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 <linux/dm-bufio.h>
18 
19 #define DM_MSG_PREFIX "persistent snapshot"
20 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32U	/* 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 op,
230 		    int op_flags, 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_op = op,
239 		.bi_op_flags = op_flags,
240 		.mem.type = DM_IO_VMA,
241 		.mem.ptr.vma = area,
242 		.client = ps->io_client,
243 		.notify.fn = NULL,
244 	};
245 	struct mdata_req req;
246 
247 	if (!metadata)
248 		return dm_io(&io_req, 1, &where, NULL);
249 
250 	req.where = &where;
251 	req.io_req = &io_req;
252 
253 	/*
254 	 * Issue the synchronous I/O from a different thread
255 	 * to avoid submit_bio_noacct recursion.
256 	 */
257 	INIT_WORK_ONSTACK(&req.work, do_metadata);
258 	queue_work(ps->metadata_wq, &req.work);
259 	flush_workqueue(ps->metadata_wq);
260 	destroy_work_on_stack(&req.work);
261 
262 	return req.result;
263 }
264 
265 /*
266  * Convert a metadata area index to a chunk index.
267  */
268 static chunk_t area_location(struct pstore *ps, chunk_t area)
269 {
270 	return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
271 }
272 
273 static void skip_metadata(struct pstore *ps)
274 {
275 	uint32_t stride = ps->exceptions_per_area + 1;
276 	chunk_t next_free = ps->next_free;
277 	if (sector_div(next_free, stride) == NUM_SNAPSHOT_HDR_CHUNKS)
278 		ps->next_free++;
279 }
280 
281 /*
282  * Read or write a metadata area.  Remembering to skip the first
283  * chunk which holds the header.
284  */
285 static int area_io(struct pstore *ps, int op, int op_flags)
286 {
287 	chunk_t chunk = area_location(ps, ps->current_area);
288 
289 	return chunk_io(ps, ps->area, chunk, op, op_flags, 0);
290 }
291 
292 static void zero_memory_area(struct pstore *ps)
293 {
294 	memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
295 }
296 
297 static int zero_disk_area(struct pstore *ps, chunk_t area)
298 {
299 	return chunk_io(ps, ps->zero_area, area_location(ps, area),
300 			REQ_OP_WRITE, 0, 0);
301 }
302 
303 static int read_header(struct pstore *ps, int *new_snapshot)
304 {
305 	int r;
306 	struct disk_header *dh;
307 	unsigned chunk_size;
308 	int chunk_size_supplied = 1;
309 	char *chunk_err;
310 
311 	/*
312 	 * Use default chunk size (or logical_block_size, if larger)
313 	 * if none supplied
314 	 */
315 	if (!ps->store->chunk_size) {
316 		ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
317 		    bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
318 					    bdev) >> 9);
319 		ps->store->chunk_mask = ps->store->chunk_size - 1;
320 		ps->store->chunk_shift = __ffs(ps->store->chunk_size);
321 		chunk_size_supplied = 0;
322 	}
323 
324 	ps->io_client = dm_io_client_create();
325 	if (IS_ERR(ps->io_client))
326 		return PTR_ERR(ps->io_client);
327 
328 	r = alloc_area(ps);
329 	if (r)
330 		return r;
331 
332 	r = chunk_io(ps, ps->header_area, 0, REQ_OP_READ, 0, 1);
333 	if (r)
334 		goto bad;
335 
336 	dh = ps->header_area;
337 
338 	if (le32_to_cpu(dh->magic) == 0) {
339 		*new_snapshot = 1;
340 		return 0;
341 	}
342 
343 	if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
344 		DMWARN("Invalid or corrupt snapshot");
345 		r = -ENXIO;
346 		goto bad;
347 	}
348 
349 	*new_snapshot = 0;
350 	ps->valid = le32_to_cpu(dh->valid);
351 	ps->version = le32_to_cpu(dh->version);
352 	chunk_size = le32_to_cpu(dh->chunk_size);
353 
