Linux Audio

Check our new training course

Loading...
v3.1
  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/slab.h>
 14#include <linux/dm-io.h>
 
 15
 16#define DM_MSG_PREFIX "persistent snapshot"
 17#define DM_CHUNK_SIZE_DEFAULT_SECTORS 32	/* 16KB */
 18
 
 
 19/*-----------------------------------------------------------------
 20 * Persistent snapshots, by persistent we mean that the snapshot
 21 * will survive a reboot.
 22 *---------------------------------------------------------------*/
 23
 24/*
 25 * We need to store a record of which parts of the origin have
 26 * been copied to the snapshot device.  The snapshot code
 27 * requires that we copy exception chunks to chunk aligned areas
 28 * of the COW store.  It makes sense therefore, to store the
 29 * metadata in chunk size blocks.
 30 *
 31 * There is no backward or forward compatibility implemented,
 32 * snapshots with different disk versions than the kernel will
 33 * not be usable.  It is expected that "lvcreate" will blank out
 34 * the start of a fresh COW device before calling the snapshot
 35 * constructor.
 36 *
 37 * The first chunk of the COW device just contains the header.
 38 * After this there is a chunk filled with exception metadata,
 39 * followed by as many exception chunks as can fit in the
 40 * metadata areas.
 41 *
 42 * All on disk structures are in little-endian format.  The end
 43 * of the exceptions info is indicated by an exception with a
 44 * new_chunk of 0, which is invalid since it would point to the
 45 * header chunk.
 46 */
 47
 48/*
 49 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
 50 */
 51#define SNAP_MAGIC 0x70416e53
 52
 53/*
 54 * The on-disk version of the metadata.
 55 */
 56#define SNAPSHOT_DISK_VERSION 1
 57
 58#define NUM_SNAPSHOT_HDR_CHUNKS 1
 59
 60struct disk_header {
 61	__le32 magic;
 62
 63	/*
 64	 * Is this snapshot valid.  There is no way of recovering
 65	 * an invalid snapshot.
 66	 */
 67	__le32 valid;
 68
 69	/*
 70	 * Simple, incrementing version. no backward
 71	 * compatibility.
 72	 */
 73	__le32 version;
 74
 75	/* In sectors */
 76	__le32 chunk_size;
 77} __packed;
 78
 79struct disk_exception {
 80	__le64 old_chunk;
 81	__le64 new_chunk;
 82} __packed;
 83
 84struct core_exception {
 85	uint64_t old_chunk;
 86	uint64_t new_chunk;
 87};
 88
 89struct commit_callback {
 90	void (*callback)(void *, int success);
 91	void *context;
 92};
 93
 94/*
 95 * The top level structure for a persistent exception store.
 96 */
 97struct pstore {
 98	struct dm_exception_store *store;
 99	int version;
100	int valid;
101	uint32_t exceptions_per_area;
102
103	/*
104	 * Now that we have an asynchronous kcopyd there is no
105	 * need for large chunk sizes, so it wont hurt to have a
106	 * whole chunks worth of metadata in memory at once.
107	 */
108	void *area;
109
110	/*
111	 * An area of zeros used to clear the next area.
112	 */
113	void *zero_area;
114
115	/*
116	 * An area used for header. The header can be written
117	 * concurrently with metadata (when invalidating the snapshot),
118	 * so it needs a separate buffer.
119	 */
120	void *header_area;
121
122	/*
123	 * Used to keep track of which metadata area the data in
124	 * 'chunk' refers to.
125	 */
126	chunk_t current_area;
127
128	/*
129	 * The next free chunk for an exception.
130	 *
131	 * When creating exceptions, all the chunks here and above are
132	 * free.  It holds the next chunk to be allocated.  On rare
133	 * occasions (e.g. after a system crash) holes can be left in
134	 * the exception store because chunks can be committed out of
135	 * order.
136	 *
137	 * When merging exceptions, it does not necessarily mean all the
138	 * chunks here and above are free.  It holds the value it would
139	 * have held if all chunks had been committed in order of
140	 * allocation.  Consequently the value may occasionally be
141	 * slightly too low, but since it's only used for 'status' and
142	 * it can never reach its minimum value too early this doesn't
143	 * matter.
144	 */
145
146	chunk_t next_free;
147
148	/*
149	 * The index of next free exception in the current
150	 * metadata area.
151	 */
152	uint32_t current_committed;
153
154	atomic_t pending_count;
155	uint32_t callback_count;
156	struct commit_callback *callbacks;
157	struct dm_io_client *io_client;
158
159	struct workqueue_struct *metadata_wq;
160};
161
162static int alloc_area(struct pstore *ps)
163{
164	int r = -ENOMEM;
165	size_t len;
166
167	len = ps->store->chunk_size << SECTOR_SHIFT;
168
169	/*
170	 * Allocate the chunk_size block of memory that will hold
171	 * a single metadata area.
172	 */
173	ps->area = vmalloc(len);
174	if (!ps->area)
175		goto err_area;
176
177	ps->zero_area = vzalloc(len);
178	if (!ps->zero_area)
179		goto err_zero_area;
180
181	ps->header_area = vmalloc(len);
182	if (!ps->header_area)
183		goto err_header_area;
184
185	return 0;
186
187err_header_area:
188	vfree(ps->zero_area);
189
190err_zero_area:
191	vfree(ps->area);
192
193err_area:
194	return r;
195}
196
197static void free_area(struct pstore *ps)
198{
199	if (ps->area)
200		vfree(ps->area);
201	ps->area = NULL;
202
203	if (ps->zero_area)
204		vfree(ps->zero_area);
205	ps->zero_area = NULL;
206
207	if (ps->header_area)
208		vfree(ps->header_area);
209	ps->header_area = NULL;
210}
211
212struct 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
219static 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 */
229static 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_work(&req.work);
 
