xref: /openbmc/linux/drivers/md/dm-era-target.c (revision ffcdf473)
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  *--------------------------------------------------------------
27  * Writeset
28  *--------------------------------------------------------------
29  */
30 struct writeset_metadata {
31 	uint32_t nr_bits;
32 	dm_block_t root;
33 };
34 
35 struct writeset {
36 	struct writeset_metadata md;
37 
38 	/*
39 	 * An in core copy of the bits to save constantly doing look ups on
40 	 * disk.
41 	 */
42 	unsigned long *bits;
43 };
44 
45 /*
46  * This does not free off the on disk bitset as this will normally be done
47  * after digesting into the era array.
48  */
49 static void writeset_free(struct writeset *ws)
50 {
51 	vfree(ws->bits);
52 	ws->bits = NULL;
53 }
54 
55 static int setup_on_disk_bitset(struct dm_disk_bitset *info,
56 				unsigned int nr_bits, dm_block_t *root)
57 {
58 	int r;
59 
60 	r = dm_bitset_empty(info, root);
61 	if (r)
62 		return r;
63 
64 	return dm_bitset_resize(info, *root, 0, nr_bits, false, root);
65 }
66 
67 static size_t bitset_size(unsigned int nr_bits)
68 {
69 	return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG);
70 }
71 
72 /*
73  * Allocates memory for the in core bitset.
74  */
75 static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks)
76 {
77 	ws->bits = vzalloc(bitset_size(nr_blocks));
78 	if (!ws->bits) {
79 		DMERR("%s: couldn't allocate in memory bitset", __func__);
80 		return -ENOMEM;
81 	}
82 
83 	return 0;
84 }
85 
86 /*
87  * Wipes the in-core bitset, and creates a new on disk bitset.
88  */
89 static int writeset_init(struct dm_disk_bitset *info, struct writeset *ws,
90 			 dm_block_t nr_blocks)
91 {
92 	int r;
93 
94 	memset(ws->bits, 0, bitset_size(nr_blocks));
95 
96 	ws->md.nr_bits = nr_blocks;
97 	r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root);
98 	if (r) {
99 		DMERR("%s: setup_on_disk_bitset failed", __func__);
100 		return r;
101 	}
102 
103 	return 0;
104 }
105 
106 static bool writeset_marked(struct writeset *ws, dm_block_t block)
107 {
108 	return test_bit(block, ws->bits);
109 }
110 
111 static int writeset_marked_on_disk(struct dm_disk_bitset *info,
112 				   struct writeset_metadata *m, dm_block_t block,
113 				   bool *result)
114 {
115 	int r;
116 	dm_block_t old = m->root;
117 
118 	/*
119 	 * The bitset was flushed when it was archived, so we know there'll
120 	 * be no change to the root.
121 	 */
122 	r = dm_bitset_test_bit(info, m->root, block, &m->root, result);
123 	if (r) {
124 		DMERR("%s: dm_bitset_test_bit failed", __func__);
125 		return r;
126 	}
127 
128 	BUG_ON(m->root != old);
129 
130 	return r;
131 }
132 
133 /*
134  * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was.
135  */
136 static int writeset_test_and_set(struct dm_disk_bitset *info,
137 				 struct writeset *ws, uint32_t block)
138 {
139 	int r;
140 
141 	if (!test_bit(block, ws->bits)) {
142 		r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root);
143 		if (r) {
144 			/* FIXME: fail mode */
145 			return r;
146 		}
147 
148 		return 0;
149 	}
150 
151 	return 1;
152 }
153 
154 /*
155  *--------------------------------------------------------------
156  * On disk metadata layout
157  *--------------------------------------------------------------
158  */
159 #define SPACE_MAP_ROOT_SIZE 128
160 #define UUID_LEN 16
161 
162 struct writeset_disk {
163 	__le32 nr_bits;
164 	__le64 root;
165 } __packed;
166 
167 struct superblock_disk {
168 	__le32 csum;
169 	__le32 flags;
170 	__le64 blocknr;
171 
172 	__u8 uuid[UUID_LEN];
173 	__le64 magic;
174 	__le32 version;
175 
176 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
177 
178 	__le32 data_block_size;
179 	__le32 metadata_block_size;
180 	__le32 nr_blocks;
181 
182 	__le32 current_era;
183 	struct writeset_disk current_writeset;
184 
185 	/*
186 	 * Only these two fields are valid within the metadata snapshot.
187 	 */
188 	__le64 writeset_tree_root;
189 	__le64 era_array_root;
190 
191 	__le64 metadata_snap;
192 } __packed;
193 
194 /*
195  *--------------------------------------------------------------
196  * Superblock validation
197  *--------------------------------------------------------------
198  */
199 static void sb_prepare_for_write(struct dm_block_validator *v,
200 				 struct dm_block *b,
201 				 size_t sb_block_size)
202 {
203 	struct superblock_disk *disk = dm_block_data(b);
204 
205 	disk->blocknr = cpu_to_le64(dm_block_location(b));
206 	disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags,
207 						sb_block_size - sizeof(__le32),
208 						SUPERBLOCK_CSUM_XOR));
209 }
210 
211 static int check_metadata_version(struct superblock_disk *disk)
212 {
213 	uint32_t metadata_version = le32_to_cpu(disk->version);
214 
215 	if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) {
216 		DMERR("Era metadata version %u found, but only versions between %u and %u supported.",
217 		      metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION);
218 		return -EINVAL;
219 	}
220 
221 	return 0;
222 }
223 
224 static int sb_check(struct dm_block_validator *v,
225 		    struct dm_block *b,
226 		    size_t sb_block_size)
227 {
228 	struct superblock_disk *disk = dm_block_data(b);
229 	__le32 csum_le;
230 
231 	if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) {
232 		DMERR("%s failed: blocknr %llu: wanted %llu",
233 		      __func__, le64_to_cpu(disk->blocknr),
234 		      (unsigned long long)dm_block_location(b));
235 		return -ENOTBLK;
236 	}
237 
238 	if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) {
239 		DMERR("%s failed: magic %llu: wanted %llu",
240 		      __func__, le64_to_cpu(disk->magic),
241 		      (unsigned long long) SUPERBLOCK_MAGIC);
242 		return -EILSEQ;
243 	}
244 
245 	csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags,
246 					     sb_block_size - sizeof(__le32),
247 					     SUPERBLOCK_CSUM_XOR));
248 	if (csum_le != disk->csum) {
249 		DMERR("%s failed: csum %u: wanted %u",
250 		      __func__, le32_to_cpu(csum_le), le32_to_cpu(disk->csum));
251 		return -EILSEQ;
252 	}
253 
254 	return check_metadata_version(disk);
255 }
256 
257 static struct dm_block_validator sb_validator = {
258 	.name = "superblock",
259 	.prepare_for_write = sb_prepare_for_write,
260 	.check = sb_check
261 };
262 
263 /*
264  *--------------------------------------------------------------
265  * Low level metadata handling
266  *--------------------------------------------------------------
267  */
268 #define DM_ERA_METADATA_BLOCK_SIZE 4096
269 #define ERA_MAX_CONCURRENT_LOCKS 5
270 
271 struct era_metadata {
272 	struct block_device *bdev;
273 	struct dm_block_manager *bm;
274 	struct dm_space_map *sm;
275 	struct dm_transaction_manager *tm;
276 
277 	dm_block_t block_size;
278 	uint32_t nr_blocks;
279 
280 	uint32_t current_era;
281 
282 	/*
283 	 * We preallocate 2 writesets.  When an era rolls over we
284 	 * switch between them. This means the allocation is done at
285 	 * preresume time, rather than on the io path.
