xref: /openbmc/linux/drivers/md/dm-era-target.c (revision 62eab49f)
1 // SPDX-License-Identifier: GPL-2.0-only
2 #include "dm.h"
3 #include "persistent-data/dm-transaction-manager.h"
4 #include "persistent-data/dm-bitset.h"
5 #include "persistent-data/dm-space-map.h"
6 
7 #include <linux/dm-io.h>
8 #include <linux/dm-kcopyd.h>
9 #include <linux/init.h>
10 #include <linux/mempool.h>
11 #include <linux/module.h>
12 #include <linux/slab.h>
13 #include <linux/vmalloc.h>
14 
15 #define DM_MSG_PREFIX "era"
16 
17 #define SUPERBLOCK_LOCATION 0
18 #define SUPERBLOCK_MAGIC 2126579579
19 #define SUPERBLOCK_CSUM_XOR 146538381
20 #define MIN_ERA_VERSION 1
21 #define MAX_ERA_VERSION 1
22 #define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION
23 #define MIN_BLOCK_SIZE 8
24 
25 /*----------------------------------------------------------------
26  * Writeset
27  *--------------------------------------------------------------*/
28 struct writeset_metadata {
29 	uint32_t nr_bits;
30 	dm_block_t root;
31 };
32 
33 struct writeset {
34 	struct writeset_metadata md;
35 
36 	/*
37 	 * An in core copy of the bits to save constantly doing look ups on
38 	 * disk.
39 	 */
40 	unsigned long *bits;
41 };
42 
43 /*
44  * This does not free off the on disk bitset as this will normally be done
45  * after digesting into the era array.
46  */
47 static void writeset_free(struct writeset *ws)
48 {
49 	vfree(ws->bits);
50 	ws->bits = NULL;
51 }
52 
53 static int setup_on_disk_bitset(struct dm_disk_bitset *info,
54 				unsigned nr_bits, dm_block_t *root)
55 {
56 	int r;
57 
58 	r = dm_bitset_empty(info, root);
59 	if (r)
60 		return r;
61 
62 	return dm_bitset_resize(info, *root, 0, nr_bits, false, root);
63 }
64 
65 static size_t bitset_size(unsigned nr_bits)
66 {
67 	return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG);
68 }
69 
70 /*
71  * Allocates memory for the in core bitset.
72  */
73 static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks)
74 {
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 			 dm_block_t nr_blocks)
89 {
90 	int r;
91 
92 	memset(ws->bits, 0, bitset_size(nr_blocks));
93 
94 	ws->md.nr_bits = nr_blocks;
95 	r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root);
96 	if (r) {
97 		DMERR("%s: setup_on_disk_bitset failed", __func__);
98 		return r;
99 	}
100 
101 	return 0;
102 }
103 
104 static bool writeset_marked(struct writeset *ws, dm_block_t block)
105 {
106 	return test_bit(block, ws->bits);
107 }
108 
109 static int writeset_marked_on_disk(struct dm_disk_bitset *info,
110 				   struct writeset_metadata *m, dm_block_t block,
111 				   bool *result)
112 {
113 	dm_block_t old = m->root;
114 
115 	/*
116 	 * The bitset was flushed when it was archived, so we know there'll
117 	 * be no change to the root.
118 	 */
119 	int r = dm_bitset_test_bit(info, m->root, block, &m->root, result);
120 	if (r) {
121 		DMERR("%s: dm_bitset_test_bit failed", __func__);
122 		return r;
123 	}
124 
125 	BUG_ON(m->root != old);
126 
127 	return r;
128 }
129 
130 /*
131  * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was.
132  */
133 static int writeset_test_and_set(struct dm_disk_bitset *info,
134 				 struct writeset *ws, uint32_t block)
135 {
136 	int r;
137 
138 	if (!test_bit(block, ws->bits)) {
139 		r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root);
140 		if (r) {
141 			/* FIXME: fail mode */
142 			return r;
143 		}
144 
145 		return 0;
146 	}
147 
148 	return 1;
149 }
150 
151 /*----------------------------------------------------------------
152  * On disk metadata layout
153  *--------------------------------------------------------------*/
154 #define SPACE_MAP_ROOT_SIZE 128
155 #define UUID_LEN 16
156 
157 struct writeset_disk {
158 	__le32 nr_bits;
159 	__le64 root;
160 } __packed;
161 
162 struct superblock_disk {
163 	__le32 csum;
164 	__le32 flags;
165 	__le64 blocknr;
166 
167 	__u8 uuid[UUID_LEN];
168 	__le64 magic;
169 	__le32 version;
170 
171 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
172 
173 	__le32 data_block_size;
174 	__le32 metadata_block_size;
175 	__le32 nr_blocks;
176 
177 	__le32 current_era;
178 	struct writeset_disk current_writeset;
179 
180 	/*
181 	 * Only these two fields are valid within the metadata snapshot.
182 	 */
183 	__le64 writeset_tree_root;
184 	__le64 era_array_root;
185 
186 	__le64 metadata_snap;
187 } __packed;
188 
189 /*----------------------------------------------------------------
190  * Superblock validation
191  *--------------------------------------------------------------*/
192 static void sb_prepare_for_write(struct dm_block_validator *v,
193 				 struct dm_block *b,
194 				 size_t sb_block_size)
195 {
196 	struct superblock_disk *disk = dm_block_data(b);
197 
198 	disk->blocknr = cpu_to_le64(dm_block_location(b));
199 	disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags,
200 						sb_block_size - sizeof(__le32),
201 						SUPERBLOCK_CSUM_XOR));
202 }
203 
204 static int check_metadata_version(struct superblock_disk *disk)
205 {
206 	uint32_t metadata_version = le32_to_cpu(disk->version);
207 	if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) {
208 		DMERR("Era metadata version %u found, but only versions between %u and %u supported.",
209 		      metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION);
210 		return -EINVAL;
211 	}
212 
213 	return 0;
214 }
215 
216 static int sb_check(struct dm_block_validator *v,
217 		    struct dm_block *b,
218 		    size_t sb_block_size)
219 {
220 	struct superblock_disk *disk = dm_block_data(b);
221 	__le32 csum_le;
222 
223 	if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) {
224 		DMERR("sb_check failed: blocknr %llu: wanted %llu",
225 		      le64_to_cpu(disk->blocknr),
226 		      (unsigned long long)dm_block_location(b));
227 		return -ENOTBLK;
228 	}
229 
230 	if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) {
231 		DMERR("sb_check failed: magic %llu: wanted %llu",
232 		      le64_to_cpu(disk->magic),
233 		      (unsigned long long) SUPERBLOCK_MAGIC);
234 		return -EILSEQ;
235 	}
236 
237 	csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags,
238 					     sb_block_size - sizeof(__le32),
239 					     SUPERBLOCK_CSUM_XOR));
240 	if (csum_le != disk->csum) {
241 		DMERR("sb_check failed: csum %u: wanted %u",
242 		      le32_to_cpu(csum_le), le32_to_cpu(disk->csum));
243 		return -EILSEQ;
244 	}
245 
246 	return check_metadata_version(disk);
247 }
248 
249 static struct dm_block_validator sb_validator = {
250 	.name = "superblock",
251 	.prepare_for_write = sb_prepare_for_write,
252 	.check = sb_check
253 };
254 
255 /*----------------------------------------------------------------
256  * Low level metadata handling
257  *--------------------------------------------------------------*/
258 #define DM_ERA_METADATA_BLOCK_SIZE 4096
259 #define ERA_MAX_CONCURRENT_LOCKS 5
260 
261 struct era_metadata {
262 	struct block_device *bdev;
263 	struct dm_block_manager *bm;
264 	struct dm_space_map *sm;
265 	struct dm_transaction_manager *tm;
266 
267 	dm_block_t block_size;
268 	uint32_t nr_blocks;
269 
270 	uint32_t current_era;
271 
272 	/*
273 	 * We preallocate 2 writesets.  When an era rolls over we
274 	 * switch between them. This means the allocation is done at
275 	 * preresume time, rather than on the io path.
