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 core_exception {
89 struct commit_callback {
90 void (*callback)(void *, int success);
95 * The top level structure for a persistent exception store.
98 struct dm_exception_store *store;
101 uint32_t exceptions_per_area;
104 * Now that we have an asynchronous kcopyd there is no
105 * need for large chunk sizes, so it wont hurt to have a
106 * whole chunks worth of metadata in memory at once.
111 * An area of zeros used to clear the next area.
116 * An area used for header. The header can be written
117 * concurrently with metadata (when invalidating the snapshot),
118 * so it needs a separate buffer.
123 * Used to keep track of which metadata area the data in
126 chunk_t current_area;
129 * The next free chunk for an exception.
131 * When creating exceptions, all the chunks here and above are
132 * free. It holds the next chunk to be allocated. On rare
133 * occasions (e.g. after a system crash) holes can be left in
134 * the exception store because chunks can be committed out of
137 * When merging exceptions, it does not necessarily mean all the
138 * chunks here and above are free. It holds the value it would
139 * have held if all chunks had been committed in order of
140 * allocation. Consequently the value may occasionally be
141 * slightly too low, but since it's only used for 'status' and
142 * it can never reach its minimum value too early this doesn't
149 * The index of next free exception in the current
152 uint32_t current_committed;
154 atomic_t pending_count;
155 uint32_t callback_count;
156 struct commit_callback *callbacks;
157 struct dm_io_client *io_client;
159 struct workqueue_struct *metadata_wq;
162 static int alloc_area(struct pstore *ps)
167 len = ps->store->chunk_size << SECTOR_SHIFT;
170 * Allocate the chunk_size block of memory that will hold
171 * a single metadata area.
173 ps->area = vmalloc(len);
177 ps->zero_area = vzalloc(len);
181 ps->header_area = vmalloc(len);
182 if (!ps->header_area)
183 goto err_header_area;
188 vfree(ps->zero_area);
197 static void free_area(struct pstore *ps)
204 vfree(ps->zero_area);
205 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, int rw,
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 generic_make_request recursion.
256 INIT_WORK_ONSTACK(&req.work, do_metadata);
257 queue_work(ps->metadata_wq, &req.work);
258 flush_work(&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);
272 * Read or write a metadata area. Remembering to skip the first
273 * chunk which holds the header.
275 static int area_io(struct pstore *ps, int rw)
280 chunk = area_location(ps, ps->current_area);
282 r = chunk_io(ps, ps->area, chunk, rw, 0);
289 static void zero_memory_area(struct pstore *ps)
291 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
294 static int zero_disk_area(struct pstore *ps, chunk_t area)
296 return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
299 static int read_header(struct pstore *ps, int *new_snapshot)
302 struct disk_header *dh;
304 int chunk_size_supplied = 1;
308 * Use default chunk size (or logical_block_size, if larger)
311 if (!ps->store->chunk_size) {
312 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
313 bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
315 ps->store->chunk_mask = ps->store->chunk_size - 1;
316 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
317 chunk_size_supplied = 0;
320 ps->io_client = dm_io_client_create();
321 if (IS_ERR(ps->io_client))
322 return PTR_ERR(ps->io_client);
328 r = chunk_io(ps, ps->header_area, 0, READ, 1);
332 dh = ps->header_area;
334 if (le32_to_cpu(dh->magic) == 0) {
339 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
340 DMWARN("Invalid or corrupt snapshot");
346 ps->valid = le32_to_cpu(dh->valid);
347 ps->version = le32_to_cpu(dh->version);
348 chunk_size = le32_to_cpu(dh->chunk_size);
350 if (ps->store->chunk_size == chunk_size)
353 if (chunk_size_supplied)
354 DMWARN("chunk size %u in device metadata overrides "
355 "table chunk size of %u.",
356 chunk_size, ps->store->chunk_size);
358 /* We had a bogus chunk_size. Fix stuff up. */
361 r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
364 DMERR("invalid on-disk chunk size %u: %s.",
365 chunk_size, chunk_err);
377 static int write_header(struct pstore *ps)
379 struct disk_header *dh;
381 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
383 dh = ps->header_area;
384 dh->magic = cpu_to_le32(SNAP_MAGIC);
385 dh->valid = cpu_to_le32(ps->valid);
386 dh->version = cpu_to_le32(ps->version);
387 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
389 return chunk_io(ps, ps->header_area, 0, WRITE, 1);
393 * Access functions for the disk exceptions, these do the endian conversions.