286 	 */
287 	struct writeset writesets[2];
288 	struct writeset *current_writeset;
289 
290 	dm_block_t writeset_tree_root;
291 	dm_block_t era_array_root;
292 
293 	struct dm_disk_bitset bitset_info;
294 	struct dm_btree_info writeset_tree_info;
295 	struct dm_array_info era_array_info;
296 
297 	dm_block_t metadata_snap;
298 
299 	/*
300 	 * A flag that is set whenever a writeset has been archived.
301 	 */
302 	bool archived_writesets;
303 
304 	/*
305 	 * Reading the space map root can fail, so we read it into this
306 	 * buffer before the superblock is locked and updated.
307 	 */
308 	__u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
309 };
310 
311 static int superblock_read_lock(struct era_metadata *md,
312 				struct dm_block **sblock)
313 {
314 	return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION,
315 			       &sb_validator, sblock);
316 }
317 
318 static int superblock_lock_zero(struct era_metadata *md,
319 				struct dm_block **sblock)
320 {
321 	return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION,
322 				     &sb_validator, sblock);
323 }
324 
325 static int superblock_lock(struct era_metadata *md,
326 			   struct dm_block **sblock)
327 {
328 	return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION,
329 				&sb_validator, sblock);
330 }
331 
332 /* FIXME: duplication with cache and thin */
333 static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
334 {
335 	int r;
336 	unsigned int i;
337 	struct dm_block *b;
338 	__le64 *data_le, zero = cpu_to_le64(0);
339 	unsigned int sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
340 
341 	/*
342 	 * We can't use a validator here - it may be all zeroes.
343 	 */
344 	r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b);
345 	if (r)
346 		return r;
347 
348 	data_le = dm_block_data(b);
349 	*result = true;
350 	for (i = 0; i < sb_block_size; i++) {
351 		if (data_le[i] != zero) {
352 			*result = false;
353 			break;
354 		}
355 	}
356 
357 	dm_bm_unlock(b);
358 
359 	return 0;
360 }
361 
362 /*----------------------------------------------------------------*/
363 
364 static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk)
365 {
366 	disk->nr_bits = cpu_to_le32(core->nr_bits);
367 	disk->root = cpu_to_le64(core->root);
368 }
369 
370 static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core)
371 {
372 	core->nr_bits = le32_to_cpu(disk->nr_bits);
373 	core->root = le64_to_cpu(disk->root);
374 }
375 
376 static void ws_inc(void *context, const void *value, unsigned int count)
377 {
378 	struct era_metadata *md = context;
379 	struct writeset_disk ws_d;
380 	dm_block_t b;
381 	unsigned int i;
382 
383 	for (i = 0; i < count; i++) {
384 		memcpy(&ws_d, value + (i * sizeof(ws_d)), sizeof(ws_d));
385 		b = le64_to_cpu(ws_d.root);
386 		dm_tm_inc(md->tm, b);
387 	}
388 }
389 
390 static void ws_dec(void *context, const void *value, unsigned int count)
391 {
392 	struct era_metadata *md = context;
393 	struct writeset_disk ws_d;
394 	dm_block_t b;
395 	unsigned int i;
396 
397 	for (i = 0; i < count; i++) {
398 		memcpy(&ws_d, value + (i * sizeof(ws_d)), sizeof(ws_d));
399 		b = le64_to_cpu(ws_d.root);
400 		dm_bitset_del(&md->bitset_info, b);
401 	}
402 }
403 
404 static int ws_eq(void *context, const void *value1, const void *value2)
405 {
406 	return !memcmp(value1, value2, sizeof(struct writeset_disk));
407 }
408 
409 /*----------------------------------------------------------------*/
410 
411 static void setup_writeset_tree_info(struct era_metadata *md)
412 {
413 	struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type;
414 
415 	md->writeset_tree_info.tm = md->tm;
416 	md->writeset_tree_info.levels = 1;
417 	vt->context = md;
418 	vt->size = sizeof(struct writeset_disk);
419 	vt->inc = ws_inc;
420 	vt->dec = ws_dec;
421 	vt->equal = ws_eq;
422 }
423 
424 static void setup_era_array_info(struct era_metadata *md)
425 {
426 	struct dm_btree_value_type vt;
427 
428 	vt.context = NULL;
429 	vt.size = sizeof(__le32);
430 	vt.inc = NULL;
431 	vt.dec = NULL;
432 	vt.equal = NULL;
433 
434 	dm_array_info_init(&md->era_array_info, md->tm, &vt);
435 }
436 
437 static void setup_infos(struct era_metadata *md)
438 {
439 	dm_disk_bitset_init(md->tm, &md->bitset_info);
440 	setup_writeset_tree_info(md);
441 	setup_era_array_info(md);
442 }
443 
444 /*----------------------------------------------------------------*/
445 
446 static int create_fresh_metadata(struct era_metadata *md)
447 {
448 	int r;
449 
450 	r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION,
451 				 &md->tm, &md->sm);
452 	if (r < 0) {
453 		DMERR("dm_tm_create_with_sm failed");
454 		return r;
455 	}
456 
457 	setup_infos(md);
458 
459 	r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root);
460 	if (r) {
461 		DMERR("couldn't create new writeset tree");
462 		goto bad;
463 	}
464 
465 	r = dm_array_empty(&md->era_array_info, &md->era_array_root);
466 	if (r) {
467 		DMERR("couldn't create era array");
468 		goto bad;
469 	}
470 
471 	return 0;
472 
473 bad:
474 	dm_sm_destroy(md->sm);
475 	dm_tm_destroy(md->tm);
476 
477 	return r;
478 }
479 
480 static int save_sm_root(struct era_metadata *md)
481 {
482 	int r;
483 	size_t metadata_len;
484 
485 	r = dm_sm_root_size(md->sm, &metadata_len);
486 	if (r < 0)
487 		return r;
488 
489 	return dm_sm_copy_root(md->sm, &md->metadata_space_map_root,
490 			       metadata_len);
491 }
492 
493 static void copy_sm_root(struct era_metadata *md, struct superblock_disk *disk)
494 {
495 	memcpy(&disk->metadata_space_map_root,
496 	       &md->metadata_space_map_root,
497 	       sizeof(md->metadata_space_map_root));
498 }
499 
500 /*
501  * Writes a superblock, including the static fields that don't get updated
502  * with every commit (possible optimisation here).  'md' should be fully
503  * constructed when this is called.
