dm: add era target
authorJoe Thornber <ejt@redhat.com>
Mon, 3 Mar 2014 15:23:15 +0000 (10:23 -0500)
committerMike Snitzer <snitzer@redhat.com>
Thu, 27 Mar 2014 20:56:23 +0000 (16:56 -0400)
dm-era is a target that behaves similar to the linear target.  In
addition it keeps track of which blocks were written within a user
defined period of time called an 'era'.  Each era target instance
maintains the current era as a monotonically increasing 32-bit
counter.

Use cases include tracking changed blocks for backup software, and
partially invalidating the contents of a cache to restore cache
coherency after rolling back a vendor snapshot.

dm-era is primarily expected to be paired with the dm-cache target.

Signed-off-by: Joe Thornber <ejt@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Documentation/device-mapper/era.txt [new file with mode: 0644]
drivers/md/Kconfig
drivers/md/Makefile
drivers/md/dm-era-target.c [new file with mode: 0644]

diff --git a/Documentation/device-mapper/era.txt b/Documentation/device-mapper/era.txt
new file mode 100644 (file)
index 0000000..3c6d01b
--- /dev/null
@@ -0,0 +1,108 @@
+Introduction
+============
+
+dm-era is a target that behaves similar to the linear target.  In
+addition it keeps track of which blocks were written within a user
+defined period of time called an 'era'.  Each era target instance
+maintains the current era as a monotonically increasing 32-bit
+counter.
+
+Use cases include tracking changed blocks for backup software, and
+partially invalidating the contents of a cache to restore cache
+coherency after rolling back a vendor snapshot.
+
+Constructor
+===========
+
+ era <metadata dev> <origin dev> <block size>
+
+ metadata dev    : fast device holding the persistent metadata
+ origin dev     : device holding data blocks that may change
+ block size      : block size of origin data device, granularity that is
+                    tracked by the target
+
+Messages
+========
+
+None of the dm messages take any arguments.
+
+checkpoint
+----------
+
+Possibly move to a new era.  You shouldn't assume the era has
+incremented.  After sending this message, you should check the
+current era via the status line.
+
+take_metadata_snap
+------------------
+
+Create a clone of the metadata, to allow a userland process to read it.
+
+drop_metadata_snap
+------------------
+
+Drop the metadata snapshot.
+
+Status
+======
+
+<metadata block size> <#used metadata blocks>/<#total metadata blocks>
+<current era> <held metadata root | '-'>
+
+metadata block size     : Fixed block size for each metadata block in
+                            sectors
+#used metadata blocks   : Number of metadata blocks used
+#total metadata blocks  : Total number of metadata blocks
+current era             : The current era
+held metadata root      : The location, in blocks, of the metadata root
+                            that has been 'held' for userspace read
+                            access. '-' indicates there is no held root
+
+Detailed use case
+=================
+
+The scenario of invalidating a cache when rolling back a vendor
+snapshot was the primary use case when developing this target:
+
+Taking a vendor snapshot
+------------------------
+
+- Send a checkpoint message to the era target
+- Make a note of the current era in its status line
+- Take vendor snapshot (the era and snapshot should be forever
+  associated now).
+
+Rolling back to an vendor snapshot
+----------------------------------
+
+- Cache enters passthrough mode (see: dm-cache's docs in cache.txt)
+- Rollback vendor storage
+- Take metadata snapshot
+- Ascertain which blocks have been written since the snapshot was taken
+  by checking each block's era
+- Invalidate those blocks in the caching software
+- Cache returns to writeback/writethrough mode
+
+Memory usage
+============
+
+The target uses a bitset to record writes in the current era.  It also
+has a spare bitset ready for switching over to a new era.  Other than
+that it uses a few 4k blocks for updating metadata.
+
+   (4 * nr_blocks) bytes + buffers
+
+Resilience
+==========
+
+Metadata is updated on disk before a write to a previously unwritten
+block is performed.  As such dm-era should not be effected by a hard
+crash such as power failure.
+
+Userland tools
+==============
+
+Userland tools are found in the increasingly poorly named
+thin-provisioning-tools project:
+
+    https://github.com/jthornber/thin-provisioning-tools
index 95ad936e60482477c1f87c0aa654fded90d06ba5..5bdedf6df153cf25d54e91a2cca5d9e3b6b8b1cf 100644 (file)
@@ -285,6 +285,17 @@ config DM_CACHE_CLEANER
          A simple cache policy that writes back all data to the
          origin.  Used when decommissioning a dm-cache.
 
+config DM_ERA
+       tristate "Era target (EXPERIMENTAL)"
+       depends on BLK_DEV_DM
+       default n
+       select DM_PERSISTENT_DATA
+       select DM_BIO_PRISON
+       ---help---
+         dm-era tracks which parts of a block device are written to
+         over time.  Useful for maintaining cache coherency when using
+         vendor snapshots.
+
 config DM_MIRROR
        tristate "Mirror target"
        depends on BLK_DEV_DM
index f26d832925791e49f624d4b19649cfc5e8312f19..a2da532b1c2bda38d18a6324c2dfdcfcc7108df5 100644 (file)
@@ -14,6 +14,7 @@ dm-thin-pool-y        += dm-thin.o dm-thin-metadata.o
 dm-cache-y     += dm-cache-target.o dm-cache-metadata.o dm-cache-policy.o
 dm-cache-mq-y   += dm-cache-policy-mq.o
 dm-cache-cleaner-y += dm-cache-policy-cleaner.o
+dm-era-y       += dm-era-target.o
 md-mod-y       += md.o bitmap.o
 raid456-y      += raid5.o
 
@@ -53,6 +54,7 @@ obj-$(CONFIG_DM_VERITY)               += dm-verity.o
 obj-$(CONFIG_DM_CACHE)         += dm-cache.o
 obj-$(CONFIG_DM_CACHE_MQ)      += dm-cache-mq.o
 obj-$(CONFIG_DM_CACHE_CLEANER) += dm-cache-cleaner.o
+obj-$(CONFIG_DM_ERA)           += dm-era.o
 
 ifeq ($(CONFIG_DM_UEVENT),y)
 dm-mod-objs                    += dm-uevent.o
diff --git a/drivers/md/dm-era-target.c b/drivers/md/dm-era-target.c
new file mode 100644 (file)
index 0000000..03d9560
--- /dev/null
@@ -0,0 +1,1730 @@
+#include "dm.h"
+#include "persistent-data/dm-transaction-manager.h"
+#include "persistent-data/dm-bitset.h"
+#include "persistent-data/dm-space-map.h"
+
+#include <linux/dm-io.h>
+#include <linux/dm-kcopyd.h>
+#include <linux/init.h>
+#include <linux/mempool.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+#include <linux/vmalloc.h>
+
+#define DM_MSG_PREFIX "era"
+
+#define SUPERBLOCK_LOCATION 0
+#define SUPERBLOCK_MAGIC 2126579579
+#define SUPERBLOCK_CSUM_XOR 146538381
+#define MIN_ERA_VERSION 1
+#define MAX_ERA_VERSION 1
+#define INVALID_WRITESET_ROOT SUPERBLOCK_LOCATION
+#define MIN_BLOCK_SIZE 8
+
+/*----------------------------------------------------------------
+ * Writeset
+ *--------------------------------------------------------------*/
+struct writeset_metadata {
+       uint32_t nr_bits;
+       dm_block_t root;
+};
+
+struct writeset {
+       struct writeset_metadata md;
+
+       /*
+        * An in core copy of the bits to save constantly doing look ups on
+        * disk.
