Merge git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-2.6-dm

* git://git.kernel.org/pub/scm/linux/kernel/git/agk/linux-2.6-dm: (80 commits)
  dm snapshot: use merge origin if snapshot invalid
  dm snapshot: report merge failure in status
  dm snapshot: merge consecutive chunks together
  dm snapshot: trigger exceptions in remaining snapshots during merge
  dm snapshot: delay merging a chunk until writes to it complete
  dm snapshot: queue writes to chunks being merged
  dm snapshot: add merging
  dm snapshot: permit only one merge at once
  dm snapshot: support barriers in snapshot merge target
  dm snapshot: avoid allocating exceptions in merge
  dm snapshot: rework writing to origin
  dm snapshot: add merge target
  dm exception store: add merge specific methods
  dm snapshot: create function for chunk_is_tracked wait
  dm snapshot: make bio optional in __origin_write
  dm mpath: reject messages when device is suspended
  dm: export suspended state to targets
  dm: rename dm_suspended to dm_suspended_md
  dm: swap target postsuspend call and setting suspended flag
  dm crypt: add plain64 iv
  ...

Documentation/device-mapper/snapshot.txt

@@ -8,13 +8,19 @@ the block device which are also writable without interfering with the
 original content;
 *) To create device "forks", i.e. multiple different versions of the
 same data stream.
+*) To merge a snapshot of a block device back into the snapshot's origin
+device.
 
+In the first two cases, dm copies only the chunks of data that get
+changed and uses a separate copy-on-write (COW) block device for
+storage.
 
-In both cases, dm copies only the chunks of data that get changed and
-uses a separate copy-on-write (COW) block device for storage.
+For snapshot merge the contents of the COW storage are merged back into
+the origin device.
 
 
-There are two dm targets available: snapshot and snapshot-origin.
+There are three dm targets available:
+snapshot, snapshot-origin, and snapshot-merge.
 
 *) snapshot-origin <origin>
 
@@ -40,8 +46,25 @@ The difference is that for transient snapshots less metadata must be
 saved on disk - they can be kept in memory by the kernel.
 
 
-How this is used by LVM2
-========================
+* snapshot-merge <origin> <COW device> <persistent> <chunksize>
+
+takes the same table arguments as the snapshot target except it only
+works with persistent snapshots.  This target assumes the role of the
+"snapshot-origin" target and must not be loaded if the "snapshot-origin"
+is still present for <origin>.
+
+Creates a merging snapshot that takes control of the changed chunks
+stored in the <COW device> of an existing snapshot, through a handover
+procedure, and merges these chunks back into the <origin>.  Once merging
+has started (in the background) the <origin> may be opened and the merge
+will continue while I/O is flowing to it.  Changes to the <origin> are
+deferred until the merging snapshot's corresponding chunk(s) have been
+merged.  Once merging has started the snapshot device, associated with
+the "snapshot" target, will return -EIO when accessed.
+
+
+How snapshot is used by LVM2
+============================
 When you create the first LVM2 snapshot of a volume, four dm devices are used:
 
 1) a device containing the original mapping table of the source volume;
@@ -72,3 +95,30 @@ brw-------  1 root root 254, 12 29 ago 18:15 /dev/mapper/volumeGroup-snap-cow
 brw-------  1 root root 254, 13 29 ago 18:15 /dev/mapper/volumeGroup-snap
 brw-------  1 root root 254, 10 29 ago 18:14 /dev/mapper/volumeGroup-base
 
+
+How snapshot-merge is used by LVM2
+==================================
+A merging snapshot assumes the role of the "snapshot-origin" while
+merging.  As such the "snapshot-origin" is replaced with
+"snapshot-merge".  The "-real" device is not changed and the "-cow"
+device is renamed to <origin name>-cow to aid LVM2's cleanup of the
+merging snapshot after it completes.  The "snapshot" that hands over its
+COW device to the "snapshot-merge" is deactivated (unless using lvchange
+--refresh); but if it is left active it will simply return I/O errors.
+
+A snapshot will merge into its origin with the following command:
+
+lvconvert --merge volumeGroup/snap
+
+we'll now have this situation:
+
+# dmsetup table|grep volumeGroup
+
+volumeGroup-base-real: 0 2097152 linear 8:19 384
+volumeGroup-base-cow: 0 204800 linear 8:19 2097536
+volumeGroup-base: 0 2097152 snapshot-merge 254:11 254:12 P 16
+
+# ls -lL /dev/mapper/volumeGroup-*
+brw-------  1 root root 254, 11 29 ago 18:15 /dev/mapper/volumeGroup-base-real
+brw-------  1 root root 254, 12 29 ago 18:16 /dev/mapper/volumeGroup-base-cow
+brw-------  1 root root 254, 10 29 ago 18:16 /dev/mapper/volumeGroup-base

drivers/md/dm-crypt.c

 /*
  * Copyright (C) 2003 Christophe Saout <christophe@saout.de>
  * Copyright (C) 2004 Clemens Fruhwirth <clemens@endorphin.org>
- * Copyright (C) 2006-2008 Red Hat, Inc. All rights reserved.
+ * Copyright (C) 2006-2009 Red Hat, Inc. All rights reserved.
  *
  * This file is released under the GPL.
  */
@@ -71,10 +71,21 @@ struct crypt_iv_operations {
 	int (*ctr)(struct crypt_config *cc, struct dm_target *ti,
 		   const char *opts);
 	void (*dtr)(struct crypt_config *cc);
-	const char *(*status)(struct crypt_config *cc);
+	int (*init)(struct crypt_config *cc);
+	int (*wipe)(struct crypt_config *cc);
 	int (*generator)(struct crypt_config *cc, u8 *iv, sector_t sector);
 };
 
+struct iv_essiv_private {
+	struct crypto_cipher *tfm;
+	struct crypto_hash *hash_tfm;
+	u8 *salt;
+};
+
+struct iv_benbi_private {
+	int shift;
+};
+
 /*
  * Crypt: maps a linear range of a block device
  * and encrypts / decrypts at the same time.
@@ -102,8 +113,8 @@ struct crypt_config {
 	struct crypt_iv_operations *iv_gen_ops;
 	char *iv_mode;
 	union {
-		struct crypto_cipher *essiv_tfm;
-		int benbi_shift;
+		struct iv_essiv_private essiv;
+		struct iv_benbi_private benbi;
 	} iv_gen_private;
 	sector_t iv_offset;
 	unsigned int iv_size;
@@ -147,6 +158,9 @@ static void kcryptd_queue_crypt(struct dm_crypt_io *io);
  * plain: the initial vector is the 32-bit little-endian version of the sector
  *        number, padded with zeros if necessary.
  *
+ * plain64: the initial vector is the 64-bit little-endian version of the sector
+ *        number, padded with zeros if necessary.
+ *
  * essiv: "encrypted sector|salt initial vector", the sector number is
  *        encrypted with the bulk cipher using a salt as key. The salt
  *        should be derived from the bulk cipher's key via hashing.
@@ -169,88 +183,123 @@ static int crypt_iv_plain_gen(struct crypt_config *cc, u8 *iv, sector_t sector)
 	return 0;
 }
 
-static int crypt_iv_essiv_ctr(struct crypt_config *cc, struct dm_target *ti,
-			      const char *opts)
+static int crypt_iv_plain64_gen(struct crypt_config *cc, u8 *iv,
+				sector_t sector)
 {
-	struct crypto_cipher *essiv_tfm;
-	struct crypto_hash *hash_tfm;
+	memset(iv, 0, cc->iv_size);
+	*(u64 *)iv = cpu_to_le64(sector);
+
+	return 0;
+}
+
+/* Initialise ESSIV - compute salt but no local memory allocations */
+static int crypt_iv_essiv_init(struct crypt_config *cc)
+{
+	struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
 	struct hash_desc desc;
 	struct scatterlist sg;
-	unsigned int saltsize;
-	u8 *salt;
 	int err;
 
-	if (opts == NULL) {
+	sg_init_one(&sg, cc->key, cc->key_size);
+	desc.tfm = essiv->hash_tfm;
+	desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;
+
+	err = crypto_hash_digest(&desc, &sg, cc->key_size, essiv->salt);
+	if (err)
+		return err;
+
+	return crypto_cipher_setkey(essiv->tfm, essiv->salt,
+				    crypto_hash_digestsize(essiv->hash_tfm));
+}
+
+/* Wipe salt and reset key derived from volume key */
+static int crypt_iv_essiv_wipe(struct crypt_config *cc)
+{
+	struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
+	unsigned salt_size = crypto_hash_digestsize(essiv->hash_tfm);
+
+	memset(essiv->salt, 0, salt_size);
+
+	return crypto_cipher_setkey(essiv->tfm, essiv->salt, salt_size);
+}
+
+static void crypt_iv_essiv_dtr(struct crypt_config *cc)
+{
+	struct iv_essiv_private *essiv = &cc->iv_gen_private.essiv;
+
+	crypto_free_cipher(essiv->tfm);
+	essiv->tfm = NULL;
+
+	crypto_free_hash(essiv->hash_tfm);
+	essiv->hash_tfm = NULL;
+
+	kzfree(essiv->salt);
+	essiv->salt = NULL;
+}
+
+static int crypt_iv_essiv_ctr(struct crypt_config *cc, struct dm_target *ti,
+			      const char *opts)
+{
+	struct crypto_cipher *essiv_tfm = NULL;
+	struct crypto_hash *hash_tfm = NULL;
+	u8 *salt = NULL;
+	int err;
+
+	if (!opts) {
 		ti->error = "Digest algorithm missing for ESSIV mode";
 		return -EINVAL;
 	}
 
-	/* Hash the cipher key with the given hash algorithm */
+	/* Allocate hash algorithm */
 	hash_tfm = crypto_alloc_hash(opts, 0, CRYPTO_ALG_ASYNC);
 	if (IS_ERR(hash_tfm)) {
 		ti->error = "Error initializing ESSIV hash";
-		return PTR_ERR(hash_tfm);
+		err = PTR_ERR(hash_tfm);
+		goto bad;
 	}
 
-	saltsize = crypto_hash_digestsize(hash_tfm);
-	salt = kmalloc(saltsize, GFP_KERNEL);
-	if (salt == NULL) {
+	salt = kzalloc(crypto_hash_digestsize(hash_tfm), GFP_KERNEL);
+	if (!salt) {
 		ti->error = "Error kmallocing salt storage in ESSIV";
-		crypto_free_hash(hash_tfm);
-		return -ENOMEM;
+		err = -ENOMEM;
+		goto bad;
 	}
 
-	sg_init_one(&sg, cc->key, cc->key_size);
-	desc.tfm = hash_tfm;
-	desc.flags = CRYPTO_TFM_REQ_MAY_SLEEP;
-	err = crypto_hash_digest(&desc, &sg, cc->key_size, salt);
-	crypto_free_hash(hash_tfm);
-
-	if (err) {
-		ti->error = "Error calculating hash in ESSIV";
-		kfree(salt);
-		return err;
-	}
-
-	/* Setup the essiv_tfm with the given salt */
+	/* Allocate essiv_tfm */
 	essiv_tfm = crypto_alloc_cipher(cc->cipher, 0, CRYPTO_ALG_ASYNC);
 	if (IS_ERR(essiv_tfm)) {
 		ti->error = "Error allocating crypto tfm for ESSIV";
-		kfree(salt);
-		return PTR_ERR(essiv_tfm);
+		err = PTR_ERR(essiv_tfm);
+		goto bad;
 	}
 	if (crypto_cipher_blocksize(essiv_tfm) !=
 	    crypto_ablkcipher_ivsize(cc->tfm)) {
 		ti->error = "Block size of ESSIV cipher does "
 			    "not match IV size of block cipher";
-		crypto_free_cipher(essiv_tfm);
-		kfree(salt);
-		return -EINVAL;
+		err = -EINVAL;
+		goto bad;
 	}
-	err = crypto_cipher_setkey(essiv_tfm, salt, saltsize);
-	if (err) {
-		ti->error = "Failed to set key for ESSIV cipher";
-		crypto_free_cipher(essiv_tfm);
-		kfree(salt);
-		return err;
-	}
-	kfree(salt);
 
-	cc->iv_gen_private.essiv_tfm = essiv_tfm;
+	cc->iv_gen_private.essiv.salt = salt;
+	cc->iv_gen_private.essiv.tfm = essiv_tfm;
+	cc->iv_gen_private.essiv.hash_tfm = hash_tfm;
+
 	return 0;
-}
 
-static void crypt_iv_essiv_dtr(struct crypt_config *cc)
-{
-	crypto_free_cipher(cc->iv_gen_private.essiv_tfm);
-	cc->iv_gen_private.essiv_tfm = NULL;
+bad:
+	if (essiv_tfm && !IS_ERR(essiv_tfm))
+		crypto_free_cipher(essiv_tfm);
+	if (hash_tfm && !IS_ERR(hash_tfm))
+		crypto_free_hash(hash_tfm);
+	kfree(salt);
+	return err;
 }
 
 static int crypt_iv_essiv_gen(struct crypt_config *cc, u8 *iv, sector_t sector)
 {
 	memset(iv, 0, cc->iv_size);
 	*(u64 *)iv = cpu_to_le64(sector);
-	crypto_cipher_encrypt_one(cc->iv_gen_private.essiv_tfm, iv, iv);
+	crypto_cipher_encrypt_one(cc->iv_gen_private.essiv.tfm, iv, iv);
 	return 0;
 }
 
@@ -273,7 +322,7 @@ static int crypt_iv_benbi_ctr(struct crypt_config *cc, struct dm_target *ti,
 		return -EINVAL;
 	}
 
-	cc->iv_gen_private.benbi_shift = 9 - log;
+	cc->iv_gen_private.benbi.shift = 9 - log;
 
 	return 0;
 }
@@ -288,7 +337,7 @@ static int crypt_iv_benbi_gen(struct crypt_config *cc, u8 *iv, sector_t sector)
 
 	memset(iv, 0, cc->iv_size - sizeof(u64)); /* rest is cleared below */
 
-	val = cpu_to_be64(((u64)sector << cc->iv_gen_private.benbi_shift) + 1);
+	val = cpu_to_be64(((u64)sector << cc->iv_gen_private.benbi.shift) + 1);
 	put_unaligned(val, (__be64 *)(iv + cc->iv_size - sizeof(u64)));
 
 	return 0;
@@ -305,9 +354,15 @@ static struct crypt_iv_operations crypt_iv_plain_ops = {
 	.generator = crypt_iv_plain_gen
 };
 
+static struct crypt_iv_operations crypt_iv_plain64_ops = {
+	.generator = crypt_iv_plain64_gen
+};
+
 static struct crypt_iv_operations crypt_iv_essiv_ops = {
 	.ctr       = crypt_iv_essiv_ctr,
 	.dtr       = crypt_iv_essiv_dtr,
+	.init      = crypt_iv_essiv_init,
+	.wipe      = crypt_iv_essiv_wipe,
 	.generator = crypt_iv_essiv_gen
 };
 
@@ -934,14 +989,14 @@ static int crypt_set_key(struct crypt_config *cc, char *key)
 
 	set_bit(DM_CRYPT_KEY_VALID, &cc->flags);
 
-	return 0;
+	return crypto_ablkcipher_setkey(cc->tfm, cc->key, cc->key_size);
 }
 
 static int crypt_wipe_key(struct crypt_config *cc)
 {
 	clear_bit(DM_CRYPT_KEY_VALID, &cc->flags);
 	memset(&cc->key, 0, cc->key_size * sizeof(u8));
-	return 0;
+	return crypto_ablkcipher_setkey(cc->tfm, cc->key, cc->key_size);
 }
 
 /*
@@ -983,11 +1038,6 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 		return -ENOMEM;
 	}
 
- 	if (crypt_set_key(cc, argv[1])) {
-		ti->error = "Error decoding key";
-		goto bad_cipher;
-	}
-
 	/* Compatibility mode for old dm-crypt cipher strings */
 	if (!chainmode || (strcmp(chainmode, "plain") == 0 && !ivmode)) {
 		chainmode = "cbc";
@@ -1015,6 +1065,11 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	strcpy(cc->chainmode, chainmode);
 	cc->tfm = tfm;
 
+	if (crypt_set_key(cc, argv[1]) < 0) {
+		ti->error = "Error decoding and setting key";
+		goto bad_ivmode;
+	}
+
 	/*
 	 * Choose ivmode. Valid modes: "plain", "essiv:<esshash>", "benbi".
 	 * See comments at iv code
@@ -1024,6 +1079,8 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 		cc->iv_gen_ops = NULL;
 	else if (strcmp(ivmode, "plain") == 0)
 		cc->iv_gen_ops = &crypt_iv_plain_ops;
+	else if (strcmp(ivmode, "plain64") == 0)
+		cc->iv_gen_ops = &crypt_iv_plain64_ops;
 	else if (strcmp(ivmode, "essiv") == 0)
 		cc->iv_gen_ops = &crypt_iv_essiv_ops;
 	else if (strcmp(ivmode, "benbi") == 0)
@@ -1039,6 +1096,12 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	    cc->iv_gen_ops->ctr(cc, ti, ivopts) < 0)
 		goto bad_ivmode;
 
+	if (cc->iv_gen_ops && cc->iv_gen_ops->init &&
+	    cc->iv_gen_ops->init(cc) < 0) {
+		ti->error = "Error initialising IV";
+		goto bad_slab_pool;
+	}
+
 	cc->iv_size = crypto_ablkcipher_ivsize(tfm);
 	if (cc->iv_size)
 		/* at least a 64 bit sector number should fit in our buffer */
@@ -1085,11 +1148,6 @@ static int crypt_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 		goto bad_bs;
 	}
 
-	if (crypto_ablkcipher_setkey(tfm, cc->key, key_size) < 0) {
-		ti->error = "Error setting key";
-		goto bad_device;
-	}
-
 	if (sscanf(argv[2], "%llu", &tmpll) != 1) {
 		ti->error = "Invalid iv_offset sector";
 		goto bad_device;
@@ -1278,6 +1336,7 @@ static void crypt_resume(struct dm_target *ti)
 static int crypt_message(struct dm_target *ti, unsigned argc, char **argv)
 {
 	struct crypt_config *cc = ti->private;
+	int ret = -EINVAL;
 
 	if (argc < 2)
 		goto error;
@@ -1287,10 +1346,22 @@ static int crypt_message(struct dm_target *ti, unsigned argc, char **argv)
 			DMWARN("not suspended during key manipulation.");
 			return -EINVAL;
 		}
-		if (argc == 3 && !strnicmp(argv[1], MESG_STR("set")))
-			return crypt_set_key(cc, argv[2]);
-		if (argc == 2 && !strnicmp(argv[1], MESG_STR("wipe")))
+		if (argc == 3 && !strnicmp(argv[1], MESG_STR("set"))) {
+			ret = crypt_set_key(cc, argv[2]);
+			if (ret)
+				return ret;
+			if (cc->iv_gen_ops && cc->iv_gen_ops->init)
+				ret = cc->iv_gen_ops->init(cc);
+			return ret;
+		}
+		if (argc == 2 && !strnicmp(argv[1], MESG_STR("wipe"))) {
+			if (cc->iv_gen_ops && cc->iv_gen_ops->wipe) {
+				ret = cc->iv_gen_ops->wipe(cc);
+				if (ret)
+					return ret;
+			}
 			return crypt_wipe_key(cc);
+		}
 	}
 
 error:

drivers/md/dm-exception-store.c

@@ -172,7 +172,8 @@ int dm_exception_store_set_chunk_size(struct dm_exception_store *store,
 	}
 
 	/* Validate the chunk size against the device block size */
-	if (chunk_size % (bdev_logical_block_size(store->cow->bdev) >> 9)) {
+	if (chunk_size %
+	    (bdev_logical_block_size(dm_snap_cow(store->snap)->bdev) >> 9)) {
 		*error = "Chunk size is not a multiple of device blocksize";
 		return -EINVAL;
 	}
@@ -190,6 +191,7 @@ int dm_exception_store_set_chunk_size(struct dm_exception_store *store,
 }
 
 int dm_exception_store_create(struct dm_target *ti, int argc, char **argv,
+			      struct dm_snapshot *snap,
 			      unsigned *args_used,
 			      struct dm_exception_store **store)
 {
@@ -198,7 +200,7 @@ int dm_exception_store_create(struct dm_target *ti, int argc, char **argv,
 	struct dm_exception_store *tmp_store;
 	char persistent;
 
-	if (argc < 3) {
+	if (argc < 2) {
 		ti->error = "Insufficient exception store arguments";
 		return -EINVAL;
 	}
@@ -209,14 +211,15 @@ int dm_exception_store_create(struct dm_target *ti, int argc, char **argv,
 		return -ENOMEM;
 	}
 
-	persistent = toupper(*argv[1]);
+	persistent = toupper(*argv[0]);
 	if (persistent == 'P')
 		type = get_type("P");
 	else if (persistent == 'N')
 		type = get_type("N");
 	else {
 		ti->error = "Persistent flag is not P or N";
-		return -EINVAL;
+		r = -EINVAL;
+		goto bad_type;
 	}
 
 	if (!type) {
@@ -226,32 +229,23 @@ int dm_exception_store_create(struct dm_target *ti, int argc, char **argv,
 	}
 
 	tmp_store->type = type;
-	tmp_store->ti = ti;
-
-	r = dm_get_device(ti, argv[0], 0, 0,
-			  FMODE_READ | FMODE_WRITE, &tmp_store->cow);
-	if (r) {
-		ti->error = "Cannot get COW device";
-		goto bad_cow;
-	}
+	tmp_store->snap = snap;
 
-	r = set_chunk_size(tmp_store, argv[2], &ti->error);
+	r = set_chunk_size(tmp_store, argv[1], &ti->error);
 	if (r)
-		goto bad_ctr;
+		goto bad;
 
 	r = type->ctr(tmp_store, 0, NULL);
 	if (r) {
 		ti->error = "Exception store type constructor failed";
-		goto bad_ctr;
+		goto bad;
 	}
 
-	*args_used = 3;
+	*args_used = 2;
 	*store = tmp_store;
 	return 0;
 
-bad_ctr:
-	dm_put_device(ti, tmp_store->cow);
-bad_cow:
+bad:
 	put_type(type);
 bad_type:
 	kfree(tmp_store);
@@ -262,7 +256,6 @@ EXPORT_SYMBOL(dm_exception_store_create);
 void dm_exception_store_destroy(struct dm_exception_store *store)
 {
 	store->type->dtr(store);
-	dm_put_device(store->ti, store->cow);
 	put_type(store->type);
 	kfree(store);
 }

drivers/md/dm-exception-store.h

@@ -26,7 +26,7 @@ typedef sector_t chunk_t;
  * of chunks that follow contiguously.  Remaining bits hold the number of the
  * chunk within the device.
  */
-struct dm_snap_exception {
+struct dm_exception {
 	struct list_head hash_list;
 
 	chunk_t old_chunk;
@@ -64,16 +64,33 @@ struct dm_exception_store_type {
 	 * Find somewhere to store the next exception.
 	 */
 	int (*prepare_exception) (struct dm_exception_store *store,
-				  struct dm_snap_exception *e);
+				  struct dm_exception *e);
 
 	/*
 	 * Update the metadata with this exception.
 	 */
 	void (*commit_exception) (struct dm_exception_store *store,
-				  struct dm_snap_exception *e,
+				  struct dm_exception *e,
 				  void (*callback) (void *, int success),
 				  void *callback_context);
 
+	/*
+	 * Returns 0 if the exception store is empty.
+	 *
+	 * If there are exceptions still to be merged, sets
+	 * *last_old_chunk and *last_new_chunk to the most recent
+	 * still-to-be-merged chunk and returns the number of
+	 * consecutive previous ones.
+	 */
+	int (*prepare_merge) (struct dm_exception_store *store,
+			      chunk_t *last_old_chunk, chunk_t *last_new_chunk);
+
+	/*
+	 * Clear the last n exceptions.
+	 * nr_merged must be <= the value returned by prepare_merge.
+	 */
+	int (*commit_merge) (struct dm_exception_store *store, int nr_merged);
+
 	/*
 	 * The snapshot is invalid, note this in the metadata.
 	 */
@@ -86,19 +103,19 @@ struct dm_exception_store_type {
 	/*
 	 * Return how full the snapshot is.
 	 */
-	void (*fraction_full) (struct dm_exception_store *store,
-			       sector_t *numerator,
-			       sector_t *denominator);
+	void (*usage) (struct dm_exception_store *store,
+		       sector_t *total_sectors, sector_t *sectors_allocated,
+		       sector_t *metadata_sectors);
 
 	/* For internal device-mapper use only. */
 	struct list_head list;
 };
 
+struct dm_snapshot;
+
 struct dm_exception_store {
 	struct dm_exception_store_type *type;
-	struct dm_target *ti;
-
-	struct dm_dev *cow;
+	struct dm_snapshot *snap;
 
 	/* Size of data blocks saved - must be a power of 2 */
 	unsigned chunk_size;
@@ -108,6 +125,11 @@ struct dm_exception_store {
 	void *context;
 };
 
+/*
+ * Obtain the cow device used by a given snapshot.
+ */
+struct dm_dev *dm_snap_cow(struct dm_snapshot *snap);
+
 /*
  * Funtions to manipulate consecutive chunks
  */
@@ -120,18 +142,25 @@ static inline chunk_t dm_chunk_number(chunk_t chunk)
 	return chunk & (chunk_t)((1ULL << DM_CHUNK_NUMBER_BITS) - 1ULL);
 }
 
