xref: /linux/drivers/md/dm-snap-persistent.c (revision 0526b56cbc3c489642bd6a5fe4b718dea7ef0ee8)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
4  * Copyright (C) 2006-2008 Red Hat GmbH
5  *
6  * This file is released under the GPL.
7  */
8 
9 #include "dm-exception-store.h"
10 
11 #include <linux/ctype.h>
12 #include <linux/mm.h>
13 #include <linux/pagemap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/export.h>
16 #include <linux/slab.h>
17 #include <linux/dm-io.h>
18 #include <linux/dm-bufio.h>
19 
20 #define DM_MSG_PREFIX "persistent snapshot"
21 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32U	/* 16KB */
22 
23 #define DM_PREFETCH_CHUNKS		12
24 
25 /*
26  *---------------------------------------------------------------
27  * Persistent snapshots, by persistent we mean that the snapshot
28  * will survive a reboot.
29  *---------------------------------------------------------------
30  */
31 
32 /*
33  * We need to store a record of which parts of the origin have
34  * been copied to the snapshot device.  The snapshot code
35  * requires that we copy exception chunks to chunk aligned areas
36  * of the COW store.  It makes sense therefore, to store the
37  * metadata in chunk size blocks.
38  *
39  * There is no backward or forward compatibility implemented,
40  * snapshots with different disk versions than the kernel will
41  * not be usable.  It is expected that "lvcreate" will blank out
42  * the start of a fresh COW device before calling the snapshot
43  * constructor.
44  *
45  * The first chunk of the COW device just contains the header.
46  * After this there is a chunk filled with exception metadata,
47  * followed by as many exception chunks as can fit in the
48  * metadata areas.
49  *
50  * All on disk structures are in little-endian format.  The end
51  * of the exceptions info is indicated by an exception with a
52  * new_chunk of 0, which is invalid since it would point to the
53  * header chunk.
54  */
55 
56 /*
57  * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
58  */
59 #define SNAP_MAGIC 0x70416e53
60 
61 /*
62  * The on-disk version of the metadata.
63  */
64 #define SNAPSHOT_DISK_VERSION 1
65 
66 #define NUM_SNAPSHOT_HDR_CHUNKS 1
67 
68 struct disk_header {
69 	__le32 magic;
70 
71 	/*
72 	 * Is this snapshot valid.  There is no way of recovering
73 	 * an invalid snapshot.
74 	 */
75 	__le32 valid;
76 
77 	/*
78 	 * Simple, incrementing version. no backward
79 	 * compatibility.
80 	 */
81 	__le32 version;
82 
83 	/* In sectors */
84 	__le32 chunk_size;
85 } __packed;
86 
87 struct disk_exception {
88 	__le64 old_chunk;
89 	__le64 new_chunk;
90 } __packed;
91 
92 struct core_exception {
93 	uint64_t old_chunk;
94 	uint64_t new_chunk;
95 };
96 
97 struct commit_callback {
98 	void (*callback)(void *ref, int success);
99 	void *context;
100 };
101 
102 /*
103  * The top level structure for a persistent exception store.
104  */
105 struct pstore {
106 	struct dm_exception_store *store;
107 	int version;
108 	int valid;
109 	uint32_t exceptions_per_area;
110 
111 	/*
112 	 * Now that we have an asynchronous kcopyd there is no
113 	 * need for large chunk sizes, so it wont hurt to have a
114 	 * whole chunks worth of metadata in memory at once.
115 	 */
116 	void *area;
117 
118 	/*
119 	 * An area of zeros used to clear the next area.
120 	 */
121 	void *zero_area;
122 
123 	/*
124 	 * An area used for header. The header can be written
125 	 * concurrently with metadata (when invalidating the snapshot),
126 	 * so it needs a separate buffer.
127 	 */
128 	void *header_area;
129 
130 	/*
131 	 * Used to keep track of which metadata area the data in
132 	 * 'chunk' refers to.
