dm: add thin provisioning target

Initial EXPERIMENTAL implementation of device-mapper thin provisioning
with snapshot support.  The 'thin' target is used to create instances of
the virtual devices that are hosted in the 'thin-pool' target.  The
thin-pool target provides data sharing among devices.  This sharing is
made possible using the persistent-data library in the previous patch.

The main highlight of this implementation, compared to the previous
implementation of snapshots, is that it allows many virtual devices to
be stored on the same data volume, simplifying administration and
allowing sharing of data between volumes (thus reducing disk usage).

Another big feature is support for arbitrary depth of recursive
snapshots (snapshots of snapshots of snapshots ...).  The previous
implementation of snapshots did this by chaining together lookup tables,
and so performance was O(depth).  This new implementation uses a single
data structure so we don't get this degradation with depth.

For further information and examples of how to use this, please read
Documentation/device-mapper/thin-provisioning.txt

Signed-off-by: Joe Thornber <thornber@redhat.com>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>

1 files changed, 2428 insertions, 0 deletions

diff --git a/drivers/md/dm-thin.c b/drivers/md/dm-thin.c
new file mode 100644
index 000000000000..c3087575fef0
--- /dev/null
+++ b/drivers/md/dm-thin.c
@@ -0,0 +1,2428 @@
+/*
+ * Copyright (C) 2011 Red Hat UK.
+ *
+ * This file is released under the GPL.
+ */
+
+#include "dm-thin-metadata.h"
+
+#include <linux/device-mapper.h>
+#include <linux/dm-io.h>
+#include <linux/dm-kcopyd.h>
+#include <linux/list.h>
+#include <linux/init.h>
+#include <linux/module.h>
+#include <linux/slab.h>
+
+#define	DM_MSG_PREFIX	"thin"
+
+/*
+ * Tunable constants
+ */
+#define ENDIO_HOOK_POOL_SIZE 10240
+#define DEFERRED_SET_SIZE 64
+#define MAPPING_POOL_SIZE 1024
+#define PRISON_CELLS 1024
+
+/*
+ * The block size of the device holding pool data must be
+ * between 64KB and 1GB.
+ */
+#define DATA_DEV_BLOCK_SIZE_MIN_SECTORS (64 * 1024 >> SECTOR_SHIFT)
+#define DATA_DEV_BLOCK_SIZE_MAX_SECTORS (1024 * 1024 * 1024 >> SECTOR_SHIFT)
+
+/*
+ * The metadata device is currently limited in size.  The limitation is
+ * checked lower down in dm-space-map-metadata, but we also check it here
+ * so we can fail early.
+ *
+ * We have one block of index, which can hold 255 index entries.  Each
+ * index entry contains allocation info about 16k metadata blocks.
+ */
+#define METADATA_DEV_MAX_SECTORS (255 * (1 << 14) * (THIN_METADATA_BLOCK_SIZE / (1 << SECTOR_SHIFT)))
+
+/*
+ * Device id is restricted to 24 bits.
+ */
+#define MAX_DEV_ID ((1 << 24) - 1)
+
+/*
+ * How do we handle breaking sharing of data blocks?
+ * =================================================
+ *
+ * We use a standard copy-on-write btree to store the mappings for the
+ * devices (note I'm talking about copy-on-write of the metadata here, not
+ * the data).  When you take an internal snapshot you clone the root node
+ * of the origin btree.  After this there is no concept of an origin or a
+ * snapshot.  They are just two device trees that happen to point to the
+ * same data blocks.
+ *
+ * When we get a write in we decide if it's to a shared data block using
+ * some timestamp magic.  If it is, we have to break sharing.
+ *
+ * Let's say we write to a shared block in what was the origin.  The
+ * steps are:
+ *
+ * i) plug io further to this physical block. (see bio_prison code).
+ *
+ * ii) quiesce any read io to that shared data block.  Obviously
+ * including all devices that share this block.  (see deferred_set code)
+ *
+ * iii) copy the data block to a newly allocate block.  This step can be
+ * missed out if the io covers the block. (schedule_copy).
+ *
+ * iv) insert the new mapping into the origin's btree
+ * (process_prepared_mappings).  This act of inserting breaks some
+ * sharing of btree nodes between the two devices.  Breaking sharing only
+ * effects the btree of that specific device.  Btrees for the other
+ * devices that share the block never change.  The btree for the origin
+ * device as it was after the last commit is untouched, ie. we're using
+ * persistent data structures in the functional programming sense.
+ *
+ * v) unplug io to this physical block, including the io that triggered
+ * the breaking of sharing.
+ *
+ * Steps (ii) and (iii) occur in parallel.
+ *
+ * The metadata _doesn't_ need to be committed before the io continues.  We
+ * get away with this because the io is always written to a _new_ block.
