4 * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
5 * Copyright (C) 2006 Red Hat GmbH
7 * This file is released under the GPL.
14 #include <linux/pagemap.h>
15 #include <linux/vmalloc.h>
16 #include <linux/slab.h>
17 #include <linux/dm-io.h>
18 #include <linux/dm-kcopyd.h>
20 #define DM_MSG_PREFIX "snapshots"
21 #define DM_CHUNK_SIZE_DEFAULT_SECTORS 32 /* 16KB */
23 /*-----------------------------------------------------------------
24 * Persistent snapshots, by persistent we mean that the snapshot
25 * will survive a reboot.
26 *---------------------------------------------------------------*/
29 * We need to store a record of which parts of the origin have
30 * been copied to the snapshot device. The snapshot code
31 * requires that we copy exception chunks to chunk aligned areas
32 * of the COW store. It makes sense therefore, to store the
33 * metadata in chunk size blocks.
35 * There is no backward or forward compatibility implemented,
36 * snapshots with different disk versions than the kernel will
37 * not be usable. It is expected that "lvcreate" will blank out
38 * the start of a fresh COW device before calling the snapshot
41 * The first chunk of the COW device just contains the header.
42 * After this there is a chunk filled with exception metadata,
43 * followed by as many exception chunks as can fit in the
46 * All on disk structures are in little-endian format. The end
47 * of the exceptions info is indicated by an exception with a
48 * new_chunk of 0, which is invalid since it would point to the
53 * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
55 #define SNAP_MAGIC 0x70416e53
58 * The on-disk version of the metadata.
60 #define SNAPSHOT_DISK_VERSION 1
66 * Is this snapshot valid. There is no way of recovering
67 * an invalid snapshot.
72 * Simple, incrementing version. no backward
81 struct disk_exception {
86 struct commit_callback {
87 void (*callback)(void *, int success);
92 * The top level structure for a persistent exception store.
95 struct dm_snapshot *snap; /* up pointer to my snapshot */
98 uint32_t exceptions_per_area;
101 * Now that we have an asynchronous kcopyd there is no
102 * need for large chunk sizes, so it wont hurt to have a
103 * whole chunks worth of metadata in memory at once.
108 * Used to keep track of which metadata area the data in
111 chunk_t current_area;
114 * The next free chunk for an exception.
119 * The index of next free exception in the current
122 uint32_t current_committed;
124 atomic_t pending_count;
125 uint32_t callback_count;
126 struct commit_callback *callbacks;
127 struct dm_io_client *io_client;
129 struct workqueue_struct *metadata_wq;
132 static unsigned sectors_to_pages(unsigned sectors)
134 return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
137 static int alloc_area(struct pstore *ps)
142 len = ps->snap->chunk_size << SECTOR_SHIFT;
145 * Allocate the chunk_size block of memory that will hold
146 * a single metadata area.
148 ps->area = vmalloc(len);
155 static void free_area(struct pstore *ps)
162 struct dm_io_region *where;
163 struct dm_io_request *io_req;
164 struct work_struct work;
168 static void do_metadata(struct work_struct *work)
170 struct mdata_req *req = container_of(work, struct mdata_req, work);
172 req->result = dm_io(req->io_req, 1, req->where, NULL);
176 * Read or write a chunk aligned and sized block of data from a device.
178 static int chunk_io(struct pstore *ps, chunk_t chunk, int rw, int metadata)
180 struct dm_io_region where = {
181 .bdev = ps->snap->cow->bdev,
182 .sector = ps->snap->chunk_size * chunk,
183 .count = ps->snap->chunk_size,
185 struct dm_io_request io_req = {
187 .mem.type = DM_IO_VMA,
188 .mem.ptr.vma = ps->area,
189 .client = ps->io_client,
192 struct mdata_req req;
195 return dm_io(&io_req, 1, &where, NULL);
198 req.io_req = &io_req;
201 * Issue the synchronous I/O from a different thread
202 * to avoid generic_make_request recursion.
204 INIT_WORK(&req.work, do_metadata);
205 queue_work(ps->metadata_wq, &req.work);
206 flush_workqueue(ps->metadata_wq);
212 * Convert a metadata area index to a chunk index.
214 static chunk_t area_location(struct pstore *ps, chunk_t area)
216 return 1 + ((ps->exceptions_per_area + 1) * area);
220 * Read or write a metadata area. Remembering to skip the first
221 * chunk which holds the header.
