Adding upstream version 16.2.11+ds.upstream/16.2.11+dsupstream

Signed-off-by: Daniel Baumann <daniel.baumann@progress-linux.org>

13 files changed, 12610 insertions, 0 deletions

diff --git a/src/spdk/lib/blob/Makefile b/src/spdk/lib/blob/Makefile
new file mode 100644
index 000000000..53ae6800b
--- /dev/null
+++ b/src/spdk/lib/blob/Makefile
@@ -0,0 +1,45 @@
+#
+#  BSD LICENSE
+#
+#  Copyright (c) Intel Corporation.
+#  All rights reserved.
+#
+#  Redistribution and use in source and binary forms, with or without
+#  modification, are permitted provided that the following conditions
+#  are met:
+#
+#    * Redistributions of source code must retain the above copyright
+#      notice, this list of conditions and the following disclaimer.
+#    * Redistributions in binary form must reproduce the above copyright
+#      notice, this list of conditions and the following disclaimer in
+#      the documentation and/or other materials provided with the
+#      distribution.
+#    * Neither the name of Intel Corporation nor the names of its
+#      contributors may be used to endorse or promote products derived
+#      from this software without specific prior written permission.
+#
+#  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+#  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+#  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+#  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+#  OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+#  SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+#  LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+#  DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+#  THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+#  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+#  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+#
+
+SPDK_ROOT_DIR := $(abspath $(CURDIR)/../..)
+include $(SPDK_ROOT_DIR)/mk/spdk.common.mk
+
+SO_VER := 3
+SO_MINOR := 0
+
+C_SRCS = blobstore.c request.c zeroes.c blob_bs_dev.c
+LIBNAME = blob
+
+SPDK_MAP_FILE = $(abspath $(CURDIR)/spdk_blob.map)
+
+include $(SPDK_ROOT_DIR)/mk/spdk.lib.mk
diff --git a/src/spdk/lib/blob/blob_bs_dev.c b/src/spdk/lib/blob/blob_bs_dev.c
new file mode 100644
index 000000000..8705a1c16
--- /dev/null
+++ b/src/spdk/lib/blob/blob_bs_dev.c
@@ -0,0 +1,150 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+#include "spdk/blob.h"
+#include "spdk/log.h"
+#include "blobstore.h"
+
+static void
+blob_bs_dev_write(struct spdk_bs_dev *dev, struct spdk_io_channel *channel, void *payload,
+		  uint64_t lba, uint32_t lba_count,
+		  struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+blob_bs_dev_writev(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+		   struct iovec *iov, int iovcnt,
+		   uint64_t lba, uint32_t lba_count,
+		   struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+blob_bs_dev_write_zeroes(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+			 uint64_t lba, uint32_t lba_count,
+			 struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+blob_bs_dev_unmap(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+		  uint64_t lba, uint32_t lba_count,
+		  struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+blob_bs_dev_read_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_dev_cb_args *cb_args = (struct spdk_bs_dev_cb_args *)cb_arg;
+
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, bserrno);
+}
+
+static inline void
+blob_bs_dev_read(struct spdk_bs_dev *dev, struct spdk_io_channel *channel, void *payload,
+		 uint64_t lba, uint32_t lba_count, struct spdk_bs_dev_cb_args *cb_args)
+{
+	struct spdk_blob_bs_dev *b = (struct spdk_blob_bs_dev *)dev;
+
+	spdk_blob_io_read(b->blob, channel, payload, lba, lba_count,
+			  blob_bs_dev_read_cpl, cb_args);
+}
+
+static inline void
+blob_bs_dev_readv(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+		  struct iovec *iov, int iovcnt,
+		  uint64_t lba, uint32_t lba_count, struct spdk_bs_dev_cb_args *cb_args)
+{
+	struct spdk_blob_bs_dev *b = (struct spdk_blob_bs_dev *)dev;
+
+	spdk_blob_io_readv(b->blob, channel, iov, iovcnt, lba, lba_count,
+			   blob_bs_dev_read_cpl, cb_args);
+}
+
+static void
+blob_bs_dev_destroy_cpl(void *cb_arg, int bserrno)
+{
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Error on blob_bs_dev destroy: %d", bserrno);
+	}
+
+	/* Free blob_bs_dev */
+	free(cb_arg);
+}
+
+static void
+blob_bs_dev_destroy(struct spdk_bs_dev *bs_dev)
+{
+	struct spdk_blob_bs_dev *b = (struct spdk_blob_bs_dev *)bs_dev;
+
+	spdk_blob_close(b->blob, blob_bs_dev_destroy_cpl, b);
+}
+
+
+struct spdk_bs_dev *
+bs_create_blob_bs_dev(struct spdk_blob *blob)
+{
+	struct spdk_blob_bs_dev  *b;
+
+	b = calloc(1, sizeof(*b));
+	if (b == NULL) {
+		return NULL;
+	}
+	/* snapshot blob */
+	b->bs_dev.blockcnt = blob->active.num_clusters *
+			     blob->bs->pages_per_cluster * bs_io_unit_per_page(blob->bs);
+	b->bs_dev.blocklen = spdk_bs_get_io_unit_size(blob->bs);
+	b->bs_dev.create_channel = NULL;
+	b->bs_dev.destroy_channel = NULL;
+	b->bs_dev.destroy = blob_bs_dev_destroy;
+	b->bs_dev.write = blob_bs_dev_write;
+	b->bs_dev.writev = blob_bs_dev_writev;
+	b->bs_dev.read = blob_bs_dev_read;
+	b->bs_dev.readv = blob_bs_dev_readv;
+	b->bs_dev.write_zeroes = blob_bs_dev_write_zeroes;
+	b->bs_dev.unmap = blob_bs_dev_unmap;
+	b->blob = blob;
+
+	return &b->bs_dev;
+}
diff --git a/src/spdk/lib/blob/blobstore.c b/src/spdk/lib/blob/blobstore.c
new file mode 100644
index 000000000..768fc5b45
--- /dev/null
+++ b/src/spdk/lib/blob/blobstore.c
@@ -0,0 +1,7461 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+
+#include "spdk/blob.h"
+#include "spdk/crc32.h"
+#include "spdk/env.h"
+#include "spdk/queue.h"
+#include "spdk/thread.h"
+#include "spdk/bit_array.h"
+#include "spdk/likely.h"
+#include "spdk/util.h"
+#include "spdk/string.h"
+
+#include "spdk_internal/assert.h"
+#include "spdk_internal/log.h"
+
+#include "blobstore.h"
+
+#define BLOB_CRC32C_INITIAL    0xffffffffUL
+
+static int bs_register_md_thread(struct spdk_blob_store *bs);
+static int bs_unregister_md_thread(struct spdk_blob_store *bs);
+static void blob_close_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno);