-static inline unsigned dm_consecutive_chunk_count(struct dm_snap_exception *e)
+static inline unsigned dm_consecutive_chunk_count(struct dm_exception *e)
 {
 	return e->new_chunk >> DM_CHUNK_NUMBER_BITS;
 }
 
-static inline void dm_consecutive_chunk_count_inc(struct dm_snap_exception *e)
+static inline void dm_consecutive_chunk_count_inc(struct dm_exception *e)
 {
 	e->new_chunk += (1ULL << DM_CHUNK_NUMBER_BITS);
 
 	BUG_ON(!dm_consecutive_chunk_count(e));
 }
 
+static inline void dm_consecutive_chunk_count_dec(struct dm_exception *e)
+{
+	BUG_ON(!dm_consecutive_chunk_count(e));
+
+	e->new_chunk -= (1ULL << DM_CHUNK_NUMBER_BITS);
+}
+
 #  else
 #    define DM_CHUNK_CONSECUTIVE_BITS 0
 
@@ -140,12 +169,16 @@ static inline chunk_t dm_chunk_number(chunk_t chunk)
 	return chunk;
 }
 
-static inline unsigned dm_consecutive_chunk_count(struct dm_snap_exception *e)
+static inline unsigned dm_consecutive_chunk_count(struct dm_exception *e)
 {
 	return 0;
 }
 
-static inline void dm_consecutive_chunk_count_inc(struct dm_snap_exception *e)
+static inline void dm_consecutive_chunk_count_inc(struct dm_exception *e)
+{
+}
+
+static inline void dm_consecutive_chunk_count_dec(struct dm_exception *e)
 {
 }
 
@@ -162,7 +195,7 @@ static inline sector_t get_dev_size(struct block_device *bdev)
 static inline chunk_t sector_to_chunk(struct dm_exception_store *store,
 				      sector_t sector)
 {
-	return (sector & ~store->chunk_mask) >> store->chunk_shift;
+	return sector >> store->chunk_shift;
 }
 
 int dm_exception_store_type_register(struct dm_exception_store_type *type);
@@ -173,6 +206,7 @@ int dm_exception_store_set_chunk_size(struct dm_exception_store *store,
 				      char **error);
 
 int dm_exception_store_create(struct dm_target *ti, int argc, char **argv,
+			      struct dm_snapshot *snap,
 			      unsigned *args_used,
 			      struct dm_exception_store **store);
 void dm_exception_store_destroy(struct dm_exception_store *store);

drivers/md/dm-io.c

@@ -5,6 +5,8 @@
  * This file is released under the GPL.
  */
 
+#include "dm.h"
+
 #include <linux/device-mapper.h>
 
 #include <linux/bio.h>
@@ -14,12 +16,19 @@
 #include <linux/slab.h>
 #include <linux/dm-io.h>
 
+#define DM_MSG_PREFIX "io"
+
+#define DM_IO_MAX_REGIONS	BITS_PER_LONG
+
 struct dm_io_client {
 	mempool_t *pool;
 	struct bio_set *bios;
 };
 
-/* FIXME: can we shrink this ? */
+/*
+ * Aligning 'struct io' reduces the number of bits required to store
+ * its address.  Refer to store_io_and_region_in_bio() below.
+ */
 struct io {
 	unsigned long error_bits;
 	unsigned long eopnotsupp_bits;
@@ -28,7 +37,9 @@ struct io {
 	struct dm_io_client *client;
 	io_notify_fn callback;
 	void *context;
-};
+} __attribute__((aligned(DM_IO_MAX_REGIONS)));
+
+static struct kmem_cache *_dm_io_cache;
 
 /*
  * io contexts are only dynamically allocated for asynchronous
@@ -53,7 +64,7 @@ struct dm_io_client *dm_io_client_create(unsigned num_pages)
 	if (!client)
 		return ERR_PTR(-ENOMEM);
 
-	client->pool = mempool_create_kmalloc_pool(ios, sizeof(struct io));
+	client->pool = mempool_create_slab_pool(ios, _dm_io_cache);
 	if (!client->pool)
 		goto bad;
 
@@ -88,18 +99,29 @@ EXPORT_SYMBOL(dm_io_client_destroy);
 
 /*-----------------------------------------------------------------
  * We need to keep track of which region a bio is doing io for.
- * In order to save a memory allocation we store this the last
- * bvec which we know is unused (blech).
- * XXX This is ugly and can OOPS with some configs... find another way.
+ * To avoid a memory allocation to store just 5 or 6 bits, we
+ * ensure the 'struct io' pointer is aligned so enough low bits are
+ * always zero and then combine it with the region number directly in
+ * bi_private.
  *---------------------------------------------------------------*/
-static inline void bio_set_region(struct bio *bio, unsigned region)
+static void store_io_and_region_in_bio(struct bio *bio, struct io *io,
+				       unsigned region)
 {
-	bio->bi_io_vec[bio->bi_max_vecs].bv_len = region;
+	if (unlikely(!IS_ALIGNED((unsigned long)io, DM_IO_MAX_REGIONS))) {
+		DMCRIT("Unaligned struct io pointer %p", io);
+		BUG();
+	}
+
+	bio->bi_private = (void *)((unsigned long)io | region);
 }
 
-static inline unsigned bio_get_region(struct bio *bio)
+static void retrieve_io_and_region_from_bio(struct bio *bio, struct io **io,
+				       unsigned *region)
 {
-	return bio->bi_io_vec[bio->bi_max_vecs].bv_len;
+	unsigned long val = (unsigned long)bio->bi_private;
+
+	*io = (void *)(val & -(unsigned long)DM_IO_MAX_REGIONS);
+	*region = val & (DM_IO_MAX_REGIONS - 1);
 }
 
 /*-----------------------------------------------------------------
@@ -140,10 +162,8 @@ static void endio(struct bio *bio, int error)
 	/*
 	 * The bio destructor in bio_put() may use the io object.
 	 */
-	io = bio->bi_private;
-	region = bio_get_region(bio);
+	retrieve_io_and_region_from_bio(bio, &io, &region);
 
-	bio->bi_max_vecs++;
 	bio_put(bio);
 
 	dec_count(io, region, error);
@@ -243,7 +263,10 @@ static void vm_dp_init(struct dpages *dp, void *data)
 
 static void dm_bio_destructor(struct bio *bio)
 {
-	struct io *io = bio->bi_private;
+	unsigned region;
+	struct io *io;
+
+	retrieve_io_and_region_from_bio(bio, &io, &region);
 
 	bio_free(bio, io->client->bios);
 }
@@ -286,26 +309,23 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where,
 	unsigned num_bvecs;
 	sector_t remaining = where->count;
 
-	while (remaining) {
+	/*
+	 * where->count may be zero if rw holds a write barrier and we
+	 * need to send a zero-sized barrier.
+	 */
+	do {
 		/*
-		 * Allocate a suitably sized-bio: we add an extra
-		 * bvec for bio_get/set_region() and decrement bi_max_vecs
-		 * to hide it from bio_add_page().
+		 * Allocate a suitably sized-bio.
 		 */
 		num_bvecs = dm_sector_div_up(remaining,
 					     (PAGE_SIZE >> SECTOR_SHIFT));
-		num_bvecs = 1 + min_t(int, bio_get_nr_vecs(where->bdev),
-				      num_bvecs);
-		if (unlikely(num_bvecs > BIO_MAX_PAGES))
-			num_bvecs = BIO_MAX_PAGES;
+		num_bvecs = min_t(int, bio_get_nr_vecs(where->bdev), num_bvecs);
 		bio = bio_alloc_bioset(GFP_NOIO, num_bvecs, io->client->bios);
 		bio->bi_sector = where->sector + (where->count - remaining);
 		bio->bi_bdev = where->bdev;
 		bio->bi_end_io = endio;
-		bio->bi_private = io;
 		bio->bi_destructor = dm_bio_destructor;
-		bio->bi_max_vecs--;
-		bio_set_region(bio, region);
+		store_io_and_region_in_bio(bio, io, region);
 
 		/*
 		 * Try and add as many pages as possible.
@@ -323,7 +343,7 @@ static void do_region(int rw, unsigned region, struct dm_io_region *where,
 
 		atomic_inc(&io->count);
 		submit_bio(rw, bio);
-	}
+	} while (remaining);
 }
 
 static void dispatch_io(int rw, unsigned int num_regions,
@@ -333,6 +353,8 @@ static void dispatch_io(int rw, unsigned int num_regions,
 	int i;
 	struct dpages old_pages = *dp;
 
+	BUG_ON(num_regions > DM_IO_MAX_REGIONS);
+
 	if (sync)
 		rw |= (1 << BIO_RW_SYNCIO) | (1 << BIO_RW_UNPLUG);
 
@@ -342,7 +364,7 @@ static void dispatch_io(int rw, unsigned int num_regions,
 	 */
 	for (i = 0; i < num_regions; i++) {
 		*dp = old_pages;
-		if (where[i].count)
+		if (where[i].count || (rw & (1 << BIO_RW_BARRIER)))
 			do_region(rw, i, where + i, dp, io);
 	}
 
@@ -357,7 +379,14 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 		   struct dm_io_region *where, int rw, struct dpages *dp,
 		   unsigned long *error_bits)
 {
-	struct io io;
+	/*
+	 * gcc <= 4.3 can't do the alignment for stack variables, so we must
+	 * align it on our own.
+	 * volatile prevents the optimizer from removing or reusing
+	 * "io_" field from the stack frame (allowed in ANSI C).
+	 */
+	volatile char io_[sizeof(struct io) + __alignof__(struct io) - 1];
+	struct io *io = (struct io *)PTR_ALIGN(&io_, __alignof__(struct io));
 
 	if (num_regions > 1 && (rw & RW_MASK) != WRITE) {
 		WARN_ON(1);
@@ -365,33 +394,33 @@ static int sync_io(struct dm_io_client *client, unsigned int num_regions,
 	}
 
 retry:
-	io.error_bits = 0;
-	io.eopnotsupp_bits = 0;
-	atomic_set(&io.count, 1); /* see dispatch_io() */
-	io.sleeper = current;
-	io.client = client;
+	io->error_bits = 0;
+	io->eopnotsupp_bits = 0;
+	atomic_set(&io->count, 1); /* see dispatch_io() */
+	io->sleeper = current;
+	io->client = client;
 
-	dispatch_io(rw, num_regions, where, dp, &io, 1);
+	dispatch_io(rw, num_regions, where, dp, io, 1);
 
 	while (1) {
 		set_current_state(TASK_UNINTERRUPTIBLE);
 
-		if (!atomic_read(&io.count))
+		if (!atomic_read(&io->count))
 			break;
 
 		io_schedule();
 	}
 	set_current_state(TASK_RUNNING);
 
-	if (io.eopnotsupp_bits && (rw & (1 << BIO_RW_BARRIER))) {
+	if (io->eopnotsupp_bits && (rw & (1 << BIO_RW_BARRIER))) {
 		rw &= ~(1 << BIO_RW_BARRIER);
 		goto retry;
 	}
 
 	if (error_bits)
-		*error_bits = io.error_bits;
+		*error_bits = io->error_bits;
 
-	return io.error_bits ? -EIO : 0;
+	return io->error_bits ? -EIO : 0;
 }
 
 static int async_io(struct dm_io_client *client, unsigned int num_regions,
@@ -472,3 +501,18 @@ int dm_io(struct dm_io_request *io_req, unsigned num_regions,
 			&dp, io_req->notify.fn, io_req->notify.context);
 }
 EXPORT_SYMBOL(dm_io);
+
+int __init dm_io_init(void)
+{
+	_dm_io_cache = KMEM_CACHE(io, 0);
+	if (!_dm_io_cache)
+		return -ENOMEM;
+
+	return 0;
+}
+
+void dm_io_exit(void)
+{
+	kmem_cache_destroy(_dm_io_cache);
+	_dm_io_cache = NULL;
+}

drivers/md/dm-ioctl.c

@@ -56,6 +56,11 @@ static void dm_hash_remove_all(int keep_open_devices);
  */
 static DECLARE_RWSEM(_hash_lock);
 
+/*
+ * Protects use of mdptr to obtain hash cell name and uuid from mapped device.
+ */
+static DEFINE_MUTEX(dm_hash_cells_mutex);
+
 static void init_buckets(struct list_head *buckets)
 {
 	unsigned int i;
@@ -206,7 +211,9 @@ static int dm_hash_insert(const char *name, const char *uuid, struct mapped_devi
 		list_add(&cell->uuid_list, _uuid_buckets + hash_str(uuid));
 	}
 	dm_get(md);
+	mutex_lock(&dm_hash_cells_mutex);
 	dm_set_mdptr(md, cell);
+	mutex_unlock(&dm_hash_cells_mutex);
 	up_write(&_hash_lock);
 
 	return 0;
@@ -224,9 +231,11 @@ static void __hash_remove(struct hash_cell *hc)
 	/* remove from the dev hash */
 	list_del(&hc->uuid_list);
 	list_del(&hc->name_list);
+	mutex_lock(&dm_hash_cells_mutex);
 	dm_set_mdptr(hc->md, NULL);
+	mutex_unlock(&dm_hash_cells_mutex);
 
-	table = dm_get_table(hc->md);
+	table = dm_get_live_table(hc->md);
 	if (table) {
 		dm_table_event(table);
 		dm_table_put(table);
@@ -321,13 +330,15 @@ static int dm_hash_rename(uint32_t cookie, const char *old, const char *new)
 	 */
 	list_del(&hc->name_list);
 	old_name = hc->name;
+	mutex_lock(&dm_hash_cells_mutex);
 	hc->name = new_name;
+	mutex_unlock(&dm_hash_cells_mutex);
 	list_add(&hc->name_list, _name_buckets + hash_str(new_name));
 
 	/*
 	 * Wake up any dm event waiters.
 	 */
-	table = dm_get_table(hc->md);
+	table = dm_get_live_table(hc->md);
 	if (table) {
 		dm_table_event(table);
 		dm_table_put(table);
@@ -512,8 +523,6 @@ static int list_versions(struct dm_ioctl *param, size_t param_size)
 	return 0;
 }
 
-
-
 static int check_name(const char *name)
 {
 	if (strchr(name, '/')) {
@@ -524,6 +533,40 @@ static int check_name(const char *name)
 	return 0;
 }
 
+/*
+ * On successful return, the caller must not attempt to acquire
+ * _hash_lock without first calling dm_table_put, because dm_table_destroy
+ * waits for this dm_table_put and could be called under this lock.
+ */
+static struct dm_table *dm_get_inactive_table(struct mapped_device *md)
+{
+	struct hash_cell *hc;
+	struct dm_table *table = NULL;
+
+	down_read(&_hash_lock);
+	hc = dm_get_mdptr(md);
+	if (!hc || hc->md != md) {
+		DMWARN("device has been removed from the dev hash table.");
+		goto out;
+	}
+
+	table = hc->new_map;
+	if (table)
+		dm_table_get(table);
+
+out:
+	up_read(&_hash_lock);
+
+	return table;
+}
+
+static struct dm_table *dm_get_live_or_inactive_table(struct mapped_device *md,
+						      struct dm_ioctl *param)
+{
+	return (param->flags & DM_QUERY_INACTIVE_TABLE_FLAG) ?
+		dm_get_inactive_table(md) : dm_get_live_table(md);
+}
+
 /*
  * Fills in a dm_ioctl structure, ready for sending back to
  * userland.
@@ -536,7 +579,7 @@ static int __dev_status(struct mapped_device *md, struct dm_ioctl *param)
 	param->flags &= ~(DM_SUSPEND_FLAG | DM_READONLY_FLAG |
 			  DM_ACTIVE_PRESENT_FLAG);
 
-	if (dm_suspended(md))
+	if (dm_suspended_md(md))
 		param->flags |= DM_SUSPEND_FLAG;
 
 	param->dev = huge_encode_dev(disk_devt(disk));
@@ -548,18 +591,30 @@ static int __dev_status(struct mapped_device *md, struct dm_ioctl *param)
 	 */
 	param->open_count = dm_open_count(md);
 
-	if (get_disk_ro(disk))
-		param->flags |= DM_READONLY_FLAG;
-
 	param->event_nr = dm_get_event_nr(md);
+	param->target_count = 0;
 
-	table = dm_get_table(md);
+	table = dm_get_live_table(md);
 	if (table) {
-		param->flags |= DM_ACTIVE_PRESENT_FLAG;
-		param->target_count = dm_table_get_num_targets(table);
+		if (!(param->flags & DM_QUERY_INACTIVE_TABLE_FLAG)) {
+			if (get_disk_ro(disk))
+				param->flags |= DM_READONLY_FLAG;
+			param->target_count = dm_table_get_num_targets(table);
+		}
 		dm_table_put(table);
-	} else
-		param->target_count = 0;
+
+		param->flags |= DM_ACTIVE_PRESENT_FLAG;
+	}
+
+	if (param->flags & DM_QUERY_INACTIVE_TABLE_FLAG) {
+		table = dm_get_inactive_table(md);
+		if (table) {
+			if (!(dm_table_get_mode(table) & FMODE_WRITE))
+				param->flags |= DM_READONLY_FLAG;
+			param->target_count = dm_table_get_num_targets(table);
+			dm_table_put(table);
+		}
+	}
 
 	return 0;
 }
@@ -634,9 +689,9 @@ static struct mapped_device *find_device(struct dm_ioctl *param)
 		 * Sneakily write in both the name and the uuid
 		 * while we have the cell.
 		 */
-		strncpy(param->name, hc->name, sizeof(param->name));
+		strlcpy(param->name, hc->name, sizeof(param->name));
 		if (hc->uuid)
-			strncpy(param->uuid, hc->uuid, sizeof(param->uuid)-1);
+			strlcpy(param->uuid, hc->uuid, sizeof(param->uuid));
 		else
 			param->uuid[0] = '\0';
 
@@ -784,7 +839,7 @@ static int do_suspend(struct dm_ioctl *param)
 	if (param->flags & DM_NOFLUSH_FLAG)
 		suspend_flags |= DM_SUSPEND_NOFLUSH_FLAG;
 
-	if (!dm_suspended(md))
+	if (!dm_suspended_md(md))
 		r = dm_suspend(md, suspend_flags);
 
 	if (!r)
@@ -800,7 +855,7 @@ static int do_resume(struct dm_ioctl *param)
 	unsigned suspend_flags = DM_SUSPEND_LOCKFS_FLAG;
 	struct hash_cell *hc;
 	struct mapped_device *md;
-	struct dm_table *new_map;
+	struct dm_table *new_map, *old_map = NULL;
 
 	down_write(&_hash_lock);
 
@@ -826,14 +881,14 @@ static int do_resume(struct dm_ioctl *param)
 			suspend_flags &= ~DM_SUSPEND_LOCKFS_FLAG;
 		if (param->flags & DM_NOFLUSH_FLAG)
 			suspend_flags |= DM_SUSPEND_NOFLUSH_FLAG;
-		if (!dm_suspended(md))
+		if (!dm_suspended_md(md))
 			dm_suspend(md, suspend_flags);
 
-		r = dm_swap_table(md, new_map);
-		if (r) {
+		old_map = dm_swap_table(md, new_map);
+		if (IS_ERR(old_map)) {
 			dm_table_destroy(new_map);
 			dm_put(md);
-			return r;
+			return PTR_ERR(old_map);
 		}
 
 		if (dm_table_get_mode(new_map) & FMODE_WRITE)
@@ -842,9 +897,11 @@ static int do_resume(struct dm_ioctl *param)
 			set_disk_ro(dm_disk(md), 1);
 	}
 
-	if (dm_suspended(md))
+	if (dm_suspended_md(md))
 		r = dm_resume(md);
 
+	if (old_map)
+		dm_table_destroy(old_map);
 
 	if (!r) {
 		dm_kobject_uevent(md, KOBJ_CHANGE, param->event_nr);
@@ -982,7 +1039,7 @@ static int dev_wait(struct dm_ioctl *param, size_t param_size)
 	if (r)
 		goto out;
 
-	table = dm_get_table(md);
+	table = dm_get_live_or_inactive_table(md, param);
 	if (table) {
 		retrieve_status(table, param, param_size);
 		dm_table_put(table);
@@ -1215,7 +1272,7 @@ static int table_deps(struct dm_ioctl *param, size_t param_size)
 	if (r)
 		goto out;
 
-	table = dm_get_table(md);
+	table = dm_get_live_or_inactive_table(md, param);
 	if (table) {
 		retrieve_deps(table, param, param_size);
 		dm_table_put(table);
@@ -1244,13 +1301,13 @@ static int table_status(struct dm_ioctl *param, size_t param_size)
 	if (r)
 		goto out;
 
-	table = dm_get_table(md);
+	table = dm_get_live_or_inactive_table(md, param);
 	if (table) {
 		retrieve_status(table, param, param_size);
 		dm_table_put(table);
 	}
 
- out:
+out:
 	dm_put(md);
 	return r;
 }
@@ -1288,10 +1345,15 @@ static int target_message(struct dm_ioctl *param, size_t param_size)
 		goto out;
 	}
 
-	table = dm_get_table(md);
+	table = dm_get_live_table(md);
 	if (!table)
 		goto out_argv;
 
+	if (dm_deleting_md(md)) {
+		r = -ENXIO;
+		goto out_table;
+	}
+
 	ti = dm_table_find_target(table, tmsg->sector);
 	if (!dm_target_is_valid(ti)) {
 		DMWARN("Target message sector outside device.");
@@ -1303,6 +1365,7 @@ static int target_message(struct dm_ioctl *param, size_t param_size)
 		r = -EINVAL;
 	}
 
+ out_table:
 	dm_table_put(table);
  out_argv:
 	kfree(argv);
@@ -1582,8 +1645,7 @@ int dm_copy_name_and_uuid(struct mapped_device *md, char *name, char *uuid)
 	if (!md)
 		return -ENXIO;
 
-	dm_get(md);
-	down_read(&_hash_lock);
+	mutex_lock(&dm_hash_cells_mutex);
 	hc = dm_get_mdptr(md);
 	if (!hc || hc->md != md) {
 		r = -ENXIO;
@@ -1596,8 +1658,7 @@ int dm_copy_name_and_uuid(struct mapped_device *md, char *name, char *uuid)
 		strcpy(uuid, hc->uuid ? : "");
 
 out:
-	up_read(&_hash_lock);
-	dm_put(md);
+	mutex_unlock(&dm_hash_cells_mutex);
 
 	return r;
 }

drivers/md/dm-kcopyd.c

@@ -450,7 +450,10 @@ static void dispatch_job(struct kcopyd_job *job)
 {
 	struct dm_kcopyd_client *kc = job->kc;
 	atomic_inc(&kc->nr_jobs);
-	push(&kc->pages_jobs, job);
+	if (unlikely(!job->source.count))
+		push(&kc->complete_jobs, job);
+	else
+		push(&kc->pages_jobs, job);
 	wake(kc);
 }
 

drivers/md/dm-log.c

@@ -145,8 +145,9 @@ int dm_dirty_log_type_unregister(struct dm_dirty_log_type *type)
 EXPORT_SYMBOL(dm_dirty_log_type_unregister);
 
 struct dm_dirty_log *dm_dirty_log_create(const char *type_name,
-					 struct dm_target *ti,
-					 unsigned int argc, char **argv)
+			struct dm_target *ti,
+			int (*flush_callback_fn)(struct dm_target *ti),
+			unsigned int argc, char **argv)
 {
 	struct dm_dirty_log_type *type;
 	struct dm_dirty_log *log;
@@ -161,6 +162,7 @@ struct dm_dirty_log *dm_dirty_log_create(const char *type_name,
 		return NULL;
 	}
 
+	log->flush_callback_fn = flush_callback_fn;
 	log->type = type;
 	if (type->ctr(log, ti, argc, argv)) {
 		kfree(log);
@@ -208,7 +210,9 @@ struct log_header {
 
 struct log_c {
 	struct dm_target *ti;
-	int touched;
+	int touched_dirtied;
+	int touched_cleaned;
+	int flush_failed;
 	uint32_t region_size;
 	unsigned int region_count;
 	region_t sync_count;
@@ -233,6 +237,7 @@ struct log_c {
 	 * Disk log fields
 	 */
 	int log_dev_failed;
+	int log_dev_flush_failed;
 	struct dm_dev *log_dev;
 	struct log_header header;
 
@@ -253,14 +258,14 @@ static inline void log_set_bit(struct log_c *l,
 			       uint32_t *bs, unsigned bit)
 {
 	ext2_set_bit(bit, (unsigned long *) bs);
-	l->touched = 1;
+	l->touched_cleaned = 1;
 }
 
 static inline void log_clear_bit(struct log_c *l,
 				 uint32_t *bs, unsigned bit)
 {
 	ext2_clear_bit(bit, (unsigned long *) bs);
-	l->touched = 1;
+	l->touched_dirtied = 1;
 }
 
 /*----------------------------------------------------------------
@@ -287,6 +292,19 @@ static int rw_header(struct log_c *lc, int rw)
 	return dm_io(&lc->io_req, 1, &lc->header_location, NULL);
 }
 