276 	 */
277 	struct writeset writesets[2];
278 	struct writeset *current_writeset;
279 
280 	dm_block_t writeset_tree_root;
281 	dm_block_t era_array_root;
282 
283 	struct dm_disk_bitset bitset_info;
284 	struct dm_btree_info writeset_tree_info;
285 	struct dm_array_info era_array_info;
286 
287 	dm_block_t metadata_snap;
288 
289 	/*
290 	 * A flag that is set whenever a writeset has been archived.
291 	 */
292 	bool archived_writesets;
293 
294 	/*
295 	 * Reading the space map root can fail, so we read it into this
296 	 * buffer before the superblock is locked and updated.
297 	 */
298 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
299 };
300 
301 static int superblock_read_lock(struct era_metadata *md,
302 				struct dm_block **sblock)
303 {
304 	return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION,
305 			       &sb_validator, sblock);
306 }
307 
308 static int superblock_lock_zero(struct era_metadata *md,
309 				struct dm_block **sblock)
310 {
311 	return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION,
312 				     &sb_validator, sblock);
313 }
314 
315 static int superblock_lock(struct era_metadata *md,
316 			   struct dm_block **sblock)
317 {
318 	return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION,
319 				&sb_validator, sblock);
320 }
321 
322 /* FIXME: duplication with cache and thin */
323 static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
324 {
325 	int r;
326 	unsigned i;
327 	struct dm_block *b;
328 	__le64 *data_le, zero = cpu_to_le64(0);
329 	unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
330 
331 	/*
332 	 * We can't use a validator here - it may be all zeroes.
333 	 */
334 	r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b);
335 	if (r)
336 		return r;
337 
338 	data_le = dm_block_data(b);
339 	*result = true;
340 	for (i = 0; i < sb_block_size; i++) {
341 		if (data_le[i] != zero) {
342 			*result = false;
343 			break;
344 		}
345 	}
346 
347 	dm_bm_unlock(b);
348 
349 	return 0;
350 }
351 
352 /*----------------------------------------------------------------*/
353 
354 static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk)
355 {
356 	disk->nr_bits = cpu_to_le32(core->nr_bits);
357 	disk->root = cpu_to_le64(core->root);
358 }
359 
360 static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core)
361 {
362 	core->nr_bits = le32_to_cpu(disk->nr_bits);
363 	core->root = le64_to_cpu(disk->root);
364 }
365 
366 static void ws_inc(void *context, const void *value)
367 {
368 	struct era_metadata *md = context;
369 	struct writeset_disk ws_d;
370 	dm_block_t b;
371 
372 	memcpy(&ws_d, value, sizeof(ws_d));
373 	b = le64_to_cpu(ws_d.root);
374 
375 	dm_tm_inc(md->tm, b);
376 }
377 
378 static void ws_dec(void *context, const void *value)
379 {
380 	struct era_metadata *md = context;
381 	struct writeset_disk ws_d;
382 	dm_block_t b;
383 
384 	memcpy(&ws_d, value, sizeof(ws_d));
385 	b = le64_to_cpu(ws_d.root);
386 
387 	dm_bitset_del(&md->bitset_info, b);
388 }
389 
390 static int ws_eq(void *context, const void *value1, const void *value2)
391 {
392 	return !memcmp(value1, value2, sizeof(struct writeset_disk));
393 }
394 
395 /*----------------------------------------------------------------*/
396 
397 static void setup_writeset_tree_info(struct era_metadata *md)
398 {
399 	struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type;
400 	md->writeset_tree_info.tm = md->tm;
401 	md->writeset_tree_info.levels = 1;
402 	vt->context = md;
403 	vt->size = sizeof(struct writeset_disk);
404 	vt->inc = ws_inc;
405 	vt->dec = ws_dec;
406 	vt->equal = ws_eq;
407 }
408 
409 static void setup_era_array_info(struct era_metadata *md)
410 
411 {
412 	struct dm_btree_value_type vt;
413 	vt.context = NULL;
414 	vt.size = sizeof(__le32);
415 	vt.inc = NULL;
416 	vt.dec = NULL;
417 	vt.equal = NULL;
418 
419 	dm_array_info_init(&md->era_array_info, md->tm, &vt);
420 }
421 
422 static void setup_infos(struct era_metadata *md)
423 {
424 	dm_disk_bitset_init(md->tm, &md->bitset_info);
425 	setup_writeset_tree_info(md);
426 	setup_era_array_info(md);
427 }
428 
429 /*----------------------------------------------------------------*/
430 
431 static int create_fresh_metadata(struct era_metadata *md)
432 {
433 	int r;
434 
435 	r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION,
436 				 &md->tm, &md->sm);
437 	if (r < 0) {
438 		DMERR("dm_tm_create_with_sm failed");
439 		return r;
440 	}
441 
442 	setup_infos(md);
443 
444 	r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root);
445 	if (r) {
446 		DMERR("couldn't create new writeset tree");
447 		goto bad;
448 	}
449 
450 	r = dm_array_empty(&md->era_array_info, &md->era_array_root);
451 	if (r) {
452 		DMERR("couldn't create era array");
453 		goto bad;
454 	}
455 
456 	return 0;
457 
458 bad:
459 	dm_sm_destroy(md->sm);
460 	dm_tm_destroy(md->tm);
461 
462 	return r;
463 }
464 
465 static int save_sm_root(struct era_metadata *md)
466 {
467 	int r;
468 	size_t metadata_len;
469 
470 	r = dm_sm_root_size(md->sm, &metadata_len);
471 	if (r < 0)
472 		return r;
473 
474 	return dm_sm_copy_root(md->sm, &md->metadata_space_map_root,
475 			       metadata_len);
476 }
477 
478 static void copy_sm_root(struct era_metadata *md, struct superblock_disk *disk)
479 {
480 	memcpy(&disk->metadata_space_map_root,
481 	       &md->metadata_space_map_root,
482 	       sizeof(md->metadata_space_map_root));
483 }
484 
485 /*
486  * Writes a superblock, including the static fields that don't get updated
487  * with every commit (possible optimisation here).  'md' should be fully
488  * constructed when this is called.