223 static int area_io(struct pstore *ps, chunk_t area, int rw)
228 chunk = area_location(ps, area);
230 r = chunk_io(ps, chunk, rw, 0);
234 ps->current_area = area;
238 static int zero_area(struct pstore *ps, chunk_t area)
240 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
241 return area_io(ps, area, WRITE);
244 static int read_header(struct pstore *ps, int *new_snapshot)
247 struct disk_header *dh;
249 int chunk_size_supplied = 1;
252 * Use default chunk size (or hardsect_size, if larger) if none supplied
254 if (!ps->snap->chunk_size) {
255 ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
256 bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
257 ps->snap->chunk_mask = ps->snap->chunk_size - 1;
258 ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
259 chunk_size_supplied = 0;
262 ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
264 if (IS_ERR(ps->io_client))
265 return PTR_ERR(ps->io_client);
271 r = chunk_io(ps, 0, READ, 1);
275 dh = (struct disk_header *) ps->area;
277 if (le32_to_cpu(dh->magic) == 0) {
282 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
283 DMWARN("Invalid or corrupt snapshot");
289 ps->valid = le32_to_cpu(dh->valid);
290 ps->version = le32_to_cpu(dh->version);
291 chunk_size = le32_to_cpu(dh->chunk_size);
293 if (!chunk_size_supplied || ps->snap->chunk_size == chunk_size)
296 DMWARN("chunk size %llu in device metadata overrides "
297 "table chunk size of %llu.",
298 (unsigned long long)chunk_size,
299 (unsigned long long)ps->snap->chunk_size);
301 /* We had a bogus chunk_size. Fix stuff up. */
304 ps->snap->chunk_size = chunk_size;
305 ps->snap->chunk_mask = chunk_size - 1;
306 ps->snap->chunk_shift = ffs(chunk_size) - 1;
308 r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
321 static int write_header(struct pstore *ps)
323 struct disk_header *dh;
325 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
327 dh = (struct disk_header *) ps->area;
328 dh->magic = cpu_to_le32(SNAP_MAGIC);
329 dh->valid = cpu_to_le32(ps->valid);
330 dh->version = cpu_to_le32(ps->version);
331 dh->chunk_size = cpu_to_le32(ps->snap->chunk_size);
333 return chunk_io(ps, 0, WRITE, 1);
337 * Access functions for the disk exceptions, these do the endian conversions.
339 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
341 BUG_ON(index >= ps->exceptions_per_area);
343 return ((struct disk_exception *) ps->area) + index;
346 static void read_exception(struct pstore *ps,
347 uint32_t index, struct disk_exception *result)
349 struct disk_exception *e = get_exception(ps, index);
352 result->old_chunk = le64_to_cpu(e->old_chunk);
353 result->new_chunk = le64_to_cpu(e->new_chunk);
356 static void write_exception(struct pstore *ps,
357 uint32_t index, struct disk_exception *de)
359 struct disk_exception *e = get_exception(ps, index);
362 e->old_chunk = cpu_to_le64(de->old_chunk);
363 e->new_chunk = cpu_to_le64(de->new_chunk);
367 * Registers the exceptions that are present in the current area.
368 * 'full' is filled in to indicate if the area has been
371 static int insert_exceptions(struct pstore *ps, int *full)
375 struct disk_exception de;
377 /* presume the area is full */
380 for (i = 0; i < ps->exceptions_per_area; i++) {
381 read_exception(ps, i, &de);
384 * If the new_chunk is pointing at the start of
385 * the COW device, where the first metadata area
386 * is we know that we've hit the end of the
387 * exceptions. Therefore the area is not full.
389 if (de.new_chunk == 0LL) {
390 ps->current_committed = i;
396 * Keep track of the start of the free chunks.
398 if (ps->next_free <= de.new_chunk)
399 ps->next_free = de.new_chunk + 1;
402 * Otherwise we add the exception to the snapshot.
404 r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk);
412 static int read_exceptions(struct pstore *ps)
418 * Keeping reading chunks and inserting exceptions until
419 * we find a partially full area.
421 for (area = 0; full; area++) {
422 r = area_io(ps, area, READ);
426 r = insert_exceptions(ps, &full);
434 static struct pstore *get_info(struct exception_store *store)
436 return (struct pstore *) store->context;
439 static void persistent_fraction_full(struct exception_store *store,
440 sector_t *numerator, sector_t *denominator)
442 *numerator = get_info(store)->next_free * store->snap->chunk_size;
443 *denominator = get_dev_size(store->snap->cow->bdev);
446 static void persistent_destroy(struct exception_store *store)
448 struct pstore *ps = get_info(store);
450 destroy_workqueue(ps->metadata_wq);
451 dm_io_client_destroy(ps->io_client);
452 vfree(ps->callbacks);
457 static int persistent_read_metadata(struct exception_store *store)
459 int r, uninitialized_var(new_snapshot);
460 struct pstore *ps = get_info(store);
463 * Read the snapshot header.
465 r = read_header(ps, &new_snapshot);
470 * Now we know correct chunk_size, complete the initialisation.
472 ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
473 sizeof(struct disk_exception);
474 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
475 sizeof(*ps->callbacks));
480 * Do we need to setup a new snapshot ?