504  */
505 static void prepare_superblock(struct era_metadata *md, struct superblock_disk *disk)
506 {
507 	disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
508 	disk->flags = cpu_to_le32(0ul);
509 
510 	/* FIXME: can't keep blanking the uuid (uuid is currently unused though) */
511 	memset(disk->uuid, 0, sizeof(disk->uuid));
512 	disk->version = cpu_to_le32(MAX_ERA_VERSION);
513 
514 	copy_sm_root(md, disk);
515 
516 	disk->data_block_size = cpu_to_le32(md->block_size);
517 	disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
518 	disk->nr_blocks = cpu_to_le32(md->nr_blocks);
519 	disk->current_era = cpu_to_le32(md->current_era);
520 
521 	ws_pack(&md->current_writeset->md, &disk->current_writeset);
522 	disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root);
523 	disk->era_array_root = cpu_to_le64(md->era_array_root);
524 	disk->metadata_snap = cpu_to_le64(md->metadata_snap);
525 }
526 
527 static int write_superblock(struct era_metadata *md)
528 {
529 	int r;
530 	struct dm_block *sblock;
531 	struct superblock_disk *disk;
532 
533 	r = save_sm_root(md);
534 	if (r) {
535 		DMERR("%s: save_sm_root failed", __func__);
536 		return r;
537 	}
538 
539 	r = superblock_lock_zero(md, &sblock);
540 	if (r)
541 		return r;
542 
543 	disk = dm_block_data(sblock);
544 	prepare_superblock(md, disk);
545 
546 	return dm_tm_commit(md->tm, sblock);
547 }
548 
549 /*
550  * Assumes block_size and the infos are set.
551  */
552 static int format_metadata(struct era_metadata *md)
553 {
554 	int r;
555 
556 	r = create_fresh_metadata(md);
557 	if (r)
558 		return r;
559 
560 	r = write_superblock(md);
561 	if (r) {
562 		dm_sm_destroy(md->sm);
563 		dm_tm_destroy(md->tm);
564 		return r;
565 	}
566 
567 	return 0;
568 }
569 
570 static int open_metadata(struct era_metadata *md)
571 {
572 	int r;
573 	struct dm_block *sblock;
574 	struct superblock_disk *disk;
575 
576 	r = superblock_read_lock(md, &sblock);
577 	if (r) {
578 		DMERR("couldn't read_lock superblock");
579 		return r;
580 	}
581 
582 	disk = dm_block_data(sblock);
583 
584 	/* Verify the data block size hasn't changed */
585 	if (le32_to_cpu(disk->data_block_size) != md->block_size) {
586 		DMERR("changing the data block size (from %u to %llu) is not supported",
587 		      le32_to_cpu(disk->data_block_size), md->block_size);
588 		r = -EINVAL;
589 		goto bad;
590 	}
591 
592 	r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION,
593 			       disk->metadata_space_map_root,
594 			       sizeof(disk->metadata_space_map_root),
595 			       &md->tm, &md->sm);
596 	if (r) {
597 		DMERR("dm_tm_open_with_sm failed");
598 		goto bad;
599 	}
600 
601 	setup_infos(md);
602 
603 	md->nr_blocks = le32_to_cpu(disk->nr_blocks);
604 	md->current_era = le32_to_cpu(disk->current_era);
605 
606 	ws_unpack(&disk->current_writeset, &md->current_writeset->md);
607 	md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root);
608 	md->era_array_root = le64_to_cpu(disk->era_array_root);
609 	md->metadata_snap = le64_to_cpu(disk->metadata_snap);
610 	md->archived_writesets = true;
611 
612 	dm_bm_unlock(sblock);
613 
614 	return 0;
615 
616 bad:
617 	dm_bm_unlock(sblock);
618 	return r;
619 }
620 
621 static int open_or_format_metadata(struct era_metadata *md,
622 				   bool may_format)
623 {
624 	int r;
625 	bool unformatted = false;
626 
627 	r = superblock_all_zeroes(md->bm, &unformatted);
628 	if (r)
629 		return r;
630 
631 	if (unformatted)
632 		return may_format ? format_metadata(md) : -EPERM;
633 
634 	return open_metadata(md);
635 }
636 
637 static int create_persistent_data_objects(struct era_metadata *md,
638 					  bool may_format)
639 {
640 	int r;
641 
642 	md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE,
643 					 ERA_MAX_CONCURRENT_LOCKS);
644 	if (IS_ERR(md->bm)) {
645 		DMERR("could not create block manager");
646 		return PTR_ERR(md->bm);
647 	}
648 
649 	r = open_or_format_metadata(md, may_format);
650 	if (r)
651 		dm_block_manager_destroy(md->bm);
652 
653 	return r;
654 }
655 
656 static void destroy_persistent_data_objects(struct era_metadata *md)
657 {
658 	dm_sm_destroy(md->sm);
659 	dm_tm_destroy(md->tm);
660 	dm_block_manager_destroy(md->bm);
661 }
662 
663 /*
664  * This waits until all era_map threads have picked up the new filter.
665  */
666 static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset)
667 {
668 	rcu_assign_pointer(md->current_writeset, new_writeset);
669 	synchronize_rcu();
670 }
671 
672 /*
673  *------------------------------------------------------------------------
674  * Writesets get 'digested' into the main era array.
675  *
676  * We're using a coroutine here so the worker thread can do the digestion,
677  * thus avoiding synchronisation of the metadata.  Digesting a whole
678  * writeset in one go would cause too much latency.
