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"
10 #include <linux/ctype.h>
12 #include <linux/pagemap.h>
13 #include <linux/vmalloc.h>
14 #include <linux/export.h>
15 #include <linux/slab.h>
16 #include <linux/dm-io.h>
17 #include <linux/dm-bufio.h>
19 #define DM_MSG_PREFIX "persistent snapshot"
20 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32U /* 16KB */
22 #define DM_PREFETCH_CHUNKS 12
24 /*-----------------------------------------------------------------
25 * Persistent snapshots, by persistent we mean that the snapshot
26 * will survive a reboot.
27 *---------------------------------------------------------------*/
30 * We need to store a record of which parts of the origin have
31 * been copied to the snapshot device. The snapshot code
32 * requires that we copy exception chunks to chunk aligned areas
33 * of the COW store. It makes sense therefore, to store the
34 * metadata in chunk size blocks.
36 * There is no backward or forward compatibility implemented,
37 * snapshots with different disk versions than the kernel will
38 * not be usable. It is expected that "lvcreate" will blank out
39 * the start of a fresh COW device before calling the snapshot
42 * The first chunk of the COW device just contains the header.
43 * After this there is a chunk filled with exception metadata,
44 * followed by as many exception chunks as can fit in the
47 * All on disk structures are in little-endian format. The end
48 * of the exceptions info is indicated by an exception with a
49 * new_chunk of 0, which is invalid since it would point to the
54 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
56 #define SNAP_MAGIC 0x70416e53
59 * The on-disk version of the metadata.
61 #define SNAPSHOT_DISK_VERSION 1
63 #define NUM_SNAPSHOT_HDR_CHUNKS 1
69 * Is this snapshot valid. There is no way of recovering
70 * an invalid snapshot.
75 * Simple, incrementing version. no backward
84 struct disk_exception {
89 struct core_exception {
94 struct commit_callback {
95 void (*callback)(void *, int success);
100 * The top level structure for a persistent exception store.
103 struct dm_exception_store *store;
106 uint32_t exceptions_per_area;
109 * Now that we have an asynchronous kcopyd there is no
110 * need for large chunk sizes, so it wont hurt to have a
111 * whole chunks worth of metadata in memory at once.
116 * An area of zeros used to clear the next area.
121 * An area used for header. The header can be written
122 * concurrently with metadata (when invalidating the snapshot),
123 * so it needs a separate buffer.
128 * Used to keep track of which metadata area the data in
131 chunk_t current_area;
134 * The next free chunk for an exception.
136 * When creating exceptions, all the chunks here and above are
137 * free. It holds the next chunk to be allocated. On rare
138 * occasions (e.g. after a system crash) holes can be left in
139 * the exception store because chunks can be committed out of
142 * When merging exceptions, it does not necessarily mean all the
143 * chunks here and above are free. It holds the value it would
144 * have held if all chunks had been committed in order of
145 * allocation. Consequently the value may occasionally be
146 * slightly too low, but since it's only used for 'status' and
147 * it can never reach its minimum value too early this doesn't
154 * The index of next free exception in the current
157 uint32_t current_committed;
159 atomic_t pending_count;
160 uint32_t callback_count;
161 struct commit_callback *callbacks;
162 struct dm_io_client *io_client;
164 struct workqueue_struct *metadata_wq;
167 static int alloc_area(struct pstore *ps)
172 len = ps->store->chunk_size << SECTOR_SHIFT;
175 * Allocate the chunk_size block of memory that will hold
176 * a single metadata area.
178 ps->area = vmalloc(len);
182 ps->zero_area = vzalloc(len);
186 ps->header_area = vmalloc(len);
187 if (!ps->header_area)
188 goto err_header_area;
193 vfree(ps->zero_area);
202 static void free_area(struct pstore *ps)
206 vfree(ps->zero_area);
207 ps->zero_area = NULL;
208 vfree(ps->header_area);
209 ps->header_area = NULL;
213 struct dm_io_region *where;
214 struct dm_io_request *io_req;
215 struct work_struct work;
219 static void do_metadata(struct work_struct *work)
221 struct mdata_req *req = container_of(work, struct mdata_req, work);
223 req->result = dm_io(req->io_req, 1, req->where, NULL);
227 * Read or write a chunk aligned and sized block of data from a device.