+        */
+       unsigned long *bits;
+};
+
+/*
+ * This does not free off the on disk bitset as this will normally be done
+ * after digesting into the era array.
+ */
+static void writeset_free(struct writeset *ws)
+{
+       vfree(ws->bits);
+}
+
+static int setup_on_disk_bitset(struct dm_disk_bitset *info,
+                               unsigned nr_bits, dm_block_t *root)
+{
+       int r;
+
+       r = dm_bitset_empty(info, root);
+       if (r)
+               return r;
+
+       return dm_bitset_resize(info, *root, 0, nr_bits, false, root);
+}
+
+static size_t bitset_size(unsigned nr_bits)
+{
+       return sizeof(unsigned long) * dm_div_up(nr_bits, BITS_PER_LONG);
+}
+
+/*
+ * Allocates memory for the in core bitset.
+ */
+static int writeset_alloc(struct writeset *ws, dm_block_t nr_blocks)
+{
+       ws->md.nr_bits = nr_blocks;
+       ws->md.root = INVALID_WRITESET_ROOT;
+       ws->bits = vzalloc(bitset_size(nr_blocks));
+       if (!ws->bits) {
+               DMERR("%s: couldn't allocate in memory bitset", __func__);
+               return -ENOMEM;
+       }
+
+       return 0;
+}
+
+/*
+ * Wipes the in-core bitset, and creates a new on disk bitset.
+ */
+static int writeset_init(struct dm_disk_bitset *info, struct writeset *ws)
+{
+       int r;
+
+       memset(ws->bits, 0, bitset_size(ws->md.nr_bits));
+
+       r = setup_on_disk_bitset(info, ws->md.nr_bits, &ws->md.root);
+       if (r) {
+               DMERR("%s: setup_on_disk_bitset failed", __func__);
+               return r;
+       }
+
+       return 0;
+}
+
+static bool writeset_marked(struct writeset *ws, dm_block_t block)
+{
+       return test_bit(block, ws->bits);
+}
+
+static int writeset_marked_on_disk(struct dm_disk_bitset *info,
+                                  struct writeset_metadata *m, dm_block_t block,
+                                  bool *result)
+{
+       dm_block_t old = m->root;
+
+       /*
+        * The bitset was flushed when it was archived, so we know there'll
+        * be no change to the root.
+        */
+       int r = dm_bitset_test_bit(info, m->root, block, &m->root, result);
+       if (r) {
+               DMERR("%s: dm_bitset_test_bit failed", __func__);
+               return r;
+       }
+
+       BUG_ON(m->root != old);
+
+       return r;
+}
+
+/*
+ * Returns < 0 on error, 0 if the bit wasn't previously set, 1 if it was.
+ */
+static int writeset_test_and_set(struct dm_disk_bitset *info,
+                                struct writeset *ws, uint32_t block)
+{
+       int r;
+
+       if (!test_and_set_bit(block, ws->bits)) {
+               r = dm_bitset_set_bit(info, ws->md.root, block, &ws->md.root);
+               if (r) {
+                       /* FIXME: fail mode */
+                       return r;
+               }
+
+               return 0;
+       }
+
+       return 1;
+}
+
+/*----------------------------------------------------------------
+ * On disk metadata layout
+ *--------------------------------------------------------------*/
+#define SPACE_MAP_ROOT_SIZE 128
+#define UUID_LEN 16
+
+struct writeset_disk {
+       __le32 nr_bits;
+       __le64 root;
+} __packed;
+
+struct superblock_disk {
+       __le32 csum;
+       __le32 flags;
+       __le64 blocknr;
+
+       __u8 uuid[UUID_LEN];
+       __le64 magic;
+       __le32 version;
+
+       __u8 metadata_space_map_root[SPACE_MAP_ROOT_SIZE];
+
+       __le32 data_block_size;
+       __le32 metadata_block_size;
+       __le32 nr_blocks;
+
+       __le32 current_era;
+       struct writeset_disk current_writeset;
+
+       /*
+        * Only these two fields are valid within the metadata snapshot.
+        */
+       __le64 writeset_tree_root;
+       __le64 era_array_root;
+
+       __le64 metadata_snap;
+} __packed;
+
+/*----------------------------------------------------------------
+ * Superblock validation
+ *--------------------------------------------------------------*/
+static void sb_prepare_for_write(struct dm_block_validator *v,
+                                struct dm_block *b,
+                                size_t sb_block_size)
+{
+       struct superblock_disk *disk = dm_block_data(b);
+
+       disk->blocknr = cpu_to_le64(dm_block_location(b));
+       disk->csum = cpu_to_le32(dm_bm_checksum(&disk->flags,
+                                               sb_block_size - sizeof(__le32),
+                                               SUPERBLOCK_CSUM_XOR));
+}
+
+static int check_metadata_version(struct superblock_disk *disk)
+{
+       uint32_t metadata_version = le32_to_cpu(disk->version);
+       if (metadata_version < MIN_ERA_VERSION || metadata_version > MAX_ERA_VERSION) {
+               DMERR("Era metadata version %u found, but only versions between %u and %u supported.",
+                     metadata_version, MIN_ERA_VERSION, MAX_ERA_VERSION);
+               return -EINVAL;
+       }
+
+       return 0;
+}
+
+static int sb_check(struct dm_block_validator *v,
+                   struct dm_block *b,
+                   size_t sb_block_size)
+{
+       struct superblock_disk *disk = dm_block_data(b);
+       __le32 csum_le;
+
+       if (dm_block_location(b) != le64_to_cpu(disk->blocknr)) {
+               DMERR("sb_check failed: blocknr %llu: wanted %llu",
+                     le64_to_cpu(disk->blocknr),
+                     (unsigned long long)dm_block_location(b));
+               return -ENOTBLK;
+       }
+
+       if (le64_to_cpu(disk->magic) != SUPERBLOCK_MAGIC) {
+               DMERR("sb_check failed: magic %llu: wanted %llu",
+                     le64_to_cpu(disk->magic),
+                     (unsigned long long) SUPERBLOCK_MAGIC);
+               return -EILSEQ;
+       }
+
+       csum_le = cpu_to_le32(dm_bm_checksum(&disk->flags,
+                                            sb_block_size - sizeof(__le32),
+                                            SUPERBLOCK_CSUM_XOR));
+       if (csum_le != disk->csum) {
+               DMERR("sb_check failed: csum %u: wanted %u",
+                     le32_to_cpu(csum_le), le32_to_cpu(disk->csum));
+               return -EILSEQ;
+       }
+
+       return check_metadata_version(disk);
+}
+
+static struct dm_block_validator sb_validator = {
+       .name = "superblock",
+       .prepare_for_write = sb_prepare_for_write,
+       .check = sb_check
+};
+
+/*----------------------------------------------------------------
+ * Low level metadata handling
+ *--------------------------------------------------------------*/
+#define DM_ERA_METADATA_BLOCK_SIZE 4096
+#define DM_ERA_METADATA_CACHE_SIZE 64
+#define ERA_MAX_CONCURRENT_LOCKS 5
+
+struct era_metadata {
+       struct block_device *bdev;
+       struct dm_block_manager *bm;
+       struct dm_space_map *sm;
+       struct dm_transaction_manager *tm;
+
+       dm_block_t block_size;
+       uint32_t nr_blocks;
+
+       uint32_t current_era;
+
+       /*
+        * We preallocate 2 writesets.  When an era rolls over we
+        * switch between them. This means the allocation is done at
+        * preresume time, rather than on the io path.