354 	if (ps->store->chunk_size == chunk_size)
355 		return 0;
356 
357 	if (chunk_size_supplied)
358 		DMWARN("chunk size %u in device metadata overrides "
359 		       "table chunk size of %u.",
360 		       chunk_size, ps->store->chunk_size);
361 
362 	/* We had a bogus chunk_size. Fix stuff up. */
363 	free_area(ps);
364 
365 	r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
366 					      &chunk_err);
367 	if (r) {
368 		DMERR("invalid on-disk chunk size %u: %s.",
369 		      chunk_size, chunk_err);
370 		return r;
371 	}
372 
373 	r = alloc_area(ps);
374 	return r;
375 
376 bad:
377 	free_area(ps);
378 	return r;
379 }
380 
381 static int write_header(struct pstore *ps)
382 {
383 	struct disk_header *dh;
384 
385 	memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
386 
387 	dh = ps->header_area;
388 	dh->magic = cpu_to_le32(SNAP_MAGIC);
389 	dh->valid = cpu_to_le32(ps->valid);
390 	dh->version = cpu_to_le32(ps->version);
391 	dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
392 
393 	return chunk_io(ps, ps->header_area, 0, REQ_OP_WRITE, 0, 1);
394 }
395 
396 /*
397  * Access functions for the disk exceptions, these do the endian conversions.
398  */
399 static struct disk_exception *get_exception(struct pstore *ps, void *ps_area,
400 					    uint32_t index)
401 {
402 	BUG_ON(index >= ps->exceptions_per_area);
403 
404 	return ((struct disk_exception *) ps_area) + index;
405 }
406 
407 static void read_exception(struct pstore *ps, void *ps_area,
408 			   uint32_t index, struct core_exception *result)
409 {
410 	struct disk_exception *de = get_exception(ps, ps_area, index);
411 
412 	/* copy it */
413 	result->old_chunk = le64_to_cpu(de->old_chunk);
414 	result->new_chunk = le64_to_cpu(de->new_chunk);
415 }
416 
417 static void write_exception(struct pstore *ps,
418 			    uint32_t index, struct core_exception *e)
419 {
420 	struct disk_exception *de = get_exception(ps, ps->area, index);
421 
422 	/* copy it */
423 	de->old_chunk = cpu_to_le64(e->old_chunk);
424 	de->new_chunk = cpu_to_le64(e->new_chunk);
425 }
426 
427 static void clear_exception(struct pstore *ps, uint32_t index)
428 {
429 	struct disk_exception *de = get_exception(ps, ps->area, index);
430 
431 	/* clear it */
432 	de->old_chunk = 0;
433 	de->new_chunk = 0;
434 }
435 
436 /*
437  * Registers the exceptions that are present in the current area.
438  * 'full' is filled in to indicate if the area has been
439  * filled.
440  */
441 static int insert_exceptions(struct pstore *ps, void *ps_area,
442 			     int (*callback)(void *callback_context,
443 					     chunk_t old, chunk_t new),
444 			     void *callback_context,
445 			     int *full)
446 {
447 	int r;
448 	unsigned int i;
449 	struct core_exception e;
450 
451 	/* presume the area is full */
452 	*full = 1;
453 
454 	for (i = 0; i < ps->exceptions_per_area; i++) {
455 		read_exception(ps, ps_area, i, &e);
456 
457 		/*
458 		 * If the new_chunk is pointing at the start of
459 		 * the COW device, where the first metadata area
460 		 * is we know that we've hit the end of the
461 		 * exceptions.  Therefore the area is not full.
462 		 */
463 		if (e.new_chunk == 0LL) {
464 			ps->current_committed = i;
465 			*full = 0;
466 			break;
467 		}
468 
469 		/*
470 		 * Keep track of the start of the free chunks.
471 		 */
472 		if (ps->next_free <= e.new_chunk)
473 			ps->next_free = e.new_chunk + 1;
474 
475 		/*
476 		 * Otherwise we add the exception to the snapshot.
