xref: /linux/drivers/md/dm-snap-persistent.c (revision ca55b2fef3a9373fcfc30f82fd26bc7fccbda732)
1 /*
2  * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
3  * Copyright (C) 2006-2008 Red Hat GmbH
4  *
5  * This file is released under the GPL.
6  */
7 
8 #include "dm-exception-store.h"
9 
10 #include <linux/ctype.h>
11 #include <linux/mm.h>
12 #include <linux/pagemap.h>
13 #include <linux/vmalloc.h>
14 #include <linux/export.h>
15 #include <linux/slab.h>
16 #include <linux/dm-io.h>
17 #include "dm-bufio.h"
18 
19 #define DM_MSG_PREFIX "persistent snapshot"
20 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32	/* 16KB */
21 
22 #define DM_PREFETCH_CHUNKS		12
23 
24 /*-----------------------------------------------------------------
25  * Persistent snapshots, by persistent we mean that the snapshot
26  * will survive a reboot.
27  *---------------------------------------------------------------*/
28 
29 /*
30  * We need to store a record of which parts of the origin have
31  * been copied to the snapshot device.  The snapshot code
32  * requires that we copy exception chunks to chunk aligned areas
33  * of the COW store.  It makes sense therefore, to store the
34  * metadata in chunk size blocks.
35  *
36  * There is no backward or forward compatibility implemented,
37  * snapshots with different disk versions than the kernel will
38  * not be usable.  It is expected that "lvcreate" will blank out
39  * the start of a fresh COW device before calling the snapshot
40  * constructor.
41  *
42  * The first chunk of the COW device just contains the header.
43  * After this there is a chunk filled with exception metadata,
44  * followed by as many exception chunks as can fit in the
45  * metadata areas.
46  *
47  * All on disk structures are in little-endian format.  The end
48  * of the exceptions info is indicated by an exception with a
49  * new_chunk of 0, which is invalid since it would point to the
50  * header chunk.
51  */
52 
53 /*
54  * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
55  */
56 #define SNAP_MAGIC 0x70416e53
57 
58 /*
59  * The on-disk version of the metadata.
60  */
61 #define SNAPSHOT_DISK_VERSION 1
62 
63 #define NUM_SNAPSHOT_HDR_CHUNKS 1
64 
65 struct disk_header {
66 	__le32 magic;
67 
68 	/*
69 	 * Is this snapshot valid.  There is no way of recovering
70 	 * an invalid snapshot.
71 	 */
72 	__le32 valid;
73 
74 	/*
75 	 * Simple, incrementing version. no backward
76 	 * compatibility.
77 	 */
78 	__le32 version;
79 
80 	/* In sectors */
81 	__le32 chunk_size;
82 } __packed;
83 
84 struct disk_exception {
85 	__le64 old_chunk;
86 	__le64 new_chunk;
87 } __packed;
88 
89 struct core_exception {
90 	uint64_t old_chunk;
91 	uint64_t new_chunk;
92 };
93 
94 struct commit_callback {
95 	void (*callback)(void *, int success);
96 	void *context;
97 };
98 
99 /*
100  * The top level structure for a persistent exception store.
101  */
102 struct pstore {
103 	struct dm_exception_store *store;
104 	int version;
105 	int valid;
106 	uint32_t exceptions_per_area;
107 
108 	/*
109 	 * Now that we have an asynchronous kcopyd there is no
110 	 * need for large chunk sizes, so it wont hurt to have a
111 	 * whole chunks worth of metadata in memory at once.
112 	 */
113 	void *area;
114 
115 	/*
116 	 * An area of zeros used to clear the next area.
117 	 */
118 	void *zero_area;
119 
120 	/*
121 	 * An area used for header. The header can be written
122 	 * concurrently with metadata (when invalidating the snapshot),
123 	 * so it needs a separate buffer.
124 	 */
125 	void *header_area;
126 
127 	/*
128 	 * Used to keep track of which metadata area the data in
129 	 * 'chunk' refers to.
130 	 */
131 	chunk_t current_area;
132 
133 	/*
134 	 * The next free chunk for an exception.
135 	 *
136 	 * When creating exceptions, all the chunks here and above are
137 	 * free.  It holds the next chunk to be allocated.  On rare
138 	 * occasions (e.g. after a system crash) holes can be left in
139 	 * the exception store because chunks can be committed out of
140 	 * order.