+static void blob_insert_cluster_on_md_thread(struct spdk_blob *blob, uint32_t cluster_num,
+		uint64_t cluster, uint32_t extent, spdk_blob_op_complete cb_fn, void *cb_arg);
+
+static int blob_set_xattr(struct spdk_blob *blob, const char *name, const void *value,
+			  uint16_t value_len, bool internal);
+static int blob_get_xattr_value(struct spdk_blob *blob, const char *name,
+				const void **value, size_t *value_len, bool internal);
+static int blob_remove_xattr(struct spdk_blob *blob, const char *name, bool internal);
+
+static void blob_insert_extent(struct spdk_blob *blob, uint32_t extent, uint64_t cluster_num,
+			       spdk_blob_op_complete cb_fn, void *cb_arg);
+
+static void
+blob_verify_md_op(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+	assert(spdk_get_thread() == blob->bs->md_thread);
+	assert(blob->state != SPDK_BLOB_STATE_LOADING);
+}
+
+static struct spdk_blob_list *
+bs_get_snapshot_entry(struct spdk_blob_store *bs, spdk_blob_id blobid)
+{
+	struct spdk_blob_list *snapshot_entry = NULL;
+
+	TAILQ_FOREACH(snapshot_entry, &bs->snapshots, link) {
+		if (snapshot_entry->id == blobid) {
+			break;
+		}
+	}
+
+	return snapshot_entry;
+}
+
+static void
+bs_claim_md_page(struct spdk_blob_store *bs, uint32_t page)
+{
+	assert(page < spdk_bit_array_capacity(bs->used_md_pages));
+	assert(spdk_bit_array_get(bs->used_md_pages, page) == false);
+
+	spdk_bit_array_set(bs->used_md_pages, page);
+}
+
+static void
+bs_release_md_page(struct spdk_blob_store *bs, uint32_t page)
+{
+	assert(page < spdk_bit_array_capacity(bs->used_md_pages));
+	assert(spdk_bit_array_get(bs->used_md_pages, page) == true);
+
+	spdk_bit_array_clear(bs->used_md_pages, page);
+}
+
+static void
+bs_claim_cluster(struct spdk_blob_store *bs, uint32_t cluster_num)
+{
+	assert(cluster_num < spdk_bit_array_capacity(bs->used_clusters));
+	assert(spdk_bit_array_get(bs->used_clusters, cluster_num) == false);
+	assert(bs->num_free_clusters > 0);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Claiming cluster %u\n", cluster_num);
+
+	spdk_bit_array_set(bs->used_clusters, cluster_num);
+	bs->num_free_clusters--;
+}
+
+static int
+blob_insert_cluster(struct spdk_blob *blob, uint32_t cluster_num, uint64_t cluster)
+{
+	uint64_t *cluster_lba = &blob->active.clusters[cluster_num];
+
+	blob_verify_md_op(blob);
+
+	if (*cluster_lba != 0) {
+		return -EEXIST;
+	}
+
+	*cluster_lba = bs_cluster_to_lba(blob->bs, cluster);
+	return 0;
+}
+
+static int
+bs_allocate_cluster(struct spdk_blob *blob, uint32_t cluster_num,
+		    uint64_t *lowest_free_cluster, uint32_t *lowest_free_md_page, bool update_map)
+{
+	uint32_t *extent_page = 0;
+
+	pthread_mutex_lock(&blob->bs->used_clusters_mutex);
+	*lowest_free_cluster = spdk_bit_array_find_first_clear(blob->bs->used_clusters,
+			       *lowest_free_cluster);
+	if (*lowest_free_cluster == UINT32_MAX) {
+		/* No more free clusters. Cannot satisfy the request */
+		pthread_mutex_unlock(&blob->bs->used_clusters_mutex);
+		return -ENOSPC;
+	}
+
+	if (blob->use_extent_table) {
+		extent_page = bs_cluster_to_extent_page(blob, cluster_num);
+		if (*extent_page == 0) {
+			/* No extent_page is allocated for the cluster */
+			*lowest_free_md_page = spdk_bit_array_find_first_clear(blob->bs->used_md_pages,
+					       *lowest_free_md_page);
+			if (*lowest_free_md_page == UINT32_MAX) {
+				/* No more free md pages. Cannot satisfy the request */
+				pthread_mutex_unlock(&blob->bs->used_clusters_mutex);
+				return -ENOSPC;
+			}
+			bs_claim_md_page(blob->bs, *lowest_free_md_page);
+		}
+	}
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Claiming cluster %lu for blob %lu\n", *lowest_free_cluster, blob->id);
+	bs_claim_cluster(blob->bs, *lowest_free_cluster);
+
+	pthread_mutex_unlock(&blob->bs->used_clusters_mutex);
+
+	if (update_map) {
+		blob_insert_cluster(blob, cluster_num, *lowest_free_cluster);
+		if (blob->use_extent_table && *extent_page == 0) {
+			*extent_page = *lowest_free_md_page;
+		}
+	}
+
+	return 0;
+}
+
+static void
+bs_release_cluster(struct spdk_blob_store *bs, uint32_t cluster_num)
+{
+	assert(cluster_num < spdk_bit_array_capacity(bs->used_clusters));
+	assert(spdk_bit_array_get(bs->used_clusters, cluster_num) == true);
+	assert(bs->num_free_clusters < bs->total_clusters);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Releasing cluster %u\n", cluster_num);
+
+	pthread_mutex_lock(&bs->used_clusters_mutex);
+	spdk_bit_array_clear(bs->used_clusters, cluster_num);
+	bs->num_free_clusters++;
+	pthread_mutex_unlock(&bs->used_clusters_mutex);
+}
+
+static void
+blob_xattrs_init(struct spdk_blob_xattr_opts *xattrs)
+{
+	xattrs->count = 0;
+	xattrs->names = NULL;
+	xattrs->ctx = NULL;
+	xattrs->get_value = NULL;
+}
+
+void
+spdk_blob_opts_init(struct spdk_blob_opts *opts)
+{
+	opts->num_clusters = 0;
+	opts->thin_provision = false;
+	opts->clear_method = BLOB_CLEAR_WITH_DEFAULT;
+	blob_xattrs_init(&opts->xattrs);
+	opts->use_extent_table = true;
+}
+
+void
+spdk_blob_open_opts_init(struct spdk_blob_open_opts *opts)
+{
+	opts->clear_method = BLOB_CLEAR_WITH_DEFAULT;
+}
+
+static struct spdk_blob *
+blob_alloc(struct spdk_blob_store *bs, spdk_blob_id id)
+{
+	struct spdk_blob *blob;
+
+	blob = calloc(1, sizeof(*blob));
+	if (!blob) {
+		return NULL;
+	}
+
+	blob->id = id;
+	blob->bs = bs;
+
+	blob->parent_id = SPDK_BLOBID_INVALID;
+
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+	blob->extent_rle_found = false;
+	blob->extent_table_found = false;
+	blob->active.num_pages = 1;
+	blob->active.pages = calloc(1, sizeof(*blob->active.pages));
+	if (!blob->active.pages) {
+		free(blob);
+		return NULL;
+	}
+
+	blob->active.pages[0] = bs_blobid_to_page(id);
+
+	TAILQ_INIT(&blob->xattrs);
+	TAILQ_INIT(&blob->xattrs_internal);
+	TAILQ_INIT(&blob->pending_persists);
+
+	return blob;
+}
+
+static void
+xattrs_free(struct spdk_xattr_tailq *xattrs)
+{
+	struct spdk_xattr	*xattr, *xattr_tmp;
+
+	TAILQ_FOREACH_SAFE(xattr, xattrs, link, xattr_tmp) {
+		TAILQ_REMOVE(xattrs, xattr, link);
+		free(xattr->name);
+		free(xattr->value);
+		free(xattr);
+	}
+}
+
+static void