477 		 */
478 		r = callback(callback_context, e.old_chunk, e.new_chunk);
479 		if (r)
480 			return r;
481 	}
482 
483 	return 0;
484 }
485 
486 static int read_exceptions(struct pstore *ps,
487 			   int (*callback)(void *callback_context, chunk_t old,
488 					   chunk_t new),
489 			   void *callback_context)
490 {
491 	int r, full = 1;
492 	struct dm_bufio_client *client;
493 	chunk_t prefetch_area = 0;
494 
495 	client = dm_bufio_client_create(dm_snap_cow(ps->store->snap)->bdev,
496 					ps->store->chunk_size << SECTOR_SHIFT,
497 					1, 0, NULL, NULL);
498 
499 	if (IS_ERR(client))
500 		return PTR_ERR(client);
501 
502 	/*
503 	 * Setup for one current buffer + desired readahead buffers.
504 	 */
505 	dm_bufio_set_minimum_buffers(client, 1 + DM_PREFETCH_CHUNKS);
506 
507 	/*
508 	 * Keeping reading chunks and inserting exceptions until
509 	 * we find a partially full area.
510 	 */
511 	for (ps->current_area = 0; full; ps->current_area++) {
512 		struct dm_buffer *bp;
513 		void *area;
514 		chunk_t chunk;
515 
516 		if (unlikely(prefetch_area < ps->current_area))
517 			prefetch_area = ps->current_area;
518 
519 		if (DM_PREFETCH_CHUNKS) do {
520 			chunk_t pf_chunk = area_location(ps, prefetch_area);
521 			if (unlikely(pf_chunk >= dm_bufio_get_device_size(client)))
522 				break;
523 			dm_bufio_prefetch(client, pf_chunk, 1);
524 			prefetch_area++;
525 			if (unlikely(!prefetch_area))
526 				break;
527 		} while (prefetch_area <= ps->current_area + DM_PREFETCH_CHUNKS);
528 
529 		chunk = area_location(ps, ps->current_area);
530 
531 		area = dm_bufio_read(client, chunk, &bp);
532 		if (IS_ERR(area)) {
533 			r = PTR_ERR(area);
534 			goto ret_destroy_bufio;
535 		}
536 
537 		r = insert_exceptions(ps, area, callback, callback_context,
538 				      &full);
539 
540 		if (!full)
541 			memcpy(ps->area, area, ps->store->chunk_size << SECTOR_SHIFT);
542 
543 		dm_bufio_release(bp);
544 
545 		dm_bufio_forget(client, chunk);
546 
547 		if (unlikely(r))
548 			goto ret_destroy_bufio;
549 	}
550 
551 	ps->current_area--;
552 
553 	skip_metadata(ps);
554 
555 	r = 0;
556 
557 ret_destroy_bufio:
558 	dm_bufio_client_destroy(client);
559 
560 	return r;
561 }
562 
563 static struct pstore *get_info(struct dm_exception_store *store)
564 {
565 	return (struct pstore *) store->context;
566 }
567 
568 static void persistent_usage(struct dm_exception_store *store,
569 			     sector_t *total_sectors,
570 			     sector_t *sectors_allocated,
571 			     sector_t *metadata_sectors)
572 {
573 	struct pstore *ps = get_info(store);
574 
575 	*sectors_allocated = ps->next_free * store->chunk_size;
576 	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
577 
578 	/*
579 	 * First chunk is the fixed header.
580 	 * Then there are (ps->current_area + 1) metadata chunks, each one
581 	 * separated from the next by ps->exceptions_per_area data chunks.
582 	 */
583 	*metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
584 			    store->chunk_size;
585 }
586 
587 static void persistent_dtr(struct dm_exception_store *store)
588 {
589 	struct pstore *ps = get_info(store);
590 
591 	destroy_workqueue(ps->metadata_wq);
592 
593 	/* Created in read_header */
594 	if (ps->io_client)
595 		dm_io_client_destroy(ps->io_client);
596 	free_area(ps);
597 
598 	/* Allocated in persistent_read_metadata */
599 	kvfree(ps->callbacks);
600 
601 	kfree(ps);
602 }
603 
604 static int persistent_read_metadata(struct dm_exception_store *store,
605 				    int (*callback)(void *callback_context,
606 						    chunk_t old, chunk_t new),
607 				    void *callback_context)
608 {
609 	int r, new_snapshot;
610 	struct pstore *ps = get_info(store);
611 
612 	/*
613 	 * Read the snapshot header.