141 	 *
142 	 * When merging exceptions, it does not necessarily mean all the
143 	 * chunks here and above are free.  It holds the value it would
144 	 * have held if all chunks had been committed in order of
145 	 * allocation.  Consequently the value may occasionally be
146 	 * slightly too low, but since it's only used for 'status' and
147 	 * it can never reach its minimum value too early this doesn't
148 	 * matter.
149 	 */
150 
151 	chunk_t next_free;
152 
153 	/*
154 	 * The index of next free exception in the current
155 	 * metadata area.
156 	 */
157 	uint32_t current_committed;
158 
159 	atomic_t pending_count;
160 	uint32_t callback_count;
161 	struct commit_callback *callbacks;
162 	struct dm_io_client *io_client;
163 
164 	struct workqueue_struct *metadata_wq;
165 };
166 
167 static int alloc_area(struct pstore *ps)
168 {
169 	int r = -ENOMEM;
170 	size_t len;
171 
172 	len = ps->store->chunk_size << SECTOR_SHIFT;
173 
174 	/*
175 	 * Allocate the chunk_size block of memory that will hold
176 	 * a single metadata area.
177 	 */
178 	ps->area = vmalloc(len);
179 	if (!ps->area)
180 		goto err_area;
181 
182 	ps->zero_area = vzalloc(len);
183 	if (!ps->zero_area)
184 		goto err_zero_area;
185 
186 	ps->header_area = vmalloc(len);
187 	if (!ps->header_area)
188 		goto err_header_area;
189 
190 	return 0;
191 
192 err_header_area:
193 	vfree(ps->zero_area);
194 
195 err_zero_area:
196 	vfree(ps->area);
197 
198 err_area:
199 	return r;
200 }
201 
202 static void free_area(struct pstore *ps)
203 {
204 	vfree(ps->area);
205 	ps->area = NULL;
206 	vfree(ps->zero_area);
207 	ps->zero_area = NULL;
208 	vfree(ps->header_area);
209 	ps->header_area = NULL;
210 }
211 
212 struct mdata_req {
213 	struct dm_io_region *where;
214 	struct dm_io_request *io_req;
215 	struct work_struct work;
216 	int result;
217 };
218 
219 static void do_metadata(struct work_struct *work)
220 {
221 	struct mdata_req *req = container_of(work, struct mdata_req, work);
222 
223 	req->result = dm_io(req->io_req, 1, req->where, NULL);
224 }
225 
226 /*
227  * Read or write a chunk aligned and sized block of data from a device.
228  */
229 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, int rw,
230 		    int metadata)
231 {
232 	struct dm_io_region where = {
233 		.bdev = dm_snap_cow(ps->store->snap)->bdev,
234 		.sector = ps->store->chunk_size * chunk,
235 		.count = ps->store->chunk_size,
236 	};
237 	struct dm_io_request io_req = {
238 		.bi_rw = rw,
239 		.mem.type = DM_IO_VMA,
240 		.mem.ptr.vma = area,
241 		.client = ps->io_client,
242 		.notify.fn = NULL,
243 	};
244 	struct mdata_req req;
245 
246 	if (!metadata)
247 		return dm_io(&io_req, 1, &where, NULL);
248 
249 	req.where = &where;
250 	req.io_req = &io_req;
251 
252 	/*
253 	 * Issue the synchronous I/O from a different thread
254 	 * to avoid generic_make_request recursion.
255 	 */
256 	INIT_WORK_ONSTACK(&req.work, do_metadata);
257 	queue_work(ps->metadata_wq, &req.work);
258 	flush_workqueue(ps->metadata_wq);
259 	destroy_work_on_stack(&req.work);
260 
261 	return req.result;
262 }
263 
264 /*
265  * Convert a metadata area index to a chunk index.
266  */
267 static chunk_t area_location(struct pstore *ps, chunk_t area)
268 {
269 	return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
270 }
271 
272 static void skip_metadata(struct pstore *ps)
273 {
274 	uint32_t stride = ps->exceptions_per_area + 1;
275 	chunk_t next_free = ps->next_free;
276 	if (sector_div(next_free, stride) == NUM_SNAPSHOT_HDR_CHUNKS)
277 		ps->next_free++;
278 }
279 
280 /*
281  * Read or write a metadata area.  Remembering to skip the first
282  * chunk which holds the header.