259
260	return req.result;
261}
262
263/*
264 * Convert a metadata area index to a chunk index.
265 */
266static chunk_t area_location(struct pstore *ps, chunk_t area)
267{
268	return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
269}
270
 
 
 
 
 
 
 
 
271/*
272 * Read or write a metadata area.  Remembering to skip the first
273 * chunk which holds the header.
274 */
275static int area_io(struct pstore *ps, int rw)
276{
277	int r;
278	chunk_t chunk;
279
280	chunk = area_location(ps, ps->current_area);
281
282	r = chunk_io(ps, ps->area, chunk, rw, 0);
283	if (r)
284		return r;
285
286	return 0;
287}
288
289static void zero_memory_area(struct pstore *ps)
290{
291	memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
292}
293
294static int zero_disk_area(struct pstore *ps, chunk_t area)
295{
296	return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
297}
298
299static int read_header(struct pstore *ps, int *new_snapshot)
300{
301	int r;
302	struct disk_header *dh;
303	unsigned chunk_size;
304	int chunk_size_supplied = 1;
305	char *chunk_err;
306
307	/*
308	 * Use default chunk size (or logical_block_size, if larger)
309	 * if none supplied
310	 */
311	if (!ps->store->chunk_size) {
312		ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
313		    bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
314					    bdev) >> 9);
315		ps->store->chunk_mask = ps->store->chunk_size - 1;
316		ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
317		chunk_size_supplied = 0;
318	}
319
320	ps->io_client = dm_io_client_create();
321	if (IS_ERR(ps->io_client))
322		return PTR_ERR(ps->io_client);
323
324	r = alloc_area(ps);
325	if (r)
326		return r;
327
328	r = chunk_io(ps, ps->header_area, 0, READ, 1);
329	if (r)
330		goto bad;
331
332	dh = ps->header_area;
333
334	if (le32_to_cpu(dh->magic) == 0) {
335		*new_snapshot = 1;
336		return 0;
337	}
338
339	if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
340		DMWARN("Invalid or corrupt snapshot");
341		r = -ENXIO;
342		goto bad;
343	}
344
345	*new_snapshot = 0;
346	ps->valid = le32_to_cpu(dh->valid);
347	ps->version = le32_to_cpu(dh->version);
348	chunk_size = le32_to_cpu(dh->chunk_size);
349
350	if (ps->store->chunk_size == chunk_size)
351		return 0;
352
353	if (chunk_size_supplied)
354		DMWARN("chunk size %u in device metadata overrides "
355		       "table chunk size of %u.",
356		       chunk_size, ps->store->chunk_size);
357
358	/* We had a bogus chunk_size. Fix stuff up. */
359	free_area(ps);
360
361	r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
362					      &chunk_err);
363	if (r) {
364		DMERR("invalid on-disk chunk size %u: %s.",
365		      chunk_size, chunk_err);
366		return r;
367	}
368
369	r = alloc_area(ps);
370	return r;
371
372bad:
373	free_area(ps);
374	return r;
375}
376
377static int write_header(struct pstore *ps)
378{
379	struct disk_header *dh;
380
381	memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
382
383	dh = ps->header_area;
384	dh->magic = cpu_to_le32(SNAP_MAGIC);
385	dh->valid = cpu_to_le32(ps->valid);
386	dh->version = cpu_to_le32(ps->version);
387	dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
388
389	return chunk_io(ps, ps->header_area, 0, WRITE, 1);
390}
391
392/*
393 * Access functions for the disk exceptions, these do the endian conversions.
394 */
395static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
 
