2 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
3 * Copyright (C) 2006-2008 Red Hat GmbH
5 * This file is released under the GPL.
8 #include "dm-exception-store.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>
18 #define DM_MSG_PREFIX "persistent snapshot"
19 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
21 #define DM_PREFETCH_CHUNKS 12
23 /*-----------------------------------------------------------------
24 * Persistent snapshots, by persistent we mean that the snapshot
25 * will survive a reboot.
26 *---------------------------------------------------------------*/
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.
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
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
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
53 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
55 #define SNAP_MAGIC 0x70416e53
58 * The on-disk version of the metadata.
60 #define SNAPSHOT_DISK_VERSION 1
62 #define NUM_SNAPSHOT_HDR_CHUNKS 1
68 * Is this snapshot valid. There is no way of recovering
69 * an invalid snapshot.
74 * Simple, incrementing version. no backward
83 struct disk_exception
{
88 struct core_exception
{
93 struct commit_callback
{
94 void (*callback
)(void *, int success
);
99 * The top level structure for a persistent exception store.
102 struct dm_exception_store
*store
;
105 uint32_t exceptions_per_area
;
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.
115 * An area of zeros used to clear the next area.
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.
127 * Used to keep track of which metadata area the data in
130 chunk_t current_area
;
133 * The next free chunk for an exception.
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
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
153 * The index of next free exception in the current
156 uint32_t current_committed
;
158 atomic_t pending_count
;
159 uint32_t callback_count
;
160 struct commit_callback
*callbacks
;
161 struct dm_io_client
*io_client
;
163 struct workqueue_struct
*metadata_wq
;
166 static int alloc_area(struct pstore
*ps
)
171 len
= ps
->store
->chunk_size
<< SECTOR_SHIFT
;
174 * Allocate the chunk_size block of memory that will hold
175 * a single metadata area.
177 ps
->area
= vmalloc(len
);
181 ps
->zero_area
= vzalloc(len
);
185 ps
->header_area
= vmalloc(len
);
186 if (!ps
->header_area
)
187 goto err_header_area
;
192 vfree(ps
->zero_area
);
201 static void free_area(struct pstore
*ps
)
205 vfree(ps
->zero_area
);
206 ps
->zero_area
= NULL
;
207 vfree(ps
->header_area
);
208 ps
->header_area
= NULL
;
212 struct dm_io_region
*where
;
213 struct dm_io_request
*io_req
;
214 struct work_struct work
;
218 static void do_metadata(struct work_struct
*work
)
220 struct mdata_req
*req
= container_of(work
, struct mdata_req
, work
);
222 req
->result
= dm_io(req
->io_req
, 1, req
->where
, NULL
);
226 * Read or write a chunk aligned and sized block of data from a device.
228 static int chunk_io(struct pstore
*ps
, void *area
, chunk_t chunk
, int rw
,
231 struct dm_io_region where
= {
232 .bdev
= dm_snap_cow(ps
->store
->snap
)->bdev
,
233 .sector
= ps
->store
->chunk_size
* chunk
,
234 .count
= ps
->store
->chunk_size
,
236 struct dm_io_request io_req
= {
238 .mem
.type
= DM_IO_VMA
,
240 .client
= ps
->io_client
,
243 struct mdata_req req
;
246 return dm_io(&io_req
, 1, &where
, NULL
);
249 req
.io_req
= &io_req
;
252 * Issue the synchronous I/O from a different thread
253 * to avoid generic_make_request recursion.
255 INIT_WORK_ONSTACK(&req
.work
, do_metadata
);
256 queue_work(ps
->metadata_wq
, &req
.work
);
257 flush_workqueue(ps
->metadata_wq
);
258 destroy_work_on_stack(&req
.work
);
264 * Convert a metadata area index to a chunk index.
266 static chunk_t
area_location(struct pstore
*ps
, chunk_t area
)
268 return NUM_SNAPSHOT_HDR_CHUNKS
+ ((ps
->exceptions_per_area
+ 1) * area
);
271 static void skip_metadata(struct pstore
*ps
)
273 uint32_t stride
= ps
->exceptions_per_area
+ 1;
274 chunk_t next_free
= ps
->next_free
;
275 if (sector_div(next_free
, stride
) == NUM_SNAPSHOT_HDR_CHUNKS
)
280 * Read or write a metadata area. Remembering to skip the first
281 * chunk which holds the header.
