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