+blob_free(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+	assert(TAILQ_EMPTY(&blob->pending_persists));
+
+	free(blob->active.extent_pages);
+	free(blob->clean.extent_pages);
+	free(blob->active.clusters);
+	free(blob->clean.clusters);
+	free(blob->active.pages);
+	free(blob->clean.pages);
+
+	xattrs_free(&blob->xattrs);
+	xattrs_free(&blob->xattrs_internal);
+
+	if (blob->back_bs_dev) {
+		blob->back_bs_dev->destroy(blob->back_bs_dev);
+	}
+
+	free(blob);
+}
+
+struct freeze_io_ctx {
+	struct spdk_bs_cpl cpl;
+	struct spdk_blob *blob;
+};
+
+static void
+blob_io_sync(struct spdk_io_channel_iter *i)
+{
+	spdk_for_each_channel_continue(i, 0);
+}
+
+static void
+blob_execute_queued_io(struct spdk_io_channel_iter *i)
+{
+	struct spdk_io_channel *_ch = spdk_io_channel_iter_get_channel(i);
+	struct spdk_bs_channel *ch = spdk_io_channel_get_ctx(_ch);
+	struct freeze_io_ctx *ctx = spdk_io_channel_iter_get_ctx(i);
+	struct spdk_bs_request_set	*set;
+	struct spdk_bs_user_op_args	*args;
+	spdk_bs_user_op_t *op, *tmp;
+
+	TAILQ_FOREACH_SAFE(op, &ch->queued_io, link, tmp) {
+		set = (struct spdk_bs_request_set *)op;
+		args = &set->u.user_op;
+
+		if (args->blob == ctx->blob) {
+			TAILQ_REMOVE(&ch->queued_io, op, link);
+			bs_user_op_execute(op);
+		}
+	}
+
+	spdk_for_each_channel_continue(i, 0);
+}
+
+static void
+blob_io_cpl(struct spdk_io_channel_iter *i, int status)
+{
+	struct freeze_io_ctx *ctx = spdk_io_channel_iter_get_ctx(i);
+
+	ctx->cpl.u.blob_basic.cb_fn(ctx->cpl.u.blob_basic.cb_arg, 0);
+
+	free(ctx);
+}
+
+static void
+blob_freeze_io(struct spdk_blob *blob, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct freeze_io_ctx *ctx;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	ctx->cpl.u.blob_basic.cb_fn = cb_fn;
+	ctx->cpl.u.blob_basic.cb_arg = cb_arg;
+	ctx->blob = blob;
+
+	/* Freeze I/O on blob */
+	blob->frozen_refcnt++;
+
+	if (blob->frozen_refcnt == 1) {
+		spdk_for_each_channel(blob->bs, blob_io_sync, ctx, blob_io_cpl);
+	} else {
+		cb_fn(cb_arg, 0);
+		free(ctx);
+	}
+}
+
+static void
+blob_unfreeze_io(struct spdk_blob *blob, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct freeze_io_ctx *ctx;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	ctx->cpl.u.blob_basic.cb_fn = cb_fn;
+	ctx->cpl.u.blob_basic.cb_arg = cb_arg;
+	ctx->blob = blob;
+
+	assert(blob->frozen_refcnt > 0);
+
+	blob->frozen_refcnt--;
+
+	if (blob->frozen_refcnt == 0) {
+		spdk_for_each_channel(blob->bs, blob_execute_queued_io, ctx, blob_io_cpl);
+	} else {
+		cb_fn(cb_arg, 0);
+		free(ctx);
+	}
+}
+
+static int
+blob_mark_clean(struct spdk_blob *blob)
+{
+	uint32_t *extent_pages = NULL;
+	uint64_t *clusters = NULL;
+	uint32_t *pages = NULL;
+
+	assert(blob != NULL);
+
+	if (blob->active.num_extent_pages) {
+		assert(blob->active.extent_pages);
+		extent_pages = calloc(blob->active.num_extent_pages, sizeof(*blob->active.extent_pages));
+		if (!extent_pages) {
+			return -ENOMEM;
+		}
+		memcpy(extent_pages, blob->active.extent_pages,
+		       blob->active.num_extent_pages * sizeof(*extent_pages));
+	}
+
+	if (blob->active.num_clusters) {
+		assert(blob->active.clusters);
+		clusters = calloc(blob->active.num_clusters, sizeof(*blob->active.clusters));
+		if (!clusters) {
+			free(extent_pages);
+			return -ENOMEM;
+		}
+		memcpy(clusters, blob->active.clusters, blob->active.num_clusters * sizeof(*blob->active.clusters));
+	}
+
+	if (blob->active.num_pages) {
+		assert(blob->active.pages);
+		pages = calloc(blob->active.num_pages, sizeof(*blob->active.pages));
+		if (!pages) {
+			free(extent_pages);
+			free(clusters);
+			return -ENOMEM;
+		}
+		memcpy(pages, blob->active.pages, blob->active.num_pages * sizeof(*blob->active.pages));
+	}
+
+	free(blob->clean.extent_pages);
+	free(blob->clean.clusters);
+	free(blob->clean.pages);
+
+	blob->clean.num_extent_pages = blob->active.num_extent_pages;
+	blob->clean.extent_pages = blob->active.extent_pages;
+	blob->clean.num_clusters = blob->active.num_clusters;
+	blob->clean.clusters = blob->active.clusters;
+	blob->clean.num_pages = blob->active.num_pages;
+	blob->clean.pages = blob->active.pages;
+
+	blob->active.extent_pages = extent_pages;
+	blob->active.clusters = clusters;
+	blob->active.pages = pages;
+
+	/* If the metadata was dirtied again while the metadata was being written to disk,
+	 *  we do not want to revert the DIRTY state back to CLEAN here.
+	 */
+	if (blob->state == SPDK_BLOB_STATE_LOADING) {
+		blob->state = SPDK_BLOB_STATE_CLEAN;
+	}
+
+	return 0;
+}
+
+static int
+blob_deserialize_xattr(struct spdk_blob *blob,
+		       struct spdk_blob_md_descriptor_xattr *desc_xattr, bool internal)
+{
+	struct spdk_xattr                       *xattr;
+
+	if (desc_xattr->length != sizeof(desc_xattr->name_length) +
+	    sizeof(desc_xattr->value_length) +
+	    desc_xattr->name_length + desc_xattr->value_length) {
+		return -EINVAL;
+	}
+
+	xattr = calloc(1, sizeof(*xattr));
+	if (xattr == NULL) {
+		return -ENOMEM;
+	}
+
+	xattr->name = malloc(desc_xattr->name_length + 1);
+	if (xattr->name == NULL) {
+		free(xattr);
+		return -ENOMEM;
+	}
+	memcpy(xattr->name, desc_xattr->name, desc_xattr->name_length);
+	xattr->name[desc_xattr->name_length] = '\0';
+
+	xattr->value = malloc(desc_xattr->value_length);
+	if (xattr->value == NULL) {
+		free(xattr->name);
+		free(xattr);
+		return -ENOMEM;
+	}
+	xattr->value_len = desc_xattr->value_length;
+	memcpy(xattr->value,
+	       (void *)((uintptr_t)desc_xattr->name + desc_xattr->name_length),
+	       desc_xattr->value_length);
+
+	TAILQ_INSERT_TAIL(internal ? &blob->xattrs_internal : &blob->xattrs, xattr, link);
+
+	return 0;
+}
+
+
+static int
+blob_parse_page(const struct spdk_blob_md_page *page, struct spdk_blob *blob)
+{
+	struct spdk_blob_md_descriptor *desc;
+	size_t	cur_desc = 0;
+	void *tmp;
+
+	desc = (struct spdk_blob_md_descriptor *)page->descriptors;
+	while (cur_desc < sizeof(page->descriptors)) {
+		if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_PADDING) {
+			if (desc->length == 0) {
+				/* If padding and length are 0, this terminates the page */
+				break;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_FLAGS) {
+			struct spdk_blob_md_descriptor_flags	*desc_flags;
+