396{
397	BUG_ON(index >= ps->exceptions_per_area);
398
399	return ((struct disk_exception *) ps->area) + index;
400}
401
402static void read_exception(struct pstore *ps,
403			   uint32_t index, struct core_exception *result)
404{
405	struct disk_exception *de = get_exception(ps, index);
406
407	/* copy it */
408	result->old_chunk = le64_to_cpu(de->old_chunk);
409	result->new_chunk = le64_to_cpu(de->new_chunk);
410}
411
412static void write_exception(struct pstore *ps,
413			    uint32_t index, struct core_exception *e)
414{
415	struct disk_exception *de = get_exception(ps, index);
416
417	/* copy it */
418	de->old_chunk = cpu_to_le64(e->old_chunk);
419	de->new_chunk = cpu_to_le64(e->new_chunk);
420}
421
422static void clear_exception(struct pstore *ps, uint32_t index)
423{
424	struct disk_exception *de = get_exception(ps, index);
425
426	/* clear it */
427	de->old_chunk = 0;
428	de->new_chunk = 0;
429}
430
431/*
432 * Registers the exceptions that are present in the current area.
433 * 'full' is filled in to indicate if the area has been
434 * filled.
435 */
436static int insert_exceptions(struct pstore *ps,
437			     int (*callback)(void *callback_context,
438					     chunk_t old, chunk_t new),
439			     void *callback_context,
440			     int *full)
441{
442	int r;
443	unsigned int i;
444	struct core_exception e;
445
446	/* presume the area is full */
447	*full = 1;
448
449	for (i = 0; i < ps->exceptions_per_area; i++) {
450		read_exception(ps, i, &e);
451
452		/*
453		 * If the new_chunk is pointing at the start of
454		 * the COW device, where the first metadata area
455		 * is we know that we've hit the end of the
456		 * exceptions.  Therefore the area is not full.
457		 */
458		if (e.new_chunk == 0LL) {
459			ps->current_committed = i;
460			*full = 0;
461			break;
462		}
463
464		/*
465		 * Keep track of the start of the free chunks.
466		 */
467		if (ps->next_free <= e.new_chunk)
468			ps->next_free = e.new_chunk + 1;
469
470		/*
471		 * Otherwise we add the exception to the snapshot.
472		 */
473		r = callback(callback_context, e.old_chunk, e.new_chunk);
474		if (r)
475			return r;
476	}
477
478	return 0;
479}
480
481static int read_exceptions(struct pstore *ps,
482			   int (*callback)(void *callback_context, chunk_t old,
483					   chunk_t new),
484			   void *callback_context)
485{
486	int r, full = 1;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
487
488	/*
489	 * Keeping reading chunks and inserting exceptions until
490	 * we find a partially full area.
491	 */
492	for (ps->current_area = 0; full; ps->current_area++) {
493		r = area_io(ps, READ);
494		if (r)
495			return r;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
496
497		r = insert_exceptions(ps, callback, callback_context, &full);
498		if (r)
499			return r;
 
 
 
 
 
 
 
 
 
500	}
501
502	ps->current_area--;
503
504	return 0;
 
 
 
 
 
 
 
