4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
5 * Copyright (C) 2006 Red Hat GmbH
7 * This file is released under the GPL.
13 #include <linux/pagemap.h>
14 #include <linux/vmalloc.h>
15 #include <linux/slab.h>
16 #include <linux/dm-io.h>
17 #include <linux/dm-kcopyd.h>
19 #define DM_MSG_PREFIX "snapshots"
20 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
22 /*-----------------------------------------------------------------
23 * Persistent snapshots, by persistent we mean that the snapshot
24 * will survive a reboot.
25 *---------------------------------------------------------------*/
28 * We need to store a record of which parts of the origin have
29 * been copied to the snapshot device. The snapshot code
30 * requires that we copy exception chunks to chunk aligned areas
31 * of the COW store. It makes sense therefore, to store the
32 * metadata in chunk size blocks.
34 * There is no backward or forward compatibility implemented,
35 * snapshots with different disk versions than the kernel will
36 * not be usable. It is expected that "lvcreate" will blank out
37 * the start of a fresh COW device before calling the snapshot
40 * The first chunk of the COW device just contains the header.
41 * After this there is a chunk filled with exception metadata,
42 * followed by as many exception chunks as can fit in the
45 * All on disk structures are in little-endian format. The end
46 * of the exceptions info is indicated by an exception with a
47 * new_chunk of 0, which is invalid since it would point to the
52 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
54 #define SNAP_MAGIC 0x70416e53
57 * The on-disk version of the metadata.
59 #define SNAPSHOT_DISK_VERSION 1
65 * Is this snapshot valid. There is no way of recovering
66 * an invalid snapshot.
71 * Simple, incrementing version. no backward
80 struct disk_exception {
85 struct commit_callback {
86 void (*callback)(void *, int success);
91 * The top level structure for a persistent exception store.
94 struct dm_snapshot *snap; /* up pointer to my snapshot */
97 uint32_t exceptions_per_area;
100 * Now that we have an asynchronous kcopyd there is no
101 * need for large chunk sizes, so it wont hurt to have a
102 * whole chunks worth of metadata in memory at once.
107 * An area of zeros used to clear the next area.
112 * Used to keep track of which metadata area the data in
115 chunk_t current_area;
118 * The next free chunk for an exception.
123 * The index of next free exception in the current
126 uint32_t current_committed;
128 atomic_t pending_count;
129 uint32_t callback_count;
130 struct commit_callback *callbacks;
131 struct dm_io_client *io_client;
133 struct workqueue_struct *metadata_wq;
136 static unsigned sectors_to_pages(unsigned sectors)
138 return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
141 static int alloc_area(struct pstore *ps)
146 len = ps->snap->chunk_size << SECTOR_SHIFT;
149 * Allocate the chunk_size block of memory that will hold
150 * a single metadata area.
152 ps->area = vmalloc(len);
156 ps->zero_area = vmalloc(len);
157 if (!ps->zero_area) {
161 memset(ps->zero_area, 0, len);
166 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->snap->cow->bdev,
195 .sector = ps->snap->chunk_size * chunk,
196 .count = ps->snap->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->snap->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->snap->cow->bdev,
259 .sector = ps->snap->chunk_size * area_location(ps, area),
260 .count = ps->snap->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->snap->chunk_size) {
284 ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
285 bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
286 ps->snap->chunk_mask = ps->snap->chunk_size - 1;
287 ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
288 chunk_size_supplied = 0;
291 ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
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->snap->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->snap->chunk_size);
330 /* We had a bogus chunk_size. Fix stuff up. */
333 ps->snap->chunk_size = chunk_size;
334 ps->snap->chunk_mask = chunk_size - 1;
335 ps->snap->chunk_shift = ffs(chunk_size) - 1;
337 r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
350 static int write_header(struct pstore *ps)
352 struct disk_header *dh;
354 memset(ps->area, 0, ps->snap->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->snap->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, int *full)
404 struct disk_exception de;
406 /* presume the area is full */
409 for (i = 0; i < ps->exceptions_per_area; i++) {
410 read_exception(ps, i, &de);
413 * If the new_chunk is pointing at the start of
414 * the COW device, where the first metadata area
415 * is we know that we've hit the end of the
416 * exceptions. Therefore the area is not full.
418 if (de.new_chunk == 0LL) {
419 ps->current_committed = i;
425 * Keep track of the start of the free chunks.
427 if (ps->next_free <= de.new_chunk)
428 ps->next_free = de.new_chunk + 1;
431 * Otherwise we add the exception to the snapshot.
433 r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk);
441 static int read_exceptions(struct pstore *ps)
446 * Keeping reading chunks and inserting exceptions until
447 * we find a partially full area.
449 for (ps->current_area = 0; full; ps->current_area++) {
450 r = area_io(ps, READ);
454 r = insert_exceptions(ps, &full);
464 static struct pstore *get_info(struct exception_store *store)
466 return (struct pstore *) store->context;
469 static void persistent_fraction_full(struct exception_store *store,
470 sector_t *numerator, sector_t *denominator)
472 *numerator = get_info(store)->next_free * store->snap->chunk_size;
473 *denominator = get_dev_size(store->snap->cow->bdev);
476 static void persistent_destroy(struct exception_store *store)
478 struct pstore *ps = get_info(store);
480 destroy_workqueue(ps->metadata_wq);
481 dm_io_client_destroy(ps->io_client);
482 vfree(ps->callbacks);
487 static int persistent_read_metadata(struct exception_store *store)
489 int r, uninitialized_var(new_snapshot);
490 struct pstore *ps = get_info(store);
493 * Read the snapshot header.
