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 uint32_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, uint32_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 * Read or write a metadata area. Remembering to skip the first
213 * chunk which holds the header.
215 static int area_io(struct pstore *ps, uint32_t area, int rw)
220 /* convert a metadata area index to a chunk index */
221 chunk = 1 + ((ps->exceptions_per_area + 1) * area);
223 r = chunk_io(ps, chunk, rw, 0);
227 ps->current_area = area;
231 static int zero_area(struct pstore *ps, uint32_t area)
233 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
234 return area_io(ps, area, WRITE);
237 static int read_header(struct pstore *ps, int *new_snapshot)
240 struct disk_header *dh;
242 int chunk_size_supplied = 1;
245 * Use default chunk size (or hardsect_size, if larger) if none supplied
247 if (!ps->snap->chunk_size) {
248 ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
249 bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
250 ps->snap->chunk_mask = ps->snap->chunk_size - 1;
251 ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
252 chunk_size_supplied = 0;
255 ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
257 if (IS_ERR(ps->io_client))
258 return PTR_ERR(ps->io_client);
264 r = chunk_io(ps, 0, READ, 1);
268 dh = (struct disk_header *) ps->area;
270 if (le32_to_cpu(dh->magic) == 0) {
275 if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
276 DMWARN("Invalid or corrupt snapshot");
282 ps->valid = le32_to_cpu(dh->valid);
283 ps->version = le32_to_cpu(dh->version);
284 chunk_size = le32_to_cpu(dh->chunk_size);
286 if (!chunk_size_supplied || ps->snap->chunk_size == chunk_size)
289 DMWARN("chunk size %llu in device metadata overrides "
290 "table chunk size of %llu.",
291 (unsigned long long)chunk_size,
292 (unsigned long long)ps->snap->chunk_size);
294 /* We had a bogus chunk_size. Fix stuff up. */
297 ps->snap->chunk_size = chunk_size;
298 ps->snap->chunk_mask = chunk_size - 1;
299 ps->snap->chunk_shift = ffs(chunk_size) - 1;
301 r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
314 static int write_header(struct pstore *ps)
316 struct disk_header *dh;
318 memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
320 dh = (struct disk_header *) ps->area;
321 dh->magic = cpu_to_le32(SNAP_MAGIC);
322 dh->valid = cpu_to_le32(ps->valid);
323 dh->version = cpu_to_le32(ps->version);
324 dh->chunk_size = cpu_to_le32(ps->snap->chunk_size);
326 return chunk_io(ps, 0, WRITE, 1);
330 * Access functions for the disk exceptions, these do the endian conversions.
332 static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
334 BUG_ON(index >= ps->exceptions_per_area);
336 return ((struct disk_exception *) ps->area) + index;
339 static void read_exception(struct pstore *ps,
340 uint32_t index, struct disk_exception *result)
342 struct disk_exception *e = get_exception(ps, index);
345 result->old_chunk = le64_to_cpu(e->old_chunk);
346 result->new_chunk = le64_to_cpu(e->new_chunk);
349 static void write_exception(struct pstore *ps,
350 uint32_t index, struct disk_exception *de)
352 struct disk_exception *e = get_exception(ps, index);
355 e->old_chunk = cpu_to_le64(de->old_chunk);
356 e->new_chunk = cpu_to_le64(de->new_chunk);
360 * Registers the exceptions that are present in the current area.
361 * 'full' is filled in to indicate if the area has been
364 static int insert_exceptions(struct pstore *ps, int *full)
368 struct disk_exception de;
370 /* presume the area is full */
373 for (i = 0; i < ps->exceptions_per_area; i++) {
374 read_exception(ps, i, &de);
377 * If the new_chunk is pointing at the start of
378 * the COW device, where the first metadata area
379 * is we know that we've hit the end of the
380 * exceptions. Therefore the area is not full.
382 if (de.new_chunk == 0LL) {
383 ps->current_committed = i;
389 * Keep track of the start of the free chunks.
391 if (ps->next_free <= de.new_chunk)
392 ps->next_free = de.new_chunk + 1;
395 * Otherwise we add the exception to the snapshot.
397 r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk);
405 static int read_exceptions(struct pstore *ps)
411 * Keeping reading chunks and inserting exceptions until
412 * we find a partially full area.
414 for (area = 0; full; area++) {
415 r = area_io(ps, area, READ);
419 r = insert_exceptions(ps, &full);
427 static struct pstore *get_info(struct exception_store *store)
429 return (struct pstore *) store->context;
432 static void persistent_fraction_full(struct exception_store *store,
433 sector_t *numerator, sector_t *denominator)
435 *numerator = get_info(store)->next_free * store->snap->chunk_size;
436 *denominator = get_dev_size(store->snap->cow->bdev);
439 static void persistent_destroy(struct exception_store *store)
441 struct pstore *ps = get_info(store);
443 destroy_workqueue(ps->metadata_wq);
444 dm_io_client_destroy(ps->io_client);
445 vfree(ps->callbacks);
450 static int persistent_read_metadata(struct exception_store *store)
452 int r, uninitialized_var(new_snapshot);
453 struct pstore *ps = get_info(store);
456 * Read the snapshot header.