+ * If there's a crash, then:
+ *
+ * - The origin mapping will point to the old origin block (the shared
+ * one).  This will contain the data as it was before the io that triggered
+ * the breaking of sharing came in.
+ *
+ * - The snap mapping still points to the old block.  As it would after
+ * the commit.
+ *
+ * The downside of this scheme is the timestamp magic isn't perfect, and
+ * will continue to think that data block in the snapshot device is shared
+ * even after the write to the origin has broken sharing.  I suspect data
+ * blocks will typically be shared by many different devices, so we're
+ * breaking sharing n + 1 times, rather than n, where n is the number of
+ * devices that reference this data block.  At the moment I think the
+ * benefits far, far outweigh the disadvantages.
+ */
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Sometimes we can't deal with a bio straight away.  We put them in prison
+ * where they can't cause any mischief.  Bios are put in a cell identified
+ * by a key, multiple bios can be in the same cell.  When the cell is
+ * subsequently unlocked the bios become available.
+ */
+struct bio_prison;
+
+struct cell_key {
+	int virtual;
+	dm_thin_id dev;
+	dm_block_t block;
+};
+
+struct cell {
+	struct hlist_node list;
+	struct bio_prison *prison;
+	struct cell_key key;
+	unsigned count;
+	struct bio_list bios;
+};
+
+struct bio_prison {
+	spinlock_t lock;
+	mempool_t *cell_pool;
+
+	unsigned nr_buckets;
+	unsigned hash_mask;
+	struct hlist_head *cells;
+};
+
+static uint32_t calc_nr_buckets(unsigned nr_cells)
+{
+	uint32_t n = 128;
+
+	nr_cells /= 4;
+	nr_cells = min(nr_cells, 8192u);
+
+	while (n < nr_cells)
+		n <<= 1;
+
+	return n;
+}
+
+/*
+ * @nr_cells should be the number of cells you want in use _concurrently_.
+ * Don't confuse it with the number of distinct keys.
+ */
+static struct bio_prison *prison_create(unsigned nr_cells)
+{
+	unsigned i;
+	uint32_t nr_buckets = calc_nr_buckets(nr_cells);
+	size_t len = sizeof(struct bio_prison) +
+		(sizeof(struct hlist_head) * nr_buckets);
+	struct bio_prison *prison = kmalloc(len, GFP_KERNEL);
+
+	if (!prison)
+		return NULL;
+
+	spin_lock_init(&prison->lock);
+	prison->cell_pool = mempool_create_kmalloc_pool(nr_cells,
+							sizeof(struct cell));
+	if (!prison->cell_pool) {
+		kfree(prison);
+		return NULL;
+	}
+
+	prison->nr_buckets = nr_buckets;
+	prison->hash_mask = nr_buckets - 1;
+	prison->cells = (struct hlist_head *) (prison + 1);
+	for (i = 0; i < nr_buckets; i++)
+		INIT_HLIST_HEAD(prison->cells + i);
+
+	return prison;
+}
+
+static void prison_destroy(struct bio_prison *prison)
+{
+	mempool_destroy(prison->cell_pool);
+	kfree(prison);
+}
+
+static uint32_t hash_key(struct bio_prison *prison, struct cell_key *key)
+{
+	const unsigned long BIG_PRIME = 4294967291UL;
+	uint64_t hash = key->block * BIG_PRIME;
+
+	return (uint32_t) (hash & prison->hash_mask);
+}
+
+static int keys_equal(struct cell_key *lhs, struct cell_key *rhs)
+{
+	       return (lhs->virtual == rhs->virtual) &&
+		       (lhs->dev == rhs->dev) &&
+		       (lhs->block == rhs->block);
+}
+
+static struct cell *__search_bucket(struct hlist_head *bucket,
+				    struct cell_key *key)
+{
+	struct cell *cell;
+	struct hlist_node *tmp;
+
+	hlist_for_each_entry(cell, tmp, bucket, list)
+		if (keys_equal(&cell->key, key))
+			return cell;
+
+	return NULL;
+}
+
+/*
+ * This may block if a new cell needs allocating.  You must ensure that
+ * cells will be unlocked even if the calling thread is blocked.
+ *
+ * Returns the number of entries in the cell prior to the new addition
+ * or < 0 on failure.
+ */
+static int bio_detain(struct bio_prison *prison, struct cell_key *key,
+		      struct bio *inmate, struct cell **ref)
+{
+	int r;
+	unsigned long flags;
+	uint32_t hash = hash_key(prison, key);
+	struct cell *uninitialized_var(cell), *cell2 = NULL;
+
+	BUG_ON(hash > prison->nr_buckets);
+
+	spin_lock_irqsave(&prison->lock, flags);
+	cell = __search_bucket(prison->cells + hash, key);
+
+	if (!cell) {
+		/*
+		 * Allocate a new cell
+		 */
+		spin_unlock_irqrestore(&prison->lock, flags);
+		cell2 = mempool_alloc(prison->cell_pool, GFP_NOIO);
+		spin_lock_irqsave(&prison->lock, flags);
+
+		/*
+		 * We've been unlocked, so we have to double check that
+		 * nobody else has inserted this cell in the meantime.