133 	 */
134 	chunk_t current_area;
135 
136 	/*
137 	 * The next free chunk for an exception.
138 	 *
139 	 * When creating exceptions, all the chunks here and above are
140 	 * free.  It holds the next chunk to be allocated.  On rare
141 	 * occasions (e.g. after a system crash) holes can be left in
142 	 * the exception store because chunks can be committed out of
143 	 * order.
144 	 *
145 	 * When merging exceptions, it does not necessarily mean all the
146 	 * chunks here and above are free.  It holds the value it would
147 	 * have held if all chunks had been committed in order of
148 	 * allocation.  Consequently the value may occasionally be
149 	 * slightly too low, but since it's only used for 'status' and
150 	 * it can never reach its minimum value too early this doesn't
151 	 * matter.
152 	 */
153 
154 	chunk_t next_free;
155 
156 	/*
157 	 * The index of next free exception in the current
158 	 * metadata area.
159 	 */
160 	uint32_t current_committed;
161 
162 	atomic_t pending_count;
163 	uint32_t callback_count;
164 	struct commit_callback *callbacks;
165 	struct dm_io_client *io_client;
166 
167 	struct workqueue_struct *metadata_wq;
168 };
169 
170 static int alloc_area(struct pstore *ps)
171 {
172 	int r = -ENOMEM;
173 	size_t len;
174 
175 	len = ps->store->chunk_size << SECTOR_SHIFT;
176 
177 	/*
178 	 * Allocate the chunk_size block of memory that will hold
179 	 * a single metadata area.
180 	 */
181 	ps->area = vmalloc(len);
182 	if (!ps->area)
183 		goto err_area;
184 
185 	ps->zero_area = vzalloc(len);
186 	if (!ps->zero_area)
187 		goto err_zero_area;
188 
189 	ps->header_area = vmalloc(len);
190 	if (!ps->header_area)
191 		goto err_header_area;
192 
193 	return 0;
194 
195 err_header_area:
196 	vfree(ps->zero_area);
197 
198 err_zero_area:
199 	vfree(ps->area);
200 
201 err_area:
202 	return r;
203 }
204 
205 static void free_area(struct pstore *ps)
206 {
207 	vfree(ps->area);
208 	ps->area = NULL;
209 	vfree(ps->zero_area);
210 	ps->zero_area = NULL;
211 	vfree(ps->header_area);
212 	ps->header_area = NULL;
213 }
214 
215 struct mdata_req {
216 	struct dm_io_region *where;
217 	struct dm_io_request *io_req;
218 	struct work_struct work;
219 	int result;
220 };
221 
222 static void do_metadata(struct work_struct *work)
223 {
224 	struct mdata_req *req = container_of(work, struct mdata_req, work);
225 
226 	req->result = dm_io(req->io_req, 1, req->where, NULL);
227 }
228 
229 /*
230  * Read or write a chunk aligned and sized block of data from a device.
231  */
232 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, blk_opf_t opf,
233 		    int metadata)
234 {
235 	struct dm_io_region where = {
236 		.bdev = dm_snap_cow(ps->store->snap)->bdev,
237 		.sector = ps->store->chunk_size * chunk,
238 		.count = ps->store->chunk_size,
239 	};
240 	struct dm_io_request io_req = {
241 		.bi_opf = opf,
242 		.mem.type = DM_IO_VMA,
243 		.mem.ptr.vma = area,
244 		.client = ps->io_client,
245 		.notify.fn = NULL,
246 	};
247 	struct mdata_req req;
248 
249 	if (!metadata)
250 		return dm_io(&io_req, 1, &where, NULL);
251 
252 	req.where = &where;
253 	req.io_req = &io_req;
254 
255 	/*
256 	 * Issue the synchronous I/O from a different thread
257 	 * to avoid submit_bio_noacct recursion.
