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
62 * Is this snapshot valid. There is no way of recovering
63 * an invalid snapshot.
68 * Simple, incrementing version. no backward
77 struct disk_exception {
82 struct commit_callback {
83 void (*callback)(void *, int success);
88 * The top level structure for a persistent exception store.
91 struct dm_exception_store *store;
94 uint32_t exceptions_per_area;
97 * Now that we have an asynchronous kcopyd there is no
98 * need for large chunk sizes, so it wont hurt to have a
99 * whole chunks worth of metadata in memory at once.
104 * An area of zeros used to clear the next area.
109 * An area used for header. The header can be written
110 * concurrently with metadata (when invalidating the snapshot),
111 * so it needs a separate buffer.
116 * Used to keep track of which metadata area the data in
119 chunk_t current_area;
122 * The next free chunk for an exception.
127 * The index of next free exception in the current
130 uint32_t current_committed;
132 atomic_t pending_count;
133 uint32_t callback_count;
134 struct commit_callback *callbacks;
135 struct dm_io_client *io_client;
137 struct workqueue_struct *metadata_wq;
140 static unsigned sectors_to_pages(unsigned sectors)
142 return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
145 static int alloc_area(struct pstore *ps)
150 len = ps->store->chunk_size << SECTOR_SHIFT;
153 * Allocate the chunk_size block of memory that will hold
154 * a single metadata area.
156 ps->area = vmalloc(len);
160 ps->zero_area = vmalloc(len);
163 memset(ps->zero_area, 0, len);
165 ps->header_area = vmalloc(len);
166 if (!ps->header_area)
167 goto err_header_area;
172 vfree(ps->zero_area);
181 static void free_area(struct pstore *ps)
188 vfree(ps->zero_area);
189 ps->zero_area = NULL;
192 vfree(ps->header_area);
193 ps->header_area = NULL;
197 struct dm_io_region *where;
198 struct dm_io_request *io_req;
199 struct work_struct work;
203 static void do_metadata(struct work_struct *work)
205 struct mdata_req *req = container_of(work, struct mdata_req, work);
207 req->result = dm_io(req->io_req, 1, req->where, NULL);
211 * Read or write a chunk aligned and sized block of data from a device.
213 static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, int rw,
216 struct dm_io_region where = {
217 .bdev = dm_snap_cow(ps->store->snap)->bdev,
218 .sector = ps->store->chunk_size * chunk,
219 .count = ps->store->chunk_size,
221 struct dm_io_request io_req = {
223 .mem.type = DM_IO_VMA,
225 .client = ps->io_client,
228 struct mdata_req req;
231 return dm_io(&io_req, 1, &where, NULL);
234 req.io_req = &io_req;
237 * Issue the synchronous I/O from a different thread
238 * to avoid generic_make_request recursion.
240 INIT_WORK(&req.work, do_metadata);
241 queue_work(ps->metadata_wq, &req.work);
242 flush_workqueue(ps->metadata_wq);
248 * Convert a metadata area index to a chunk index.
250 static chunk_t area_location(struct pstore *ps, chunk_t area)
252 return 1 + ((ps->exceptions_per_area + 1) * area);
256 * Read or write a metadata area. Remembering to skip the first
257 * chunk which holds the header.
259 static int area_io(struct pstore *ps, int rw)
264 chunk = area_location(ps, ps->current_area);
266 r = chunk_io(ps, ps->area, chunk, rw, 0);
273 static void zero_memory_area(struct pstore *ps)
275 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
278 static int zero_disk_area(struct pstore *ps, chunk_t area)
280 return chunk_io(ps, ps->zero_area, area_location(ps, area), WRITE, 0);
283 static int read_header(struct pstore *ps, int *new_snapshot)
286 struct disk_header *dh;
288 int chunk_size_supplied = 1;
292 * Use default chunk size (or logical_block_size, if larger)
295 if (!ps->store->chunk_size) {
296 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
297 bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
299 ps->store->chunk_mask = ps->store->chunk_size - 1;
300 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
301 chunk_size_supplied = 0;
304 ps->io_client = dm_io_client_create(sectors_to_pages(ps->store->
306 if (IS_ERR(ps->io_client))
307 return PTR_ERR(ps->io_client);
313 r = chunk_io(ps, ps->header_area, 0, READ, 1);
317 dh = ps->header_area;
319 if (le32_to_cpu(dh->magic) == 0) {
324 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
325 DMWARN("Invalid or corrupt snapshot");
331 ps->valid = le32_to_cpu(dh->valid);
332 ps->version = le32_to_cpu(dh->version);
333 chunk_size = le32_to_cpu(dh->chunk_size);
335 if (ps->store->chunk_size == chunk_size)
338 if (chunk_size_supplied)
339 DMWARN("chunk size %u in device metadata overrides "
340 "table chunk size of %u.",
341 chunk_size, ps->store->chunk_size);
343 /* We had a bogus chunk_size. Fix stuff up. */
346 r = dm_exception_store_set_chunk_size(ps->store, chunk_size,
349 DMERR("invalid on-disk chunk size %u: %s.",
350 chunk_size, chunk_err);
354 r = dm_io_client_resize(sectors_to_pages(ps->store->chunk_size),
367 static int write_header(struct pstore *ps)
369 struct disk_header *dh;
371 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
373 dh = ps->header_area;
374 dh->magic = cpu_to_le32(SNAP_MAGIC);
375 dh->valid = cpu_to_le32(ps->valid);
376 dh->version = cpu_to_le32(ps->version);
377 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
379 return chunk_io(ps, ps->header_area, 0, WRITE, 1);
383 * Access functions for the disk exceptions, these do the endian conversions.