283  */
284 static int area_io(struct pstore *ps, int rw)
285 {
286 	int r;
287 	chunk_t chunk;
288 
289 	chunk = area_location(ps, ps->current_area);
290 
291 	r = chunk_io(ps, ps->area, chunk, rw, 0);
292 	if (r)
293 		return r;
294 
295 	return 0;
296 }
297 
298 static void zero_memory_area(struct pstore *ps)
299 {
300 	memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
301 }
302 
303 static int zero_disk_area(struct pstore *ps, chunk_t area)
304 {
305 	return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
306 }
307 
308 static int read_header(struct pstore *ps, int *new_snapshot)
309 {
310 	int r;
311 	struct disk_header *dh;
312 	unsigned chunk_size;
313 	int chunk_size_supplied = 1;
314 	char *chunk_err;
315 
316 	/*
317 	 * Use default chunk size (or logical_block_size, if larger)
318 	 * if none supplied
319 	 */
320 	if (!ps->store->chunk_size) {
321 		ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
322 		    bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
323 					    bdev) >> 9);
324 		ps->store->chunk_mask = ps->store->chunk_size - 1;
325 		ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
326 		chunk_size_supplied = 0;
327 	}
328 
329 	ps->io_client = dm_io_client_create();
330 	if (IS_ERR(ps->io_client))
331 		return PTR_ERR(ps->io_client);
332 
333 	r = alloc_area(ps);
334 	if (r)
335 		return r;
336 
337 	r = chunk_io(ps, ps->header_area, 0, READ, 1);
338 	if (r)
339 		goto bad;
340 
341 	dh = ps->header_area;
342 
343 	if (le32_to_cpu(dh->magic) == 0) {
344 		*new_snapshot = 1;
345 		return 0;
346 	}
347 
348 	if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
349 		DMWARN("Invalid or corrupt snapshot");
350 		r = -ENXIO;
351 		goto bad;
352 	}
353 
354 	*new_snapshot = 0;
355 	ps->valid = le32_to_cpu(dh->valid);
356 	ps->version = le32_to_cpu(dh->version);
357 	chunk_size = le32_to_cpu(dh->chunk_size);
358 
359 	if (ps->store->chunk_size == chunk_size)
360 		return 0;
361 
362 	if (chunk_size_supplied)
363 		DMWARN("chunk size %u in device metadata overrides "
364 		       "table chunk size of %u.",
365 		       chunk_size, ps->store->chunk_size);
366 
367 	/* We had a bogus chunk_size. Fix stuff up. */
368 	free_area(ps);
369 
370 	r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
371 					      &chunk_err);
372 	if (r) {
373 		DMERR("invalid on-disk chunk size %u: %s.",
374 		      chunk_size, chunk_err);
375 		return r;
376 	}
377 
378 	r = alloc_area(ps);
379 	return r;
380 
381 bad:
382 	free_area(ps);
383 	return r;
384 }
385 
386 static int write_header(struct pstore *ps)
387 {
388 	struct disk_header *dh;
389 
390 	memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
391 
392 	dh = ps->header_area;
393 	dh->magic = cpu_to_le32(SNAP_MAGIC);
394 	dh->valid = cpu_to_le32(ps->valid);
395 	dh->version = cpu_to_le32(ps->version);
396 	dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
397 
398 	return chunk_io(ps, ps->header_area, 0, WRITE, 1);
399 }
400 
401 /*
402  * Access functions for the disk exceptions, these do the endian conversions.