679  *------------------------------------------------------------------------
680  */
681 struct digest {
682 	uint32_t era;
683 	unsigned int nr_bits, current_bit;
684 	struct writeset_metadata writeset;
685 	__le32 value;
686 	struct dm_disk_bitset info;
687 
688 	int (*step)(struct era_metadata *md, struct digest *d);
689 };
690 
691 static int metadata_digest_lookup_writeset(struct era_metadata *md,
692 					   struct digest *d);
693 
694 static int metadata_digest_remove_writeset(struct era_metadata *md,
695 					   struct digest *d)
696 {
697 	int r;
698 	uint64_t key = d->era;
699 
700 	r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root,
701 			    &key, &md->writeset_tree_root);
702 	if (r) {
703 		DMERR("%s: dm_btree_remove failed", __func__);
704 		return r;
705 	}
706 
707 	d->step = metadata_digest_lookup_writeset;
708 	return 0;
709 }
710 
711 #define INSERTS_PER_STEP 100
712 
713 static int metadata_digest_transcribe_writeset(struct era_metadata *md,
714 					       struct digest *d)
715 {
716 	int r;
717 	bool marked;
718 	unsigned int b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits);
719 
720 	for (b = d->current_bit; b < e; b++) {
721 		r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked);
722 		if (r) {
723 			DMERR("%s: writeset_marked_on_disk failed", __func__);
724 			return r;
725 		}
726 
727 		if (!marked)
728 			continue;
729 
730 		__dm_bless_for_disk(&d->value);
731 		r = dm_array_set_value(&md->era_array_info, md->era_array_root,
732 				       b, &d->value, &md->era_array_root);
733 		if (r) {
734 			DMERR("%s: dm_array_set_value failed", __func__);
735 			return r;
736 		}
737 	}
738 
739 	if (b == d->nr_bits)
740 		d->step = metadata_digest_remove_writeset;
741 	else
742 		d->current_bit = b;
743 
744 	return 0;
745 }
746 
747 static int metadata_digest_lookup_writeset(struct era_metadata *md,
748 					   struct digest *d)
749 {
750 	int r;
751 	uint64_t key;
752 	struct writeset_disk disk;
753 
754 	r = dm_btree_find_lowest_key(&md->writeset_tree_info,
755 				     md->writeset_tree_root, &key);
756 	if (r < 0)
757 		return r;
758 
759 	d->era = key;
760 
761 	r = dm_btree_lookup(&md->writeset_tree_info,
762 			    md->writeset_tree_root, &key, &disk);
763 	if (r) {
764 		if (r == -ENODATA) {
765 			d->step = NULL;
766 			return 0;
767 		}
768 
769 		DMERR("%s: dm_btree_lookup failed", __func__);
770 		return r;
771 	}
772 
773 	ws_unpack(&disk, &d->writeset);
774 	d->value = cpu_to_le32(key);
775 
776 	/*
777 	 * We initialise another bitset info to avoid any caching side effects
778 	 * with the previous one.
779 	 */
780 	dm_disk_bitset_init(md->tm, &d->info);
781 
782 	d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks);
783 	d->current_bit = 0;
784 	d->step = metadata_digest_transcribe_writeset;
785 
786 	return 0;
787 }
788 
789 static int metadata_digest_start(struct era_metadata *md, struct digest *d)
790 {
791 	if (d->step)
792 		return 0;
793 
794 	memset(d, 0, sizeof(*d));
795 	d->step = metadata_digest_lookup_writeset;
796 
797 	return 0;
798 }
799 
800 /*
801  *-----------------------------------------------------------------
802  * High level metadata interface.  Target methods should use these,
803  * and not the lower level ones.
804  *-----------------------------------------------------------------
805  */
806 static struct era_metadata *metadata_open(struct block_device *bdev,
807 					  sector_t block_size,
808 					  bool may_format)
809 {
810 	int r;
811 	struct era_metadata *md = kzalloc(sizeof(*md), GFP_KERNEL);
812 
813 	if (!md)
814 		return NULL;
815 
816 	md->bdev = bdev;
817 	md->block_size = block_size;
818 
819 	md->writesets[0].md.root = INVALID_WRITESET_ROOT;
820 	md->writesets[1].md.root = INVALID_WRITESET_ROOT;
821 	md->current_writeset = &md->writesets[0];
822 
823 	r = create_persistent_data_objects(md, may_format);
824 	if (r) {
825 		kfree(md);
826 		return ERR_PTR(r);
827 	}
828 
829 	return md;
830 }
831 
832 static void metadata_close(struct era_metadata *md)
833 {
834 	writeset_free(&md->writesets[0]);
835 	writeset_free(&md->writesets[1]);
836 	destroy_persistent_data_objects(md);
837 	kfree(md);
838 }
839 
840 static bool valid_nr_blocks(dm_block_t n)
841 {
842 	/*
843 	 * dm_bitset restricts us to 2^32.  test_bit & co. restrict us
844 	 * further to 2^31 - 1
845 	 */
846 	return n < (1ull << 31);
847 }
848 
849 static int metadata_resize(struct era_metadata *md, void *arg)
850 {
851 	int r;
852 	dm_block_t *new_size = arg;
853 	__le32 value;
854 
855 	if (!valid_nr_blocks(*new_size)) {
856 		DMERR("Invalid number of origin blocks %llu",
857 		      (unsigned long long) *new_size);
858 		return -EINVAL;
859 	}
860 
861 	writeset_free(&md->writesets[0]);
862 	writeset_free(&md->writesets[1]);
863 
864 	r = writeset_alloc(&md->writesets[0], *new_size);
865 	if (r) {
866 		DMERR("%s: writeset_alloc failed for writeset 0", __func__);
867 		return r;
868 	}
869 
870 	r = writeset_alloc(&md->writesets[1], *new_size);
871 	if (r) {
872 		DMERR("%s: writeset_alloc failed for writeset 1", __func__);
873 		writeset_free(&md->writesets[0]);
874 		return r;
875 	}
876 
877 	value = cpu_to_le32(0u);
878 	__dm_bless_for_disk(&value);
879 	r = dm_array_resize(&md->era_array_info, md->era_array_root,
880 			    md->nr_blocks, *new_size,
881 			    &value, &md->era_array_root);
882 	if (r) {
883 		DMERR("%s: dm_array_resize failed", __func__);
884 		writeset_free(&md->writesets[0]);
885 		writeset_free(&md->writesets[1]);
886 		return r;
887 	}
888 
889 	md->nr_blocks = *new_size;
890 	return 0;
891 }
892 
893 static int metadata_era_archive(struct era_metadata *md)
894 {
895 	int r;
896 	uint64_t keys[1];
897 	struct writeset_disk value;
898 
899 	r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
900 			    &md->current_writeset->md.root);
901 	if (r) {
902 		DMERR("%s: dm_bitset_flush failed", __func__);
903 		return r;
904 	}
905 
906 	ws_pack(&md->current_writeset->md, &value);
907 
908 	keys[0] = md->current_era;
909 	__dm_bless_for_disk(&value);
910 	r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root,
911 			    keys, &value, &md->writeset_tree_root);
912 	if (r) {
913 		DMERR("%s: couldn't insert writeset into btree", __func__);
914 		/* FIXME: fail mode */
915 		return r;
916 	}
917 
918 	md->current_writeset->md.root = INVALID_WRITESET_ROOT;
919 	md->archived_writesets = true;
920 
921 	return 0;
922 }
923 
924 static struct writeset *next_writeset(struct era_metadata *md)
925 {
926 	return (md->current_writeset == &md->writesets[0]) ?