614 	 */
615 	r = read_header(ps, &new_snapshot);
616 	if (r)
617 		return r;
618 
619 	/*
620 	 * Now we know correct chunk_size, complete the initialisation.
621 	 */
622 	ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
623 				  sizeof(struct disk_exception);
624 	ps->callbacks = kvcalloc(ps->exceptions_per_area,
625 				 sizeof(*ps->callbacks), GFP_KERNEL);
626 	if (!ps->callbacks)
627 		return -ENOMEM;
628 
629 	/*
630 	 * Do we need to setup a new snapshot ?
631 	 */
632 	if (new_snapshot) {
633 		r = write_header(ps);
634 		if (r) {
635 			DMWARN("write_header failed");
636 			return r;
637 		}
638 
639 		ps->current_area = 0;
640 		zero_memory_area(ps);
641 		r = zero_disk_area(ps, 0);
642 		if (r)
643 			DMWARN("zero_disk_area(0) failed");
644 		return r;
645 	}
646 	/*
647 	 * Sanity checks.
648 	 */
649 	if (ps->version != SNAPSHOT_DISK_VERSION) {
650 		DMWARN("unable to handle snapshot disk version %d",
651 		       ps->version);
652 		return -EINVAL;
653 	}
654 
655 	/*
656 	 * Metadata are valid, but snapshot is invalidated
657 	 */
658 	if (!ps->valid)
659 		return 1;
660 
661 	/*
662 	 * Read the metadata.
663 	 */
664 	r = read_exceptions(ps, callback, callback_context);
665 
666 	return r;
667 }
668 
669 static int persistent_prepare_exception(struct dm_exception_store *store,
670 					struct dm_exception *e)
671 {
672 	struct pstore *ps = get_info(store);
673 	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
674 
675 	/* Is there enough room ? */
676 	if (size < ((ps->next_free + 1) * store->chunk_size))
677 		return -ENOSPC;
678 
679 	e->new_chunk = ps->next_free;
680 
681 	/*
682 	 * Move onto the next free pending, making sure to take
683 	 * into account the location of the metadata chunks.
684 	 */
685 	ps->next_free++;
686 	skip_metadata(ps);
687 
688 	atomic_inc(&ps->pending_count);
689 	return 0;
690 }
691 
692 static void persistent_commit_exception(struct dm_exception_store *store,
693 					struct dm_exception *e, int valid,
694 					void (*callback) (void *, int success),
695 					void *callback_context)
696 {
697 	unsigned int i;
698 	struct pstore *ps = get_info(store);
699 	struct core_exception ce;
700 	struct commit_callback *cb;
701 
702 	if (!valid)
703 		ps->valid = 0;
704 
705 	ce.old_chunk = e->old_chunk;
706 	ce.new_chunk = e->new_chunk;
707 	write_exception(ps, ps->current_committed++, &ce);
708 
709 	/*
710 	 * Add the callback to the back of the array.  This code
711 	 * is the only place where the callback array is
712 	 * manipulated, and we know that it will never be called
713 	 * multiple times concurrently.
714 	 */
715 	cb = ps->callbacks + ps->callback_count++;
716 	cb->callback = callback;
717 	cb->context = callback_context;
718 
719 	/*
720 	 * If there are exceptions in flight and we have not yet
721 	 * filled this metadata area there's nothing more to do.
722 	 */
723 	if (!atomic_dec_and_test(&ps->pending_count) &&
724 	    (ps->current_committed != ps->exceptions_per_area))
725 		return;
726 
727 	/*
728 	 * If we completely filled the current area, then wipe the next one.