489  */
490 static void prepare_superblock(struct era_metadata *md, struct superblock_disk *disk)
491 {
492 	disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
493 	disk->flags = cpu_to_le32(0ul);
494 
495 	/* FIXME: can't keep blanking the uuid (uuid is currently unused though) */
496 	memset(disk->uuid, 0, sizeof(disk->uuid));
497 	disk->version = cpu_to_le32(MAX_ERA_VERSION);
498 
499 	copy_sm_root(md, disk);
500 
501 	disk->data_block_size = cpu_to_le32(md->block_size);
502 	disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
503 	disk->nr_blocks = cpu_to_le32(md->nr_blocks);
504 	disk->current_era = cpu_to_le32(md->current_era);
505 
506 	ws_pack(&md->current_writeset->md, &disk->current_writeset);
507 	disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root);
508 	disk->era_array_root = cpu_to_le64(md->era_array_root);
509 	disk->metadata_snap = cpu_to_le64(md->metadata_snap);
510 }
511 
512 static int write_superblock(struct era_metadata *md)
513 {
514 	int r;
515 	struct dm_block *sblock;
516 	struct superblock_disk *disk;
517 
518 	r = save_sm_root(md);
519 	if (r) {
520 		DMERR("%s: save_sm_root failed", __func__);
521 		return r;
522 	}
523 
524 	r = superblock_lock_zero(md, &sblock);
525 	if (r)
526 		return r;
527 
528 	disk = dm_block_data(sblock);
529 	prepare_superblock(md, disk);
530 
531 	return dm_tm_commit(md->tm, sblock);
532 }
533 
534 /*
535  * Assumes block_size and the infos are set.
536  */
537 static int format_metadata(struct era_metadata *md)
538 {
539 	int r;
540 
541 	r = create_fresh_metadata(md);
542 	if (r)
543 		return r;
544 
545 	r = write_superblock(md);
546 	if (r) {
547 		dm_sm_destroy(md->sm);
548 		dm_tm_destroy(md->tm);
549 		return r;
550 	}
551 
552 	return 0;
553 }
554 
555 static int open_metadata(struct era_metadata *md)
556 {
557 	int r;
558 	struct dm_block *sblock;
559 	struct superblock_disk *disk;
560 
561 	r = superblock_read_lock(md, &sblock);
562 	if (r) {
563 		DMERR("couldn't read_lock superblock");
564 		return r;
565 	}
566 
567 	disk = dm_block_data(sblock);
568 
569 	/* Verify the data block size hasn't changed */
570 	if (le32_to_cpu(disk->data_block_size) != md->block_size) {
571 		DMERR("changing the data block size (from %u to %llu) is not supported",
572 		      le32_to_cpu(disk->data_block_size), md->block_size);
573 		r = -EINVAL;
574 		goto bad;
575 	}
576 
577 	r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION,
578 			       disk->metadata_space_map_root,
579 			       sizeof(disk->metadata_space_map_root),
580 			       &md->tm, &md->sm);
581 	if (r) {
582 		DMERR("dm_tm_open_with_sm failed");
583 		goto bad;
584 	}
585 
586 	setup_infos(md);
587 
588 	md->nr_blocks = le32_to_cpu(disk->nr_blocks);
589 	md->current_era = le32_to_cpu(disk->current_era);
590 
591 	ws_unpack(&disk->current_writeset, &md->current_writeset->md);
592 	md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root);
593 	md->era_array_root = le64_to_cpu(disk->era_array_root);
594 	md->metadata_snap = le64_to_cpu(disk->metadata_snap);
595 	md->archived_writesets = true;
596 
597 	dm_bm_unlock(sblock);
598 
599 	return 0;
600 
601 bad:
602 	dm_bm_unlock(sblock);
603 	return r;
604 }
605 
606 static int open_or_format_metadata(struct era_metadata *md,
607 				   bool may_format)
608 {
609 	int r;
610 	bool unformatted = false;
611 
612 	r = superblock_all_zeroes(md->bm, &unformatted);
613 	if (r)
614 		return r;
615 
616 	if (unformatted)
617 		return may_format ? format_metadata(md) : -EPERM;
618 
619 	return open_metadata(md);
620 }
621 
622 static int create_persistent_data_objects(struct era_metadata *md,
623 					  bool may_format)
624 {
625 	int r;
626 
627 	md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE,
628 					 ERA_MAX_CONCURRENT_LOCKS);
629 	if (IS_ERR(md->bm)) {
630 		DMERR("could not create block manager");
631 		return PTR_ERR(md->bm);
632 	}
633 
634 	r = open_or_format_metadata(md, may_format);
635 	if (r)
636 		dm_block_manager_destroy(md->bm);
637 
638 	return r;
639 }
640 
641 static void destroy_persistent_data_objects(struct era_metadata *md)
642 {
643 	dm_sm_destroy(md->sm);
644 	dm_tm_destroy(md->tm);
645 	dm_block_manager_destroy(md->bm);
646 }
647 
648 /*
649  * This waits until all era_map threads have picked up the new filter.
650  */
651 static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset)
652 {
653 	rcu_assign_pointer(md->current_writeset, new_writeset);
654 	synchronize_rcu();
655 }
656 
657 /*----------------------------------------------------------------
658  * Writesets get 'digested' into the main era array.
659  *
660  * We're using a coroutine here so the worker thread can do the digestion,
661  * thus avoiding synchronisation of the metadata.  Digesting a whole
662  * writeset in one go would cause too much latency.