+		 */
+		cell = __search_bucket(prison->cells + hash, key);
+
+		if (!cell) {
+			cell = cell2;
+			cell2 = NULL;
+
+			cell->prison = prison;
+			memcpy(&cell->key, key, sizeof(cell->key));
+			cell->count = 0;
+			bio_list_init(&cell->bios);
+			hlist_add_head(&cell->list, prison->cells + hash);
+		}
+	}
+
+	r = cell->count++;
+	bio_list_add(&cell->bios, inmate);
+	spin_unlock_irqrestore(&prison->lock, flags);
+
+	if (cell2)
+		mempool_free(cell2, prison->cell_pool);
+
+	*ref = cell;
+
+	return r;
+}
+
+/*
+ * @inmates must have been initialised prior to this call
+ */
+static void __cell_release(struct cell *cell, struct bio_list *inmates)
+{
+	struct bio_prison *prison = cell->prison;
+
+	hlist_del(&cell->list);
+
+	if (inmates)
+		bio_list_merge(inmates, &cell->bios);
+
+	mempool_free(cell, prison->cell_pool);
+}
+
+static void cell_release(struct cell *cell, struct bio_list *bios)
+{
+	unsigned long flags;
+	struct bio_prison *prison = cell->prison;
+
+	spin_lock_irqsave(&prison->lock, flags);
+	__cell_release(cell, bios);
+	spin_unlock_irqrestore(&prison->lock, flags);
+}
+
+/*
+ * There are a couple of places where we put a bio into a cell briefly
+ * before taking it out again.  In these situations we know that no other
+ * bio may be in the cell.  This function releases the cell, and also does
+ * a sanity check.
+ */
+static void cell_release_singleton(struct cell *cell, struct bio *bio)
+{
+	struct bio_prison *prison = cell->prison;
+	struct bio_list bios;
+	struct bio *b;
+	unsigned long flags;
+
+	bio_list_init(&bios);
+
+	spin_lock_irqsave(&prison->lock, flags);
+	__cell_release(cell, &bios);
+	spin_unlock_irqrestore(&prison->lock, flags);
+
+	b = bio_list_pop(&bios);
+	BUG_ON(b != bio);
+	BUG_ON(!bio_list_empty(&bios));
+}
+
+static void cell_error(struct cell *cell)
+{
+	struct bio_prison *prison = cell->prison;
+	struct bio_list bios;
+	struct bio *bio;
+	unsigned long flags;
+
+	bio_list_init(&bios);
+
+	spin_lock_irqsave(&prison->lock, flags);
+	__cell_release(cell, &bios);
+	spin_unlock_irqrestore(&prison->lock, flags);
+
+	while ((bio = bio_list_pop(&bios)))
+		bio_io_error(bio);
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * We use the deferred set to keep track of pending reads to shared blocks.
+ * We do this to ensure the new mapping caused by a write isn't performed
+ * until these prior reads have completed.  Otherwise the insertion of the
+ * new mapping could free the old block that the read bios are mapped to.
+ */
+
+struct deferred_set;
+struct deferred_entry {
+	struct deferred_set *ds;
+	unsigned count;
+	struct list_head work_items;
+};
+
+struct deferred_set {
+	spinlock_t lock;
+	unsigned current_entry;
+	unsigned sweeper;
+	struct deferred_entry entries[DEFERRED_SET_SIZE];
+};
+
+static void ds_init(struct deferred_set *ds)
+{
+	int i;
+
+	spin_lock_init(&ds->lock);
+	ds->current_entry = 0;
+	ds->sweeper = 0;
+	for (i = 0; i < DEFERRED_SET_SIZE; i++) {
+		ds->entries[i].ds = ds;
+		ds->entries[i].count = 0;
+		INIT_LIST_HEAD(&ds->entries[i].work_items);
+	}
+}
+
+static struct deferred_entry *ds_inc(struct deferred_set *ds)
+{
+	unsigned long flags;
+	struct deferred_entry *entry;
+
+	spin_lock_irqsave(&ds->lock, flags);
+	entry = ds->entries + ds->current_entry;
+	entry->count++;
+	spin_unlock_irqrestore(&ds->lock, flags);
+
+	return entry;
+}
+
+static unsigned ds_next(unsigned index)
+{
+	return (index + 1) % DEFERRED_SET_SIZE;
+}
+
+static void __sweep(struct deferred_set *ds, struct list_head *head)
+{
+	while ((ds->sweeper != ds->current_entry) &&
+	       !ds->entries[ds->sweeper].count) {
+		list_splice_init(&ds->entries[ds->sweeper].work_items, head);
+		ds->sweeper = ds_next(ds->sweeper);
+	}
+
+	if ((ds->sweeper == ds->current_entry) && !ds->entries[ds->sweeper].count)
+		list_splice_init(&ds->entries[ds->sweeper].work_items, head);
+}
+
+static void ds_dec(struct deferred_entry *entry, struct list_head *head)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&entry->ds->lock, flags);
+	BUG_ON(!entry->count);
+	--entry->count;
+	__sweep(entry->ds, head);
+	spin_unlock_irqrestore(&entry->ds->lock, flags);
+}
+
+/*
+ * Returns 1 if deferred or 0 if no pending items to delay job.
