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>
17 #define DM_MSG_PREFIX "persistent snapshot"
18 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
20 /*-----------------------------------------------------------------
21 * Persistent snapshots, by persistent we mean that the snapshot
22 * will survive a reboot.
23 *---------------------------------------------------------------*/
26 * We need to store a record of which parts of the origin have
27 * been copied to the snapshot device. The snapshot code
28 * requires that we copy exception chunks to chunk aligned areas
29 * of the COW store. It makes sense therefore, to store the
30 * metadata in chunk size blocks.
32 * There is no backward or forward compatibility implemented,
33 * snapshots with different disk versions than the kernel will
34 * not be usable. It is expected that "lvcreate" will blank out
35 * the start of a fresh COW device before calling the snapshot
38 * The first chunk of the COW device just contains the header.
39 * After this there is a chunk filled with exception metadata,
40 * followed by as many exception chunks as can fit in the
43 * All on disk structures are in little-endian format. The end
44 * of the exceptions info is indicated by an exception with a
45 * new_chunk of 0, which is invalid since it would point to the
50 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
52 #define SNAP_MAGIC 0x70416e53
55 * The on-disk version of the metadata.
57 #define SNAPSHOT_DISK_VERSION 1
59 #define NUM_SNAPSHOT_HDR_CHUNKS 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 core_exception {
90 struct commit_callback {
91 void (*callback)(void *, int success);
96 * The top level structure for a persistent exception store.
99 struct dm_exception_store *store;
102 uint32_t exceptions_per_area;
105 * Now that we have an asynchronous kcopyd there is no
106 * need for large chunk sizes, so it wont hurt to have a
107 * whole chunks worth of metadata in memory at once.
112 * An area of zeros used to clear the next area.
117 * An area used for header. The header can be written
118 * concurrently with metadata (when invalidating the snapshot),
119 * so it needs a separate buffer.
124 * Used to keep track of which metadata area the data in
127 chunk_t current_area;
130 * The next free chunk for an exception.
132 * When creating exceptions, all the chunks here and above are
133 * free. It holds the next chunk to be allocated. On rare
134 * occasions (e.g. after a system crash) holes can be left in
135 * the exception store because chunks can be committed out of
138 * When merging exceptions, it does not necessarily mean all the
139 * chunks here and above are free. It holds the value it would
140 * have held if all chunks had been committed in order of
141 * allocation. Consequently the value may occasionally be
142 * slightly too low, but since it's only used for 'status' and
143 * it can never reach its minimum value too early this doesn't
150 * The index of next free exception in the current
153 uint32_t current_committed;
155 atomic_t pending_count;
156 uint32_t callback_count;
157 struct commit_callback *callbacks;
158 struct dm_io_client *io_client;
160 struct workqueue_struct *metadata_wq;
163 static int alloc_area(struct pstore *ps)
168 len = ps->store->chunk_size << SECTOR_SHIFT;
171 * Allocate the chunk_size block of memory that will hold
172 * a single metadata area.
174 ps->area = vmalloc(len);
178 ps->zero_area = vzalloc(len);
182 ps->header_area = vmalloc(len);
183 if (!ps->header_area)
184 goto err_header_area;
189 vfree(ps->zero_area);
198 static void free_area(struct pstore *ps)
205 vfree(ps->zero_area);
206 ps->zero_area = NULL;
209 vfree(ps->header_area);
210 ps->header_area = NULL;
214 struct dm_io_region *where;
215 struct dm_io_request *io_req;
216 struct work_struct work;
220 static void do_metadata(struct work_struct *work)
222 struct mdata_req *req = container_of(work, struct mdata_req, work);
224 req->result = dm_io(req->io_req, 1, req->where, NULL);
228 * Read or write a chunk aligned and sized block of data from a device.
230 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, int rw,
233 struct dm_io_region where = {
234 .bdev = dm_snap_cow(ps->store->snap)->bdev,
235 .sector = ps->store->chunk_size * chunk,
236 .count = ps->store->chunk_size,
238 struct dm_io_request io_req = {
240 .mem.type = DM_IO_VMA,
242 .client = ps->io_client,
245 struct mdata_req req;
248 return dm_io(&io_req, 1, &where, NULL);
251 req.io_req = &io_req;
254 * Issue the synchronous I/O from a different thread
255 * to avoid generic_make_request recursion.
257 INIT_WORK_ONSTACK(&req.work, do_metadata);
258 queue_work(ps->metadata_wq, &req.work);
259 flush_work(&req.work);
265 * Convert a metadata area index to a chunk index.