458 r = read_header(ps, &new_snapshot);
463 * Now we know correct chunk_size, complete the initialisation.
465 ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
466 sizeof(struct disk_exception);
467 ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
468 sizeof(*ps->callbacks));
473 * Do we need to setup a new snapshot ?
476 r = write_header(ps);
478 DMWARN("write_header failed");
482 r = zero_area(ps, 0);
484 DMWARN("zero_area(0) failed");
492 if (ps->version != SNAPSHOT_DISK_VERSION) {
493 DMWARN("unable to handle snapshot disk version %d",
499 * Metadata are valid, but snapshot is invalidated
507 r = read_exceptions(ps);
515 static int persistent_prepare(struct exception_store *store,
516 struct dm_snap_exception *e)
518 struct pstore *ps = get_info(store);
520 sector_t size = get_dev_size(store->snap->cow->bdev);
522 /* Is there enough room ? */
523 if (size < ((ps->next_free + 1) * store->snap->chunk_size))
526 e->new_chunk = ps->next_free;
529 * Move onto the next free pending, making sure to take
530 * into account the location of the metadata chunks.
532 stride = (ps->exceptions_per_area + 1);
533 if ((++ps->next_free % stride) == 1)
536 atomic_inc(&ps->pending_count);
540 static void persistent_commit(struct exception_store *store,
541 struct dm_snap_exception *e,
542 void (*callback) (void *, int success),
543 void *callback_context)
547 struct pstore *ps = get_info(store);
548 struct disk_exception de;
549 struct commit_callback *cb;
551 de.old_chunk = e->old_chunk;
552 de.new_chunk = e->new_chunk;
553 write_exception(ps, ps->current_committed++, &de);
556 * Add the callback to the back of the array. This code
557 * is the only place where the callback array is
558 * manipulated, and we know that it will never be called
559 * multiple times concurrently.
561 cb = ps->callbacks + ps->callback_count++;
562 cb->callback = callback;
563 cb->context = callback_context;
566 * If there are no more exceptions in flight, or we have
567 * filled this metadata area we commit the exceptions to
570 if (atomic_dec_and_test(&ps->pending_count) ||
571 (ps->current_committed == ps->exceptions_per_area)) {
572 r = area_io(ps, ps->current_area, WRITE);
577 * Have we completely filled the current area ?
579 if (ps->current_committed == ps->exceptions_per_area) {
580 ps->current_committed = 0;
581 r = zero_area(ps, ps->current_area + 1);
586 for (i = 0; i < ps->callback_count; i++) {
587 cb = ps->callbacks + i;
588 cb->callback(cb->context, r == 0 ? 1 : 0);
591 ps->callback_count = 0;
595 static void persistent_drop(struct exception_store *store)
597 struct pstore *ps = get_info(store);
600 if (write_header(ps))
601 DMWARN("write header failed");
604 int dm_create_persistent(struct exception_store *store)
608 /* allocate the pstore */
609 ps = kmalloc(sizeof(*ps), GFP_KERNEL);
613 ps->snap = store->snap;
615 ps->version = SNAPSHOT_DISK_VERSION;
617 ps->next_free = 2; /* skipping the header and first area */
618 ps->current_committed = 0;
620 ps->callback_count = 0;
621 atomic_set(&ps->pending_count, 0);
622 ps->callbacks = NULL;
624 ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
625 if (!ps->metadata_wq) {
627 DMERR("couldn't start header metadata update thread");
631 store->destroy = persistent_destroy;
632 store->read_metadata = persistent_read_metadata;
633 store->prepare_exception = persistent_prepare;
634 store->commit_exception = persistent_commit;
635 store->drop_snapshot = persistent_drop;
636 store->fraction_full = persistent_fraction_full;
642 /*-----------------------------------------------------------------
643 * Implementation of the store for non-persistent snapshots.
644 *---------------------------------------------------------------*/
649 static void transient_destroy(struct exception_store *store)
651 kfree(store->context);
654 static int transient_read_metadata(struct exception_store *store)
659 static int transient_prepare(struct exception_store *store,
660 struct dm_snap_exception *e)
662 struct transient_c *tc = (struct transient_c *) store->context;
663 sector_t size = get_dev_size(store->snap->cow->bdev);
665 if (size < (tc->next_free + store->snap->chunk_size))
668 e->new_chunk = sector_to_chunk(store->snap, tc->next_free);
669 tc->next_free += store->snap->chunk_size;
674 static void transient_commit(struct exception_store *store,
675 struct dm_snap_exception *e,
676 void (*callback) (void *, int success),
677 void *callback_context)
680 callback(callback_context, 1);
683 static void transient_fraction_full(struct exception_store *store,
684 sector_t *numerator, sector_t *denominator)
686 *numerator = ((struct transient_c *) store->context)->next_free;
687 *denominator = get_dev_size(store->snap->cow->bdev);
690 int dm_create_transient(struct exception_store *store)
692 struct transient_c *tc;
694 store->destroy = transient_destroy;
695 store->read_metadata = transient_read_metadata;
696 store->prepare_exception = transient_prepare;
697 store->commit_exception = transient_commit;
698 store->drop_snapshot = NULL;
699 store->fraction_full = transient_fraction_full;
701 tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);