+static int flush_header(struct log_c *lc)
+{
+	struct dm_io_region null_location = {
+		.bdev = lc->header_location.bdev,
+		.sector = 0,
+		.count = 0,
+	};
+
+	lc->io_req.bi_rw = WRITE_BARRIER;
+
+	return dm_io(&lc->io_req, 1, &null_location, NULL);
+}
+
 static int read_header(struct log_c *log)
 {
 	int r;
@@ -378,7 +396,9 @@ static int create_log_context(struct dm_dirty_log *log, struct dm_target *ti,
 	}
 
 	lc->ti = ti;
-	lc->touched = 0;
+	lc->touched_dirtied = 0;
+	lc->touched_cleaned = 0;
+	lc->flush_failed = 0;
 	lc->region_size = region_size;
 	lc->region_count = region_count;
 	lc->sync = sync;
@@ -406,6 +426,7 @@ static int create_log_context(struct dm_dirty_log *log, struct dm_target *ti,
 	} else {
 		lc->log_dev = dev;
 		lc->log_dev_failed = 0;
+		lc->log_dev_flush_failed = 0;
 		lc->header_location.bdev = lc->log_dev->bdev;
 		lc->header_location.sector = 0;
 
@@ -614,6 +635,11 @@ static int disk_resume(struct dm_dirty_log *log)
 
 	/* write the new header */
 	r = rw_header(lc, WRITE);
+	if (!r) {
+		r = flush_header(lc);
+		if (r)
+			lc->log_dev_flush_failed = 1;
+	}
 	if (r) {
 		DMWARN("%s: Failed to write header on dirty region log device",
 		       lc->log_dev->name);
@@ -656,18 +682,40 @@ static int core_flush(struct dm_dirty_log *log)
 
 static int disk_flush(struct dm_dirty_log *log)
 {
-	int r;
-	struct log_c *lc = (struct log_c *) log->context;
+	int r, i;
+	struct log_c *lc = log->context;
 
 	/* only write if the log has changed */
-	if (!lc->touched)
+	if (!lc->touched_cleaned && !lc->touched_dirtied)
 		return 0;
 
+	if (lc->touched_cleaned && log->flush_callback_fn &&
+	    log->flush_callback_fn(lc->ti)) {
+		/*
+		 * At this point it is impossible to determine which
+		 * regions are clean and which are dirty (without
+		 * re-reading the log off disk). So mark all of them
+		 * dirty.
+		 */
+		lc->flush_failed = 1;
+		for (i = 0; i < lc->region_count; i++)
+			log_clear_bit(lc, lc->clean_bits, i);
+	}
+
 	r = rw_header(lc, WRITE);
 	if (r)
 		fail_log_device(lc);
-	else
-		lc->touched = 0;
+	else {
+		if (lc->touched_dirtied) {
+			r = flush_header(lc);
+			if (r) {
+				lc->log_dev_flush_failed = 1;
+				fail_log_device(lc);
+			} else
+				lc->touched_dirtied = 0;
+		}
+		lc->touched_cleaned = 0;
+	}
 
 	return r;
 }
@@ -681,7 +729,8 @@ static void core_mark_region(struct dm_dirty_log *log, region_t region)
 static void core_clear_region(struct dm_dirty_log *log, region_t region)
 {
 	struct log_c *lc = (struct log_c *) log->context;
-	log_set_bit(lc, lc->clean_bits, region);
+	if (likely(!lc->flush_failed))
+		log_set_bit(lc, lc->clean_bits, region);
 }
 
 static int core_get_resync_work(struct dm_dirty_log *log, region_t *region)
@@ -762,7 +811,9 @@ static int disk_status(struct dm_dirty_log *log, status_type_t status,
 	switch(status) {
 	case STATUSTYPE_INFO:
 		DMEMIT("3 %s %s %c", log->type->name, lc->log_dev->name,
-		       lc->log_dev_failed ? 'D' : 'A');
+		       lc->log_dev_flush_failed ? 'F' :
+		       lc->log_dev_failed ? 'D' :
+		       'A');
 		break;
 
 	case STATUSTYPE_TABLE:

drivers/md/dm-mpath.c

@@ -93,6 +93,10 @@ struct multipath {
 	 * can resubmit bios on error.
 	 */
 	mempool_t *mpio_pool;
+
+	struct mutex work_mutex;
+
+	unsigned suspended;	/* Don't create new I/O internally when set. */
 };
 
 /*
@@ -198,6 +202,7 @@ static struct multipath *alloc_multipath(struct dm_target *ti)
 		m->queue_io = 1;
 		INIT_WORK(&m->process_queued_ios, process_queued_ios);
 		INIT_WORK(&m->trigger_event, trigger_event);
+		mutex_init(&m->work_mutex);
 		m->mpio_pool = mempool_create_slab_pool(MIN_IOS, _mpio_cache);
 		if (!m->mpio_pool) {
 			kfree(m);
@@ -885,13 +890,18 @@ static int multipath_ctr(struct dm_target *ti, unsigned int argc,
 	return r;
 }
 
-static void multipath_dtr(struct dm_target *ti)
+static void flush_multipath_work(void)
 {
-	struct multipath *m = (struct multipath *) ti->private;
-
 	flush_workqueue(kmpath_handlerd);
 	flush_workqueue(kmultipathd);
 	flush_scheduled_work();
+}
+
+static void multipath_dtr(struct dm_target *ti)
+{
+	struct multipath *m = ti->private;
+
+	flush_multipath_work();
 	free_multipath(m);
 }
 
@@ -1261,6 +1271,16 @@ static void multipath_presuspend(struct dm_target *ti)
 	queue_if_no_path(m, 0, 1);
 }
 
+static void multipath_postsuspend(struct dm_target *ti)
+{
+	struct multipath *m = ti->private;
+
+	mutex_lock(&m->work_mutex);
+	m->suspended = 1;
+	flush_multipath_work();
+	mutex_unlock(&m->work_mutex);
+}
+
 /*
  * Restore the queue_if_no_path setting.
  */
@@ -1269,6 +1289,10 @@ static void multipath_resume(struct dm_target *ti)
 	struct multipath *m = (struct multipath *) ti->private;
 	unsigned long flags;
 
+	mutex_lock(&m->work_mutex);
+	m->suspended = 0;
+	mutex_unlock(&m->work_mutex);
+
 	spin_lock_irqsave(&m->lock, flags);
 	m->queue_if_no_path = m->saved_queue_if_no_path;
 	spin_unlock_irqrestore(&m->lock, flags);
@@ -1397,51 +1421,71 @@ static int multipath_status(struct dm_target *ti, status_type_t type,
 
 static int multipath_message(struct dm_target *ti, unsigned argc, char **argv)
 {
-	int r;
+	int r = -EINVAL;
 	struct dm_dev *dev;
 	struct multipath *m = (struct multipath *) ti->private;
 	action_fn action;
 
+	mutex_lock(&m->work_mutex);
+
+	if (m->suspended) {
+		r = -EBUSY;
+		goto out;
+	}
+
+	if (dm_suspended(ti)) {
+		r = -EBUSY;
+		goto out;
+	}
+
 	if (argc == 1) {
-		if (!strnicmp(argv[0], MESG_STR("queue_if_no_path")))
-			return queue_if_no_path(m, 1, 0);
-		else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path")))
-			return queue_if_no_path(m, 0, 0);
+		if (!strnicmp(argv[0], MESG_STR("queue_if_no_path"))) {
+			r = queue_if_no_path(m, 1, 0);
+			goto out;
+		} else if (!strnicmp(argv[0], MESG_STR("fail_if_no_path"))) {
+			r = queue_if_no_path(m, 0, 0);
+			goto out;
+		}
 	}
 
-	if (argc != 2)
-		goto error;
+	if (argc != 2) {
+		DMWARN("Unrecognised multipath message received.");
+		goto out;
+	}
 
-	if (!strnicmp(argv[0], MESG_STR("disable_group")))
-		return bypass_pg_num(m, argv[1], 1);
-	else if (!strnicmp(argv[0], MESG_STR("enable_group")))
-		return bypass_pg_num(m, argv[1], 0);
-	else if (!strnicmp(argv[0], MESG_STR("switch_group")))
-		return switch_pg_num(m, argv[1]);
-	else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
+	if (!strnicmp(argv[0], MESG_STR("disable_group"))) {
+		r = bypass_pg_num(m, argv[1], 1);
+		goto out;
+	} else if (!strnicmp(argv[0], MESG_STR("enable_group"))) {
+		r = bypass_pg_num(m, argv[1], 0);
+		goto out;
+	} else if (!strnicmp(argv[0], MESG_STR("switch_group"))) {
+		r = switch_pg_num(m, argv[1]);
+		goto out;
+	} else if (!strnicmp(argv[0], MESG_STR("reinstate_path")))
 		action = reinstate_path;
 	else if (!strnicmp(argv[0], MESG_STR("fail_path")))
 		action = fail_path;
-	else
-		goto error;
+	else {
+		DMWARN("Unrecognised multipath message received.");
+		goto out;
+	}
 
 	r = dm_get_device(ti, argv[1], ti->begin, ti->len,
 			  dm_table_get_mode(ti->table), &dev);
 	if (r) {
 		DMWARN("message: error getting device %s",
 		       argv[1]);
-		return -EINVAL;
+		goto out;
 	}
 
 	r = action_dev(m, dev, action);
 
 	dm_put_device(ti, dev);
 
+out:
+	mutex_unlock(&m->work_mutex);
 	return r;
-
-error:
-	DMWARN("Unrecognised multipath message received.");
-	return -EINVAL;
 }
 
 static int multipath_ioctl(struct dm_target *ti, unsigned int cmd,
@@ -1567,13 +1611,14 @@ static int multipath_busy(struct dm_target *ti)
  *---------------------------------------------------------------*/
 static struct target_type multipath_target = {
 	.name = "multipath",
-	.version = {1, 1, 0},
+	.version = {1, 1, 1},
 	.module = THIS_MODULE,
 	.ctr = multipath_ctr,
 	.dtr = multipath_dtr,
 	.map_rq = multipath_map,
 	.rq_end_io = multipath_end_io,
 	.presuspend = multipath_presuspend,
+	.postsuspend = multipath_postsuspend,
 	.resume = multipath_resume,
 	.status = multipath_status,
 	.message = multipath_message,

drivers/md/dm-raid1.c

@@ -35,6 +35,7 @@ static DECLARE_WAIT_QUEUE_HEAD(_kmirrord_recovery_stopped);
  *---------------------------------------------------------------*/
 enum dm_raid1_error {
 	DM_RAID1_WRITE_ERROR,
+	DM_RAID1_FLUSH_ERROR,
 	DM_RAID1_SYNC_ERROR,
 	DM_RAID1_READ_ERROR
 };
@@ -57,6 +58,7 @@ struct mirror_set {
 	struct bio_list reads;
 	struct bio_list writes;
 	struct bio_list failures;
+	struct bio_list holds;	/* bios are waiting until suspend */
 
 	struct dm_region_hash *rh;
 	struct dm_kcopyd_client *kcopyd_client;
@@ -67,6 +69,7 @@ struct mirror_set {
 	region_t nr_regions;
 	int in_sync;
 	int log_failure;
+	int leg_failure;
 	atomic_t suspend;
 
 	atomic_t default_mirror;	/* Default mirror */
@@ -179,6 +182,17 @@ static void set_default_mirror(struct mirror *m)
 	atomic_set(&ms->default_mirror, m - m0);
 }
 
+static struct mirror *get_valid_mirror(struct mirror_set *ms)
+{
+	struct mirror *m;
+
+	for (m = ms->mirror; m < ms->mirror + ms->nr_mirrors; m++)
+		if (!atomic_read(&m->error_count))
+			return m;
+
+	return NULL;
+}
+
 /* fail_mirror
  * @m: mirror device to fail
  * @error_type: one of the enum's, DM_RAID1_*_ERROR
@@ -198,6 +212,8 @@ static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
 	struct mirror_set *ms = m->ms;
 	struct mirror *new;
 
+	ms->leg_failure = 1;
+
 	/*
 	 * error_count is used for nothing more than a
 	 * simple way to tell if a device has encountered
@@ -224,19 +240,50 @@ static void fail_mirror(struct mirror *m, enum dm_raid1_error error_type)
 		goto out;
 	}
 
-	for (new = ms->mirror; new < ms->mirror + ms->nr_mirrors; new++)
-		if (!atomic_read(&new->error_count)) {
-			set_default_mirror(new);
-			break;
-		}
-
-	if (unlikely(new == ms->mirror + ms->nr_mirrors))
+	new = get_valid_mirror(ms);
+	if (new)
+		set_default_mirror(new);
+	else
 		DMWARN("All sides of mirror have failed.");
 
 out:
 	schedule_work(&ms->trigger_event);
 }
 
+static int mirror_flush(struct dm_target *ti)
+{
+	struct mirror_set *ms = ti->private;
+	unsigned long error_bits;
+
+	unsigned int i;
+	struct dm_io_region io[ms->nr_mirrors];
+	struct mirror *m;
+	struct dm_io_request io_req = {
+		.bi_rw = WRITE_BARRIER,
+		.mem.type = DM_IO_KMEM,
+		.mem.ptr.bvec = NULL,
+		.client = ms->io_client,
+	};
+
+	for (i = 0, m = ms->mirror; i < ms->nr_mirrors; i++, m++) {
+		io[i].bdev = m->dev->bdev;
+		io[i].sector = 0;
+		io[i].count = 0;
+	}
+
+	error_bits = -1;
+	dm_io(&io_req, ms->nr_mirrors, io, &error_bits);
+	if (unlikely(error_bits != 0)) {
+		for (i = 0; i < ms->nr_mirrors; i++)
+			if (test_bit(i, &error_bits))
+				fail_mirror(ms->mirror + i,
+					    DM_RAID1_FLUSH_ERROR);
+		return -EIO;
+	}
+
+	return 0;
+}
+
 /*-----------------------------------------------------------------
  * Recovery.
  *
@@ -396,6 +443,8 @@ static int mirror_available(struct mirror_set *ms, struct bio *bio)
  */
 static sector_t map_sector(struct mirror *m, struct bio *bio)
 {
+	if (unlikely(!bio->bi_size))
+		return 0;
 	return m->offset + (bio->bi_sector - m->ms->ti->begin);
 }
 
@@ -413,6 +462,27 @@ static void map_region(struct dm_io_region *io, struct mirror *m,
 	io->count = bio->bi_size >> 9;
 }
 
+static void hold_bio(struct mirror_set *ms, struct bio *bio)
+{
+	/*
+	 * If device is suspended, complete the bio.
+	 */
+	if (atomic_read(&ms->suspend)) {
+		if (dm_noflush_suspending(ms->ti))
+			bio_endio(bio, DM_ENDIO_REQUEUE);
+		else
+			bio_endio(bio, -EIO);
+		return;
+	}
+
+	/*
+	 * Hold bio until the suspend is complete.
+	 */
+	spin_lock_irq(&ms->lock);
+	bio_list_add(&ms->holds, bio);
+	spin_unlock_irq(&ms->lock);
+}
+
 /*-----------------------------------------------------------------
  * Reads
  *---------------------------------------------------------------*/
@@ -511,7 +581,6 @@ static void write_callback(unsigned long error, void *context)
 	unsigned i, ret = 0;
 	struct bio *bio = (struct bio *) context;
 	struct mirror_set *ms;
-	int uptodate = 0;
 	int should_wake = 0;
 	unsigned long flags;
 
@@ -524,36 +593,27 @@ static void write_callback(unsigned long error, void *context)
 	 * This way we handle both writes to SYNC and NOSYNC
 	 * regions with the same code.
 	 */
-	if (likely(!error))
-		goto out;
+	if (likely(!error)) {
+		bio_endio(bio, ret);
+		return;
+	}
 
 	for (i = 0; i < ms->nr_mirrors; i++)
 		if (test_bit(i, &error))
 			fail_mirror(ms->mirror + i, DM_RAID1_WRITE_ERROR);
-		else
-			uptodate = 1;
 
-	if (unlikely(!uptodate)) {
-		DMERR("All replicated volumes dead, failing I/O");
-		/* None of the writes succeeded, fail the I/O. */
-		ret = -EIO;
-	} else if (errors_handled(ms)) {
-		/*
-		 * Need to raise event.  Since raising
-		 * events can block, we need to do it in
-		 * the main thread.
-		 */
-		spin_lock_irqsave(&ms->lock, flags);
-		if (!ms->failures.head)
-			should_wake = 1;
-		bio_list_add(&ms->failures, bio);
-		spin_unlock_irqrestore(&ms->lock, flags);
-		if (should_wake)
-			wakeup_mirrord(ms);
-		return;
-	}
-out:
-	bio_endio(bio, ret);
+	/*
+	 * Need to raise event.  Since raising
+	 * events can block, we need to do it in
+	 * the main thread.
+	 */
+	spin_lock_irqsave(&ms->lock, flags);
+	if (!ms->failures.head)
+		should_wake = 1;
+	bio_list_add(&ms->failures, bio);
+	spin_unlock_irqrestore(&ms->lock, flags);
+	if (should_wake)
+		wakeup_mirrord(ms);
 }
 
 static void do_write(struct mirror_set *ms, struct bio *bio)
@@ -562,7 +622,7 @@ static void do_write(struct mirror_set *ms, struct bio *bio)
 	struct dm_io_region io[ms->nr_mirrors], *dest = io;
 	struct mirror *m;
 	struct dm_io_request io_req = {
-		.bi_rw = WRITE,
+		.bi_rw = WRITE | (bio->bi_rw & WRITE_BARRIER),
 		.mem.type = DM_IO_BVEC,
 		.mem.ptr.bvec = bio->bi_io_vec + bio->bi_idx,
 		.notify.fn = write_callback,
@@ -603,6 +663,11 @@ static void do_writes(struct mirror_set *ms, struct bio_list *writes)
 	bio_list_init(&requeue);
 
 	while ((bio = bio_list_pop(writes))) {
+		if (unlikely(bio_empty_barrier(bio))) {
+			bio_list_add(&sync, bio);
+			continue;
+		}
+
 		region = dm_rh_bio_to_region(ms->rh, bio);
 
 		if (log->type->is_remote_recovering &&
@@ -672,8 +737,12 @@ static void do_writes(struct mirror_set *ms, struct bio_list *writes)
 		dm_rh_delay(ms->rh, bio);
 
 	while ((bio = bio_list_pop(&nosync))) {
-		map_bio(get_default_mirror(ms), bio);
-		generic_make_request(bio);
+		if (unlikely(ms->leg_failure) && errors_handled(ms))
+			hold_bio(ms, bio);
+		else {
+			map_bio(get_default_mirror(ms), bio);
+			generic_make_request(bio);
+		}
 	}
 }
 
@@ -681,20 +750,12 @@ static void do_failures(struct mirror_set *ms, struct bio_list *failures)
 {
 	struct bio *bio;
 
-	if (!failures->head)
-		return;
-
-	if (!ms->log_failure) {
-		while ((bio = bio_list_pop(failures))) {
-			ms->in_sync = 0;
-			dm_rh_mark_nosync(ms->rh, bio, bio->bi_size, 0);
-		}
+	if (likely(!failures->head))
 		return;
-	}
 
 	/*
 	 * If the log has failed, unattempted writes are being
-	 * put on the failures list.  We can't issue those writes
+	 * put on the holds list.  We can't issue those writes
 	 * until a log has been marked, so we must store them.
 	 *
 	 * If a 'noflush' suspend is in progress, we can requeue
@@ -709,23 +770,27 @@ static void do_failures(struct mirror_set *ms, struct bio_list *failures)
 	 * for us to treat them the same and requeue them
 	 * as well.
 	 */
-	if (dm_noflush_suspending(ms->ti)) {
-		while ((bio = bio_list_pop(failures)))
-			bio_endio(bio, DM_ENDIO_REQUEUE);
-		return;
-	}
+	while ((bio = bio_list_pop(failures))) {
+		if (!ms->log_failure) {
+			ms->in_sync = 0;
+			dm_rh_mark_nosync(ms->rh, bio);
+		}
 
-	if (atomic_read(&ms->suspend)) {
-		while ((bio = bio_list_pop(failures)))
+		/*
+		 * If all the legs are dead, fail the I/O.
+		 * If we have been told to handle errors, hold the bio
+		 * and wait for userspace to deal with the problem.
+		 * Otherwise pretend that the I/O succeeded. (This would
+		 * be wrong if the failed leg returned after reboot and
+		 * got replicated back to the good legs.)
+		 */
+		if (!get_valid_mirror(ms))
 			bio_endio(bio, -EIO);
-		return;
+		else if (errors_handled(ms))
+			hold_bio(ms, bio);
+		else
+			bio_endio(bio, 0);
 	}
-
-	spin_lock_irq(&ms->lock);
-	bio_list_merge(&ms->failures, failures);
-	spin_unlock_irq(&ms->lock);
-
-	delayed_wake(ms);
 }
 
 static void trigger_event(struct work_struct *work)
@@ -784,12 +849,17 @@ static struct mirror_set *alloc_context(unsigned int nr_mirrors,
 	}
 
 	spin_lock_init(&ms->lock);
+	bio_list_init(&ms->reads);
+	bio_list_init(&ms->writes);
+	bio_list_init(&ms->failures);
+	bio_list_init(&ms->holds);
 
 	ms->ti = ti;
 	ms->nr_mirrors = nr_mirrors;
 	ms->nr_regions = dm_sector_div_up(ti->len, region_size);
 	ms->in_sync = 0;
 	ms->log_failure = 0;
+	ms->leg_failure = 0;
 	atomic_set(&ms->suspend, 0);
 	atomic_set(&ms->default_mirror, DEFAULT_MIRROR);
 
@@ -889,7 +959,8 @@ static struct dm_dirty_log *create_dirty_log(struct dm_target *ti,
 		return NULL;
 	}
 
-	dl = dm_dirty_log_create(argv[0], ti, param_count, argv + 2);
+	dl = dm_dirty_log_create(argv[0], ti, mirror_flush, param_count,
+				 argv + 2);
 	if (!dl) {
 		ti->error = "Error creating mirror dirty log";
 		return NULL;
@@ -995,6 +1066,7 @@ static int mirror_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 
 	ti->private = ms;
 	ti->split_io = dm_rh_get_region_size(ms->rh);
+	ti->num_flush_requests = 1;
 
 	ms->kmirrord_wq = create_singlethread_workqueue("kmirrord");
 	if (!ms->kmirrord_wq) {
@@ -1122,7 +1194,8 @@ static int mirror_end_io(struct dm_target *ti, struct bio *bio,
 	 * We need to dec pending if this was a write.
 	 */
 	if (rw == WRITE) {
-		dm_rh_dec(ms->rh, map_context->ll);
+		if (likely(!bio_empty_barrier(bio)))
+			dm_rh_dec(ms->rh, map_context->ll);
 		return error;
 	}
 
@@ -1180,6 +1253,9 @@ static void mirror_presuspend(struct dm_target *ti)
 	struct mirror_set *ms = (struct mirror_set *) ti->private;
 	struct dm_dirty_log *log = dm_rh_dirty_log(ms->rh);
 
+	struct bio_list holds;
+	struct bio *bio;
+
 	atomic_set(&ms->suspend, 1);
 
 	/*
@@ -1202,6 +1278,22 @@ static void mirror_presuspend(struct dm_target *ti)
 	 * we know that all of our I/O has been pushed.
 	 */
 	flush_workqueue(ms->kmirrord_wq);
+
+	/*
+	 * Now set ms->suspend is set and the workqueue flushed, no more
+	 * entries can be added to ms->hold list, so process it.
+	 *
+	 * Bios can still arrive concurrently with or after this
+	 * presuspend function, but they cannot join the hold list
+	 * because ms->suspend is set.
+	 */
+	spin_lock_irq(&ms->lock);
+	holds = ms->holds;
+	bio_list_init(&ms->holds);
+	spin_unlock_irq(&ms->lock);
+
+	while ((bio = bio_list_pop(&holds)))
+		hold_bio(ms, bio);
 }
 
 static void mirror_postsuspend(struct dm_target *ti)
@@ -1244,7 +1336,8 @@ static char device_status_char(struct mirror *m)
 	if (!atomic_read(&(m->error_count)))
 		return 'A';
 
-	return (test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
+	return (test_bit(DM_RAID1_FLUSH_ERROR, &(m->error_type))) ? 'F' :
+		(test_bit(DM_RAID1_WRITE_ERROR, &(m->error_type))) ? 'D' :
 		(test_bit(DM_RAID1_SYNC_ERROR, &(m->error_type))) ? 'S' :
 		(test_bit(DM_RAID1_READ_ERROR, &(m->error_type))) ? 'R' : 'U';
 }

drivers/md/dm-region-hash.c

@@ -79,6 +79,11 @@ struct dm_region_hash {
 	struct list_head recovered_regions;
 	struct list_head failed_recovered_regions;
 
+	/*
+	 * If there was a barrier failure no regions can be marked clean.
+	 */
+	int barrier_failure;
+
 	void *context;
 	sector_t target_begin;
 
