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
)
208 vfree(ps
->zero_area
);
209 ps
->zero_area
= NULL
;
212 vfree(ps
->header_area
);
213 ps
->header_area
= NULL
;
217 struct dm_io_region
*where
;
218 struct dm_io_request
*io_req
;
219 struct work_struct work
;
223 static void do_metadata(struct work_struct
*work
)
225 struct mdata_req
*req
= container_of(work
, struct mdata_req
, work
);
227 req
->result
= dm_io(req
->io_req
, 1, req
->where
, NULL
);
231 * Read or write a chunk aligned and sized block of data from a device.
233 static int chunk_io(struct pstore
*ps
, void *area
, chunk_t chunk
, int rw
,
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
,
241 struct dm_io_request io_req
= {
243 .mem
.type
= DM_IO_VMA
,
245 .client
= ps
->io_client
,
248 struct mdata_req req
;
251 return dm_io(&io_req
, 1, &where
, NULL
);
254 req
.io_req
= &io_req
;
257 * Issue the synchronous I/O from a different thread
258 * to avoid generic_make_request recursion.
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
);
269 * Convert a metadata area index to a chunk index.
271 static chunk_t
area_location(struct pstore
*ps
, chunk_t area
)
273 return NUM_SNAPSHOT_HDR_CHUNKS
+ ((ps
->exceptions_per_area
+ 1) * area
);
276 static void skip_metadata(struct pstore
*ps
)
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
)
285 * Read or write a metadata area. Remembering to skip the first
286 * chunk which holds the header.
288 static int area_io(struct pstore
*ps
, int rw
)
293 chunk
= area_location(ps
, ps
->current_area
);
295 r
= chunk_io(ps
, ps
->area
, chunk
, rw
, 0);
302 static void zero_memory_area(struct pstore
*ps
)
304 memset(ps
->area
, 0, ps
->store
->chunk_size
<< SECTOR_SHIFT
);
307 static int zero_disk_area(struct pstore
*ps
, chunk_t area
)
309 return chunk_io(ps
, ps
->zero_area
, area_location(ps
, area
), WRITE
, 0);
312 static int read_header(struct pstore
*ps
, int *new_snapshot
)
315 struct disk_header
*dh
;
317 int chunk_size_supplied
= 1;
321 * Use default chunk size (or logical_block_size, if larger)
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
)->
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;
333 ps
->io_client
= dm_io_client_create();
334 if (IS_ERR(ps
->io_client
))
335 return PTR_ERR(ps
->io_client
);
341 r
= chunk_io(ps
, ps
->header_area
, 0, READ
, 1);
345 dh
= ps
->header_area
;
347 if (le32_to_cpu(dh
->magic
) == 0) {
352 if (le32_to_cpu(dh
->magic
) != SNAP_MAGIC
) {
353 DMWARN("Invalid or corrupt snapshot");
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
);
363 if (ps
->store
->chunk_size
== chunk_size
)
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
);
371 /* We had a bogus chunk_size. Fix stuff up. */
374 r
= dm_exception_store_set_chunk_size(ps
->store
, chunk_size
,
377 DMERR("invalid on-disk chunk size %u: %s.",
378 chunk_size
, chunk_err
);
390 static int write_header(struct pstore
*ps
)
392 struct disk_header
*dh
;
394 memset(ps
->header_area
, 0, ps
->store
->chunk_size
<< SECTOR_SHIFT
);
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
);
402 return chunk_io(ps
, ps
->header_area
, 0, WRITE
, 1);
406 * Access functions for the disk exceptions, these do the endian conversions.
408 static struct disk_exception
*get_exception(struct pstore
*ps
, void *ps_area
,
411 BUG_ON(index
>= ps
->exceptions_per_area
);
413 return ((struct disk_exception
*) ps_area
) + index
;
416 static void read_exception(struct pstore
*ps
, void *ps_area
,
417 uint32_t index
, struct core_exception
*result
)
419 struct disk_exception
*de
= get_exception(ps
, ps_area
, index
);
422 result
->old_chunk
= le64_to_cpu(de
->old_chunk
);
423 result
->new_chunk
= le64_to_cpu(de
->new_chunk
);
426 static void write_exception(struct pstore
*ps
,
427 uint32_t index
, struct core_exception
*e
)
429 struct disk_exception
*de
= get_exception(ps
, ps
->area
, index
);
432 de
->old_chunk
= cpu_to_le64(e
->old_chunk
);
433 de
->new_chunk
= cpu_to_le64(e
->new_chunk
);
436 static void clear_exception(struct pstore
*ps
, uint32_t index
)
438 struct disk_exception
*de
= get_exception(ps
, ps
->area
, index
);
446 * Registers the exceptions that are present in the current area.
