4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
6 * This file is released under the GPL.
15 #include <linux/pagemap.h>
16 #include <linux/vmalloc.h>
17 #include <linux/slab.h>
19 #define DM_MSG_PREFIX "snapshots"
20 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
22 /*-----------------------------------------------------------------
23 * Persistent snapshots, by persistent we mean that the snapshot
24 * will survive a reboot.
25 *---------------------------------------------------------------*/
28 * We need to store a record of which parts of the origin have
29 * been copied to the snapshot device. The snapshot code
30 * requires that we copy exception chunks to chunk aligned areas
31 * of the COW store. It makes sense therefore, to store the
32 * metadata in chunk size blocks.
34 * There is no backward or forward compatibility implemented,
35 * snapshots with different disk versions than the kernel will
36 * not be usable. It is expected that "lvcreate" will blank out
37 * the start of a fresh COW device before calling the snapshot
40 * The first chunk of the COW device just contains the header.
41 * After this there is a chunk filled with exception metadata,
42 * followed by as many exception chunks as can fit in the
45 * All on disk structures are in little-endian format. The end
46 * of the exceptions info is indicated by an exception with a
47 * new_chunk of 0, which is invalid since it would point to the
52 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
54 #define SNAP_MAGIC 0x70416e53
57 * The on-disk version of the metadata.
59 #define SNAPSHOT_DISK_VERSION 1
65 * Is this snapshot valid. There is no way of recovering
66 * an invalid snapshot.
71 * Simple, incrementing version. no backward
80 struct disk_exception
{
85 struct commit_callback
{
86 void (*callback
)(void *, int success
);
91 * The top level structure for a persistent exception store.
94 struct dm_snapshot
*snap
; /* up pointer to my snapshot */
97 uint32_t exceptions_per_area
;
100 * Now that we have an asynchronous kcopyd there is no
101 * need for large chunk sizes, so it wont hurt to have a
102 * whole chunks worth of metadata in memory at once.
107 * Used to keep track of which metadata area the data in
110 uint32_t current_area
;
113 * The next free chunk for an exception.
118 * The index of next free exception in the current
121 uint32_t current_committed
;
123 atomic_t pending_count
;
124 uint32_t callback_count
;
125 struct commit_callback
*callbacks
;
128 static inline unsigned int sectors_to_pages(unsigned int sectors
)
130 return sectors
/ (PAGE_SIZE
>> 9);
133 static int alloc_area(struct pstore
*ps
)
138 len
= ps
->snap
->chunk_size
<< SECTOR_SHIFT
;
141 * Allocate the chunk_size block of memory that will hold
142 * a single metadata area.
144 ps
->area
= vmalloc(len
);
151 static void free_area(struct pstore
*ps
)
158 * Read or write a chunk aligned and sized block of data from a device.
160 static int chunk_io(struct pstore
*ps
, uint32_t chunk
, int rw
)
162 struct io_region where
;
165 where
.bdev
= ps
->snap
->cow
->bdev
;
166 where
.sector
= ps
->snap
->chunk_size
* chunk
;
167 where
.count
= ps
->snap
->chunk_size
;
169 return dm_io_sync_vm(1, &where
, rw
, ps
->area
, &bits
);
173 * Read or write a metadata area. Remembering to skip the first
174 * chunk which holds the header.