258 	 */
259 	INIT_WORK_ONSTACK(&req.work, do_metadata);
260 	queue_work(ps->metadata_wq, &req.work);
261 	flush_workqueue(ps->metadata_wq);
262 	destroy_work_on_stack(&req.work);
263 
264 	return req.result;
265 }
266 
267 /*
268  * Convert a metadata area index to a chunk index.
269  */
270 static chunk_t area_location(struct pstore *ps, chunk_t area)
271 {
272 	return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
273 }
274 
275 static void skip_metadata(struct pstore *ps)
276 {
277 	uint32_t stride = ps->exceptions_per_area + 1;
278 	chunk_t next_free = ps->next_free;
279 
280 	if (sector_div(next_free, stride) == NUM_SNAPSHOT_HDR_CHUNKS)
281 		ps->next_free++;
282 }
283 
284 /*
285  * Read or write a metadata area.  Remembering to skip the first
286  * chunk which holds the header.
287  */
288 static int area_io(struct pstore *ps, blk_opf_t opf)
289 {
290 	chunk_t chunk = area_location(ps, ps->current_area);
291 
292 	return chunk_io(ps, ps->area, chunk, opf, 0);
293 }
294 
295 static void zero_memory_area(struct pstore *ps)
296 {
297 	memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
298 }
299 
300 static int zero_disk_area(struct pstore *ps, chunk_t area)
301 {
302 	return chunk_io(ps, ps->zero_area, area_location(ps, area),
303 			REQ_OP_WRITE, 0);
304 }
305 
306 static int read_header(struct pstore *ps, int *new_snapshot)
307 {
308 	int r;
309 	struct disk_header *dh;
310 	unsigned int chunk_size;
311 	int chunk_size_supplied = 1;
312 	char *chunk_err;
313 
314 	/*
315 	 * Use default chunk size (or logical_block_size, if larger)
316 	 * if none supplied
317 	 */
318 	if (!ps->store->chunk_size) {
319 		ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
320 		    bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
321 					    bdev) >> 9);
322 		ps->store->chunk_mask = ps->store->chunk_size - 1;
323 		ps->store->chunk_shift = __ffs(ps->store->chunk_size);
324 		chunk_size_supplied = 0;
325 	}
326 
327 	ps->io_client = dm_io_client_create();
328 	if (IS_ERR(ps->io_client))
329 		return PTR_ERR(ps->io_client);
330 
331 	r = alloc_area(ps);
332 	if (r)
333 		return r;
334 
335 	r = chunk_io(ps, ps->header_area, 0, REQ_OP_READ, 1);
336 	if (r)
337 		goto bad;
338 
339 	dh = ps->header_area;
340 
341 	if (le32_to_cpu(dh->magic) == 0) {
342 		*new_snapshot = 1;
343 		return 0;
344 	}
345 
346 	if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
347 		DMWARN("Invalid or corrupt snapshot");
348 		r = -ENXIO;
349 		goto bad;
350 	}
351 
352 	*new_snapshot = 0;
353 	ps->valid = le32_to_cpu(dh->valid);
354 	ps->version = le32_to_cpu(dh->version);
355 	chunk_size = le32_to_cpu(dh->chunk_size);
356 
357 	if (ps->store->chunk_size == chunk_size)
358 		return 0;
359 
360 	if (chunk_size_supplied)
361 		DMWARN("chunk size %u in device metadata overrides table chunk size of %u.",
362 		       chunk_size, ps->store->chunk_size);
363 
364 	/* We had a bogus chunk_size. Fix stuff up. */
365 	free_area(ps);
366 
367 	r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
368 					      &chunk_err);
369 	if (r) {
370 		DMERR("invalid on-disk chunk size %u: %s.",
371 		      chunk_size, chunk_err);
372 		return r;
373 	}
374 
375 	r = alloc_area(ps);
376 	return r;
377 
378 bad:
379 	free_area(ps);
380 	return r;
381 }
382 
383 static int write_header(struct pstore *ps)
384 {
385 	struct disk_header *dh;
386 
387 	memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
388 
389 	dh = ps->header_area;
390 	dh->magic = cpu_to_le32(SNAP_MAGIC);
391 	dh->valid = cpu_to_le32(ps->valid);
392 	dh->version = cpu_to_le32(ps->version);
393 	dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
394 
395 	return chunk_io(ps, ps->header_area, 0, REQ_OP_WRITE, 1);
396 }
397 
398 /*
399  * Access functions for the disk exceptions, these do the endian conversions.