+ */
+static int ds_add_work(struct deferred_set *ds, struct list_head *work)
+{
+	int r = 1;
+	unsigned long flags;
+	unsigned next_entry;
+
+	spin_lock_irqsave(&ds->lock, flags);
+	if ((ds->sweeper == ds->current_entry) &&
+	    !ds->entries[ds->current_entry].count)
+		r = 0;
+	else {
+		list_add(work, &ds->entries[ds->current_entry].work_items);
+		next_entry = ds_next(ds->current_entry);
+		if (!ds->entries[next_entry].count)
+			ds->current_entry = next_entry;
+	}
+	spin_unlock_irqrestore(&ds->lock, flags);
+
+	return r;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Key building.
+ */
+static void build_data_key(struct dm_thin_device *td,
+			   dm_block_t b, struct cell_key *key)
+{
+	key->virtual = 0;
+	key->dev = dm_thin_dev_id(td);
+	key->block = b;
+}
+
+static void build_virtual_key(struct dm_thin_device *td, dm_block_t b,
+			      struct cell_key *key)
+{
+	key->virtual = 1;
+	key->dev = dm_thin_dev_id(td);
+	key->block = b;
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * A pool device ties together a metadata device and a data device.  It
+ * also provides the interface for creating and destroying internal
+ * devices.
+ */
+struct new_mapping;
+struct pool {
+	struct list_head list;
+	struct dm_target *ti;	/* Only set if a pool target is bound */
+
+	struct mapped_device *pool_md;
+	struct block_device *md_dev;
+	struct dm_pool_metadata *pmd;
+
+	uint32_t sectors_per_block;
+	unsigned block_shift;
+	dm_block_t offset_mask;
+	dm_block_t low_water_blocks;
+
+	unsigned zero_new_blocks:1;
+	unsigned low_water_triggered:1;	/* A dm event has been sent */
+	unsigned no_free_space:1;	/* A -ENOSPC warning has been issued */
+
+	struct bio_prison *prison;
+	struct dm_kcopyd_client *copier;
+
+	struct workqueue_struct *wq;
+	struct work_struct worker;
+
+	unsigned ref_count;
+
+	spinlock_t lock;
+	struct bio_list deferred_bios;
+	struct bio_list deferred_flush_bios;
+	struct list_head prepared_mappings;
+
+	struct bio_list retry_on_resume_list;
+
+	struct deferred_set ds;	/* FIXME: move to thin_c */
+
+	struct new_mapping *next_mapping;
+	mempool_t *mapping_pool;
+	mempool_t *endio_hook_pool;
+};
+
+/*
+ * Target context for a pool.
+ */
+struct pool_c {
+	struct dm_target *ti;
+	struct pool *pool;
+	struct dm_dev *data_dev;
+	struct dm_dev *metadata_dev;
+	struct dm_target_callbacks callbacks;
+
+	dm_block_t low_water_blocks;
+	unsigned zero_new_blocks:1;
+};
+
+/*
+ * Target context for a thin.
+ */
+struct thin_c {
+	struct dm_dev *pool_dev;
+	dm_thin_id dev_id;
+
+	struct pool *pool;
+	struct dm_thin_device *td;
+};
+
+/*----------------------------------------------------------------*/
+
+/*
+ * A global list of pools that uses a struct mapped_device as a key.
