dm snapshot: split out exception store implementations
Alasdair G Kergon [Tue, 6 Jan 2009 03:05:17 +0000 (03:05 +0000)]
Move the existing snapshot exception store implementations out into
separate files.  Later patches will place these behind a new
interface in preparation for alternative implementations.

Signed-off-by: Alasdair G Kergon <agk@redhat.com>

drivers/md/Makefile
drivers/md/dm-exception-store.c
drivers/md/dm-exception-store.h
drivers/md/dm-snap-persistent.c [new file with mode: 0644]
drivers/md/dm-snap-transient.c [new file with mode: 0644]
drivers/md/dm-snap.c

index 63f0ae9..72880b7 100644 (file)
@@ -5,7 +5,8 @@
 dm-mod-objs    := dm.o dm-table.o dm-target.o dm-linear.o dm-stripe.o \
                   dm-ioctl.o dm-io.o dm-kcopyd.o dm-sysfs.o
 dm-multipath-objs := dm-path-selector.o dm-mpath.o
-dm-snapshot-objs := dm-snap.o dm-exception-store.o
+dm-snapshot-objs := dm-snap.o dm-exception-store.o dm-snap-transient.o \
+                   dm-snap-persistent.o
 dm-mirror-objs := dm-raid1.o
 md-mod-objs     := md.o bitmap.o
 raid456-objs   := raid5.o raid6algos.o raid6recov.o raid6tables.o \
index c5c9a26..74777e0 100644 (file)
 /*
- * dm-exception-store.c
- *
  * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
- * Copyright (C) 2006 Red Hat GmbH
+ * Copyright (C) 2006-2008 Red Hat GmbH
  *
  * This file is released under the GPL.
  */
 
 #include "dm-exception-store.h"
-#include "dm-snap.h"
 