447 * 'full' is filled in to indicate if the area has been
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
,
458 struct core_exception e
;
460 /* presume the area is full */
463 for (i
= 0; i
< ps
->exceptions_per_area
; i
++) {
464 read_exception(ps
, ps_area
, i
, &e
);
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.
472 if (e
.new_chunk
== 0LL) {
473 ps
->current_committed
= i
;
479 * Keep track of the start of the free chunks.
481 if (ps
->next_free
<= e
.new_chunk
)
482 ps
->next_free
= e
.new_chunk
+ 1;
485 * Otherwise we add the exception to the snapshot.
487 r
= callback(callback_context
, e
.old_chunk
, e
.new_chunk
);
495 static int read_exceptions(struct pstore
*ps
,
496 int (*callback
)(void *callback_context
, chunk_t old
,
498 void *callback_context
)
501 struct dm_bufio_client
*client
;
502 chunk_t prefetch_area
= 0;
504 client
= dm_bufio_client_create(dm_snap_cow(ps
->store
->snap
)->bdev
,
505 ps
->store
->chunk_size
<< SECTOR_SHIFT
,
509 return PTR_ERR(client
);
512 * Setup for one current buffer + desired readahead buffers.
514 dm_bufio_set_minimum_buffers(client
, 1 + DM_PREFETCH_CHUNKS
);
517 * Keeping reading chunks and inserting exceptions until
518 * we find a partially full area.
520 for (ps
->current_area
= 0; full
; ps
->current_area
++) {
521 struct dm_buffer
*bp
;
525 if (unlikely(prefetch_area
< ps
->current_area
))
526 prefetch_area
= ps
->current_area
;
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
)))
532 dm_bufio_prefetch(client
, pf_chunk
, 1);
534 if (unlikely(!prefetch_area
))
536 } while (prefetch_area
<= ps
->current_area
+ DM_PREFETCH_CHUNKS
);
538 chunk
= area_location(ps
, ps
->current_area
);
540 area
= dm_bufio_read(client
, chunk
, &bp
);
541 if (unlikely(IS_ERR(area
))) {
543 goto ret_destroy_bufio
;
546 r
= insert_exceptions(ps
, area
, callback
, callback_context
,
549 dm_bufio_release(bp
);
551 dm_bufio_forget(client
, chunk
);
554 goto ret_destroy_bufio
;
564 dm_bufio_client_destroy(client
);
569 static struct pstore
*get_info(struct dm_exception_store
*store
)
571 return (struct pstore
*) store
->context
;
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
)
579 struct pstore
*ps
= get_info(store
);
581 *sectors_allocated
= ps
->next_free
* store
->chunk_size
;
582 *total_sectors
= get_dev_size(dm_snap_cow(store
->snap
)->bdev
);
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.
589 *metadata_sectors
= (ps
->current_area
+ 1 + NUM_SNAPSHOT_HDR_CHUNKS
) *
593 static void persistent_dtr(struct dm_exception_store
*store
)
595 struct pstore
*ps
= get_info(store
);
597 destroy_workqueue(ps
->metadata_wq
);
599 /* Created in read_header */
601 dm_io_client_destroy(ps
->io_client
);
604 /* Allocated in persistent_read_metadata */
606 vfree(ps
->callbacks
);
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
)
616 int r
, uninitialized_var(new_snapshot
);
617 struct pstore
*ps
= get_info(store
);
620 * Read the snapshot header.
622 r
= read_header(ps
, &new_snapshot
);
627 * Now we know correct chunk_size, complete the initialisation.
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
));
637 * Do we need to setup a new snapshot ?
640 r
= write_header(ps
);
642 DMWARN("write_header failed");
646 ps
->current_area
= 0;
647 zero_memory_area(ps
);
648 r
= zero_disk_area(ps
, 0);
650 DMWARN("zero_disk_area(0) failed");
656 if (ps
->version
!= SNAPSHOT_DISK_VERSION
) {
657 DMWARN("unable to handle snapshot disk version %d",
663 * Metadata are valid, but snapshot is invalidated
671 r
= read_exceptions(ps
, callback
, callback_context
);
676 static int persistent_prepare_exception(struct dm_exception_store
*store
,
677 struct dm_exception
*e
)
679 struct pstore
*ps
= get_info(store
);
680 sector_t size
= get_dev_size(dm_snap_cow(store
->snap
)->bdev
);
682 /* Is there enough room ? */
683 if (size
< ((ps
->next_free
+ 1) * store
->chunk_size
))
686 e
->new_chunk
= ps
->next_free
;
689 * Move onto the next free pending, making sure to take
690 * into account the location of the metadata chunks.