483 r = write_header(ps);
485 DMWARN("write_header failed");
489 r = zero_area(ps, 0);
491 DMWARN("zero_area(0) failed");
499 if (ps->version != SNAPSHOT_DISK_VERSION) {
500 DMWARN("unable to handle snapshot disk version %d",
506 * Metadata are valid, but snapshot is invalidated
514 r = read_exceptions(ps);
522 static int persistent_prepare(struct exception_store *store,
523 struct dm_snap_exception *e)
525 struct pstore *ps = get_info(store);
528 sector_t size = get_dev_size(store->snap->cow->bdev);
530 /* Is there enough room ? */
531 if (size < ((ps->next_free + 1) * store->snap->chunk_size))
534 e->new_chunk = ps->next_free;
537 * Move onto the next free pending, making sure to take
538 * into account the location of the metadata chunks.
540 stride = (ps->exceptions_per_area + 1);
541 next_free = ++ps->next_free;
542 if (sector_div(next_free, stride) == 1)
545 atomic_inc(&ps->pending_count);
549 static void persistent_commit(struct exception_store *store,
550 struct dm_snap_exception *e,
551 void (*callback) (void *, int success),
552 void *callback_context)
556 struct pstore *ps = get_info(store);
557 struct disk_exception de;
558 struct commit_callback *cb;
560 de.old_chunk = e->old_chunk;
561 de.new_chunk = e->new_chunk;
562 write_exception(ps, ps->current_committed++, &de);
565 * Add the callback to the back of the array. This code
566 * is the only place where the callback array is
567 * manipulated, and we know that it will never be called
568 * multiple times concurrently.
570 cb = ps->callbacks + ps->callback_count++;
571 cb->callback = callback;
572 cb->context = callback_context;
575 * If there are no more exceptions in flight, or we have
576 * filled this metadata area we commit the exceptions to
579 if (atomic_dec_and_test(&ps->pending_count) ||
580 (ps->current_committed == ps->exceptions_per_area)) {
581 r = area_io(ps, ps->current_area, WRITE);
586 * Have we completely filled the current area ?
588 if (ps->current_committed == ps->exceptions_per_area) {
589 ps->current_committed = 0;
590 r = zero_area(ps, ps->current_area + 1);
595 for (i = 0; i < ps->callback_count; i++) {
596 cb = ps->callbacks + i;
597 cb->callback(cb->context, r == 0 ? 1 : 0);
600 ps->callback_count = 0;
604 static void persistent_drop(struct exception_store *store)
606 struct pstore *ps = get_info(store);
609 if (write_header(ps))
610 DMWARN("write header failed");
613 int dm_create_persistent(struct exception_store *store)
617 /* allocate the pstore */
618 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
622 ps->snap = store->snap;
624 ps->version = SNAPSHOT_DISK_VERSION;
626 ps->next_free = 2; /* skipping the header and first area */
627 ps->current_committed = 0;
629 ps->callback_count = 0;
630 atomic_set(&ps->pending_count, 0);
631 ps->callbacks = NULL;
633 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
634 if (!ps->metadata_wq) {
636 DMERR("couldn't start header metadata update thread");
640 store->destroy = persistent_destroy;
641 store->read_metadata = persistent_read_metadata;
642 store->prepare_exception = persistent_prepare;
643 store->commit_exception = persistent_commit;
644 store->drop_snapshot = persistent_drop;
645 store->fraction_full = persistent_fraction_full;
651 /*-----------------------------------------------------------------
652 * Implementation of the store for non-persistent snapshots.
653 *---------------------------------------------------------------*/
658 static void transient_destroy(struct exception_store *store)
660 kfree(store->context);
663 static int transient_read_metadata(struct exception_store *store)
668 static int transient_prepare(struct exception_store *store,
669 struct dm_snap_exception *e)
671 struct transient_c *tc = (struct transient_c *) store->context;
672 sector_t size = get_dev_size(store->snap->cow->bdev);
674 if (size < (tc->next_free + store->snap->chunk_size))
677 e->new_chunk = sector_to_chunk(store->snap, tc->next_free);
678 tc->next_free += store->snap->chunk_size;
683 static void transient_commit(struct exception_store *store,
684 struct dm_snap_exception *e,
685 void (*callback) (void *, int success),
686 void *callback_context)
689 callback(callback_context, 1);
692 static void transient_fraction_full(struct exception_store *store,
693 sector_t *numerator, sector_t *denominator)
695 *numerator = ((struct transient_c *) store->context)->next_free;
696 *denominator = get_dev_size(store->snap->cow->bdev);
699 int dm_create_transient(struct exception_store *store)
701 struct transient_c *tc;
703 store->destroy = transient_destroy;
704 store->read_metadata = transient_read_metadata;
705 store->prepare_exception = transient_prepare;
706 store->commit_exception = transient_commit;
707 store->drop_snapshot = NULL;
708 store->fraction_full = transient_fraction_full;
710 tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);