400  */
401 static struct disk_exception *get_exception(struct pstore *ps, void *ps_area,
402 					    uint32_t index)
403 {
404 	BUG_ON(index >= ps->exceptions_per_area);
405 
406 	return ((struct disk_exception *) ps_area) + index;
407 }
408 
409 static void read_exception(struct pstore *ps, void *ps_area,
410 			   uint32_t index, struct core_exception *result)
411 {
412 	struct disk_exception *de = get_exception(ps, ps_area, index);
413 
414 	/* copy it */
415 	result->old_chunk = le64_to_cpu(de->old_chunk);
416 	result->new_chunk = le64_to_cpu(de->new_chunk);
417 }
418 
419 static void write_exception(struct pstore *ps,
420 			    uint32_t index, struct core_exception *e)
421 {
422 	struct disk_exception *de = get_exception(ps, ps->area, index);
423 
424 	/* copy it */
425 	de->old_chunk = cpu_to_le64(e->old_chunk);
426 	de->new_chunk = cpu_to_le64(e->new_chunk);
427 }
428 
429 static void clear_exception(struct pstore *ps, uint32_t index)
430 {
431 	struct disk_exception *de = get_exception(ps, ps->area, index);
432 
433 	/* clear it */
434 	de->old_chunk = 0;
435 	de->new_chunk = 0;
436 }
437 
438 /*
439  * Registers the exceptions that are present in the current area.
440  * 'full' is filled in to indicate if the area has been
441  * filled.
442  */
443 static int insert_exceptions(struct pstore *ps, void *ps_area,
444 			     int (*callback)(void *callback_context,
445 					     chunk_t old, chunk_t new),
446 			     void *callback_context,
447 			     int *full)
448 {
449 	int r;
450 	unsigned int i;
451 	struct core_exception e;
452 
453 	/* presume the area is full */
454 	*full = 1;
455 
456 	for (i = 0; i < ps->exceptions_per_area; i++) {
457 		read_exception(ps, ps_area, i, &e);
458 
459 		/*
460 		 * If the new_chunk is pointing at the start of
461 		 * the COW device, where the first metadata area
462 		 * is we know that we've hit the end of the
463 		 * exceptions.  Therefore the area is not full.
464 		 */
465 		if (e.new_chunk == 0LL) {
466 			ps->current_committed = i;
467 			*full = 0;
468 			break;
469 		}
470 
471 		/*
472 		 * Keep track of the start of the free chunks.
473 		 */
474 		if (ps->next_free <= e.new_chunk)
475 			ps->next_free = e.new_chunk + 1;
476 
477 		/*
478 		 * Otherwise we add the exception to the snapshot.
479 		 */
480 		r = callback(callback_context, e.old_chunk, e.new_chunk);
481 		if (r)
482 			return r;
483 	}
484 
485 	return 0;
486 }
487 
488 static int read_exceptions(struct pstore *ps,
489 			   int (*callback)(void *callback_context, chunk_t old,
490 					   chunk_t new),
491 			   void *callback_context)
492 {
493 	int r, full = 1;
494 	struct dm_bufio_client *client;
495 	chunk_t prefetch_area = 0;
496 
497 	client = dm_bufio_client_create(dm_snap_cow(ps->store->snap)->bdev,
498 					ps->store->chunk_size << SECTOR_SHIFT,
499 					1, 0, NULL, NULL, 0);
500 
501 	if (IS_ERR(client))
502 		return PTR_ERR(client);
503 
504 	/*
505 	 * Setup for one current buffer + desired readahead buffers.
