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 * Used to keep track of which metadata area the data in
112 chunk_t current_area;
115 * The next free chunk for an exception.
120 * The index of next free exception in the current
123 uint32_t current_committed;
125 atomic_t pending_count;
126 uint32_t callback_count;
127 struct commit_callback *callbacks;
128 struct dm_io_client *io_client;
130 struct workqueue_struct *metadata_wq;
133 static unsigned sectors_to_pages(unsigned sectors)
135 return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
138 static int alloc_area(struct pstore *ps)
143 len = ps->store->chunk_size << SECTOR_SHIFT;
146 * Allocate the chunk_size block of memory that will hold
147 * a single metadata area.
149 ps->area = vmalloc(len);
153 ps->zero_area = vmalloc(len);
154 if (!ps->zero_area) {
158 memset(ps->zero_area, 0, len);
163 static void free_area(struct pstore *ps)
170 vfree(ps->zero_area);
171 ps->zero_area = NULL;
175 struct dm_io_region *where;
176 struct dm_io_request *io_req;
177 struct work_struct work;
181 static void do_metadata(struct work_struct *work)
183 struct mdata_req *req = container_of(work, struct mdata_req, work);
185 req->result = dm_io(req->io_req, 1, req->where, NULL);
189 * Read or write a chunk aligned and sized block of data from a device.
191 static int chunk_io(struct pstore *ps, chunk_t chunk, int rw, int metadata)
193 struct dm_io_region where = {
194 .bdev = ps->store->cow->bdev,
195 .sector = ps->store->chunk_size * chunk,
196 .count = ps->store->chunk_size,
198 struct dm_io_request io_req = {
200 .mem.type = DM_IO_VMA,
201 .mem.ptr.vma = ps->area,
202 .client = ps->io_client,
205 struct mdata_req req;
208 return dm_io(&io_req, 1, &where, NULL);
211 req.io_req = &io_req;
214 * Issue the synchronous I/O from a different thread
215 * to avoid generic_make_request recursion.
217 INIT_WORK(&req.work, do_metadata);
218 queue_work(ps->metadata_wq, &req.work);
219 flush_workqueue(ps->metadata_wq);
225 * Convert a metadata area index to a chunk index.
227 static chunk_t area_location(struct pstore *ps, chunk_t area)
229 return 1 + ((ps->exceptions_per_area + 1) * area);
233 * Read or write a metadata area. Remembering to skip the first
234 * chunk which holds the header.
236 static int area_io(struct pstore *ps, int rw)
241 chunk = area_location(ps, ps->current_area);
243 r = chunk_io(ps, chunk, rw, 0);
250 static void zero_memory_area(struct pstore *ps)
252 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
255 static int zero_disk_area(struct pstore *ps, chunk_t area)
257 struct dm_io_region where = {
258 .bdev = ps->store->cow->bdev,
259 .sector = ps->store->chunk_size * area_location(ps, area),
260 .count = ps->store->chunk_size,
262 struct dm_io_request io_req = {
264 .mem.type = DM_IO_VMA,
265 .mem.ptr.vma = ps->zero_area,
266 .client = ps->io_client,
270 return dm_io(&io_req, 1, &where, NULL);
273 static int read_header(struct pstore *ps, int *new_snapshot)
276 struct disk_header *dh;
278 int chunk_size_supplied = 1;
281 * Use default chunk size (or hardsect_size, if larger) if none supplied
283 if (!ps->store->chunk_size) {
284 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
285 bdev_logical_block_size(ps->store->cow->bdev) >> 9);
286 ps->store->chunk_mask = ps->store->chunk_size - 1;
287 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
288 chunk_size_supplied = 0;
291 ps->io_client = dm_io_client_create(sectors_to_pages(ps->store->
293 if (IS_ERR(ps->io_client))
294 return PTR_ERR(ps->io_client);
300 r = chunk_io(ps, 0, READ, 1);
304 dh = (struct disk_header *) ps->area;
306 if (le32_to_cpu(dh->magic) == 0) {
311 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
312 DMWARN("Invalid or corrupt snapshot");
318 ps->valid = le32_to_cpu(dh->valid);
319 ps->version = le32_to_cpu(dh->version);
320 chunk_size = le32_to_cpu(dh->chunk_size);
322 if (!chunk_size_supplied || ps->store->chunk_size == chunk_size)
325 DMWARN("chunk size %llu in device metadata overrides "
326 "table chunk size of %llu.",
327 (unsigned long long)chunk_size,
328 (unsigned long long)ps->store->chunk_size);
330 /* We had a bogus chunk_size. Fix stuff up. */
333 ps->store->chunk_size = chunk_size;
334 ps->store->chunk_mask = chunk_size - 1;
335 ps->store->chunk_shift = ffs(chunk_size) - 1;
337 r = dm_io_client_resize(sectors_to_pages(ps->store->chunk_size),
350 static int write_header(struct pstore *ps)
352 struct disk_header *dh;
354 memset(ps->area, 0, ps->store->chunk_size << SECTOR_SHIFT);
356 dh = (struct disk_header *) ps->area;
357 dh->magic = cpu_to_le32(SNAP_MAGIC);
358 dh->valid = cpu_to_le32(ps->valid);
359 dh->version = cpu_to_le32(ps->version);
360 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
362 return chunk_io(ps, 0, WRITE, 1);
366 * Access functions for the disk exceptions, these do the endian conversions.