505}
506
507static struct pstore *get_info(struct dm_exception_store *store)
508{
509	return (struct pstore *) store->context;
510}
511
512static void persistent_usage(struct dm_exception_store *store,
513			     sector_t *total_sectors,
514			     sector_t *sectors_allocated,
515			     sector_t *metadata_sectors)
516{
517	struct pstore *ps = get_info(store);
518
519	*sectors_allocated = ps->next_free * store->chunk_size;
520	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
521
522	/*
523	 * First chunk is the fixed header.
524	 * Then there are (ps->current_area + 1) metadata chunks, each one
525	 * separated from the next by ps->exceptions_per_area data chunks.
526	 */
527	*metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
528			    store->chunk_size;
529}
530
531static void persistent_dtr(struct dm_exception_store *store)
532{
533	struct pstore *ps = get_info(store);
534
535	destroy_workqueue(ps->metadata_wq);
536
537	/* Created in read_header */
538	if (ps->io_client)
539		dm_io_client_destroy(ps->io_client);
540	free_area(ps);
541
542	/* Allocated in persistent_read_metadata */
543	if (ps->callbacks)
544		vfree(ps->callbacks);
545
546	kfree(ps);
547}
548
549static int persistent_read_metadata(struct dm_exception_store *store,
550				    int (*callback)(void *callback_context,
551						    chunk_t old, chunk_t new),
552				    void *callback_context)
553{
554	int r, uninitialized_var(new_snapshot);
555	struct pstore *ps = get_info(store);
556
557	/*
558	 * Read the snapshot header.
559	 */
560	r = read_header(ps, &new_snapshot);
561	if (r)
562		return r;
563
564	/*
565	 * Now we know correct chunk_size, complete the initialisation.
566	 */
567	ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
568				  sizeof(struct disk_exception);
569	ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
570				   sizeof(*ps->callbacks));
571	if (!ps->callbacks)
572		return -ENOMEM;
573
574	/*
575	 * Do we need to setup a new snapshot ?
576	 */
577	if (new_snapshot) {
578		r = write_header(ps);
579		if (r) {
580			DMWARN("write_header failed");
581			return r;
582		}
583
584		ps->current_area = 0;
585		zero_memory_area(ps);
586		r = zero_disk_area(ps, 0);
587		if (r)
588			DMWARN("zero_disk_area(0) failed");
589		return r;
590	}
591	/*
592	 * Sanity checks.
593	 */
594	if (ps->version != SNAPSHOT_DISK_VERSION) {
595		DMWARN("unable to handle snapshot disk version %d",
596		       ps->version);
597		return -EINVAL;
598	}
599
600	/*
601	 * Metadata are valid, but snapshot is invalidated
602	 */
603	if (!ps->valid)
604		return 1;
605
606	/*
607	 * Read the metadata.
608	 */
609	r = read_exceptions(ps, callback, callback_context);
610
611	return r;
612}
613
614static int persistent_prepare_exception(struct dm_exception_store *store,
615					struct dm_exception *e)
616{
617	struct pstore *ps = get_info(store);
618	uint32_t stride;
619	chunk_t next_free;
620	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
621
622	/* Is there enough room ? */
623	if (size < ((ps->next_free + 1) * store->chunk_size))
624		return -ENOSPC;
625
626	e->new_chunk = ps->next_free;
627
628	/*
629	 * Move onto the next free pending, making sure to take
630	 * into account the location of the metadata chunks.
631	 */
632	stride = (ps->exceptions_per_area + 1);
633	next_free = ++ps->next_free;
634	if (sector_div(next_free, stride) == 1)
635		ps->next_free++;
636
637	atomic_inc(&ps->pending_count);
638	return 0;
639}
640
641static void persistent_commit_exception(struct dm_exception_store *store,
642					struct dm_exception *e,
643					void (*callback) (void *, int success),
644					void *callback_context)
645{
646	unsigned int i;
647	struct pstore *ps = get_info(store);
648	struct core_exception ce;
649	struct commit_callback *cb;
650
651	ce.old_chunk = e->old_chunk;
652	ce.new_chunk = e->new_chunk;
653	write_exception(ps, ps->current_committed++, &ce);
654
655	/*
656	 * Add the callback to the back of the array.  This code
657	 * is the only place where the callback array is
658	 * manipulated, and we know that it will never be called
659	 * multiple times concurrently.
660	 */
661	cb = ps->callbacks + ps->callback_count++;
662	cb->callback = callback;
663	cb->context = callback_context;
664
665	/*
666	 * If there are exceptions in flight and we have not yet
667	 * filled this metadata area there's nothing more to do.
668	 */
669	if (!atomic_dec_and_test(&ps->pending_count) &&
670	    (ps->current_committed != ps->exceptions_per_area))
671		return;
672
673	/*
674	 * If we completely filled the current area, then wipe the next one.
675	 */
676	if ((ps->current_committed == ps->exceptions_per_area) &&
677	    zero_disk_area(ps, ps->current_area + 1))
678		ps->valid = 0;
679
680	/*
681	 * Commit exceptions to disk.
682	 */
683	if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
684		ps->valid = 0;
685
686	/*
687	 * Advance to the next area if this one is full.
688	 */
689	if (ps->current_committed == ps->exceptions_per_area) {
690		ps->current_committed = 0;
691		ps->current_area++;
692		zero_memory_area(ps);
693	}
694
695	for (i = 0; i < ps->callback_count; i++) {
696		cb = ps->callbacks + i;
697		cb->callback(cb->context, ps->valid);
698	}
699
700	ps->callback_count = 0;
701}