 #include <linux/mm.h>
 #include <linux/pagemap.h>
 #include <linux/vmalloc.h>
 #include <linux/slab.h>
-#include <linux/dm-io.h>
-#include <linux/dm-kcopyd.h>
-
-#define DM_MSG_PREFIX "snapshots"
-#define DM_CHUNK_SIZE_DEFAULT_SECTORS 32       /* 16KB */
-
-/*-----------------------------------------------------------------
- * Persistent snapshots, by persistent we mean that the snapshot
- * will survive a reboot.
- *---------------------------------------------------------------*/
-
-/*
- * We need to store a record of which parts of the origin have
- * been copied to the snapshot device.  The snapshot code
- * requires that we copy exception chunks to chunk aligned areas
- * of the COW store.  It makes sense therefore, to store the
- * metadata in chunk size blocks.
- *
- * There is no backward or forward compatibility implemented,
- * snapshots with different disk versions than the kernel will
- * not be usable.  It is expected that "lvcreate" will blank out
- * the start of a fresh COW device before calling the snapshot
- * constructor.
- *
- * The first chunk of the COW device just contains the header.
- * After this there is a chunk filled with exception metadata,
- * followed by as many exception chunks as can fit in the
- * metadata areas.
- *
- * All on disk structures are in little-endian format.  The end
- * of the exceptions info is indicated by an exception with a
- * new_chunk of 0, which is invalid since it would point to the
- * header chunk.
- */
-
-/*
- * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
- */
-#define SNAP_MAGIC 0x70416e53
-
-/*
- * The on-disk version of the metadata.
- */
-#define SNAPSHOT_DISK_VERSION 1
-
-struct disk_header {
-       uint32_t magic;
-
-       /*
-        * Is this snapshot valid.  There is no way of recovering
-        * an invalid snapshot.
-        */
-       uint32_t valid;
-
-       /*
-        * Simple, incrementing version. no backward
-        * compatibility.
-        */
-       uint32_t version;
-
-       /* In sectors */
-       uint32_t chunk_size;
-};
-
-struct disk_exception {
-       uint64_t old_chunk;
-       uint64_t new_chunk;
-};
-
-struct commit_callback {
-       void (*callback)(void *, int success);
-       void *context;
-};
-
-/*
- * The top level structure for a persistent exception store.
- */
-struct pstore {
-       struct dm_snapshot *snap;       /* up pointer to my snapshot */
-       int version;
-       int valid;
-       uint32_t exceptions_per_area;
-
-       /*
-        * Now that we have an asynchronous kcopyd there is no
-        * need for large chunk sizes, so it wont hurt to have a
-        * whole chunks worth of metadata in memory at once.
-        */
-       void *area;
-
-       /*
-        * An area of zeros used to clear the next area.
-        */
-       void *zero_area;
-
-       /*
-        * Used to keep track of which metadata area the data in
-        * 'chunk' refers to.
-        */
-       chunk_t current_area;
-
-       /*
-        * The next free chunk for an exception.
-        */
-       chunk_t next_free;
-
-       /*
-        * The index of next free exception in the current
-        * metadata area.
-        */
-       uint32_t current_committed;
-
-       atomic_t pending_count;
-       uint32_t callback_count;
-       struct commit_callback *callbacks;
-       struct dm_io_client *io_client;
-
-       struct workqueue_struct *metadata_wq;
-};
-
-static unsigned sectors_to_pages(unsigned sectors)
-{
-       return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
-}
-
-static int alloc_area(struct pstore *ps)
-{
-       int r = -ENOMEM;
-       size_t len;
-
-       len = ps->snap->chunk_size << SECTOR_SHIFT;
-
-       /*
-        * Allocate the chunk_size block of memory that will hold
-        * a single metadata area.
-        */
-       ps->area = vmalloc(len);
-       if (!ps->area)
-               return r;
-
-       ps->zero_area = vmalloc(len);
-       if (!ps->zero_area) {
-               vfree(ps->area);
-               return r;
-       }
-       memset(ps->zero_area, 0, len);
-
-       return 0;
-}
-
-static void free_area(struct pstore *ps)
-{
-       vfree(ps->area);
-       ps->area = NULL;
-       vfree(ps->zero_area);
-       ps->zero_area = NULL;
-}
-
-struct mdata_req {
-       struct dm_io_region *where;
-       struct dm_io_request *io_req;
-       struct work_struct work;
-       int result;
-};
-
-static void do_metadata(struct work_struct *work)
-{
-       struct mdata_req *req = container_of(work, struct mdata_req, work);
-
-       req->result = dm_io(req->io_req, 1, req->where, NULL);
-}
-
-/*
- * Read or write a chunk aligned and sized block of data from a device.
- */
-static int chunk_io(struct pstore *ps, chunk_t chunk, int rw, int metadata)
-{
-       struct dm_io_region where = {
-               .bdev = ps->snap->cow->bdev,
-               .sector = ps->snap->chunk_size * chunk,
-               .count = ps->snap->chunk_size,
-       };
-       struct dm_io_request io_req = {
-               .bi_rw = rw,
-               .mem.type = DM_IO_VMA,
-               .mem.ptr.vma = ps->area,
-               .client = ps->io_client,
-               .notify.fn = NULL,
-       };
-       struct mdata_req req;
-
-       if (!metadata)
-               return dm_io(&io_req, 1, &where, NULL);
-
-       req.where = &where;
-       req.io_req = &io_req;
-
-       /*
-        * Issue the synchronous I/O from a different thread
-        * to avoid generic_make_request recursion.
-        */
-       INIT_WORK(&req.work, do_metadata);
-       queue_work(ps->metadata_wq, &req.work);
-       flush_workqueue(ps->metadata_wq);
-
-       return req.result;
-}
-
-/*
- * Convert a metadata area index to a chunk index.
- */
-static chunk_t area_location(struct pstore *ps, chunk_t area)
-{
-       return 1 + ((ps->exceptions_per_area + 1) * area);
-}
-
-/*
- * Read or write a metadata area.  Remembering to skip the first
- * chunk which holds the header.
- */
-static int area_io(struct pstore *ps, int rw)
-{
-       int r;
-       chunk_t chunk;
-
-       chunk = area_location(ps, ps->current_area);
-
-       r = chunk_io(ps, chunk, rw, 0);
-       if (r)
-               return r;
-
-       return 0;
-}
-
-static void zero_memory_area(struct pstore *ps)
-{
-       memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
-}
-
-static int zero_disk_area(struct pstore *ps, chunk_t area)
-{
-       struct dm_io_region where = {
-               .bdev = ps->snap->cow->bdev,
-               .sector = ps->snap->chunk_size * area_location(ps, area),
-               .count = ps->snap->chunk_size,
-       };
-       struct dm_io_request io_req = {
-               .bi_rw = WRITE,
-               .mem.type = DM_IO_VMA,
-               .mem.ptr.vma = ps->zero_area,
-               .client = ps->io_client,
-               .notify.fn = NULL,
-       };
-
-       return dm_io(&io_req, 1, &where, NULL);
-}
-
-static int read_header(struct pstore *ps, int *new_snapshot)
-{
-       int r;
-       struct disk_header *dh;
-       chunk_t chunk_size;
-       int chunk_size_supplied = 1;
-
-       /*
-        * Use default chunk size (or hardsect_size, if larger) if none supplied
-        */
-       if (!ps->snap->chunk_size) {
-               ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
-                   bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
-               ps->snap->chunk_mask = ps->snap->chunk_size - 1;
-               ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
-               chunk_size_supplied = 0;
-       }
-
-       ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
-                                                            chunk_size));
-       if (IS_ERR(ps->io_client))
-               return PTR_ERR(ps->io_client);
-
-       r = alloc_area(ps);
-       if (r)
-               return r;
-
-       r = chunk_io(ps, 0, READ, 1);
-       if (r)
-               goto bad;
-
-       dh = (struct disk_header *) ps->area;
-
-       if (le32_to_cpu(dh->magic) == 0) {
-               *new_snapshot = 1;
-               return 0;
-       }
-
-       if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
-               DMWARN("Invalid or corrupt snapshot");
-               r = -ENXIO;
-               goto bad;
-       }
-
-       *new_snapshot = 0;
-       ps->valid = le32_to_cpu(dh->valid);
-       ps->version = le32_to_cpu(dh->version);
-       chunk_size = le32_to_cpu(dh->chunk_size);
-
-       if (!chunk_size_supplied || ps->snap->chunk_size == chunk_size)
-               return 0;
-
-       DMWARN("chunk size %llu in device metadata overrides "
-              "table chunk size of %llu.",
-              (unsigned long long)chunk_size,
-              (unsigned long long)ps->snap->chunk_size);
-
-       /* We had a bogus chunk_size. Fix stuff up. */
-       free_area(ps);
-
-       ps->snap->chunk_size = chunk_size;
-       ps->snap->chunk_mask = chunk_size - 1;
-       ps->snap->chunk_shift = ffs(chunk_size) - 1;
-
-       r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
-                               ps->io_client);
-       if (r)
-               return r;
-
-       r = alloc_area(ps);
-       return r;
-
-bad:
-       free_area(ps);
-       return r;
-}
-
-static int write_header(struct pstore *ps)
-{
-       struct disk_header *dh;
-
-       memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
-
-       dh = (struct disk_header *) ps->area;
-       dh->magic = cpu_to_le32(SNAP_MAGIC);
-       dh->valid = cpu_to_le32(ps->valid);
-       dh->version = cpu_to_le32(ps->version);
-       dh->chunk_size = cpu_to_le32(ps->snap->chunk_size);
-
-       return chunk_io(ps, 0, WRITE, 1);
-}
-
-/*
- * Access functions for the disk exceptions, these do the endian conversions.
- */
-static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
-{
-       BUG_ON(index >= ps->exceptions_per_area);
-
-       return ((struct disk_exception *) ps->area) + index;
-}
+#include <linux/device-mapper.h>
 