663  *--------------------------------------------------------------*/
664 struct digest {
665 	uint32_t era;
666 	unsigned nr_bits, current_bit;
667 	struct writeset_metadata writeset;
668 	__le32 value;
669 	struct dm_disk_bitset info;
670 
671 	int (*step)(struct era_metadata *, struct digest *);
672 };
673 
674 static int metadata_digest_lookup_writeset(struct era_metadata *md,
675 					   struct digest *d);
676 
677 static int metadata_digest_remove_writeset(struct era_metadata *md,
678 					   struct digest *d)
679 {
680 	int r;
681 	uint64_t key = d->era;
682 
683 	r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root,
684 			    &key, &md->writeset_tree_root);
685 	if (r) {
686 		DMERR("%s: dm_btree_remove failed", __func__);
687 		return r;
688 	}
689 
690 	d->step = metadata_digest_lookup_writeset;
691 	return 0;
692 }
693 
694 #define INSERTS_PER_STEP 100
695 
696 static int metadata_digest_transcribe_writeset(struct era_metadata *md,
697 					       struct digest *d)
698 {
699 	int r;
700 	bool marked;
701 	unsigned b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits);
702 
703 	for (b = d->current_bit; b < e; b++) {
704 		r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked);
705 		if (r) {
706 			DMERR("%s: writeset_marked_on_disk failed", __func__);
707 			return r;
708 		}
709 
710 		if (!marked)
711 			continue;
712 
713 		__dm_bless_for_disk(&d->value);
714 		r = dm_array_set_value(&md->era_array_info, md->era_array_root,
715 				       b, &d->value, &md->era_array_root);
716 		if (r) {
717 			DMERR("%s: dm_array_set_value failed", __func__);
718 			return r;
719 		}
720 	}
721 
722 	if (b == d->nr_bits)
723 		d->step = metadata_digest_remove_writeset;
724 	else
725 		d->current_bit = b;
726 
727 	return 0;
728 }
729 
730 static int metadata_digest_lookup_writeset(struct era_metadata *md,
731 					   struct digest *d)
732 {
733 	int r;
734 	uint64_t key;
735 	struct writeset_disk disk;
736 
737 	r = dm_btree_find_lowest_key(&md->writeset_tree_info,
738 				     md->writeset_tree_root, &key);
739 	if (r < 0)
740 		return r;
741 
742 	d->era = key;
743 
744 	r = dm_btree_lookup(&md->writeset_tree_info,
745 			    md->writeset_tree_root, &key, &disk);
746 	if (r) {
747 		if (r == -ENODATA) {
748 			d->step = NULL;
749 			return 0;
750 		}
751 
752 		DMERR("%s: dm_btree_lookup failed", __func__);
753 		return r;
754 	}
755 
756 	ws_unpack(&disk, &d->writeset);
757 	d->value = cpu_to_le32(key);
758 
759 	/*
760 	 * We initialise another bitset info to avoid any caching side effects
761 	 * with the previous one.
762 	 */
763 	dm_disk_bitset_init(md->tm, &d->info);
764 
765 	d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks);
766 	d->current_bit = 0;
767 	d->step = metadata_digest_transcribe_writeset;
768 
769 	return 0;
770 }
771 
772 static int metadata_digest_start(struct era_metadata *md, struct digest *d)
773 {
774 	if (d->step)
775 		return 0;
776 
777 	memset(d, 0, sizeof(*d));
778 	d->step = metadata_digest_lookup_writeset;
779 
780 	return 0;
781 }
782 
783 /*----------------------------------------------------------------
784  * High level metadata interface.  Target methods should use these, and not
785  * the lower level ones.
786  *--------------------------------------------------------------*/
787 static struct era_metadata *metadata_open(struct block_device *bdev,
788 					  sector_t block_size,
789 					  bool may_format)
790 {
791 	int r;
792 	struct era_metadata *md = kzalloc(sizeof(*md), GFP_KERNEL);
793 
794 	if (!md)
795 		return NULL;
796 
797 	md->bdev = bdev;
798 	md->block_size = block_size;
799 
800 	md->writesets[0].md.root = INVALID_WRITESET_ROOT;
801 	md->writesets[1].md.root = INVALID_WRITESET_ROOT;
802 	md->current_writeset = &md->writesets[0];
803 
804 	r = create_persistent_data_objects(md, may_format);
805 	if (r) {
806 		kfree(md);
807 		return ERR_PTR(r);
808 	}
809 
810 	return md;
811 }
812 
813 static void metadata_close(struct era_metadata *md)
814 {
815 	writeset_free(&md->writesets[0]);
816 	writeset_free(&md->writesets[1]);
817 	destroy_persistent_data_objects(md);
818 	kfree(md);
819 }
820 
821 static bool valid_nr_blocks(dm_block_t n)
822 {
823 	/*
824 	 * dm_bitset restricts us to 2^32.  test_bit & co. restrict us
825 	 * further to 2^31 - 1
826 	 */
827 	return n < (1ull << 31);
828 }
829 
830 static int metadata_resize(struct era_metadata *md, void *arg)
831 {
832 	int r;
833 	dm_block_t *new_size = arg;
834 	__le32 value;
835 
836 	if (!valid_nr_blocks(*new_size)) {
837 		DMERR("Invalid number of origin blocks %llu",
838 		      (unsigned long long) *new_size);
839 		return -EINVAL;
840 	}
841 
842 	writeset_free(&md->writesets[0]);
843 	writeset_free(&md->writesets[1]);
844 
845 	r = writeset_alloc(&md->writesets[0], *new_size);
846 	if (r) {
847 		DMERR("%s: writeset_alloc failed for writeset 0", __func__);
848 		return r;
849 	}
850 
851 	r = writeset_alloc(&md->writesets[1], *new_size);
852 	if (r) {
853 		DMERR("%s: writeset_alloc failed for writeset 1", __func__);
854 		writeset_free(&md->writesets[0]);
855 		return r;
856 	}
857 
858 	value = cpu_to_le32(0u);
859 	__dm_bless_for_disk(&value);
860 	r = dm_array_resize(&md->era_array_info, md->era_array_root,
861 			    md->nr_blocks, *new_size,
862 			    &value, &md->era_array_root);
863 	if (r) {
864 		DMERR("%s: dm_array_resize failed", __func__);
865 		writeset_free(&md->writesets[0]);
866 		writeset_free(&md->writesets[1]);
867 		return r;
868 	}
869 
870 	md->nr_blocks = *new_size;
871 	return 0;
872 }
873 
874 static int metadata_era_archive(struct era_metadata *md)
875 {
876 	int r;
877 	uint64_t keys[1];
878 	struct writeset_disk value;
879 
880 	r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
881 			    &md->current_writeset->md.root);
882 	if (r) {
883 		DMERR("%s: dm_bitset_flush failed", __func__);
884 		return r;
885 	}
886 
887 	ws_pack(&md->current_writeset->md, &value);
888 
889 	keys[0] = md->current_era;
890 	__dm_bless_for_disk(&value);
891 	r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root,
892 			    keys, &value, &md->writeset_tree_root);
893 	if (r) {
894 		DMERR("%s: couldn't insert writeset into btree", __func__);
895 		/* FIXME: fail mode */
896 		return r;
897 	}
898 
899 	md->current_writeset->md.root = INVALID_WRITESET_ROOT;
900 	md->archived_writesets = true;
901 
902 	return 0;
903 }
904 
905 static struct writeset *next_writeset(struct era_metadata *md)
906 {
907 	return (md->current_writeset == &md->writesets[0]) ?