368 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
370 BUG_ON(index >= ps->exceptions_per_area);
372 return ((struct disk_exception *) ps->area) + index;
375 static void read_exception(struct pstore *ps,
376 uint32_t index, struct disk_exception *result)
378 struct disk_exception *e = get_exception(ps, index);
381 result->old_chunk = le64_to_cpu(e->old_chunk);
382 result->new_chunk = le64_to_cpu(e->new_chunk);
385 static void write_exception(struct pstore *ps,
386 uint32_t index, struct disk_exception *de)
388 struct disk_exception *e = get_exception(ps, index);
391 e->old_chunk = cpu_to_le64(de->old_chunk);
392 e->new_chunk = cpu_to_le64(de->new_chunk);
396 * Registers the exceptions that are present in the current area.
397 * 'full' is filled in to indicate if the area has been
400 static int insert_exceptions(struct pstore *ps,
401 int (*callback)(void *callback_context,
402 chunk_t old, chunk_t new),
403 void *callback_context,
408 struct disk_exception de;
410 /* presume the area is full */
413 for (i = 0; i < ps->exceptions_per_area; i++) {
414 read_exception(ps, i, &de);
417 * If the new_chunk is pointing at the start of
418 * the COW device, where the first metadata area
419 * is we know that we've hit the end of the
420 * exceptions. Therefore the area is not full.
422 if (de.new_chunk == 0LL) {
423 ps->current_committed = i;
429 * Keep track of the start of the free chunks.
431 if (ps->next_free <= de.new_chunk)
432 ps->next_free = de.new_chunk + 1;
435 * Otherwise we add the exception to the snapshot.
437 r = callback(callback_context, de.old_chunk, de.new_chunk);
445 static int read_exceptions(struct pstore *ps,
446 int (*callback)(void *callback_context, chunk_t old,
448 void *callback_context)
453 * Keeping reading chunks and inserting exceptions until
454 * we find a partially full area.
456 for (ps->current_area = 0; full; ps->current_area++) {
457 r = area_io(ps, READ);
461 r = insert_exceptions(ps, callback, callback_context, &full);
471 static struct pstore *get_info(struct dm_exception_store *store)
473 return (struct pstore *) store->context;
476 static void persistent_fraction_full(struct dm_exception_store *store,
477 sector_t *numerator, sector_t *denominator)
479 *numerator = get_info(store)->next_free * store->chunk_size;
480 *denominator = get_dev_size(store->cow->bdev);
483 static void persistent_dtr(struct dm_exception_store *store)
485 struct pstore *ps = get_info(store);
487 destroy_workqueue(ps->metadata_wq);
489 /* Created in read_header */
491 dm_io_client_destroy(ps->io_client);
494 /* Allocated in persistent_read_metadata */
496 vfree(ps->callbacks);
501 static int persistent_read_metadata(struct dm_exception_store *store,
502 int (*callback)(void *callback_context,
503 chunk_t old, chunk_t new),
504 void *callback_context)
506 int r, uninitialized_var(new_snapshot);
507 struct pstore *ps = get_info(store);
510 * Read the snapshot header.
512 r = read_header(ps, &new_snapshot);
517 * Now we know correct chunk_size, complete the initialisation.
519 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
520 sizeof(struct disk_exception);
521 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
522 sizeof(*ps->callbacks));
527 * Do we need to setup a new snapshot ?
530 r = write_header(ps);
532 DMWARN("write_header failed");
536 ps->current_area = 0;
537 zero_memory_area(ps);
538 r = zero_disk_area(ps, 0);
540 DMWARN("zero_disk_area(0) failed");
547 if (ps->version != SNAPSHOT_DISK_VERSION) {
548 DMWARN("unable to handle snapshot disk version %d",
554 * Metadata are valid, but snapshot is invalidated
562 r = read_exceptions(ps, callback, callback_context);
570 static int persistent_prepare_exception(struct dm_exception_store *store,
571 struct dm_snap_exception *e)
573 struct pstore *ps = get_info(store);
576 sector_t size = get_dev_size(store->cow->bdev);
578 /* Is there enough room ? */
579 if (size < ((ps->next_free + 1) * store->chunk_size))
582 e->new_chunk = ps->next_free;
585 * Move onto the next free pending, making sure to take
586 * into account the location of the metadata chunks.