267 static chunk_t area_location(struct pstore *ps, chunk_t area)
269 return NUM_SNAPSHOT_HDR_CHUNKS + ((ps->exceptions_per_area + 1) * area);
273 * Read or write a metadata area. Remembering to skip the first
274 * chunk which holds the header.
276 static int area_io(struct pstore *ps, int rw)
281 chunk = area_location(ps, ps->current_area);
283 r = chunk_io(ps, ps->area, chunk, rw, 0);
290 static void zero_memory_area(struct pstore *ps)
292 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
295 static int zero_disk_area(struct pstore *ps, chunk_t area)
297 return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
300 static int read_header(struct pstore *ps, int *new_snapshot)
303 struct disk_header *dh;
305 int chunk_size_supplied = 1;
309 * Use default chunk size (or logical_block_size, if larger)
312 if (!ps->store->chunk_size) {
313 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
314 bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
316 ps->store->chunk_mask = ps->store->chunk_size - 1;
317 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
318 chunk_size_supplied = 0;
321 ps->io_client = dm_io_client_create();
322 if (IS_ERR(ps->io_client))
323 return PTR_ERR(ps->io_client);
329 r = chunk_io(ps, ps->header_area, 0, READ, 1);
333 dh = ps->header_area;
335 if (le32_to_cpu(dh->magic) == 0) {
340 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
341 DMWARN("Invalid or corrupt snapshot");
347 ps->valid = le32_to_cpu(dh->valid);
348 ps->version = le32_to_cpu(dh->version);
349 chunk_size = le32_to_cpu(dh->chunk_size);
351 if (ps->store->chunk_size == chunk_size)
354 if (chunk_size_supplied)
355 DMWARN("chunk size %u in device metadata overrides "
356 "table chunk size of %u.",
357 chunk_size, ps->store->chunk_size);
359 /* We had a bogus chunk_size. Fix stuff up. */
362 r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
365 DMERR("invalid on-disk chunk size %u: %s.",
366 chunk_size, chunk_err);
378 static int write_header(struct pstore *ps)
380 struct disk_header *dh;
382 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
384 dh = ps->header_area;
385 dh->magic = cpu_to_le32(SNAP_MAGIC);
386 dh->valid = cpu_to_le32(ps->valid);
387 dh->version = cpu_to_le32(ps->version);
388 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
390 return chunk_io(ps, ps->header_area, 0, WRITE, 1);
394 * Access functions for the disk exceptions, these do the endian conversions.
396 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
398 BUG_ON(index >= ps->exceptions_per_area);
400 return ((struct disk_exception *) ps->area) + index;
403 static void read_exception(struct pstore *ps,
404 uint32_t index, struct core_exception *result)
406 struct disk_exception *de = get_exception(ps, index);
409 result->old_chunk = le64_to_cpu(de->old_chunk);
410 result->new_chunk = le64_to_cpu(de->new_chunk);
413 static void write_exception(struct pstore *ps,
414 uint32_t index, struct core_exception *e)
416 struct disk_exception *de = get_exception(ps, index);
419 de->old_chunk = cpu_to_le64(e->old_chunk);
420 de->new_chunk = cpu_to_le64(e->new_chunk);
423 static void clear_exception(struct pstore *ps, uint32_t index)
425 struct disk_exception *de = get_exception(ps, index);
433 * Registers the exceptions that are present in the current area.
434 * 'full' is filled in to indicate if the area has been
437 static int insert_exceptions(struct pstore *ps,
438 int (*callback)(void *callback_context,
439 chunk_t old, chunk_t new),
440 void *callback_context,
445 struct core_exception e;
447 /* presume the area is full */
450 for (i = 0; i < ps->exceptions_per_area; i++) {
451 read_exception(ps, i, &e);
454 * If the new_chunk is pointing at the start of
455 * the COW device, where the first metadata area
456 * is we know that we've hit the end of the
457 * exceptions. Therefore the area is not full.
459 if (e.new_chunk == 0LL) {
460 ps->current_committed = i;
466 * Keep track of the start of the free chunks.
468 if (ps->next_free <= e.new_chunk)
469 ps->next_free = e.new_chunk + 1;
472 * Otherwise we add the exception to the snapshot.
474 r = callback(callback_context, e.old_chunk, e.new_chunk);
482 static int read_exceptions(struct pstore *ps,
483 int (*callback)(void *callback_context, chunk_t old,
485 void *callback_context)
490 * Keeping reading chunks and inserting exceptions until
491 * we find a partially full area.