+			desc_flags = (struct spdk_blob_md_descriptor_flags *)desc;
+
+			if (desc_flags->length != sizeof(*desc_flags) - sizeof(*desc)) {
+				return -EINVAL;
+			}
+
+			if ((desc_flags->invalid_flags | SPDK_BLOB_INVALID_FLAGS_MASK) !=
+			    SPDK_BLOB_INVALID_FLAGS_MASK) {
+				return -EINVAL;
+			}
+
+			if ((desc_flags->data_ro_flags | SPDK_BLOB_DATA_RO_FLAGS_MASK) !=
+			    SPDK_BLOB_DATA_RO_FLAGS_MASK) {
+				blob->data_ro = true;
+				blob->md_ro = true;
+			}
+
+			if ((desc_flags->md_ro_flags | SPDK_BLOB_MD_RO_FLAGS_MASK) !=
+			    SPDK_BLOB_MD_RO_FLAGS_MASK) {
+				blob->md_ro = true;
+			}
+
+			if ((desc_flags->data_ro_flags & SPDK_BLOB_READ_ONLY)) {
+				blob->data_ro = true;
+				blob->md_ro = true;
+			}
+
+			blob->invalid_flags = desc_flags->invalid_flags;
+			blob->data_ro_flags = desc_flags->data_ro_flags;
+			blob->md_ro_flags = desc_flags->md_ro_flags;
+
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_RLE) {
+			struct spdk_blob_md_descriptor_extent_rle	*desc_extent_rle;
+			unsigned int				i, j;
+			unsigned int				cluster_count = blob->active.num_clusters;
+
+			if (blob->extent_table_found) {
+				/* Extent Table already present in the md,
+				 * both descriptors should never be at the same time. */
+				return -EINVAL;
+			}
+			blob->extent_rle_found = true;
+
+			desc_extent_rle = (struct spdk_blob_md_descriptor_extent_rle *)desc;
+
+			if (desc_extent_rle->length == 0 ||
+			    (desc_extent_rle->length % sizeof(desc_extent_rle->extents[0]) != 0)) {
+				return -EINVAL;
+			}
+
+			for (i = 0; i < desc_extent_rle->length / sizeof(desc_extent_rle->extents[0]); i++) {
+				for (j = 0; j < desc_extent_rle->extents[i].length; j++) {
+					if (desc_extent_rle->extents[i].cluster_idx != 0) {
+						if (!spdk_bit_array_get(blob->bs->used_clusters,
+									desc_extent_rle->extents[i].cluster_idx + j)) {
+							return -EINVAL;
+						}
+					}
+					cluster_count++;
+				}
+			}
+
+			if (cluster_count == 0) {
+				return -EINVAL;
+			}
+			tmp = realloc(blob->active.clusters, cluster_count * sizeof(*blob->active.clusters));
+			if (tmp == NULL) {
+				return -ENOMEM;
+			}
+			blob->active.clusters = tmp;
+			blob->active.cluster_array_size = cluster_count;
+
+			for (i = 0; i < desc_extent_rle->length / sizeof(desc_extent_rle->extents[0]); i++) {
+				for (j = 0; j < desc_extent_rle->extents[i].length; j++) {
+					if (desc_extent_rle->extents[i].cluster_idx != 0) {
+						blob->active.clusters[blob->active.num_clusters++] = bs_cluster_to_lba(blob->bs,
+								desc_extent_rle->extents[i].cluster_idx + j);
+					} else if (spdk_blob_is_thin_provisioned(blob)) {
+						blob->active.clusters[blob->active.num_clusters++] = 0;
+					} else {
+						return -EINVAL;
+					}
+				}
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_TABLE) {
+			struct spdk_blob_md_descriptor_extent_table *desc_extent_table;
+			uint32_t num_extent_pages = blob->active.num_extent_pages;
+			uint32_t i, j;
+			size_t extent_pages_length;
+
+			desc_extent_table = (struct spdk_blob_md_descriptor_extent_table *)desc;
+			extent_pages_length = desc_extent_table->length - sizeof(desc_extent_table->num_clusters);
+
+			if (blob->extent_rle_found) {
+				/* This means that Extent RLE is present in MD,
+				 * both should never be at the same time. */
+				return -EINVAL;
+			} else if (blob->extent_table_found &&
+				   desc_extent_table->num_clusters != blob->remaining_clusters_in_et) {
+				/* Number of clusters in this ET does not match number
+				 * from previously read EXTENT_TABLE. */
+				return -EINVAL;
+			}
+
+			blob->extent_table_found = true;
+
+			if (desc_extent_table->length == 0 ||
+			    (extent_pages_length % sizeof(desc_extent_table->extent_page[0]) != 0)) {
+				return -EINVAL;
+			}
+
+			for (i = 0; i < extent_pages_length / sizeof(desc_extent_table->extent_page[0]); i++) {
+				num_extent_pages += desc_extent_table->extent_page[i].num_pages;
+			}
+
+			tmp = realloc(blob->active.extent_pages, num_extent_pages * sizeof(uint32_t));
+			if (tmp == NULL) {
+				return -ENOMEM;
+			}
+			blob->active.extent_pages = tmp;
+			blob->active.extent_pages_array_size = num_extent_pages;
+
+			blob->remaining_clusters_in_et = desc_extent_table->num_clusters;
+
+			/* Extent table entries contain md page numbers for extent pages.
+			 * Zeroes represent unallocated extent pages, those are run-length-encoded.
+			 */
+			for (i = 0; i < extent_pages_length / sizeof(desc_extent_table->extent_page[0]); i++) {
+				if (desc_extent_table->extent_page[i].page_idx != 0) {
+					assert(desc_extent_table->extent_page[i].num_pages == 1);
+					blob->active.extent_pages[blob->active.num_extent_pages++] =
+						desc_extent_table->extent_page[i].page_idx;
+				} else if (spdk_blob_is_thin_provisioned(blob)) {
+					for (j = 0; j < desc_extent_table->extent_page[i].num_pages; j++) {
+						blob->active.extent_pages[blob->active.num_extent_pages++] = 0;
+					}
+				} else {
+					return -EINVAL;
+				}
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE) {
+			struct spdk_blob_md_descriptor_extent_page	*desc_extent;
+			unsigned int					i;
+			unsigned int					cluster_count = 0;
+			size_t						cluster_idx_length;
+
+			if (blob->extent_rle_found) {
+				/* This means that Extent RLE is present in MD,
+				 * both should never be at the same time. */
+				return -EINVAL;
+			}
+
+			desc_extent = (struct spdk_blob_md_descriptor_extent_page *)desc;
+			cluster_idx_length = desc_extent->length - sizeof(desc_extent->start_cluster_idx);
+
+			if (desc_extent->length <= sizeof(desc_extent->start_cluster_idx) ||
+			    (cluster_idx_length % sizeof(desc_extent->cluster_idx[0]) != 0)) {
+				return -EINVAL;
+			}
+
+			for (i = 0; i < cluster_idx_length / sizeof(desc_extent->cluster_idx[0]); i++) {
+				if (desc_extent->cluster_idx[i] != 0) {
+					if (!spdk_bit_array_get(blob->bs->used_clusters, desc_extent->cluster_idx[i])) {
+						return -EINVAL;
+					}
+				}
+				cluster_count++;
+			}
+
+			if (cluster_count == 0) {
+				return -EINVAL;
+			}
+
+			/* When reading extent pages sequentially starting cluster idx should match
+			 * current size of a blob.