+ */
+static struct dm_thin_pool_table {
+	struct mutex mutex;
+	struct list_head pools;
+} dm_thin_pool_table;
+
+static void pool_table_init(void)
+{
+	mutex_init(&dm_thin_pool_table.mutex);
+	INIT_LIST_HEAD(&dm_thin_pool_table.pools);
+}
+
+static void __pool_table_insert(struct pool *pool)
+{
+	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
+	list_add(&pool->list, &dm_thin_pool_table.pools);
+}
+
+static void __pool_table_remove(struct pool *pool)
+{
+	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
+	list_del(&pool->list);
+}
+
+static struct pool *__pool_table_lookup(struct mapped_device *md)
+{
+	struct pool *pool = NULL, *tmp;
+
+	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
+
+	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
+		if (tmp->pool_md == md) {
+			pool = tmp;
+			break;
+		}
+	}
+
+	return pool;
+}
+
+static struct pool *__pool_table_lookup_metadata_dev(struct block_device *md_dev)
+{
+	struct pool *pool = NULL, *tmp;
+
+	BUG_ON(!mutex_is_locked(&dm_thin_pool_table.mutex));
+
+	list_for_each_entry(tmp, &dm_thin_pool_table.pools, list) {
+		if (tmp->md_dev == md_dev) {
+			pool = tmp;
+			break;
+		}
+	}
+
+	return pool;
+}
+
+/*----------------------------------------------------------------*/
+
+static void __requeue_bio_list(struct thin_c *tc, struct bio_list *master)
+{
+	struct bio *bio;
+	struct bio_list bios;
+
+	bio_list_init(&bios);
+	bio_list_merge(&bios, master);
+	bio_list_init(master);
+
+	while ((bio = bio_list_pop(&bios))) {
+		if (dm_get_mapinfo(bio)->ptr == tc)
+			bio_endio(bio, DM_ENDIO_REQUEUE);
+		else
+			bio_list_add(master, bio);
+	}
+}
+
+static void requeue_io(struct thin_c *tc)
+{
+	struct pool *pool = tc->pool;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pool->lock, flags);
+	__requeue_bio_list(tc, &pool->deferred_bios);
+	__requeue_bio_list(tc, &pool->retry_on_resume_list);
+	spin_unlock_irqrestore(&pool->lock, flags);
+}
+
+/*
+ * This section of code contains the logic for processing a thin device's IO.
+ * Much of the code depends on pool object resources (lists, workqueues, etc)
+ * but most is exclusively called from the thin target rather than the thin-pool
+ * target.
+ */
+
+static dm_block_t get_bio_block(struct thin_c *tc, struct bio *bio)
+{
+	return bio->bi_sector >> tc->pool->block_shift;
+}
+
+static void remap(struct thin_c *tc, struct bio *bio, dm_block_t block)
+{
+	struct pool *pool = tc->pool;
+
+	bio->bi_bdev = tc->pool_dev->bdev;
+	bio->bi_sector = (block << pool->block_shift) +
+		(bio->bi_sector & pool->offset_mask);
+}
+
+static void remap_and_issue(struct thin_c *tc, struct bio *bio,
+			    dm_block_t block)
+{
+	struct pool *pool = tc->pool;
+	unsigned long flags;
+
+	remap(tc, bio, block);
+
+	/*
+	 * Batch together any FUA/FLUSH bios we find and then issue
+	 * a single commit for them in process_deferred_bios().
+	 */
+	if (bio->bi_rw & (REQ_FLUSH | REQ_FUA)) {
+		spin_lock_irqsave(&pool->lock, flags);
+		bio_list_add(&pool->deferred_flush_bios, bio);
+		spin_unlock_irqrestore(&pool->lock, flags);
+	} else
+		generic_make_request(bio);
+}
+
+/*
+ * wake_worker() is used when new work is queued and when pool_resume is
+ * ready to continue deferred IO processing.
+ */
+static void wake_worker(struct pool *pool)
+{
+	queue_work(pool->wq, &pool->worker);
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Bio endio functions.
+ */
+struct endio_hook {
+	struct thin_c *tc;
+	bio_end_io_t *saved_bi_end_io;
+	struct deferred_entry *entry;
+};
+
+struct new_mapping {
+	struct list_head list;
+
+	int prepared;
+
+	struct thin_c *tc;
+	dm_block_t virt_block;
+	dm_block_t data_block;
+	struct cell *cell;
+	int err;
+
+	/*
+	 * If the bio covers the whole area of a block then we can avoid
+	 * zeroing or copying.  Instead this bio is hooked.  The bio will
+	 * still be in the cell, so care has to be taken to avoid issuing
+	 * the bio twice.