927 		&md->writesets[1] : &md->writesets[0];
928 }
929 
930 static int metadata_new_era(struct era_metadata *md)
931 {
932 	int r;
933 	struct writeset *new_writeset = next_writeset(md);
934 
935 	r = writeset_init(&md->bitset_info, new_writeset, md->nr_blocks);
936 	if (r) {
937 		DMERR("%s: writeset_init failed", __func__);
938 		return r;
939 	}
940 
941 	swap_writeset(md, new_writeset);
942 	md->current_era++;
943 
944 	return 0;
945 }
946 
947 static int metadata_era_rollover(struct era_metadata *md)
948 {
949 	int r;
950 
951 	if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
952 		r = metadata_era_archive(md);
953 		if (r) {
954 			DMERR("%s: metadata_archive_era failed", __func__);
955 			/* FIXME: fail mode? */
956 			return r;
957 		}
958 	}
959 
960 	r = metadata_new_era(md);
961 	if (r) {
962 		DMERR("%s: new era failed", __func__);
963 		/* FIXME: fail mode */
964 		return r;
965 	}
966 
967 	return 0;
968 }
969 
970 static bool metadata_current_marked(struct era_metadata *md, dm_block_t block)
971 {
972 	bool r;
973 	struct writeset *ws;
974 
975 	rcu_read_lock();
976 	ws = rcu_dereference(md->current_writeset);
977 	r = writeset_marked(ws, block);
978 	rcu_read_unlock();
979 
980 	return r;
981 }
982 
983 static int metadata_commit(struct era_metadata *md)
984 {
985 	int r;
986 	struct dm_block *sblock;
987 
988 	if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
989 		r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
990 				    &md->current_writeset->md.root);
991 		if (r) {
992 			DMERR("%s: bitset flush failed", __func__);
993 			return r;
994 		}
995 	}
996 
997 	r = dm_tm_pre_commit(md->tm);
998 	if (r) {
999 		DMERR("%s: pre commit failed", __func__);
1000 		return r;
1001 	}
1002 
1003 	r = save_sm_root(md);
1004 	if (r) {
1005 		DMERR("%s: save_sm_root failed", __func__);
1006 		return r;
1007 	}
1008 
1009 	r = superblock_lock(md, &sblock);
1010 	if (r) {
1011 		DMERR("%s: superblock lock failed", __func__);
1012 		return r;
1013 	}
1014 
1015 	prepare_superblock(md, dm_block_data(sblock));
1016 
1017 	return dm_tm_commit(md->tm, sblock);
1018 }
1019 
1020 static int metadata_checkpoint(struct era_metadata *md)
1021 {
1022 	/*
1023 	 * For now we just rollover, but later I want to put a check in to
1024 	 * avoid this if the filter is still pretty fresh.
1025 	 */
1026 	return metadata_era_rollover(md);
1027 }
1028 
1029 /*
1030  * Metadata snapshots allow userland to access era data.
1031  */
1032 static int metadata_take_snap(struct era_metadata *md)
1033 {
1034 	int r, inc;
1035 	struct dm_block *clone;
1036 
1037 	if (md->metadata_snap != SUPERBLOCK_LOCATION) {
1038 		DMERR("%s: metadata snapshot already exists", __func__);
1039 		return -EINVAL;
1040 	}
1041 
1042 	r = metadata_era_rollover(md);
1043 	if (r) {
1044 		DMERR("%s: era rollover failed", __func__);
1045 		return r;
1046 	}
1047 
1048 	r = metadata_commit(md);
1049 	if (r) {
1050 		DMERR("%s: pre commit failed", __func__);
1051 		return r;
1052 	}
1053 
1054 	r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION);
1055 	if (r) {
1056 		DMERR("%s: couldn't increment superblock", __func__);
1057 		return r;
1058 	}
1059 
1060 	r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION,
1061 			       &sb_validator, &clone, &inc);
1062 	if (r) {
1063 		DMERR("%s: couldn't shadow superblock", __func__);
1064 		dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION);
1065 		return r;
1066 	}
1067 	BUG_ON(!inc);
1068 
1069 	r = dm_sm_inc_block(md->sm, md->writeset_tree_root);
1070 	if (r) {
1071 		DMERR("%s: couldn't inc writeset tree root", __func__);
1072 		dm_tm_unlock(md->tm, clone);
1073 		return r;
1074 	}
1075 
1076 	r = dm_sm_inc_block(md->sm, md->era_array_root);
1077 	if (r) {
1078 		DMERR("%s: couldn't inc era tree root", __func__);
1079 		dm_sm_dec_block(md->sm, md->writeset_tree_root);
1080 		dm_tm_unlock(md->tm, clone);
1081 		return r;
1082 	}
1083 
1084 	md->metadata_snap = dm_block_location(clone);
1085 
1086 	dm_tm_unlock(md->tm, clone);
1087 
1088 	return 0;
1089 }
1090 
1091 static int metadata_drop_snap(struct era_metadata *md)
1092 {
1093 	int r;
1094 	dm_block_t location;
1095 	struct dm_block *clone;
1096 	struct superblock_disk *disk;
1097 
1098 	if (md->metadata_snap == SUPERBLOCK_LOCATION) {
1099 		DMERR("%s: no snap to drop", __func__);
1100 		return -EINVAL;
1101 	}
1102 
1103 	r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone);
1104 	if (r) {
1105 		DMERR("%s: couldn't read lock superblock clone", __func__);
1106 		return r;
1107 	}
1108 
1109 	/*
1110 	 * Whatever happens now we'll commit with no record of the metadata
1111 	 * snap.