@@ -211,6 +216,7 @@ struct dm_region_hash *dm_region_hash_create(
 	INIT_LIST_HEAD(&rh->quiesced_regions);
 	INIT_LIST_HEAD(&rh->recovered_regions);
 	INIT_LIST_HEAD(&rh->failed_recovered_regions);
+	rh->barrier_failure = 0;
 
 	rh->region_pool = mempool_create_kmalloc_pool(MIN_REGIONS,
 						      sizeof(struct dm_region));
@@ -377,8 +383,6 @@ static void complete_resync_work(struct dm_region *reg, int success)
 /* dm_rh_mark_nosync
  * @ms
  * @bio
- * @done
- * @error
  *
  * The bio was written on some mirror(s) but failed on other mirror(s).
  * We can successfully endio the bio but should avoid the region being
@@ -386,8 +390,7 @@ static void complete_resync_work(struct dm_region *reg, int success)
  *
  * This function is _not_ safe in interrupt context!
  */
-void dm_rh_mark_nosync(struct dm_region_hash *rh,
-		       struct bio *bio, unsigned done, int error)
+void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio)
 {
 	unsigned long flags;
 	struct dm_dirty_log *log = rh->log;
@@ -395,6 +398,11 @@ void dm_rh_mark_nosync(struct dm_region_hash *rh,
 	region_t region = dm_rh_bio_to_region(rh, bio);
 	int recovering = 0;
 
+	if (bio_empty_barrier(bio)) {
+		rh->barrier_failure = 1;
+		return;
+	}
+
 	/* We must inform the log that the sync count has changed. */
 	log->type->set_region_sync(log, region, 0);
 
@@ -419,7 +427,6 @@ void dm_rh_mark_nosync(struct dm_region_hash *rh,
 	BUG_ON(!list_empty(&reg->list));
 	spin_unlock_irqrestore(&rh->region_lock, flags);
 
-	bio_endio(bio, error);
 	if (recovering)
 		complete_resync_work(reg, 0);
 }
@@ -515,8 +522,11 @@ void dm_rh_inc_pending(struct dm_region_hash *rh, struct bio_list *bios)
 {
 	struct bio *bio;
 
-	for (bio = bios->head; bio; bio = bio->bi_next)
+	for (bio = bios->head; bio; bio = bio->bi_next) {
+		if (bio_empty_barrier(bio))
+			continue;
 		rh_inc(rh, dm_rh_bio_to_region(rh, bio));
+	}
 }
 EXPORT_SYMBOL_GPL(dm_rh_inc_pending);
 
@@ -544,7 +554,14 @@ void dm_rh_dec(struct dm_region_hash *rh, region_t region)
 		 */
 
 		/* do nothing for DM_RH_NOSYNC */
-		if (reg->state == DM_RH_RECOVERING) {
+		if (unlikely(rh->barrier_failure)) {
+			/*
+			 * If a write barrier failed some time ago, we
+			 * don't know whether or not this write made it
+			 * to the disk, so we must resync the device.
+			 */
+			reg->state = DM_RH_NOSYNC;
+		} else if (reg->state == DM_RH_RECOVERING) {
 			list_add_tail(&reg->list, &rh->quiesced_regions);
 		} else if (reg->state == DM_RH_DIRTY) {
 			reg->state = DM_RH_CLEAN;

drivers/md/dm-snap-persistent.c

@@ -55,6 +55,8 @@
  */
 #define SNAPSHOT_DISK_VERSION 1
 
+#define NUM_SNAPSHOT_HDR_CHUNKS 1
+
 struct disk_header {
 	uint32_t magic;
 
@@ -120,7 +122,22 @@ struct pstore {
 
 	/*
 	 * The next free chunk for an exception.
+	 *
+	 * When creating exceptions, all the chunks here and above are
+	 * free.  It holds the next chunk to be allocated.  On rare
+	 * occasions (e.g. after a system crash) holes can be left in
+	 * the exception store because chunks can be committed out of
+	 * order.
+	 *
+	 * When merging exceptions, it does not necessarily mean all the
+	 * chunks here and above are free.  It holds the value it would
+	 * have held if all chunks had been committed in order of
+	 * allocation.  Consequently the value may occasionally be
+	 * slightly too low, but since it's only used for 'status' and
+	 * it can never reach its minimum value too early this doesn't
+	 * matter.
 	 */
+
 	chunk_t next_free;
 
 	/*
@@ -214,7 +231,7 @@ static int chunk_io(struct pstore *ps, void *area, chunk_t chunk, int rw,
 		    int metadata)
 {
 	struct dm_io_region where = {
-		.bdev = ps->store->cow->bdev,
+		.bdev = dm_snap_cow(ps->store->snap)->bdev,
 		.sector = ps->store->chunk_size * chunk,
 		.count = ps->store->chunk_size,
 	};
@@ -294,7 +311,8 @@ static int read_header(struct pstore *ps, int *new_snapshot)
 	 */
 	if (!ps->store->chunk_size) {
 		ps->store->chunk_size = max(DM_CHUNK_SIZE_DEFAULT_SECTORS,
-		    bdev_logical_block_size(ps->store->cow->bdev) >> 9);
+		    bdev_logical_block_size(dm_snap_cow(ps->store->snap)->
+					    bdev) >> 9);
 		ps->store->chunk_mask = ps->store->chunk_size - 1;
 		ps->store->chunk_shift = ffs(ps->store->chunk_size) - 1;
 		chunk_size_supplied = 0;
@@ -408,6 +426,15 @@ static void write_exception(struct pstore *ps,
 	e->new_chunk = cpu_to_le64(de->new_chunk);
 }
 
+static void clear_exception(struct pstore *ps, uint32_t index)
+{
+	struct disk_exception *e = get_exception(ps, index);
+
+	/* clear it */
+	e->old_chunk = 0;
+	e->new_chunk = 0;
+}
+
 /*
  * Registers the exceptions that are present in the current area.
  * 'full' is filled in to indicate if the area has been
@@ -489,11 +516,23 @@ 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)
+static void persistent_usage(struct dm_exception_store *store,
+			     sector_t *total_sectors,
+			     sector_t *sectors_allocated,
+			     sector_t *metadata_sectors)
 {
-	*numerator = get_info(store)->next_free * store->chunk_size;
-	*denominator = get_dev_size(store->cow->bdev);
+	struct pstore *ps = get_info(store);
+
+	*sectors_allocated = ps->next_free * store->chunk_size;
+	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
+
+	/*
+	 * First chunk is the fixed header.
+	 * Then there are (ps->current_area + 1) metadata chunks, each one
+	 * separated from the next by ps->exceptions_per_area data chunks.
+	 */
+	*metadata_sectors = (ps->current_area + 1 + NUM_SNAPSHOT_HDR_CHUNKS) *
+			    store->chunk_size;
 }
 
 static void persistent_dtr(struct dm_exception_store *store)
@@ -552,44 +591,40 @@ static int persistent_read_metadata(struct dm_exception_store *store,
 		ps->current_area = 0;
 		zero_memory_area(ps);
 		r = zero_disk_area(ps, 0);
-		if (r) {
+		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;
-		}
+		return r;
+	}
+	/*
+	 * 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;
+	/*
+	 * Metadata are valid, but snapshot is invalidated
+	 */
+	if (!ps->valid)
+		return 1;
 
-		/*
-		 * Read the metadata.
-		 */
-		r = read_exceptions(ps, callback, callback_context);
-		if (r)
-			return r;
-	}
+	/*
+	 * Read the metadata.
+	 */
+	r = read_exceptions(ps, callback, callback_context);
 
-	return 0;
+	return r;
 }
 
 static int persistent_prepare_exception(struct dm_exception_store *store,
-					struct dm_snap_exception *e)
+					struct dm_exception *e)
 {
 	struct pstore *ps = get_info(store);
 	uint32_t stride;
 	chunk_t next_free;
-	sector_t size = get_dev_size(store->cow->bdev);
+	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
 
 	/* Is there enough room ? */
 	if (size < ((ps->next_free + 1) * store->chunk_size))
@@ -611,7 +646,7 @@ static int persistent_prepare_exception(struct dm_exception_store *store,
 }
 
 static void persistent_commit_exception(struct dm_exception_store *store,
-					struct dm_snap_exception *e,
+					struct dm_exception *e,
 					void (*callback) (void *, int success),
 					void *callback_context)
 {
@@ -672,6 +707,85 @@ static void persistent_commit_exception(struct dm_exception_store *store,
 	ps->callback_count = 0;
 }
 
+static int persistent_prepare_merge(struct dm_exception_store *store,
+				    chunk_t *last_old_chunk,
+				    chunk_t *last_new_chunk)
+{
+	struct pstore *ps = get_info(store);
+	struct disk_exception de;
+	int nr_consecutive;
+	int r;
+
+	/*
+	 * When current area is empty, move back to preceding area.
+	 */
+	if (!ps->current_committed) {
+		/*
+		 * Have we finished?
+		 */
+		if (!ps->current_area)
+			return 0;
+
+		ps->current_area--;
+		r = area_io(ps, READ);
+		if (r < 0)
+			return r;
+		ps->current_committed = ps->exceptions_per_area;
+	}
+
+	read_exception(ps, ps->current_committed - 1, &de);
+	*last_old_chunk = de.old_chunk;
+	*last_new_chunk = de.new_chunk;
+
+	/*
+	 * Find number of consecutive chunks within the current area,
+	 * working backwards.
+	 */
+	for (nr_consecutive = 1; nr_consecutive < ps->current_committed;
+	     nr_consecutive++) {
+		read_exception(ps, ps->current_committed - 1 - nr_consecutive,
+			       &de);
+		if (de.old_chunk != *last_old_chunk - nr_consecutive ||
+		    de.new_chunk != *last_new_chunk - nr_consecutive)
+			break;
+	}
+
+	return nr_consecutive;
+}
+
+static int persistent_commit_merge(struct dm_exception_store *store,
+				   int nr_merged)
+{
+	int r, i;
+	struct pstore *ps = get_info(store);
+
+	BUG_ON(nr_merged > ps->current_committed);
+
+	for (i = 0; i < nr_merged; i++)
+		clear_exception(ps, ps->current_committed - 1 - i);
+
+	r = area_io(ps, WRITE);
+	if (r < 0)
+		return r;
+
+	ps->current_committed -= nr_merged;
+
+	/*
+	 * At this stage, only persistent_usage() uses ps->next_free, so
+	 * we make no attempt to keep ps->next_free strictly accurate
+	 * as exceptions may have been committed out-of-order originally.
+	 * Once a snapshot has become merging, we set it to the value it
+	 * would have held had all the exceptions been committed in order.
+	 *
+	 * ps->current_area does not get reduced by prepare_merge() until
+	 * after commit_merge() has removed the nr_merged previous exceptions.
+	 */
+	ps->next_free = (area_location(ps, ps->current_area) - 1) +
+			(ps->current_committed + 1) + NUM_SNAPSHOT_HDR_CHUNKS;
+
+	return 0;
+}
+
 static void persistent_drop_snapshot(struct dm_exception_store *store)
 {
 	struct pstore *ps = get_info(store);
@@ -697,7 +811,7 @@ static int persistent_ctr(struct dm_exception_store *store,
 	ps->area = NULL;
 	ps->zero_area = NULL;
 	ps->header_area = NULL;
-	ps->next_free = 2;	/* skipping the header and first area */
+	ps->next_free = NUM_SNAPSHOT_HDR_CHUNKS + 1; /* header and 1st area */
 	ps->current_committed = 0;
 
 	ps->callback_count = 0;
@@ -726,8 +840,7 @@ static unsigned persistent_status(struct dm_exception_store *store,
 	case STATUSTYPE_INFO:
 		break;
 	case STATUSTYPE_TABLE:
-		DMEMIT(" %s P %llu", store->cow->name,
-		       (unsigned long long)store->chunk_size);
+		DMEMIT(" P %llu", (unsigned long long)store->chunk_size);
 	}
 
 	return sz;
@@ -741,8 +854,10 @@ static struct dm_exception_store_type _persistent_type = {
 	.read_metadata = persistent_read_metadata,
 	.prepare_exception = persistent_prepare_exception,
 	.commit_exception = persistent_commit_exception,
+	.prepare_merge = persistent_prepare_merge,
+	.commit_merge = persistent_commit_merge,
 	.drop_snapshot = persistent_drop_snapshot,
-	.fraction_full = persistent_fraction_full,
+	.usage = persistent_usage,
 	.status = persistent_status,
 };
 
@@ -754,8 +869,10 @@ static struct dm_exception_store_type _persistent_compat_type = {
 	.read_metadata = persistent_read_metadata,
 	.prepare_exception = persistent_prepare_exception,
 	.commit_exception = persistent_commit_exception,
+	.prepare_merge = persistent_prepare_merge,
+	.commit_merge = persistent_commit_merge,
 	.drop_snapshot = persistent_drop_snapshot,
-	.fraction_full = persistent_fraction_full,
+	.usage = persistent_usage,
 	.status = persistent_status,
 };
 

drivers/md/dm-snap-transient.c

@@ -36,10 +36,10 @@ static int transient_read_metadata(struct dm_exception_store *store,
 }
 
 static int transient_prepare_exception(struct dm_exception_store *store,
-				       struct dm_snap_exception *e)
+				       struct dm_exception *e)
 {
 	struct transient_c *tc = store->context;
-	sector_t size = get_dev_size(store->cow->bdev);
+	sector_t size = get_dev_size(dm_snap_cow(store->snap)->bdev);
 
 	if (size < (tc->next_free + store->chunk_size))
 		return -1;
@@ -51,7 +51,7 @@ static int transient_prepare_exception(struct dm_exception_store *store,
 }
 
 static void transient_commit_exception(struct dm_exception_store *store,
-				       struct dm_snap_exception *e,
+				       struct dm_exception *e,
 				       void (*callback) (void *, int success),
 				       void *callback_context)
 {
@@ -59,11 +59,14 @@ static void transient_commit_exception(struct dm_exception_store *store,
 	callback(callback_context, 1);
 }
 
-static void transient_fraction_full(struct dm_exception_store *store,
-				    sector_t *numerator, sector_t *denominator)
+static void transient_usage(struct dm_exception_store *store,
+			    sector_t *total_sectors,
+			    sector_t *sectors_allocated,
+			    sector_t *metadata_sectors)
 {
-	*numerator = ((struct transient_c *) store->context)->next_free;
-	*denominator = get_dev_size(store->cow->bdev);
+	*sectors_allocated = ((struct transient_c *) store->context)->next_free;
+	*total_sectors = get_dev_size(dm_snap_cow(store->snap)->bdev);
+	*metadata_sectors = 0;
 }
 
 static int transient_ctr(struct dm_exception_store *store,
@@ -91,8 +94,7 @@ static unsigned transient_status(struct dm_exception_store *store,
 	case STATUSTYPE_INFO:
 		break;
 	case STATUSTYPE_TABLE:
-		DMEMIT(" %s N %llu", store->cow->name,
-		       (unsigned long long)store->chunk_size);
+		DMEMIT(" N %llu", (unsigned long long)store->chunk_size);
 	}
 
 	return sz;
@@ -106,7 +108,7 @@ static struct dm_exception_store_type _transient_type = {
 	.read_metadata = transient_read_metadata,
 	.prepare_exception = transient_prepare_exception,
 	.commit_exception = transient_commit_exception,
-	.fraction_full = transient_fraction_full,
+	.usage = transient_usage,
 	.status = transient_status,
 };
 
@@ -118,7 +120,7 @@ static struct dm_exception_store_type _transient_compat_type = {
 	.read_metadata = transient_read_metadata,
 	.prepare_exception = transient_prepare_exception,
 	.commit_exception = transient_commit_exception,
-	.fraction_full = transient_fraction_full,
+	.usage = transient_usage,
 	.status = transient_status,
 };
 

drivers/md/dm-snap.c

@@ -25,6 +25,11 @@
 
 #define DM_MSG_PREFIX "snapshots"
 
+static const char dm_snapshot_merge_target_name[] = "snapshot-merge";
+
+#define dm_target_is_snapshot_merge(ti) \
+	((ti)->type->name == dm_snapshot_merge_target_name)
+
 /*
  * The percentage increment we will wake up users at
  */
@@ -49,7 +54,7 @@
 #define DM_TRACKED_CHUNK_HASH(x)	((unsigned long)(x) & \
 					 (DM_TRACKED_CHUNK_HASH_SIZE - 1))
 
-struct exception_table {
+struct dm_exception_table {
 	uint32_t hash_mask;
 	unsigned hash_shift;
 	struct list_head *table;
@@ -59,22 +64,31 @@ struct dm_snapshot {
 	struct rw_semaphore lock;
 
 	struct dm_dev *origin;
+	struct dm_dev *cow;
+
+	struct dm_target *ti;
 
 	/* List of snapshots per Origin */
 	struct list_head list;
 
-	/* You can't use a snapshot if this is 0 (e.g. if full) */
+	/*
+	 * You can't use a snapshot if this is 0 (e.g. if full).
+	 * A snapshot-merge target never clears this.
+	 */
 	int valid;
 
 	/* Origin writes don't trigger exceptions until this is set */
 	int active;
 
+	/* Whether or not owning mapped_device is suspended */
+	int suspended;
+
 	mempool_t *pending_pool;
 
 	atomic_t pending_exceptions_count;
 
-	struct exception_table pending;
-	struct exception_table complete;
+	struct dm_exception_table pending;
+	struct dm_exception_table complete;
 
 	/*
 	 * pe_lock protects all pending_exception operations and access
@@ -95,8 +109,51 @@ struct dm_snapshot {
 	mempool_t *tracked_chunk_pool;
 	spinlock_t tracked_chunk_lock;
 	struct hlist_head tracked_chunk_hash[DM_TRACKED_CHUNK_HASH_SIZE];
+
+	/*
+	 * The merge operation failed if this flag is set.
+	 * Failure modes are handled as follows:
+	 * - I/O error reading the header
+	 *   	=> don't load the target; abort.
+	 * - Header does not have "valid" flag set
+	 *   	=> use the origin; forget about the snapshot.
+	 * - I/O error when reading exceptions
+	 *   	=> don't load the target; abort.
+	 *         (We can't use the intermediate origin state.)
+	 * - I/O error while merging
+	 *	=> stop merging; set merge_failed; process I/O normally.
+	 */
+	int merge_failed;
+
+	/* Wait for events based on state_bits */
+	unsigned long state_bits;
+
+	/* Range of chunks currently being merged. */
+	chunk_t first_merging_chunk;
+	int num_merging_chunks;
+
+	/*
+	 * Incoming bios that overlap with chunks being merged must wait
+	 * for them to be committed.
+	 */
+	struct bio_list bios_queued_during_merge;
 };
 
+/*
+ * state_bits:
+ *   RUNNING_MERGE  - Merge operation is in progress.
+ *   SHUTDOWN_MERGE - Set to signal that merge needs to be stopped;
+ *                    cleared afterwards.
+ */
+#define RUNNING_MERGE          0
+#define SHUTDOWN_MERGE         1
+
+struct dm_dev *dm_snap_cow(struct dm_snapshot *s)
+{
+	return s->cow;
+}
+EXPORT_SYMBOL(dm_snap_cow);
+
 static struct workqueue_struct *ksnapd;
 static void flush_queued_bios(struct work_struct *work);
 
@@ -116,7 +173,7 @@ static int bdev_equal(struct block_device *lhs, struct block_device *rhs)
 }
 
 struct dm_snap_pending_exception {
-	struct dm_snap_exception e;
+	struct dm_exception e;
 
 	/*
 	 * Origin buffers waiting for this to complete are held
@@ -125,28 +182,6 @@ struct dm_snap_pending_exception {
 	struct bio_list origin_bios;
 	struct bio_list snapshot_bios;
 
-	/*
-	 * Short-term queue of pending exceptions prior to submission.
-	 */
-	struct list_head list;
-
-	/*
-	 * The primary pending_exception is the one that holds
-	 * the ref_count and the list of origin_bios for a
-	 * group of pending_exceptions.  It is always last to get freed.
-	 * These fields get set up when writing to the origin.
-	 */
-	struct dm_snap_pending_exception *primary_pe;
-
-	/*
-	 * Number of pending_exceptions processing this chunk.
-	 * When this drops to zero we must complete the origin bios.
-	 * If incrementing or decrementing this, hold pe->snap->lock for
-	 * the sibling concerned and not pe->primary_pe->snap->lock unless
-	 * they are the same.
-	 */
-	atomic_t ref_count;
-
 	/* Pointer back to snapshot context */
 	struct dm_snapshot *snap;
 
@@ -221,6 +256,16 @@ static int __chunk_is_tracked(struct dm_snapshot *s, chunk_t chunk)
 	return found;
 }
 
+/*
+ * This conflicting I/O is extremely improbable in the caller,
+ * so msleep(1) is sufficient and there is no need for a wait queue.
+ */
+static void __check_for_conflicting_io(struct dm_snapshot *s, chunk_t chunk)
+{
+	while (__chunk_is_tracked(s, chunk))
+		msleep(1);
+}
+
 /*
  * One of these per registered origin, held in the snapshot_origins hash
  */
@@ -243,6 +288,10 @@ struct origin {
 static struct list_head *_origins;
 static struct rw_semaphore _origins_lock;
 
+static DECLARE_WAIT_QUEUE_HEAD(_pending_exceptions_done);
+static DEFINE_SPINLOCK(_pending_exceptions_done_spinlock);
+static uint64_t _pending_exceptions_done_count;
+
 static int init_origin_hash(void)
 {
 	int i;
@@ -290,23 +339,145 @@ static void __insert_origin(struct origin *o)
 	list_add_tail(&o->hash_list, sl);
 }
 
+/*
+ * _origins_lock must be held when calling this function.
+ * Returns number of snapshots registered using the supplied cow device, plus:
+ * snap_src - a snapshot suitable for use as a source of exception handover
+ * snap_dest - a snapshot capable of receiving exception handover.
+ * snap_merge - an existing snapshot-merge target linked to the same origin.
+ *   There can be at most one snapshot-merge target. The parameter is optional.
+ *
+ * Possible return values and states of snap_src and snap_dest.
+ *   0: NULL, NULL  - first new snapshot
+ *   1: snap_src, NULL - normal snapshot
+ *   2: snap_src, snap_dest  - waiting for handover
+ *   2: snap_src, NULL - handed over, waiting for old to be deleted
+ *   1: NULL, snap_dest - source got destroyed without handover
+ */
+static int __find_snapshots_sharing_cow(struct dm_snapshot *snap,
+					struct dm_snapshot **snap_src,
+					struct dm_snapshot **snap_dest,
+					struct dm_snapshot **snap_merge)
+{
+	struct dm_snapshot *s;
+	struct origin *o;
+	int count = 0;
+	int active;
+
+	o = __lookup_origin(snap->origin->bdev);
+	if (!o)
+		goto out;
+
+	list_for_each_entry(s, &o->snapshots, list) {
+		if (dm_target_is_snapshot_merge(s->ti) && snap_merge)
+			*snap_merge = s;
+		if (!bdev_equal(s->cow->bdev, snap->cow->bdev))
+			continue;
+
+		down_read(&s->lock);
+		active = s->active;
+		up_read(&s->lock);
+
+		if (active) {
+			if (snap_src)
+				*snap_src = s;
+		} else if (snap_dest)
+			*snap_dest = s;
+
+		count++;
+	}
+
+out:
+	return count;
+}
+
+/*
+ * On success, returns 1 if this snapshot is a handover destination,
+ * otherwise returns 0.
+ */
+static int __validate_exception_handover(struct dm_snapshot *snap)
+{
+	struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
+	struct dm_snapshot *snap_merge = NULL;
+
+	/* Does snapshot need exceptions handed over to it? */
+	if ((__find_snapshots_sharing_cow(snap, &snap_src, &snap_dest,
+					  &snap_merge) == 2) ||
+	    snap_dest) {
+		snap->ti->error = "Snapshot cow pairing for exception "
+				  "table handover failed";
+		return -EINVAL;
+	}
+
+	/*
+	 * If no snap_src was found, snap cannot become a handover
+	 * destination.
+	 */
+	if (!snap_src)
+		return 0;
+
+	/*
+	 * Non-snapshot-merge handover?
+	 */
+	if (!dm_target_is_snapshot_merge(snap->ti))
+		return 1;
+
+	/*
+	 * Do not allow more than one merging snapshot.
+	 */
+	if (snap_merge) {
+		snap->ti->error = "A snapshot is already merging.";
+		return -EINVAL;
+	}
+
+	if (!snap_src->store->type->prepare_merge ||
+	    !snap_src->store->type->commit_merge) {
+		snap->ti->error = "Snapshot exception store does not "
+				  "support snapshot-merge.";
+		return -EINVAL;
+	}
+
+	return 1;
+}
+
+static void __insert_snapshot(struct origin *o, struct dm_snapshot *s)
+{
+	struct dm_snapshot *l;
+
+	/* Sort the list according to chunk size, largest-first smallest-last */
+	list_for_each_entry(l, &o->snapshots, list)
+		if (l->store->chunk_size < s->store->chunk_size)
+			break;
+	list_add_tail(&s->list, &l->list);
+}
+
 /*
  * Make a note of the snapshot and its origin so we can look it
  * up when the origin has a write on it.
+ *
+ * Also validate snapshot exception store handovers.
+ * On success, returns 1 if this registration is a handover destination,
+ * otherwise returns 0.
  */
 static int register_snapshot(struct dm_snapshot *snap)
 {
-	struct dm_snapshot *l;
-	struct origin *o, *new_o;
+	struct origin *o, *new_o = NULL;
 	struct block_device *bdev = snap->origin->bdev;
+	int r = 0;
 
 	new_o = kmalloc(sizeof(*new_o), GFP_KERNEL);
 	if (!new_o)
 		return -ENOMEM;
 
 	down_write(&_origins_lock);
-	o = __lookup_origin(bdev);
 