229 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, blk_opf_t opf,
232 struct dm_io_region where = {
233 .bdev = dm_snap_cow(ps->store->snap)->bdev,
234 .sector = ps->store->chunk_size * chunk,
235 .count = ps->store->chunk_size,
237 struct dm_io_request io_req = {
239 .mem.type = DM_IO_VMA,
241 .client = ps->io_client,
244 struct mdata_req req;
247 return dm_io(&io_req, 1, &where, NULL);
250 req.io_req = &io_req;
253 * Issue the synchronous I/O from a different thread
254 * to avoid submit_bio_noacct recursion.
256 INIT_WORK_ONSTACK(&req.work, do_metadata);
257 queue_work(ps->metadata_wq, &req.work);
258 flush_workqueue(ps->metadata_wq);
259 destroy_work_on_stack(&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);
272 static void skip_metadata(struct pstore *ps)
274 uint32_t stride = ps->exceptions_per_area + 1;
275 chunk_t next_free = ps->next_free;
276 if (sector_div(next_free, stride) == NUM_SNAPSHOT_HDR_CHUNKS)
281 * Read or write a metadata area. Remembering to skip the first
282 * chunk which holds the header.
284 static int area_io(struct pstore *ps, blk_opf_t opf)
286 chunk_t chunk = area_location(ps, ps->current_area);
288 return chunk_io(ps, ps->area, chunk, opf, 0);
291 static void zero_memory_area(struct pstore *ps)
293 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
296 static int zero_disk_area(struct pstore *ps, chunk_t area)
298 return chunk_io(ps, ps->zero_area, area_location(ps, area),
302 static int read_header(struct pstore *ps, int *new_snapshot)
305 struct disk_header *dh;
307 int chunk_size_supplied = 1;
311 * Use default chunk size (or logical_block_size, if larger)
314 if (!ps->store->chunk_size) {
315 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
316 bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
318 ps->store->chunk_mask = ps->store->chunk_size - 1;
319 ps->store->chunk_shift = __ffs(ps->store->chunk_size);
320 chunk_size_supplied = 0;
323 ps->io_client = dm_io_client_create();
324 if (IS_ERR(ps->io_client))
325 return PTR_ERR(ps->io_client);
331 r = chunk_io(ps, ps->header_area, 0, REQ_OP_READ, 1);
335 dh = ps->header_area;
337 if (le32_to_cpu(dh->magic) == 0) {
342 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
343 DMWARN("Invalid or corrupt snapshot");
349 ps->valid = le32_to_cpu(dh->valid);
350 ps->version = le32_to_cpu(dh->version);
351 chunk_size = le32_to_cpu(dh->chunk_size);
353 if (ps->store->chunk_size == chunk_size)
356 if (chunk_size_supplied)
357 DMWARN("chunk size %u in device metadata overrides "
358 "table chunk size of %u.",
359 chunk_size, ps->store->chunk_size);
361 /* We had a bogus chunk_size. Fix stuff up. */
364 r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
367 DMERR("invalid on-disk chunk size %u: %s.",
368 chunk_size, chunk_err);
380 static int write_header(struct pstore *ps)
382 struct disk_header *dh;
384 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
386 dh = ps->header_area;
387 dh->magic = cpu_to_le32(SNAP_MAGIC);
388 dh->valid = cpu_to_le32(ps->valid);
389 dh->version = cpu_to_le32(ps->version);
390 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
392 return chunk_io(ps, ps->header_area, 0, REQ_OP_WRITE, 1);
396 * Access functions for the disk exceptions, these do the endian conversions.