-static void read_exception(struct pstore *ps,
-                          uint32_t index, struct disk_exception *result)
-{
-       struct disk_exception *e = get_exception(ps, index);
-
-       /* copy it */
-       result->old_chunk = le64_to_cpu(e->old_chunk);
-       result->new_chunk = le64_to_cpu(e->new_chunk);
-}
-
-static void write_exception(struct pstore *ps,
-                           uint32_t index, struct disk_exception *de)
-{
-       struct disk_exception *e = get_exception(ps, index);
-
-       /* copy it */
-       e->old_chunk = cpu_to_le64(de->old_chunk);
-       e->new_chunk = cpu_to_le64(de->new_chunk);
-}
+#define DM_MSG_PREFIX "snapshot exception stores"
 
-/*
- * Registers the exceptions that are present in the current area.
- * 'full' is filled in to indicate if the area has been
- * filled.
- */
-static int insert_exceptions(struct pstore *ps, int *full)
+int dm_exception_store_init(void)
 {
        int r;
-       unsigned int i;
-       struct disk_exception de;
-
-       /* presume the area is full */
-       *full = 1;
-
-       for (i = 0; i < ps->exceptions_per_area; i++) {
-               read_exception(ps, i, &de);
-
-               /*
-                * If the new_chunk is pointing at the start of
-                * the COW device, where the first metadata area
-                * is we know that we've hit the end of the
-                * exceptions.  Therefore the area is not full.
-                */
-               if (de.new_chunk == 0LL) {
-                       ps->current_committed = i;
-                       *full = 0;
-                       break;
-               }
-
-               /*
-                * Keep track of the start of the free chunks.
-                */
-               if (ps->next_free <= de.new_chunk)
-                       ps->next_free = de.new_chunk + 1;
-
-               /*
-                * Otherwise we add the exception to the snapshot.
-                */
-               r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk);
-               if (r)
-                       return r;
-       }
-
-       return 0;
-}
-
-static int read_exceptions(struct pstore *ps)
-{
-       int r, full = 1;
-
-       /*
-        * Keeping reading chunks and inserting exceptions until
-        * we find a partially full area.
-        */
-       for (ps->current_area = 0; full; ps->current_area++) {
-               r = area_io(ps, READ);
-               if (r)
-                       return r;
 
-               r = insert_exceptions(ps, &full);
-               if (r)
-                       return r;
+       r = dm_transient_snapshot_init();
+       if (r) {
+               DMERR("Unable to register transient exception store type.");
+               goto transient_fail;
        }
 
-       ps->current_area--;
-
-       return 0;
-}
-
-static struct pstore *get_info(struct dm_exception_store *store)
-{
-       return (struct pstore *) store->context;
-}
-
-static void persistent_fraction_full(struct dm_exception_store *store,
-                                    sector_t *numerator, sector_t *denominator)
-{
-       *numerator = get_info(store)->next_free * store->snap->chunk_size;
-       *denominator = get_dev_size(store->snap->cow->bdev);
-}
-
-static void persistent_destroy(struct dm_exception_store *store)
-{
-       struct pstore *ps = get_info(store);
-
-       destroy_workqueue(ps->metadata_wq);
-       dm_io_client_destroy(ps->io_client);
-       vfree(ps->callbacks);
-       free_area(ps);
-       kfree(ps);
-}
-
-static int persistent_read_metadata(struct dm_exception_store *store)
-{
-       int r, uninitialized_var(new_snapshot);
-       struct pstore *ps = get_info(store);
-
-       /*
-        * Read the snapshot header.
-        */
-       r = read_header(ps, &new_snapshot);
-       if (r)
-               return r;
-
-       /*
-        * Now we know correct chunk_size, complete the initialisation.
-        */
-       ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
-                                 sizeof(struct disk_exception);
-       ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
-                       sizeof(*ps->callbacks));
-       if (!ps->callbacks)
-               return -ENOMEM;
-
-       /*
-        * Do we need to setup a new snapshot ?
-        */
-       if (new_snapshot) {
-               r = write_header(ps);
-               if (r) {
-                       DMWARN("write_header failed");
-                       return r;
-               }
-
-               ps->current_area = 0;
-               zero_memory_area(ps);
-               r = zero_disk_area(ps, 0);
-               if (r) {
-                       DMWARN("zero_disk_area(0) failed");
-                       return r;
-               }
-       } else {
-               /*
-                * Sanity checks.
-                */
-               if (ps->version != SNAPSHOT_DISK_VERSION) {
-                       DMWARN("unable to handle snapshot disk version %d",
-                              ps->version);
-                       return -EINVAL;
-               }
-
-               /*
-                * Metadata are valid, but snapshot is invalidated
-                */
-               if (!ps->valid)
-                       return 1;
-
-               /*
-                * Read the metadata.
-                */
-               r = read_exceptions(ps);
-               if (r)
-                       return r;
+       r = dm_persistent_snapshot_init();
+       if (r) {
+               DMERR("Unable to register persistent exception store type");
+               goto persistent_fail;
        }
 
        return 0;
-}
-
-static int persistent_prepare(struct dm_exception_store *store,
-                             struct dm_snap_exception *e)
-{
-       struct pstore *ps = get_info(store);
-       uint32_t stride;
-       chunk_t next_free;
-       sector_t size = get_dev_size(store->snap->cow->bdev);
-
-       /* Is there enough room ? */
-       if (size < ((ps->next_free + 1) * store->snap->chunk_size))
-               return -ENOSPC;
 