283 static int area_io(struct pstore
*ps
, int rw
)
288 chunk
= area_location(ps
, ps
->current_area
);
290 r
= chunk_io(ps
, ps
->area
, chunk
, rw
, 0);
297 static void zero_memory_area(struct pstore
*ps
)
299 memset(ps
->area
, 0, ps
->store
->chunk_size
<< SECTOR_SHIFT
);
302 static int zero_disk_area(struct pstore
*ps
, chunk_t area
)
304 return chunk_io(ps
, ps
->zero_area
, area_location(ps
, area
), WRITE
, 0);
307 static int read_header(struct pstore
*ps
, int *new_snapshot
)
310 struct disk_header
*dh
;
312 int chunk_size_supplied
= 1;
316 * Use default chunk size (or logical_block_size, if larger)
319 if (!ps
->store
->chunk_size
) {
320 ps
->store
->chunk_size
= max(DM_CHUNK_SIZE_DEFAULT_SECTORS
,
321 bdev_logical_block_size(dm_snap_cow(ps
->store
->snap
)->
323 ps
->store
->chunk_mask
= ps
->store
->chunk_size
- 1;
324 ps
->store
->chunk_shift
= ffs(ps
->store
->chunk_size
) - 1;
325 chunk_size_supplied
= 0;
328 ps
->io_client
= dm_io_client_create();
329 if (IS_ERR(ps
->io_client
))
330 return PTR_ERR(ps
->io_client
);
336 r
= chunk_io(ps
, ps
->header_area
, 0, READ
, 1);
340 dh
= ps
->header_area
;
342 if (le32_to_cpu(dh
->magic
) == 0) {
347 if (le32_to_cpu(dh
->magic
) != SNAP_MAGIC
) {
348 DMWARN("Invalid or corrupt snapshot");
354 ps
->valid
= le32_to_cpu(dh
->valid
);
355 ps
->version
= le32_to_cpu(dh
->version
);
356 chunk_size
= le32_to_cpu(dh
->chunk_size
);
358 if (ps
->store
->chunk_size
== chunk_size
)
361 if (chunk_size_supplied
)
362 DMWARN("chunk size %u in device metadata overrides "
363 "table chunk size of %u.",
364 chunk_size
, ps
->store
->chunk_size
);
366 /* We had a bogus chunk_size. Fix stuff up. */
369 r
= dm_exception_store_set_chunk_size(ps
->store
, chunk_size
,
372 DMERR("invalid on-disk chunk size %u: %s.",
373 chunk_size
, chunk_err
);
385 static int write_header(struct pstore
*ps
)
387 struct disk_header
*dh
;
389 memset(ps
->header_area
, 0, ps
->store
->chunk_size
<< SECTOR_SHIFT
);
391 dh
= ps
->header_area
;
392 dh
->magic
= cpu_to_le32(SNAP_MAGIC
);
393 dh
->valid
= cpu_to_le32(ps
->valid
);
394 dh
->version
= cpu_to_le32(ps
->version
);
395 dh
->chunk_size
= cpu_to_le32(ps
->store
->chunk_size
);
397 return chunk_io(ps
, ps
->header_area
, 0, WRITE
, 1);
401 * Access functions for the disk exceptions, these do the endian conversions.
403 static struct disk_exception
*get_exception(struct pstore
*ps
, void *ps_area
,
406 BUG_ON(index
>= ps
->exceptions_per_area
);
408 return ((struct disk_exception
*) ps_area
) + index
;
411 static void read_exception(struct pstore
*ps
, void *ps_area
,
412 uint32_t index
, struct core_exception
*result
)
414 struct disk_exception
*de
= get_exception(ps
, ps_area
, index
);
417 result
->old_chunk
= le64_to_cpu(de
->old_chunk
);
418 result
->new_chunk
= le64_to_cpu(de
->new_chunk
);
421 static void write_exception(struct pstore
*ps
,
422 uint32_t index
, struct core_exception
*e
)
424 struct disk_exception
*de
= get_exception(ps
, ps
->area
, index
);
427 de
->old_chunk
= cpu_to_le64(e
->old_chunk
);
428 de
->new_chunk
= cpu_to_le64(e
->new_chunk
);
431 static void clear_exception(struct pstore
*ps
, uint32_t index
)
433 struct disk_exception
*de
= get_exception(ps
, ps
->area
, index
);
441 * Registers the exceptions that are present in the current area.
