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/slab.h>
14 #include <linux/dm-io.h>
16 #define DM_MSG_PREFIX "persistent snapshot"
17 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
19 /*-----------------------------------------------------------------
20 * Persistent snapshots, by persistent we mean that the snapshot
21 * will survive a reboot.
22 *---------------------------------------------------------------*/
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.
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
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
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
49 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
51 #define SNAP_MAGIC 0x70416e53
54 * The on-disk version of the metadata.
56 #define SNAPSHOT_DISK_VERSION 1
58 #define NUM_SNAPSHOT_HDR_CHUNKS 1
64 * Is this snapshot valid. There is no way of recovering
65 * an invalid snapshot.
70 * Simple, incrementing version. no backward
79 struct disk_exception {
84 struct commit_callback {
85 void (*callback)(void *, int success);
90 * The top level structure for a persistent exception store.
93 struct dm_exception_store *store;
96 uint32_t exceptions_per_area;
99 * Now that we have an asynchronous kcopyd there is no
100 * need for large chunk sizes, so it wont hurt to have a
101 * whole chunks worth of metadata in memory at once.
106 * An area of zeros used to clear the next area.
111 * An area used for header. The header can be written
112 * concurrently with metadata (when invalidating the snapshot),
113 * so it needs a separate buffer.
118 * Used to keep track of which metadata area the data in
121 chunk_t current_area;
124 * The next free chunk for an exception.
126 * When creating exceptions, all the chunks here and above are
127 * free. It holds the next chunk to be allocated. On rare
128 * occasions (e.g. after a system crash) holes can be left in
129 * the exception store because chunks can be committed out of
132 * When merging exceptions, it does not necessarily mean all the
133 * chunks here and above are free. It holds the value it would
134 * have held if all chunks had been committed in order of
135 * allocation. Consequently the value may occasionally be
136 * slightly too low, but since it's only used for 'status' and
137 * it can never reach its minimum value too early this doesn't
144 * The index of next free exception in the current
147 uint32_t current_committed;
149 atomic_t pending_count;
150 uint32_t callback_count;
151 struct commit_callback *callbacks;
152 struct dm_io_client *io_client;
154 struct workqueue_struct *metadata_wq;
157 static int alloc_area(struct pstore *ps)
162 len = ps->store->chunk_size << SECTOR_SHIFT;
165 * Allocate the chunk_size block of memory that will hold
166 * a single metadata area.
168 ps->area = vmalloc(len);
172 ps->zero_area = vmalloc(len);
175 memset(ps->zero_area, 0, len);
177 ps->header_area = vmalloc(len);
178 if (!ps->header_area)
179 goto err_header_area;
184 vfree(ps->zero_area);
193 static void free_area(struct pstore *ps)
200 vfree(ps->zero_area);
201 ps->zero_area = NULL;
204 vfree(ps->header_area);
205 ps->header_area = NULL;
209 struct dm_io_region *where;
210 struct dm_io_request *io_req;
211 struct work_struct work;
215 static void do_metadata(struct work_struct *work)
217 struct mdata_req *req = container_of(work, struct mdata_req, work);
219 req->result = dm_io(req->io_req, 1, req->where, NULL);
223 * Read or write a chunk aligned and sized block of data from a device.
225 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, int rw,
228 struct dm_io_region where = {
229 .bdev = dm_snap_cow(ps->store->snap)->bdev,
230 .sector = ps->store->chunk_size * chunk,
231 .count = ps->store->chunk_size,
233 struct dm_io_request io_req = {
235 .mem.type = DM_IO_VMA,
237 .client = ps->io_client,
240 struct mdata_req req;
243 return dm_io(&io_req, 1, &where, NULL);
246 req.io_req = &io_req;
249 * Issue the synchronous I/O from a different thread
250 * to avoid generic_make_request recursion.
252 INIT_WORK_ONSTACK(&req.work, do_metadata);
253 queue_work(ps->metadata_wq, &req.work);
254 flush_work(&req.work);
260 * Convert a metadata area index to a chunk index.
262 static chunk_t area_location(struct pstore *ps, chunk_t area)
264 return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
268 * Read or write a metadata area. Remembering to skip the first
269 * chunk which holds the header.