702
703static int persistent_prepare_merge(struct dm_exception_store *store,
704				    chunk_t *last_old_chunk,
705				    chunk_t *last_new_chunk)
706{
707	struct pstore *ps = get_info(store);
708	struct core_exception ce;
709	int nr_consecutive;
710	int r;
711
712	/*
713	 * When current area is empty, move back to preceding area.
714	 */
715	if (!ps->current_committed) {
716		/*
717		 * Have we finished?
718		 */
719		if (!ps->current_area)
720			return 0;
721
722		ps->current_area--;
723		r = area_io(ps, READ);
724		if (r < 0)
725			return r;
726		ps->current_committed = ps->exceptions_per_area;
727	}
728
729	read_exception(ps, ps->current_committed - 1, &ce);
730	*last_old_chunk = ce.old_chunk;
731	*last_new_chunk = ce.new_chunk;
732
733	/*
734	 * Find number of consecutive chunks within the current area,
735	 * working backwards.
736	 */
737	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
738	     nr_consecutive++) {
739		read_exception(ps, ps->current_committed - 1 - nr_consecutive,
740			       &ce);
741		if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
742		    ce.new_chunk != *last_new_chunk - nr_consecutive)
743			break;
744	}
745
746	return nr_consecutive;
747}
748
749static int persistent_commit_merge(struct dm_exception_store *store,
750				   int nr_merged)
751{
752	int r, i;
753	struct pstore *ps = get_info(store);
754
755	BUG_ON(nr_merged > ps->current_committed);
756
757	for (i = 0; i < nr_merged; i++)
758		clear_exception(ps, ps->current_committed - 1 - i);
759
760	r = area_io(ps, WRITE_FLUSH_FUA);
761	if (r < 0)
762		return r;
763
764	ps->current_committed -= nr_merged;
765
766	/*
767	 * At this stage, only persistent_usage() uses ps->next_free, so
768	 * we make no attempt to keep ps->next_free strictly accurate
769	 * as exceptions may have been committed out-of-order originally.
770	 * Once a snapshot has become merging, we set it to the value it
771	 * would have held had all the exceptions been committed in order.
772	 *
773	 * ps->current_area does not get reduced by prepare_merge() until
774	 * after commit_merge() has removed the nr_merged previous exceptions.
775	 */
776	ps->next_free = area_location(ps, ps->current_area) +
777			ps->current_committed + 1;
778
779	return 0;
780}
781
782static void persistent_drop_snapshot(struct dm_exception_store *store)
783{
784	struct pstore *ps = get_info(store);
785
786	ps->valid = 0;
787	if (write_header(ps))
788		DMWARN("write header failed");
789}
790
791static int persistent_ctr(struct dm_exception_store *store,
792			  unsigned argc, char **argv)
793{
794	struct pstore *ps;
795
796	/* allocate the pstore */
797	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
798	if (!ps)
799		return -ENOMEM;
800
801	ps->store = store;
802	ps->valid = 1;
803	ps->version = SNAPSHOT_DISK_VERSION;
804	ps->area = NULL;
805	ps->zero_area = NULL;
806	ps->header_area = NULL;
807	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
808	ps->current_committed = 0;
809
810	ps->callback_count = 0;
811	atomic_set(&ps->pending_count, 0);
812	ps->callbacks = NULL;
813
814	ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
815	if (!ps->metadata_wq) {
816		kfree(ps);
817		DMERR("couldn't start header metadata update thread");
818		return -ENOMEM;
819	}
820
821	store->context = ps;
822
823	return 0;
824}
825
826static unsigned persistent_status(struct dm_exception_store *store,
827				  status_type_t status, char *result,
828				  unsigned maxlen)
829{
830	unsigned sz = 0;
831
832	switch (status) {
833	case STATUSTYPE_INFO:
834		break;
835	case STATUSTYPE_TABLE:
836		DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
837	}
838
839	return sz;
840}
841
842static struct dm_exception_store_type _persistent_type = {
843	.name = "persistent",
844	.module = THIS_MODULE,
845	.ctr = persistent_ctr,
846	.dtr = persistent_dtr,
847	.read_metadata = persistent_read_metadata,
848	.prepare_exception = persistent_prepare_exception,
849	.commit_exception = persistent_commit_exception,
850	.prepare_merge = persistent_prepare_merge,
851	.commit_merge = persistent_commit_merge,
852	.drop_snapshot = persistent_drop_snapshot,
853	.usage = persistent_usage,
854	.status = persistent_status,
855};
856
857static struct dm_exception_store_type _persistent_compat_type = {
858	.name = "P",
859	.module = THIS_MODULE,
860	.ctr = persistent_ctr,
861	.dtr = persistent_dtr,
862	.read_metadata = persistent_read_metadata,
863	.prepare_exception = persistent_prepare_exception,
864	.commit_exception = persistent_commit_exception,
865	.prepare_merge = persistent_prepare_merge,
866	.commit_merge = persistent_commit_merge,
867	.drop_snapshot = persistent_drop_snapshot,
868	.usage = persistent_usage,
869	.status = persistent_status,
870};
871
872int dm_persistent_snapshot_init(void)
873{
874	int r;
875
876	r = dm_exception_store_type_register(&_persistent_type);
877	if (r) {
878		DMERR("Unable to register persistent exception store type");
879		return r;
880	}
881
882	r = dm_exception_store_type_register(&_persistent_compat_type);
883	if (r) {
884		DMERR("Unable to register old-style persistent exception "
885		      "store type");
886		dm_exception_store_type_unregister(&_persistent_type);
887		return r;
888	}
889
890	return r;
891}