495 r = read_header(ps, &new_snapshot);
500 * Now we know correct chunk_size, complete the initialisation.
502 ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
503 sizeof(struct disk_exception);
504 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
505 sizeof(*ps->callbacks));
510 * Do we need to setup a new snapshot ?
513 r = write_header(ps);
515 DMWARN("write_header failed");
519 ps->current_area = 0;
520 zero_memory_area(ps);
521 r = zero_disk_area(ps, 0);
523 DMWARN("zero_disk_area(0) failed");
530 if (ps->version != SNAPSHOT_DISK_VERSION) {
531 DMWARN("unable to handle snapshot disk version %d",
537 * Metadata are valid, but snapshot is invalidated
545 r = read_exceptions(ps);
553 static int persistent_prepare(struct exception_store *store,
554 struct dm_snap_exception *e)
556 struct pstore *ps = get_info(store);
559 sector_t size = get_dev_size(store->snap->cow->bdev);
561 /* Is there enough room ? */
562 if (size < ((ps->next_free + 1) * store->snap->chunk_size))
565 e->new_chunk = ps->next_free;
568 * Move onto the next free pending, making sure to take
569 * into account the location of the metadata chunks.
571 stride = (ps->exceptions_per_area + 1);
572 next_free = ++ps->next_free;
573 if (sector_div(next_free, stride) == 1)
576 atomic_inc(&ps->pending_count);
580 static void persistent_commit(struct exception_store *store,
581 struct dm_snap_exception *e,
582 void (*callback) (void *, int success),
583 void *callback_context)
586 struct pstore *ps = get_info(store);
587 struct disk_exception de;
588 struct commit_callback *cb;
590 de.old_chunk = e->old_chunk;
591 de.new_chunk = e->new_chunk;
592 write_exception(ps, ps->current_committed++, &de);
595 * Add the callback to the back of the array. This code
596 * is the only place where the callback array is
597 * manipulated, and we know that it will never be called
598 * multiple times concurrently.
600 cb = ps->callbacks + ps->callback_count++;
601 cb->callback = callback;
602 cb->context = callback_context;
605 * If there are exceptions in flight and we have not yet
606 * filled this metadata area there's nothing more to do.
608 if (!atomic_dec_and_test(&ps->pending_count) &&
609 (ps->current_committed != ps->exceptions_per_area))
613 * If we completely filled the current area, then wipe the next one.
615 if ((ps->current_committed == ps->exceptions_per_area) &&
616 zero_disk_area(ps, ps->current_area + 1))
620 * Commit exceptions to disk.
622 if (ps->valid && area_io(ps, WRITE))
626 * Advance to the next area if this one is full.
628 if (ps->current_committed == ps->exceptions_per_area) {
629 ps->current_committed = 0;
631 zero_memory_area(ps);
634 for (i = 0; i < ps->callback_count; i++) {
635 cb = ps->callbacks + i;
636 cb->callback(cb->context, ps->valid);
639 ps->callback_count = 0;
642 static void persistent_drop(struct exception_store *store)
644 struct pstore *ps = get_info(store);
647 if (write_header(ps))
648 DMWARN("write header failed");
651 int dm_create_persistent(struct exception_store *store)
655 /* allocate the pstore */
656 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
660 ps->snap = store->snap;
662 ps->version = SNAPSHOT_DISK_VERSION;
664 ps->next_free = 2; /* skipping the header and first area */
665 ps->current_committed = 0;
667 ps->callback_count = 0;
668 atomic_set(&ps->pending_count, 0);
669 ps->callbacks = NULL;
671 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
672 if (!ps->metadata_wq) {
674 DMERR("couldn't start header metadata update thread");
678 store->destroy = persistent_destroy;
679 store->read_metadata = persistent_read_metadata;
680 store->prepare_exception = persistent_prepare;
681 store->commit_exception = persistent_commit;
682 store->drop_snapshot = persistent_drop;
683 store->fraction_full = persistent_fraction_full;
689 /*-----------------------------------------------------------------
690 * Implementation of the store for non-persistent snapshots.
691 *---------------------------------------------------------------*/
696 static void transient_destroy(struct exception_store *store)
698 kfree(store->context);
701 static int transient_read_metadata(struct exception_store *store)
706 static int transient_prepare(struct exception_store *store,
707 struct dm_snap_exception *e)
709 struct transient_c *tc = (struct transient_c *) store->context;
710 sector_t size = get_dev_size(store->snap->cow->bdev);
712 if (size < (tc->next_free + store->snap->chunk_size))
715 e->new_chunk = sector_to_chunk(store->snap, tc->next_free);
716 tc->next_free += store->snap->chunk_size;
721 static void transient_commit(struct exception_store *store,
722 struct dm_snap_exception *e,
723 void (*callback) (void *, int success),
724 void *callback_context)
727 callback(callback_context, 1);
730 static void transient_fraction_full(struct exception_store *store,
731 sector_t *numerator, sector_t *denominator)
733 *numerator = ((struct transient_c *) store->context)->next_free;
734 *denominator = get_dev_size(store->snap->cow->bdev);
737 int dm_create_transient(struct exception_store *store)
739 struct transient_c *tc;
741 store->destroy = transient_destroy;
742 store->read_metadata = transient_read_metadata;
743 store->prepare_exception = transient_prepare;
744 store->commit_exception = transient_commit;
745 store->drop_snapshot = NULL;
746 store->fraction_full = transient_fraction_full;
748 tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);