506 	 */
507 	dm_bufio_set_minimum_buffers(client, 1 + DM_PREFETCH_CHUNKS);
508 
509 	/*
510 	 * Keeping reading chunks and inserting exceptions until
511 	 * we find a partially full area.
512 	 */
513 	for (ps->current_area = 0; full; ps->current_area++) {
514 		struct dm_buffer *bp;
515 		void *area;
516 		chunk_t chunk;
517 
518 		if (unlikely(prefetch_area < ps->current_area))
519 			prefetch_area = ps->current_area;
520 
521 		if (DM_PREFETCH_CHUNKS) {
522 			do {
523 				chunk_t pf_chunk = area_location(ps, prefetch_area);
524 
525 				if (unlikely(pf_chunk >= dm_bufio_get_device_size(client)))
526 					break;
527 				dm_bufio_prefetch(client, pf_chunk, 1);
528 				prefetch_area++;
529 				if (unlikely(!prefetch_area))
530 					break;
531 			} while (prefetch_area <= ps->current_area + DM_PREFETCH_CHUNKS);
532 		}
533 
534 		chunk = area_location(ps, ps->current_area);
535 
536 		area = dm_bufio_read(client, chunk, &bp);
537 		if (IS_ERR(area)) {
538 			r = PTR_ERR(area);
539 			goto ret_destroy_bufio;
540 		}
541 
542 		r = insert_exceptions(ps, area, callback, callback_context,
543 				      &full);
544 
545 		if (!full)
546 			memcpy(ps->area, area, ps->store->chunk_size << SECTOR_SHIFT);
547 
548 		dm_bufio_release(bp);
549 
550 		dm_bufio_forget(client, chunk);
551 
552 		if (unlikely(r))
553 			goto ret_destroy_bufio;
554 	}
555 
556 	ps->current_area--;
557 
558 	skip_metadata(ps);
559 
560 	r = 0;
561 
562 ret_destroy_bufio:
563 	dm_bufio_client_destroy(client);
564 
565 	return r;
566 }
567 
568 static struct pstore *get_info(struct dm_exception_store *store)
569 {
570 	return store->context;
571 }
572 
573 static void persistent_usage(struct dm_exception_store *store,
574 			     sector_t *total_sectors,
575 			     sector_t *sectors_allocated,
576 			     sector_t *metadata_sectors)
577 {
578 	struct pstore *ps = get_info(store);
579 
580 	*sectors_allocated = ps->next_free * store->chunk_size;
581 	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
582 
583 	/*
584 	 * First chunk is the fixed header.
585 	 * Then there are (ps->current_area + 1) metadata chunks, each one
586 	 * separated from the next by ps->exceptions_per_area data chunks.
587 	 */
588 	*metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
589 			    store->chunk_size;
590 }
591 
592 static void persistent_dtr(struct dm_exception_store *store)
593 {
594 	struct pstore *ps = get_info(store);
595 
596 	destroy_workqueue(ps->metadata_wq);
597 
598 	/* Created in read_header */
599 	if (ps->io_client)
600 		dm_io_client_destroy(ps->io_client);
601 	free_area(ps);
602 
603 	/* Allocated in persistent_read_metadata */
604 	kvfree(ps->callbacks);
605 
606 	kfree(ps);
607 }
608 
609 static int persistent_read_metadata(struct dm_exception_store *store,
610 				    int (*callback)(void *callback_context,
611 						    chunk_t old, chunk_t new),
612 				    void *callback_context)
613 {
614 	int r, new_snapshot;
615 	struct pstore *ps = get_info(store);
616 
617 	/*
618 	 * Read the snapshot header.