+	r = __validate_exception_handover(snap);
+	if (r < 0) {
+		kfree(new_o);
+		goto out;
+	}
+
+	o = __lookup_origin(bdev);
 	if (o)
 		kfree(new_o);
 	else {
@@ -320,14 +491,27 @@ static int register_snapshot(struct dm_snapshot *snap)
 		__insert_origin(o);
 	}
 
-	/* Sort the list according to chunk size, largest-first smallest-last */
-	list_for_each_entry(l, &o->snapshots, list)
-		if (l->store->chunk_size < snap->store->chunk_size)
-			break;
-	list_add_tail(&snap->list, &l->list);
+	__insert_snapshot(o, snap);
+
+out:
+	up_write(&_origins_lock);
+
+	return r;
+}
+
+/*
+ * Move snapshot to correct place in list according to chunk size.
+ */
+static void reregister_snapshot(struct dm_snapshot *s)
+{
+	struct block_device *bdev = s->origin->bdev;
+
+	down_write(&_origins_lock);
+
+	list_del(&s->list);
+	__insert_snapshot(__lookup_origin(bdev), s);
 
 	up_write(&_origins_lock);
-	return 0;
 }
 
 static void unregister_snapshot(struct dm_snapshot *s)
@@ -338,7 +522,7 @@ static void unregister_snapshot(struct dm_snapshot *s)
 	o = __lookup_origin(s->origin->bdev);
 
 	list_del(&s->list);
-	if (list_empty(&o->snapshots)) {
+	if (o && list_empty(&o->snapshots)) {
 		list_del(&o->hash_list);
 		kfree(o);
 	}
@@ -351,8 +535,8 @@ static void unregister_snapshot(struct dm_snapshot *s)
  * The lowest hash_shift bits of the chunk number are ignored, allowing
  * some consecutive chunks to be grouped together.
  */
-static int init_exception_table(struct exception_table *et, uint32_t size,
-				unsigned hash_shift)
+static int dm_exception_table_init(struct dm_exception_table *et,
+				   uint32_t size, unsigned hash_shift)
 {
 	unsigned int i;
 
@@ -368,10 +552,11 @@ static int init_exception_table(struct exception_table *et, uint32_t size,
 	return 0;
 }
 
-static void exit_exception_table(struct exception_table *et, struct kmem_cache *mem)
+static void dm_exception_table_exit(struct dm_exception_table *et,
+				    struct kmem_cache *mem)
 {
 	struct list_head *slot;
-	struct dm_snap_exception *ex, *next;
+	struct dm_exception *ex, *next;
 	int i, size;
 
 	size = et->hash_mask + 1;
@@ -385,19 +570,12 @@ static void exit_exception_table(struct exception_table *et, struct kmem_cache *
 	vfree(et->table);
 }
 
-static uint32_t exception_hash(struct exception_table *et, chunk_t chunk)
+static uint32_t exception_hash(struct dm_exception_table *et, chunk_t chunk)
 {
 	return (chunk >> et->hash_shift) & et->hash_mask;
 }
 
-static void insert_exception(struct exception_table *eh,
-			     struct dm_snap_exception *e)
-{
-	struct list_head *l = &eh->table[exception_hash(eh, e->old_chunk)];
-	list_add(&e->hash_list, l);
-}
-
-static void remove_exception(struct dm_snap_exception *e)
+static void dm_remove_exception(struct dm_exception *e)
 {
 	list_del(&e->hash_list);
 }
@@ -406,11 +584,11 @@ static void remove_exception(struct dm_snap_exception *e)
  * Return the exception data for a sector, or NULL if not
  * remapped.
  */
-static struct dm_snap_exception *lookup_exception(struct exception_table *et,
-						  chunk_t chunk)
+static struct dm_exception *dm_lookup_exception(struct dm_exception_table *et,
+						chunk_t chunk)
 {
 	struct list_head *slot;
-	struct dm_snap_exception *e;
+	struct dm_exception *e;
 
 	slot = &et->table[exception_hash(et, chunk)];
 	list_for_each_entry (e, slot, hash_list)
@@ -421,9 +599,9 @@ static struct dm_snap_exception *lookup_exception(struct exception_table *et,
 	return NULL;
 }
 
-static struct dm_snap_exception *alloc_exception(void)
+static struct dm_exception *alloc_completed_exception(void)
 {
-	struct dm_snap_exception *e;
+	struct dm_exception *e;
 
 	e = kmem_cache_alloc(exception_cache, GFP_NOIO);
 	if (!e)
@@ -432,7 +610,7 @@ static struct dm_snap_exception *alloc_exception(void)
 	return e;
 }
 
-static void free_exception(struct dm_snap_exception *e)
+static void free_completed_exception(struct dm_exception *e)
 {
 	kmem_cache_free(exception_cache, e);
 }
@@ -457,12 +635,11 @@ static void free_pending_exception(struct dm_snap_pending_exception *pe)
 	atomic_dec(&s->pending_exceptions_count);
 }
 
-static void insert_completed_exception(struct dm_snapshot *s,
-				       struct dm_snap_exception *new_e)
+static void dm_insert_exception(struct dm_exception_table *eh,
+				struct dm_exception *new_e)
 {
-	struct exception_table *eh = &s->complete;
 	struct list_head *l;
-	struct dm_snap_exception *e = NULL;
+	struct dm_exception *e = NULL;
 
 	l = &eh->table[exception_hash(eh, new_e->old_chunk)];
 
@@ -478,7 +655,7 @@ static void insert_completed_exception(struct dm_snapshot *s,
 		    new_e->new_chunk == (dm_chunk_number(e->new_chunk) +
 					 dm_consecutive_chunk_count(e) + 1)) {
 			dm_consecutive_chunk_count_inc(e);
-			free_exception(new_e);
+			free_completed_exception(new_e);
 			return;
 		}
 
@@ -488,7 +665,7 @@ static void insert_completed_exception(struct dm_snapshot *s,
 			dm_consecutive_chunk_count_inc(e);
 			e->old_chunk--;
 			e->new_chunk--;
-			free_exception(new_e);
+			free_completed_exception(new_e);
 			return;
 		}
 
@@ -507,9 +684,9 @@ static void insert_completed_exception(struct dm_snapshot *s,
 static int dm_add_exception(void *context, chunk_t old, chunk_t new)
 {
 	struct dm_snapshot *s = context;
-	struct dm_snap_exception *e;
+	struct dm_exception *e;
 
-	e = alloc_exception();
+	e = alloc_completed_exception();
 	if (!e)
 		return -ENOMEM;
 
@@ -518,11 +695,30 @@ static int dm_add_exception(void *context, chunk_t old, chunk_t new)
 	/* Consecutive_count is implicitly initialised to zero */
 	e->new_chunk = new;
 
-	insert_completed_exception(s, e);
+	dm_insert_exception(&s->complete, e);
 
 	return 0;
 }
 
+#define min_not_zero(l, r) (((l) == 0) ? (r) : (((r) == 0) ? (l) : min(l, r)))
+
+/*
+ * Return a minimum chunk size of all snapshots that have the specified origin.
+ * Return zero if the origin has no snapshots.
+ */
+static sector_t __minimum_chunk_size(struct origin *o)
+{
+	struct dm_snapshot *snap;
+	unsigned chunk_size = 0;
+
+	if (o)
+		list_for_each_entry(snap, &o->snapshots, list)
+			chunk_size = min_not_zero(chunk_size,
+						  snap->store->chunk_size);
+
+	return chunk_size;
+}
+
 /*
  * Hard coded magic.
  */
@@ -546,16 +742,18 @@ static int init_hash_tables(struct dm_snapshot *s)
 	 * Calculate based on the size of the original volume or
 	 * the COW volume...
 	 */
-	cow_dev_size = get_dev_size(s->store->cow->bdev);
+	cow_dev_size = get_dev_size(s->cow->bdev);
 	origin_dev_size = get_dev_size(s->origin->bdev);
 	max_buckets = calc_max_buckets();
 
 	hash_size = min(origin_dev_size, cow_dev_size) >> s->store->chunk_shift;
 	hash_size = min(hash_size, max_buckets);
 
+	if (hash_size < 64)
+		hash_size = 64;
 	hash_size = rounddown_pow_of_two(hash_size);
-	if (init_exception_table(&s->complete, hash_size,
-				 DM_CHUNK_CONSECUTIVE_BITS))
+	if (dm_exception_table_init(&s->complete, hash_size,
+				    DM_CHUNK_CONSECUTIVE_BITS))
 		return -ENOMEM;
 
 	/*
@@ -566,14 +764,284 @@ static int init_hash_tables(struct dm_snapshot *s)
 	if (hash_size < 64)
 		hash_size = 64;
 
-	if (init_exception_table(&s->pending, hash_size, 0)) {
-		exit_exception_table(&s->complete, exception_cache);
+	if (dm_exception_table_init(&s->pending, hash_size, 0)) {
+		dm_exception_table_exit(&s->complete, exception_cache);
 		return -ENOMEM;
 	}
 
 	return 0;
 }
 
+static void merge_shutdown(struct dm_snapshot *s)
+{
+	clear_bit_unlock(RUNNING_MERGE, &s->state_bits);
+	smp_mb__after_clear_bit();
+	wake_up_bit(&s->state_bits, RUNNING_MERGE);
+}
+
+static struct bio *__release_queued_bios_after_merge(struct dm_snapshot *s)
+{
+	s->first_merging_chunk = 0;
+	s->num_merging_chunks = 0;
+
+	return bio_list_get(&s->bios_queued_during_merge);
+}
+
+/*
+ * Remove one chunk from the index of completed exceptions.
+ */
+static int __remove_single_exception_chunk(struct dm_snapshot *s,
+					   chunk_t old_chunk)
+{
+	struct dm_exception *e;
+
+	e = dm_lookup_exception(&s->complete, old_chunk);
+	if (!e) {
+		DMERR("Corruption detected: exception for block %llu is "
+		      "on disk but not in memory",
+		      (unsigned long long)old_chunk);
+		return -EINVAL;
+	}
+
+	/*
+	 * If this is the only chunk using this exception, remove exception.
+	 */
+	if (!dm_consecutive_chunk_count(e)) {
+		dm_remove_exception(e);
+		free_completed_exception(e);
+		return 0;
+	}
+
+	/*
+	 * The chunk may be either at the beginning or the end of a
+	 * group of consecutive chunks - never in the middle.  We are
+	 * removing chunks in the opposite order to that in which they
+	 * were added, so this should always be true.
+	 * Decrement the consecutive chunk counter and adjust the
+	 * starting point if necessary.
+	 */
+	if (old_chunk == e->old_chunk) {
+		e->old_chunk++;
+		e->new_chunk++;
+	} else if (old_chunk != e->old_chunk +
+		   dm_consecutive_chunk_count(e)) {
+		DMERR("Attempt to merge block %llu from the "
+		      "middle of a chunk range [%llu - %llu]",
+		      (unsigned long long)old_chunk,
+		      (unsigned long long)e->old_chunk,
+		      (unsigned long long)
+		      e->old_chunk + dm_consecutive_chunk_count(e));
+		return -EINVAL;
+	}
+
+	dm_consecutive_chunk_count_dec(e);
+
+	return 0;
+}
+
+static void flush_bios(struct bio *bio);
+
+static int remove_single_exception_chunk(struct dm_snapshot *s)
+{
+	struct bio *b = NULL;
+	int r;
+	chunk_t old_chunk = s->first_merging_chunk + s->num_merging_chunks - 1;
+
+	down_write(&s->lock);
+
+	/*
+	 * Process chunks (and associated exceptions) in reverse order
+	 * so that dm_consecutive_chunk_count_dec() accounting works.
+	 */
+	do {
+		r = __remove_single_exception_chunk(s, old_chunk);
+		if (r)
+			goto out;
+	} while (old_chunk-- > s->first_merging_chunk);
+
+	b = __release_queued_bios_after_merge(s);
+
+out:
+	up_write(&s->lock);
+	if (b)
+		flush_bios(b);
+
+	return r;
+}
+
+static int origin_write_extent(struct dm_snapshot *merging_snap,
+			       sector_t sector, unsigned chunk_size);
+
+static void merge_callback(int read_err, unsigned long write_err,
+			   void *context);
+
+static uint64_t read_pending_exceptions_done_count(void)
+{
+	uint64_t pending_exceptions_done;
+
+	spin_lock(&_pending_exceptions_done_spinlock);
+	pending_exceptions_done = _pending_exceptions_done_count;
+	spin_unlock(&_pending_exceptions_done_spinlock);
+
+	return pending_exceptions_done;
+}
+
+static void increment_pending_exceptions_done_count(void)
+{
+	spin_lock(&_pending_exceptions_done_spinlock);
+	_pending_exceptions_done_count++;
+	spin_unlock(&_pending_exceptions_done_spinlock);
+
+	wake_up_all(&_pending_exceptions_done);
+}
+
+static void snapshot_merge_next_chunks(struct dm_snapshot *s)
+{
+	int i, linear_chunks;
+	chunk_t old_chunk, new_chunk;
+	struct dm_io_region src, dest;
+	sector_t io_size;
+	uint64_t previous_count;
+
+	BUG_ON(!test_bit(RUNNING_MERGE, &s->state_bits));
+	if (unlikely(test_bit(SHUTDOWN_MERGE, &s->state_bits)))
+		goto shut;
+
+	/*
+	 * valid flag never changes during merge, so no lock required.
+	 */
+	if (!s->valid) {
+		DMERR("Snapshot is invalid: can't merge");
+		goto shut;
+	}
+
+	linear_chunks = s->store->type->prepare_merge(s->store, &old_chunk,
+						      &new_chunk);
+	if (linear_chunks <= 0) {
+		if (linear_chunks < 0) {
+			DMERR("Read error in exception store: "
+			      "shutting down merge");
+			down_write(&s->lock);
+			s->merge_failed = 1;
+			up_write(&s->lock);
+		}
+		goto shut;
+	}
+
+	/* Adjust old_chunk and new_chunk to reflect start of linear region */
+	old_chunk = old_chunk + 1 - linear_chunks;
+	new_chunk = new_chunk + 1 - linear_chunks;
+
+	/*
+	 * Use one (potentially large) I/O to copy all 'linear_chunks'
+	 * from the exception store to the origin
+	 */
+	io_size = linear_chunks * s->store->chunk_size;
+
+	dest.bdev = s->origin->bdev;
+	dest.sector = chunk_to_sector(s->store, old_chunk);
+	dest.count = min(io_size, get_dev_size(dest.bdev) - dest.sector);
+
+	src.bdev = s->cow->bdev;
+	src.sector = chunk_to_sector(s->store, new_chunk);
+	src.count = dest.count;
+
+	/*
+	 * Reallocate any exceptions needed in other snapshots then
+	 * wait for the pending exceptions to complete.
+	 * Each time any pending exception (globally on the system)
+	 * completes we are woken and repeat the process to find out
+	 * if we can proceed.  While this may not seem a particularly
+	 * efficient algorithm, it is not expected to have any
+	 * significant impact on performance.
+	 */
+	previous_count = read_pending_exceptions_done_count();
+	while (origin_write_extent(s, dest.sector, io_size)) {
+		wait_event(_pending_exceptions_done,
+			   (read_pending_exceptions_done_count() !=
+			    previous_count));
+		/* Retry after the wait, until all exceptions are done. */
+		previous_count = read_pending_exceptions_done_count();
+	}
+
+	down_write(&s->lock);
+	s->first_merging_chunk = old_chunk;
+	s->num_merging_chunks = linear_chunks;
+	up_write(&s->lock);
+
+	/* Wait until writes to all 'linear_chunks' drain */
+	for (i = 0; i < linear_chunks; i++)
+		__check_for_conflicting_io(s, old_chunk + i);
+
+	dm_kcopyd_copy(s->kcopyd_client, &src, 1, &dest, 0, merge_callback, s);
+	return;
+
+shut:
+	merge_shutdown(s);
+}
+
+static void error_bios(struct bio *bio);
+
+static void merge_callback(int read_err, unsigned long write_err, void *context)
+{
+	struct dm_snapshot *s = context;
+	struct bio *b = NULL;
+
+	if (read_err || write_err) {
+		if (read_err)
+			DMERR("Read error: shutting down merge.");
+		else
+			DMERR("Write error: shutting down merge.");
+		goto shut;
+	}
+
+	if (s->store->type->commit_merge(s->store,
+					 s->num_merging_chunks) < 0) {
+		DMERR("Write error in exception store: shutting down merge");
+		goto shut;
+	}
+
+	if (remove_single_exception_chunk(s) < 0)
+		goto shut;
+
+	snapshot_merge_next_chunks(s);
+
+	return;
+
+shut:
+	down_write(&s->lock);
+	s->merge_failed = 1;
+	b = __release_queued_bios_after_merge(s);
+	up_write(&s->lock);
+	error_bios(b);
+
+	merge_shutdown(s);
+}
+
+static void start_merge(struct dm_snapshot *s)
+{
+	if (!test_and_set_bit(RUNNING_MERGE, &s->state_bits))
+		snapshot_merge_next_chunks(s);
+}
+
+static int wait_schedule(void *ptr)
+{
+	schedule();
+
+	return 0;
+}
+
+/*
+ * Stop the merging process and wait until it finishes.
+ */
+static void stop_merge(struct dm_snapshot *s)
+{
+	set_bit(SHUTDOWN_MERGE, &s->state_bits);
+	wait_on_bit(&s->state_bits, RUNNING_MERGE, wait_schedule,
+		    TASK_UNINTERRUPTIBLE);
+	clear_bit(SHUTDOWN_MERGE, &s->state_bits);
+}
+
 /*
  * Construct a snapshot mapping: <origin_dev> <COW-dev> <p/n> <chunk-size>
  */
@@ -582,50 +1050,73 @@ static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	struct dm_snapshot *s;
 	int i;
 	int r = -EINVAL;
-	char *origin_path;
-	struct dm_exception_store *store;
-	unsigned args_used;
+	char *origin_path, *cow_path;
+	unsigned args_used, num_flush_requests = 1;
+	fmode_t origin_mode = FMODE_READ;
 
 	if (argc != 4) {
 		ti->error = "requires exactly 4 arguments";
 		r = -EINVAL;
-		goto bad_args;
+		goto bad;
+	}
+
+	if (dm_target_is_snapshot_merge(ti)) {
+		num_flush_requests = 2;
+		origin_mode = FMODE_WRITE;
 	}
 
 	origin_path = argv[0];
 	argv++;
 	argc--;
 
-	r = dm_exception_store_create(ti, argc, argv, &args_used, &store);
+	s = kmalloc(sizeof(*s), GFP_KERNEL);
+	if (!s) {
+		ti->error = "Cannot allocate snapshot context private "
+		    "structure";
+		r = -ENOMEM;
+		goto bad;
+	}
+
+	cow_path = argv[0];
+	argv++;
+	argc--;
+
+	r = dm_get_device(ti, cow_path, 0, 0,
+			  FMODE_READ | FMODE_WRITE, &s->cow);
+	if (r) {
+		ti->error = "Cannot get COW device";
+		goto bad_cow;
+	}
+
+	r = dm_exception_store_create(ti, argc, argv, s, &args_used, &s->store);
 	if (r) {
 		ti->error = "Couldn't create exception store";
 		r = -EINVAL;
-		goto bad_args;
+		goto bad_store;
 	}
 
 	argv += args_used;
 	argc -= args_used;
 
-	s = kmalloc(sizeof(*s), GFP_KERNEL);
-	if (!s) {
-		ti->error = "Cannot allocate snapshot context private "
-		    "structure";
-		r = -ENOMEM;
-		goto bad_snap;
-	}
-
-	r = dm_get_device(ti, origin_path, 0, ti->len, FMODE_READ, &s->origin);
+	r = dm_get_device(ti, origin_path, 0, ti->len, origin_mode, &s->origin);
 	if (r) {
 		ti->error = "Cannot get origin device";
 		goto bad_origin;
 	}
 
-	s->store = store;
+	s->ti = ti;
 	s->valid = 1;
 	s->active = 0;
+	s->suspended = 0;
 	atomic_set(&s->pending_exceptions_count, 0);
 	init_rwsem(&s->lock);
+	INIT_LIST_HEAD(&s->list);
 	spin_lock_init(&s->pe_lock);
+	s->state_bits = 0;
+	s->merge_failed = 0;
+	s->first_merging_chunk = 0;
+	s->num_merging_chunks = 0;
+	bio_list_init(&s->bios_queued_during_merge);
 
 	/* Allocate hash table for COW data */
 	if (init_hash_tables(s)) {
@@ -659,39 +1150,55 @@ static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 
 	spin_lock_init(&s->tracked_chunk_lock);
 
-	/* Metadata must only be loaded into one table at once */
+	bio_list_init(&s->queued_bios);
+	INIT_WORK(&s->queued_bios_work, flush_queued_bios);
+
+	ti->private = s;
+	ti->num_flush_requests = num_flush_requests;
+
+	/* Add snapshot to the list of snapshots for this origin */
+	/* Exceptions aren't triggered till snapshot_resume() is called */
+	r = register_snapshot(s);
+	if (r == -ENOMEM) {
+		ti->error = "Snapshot origin struct allocation failed";
+		goto bad_load_and_register;
+	} else if (r < 0) {
+		/* invalid handover, register_snapshot has set ti->error */
+		goto bad_load_and_register;
+	}
+
+	/*
+	 * Metadata must only be loaded into one table at once, so skip this
+	 * if metadata will be handed over during resume.
+	 * Chunk size will be set during the handover - set it to zero to
+	 * ensure it's ignored.
+	 */
+	if (r > 0) {
+		s->store->chunk_size = 0;
+		return 0;
+	}
+
 	r = s->store->type->read_metadata(s->store, dm_add_exception,
 					  (void *)s);
 	if (r < 0) {
 		ti->error = "Failed to read snapshot metadata";
-		goto bad_load_and_register;
+		goto bad_read_metadata;
 	} else if (r > 0) {
 		s->valid = 0;
 		DMWARN("Snapshot is marked invalid.");
 	}
 
-	bio_list_init(&s->queued_bios);
-	INIT_WORK(&s->queued_bios_work, flush_queued_bios);
-
 	if (!s->store->chunk_size) {
 		ti->error = "Chunk size not set";
-		goto bad_load_and_register;
-	}
-
-	/* Add snapshot to the list of snapshots for this origin */
-	/* Exceptions aren't triggered till snapshot_resume() is called */
-	if (register_snapshot(s)) {
-		r = -EINVAL;
-		ti->error = "Cannot register snapshot origin";
-		goto bad_load_and_register;
+		goto bad_read_metadata;
 	}
-
-	ti->private = s;
 	ti->split_io = s->store->chunk_size;
-	ti->num_flush_requests = 1;
 
 	return 0;
 
+bad_read_metadata:
+	unregister_snapshot(s);
+
 bad_load_and_register:
 	mempool_destroy(s->tracked_chunk_pool);
 
@@ -702,19 +1209,22 @@ static int snapshot_ctr(struct dm_target *ti, unsigned int argc, char **argv)
 	dm_kcopyd_client_destroy(s->kcopyd_client);
 
 bad_kcopyd:
-	exit_exception_table(&s->pending, pending_cache);
-	exit_exception_table(&s->complete, exception_cache);
+	dm_exception_table_exit(&s->pending, pending_cache);
+	dm_exception_table_exit(&s->complete, exception_cache);
 
 bad_hash_tables:
 	dm_put_device(ti, s->origin);
 
 bad_origin:
-	kfree(s);
+	dm_exception_store_destroy(s->store);
 
-bad_snap:
-	dm_exception_store_destroy(store);
+bad_store:
+	dm_put_device(ti, s->cow);
+
+bad_cow:
+	kfree(s);
 
-bad_args:
+bad:
 	return r;
 }
 
@@ -723,8 +1233,39 @@ static void __free_exceptions(struct dm_snapshot *s)
 	dm_kcopyd_client_destroy(s->kcopyd_client);
 	s->kcopyd_client = NULL;
 
-	exit_exception_table(&s->pending, pending_cache);
-	exit_exception_table(&s->complete, exception_cache);
+	dm_exception_table_exit(&s->pending, pending_cache);
+	dm_exception_table_exit(&s->complete, exception_cache);
+}
+
+static void __handover_exceptions(struct dm_snapshot *snap_src,
+				  struct dm_snapshot *snap_dest)
+{
+	union {
+		struct dm_exception_table table_swap;
+		struct dm_exception_store *store_swap;
+	} u;
+
+	/*
+	 * Swap all snapshot context information between the two instances.
+	 */
+	u.table_swap = snap_dest->complete;
+	snap_dest->complete = snap_src->complete;
+	snap_src->complete = u.table_swap;
+
+	u.store_swap = snap_dest->store;
+	snap_dest->store = snap_src->store;
+	snap_src->store = u.store_swap;
+
+	snap_dest->store->snap = snap_dest;
+	snap_src->store->snap = snap_src;
+
+	snap_dest->ti->split_io = snap_dest->store->chunk_size;
+	snap_dest->valid = snap_src->valid;
+
+	/*
+	 * Set source invalid to ensure it receives no further I/O.
+	 */
+	snap_src->valid = 0;
 }
 
 static void snapshot_dtr(struct dm_target *ti)
@@ -733,9 +1274,24 @@ static void snapshot_dtr(struct dm_target *ti)
 	int i;
 #endif
 	struct dm_snapshot *s = ti->private;
+	struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
 
 	flush_workqueue(ksnapd);
 