271 static int area_io(struct pstore *ps, int rw)
276 chunk = area_location(ps, ps->current_area);
278 r = chunk_io(ps, ps->area, chunk, rw, 0);
285 static void zero_memory_area(struct pstore *ps)
287 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
290 static int zero_disk_area(struct pstore *ps, chunk_t area)
292 return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
295 static int read_header(struct pstore *ps, int *new_snapshot)
298 struct disk_header *dh;
300 int chunk_size_supplied = 1;
304 * Use default chunk size (or logical_block_size, if larger)
307 if (!ps->store->chunk_size) {
308 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
309 bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
311 ps->store->chunk_mask = ps->store->chunk_size - 1;
312 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
313 chunk_size_supplied = 0;
316 ps->io_client = dm_io_client_create();
317 if (IS_ERR(ps->io_client))
318 return PTR_ERR(ps->io_client);
324 r = chunk_io(ps, ps->header_area, 0, READ, 1);
328 dh = ps->header_area;
330 if (le32_to_cpu(dh->magic) == 0) {
335 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
336 DMWARN("Invalid or corrupt snapshot");
342 ps->valid = le32_to_cpu(dh->valid);
343 ps->version = le32_to_cpu(dh->version);
344 chunk_size = le32_to_cpu(dh->chunk_size);
346 if (ps->store->chunk_size == chunk_size)
349 if (chunk_size_supplied)
350 DMWARN("chunk size %u in device metadata overrides "
351 "table chunk size of %u.",
352 chunk_size, ps->store->chunk_size);
354 /* We had a bogus chunk_size. Fix stuff up. */
357 r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
360 DMERR("invalid on-disk chunk size %u: %s.",
361 chunk_size, chunk_err);
373 static int write_header(struct pstore *ps)
375 struct disk_header *dh;
377 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
379 dh = ps->header_area;
380 dh->magic = cpu_to_le32(SNAP_MAGIC);
381 dh->valid = cpu_to_le32(ps->valid);
382 dh->version = cpu_to_le32(ps->version);
383 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
385 return chunk_io(ps, ps->header_area, 0, WRITE, 1);
389 * Access functions for the disk exceptions, these do the endian conversions.
391 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
393 BUG_ON(index >= ps->exceptions_per_area);
395 return ((struct disk_exception *) ps->area) + index;
398 static void read_exception(struct pstore *ps,
399 uint32_t index, struct disk_exception *result)
401 struct disk_exception *e = get_exception(ps, index);
404 result->old_chunk = le64_to_cpu(e->old_chunk);
405 result->new_chunk = le64_to_cpu(e->new_chunk);
408 static void write_exception(struct pstore *ps,
409 uint32_t index, struct disk_exception *de)
411 struct disk_exception *e = get_exception(ps, index);
414 e->old_chunk = cpu_to_le64(de->old_chunk);
415 e->new_chunk = cpu_to_le64(de->new_chunk);
418 static void clear_exception(struct pstore *ps, uint32_t index)
420 struct disk_exception *e = get_exception(ps, index);
428 * Registers the exceptions that are present in the current area.
429 * 'full' is filled in to indicate if the area has been
432 static int insert_exceptions(struct pstore *ps,
433 int (*callback)(void *callback_context,
434 chunk_t old, chunk_t new),
435 void *callback_context,
440 struct disk_exception de;
442 /* presume the area is full */
445 for (i = 0; i < ps->exceptions_per_area; i++) {
446 read_exception(ps, i, &de);
449 * If the new_chunk is pointing at the start of
450 * the COW device, where the first metadata area
451 * is we know that we've hit the end of the
452 * exceptions. Therefore the area is not full.
454 if (de.new_chunk == 0LL) {
455 ps->current_committed = i;
461 * Keep track of the start of the free chunks.
463 if (ps->next_free <= de.new_chunk)
464 ps->next_free = de.new_chunk + 1;
467 * Otherwise we add the exception to the snapshot.
469 r = callback(callback_context, de.old_chunk, de.new_chunk);
477 static int read_exceptions(struct pstore *ps,
478 int (*callback)(void *callback_context, chunk_t old,
480 void *callback_context)
485 * Keeping reading chunks and inserting exceptions until
486 * we find a partially full area.