442 * 'full' is filled in to indicate if the area has been
445 static int insert_exceptions(struct pstore
*ps
, void *ps_area
,
446 int (*callback
)(void *callback_context
,
447 chunk_t old
, chunk_t
new),
448 void *callback_context
,
453 struct core_exception e
;
455 /* presume the area is full */
458 for (i
= 0; i
< ps
->exceptions_per_area
; i
++) {
459 read_exception(ps
, ps_area
, i
, &e
);
462 * If the new_chunk is pointing at the start of
463 * the COW device, where the first metadata area
464 * is we know that we've hit the end of the
465 * exceptions. Therefore the area is not full.
467 if (e
.new_chunk
== 0LL) {
468 ps
->current_committed
= i
;
474 * Keep track of the start of the free chunks.
476 if (ps
->next_free
<= e
.new_chunk
)
477 ps
->next_free
= e
.new_chunk
+ 1;
480 * Otherwise we add the exception to the snapshot.
482 r
= callback(callback_context
, e
.old_chunk
, e
.new_chunk
);
490 static int read_exceptions(struct pstore
*ps
,
491 int (*callback
)(void *callback_context
, chunk_t old
,
493 void *callback_context
)
496 struct dm_bufio_client
*client
;
497 chunk_t prefetch_area
= 0;
499 client
= dm_bufio_client_create(dm_snap_cow(ps
->store
->snap
)->bdev
,
500 ps
->store
->chunk_size
<< SECTOR_SHIFT
,
504 return PTR_ERR(client
);
507 * Setup for one current buffer + desired readahead buffers.
509 dm_bufio_set_minimum_buffers(client
, 1 + DM_PREFETCH_CHUNKS
);
512 * Keeping reading chunks and inserting exceptions until
513 * we find a partially full area.
515 for (ps
->current_area
= 0; full
; ps
->current_area
++) {
516 struct dm_buffer
*bp
;
520 if (unlikely(prefetch_area
< ps
->current_area
))
521 prefetch_area
= ps
->current_area
;
523 if (DM_PREFETCH_CHUNKS
) do {
524 chunk_t pf_chunk
= area_location(ps
, prefetch_area
);
525 if (unlikely(pf_chunk
>= dm_bufio_get_device_size(client
)))
527 dm_bufio_prefetch(client
, pf_chunk
, 1);
529 if (unlikely(!prefetch_area
))
531 } while (prefetch_area
<= ps
->current_area
+ DM_PREFETCH_CHUNKS
);
533 chunk
= area_location(ps
, ps
->current_area
);
535 area
= dm_bufio_read(client
, chunk
, &bp
);
536 if (unlikely(IS_ERR(area
))) {
538 goto ret_destroy_bufio
;
541 r
= insert_exceptions(ps
, area
, callback
, callback_context
,
545 memcpy(ps
->area
, area
, ps
->store
->chunk_size
<< SECTOR_SHIFT
);
547 dm_bufio_release(bp
);
549 dm_bufio_forget(client
, chunk
);
552 goto ret_destroy_bufio
;
562 dm_bufio_client_destroy(client
);
567 static struct pstore
*get_info(struct dm_exception_store
*store
)
569 return (struct pstore
*) store
->context
;
572 static void persistent_usage(struct dm_exception_store
*store
,
573 sector_t
*total_sectors
,
574 sector_t
*sectors_allocated
,
575 sector_t
*metadata_sectors
)
577 struct pstore
*ps
= get_info(store
);
579 *sectors_allocated
= ps
->next_free
* store
->chunk_size
;
580 *total_sectors
= get_dev_size(dm_snap_cow(store
->snap
)->bdev
);
583 * First chunk is the fixed header.
584 * Then there are (ps->current_area + 1) metadata chunks, each one
585 * separated from the next by ps->exceptions_per_area data chunks.