403  */
404 static struct disk_exception *get_exception(struct pstore *ps, void *ps_area,
405 					    uint32_t index)
406 {
407 	BUG_ON(index >= ps->exceptions_per_area);
408 
409 	return ((struct disk_exception *) ps_area) + index;
410 }
411 
412 static void read_exception(struct pstore *ps, void *ps_area,
413 			   uint32_t index, struct core_exception *result)
414 {
415 	struct disk_exception *de = get_exception(ps, ps_area, index);
416 
417 	/* copy it */
418 	result->old_chunk = le64_to_cpu(de->old_chunk);
419 	result->new_chunk = le64_to_cpu(de->new_chunk);
420 }
421 
422 static void write_exception(struct pstore *ps,
423 			    uint32_t index, struct core_exception *e)
424 {
425 	struct disk_exception *de = get_exception(ps, ps->area, index);
426 
427 	/* copy it */
428 	de->old_chunk = cpu_to_le64(e->old_chunk);
429 	de->new_chunk = cpu_to_le64(e->new_chunk);
430 }
431 
432 static void clear_exception(struct pstore *ps, uint32_t index)
433 {
434 	struct disk_exception *de = get_exception(ps, ps->area, index);
435 
436 	/* clear it */
437 	de->old_chunk = 0;
438 	de->new_chunk = 0;
439 }
440 
441 /*
442  * Registers the exceptions that are present in the current area.
443  * 'full' is filled in to indicate if the area has been
444  * filled.
445  */
446 static int insert_exceptions(struct pstore *ps, void *ps_area,
447 			     int (*callback)(void *callback_context,
448 					     chunk_t old, chunk_t new),
449 			     void *callback_context,
450 			     int *full)
451 {
452 	int r;
453 	unsigned int i;
454 	struct core_exception e;
455 
456 	/* presume the area is full */
457 	*full = 1;
458 
459 	for (i = 0; i < ps->exceptions_per_area; i++) {
460 		read_exception(ps, ps_area, i, &e);
461 
462 		/*
463 		 * If the new_chunk is pointing at the start of
464 		 * the COW device, where the first metadata area
465 		 * is we know that we've hit the end of the
466 		 * exceptions.  Therefore the area is not full.
467 		 */
468 		if (e.new_chunk == 0LL) {
469 			ps->current_committed = i;
470 			*full = 0;
471 			break;
472 		}
473 
474 		/*
475 		 * Keep track of the start of the free chunks.
476 		 */
477 		if (ps->next_free <= e.new_chunk)
478 			ps->next_free = e.new_chunk + 1;
479 
480 		/*
481 		 * Otherwise we add the exception to the snapshot.
482 		 */
483 		r = callback(callback_context, e.old_chunk, e.new_chunk);
484 		if (r)
485 			return r;
486 	}
487 
488 	return 0;
489 }
490 
491 static int read_exceptions(struct pstore *ps,
492 			   int (*callback)(void *callback_context, chunk_t old,
493 					   chunk_t new),
494 			   void *callback_context)
495 {
496 	int r, full = 1;
497 	struct dm_bufio_client *client;
498 	chunk_t prefetch_area = 0;
499 
500 	client = dm_bufio_client_create(dm_snap_cow(ps->store->snap)->bdev,
501 					ps->store->chunk_size << SECTOR_SHIFT,
502 					1, 0, NULL, NULL);
503 
504 	if (IS_ERR(client))
505 		return PTR_ERR(client);
506 
507 	/*
508 	 * Setup for one current buffer + desired readahead buffers.
509 	 */
510 	dm_bufio_set_minimum_buffers(client, 1 + DM_PREFETCH_CHUNKS);
511 
512 	/*
513 	 * Keeping reading chunks and inserting exceptions until
514 	 * we find a partially full area.
515 	 */
516 	for (ps->current_area = 0; full; ps->current_area++) {
517 		struct dm_buffer *bp;
518 		void *area;
519 		chunk_t chunk;
520 
521 		if (unlikely(prefetch_area < ps->current_area))
522 			prefetch_area = ps->current_area;
523 
524 		if (DM_PREFETCH_CHUNKS) do {
525 			chunk_t pf_chunk = area_location(ps, prefetch_area);
526 			if (unlikely(pf_chunk >= dm_bufio_get_device_size(client)))
527 				break;
528 			dm_bufio_prefetch(client, pf_chunk, 1);
529 			prefetch_area++;
530 			if (unlikely(!prefetch_area))
531 				break;
532 		} while (prefetch_area <= ps->current_area + DM_PREFETCH_CHUNKS);
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 (struct pstore *) 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 	vfree(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, uninitialized_var(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 = dm_vcalloc(ps->exceptions_per_area,
630 				   sizeof(*ps->callbacks));
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,
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 	ce.old_chunk = e->old_chunk;
708 	ce.new_chunk = e->new_chunk;
709 	write_exception(ps, ps->current_committed++, &ce);
710 
711 	/*
712 	 * Add the callback to the back of the array.  This code
713 	 * is the only place where the callback array is
714 	 * manipulated, and we know that it will never be called
715 	 * multiple times concurrently.