398 static struct disk_exception *get_exception(struct pstore *ps, void *ps_area,
401 BUG_ON(index >= ps->exceptions_per_area);
403 return ((struct disk_exception *) ps_area) + index;
406 static void read_exception(struct pstore *ps, void *ps_area,
407 uint32_t index, struct core_exception *result)
409 struct disk_exception *de = get_exception(ps, ps_area, index);
412 result->old_chunk = le64_to_cpu(de->old_chunk);
413 result->new_chunk = le64_to_cpu(de->new_chunk);
416 static void write_exception(struct pstore *ps,
417 uint32_t index, struct core_exception *e)
419 struct disk_exception *de = get_exception(ps, ps->area, index);
422 de->old_chunk = cpu_to_le64(e->old_chunk);
423 de->new_chunk = cpu_to_le64(e->new_chunk);
426 static void clear_exception(struct pstore *ps, uint32_t index)
428 struct disk_exception *de = get_exception(ps, ps->area, index);
436 * Registers the exceptions that are present in the current area.
437 * 'full' is filled in to indicate if the area has been
440 static int insert_exceptions(struct pstore *ps, void *ps_area,
441 int (*callback)(void *callback_context,
442 chunk_t old, chunk_t new),
443 void *callback_context,
448 struct core_exception e;
450 /* presume the area is full */
453 for (i = 0; i < ps->exceptions_per_area; i++) {
454 read_exception(ps, ps_area, i, &e);
457 * If the new_chunk is pointing at the start of
458 * the COW device, where the first metadata area
459 * is we know that we've hit the end of the
460 * exceptions. Therefore the area is not full.
462 if (e.new_chunk == 0LL) {
463 ps->current_committed = i;
469 * Keep track of the start of the free chunks.
471 if (ps->next_free <= e.new_chunk)
472 ps->next_free = e.new_chunk + 1;
475 * Otherwise we add the exception to the snapshot.
477 r = callback(callback_context, e.old_chunk, e.new_chunk);
485 static int read_exceptions(struct pstore *ps,
486 int (*callback)(void *callback_context, chunk_t old,
488 void *callback_context)
491 struct dm_bufio_client *client;
492 chunk_t prefetch_area = 0;
494 client = dm_bufio_client_create(dm_snap_cow(ps->store->snap)->bdev,
495 ps->store->chunk_size << SECTOR_SHIFT,
496 1, 0, NULL, NULL, 0);
499 return PTR_ERR(client);
502 * Setup for one current buffer + desired readahead buffers.
504 dm_bufio_set_minimum_buffers(client, 1 + DM_PREFETCH_CHUNKS);
507 * Keeping reading chunks and inserting exceptions until
508 * we find a partially full area.
510 for (ps->current_area = 0; full; ps->current_area++) {
511 struct dm_buffer *bp;
515 if (unlikely(prefetch_area < ps->current_area))
516 prefetch_area = ps->current_area;
518 if (DM_PREFETCH_CHUNKS) do {
519 chunk_t pf_chunk = area_location(ps, prefetch_area);
520 if (unlikely(pf_chunk >= dm_bufio_get_device_size(client)))
522 dm_bufio_prefetch(client, pf_chunk, 1);
524 if (unlikely(!prefetch_area))
526 } while (prefetch_area <= ps->current_area + DM_PREFETCH_CHUNKS);
528 chunk = area_location(ps, ps->current_area);
530 area = dm_bufio_read(client, chunk, &bp);
533 goto ret_destroy_bufio;
536 r = insert_exceptions(ps, area, callback, callback_context,
540 memcpy(ps->area, area, ps->store->chunk_size << SECTOR_SHIFT);
542 dm_bufio_release(bp);
544 dm_bufio_forget(client, chunk);
547 goto ret_destroy_bufio;
557 dm_bufio_client_destroy(client);
562 static struct pstore *get_info(struct dm_exception_store *store)
564 return (struct pstore *) store->context;
567 static void persistent_usage(struct dm_exception_store *store,
568 sector_t *total_sectors,
569 sector_t *sectors_allocated,
570 sector_t *metadata_sectors)
572 struct pstore *ps = get_info(store);
574 *sectors_allocated = ps->next_free * store->chunk_size;
575 *total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
578 * First chunk is the fixed header.
579 * Then there are (ps->current_area + 1) metadata chunks, each one
580 * separated from the next by ps->exceptions_per_area data chunks.