619 	 */
620 	r = read_header(ps, &new_snapshot);
621 	if (r)
622 		return r;
623 
624 	/*
625 	 * Now we know correct chunk_size, complete the initialisation.
626 	 */
627 	ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
628 				  sizeof(struct disk_exception);
629 	ps->callbacks = kvcalloc(ps->exceptions_per_area,
630 				 sizeof(*ps->callbacks), GFP_KERNEL);
631 	if (!ps->callbacks)
632 		return -ENOMEM;
633 
634 	/*
635 	 * Do we need to setup a new snapshot ?
636 	 */
637 	if (new_snapshot) {
638 		r = write_header(ps);
639 		if (r) {
640 			DMWARN("write_header failed");
641 			return r;
642 		}
643 
644 		ps->current_area = 0;
645 		zero_memory_area(ps);
646 		r = zero_disk_area(ps, 0);
647 		if (r)
648 			DMWARN("zero_disk_area(0) failed");
649 		return r;
650 	}
651 	/*
652 	 * Sanity checks.
653 	 */
654 	if (ps->version != SNAPSHOT_DISK_VERSION) {
655 		DMWARN("unable to handle snapshot disk version %d",
656 		       ps->version);
657 		return -EINVAL;
658 	}
659 
660 	/*
661 	 * Metadata are valid, but snapshot is invalidated
662 	 */
663 	if (!ps->valid)
664 		return 1;
665 
666 	/*
667 	 * Read the metadata.
668 	 */
669 	r = read_exceptions(ps, callback, callback_context);
670 
671 	return r;
672 }
673 
674 static int persistent_prepare_exception(struct dm_exception_store *store,
675 					struct dm_exception *e)
676 {
677 	struct pstore *ps = get_info(store);
678 	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
679 
680 	/* Is there enough room ? */
681 	if (size < ((ps->next_free + 1) * store->chunk_size))
682 		return -ENOSPC;
683 
684 	e->new_chunk = ps->next_free;
685 
686 	/*
687 	 * Move onto the next free pending, making sure to take
688 	 * into account the location of the metadata chunks.
689 	 */
690 	ps->next_free++;
691 	skip_metadata(ps);
692 
693 	atomic_inc(&ps->pending_count);
694 	return 0;
695 }
696 
697 static void persistent_commit_exception(struct dm_exception_store *store,
698 					struct dm_exception *e, int valid,
699 					void (*callback)(void *, int success),
700 					void *callback_context)
701 {
702 	unsigned int i;
703 	struct pstore *ps = get_info(store);
704 	struct core_exception ce;
705 	struct commit_callback *cb;
706 
707 	if (!valid)
708 		ps->valid = 0;
709 
710 	ce.old_chunk = e->old_chunk;
711 	ce.new_chunk = e->new_chunk;
712 	write_exception(ps, ps->current_committed++, &ce);
713 
714 	/*
715 	 * Add the callback to the back of the array.  This code
716 	 * is the only place where the callback array is
717 	 * manipulated, and we know that it will never be called
718 	 * multiple times concurrently.
719 	 */
720 	cb = ps->callbacks + ps->callback_count++;
721 	cb->callback = callback;
722 	cb->context = callback_context;
723 
724 	/*
725 	 * If there are exceptions in flight and we have not yet
726 	 * filled this metadata area there's nothing more to do.
727 	 */
728 	if (!atomic_dec_and_test(&ps->pending_count) &&
729 	    (ps->current_committed != ps->exceptions_per_area))
730 		return;
731 
732 	/*
733 	 * If we completely filled the current area, then wipe the next one.
734 	 */
735 	if ((ps->current_committed == ps->exceptions_per_area) &&
736 	    zero_disk_area(ps, ps->current_area + 1))
737 		ps->valid = 0;
738 
739 	/*
740 	 * Commit exceptions to disk.