1112 	 */
1113 	md->metadata_snap = SUPERBLOCK_LOCATION;
1114 
1115 	disk = dm_block_data(clone);
1116 	r = dm_btree_del(&md->writeset_tree_info,
1117 			 le64_to_cpu(disk->writeset_tree_root));
1118 	if (r) {
1119 		DMERR("%s: error deleting writeset tree clone", __func__);
1120 		dm_tm_unlock(md->tm, clone);
1121 		return r;
1122 	}
1123 
1124 	r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root));
1125 	if (r) {
1126 		DMERR("%s: error deleting era array clone", __func__);
1127 		dm_tm_unlock(md->tm, clone);
1128 		return r;
1129 	}
1130 
1131 	location = dm_block_location(clone);
1132 	dm_tm_unlock(md->tm, clone);
1133 
1134 	return dm_sm_dec_block(md->sm, location);
1135 }
1136 
1137 struct metadata_stats {
1138 	dm_block_t used;
1139 	dm_block_t total;
1140 	dm_block_t snap;
1141 	uint32_t era;
1142 };
1143 
1144 static int metadata_get_stats(struct era_metadata *md, void *ptr)
1145 {
1146 	int r;
1147 	struct metadata_stats *s = ptr;
1148 	dm_block_t nr_free, nr_total;
1149 
1150 	r = dm_sm_get_nr_free(md->sm, &nr_free);
1151 	if (r) {
1152 		DMERR("dm_sm_get_nr_free returned %d", r);
1153 		return r;
1154 	}
1155 
1156 	r = dm_sm_get_nr_blocks(md->sm, &nr_total);
1157 	if (r) {
1158 		DMERR("dm_pool_get_metadata_dev_size returned %d", r);
1159 		return r;
1160 	}
1161 
1162 	s->used = nr_total - nr_free;
1163 	s->total = nr_total;
1164 	s->snap = md->metadata_snap;
1165 	s->era = md->current_era;
1166 
1167 	return 0;
1168 }
1169 
1170 /*----------------------------------------------------------------*/
1171 
1172 struct era {
1173 	struct dm_target *ti;
1174 
1175 	struct dm_dev *metadata_dev;
1176 	struct dm_dev *origin_dev;
1177 
1178 	dm_block_t nr_blocks;
1179 	uint32_t sectors_per_block;
1180 	int sectors_per_block_shift;
1181 	struct era_metadata *md;
1182 
1183 	struct workqueue_struct *wq;
1184 	struct work_struct worker;
1185 
1186 	spinlock_t deferred_lock;
1187 	struct bio_list deferred_bios;
1188 
1189 	spinlock_t rpc_lock;
1190 	struct list_head rpc_calls;
1191 
1192 	struct digest digest;
1193 	atomic_t suspended;
1194 };
1195 
1196 struct rpc {
1197 	struct list_head list;
1198 
1199 	int (*fn0)(struct era_metadata *md);
1200 	int (*fn1)(struct era_metadata *md, void *ref);
1201 	void *arg;
1202 	int result;
1203 
1204 	struct completion complete;
1205 };
1206 
1207 /*
1208  *---------------------------------------------------------------
1209  * Remapping.
1210  *---------------------------------------------------------------
1211  */
1212 static bool block_size_is_power_of_two(struct era *era)
1213 {
1214 	return era->sectors_per_block_shift >= 0;
1215 }
1216 
1217 static dm_block_t get_block(struct era *era, struct bio *bio)
1218 {
1219 	sector_t block_nr = bio->bi_iter.bi_sector;
1220 
1221 	if (!block_size_is_power_of_two(era))
1222 		(void) sector_div(block_nr, era->sectors_per_block);
1223 	else
1224 		block_nr >>= era->sectors_per_block_shift;
1225 
1226 	return block_nr;
1227 }
1228 
1229 static void remap_to_origin(struct era *era, struct bio *bio)
1230 {
1231 	bio_set_dev(bio, era->origin_dev->bdev);
1232 }
1233 
1234 /*
1235  *--------------------------------------------------------------
1236  * Worker thread
1237  *--------------------------------------------------------------
1238  */
1239 static void wake_worker(struct era *era)
1240 {
1241 	if (!atomic_read(&era->suspended))
1242 		queue_work(era->wq, &era->worker);
1243 }
1244 
1245 static void process_old_eras(struct era *era)
1246 {
1247 	int r;
1248 
1249 	if (!era->digest.step)
1250 		return;
1251 
1252 	r = era->digest.step(era->md, &era->digest);
1253 	if (r < 0) {
1254 		DMERR("%s: digest step failed, stopping digestion", __func__);
1255 		era->digest.step = NULL;
1256 
1257 	} else if (era->digest.step)
1258 		wake_worker(era);
1259 }
1260 
1261 static void process_deferred_bios(struct era *era)
1262 {
1263 	int r;
1264 	struct bio_list deferred_bios, marked_bios;
1265 	struct bio *bio;
1266 	struct blk_plug plug;
1267 	bool commit_needed = false;
1268 	bool failed = false;
1269 	struct writeset *ws = era->md->current_writeset;
1270 
1271 	bio_list_init(&deferred_bios);
1272 	bio_list_init(&marked_bios);
1273 
1274 	spin_lock(&era->deferred_lock);
1275 	bio_list_merge(&deferred_bios, &era->deferred_bios);
1276 	bio_list_init(&era->deferred_bios);
1277 	spin_unlock(&era->deferred_lock);
1278 
1279 	if (bio_list_empty(&deferred_bios))
1280 		return;
1281 
1282 	while ((bio = bio_list_pop(&deferred_bios))) {
1283 		r = writeset_test_and_set(&era->md->bitset_info, ws,
1284 					  get_block(era, bio));
1285 		if (r < 0) {
1286 			/*
1287 			 * This is bad news, we need to rollback.
1288 			 * FIXME: finish.
1289 			 */
1290 			failed = true;
1291 		} else if (r == 0)
1292 			commit_needed = true;
1293 
1294 		bio_list_add(&marked_bios, bio);
1295 	}
1296 
1297 	if (commit_needed) {
1298 		r = metadata_commit(era->md);
1299 		if (r)
1300 			failed = true;
1301 	}
1302 
1303 	if (failed)
1304 		while ((bio = bio_list_pop(&marked_bios)))
1305 			bio_io_error(bio);
1306 	else {
1307 		blk_start_plug(&plug);
1308 		while ((bio = bio_list_pop(&marked_bios))) {
1309 			/*
1310 			 * Only update the in-core writeset if the on-disk one
1311 			 * was updated too.
1312 			 */
1313 			if (commit_needed)
1314 				set_bit(get_block(era, bio), ws->bits);
1315 			submit_bio_noacct(bio);
1316 		}
1317 		blk_finish_plug(&plug);
1318 	}
1319 }
1320 
1321 static void process_rpc_calls(struct era *era)
1322 {
1323 	int r;
1324 	bool need_commit = false;
1325 	struct list_head calls;
1326 	struct rpc *rpc, *tmp;
1327 
1328 	INIT_LIST_HEAD(&calls);
1329 	spin_lock(&era->rpc_lock);
1330 	list_splice_init(&era->rpc_calls, &calls);
1331 	spin_unlock(&era->rpc_lock);
1332 
1333 	list_for_each_entry_safe(rpc, tmp, &calls, list) {
1334 		rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg);
1335 		need_commit = true;
1336 	}
1337 
1338 	if (need_commit) {
1339 		r = metadata_commit(era->md);
1340 		if (r)
1341 			list_for_each_entry_safe(rpc, tmp, &calls, list)
1342 				rpc->result = r;
1343 	}
1344 
1345 	list_for_each_entry_safe(rpc, tmp, &calls, list)
1346 		complete(&rpc->complete);
1347 }
1348 
1349 static void kick_off_digest(struct era *era)
1350 {
1351 	if (era->md->archived_writesets) {
1352 		era->md->archived_writesets = false;
1353 		metadata_digest_start(era->md, &era->digest);
1354 	}
1355 }
1356 
1357 static void do_work(struct work_struct *ws)
1358 {
1359 	struct era *era = container_of(ws, struct era, worker);
1360 
1361 	kick_off_digest(era);
1362 	process_old_eras(era);
1363 	process_deferred_bios(era);
1364 	process_rpc_calls(era);
1365 }
1366 
1367 static void defer_bio(struct era *era, struct bio *bio)
1368 {
1369 	spin_lock(&era->deferred_lock);
1370 	bio_list_add(&era->deferred_bios, bio);
1371 	spin_unlock(&era->deferred_lock);
1372 
1373 	wake_worker(era);
1374 }
1375 
1376 /*
1377  * Make an rpc call to the worker to change the metadata.