908 		&md->writesets[1] : &md->writesets[0];
909 }
910 
911 static int metadata_new_era(struct era_metadata *md)
912 {
913 	int r;
914 	struct writeset *new_writeset = next_writeset(md);
915 
916 	r = writeset_init(&md->bitset_info, new_writeset, md->nr_blocks);
917 	if (r) {
918 		DMERR("%s: writeset_init failed", __func__);
919 		return r;
920 	}
921 
922 	swap_writeset(md, new_writeset);
923 	md->current_era++;
924 
925 	return 0;
926 }
927 
928 static int metadata_era_rollover(struct era_metadata *md)
929 {
930 	int r;
931 
932 	if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
933 		r = metadata_era_archive(md);
934 		if (r) {
935 			DMERR("%s: metadata_archive_era failed", __func__);
936 			/* FIXME: fail mode? */
937 			return r;
938 		}
939 	}
940 
941 	r = metadata_new_era(md);
942 	if (r) {
943 		DMERR("%s: new era failed", __func__);
944 		/* FIXME: fail mode */
945 		return r;
946 	}
947 
948 	return 0;
949 }
950 
951 static bool metadata_current_marked(struct era_metadata *md, dm_block_t block)
952 {
953 	bool r;
954 	struct writeset *ws;
955 
956 	rcu_read_lock();
957 	ws = rcu_dereference(md->current_writeset);
958 	r = writeset_marked(ws, block);
959 	rcu_read_unlock();
960 
961 	return r;
962 }
963 
964 static int metadata_commit(struct era_metadata *md)
965 {
966 	int r;
967 	struct dm_block *sblock;
968 
969 	if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
970 		r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
971 				    &md->current_writeset->md.root);
972 		if (r) {
973 			DMERR("%s: bitset flush failed", __func__);
974 			return r;
975 		}
976 	}
977 
978 	r = dm_tm_pre_commit(md->tm);
979 	if (r) {
980 		DMERR("%s: pre commit failed", __func__);
981 		return r;
982 	}
983 
984 	r = save_sm_root(md);
985 	if (r) {
986 		DMERR("%s: save_sm_root failed", __func__);
987 		return r;
988 	}
989 
990 	r = superblock_lock(md, &sblock);
991 	if (r) {
992 		DMERR("%s: superblock lock failed", __func__);
993 		return r;
994 	}
995 
996 	prepare_superblock(md, dm_block_data(sblock));
997 
998 	return dm_tm_commit(md->tm, sblock);
999 }
1000 
1001 static int metadata_checkpoint(struct era_metadata *md)
1002 {
1003 	/*
1004 	 * For now we just rollover, but later I want to put a check in to
1005 	 * avoid this if the filter is still pretty fresh.
1006 	 */
1007 	return metadata_era_rollover(md);
1008 }
1009 
1010 /*
1011  * Metadata snapshots allow userland to access era data.
1012  */
1013 static int metadata_take_snap(struct era_metadata *md)
1014 {
1015 	int r, inc;
1016 	struct dm_block *clone;
1017 
1018 	if (md->metadata_snap != SUPERBLOCK_LOCATION) {
1019 		DMERR("%s: metadata snapshot already exists", __func__);
1020 		return -EINVAL;
1021 	}
1022 
1023 	r = metadata_era_rollover(md);
1024 	if (r) {
1025 		DMERR("%s: era rollover failed", __func__);
1026 		return r;
1027 	}
1028 
1029 	r = metadata_commit(md);
1030 	if (r) {
1031 		DMERR("%s: pre commit failed", __func__);
1032 		return r;
1033 	}
1034 
1035 	r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION);
1036 	if (r) {
1037 		DMERR("%s: couldn't increment superblock", __func__);
1038 		return r;
1039 	}
1040 
1041 	r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION,
1042 			       &sb_validator, &clone, &inc);
1043 	if (r) {
1044 		DMERR("%s: couldn't shadow superblock", __func__);
1045 		dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION);
1046 		return r;
1047 	}
1048 	BUG_ON(!inc);
1049 
1050 	r = dm_sm_inc_block(md->sm, md->writeset_tree_root);
1051 	if (r) {
1052 		DMERR("%s: couldn't inc writeset tree root", __func__);
1053 		dm_tm_unlock(md->tm, clone);
1054 		return r;
1055 	}
1056 
1057 	r = dm_sm_inc_block(md->sm, md->era_array_root);
1058 	if (r) {
1059 		DMERR("%s: couldn't inc era tree root", __func__);
1060 		dm_sm_dec_block(md->sm, md->writeset_tree_root);
1061 		dm_tm_unlock(md->tm, clone);
1062 		return r;
1063 	}
1064 
1065 	md->metadata_snap = dm_block_location(clone);
1066 
1067 	dm_tm_unlock(md->tm, clone);
1068 
1069 	return 0;
1070 }
1071 
1072 static int metadata_drop_snap(struct era_metadata *md)
1073 {
1074 	int r;
1075 	dm_block_t location;
1076 	struct dm_block *clone;
1077 	struct superblock_disk *disk;
1078 
1079 	if (md->metadata_snap == SUPERBLOCK_LOCATION) {
1080 		DMERR("%s: no snap to drop", __func__);
1081 		return -EINVAL;
1082 	}
1083 
1084 	r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone);
1085 	if (r) {
1086 		DMERR("%s: couldn't read lock superblock clone", __func__);
1087 		return r;
1088 	}
1089 
1090 	/*
1091 	 * Whatever happens now we'll commit with no record of the metadata
1092 	 * snap.