587 *metadata_sectors
= (ps
->current_area
+ 1 + NUM_SNAPSHOT_HDR_CHUNKS
) *
591 static void persistent_dtr(struct dm_exception_store
*store
)
593 struct pstore
*ps
= get_info(store
);
595 destroy_workqueue(ps
->metadata_wq
);
597 /* Created in read_header */
599 dm_io_client_destroy(ps
->io_client
);
602 /* Allocated in persistent_read_metadata */
603 vfree(ps
->callbacks
);
608 static int persistent_read_metadata(struct dm_exception_store
*store
,
609 int (*callback
)(void *callback_context
,
610 chunk_t old
, chunk_t
new),
611 void *callback_context
)
613 int r
, uninitialized_var(new_snapshot
);
614 struct pstore
*ps
= get_info(store
);
617 * Read the snapshot header.
619 r
= read_header(ps
, &new_snapshot
);
624 * Now we know correct chunk_size, complete the initialisation.
626 ps
->exceptions_per_area
= (ps
->store
->chunk_size
<< SECTOR_SHIFT
) /
627 sizeof(struct disk_exception
);
628 ps
->callbacks
= dm_vcalloc(ps
->exceptions_per_area
,
629 sizeof(*ps
->callbacks
));
634 * Do we need to setup a new snapshot ?
637 r
= write_header(ps
);
639 DMWARN("write_header failed");
643 ps
->current_area
= 0;
644 zero_memory_area(ps
);
645 r
= zero_disk_area(ps
, 0);
647 DMWARN("zero_disk_area(0) failed");
653 if (ps
->version
!= SNAPSHOT_DISK_VERSION
) {
654 DMWARN("unable to handle snapshot disk version %d",
660 * Metadata are valid, but snapshot is invalidated
668 r
= read_exceptions(ps
, callback
, callback_context
);
673 static int persistent_prepare_exception(struct dm_exception_store
*store
,
674 struct dm_exception
*e
)
676 struct pstore
*ps
= get_info(store
);
677 sector_t size
= get_dev_size(dm_snap_cow(store
->snap
)->bdev
);
679 /* Is there enough room ? */
680 if (size
< ((ps
->next_free
+ 1) * store
->chunk_size
))
683 e
->new_chunk
= ps
->next_free
;
686 * Move onto the next free pending, making sure to take
687 * into account the location of the metadata chunks.
692 atomic_inc(&ps
->pending_count
);
696 static void persistent_commit_exception(struct dm_exception_store
*store
,
697 struct dm_exception
*e
,
698 void (*callback
) (void *, int success
),
699 void *callback_context
)
702 struct pstore
*ps
= get_info(store
);
703 struct core_exception ce
;
704 struct commit_callback
*cb
;
706 ce
.old_chunk
= e
->old_chunk
;
707 ce
.new_chunk
= e
->new_chunk
;
708 write_exception(ps
, ps
->current_committed
++, &ce
);
711 * Add the callback to the back of the array. This code
712 * is the only place where the callback array is
713 * manipulated, and we know that it will never be called
714 * multiple times concurrently.
716 cb
= ps
->callbacks
+ ps
->callback_count
++;
717 cb
->callback
= callback
;
718 cb
->context
= callback_context
;
721 * If there are exceptions in flight and we have not yet
722 * filled this metadata area there's nothing more to do.
724 if (!atomic_dec_and_test(&ps
->pending_count
) &&
725 (ps
->current_committed
!= ps
->exceptions_per_area
))
729 * If we completely filled the current area, then wipe the next one.
731 if ((ps
->current_committed
== ps
->exceptions_per_area
) &&
732 zero_disk_area(ps
, ps
->current_area
+ 1))
736 * Commit exceptions to disk.
738 if (ps
->valid
&& area_io(ps
, WRITE_FLUSH_FUA
))
742 * Advance to the next area if this one is full.
744 if (ps
->current_committed
== ps
->exceptions_per_area
) {
745 ps
->current_committed
= 0;
747 zero_memory_area(ps
);
750 for (i
= 0; i
< ps
->callback_count
; i
++) {
751 cb
= ps
->callbacks
+ i
;
752 cb
->callback(cb
->context
, ps
->valid
);
755 ps
->callback_count
= 0;
758 static int persistent_prepare_merge(struct dm_exception_store
*store
,
759 chunk_t
*last_old_chunk
,
760 chunk_t
*last_new_chunk
)
762 struct pstore
*ps
= get_info(store
);
763 struct core_exception ce
;
768 * When current area is empty, move back to preceding area.