-       e->new_chunk = ps->next_free;
-
-       /*
-        * Move onto the next free pending, making sure to take
-        * into account the location of the metadata chunks.
-        */
-       stride = (ps->exceptions_per_area + 1);
-       next_free = ++ps->next_free;
-       if (sector_div(next_free, stride) == 1)
-               ps->next_free++;
-
-       atomic_inc(&ps->pending_count);
-       return 0;
-}
-
-static void persistent_commit(struct dm_exception_store *store,
-                             struct dm_snap_exception *e,
-                             void (*callback) (void *, int success),
-                             void *callback_context)
-{
-       unsigned int i;
-       struct pstore *ps = get_info(store);
-       struct disk_exception de;
-       struct commit_callback *cb;
-
-       de.old_chunk = e->old_chunk;
-       de.new_chunk = e->new_chunk;
-       write_exception(ps, ps->current_committed++, &de);
-
-       /*
-        * Add the callback to the back of the array.  This code
-        * is the only place where the callback array is
-        * manipulated, and we know that it will never be called
-        * multiple times concurrently.
-        */
-       cb = ps->callbacks + ps->callback_count++;
-       cb->callback = callback;
-       cb->context = callback_context;
-
-       /*
-        * If there are exceptions in flight and we have not yet
-        * filled this metadata area there's nothing more to do.
-        */
-       if (!atomic_dec_and_test(&ps->pending_count) &&
-           (ps->current_committed != ps->exceptions_per_area))
-               return;
-
-       /*
-        * If we completely filled the current area, then wipe the next one.
-        */
-       if ((ps->current_committed == ps->exceptions_per_area) &&
-            zero_disk_area(ps, ps->current_area + 1))
-               ps->valid = 0;
-
-       /*
-        * Commit exceptions to disk.
-        */
-       if (ps->valid && area_io(ps, WRITE))
-               ps->valid = 0;
-
-       /*
-        * Advance to the next area if this one is full.
-        */
-       if (ps->current_committed == ps->exceptions_per_area) {
-               ps->current_committed = 0;
-               ps->current_area++;
-               zero_memory_area(ps);
-       }
-
-       for (i = 0; i < ps->callback_count; i++) {
-               cb = ps->callbacks + i;
-               cb->callback(cb->context, ps->valid);
-       }
-
-       ps->callback_count = 0;
-}
-
-static void persistent_drop(struct dm_exception_store *store)
-{
-       struct pstore *ps = get_info(store);
-
-       ps->valid = 0;
-       if (write_header(ps))
-               DMWARN("write header failed");
-}
-
-int dm_create_persistent(struct dm_exception_store *store)
-{
-       struct pstore *ps;
-
-       /* allocate the pstore */
-       ps = kmalloc(sizeof(*ps), GFP_KERNEL);
-       if (!ps)
-               return -ENOMEM;
-
-       ps->snap = store->snap;
-       ps->valid = 1;
-       ps->version = SNAPSHOT_DISK_VERSION;
-       ps->area = NULL;
-       ps->next_free = 2;      /* skipping the header and first area */
-       ps->current_committed = 0;
-
-       ps->callback_count = 0;
-       atomic_set(&ps->pending_count, 0);
-       ps->callbacks = NULL;
-
-       ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
-       if (!ps->metadata_wq) {
-               kfree(ps);
-               DMERR("couldn't start header metadata update thread");
-               return -ENOMEM;
-       }
-
-       store->destroy = persistent_destroy;
-       store->read_metadata = persistent_read_metadata;
-       store->prepare_exception = persistent_prepare;
-       store->commit_exception = persistent_commit;
-       store->drop_snapshot = persistent_drop;
-       store->fraction_full = persistent_fraction_full;
-       store->context = ps;
-
-       return 0;
-}
-
-/*-----------------------------------------------------------------
- * Implementation of the store for non-persistent snapshots.
- *---------------------------------------------------------------*/
-struct transient_c {
-       sector_t next_free;
-};
-
-static void transient_destroy(struct dm_exception_store *store)
-{
-       kfree(store->context);
-}
-
-static int transient_read_metadata(struct dm_exception_store *store)
-{
-       return 0;
-}
-
-static int transient_prepare(struct dm_exception_store *store,
-                            struct dm_snap_exception *e)
-{
-       struct transient_c *tc = (struct transient_c *) store->context;
-       sector_t size = get_dev_size(store->snap->cow->bdev);
-
-       if (size < (tc->next_free + store->snap->chunk_size))
-               return -1;
-
-       e->new_chunk = sector_to_chunk(store->snap, tc->next_free);
-       tc->next_free += store->snap->chunk_size;
-
-       return 0;
-}
-
-static void transient_commit(struct dm_exception_store *store,
-                            struct dm_snap_exception *e,
-                            void (*callback) (void *, int success),
-                            void *callback_context)
-{
-       /* Just succeed */
-       callback(callback_context, 1);
-}
-
-static void transient_fraction_full(struct dm_exception_store *store,
-                                   sector_t *numerator, sector_t *denominator)
-{
-       *numerator = ((struct transient_c *) store->context)->next_free;
-       *denominator = get_dev_size(store->snap->cow->bdev);
+persistent_fail:
+       dm_persistent_snapshot_exit();
+transient_fail:
+       return r;
 }
 
-int dm_create_transient(struct dm_exception_store *store)
+void dm_exception_store_exit(void)
 {
-       struct transient_c *tc;
-
-       store->destroy = transient_destroy;
-       store->read_metadata = transient_read_metadata;
-       store->prepare_exception = transient_prepare;
-       store->commit_exception = transient_commit;
-       store->drop_snapshot = NULL;
-       store->fraction_full = transient_fraction_full;
-
-       tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);
-       if (!tc)
-               return -ENOMEM;
-
-       tc->next_free = 0;
-       store->context = tc;
-
-       return 0;
+       dm_persistent_snapshot_exit();
+       dm_transient_snapshot_exit();
 }
index 25677df..78d1ace 100644 (file)
@@ -122,9 +122,18 @@ static inline void dm_consecutive_chunk_count_inc(struct dm_snap_exception *e)
 