+        */
+       struct writeset writesets[2];
+       struct writeset *current_writeset;
+
+       dm_block_t writeset_tree_root;
+       dm_block_t era_array_root;
+
+       struct dm_disk_bitset bitset_info;
+       struct dm_btree_info writeset_tree_info;
+       struct dm_array_info era_array_info;
+
+       dm_block_t metadata_snap;
+
+       /*
+        * A flag that is set whenever a writeset has been archived.
+        */
+       bool archived_writesets;
+};
+
+static int superblock_read_lock(struct era_metadata *md,
+                               struct dm_block **sblock)
+{
+       return dm_bm_read_lock(md->bm, SUPERBLOCK_LOCATION,
+                              &sb_validator, sblock);
+}
+
+static int superblock_lock_zero(struct era_metadata *md,
+                               struct dm_block **sblock)
+{
+       return dm_bm_write_lock_zero(md->bm, SUPERBLOCK_LOCATION,
+                                    &sb_validator, sblock);
+}
+
+static int superblock_lock(struct era_metadata *md,
+                          struct dm_block **sblock)
+{
+       return dm_bm_write_lock(md->bm, SUPERBLOCK_LOCATION,
+                               &sb_validator, sblock);
+}
+
+/* FIXME: duplication with cache and thin */
+static int superblock_all_zeroes(struct dm_block_manager *bm, bool *result)
+{
+       int r;
+       unsigned i;
+       struct dm_block *b;
+       __le64 *data_le, zero = cpu_to_le64(0);
+       unsigned sb_block_size = dm_bm_block_size(bm) / sizeof(__le64);
+
+       /*
+        * We can't use a validator here - it may be all zeroes.
+        */
+       r = dm_bm_read_lock(bm, SUPERBLOCK_LOCATION, NULL, &b);
+       if (r)
+               return r;
+
+       data_le = dm_block_data(b);
+       *result = true;
+       for (i = 0; i < sb_block_size; i++) {
+               if (data_le[i] != zero) {
+                       *result = false;
+                       break;
+               }
+       }
+
+       return dm_bm_unlock(b);
+}
+
+/*----------------------------------------------------------------*/
+
+static void ws_pack(const struct writeset_metadata *core, struct writeset_disk *disk)
+{
+       disk->nr_bits = cpu_to_le32(core->nr_bits);
+       disk->root = cpu_to_le64(core->root);
+}
+
+static void ws_unpack(const struct writeset_disk *disk, struct writeset_metadata *core)
+{
+       core->nr_bits = le32_to_cpu(disk->nr_bits);
+       core->root = le64_to_cpu(disk->root);
+}
+
+static void ws_inc(void *context, const void *value)
+{
+       struct era_metadata *md = context;
+       struct writeset_disk ws_d;
+       dm_block_t b;
+
+       memcpy(&ws_d, value, sizeof(ws_d));
+       b = le64_to_cpu(ws_d.root);
+
+       dm_tm_inc(md->tm, b);
+}
+
+static void ws_dec(void *context, const void *value)
+{
+       struct era_metadata *md = context;
+       struct writeset_disk ws_d;
+       dm_block_t b;
+
+       memcpy(&ws_d, value, sizeof(ws_d));
+       b = le64_to_cpu(ws_d.root);
+
+       dm_bitset_del(&md->bitset_info, b);
+}
+
+static int ws_eq(void *context, const void *value1, const void *value2)
+{
+       return !memcmp(value1, value2, sizeof(struct writeset_metadata));
+}
+
+/*----------------------------------------------------------------*/
+
+static void setup_writeset_tree_info(struct era_metadata *md)
+{
+       struct dm_btree_value_type *vt = &md->writeset_tree_info.value_type;
+       md->writeset_tree_info.tm = md->tm;
+       md->writeset_tree_info.levels = 1;
+       vt->context = md;
+       vt->size = sizeof(struct writeset_disk);
+       vt->inc = ws_inc;
+       vt->dec = ws_dec;
+       vt->equal = ws_eq;
+}
+
+static void setup_era_array_info(struct era_metadata *md)
+
+{
+       struct dm_btree_value_type vt;
+       vt.context = NULL;
+       vt.size = sizeof(__le32);
+       vt.inc = NULL;
+       vt.dec = NULL;
+       vt.equal = NULL;
+
+       dm_array_info_init(&md->era_array_info, md->tm, &vt);
+}
+
+static void setup_infos(struct era_metadata *md)
+{
+       dm_disk_bitset_init(md->tm, &md->bitset_info);
+       setup_writeset_tree_info(md);
+       setup_era_array_info(md);
+}
+
+/*----------------------------------------------------------------*/
+
+static int create_fresh_metadata(struct era_metadata *md)
+{
+       int r;
+
+       r = dm_tm_create_with_sm(md->bm, SUPERBLOCK_LOCATION,
+                                &md->tm, &md->sm);
+       if (r < 0) {
+               DMERR("dm_tm_create_with_sm failed");
+               return r;
+       }
+
+       setup_infos(md);
+
+       r = dm_btree_empty(&md->writeset_tree_info, &md->writeset_tree_root);
+       if (r) {
+               DMERR("couldn't create new writeset tree");
+               goto bad;
+       }
+
+       r = dm_array_empty(&md->era_array_info, &md->era_array_root);
+       if (r) {
+               DMERR("couldn't create era array");
+               goto bad;
+       }
+
+       return 0;
+
+bad:
+       dm_sm_destroy(md->sm);
+       dm_tm_destroy(md->tm);
+
+       return r;
+}
+
+/*
+ * Writes a superblock, including the static fields that don't get updated
+ * with every commit (possible optimisation here).  'md' should be fully
+ * constructed when this is called.
+ */
+static int prepare_superblock(struct era_metadata *md, struct superblock_disk *disk)
+{
+       int r;
+       size_t metadata_len;
+
+       disk->magic = cpu_to_le64(SUPERBLOCK_MAGIC);
+       disk->flags = cpu_to_le32(0ul);
+
+       /* FIXME: can't keep blanking the uuid (uuid is currently unused though) */
+       memset(disk->uuid, 0, sizeof(disk->uuid));
+       disk->version = cpu_to_le32(MAX_ERA_VERSION);
+
+       r = dm_sm_root_size(md->sm, &metadata_len);
+       if (r < 0)
+               return r;
+
+       r = dm_sm_copy_root(md->sm, &disk->metadata_space_map_root,
+                           metadata_len);
+       if (r < 0)
+               return r;
+
+       disk->data_block_size = cpu_to_le32(md->block_size);
+       disk->metadata_block_size = cpu_to_le32(DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT);
+       disk->nr_blocks = cpu_to_le32(md->nr_blocks);
+       disk->current_era = cpu_to_le32(md->current_era);
+
+       ws_pack(&md->current_writeset->md, &disk->current_writeset);
+       disk->writeset_tree_root = cpu_to_le64(md->writeset_tree_root);
+       disk->era_array_root = cpu_to_le64(md->era_array_root);
+       disk->metadata_snap = cpu_to_le64(md->metadata_snap);
+
+       return 0;
+}
+
+static int write_superblock(struct era_metadata *md)
+{
+       int r;
+       struct dm_block *sblock;
+       struct superblock_disk *disk;
+
+       r = superblock_lock_zero(md, &sblock);
+       if (r)
+               return r;
+
+       disk = dm_block_data(sblock);
+       r = prepare_superblock(md, disk);
+       if (r) {
+               DMERR("%s: prepare_superblock failed", __func__);
+               dm_bm_unlock(sblock); /* FIXME: does this commit? */
+               return r;
+       }
+
+       return dm_tm_commit(md->tm, sblock);
+}
+
+/*
+ * Assumes block_size and the infos are set.