892
893void dm_persistent_snapshot_exit(void)
894{
895	dm_exception_store_type_unregister(&_persistent_type);
896	dm_exception_store_type_unregister(&_persistent_compat_type);
897}
v3.15
  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
 64struct 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
 83struct disk_exception {
 84	__le64 old_chunk;
 85	__le64 new_chunk;
 86} __packed;
 87
 88struct core_exception {
 89	uint64_t old_chunk;
 90	uint64_t new_chunk;
 91};
 92
 93struct 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 */
101struct 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
166static 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
191err_header_area:
192	vfree(ps->zero_area);
193
194err_zero_area:
195	vfree(ps->area);
196
197err_area:
198	return r;
199}
200
201static 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
216struct 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
223static 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 */
233static 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 */
271static 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
276static 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 */
288static 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
302static void zero_memory_area(struct pstore *ps)
303{
304	memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
305}
306
307static 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
312static 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
385bad:
386	free_area(ps);
387	return r;
388}
389
390static 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 */
408static 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
416static 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
426static 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
436static 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 */
450static 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
495static 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		if (!full)
550			memcpy(ps->area, area, ps->store->chunk_size << SECTOR_SHIFT);
551
552		dm_bufio_release(bp);
553
554		dm_bufio_forget(client, chunk);
555
556		if (unlikely(r))
557			goto ret_destroy_bufio;
558	}
559
560	ps->current_area--;
561
562	skip_metadata(ps);
563
564	r = 0;
565
566ret_destroy_bufio:
567	dm_bufio_client_destroy(client);
568
569	return r;
570}
571
572static struct pstore *get_info(struct dm_exception_store *store)
573{
574	return (struct pstore *) store->context;
575}
576
577static void persistent_usage(struct dm_exception_store *store,
578			     sector_t *total_sectors,
579			     sector_t *sectors_allocated,
580			     sector_t *metadata_sectors)
581{
582	struct pstore *ps = get_info(store);
583
584	*sectors_allocated = ps->next_free * store->chunk_size;
585	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
586
587	/*
588	 * First chunk is the fixed header.
589	 * Then there are (ps->current_area + 1) metadata chunks, each one
590	 * separated from the next by ps->exceptions_per_area data chunks.
591	 */
592	*metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
593			    store->chunk_size;
594}
595
596static void persistent_dtr(struct dm_exception_store *store)
597{
598	struct pstore *ps = get_info(store);
599
600	destroy_workqueue(ps->metadata_wq);
601
602	/* Created in read_header */
603	if (ps->io_client)
604		dm_io_client_destroy(ps->io_client);
605	free_area(ps);
606
607	/* Allocated in persistent_read_metadata */
608	if (ps->callbacks)
609		vfree(ps->callbacks);
610
611	kfree(ps);
612}
613
614static int persistent_read_metadata(struct dm_exception_store *store,
615				    int (*callback)(void *callback_context,
616						    chunk_t old, chunk_t new),
617				    void *callback_context)
618{
619	int r, uninitialized_var(new_snapshot);
620	struct pstore *ps = get_info(store);
621
622	/*
623	 * Read the snapshot header.
624	 */
625	r = read_header(ps, &new_snapshot);
626	if (r)
627		return r;
628
629	/*
630	 * Now we know correct chunk_size, complete the initialisation.
631	 */
632	ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
633				  sizeof(struct disk_exception);
634	ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
635				   sizeof(*ps->callbacks));
636	if (!ps->callbacks)
637		return -ENOMEM;
638
639	/*
640	 * Do we need to setup a new snapshot ?
641	 */
642	if (new_snapshot) {
643		r = write_header(ps);
644		if (r) {
645			DMWARN("write_header failed");
646			return r;
647		}
648
649		ps->current_area = 0;
650		zero_memory_area(ps);
651		r = zero_disk_area(ps, 0);
652		if (r)
653			DMWARN("zero_disk_area(0) failed");
654		return r;
655	}
656	/*
657	 * Sanity checks.
658	 */
659	if (ps->version != SNAPSHOT_DISK_VERSION) {
660		DMWARN("unable to handle snapshot disk version %d",
661		       ps->version);
662		return -EINVAL;
663	}
664
665	/*
666	 * Metadata are valid, but snapshot is invalidated
667	 */
668	if (!ps->valid)
669		return 1;
670
671	/*
672	 * Read the metadata.
673	 */
674	r = read_exceptions(ps, callback, callback_context);
675
676	return r;
677}
678
679static int persistent_prepare_exception(struct dm_exception_store *store,
680					struct dm_exception *e)
681{
682	struct pstore *ps = get_info(store);
 