1093 	 */
1094 	md->metadata_snap = SUPERBLOCK_LOCATION;
1095 
1096 	disk = dm_block_data(clone);
1097 	r = dm_btree_del(&md->writeset_tree_info,
1098 			 le64_to_cpu(disk->writeset_tree_root));
1099 	if (r) {
1100 		DMERR("%s: error deleting writeset tree clone", __func__);
1101 		dm_tm_unlock(md->tm, clone);
1102 		return r;
1103 	}
1104 
1105 	r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root));
1106 	if (r) {
1107 		DMERR("%s: error deleting era array clone", __func__);
1108 		dm_tm_unlock(md->tm, clone);
1109 		return r;
1110 	}
1111 
1112 	location = dm_block_location(clone);
1113 	dm_tm_unlock(md->tm, clone);
1114 
1115 	return dm_sm_dec_block(md->sm, location);
1116 }
1117 
1118 struct metadata_stats {
1119 	dm_block_t used;
1120 	dm_block_t total;
1121 	dm_block_t snap;
1122 	uint32_t era;
1123 };
1124 
1125 static int metadata_get_stats(struct era_metadata *md, void *ptr)
1126 {
1127 	int r;
1128 	struct metadata_stats *s = ptr;
1129 	dm_block_t nr_free, nr_total;
1130 
1131 	r = dm_sm_get_nr_free(md->sm, &nr_free);
1132 	if (r) {
1133 		DMERR("dm_sm_get_nr_free returned %d", r);
1134 		return r;
1135 	}
1136 
1137 	r = dm_sm_get_nr_blocks(md->sm, &nr_total);
1138 	if (r) {
1139 		DMERR("dm_pool_get_metadata_dev_size returned %d", r);
1140 		return r;
1141 	}
1142 
1143 	s->used = nr_total - nr_free;
1144 	s->total = nr_total;
1145 	s->snap = md->metadata_snap;
1146 	s->era = md->current_era;
1147 
1148 	return 0;
1149 }
1150 
1151 /*----------------------------------------------------------------*/
1152 
1153 struct era {
1154 	struct dm_target *ti;
1155 
1156 	struct dm_dev *metadata_dev;
1157 	struct dm_dev *origin_dev;
1158 
1159 	dm_block_t nr_blocks;
1160 	uint32_t sectors_per_block;
1161 	int sectors_per_block_shift;
1162 	struct era_metadata *md;
1163 
1164 	struct workqueue_struct *wq;
1165 	struct work_struct worker;
1166 
1167 	spinlock_t deferred_lock;
1168 	struct bio_list deferred_bios;
1169 
1170 	spinlock_t rpc_lock;
1171 	struct list_head rpc_calls;
1172 
1173 	struct digest digest;
1174 	atomic_t suspended;
1175 };
1176 
1177 struct rpc {
1178 	struct list_head list;
1179 
1180 	int (*fn0)(struct era_metadata *);
1181 	int (*fn1)(struct era_metadata *, void *);
1182 	void *arg;
1183 	int result;
1184 
1185 	struct completion complete;
1186 };
1187 
1188 /*----------------------------------------------------------------
1189  * Remapping.
1190  *---------------------------------------------------------------*/
1191 static bool block_size_is_power_of_two(struct era *era)
1192 {
1193 	return era->sectors_per_block_shift >= 0;
1194 }
1195 
1196 static dm_block_t get_block(struct era *era, struct bio *bio)
1197 {
1198 	sector_t block_nr = bio->bi_iter.bi_sector;
1199 
1200 	if (!block_size_is_power_of_two(era))
1201 		(void) sector_div(block_nr, era->sectors_per_block);
1202 	else
1203 		block_nr >>= era->sectors_per_block_shift;
1204 
1205 	return block_nr;
1206 }
1207 
1208 static void remap_to_origin(struct era *era, struct bio *bio)
1209 {
1210 	bio_set_dev(bio, era->origin_dev->bdev);
1211 }
1212 
1213 /*----------------------------------------------------------------
1214  * Worker thread
1215  *--------------------------------------------------------------*/
1216 static void wake_worker(struct era *era)
1217 {
1218 	if (!atomic_read(&era->suspended))
1219 		queue_work(era->wq, &era->worker);
1220 }
1221 
1222 static void process_old_eras(struct era *era)
1223 {
1224 	int r;
1225 
1226 	if (!era->digest.step)
1227 		return;
1228 
1229 	r = era->digest.step(era->md, &era->digest);
1230 	if (r < 0) {
1231 		DMERR("%s: digest step failed, stopping digestion", __func__);
1232 		era->digest.step = NULL;
1233 
1234 	} else if (era->digest.step)
1235 		wake_worker(era);
1236 }
1237 
1238 static void process_deferred_bios(struct era *era)
1239 {
1240 	int r;
1241 	struct bio_list deferred_bios, marked_bios;
1242 	struct bio *bio;
1243 	struct blk_plug plug;
1244 	bool commit_needed = false;
1245 	bool failed = false;
1246 	struct writeset *ws = era->md->current_writeset;
1247 
1248 	bio_list_init(&deferred_bios);
1249 	bio_list_init(&marked_bios);
1250 
1251 	spin_lock(&era->deferred_lock);
1252 	bio_list_merge(&deferred_bios, &era->deferred_bios);
1253 	bio_list_init(&era->deferred_bios);
1254 	spin_unlock(&era->deferred_lock);
1255 
1256 	if (bio_list_empty(&deferred_bios))
1257 		return;
1258 
1259 	while ((bio = bio_list_pop(&deferred_bios))) {
1260 		r = writeset_test_and_set(&era->md->bitset_info, ws,
1261 					  get_block(era, bio));
1262 		if (r < 0) {
1263 			/*
1264 			 * This is bad news, we need to rollback.
1265 			 * FIXME: finish.
1266 			 */
1267 			failed = true;
1268 		} else if (r == 0)
1269 			commit_needed = true;
1270 
1271 		bio_list_add(&marked_bios, bio);
1272 	}
1273 
1274 	if (commit_needed) {
1275 		r = metadata_commit(era->md);
1276 		if (r)
1277 			failed = true;
1278 	}
1279 
1280 	if (failed)
1281 		while ((bio = bio_list_pop(&marked_bios)))
1282 			bio_io_error(bio);
1283 	else {
1284 		blk_start_plug(&plug);
1285 		while ((bio = bio_list_pop(&marked_bios))) {
1286 			/*
1287 			 * Only update the in-core writeset if the on-disk one
1288 			 * was updated too.
1289 			 */
1290 			if (commit_needed)
1291 				set_bit(get_block(era, bio), ws->bits);
1292 			submit_bio_noacct(bio);
1293 		}
1294 		blk_finish_plug(&plug);
1295 	}
1296 }
1297 
1298 static void process_rpc_calls(struct era *era)
1299 {
1300 	int r;
1301 	bool need_commit = false;
1302 	struct list_head calls;
1303 	struct rpc *rpc, *tmp;
1304 
1305 	INIT_LIST_HEAD(&calls);
1306 	spin_lock(&era->rpc_lock);
1307 	list_splice_init(&era->rpc_calls, &calls);
1308 	spin_unlock(&era->rpc_lock);
1309 
1310 	list_for_each_entry_safe(rpc, tmp, &calls, list) {
1311 		rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg);
1312 		need_commit = true;
1313 	}
1314 
1315 	if (need_commit) {
1316 		r = metadata_commit(era->md);
1317 		if (r)
1318 			list_for_each_entry_safe(rpc, tmp, &calls, list)
1319 				rpc->result = r;
1320 	}
1321 
1322 	list_for_each_entry_safe(rpc, tmp, &calls, list)
1323 		complete(&rpc->complete);
1324 }
1325 
1326 static void kick_off_digest(struct era *era)
1327 {
1328 	if (era->md->archived_writesets) {
1329 		era->md->archived_writesets = false;
1330 		metadata_digest_start(era->md, &era->digest);
1331 	}
1332 }
1333 
1334 static void do_work(struct work_struct *ws)
1335 {
1336 	struct era *era = container_of(ws, struct era, worker);
1337 
1338 	kick_off_digest(era);
1339 	process_old_eras(era);
1340 	process_deferred_bios(era);
1341 	process_rpc_calls(era);
1342 }
1343 
1344 static void defer_bio(struct era *era, struct bio *bio)
1345 {
1346 	spin_lock(&era->deferred_lock);
1347 	bio_list_add(&era->deferred_bios, bio);
1348 	spin_unlock(&era->deferred_lock);
1349 
1350 	wake_worker(era);
1351 }
1352 
1353 /*
1354  * Make an rpc call to the worker to change the metadata.