395 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
397 BUG_ON(index >= ps->exceptions_per_area);
399 return ((struct disk_exception *) ps->area) + index;
402 static void read_exception(struct pstore *ps,
403 uint32_t index, struct core_exception *result)
405 struct disk_exception *de = get_exception(ps, index);
408 result->old_chunk = le64_to_cpu(de->old_chunk);
409 result->new_chunk = le64_to_cpu(de->new_chunk);
412 static void write_exception(struct pstore *ps,
413 uint32_t index, struct core_exception *e)
415 struct disk_exception *de = get_exception(ps, index);
418 de->old_chunk = cpu_to_le64(e->old_chunk);
419 de->new_chunk = cpu_to_le64(e->new_chunk);
422 static void clear_exception(struct pstore *ps, uint32_t index)
424 struct disk_exception *de = get_exception(ps, index);
432 * Registers the exceptions that are present in the current area.
433 * 'full' is filled in to indicate if the area has been
436 static int insert_exceptions(struct pstore *ps,
437 int (*callback)(void *callback_context,
438 chunk_t old, chunk_t new),
439 void *callback_context,
444 struct core_exception e;
446 /* presume the area is full */
449 for (i = 0; i < ps->exceptions_per_area; i++) {
450 read_exception(ps, i, &e);
453 * If the new_chunk is pointing at the start of
454 * the COW device, where the first metadata area
455 * is we know that we've hit the end of the
456 * exceptions. Therefore the area is not full.
458 if (e.new_chunk == 0LL) {
459 ps->current_committed = i;
465 * Keep track of the start of the free chunks.
467 if (ps->next_free <= e.new_chunk)
468 ps->next_free = e.new_chunk + 1;
471 * Otherwise we add the exception to the snapshot.
473 r = callback(callback_context, e.old_chunk, e.new_chunk);
481 static int read_exceptions(struct pstore *ps,
482 int (*callback)(void *callback_context, chunk_t old,
484 void *callback_context)
489 * Keeping reading chunks and inserting exceptions until
490 * we find a partially full area.
492 for (ps->current_area = 0; full; ps->current_area++) {
493 r = area_io(ps, READ);
497 r = insert_exceptions(ps, callback, callback_context, &full);
507 static struct pstore *get_info(struct dm_exception_store *store)
509 return (struct pstore *) store->context;
512 static void persistent_usage(struct dm_exception_store *store,
513 sector_t *total_sectors,
514 sector_t *sectors_allocated,
515 sector_t *metadata_sectors)
517 struct pstore *ps = get_info(store);
519 *sectors_allocated = ps->next_free * store->chunk_size;
520 *total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
523 * First chunk is the fixed header.
524 * Then there are (ps->current_area + 1) metadata chunks, each one
525 * separated from the next by ps->exceptions_per_area data chunks.
527 *metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
531 static void persistent_dtr(struct dm_exception_store *store)
533 struct pstore *ps = get_info(store);
535 destroy_workqueue(ps->metadata_wq);
537 /* Created in read_header */
539 dm_io_client_destroy(ps->io_client);
542 /* Allocated in persistent_read_metadata */
544 vfree(ps->callbacks);
549 static int persistent_read_metadata(struct dm_exception_store *store,
550 int (*callback)(void *callback_context,
551 chunk_t old, chunk_t new),
552 void *callback_context)
554 int r, uninitialized_var(new_snapshot);
555 struct pstore *ps = get_info(store);
558 * Read the snapshot header.