1378  */
1379 static int perform_rpc(struct era *era, struct rpc *rpc)
1380 {
1381 	rpc->result = 0;
1382 	init_completion(&rpc->complete);
1383 
1384 	spin_lock(&era->rpc_lock);
1385 	list_add(&rpc->list, &era->rpc_calls);
1386 	spin_unlock(&era->rpc_lock);
1387 
1388 	wake_worker(era);
1389 	wait_for_completion(&rpc->complete);
1390 
1391 	return rpc->result;
1392 }
1393 
1394 static int in_worker0(struct era *era, int (*fn)(struct era_metadata *md))
1395 {
1396 	struct rpc rpc;
1397 
1398 	rpc.fn0 = fn;
1399 	rpc.fn1 = NULL;
1400 
1401 	return perform_rpc(era, &rpc);
1402 }
1403 
1404 static int in_worker1(struct era *era,
1405 		      int (*fn)(struct era_metadata *md, void *ref), void *arg)
1406 {
1407 	struct rpc rpc;
1408 
1409 	rpc.fn0 = NULL;
1410 	rpc.fn1 = fn;
1411 	rpc.arg = arg;
1412 
1413 	return perform_rpc(era, &rpc);
1414 }
1415 
1416 static void start_worker(struct era *era)
1417 {
1418 	atomic_set(&era->suspended, 0);
1419 }
1420 
1421 static void stop_worker(struct era *era)
1422 {
1423 	atomic_set(&era->suspended, 1);
1424 	drain_workqueue(era->wq);
1425 }
1426 
1427 /*
1428  *--------------------------------------------------------------
1429  * Target methods
1430  *--------------------------------------------------------------
1431  */
1432 static void era_destroy(struct era *era)
1433 {
1434 	if (era->md)
1435 		metadata_close(era->md);
1436 
1437 	if (era->wq)
1438 		destroy_workqueue(era->wq);
1439 
1440 	if (era->origin_dev)
1441 		dm_put_device(era->ti, era->origin_dev);
1442 
1443 	if (era->metadata_dev)
1444 		dm_put_device(era->ti, era->metadata_dev);
1445 
1446 	kfree(era);
1447 }
1448 
1449 static dm_block_t calc_nr_blocks(struct era *era)
1450 {
1451 	return dm_sector_div_up(era->ti->len, era->sectors_per_block);
1452 }
1453 
1454 static bool valid_block_size(dm_block_t block_size)
1455 {
1456 	bool greater_than_zero = block_size > 0;
1457 	bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0;
1458 
1459 	return greater_than_zero && multiple_of_min_block_size;
1460 }
1461 
1462 /*
1463  * <metadata dev> <data dev> <data block size (sectors)>
1464  */
1465 static int era_ctr(struct dm_target *ti, unsigned int argc, char **argv)
1466 {
1467 	int r;
1468 	char dummy;
1469 	struct era *era;
1470 	struct era_metadata *md;
1471 
1472 	if (argc != 3) {
1473 		ti->error = "Invalid argument count";
1474 		return -EINVAL;
1475 	}
1476 
1477 	era = kzalloc(sizeof(*era), GFP_KERNEL);
1478 	if (!era) {
1479 		ti->error = "Error allocating era structure";
1480 		return -ENOMEM;
1481 	}
1482 
1483 	era->ti = ti;
1484 
1485 	r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &era->metadata_dev);
1486 	if (r) {
1487 		ti->error = "Error opening metadata device";
1488 		era_destroy(era);
1489 		return -EINVAL;
1490 	}
1491 
1492 	r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &era->origin_dev);
1493 	if (r) {
1494 		ti->error = "Error opening data device";
1495 		era_destroy(era);
1496 		return -EINVAL;
1497 	}
1498 
1499 	r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy);
1500 	if (r != 1) {
1501 		ti->error = "Error parsing block size";
1502 		era_destroy(era);
1503 		return -EINVAL;
1504 	}
1505 
1506 	r = dm_set_target_max_io_len(ti, era->sectors_per_block);
1507 	if (r) {
1508 		ti->error = "could not set max io len";
1509 		era_destroy(era);
1510 		return -EINVAL;
1511 	}
1512 
1513 	if (!valid_block_size(era->sectors_per_block)) {
1514 		ti->error = "Invalid block size";
1515 		era_destroy(era);
1516 		return -EINVAL;
1517 	}
1518 	if (era->sectors_per_block & (era->sectors_per_block - 1))
1519 		era->sectors_per_block_shift = -1;
1520 	else
1521 		era->sectors_per_block_shift = __ffs(era->sectors_per_block);
1522 
1523 	md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true);
1524 	if (IS_ERR(md)) {
1525 		ti->error = "Error reading metadata";
1526 		era_destroy(era);
1527 		return PTR_ERR(md);
1528 	}
1529 	era->md = md;
1530 
1531 	era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
1532 	if (!era->wq) {
1533 		ti->error = "could not create workqueue for metadata object";
1534 		era_destroy(era);
1535 		return -ENOMEM;
1536 	}
1537 	INIT_WORK(&era->worker, do_work);
1538 
1539 	spin_lock_init(&era->deferred_lock);
1540 	bio_list_init(&era->deferred_bios);
1541 
1542 	spin_lock_init(&era->rpc_lock);
1543 	INIT_LIST_HEAD(&era->rpc_calls);
1544 
1545 	ti->private = era;
1546 	ti->num_flush_bios = 1;
1547 	ti->flush_supported = true;
1548 
1549 	ti->num_discard_bios = 1;
1550 
1551 	return 0;
1552 }
1553 
1554 static void era_dtr(struct dm_target *ti)
1555 {
1556 	era_destroy(ti->private);
1557 }
1558 
1559 static int era_map(struct dm_target *ti, struct bio *bio)
1560 {
1561 	struct era *era = ti->private;
1562 	dm_block_t block = get_block(era, bio);
1563 
1564 	/*
1565 	 * All bios get remapped to the origin device.  We do this now, but
1566 	 * it may not get issued until later.  Depending on whether the
1567 	 * block is marked in this era.