+	down_read(&_origins_lock);
+	/* Check whether exception handover must be cancelled */
+	(void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
+	if (snap_src && snap_dest && (s == snap_src)) {
+		down_write(&snap_dest->lock);
+		snap_dest->valid = 0;
+		up_write(&snap_dest->lock);
+		DMERR("Cancelling snapshot handover.");
+	}
+	up_read(&_origins_lock);
+
+	if (dm_target_is_snapshot_merge(ti))
+		stop_merge(s);
+
 	/* Prevent further origin writes from using this snapshot. */
 	/* After this returns there can be no new kcopyd jobs. */
 	unregister_snapshot(s);
@@ -763,6 +1319,8 @@ static void snapshot_dtr(struct dm_target *ti)
 
 	dm_exception_store_destroy(s->store);
 
+	dm_put_device(ti, s->cow);
+
 	kfree(s);
 }
 
@@ -795,6 +1353,26 @@ static void flush_queued_bios(struct work_struct *work)
 	flush_bios(queued_bios);
 }
 
+static int do_origin(struct dm_dev *origin, struct bio *bio);
+
+/*
+ * Flush a list of buffers.
+ */
+static void retry_origin_bios(struct dm_snapshot *s, struct bio *bio)
+{
+	struct bio *n;
+	int r;
+
+	while (bio) {
+		n = bio->bi_next;
+		bio->bi_next = NULL;
+		r = do_origin(s->origin, bio);
+		if (r == DM_MAPIO_REMAPPED)
+			generic_make_request(bio);
+		bio = n;
+	}
+}
+
 /*
  * Error a list of buffers.
  */
@@ -825,45 +1403,12 @@ static void __invalidate_snapshot(struct dm_snapshot *s, int err)
 
 	s->valid = 0;
 
-	dm_table_event(s->store->ti->table);
-}
-
-static void get_pending_exception(struct dm_snap_pending_exception *pe)
-{
-	atomic_inc(&pe->ref_count);
-}
-
-static struct bio *put_pending_exception(struct dm_snap_pending_exception *pe)
-{
-	struct dm_snap_pending_exception *primary_pe;
-	struct bio *origin_bios = NULL;
-
-	primary_pe = pe->primary_pe;
-
-	/*
-	 * If this pe is involved in a write to the origin and
-	 * it is the last sibling to complete then release
-	 * the bios for the original write to the origin.
-	 */
-	if (primary_pe &&
-	    atomic_dec_and_test(&primary_pe->ref_count)) {
-		origin_bios = bio_list_get(&primary_pe->origin_bios);
-		free_pending_exception(primary_pe);
-	}
-
-	/*
-	 * Free the pe if it's not linked to an origin write or if
-	 * it's not itself a primary pe.
-	 */
-	if (!primary_pe || primary_pe != pe)
-		free_pending_exception(pe);
-
-	return origin_bios;
+	dm_table_event(s->ti->table);
 }
 
 static void pending_complete(struct dm_snap_pending_exception *pe, int success)
 {
-	struct dm_snap_exception *e;
+	struct dm_exception *e;
 	struct dm_snapshot *s = pe->snap;
 	struct bio *origin_bios = NULL;
 	struct bio *snapshot_bios = NULL;
@@ -877,7 +1422,7 @@ static void pending_complete(struct dm_snap_pending_exception *pe, int success)
 		goto out;
 	}
 
-	e = alloc_exception();
+	e = alloc_completed_exception();
 	if (!e) {
 		down_write(&s->lock);
 		__invalidate_snapshot(s, -ENOMEM);
@@ -888,28 +1433,27 @@ static void pending_complete(struct dm_snap_pending_exception *pe, int success)
 
 	down_write(&s->lock);
 	if (!s->valid) {
-		free_exception(e);
+		free_completed_exception(e);
 		error = 1;
 		goto out;
 	}
 
-	/*
-	 * Check for conflicting reads. This is extremely improbable,
-	 * so msleep(1) is sufficient and there is no need for a wait queue.
-	 */
-	while (__chunk_is_tracked(s, pe->e.old_chunk))
-		msleep(1);
+	/* Check for conflicting reads */
+	__check_for_conflicting_io(s, pe->e.old_chunk);
 
 	/*
 	 * Add a proper exception, and remove the
 	 * in-flight exception from the list.
 	 */
-	insert_completed_exception(s, e);
+	dm_insert_exception(&s->complete, e);
 
  out:
-	remove_exception(&pe->e);
+	dm_remove_exception(&pe->e);
 	snapshot_bios = bio_list_get(&pe->snapshot_bios);
-	origin_bios = put_pending_exception(pe);
+	origin_bios = bio_list_get(&pe->origin_bios);
+	free_pending_exception(pe);
+
+	increment_pending_exceptions_done_count();
 
 	up_write(&s->lock);
 
@@ -919,7 +1463,7 @@ static void pending_complete(struct dm_snap_pending_exception *pe, int success)
 	else
 		flush_bios(snapshot_bios);
 
-	flush_bios(origin_bios);
+	retry_origin_bios(s, origin_bios);
 }
 
 static void commit_callback(void *context, int success)
@@ -963,7 +1507,7 @@ static void start_copy(struct dm_snap_pending_exception *pe)
 	src.sector = chunk_to_sector(s->store, pe->e.old_chunk);
 	src.count = min((sector_t)s->store->chunk_size, dev_size - src.sector);
 
-	dest.bdev = s->store->cow->bdev;
+	dest.bdev = s->cow->bdev;
 	dest.sector = chunk_to_sector(s->store, pe->e.new_chunk);
 	dest.count = src.count;
 
@@ -975,7 +1519,7 @@ static void start_copy(struct dm_snap_pending_exception *pe)
 static struct dm_snap_pending_exception *
 __lookup_pending_exception(struct dm_snapshot *s, chunk_t chunk)
 {
-	struct dm_snap_exception *e = lookup_exception(&s->pending, chunk);
+	struct dm_exception *e = dm_lookup_exception(&s->pending, chunk);
 
 	if (!e)
 		return NULL;
@@ -1006,8 +1550,6 @@ __find_pending_exception(struct dm_snapshot *s,
 	pe->e.old_chunk = chunk;
 	bio_list_init(&pe->origin_bios);
 	bio_list_init(&pe->snapshot_bios);
-	pe->primary_pe = NULL;
-	atomic_set(&pe->ref_count, 0);
 	pe->started = 0;
 
 	if (s->store->type->prepare_exception(s->store, &pe->e)) {
@@ -1015,16 +1557,15 @@ __find_pending_exception(struct dm_snapshot *s,
 		return NULL;
 	}
 
-	get_pending_exception(pe);
-	insert_exception(&s->pending, &pe->e);
+	dm_insert_exception(&s->pending, &pe->e);
 
 	return pe;
 }
 
-static void remap_exception(struct dm_snapshot *s, struct dm_snap_exception *e,
+static void remap_exception(struct dm_snapshot *s, struct dm_exception *e,
 			    struct bio *bio, chunk_t chunk)
 {
-	bio->bi_bdev = s->store->cow->bdev;
+	bio->bi_bdev = s->cow->bdev;
 	bio->bi_sector = chunk_to_sector(s->store,
 					 dm_chunk_number(e->new_chunk) +
 					 (chunk - e->old_chunk)) +
@@ -1035,14 +1576,14 @@ static void remap_exception(struct dm_snapshot *s, struct dm_snap_exception *e,
 static int snapshot_map(struct dm_target *ti, struct bio *bio,
 			union map_info *map_context)
 {
-	struct dm_snap_exception *e;
+	struct dm_exception *e;
 	struct dm_snapshot *s = ti->private;
 	int r = DM_MAPIO_REMAPPED;
 	chunk_t chunk;
 	struct dm_snap_pending_exception *pe = NULL;
 
 	if (unlikely(bio_empty_barrier(bio))) {
-		bio->bi_bdev = s->store->cow->bdev;
+		bio->bi_bdev = s->cow->bdev;
 		return DM_MAPIO_REMAPPED;
 	}
 
@@ -1063,7 +1604,7 @@ static int snapshot_map(struct dm_target *ti, struct bio *bio,
 	}
 
 	/* If the block is already remapped - use that, else remap it */
-	e = lookup_exception(&s->complete, chunk);
+	e = dm_lookup_exception(&s->complete, chunk);
 	if (e) {
 		remap_exception(s, e, bio, chunk);
 		goto out_unlock;
@@ -1087,7 +1628,7 @@ static int snapshot_map(struct dm_target *ti, struct bio *bio,
 				goto out_unlock;
 			}
 
-			e = lookup_exception(&s->complete, chunk);
+			e = dm_lookup_exception(&s->complete, chunk);
 			if (e) {
 				free_pending_exception(pe);
 				remap_exception(s, e, bio, chunk);
@@ -1125,6 +1666,78 @@ static int snapshot_map(struct dm_target *ti, struct bio *bio,
 	return r;
 }
 
+/*
+ * A snapshot-merge target behaves like a combination of a snapshot
+ * target and a snapshot-origin target.  It only generates new
+ * exceptions in other snapshots and not in the one that is being
+ * merged.
+ *
+ * For each chunk, if there is an existing exception, it is used to
+ * redirect I/O to the cow device.  Otherwise I/O is sent to the origin,
+ * which in turn might generate exceptions in other snapshots.
+ * If merging is currently taking place on the chunk in question, the
+ * I/O is deferred by adding it to s->bios_queued_during_merge.
+ */
+static int snapshot_merge_map(struct dm_target *ti, struct bio *bio,
+			      union map_info *map_context)
+{
+	struct dm_exception *e;
+	struct dm_snapshot *s = ti->private;
+	int r = DM_MAPIO_REMAPPED;
+	chunk_t chunk;
+
+	if (unlikely(bio_empty_barrier(bio))) {
+		if (!map_context->flush_request)
+			bio->bi_bdev = s->origin->bdev;
+		else
+			bio->bi_bdev = s->cow->bdev;
+		map_context->ptr = NULL;
+		return DM_MAPIO_REMAPPED;
+	}
+
+	chunk = sector_to_chunk(s->store, bio->bi_sector);
+
+	down_write(&s->lock);
+
+	/* Full merging snapshots are redirected to the origin */
+	if (!s->valid)
+		goto redirect_to_origin;
+
+	/* If the block is already remapped - use that */
+	e = dm_lookup_exception(&s->complete, chunk);
+	if (e) {
+		/* Queue writes overlapping with chunks being merged */
+		if (bio_rw(bio) == WRITE &&
+		    chunk >= s->first_merging_chunk &&
+		    chunk < (s->first_merging_chunk +
+			     s->num_merging_chunks)) {
+			bio->bi_bdev = s->origin->bdev;
+			bio_list_add(&s->bios_queued_during_merge, bio);
+			r = DM_MAPIO_SUBMITTED;
+			goto out_unlock;
+		}
+
+		remap_exception(s, e, bio, chunk);
+
+		if (bio_rw(bio) == WRITE)
+			map_context->ptr = track_chunk(s, chunk);
+		goto out_unlock;
+	}
+
+redirect_to_origin:
+	bio->bi_bdev = s->origin->bdev;
+
+	if (bio_rw(bio) == WRITE) {
+		up_write(&s->lock);
+		return do_origin(s->origin, bio);
+	}
+
+out_unlock:
+	up_write(&s->lock);
+
+	return r;
+}
+
 static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
 			   int error, union map_info *map_context)
 {
@@ -1137,40 +1750,135 @@ static int snapshot_end_io(struct dm_target *ti, struct bio *bio,
 	return 0;
 }
 
+static void snapshot_merge_presuspend(struct dm_target *ti)
+{
+	struct dm_snapshot *s = ti->private;
+
+	stop_merge(s);
+}
+
+static void snapshot_postsuspend(struct dm_target *ti)
+{
+	struct dm_snapshot *s = ti->private;
+
+	down_write(&s->lock);
+	s->suspended = 1;
+	up_write(&s->lock);
+}
+
+static int snapshot_preresume(struct dm_target *ti)
+{
+	int r = 0;
+	struct dm_snapshot *s = ti->private;
+	struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
+
+	down_read(&_origins_lock);
+	(void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
+	if (snap_src && snap_dest) {
+		down_read(&snap_src->lock);
+		if (s == snap_src) {
+			DMERR("Unable to resume snapshot source until "
+			      "handover completes.");
+			r = -EINVAL;
+		} else if (!snap_src->suspended) {
+			DMERR("Unable to perform snapshot handover until "
+			      "source is suspended.");
+			r = -EINVAL;
+		}
+		up_read(&snap_src->lock);
+	}
+	up_read(&_origins_lock);
+
+	return r;
+}
+
 static void snapshot_resume(struct dm_target *ti)
 {
 	struct dm_snapshot *s = ti->private;
+	struct dm_snapshot *snap_src = NULL, *snap_dest = NULL;
+
+	down_read(&_origins_lock);
+	(void) __find_snapshots_sharing_cow(s, &snap_src, &snap_dest, NULL);
+	if (snap_src && snap_dest) {
+		down_write(&snap_src->lock);
+		down_write_nested(&snap_dest->lock, SINGLE_DEPTH_NESTING);
+		__handover_exceptions(snap_src, snap_dest);
+		up_write(&snap_dest->lock);
+		up_write(&snap_src->lock);
+	}
+	up_read(&_origins_lock);
+
+	/* Now we have correct chunk size, reregister */
+	reregister_snapshot(s);
 
 	down_write(&s->lock);
 	s->active = 1;
+	s->suspended = 0;
 	up_write(&s->lock);
 }
 
+static sector_t get_origin_minimum_chunksize(struct block_device *bdev)
+{
+	sector_t min_chunksize;
+
+	down_read(&_origins_lock);
+	min_chunksize = __minimum_chunk_size(__lookup_origin(bdev));
+	up_read(&_origins_lock);
+
+	return min_chunksize;
+}
+
+static void snapshot_merge_resume(struct dm_target *ti)
+{
+	struct dm_snapshot *s = ti->private;
+
+	/*
+	 * Handover exceptions from existing snapshot.
+	 */
+	snapshot_resume(ti);
+
+	/*
+	 * snapshot-merge acts as an origin, so set ti->split_io
+	 */
+	ti->split_io = get_origin_minimum_chunksize(s->origin->bdev);
+
+	start_merge(s);
+}
+
 static int snapshot_status(struct dm_target *ti, status_type_t type,
 			   char *result, unsigned int maxlen)
 {
 	unsigned sz = 0;
 	struct dm_snapshot *snap = ti->private;
 
-	down_write(&snap->lock);
-
 	switch (type) {
 	case STATUSTYPE_INFO:
+
+		down_write(&snap->lock);
+
 		if (!snap->valid)
 			DMEMIT("Invalid");
+		else if (snap->merge_failed)
+			DMEMIT("Merge failed");
 		else {
-			if (snap->store->type->fraction_full) {
-				sector_t numerator, denominator;
-				snap->store->type->fraction_full(snap->store,
-								 &numerator,
-								 &denominator);
-				DMEMIT("%llu/%llu",
-				       (unsigned long long)numerator,
-				       (unsigned long long)denominator);
+			if (snap->store->type->usage) {
+				sector_t total_sectors, sectors_allocated,
+					 metadata_sectors;
+				snap->store->type->usage(snap->store,
+							 &total_sectors,
+							 &sectors_allocated,
+							 &metadata_sectors);
+				DMEMIT("%llu/%llu %llu",
+				       (unsigned long long)sectors_allocated,
+				       (unsigned long long)total_sectors,
+				       (unsigned long long)metadata_sectors);
 			}
 			else
 				DMEMIT("Unknown");
 		}
+
+		up_write(&snap->lock);
+
 		break;
 
 	case STATUSTYPE_TABLE:
@@ -1179,14 +1887,12 @@ static int snapshot_status(struct dm_target *ti, status_type_t type,
 		 * to make private copies if the output is to
 		 * make sense.
 		 */
-		DMEMIT("%s", snap->origin->name);
+		DMEMIT("%s %s", snap->origin->name, snap->cow->name);
 		snap->store->type->status(snap->store, type, result + sz,
 					  maxlen - sz);
 		break;
 	}
 
-	up_write(&snap->lock);
-
 	return 0;
 }
 
@@ -1202,17 +1908,36 @@ static int snapshot_iterate_devices(struct dm_target *ti,
 /*-----------------------------------------------------------------
  * Origin methods
  *---------------------------------------------------------------*/
-static int __origin_write(struct list_head *snapshots, struct bio *bio)
+
+/*
+ * If no exceptions need creating, DM_MAPIO_REMAPPED is returned and any
+ * supplied bio was ignored.  The caller may submit it immediately.
+ * (No remapping actually occurs as the origin is always a direct linear
+ * map.)
+ *
+ * If further exceptions are required, DM_MAPIO_SUBMITTED is returned
+ * and any supplied bio is added to a list to be submitted once all
+ * the necessary exceptions exist.
+ */
+static int __origin_write(struct list_head *snapshots, sector_t sector,
+			  struct bio *bio)
 {
-	int r = DM_MAPIO_REMAPPED, first = 0;
+	int r = DM_MAPIO_REMAPPED;
 	struct dm_snapshot *snap;
-	struct dm_snap_exception *e;
-	struct dm_snap_pending_exception *pe, *next_pe, *primary_pe = NULL;
+	struct dm_exception *e;
+	struct dm_snap_pending_exception *pe;
+	struct dm_snap_pending_exception *pe_to_start_now = NULL;
+	struct dm_snap_pending_exception *pe_to_start_last = NULL;
 	chunk_t chunk;
-	LIST_HEAD(pe_queue);
 
 	/* Do all the snapshots on this origin */
 	list_for_each_entry (snap, snapshots, list) {
+		/*
+		 * Don't make new exceptions in a merging snapshot
+		 * because it has effectively been deleted
+		 */
+		if (dm_target_is_snapshot_merge(snap->ti))
+			continue;
 
 		down_write(&snap->lock);
 
@@ -1221,24 +1946,21 @@ static int __origin_write(struct list_head *snapshots, struct bio *bio)
 			goto next_snapshot;
 
 		/* Nothing to do if writing beyond end of snapshot */
-		if (bio->bi_sector >= dm_table_get_size(snap->store->ti->table))
+		if (sector >= dm_table_get_size(snap->ti->table))
 			goto next_snapshot;
 
 		/*
 		 * Remember, different snapshots can have
 		 * different chunk sizes.
 		 */
-		chunk = sector_to_chunk(snap->store, bio->bi_sector);
+		chunk = sector_to_chunk(snap->store, sector);
 
 		/*
 		 * Check exception table to see if block
 		 * is already remapped in this snapshot
 		 * and trigger an exception if not.
-		 *
-		 * ref_count is initialised to 1 so pending_complete()
-		 * won't destroy the primary_pe while we're inside this loop.
 		 */
-		e = lookup_exception(&snap->complete, chunk);
+		e = dm_lookup_exception(&snap->complete, chunk);
 		if (e)
 			goto next_snapshot;
 
@@ -1253,7 +1975,7 @@ static int __origin_write(struct list_head *snapshots, struct bio *bio)
 				goto next_snapshot;
 			}
 
-			e = lookup_exception(&snap->complete, chunk);
+			e = dm_lookup_exception(&snap->complete, chunk);
 			if (e) {
 				free_pending_exception(pe);
 				goto next_snapshot;
@@ -1266,59 +1988,43 @@ static int __origin_write(struct list_head *snapshots, struct bio *bio)
 			}
 		}
 
-		if (!primary_pe) {
-			/*
-			 * Either every pe here has same
-			 * primary_pe or none has one yet.
-			 */
-			if (pe->primary_pe)
-				primary_pe = pe->primary_pe;
-			else {
-				primary_pe = pe;
-				first = 1;
-			}
-
-			bio_list_add(&primary_pe->origin_bios, bio);
+		r = DM_MAPIO_SUBMITTED;
 
-			r = DM_MAPIO_SUBMITTED;
-		}
+		/*
+		 * If an origin bio was supplied, queue it to wait for the
+		 * completion of this exception, and start this one last,
+		 * at the end of the function.
+		 */
+		if (bio) {
+			bio_list_add(&pe->origin_bios, bio);
+			bio = NULL;
 
-		if (!pe->primary_pe) {
-			pe->primary_pe = primary_pe;
-			get_pending_exception(primary_pe);
+			if (!pe->started) {
+				pe->started = 1;
+				pe_to_start_last = pe;
+			}
 		}
 
 		if (!pe->started) {
 			pe->started = 1;
-			list_add_tail(&pe->list, &pe_queue);
+			pe_to_start_now = pe;
 		}
 
  next_snapshot:
 		up_write(&snap->lock);
-	}
 
-	if (!primary_pe)
-		return r;
-
-	/*
-	 * If this is the first time we're processing this chunk and
-	 * ref_count is now 1 it means all the pending exceptions
-	 * got completed while we were in the loop above, so it falls to
-	 * us here to remove the primary_pe and submit any origin_bios.
-	 */
-
-	if (first && atomic_dec_and_test(&primary_pe->ref_count)) {
-		flush_bios(bio_list_get(&primary_pe->origin_bios));
-		free_pending_exception(primary_pe);
-		/* If we got here, pe_queue is necessarily empty. */
-		return r;
+		if (pe_to_start_now) {
+			start_copy(pe_to_start_now);
+			pe_to_start_now = NULL;
+		}
 	}
 
 	/*
-	 * Now that we have a complete pe list we can start the copying.
+	 * Submit the exception against which the bio is queued last,
+	 * to give the other exceptions a head start.
 	 */
-	list_for_each_entry_safe(pe, next_pe, &pe_queue, list)
-		start_copy(pe);
+	if (pe_to_start_last)
+		start_copy(pe_to_start_last);
 
 	return r;
 }
@@ -1334,12 +2040,47 @@ static int do_origin(struct dm_dev *origin, struct bio *bio)
 	down_read(&_origins_lock);
 	o = __lookup_origin(origin->bdev);
 	if (o)
-		r = __origin_write(&o->snapshots, bio);
+		r = __origin_write(&o->snapshots, bio->bi_sector, bio);
 	up_read(&_origins_lock);
 
 	return r;
 }
 
+/*
+ * Trigger exceptions in all non-merging snapshots.
+ *
+ * The chunk size of the merging snapshot may be larger than the chunk
+ * size of some other snapshot so we may need to reallocate multiple
+ * chunks in other snapshots.
+ *
+ * We scan all the overlapping exceptions in the other snapshots.
+ * Returns 1 if anything was reallocated and must be waited for,
+ * otherwise returns 0.
+ *
+ * size must be a multiple of merging_snap's chunk_size.
+ */
+static int origin_write_extent(struct dm_snapshot *merging_snap,
+			       sector_t sector, unsigned size)
+{
+	int must_wait = 0;
+	sector_t n;
+	struct origin *o;
+
+	/*
+	 * The origin's __minimum_chunk_size() got stored in split_io
+	 * by snapshot_merge_resume().
+	 */
+	down_read(&_origins_lock);
+	o = __lookup_origin(merging_snap->origin->bdev);
+	for (n = 0; n < size; n += merging_snap->ti->split_io)
+		if (__origin_write(&o->snapshots, sector + n, NULL) ==
+		    DM_MAPIO_SUBMITTED)
+			must_wait = 1;
+	up_read(&_origins_lock);
+
+	return must_wait;
+}
+
 /*
  * Origin: maps a linear range of a device, with hooks for snapshotting.
  */
@@ -1391,8 +2132,6 @@ static int origin_map(struct dm_target *ti, struct bio *bio,
 	return (bio_rw(bio) == WRITE) ? do_origin(dev, bio) : DM_MAPIO_REMAPPED;
 }
 
-#define min_not_zero(l, r) (l == 0) ? r : ((r == 0) ? l : min(l, r))
-
 /*
  * Set the target "split_io" field to the minimum of all the snapshots'
  * chunk sizes.
@@ -1400,19 +2139,8 @@ static int origin_map(struct dm_target *ti, struct bio *bio,
 static void origin_resume(struct dm_target *ti)
 {
 	struct dm_dev *dev = ti->private;
-	struct dm_snapshot *snap;
-	struct origin *o;
-	unsigned chunk_size = 0;
-
-	down_read(&_origins_lock);
-	o = __lookup_origin(dev->bdev);
-	if (o)
-		list_for_each_entry (snap, &o->snapshots, list)
-			chunk_size = min_not_zero(chunk_size,
-						  snap->store->chunk_size);
-	up_read(&_origins_lock);
 
-	ti->split_io = chunk_size;
+	ti->split_io = get_origin_minimum_chunksize(dev->bdev);
 }
 
 static int origin_status(struct dm_target *ti, status_type_t type, char *result,
@@ -1455,17 +2183,35 @@ static struct target_type origin_target = {
 
 static struct target_type snapshot_target = {
 	.name    = "snapshot",
-	.version = {1, 7, 0},
+	.version = {1, 9, 0},
 	.module  = THIS_MODULE,
 	.ctr     = snapshot_ctr,
 	.dtr     = snapshot_dtr,
 	.map     = snapshot_map,
 	.end_io  = snapshot_end_io,
+	.postsuspend = snapshot_postsuspend,
+	.preresume  = snapshot_preresume,
 	.resume  = snapshot_resume,
 	.status  = snapshot_status,
 	.iterate_devices = snapshot_iterate_devices,
 };
 