582 *metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
586 static void persistent_dtr(struct dm_exception_store *store)
588 struct pstore *ps = get_info(store);
590 destroy_workqueue(ps->metadata_wq);
592 /* Created in read_header */
594 dm_io_client_destroy(ps->io_client);
597 /* Allocated in persistent_read_metadata */
598 kvfree(ps->callbacks);
603 static int persistent_read_metadata(struct dm_exception_store *store,
604 int (*callback)(void *callback_context,
605 chunk_t old, chunk_t new),
606 void *callback_context)
609 struct pstore *ps = get_info(store);
612 * Read the snapshot header.
614 r = read_header(ps, &new_snapshot);
619 * Now we know correct chunk_size, complete the initialisation.
621 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
622 sizeof(struct disk_exception);
623 ps->callbacks = kvcalloc(ps->exceptions_per_area,
624 sizeof(*ps->callbacks), GFP_KERNEL);
629 * Do we need to setup a new snapshot ?
632 r = write_header(ps);
634 DMWARN("write_header failed");
638 ps->current_area = 0;
639 zero_memory_area(ps);
640 r = zero_disk_area(ps, 0);
642 DMWARN("zero_disk_area(0) failed");
648 if (ps->version != SNAPSHOT_DISK_VERSION) {
649 DMWARN("unable to handle snapshot disk version %d",
655 * Metadata are valid, but snapshot is invalidated
663 r = read_exceptions(ps, callback, callback_context);
668 static int persistent_prepare_exception(struct dm_exception_store *store,
669 struct dm_exception *e)
671 struct pstore *ps = get_info(store);
672 sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
674 /* Is there enough room ? */
675 if (size < ((ps->next_free + 1) * store->chunk_size))
678 e->new_chunk = ps->next_free;
681 * Move onto the next free pending, making sure to take
682 * into account the location of the metadata chunks.
687 atomic_inc(&ps->pending_count);
691 static void persistent_commit_exception(struct dm_exception_store *store,
692 struct dm_exception *e, int valid,
693 void (*callback) (void *, int success),
694 void *callback_context)
697 struct pstore *ps = get_info(store);
698 struct core_exception ce;
699 struct commit_callback *cb;
704 ce.old_chunk = e->old_chunk;
705 ce.new_chunk = e->new_chunk;
706 write_exception(ps, ps->current_committed++, &ce);
709 * Add the callback to the back of the array. This code
710 * is the only place where the callback array is
711 * manipulated, and we know that it will never be called
712 * multiple times concurrently.
714 cb = ps->callbacks + ps->callback_count++;
715 cb->callback = callback;
716 cb->context = callback_context;
719 * If there are exceptions in flight and we have not yet
720 * filled this metadata area there's nothing more to do.
722 if (!atomic_dec_and_test(&ps->pending_count) &&
723 (ps->current_committed != ps->exceptions_per_area))
727 * If we completely filled the current area, then wipe the next one.
729 if ((ps->current_committed == ps->exceptions_per_area) &&
730 zero_disk_area(ps, ps->current_area + 1))
734 * Commit exceptions to disk.
736 if (ps->valid && area_io(ps, REQ_OP_WRITE | REQ_PREFLUSH | REQ_FUA |
741 * Advance to the next area if this one is full.
743 if (ps->current_committed == ps->exceptions_per_area) {
744 ps->current_committed = 0;
746 zero_memory_area(ps);
749 for (i = 0; i < ps->callback_count; i++) {
750 cb = ps->callbacks + i;
751 cb->callback(cb->context, ps->valid);
754 ps->callback_count = 0;
757 static int persistent_prepare_merge(struct dm_exception_store *store,
758 chunk_t *last_old_chunk,
759 chunk_t *last_new_chunk)
761 struct pstore *ps = get_info(store);
762 struct core_exception ce;
767 * When current area is empty, move back to preceding area.