 #  endif
 
+int dm_exception_store_init(void);
+void dm_exception_store_exit(void);
+
 /*
  * Two exception store implementations.
  */
+int dm_persistent_snapshot_init(void);
+void dm_persistent_snapshot_exit(void);
+
+int dm_transient_snapshot_init(void);
+void dm_transient_snapshot_exit(void);
+
 int dm_create_persistent(struct dm_exception_store *store);
 
 int dm_create_transient(struct dm_exception_store *store);
diff --git a/drivers/md/dm-snap-persistent.c b/drivers/md/dm-snap-persistent.c
new file mode 100644 (file)
index 0000000..57c946c
--- /dev/null
@@ -0,0 +1,694 @@
+/*
+ * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
+ * Copyright (C) 2006-2008 Red Hat GmbH
+ *
+ * This file is released under the GPL.
+ */
+
+#include "dm-exception-store.h"
+#include "dm-snap.h"
+
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/vmalloc.h>
+#include <linux/slab.h>
+#include <linux/dm-io.h>
+
+#define DM_MSG_PREFIX "persistent snapshot"
+#define DM_CHUNK_SIZE_DEFAULT_SECTORS 32       /* 16KB */
+
+/*-----------------------------------------------------------------
+ * Persistent snapshots, by persistent we mean that the snapshot
+ * will survive a reboot.
+ *---------------------------------------------------------------*/
+
+/*
+ * We need to store a record of which parts of the origin have
+ * been copied to the snapshot device.  The snapshot code
+ * requires that we copy exception chunks to chunk aligned areas
+ * of the COW store.  It makes sense therefore, to store the
+ * metadata in chunk size blocks.
+ *
+ * There is no backward or forward compatibility implemented,
+ * snapshots with different disk versions than the kernel will
+ * not be usable.  It is expected that "lvcreate" will blank out
+ * the start of a fresh COW device before calling the snapshot
+ * constructor.
+ *
+ * The first chunk of the COW device just contains the header.
+ * After this there is a chunk filled with exception metadata,
+ * followed by as many exception chunks as can fit in the
+ * metadata areas.
+ *
+ * All on disk structures are in little-endian format.  The end
+ * of the exceptions info is indicated by an exception with a
+ * new_chunk of 0, which is invalid since it would point to the
+ * header chunk.
+ */
+
+/*
+ * Magic for persistent snapshots: "SnAp" - Feeble isn't it.
+ */
+#define SNAP_MAGIC 0x70416e53
+
+/*
+ * The on-disk version of the metadata.
+ */
+#define SNAPSHOT_DISK_VERSION 1
+
+struct disk_header {
+       uint32_t magic;
+
+       /*
+        * Is this snapshot valid.  There is no way of recovering
+        * an invalid snapshot.
+        */
+       uint32_t valid;
+
+       /*
+        * Simple, incrementing version. no backward
+        * compatibility.
+        */
+       uint32_t version;
+
+       /* In sectors */
+       uint32_t chunk_size;
+};
+
+struct disk_exception {
+       uint64_t old_chunk;
+       uint64_t new_chunk;
+};
+
+struct commit_callback {
+       void (*callback)(void *, int success);
+       void *context;
+};
+
+/*
+ * The top level structure for a persistent exception store.
+ */
+struct pstore {
+       struct dm_snapshot *snap;       /* up pointer to my snapshot */
+       int version;
+       int valid;
+       uint32_t exceptions_per_area;
+
+       /*
+        * Now that we have an asynchronous kcopyd there is no
+        * need for large chunk sizes, so it wont hurt to have a
+        * whole chunks worth of metadata in memory at once.
+        */
+       void *area;
+
+       /*
+        * An area of zeros used to clear the next area.
+        */
+       void *zero_area;
+
+       /*
+        * Used to keep track of which metadata area the data in
+        * 'chunk' refers to.
+        */
+       chunk_t current_area;
+
+       /*
+        * The next free chunk for an exception.
+        */
+       chunk_t next_free;
+
+       /*
+        * The index of next free exception in the current
+        * metadata area.
+        */
+       uint32_t current_committed;
+
+       atomic_t pending_count;
+       uint32_t callback_count;
+       struct commit_callback *callbacks;
+       struct dm_io_client *io_client;
+
+       struct workqueue_struct *metadata_wq;
+};
+
+static unsigned sectors_to_pages(unsigned sectors)
+{
+       return DIV_ROUND_UP(sectors, PAGE_SIZE >> 9);
+}
+
+static int alloc_area(struct pstore *ps)
+{
+       int r = -ENOMEM;
+       size_t len;
+
+       len = ps->snap->chunk_size << SECTOR_SHIFT;
+
+       /*
+        * Allocate the chunk_size block of memory that will hold
+        * a single metadata area.
+        */
+       ps->area = vmalloc(len);
+       if (!ps->area)
+               return r;
+
+       ps->zero_area = vmalloc(len);
+       if (!ps->zero_area) {
+               vfree(ps->area);
+               return r;
+       }
+       memset(ps->zero_area, 0, len);
+
+       return 0;
+}
+
+static void free_area(struct pstore *ps)
+{
+       vfree(ps->area);
+       ps->area = NULL;
+       vfree(ps->zero_area);
+       ps->zero_area = NULL;
+}
+
+struct mdata_req {
+       struct dm_io_region *where;
+       struct dm_io_request *io_req;
+       struct work_struct work;
+       int result;
+};
+
+static void do_metadata(struct work_struct *work)
+{
+       struct mdata_req *req = container_of(work, struct mdata_req, work);
+
+       req->result = dm_io(req->io_req, 1, req->where, NULL);
+}
+
+/*
+ * Read or write a chunk aligned and sized block of data from a device.
+ */
+static int chunk_io(struct pstore *ps, chunk_t chunk, int rw, int metadata)
+{
+       struct dm_io_region where = {
+               .bdev = ps->snap->cow->bdev,
+               .sector = ps->snap->chunk_size * chunk,
+               .count = ps->snap->chunk_size,
+       };
+       struct dm_io_request io_req = {
+               .bi_rw = rw,
+               .mem.type = DM_IO_VMA,
+               .mem.ptr.vma = ps->area,
+               .client = ps->io_client,
+               .notify.fn = NULL,
+       };
+       struct mdata_req req;
+
+       if (!metadata)
+               return dm_io(&io_req, 1, &where, NULL);
+
+       req.where = &where;