493 for (ps->current_area = 0; full; ps->current_area++) {
494 r = area_io(ps, READ);
498 r = insert_exceptions(ps, callback, callback_context, &full);
508 static struct pstore *get_info(struct dm_exception_store *store)
510 return (struct pstore *) store->context;
513 static void persistent_usage(struct dm_exception_store *store,
514 sector_t *total_sectors,
515 sector_t *sectors_allocated,
516 sector_t *metadata_sectors)
518 struct pstore *ps = get_info(store);
520 *sectors_allocated = ps->next_free * store->chunk_size;
521 *total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
524 * First chunk is the fixed header.
525 * Then there are (ps->current_area + 1) metadata chunks, each one
526 * separated from the next by ps->exceptions_per_area data chunks.
528 *metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
532 static void persistent_dtr(struct dm_exception_store *store)
534 struct pstore *ps = get_info(store);
536 destroy_workqueue(ps->metadata_wq);
538 /* Created in read_header */
540 dm_io_client_destroy(ps->io_client);
543 /* Allocated in persistent_read_metadata */
545 vfree(ps->callbacks);
550 static int persistent_read_metadata(struct dm_exception_store *store,
551 int (*callback)(void *callback_context,
552 chunk_t old, chunk_t new),
553 void *callback_context)
555 int r, uninitialized_var(new_snapshot);
556 struct pstore *ps = get_info(store);
559 * Read the snapshot header.
561 r = read_header(ps, &new_snapshot);
566 * Now we know correct chunk_size, complete the initialisation.
568 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
569 sizeof(struct disk_exception);
570 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
571 sizeof(*ps->callbacks));
576 * Do we need to setup a new snapshot ?
579 r = write_header(ps);
581 DMWARN("write_header failed");
585 ps->current_area = 0;
586 zero_memory_area(ps);
587 r = zero_disk_area(ps, 0);
589 DMWARN("zero_disk_area(0) failed");
595 if (ps->version != SNAPSHOT_DISK_VERSION) {
596 DMWARN("unable to handle snapshot disk version %d",
602 * Metadata are valid, but snapshot is invalidated
610 r = read_exceptions(ps, callback, callback_context);
615 static int persistent_prepare_exception(struct dm_exception_store *store,
616 struct dm_exception *e)
618 struct pstore *ps = get_info(store);
621 sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
623 /* Is there enough room ? */
624 if (size < ((ps->next_free + 1) * store->chunk_size))
627 e->new_chunk = ps->next_free;
630 * Move onto the next free pending, making sure to take
631 * into account the location of the metadata chunks.
633 stride = (ps->exceptions_per_area + 1);
634 next_free = ++ps->next_free;
635 if (sector_div(next_free, stride) == 1)
638 atomic_inc(&ps->pending_count);
642 static void persistent_commit_exception(struct dm_exception_store *store,
643 struct dm_exception *e,
644 void (*callback) (void *, int success),
645 void *callback_context)
648 struct pstore *ps = get_info(store);
649 struct core_exception ce;
650 struct commit_callback *cb;
652 ce.old_chunk = e->old_chunk;
653 ce.new_chunk = e->new_chunk;
654 write_exception(ps, ps->current_committed++, &ce);
657 * Add the callback to the back of the array. This code
658 * is the only place where the callback array is
659 * manipulated, and we know that it will never be called
660 * multiple times concurrently.
662 cb = ps->callbacks + ps->callback_count++;
663 cb->callback = callback;
664 cb->context = callback_context;
667 * If there are exceptions in flight and we have not yet
668 * filled this metadata area there's nothing more to do.
670 if (!atomic_dec_and_test(&ps->pending_count) &&
671 (ps->current_committed != ps->exceptions_per_area))
675 * If we completely filled the current area, then wipe the next one.
677 if ((ps->current_committed == ps->exceptions_per_area) &&
678 zero_disk_area(ps, ps->current_area + 1))
682 * Commit exceptions to disk.
684 if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
688 * Advance to the next area if this one is full.
690 if (ps->current_committed == ps->exceptions_per_area) {
691 ps->current_committed = 0;
693 zero_memory_area(ps);
696 for (i = 0; i < ps->callback_count; i++) {
697 cb = ps->callbacks + i;
698 cb->callback(cb->context, ps->valid);
701 ps->callback_count = 0;
704 static int persistent_prepare_merge(struct dm_exception_store *store,
705 chunk_t *last_old_chunk,
706 chunk_t *last_new_chunk)
708 struct pstore *ps = get_info(store);
709 struct core_exception ce;
714 * When current area is empty, move back to preceding area.