 
683	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
684
685	/* Is there enough room ? */
686	if (size < ((ps->next_free + 1) * store->chunk_size))
687		return -ENOSPC;
688
689	e->new_chunk = ps->next_free;
690
691	/*
692	 * Move onto the next free pending, making sure to take
693	 * into account the location of the metadata chunks.
694	 */
695	ps->next_free++;
696	skip_metadata(ps);
 
 
697
698	atomic_inc(&ps->pending_count);
699	return 0;
700}
701
702static void persistent_commit_exception(struct dm_exception_store *store,
703					struct dm_exception *e,
704					void (*callback) (void *, int success),
705					void *callback_context)
706{
707	unsigned int i;
708	struct pstore *ps = get_info(store);
709	struct core_exception ce;
710	struct commit_callback *cb;
711
712	ce.old_chunk = e->old_chunk;
713	ce.new_chunk = e->new_chunk;
714	write_exception(ps, ps->current_committed++, &ce);
715
716	/*
717	 * Add the callback to the back of the array.  This code
718	 * is the only place where the callback array is
719	 * manipulated, and we know that it will never be called
720	 * multiple times concurrently.
721	 */
722	cb = ps->callbacks + ps->callback_count++;
723	cb->callback = callback;
724	cb->context = callback_context;
725
726	/*
727	 * If there are exceptions in flight and we have not yet
728	 * filled this metadata area there's nothing more to do.
729	 */
730	if (!atomic_dec_and_test(&ps->pending_count) &&
731	    (ps->current_committed != ps->exceptions_per_area))
732		return;
733
734	/*
735	 * If we completely filled the current area, then wipe the next one.
736	 */
737	if ((ps->current_committed == ps->exceptions_per_area) &&
738	    zero_disk_area(ps, ps->current_area + 1))
739		ps->valid = 0;
740
741	/*
742	 * Commit exceptions to disk.
743	 */
744	if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
745		ps->valid = 0;
746
747	/*
748	 * Advance to the next area if this one is full.
749	 */
750	if (ps->current_committed == ps->exceptions_per_area) {
751		ps->current_committed = 0;
752		ps->current_area++;
753		zero_memory_area(ps);
754	}
755
756	for (i = 0; i < ps->callback_count; i++) {
757		cb = ps->callbacks + i;
758		cb->callback(cb->context, ps->valid);
759	}
760
761	ps->callback_count = 0;
762}
763
764static int persistent_prepare_merge(struct dm_exception_store *store,
765				    chunk_t *last_old_chunk,
766				    chunk_t *last_new_chunk)
767{
768	struct pstore *ps = get_info(store);
769	struct core_exception ce;
770	int nr_consecutive;
771	int r;
772
773	/*
774	 * When current area is empty, move back to preceding area.
775	 */
776	if (!ps->current_committed) {
777		/*
778		 * Have we finished?
779		 */
780		if (!ps->current_area)
781			return 0;
782
783		ps->current_area--;
784		r = area_io(ps, READ);
785		if (r < 0)
786			return r;
787		ps->current_committed = ps->exceptions_per_area;
788	}
789
790	read_exception(ps, ps->area, ps->current_committed - 1, &ce);
791	*last_old_chunk = ce.old_chunk;
792	*last_new_chunk = ce.new_chunk;
793
794	/*
795	 * Find number of consecutive chunks within the current area,
796	 * working backwards.
797	 */
798	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
799	     nr_consecutive++) {
800		read_exception(ps, ps->area,
801			       ps->current_committed - 1 - nr_consecutive, &ce);
802		if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
803		    ce.new_chunk != *last_new_chunk - nr_consecutive)
804			break;
805	}
806
807	return nr_consecutive;
808}
809
810static int persistent_commit_merge(struct dm_exception_store *store,
811				   int nr_merged)
812{
813	int r, i;
814	struct pstore *ps = get_info(store);
815
816	BUG_ON(nr_merged > ps->current_committed);
817
818	for (i = 0; i < nr_merged; i++)
819		clear_exception(ps, ps->current_committed - 1 - i);
820
821	r = area_io(ps, WRITE_FLUSH_FUA);
822	if (r < 0)
823		return r;