+       req.io_req = &io_req;
+
+       /*
+        * Issue the synchronous I/O from a different thread
+        * to avoid generic_make_request recursion.
+        */
+       INIT_WORK(&req.work, do_metadata);
+       queue_work(ps->metadata_wq, &req.work);
+       flush_workqueue(ps->metadata_wq);
+
+       return req.result;
+}
+
+/*
+ * Convert a metadata area index to a chunk index.
+ */
+static chunk_t area_location(struct pstore *ps, chunk_t area)
+{
+       return 1 + ((ps->exceptions_per_area + 1) * area);
+}
+
+/*
+ * Read or write a metadata area.  Remembering to skip the first
+ * chunk which holds the header.
+ */
+static int area_io(struct pstore *ps, int rw)
+{
+       int r;
+       chunk_t chunk;
+
+       chunk = area_location(ps, ps->current_area);
+
+       r = chunk_io(ps, chunk, rw, 0);
+       if (r)
+               return r;
+
+       return 0;
+}
+
+static void zero_memory_area(struct pstore *ps)
+{
+       memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
+}
+
+static int zero_disk_area(struct pstore *ps, chunk_t area)
+{
+       struct dm_io_region where = {
+               .bdev = ps->snap->cow->bdev,
+               .sector = ps->snap->chunk_size * area_location(ps, area),
+               .count = ps->snap->chunk_size,
+       };
+       struct dm_io_request io_req = {
+               .bi_rw = WRITE,
+               .mem.type = DM_IO_VMA,
+               .mem.ptr.vma = ps->zero_area,
+               .client = ps->io_client,
+               .notify.fn = NULL,
+       };
+
+       return dm_io(&io_req, 1, &where, NULL);
+}
+
+static int read_header(struct pstore *ps, int *new_snapshot)
+{
+       int r;
+       struct disk_header *dh;
+       chunk_t chunk_size;
+       int chunk_size_supplied = 1;
+
+       /*
+        * Use default chunk size (or hardsect_size, if larger) if none supplied
+        */
+       if (!ps->snap->chunk_size) {
+               ps->snap->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
+                   bdev_hardsect_size(ps->snap->cow->bdev) >> 9);
+               ps->snap->chunk_mask = ps->snap->chunk_size - 1;
+               ps->snap->chunk_shift = ffs(ps->snap->chunk_size) - 1;
+               chunk_size_supplied = 0;
+       }
+
+       ps->io_client = dm_io_client_create(sectors_to_pages(ps->snap->
+                                                            chunk_size));
+       if (IS_ERR(ps->io_client))
+               return PTR_ERR(ps->io_client);
+
+       r = alloc_area(ps);
+       if (r)
+               return r;
+
+       r = chunk_io(ps, 0, READ, 1);
+       if (r)
+               goto bad;
+
+       dh = (struct disk_header *) ps->area;
+
+       if (le32_to_cpu(dh->magic) == 0) {
+               *new_snapshot = 1;
+               return 0;
+       }
+
+       if (le32_to_cpu(dh->magic) != SNAP_MAGIC) {
+               DMWARN("Invalid or corrupt snapshot");
+               r = -ENXIO;
+               goto bad;
+       }
+
+       *new_snapshot = 0;
+       ps->valid = le32_to_cpu(dh->valid);
+       ps->version = le32_to_cpu(dh->version);
+       chunk_size = le32_to_cpu(dh->chunk_size);
+
+       if (!chunk_size_supplied || ps->snap->chunk_size == chunk_size)
+               return 0;
+
+       DMWARN("chunk size %llu in device metadata overrides "
+              "table chunk size of %llu.",
+              (unsigned long long)chunk_size,
+              (unsigned long long)ps->snap->chunk_size);
+
+       /* We had a bogus chunk_size. Fix stuff up. */
+       free_area(ps);
+
+       ps->snap->chunk_size = chunk_size;
+       ps->snap->chunk_mask = chunk_size - 1;
+       ps->snap->chunk_shift = ffs(chunk_size) - 1;
+
+       r = dm_io_client_resize(sectors_to_pages(ps->snap->chunk_size),
+                               ps->io_client);
+       if (r)
+               return r;
+
+       r = alloc_area(ps);
+       return r;
+
+bad:
+       free_area(ps);
+       return r;
+}
+
+static int write_header(struct pstore *ps)
+{
+       struct disk_header *dh;
+
+       memset(ps->area, 0, ps->snap->chunk_size << SECTOR_SHIFT);
+
+       dh = (struct disk_header *) ps->area;
+       dh->magic = cpu_to_le32(SNAP_MAGIC);
+       dh->valid = cpu_to_le32(ps->valid);
+       dh->version = cpu_to_le32(ps->version);
+       dh->chunk_size = cpu_to_le32(ps->snap->chunk_size);
+
+       return chunk_io(ps, 0, WRITE, 1);
+}
+
+/*
+ * Access functions for the disk exceptions, these do the endian conversions.
+ */
+static struct disk_exception *get_exception(struct pstore *ps, uint32_t index)
+{
+       BUG_ON(index >= ps->exceptions_per_area);
+
+       return ((struct disk_exception *) ps->area) + index;
+}
+
+static void read_exception(struct pstore *ps,
+                          uint32_t index, struct disk_exception *result)
+{
+       struct disk_exception *e = get_exception(ps, index);
+
+       /* copy it */
+       result->old_chunk = le64_to_cpu(e->old_chunk);
+       result->new_chunk = le64_to_cpu(e->new_chunk);
+}
+
+static void write_exception(struct pstore *ps,
+                           uint32_t index, struct disk_exception *de)
+{
+       struct disk_exception *e = get_exception(ps, index);
+
+       /* copy it */
+       e->old_chunk = cpu_to_le64(de->old_chunk);
+       e->new_chunk = cpu_to_le64(de->new_chunk);
+}
+
+/*
+ * Registers the exceptions that are present in the current area.
+ * 'full' is filled in to indicate if the area has been
+ * filled.
+ */
+static int insert_exceptions(struct pstore *ps, int *full)
+{
+       int r;
+       unsigned int i;
+       struct disk_exception de;
+
+       /* presume the area is full */
+       *full = 1;
+
+       for (i = 0; i < ps->exceptions_per_area; i++) {
+               read_exception(ps, i, &de);
+
+               /*
+                * If the new_chunk is pointing at the start of
+                * the COW device, where the first metadata area
+                * is we know that we've hit the end of the
+                * exceptions.  Therefore the area is not full.
+                */
+               if (de.new_chunk == 0LL) {