+ */
+static int format_metadata(struct era_metadata *md)
+{
+       int r;
+
+       r = create_fresh_metadata(md);
+       if (r)
+               return r;
+
+       r = write_superblock(md);
+       if (r) {
+               dm_sm_destroy(md->sm);
+               dm_tm_destroy(md->tm);
+               return r;
+       }
+
+       return 0;
+}
+
+static int open_metadata(struct era_metadata *md)
+{
+       int r;
+       struct dm_block *sblock;
+       struct superblock_disk *disk;
+
+       r = superblock_read_lock(md, &sblock);
+       if (r) {
+               DMERR("couldn't read_lock superblock");
+               return r;
+       }
+
+       disk = dm_block_data(sblock);
+       r = dm_tm_open_with_sm(md->bm, SUPERBLOCK_LOCATION,
+                              disk->metadata_space_map_root,
+                              sizeof(disk->metadata_space_map_root),
+                              &md->tm, &md->sm);
+       if (r) {
+               DMERR("dm_tm_open_with_sm failed");
+               goto bad;
+       }
+
+       setup_infos(md);
+
+       md->block_size = le32_to_cpu(disk->data_block_size);
+       md->nr_blocks = le32_to_cpu(disk->nr_blocks);
+       md->current_era = le32_to_cpu(disk->current_era);
+
+       md->writeset_tree_root = le64_to_cpu(disk->writeset_tree_root);
+       md->era_array_root = le64_to_cpu(disk->era_array_root);
+       md->metadata_snap = le64_to_cpu(disk->metadata_snap);
+       md->archived_writesets = true;
+
+       return dm_bm_unlock(sblock);
+
+bad:
+       dm_bm_unlock(sblock);
+       return r;
+}
+
+static int open_or_format_metadata(struct era_metadata *md,
+                                  bool may_format)
+{
+       int r;
+       bool unformatted = false;
+
+       r = superblock_all_zeroes(md->bm, &unformatted);
+       if (r)
+               return r;
+
+       if (unformatted)
+               return may_format ? format_metadata(md) : -EPERM;
+
+       return open_metadata(md);
+}
+
+static int create_persistent_data_objects(struct era_metadata *md,
+                                         bool may_format)
+{
+       int r;
+
+       md->bm = dm_block_manager_create(md->bdev, DM_ERA_METADATA_BLOCK_SIZE,
+                                        DM_ERA_METADATA_CACHE_SIZE,
+                                        ERA_MAX_CONCURRENT_LOCKS);
+       if (IS_ERR(md->bm)) {
+               DMERR("could not create block manager");
+               return PTR_ERR(md->bm);
+       }
+
+       r = open_or_format_metadata(md, may_format);
+       if (r)
+               dm_block_manager_destroy(md->bm);
+
+       return r;
+}
+
+static void destroy_persistent_data_objects(struct era_metadata *md)
+{
+       dm_sm_destroy(md->sm);
+       dm_tm_destroy(md->tm);
+       dm_block_manager_destroy(md->bm);
+}
+
+/*
+ * This waits until all era_map threads have picked up the new filter.
+ */
+static void swap_writeset(struct era_metadata *md, struct writeset *new_writeset)
+{
+       rcu_assign_pointer(md->current_writeset, new_writeset);
+       synchronize_rcu();
+}
+
+/*----------------------------------------------------------------
+ * Writesets get 'digested' into the main era array.
+ *
+ * We're using a coroutine here so the worker thread can do the digestion,
+ * thus avoiding synchronisation of the metadata.  Digesting a whole
+ * writeset in one go would cause too much latency.
+ *--------------------------------------------------------------*/
+struct digest {
+       uint32_t era;
+       unsigned nr_bits, current_bit;
+       struct writeset_metadata writeset;
+       __le32 value;
+       struct dm_disk_bitset info;
+
+       int (*step)(struct era_metadata *, struct digest *);
+};
+
+static int metadata_digest_lookup_writeset(struct era_metadata *md,
+                                          struct digest *d);
+
+static int metadata_digest_remove_writeset(struct era_metadata *md,
+                                          struct digest *d)
+{
+       int r;
+       uint64_t key = d->era;
+
+       r = dm_btree_remove(&md->writeset_tree_info, md->writeset_tree_root,
+                           &key, &md->writeset_tree_root);
+       if (r) {
+               DMERR("%s: dm_btree_remove failed", __func__);
+               return r;
+       }
+
+       d->step = metadata_digest_lookup_writeset;
+       return 0;
+}
+
+#define INSERTS_PER_STEP 100
+
+static int metadata_digest_transcribe_writeset(struct era_metadata *md,
+                                              struct digest *d)
+{
+       int r;
+       bool marked;
+       unsigned b, e = min(d->current_bit + INSERTS_PER_STEP, d->nr_bits);
+
+       for (b = d->current_bit; b < e; b++) {
+               r = writeset_marked_on_disk(&d->info, &d->writeset, b, &marked);
+               if (r) {
+                       DMERR("%s: writeset_marked_on_disk failed", __func__);
+                       return r;
+               }
+
+               if (!marked)
+                       continue;
+
+               __dm_bless_for_disk(&d->value);
+               r = dm_array_set_value(&md->era_array_info, md->era_array_root,
+                                      b, &d->value, &md->era_array_root);
+               if (r) {
+                       DMERR("%s: dm_array_set_value failed", __func__);
+                       return r;
+               }
+       }
+
+       if (b == d->nr_bits)
+               d->step = metadata_digest_remove_writeset;
+       else
+               d->current_bit = b;
+
+       return 0;
+}
+
+static int metadata_digest_lookup_writeset(struct era_metadata *md,
+                                          struct digest *d)
+{
+       int r;
+       uint64_t key;
+       struct writeset_disk disk;
+
+       r = dm_btree_find_lowest_key(&md->writeset_tree_info,
+                                    md->writeset_tree_root, &key);
+       if (r < 0)
+               return r;
+
+       d->era = key;
+
+       r = dm_btree_lookup(&md->writeset_tree_info,
+                           md->writeset_tree_root, &key, &disk);
+       if (r) {
+               if (r == -ENODATA) {
+                       d->step = NULL;
+                       return 0;
+               }
+
+               DMERR("%s: dm_btree_lookup failed", __func__);
+               return r;
+       }
+
+       ws_unpack(&disk, &d->writeset);
+       d->value = cpu_to_le32(key);
+
+       d->nr_bits = min(d->writeset.nr_bits, md->nr_blocks);
+       d->current_bit = 0;
+       d->step = metadata_digest_transcribe_writeset;
+
+       return 0;
+}
+
+static int metadata_digest_start(struct era_metadata *md, struct digest *d)
+{
+       if (d->step)
+               return 0;
+
+       memset(d, 0, sizeof(*d));
+
+       /*
+        * We initialise another bitset info to avoid any caching side
+        * effects with the previous one.