488 for (ps->current_area = 0; full; ps->current_area++) {
489 r = area_io(ps, READ);
493 r = insert_exceptions(ps, callback, callback_context, &full);
503 static struct pstore *get_info(struct dm_exception_store *store)
505 return (struct pstore *) store->context;
508 static void persistent_usage(struct dm_exception_store *store,
509 sector_t *total_sectors,
510 sector_t *sectors_allocated,
511 sector_t *metadata_sectors)
513 struct pstore *ps = get_info(store);
515 *sectors_allocated = ps->next_free * store->chunk_size;
516 *total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
519 * First chunk is the fixed header.
520 * Then there are (ps->current_area + 1) metadata chunks, each one
521 * separated from the next by ps->exceptions_per_area data chunks.
523 *metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
527 static void persistent_dtr(struct dm_exception_store *store)
529 struct pstore *ps = get_info(store);
531 destroy_workqueue(ps->metadata_wq);
533 /* Created in read_header */
535 dm_io_client_destroy(ps->io_client);
538 /* Allocated in persistent_read_metadata */
540 vfree(ps->callbacks);
545 static int persistent_read_metadata(struct dm_exception_store *store,
546 int (*callback)(void *callback_context,
547 chunk_t old, chunk_t new),
548 void *callback_context)
550 int r, uninitialized_var(new_snapshot);
551 struct pstore *ps = get_info(store);
554 * Read the snapshot header.
556 r = read_header(ps, &new_snapshot);
561 * Now we know correct chunk_size, complete the initialisation.
563 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
564 sizeof(struct disk_exception);
565 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
566 sizeof(*ps->callbacks));
571 * Do we need to setup a new snapshot ?
574 r = write_header(ps);
576 DMWARN("write_header failed");
580 ps->current_area = 0;
581 zero_memory_area(ps);
582 r = zero_disk_area(ps, 0);
584 DMWARN("zero_disk_area(0) failed");
590 if (ps->version != SNAPSHOT_DISK_VERSION) {
591 DMWARN("unable to handle snapshot disk version %d",
597 * Metadata are valid, but snapshot is invalidated
605 r = read_exceptions(ps, callback, callback_context);
610 static int persistent_prepare_exception(struct dm_exception_store *store,
611 struct dm_exception *e)
613 struct pstore *ps = get_info(store);
616 sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
618 /* Is there enough room ? */
619 if (size < ((ps->next_free + 1) * store->chunk_size))
622 e->new_chunk = ps->next_free;
625 * Move onto the next free pending, making sure to take
626 * into account the location of the metadata chunks.
628 stride = (ps->exceptions_per_area + 1);
629 next_free = ++ps->next_free;
630 if (sector_div(next_free, stride) == 1)
633 atomic_inc(&ps->pending_count);
637 static void persistent_commit_exception(struct dm_exception_store *store,
638 struct dm_exception *e,
639 void (*callback) (void *, int success),
640 void *callback_context)
643 struct pstore *ps = get_info(store);
644 struct disk_exception de;
645 struct commit_callback *cb;
647 de.old_chunk = e->old_chunk;
648 de.new_chunk = e->new_chunk;
649 write_exception(ps, ps->current_committed++, &de);
652 * Add the callback to the back of the array. This code
653 * is the only place where the callback array is
654 * manipulated, and we know that it will never be called
655 * multiple times concurrently.
657 cb = ps->callbacks + ps->callback_count++;
658 cb->callback = callback;
659 cb->context = callback_context;
662 * If there are exceptions in flight and we have not yet
663 * filled this metadata area there's nothing more to do.
665 if (!atomic_dec_and_test(&ps->pending_count) &&
666 (ps->current_committed != ps->exceptions_per_area))
670 * If we completely filled the current area, then wipe the next one.
672 if ((ps->current_committed == ps->exceptions_per_area) &&
673 zero_disk_area(ps, ps->current_area + 1))
677 * Commit exceptions to disk.
679 if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
683 * Advance to the next area if this one is full.
685 if (ps->current_committed == ps->exceptions_per_area) {
686 ps->current_committed = 0;
688 zero_memory_area(ps);
691 for (i = 0; i < ps->callback_count; i++) {
692 cb = ps->callbacks + i;
693 cb->callback(cb->context, ps->valid);
696 ps->callback_count = 0;
699 static int persistent_prepare_merge(struct dm_exception_store *store,
700 chunk_t *last_old_chunk,
701 chunk_t *last_new_chunk)
703 struct pstore *ps = get_info(store);
704 struct disk_exception de;
709 * When current area is empty, move back to preceding area.