+                       ps->current_committed = i;
+                       *full = 0;
+                       break;
+               }
+
+               /*
+                * Keep track of the start of the free chunks.
+                */
+               if (ps->next_free <= de.new_chunk)
+                       ps->next_free = de.new_chunk + 1;
+
+               /*
+                * Otherwise we add the exception to the snapshot.
+                */
+               r = dm_add_exception(ps->snap, de.old_chunk, de.new_chunk);
+               if (r)
+                       return r;
+       }
+
+       return 0;
+}
+
+static int read_exceptions(struct pstore *ps)
+{
+       int r, full = 1;
+
+       /*
+        * Keeping reading chunks and inserting exceptions until
+        * we find a partially full area.
+        */
+       for (ps->current_area = 0; full; ps->current_area++) {
+               r = area_io(ps, READ);
+               if (r)
+                       return r;
+
+               r = insert_exceptions(ps, &full);
+               if (r)
+                       return r;
+       }
+
+       ps->current_area--;
+
+       return 0;
+}
+
+static struct pstore *get_info(struct dm_exception_store *store)
+{
+       return (struct pstore *) store->context;
+}
+
+static void persistent_fraction_full(struct dm_exception_store *store,
+                                    sector_t *numerator, sector_t *denominator)
+{
+       *numerator = get_info(store)->next_free * store->snap->chunk_size;
+       *denominator = get_dev_size(store->snap->cow->bdev);
+}
+
+static void persistent_destroy(struct dm_exception_store *store)
+{
+       struct pstore *ps = get_info(store);
+
+       destroy_workqueue(ps->metadata_wq);
+       dm_io_client_destroy(ps->io_client);
+       vfree(ps->callbacks);
+       free_area(ps);
+       kfree(ps);
+}
+
+static int persistent_read_metadata(struct dm_exception_store *store)
+{
+       int r, uninitialized_var(new_snapshot);
+       struct pstore *ps = get_info(store);
+
+       /*
+        * Read the snapshot header.
+        */
+       r = read_header(ps, &new_snapshot);
+       if (r)
+               return r;
+
+       /*
+        * Now we know correct chunk_size, complete the initialisation.
+        */
+       ps->exceptions_per_area = (ps->snap->chunk_size << SECTOR_SHIFT) /
+                                 sizeof(struct disk_exception);
+       ps->callbacks = dm_vcalloc(ps->exceptions_per_area,
+                       sizeof(*ps->callbacks));
+       if (!ps->callbacks)
+               return -ENOMEM;
+
+       /*
+        * Do we need to setup a new snapshot ?
+        */
+       if (new_snapshot) {
+               r = write_header(ps);
+               if (r) {
+                       DMWARN("write_header failed");
+                       return r;
+               }
+
+               ps->current_area = 0;
+               zero_memory_area(ps);
+               r = zero_disk_area(ps, 0);
+               if (r) {
+                       DMWARN("zero_disk_area(0) failed");
+                       return r;
+               }
+       } else {
+               /*
+                * Sanity checks.
+                */
+               if (ps->version != SNAPSHOT_DISK_VERSION) {
+                       DMWARN("unable to handle snapshot disk version %d",
+                              ps->version);
+                       return -EINVAL;
+               }
+
+               /*
+                * Metadata are valid, but snapshot is invalidated
+                */
+               if (!ps->valid)
+                       return 1;
+
+               /*
+                * Read the metadata.
+                */
+               r = read_exceptions(ps);
+               if (r)
+                       return r;
+       }
+
+       return 0;
+}
+
+static int persistent_prepare(struct dm_exception_store *store,
+                             struct dm_snap_exception *e)
+{
+       struct pstore *ps = get_info(store);
+       uint32_t stride;
+       chunk_t next_free;
+       sector_t size = get_dev_size(store->snap->cow->bdev);
+
+       /* Is there enough room ? */
+       if (size < ((ps->next_free + 1) * store->snap->chunk_size))
+               return -ENOSPC;
+
+       e->new_chunk = ps->next_free;
+
+       /*
+        * Move onto the next free pending, making sure to take
+        * into account the location of the metadata chunks.
+        */
+       stride = (ps->exceptions_per_area + 1);
+       next_free = ++ps->next_free;
+       if (sector_div(next_free, stride) == 1)
+               ps->next_free++;
+
+       atomic_inc(&ps->pending_count);
+       return 0;
+}
+
+static void persistent_commit(struct dm_exception_store *store,
+                             struct dm_snap_exception *e,
+                             void (*callback) (void *, int success),
+                             void *callback_context)
+{
+       unsigned int i;
+       struct pstore *ps = get_info(store);
+       struct disk_exception de;
+       struct commit_callback *cb;
+
+       de.old_chunk = e->old_chunk;
+       de.new_chunk = e->new_chunk;
+       write_exception(ps, ps->current_committed++, &de);
+
+       /*
+        * Add the callback to the back of the array.  This code
+        * is the only place where the callback array is
+        * manipulated, and we know that it will never be called
+        * multiple times concurrently.
+        */
+       cb = ps->callbacks + ps->callback_count++;
+       cb->callback = callback;
+       cb->context = callback_context;
+
+       /*
+        * If there are exceptions in flight and we have not yet
+        * filled this metadata area there's nothing more to do.
+        */
+       if (!atomic_dec_and_test(&ps->pending_count) &&
+           (ps->current_committed != ps->exceptions_per_area))
+               return;
+
+       /*
+        * If we completely filled the current area, then wipe the next one.
+        */
+       if ((ps->current_committed == ps->exceptions_per_area) &&
+            zero_disk_area(ps, ps->current_area + 1))
+               ps->valid = 0;