729 	 */
730 	if ((ps->current_committed == ps->exceptions_per_area) &&
731 	    zero_disk_area(ps, ps->current_area + 1))
732 		ps->valid = 0;
733 
734 	/*
735 	 * Commit exceptions to disk.
736 	 */
737 	if (ps->valid && area_io(ps, REQ_OP_WRITE,
738 				 REQ_PREFLUSH | REQ_FUA | REQ_SYNC))
739 		ps->valid = 0;
740 
741 	/*
742 	 * Advance to the next area if this one is full.
743 	 */
744 	if (ps->current_committed == ps->exceptions_per_area) {
745 		ps->current_committed = 0;
746 		ps->current_area++;
747 		zero_memory_area(ps);
748 	}
749 
750 	for (i = 0; i < ps->callback_count; i++) {
751 		cb = ps->callbacks + i;
752 		cb->callback(cb->context, ps->valid);
753 	}
754 
755 	ps->callback_count = 0;
756 }
757 
758 static int persistent_prepare_merge(struct dm_exception_store *store,
759 				    chunk_t *last_old_chunk,
760 				    chunk_t *last_new_chunk)
761 {
762 	struct pstore *ps = get_info(store);
763 	struct core_exception ce;
764 	int nr_consecutive;
765 	int r;
766 
767 	/*
768 	 * When current area is empty, move back to preceding area.
769 	 */
770 	if (!ps->current_committed) {
771 		/*
772 		 * Have we finished?
773 		 */
774 		if (!ps->current_area)
775 			return 0;
776 
777 		ps->current_area--;
778 		r = area_io(ps, REQ_OP_READ, 0);
779 		if (r < 0)
780 			return r;
781 		ps->current_committed = ps->exceptions_per_area;
782 	}
783 
784 	read_exception(ps, ps->area, ps->current_committed - 1, &ce);
785 	*last_old_chunk = ce.old_chunk;
786 	*last_new_chunk = ce.new_chunk;
787 
788 	/*
789 	 * Find number of consecutive chunks within the current area,
790 	 * working backwards.
791 	 */
792 	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
793 	     nr_consecutive++) {
794 		read_exception(ps, ps->area,
795 			       ps->current_committed - 1 - nr_consecutive, &ce);
796 		if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
797 		    ce.new_chunk != *last_new_chunk - nr_consecutive)
798 			break;
799 	}
800 
801 	return nr_consecutive;
802 }
803 
804 static int persistent_commit_merge(struct dm_exception_store *store,
805 				   int nr_merged)
806 {
807 	int r, i;
808 	struct pstore *ps = get_info(store);
809 
810 	BUG_ON(nr_merged > ps->current_committed);
811 
812 	for (i = 0; i < nr_merged; i++)
813 		clear_exception(ps, ps->current_committed - 1 - i);
814 
815 	r = area_io(ps, REQ_OP_WRITE, REQ_PREFLUSH | REQ_FUA);
816 	if (r < 0)
817 		return r;
818 
819 	ps->current_committed -= nr_merged;
820 
821 	/*
822 	 * At this stage, only persistent_usage() uses ps->next_free, so
823 	 * we make no attempt to keep ps->next_free strictly accurate
824 	 * as exceptions may have been committed out-of-order originally.
825 	 * Once a snapshot has become merging, we set it to the value it
826 	 * would have held had all the exceptions been committed in order.
827 	 *
828 	 * ps->current_area does not get reduced by prepare_merge() until
829 	 * after commit_merge() has removed the nr_merged previous exceptions.