716 	 */
717 	cb = ps->callbacks + ps->callback_count++;
718 	cb->callback = callback;
719 	cb->context = callback_context;
720 
721 	/*
722 	 * If there are exceptions in flight and we have not yet
723 	 * filled this metadata area there's nothing more to do.
724 	 */
725 	if (!atomic_dec_and_test(&ps->pending_count) &&
726 	    (ps->current_committed != ps->exceptions_per_area))
727 		return;
728 
729 	/*
730 	 * If we completely filled the current area, then wipe the next one.
731 	 */
732 	if ((ps->current_committed == ps->exceptions_per_area) &&
733 	    zero_disk_area(ps, ps->current_area + 1))
734 		ps->valid = 0;
735 
736 	/*
737 	 * Commit exceptions to disk.
738 	 */
739 	if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
740 		ps->valid = 0;
741 
742 	/*
743 	 * Advance to the next area if this one is full.
744 	 */
745 	if (ps->current_committed == ps->exceptions_per_area) {
746 		ps->current_committed = 0;
747 		ps->current_area++;
748 		zero_memory_area(ps);
749 	}
750 
751 	for (i = 0; i < ps->callback_count; i++) {
752 		cb = ps->callbacks + i;
753 		cb->callback(cb->context, ps->valid);
754 	}
755 
756 	ps->callback_count = 0;
757 }
758 
759 static int persistent_prepare_merge(struct dm_exception_store *store,
760 				    chunk_t *last_old_chunk,
761 				    chunk_t *last_new_chunk)
762 {
763 	struct pstore *ps = get_info(store);
764 	struct core_exception ce;
765 	int nr_consecutive;
766 	int r;
767 
768 	/*
769 	 * When current area is empty, move back to preceding area.
770 	 */
771 	if (!ps->current_committed) {
772 		/*
773 		 * Have we finished?
774 		 */
775 		if (!ps->current_area)
776 			return 0;
777 
778 		ps->current_area--;
779 		r = area_io(ps, READ);
780 		if (r < 0)
781 			return r;
782 		ps->current_committed = ps->exceptions_per_area;
783 	}
784 
785 	read_exception(ps, ps->area, ps->current_committed - 1, &ce);
786 	*last_old_chunk = ce.old_chunk;
787 	*last_new_chunk = ce.new_chunk;
788 
789 	/*
790 	 * Find number of consecutive chunks within the current area,
791 	 * working backwards.
792 	 */
793 	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
794 	     nr_consecutive++) {
795 		read_exception(ps, ps->area,
796 			       ps->current_committed - 1 - nr_consecutive, &ce);
797 		if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
798 		    ce.new_chunk != *last_new_chunk - nr_consecutive)
799 			break;
800 	}
801 
802 	return nr_consecutive;
803 }
804 
805 static int persistent_commit_merge(struct dm_exception_store *store,
806 				   int nr_merged)
807 {
808 	int r, i;
809 	struct pstore *ps = get_info(store);
810 
811 	BUG_ON(nr_merged > ps->current_committed);
812 
813 	for (i = 0; i < nr_merged; i++)
814 		clear_exception(ps, ps->current_committed - 1 - i);
815 
816 	r = area_io(ps, WRITE_FLUSH_FUA);
817 	if (r < 0)
818 		return r;
819 
820 	ps->current_committed -= nr_merged;
821 
822 	/*
823 	 * At this stage, only persistent_usage() uses ps->next_free, so
824 	 * we make no attempt to keep ps->next_free strictly accurate
825 	 * as exceptions may have been committed out-of-order originally.
826 	 * Once a snapshot has become merging, we set it to the value it
827 	 * would have held had all the exceptions been committed in order.
828 	 *
829 	 * ps->current_area does not get reduced by prepare_merge() until
830 	 * after commit_merge() has removed the nr_merged previous exceptions.