+
+       /*
+        * Commit exceptions to disk.
+        */
+       if (ps->valid && area_io(ps, WRITE))
+               ps->valid = 0;
+
+       /*
+        * Advance to the next area if this one is full.
+        */
+       if (ps->current_committed == ps->exceptions_per_area) {
+               ps->current_committed = 0;
+               ps->current_area++;
+               zero_memory_area(ps);
+       }
+
+       for (i = 0; i < ps->callback_count; i++) {
+               cb = ps->callbacks + i;
+               cb->callback(cb->context, ps->valid);
+       }
+
+       ps->callback_count = 0;
+}
+
+static void persistent_drop(struct dm_exception_store *store)
+{
+       struct pstore *ps = get_info(store);
+
+       ps->valid = 0;
+       if (write_header(ps))
+               DMWARN("write header failed");
+}
+
+int dm_create_persistent(struct dm_exception_store *store)
+{
+       struct pstore *ps;
+
+       /* allocate the pstore */
+       ps = kmalloc(sizeof(*ps), GFP_KERNEL);
+       if (!ps)
+               return -ENOMEM;
+
+       ps->snap = store->snap;
+       ps->valid = 1;
+       ps->version = SNAPSHOT_DISK_VERSION;
+       ps->area = NULL;
+       ps->next_free = 2;      /* skipping the header and first area */
+       ps->current_committed = 0;
+
+       ps->callback_count = 0;
+       atomic_set(&ps->pending_count, 0);
+       ps->callbacks = NULL;
+
+       ps->metadata_wq = create_singlethread_workqueue("ksnaphd");
+       if (!ps->metadata_wq) {
+               kfree(ps);
+               DMERR("couldn't start header metadata update thread");
+               return -ENOMEM;
+       }
+
+       store->destroy = persistent_destroy;
+       store->read_metadata = persistent_read_metadata;
+       store->prepare_exception = persistent_prepare;
+       store->commit_exception = persistent_commit;
+       store->drop_snapshot = persistent_drop;
+       store->fraction_full = persistent_fraction_full;
+       store->context = ps;
+
+       return 0;
+}
+
+int dm_persistent_snapshot_init(void)
+{
+       return 0;
+}
+
+void dm_persistent_snapshot_exit(void)
+{
+}
diff --git a/drivers/md/dm-snap-transient.c b/drivers/md/dm-snap-transient.c
new file mode 100644 (file)
index 0000000..2a781df
--- /dev/null
@@ -0,0 +1,95 @@
+/*
+ * Copyright (C) 2001-2002 Sistina Software (UK) Limited.
+ * Copyright (C) 2006-2008 Red Hat GmbH
+ *
+ * This file is released under the GPL.
+ */
+
+#include "dm-exception-store.h"
+#include "dm-snap.h"
+
+#include <linux/mm.h>
+#include <linux/pagemap.h>
+#include <linux/vmalloc.h>
+#include <linux/slab.h>
+#include <linux/dm-io.h>
+
+#define DM_MSG_PREFIX "transient snapshot"
+
+/*-----------------------------------------------------------------
+ * Implementation of the store for non-persistent snapshots.
+ *---------------------------------------------------------------*/
+struct transient_c {
+       sector_t next_free;
+};
+
+static void transient_destroy(struct dm_exception_store *store)
+{
+       kfree(store->context);
+}
+
+static int transient_read_metadata(struct dm_exception_store *store)
+{
+       return 0;
+}
+
+static int transient_prepare(struct dm_exception_store *store,
+                            struct dm_snap_exception *e)
+{
+       struct transient_c *tc = (struct transient_c *) store->context;
+       sector_t size = get_dev_size(store->snap->cow->bdev);
+
+       if (size < (tc->next_free + store->snap->chunk_size))
+               return -1;
+
+       e->new_chunk = sector_to_chunk(store->snap, tc->next_free);
+       tc->next_free += store->snap->chunk_size;
+
+       return 0;
+}
+
+static void transient_commit(struct dm_exception_store *store,
+                            struct dm_snap_exception *e,
+                            void (*callback) (void *, int success),
+                            void *callback_context)
+{
+       /* Just succeed */
+       callback(callback_context, 1);
+}
+
+static void transient_fraction_full(struct dm_exception_store *store,
+                                   sector_t *numerator, sector_t *denominator)
+{
+       *numerator = ((struct transient_c *) store->context)->next_free;
+       *denominator = get_dev_size(store->snap->cow->bdev);
+}
+
+int dm_create_transient(struct dm_exception_store *store)
+{
+       struct transient_c *tc;
+
+       store->destroy = transient_destroy;
+       store->read_metadata = transient_read_metadata;
+       store->prepare_exception = transient_prepare;
+       store->commit_exception = transient_commit;
+       store->drop_snapshot = NULL;
+       store->fraction_full = transient_fraction_full;
+
+       tc = kmalloc(sizeof(struct transient_c), GFP_KERNEL);
+       if (!tc)
+               return -ENOMEM;
+
+       tc->next_free = 0;
+       store->context = tc;
+
+       return 0;
+}
+
+int dm_transient_snapshot_init(void)
+{
+       return 0;
+}
+
+void dm_transient_snapshot_exit(void)
+{
+}
index 81f03a0..018b567 100644 (file)
@@ -1406,6 +1406,12 @@ static int __init dm_snapshot_init(void)
 {
        int r;
 
+       r = dm_exception_store_init();
+       if (r) {
+               DMERR("Failed to initialize exception stores");
+               return r;
+       }
+
        r = dm_register_target(&snapshot_target);
        if (r) {
                DMERR("snapshot target register failed %d", r);
@@ -1454,17 +1460,17 @@ static int __init dm_snapshot_init(void)
 
        return 0;
 
-      bad_pending_pool:
+bad_pending_pool:
        kmem_cache_destroy(tracked_chunk_cache);
-      bad5:
+bad5:
        kmem_cache_destroy(pending_cache);
-      bad4:
+bad4:
        kmem_cache_destroy(exception_cache);
-      bad3:
+bad3:
        exit_origin_hash();
-      bad2:
+bad2:
        dm_unregister_target(&origin_target);
-      bad1:
+bad1:
        dm_unregister_target(&snapshot_target);
        return r;
 }
@@ -1480,6 +1486,8 @@ static void __exit dm_snapshot_exit(void)
        kmem_cache_destroy(pending_cache);
        kmem_cache_destroy(exception_cache);
        kmem_cache_destroy(tracked_chunk_cache);
+
+       dm_exception_store_exit();
 }
 
 /* Module hooks */