824
825	ps->current_committed -= nr_merged;
826
827	/*
828	 * At this stage, only persistent_usage() uses ps->next_free, so
829	 * we make no attempt to keep ps->next_free strictly accurate
830	 * as exceptions may have been committed out-of-order originally.
831	 * Once a snapshot has become merging, we set it to the value it
832	 * would have held had all the exceptions been committed in order.
833	 *
834	 * ps->current_area does not get reduced by prepare_merge() until
835	 * after commit_merge() has removed the nr_merged previous exceptions.
836	 */
837	ps->next_free = area_location(ps, ps->current_area) +
838			ps->current_committed + 1;
839
840	return 0;
841}
842
843static void persistent_drop_snapshot(struct dm_exception_store *store)
844{
845	struct pstore *ps = get_info(store);
846
847	ps->valid = 0;
848	if (write_header(ps))
849		DMWARN("write header failed");
850}
851
852static int persistent_ctr(struct dm_exception_store *store,
853			  unsigned argc, char **argv)
854{
855	struct pstore *ps;
856
857	/* allocate the pstore */
858	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
859	if (!ps)
860		return -ENOMEM;
861
862	ps->store = store;
863	ps->valid = 1;
864	ps->version = SNAPSHOT_DISK_VERSION;
865	ps->area = NULL;
866	ps->zero_area = NULL;
867	ps->header_area = NULL;
868	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
869	ps->current_committed = 0;
870
871	ps->callback_count = 0;
872	atomic_set(&ps->pending_count, 0);
873	ps->callbacks = NULL;
874
875	ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
876	if (!ps->metadata_wq) {
877		kfree(ps);
878		DMERR("couldn't start header metadata update thread");
879		return -ENOMEM;
880	}
881
882	store->context = ps;
883
884	return 0;
885}
886
887static unsigned persistent_status(struct dm_exception_store *store,
888				  status_type_t status, char *result,
889				  unsigned maxlen)
890{
891	unsigned sz = 0;
892
893	switch (status) {
894	case STATUSTYPE_INFO:
895		break;
896	case STATUSTYPE_TABLE:
897		DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
898	}
899
900	return sz;
901}
902
903static struct dm_exception_store_type _persistent_type = {
904	.name = "persistent",
905	.module = THIS_MODULE,
906	.ctr = persistent_ctr,
907	.dtr = persistent_dtr,
908	.read_metadata = persistent_read_metadata,
909	.prepare_exception = persistent_prepare_exception,
910	.commit_exception = persistent_commit_exception,
911	.prepare_merge = persistent_prepare_merge,
912	.commit_merge = persistent_commit_merge,
913	.drop_snapshot = persistent_drop_snapshot,
914	.usage = persistent_usage,
915	.status = persistent_status,
916};
917
918static struct dm_exception_store_type _persistent_compat_type = {
919	.name = "P",
920	.module = THIS_MODULE,
921	.ctr = persistent_ctr,
922	.dtr = persistent_dtr,
923	.read_metadata = persistent_read_metadata,
924	.prepare_exception = persistent_prepare_exception,
925	.commit_exception = persistent_commit_exception,
926	.prepare_merge = persistent_prepare_merge,
927	.commit_merge = persistent_commit_merge,
928	.drop_snapshot = persistent_drop_snapshot,
929	.usage = persistent_usage,
930	.status = persistent_status,
931};
932
933int dm_persistent_snapshot_init(void)
934{
935	int r;
936
937	r = dm_exception_store_type_register(&_persistent_type);
938	if (r) {
939		DMERR("Unable to register persistent exception store type");
940		return r;
941	}
942
943	r = dm_exception_store_type_register(&_persistent_compat_type);
944	if (r) {
945		DMERR("Unable to register old-style persistent exception "
946		      "store type");
947		dm_exception_store_type_unregister(&_persistent_type);
948		return r;
949	}
950
951	return r;
952}
953
954void dm_persistent_snapshot_exit(void)
955{
956	dm_exception_store_type_unregister(&_persistent_type);
957	dm_exception_store_type_unregister(&_persistent_compat_type);
958}