+static struct target_type merge_target = {
+	.name    = dm_snapshot_merge_target_name,
+	.version = {1, 0, 0},
+	.module  = THIS_MODULE,
+	.ctr     = snapshot_ctr,
+	.dtr     = snapshot_dtr,
+	.map     = snapshot_merge_map,
+	.end_io  = snapshot_end_io,
+	.presuspend = snapshot_merge_presuspend,
+	.postsuspend = snapshot_postsuspend,
+	.preresume  = snapshot_preresume,
+	.resume  = snapshot_merge_resume,
+	.status  = snapshot_status,
+	.iterate_devices = snapshot_iterate_devices,
+};
+
 static int __init dm_snapshot_init(void)
 {
 	int r;
@@ -1477,7 +2223,7 @@ static int __init dm_snapshot_init(void)
 	}
 
 	r = dm_register_target(&snapshot_target);
-	if (r) {
+	if (r < 0) {
 		DMERR("snapshot target register failed %d", r);
 		goto bad_register_snapshot_target;
 	}
@@ -1485,34 +2231,40 @@ static int __init dm_snapshot_init(void)
 	r = dm_register_target(&origin_target);
 	if (r < 0) {
 		DMERR("Origin target register failed %d", r);
-		goto bad1;
+		goto bad_register_origin_target;
+	}
+
+	r = dm_register_target(&merge_target);
+	if (r < 0) {
+		DMERR("Merge target register failed %d", r);
+		goto bad_register_merge_target;
 	}
 
 	r = init_origin_hash();
 	if (r) {
 		DMERR("init_origin_hash failed.");
-		goto bad2;
+		goto bad_origin_hash;
 	}
 
-	exception_cache = KMEM_CACHE(dm_snap_exception, 0);
+	exception_cache = KMEM_CACHE(dm_exception, 0);
 	if (!exception_cache) {
 		DMERR("Couldn't create exception cache.");
 		r = -ENOMEM;
-		goto bad3;
+		goto bad_exception_cache;
 	}
 
 	pending_cache = KMEM_CACHE(dm_snap_pending_exception, 0);
 	if (!pending_cache) {
 		DMERR("Couldn't create pending cache.");
 		r = -ENOMEM;
-		goto bad4;
+		goto bad_pending_cache;
 	}
 
 	tracked_chunk_cache = KMEM_CACHE(dm_snap_tracked_chunk, 0);
 	if (!tracked_chunk_cache) {
 		DMERR("Couldn't create cache to track chunks in use.");
 		r = -ENOMEM;
-		goto bad5;
+		goto bad_tracked_chunk_cache;
 	}
 
 	ksnapd = create_singlethread_workqueue("ksnapd");
@@ -1526,19 +2278,21 @@ static int __init dm_snapshot_init(void)
 
 bad_pending_pool:
 	kmem_cache_destroy(tracked_chunk_cache);
-bad5:
+bad_tracked_chunk_cache:
 	kmem_cache_destroy(pending_cache);
-bad4:
+bad_pending_cache:
 	kmem_cache_destroy(exception_cache);
-bad3:
+bad_exception_cache:
 	exit_origin_hash();
-bad2:
+bad_origin_hash:
+	dm_unregister_target(&merge_target);
+bad_register_merge_target:
 	dm_unregister_target(&origin_target);
-bad1:
+bad_register_origin_target:
 	dm_unregister_target(&snapshot_target);
-
 bad_register_snapshot_target:
 	dm_exception_store_exit();
+
 	return r;
 }
 
@@ -1548,6 +2302,7 @@ static void __exit dm_snapshot_exit(void)
 
 	dm_unregister_target(&snapshot_target);
 	dm_unregister_target(&origin_target);
+	dm_unregister_target(&merge_target);
 
 	exit_origin_hash();
 	kmem_cache_destroy(pending_cache);

drivers/md/dm-sysfs.c

@@ -59,7 +59,7 @@ static ssize_t dm_attr_uuid_show(struct mapped_device *md, char *buf)
 
 static ssize_t dm_attr_suspended_show(struct mapped_device *md, char *buf)
 {
-	sprintf(buf, "%d\n", dm_suspended(md));
+	sprintf(buf, "%d\n", dm_suspended_md(md));
 
 	return strlen(buf);
 }
@@ -79,6 +79,13 @@ static struct sysfs_ops dm_sysfs_ops = {
 	.show	= dm_attr_show,
 };
 
+/*
+ * The sysfs structure is embedded in md struct, nothing to do here
+ */
+static void dm_sysfs_release(struct kobject *kobj)
+{
+}
+
 /*
  * dm kobject is embedded in mapped_device structure
  * no need to define release function here
@@ -86,6 +93,7 @@ static struct sysfs_ops dm_sysfs_ops = {
 static struct kobj_type dm_ktype = {
 	.sysfs_ops	= &dm_sysfs_ops,
 	.default_attrs	= dm_attrs,
+	.release	= dm_sysfs_release
 };
 
 /*

drivers/md/dm-table.c

@@ -238,6 +238,9 @@ void dm_table_destroy(struct dm_table *t)
 {
 	unsigned int i;
 
+	if (!t)
+		return;
+
 	while (atomic_read(&t->holders))
 		msleep(1);
 	smp_mb();

drivers/md/dm-uevent.c

@@ -139,14 +139,13 @@ void dm_send_uevents(struct list_head *events, struct kobject *kobj)
 		list_del_init(&event->elist);
 
 		/*
-		 * Need to call dm_copy_name_and_uuid from here for now.
-		 * Context of previous var adds and locking used for
-		 * hash_cell not compatable.
+		 * When a device is being removed this copy fails and we
+		 * discard these unsent events.
 		 */
 		if (dm_copy_name_and_uuid(event->md, event->name,
 					  event->uuid)) {
-			DMERR("%s: dm_copy_name_and_uuid() failed",
-			      __func__);
+			DMINFO("%s: skipping sending uevent for lost device",
+			       __func__);
 			goto uevent_free;
 		}
 

drivers/md/dm.c

@@ -142,10 +142,20 @@ struct mapped_device {
 	 */
 	int barrier_error;
 
+	/*
+	 * Protect barrier_error from concurrent endio processing
+	 * in request-based dm.
+	 */
+	spinlock_t barrier_error_lock;
+
 	/*
 	 * Processing queue (flush/barriers)
 	 */
 	struct workqueue_struct *wq;
+	struct work_struct barrier_work;
+
+	/* A pointer to the currently processing pre/post flush request */
+	struct request *flush_request;
 
 	/*
 	 * The current mapping.
@@ -178,9 +188,6 @@ struct mapped_device {
 	/* forced geometry settings */
 	struct hd_geometry geometry;
 
-	/* marker of flush suspend for request-based dm */
-	struct request suspend_rq;
-
 	/* For saving the address of __make_request for request based dm */
 	make_request_fn *saved_make_request_fn;
 
@@ -275,6 +282,7 @@ static int (*_inits[])(void) __initdata = {
 	dm_target_init,
 	dm_linear_init,
 	dm_stripe_init,
+	dm_io_init,
 	dm_kcopyd_init,
 	dm_interface_init,
 };
@@ -284,6 +292,7 @@ static void (*_exits[])(void) = {
 	dm_target_exit,
 	dm_linear_exit,
 	dm_stripe_exit,
+	dm_io_exit,
 	dm_kcopyd_exit,
 	dm_interface_exit,
 };
@@ -320,6 +329,11 @@ static void __exit dm_exit(void)
 /*
  * Block device functions
  */
+int dm_deleting_md(struct mapped_device *md)
+{
+	return test_bit(DMF_DELETING, &md->flags);
+}
+
 static int dm_blk_open(struct block_device *bdev, fmode_t mode)
 {
 	struct mapped_device *md;
@@ -331,7 +345,7 @@ static int dm_blk_open(struct block_device *bdev, fmode_t mode)
 		goto out;
 
 	if (test_bit(DMF_FREEING, &md->flags) ||
-	    test_bit(DMF_DELETING, &md->flags)) {
+	    dm_deleting_md(md)) {
 		md = NULL;
 		goto out;
 	}
@@ -388,7 +402,7 @@ static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
 			unsigned int cmd, unsigned long arg)
 {
 	struct mapped_device *md = bdev->bd_disk->private_data;
-	struct dm_table *map = dm_get_table(md);
+	struct dm_table *map = dm_get_live_table(md);
 	struct dm_target *tgt;
 	int r = -ENOTTY;
 
@@ -401,7 +415,7 @@ static int dm_blk_ioctl(struct block_device *bdev, fmode_t mode,
 
 	tgt = dm_table_get_target(map, 0);
 
-	if (dm_suspended(md)) {
+	if (dm_suspended_md(md)) {
 		r = -EAGAIN;
 		goto out;
 	}
@@ -430,9 +444,10 @@ static void free_tio(struct mapped_device *md, struct dm_target_io *tio)
 	mempool_free(tio, md->tio_pool);
 }
 
-static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md)
+static struct dm_rq_target_io *alloc_rq_tio(struct mapped_device *md,
+					    gfp_t gfp_mask)
 {
-	return mempool_alloc(md->tio_pool, GFP_ATOMIC);
+	return mempool_alloc(md->tio_pool, gfp_mask);
 }
 
 static void free_rq_tio(struct dm_rq_target_io *tio)
@@ -450,6 +465,12 @@ static void free_bio_info(struct dm_rq_clone_bio_info *info)
 	mempool_free(info, info->tio->md->io_pool);
 }
 
+static int md_in_flight(struct mapped_device *md)
+{
+	return atomic_read(&md->pending[READ]) +
+	       atomic_read(&md->pending[WRITE]);
+}
+
 static void start_io_acct(struct dm_io *io)
 {
 	struct mapped_device *md = io->md;
@@ -512,7 +533,7 @@ static void queue_io(struct mapped_device *md, struct bio *bio)
  * function to access the md->map field, and make sure they call
  * dm_table_put() when finished.
  */
-struct dm_table *dm_get_table(struct mapped_device *md)
+struct dm_table *dm_get_live_table(struct mapped_device *md)
 {
 	struct dm_table *t;
 	unsigned long flags;
@@ -716,28 +737,38 @@ static void end_clone_bio(struct bio *clone, int error)
 	blk_update_request(tio->orig, 0, nr_bytes);
 }
 
+static void store_barrier_error(struct mapped_device *md, int error)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&md->barrier_error_lock, flags);
+	/*
+	 * Basically, the first error is taken, but:
+	 *   -EOPNOTSUPP supersedes any I/O error.
+	 *   Requeue request supersedes any I/O error but -EOPNOTSUPP.
+	 */
+	if (!md->barrier_error || error == -EOPNOTSUPP ||
+	    (md->barrier_error != -EOPNOTSUPP &&
+	     error == DM_ENDIO_REQUEUE))
+		md->barrier_error = error;
+	spin_unlock_irqrestore(&md->barrier_error_lock, flags);
+}
+
 /*
  * Don't touch any member of the md after calling this function because
  * the md may be freed in dm_put() at the end of this function.
  * Or do dm_get() before calling this function and dm_put() later.
  */
-static void rq_completed(struct mapped_device *md, int run_queue)
+static void rq_completed(struct mapped_device *md, int rw, int run_queue)
 {
-	int wakeup_waiters = 0;
-	struct request_queue *q = md->queue;
-	unsigned long flags;
-
-	spin_lock_irqsave(q->queue_lock, flags);
-	if (!queue_in_flight(q))
-		wakeup_waiters = 1;
-	spin_unlock_irqrestore(q->queue_lock, flags);
+	atomic_dec(&md->pending[rw]);
 
 	/* nudge anyone waiting on suspend queue */
-	if (wakeup_waiters)
+	if (!md_in_flight(md))
 		wake_up(&md->wait);
 
 	if (run_queue)
-		blk_run_queue(q);
+		blk_run_queue(md->queue);
 
 	/*
 	 * dm_put() must be at the end of this function. See the comment above
@@ -753,6 +784,44 @@ static void free_rq_clone(struct request *clone)
 	free_rq_tio(tio);
 }
 
+/*
+ * Complete the clone and the original request.
+ * Must be called without queue lock.
+ */
+static void dm_end_request(struct request *clone, int error)
+{
+	int rw = rq_data_dir(clone);
+	int run_queue = 1;
+	bool is_barrier = blk_barrier_rq(clone);
+	struct dm_rq_target_io *tio = clone->end_io_data;
+	struct mapped_device *md = tio->md;
+	struct request *rq = tio->orig;
+
+	if (blk_pc_request(rq) && !is_barrier) {
+		rq->errors = clone->errors;
+		rq->resid_len = clone->resid_len;
+
+		if (rq->sense)
+			/*
+			 * We are using the sense buffer of the original
+			 * request.
+			 * So setting the length of the sense data is enough.
+			 */
+			rq->sense_len = clone->sense_len;
+	}
+
+	free_rq_clone(clone);
+
+	if (unlikely(is_barrier)) {
+		if (unlikely(error))
+			store_barrier_error(md, error);
+		run_queue = 0;
+	} else
+		blk_end_request_all(rq, error);
+
+	rq_completed(md, rw, run_queue);
+}
+
 static void dm_unprep_request(struct request *rq)
 {
 	struct request *clone = rq->special;
@@ -768,12 +837,23 @@ static void dm_unprep_request(struct request *rq)
  */
 void dm_requeue_unmapped_request(struct request *clone)
 {
+	int rw = rq_data_dir(clone);
 	struct dm_rq_target_io *tio = clone->end_io_data;
 	struct mapped_device *md = tio->md;
 	struct request *rq = tio->orig;
 	struct request_queue *q = rq->q;
 	unsigned long flags;
 
+	if (unlikely(blk_barrier_rq(clone))) {
+		/*
+		 * Barrier clones share an original request.
+		 * Leave it to dm_end_request(), which handles this special
+		 * case.
+		 */
+		dm_end_request(clone, DM_ENDIO_REQUEUE);
+		return;
+	}
+
 	dm_unprep_request(rq);
 
 	spin_lock_irqsave(q->queue_lock, flags);
@@ -782,7 +862,7 @@ void dm_requeue_unmapped_request(struct request *clone)
 	blk_requeue_request(q, rq);
 	spin_unlock_irqrestore(q->queue_lock, flags);
 
-	rq_completed(md, 0);
+	rq_completed(md, rw, 0);
 }
 EXPORT_SYMBOL_GPL(dm_requeue_unmapped_request);
 
@@ -815,34 +895,28 @@ static void start_queue(struct request_queue *q)
 	spin_unlock_irqrestore(q->queue_lock, flags);
 }
 
-/*
- * Complete the clone and the original request.
- * Must be called without queue lock.
- */
-static void dm_end_request(struct request *clone, int error)
+static void dm_done(struct request *clone, int error, bool mapped)
 {
+	int r = error;
 	struct dm_rq_target_io *tio = clone->end_io_data;
-	struct mapped_device *md = tio->md;
-	struct request *rq = tio->orig;
+	dm_request_endio_fn rq_end_io = tio->ti->type->rq_end_io;
 
-	if (blk_pc_request(rq)) {
-		rq->errors = clone->errors;
-		rq->resid_len = clone->resid_len;
+	if (mapped && rq_end_io)
+		r = rq_end_io(tio->ti, clone, error, &tio->info);
 
-		if (rq->sense)
-			/*
-			 * We are using the sense buffer of the original
-			 * request.
-			 * So setting the length of the sense data is enough.
-			 */
-			rq->sense_len = clone->sense_len;
+	if (r <= 0)
+		/* The target wants to complete the I/O */
+		dm_end_request(clone, r);
+	else if (r == DM_ENDIO_INCOMPLETE)
+		/* The target will handle the I/O */
+		return;
+	else if (r == DM_ENDIO_REQUEUE)
+		/* The target wants to requeue the I/O */
+		dm_requeue_unmapped_request(clone);
+	else {
+		DMWARN("unimplemented target endio return value: %d", r);
+		BUG();
 	}
-
-	free_rq_clone(clone);
-
-	blk_end_request_all(rq, error);
-
-	rq_completed(md, 1);
 }
 
 /*
@@ -850,27 +924,14 @@ static void dm_end_request(struct request *clone, int error)
  */
 static void dm_softirq_done(struct request *rq)
 {
+	bool mapped = true;
 	struct request *clone = rq->completion_data;
 	struct dm_rq_target_io *tio = clone->end_io_data;
-	dm_request_endio_fn rq_end_io = tio->ti->type->rq_end_io;
-	int error = tio->error;
 
-	if (!(rq->cmd_flags & REQ_FAILED) && rq_end_io)
-		error = rq_end_io(tio->ti, clone, error, &tio->info);
+	if (rq->cmd_flags & REQ_FAILED)
+		mapped = false;
 
-	if (error <= 0)
-		/* The target wants to complete the I/O */
-		dm_end_request(clone, error);
-	else if (error == DM_ENDIO_INCOMPLETE)
-		/* The target will handle the I/O */
-		return;
-	else if (error == DM_ENDIO_REQUEUE)
-		/* The target wants to requeue the I/O */
-		dm_requeue_unmapped_request(clone);
-	else {
-		DMWARN("unimplemented target endio return value: %d", error);
-		BUG();
-	}
+	dm_done(clone, tio->error, mapped);
 }
 
 /*
@@ -882,6 +943,19 @@ static void dm_complete_request(struct request *clone, int error)
 	struct dm_rq_target_io *tio = clone->end_io_data;
 	struct request *rq = tio->orig;
 
+	if (unlikely(blk_barrier_rq(clone))) {
+		/*
+		 * Barrier clones share an original request.  So can't use
+		 * softirq_done with the original.
+		 * Pass the clone to dm_done() directly in this special case.
+		 * It is safe (even if clone->q->queue_lock is held here)
+		 * because there is no I/O dispatching during the completion
+		 * of barrier clone.
+		 */
+		dm_done(clone, error, true);
+		return;
+	}
+
 	tio->error = error;
 	rq->completion_data = clone;
 	blk_complete_request(rq);
@@ -898,6 +972,17 @@ void dm_kill_unmapped_request(struct request *clone, int error)
 	struct dm_rq_target_io *tio = clone->end_io_data;
 	struct request *rq = tio->orig;
 
+	if (unlikely(blk_barrier_rq(clone))) {
+		/*
+		 * Barrier clones share an original request.
+		 * Leave it to dm_end_request(), which handles this special
+		 * case.
+		 */
+		BUG_ON(error > 0);
+		dm_end_request(clone, error);
+		return;
+	}
+
 	rq->cmd_flags |= REQ_FAILED;
 	dm_complete_request(clone, error);
 }
@@ -1214,7 +1299,7 @@ static void __split_and_process_bio(struct mapped_device *md, struct bio *bio)
 	struct clone_info ci;
 	int error = 0;
 
-	ci.map = dm_get_table(md);
+	ci.map = dm_get_live_table(md);
 	if (unlikely(!ci.map)) {
 		if (!bio_rw_flagged(bio, BIO_RW_BARRIER))
 			bio_io_error(bio);
@@ -1255,7 +1340,7 @@ static int dm_merge_bvec(struct request_queue *q,
 			 struct bio_vec *biovec)
 {
 	struct mapped_device *md = q->queuedata;
-	struct dm_table *map = dm_get_table(md);
+	struct dm_table *map = dm_get_live_table(md);
 	struct dm_target *ti;
 	sector_t max_sectors;
 	int max_size = 0;
@@ -1352,11 +1437,6 @@ static int dm_make_request(struct request_queue *q, struct bio *bio)
 {
 	struct mapped_device *md = q->queuedata;
 
-	if (unlikely(bio_rw_flagged(bio, BIO_RW_BARRIER))) {
-		bio_endio(bio, -EOPNOTSUPP);
-		return 0;
-	}
-
 	return md->saved_make_request_fn(q, bio); /* call __make_request() */
 }
 
@@ -1375,6 +1455,25 @@ static int dm_request(struct request_queue *q, struct bio *bio)
 	return _dm_request(q, bio);
 }
 
+/*
+ * Mark this request as flush request, so that dm_request_fn() can
+ * recognize.
+ */
+static void dm_rq_prepare_flush(struct request_queue *q, struct request *rq)
+{
+	rq->cmd_type = REQ_TYPE_LINUX_BLOCK;
+	rq->cmd[0] = REQ_LB_OP_FLUSH;
+}
+
+static bool dm_rq_is_flush_request(struct request *rq)
+{
+	if (rq->cmd_type == REQ_TYPE_LINUX_BLOCK &&
+	    rq->cmd[0] == REQ_LB_OP_FLUSH)
+		return true;
+	else
+		return false;
+}
+
 void dm_dispatch_request(struct request *rq)
 {
 	int r;
@@ -1420,25 +1519,54 @@ static int dm_rq_bio_constructor(struct bio *bio, struct bio *bio_orig,
 static int setup_clone(struct request *clone, struct request *rq,
 		       struct dm_rq_target_io *tio)
 {
-	int r = blk_rq_prep_clone(clone, rq, tio->md->bs, GFP_ATOMIC,
-				  dm_rq_bio_constructor, tio);
+	int r;
 
-	if (r)
-		return r;
+	if (dm_rq_is_flush_request(rq)) {
+		blk_rq_init(NULL, clone);
+		clone->cmd_type = REQ_TYPE_FS;
+		clone->cmd_flags |= (REQ_HARDBARRIER | WRITE);
+	} else {
+		r = blk_rq_prep_clone(clone, rq, tio->md->bs, GFP_ATOMIC,
+				      dm_rq_bio_constructor, tio);
+		if (r)
+			return r;
+
+		clone->cmd = rq->cmd;
+		clone->cmd_len = rq->cmd_len;
+		clone->sense = rq->sense;
+		clone->buffer = rq->buffer;
+	}
 
-	clone->cmd = rq->cmd;
-	clone->cmd_len = rq->cmd_len;
-	clone->sense = rq->sense;
-	clone->buffer = rq->buffer;
 	clone->end_io = end_clone_request;
 	clone->end_io_data = tio;
 
 	return 0;
 }
 
-static int dm_rq_flush_suspending(struct mapped_device *md)
+static struct request *clone_rq(struct request *rq, struct mapped_device *md,
+				gfp_t gfp_mask)
 {
-	return !md->suspend_rq.special;
+	struct request *clone;
+	struct dm_rq_target_io *tio;
+
+	tio = alloc_rq_tio(md, gfp_mask);
+	if (!tio)
+		return NULL;
+
+	tio->md = md;
+	tio->ti = NULL;
+	tio->orig = rq;
+	tio->error = 0;
+	memset(&tio->info, 0, sizeof(tio->info));
+
+	clone = &tio->clone;
+	if (setup_clone(clone, rq, tio)) {
+		/* -ENOMEM */
+		free_rq_tio(tio);
+		return NULL;
+	}
+
+	return clone;
 }
 
 /*
@@ -1447,39 +1575,19 @@ static int dm_rq_flush_suspending(struct mapped_device *md)
 static int dm_prep_fn(struct request_queue *q, struct request *rq)
 {
 	struct mapped_device *md = q->queuedata;
-	struct dm_rq_target_io *tio;
 	struct request *clone;
 
-	if (unlikely(rq == &md->suspend_rq)) {
-		if (dm_rq_flush_suspending(md))
-			return BLKPREP_OK;
-		else
-			/* The flush suspend was interrupted */
-			return BLKPREP_KILL;
-	}
+	if (unlikely(dm_rq_is_flush_request(rq)))
+		return BLKPREP_OK;
 
 	if (unlikely(rq->special)) {
 		DMWARN("Already has something in rq->special.");
 		return BLKPREP_KILL;
 	}
 
-	tio = alloc_rq_tio(md); /* Only one for each original request */
-	if (!tio)
-		/* -ENOMEM */
-		return BLKPREP_DEFER;
-
-	tio->md = md;
-	tio->ti = NULL;
-	tio->orig = rq;
-	tio->error = 0;
-	memset(&tio->info, 0, sizeof(tio->info));
-
-	clone = &tio->clone;
-	if (setup_clone(clone, rq, tio)) {
-		/* -ENOMEM */
-		free_rq_tio(tio);
+	clone = clone_rq(rq, md, GFP_ATOMIC);
+	if (!clone)
 		return BLKPREP_DEFER;
-	}
 
 	rq->special = clone;
 	rq->cmd_flags |= REQ_DONTPREP;
@@ -1487,11 +1595,10 @@ static int dm_prep_fn(struct request_queue *q, struct request *rq)
 	return BLKPREP_OK;
 }
 
-static void map_request(struct dm_target *ti, struct request *rq,
+static void map_request(struct dm_target *ti, struct request *clone,
 			struct mapped_device *md)
 {
 	int r;
-	struct request *clone = rq->special;
 	struct dm_rq_target_io *tio = clone->end_io_data;
 
 	/*
@@ -1511,6 +1618,8 @@ static void map_request(struct dm_target *ti, struct request *rq,
 		break;
 	case DM_MAPIO_REMAPPED:
 		/* The target has remapped the I/O so dispatch it */
+		trace_block_rq_remap(clone->q, clone, disk_devt(dm_disk(md)),
+				     blk_rq_pos(tio->orig));
 		dm_dispatch_request(clone);
 		break;
 	case DM_MAPIO_REQUEUE:
@@ -1536,29 +1645,26 @@ static void map_request(struct dm_target *ti, struct request *rq,
 static void dm_request_fn(struct request_queue *q)
 {
 	struct mapped_device *md = q->queuedata;
-	struct dm_table *map = dm_get_table(md);
+	struct dm_table *map = dm_get_live_table(md);
 	struct dm_target *ti;
-	struct request *rq;
+	struct request *rq, *clone;
 