+        */
+       dm_disk_bitset_init(md->tm, &d->info);
+       d->step = metadata_digest_lookup_writeset;
+
+       return 0;
+}
+
+/*----------------------------------------------------------------
+ * High level metadata interface.  Target methods should use these, and not
+ * the lower level ones.
+ *--------------------------------------------------------------*/
+static struct era_metadata *metadata_open(struct block_device *bdev,
+                                         sector_t block_size,
+                                         bool may_format)
+{
+       int r;
+       struct era_metadata *md = kzalloc(sizeof(*md), GFP_KERNEL);
+
+       if (!md)
+               return NULL;
+
+       md->bdev = bdev;
+       md->block_size = block_size;
+
+       md->writesets[0].md.root = INVALID_WRITESET_ROOT;
+       md->writesets[1].md.root = INVALID_WRITESET_ROOT;
+       md->current_writeset = &md->writesets[0];
+
+       r = create_persistent_data_objects(md, may_format);
+       if (r) {
+               kfree(md);
+               return ERR_PTR(r);
+       }
+
+       return md;
+}
+
+static void metadata_close(struct era_metadata *md)
+{
+       destroy_persistent_data_objects(md);
+       kfree(md);
+}
+
+static bool valid_nr_blocks(dm_block_t n)
+{
+       /*
+        * dm_bitset restricts us to 2^32.  test_bit & co. restrict us
+        * further to 2^31 - 1
+        */
+       return n < (1ull << 31);
+}
+
+static int metadata_resize(struct era_metadata *md, void *arg)
+{
+       int r;
+       dm_block_t *new_size = arg;
+       __le32 value;
+
+       if (!valid_nr_blocks(*new_size)) {
+               DMERR("Invalid number of origin blocks %llu",
+                     (unsigned long long) *new_size);
+               return -EINVAL;
+       }
+
+       writeset_free(&md->writesets[0]);
+       writeset_free(&md->writesets[1]);
+
+       r = writeset_alloc(&md->writesets[0], *new_size);
+       if (r) {
+               DMERR("%s: writeset_alloc failed for writeset 0", __func__);
+               return r;
+       }
+
+       r = writeset_alloc(&md->writesets[1], *new_size);
+       if (r) {
+               DMERR("%s: writeset_alloc failed for writeset 1", __func__);
+               return r;
+       }
+
+       value = cpu_to_le32(0u);
+       __dm_bless_for_disk(&value);
+       r = dm_array_resize(&md->era_array_info, md->era_array_root,
+                           md->nr_blocks, *new_size,
+                           &value, &md->era_array_root);
+       if (r) {
+               DMERR("%s: dm_array_resize failed", __func__);
+               return r;
+       }
+
+       md->nr_blocks = *new_size;
+       return 0;
+}
+
+static int metadata_era_archive(struct era_metadata *md)
+{
+       int r;
+       uint64_t keys[1];
+       struct writeset_disk value;
+
+       r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
+                           &md->current_writeset->md.root);
+       if (r) {
+               DMERR("%s: dm_bitset_flush failed", __func__);
+               return r;
+       }
+
+       ws_pack(&md->current_writeset->md, &value);
+       md->current_writeset->md.root = INVALID_WRITESET_ROOT;
+
+       keys[0] = md->current_era;
+       __dm_bless_for_disk(&value);
+       r = dm_btree_insert(&md->writeset_tree_info, md->writeset_tree_root,
+                           keys, &value, &md->writeset_tree_root);
+       if (r) {
+               DMERR("%s: couldn't insert writeset into btree", __func__);
+               /* FIXME: fail mode */
+               return r;
+       }
+
+       md->archived_writesets = true;
+
+       return 0;
+}
+
+static struct writeset *next_writeset(struct era_metadata *md)
+{
+       return (md->current_writeset == &md->writesets[0]) ?
+               &md->writesets[1] : &md->writesets[0];
+}
+
+static int metadata_new_era(struct era_metadata *md)
+{
+       int r;
+       struct writeset *new_writeset = next_writeset(md);
+
+       r = writeset_init(&md->bitset_info, new_writeset);
+       if (r) {
+               DMERR("%s: writeset_init failed", __func__);
+               return r;
+       }
+
+       swap_writeset(md, new_writeset);
+       md->current_era++;
+
+       return 0;
+}
+
+static int metadata_era_rollover(struct era_metadata *md)
+{
+       int r;
+
+       if (md->current_writeset->md.root != INVALID_WRITESET_ROOT) {
+               r = metadata_era_archive(md);
+               if (r) {
+                       DMERR("%s: metadata_archive_era failed", __func__);
+                       /* FIXME: fail mode? */
+                       return r;
+               }
+       }
+
+       r = metadata_new_era(md);
+       if (r) {
+               DMERR("%s: new era failed", __func__);
+               /* FIXME: fail mode */
+               return r;
+       }
+
+       return 0;
+}
+
+static bool metadata_current_marked(struct era_metadata *md, dm_block_t block)
+{
+       bool r;
+       struct writeset *ws;
+
+       rcu_read_lock();
+       ws = rcu_dereference(md->current_writeset);
+       r = writeset_marked(ws, block);
+       rcu_read_unlock();
+
+       return r;
+}
+
+static int metadata_commit(struct era_metadata *md)
+{
+       int r;
+       struct dm_block *sblock;
+
+       if (md->current_writeset->md.root != SUPERBLOCK_LOCATION) {
+               r = dm_bitset_flush(&md->bitset_info, md->current_writeset->md.root,
+                                   &md->current_writeset->md.root);
+               if (r) {
+                       DMERR("%s: bitset flush failed", __func__);
+                       return r;
+               }
+       }
+
+       r = dm_tm_pre_commit(md->tm);
+       if (r) {
+               DMERR("%s: pre commit failed", __func__);
+               return r;
+       }
+
+       r = superblock_lock(md, &sblock);
+       if (r) {
+               DMERR("%s: superblock lock failed", __func__);
+               return r;
+       }
+
+       r = prepare_superblock(md, dm_block_data(sblock));
+       if (r) {
+               DMERR("%s: prepare_superblock failed", __func__);
+               dm_bm_unlock(sblock); /* FIXME: does this commit? */
+               return r;
+       }
+
+       return dm_tm_commit(md->tm, sblock);
+}
+
+static int metadata_checkpoint(struct era_metadata *md)
+{
+       /*
+        * For now we just rollover, but later I want to put a check in to
+        * avoid this if the filter is still pretty fresh.
+        */
+       return metadata_era_rollover(md);
+}
+
+/*
+ * Metadata snapshots allow userland to access era data.