385 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
387 BUG_ON(index >= ps->exceptions_per_area);
389 return ((struct disk_exception *) ps->area) + index;
392 static void read_exception(struct pstore *ps,
393 uint32_t index, struct disk_exception *result)
395 struct disk_exception *e = get_exception(ps, index);
398 result->old_chunk = le64_to_cpu(e->old_chunk);
399 result->new_chunk = le64_to_cpu(e->new_chunk);
402 static void write_exception(struct pstore *ps,
403 uint32_t index, struct disk_exception *de)
405 struct disk_exception *e = get_exception(ps, index);
408 e->old_chunk = cpu_to_le64(de->old_chunk);
409 e->new_chunk = cpu_to_le64(de->new_chunk);
413 * Registers the exceptions that are present in the current area.
414 * 'full' is filled in to indicate if the area has been
417 static int insert_exceptions(struct pstore *ps,
418 int (*callback)(void *callback_context,
419 chunk_t old, chunk_t new),
420 void *callback_context,
425 struct disk_exception de;
427 /* presume the area is full */
430 for (i = 0; i < ps->exceptions_per_area; i++) {
431 read_exception(ps, i, &de);
434 * If the new_chunk is pointing at the start of
435 * the COW device, where the first metadata area
436 * is we know that we've hit the end of the
437 * exceptions. Therefore the area is not full.
439 if (de.new_chunk == 0LL) {
440 ps->current_committed = i;
446 * Keep track of the start of the free chunks.
448 if (ps->next_free <= de.new_chunk)
449 ps->next_free = de.new_chunk + 1;
452 * Otherwise we add the exception to the snapshot.
454 r = callback(callback_context, de.old_chunk, de.new_chunk);
462 static int read_exceptions(struct pstore *ps,
463 int (*callback)(void *callback_context, chunk_t old,
465 void *callback_context)
470 * Keeping reading chunks and inserting exceptions until
471 * we find a partially full area.
473 for (ps->current_area = 0; full; ps->current_area++) {
474 r = area_io(ps, READ);
478 r = insert_exceptions(ps, callback, callback_context, &full);
488 static struct pstore *get_info(struct dm_exception_store *store)
490 return (struct pstore *) store->context;
493 static void persistent_usage(struct dm_exception_store *store,
494 sector_t *total_sectors,
495 sector_t *sectors_allocated,
496 sector_t *metadata_sectors)
498 struct pstore *ps = get_info(store);
500 *sectors_allocated = ps->next_free * store->chunk_size;
501 *total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
504 * First chunk is the fixed header.
505 * Then there are (ps->current_area + 1) metadata chunks, each one
506 * separated from the next by ps->exceptions_per_area data chunks.
508 *metadata_sectors = (ps->current_area + 2) * store->chunk_size;
511 static void persistent_dtr(struct dm_exception_store *store)
513 struct pstore *ps = get_info(store);
515 destroy_workqueue(ps->metadata_wq);
517 /* Created in read_header */
519 dm_io_client_destroy(ps->io_client);
522 /* Allocated in persistent_read_metadata */
524 vfree(ps->callbacks);
529 static int persistent_read_metadata(struct dm_exception_store *store,
530 int (*callback)(void *callback_context,
531 chunk_t old, chunk_t new),
532 void *callback_context)
534 int r, uninitialized_var(new_snapshot);
535 struct pstore *ps = get_info(store);
538 * Read the snapshot header.
540 r = read_header(ps, &new_snapshot);
545 * Now we know correct chunk_size, complete the initialisation.
547 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
548 sizeof(struct disk_exception);
549 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
550 sizeof(*ps->callbacks));
555 * Do we need to setup a new snapshot ?