+	 */
+	struct bio *bio;
+	bio_end_io_t *saved_bi_end_io;
+};
+
+static void __maybe_add_mapping(struct new_mapping *m)
+{
+	struct pool *pool = m->tc->pool;
+
+	if (list_empty(&m->list) && m->prepared) {
+		list_add(&m->list, &pool->prepared_mappings);
+		wake_worker(pool);
+	}
+}
+
+static void copy_complete(int read_err, unsigned long write_err, void *context)
+{
+	unsigned long flags;
+	struct new_mapping *m = context;
+	struct pool *pool = m->tc->pool;
+
+	m->err = read_err || write_err ? -EIO : 0;
+
+	spin_lock_irqsave(&pool->lock, flags);
+	m->prepared = 1;
+	__maybe_add_mapping(m);
+	spin_unlock_irqrestore(&pool->lock, flags);
+}
+
+static void overwrite_endio(struct bio *bio, int err)
+{
+	unsigned long flags;
+	struct new_mapping *m = dm_get_mapinfo(bio)->ptr;
+	struct pool *pool = m->tc->pool;
+
+	m->err = err;
+
+	spin_lock_irqsave(&pool->lock, flags);
+	m->prepared = 1;
+	__maybe_add_mapping(m);
+	spin_unlock_irqrestore(&pool->lock, flags);
+}
+
+static void shared_read_endio(struct bio *bio, int err)
+{
+	struct list_head mappings;
+	struct new_mapping *m, *tmp;
+	struct endio_hook *h = dm_get_mapinfo(bio)->ptr;
+	unsigned long flags;
+	struct pool *pool = h->tc->pool;
+
+	bio->bi_end_io = h->saved_bi_end_io;
+	bio_endio(bio, err);
+
+	INIT_LIST_HEAD(&mappings);
+	ds_dec(h->entry, &mappings);
+
+	spin_lock_irqsave(&pool->lock, flags);
+	list_for_each_entry_safe(m, tmp, &mappings, list) {
+		list_del(&m->list);
+		INIT_LIST_HEAD(&m->list);
+		__maybe_add_mapping(m);
+	}
+	spin_unlock_irqrestore(&pool->lock, flags);
+
+	mempool_free(h, pool->endio_hook_pool);
+}
+
+/*----------------------------------------------------------------*/
+
+/*
+ * Workqueue.
+ */
+
+/*
+ * Prepared mapping jobs.
+ */
+
+/*
+ * This sends the bios in the cell back to the deferred_bios list.
+ */
+static void cell_defer(struct thin_c *tc, struct cell *cell,
+		       dm_block_t data_block)
+{
+	struct pool *pool = tc->pool;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pool->lock, flags);
+	cell_release(cell, &pool->deferred_bios);
+	spin_unlock_irqrestore(&tc->pool->lock, flags);
+
+	wake_worker(pool);
+}
+
+/*
+ * Same as cell_defer above, except it omits one particular detainee,
+ * a write bio that covers the block and has already been processed.
+ */
+static void cell_defer_except(struct thin_c *tc, struct cell *cell,
+			      struct bio *exception)
+{
+	struct bio_list bios;
+	struct bio *bio;
+	struct pool *pool = tc->pool;
+	unsigned long flags;
+
+	bio_list_init(&bios);
+	cell_release(cell, &bios);
+
+	spin_lock_irqsave(&pool->lock, flags);
+	while ((bio = bio_list_pop(&bios)))
+		if (bio != exception)
+			bio_list_add(&pool->deferred_bios, bio);
+	spin_unlock_irqrestore(&pool->lock, flags);
+
+	wake_worker(pool);
+}
+
+static void process_prepared_mapping(struct new_mapping *m)
+{
+	struct thin_c *tc = m->tc;
+	struct bio *bio;
+	int r;
+
+	bio = m->bio;
+	if (bio)
+		bio->bi_end_io = m->saved_bi_end_io;
+
+	if (m->err) {
+		cell_error(m->cell);
+		return;
+	}
+
+	/*
+	 * Commit the prepared block into the mapping btree.
+	 * Any I/O for this block arriving after this point will get
+	 * remapped to it directly.
+	 */
+	r = dm_thin_insert_block(tc->td, m->virt_block, m->data_block);
+	if (r) {
+		DMERR("dm_thin_insert_block() failed");
+		cell_error(m->cell);
+		return;
+	}
+
+	/*
+	 * Release any bios held while the block was being provisioned.
+	 * If we are processing a write bio that completely covers the block,
+	 * we already processed it so can ignore it now when processing
+	 * the bios in the cell.
+	 */
+	if (bio) {
+		cell_defer_except(tc, m->cell, bio);
+		bio_endio(bio, 0);
+	} else
+		cell_defer(tc, m->cell, m->data_block);
+
+	list_del(&m->list);
+	mempool_free(m, tc->pool->mapping_pool);
+}
+
+static void process_prepared_mappings(struct pool *pool)
+{
+	unsigned long flags;
+	struct list_head maps;
+	struct new_mapping *m, *tmp;
+
+	INIT_LIST_HEAD(&maps);
+	spin_lock_irqsave(&pool->lock, flags);
+	list_splice_init(&pool->prepared_mappings, &maps);
+	spin_unlock_irqrestore(&pool->lock, flags);
+
+	list_for_each_entry_safe(m, tmp, &maps, list)
+		process_prepared_mapping(m);
+}
+
+/*
+ * Deferred bio jobs.