+ */
+static int metadata_take_snap(struct era_metadata *md)
+{
+       int r, inc;
+       struct dm_block *clone;
+
+       if (md->metadata_snap != SUPERBLOCK_LOCATION) {
+               DMERR("%s: metadata snapshot already exists", __func__);
+               return -EINVAL;
+       }
+
+       r = metadata_era_rollover(md);
+       if (r) {
+               DMERR("%s: era rollover failed", __func__);
+               return r;
+       }
+
+       r = metadata_commit(md);
+       if (r) {
+               DMERR("%s: pre commit failed", __func__);
+               return r;
+       }
+
+       r = dm_sm_inc_block(md->sm, SUPERBLOCK_LOCATION);
+       if (r) {
+               DMERR("%s: couldn't increment superblock", __func__);
+               return r;
+       }
+
+       r = dm_tm_shadow_block(md->tm, SUPERBLOCK_LOCATION,
+                              &sb_validator, &clone, &inc);
+       if (r) {
+               DMERR("%s: couldn't shadow superblock", __func__);
+               dm_sm_dec_block(md->sm, SUPERBLOCK_LOCATION);
+               return r;
+       }
+       BUG_ON(!inc);
+
+       r = dm_sm_inc_block(md->sm, md->writeset_tree_root);
+       if (r) {
+               DMERR("%s: couldn't inc writeset tree root", __func__);
+               dm_tm_unlock(md->tm, clone);
+               return r;
+       }
+
+       r = dm_sm_inc_block(md->sm, md->era_array_root);
+       if (r) {
+               DMERR("%s: couldn't inc era tree root", __func__);
+               dm_sm_dec_block(md->sm, md->writeset_tree_root);
+               dm_tm_unlock(md->tm, clone);
+               return r;
+       }
+
+       md->metadata_snap = dm_block_location(clone);
+
+       r = dm_tm_unlock(md->tm, clone);
+       if (r) {
+               DMERR("%s: couldn't unlock clone", __func__);
+               md->metadata_snap = SUPERBLOCK_LOCATION;
+               return r;
+       }
+
+       return 0;
+}
+
+static int metadata_drop_snap(struct era_metadata *md)
+{
+       int r;
+       dm_block_t location;
+       struct dm_block *clone;
+       struct superblock_disk *disk;
+
+       if (md->metadata_snap == SUPERBLOCK_LOCATION) {
+               DMERR("%s: no snap to drop", __func__);
+               return -EINVAL;
+       }
+
+       r = dm_tm_read_lock(md->tm, md->metadata_snap, &sb_validator, &clone);
+       if (r) {
+               DMERR("%s: couldn't read lock superblock clone", __func__);
+               return r;
+       }
+
+       /*
+        * Whatever happens now we'll commit with no record of the metadata
+        * snap.
+        */
+       md->metadata_snap = SUPERBLOCK_LOCATION;
+
+       disk = dm_block_data(clone);
+       r = dm_btree_del(&md->writeset_tree_info,
+                        le64_to_cpu(disk->writeset_tree_root));
+       if (r) {
+               DMERR("%s: error deleting writeset tree clone", __func__);
+               dm_tm_unlock(md->tm, clone);
+               return r;
+       }
+
+       r = dm_array_del(&md->era_array_info, le64_to_cpu(disk->era_array_root));
+       if (r) {
+               DMERR("%s: error deleting era array clone", __func__);
+               dm_tm_unlock(md->tm, clone);
+               return r;
+       }
+
+       location = dm_block_location(clone);
+       dm_tm_unlock(md->tm, clone);
+
+       return dm_sm_dec_block(md->sm, location);
+}
+
+struct metadata_stats {
+       dm_block_t used;
+       dm_block_t total;
+       dm_block_t snap;
+       uint32_t era;
+};
+
+static int metadata_get_stats(struct era_metadata *md, void *ptr)
+{
+       int r;
+       struct metadata_stats *s = ptr;
+       dm_block_t nr_free, nr_total;
+
+       r = dm_sm_get_nr_free(md->sm, &nr_free);
+       if (r) {
+               DMERR("dm_sm_get_nr_free returned %d", r);
+               return r;
+       }
+
+       r = dm_sm_get_nr_blocks(md->sm, &nr_total);
+       if (r) {
+               DMERR("dm_pool_get_metadata_dev_size returned %d", r);
+               return r;
+       }
+
+       s->used = nr_total - nr_free;
+       s->total = nr_total;
+       s->snap = md->metadata_snap;
+       s->era = md->current_era;
+
+       return 0;
+}
+
+/*----------------------------------------------------------------*/
+
+struct era {
+       struct dm_target *ti;
+       struct dm_target_callbacks callbacks;
+
+       struct dm_dev *metadata_dev;
+       struct dm_dev *origin_dev;
+
+       dm_block_t nr_blocks;
+       uint32_t sectors_per_block;
+       int sectors_per_block_shift;
+       struct era_metadata *md;
+
+       struct workqueue_struct *wq;
+       struct work_struct worker;
+
+       spinlock_t deferred_lock;
+       struct bio_list deferred_bios;
+
+       spinlock_t rpc_lock;
+       struct list_head rpc_calls;
+
+       struct digest digest;
+       atomic_t suspended;
+};
+
+struct rpc {
+       struct list_head list;
+
+       int (*fn0)(struct era_metadata *);
+       int (*fn1)(struct era_metadata *, void *);
+       void *arg;
+       int result;
+
+       struct completion complete;
+};
+
+/*----------------------------------------------------------------
+ * Remapping.
+ *---------------------------------------------------------------*/
+static bool block_size_is_power_of_two(struct era *era)
+{
+       return era->sectors_per_block_shift >= 0;
+}
+
+static dm_block_t get_block(struct era *era, struct bio *bio)
+{
+       sector_t block_nr = bio->bi_iter.bi_sector;
+
+       if (!block_size_is_power_of_two(era))
+               (void) sector_div(block_nr, era->sectors_per_block);
+       else
+               block_nr >>= era->sectors_per_block_shift;
+
+       return block_nr;
+}
+
+static void remap_to_origin(struct era *era, struct bio *bio)
+{
+       bio->bi_bdev = era->origin_dev->bdev;
+}
+
+/*----------------------------------------------------------------
+ * Worker thread
+ *--------------------------------------------------------------*/
+static void wake_worker(struct era *era)
+{
+       if (!atomic_read(&era->suspended))
+               queue_work(era->wq, &era->worker);
+}
+
+static void process_old_eras(struct era *era)
+{
+       int r;
+
+       if (!era->digest.step)
+               return;
+
+       r = era->digest.step(era->md, &era->digest);
+       if (r < 0) {
+               DMERR("%s: digest step failed, stopping digestion", __func__);
+               era->digest.step = NULL;
+
+       } else if (era->digest.step)
+               wake_worker(era);
+}
+
+static void process_deferred_bios(struct era *era)
+{
+       int r;
+       struct bio_list deferred_bios, marked_bios;
+       struct bio *bio;
+       bool commit_needed = false;
+       bool failed = false;
+
+       bio_list_init(&deferred_bios);
+       bio_list_init(&marked_bios);
+
+       spin_lock(&era->deferred_lock);
+       bio_list_merge(&deferred_bios, &era->deferred_bios);
+       bio_list_init(&era->deferred_bios);
+       spin_unlock(&era->deferred_lock);
+
+       while ((bio = bio_list_pop(&deferred_bios))) {
+               r = writeset_test_and_set(&era->md->bitset_info,
+                                         era->md->current_writeset,
+                                         get_block(era, bio));
+               if (r < 0) {
+                       /*
+                        * This is bad news, we need to rollback.
+                        * FIXME: finish.