558 r = write_header(ps);
560 DMWARN("write_header failed");
564 ps->current_area = 0;
565 zero_memory_area(ps);
566 r = zero_disk_area(ps, 0);
568 DMWARN("zero_disk_area(0) failed");
574 if (ps->version != SNAPSHOT_DISK_VERSION) {
575 DMWARN("unable to handle snapshot disk version %d",
581 * Metadata are valid, but snapshot is invalidated
589 r = read_exceptions(ps, callback, callback_context);
594 static int persistent_prepare_exception(struct dm_exception_store *store,
595 struct dm_exception *e)
597 struct pstore *ps = get_info(store);
600 sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
602 /* Is there enough room ? */
603 if (size < ((ps->next_free + 1) * store->chunk_size))
606 e->new_chunk = ps->next_free;
609 * Move onto the next free pending, making sure to take
610 * into account the location of the metadata chunks.
612 stride = (ps->exceptions_per_area + 1);
613 next_free = ++ps->next_free;
614 if (sector_div(next_free, stride) == 1)
617 atomic_inc(&ps->pending_count);
621 static void persistent_commit_exception(struct dm_exception_store *store,
622 struct dm_exception *e,
623 void (*callback) (void *, int success),
624 void *callback_context)
627 struct pstore *ps = get_info(store);
628 struct disk_exception de;
629 struct commit_callback *cb;
631 de.old_chunk = e->old_chunk;
632 de.new_chunk = e->new_chunk;
633 write_exception(ps, ps->current_committed++, &de);
636 * Add the callback to the back of the array. This code
637 * is the only place where the callback array is
638 * manipulated, and we know that it will never be called
639 * multiple times concurrently.
641 cb = ps->callbacks + ps->callback_count++;
642 cb->callback = callback;
643 cb->context = callback_context;
646 * If there are exceptions in flight and we have not yet
647 * filled this metadata area there's nothing more to do.
649 if (!atomic_dec_and_test(&ps->pending_count) &&
650 (ps->current_committed != ps->exceptions_per_area))
654 * If we completely filled the current area, then wipe the next one.
656 if ((ps->current_committed == ps->exceptions_per_area) &&
657 zero_disk_area(ps, ps->current_area + 1))
661 * Commit exceptions to disk.
663 if (ps->valid && area_io(ps, WRITE_BARRIER))
667 * Advance to the next area if this one is full.
669 if (ps->current_committed == ps->exceptions_per_area) {
670 ps->current_committed = 0;
672 zero_memory_area(ps);
675 for (i = 0; i < ps->callback_count; i++) {
676 cb = ps->callbacks + i;
677 cb->callback(cb->context, ps->valid);
680 ps->callback_count = 0;
683 static void persistent_drop_snapshot(struct dm_exception_store *store)
685 struct pstore *ps = get_info(store);
688 if (write_header(ps))
689 DMWARN("write header failed");
692 static int persistent_ctr(struct dm_exception_store *store,
693 unsigned argc, char **argv)
697 /* allocate the pstore */
698 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
704 ps->version = SNAPSHOT_DISK_VERSION;
706 ps->zero_area = NULL;
707 ps->header_area = NULL;
708 ps->next_free = 2; /* skipping the header and first area */
709 ps->current_committed = 0;
711 ps->callback_count = 0;
712 atomic_set(&ps->pending_count, 0);
713 ps->callbacks = NULL;
715 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
716 if (!ps->metadata_wq) {
718 DMERR("couldn't start header metadata update thread");
727 static unsigned persistent_status(struct dm_exception_store *store,
728 status_type_t status, char *result,
734 case STATUSTYPE_INFO:
736 case STATUSTYPE_TABLE:
737 DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
743 static struct dm_exception_store_type _persistent_type = {
744 .name = "persistent",
745 .module = THIS_MODULE,
746 .ctr = persistent_ctr,
747 .dtr = persistent_dtr,
748 .read_metadata = persistent_read_metadata,
749 .prepare_exception = persistent_prepare_exception,
750 .commit_exception = persistent_commit_exception,
751 .drop_snapshot = persistent_drop_snapshot,
752 .usage = persistent_usage,
753 .status = persistent_status,
756 static struct dm_exception_store_type _persistent_compat_type = {
758 .module = THIS_MODULE,
759 .ctr = persistent_ctr,
760 .dtr = persistent_dtr,
761 .read_metadata = persistent_read_metadata,
762 .prepare_exception = persistent_prepare_exception,
763 .commit_exception = persistent_commit_exception,
764 .drop_snapshot = persistent_drop_snapshot,
765 .usage = persistent_usage,
766 .status = persistent_status,
769 int dm_persistent_snapshot_init(void)
773 r = dm_exception_store_type_register(&_persistent_type);
775 DMERR("Unable to register persistent exception store type");
779 r = dm_exception_store_type_register(&_persistent_compat_type);
781 DMERR("Unable to register old-style persistent exception "
783 dm_exception_store_type_unregister(&_persistent_type);
790 void dm_persistent_snapshot_exit(void)
792 dm_exception_store_type_unregister(&_persistent_type);
793 dm_exception_store_type_unregister(&_persistent_compat_type);