+ */
+static int io_overwrites_block(struct pool *pool, struct bio *bio)
+{
+	return ((bio_data_dir(bio) == WRITE) &&
+		!(bio->bi_sector & pool->offset_mask)) &&
+		(bio->bi_size == (pool->sectors_per_block << SECTOR_SHIFT));
+}
+
+static void save_and_set_endio(struct bio *bio, bio_end_io_t **save,
+			       bio_end_io_t *fn)
+{
+	*save = bio->bi_end_io;
+	bio->bi_end_io = fn;
+}
+
+static int ensure_next_mapping(struct pool *pool)
+{
+	if (pool->next_mapping)
+		return 0;
+
+	pool->next_mapping = mempool_alloc(pool->mapping_pool, GFP_ATOMIC);
+
+	return pool->next_mapping ? 0 : -ENOMEM;
+}
+
+static struct new_mapping *get_next_mapping(struct pool *pool)
+{
+	struct new_mapping *r = pool->next_mapping;
+
+	BUG_ON(!pool->next_mapping);
+
+	pool->next_mapping = NULL;
+
+	return r;
+}
+
+static void schedule_copy(struct thin_c *tc, dm_block_t virt_block,
+			  dm_block_t data_origin, dm_block_t data_dest,
+			  struct cell *cell, struct bio *bio)
+{
+	int r;
+	struct pool *pool = tc->pool;
+	struct new_mapping *m = get_next_mapping(pool);
+
+	INIT_LIST_HEAD(&m->list);
+	m->prepared = 0;
+	m->tc = tc;
+	m->virt_block = virt_block;
+	m->data_block = data_dest;
+	m->cell = cell;
+	m->err = 0;
+	m->bio = NULL;
+
+	ds_add_work(&pool->ds, &m->list);
+
+	/*
+	 * IO to pool_dev remaps to the pool target's data_dev.
+	 *
+	 * If the whole block of data is being overwritten, we can issue the
+	 * bio immediately. Otherwise we use kcopyd to clone the data first.
+	 */
+	if (io_overwrites_block(pool, bio)) {
+		m->bio = bio;
+		save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
+		dm_get_mapinfo(bio)->ptr = m;
+		remap_and_issue(tc, bio, data_dest);
+	} else {
+		struct dm_io_region from, to;
+
+		from.bdev = tc->pool_dev->bdev;
+		from.sector = data_origin * pool->sectors_per_block;
+		from.count = pool->sectors_per_block;
+
+		to.bdev = tc->pool_dev->bdev;
+		to.sector = data_dest * pool->sectors_per_block;
+		to.count = pool->sectors_per_block;
+
+		r = dm_kcopyd_copy(pool->copier, &from, 1, &to,
+				   0, copy_complete, m);
+		if (r < 0) {
+			mempool_free(m, pool->mapping_pool);
+			DMERR("dm_kcopyd_copy() failed");
+			cell_error(cell);
+		}
+	}
+}
+
+static void schedule_zero(struct thin_c *tc, dm_block_t virt_block,
+			  dm_block_t data_block, struct cell *cell,
+			  struct bio *bio)
+{
+	struct pool *pool = tc->pool;
+	struct new_mapping *m = get_next_mapping(pool);
+
+	INIT_LIST_HEAD(&m->list);
+	m->prepared = 0;
+	m->tc = tc;
+	m->virt_block = virt_block;
+	m->data_block = data_block;
+	m->cell = cell;
+	m->err = 0;
+	m->bio = NULL;
+
+	/*
+	 * If the whole block of data is being overwritten or we are not
+	 * zeroing pre-existing data, we can issue the bio immediately.
+	 * Otherwise we use kcopyd to zero the data first.
+	 */
+	if (!pool->zero_new_blocks)
+		process_prepared_mapping(m);
+
+	else if (io_overwrites_block(pool, bio)) {
+		m->bio = bio;
+		save_and_set_endio(bio, &m->saved_bi_end_io, overwrite_endio);
+		dm_get_mapinfo(bio)->ptr = m;
+		remap_and_issue(tc, bio, data_block);
+
+	} else {
+		int r;
+		struct dm_io_region to;
+
+		to.bdev = tc->pool_dev->bdev;
+		to.sector = data_block * pool->sectors_per_block;
+		to.count = pool->sectors_per_block;
+
+		r = dm_kcopyd_zero(pool->copier, 1, &to, 0, copy_complete, m);
+		if (r < 0) {
+			mempool_free(m, pool->mapping_pool);
+			DMERR("dm_kcopyd_zero() failed");
+			cell_error(cell);
+		}
+	}
+}
+
+static int alloc_data_block(struct thin_c *tc, dm_block_t *result)
+{
+	int r;
+	dm_block_t free_blocks;
+	unsigned long flags;
+	struct pool *pool = tc->pool;
+
+	r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
+	if (r)
+		return r;
+
+	if (free_blocks <= pool->low_water_blocks && !pool->low_water_triggered) {
+		DMWARN("%s: reached low water mark, sending event.",
+		       dm_device_name(pool->pool_md));
+		spin_lock_irqsave(&pool->lock, flags);
+		pool->low_water_triggered = 1;
+		spin_unlock_irqrestore(&pool->lock, flags);
+		dm_table_event(pool->ti->table);
+	}
+
+	if (!free_blocks) {
+		if (pool->no_free_space)
+			return -ENOSPC;
+		else {
+			/*
+			 * Try to commit to see if that will free up some
+			 * more space.