+			 * If changed to batch reading, this check shall be removed. */
+			if (desc_extent->start_cluster_idx != blob->active.num_clusters) {
+				return -EINVAL;
+			}
+
+			tmp = realloc(blob->active.clusters,
+				      (cluster_count + blob->active.num_clusters) * sizeof(*blob->active.clusters));
+			if (tmp == NULL) {
+				return -ENOMEM;
+			}
+			blob->active.clusters = tmp;
+			blob->active.cluster_array_size = (cluster_count + blob->active.num_clusters);
+
+			for (i = 0; i < cluster_idx_length / sizeof(desc_extent->cluster_idx[0]); i++) {
+				if (desc_extent->cluster_idx[i] != 0) {
+					blob->active.clusters[blob->active.num_clusters++] = bs_cluster_to_lba(blob->bs,
+							desc_extent->cluster_idx[i]);
+				} else if (spdk_blob_is_thin_provisioned(blob)) {
+					blob->active.clusters[blob->active.num_clusters++] = 0;
+				} else {
+					return -EINVAL;
+				}
+			}
+			assert(desc_extent->start_cluster_idx + cluster_count == blob->active.num_clusters);
+			assert(blob->remaining_clusters_in_et >= cluster_count);
+			blob->remaining_clusters_in_et -= cluster_count;
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR) {
+			int rc;
+
+			rc = blob_deserialize_xattr(blob,
+						    (struct spdk_blob_md_descriptor_xattr *) desc, false);
+			if (rc != 0) {
+				return rc;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR_INTERNAL) {
+			int rc;
+
+			rc = blob_deserialize_xattr(blob,
+						    (struct spdk_blob_md_descriptor_xattr *) desc, true);
+			if (rc != 0) {
+				return rc;
+			}
+		} else {
+			/* Unrecognized descriptor type.  Do not fail - just continue to the
+			 *  next descriptor.  If this descriptor is associated with some feature
+			 *  defined in a newer version of blobstore, that version of blobstore
+			 *  should create and set an associated feature flag to specify if this
+			 *  blob can be loaded or not.
+			 */
+		}
+
+		/* Advance to the next descriptor */
+		cur_desc += sizeof(*desc) + desc->length;
+		if (cur_desc + sizeof(*desc) > sizeof(page->descriptors)) {
+			break;
+		}
+		desc = (struct spdk_blob_md_descriptor *)((uintptr_t)page->descriptors + cur_desc);
+	}
+
+	return 0;
+}
+
+static bool bs_load_cur_extent_page_valid(struct spdk_blob_md_page *page);
+
+static int
+blob_parse_extent_page(struct spdk_blob_md_page *extent_page, struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+	assert(blob->state == SPDK_BLOB_STATE_LOADING);
+
+	if (bs_load_cur_extent_page_valid(extent_page) == false) {
+		return -ENOENT;
+	}
+
+	return blob_parse_page(extent_page, blob);
+}
+
+static int
+blob_parse(const struct spdk_blob_md_page *pages, uint32_t page_count,
+	   struct spdk_blob *blob)
+{
+	const struct spdk_blob_md_page *page;
+	uint32_t i;
+	int rc;
+
+	assert(page_count > 0);
+	assert(pages[0].sequence_num == 0);
+	assert(blob != NULL);
+	assert(blob->state == SPDK_BLOB_STATE_LOADING);
+	assert(blob->active.clusters == NULL);
+
+	/* The blobid provided doesn't match what's in the MD, this can
+	 * happen for example if a bogus blobid is passed in through open.
+	 */
+	if (blob->id != pages[0].id) {
+		SPDK_ERRLOG("Blobid (%lu) doesn't match what's in metadata (%lu)\n",
+			    blob->id, pages[0].id);
+		return -ENOENT;
+	}
+
+	for (i = 0; i < page_count; i++) {
+		page = &pages[i];
+
+		assert(page->id == blob->id);
+		assert(page->sequence_num == i);
+
+		rc = blob_parse_page(page, blob);
+		if (rc != 0) {
+			return rc;
+		}
+	}
+
+	return 0;
+}
+
+static int
+blob_serialize_add_page(const struct spdk_blob *blob,
+			struct spdk_blob_md_page **pages,
+			uint32_t *page_count,
+			struct spdk_blob_md_page **last_page)
+{
+	struct spdk_blob_md_page *page;
+
+	assert(pages != NULL);
+	assert(page_count != NULL);
+
+	if (*page_count == 0) {
+		assert(*pages == NULL);
+		*page_count = 1;
+		*pages = spdk_malloc(SPDK_BS_PAGE_SIZE, SPDK_BS_PAGE_SIZE,
+				     NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	} else {
+		assert(*pages != NULL);
+		(*page_count)++;
+		*pages = spdk_realloc(*pages,
+				      SPDK_BS_PAGE_SIZE * (*page_count),
+				      SPDK_BS_PAGE_SIZE);
+	}
+
+	if (*pages == NULL) {
+		*page_count = 0;
+		*last_page = NULL;
+		return -ENOMEM;
+	}
+
+	page = &(*pages)[*page_count - 1];
+	memset(page, 0, sizeof(*page));
+	page->id = blob->id;
+	page->sequence_num = *page_count - 1;
+	page->next = SPDK_INVALID_MD_PAGE;
+	*last_page = page;
+
+	return 0;
+}
+
+/* Transform the in-memory representation 'xattr' into an on-disk xattr descriptor.
+ * Update required_sz on both success and failure.
+ *
+ */
+static int
+blob_serialize_xattr(const struct spdk_xattr *xattr,
+		     uint8_t *buf, size_t buf_sz,
+		     size_t *required_sz, bool internal)
+{
+	struct spdk_blob_md_descriptor_xattr	*desc;
+
+	*required_sz = sizeof(struct spdk_blob_md_descriptor_xattr) +
+		       strlen(xattr->name) +
+		       xattr->value_len;
+
+	if (buf_sz < *required_sz) {
+		return -1;
+	}
+
+	desc = (struct spdk_blob_md_descriptor_xattr *)buf;
+
+	desc->type = internal ? SPDK_MD_DESCRIPTOR_TYPE_XATTR_INTERNAL : SPDK_MD_DESCRIPTOR_TYPE_XATTR;
+	desc->length = sizeof(desc->name_length) +
+		       sizeof(desc->value_length) +
+		       strlen(xattr->name) +
+		       xattr->value_len;
+	desc->name_length = strlen(xattr->name);
+	desc->value_length = xattr->value_len;
+
+	memcpy(desc->name, xattr->name, desc->name_length);
+	memcpy((void *)((uintptr_t)desc->name + desc->name_length),
+	       xattr->value,
+	       desc->value_length);
+
+	return 0;
+}
+
+static void
+blob_serialize_extent_table_entry(const struct spdk_blob *blob,
+				  uint64_t start_ep, uint64_t *next_ep,
+				  uint8_t **buf, size_t *remaining_sz)
+{
+	struct spdk_blob_md_descriptor_extent_table *desc;
+	size_t cur_sz;
+	uint64_t i, et_idx;
+	uint32_t extent_page, ep_len;
+
+	/* The buffer must have room for at least num_clusters entry */
+	cur_sz = sizeof(struct spdk_blob_md_descriptor) + sizeof(desc->num_clusters);
+	if (*remaining_sz < cur_sz) {
+		*next_ep = start_ep;
+		return;
+	}
+
+	desc = (struct spdk_blob_md_descriptor_extent_table *)*buf;
+	desc->type = SPDK_MD_DESCRIPTOR_TYPE_EXTENT_TABLE;
+
+	desc->num_clusters = blob->active.num_clusters;
+
+	ep_len = 1;
+	et_idx = 0;
+	for (i = start_ep; i < blob->active.num_extent_pages; i++) {
+		if (*remaining_sz < cur_sz  + sizeof(desc->extent_page[0])) {
+			/* If we ran out of buffer space, return */
+			break;
+		}
+
+		extent_page = blob->active.extent_pages[i];
+		/* Verify that next extent_page is unallocated */
+		if (extent_page == 0 &&
+		    (i + 1 < blob->active.num_extent_pages && blob->active.extent_pages[i + 1] == 0)) {
+			ep_len++;
+			continue;
+		}
+		desc->extent_page[et_idx].page_idx = extent_page;
+		desc->extent_page[et_idx].num_pages = ep_len;
+		et_idx++;
+
+		ep_len = 1;
+		cur_sz += sizeof(desc->extent_page[et_idx]);
+	}
+	*next_ep = i;
+
+	desc->length = sizeof(desc->num_clusters) + sizeof(desc->extent_page[0]) * et_idx;
+	*remaining_sz -= sizeof(struct spdk_blob_md_descriptor) + desc->length;
+	*buf += sizeof(struct spdk_blob_md_descriptor) + desc->length;
+}
+
+static int
+blob_serialize_extent_table(const struct spdk_blob *blob,
+			    struct spdk_blob_md_page **pages,
+			    struct spdk_blob_md_page *cur_page,
+			    uint32_t *page_count, uint8_t **buf,
+			    size_t *remaining_sz)
+{
+	uint64_t				last_extent_page;
+	int					rc;
+
+	last_extent_page = 0;
+	/* At least single extent table entry has to be always persisted.