176 static int area_io(struct pstore
*ps
, uint32_t area
, int rw
)
181 /* convert a metadata area index to a chunk index */
182 chunk
= 1 + ((ps
->exceptions_per_area
+ 1) * area
);
184 r
= chunk_io(ps
, chunk
, rw
);
188 ps
->current_area
= area
;
192 static int zero_area(struct pstore
*ps
, uint32_t area
)
194 memset(ps
->area
, 0, ps
->snap
->chunk_size
<< SECTOR_SHIFT
);
195 return area_io(ps
, area
, WRITE
);
198 static int read_header(struct pstore
*ps
, int *new_snapshot
)
201 struct disk_header
*dh
;
203 int chunk_size_supplied
= 1;
206 * Use default chunk size (or hardsect_size, if larger) if none supplied
208 if (!ps
->snap
->chunk_size
) {
209 ps
->snap
->chunk_size
= max(DM_CHUNK_SIZE_DEFAULT_SECTORS
,
210 bdev_hardsect_size(ps
->snap
->cow
->bdev
) >> 9);
211 ps
->snap
->chunk_mask
= ps
->snap
->chunk_size
- 1;
212 ps
->snap
->chunk_shift
= ffs(ps
->snap
->chunk_size
) - 1;
213 chunk_size_supplied
= 0;
216 r
= dm_io_get(sectors_to_pages(ps
->snap
->chunk_size
));
224 r
= chunk_io(ps
, 0, READ
);
228 dh
= (struct disk_header
*) ps
->area
;
230 if (le32_to_cpu(dh
->magic
) == 0) {
235 if (le32_to_cpu(dh
->magic
) != SNAP_MAGIC
) {
236 DMWARN("Invalid or corrupt snapshot");
242 ps
->valid
= le32_to_cpu(dh
->valid
);
243 ps
->version
= le32_to_cpu(dh
->version
);
244 chunk_size
= le32_to_cpu(dh
->chunk_size
);
246 if (!chunk_size_supplied
|| ps
->snap
->chunk_size
== chunk_size
)
249 DMWARN("chunk size %llu in device metadata overrides "
250 "table chunk size of %llu.",
251 (unsigned long long)chunk_size
,
252 (unsigned long long)ps
->snap
->chunk_size
);
254 /* We had a bogus chunk_size. Fix stuff up. */
255 dm_io_put(sectors_to_pages(ps
->snap
->chunk_size
));
258 ps
->snap
->chunk_size
= chunk_size
;
259 ps
->snap
->chunk_mask
= chunk_size
- 1;
260 ps
->snap
->chunk_shift
= ffs(chunk_size
) - 1;
262 r
= dm_io_get(sectors_to_pages(chunk_size
));
275 dm_io_put(sectors_to_pages(ps
->snap
->chunk_size
));
279 static int write_header(struct pstore
*ps
)
281 struct disk_header
*dh
;
283 memset(ps
->area
, 0, ps
->snap
->chunk_size
<< SECTOR_SHIFT
);
285 dh
= (struct disk_header
*) ps
->area
;
286 dh
->magic
= cpu_to_le32(SNAP_MAGIC
);
287 dh
->valid
= cpu_to_le32(ps
->valid
);
288 dh
->version
= cpu_to_le32(ps
->version
);
289 dh
->chunk_size
= cpu_to_le32(ps
->snap
->chunk_size
);
291 return chunk_io(ps
, 0, WRITE
);
295 * Access functions for the disk exceptions, these do the endian conversions.
297 static struct disk_exception
*get_exception(struct pstore
*ps
, uint32_t index
)
299 if (index
>= ps
->exceptions_per_area
)
302 return ((struct disk_exception
*) ps
->area
) + index
;
305 static int read_exception(struct pstore
*ps
,
306 uint32_t index
, struct disk_exception
*result
)
308 struct disk_exception
*e
;
310 e
= get_exception(ps
, index
);
315 result
->old_chunk
= le64_to_cpu(e
->old_chunk
);
316 result
->new_chunk
= le64_to_cpu(e
->new_chunk
);
321 static int write_exception(struct pstore
*ps
,
322 uint32_t index
, struct disk_exception
*de
)
324 struct disk_exception
*e
;
326 e
= get_exception(ps
, index
);
331 e
->old_chunk
= cpu_to_le64(de
->old_chunk
);
332 e
->new_chunk
= cpu_to_le64(de
->new_chunk
);
338 * Registers the exceptions that are present in the current area.