+                        */
+                       failed = true;
+
+               } else if (r == 0)
+                       commit_needed = true;
+
+               bio_list_add(&marked_bios, bio);
+       }
+
+       if (commit_needed) {
+               r = metadata_commit(era->md);
+               if (r)
+                       failed = true;
+       }
+
+       if (failed)
+               while ((bio = bio_list_pop(&marked_bios)))
+                       bio_io_error(bio);
+       else
+               while ((bio = bio_list_pop(&marked_bios)))
+                       generic_make_request(bio);
+}
+
+static void process_rpc_calls(struct era *era)
+{
+       int r;
+       bool need_commit = false;
+       struct list_head calls;
+       struct rpc *rpc, *tmp;
+
+       INIT_LIST_HEAD(&calls);
+       spin_lock(&era->rpc_lock);
+       list_splice_init(&era->rpc_calls, &calls);
+       spin_unlock(&era->rpc_lock);
+
+       list_for_each_entry_safe(rpc, tmp, &calls, list) {
+               rpc->result = rpc->fn0 ? rpc->fn0(era->md) : rpc->fn1(era->md, rpc->arg);
+               need_commit = true;
+       }
+
+       if (need_commit) {
+               r = metadata_commit(era->md);
+               if (r)
+                       list_for_each_entry_safe(rpc, tmp, &calls, list)
+                               rpc->result = r;
+       }
+
+       list_for_each_entry_safe(rpc, tmp, &calls, list)
+               complete(&rpc->complete);
+}
+
+static void kick_off_digest(struct era *era)
+{
+       if (era->md->archived_writesets) {
+               era->md->archived_writesets = false;
+               metadata_digest_start(era->md, &era->digest);
+       }
+}
+
+static void do_work(struct work_struct *ws)
+{
+       struct era *era = container_of(ws, struct era, worker);
+
+       kick_off_digest(era);
+       process_old_eras(era);
+       process_deferred_bios(era);
+       process_rpc_calls(era);
+}
+
+static void defer_bio(struct era *era, struct bio *bio)
+{
+       spin_lock(&era->deferred_lock);
+       bio_list_add(&era->deferred_bios, bio);
+       spin_unlock(&era->deferred_lock);
+
+       wake_worker(era);
+}
+
+/*
+ * Make an rpc call to the worker to change the metadata.
+ */
+static int perform_rpc(struct era *era, struct rpc *rpc)
+{
+       rpc->result = 0;
+       init_completion(&rpc->complete);
+
+       spin_lock(&era->rpc_lock);
+       list_add(&rpc->list, &era->rpc_calls);
+       spin_unlock(&era->rpc_lock);
+
+       wake_worker(era);
+       wait_for_completion(&rpc->complete);
+
+       return rpc->result;
+}
+
+static int in_worker0(struct era *era, int (*fn)(struct era_metadata *))
+{
+       struct rpc rpc;
+       rpc.fn0 = fn;
+       rpc.fn1 = NULL;
+
+       return perform_rpc(era, &rpc);
+}
+
+static int in_worker1(struct era *era,
+                     int (*fn)(struct era_metadata *, void *), void *arg)
+{
+       struct rpc rpc;
+       rpc.fn0 = NULL;
+       rpc.fn1 = fn;
+       rpc.arg = arg;
+
+       return perform_rpc(era, &rpc);
+}
+
+static void start_worker(struct era *era)
+{
+       atomic_set(&era->suspended, 0);
+}
+
+static void stop_worker(struct era *era)
+{
+       atomic_set(&era->suspended, 1);
+       flush_workqueue(era->wq);
+}
+
+/*----------------------------------------------------------------
+ * Target methods
+ *--------------------------------------------------------------*/
+static int dev_is_congested(struct dm_dev *dev, int bdi_bits)
+{
+       struct request_queue *q = bdev_get_queue(dev->bdev);
+       return bdi_congested(&q->backing_dev_info, bdi_bits);
+}
+
+static int era_is_congested(struct dm_target_callbacks *cb, int bdi_bits)
+{
+       struct era *era = container_of(cb, struct era, callbacks);
+       return dev_is_congested(era->origin_dev, bdi_bits);
+}
+
+static void era_destroy(struct era *era)
+{
+       metadata_close(era->md);
+
+       if (era->wq)
+               destroy_workqueue(era->wq);
+
+       if (era->origin_dev)
+               dm_put_device(era->ti, era->origin_dev);
+
+       if (era->metadata_dev)
+               dm_put_device(era->ti, era->metadata_dev);
+
+       kfree(era);
+}
+
+static dm_block_t calc_nr_blocks(struct era *era)
+{
+       return dm_sector_div_up(era->ti->len, era->sectors_per_block);
+}
+
+static bool valid_block_size(dm_block_t block_size)
+{
+       bool greater_than_zero = block_size > 0;
+       bool multiple_of_min_block_size = (block_size & (MIN_BLOCK_SIZE - 1)) == 0;
+
+       return greater_than_zero && multiple_of_min_block_size;
+}
+
+/*
+ * <metadata dev> <data dev> <data block size (sectors)>
+ */
+static int era_ctr(struct dm_target *ti, unsigned argc, char **argv)
+{
+       int r;
+       char dummy;
+       struct era *era;
+       struct era_metadata *md;
+
+       if (argc != 3) {
+               ti->error = "Invalid argument count";
+               return -EINVAL;
+       }
+
+       era = kzalloc(sizeof(*era), GFP_KERNEL);
+       if (!era) {
+               ti->error = "Error allocating era structure";
+               return -ENOMEM;
+       }
+
+       era->ti = ti;
+
+       r = dm_get_device(ti, argv[0], FMODE_READ | FMODE_WRITE, &era->metadata_dev);
+       if (r) {
+               ti->error = "Error opening metadata device";
+               era_destroy(era);
+               return -EINVAL;
+       }
+
+       r = dm_get_device(ti, argv[1], FMODE_READ | FMODE_WRITE, &era->origin_dev);
+       if (r) {
+               ti->error = "Error opening data device";
+               era_destroy(era);
+               return -EINVAL;
+       }
+
+       r = sscanf(argv[2], "%u%c", &era->sectors_per_block, &dummy);
+       if (r != 1) {
+               ti->error = "Error parsing block size";
+               era_destroy(era);
+               return -EINVAL;
+       }
+
+       r = dm_set_target_max_io_len(ti, era->sectors_per_block);
+       if (r) {
+               ti->error = "could not set max io len";
+               era_destroy(era);
+               return -EINVAL;
+       }
+
+       if (!valid_block_size(era->sectors_per_block)) {
+               ti->error = "Invalid block size";
+               era_destroy(era);
+               return -EINVAL;
+       }
+       if (era->sectors_per_block & (era->sectors_per_block - 1))
+               era->sectors_per_block_shift = -1;
+       else
+               era->sectors_per_block_shift = __ffs(era->sectors_per_block);
+
+       md = metadata_open(era->metadata_dev->bdev, era->sectors_per_block, true);
+       if (IS_ERR(md)) {
+               ti->error = "Error reading metadata";
+               era_destroy(era);
+               return PTR_ERR(md);
+       }
+       era->md = md;
+
+       era->nr_blocks = calc_nr_blocks(era);
+
+       r = metadata_resize(era->md, &era->nr_blocks);
+       if (r) {
+               ti->error = "couldn't resize metadata";
+               era_destroy(era);
+               return -ENOMEM;
+       }
+
+       era->wq = alloc_ordered_workqueue("dm-" DM_MSG_PREFIX, WQ_MEM_RECLAIM);
+       if (!era->wq) {
+               ti->error = "could not create workqueue for metadata object";
+               era_destroy(era);
+               return -ENOMEM;
+       }
+       INIT_WORK(&era->worker, do_work);
+
+       spin_lock_init(&era->deferred_lock);
+       bio_list_init(&era->deferred_bios);
+
+       spin_lock_init(&era->rpc_lock);
+       INIT_LIST_HEAD(&era->rpc_calls);
+
+       ti->private = era;
+       ti->num_flush_bios = 1;
+       ti->flush_supported = true;
+
+       ti->num_discard_bios = 1;
+       ti->discards_supported = true;
+       era->callbacks.congested_fn = era_is_congested;
+       dm_table_add_target_callbacks(ti->table, &era->callbacks);
+
+       return 0;
+}
+
+static void era_dtr(struct dm_target *ti)
+{
+       era_destroy(ti->private);
+}
+
+static int era_map(struct dm_target *ti, struct bio *bio)
+{
+       struct era *era = ti->private;
+       dm_block_t block = get_block(era, bio);
+
+       /*
+        * All bios get remapped to the origin device.  We do this now, but
+        * it may not get issued until later.  Depending on whether the
+        * block is marked in this era.