+	 * Such case occurs with num_extent_pages == 0. */
+	while (last_extent_page <= blob->active.num_extent_pages) {
+		blob_serialize_extent_table_entry(blob, last_extent_page, &last_extent_page, buf,
+						  remaining_sz);
+
+		if (last_extent_page == blob->active.num_extent_pages) {
+			break;
+		}
+
+		rc = blob_serialize_add_page(blob, pages, page_count, &cur_page);
+		if (rc < 0) {
+			return rc;
+		}
+
+		*buf = (uint8_t *)cur_page->descriptors;
+		*remaining_sz = sizeof(cur_page->descriptors);
+	}
+
+	return 0;
+}
+
+static void
+blob_serialize_extent_rle(const struct spdk_blob *blob,
+			  uint64_t start_cluster, uint64_t *next_cluster,
+			  uint8_t **buf, size_t *buf_sz)
+{
+	struct spdk_blob_md_descriptor_extent_rle *desc_extent_rle;
+	size_t cur_sz;
+	uint64_t i, extent_idx;
+	uint64_t lba, lba_per_cluster, lba_count;
+
+	/* The buffer must have room for at least one extent */
+	cur_sz = sizeof(struct spdk_blob_md_descriptor) + sizeof(desc_extent_rle->extents[0]);
+	if (*buf_sz < cur_sz) {
+		*next_cluster = start_cluster;
+		return;
+	}
+
+	desc_extent_rle = (struct spdk_blob_md_descriptor_extent_rle *)*buf;
+	desc_extent_rle->type = SPDK_MD_DESCRIPTOR_TYPE_EXTENT_RLE;
+
+	lba_per_cluster = bs_cluster_to_lba(blob->bs, 1);
+
+	lba = blob->active.clusters[start_cluster];
+	lba_count = lba_per_cluster;
+	extent_idx = 0;
+	for (i = start_cluster + 1; i < blob->active.num_clusters; i++) {
+		if ((lba + lba_count) == blob->active.clusters[i] && lba != 0) {
+			/* Run-length encode sequential non-zero LBA */
+			lba_count += lba_per_cluster;
+			continue;
+		} else if (lba == 0 && blob->active.clusters[i] == 0) {
+			/* Run-length encode unallocated clusters */
+			lba_count += lba_per_cluster;
+			continue;
+		}
+		desc_extent_rle->extents[extent_idx].cluster_idx = lba / lba_per_cluster;
+		desc_extent_rle->extents[extent_idx].length = lba_count / lba_per_cluster;
+		extent_idx++;
+
+		cur_sz += sizeof(desc_extent_rle->extents[extent_idx]);
+
+		if (*buf_sz < cur_sz) {
+			/* If we ran out of buffer space, return */
+			*next_cluster = i;
+			break;
+		}
+
+		lba = blob->active.clusters[i];
+		lba_count = lba_per_cluster;
+	}
+
+	if (*buf_sz >= cur_sz) {
+		desc_extent_rle->extents[extent_idx].cluster_idx = lba / lba_per_cluster;
+		desc_extent_rle->extents[extent_idx].length = lba_count / lba_per_cluster;
+		extent_idx++;
+
+		*next_cluster = blob->active.num_clusters;
+	}
+
+	desc_extent_rle->length = sizeof(desc_extent_rle->extents[0]) * extent_idx;
+	*buf_sz -= sizeof(struct spdk_blob_md_descriptor) + desc_extent_rle->length;
+	*buf += sizeof(struct spdk_blob_md_descriptor) + desc_extent_rle->length;
+}
+
+static int
+blob_serialize_extents_rle(const struct spdk_blob *blob,
+			   struct spdk_blob_md_page **pages,
+			   struct spdk_blob_md_page *cur_page,
+			   uint32_t *page_count, uint8_t **buf,
+			   size_t *remaining_sz)
+{
+	uint64_t				last_cluster;
+	int					rc;
+
+	last_cluster = 0;
+	while (last_cluster < blob->active.num_clusters) {
+		blob_serialize_extent_rle(blob, last_cluster, &last_cluster, buf, remaining_sz);
+
+		if (last_cluster == blob->active.num_clusters) {
+			break;
+		}
+
+		rc = blob_serialize_add_page(blob, pages, page_count, &cur_page);
+		if (rc < 0) {
+			return rc;
+		}
+
+		*buf = (uint8_t *)cur_page->descriptors;
+		*remaining_sz = sizeof(cur_page->descriptors);
+	}
+
+	return 0;
+}
+
+static void
+blob_serialize_extent_page(const struct spdk_blob *blob,
+			   uint64_t cluster, struct spdk_blob_md_page *page)
+{
+	struct spdk_blob_md_descriptor_extent_page *desc_extent;
+	uint64_t i, extent_idx;
+	uint64_t lba, lba_per_cluster;
+	uint64_t start_cluster_idx = (cluster / SPDK_EXTENTS_PER_EP) * SPDK_EXTENTS_PER_EP;
+
+	desc_extent = (struct spdk_blob_md_descriptor_extent_page *) page->descriptors;
+	desc_extent->type = SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE;
+
+	lba_per_cluster = bs_cluster_to_lba(blob->bs, 1);
+
+	desc_extent->start_cluster_idx = start_cluster_idx;
+	extent_idx = 0;
+	for (i = start_cluster_idx; i < blob->active.num_clusters; i++) {
+		lba = blob->active.clusters[i];
+		desc_extent->cluster_idx[extent_idx++] = lba / lba_per_cluster;
+		if (extent_idx >= SPDK_EXTENTS_PER_EP) {
+			break;
+		}
+	}
+	desc_extent->length = sizeof(desc_extent->start_cluster_idx) +
+			      sizeof(desc_extent->cluster_idx[0]) * extent_idx;
+}
+
+static void
+blob_serialize_flags(const struct spdk_blob *blob,
+		     uint8_t *buf, size_t *buf_sz)
+{
+	struct spdk_blob_md_descriptor_flags *desc;
+
+	/*
+	 * Flags get serialized first, so we should always have room for the flags
+	 *  descriptor.