588 stride = (ps->exceptions_per_area + 1);
589 next_free = ++ps->next_free;
590 if (sector_div(next_free, stride) == 1)
593 atomic_inc(&ps->pending_count);
597 static void persistent_commit_exception(struct dm_exception_store *store,
598 struct dm_snap_exception *e,
599 void (*callback) (void *, int success),
600 void *callback_context)
603 struct pstore *ps = get_info(store);
604 struct disk_exception de;
605 struct commit_callback *cb;
607 de.old_chunk = e->old_chunk;
608 de.new_chunk = e->new_chunk;
609 write_exception(ps, ps->current_committed++, &de);
612 * Add the callback to the back of the array. This code
613 * is the only place where the callback array is
614 * manipulated, and we know that it will never be called
615 * multiple times concurrently.
617 cb = ps->callbacks + ps->callback_count++;
618 cb->callback = callback;
619 cb->context = callback_context;
622 * If there are exceptions in flight and we have not yet
623 * filled this metadata area there's nothing more to do.
625 if (!atomic_dec_and_test(&ps->pending_count) &&
626 (ps->current_committed != ps->exceptions_per_area))
630 * If we completely filled the current area, then wipe the next one.
632 if ((ps->current_committed == ps->exceptions_per_area) &&
633 zero_disk_area(ps, ps->current_area + 1))
637 * Commit exceptions to disk.
639 if (ps->valid && area_io(ps, WRITE_BARRIER))
643 * Advance to the next area if this one is full.
645 if (ps->current_committed == ps->exceptions_per_area) {
646 ps->current_committed = 0;
648 zero_memory_area(ps);
651 for (i = 0; i < ps->callback_count; i++) {
652 cb = ps->callbacks + i;
653 cb->callback(cb->context, ps->valid);
656 ps->callback_count = 0;
659 static void persistent_drop_snapshot(struct dm_exception_store *store)
661 struct pstore *ps = get_info(store);
664 if (write_header(ps))
665 DMWARN("write header failed");
668 static int persistent_ctr(struct dm_exception_store *store,
669 unsigned argc, char **argv)
673 /* allocate the pstore */
674 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
680 ps->version = SNAPSHOT_DISK_VERSION;
682 ps->next_free = 2; /* skipping the header and first area */
683 ps->current_committed = 0;
685 ps->callback_count = 0;
686 atomic_set(&ps->pending_count, 0);
687 ps->callbacks = NULL;
689 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
690 if (!ps->metadata_wq) {
692 DMERR("couldn't start header metadata update thread");
701 static unsigned persistent_status(struct dm_exception_store *store,
702 status_type_t status, char *result,
708 case STATUSTYPE_INFO:
710 case STATUSTYPE_TABLE:
711 DMEMIT(" %s P %llu", store->cow->name,
712 (unsigned long long)store->chunk_size);
718 static struct dm_exception_store_type _persistent_type = {
719 .name = "persistent",
720 .module = THIS_MODULE,
721 .ctr = persistent_ctr,
722 .dtr = persistent_dtr,
723 .read_metadata = persistent_read_metadata,
724 .prepare_exception = persistent_prepare_exception,
725 .commit_exception = persistent_commit_exception,
726 .drop_snapshot = persistent_drop_snapshot,
727 .fraction_full = persistent_fraction_full,
728 .status = persistent_status,
731 static struct dm_exception_store_type _persistent_compat_type = {
733 .module = THIS_MODULE,
734 .ctr = persistent_ctr,
735 .dtr = persistent_dtr,
736 .read_metadata = persistent_read_metadata,
737 .prepare_exception = persistent_prepare_exception,
738 .commit_exception = persistent_commit_exception,
739 .drop_snapshot = persistent_drop_snapshot,
740 .fraction_full = persistent_fraction_full,
741 .status = persistent_status,
744 int dm_persistent_snapshot_init(void)
748 r = dm_exception_store_type_register(&_persistent_type);
750 DMERR("Unable to register persistent exception store type");
754 r = dm_exception_store_type_register(&_persistent_compat_type);
756 DMERR("Unable to register old-style persistent exception "
758 dm_exception_store_type_unregister(&_persistent_type);
765 void dm_persistent_snapshot_exit(void)
767 dm_exception_store_type_unregister(&_persistent_type);
768 dm_exception_store_type_unregister(&_persistent_compat_type);