560 r = read_header(ps, &new_snapshot);
565 * Now we know correct chunk_size, complete the initialisation.
567 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
568 sizeof(struct disk_exception);
569 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
570 sizeof(*ps->callbacks));
575 * Do we need to setup a new snapshot ?
578 r = write_header(ps);
580 DMWARN("write_header failed");
584 ps->current_area = 0;
585 zero_memory_area(ps);
586 r = zero_disk_area(ps, 0);
588 DMWARN("zero_disk_area(0) failed");
594 if (ps->version != SNAPSHOT_DISK_VERSION) {
595 DMWARN("unable to handle snapshot disk version %d",
601 * Metadata are valid, but snapshot is invalidated
609 r = read_exceptions(ps, callback, callback_context);
614 static int persistent_prepare_exception(struct dm_exception_store *store,
615 struct dm_exception *e)
617 struct pstore *ps = get_info(store);
620 sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
622 /* Is there enough room ? */
623 if (size < ((ps->next_free + 1) * store->chunk_size))
626 e->new_chunk = ps->next_free;
629 * Move onto the next free pending, making sure to take
630 * into account the location of the metadata chunks.
632 stride = (ps->exceptions_per_area + 1);
633 next_free = ++ps->next_free;
634 if (sector_div(next_free, stride) == 1)
637 atomic_inc(&ps->pending_count);
641 static void persistent_commit_exception(struct dm_exception_store *store,
642 struct dm_exception *e,
643 void (*callback) (void *, int success),
644 void *callback_context)
647 struct pstore *ps = get_info(store);
648 struct core_exception ce;
649 struct commit_callback *cb;
651 ce.old_chunk = e->old_chunk;
652 ce.new_chunk = e->new_chunk;
653 write_exception(ps, ps->current_committed++, &ce);
656 * Add the callback to the back of the array. This code
657 * is the only place where the callback array is
658 * manipulated, and we know that it will never be called
659 * multiple times concurrently.
661 cb = ps->callbacks + ps->callback_count++;
662 cb->callback = callback;
663 cb->context = callback_context;
666 * If there are exceptions in flight and we have not yet
667 * filled this metadata area there's nothing more to do.
669 if (!atomic_dec_and_test(&ps->pending_count) &&
670 (ps->current_committed != ps->exceptions_per_area))
674 * If we completely filled the current area, then wipe the next one.
676 if ((ps->current_committed == ps->exceptions_per_area) &&
677 zero_disk_area(ps, ps->current_area + 1))
681 * Commit exceptions to disk.
683 if (ps->valid && area_io(ps, WRITE_FLUSH_FUA))
687 * Advance to the next area if this one is full.
689 if (ps->current_committed == ps->exceptions_per_area) {
690 ps->current_committed = 0;
692 zero_memory_area(ps);
695 for (i = 0; i < ps->callback_count; i++) {
696 cb = ps->callbacks + i;
697 cb->callback(cb->context, ps->valid);
700 ps->callback_count = 0;
703 static int persistent_prepare_merge(struct dm_exception_store *store,
704 chunk_t *last_old_chunk,
705 chunk_t *last_new_chunk)
707 struct pstore *ps = get_info(store);
708 struct core_exception ce;
713 * When current area is empty, move back to preceding area.