+	 */
+	assert(*buf_sz >= sizeof(*desc));
+
+	desc = (struct spdk_blob_md_descriptor_flags *)buf;
+	desc->type = SPDK_MD_DESCRIPTOR_TYPE_FLAGS;
+	desc->length = sizeof(*desc) - sizeof(struct spdk_blob_md_descriptor);
+	desc->invalid_flags = blob->invalid_flags;
+	desc->data_ro_flags = blob->data_ro_flags;
+	desc->md_ro_flags = blob->md_ro_flags;
+
+	*buf_sz -= sizeof(*desc);
+}
+
+static int
+blob_serialize_xattrs(const struct spdk_blob *blob,
+		      const struct spdk_xattr_tailq *xattrs, bool internal,
+		      struct spdk_blob_md_page **pages,
+		      struct spdk_blob_md_page *cur_page,
+		      uint32_t *page_count, uint8_t **buf,
+		      size_t *remaining_sz)
+{
+	const struct spdk_xattr	*xattr;
+	int	rc;
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		size_t required_sz = 0;
+
+		rc = blob_serialize_xattr(xattr,
+					  *buf, *remaining_sz,
+					  &required_sz, internal);
+		if (rc < 0) {
+			/* Need to add a new page to the chain */
+			rc = blob_serialize_add_page(blob, pages, page_count,
+						     &cur_page);
+			if (rc < 0) {
+				spdk_free(*pages);
+				*pages = NULL;
+				*page_count = 0;
+				return rc;
+			}
+
+			*buf = (uint8_t *)cur_page->descriptors;
+			*remaining_sz = sizeof(cur_page->descriptors);
+
+			/* Try again */
+			required_sz = 0;
+			rc = blob_serialize_xattr(xattr,
+						  *buf, *remaining_sz,
+						  &required_sz, internal);
+
+			if (rc < 0) {
+				spdk_free(*pages);
+				*pages = NULL;
+				*page_count = 0;
+				return rc;
+			}
+		}
+
+		*remaining_sz -= required_sz;
+		*buf += required_sz;
+	}
+
+	return 0;
+}
+
+static int
+blob_serialize(const struct spdk_blob *blob, struct spdk_blob_md_page **pages,
+	       uint32_t *page_count)
+{
+	struct spdk_blob_md_page		*cur_page;
+	int					rc;
+	uint8_t					*buf;
+	size_t					remaining_sz;
+
+	assert(pages != NULL);
+	assert(page_count != NULL);
+	assert(blob != NULL);
+	assert(blob->state == SPDK_BLOB_STATE_DIRTY);
+
+	*pages = NULL;
+	*page_count = 0;
+
+	/* A blob always has at least 1 page, even if it has no descriptors */
+	rc = blob_serialize_add_page(blob, pages, page_count, &cur_page);
+	if (rc < 0) {
+		return rc;
+	}
+
+	buf = (uint8_t *)cur_page->descriptors;
+	remaining_sz = sizeof(cur_page->descriptors);
+
+	/* Serialize flags */
+	blob_serialize_flags(blob, buf, &remaining_sz);
+	buf += sizeof(struct spdk_blob_md_descriptor_flags);
+
+	/* Serialize xattrs */
+	rc = blob_serialize_xattrs(blob, &blob->xattrs, false,
+				   pages, cur_page, page_count, &buf, &remaining_sz);
+	if (rc < 0) {
+		return rc;
+	}
+
+	/* Serialize internal xattrs */
+	rc = blob_serialize_xattrs(blob, &blob->xattrs_internal, true,
+				   pages, cur_page, page_count, &buf, &remaining_sz);
+	if (rc < 0) {
+		return rc;
+	}
+
+	if (blob->use_extent_table) {
+		/* Serialize extent table */
+		rc = blob_serialize_extent_table(blob, pages, cur_page, page_count, &buf, &remaining_sz);
+	} else {
+		/* Serialize extents */
+		rc = blob_serialize_extents_rle(blob, pages, cur_page, page_count, &buf, &remaining_sz);
+	}
+
+	return rc;
+}
+
+struct spdk_blob_load_ctx {
+	struct spdk_blob		*blob;
+
+	struct spdk_blob_md_page	*pages;
+	uint32_t			num_pages;
+	uint32_t			next_extent_page;
+	spdk_bs_sequence_t	        *seq;
+
+	spdk_bs_sequence_cpl		cb_fn;
+	void				*cb_arg;
+};
+
+static uint32_t
+blob_md_page_calc_crc(void *page)
+{
+	uint32_t		crc;
+
+	crc = BLOB_CRC32C_INITIAL;
+	crc = spdk_crc32c_update(page, SPDK_BS_PAGE_SIZE - 4, crc);
+	crc ^= BLOB_CRC32C_INITIAL;
+
+	return crc;
+
+}
+
+static void
+blob_load_final(void *cb_arg, int bserrno)
+{
+	struct spdk_blob_load_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+
+	if (bserrno == 0) {
+		blob_mark_clean(blob);
+	}
+
+	ctx->cb_fn(ctx->seq, ctx->cb_arg, bserrno);
+
+	/* Free the memory */
+	spdk_free(ctx->pages);
+	free(ctx);
+}
+
+static void
+blob_load_snapshot_cpl(void *cb_arg, struct spdk_blob *snapshot, int bserrno)
+{
+	struct spdk_blob_load_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+
+	if (bserrno == 0) {
+		blob->back_bs_dev = bs_create_blob_bs_dev(snapshot);
+		if (blob->back_bs_dev == NULL) {
+			bserrno = -ENOMEM;
+		}
+	}
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Snapshot fail\n");
+	}
+
+	blob_load_final(ctx, bserrno);
+}
+
+static void blob_update_clear_method(struct spdk_blob *blob);
+
+static void
+blob_load_backing_dev(void *cb_arg)
+{
+	struct spdk_blob_load_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	const void			*value;
+	size_t				len;
+	int				rc;
+
+	if (spdk_blob_is_thin_provisioned(blob)) {
+		rc = blob_get_xattr_value(blob, BLOB_SNAPSHOT, &value, &len, true);
+		if (rc == 0) {
+			if (len != sizeof(spdk_blob_id)) {
+				blob_load_final(ctx, -EINVAL);
+				return;
+			}
+			/* open snapshot blob and continue in the callback function */
+			blob->parent_id = *(spdk_blob_id *)value;
+			spdk_bs_open_blob(blob->bs, blob->parent_id,
+					  blob_load_snapshot_cpl, ctx);
+			return;
+		} else {
+			/* add zeroes_dev for thin provisioned blob */
+			blob->back_bs_dev = bs_create_zeroes_dev();
+		}
+	} else {
+		/* standard blob */
+		blob->back_bs_dev = NULL;
+	}
+	blob_load_final(ctx, 0);
+}
+
+static void
+blob_load_cpl_extents_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_load_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_md_page	*page;
+	uint64_t			i;
+	uint32_t			crc;
+	uint64_t			lba;
+	void				*tmp;
+	uint64_t			sz;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Extent page read failed: %d\n", bserrno);
+		blob_load_final(ctx, bserrno);
+		return;
+	}
+
+	if (ctx->pages == NULL) {
+		/* First iteration of this function, allocate buffer for single EXTENT_PAGE */
+		ctx->pages = spdk_zmalloc(SPDK_BS_PAGE_SIZE, SPDK_BS_PAGE_SIZE, NULL, SPDK_ENV_SOCKET_ID_ANY,
+					  SPDK_MALLOC_DMA);
+		if (!ctx->pages) {
+			blob_load_final(ctx, -ENOMEM);
+			return;
+		}
+		ctx->num_pages = 1;
+		ctx->next_extent_page = 0;
+	} else {
+		page = &ctx->pages[0];
+		crc = blob_md_page_calc_crc(page);
+		if (crc != page->crc) {
+			blob_load_final(ctx, -EINVAL);
+			return;
+		}
+
+		if (page->next != SPDK_INVALID_MD_PAGE) {
+			blob_load_final(ctx, -EINVAL);
+			return;
+		}
+
+		bserrno = blob_parse_extent_page(page, blob);
+		if (bserrno) {
+			blob_load_final(ctx, bserrno);
+			return;
+		}
+	}
+
+	for (i = ctx->next_extent_page; i < blob->active.num_extent_pages; i++) {
+		if (blob->active.extent_pages[i] != 0) {
+			/* Extent page was allocated, read and parse it. */
+			lba = bs_md_page_to_lba(blob->bs, blob->active.extent_pages[i]);
+			ctx->next_extent_page = i + 1;
+
+			bs_sequence_read_dev(seq, &ctx->pages[0], lba,
+					     bs_byte_to_lba(blob->bs, SPDK_BS_PAGE_SIZE),
+					     blob_load_cpl_extents_cpl, ctx);
+			return;
+		} else {
+			/* Thin provisioned blobs can point to unallocated extent pages.