+			 */
+			r = dm_pool_commit_metadata(pool->pmd);
+			if (r) {
+				DMERR("%s: dm_pool_commit_metadata() failed, error = %d",
+				      __func__, r);
+				return r;
+			}
+
+			r = dm_pool_get_free_block_count(pool->pmd, &free_blocks);
+			if (r)
+				return r;
+
+			/*
+			 * If we still have no space we set a flag to avoid
+			 * doing all this checking and return -ENOSPC.
+			 */
+			if (!free_blocks) {
+				DMWARN("%s: no free space available.",
+				       dm_device_name(pool->pool_md));
+				spin_lock_irqsave(&pool->lock, flags);
+				pool->no_free_space = 1;
+				spin_unlock_irqrestore(&pool->lock, flags);
+				return -ENOSPC;
+			}
+		}
+	}
+
+	r = dm_pool_alloc_data_block(pool->pmd, result);
+	if (r)
+		return r;
+
+	return 0;
+}
+
+/*
+ * If we have run out of space, queue bios until the device is
+ * resumed, presumably after having been reloaded with more space.
+ */
+static void retry_on_resume(struct bio *bio)
+{
+	struct thin_c *tc = dm_get_mapinfo(bio)->ptr;
+	struct pool *pool = tc->pool;
+	unsigned long flags;
+
+	spin_lock_irqsave(&pool->lock, flags);
+	bio_list_add(&pool->retry_on_resume_list, bio);
+	spin_unlock_irqrestore(&pool->lock, flags);
+}
+
+static void no_space(struct cell *cell)
+{
+	struct bio *bio;
+	struct bio_list bios;
+
+	bio_list_init(&bios);
+	cell_release(cell, &bios);
+
+	while ((bio = bio_list_pop(&bios)))
+		retry_on_resume(bio);
+}
+
+static void break_sharing(struct thin_c *tc, struct bio *bio, dm_block_t block,
+			  struct cell_key *key,
+			  struct dm_thin_lookup_result *lookup_result,
+			  struct cell *cell)
+{
+	int r;
+	dm_block_t data_block;
+
+	r = alloc_data_block(tc, &data_block);
+	switch (r) {
+	case 0:
+		schedule_copy(tc, block, lookup_result->block,
+			      data_block, cell, bio);
+		break;
+
+	case -ENOSPC:
+		no_space(cell);
+		break;
+
+	default:
+		DMERR("%s: alloc_data_block() failed, error = %d", __func__, r);
+		cell_error(cell);
+		break;
+	}
+}
+
+static void process_shared_bio(struct thin_c *tc, struct bio *bio,
+			       dm_block_t block,
+			       struct dm_thin_lookup_result *lookup_result)
+{
+	struct cell *cell;
+	struct pool *pool = tc->pool;
+	struct cell_key key;
+
+	/*
+	 * If cell is already occupied, then sharing is already in the process
+	 * of being broken so we have nothing further to do here.
+	 */
+	build_data_key(tc->td, lookup_result->block, &key);
+	if (bio_detain(pool->prison, &key, bio, &cell))
+		return;
+
+	if (bio_data_dir(bio) == WRITE)
+		break_sharing(tc, bio, block, &key, lookup_result, cell);
+	else {
+		struct endio_hook *h;
+		h = mempool_alloc(pool->endio_hook_pool, GFP_NOIO);
+
+		h->tc = tc;
+		h->entry = ds_inc(&pool->ds);
+		save_and_set_endio(bio, &h->saved_bi_end_io, shared_read_endio);
+		dm_get_mapinfo(bio)->ptr = h;
+
+		cell_release_singleton(cell, bio);
+		remap_and_issue(tc, bio, lookup_result->block);
+	}
+}
+
+static void provision_block(struct thin_c *tc, struct bio *bio, dm_block_t block,
+			    struct cell *cell)
+{
+	int r;
+	dm_block_t data_block;
+
+	/*
+	 * Remap empty bios (flushes) immediately, without provisioning.
+	 */
+	if (!bio->bi_size) {
+		cell_release_singleton(cell, bio);
+		remap_and_issue(tc, bio, 0);
+		return;