1355  */
1356 static int perform_rpc(struct era *era, struct rpc *rpc)
1357 {
1358 	rpc->result = 0;
1359 	init_completion(&rpc->complete);
1360 
1361 	spin_lock(&era->rpc_lock);
1362 	list_add(&rpc->list, &era->rpc_calls);
1363 	spin_unlock(&era->rpc_lock);
1364 
1365 	wake_worker(era);
1366 	wait_for_completion(&rpc->complete);
1367 
1368 	return rpc->result;
1369 }
1370 
1371 static int in_worker0(struct era *era, int (*fn)(struct era_metadata *))
1372 {
1373 	struct rpc rpc;
1374 	rpc.fn0 = fn;
1375 	rpc.fn1 = NULL;
1376 
1377 	return perform_rpc(era, &rpc);
1378 }
1379 
1380 static int in_worker1(struct era *era,
1381 		      int (*fn)(struct era_metadata *, void *), void *arg)
1382 {
1383 	struct rpc rpc;
1384 	rpc.fn0 = NULL;
1385 	rpc.fn1 = fn;
1386 	rpc.arg = arg;
1387 
1388 	return perform_rpc(era, &rpc);
1389 }
1390 
1391 static void start_worker(struct era *era)
1392 {
1393 	atomic_set(&era->suspended, 0);
1394 }
1395 
1396 static void stop_worker(struct era *era)
1397 {
1398 	atomic_set(&era->suspended, 1);
1399 	flush_workqueue(era->wq);
1400 }
1401 
1402 /*----------------------------------------------------------------
1403  * Target methods
1404  *--------------------------------------------------------------*/
1405 static void era_destroy(struct era *era)
1406 {
1407 	if (era->md)
1408 		metadata_close(era->md);
1409 
1410 	if (era->wq)
1411 		destroy_workqueue(era->wq);
1412 
1413 	if (era->origin_dev)
1414 		dm_put_device(era->ti, era->origin_dev);
1415 
1416 	if (era->metadata_dev)
1417 		dm_put_device(era->ti, era->metadata_dev);
1418 
1419 	kfree(era);
1420 }
1421 
1422 static dm_block_t calc_nr_blocks(struct era *era)
1423 {
1424 	return dm_sector_div_up(era->ti->len, era->sectors_per_block);
1425 }
1426 
1427 static bool valid_block_size(dm_block_t block_size)
1428 {
1429 	bool greater_than_zero = block_size > 0;
1430 	bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0;
1431 
1432 	return greater_than_zero && multiple_of_min_block_size;
1433 }
1434 
1435 /*
1436  * <metadata dev> <data dev> <data block size (sectors)>
1437  */
1438 static int era_ctr(struct dm_target *ti, unsigned argc, char **argv)
1439 {
1440 	int r;
1441 	char dummy;
1442 	struct era *era;
1443 	struct era_metadata *md;
1444 
1445 	if (argc != 3) {
1446 		ti->error = "Invalid argument count";
1447 		return -EINVAL;
1448 	}
1449 
1450 	era = kzalloc(sizeof(*era), GFP_KERNEL);
1451 	if (!era) {
1452 		ti->error = "Error allocating era structure";
1453 		return -ENOMEM;
1454 	}
1455 
1456 	era->ti = ti;
1457 
1458 	r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &era->metadata_dev);
1459 	if (r) {
1460 		ti->error = "Error opening metadata device";
1461 		era_destroy(era);
1462 		return -EINVAL;
1463 	}
1464 
1465 	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &era->origin_dev);
1466 	if (r) {
1467 		ti->error = "Error opening data device";
1468 		era_destroy(era);
1469 		return -EINVAL;
1470 	}
1471 
1472 	r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy);
1473 	if (r != 1) {
1474 		ti->error = "Error parsing block size";
1475 		era_destroy(era);
1476 		return -EINVAL;
1477 	}
1478 
1479 	r = dm_set_target_max_io_len(ti, era->sectors_per_block);
1480 	if (r) {
1481 		ti->error = "could not set max io len";
1482 		era_destroy(era);
1483 		return -EINVAL;
1484 	}
1485 
1486 	if (!valid_block_size(era->sectors_per_block)) {
1487 		ti->error = "Invalid block size";
1488 		era_destroy(era);
1489 		return -EINVAL;
1490 	}
1491 	if (era->sectors_per_block & (era->sectors_per_block - 1))
1492 		era->sectors_per_block_shift = -1;
1493 	else
1494 		era->sectors_per_block_shift = __ffs(era->sectors_per_block);
1495 
1496 	md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true);
1497 	if (IS_ERR(md)) {
1498 		ti->error = "Error reading metadata";
1499 		era_destroy(era);
1500 		return PTR_ERR(md);
1501 	}
1502 	era->md = md;
1503 
1504 	era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1505 	if (!era->wq) {
1506 		ti->error = "could not create workqueue for metadata object";
1507 		era_destroy(era);
1508 		return -ENOMEM;
1509 	}
1510 	INIT_WORK(&era->worker, do_work);
1511 
1512 	spin_lock_init(&era->deferred_lock);
1513 	bio_list_init(&era->deferred_bios);
1514 
1515 	spin_lock_init(&era->rpc_lock);
1516 	INIT_LIST_HEAD(&era->rpc_calls);
1517 
1518 	ti->private = era;
1519 	ti->num_flush_bios = 1;
1520 	ti->flush_supported = true;
1521 
1522 	ti->num_discard_bios = 1;
1523 
1524 	return 0;
1525 }
1526 
1527 static void era_dtr(struct dm_target *ti)
1528 {
1529 	era_destroy(ti->private);
1530 }
1531 
1532 static int era_map(struct dm_target *ti, struct bio *bio)
1533 {
1534 	struct era *era = ti->private;
1535 	dm_block_t block = get_block(era, bio);
1536 
1537 	/*
1538 	 * All bios get remapped to the origin device.  We do this now, but
1539 	 * it may not get issued until later.  Depending on whether the
1540 	 * block is marked in this era.