830 	 */
831 	ps->next_free = area_location(ps, ps->current_area) +
832 			ps->current_committed + 1;
833 
834 	return 0;
835 }
836 
837 static void persistent_drop_snapshot(struct dm_exception_store *store)
838 {
839 	struct pstore *ps = get_info(store);
840 
841 	ps->valid = 0;
842 	if (write_header(ps))
843 		DMWARN("write header failed");
844 }
845 
846 static int persistent_ctr(struct dm_exception_store *store, char *options)
847 {
848 	struct pstore *ps;
849 	int r;
850 
851 	/* allocate the pstore */
852 	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
853 	if (!ps)
854 		return -ENOMEM;
855 
856 	ps->store = store;
857 	ps->valid = 1;
858 	ps->version = SNAPSHOT_DISK_VERSION;
859 	ps->area = NULL;
860 	ps->zero_area = NULL;
861 	ps->header_area = NULL;
862 	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
863 	ps->current_committed = 0;
864 
865 	ps->callback_count = 0;
866 	atomic_set(&ps->pending_count, 0);
867 	ps->callbacks = NULL;
868 
869 	ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
870 	if (!ps->metadata_wq) {
871 		DMERR("couldn't start header metadata update thread");
872 		r = -ENOMEM;
873 		goto err_workqueue;
874 	}
875 
876 	if (options) {
877 		char overflow = toupper(options[0]);
878 		if (overflow == 'O')
879 			store->userspace_supports_overflow = true;
880 		else {
881 			DMERR("Unsupported persistent store option: %s", options);
882 			r = -EINVAL;
883 			goto err_options;
884 		}
885 	}
886 
887 	store->context = ps;
888 
889 	return 0;
890 
891 err_options:
892 	destroy_workqueue(ps->metadata_wq);
893 err_workqueue:
894 	kfree(ps);
895 
896 	return r;
897 }
898 
899 static unsigned persistent_status(struct dm_exception_store *store,
900 				  status_type_t status, char *result,
901 				  unsigned maxlen)
902 {
903 	unsigned sz = 0;
904 
905 	switch (status) {
906 	case STATUSTYPE_INFO:
907 		break;
908 	case STATUSTYPE_TABLE:
909 		DMEMIT(" %s %llu", store->userspace_supports_overflow ? "PO" : "P",
910 		       (unsigned long long)store->chunk_size);
911 	}
912 
913 	return sz;
914 }
915 
916 static struct dm_exception_store_type _persistent_type = {
917 	.name = "persistent",
918 	.module = THIS_MODULE,
919 	.ctr = persistent_ctr,
920 	.dtr = persistent_dtr,
921 	.read_metadata = persistent_read_metadata,
922 	.prepare_exception = persistent_prepare_exception,
923 	.commit_exception = persistent_commit_exception,
924 	.prepare_merge = persistent_prepare_merge,
925 	.commit_merge = persistent_commit_merge,
926 	.drop_snapshot = persistent_drop_snapshot,
927 	.usage = persistent_usage,
928 	.status = persistent_status,
929 };
930 
931 static struct dm_exception_store_type _persistent_compat_type = {
932 	.name = "P",
933 	.module = THIS_MODULE,
934 	.ctr = persistent_ctr,
935 	.dtr = persistent_dtr,
936 	.read_metadata = persistent_read_metadata,
937 	.prepare_exception = persistent_prepare_exception,
938 	.commit_exception = persistent_commit_exception,
939 	.prepare_merge = persistent_prepare_merge,
940 	.commit_merge = persistent_commit_merge,
941 	.drop_snapshot = persistent_drop_snapshot,
942 	.usage = persistent_usage,
943 	.status = persistent_status,
944 };
945 
946 int dm_persistent_snapshot_init(void)
947 {
948 	int r;
949 
950 	r = dm_exception_store_type_register(&_persistent_type);
951 	if (r) {
952 		DMERR("Unable to register persistent exception store type");
953 		return r;
954 	}
955 
956 	r = dm_exception_store_type_register(&_persistent_compat_type);
957 	if (r) {
958 		DMERR("Unable to register old-style persistent exception "
959 		      "store type");
960 		dm_exception_store_type_unregister(&_persistent_type);
961 		return r;
962 	}
963 
964 	return r;
965 }
966 
967 void dm_persistent_snapshot_exit(void)
968 {
969 	dm_exception_store_type_unregister(&_persistent_type);
970 	dm_exception_store_type_unregister(&_persistent_compat_type);
971 }
972