339 * 'full' is filled in to indicate if the area has been
342 static int insert_exceptions(struct pstore
*ps
, int *full
)
346 struct disk_exception de
;
348 /* presume the area is full */
351 for (i
= 0; i
< ps
->exceptions_per_area
; i
++) {
352 r
= read_exception(ps
, i
, &de
);
358 * If the new_chunk is pointing at the start of
359 * the COW device, where the first metadata area
360 * is we know that we've hit the end of the
361 * exceptions. Therefore the area is not full.
363 if (de
.new_chunk
== 0LL) {
364 ps
->current_committed
= i
;
370 * Keep track of the start of the free chunks.
372 if (ps
->next_free
<= de
.new_chunk
)
373 ps
->next_free
= de
.new_chunk
+ 1;
376 * Otherwise we add the exception to the snapshot.
378 r
= dm_add_exception(ps
->snap
, de
.old_chunk
, de
.new_chunk
);
386 static int read_exceptions(struct pstore
*ps
)
392 * Keeping reading chunks and inserting exceptions until
393 * we find a partially full area.
395 for (area
= 0; full
; area
++) {
396 r
= area_io(ps
, area
, READ
);
400 r
= insert_exceptions(ps
, &full
);
408 static inline struct pstore
*get_info(struct exception_store
*store
)
410 return (struct pstore
*) store
->context
;
413 static void persistent_fraction_full(struct exception_store
*store
,
414 sector_t
*numerator
, sector_t
*denominator
)
416 *numerator
= get_info(store
)->next_free
* store
->snap
->chunk_size
;
417 *denominator
= get_dev_size(store
->snap
->cow
->bdev
);
420 static void persistent_destroy(struct exception_store
*store
)
422 struct pstore
*ps
= get_info(store
);
424 dm_io_put(sectors_to_pages(ps
->snap
->chunk_size
));
425 vfree(ps
->callbacks
);
430 static int persistent_read_metadata(struct exception_store
*store
)
433 struct pstore
*ps
= get_info(store
);
436 * Read the snapshot header.
438 r
= read_header(ps
, &new_snapshot
);
443 * Now we know correct chunk_size, complete the initialisation.
445 ps
->exceptions_per_area
= (ps
->snap
->chunk_size
<< SECTOR_SHIFT
) /
446 sizeof(struct disk_exception
);
447 ps
->callbacks
= dm_vcalloc(ps
->exceptions_per_area
,
448 sizeof(*ps
->callbacks
));
453 * Do we need to setup a new snapshot ?
456 r
= write_header(ps
);
458 DMWARN("write_header failed");
462 r
= zero_area(ps
, 0);
464 DMWARN("zero_area(0) failed");
473 DMWARN("snapshot is marked invalid");
477 if (ps
->version
!= SNAPSHOT_DISK_VERSION
) {
478 DMWARN("unable to handle snapshot disk version %d",
486 r
= read_exceptions(ps
);
494 static int persistent_prepare(struct exception_store
*store
,
497 struct pstore
*ps
= get_info(store
);
499 sector_t size
= get_dev_size(store
->snap
->cow
->bdev
);
501 /* Is there enough room ? */
502 if (size
< ((ps
->next_free
+ 1) * store
->snap
->chunk_size
))
505 e
->new_chunk
= ps
->next_free
;
508 * Move onto the next free pending, making sure to take
509 * into account the location of the metadata chunks.
511 stride
= (ps
->exceptions_per_area
+ 1);
512 if ((++ps
->next_free
% stride
) == 1)
515 atomic_inc(&ps
->pending_count
);
519 static void persistent_commit(struct exception_store
*store
,
521 void (*callback
) (void *, int success
),
522 void *callback_context
)
526 struct pstore
*ps
= get_info(store
);
527 struct disk_exception de
;
528 struct commit_callback
*cb
;
530 de
.old_chunk
= e
->old_chunk
;
531 de
.new_chunk
= e
->new_chunk
;
532 write_exception(ps
, ps
->current_committed
++, &de
);
535 * Add the callback to the back of the array. This code
536 * is the only place where the callback array is
537 * manipulated, and we know that it will never be called
538 * multiple times concurrently.