1541 	 */
1542 	remap_to_origin(era, bio);
1543 
1544 	/*
1545 	 * REQ_PREFLUSH bios carry no data, so we're not interested in them.
1546 	 */
1547 	if (!(bio->bi_opf & REQ_PREFLUSH) &&
1548 	    (bio_data_dir(bio) == WRITE) &&
1549 	    !metadata_current_marked(era->md, block)) {
1550 		defer_bio(era, bio);
1551 		return DM_MAPIO_SUBMITTED;
1552 	}
1553 
1554 	return DM_MAPIO_REMAPPED;
1555 }
1556 
1557 static void era_postsuspend(struct dm_target *ti)
1558 {
1559 	int r;
1560 	struct era *era = ti->private;
1561 
1562 	r = in_worker0(era, metadata_era_archive);
1563 	if (r) {
1564 		DMERR("%s: couldn't archive current era", __func__);
1565 		/* FIXME: fail mode */
1566 	}
1567 
1568 	stop_worker(era);
1569 }
1570 
1571 static int era_preresume(struct dm_target *ti)
1572 {
1573 	int r;
1574 	struct era *era = ti->private;
1575 	dm_block_t new_size = calc_nr_blocks(era);
1576 
1577 	if (era->nr_blocks != new_size) {
1578 		r = metadata_resize(era->md, &new_size);
1579 		if (r) {
1580 			DMERR("%s: metadata_resize failed", __func__);
1581 			return r;
1582 		}
1583 
1584 		r = metadata_commit(era->md);
1585 		if (r) {
1586 			DMERR("%s: metadata_commit failed", __func__);
1587 			return r;
1588 		}
1589 
1590 		era->nr_blocks = new_size;
1591 	}
1592 
1593 	start_worker(era);
1594 
1595 	r = in_worker0(era, metadata_era_rollover);
1596 	if (r) {
1597 		DMERR("%s: metadata_era_rollover failed", __func__);
1598 		return r;
1599 	}
1600 
1601 	return 0;
1602 }
1603 
1604 /*
1605  * Status format:
1606  *
1607  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
1608  * <current era> <held metadata root | '-'>
1609  */
1610 static void era_status(struct dm_target *ti, status_type_t type,
1611 		       unsigned status_flags, char *result, unsigned maxlen)
1612 {
1613 	int r;
1614 	struct era *era = ti->private;
1615 	ssize_t sz = 0;
1616 	struct metadata_stats stats;
1617 	char buf[BDEVNAME_SIZE];
1618 
1619 	switch (type) {
1620 	case STATUSTYPE_INFO:
1621 		r = in_worker1(era, metadata_get_stats, &stats);
1622 		if (r)
1623 			goto err;
1624 
1625 		DMEMIT("%u %llu/%llu %u",
1626 		       (unsigned) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
1627 		       (unsigned long long) stats.used,
1628 		       (unsigned long long) stats.total,
1629 		       (unsigned) stats.era);
1630 
1631 		if (stats.snap != SUPERBLOCK_LOCATION)
1632 			DMEMIT(" %llu", stats.snap);
1633 		else
1634 			DMEMIT(" -");
1635 		break;
1636 
1637 	case STATUSTYPE_TABLE:
1638 		format_dev_t(buf, era->metadata_dev->bdev->bd_dev);
1639 		DMEMIT("%s ", buf);
1640 		format_dev_t(buf, era->origin_dev->bdev->bd_dev);
1641 		DMEMIT("%s %u", buf, era->sectors_per_block);
1642 		break;
1643 	}
1644 
1645 	return;
1646 
1647 err:
1648 	DMEMIT("Error");
1649 }
1650 
1651 static int era_message(struct dm_target *ti, unsigned argc, char **argv,
1652 		       char *result, unsigned maxlen)
1653 {
1654 	struct era *era = ti->private;
1655 
1656 	if (argc != 1) {
1657 		DMERR("incorrect number of message arguments");
1658 		return -EINVAL;
1659 	}
1660 
1661 	if (!strcasecmp(argv[0], "checkpoint"))
1662 		return in_worker0(era, metadata_checkpoint);
1663 
1664 	if (!strcasecmp(argv[0], "take_metadata_snap"))
1665 		return in_worker0(era, metadata_take_snap);
1666 
1667 	if (!strcasecmp(argv[0], "drop_metadata_snap"))
1668 		return in_worker0(era, metadata_drop_snap);
1669 
1670 	DMERR("unsupported message '%s'", argv[0]);
1671 	return -EINVAL;
1672 }
1673 
1674 static sector_t get_dev_size(struct dm_dev *dev)
1675 {
1676 	return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
1677 }
1678 
1679 static int era_iterate_devices(struct dm_target *ti,
1680 			       iterate_devices_callout_fn fn, void *data)
1681 {
1682 	struct era *era = ti->private;
1683 	return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data);
1684 }
1685 
1686 static void era_io_hints(struct dm_target *ti, struct queue_limits *limits)
1687 {
1688 	struct era *era = ti->private;
1689 	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
1690 
1691 	/*
1692 	 * If the system-determined stacked limits are compatible with the
1693 	 * era device's blocksize (io_opt is a factor) do not override them.
1694 	 */
1695 	if (io_opt_sectors < era->sectors_per_block ||
1696 	    do_div(io_opt_sectors, era->sectors_per_block)) {
1697 		blk_limits_io_min(limits, 0);
1698 		blk_limits_io_opt(limits, era->sectors_per_block << SECTOR_SHIFT);
1699 	}
1700 }
1701 
1702 /*----------------------------------------------------------------*/
1703 
1704 static struct target_type era_target = {
1705 	.name = "era",
1706 	.version = {1, 0, 0},
1707 	.module = THIS_MODULE,
1708 	.ctr = era_ctr,
1709 	.dtr = era_dtr,
1710 	.map = era_map,
1711 	.postsuspend = era_postsuspend,
1712 	.preresume = era_preresume,
1713 	.status = era_status,
1714 	.message = era_message,
1715 	.iterate_devices = era_iterate_devices,
1716 	.io_hints = era_io_hints
1717 };
1718 
1719 static int __init dm_era_init(void)
1720 {
1721 	int r;
1722 
1723 	r = dm_register_target(&era_target);
1724 	if (r) {
1725 		DMERR("era target registration failed: %d", r);
1726 		return r;
1727 	}
1728 
1729 	return 0;
1730 }
1731 
1732 static void __exit dm_era_exit(void)
1733 {
1734 	dm_unregister_target(&era_target);
1735 }
1736 
1737 module_init(dm_era_init);
1738 module_exit(dm_era_exit);
1739 
1740 MODULE_DESCRIPTION(DM_NAME " era target");
1741 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
1742 MODULE_LICENSE("GPL");
1743