715 if (!ps->current_committed) {
719 if (!ps->current_area)
723 r = area_io(ps, READ);
726 ps->current_committed = ps->exceptions_per_area;
729 read_exception(ps, ps->current_committed - 1, &ce);
730 *last_old_chunk = ce.old_chunk;
731 *last_new_chunk = ce.new_chunk;
734 * Find number of consecutive chunks within the current area,
737 for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
739 read_exception(ps, ps->current_committed - 1 - nr_consecutive,
741 if (ce.old_chunk != *last_old_chunk - nr_consecutive ||
742 ce.new_chunk != *last_new_chunk - nr_consecutive)
746 return nr_consecutive;
749 static int persistent_commit_merge(struct dm_exception_store *store,
753 struct pstore *ps = get_info(store);
755 BUG_ON(nr_merged > ps->current_committed);
757 for (i = 0; i < nr_merged; i++)
758 clear_exception(ps, ps->current_committed - 1 - i);
760 r = area_io(ps, WRITE_FLUSH_FUA);
764 ps->current_committed -= nr_merged;
767 * At this stage, only persistent_usage() uses ps->next_free, so
768 * we make no attempt to keep ps->next_free strictly accurate
769 * as exceptions may have been committed out-of-order originally.
770 * Once a snapshot has become merging, we set it to the value it
771 * would have held had all the exceptions been committed in order.
773 * ps->current_area does not get reduced by prepare_merge() until
774 * after commit_merge() has removed the nr_merged previous exceptions.
776 ps->next_free = area_location(ps, ps->current_area) +
777 ps->current_committed + 1;
782 static void persistent_drop_snapshot(struct dm_exception_store *store)
784 struct pstore *ps = get_info(store);
787 if (write_header(ps))
788 DMWARN("write header failed");
791 static int persistent_ctr(struct dm_exception_store *store,
792 unsigned argc, char **argv)
796 /* allocate the pstore */
797 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
803 ps->version = SNAPSHOT_DISK_VERSION;
805 ps->zero_area = NULL;
806 ps->header_area = NULL;
807 ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
808 ps->current_committed = 0;
810 ps->callback_count = 0;
811 atomic_set(&ps->pending_count, 0);
812 ps->callbacks = NULL;
814 ps->metadata_wq = alloc_workqueue("ksnaphd", WQ_MEM_RECLAIM, 0);
815 if (!ps->metadata_wq) {
817 DMERR("couldn't start header metadata update thread");
826 static unsigned persistent_status(struct dm_exception_store *store,
827 status_type_t status, char *result,
833 case STATUSTYPE_INFO:
835 case STATUSTYPE_TABLE:
836 DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
842 static struct dm_exception_store_type _persistent_type = {
843 .name = "persistent",
844 .module = THIS_MODULE,
845 .ctr = persistent_ctr,
846 .dtr = persistent_dtr,
847 .read_metadata = persistent_read_metadata,
848 .prepare_exception = persistent_prepare_exception,
849 .commit_exception = persistent_commit_exception,
850 .prepare_merge = persistent_prepare_merge,
851 .commit_merge = persistent_commit_merge,
852 .drop_snapshot = persistent_drop_snapshot,
853 .usage = persistent_usage,
854 .status = persistent_status,
857 static struct dm_exception_store_type _persistent_compat_type = {
859 .module = THIS_MODULE,
860 .ctr = persistent_ctr,
861 .dtr = persistent_dtr,
862 .read_metadata = persistent_read_metadata,
863 .prepare_exception = persistent_prepare_exception,
864 .commit_exception = persistent_commit_exception,
865 .prepare_merge = persistent_prepare_merge,
866 .commit_merge = persistent_commit_merge,
867 .drop_snapshot = persistent_drop_snapshot,
868 .usage = persistent_usage,
869 .status = persistent_status,
872 int dm_persistent_snapshot_init(void)
876 r = dm_exception_store_type_register(&_persistent_type);
878 DMERR("Unable to register persistent exception store type");
882 r = dm_exception_store_type_register(&_persistent_compat_type);
884 DMERR("Unable to register old-style persistent exception "
886 dm_exception_store_type_unregister(&_persistent_type);
893 void dm_persistent_snapshot_exit(void)
895 dm_exception_store_type_unregister(&_persistent_type);
896 dm_exception_store_type_unregister(&_persistent_compat_type);