695 atomic_inc(&ps
->pending_count
);
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
)
705 struct pstore
*ps
= get_info(store
);
706 struct core_exception ce
;
707 struct commit_callback
*cb
;
709 ce
.old_chunk
= e
->old_chunk
;
710 ce
.new_chunk
= e
->new_chunk
;
711 write_exception(ps
, ps
->current_committed
++, &ce
);
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.
719 cb
= ps
->callbacks
+ ps
->callback_count
++;
720 cb
->callback
= callback
;
721 cb
->context
= callback_context
;
724 * If there are exceptions in flight and we have not yet
725 * filled this metadata area there's nothing more to do.
727 if (!atomic_dec_and_test(&ps
->pending_count
) &&
728 (ps
->current_committed
!= ps
->exceptions_per_area
))
732 * If we completely filled the current area, then wipe the next one.
734 if ((ps
->current_committed
== ps
->exceptions_per_area
) &&
735 zero_disk_area(ps
, ps
->current_area
+ 1))
739 * Commit exceptions to disk.
741 if (ps
->valid
&& area_io(ps
, WRITE_FLUSH_FUA
))
745 * Advance to the next area if this one is full.
747 if (ps
->current_committed
== ps
->exceptions_per_area
) {
748 ps
->current_committed
= 0;
750 zero_memory_area(ps
);
753 for (i
= 0; i
< ps
->callback_count
; i
++) {
754 cb
= ps
->callbacks
+ i
;
755 cb
->callback(cb
->context
, ps
->valid
);
758 ps
->callback_count
= 0;
761 static int persistent_prepare_merge(struct dm_exception_store
*store
,
762 chunk_t
*last_old_chunk
,
763 chunk_t
*last_new_chunk
)
765 struct pstore
*ps
= get_info(store
);
766 struct core_exception ce
;
771 * When current area is empty, move back to preceding area.
773 if (!ps
->current_committed
) {
777 if (!ps
->current_area
)
781 r
= area_io(ps
, READ
);
784 ps
->current_committed
= ps
->exceptions_per_area
;
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
;
792 * Find number of consecutive chunks within the current area,
795 for (nr_consecutive
= 1; nr_consecutive
< ps
->current_committed
;
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
)
804 return nr_consecutive
;
807 static int persistent_commit_merge(struct dm_exception_store
*store
,
811 struct pstore
*ps
= get_info(store
);
813 BUG_ON(nr_merged
> ps
->current_committed
);
815 for (i
= 0; i
< nr_merged
; i
++)
816 clear_exception(ps
, ps
->current_committed
- 1 - i
);
818 r
= area_io(ps
, WRITE_FLUSH_FUA
);
822 ps
->current_committed
-= nr_merged
;
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.
831 * ps->current_area does not get reduced by prepare_merge() until
832 * after commit_merge() has removed the nr_merged previous exceptions.
834 ps
->next_free
= area_location(ps
, ps
->current_area
) +
835 ps
->current_committed
+ 1;
840 static void persistent_drop_snapshot(struct dm_exception_store
*store
)
842 struct pstore
*ps
= get_info(store
);
845 if (write_header(ps
))
846 DMWARN("write header failed");
849 static int persistent_ctr(struct dm_exception_store
*store
,
850 unsigned argc
, char **argv
)
854 /* allocate the pstore */
855 ps
= kzalloc(sizeof(*ps
), GFP_KERNEL
);
861 ps
->version
= SNAPSHOT_DISK_VERSION
;
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;
868 ps
->callback_count
= 0;
869 atomic_set(&ps
->pending_count
, 0);
870 ps
->callbacks
= NULL
;
872 ps
->metadata_wq
= alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM
, 0);
873 if (!ps
->metadata_wq
) {
875 DMERR("couldn't start header metadata update thread");
884 static unsigned persistent_status(struct dm_exception_store
*store
,
885 status_type_t status
, char *result
,
891 case STATUSTYPE_INFO
:
893 case STATUSTYPE_TABLE
:
894 DMEMIT(" P %llu", (unsigned long long)store
->chunk_size
);
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
,
915 static struct dm_exception_store_type _persistent_compat_type
= {
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
,
930 int dm_persistent_snapshot_init(void)
934 r
= dm_exception_store_type_register(&_persistent_type
);
936 DMERR("Unable to register persistent exception store type");
940 r
= dm_exception_store_type_register(&_persistent_compat_type
);
942 DMERR("Unable to register old-style persistent exception "
944 dm_exception_store_type_unregister(&_persistent_type
);
951 void dm_persistent_snapshot_exit(void)
953 dm_exception_store_type_unregister(&_persistent_type
);
954 dm_exception_store_type_unregister(&_persistent_compat_type
);