769 if (!ps->current_committed) {
773 if (!ps->current_area)
777 r = area_io(ps, REQ_OP_READ);
780 ps->current_committed = ps->exceptions_per_area;
783 read_exception(ps, ps->area, ps->current_committed - 1, &ce);
784 *last_old_chunk = ce.old_chunk;
785 *last_new_chunk = ce.new_chunk;
788 * Find number of consecutive chunks within the current area,
791 for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
793 read_exception(ps, ps->area,
794 ps->current_committed - 1 - nr_consecutive, &ce);
795 if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
796 ce.new_chunk != *last_new_chunk - nr_consecutive)
800 return nr_consecutive;
803 static int persistent_commit_merge(struct dm_exception_store *store,
807 struct pstore *ps = get_info(store);
809 BUG_ON(nr_merged > ps->current_committed);
811 for (i = 0; i < nr_merged; i++)
812 clear_exception(ps, ps->current_committed - 1 - i);
814 r = area_io(ps, REQ_OP_WRITE | REQ_PREFLUSH | REQ_FUA);
818 ps->current_committed -= nr_merged;
821 * At this stage, only persistent_usage() uses ps->next_free, so
822 * we make no attempt to keep ps->next_free strictly accurate
823 * as exceptions may have been committed out-of-order originally.
824 * Once a snapshot has become merging, we set it to the value it
825 * would have held had all the exceptions been committed in order.
827 * ps->current_area does not get reduced by prepare_merge() until
828 * after commit_merge() has removed the nr_merged previous exceptions.
830 ps->next_free = area_location(ps, ps->current_area) +
831 ps->current_committed + 1;
836 static void persistent_drop_snapshot(struct dm_exception_store *store)
838 struct pstore *ps = get_info(store);
841 if (write_header(ps))
842 DMWARN("write header failed");
845 static int persistent_ctr(struct dm_exception_store *store, char *options)
850 /* allocate the pstore */
851 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
857 ps->version = SNAPSHOT_DISK_VERSION;
859 ps->zero_area = NULL;
860 ps->header_area = NULL;
861 ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
862 ps->current_committed = 0;
864 ps->callback_count = 0;
865 atomic_set(&ps->pending_count, 0);
866 ps->callbacks = NULL;
868 ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
869 if (!ps->metadata_wq) {
870 DMERR("couldn't start header metadata update thread");
876 char overflow = toupper(options[0]);
878 store->userspace_supports_overflow = true;
880 DMERR("Unsupported persistent store option: %s", options);
891 destroy_workqueue(ps->metadata_wq);
898 static unsigned persistent_status(struct dm_exception_store *store,
899 status_type_t status, char *result,
905 case STATUSTYPE_INFO:
907 case STATUSTYPE_TABLE:
908 DMEMIT(" %s %llu", store->userspace_supports_overflow ? "PO" : "P",
909 (unsigned long long)store->chunk_size);
919 static struct dm_exception_store_type _persistent_type = {
920 .name = "persistent",
921 .module = THIS_MODULE,
922 .ctr = persistent_ctr,
923 .dtr = persistent_dtr,
924 .read_metadata = persistent_read_metadata,
925 .prepare_exception = persistent_prepare_exception,
926 .commit_exception = persistent_commit_exception,
927 .prepare_merge = persistent_prepare_merge,
928 .commit_merge = persistent_commit_merge,
929 .drop_snapshot = persistent_drop_snapshot,
930 .usage = persistent_usage,
931 .status = persistent_status,
934 static struct dm_exception_store_type _persistent_compat_type = {
936 .module = THIS_MODULE,
937 .ctr = persistent_ctr,
938 .dtr = persistent_dtr,
939 .read_metadata = persistent_read_metadata,
940 .prepare_exception = persistent_prepare_exception,
941 .commit_exception = persistent_commit_exception,
942 .prepare_merge = persistent_prepare_merge,
943 .commit_merge = persistent_commit_merge,
944 .drop_snapshot = persistent_drop_snapshot,
945 .usage = persistent_usage,
946 .status = persistent_status,
949 int dm_persistent_snapshot_init(void)
953 r = dm_exception_store_type_register(&_persistent_type);
955 DMERR("Unable to register persistent exception store type");
959 r = dm_exception_store_type_register(&_persistent_compat_type);
961 DMERR("Unable to register old-style persistent exception "
963 dm_exception_store_type_unregister(&_persistent_type);
970 void dm_persistent_snapshot_exit(void)
972 dm_exception_store_type_unregister(&_persistent_type);
973 dm_exception_store_type_unregister(&_persistent_compat_type);