741 	 */
742 	if (ps->valid && area_io(ps, REQ_OP_WRITE | REQ_PREFLUSH | REQ_FUA |
743 				 REQ_SYNC))
744 		ps->valid = 0;
745 
746 	/*
747 	 * Advance to the next area if this one is full.
748 	 */
749 	if (ps->current_committed == ps->exceptions_per_area) {
750 		ps->current_committed = 0;
751 		ps->current_area++;
752 		zero_memory_area(ps);
753 	}
754 
755 	for (i = 0; i < ps->callback_count; i++) {
756 		cb = ps->callbacks + i;
757 		cb->callback(cb->context, ps->valid);
758 	}
759 
760 	ps->callback_count = 0;
761 }
762 
763 static int persistent_prepare_merge(struct dm_exception_store *store,
764 				    chunk_t *last_old_chunk,
765 				    chunk_t *last_new_chunk)
766 {
767 	struct pstore *ps = get_info(store);
768 	struct core_exception ce;
769 	int nr_consecutive;
770 	int r;
771 
772 	/*
773 	 * When current area is empty, move back to preceding area.
774 	 */
775 	if (!ps->current_committed) {
776 		/*
777 		 * Have we finished?
778 		 */
779 		if (!ps->current_area)
780 			return 0;
781 
782 		ps->current_area--;
783 		r = area_io(ps, REQ_OP_READ);
784 		if (r < 0)
785 			return r;
786 		ps->current_committed = ps->exceptions_per_area;
787 	}
788 
789 	read_exception(ps, ps->area, ps->current_committed - 1, &ce);
790 	*last_old_chunk = ce.old_chunk;
791 	*last_new_chunk = ce.new_chunk;
792 
793 	/*
794 	 * Find number of consecutive chunks within the current area,
795 	 * working backwards.
796 	 */
797 	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
798 	     nr_consecutive++) {
799 		read_exception(ps, ps->area,
800 			       ps->current_committed - 1 - nr_consecutive, &ce);
801 		if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
802 		    ce.new_chunk != *last_new_chunk - nr_consecutive)
803 			break;
804 	}
805 
806 	return nr_consecutive;
807 }
808 
809 static int persistent_commit_merge(struct dm_exception_store *store,
810 				   int nr_merged)
811 {
812 	int r, i;
813 	struct pstore *ps = get_info(store);
814 
815 	BUG_ON(nr_merged > ps->current_committed);
816 
817 	for (i = 0; i < nr_merged; i++)
818 		clear_exception(ps, ps->current_committed - 1 - i);
819 
820 	r = area_io(ps, REQ_OP_WRITE | REQ_PREFLUSH | REQ_FUA);
821 	if (r < 0)
822 		return r;
823 
824 	ps->current_committed -= nr_merged;
825 
826 	/*
827 	 * At this stage, only persistent_usage() uses ps->next_free, so
828 	 * we make no attempt to keep ps->next_free strictly accurate
829 	 * as exceptions may have been committed out-of-order originally.
830 	 * Once a snapshot has become merging, we set it to the value it
831 	 * would have held had all the exceptions been committed in order.
832 	 *
833 	 * ps->current_area does not get reduced by prepare_merge() until
834 	 * after commit_merge() has removed the nr_merged previous exceptions.
