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 = ps->store->cow->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;
291 * Use default chunk size (or hardsect_size, if larger) if none supplied
293 if (!ps->store->chunk_size) {
294 ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
295 bdev_logical_block_size(ps->store->cow->bdev) >> 9);
296 ps->store->chunk_mask = ps->store->chunk_size - 1;
297 ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
298 chunk_size_supplied = 0;
301 ps->io_client = dm_io_client_create(sectors_to_pages(ps->store->
303 if (IS_ERR(ps->io_client))
304 return PTR_ERR(ps->io_client);
310 r = chunk_io(ps, ps->header_area, 0, READ, 1);
314 dh = ps->header_area;
316 if (le32_to_cpu(dh->magic) == 0) {
321 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
322 DMWARN("Invalid or corrupt snapshot");
328 ps->valid = le32_to_cpu(dh->valid);
329 ps->version = le32_to_cpu(dh->version);
330 chunk_size = le32_to_cpu(dh->chunk_size);
332 if (!chunk_size_supplied || ps->store->chunk_size == chunk_size)
335 DMWARN("chunk size %llu in device metadata overrides "
336 "table chunk size of %llu.",
337 (unsigned long long)chunk_size,
338 (unsigned long long)ps->store->chunk_size);
340 /* We had a bogus chunk_size. Fix stuff up. */
343 ps->store->chunk_size = chunk_size;
344 ps->store->chunk_mask = chunk_size - 1;
345 ps->store->chunk_shift = ffs(chunk_size) - 1;
347 r = dm_io_client_resize(sectors_to_pages(ps->store->chunk_size),
360 static int write_header(struct pstore *ps)
362 struct disk_header *dh;
364 memset(ps->header_area, 0, ps->store->chunk_size << SECTOR_SHIFT);
366 dh = ps->header_area;
367 dh->magic = cpu_to_le32(SNAP_MAGIC);
368 dh->valid = cpu_to_le32(ps->valid);
369 dh->version = cpu_to_le32(ps->version);
370 dh->chunk_size = cpu_to_le32(ps->store->chunk_size);
372 return chunk_io(ps, ps->header_area, 0, WRITE, 1);
376 * Access functions for the disk exceptions, these do the endian conversions.
378 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
380 BUG_ON(index >= ps->exceptions_per_area);
382 return ((struct disk_exception *) ps->area) + index;
385 static void read_exception(struct pstore *ps,
386 uint32_t index, struct disk_exception *result)
388 struct disk_exception *e = get_exception(ps, index);
391 result->old_chunk = le64_to_cpu(e->old_chunk);
392 result->new_chunk = le64_to_cpu(e->new_chunk);
395 static void write_exception(struct pstore *ps,
396 uint32_t index, struct disk_exception *de)
398 struct disk_exception *e = get_exception(ps, index);
401 e->old_chunk = cpu_to_le64(de->old_chunk);
402 e->new_chunk = cpu_to_le64(de->new_chunk);
406 * Registers the exceptions that are present in the current area.
407 * 'full' is filled in to indicate if the area has been
410 static int insert_exceptions(struct pstore *ps,
411 int (*callback)(void *callback_context,
412 chunk_t old, chunk_t new),
413 void *callback_context,
418 struct disk_exception de;
420 /* presume the area is full */
423 for (i = 0; i < ps->exceptions_per_area; i++) {
424 read_exception(ps, i, &de);
427 * If the new_chunk is pointing at the start of
428 * the COW device, where the first metadata area
429 * is we know that we've hit the end of the
430 * exceptions. Therefore the area is not full.
432 if (de.new_chunk == 0LL) {
433 ps->current_committed = i;
439 * Keep track of the start of the free chunks.
441 if (ps->next_free <= de.new_chunk)
442 ps->next_free = de.new_chunk + 1;
445 * Otherwise we add the exception to the snapshot.
447 r = callback(callback_context, de.old_chunk, de.new_chunk);
455 static int read_exceptions(struct pstore *ps,
456 int (*callback)(void *callback_context, chunk_t old,
458 void *callback_context)
463 * Keeping reading chunks and inserting exceptions until
464 * we find a partially full area.
466 for (ps->current_area = 0; full; ps->current_area++) {
467 r = area_io(ps, READ);
471 r = insert_exceptions(ps, callback, callback_context, &full);
481 static struct pstore *get_info(struct dm_exception_store *store)
483 return (struct pstore *) store->context;
486 static void persistent_fraction_full(struct dm_exception_store *store,
487 sector_t *numerator, sector_t *denominator)
489 *numerator = get_info(store)->next_free * store->chunk_size;
490 *denominator = get_dev_size(store->cow->bdev);
493 static void persistent_dtr(struct dm_exception_store *store)
495 struct pstore *ps = get_info(store);
497 destroy_workqueue(ps->metadata_wq);
499 /* Created in read_header */
501 dm_io_client_destroy(ps->io_client);
504 /* Allocated in persistent_read_metadata */
506 vfree(ps->callbacks);
511 static int persistent_read_metadata(struct dm_exception_store *store,
512 int (*callback)(void *callback_context,
513 chunk_t old, chunk_t new),
514 void *callback_context)
516 int r, uninitialized_var(new_snapshot);
517 struct pstore *ps = get_info(store);
520 * Read the snapshot header.