711 if (!ps->current_committed) {
715 if (!ps->current_area)
719 r = area_io(ps, READ);
722 ps->current_committed = ps->exceptions_per_area;
725 read_exception(ps, ps->current_committed - 1, &de);
726 *last_old_chunk = de.old_chunk;
727 *last_new_chunk = de.new_chunk;
730 * Find number of consecutive chunks within the current area,
733 for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
735 read_exception(ps, ps->current_committed - 1 - nr_consecutive,
737 if (de.old_chunk != *last_old_chunk - nr_consecutive ||
738 de.new_chunk != *last_new_chunk - nr_consecutive)
742 return nr_consecutive;
745 static int persistent_commit_merge(struct dm_exception_store *store,
749 struct pstore *ps = get_info(store);
751 BUG_ON(nr_merged > ps->current_committed);
753 for (i = 0; i < nr_merged; i++)
754 clear_exception(ps, ps->current_committed - 1 - i);
756 r = area_io(ps, WRITE);
760 ps->current_committed -= nr_merged;
763 * At this stage, only persistent_usage() uses ps->next_free, so
764 * we make no attempt to keep ps->next_free strictly accurate
765 * as exceptions may have been committed out-of-order originally.
766 * Once a snapshot has become merging, we set it to the value it
767 * would have held had all the exceptions been committed in order.
769 * ps->current_area does not get reduced by prepare_merge() until
770 * after commit_merge() has removed the nr_merged previous exceptions.
772 ps->next_free = area_location(ps, ps->current_area) +
773 ps->current_committed + 1;
778 static void persistent_drop_snapshot(struct dm_exception_store *store)
780 struct pstore *ps = get_info(store);
783 if (write_header(ps))
784 DMWARN("write header failed");
787 static int persistent_ctr(struct dm_exception_store *store,
788 unsigned argc, char **argv)
792 /* allocate the pstore */
793 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
799 ps->version = SNAPSHOT_DISK_VERSION;
801 ps->zero_area = NULL;
802 ps->header_area = NULL;
803 ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
804 ps->current_committed = 0;
806 ps->callback_count = 0;
807 atomic_set(&ps->pending_count, 0);
808 ps->callbacks = NULL;
810 ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
811 if (!ps->metadata_wq) {
813 DMERR("couldn't start header metadata update thread");
822 static unsigned persistent_status(struct dm_exception_store *store,
823 status_type_t status, char *result,
829 case STATUSTYPE_INFO:
831 case STATUSTYPE_TABLE:
832 DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
838 static struct dm_exception_store_type _persistent_type = {
839 .name = "persistent",
840 .module = THIS_MODULE,
841 .ctr = persistent_ctr,
842 .dtr = persistent_dtr,
843 .read_metadata = persistent_read_metadata,
844 .prepare_exception = persistent_prepare_exception,
845 .commit_exception = persistent_commit_exception,
846 .prepare_merge = persistent_prepare_merge,
847 .commit_merge = persistent_commit_merge,
848 .drop_snapshot = persistent_drop_snapshot,
849 .usage = persistent_usage,
850 .status = persistent_status,
853 static struct dm_exception_store_type _persistent_compat_type = {
855 .module = THIS_MODULE,
856 .ctr = persistent_ctr,
857 .dtr = persistent_dtr,
858 .read_metadata = persistent_read_metadata,
859 .prepare_exception = persistent_prepare_exception,
860 .commit_exception = persistent_commit_exception,
861 .prepare_merge = persistent_prepare_merge,
862 .commit_merge = persistent_commit_merge,
863 .drop_snapshot = persistent_drop_snapshot,
864 .usage = persistent_usage,
865 .status = persistent_status,
868 int dm_persistent_snapshot_init(void)
872 r = dm_exception_store_type_register(&_persistent_type);
874 DMERR("Unable to register persistent exception store type");
878 r = dm_exception_store_type_register(&_persistent_compat_type);
880 DMERR("Unable to register old-style persistent exception "
882 dm_exception_store_type_unregister(&_persistent_type);
889 void dm_persistent_snapshot_exit(void)
891 dm_exception_store_type_unregister(&_persistent_type);
892 dm_exception_store_type_unregister(&_persistent_compat_type);