 	/*
-	 * For noflush suspend, check blk_queue_stopped() to immediately
-	 * quit I/O dispatching.
+	 * For suspend, check blk_queue_stopped() and increment
+	 * ->pending within a single queue_lock not to increment the
+	 * number of in-flight I/Os after the queue is stopped in
+	 * dm_suspend().
 	 */
 	while (!blk_queue_plugged(q) && !blk_queue_stopped(q)) {
 		rq = blk_peek_request(q);
 		if (!rq)
 			goto plug_and_out;
 
-		if (unlikely(rq == &md->suspend_rq)) { /* Flush suspend maker */
-			if (queue_in_flight(q))
-				/* Not quiet yet.  Wait more */
-				goto plug_and_out;
-
-			/* This device should be quiet now */
-			__stop_queue(q);
+		if (unlikely(dm_rq_is_flush_request(rq))) {
+			BUG_ON(md->flush_request);
+			md->flush_request = rq;
 			blk_start_request(rq);
-			__blk_end_request_all(rq, 0);
-			wake_up(&md->wait);
+			queue_work(md->wq, &md->barrier_work);
 			goto out;
 		}
 
@@ -1567,8 +1673,11 @@ static void dm_request_fn(struct request_queue *q)
 			goto plug_and_out;
 
 		blk_start_request(rq);
+		clone = rq->special;
+		atomic_inc(&md->pending[rq_data_dir(clone)]);
+
 		spin_unlock(q->queue_lock);
-		map_request(ti, rq, md);
+		map_request(ti, clone, md);
 		spin_lock_irq(q->queue_lock);
 	}
 
@@ -1595,7 +1704,7 @@ static int dm_lld_busy(struct request_queue *q)
 {
 	int r;
 	struct mapped_device *md = q->queuedata;
-	struct dm_table *map = dm_get_table(md);
+	struct dm_table *map = dm_get_live_table(md);
 
 	if (!map || test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags))
 		r = 1;
@@ -1610,7 +1719,7 @@ static int dm_lld_busy(struct request_queue *q)
 static void dm_unplug_all(struct request_queue *q)
 {
 	struct mapped_device *md = q->queuedata;
-	struct dm_table *map = dm_get_table(md);
+	struct dm_table *map = dm_get_live_table(md);
 
 	if (map) {
 		if (dm_request_based(md))
@@ -1628,7 +1737,7 @@ static int dm_any_congested(void *congested_data, int bdi_bits)
 	struct dm_table *map;
 
 	if (!test_bit(DMF_BLOCK_IO_FOR_SUSPEND, &md->flags)) {
-		map = dm_get_table(md);
+		map = dm_get_live_table(md);
 		if (map) {
 			/*
 			 * Request-based dm cares about only own queue for
@@ -1725,6 +1834,7 @@ static int next_free_minor(int *minor)
 static const struct block_device_operations dm_blk_dops;
 
 static void dm_wq_work(struct work_struct *work);
+static void dm_rq_barrier_work(struct work_struct *work);
 
 /*
  * Allocate and initialise a blank device with a given minor.
@@ -1754,6 +1864,7 @@ static struct mapped_device *alloc_dev(int minor)
 	init_rwsem(&md->io_lock);
 	mutex_init(&md->suspend_lock);
 	spin_lock_init(&md->deferred_lock);
+	spin_lock_init(&md->barrier_error_lock);
 	rwlock_init(&md->map_lock);
 	atomic_set(&md->holders, 1);
 	atomic_set(&md->open_count, 0);
@@ -1788,6 +1899,8 @@ static struct mapped_device *alloc_dev(int minor)
 	blk_queue_softirq_done(md->queue, dm_softirq_done);
 	blk_queue_prep_rq(md->queue, dm_prep_fn);
 	blk_queue_lld_busy(md->queue, dm_lld_busy);
+	blk_queue_ordered(md->queue, QUEUE_ORDERED_DRAIN_FLUSH,
+			  dm_rq_prepare_flush);
 
 	md->disk = alloc_disk(1);
 	if (!md->disk)
@@ -1797,6 +1910,7 @@ static struct mapped_device *alloc_dev(int minor)
 	atomic_set(&md->pending[1], 0);
 	init_waitqueue_head(&md->wait);
 	INIT_WORK(&md->work, dm_wq_work);
+	INIT_WORK(&md->barrier_work, dm_rq_barrier_work);
 	init_waitqueue_head(&md->eventq);
 
 	md->disk->major = _major;
@@ -1921,9 +2035,13 @@ static void __set_size(struct mapped_device *md, sector_t size)
 	mutex_unlock(&md->bdev->bd_inode->i_mutex);
 }
 
-static int __bind(struct mapped_device *md, struct dm_table *t,
-		  struct queue_limits *limits)
+/*
+ * Returns old map, which caller must destroy.
+ */
+static struct dm_table *__bind(struct mapped_device *md, struct dm_table *t,
+			       struct queue_limits *limits)
 {
+	struct dm_table *old_map;
 	struct request_queue *q = md->queue;
 	sector_t size;
 	unsigned long flags;
@@ -1938,11 +2056,6 @@ static int __bind(struct mapped_device *md, struct dm_table *t,
 
 	__set_size(md, size);
 
-	if (!size) {
-		dm_table_destroy(t);
-		return 0;
-	}
-
 	dm_table_event_callback(t, event_callback, md);
 
 	/*
@@ -1958,26 +2071,31 @@ static int __bind(struct mapped_device *md, struct dm_table *t,
 	__bind_mempools(md, t);
 
 	write_lock_irqsave(&md->map_lock, flags);
+	old_map = md->map;
 	md->map = t;
 	dm_table_set_restrictions(t, q, limits);
 	write_unlock_irqrestore(&md->map_lock, flags);
 
-	return 0;
+	return old_map;
 }
 
-static void __unbind(struct mapped_device *md)
+/*
+ * Returns unbound table for the caller to free.
+ */
+static struct dm_table *__unbind(struct mapped_device *md)
 {
 	struct dm_table *map = md->map;
 	unsigned long flags;
 
 	if (!map)
-		return;
+		return NULL;
 
 	dm_table_event_callback(map, NULL, NULL);
 	write_lock_irqsave(&md->map_lock, flags);
 	md->map = NULL;
 	write_unlock_irqrestore(&md->map_lock, flags);
-	dm_table_destroy(map);
+
+	return map;
 }
 
 /*
@@ -2059,18 +2177,18 @@ void dm_put(struct mapped_device *md)
 	BUG_ON(test_bit(DMF_FREEING, &md->flags));
 
 	if (atomic_dec_and_lock(&md->holders, &_minor_lock)) {
-		map = dm_get_table(md);
+		map = dm_get_live_table(md);
 		idr_replace(&_minor_idr, MINOR_ALLOCED,
 			    MINOR(disk_devt(dm_disk(md))));
 		set_bit(DMF_FREEING, &md->flags);
 		spin_unlock(&_minor_lock);
-		if (!dm_suspended(md)) {
+		if (!dm_suspended_md(md)) {
 			dm_table_presuspend_targets(map);
 			dm_table_postsuspend_targets(map);
 		}
 		dm_sysfs_exit(md);
 		dm_table_put(map);
-		__unbind(md);
+		dm_table_destroy(__unbind(md));
 		free_dev(md);
 	}
 }
@@ -2080,8 +2198,6 @@ static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
 {
 	int r = 0;
 	DECLARE_WAITQUEUE(wait, current);
-	struct request_queue *q = md->queue;
-	unsigned long flags;
 
 	dm_unplug_all(md->queue);
 
@@ -2091,15 +2207,7 @@ static int dm_wait_for_completion(struct mapped_device *md, int interruptible)
 		set_current_state(interruptible);
 
 		smp_mb();
-		if (dm_request_based(md)) {
-			spin_lock_irqsave(q->queue_lock, flags);
-			if (!queue_in_flight(q) && blk_queue_stopped(q)) {
-				spin_unlock_irqrestore(q->queue_lock, flags);
-				break;
-			}
-			spin_unlock_irqrestore(q->queue_lock, flags);
-		} else if (!atomic_read(&md->pending[0]) &&
-					!atomic_read(&md->pending[1]))
+		if (!md_in_flight(md))
 			break;
 
 		if (interruptible == TASK_INTERRUPTIBLE &&
@@ -2194,98 +2302,106 @@ static void dm_queue_flush(struct mapped_device *md)
 	queue_work(md->wq, &md->work);
 }
 
-/*
- * Swap in a new table (destroying old one).
- */
-int dm_swap_table(struct mapped_device *md, struct dm_table *table)
+static void dm_rq_set_flush_nr(struct request *clone, unsigned flush_nr)
 {
-	struct queue_limits limits;
-	int r = -EINVAL;
+	struct dm_rq_target_io *tio = clone->end_io_data;
 
-	mutex_lock(&md->suspend_lock);
+	tio->info.flush_request = flush_nr;
+}
 
-	/* device must be suspended */
-	if (!dm_suspended(md))
-		goto out;
+/* Issue barrier requests to targets and wait for their completion. */
+static int dm_rq_barrier(struct mapped_device *md)
+{
+	int i, j;
+	struct dm_table *map = dm_get_live_table(md);
+	unsigned num_targets = dm_table_get_num_targets(map);
+	struct dm_target *ti;
+	struct request *clone;
 
-	r = dm_calculate_queue_limits(table, &limits);
-	if (r)
-		goto out;
+	md->barrier_error = 0;
 
-	/* cannot change the device type, once a table is bound */
-	if (md->map &&
-	    (dm_table_get_type(md->map) != dm_table_get_type(table))) {
-		DMWARN("can't change the device type after a table is bound");
-		goto out;
+	for (i = 0; i < num_targets; i++) {
+		ti = dm_table_get_target(map, i);
+		for (j = 0; j < ti->num_flush_requests; j++) {
+			clone = clone_rq(md->flush_request, md, GFP_NOIO);
+			dm_rq_set_flush_nr(clone, j);
+			atomic_inc(&md->pending[rq_data_dir(clone)]);
+			map_request(ti, clone, md);
+		}
 	}
 
-	__unbind(md);
-	r = __bind(md, table, &limits);
-
-out:
-	mutex_unlock(&md->suspend_lock);
-	return r;
-}
+	dm_wait_for_completion(md, TASK_UNINTERRUPTIBLE);
+	dm_table_put(map);
 
-static void dm_rq_invalidate_suspend_marker(struct mapped_device *md)
-{
-	md->suspend_rq.special = (void *)0x1;
+	return md->barrier_error;
 }
 
-static void dm_rq_abort_suspend(struct mapped_device *md, int noflush)
+static void dm_rq_barrier_work(struct work_struct *work)
 {
+	int error;
+	struct mapped_device *md = container_of(work, struct mapped_device,
+						barrier_work);
 	struct request_queue *q = md->queue;
+	struct request *rq;
 	unsigned long flags;
 
-	spin_lock_irqsave(q->queue_lock, flags);
-	if (!noflush)
-		dm_rq_invalidate_suspend_marker(md);
-	__start_queue(q);
-	spin_unlock_irqrestore(q->queue_lock, flags);
-}
+	/*
+	 * Hold the md reference here and leave it at the last part so that
+	 * the md can't be deleted by device opener when the barrier request
+	 * completes.
+	 */
+	dm_get(md);
 
-static void dm_rq_start_suspend(struct mapped_device *md, int noflush)
-{
-	struct request *rq = &md->suspend_rq;
-	struct request_queue *q = md->queue;
+	error = dm_rq_barrier(md);
 
-	if (noflush)
-		stop_queue(q);
-	else {
-		blk_rq_init(q, rq);
-		blk_insert_request(q, rq, 0, NULL);
-	}
+	rq = md->flush_request;
+	md->flush_request = NULL;
+
+	if (error == DM_ENDIO_REQUEUE) {
+		spin_lock_irqsave(q->queue_lock, flags);
+		blk_requeue_request(q, rq);
+		spin_unlock_irqrestore(q->queue_lock, flags);
+	} else
+		blk_end_request_all(rq, error);
+
+	blk_run_queue(q);
+
+	dm_put(md);
 }
 
-static int dm_rq_suspend_available(struct mapped_device *md, int noflush)
+/*
+ * Swap in a new table, returning the old one for the caller to destroy.
+ */
+struct dm_table *dm_swap_table(struct mapped_device *md, struct dm_table *table)
 {
-	int r = 1;
-	struct request *rq = &md->suspend_rq;
-	struct request_queue *q = md->queue;
-	unsigned long flags;
+	struct dm_table *map = ERR_PTR(-EINVAL);
+	struct queue_limits limits;
+	int r;
 
-	if (noflush)
-		return r;
+	mutex_lock(&md->suspend_lock);
 
-	/* The marker must be protected by queue lock if it is in use */
-	spin_lock_irqsave(q->queue_lock, flags);
-	if (unlikely(rq->ref_count)) {
-		/*
-		 * This can happen, when the previous flush suspend was
-		 * interrupted, the marker is still in the queue and
-		 * this flush suspend has been invoked, because we don't
-		 * remove the marker at the time of suspend interruption.
-		 * We have only one marker per mapped_device, so we can't
-		 * start another flush suspend while it is in use.
-		 */
-		BUG_ON(!rq->special); /* The marker should be invalidated */
-		DMWARN("Invalidating the previous flush suspend is still in"
-		       " progress.  Please retry later.");
-		r = 0;
+	/* device must be suspended */
+	if (!dm_suspended_md(md))
+		goto out;
+
+	r = dm_calculate_queue_limits(table, &limits);
+	if (r) {
+		map = ERR_PTR(r);
+		goto out;
 	}
-	spin_unlock_irqrestore(q->queue_lock, flags);
 
-	return r;
+	/* cannot change the device type, once a table is bound */
+	if (md->map &&
+	    (dm_table_get_type(md->map) != dm_table_get_type(table))) {
+		DMWARN("can't change the device type after a table is bound");
+		goto out;
+	}
+
+	map = __bind(md, table, &limits);
+
+out:
+	mutex_unlock(&md->suspend_lock);
+	return map;
 }
 
 /*
@@ -2330,49 +2446,11 @@ static void unlock_fs(struct mapped_device *md)
 /*
  * Suspend mechanism in request-based dm.
  *
- * After the suspend starts, further incoming requests are kept in
- * the request_queue and deferred.
- * Remaining requests in the request_queue at the start of suspend are flushed
- * if it is flush suspend.
- * The suspend completes when the following conditions have been satisfied,
- * so wait for it:
- *    1. q->in_flight is 0 (which means no in_flight request)
- *    2. queue has been stopped (which means no request dispatching)
- *
+ * 1. Flush all I/Os by lock_fs() if needed.
+ * 2. Stop dispatching any I/O by stopping the request_queue.
+ * 3. Wait for all in-flight I/Os to be completed or requeued.
  *
- * Noflush suspend
- * ---------------
- * Noflush suspend doesn't need to dispatch remaining requests.
- * So stop the queue immediately.  Then, wait for all in_flight requests
- * to be completed or requeued.
- *
- * To abort noflush suspend, start the queue.
- *
- *
- * Flush suspend
- * -------------
- * Flush suspend needs to dispatch remaining requests.  So stop the queue
- * after the remaining requests are completed. (Requeued request must be also
- * re-dispatched and completed.  Until then, we can't stop the queue.)
- *
- * During flushing the remaining requests, further incoming requests are also
- * inserted to the same queue.  To distinguish which requests are to be
- * flushed, we insert a marker request to the queue at the time of starting
- * flush suspend, like a barrier.
- * The dispatching is blocked when the marker is found on the top of the queue.
- * And the queue is stopped when all in_flight requests are completed, since
- * that means the remaining requests are completely flushed.
- * Then, the marker is removed from the queue.
- *
- * To abort flush suspend, we also need to take care of the marker, not only
- * starting the queue.
- * We don't remove the marker forcibly from the queue since it's against
- * the block-layer manner.  Instead, we put a invalidated mark on the marker.
- * When the invalidated marker is found on the top of the queue, it is
- * immediately removed from the queue, so it doesn't block dispatching.
- * Because we have only one marker per mapped_device, we can't start another
- * flush suspend until the invalidated marker is removed from the queue.
- * So fail and return with -EBUSY in such a case.
+ * To abort suspend, start the request_queue.
  */
 int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 {
@@ -2383,17 +2461,12 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 
 	mutex_lock(&md->suspend_lock);
 
-	if (dm_suspended(md)) {
+	if (dm_suspended_md(md)) {
 		r = -EINVAL;
 		goto out_unlock;
 	}
 
-	if (dm_request_based(md) && !dm_rq_suspend_available(md, noflush)) {
-		r = -EBUSY;
-		goto out_unlock;
-	}
-
-	map = dm_get_table(md);
+	map = dm_get_live_table(md);
 
 	/*
 	 * DMF_NOFLUSH_SUSPENDING must be set before presuspend.
@@ -2406,8 +2479,10 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 	dm_table_presuspend_targets(map);
 
 	/*
-	 * Flush I/O to the device. noflush supersedes do_lockfs,
-	 * because lock_fs() needs to flush I/Os.
+	 * Flush I/O to the device.
+	 * Any I/O submitted after lock_fs() may not be flushed.
+	 * noflush takes precedence over do_lockfs.
+	 * (lock_fs() flushes I/Os and waits for them to complete.)
 	 */
 	if (!noflush && do_lockfs) {
 		r = lock_fs(md);
@@ -2436,10 +2511,15 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 	set_bit(DMF_QUEUE_IO_TO_THREAD, &md->flags);
 	up_write(&md->io_lock);
 
-	flush_workqueue(md->wq);
-
+	/*
+	 * Request-based dm uses md->wq for barrier (dm_rq_barrier_work) which
+	 * can be kicked until md->queue is stopped.  So stop md->queue before
+	 * flushing md->wq.
+	 */
 	if (dm_request_based(md))
-		dm_rq_start_suspend(md, noflush);
+		stop_queue(md->queue);
+
+	flush_workqueue(md->wq);
 
 	/*
 	 * At this point no more requests are entering target request routines.
@@ -2458,7 +2538,7 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 		dm_queue_flush(md);
 
 		if (dm_request_based(md))
-			dm_rq_abort_suspend(md, noflush);
+			start_queue(md->queue);
 
 		unlock_fs(md);
 		goto out; /* pushback list is already flushed, so skip flush */
@@ -2470,10 +2550,10 @@ int dm_suspend(struct mapped_device *md, unsigned suspend_flags)
 	 * requests are being added to md->deferred list.
 	 */
 
-	dm_table_postsuspend_targets(map);
-
 	set_bit(DMF_SUSPENDED, &md->flags);
 
+	dm_table_postsuspend_targets(map);
+
 out:
 	dm_table_put(map);
 
@@ -2488,10 +2568,10 @@ int dm_resume(struct mapped_device *md)
 	struct dm_table *map = NULL;
 
 	mutex_lock(&md->suspend_lock);
-	if (!dm_suspended(md))
+	if (!dm_suspended_md(md))
 		goto out;
 
-	map = dm_get_table(md);
+	map = dm_get_live_table(md);
 	if (!map || !dm_table_get_size(map))
 		goto out;
 
@@ -2592,18 +2672,29 @@ struct mapped_device *dm_get_from_kobject(struct kobject *kobj)
 		return NULL;
 
 	if (test_bit(DMF_FREEING, &md->flags) ||
-	    test_bit(DMF_DELETING, &md->flags))
+	    dm_deleting_md(md))
 		return NULL;
 
 	dm_get(md);
 	return md;
 }
 
-int dm_suspended(struct mapped_device *md)
+int dm_suspended_md(struct mapped_device *md)
 {
 	return test_bit(DMF_SUSPENDED, &md->flags);
 }
 
+int dm_suspended(struct dm_target *ti)
+{
+	struct mapped_device *md = dm_table_get_md(ti->table);
+	int r = dm_suspended_md(md);
+
+	dm_put(md);
+
+	return r;
+}
+EXPORT_SYMBOL_GPL(dm_suspended);
+
 int dm_noflush_suspending(struct dm_target *ti)
 {
 	struct mapped_device *md = dm_table_get_md(ti->table);

drivers/md/dm.h

@@ -88,6 +88,16 @@ int dm_target_iterate(void (*iter_func)(struct target_type *tt,
 
 int dm_split_args(int *argc, char ***argvp, char *input);
 
+/*
+ * Is this mapped_device being deleted?
+ */
+int dm_deleting_md(struct mapped_device *md);
+
+/*
+ * Is this mapped_device suspended?
+ */
+int dm_suspended_md(struct mapped_device *md);
+
 /*
  * The device-mapper can be driven through one of two interfaces;
  * ioctl or filesystem, depending which patch you have applied.
@@ -118,6 +128,9 @@ int dm_lock_for_deletion(struct mapped_device *md);
 void dm_kobject_uevent(struct mapped_device *md, enum kobject_action action,
 		       unsigned cookie);
 
+int dm_io_init(void);
+void dm_io_exit(void);
+
 int dm_kcopyd_init(void);
 void dm_kcopyd_exit(void);
 

include/linux/device-mapper.h

@@ -235,7 +235,7 @@ void dm_uevent_add(struct mapped_device *md, struct list_head *elist);
 const char *dm_device_name(struct mapped_device *md);
 int dm_copy_name_and_uuid(struct mapped_device *md, char *name, char *uuid);
 struct gendisk *dm_disk(struct mapped_device *md);
-int dm_suspended(struct mapped_device *md);
+int dm_suspended(struct dm_target *ti);
 int dm_noflush_suspending(struct dm_target *ti);
 union map_info *dm_get_mapinfo(struct bio *bio);
 union map_info *dm_get_rq_mapinfo(struct request *rq);
@@ -276,7 +276,7 @@ void dm_table_unplug_all(struct dm_table *t);
 /*
  * Table reference counting.
  */
-struct dm_table *dm_get_table(struct mapped_device *md);
+struct dm_table *dm_get_live_table(struct mapped_device *md);
 void dm_table_get(struct dm_table *t);
 void dm_table_put(struct dm_table *t);
 
@@ -295,8 +295,10 @@ void dm_table_event(struct dm_table *t);
 
 /*
  * The device must be suspended before calling this method.
+ * Returns the previous table, which the caller must destroy.
  */
-int dm_swap_table(struct mapped_device *md, struct dm_table *t);
+struct dm_table *dm_swap_table(struct mapped_device *md,
+			       struct dm_table *t);
 
 /*
  * A wrapper around vmalloc.

include/linux/dm-dirty-log.h

@@ -21,6 +21,7 @@ struct dm_dirty_log_type;
 
 struct dm_dirty_log {
 	struct dm_dirty_log_type *type;
+	int (*flush_callback_fn)(struct dm_target *ti);
 	void *context;
 };
 
@@ -136,8 +137,9 @@ int dm_dirty_log_type_unregister(struct dm_dirty_log_type *type);
  * type->constructor/destructor() directly.
  */
 struct dm_dirty_log *dm_dirty_log_create(const char *type_name,
-					 struct dm_target *ti,
-					 unsigned argc, char **argv);
+			struct dm_target *ti,
+			int (*flush_callback_fn)(struct dm_target *ti),
+			unsigned argc, char **argv);
 void dm_dirty_log_destroy(struct dm_dirty_log *log);
 
 #endif	/* __KERNEL__ */

include/linux/dm-ioctl.h

 /*
  * Copyright (C) 2001 - 2003 Sistina Software (UK) Limited.
- * Copyright (C) 2004 - 2005 Red Hat, Inc. All rights reserved.
+ * Copyright (C) 2004 - 2009 Red Hat, Inc. All rights reserved.
  *
  * This file is released under the LGPL.
  */
@@ -266,9 +266,9 @@ enum {
 #define DM_DEV_SET_GEOMETRY	_IOWR(DM_IOCTL, DM_DEV_SET_GEOMETRY_CMD, struct dm_ioctl)
 
 #define DM_VERSION_MAJOR	4
-#define DM_VERSION_MINOR	15
+#define DM_VERSION_MINOR	16
 #define DM_VERSION_PATCHLEVEL	0
-#define DM_VERSION_EXTRA	"-ioctl (2009-04-01)"
+#define DM_VERSION_EXTRA	"-ioctl (2009-11-05)"
 
 /* Status bits */
 #define DM_READONLY_FLAG	(1 << 0) /* In/Out */
@@ -309,4 +309,11 @@ enum {
  */
 #define DM_NOFLUSH_FLAG		(1 << 11) /* In */
 
+/*
+ * If set, any table information returned will relate to the inactive
+ * table instead of the live one.  Always check DM_INACTIVE_PRESENT_FLAG
+ * is set before using the data returned.
+ */
+#define DM_QUERY_INACTIVE_TABLE_FLAG	(1 << 12) /* In */
+
 #endif				/* _LINUX_DM_IOCTL_H */

include/linux/dm-region-hash.h

@@ -78,8 +78,7 @@ void dm_rh_dec(struct dm_region_hash *rh, region_t region);
 /* Delay bios on regions. */
 void dm_rh_delay(struct dm_region_hash *rh, struct bio *bio);
 
-void dm_rh_mark_nosync(struct dm_region_hash *rh,
-		       struct bio *bio, unsigned done, int error);
+void dm_rh_mark_nosync(struct dm_region_hash *rh, struct bio *bio);
 
 /*
  * Region recovery control.