716 if (!ps->current_committed) {
720 if (!ps->current_area)
724 r = area_io(ps, READ);
727 ps->current_committed = ps->exceptions_per_area;
730 read_exception(ps, ps->current_committed - 1, &ce);
731 *last_old_chunk = ce.old_chunk;
732 *last_new_chunk = ce.new_chunk;
735 * Find number of consecutive chunks within the current area,
738 for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
740 read_exception(ps, ps->current_committed - 1 - nr_consecutive,
742 if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
743 ce.new_chunk != *last_new_chunk - nr_consecutive)
747 return nr_consecutive;
750 static int persistent_commit_merge(struct dm_exception_store *store,
754 struct pstore *ps = get_info(store);
756 BUG_ON(nr_merged > ps->current_committed);
758 for (i = 0; i < nr_merged; i++)
759 clear_exception(ps, ps->current_committed - 1 - i);
761 r = area_io(ps, WRITE_FLUSH_FUA);
765 ps->current_committed -= nr_merged;
768 * At this stage, only persistent_usage() uses ps->next_free, so
769 * we make no attempt to keep ps->next_free strictly accurate
770 * as exceptions may have been committed out-of-order originally.
771 * Once a snapshot has become merging, we set it to the value it
772 * would have held had all the exceptions been committed in order.
774 * ps->current_area does not get reduced by prepare_merge() until
775 * after commit_merge() has removed the nr_merged previous exceptions.
777 ps->next_free = area_location(ps, ps->current_area) +
778 ps->current_committed + 1;
783 static void persistent_drop_snapshot(struct dm_exception_store *store)
785 struct pstore *ps = get_info(store);
788 if (write_header(ps))
789 DMWARN("write header failed");
792 static int persistent_ctr(struct dm_exception_store *store,
793 unsigned argc, char **argv)
797 /* allocate the pstore */
798 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
804 ps->version = SNAPSHOT_DISK_VERSION;
806 ps->zero_area = NULL;
807 ps->header_area = NULL;
808 ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
809 ps->current_committed = 0;
811 ps->callback_count = 0;
812 atomic_set(&ps->pending_count, 0);
813 ps->callbacks = NULL;
815 ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
816 if (!ps->metadata_wq) {
818 DMERR("couldn't start header metadata update thread");
827 static unsigned persistent_status(struct dm_exception_store *store,
828 status_type_t status, char *result,
834 case STATUSTYPE_INFO:
836 case STATUSTYPE_TABLE:
837 DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
843 static struct dm_exception_store_type _persistent_type = {
844 .name = "persistent",
845 .module = THIS_MODULE,
846 .ctr = persistent_ctr,
847 .dtr = persistent_dtr,
848 .read_metadata = persistent_read_metadata,
849 .prepare_exception = persistent_prepare_exception,
850 .commit_exception = persistent_commit_exception,
851 .prepare_merge = persistent_prepare_merge,
852 .commit_merge = persistent_commit_merge,
853 .drop_snapshot = persistent_drop_snapshot,
854 .usage = persistent_usage,
855 .status = persistent_status,
858 static struct dm_exception_store_type _persistent_compat_type = {
860 .module = THIS_MODULE,
861 .ctr = persistent_ctr,
862 .dtr = persistent_dtr,
863 .read_metadata = persistent_read_metadata,
864 .prepare_exception = persistent_prepare_exception,
865 .commit_exception = persistent_commit_exception,
866 .prepare_merge = persistent_prepare_merge,
867 .commit_merge = persistent_commit_merge,
868 .drop_snapshot = persistent_drop_snapshot,
869 .usage = persistent_usage,
870 .status = persistent_status,
873 int dm_persistent_snapshot_init(void)
877 r = dm_exception_store_type_register(&_persistent_type);
879 DMERR("Unable to register persistent exception store type");
883 r = dm_exception_store_type_register(&_persistent_compat_type);
885 DMERR("Unable to register old-style persistent exception "
887 dm_exception_store_type_unregister(&_persistent_type);
894 void dm_persistent_snapshot_exit(void)
896 dm_exception_store_type_unregister(&_persistent_type);
897 dm_exception_store_type_unregister(&_persistent_compat_type);