522 r = read_header(ps, &new_snapshot);
527 * Now we know correct chunk_size, complete the initialisation.
529 ps->exceptions_per_area = (ps->store->chunk_size << SECTOR_SHIFT) /
530 sizeof(struct disk_exception);
531 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
532 sizeof(*ps->callbacks));
537 * Do we need to setup a new snapshot ?
540 r = write_header(ps);
542 DMWARN("write_header failed");
546 ps->current_area = 0;
547 zero_memory_area(ps);
548 r = zero_disk_area(ps, 0);
550 DMWARN("zero_disk_area(0) failed");
557 if (ps->version != SNAPSHOT_DISK_VERSION) {
558 DMWARN("unable to handle snapshot disk version %d",
564 * Metadata are valid, but snapshot is invalidated
572 r = read_exceptions(ps, callback, callback_context);
580 static int persistent_prepare_exception(struct dm_exception_store *store,
581 struct dm_snap_exception *e)
583 struct pstore *ps = get_info(store);
586 sector_t size = get_dev_size(store->cow->bdev);
588 /* Is there enough room ? */
589 if (size < ((ps->next_free + 1) * store->chunk_size))
592 e->new_chunk = ps->next_free;
595 * Move onto the next free pending, making sure to take
596 * into account the location of the metadata chunks.
598 stride = (ps->exceptions_per_area + 1);
599 next_free = ++ps->next_free;
600 if (sector_div(next_free, stride) == 1)
603 atomic_inc(&ps->pending_count);
607 static void persistent_commit_exception(struct dm_exception_store *store,
608 struct dm_snap_exception *e,
609 void (*callback) (void *, int success),
610 void *callback_context)
613 struct pstore *ps = get_info(store);
614 struct disk_exception de;
615 struct commit_callback *cb;
617 de.old_chunk = e->old_chunk;
618 de.new_chunk = e->new_chunk;
619 write_exception(ps, ps->current_committed++, &de);
622 * Add the callback to the back of the array. This code
623 * is the only place where the callback array is
624 * manipulated, and we know that it will never be called
625 * multiple times concurrently.
627 cb = ps->callbacks + ps->callback_count++;
628 cb->callback = callback;
629 cb->context = callback_context;
632 * If there are exceptions in flight and we have not yet
633 * filled this metadata area there's nothing more to do.
635 if (!atomic_dec_and_test(&ps->pending_count) &&
636 (ps->current_committed != ps->exceptions_per_area))
640 * If we completely filled the current area, then wipe the next one.
642 if ((ps->current_committed == ps->exceptions_per_area) &&
643 zero_disk_area(ps, ps->current_area + 1))
647 * Commit exceptions to disk.
649 if (ps->valid && area_io(ps, WRITE_BARRIER))
653 * Advance to the next area if this one is full.
655 if (ps->current_committed == ps->exceptions_per_area) {
656 ps->current_committed = 0;
658 zero_memory_area(ps);
661 for (i = 0; i < ps->callback_count; i++) {
662 cb = ps->callbacks + i;
663 cb->callback(cb->context, ps->valid);
666 ps->callback_count = 0;
669 static void persistent_drop_snapshot(struct dm_exception_store *store)
671 struct pstore *ps = get_info(store);
674 if (write_header(ps))
675 DMWARN("write header failed");
678 static int persistent_ctr(struct dm_exception_store *store,
679 unsigned argc, char **argv)
683 /* allocate the pstore */
684 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
690 ps->version = SNAPSHOT_DISK_VERSION;
692 ps->zero_area = NULL;
693 ps->header_area = NULL;
694 ps->next_free = 2; /* skipping the header and first area */
695 ps->current_committed = 0;
697 ps->callback_count = 0;
698 atomic_set(&ps->pending_count, 0);
699 ps->callbacks = NULL;
701 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
702 if (!ps->metadata_wq) {
704 DMERR("couldn't start header metadata update thread");
713 static unsigned persistent_status(struct dm_exception_store *store,
714 status_type_t status, char *result,
720 case STATUSTYPE_INFO:
722 case STATUSTYPE_TABLE:
723 DMEMIT(" %s P %llu", store->cow->name,
724 (unsigned long long)store->chunk_size);
730 static struct dm_exception_store_type _persistent_type = {
731 .name = "persistent",
732 .module = THIS_MODULE,
733 .ctr = persistent_ctr,
734 .dtr = persistent_dtr,
735 .read_metadata = persistent_read_metadata,
736 .prepare_exception = persistent_prepare_exception,
737 .commit_exception = persistent_commit_exception,
738 .drop_snapshot = persistent_drop_snapshot,
739 .fraction_full = persistent_fraction_full,
740 .status = persistent_status,
743 static struct dm_exception_store_type _persistent_compat_type = {
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 .fraction_full = persistent_fraction_full,
753 .status = persistent_status,
756 int dm_persistent_snapshot_init(void)
760 r = dm_exception_store_type_register(&_persistent_type);
762 DMERR("Unable to register persistent exception store type");
766 r = dm_exception_store_type_register(&_persistent_compat_type);
768 DMERR("Unable to register old-style persistent exception "
770 dm_exception_store_type_unregister(&_persistent_type);
777 void dm_persistent_snapshot_exit(void)
779 dm_exception_store_type_unregister(&_persistent_type);
780 dm_exception_store_type_unregister(&_persistent_compat_type);