770 if (!ps
->current_committed
) {
774 if (!ps
->current_area
)
778 r
= area_io(ps
, READ
);
781 ps
->current_committed
= ps
->exceptions_per_area
;
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
;
789 * Find number of consecutive chunks within the current area,
792 for (nr_consecutive
= 1; nr_consecutive
< ps
->current_committed
;
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
)
801 return nr_consecutive
;
804 static int persistent_commit_merge(struct dm_exception_store
*store
,
808 struct pstore
*ps
= get_info(store
);
810 BUG_ON(nr_merged
> ps
->current_committed
);
812 for (i
= 0; i
< nr_merged
; i
++)
813 clear_exception(ps
, ps
->current_committed
- 1 - i
);
815 r
= area_io(ps
, WRITE_FLUSH_FUA
);
819 ps
->current_committed
-= nr_merged
;
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.
828 * ps->current_area does not get reduced by prepare_merge() until
829 * after commit_merge() has removed the nr_merged previous exceptions.
831 ps
->next_free
= area_location(ps
, ps
->current_area
) +
832 ps
->current_committed
+ 1;
837 static void persistent_drop_snapshot(struct dm_exception_store
*store
)
839 struct pstore
*ps
= get_info(store
);
842 if (write_header(ps
))
843 DMWARN("write header failed");
846 static int persistent_ctr(struct dm_exception_store
*store
,
847 unsigned argc
, char **argv
)
851 /* allocate the pstore */
852 ps
= kzalloc(sizeof(*ps
), GFP_KERNEL
);
858 ps
->version
= SNAPSHOT_DISK_VERSION
;
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;
865 ps
->callback_count
= 0;
866 atomic_set(&ps
->pending_count
, 0);
867 ps
->callbacks
= NULL
;
869 ps
->metadata_wq
= alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM
, 0);
870 if (!ps
->metadata_wq
) {
872 DMERR("couldn't start header metadata update thread");
881 static unsigned persistent_status(struct dm_exception_store
*store
,
882 status_type_t status
, char *result
,
888 case STATUSTYPE_INFO
:
890 case STATUSTYPE_TABLE
:
891 DMEMIT(" P %llu", (unsigned long long)store
->chunk_size
);
897 static struct dm_exception_store_type _persistent_type
= {
898 .name
= "persistent",
899 .module
= THIS_MODULE
,
900 .ctr
= persistent_ctr
,
901 .dtr
= persistent_dtr
,
902 .read_metadata
= persistent_read_metadata
,
903 .prepare_exception
= persistent_prepare_exception
,
904 .commit_exception
= persistent_commit_exception
,
905 .prepare_merge
= persistent_prepare_merge
,
906 .commit_merge
= persistent_commit_merge
,
907 .drop_snapshot
= persistent_drop_snapshot
,
908 .usage
= persistent_usage
,
909 .status
= persistent_status
,
912 static struct dm_exception_store_type _persistent_compat_type
= {
914 .module
= THIS_MODULE
,
915 .ctr
= persistent_ctr
,
916 .dtr
= persistent_dtr
,
917 .read_metadata
= persistent_read_metadata
,
918 .prepare_exception
= persistent_prepare_exception
,
919 .commit_exception
= persistent_commit_exception
,
920 .prepare_merge
= persistent_prepare_merge
,
921 .commit_merge
= persistent_commit_merge
,
922 .drop_snapshot
= persistent_drop_snapshot
,
923 .usage
= persistent_usage
,
924 .status
= persistent_status
,
927 int dm_persistent_snapshot_init(void)
931 r
= dm_exception_store_type_register(&_persistent_type
);
933 DMERR("Unable to register persistent exception store type");
937 r
= dm_exception_store_type_register(&_persistent_compat_type
);
939 DMERR("Unable to register old-style persistent exception "
941 dm_exception_store_type_unregister(&_persistent_type
);
948 void dm_persistent_snapshot_exit(void)
950 dm_exception_store_type_unregister(&_persistent_type
);
951 dm_exception_store_type_unregister(&_persistent_compat_type
);