835 	 */
836 	ps->next_free = area_location(ps, ps->current_area) +
837 			ps->current_committed + 1;
838 
839 	return 0;
840 }
841 
842 static void persistent_drop_snapshot(struct dm_exception_store *store)
843 {
844 	struct pstore *ps = get_info(store);
845 
846 	ps->valid = 0;
847 	if (write_header(ps))
848 		DMWARN("write header failed");
849 }
850 
851 static int persistent_ctr(struct dm_exception_store *store, char *options)
852 {
853 	struct pstore *ps;
854 	int r;
855 
856 	/* allocate the pstore */
857 	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
858 	if (!ps)
859 		return -ENOMEM;
860 
861 	ps->store = store;
862 	ps->valid = 1;
863 	ps->version = SNAPSHOT_DISK_VERSION;
864 	ps->area = NULL;
865 	ps->zero_area = NULL;
866 	ps->header_area = NULL;
867 	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
868 	ps->current_committed = 0;
869 
870 	ps->callback_count = 0;
871 	atomic_set(&ps->pending_count, 0);
872 	ps->callbacks = NULL;
873 
874 	ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
875 	if (!ps->metadata_wq) {
876 		DMERR("couldn't start header metadata update thread");
877 		r = -ENOMEM;
878 		goto err_workqueue;
879 	}
880 
881 	if (options) {
882 		char overflow = toupper(options[0]);
883 
884 		if (overflow == 'O')
885 			store->userspace_supports_overflow = true;
886 		else {
887 			DMERR("Unsupported persistent store option: %s", options);
888 			r = -EINVAL;
889 			goto err_options;
890 		}
891 	}
892 
893 	store->context = ps;
894 
895 	return 0;
896 
897 err_options:
898 	destroy_workqueue(ps->metadata_wq);
899 err_workqueue:
900 	kfree(ps);
901 
902 	return r;
903 }
904 
905 static unsigned int persistent_status(struct dm_exception_store *store,
906 				  status_type_t status, char *result,
907 				  unsigned int maxlen)
908 {
909 	unsigned int sz = 0;
910 
911 	switch (status) {
912 	case STATUSTYPE_INFO:
913 		break;
914 	case STATUSTYPE_TABLE:
915 		DMEMIT(" %s %llu", store->userspace_supports_overflow ? "PO" : "P",
916 		       (unsigned long long)store->chunk_size);
917 		break;
918 	case STATUSTYPE_IMA:
919 		*result = '\0';
920 		break;
921 	}
922 
923 	return sz;
924 }
925 
926 static struct dm_exception_store_type _persistent_type = {
927 	.name = "persistent",
928 	.module = THIS_MODULE,
929 	.ctr = persistent_ctr,
930 	.dtr = persistent_dtr,
931 	.read_metadata = persistent_read_metadata,
932 	.prepare_exception = persistent_prepare_exception,
933 	.commit_exception = persistent_commit_exception,
934 	.prepare_merge = persistent_prepare_merge,
935 	.commit_merge = persistent_commit_merge,
936 	.drop_snapshot = persistent_drop_snapshot,
937 	.usage = persistent_usage,
938 	.status = persistent_status,
939 };
940 
941 static struct dm_exception_store_type _persistent_compat_type = {
942 	.name = "P",
943 	.module = THIS_MODULE,
944 	.ctr = persistent_ctr,
945 	.dtr = persistent_dtr,
946 	.read_metadata = persistent_read_metadata,
947 	.prepare_exception = persistent_prepare_exception,
948 	.commit_exception = persistent_commit_exception,
949 	.prepare_merge = persistent_prepare_merge,
950 	.commit_merge = persistent_commit_merge,
951 	.drop_snapshot = persistent_drop_snapshot,
952 	.usage = persistent_usage,
953 	.status = persistent_status,
954 };
955 
956 int dm_persistent_snapshot_init(void)
957 {
958 	int r;
959 
960 	r = dm_exception_store_type_register(&_persistent_type);
961 	if (r) {
962 		DMERR("Unable to register persistent exception store type");
963 		return r;
964 	}
965 
966 	r = dm_exception_store_type_register(&_persistent_compat_type);
967 	if (r) {
968 		DMERR("Unable to register old-style persistent exception store type");
969 		dm_exception_store_type_unregister(&_persistent_type);
970 		return r;
971 	}
972 
973 	return r;
974 }
975 
976 void dm_persistent_snapshot_exit(void)
977 {
978 	dm_exception_store_type_unregister(&_persistent_type);
979 	dm_exception_store_type_unregister(&_persistent_compat_type);
980 }
981