831 	 */
832 	ps->next_free = area_location(ps, ps->current_area) +
833 			ps->current_committed + 1;
834 
835 	return 0;
836 }
837 
838 static void persistent_drop_snapshot(struct dm_exception_store *store)
839 {
840 	struct pstore *ps = get_info(store);
841 
842 	ps->valid = 0;
843 	if (write_header(ps))
844 		DMWARN("write header failed");
845 }
846 
847 static int persistent_ctr(struct dm_exception_store *store, char *options)
848 {
849 	struct pstore *ps;
850 	int r;
851 
852 	/* allocate the pstore */
853 	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
854 	if (!ps)
855 		return -ENOMEM;
856 
857 	ps->store = store;
858 	ps->valid = 1;
859 	ps->version = SNAPSHOT_DISK_VERSION;
860 	ps->area = NULL;
861 	ps->zero_area = NULL;
862 	ps->header_area = NULL;
863 	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
864 	ps->current_committed = 0;
865 
866 	ps->callback_count = 0;
867 	atomic_set(&ps->pending_count, 0);
868 	ps->callbacks = NULL;
869 
870 	ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
871 	if (!ps->metadata_wq) {
872 		DMERR("couldn't start header metadata update thread");
873 		r = -ENOMEM;
874 		goto err_workqueue;
875 	}
876 
877 	if (options) {
878 		char overflow = toupper(options[0]);
879 		if (overflow == 'O')
880 			store->userspace_supports_overflow = true;
881 		else {
882 			DMERR("Unsupported persistent store option: %s", options);
883 			r = -EINVAL;
884 			goto err_options;
885 		}
886 	}
887 
888 	store->context = ps;
889 
890 	return 0;
891 
892 err_options:
893 	destroy_workqueue(ps->metadata_wq);
894 err_workqueue:
895 	kfree(ps);
896 
897 	return r;
898 }
899 
900 static unsigned persistent_status(struct dm_exception_store *store,
901 				  status_type_t status, char *result,
902 				  unsigned maxlen)
903 {
904 	unsigned sz = 0;
905 
906 	switch (status) {
907 	case STATUSTYPE_INFO:
908 		break;
909 	case STATUSTYPE_TABLE:
910 		DMEMIT(" %s %llu", store->userspace_supports_overflow ? "PO" : "P",
911 		       (unsigned long long)store->chunk_size);
912 	}
913 
914 	return sz;
915 }
916 
917 static struct dm_exception_store_type _persistent_type = {
918 	.name = "persistent",
919 	.module = THIS_MODULE,
920 	.ctr = persistent_ctr,
921 	.dtr = persistent_dtr,
922 	.read_metadata = persistent_read_metadata,
923 	.prepare_exception = persistent_prepare_exception,
924 	.commit_exception = persistent_commit_exception,
925 	.prepare_merge = persistent_prepare_merge,
926 	.commit_merge = persistent_commit_merge,
927 	.drop_snapshot = persistent_drop_snapshot,
928 	.usage = persistent_usage,
929 	.status = persistent_status,
930 };
931 
932 static struct dm_exception_store_type _persistent_compat_type = {
933 	.name = "P",
934 	.module = THIS_MODULE,
935 	.ctr = persistent_ctr,
936 	.dtr = persistent_dtr,
937 	.read_metadata = persistent_read_metadata,
938 	.prepare_exception = persistent_prepare_exception,
939 	.commit_exception = persistent_commit_exception,
940 	.prepare_merge = persistent_prepare_merge,
941 	.commit_merge = persistent_commit_merge,
942 	.drop_snapshot = persistent_drop_snapshot,
943 	.usage = persistent_usage,
944 	.status = persistent_status,
945 };
946 
947 int dm_persistent_snapshot_init(void)
948 {
949 	int r;
950 
951 	r = dm_exception_store_type_register(&_persistent_type);
952 	if (r) {
953 		DMERR("Unable to register persistent exception store type");
954 		return r;
955 	}
956 
957 	r = dm_exception_store_type_register(&_persistent_compat_type);
958 	if (r) {
959 		DMERR("Unable to register old-style persistent exception "
960 		      "store type");
961 		dm_exception_store_type_unregister(&_persistent_type);
962 		return r;
963 	}
964 
965 	return r;
966 }
967 
968 void dm_persistent_snapshot_exit(void)
969 {
970 	dm_exception_store_type_unregister(&_persistent_type);
971 	dm_exception_store_type_unregister(&_persistent_compat_type);
972 }
973