1568 	 */
1569 	remap_to_origin(era, bio);
1570 
1571 	/*
1572 	 * REQ_PREFLUSH bios carry no data, so we're not interested in them.
1573 	 */
1574 	if (!(bio->bi_opf & REQ_PREFLUSH) &&
1575 	    (bio_data_dir(bio) == WRITE) &&
1576 	    !metadata_current_marked(era->md, block)) {
1577 		defer_bio(era, bio);
1578 		return DM_MAPIO_SUBMITTED;
1579 	}
1580 
1581 	return DM_MAPIO_REMAPPED;
1582 }
1583 
1584 static void era_postsuspend(struct dm_target *ti)
1585 {
1586 	int r;
1587 	struct era *era = ti->private;
1588 
1589 	r = in_worker0(era, metadata_era_archive);
1590 	if (r) {
1591 		DMERR("%s: couldn't archive current era", __func__);
1592 		/* FIXME: fail mode */
1593 	}
1594 
1595 	stop_worker(era);
1596 
1597 	r = metadata_commit(era->md);
1598 	if (r) {
1599 		DMERR("%s: metadata_commit failed", __func__);
1600 		/* FIXME: fail mode */
1601 	}
1602 }
1603 
1604 static int era_preresume(struct dm_target *ti)
1605 {
1606 	int r;
1607 	struct era *era = ti->private;
1608 	dm_block_t new_size = calc_nr_blocks(era);
1609 
1610 	if (era->nr_blocks != new_size) {
1611 		r = metadata_resize(era->md, &new_size);
1612 		if (r) {
1613 			DMERR("%s: metadata_resize failed", __func__);
1614 			return r;
1615 		}
1616 
1617 		r = metadata_commit(era->md);
1618 		if (r) {
1619 			DMERR("%s: metadata_commit failed", __func__);
1620 			return r;
1621 		}
1622 
1623 		era->nr_blocks = new_size;
1624 	}
1625 
1626 	start_worker(era);
1627 
1628 	r = in_worker0(era, metadata_era_rollover);
1629 	if (r) {
1630 		DMERR("%s: metadata_era_rollover failed", __func__);
1631 		return r;
1632 	}
1633 
1634 	return 0;
1635 }
1636 
1637 /*
1638  * Status format:
1639  *
1640  * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
1641  * <current era> <held metadata root | '-'>
1642  */
1643 static void era_status(struct dm_target *ti, status_type_t type,
1644 		       unsigned int status_flags, char *result, unsigned int maxlen)
1645 {
1646 	int r;
1647 	struct era *era = ti->private;
1648 	ssize_t sz = 0;
1649 	struct metadata_stats stats;
1650 	char buf[BDEVNAME_SIZE];
1651 
1652 	switch (type) {
1653 	case STATUSTYPE_INFO:
1654 		r = in_worker1(era, metadata_get_stats, &stats);
1655 		if (r)
1656 			goto err;
1657 
1658 		DMEMIT("%u %llu/%llu %u",
1659 		       (unsigned int) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
1660 		       (unsigned long long) stats.used,
1661 		       (unsigned long long) stats.total,
1662 		       (unsigned int) stats.era);
1663 
1664 		if (stats.snap != SUPERBLOCK_LOCATION)
1665 			DMEMIT(" %llu", stats.snap);
1666 		else
1667 			DMEMIT(" -");
1668 		break;
1669 
1670 	case STATUSTYPE_TABLE:
1671 		format_dev_t(buf, era->metadata_dev->bdev->bd_dev);
1672 		DMEMIT("%s ", buf);
1673 		format_dev_t(buf, era->origin_dev->bdev->bd_dev);
1674 		DMEMIT("%s %u", buf, era->sectors_per_block);
1675 		break;
1676 
1677 	case STATUSTYPE_IMA:
1678 		*result = '\0';
1679 		break;
1680 	}
1681 
1682 	return;
1683 
1684 err:
1685 	DMEMIT("Error");
1686 }
1687 
1688 static int era_message(struct dm_target *ti, unsigned int argc, char **argv,
1689 		       char *result, unsigned int maxlen)
1690 {
1691 	struct era *era = ti->private;
1692 
1693 	if (argc != 1) {
1694 		DMERR("incorrect number of message arguments");
1695 		return -EINVAL;
1696 	}
1697 
1698 	if (!strcasecmp(argv[0], "checkpoint"))
1699 		return in_worker0(era, metadata_checkpoint);
1700 
1701 	if (!strcasecmp(argv[0], "take_metadata_snap"))
1702 		return in_worker0(era, metadata_take_snap);
1703 
1704 	if (!strcasecmp(argv[0], "drop_metadata_snap"))
1705 		return in_worker0(era, metadata_drop_snap);
1706 
1707 	DMERR("unsupported message '%s'", argv[0]);
1708 	return -EINVAL;
1709 }
1710 
1711 static sector_t get_dev_size(struct dm_dev *dev)
1712 {
1713 	return bdev_nr_sectors(dev->bdev);
1714 }
1715 
1716 static int era_iterate_devices(struct dm_target *ti,
1717 			       iterate_devices_callout_fn fn, void *data)
1718 {
1719 	struct era *era = ti->private;
1720 
1721 	return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data);
1722 }
1723 
1724 static void era_io_hints(struct dm_target *ti, struct queue_limits *limits)
1725 {
1726 	struct era *era = ti->private;
1727 	uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
1728 
1729 	/*
1730 	 * If the system-determined stacked limits are compatible with the
1731 	 * era device's blocksize (io_opt is a factor) do not override them.
1732 	 */
1733 	if (io_opt_sectors < era->sectors_per_block ||
1734 	    do_div(io_opt_sectors, era->sectors_per_block)) {
1735 		blk_limits_io_min(limits, 0);
1736 		blk_limits_io_opt(limits, era->sectors_per_block << SECTOR_SHIFT);
1737 	}
1738 }
1739 
1740 /*----------------------------------------------------------------*/
1741 
1742 static struct target_type era_target = {
1743 	.name = "era",
1744 	.version = {1, 0, 0},
1745 	.module = THIS_MODULE,
1746 	.ctr = era_ctr,
1747 	.dtr = era_dtr,
1748 	.map = era_map,
1749 	.postsuspend = era_postsuspend,
1750 	.preresume = era_preresume,
1751 	.status = era_status,
1752 	.message = era_message,
1753 	.iterate_devices = era_iterate_devices,
1754 	.io_hints = era_io_hints
1755 };
1756 module_dm(era);
1757 
1758 MODULE_DESCRIPTION(DM_NAME " era target");
1759 MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
1760 MODULE_LICENSE("GPL");
1761