540 cb
= ps
->callbacks
+ ps
->callback_count
++;
541 cb
->callback
= callback
;
542 cb
->context
= callback_context
;
545 * If there are no more exceptions in flight, or we have
546 * filled this metadata area we commit the exceptions to
549 if (atomic_dec_and_test(&ps
->pending_count
) ||
550 (ps
->current_committed
== ps
->exceptions_per_area
)) {
551 r
= area_io(ps
, ps
->current_area
, WRITE
);
555 for (i
= 0; i
< ps
->callback_count
; i
++) {
556 cb
= ps
->callbacks
+ i
;
557 cb
->callback(cb
->context
, r
== 0 ? 1 : 0);
560 ps
->callback_count
= 0;
564 * Have we completely filled the current area ?
566 if (ps
->current_committed
== ps
->exceptions_per_area
) {
567 ps
->current_committed
= 0;
568 r
= zero_area(ps
, ps
->current_area
+ 1);
574 static void persistent_drop(struct exception_store
*store
)
576 struct pstore
*ps
= get_info(store
);
579 if (write_header(ps
))
580 DMWARN("write header failed");
583 int dm_create_persistent(struct exception_store
*store
)
587 /* allocate the pstore */
588 ps
= kmalloc(sizeof(*ps
), GFP_KERNEL
);
592 ps
->snap
= store
->snap
;
594 ps
->version
= SNAPSHOT_DISK_VERSION
;
596 ps
->next_free
= 2; /* skipping the header and first area */
597 ps
->current_committed
= 0;
599 ps
->callback_count
= 0;
600 atomic_set(&ps
->pending_count
, 0);
601 ps
->callbacks
= NULL
;
603 store
->destroy
= persistent_destroy
;
604 store
->read_metadata
= persistent_read_metadata
;
605 store
->prepare_exception
= persistent_prepare
;
606 store
->commit_exception
= persistent_commit
;
607 store
->drop_snapshot
= persistent_drop
;
608 store
->fraction_full
= persistent_fraction_full
;
614 /*-----------------------------------------------------------------
615 * Implementation of the store for non-persistent snapshots.
616 *---------------------------------------------------------------*/
621 static void transient_destroy(struct exception_store
*store
)
623 kfree(store
->context
);
626 static int transient_read_metadata(struct exception_store
*store
)
631 static int transient_prepare(struct exception_store
*store
, struct exception
*e
)
633 struct transient_c
*tc
= (struct transient_c
*) store
->context
;
634 sector_t size
= get_dev_size(store
->snap
->cow
->bdev
);
636 if (size
< (tc
->next_free
+ store
->snap
->chunk_size
))
639 e
->new_chunk
= sector_to_chunk(store
->snap
, tc
->next_free
);
640 tc
->next_free
+= store
->snap
->chunk_size
;
645 static void transient_commit(struct exception_store
*store
,
647 void (*callback
) (void *, int success
),
648 void *callback_context
)
651 callback(callback_context
, 1);
654 static void transient_fraction_full(struct exception_store
*store
,
655 sector_t
*numerator
, sector_t
*denominator
)
657 *numerator
= ((struct transient_c
*) store
->context
)->next_free
;
658 *denominator
= get_dev_size(store
->snap
->cow
->bdev
);
661 int dm_create_transient(struct exception_store
*store
)
663 struct transient_c
*tc
;
665 store
->destroy
= transient_destroy
;
666 store
->read_metadata
= transient_read_metadata
;
667 store
->prepare_exception
= transient_prepare
;
668 store
->commit_exception
= transient_commit
;
669 store
->drop_snapshot
= NULL
;
670 store
->fraction_full
= transient_fraction_full
;
672 tc
= kmalloc(sizeof(struct transient_c
), GFP_KERNEL
);