1682 	 * If the system-determined stacked limits are compatible with the
1683 	 * era device's blocksize (io_opt is a factor) do not override them.
1684 	 */
1685 	if (io_opt_sectors < era->sectors_per_block ||
1686 	    do_div(io_opt_sectors, era->sectors_per_block)) {
1687 		blk_limits_io_min(limits, 0);
1688 		blk_limits_io_opt(limits, era->sectors_per_block << SECTOR_SHIFT);
1689 	}
1690 }
1691 
1692 /*----------------------------------------------------------------*/
1693 
1694 static struct target_type era_target = {
1695 	.name = "era",
1696 	.version = {1, 0, 0},
1697 	.module = THIS_MODULE,
1698 	.ctr = era_ctr,
1699 	.dtr = era_dtr,
1700 	.map = era_map,
1701 	.postsuspend = era_postsuspend,
1702 	.preresume = era_preresume,
1703 	.status = era_status,
1704 	.message = era_message,
1705 	.iterate_devices = era_iterate_devices,
1706 	.io_hints = era_io_hints
1707 };
1708 
1709 static int __init dm_era_init(void)
1710 {
1711 	int r;
1712 
1713 	r = dm_register_target(&era_target);
1714 	if (r) {
1715 		DMERR("era target registration failed: %d", r);
1716 		return r;
1717 	}
1718 
1719 	return 0;
1720 }
1721 
1722 static void __exit dm_era_exit(void)
1723 {
1724 	dm_unregister_target(&era_target);
1725 }
1726 
1727 module_init(dm_era_init);
1728 module_exit(dm_era_exit);
1729 
1730 MODULE_DESCRIPTION(DM_NAME " era target");
1731 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
1732 MODULE_LICENSE("GPL");
1733