+        */
+       remap_to_origin(era, bio);
+
+       /*
+        * REQ_FLUSH bios carry no data, so we're not interested in them.
+        */
+       if (!(bio->bi_rw & REQ_FLUSH) &&
+           (bio_data_dir(bio) == WRITE) &&
+           !metadata_current_marked(era->md, block)) {
+               defer_bio(era, bio);
+               return DM_MAPIO_SUBMITTED;
+       }
+
+       return DM_MAPIO_REMAPPED;
+}
+
+static void era_postsuspend(struct dm_target *ti)
+{
+       int r;
+       struct era *era = ti->private;
+
+       r = in_worker0(era, metadata_era_archive);
+       if (r) {
+               DMERR("%s: couldn't archive current era", __func__);
+               /* FIXME: fail mode */
+       }
+
+       stop_worker(era);
+}
+
+static int era_preresume(struct dm_target *ti)
+{
+       int r;
+       struct era *era = ti->private;
+       dm_block_t new_size = calc_nr_blocks(era);
+
+       if (era->nr_blocks != new_size) {
+               r = in_worker1(era, metadata_resize, &new_size);
+               if (r)
+                       return r;
+
+               era->nr_blocks = new_size;
+       }
+
+       start_worker(era);
+
+       r = in_worker0(era, metadata_new_era);
+       if (r) {
+               DMERR("%s: metadata_era_rollover failed", __func__);
+               return r;
+       }
+
+       return 0;
+}
+
+/*
+ * Status format:
+ *
+ * <metadata block size> <#used metadata blocks>/<#total metadata blocks>
+ * <current era> <held metadata root | '-'>
+ */
+static void era_status(struct dm_target *ti, status_type_t type,
+                      unsigned status_flags, char *result, unsigned maxlen)
+{
+       int r;
+       struct era *era = ti->private;
+       ssize_t sz = 0;
+       struct metadata_stats stats;
+       char buf[BDEVNAME_SIZE];
+
+       switch (type) {
+       case STATUSTYPE_INFO:
+               r = in_worker1(era, metadata_get_stats, &stats);
+               if (r)
+                       goto err;
+
+               DMEMIT("%u %llu/%llu %u",
+                      (unsigned) (DM_ERA_METADATA_BLOCK_SIZE >> SECTOR_SHIFT),
+                      (unsigned long long) stats.used,
+                      (unsigned long long) stats.total,
+                      (unsigned) stats.era);
+
+               if (stats.snap != SUPERBLOCK_LOCATION)
+                       DMEMIT(" %llu", stats.snap);
+               else
+                       DMEMIT(" -");
+               break;
+
+       case STATUSTYPE_TABLE:
+               format_dev_t(buf, era->metadata_dev->bdev->bd_dev);
+               DMEMIT("%s ", buf);
+               format_dev_t(buf, era->origin_dev->bdev->bd_dev);
+               DMEMIT("%s %u", buf, era->sectors_per_block);
+               break;
+       }
+
+       return;
+
+err:
+       DMEMIT("Error");
+}
+
+static int era_message(struct dm_target *ti, unsigned argc, char **argv)
+{
+       struct era *era = ti->private;
+
+       if (argc != 1) {
+               DMERR("incorrect number of message arguments");
+               return -EINVAL;
+       }
+
+       if (!strcasecmp(argv[0], "checkpoint"))
+               return in_worker0(era, metadata_checkpoint);
+
+       if (!strcasecmp(argv[0], "take_metadata_snap"))
+               return in_worker0(era, metadata_take_snap);
+
+       if (!strcasecmp(argv[0], "drop_metadata_snap"))
+               return in_worker0(era, metadata_drop_snap);
+
+       DMERR("unsupported message '%s'", argv[0]);
+       return -EINVAL;
+}
+
+static sector_t get_dev_size(struct dm_dev *dev)
+{
+       return i_size_read(dev->bdev->bd_inode) >> SECTOR_SHIFT;
+}
+
+static int era_iterate_devices(struct dm_target *ti,
+                              iterate_devices_callout_fn fn, void *data)
+{
+       struct era *era = ti->private;
+       return fn(ti, era->origin_dev, 0, get_dev_size(era->origin_dev), data);
+}
+
+static int era_merge(struct dm_target *ti, struct bvec_merge_data *bvm,
+                    struct bio_vec *biovec, int max_size)
+{
+       struct era *era = ti->private;
+       struct request_queue *q = bdev_get_queue(era->origin_dev->bdev);
+
+       if (!q->merge_bvec_fn)
+               return max_size;
+
+       bvm->bi_bdev = era->origin_dev->bdev;
+
+       return min(max_size, q->merge_bvec_fn(q, bvm, biovec));
+}
+
+static void era_io_hints(struct dm_target *ti, struct queue_limits *limits)
+{
+       struct era *era = ti->private;
+       uint64_t io_opt_sectors = limits->io_opt >> SECTOR_SHIFT;
+
+       /*
+        * If the system-determined stacked limits are compatible with the
+        * era device's blocksize (io_opt is a factor) do not override them.
+        */
+       if (io_opt_sectors < era->sectors_per_block ||
+           do_div(io_opt_sectors, era->sectors_per_block)) {
+               blk_limits_io_min(limits, 0);
+               blk_limits_io_opt(limits, era->sectors_per_block << SECTOR_SHIFT);
+       }
+}
+
+/*----------------------------------------------------------------*/
+
+static struct target_type era_target = {
+       .name = "era",
+       .version = {1, 0, 0},
+       .module = THIS_MODULE,
+       .ctr = era_ctr,
+       .dtr = era_dtr,
+       .map = era_map,
+       .postsuspend = era_postsuspend,
+       .preresume = era_preresume,
+       .status = era_status,
+       .message = era_message,
+       .iterate_devices = era_iterate_devices,
+       .merge = era_merge,
+       .io_hints = era_io_hints
+};
+
+static int __init dm_era_init(void)
+{
+       int r;
+
+       r = dm_register_target(&era_target);
+       if (r) {
+               DMERR("era target registration failed: %d", r);
+               return r;
+       }
+
+       return 0;
+}
+
+static void __exit dm_era_exit(void)
+{
+       dm_unregister_target(&era_target);
+}
+
+module_init(dm_era_init);
+module_exit(dm_era_exit);
+
+MODULE_DESCRIPTION(DM_NAME " era target");
+MODULE_AUTHOR("Joe Thornber <ejt@redhat.com>");
+MODULE_LICENSE("GPL");