+			 * In this case blob size should be increased by up to the amount left in remaining_clusters_in_et. */
+
+			sz = spdk_min(blob->remaining_clusters_in_et, SPDK_EXTENTS_PER_EP);
+			blob->active.num_clusters += sz;
+			blob->remaining_clusters_in_et -= sz;
+
+			assert(spdk_blob_is_thin_provisioned(blob));
+			assert(i + 1 < blob->active.num_extent_pages || blob->remaining_clusters_in_et == 0);
+
+			tmp = realloc(blob->active.clusters, blob->active.num_clusters * sizeof(*blob->active.clusters));
+			if (tmp == NULL) {
+				blob_load_final(ctx, -ENOMEM);
+				return;
+			}
+			memset(tmp + sizeof(*blob->active.clusters) * blob->active.cluster_array_size, 0,
+			       sizeof(*blob->active.clusters) * (blob->active.num_clusters - blob->active.cluster_array_size));
+			blob->active.clusters = tmp;
+			blob->active.cluster_array_size = blob->active.num_clusters;
+		}
+	}
+
+	blob_load_backing_dev(ctx);
+}
+
+static void
+blob_load_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_load_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_md_page	*page;
+	int				rc;
+	uint32_t			crc;
+	uint32_t			current_page;
+
+	if (ctx->num_pages == 1) {
+		current_page = bs_blobid_to_page(blob->id);
+	} else {
+		assert(ctx->num_pages != 0);
+		page = &ctx->pages[ctx->num_pages - 2];
+		current_page = page->next;
+	}
+
+	if (bserrno) {
+		SPDK_ERRLOG("Metadata page %d read failed for blobid %lu: %d\n",
+			    current_page, blob->id, bserrno);
+		blob_load_final(ctx, bserrno);
+		return;
+	}
+
+	page = &ctx->pages[ctx->num_pages - 1];
+	crc = blob_md_page_calc_crc(page);
+	if (crc != page->crc) {
+		SPDK_ERRLOG("Metadata page %d crc mismatch for blobid %lu\n",
+			    current_page, blob->id);
+		blob_load_final(ctx, -EINVAL);
+		return;
+	}
+
+	if (page->next != SPDK_INVALID_MD_PAGE) {
+		uint32_t next_page = page->next;
+		uint64_t next_lba = bs_md_page_to_lba(blob->bs, next_page);
+
+		/* Read the next page */
+		ctx->num_pages++;
+		ctx->pages = spdk_realloc(ctx->pages, (sizeof(*page) * ctx->num_pages),
+					  sizeof(*page));
+		if (ctx->pages == NULL) {
+			blob_load_final(ctx, -ENOMEM);
+			return;
+		}
+
+		bs_sequence_read_dev(seq, &ctx->pages[ctx->num_pages - 1],
+				     next_lba,
+				     bs_byte_to_lba(blob->bs, sizeof(*page)),
+				     blob_load_cpl, ctx);
+		return;
+	}
+
+	/* Parse the pages */
+	rc = blob_parse(ctx->pages, ctx->num_pages, blob);
+	if (rc) {
+		blob_load_final(ctx, rc);
+		return;
+	}
+
+	if (blob->extent_table_found == true) {
+		/* If EXTENT_TABLE was found, that means support for it should be enabled. */
+		assert(blob->extent_rle_found == false);
+		blob->use_extent_table = true;
+	} else {
+		/* If EXTENT_RLE or no extent_* descriptor was found disable support
+		 * for extent table. No extent_* descriptors means that blob has length of 0
+		 * and no extent_rle descriptors were persisted for it.
+		 * EXTENT_TABLE if used, is always present in metadata regardless of length. */
+		blob->use_extent_table = false;
+	}
+
+	/* Check the clear_method stored in metadata vs what may have been passed
+	 * via spdk_bs_open_blob_ext() and update accordingly.
+	 */
+	blob_update_clear_method(blob);
+
+	spdk_free(ctx->pages);
+	ctx->pages = NULL;
+
+	if (blob->extent_table_found) {
+		blob_load_cpl_extents_cpl(seq, ctx, 0);
+	} else {
+		blob_load_backing_dev(ctx);
+	}
+}
+
+/* Load a blob from disk given a blobid */
+static void
+blob_load(spdk_bs_sequence_t *seq, struct spdk_blob *blob,
+	  spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_blob_load_ctx *ctx;
+	struct spdk_blob_store *bs;
+	uint32_t page_num;
+	uint64_t lba;
+
+	blob_verify_md_op(blob);
+
+	bs = blob->bs;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(seq, cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->blob = blob;
+	ctx->pages = spdk_realloc(ctx->pages, SPDK_BS_PAGE_SIZE, SPDK_BS_PAGE_SIZE);
+	if (!ctx->pages) {
+		free(ctx);
+		cb_fn(seq, cb_arg, -ENOMEM);
+		return;
+	}
+	ctx->num_pages = 1;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+	ctx->seq = seq;
+
+	page_num = bs_blobid_to_page(blob->id);
+	lba = bs_md_page_to_lba(blob->bs, page_num);
+
+	blob->state = SPDK_BLOB_STATE_LOADING;
+
+	bs_sequence_read_dev(seq, &ctx->pages[0], lba,
+			     bs_byte_to_lba(bs, SPDK_BS_PAGE_SIZE),
+			     blob_load_cpl, ctx);
+}
+
+struct spdk_blob_persist_ctx {
+	struct spdk_blob		*blob;
+
+	struct spdk_bs_super_block	*super;
+
+	struct spdk_blob_md_page	*pages;
+	uint32_t			next_extent_page;
+	struct spdk_blob_md_page	*extent_page;
+
+	spdk_bs_sequence_t		*seq;
+	spdk_bs_sequence_cpl		cb_fn;
+	void				*cb_arg;
+	TAILQ_ENTRY(spdk_blob_persist_ctx) link;
+};
+
+static void
+bs_batch_clear_dev(struct spdk_blob_persist_ctx *ctx, spdk_bs_batch_t *batch, uint64_t lba,
+		   uint32_t lba_count)
+{
+	switch (ctx->blob->clear_method) {
+	case BLOB_CLEAR_WITH_DEFAULT:
+	case BLOB_CLEAR_WITH_UNMAP:
+		bs_batch_unmap_dev(batch, lba, lba_count);
+		break;
+	case BLOB_CLEAR_WITH_WRITE_ZEROES:
+		bs_batch_write_zeroes_dev(batch, lba, lba_count);
+		break;
+	case BLOB_CLEAR_WITH_NONE:
+	default:
+		break;
+	}
+}
+
+static void blob_persist_check_dirty(struct spdk_blob_persist_ctx *ctx);
+
+static void
+blob_persist_complete(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob_persist_ctx	*next_persist;
+	struct spdk_blob		*blob = ctx->blob;
+
+	if (bserrno == 0) {
+		blob_mark_clean(blob);
+	}
+
+	assert(ctx == TAILQ_FIRST(&blob->pending_persists));
+	TAILQ_REMOVE(&blob->pending_persists, ctx, link);
+
+	next_persist = TAILQ_FIRST(&blob->pending_persists);
+
+	/* Call user callback */
+	ctx->cb_fn(seq, ctx->cb_arg, bserrno);
+
+	/* Free the memory */
+	spdk_free(ctx->pages);
+	free(ctx);
+
+	if (next_persist != NULL) {
+		blob_persist_check_dirty(next_persist);
+	}
+}
+
+static void
+blob_persist_clear_clusters_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	size_t				i;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	/* Release all clusters that were truncated */
+	for (i = blob->active.num_clusters; i < blob->active.cluster_array_size; i++) {
+		uint32_t cluster_num = bs_lba_to_cluster(bs, blob->active.clusters[i]);
+
+		/* Nothing to release if it was not allocated */
+		if (blob->active.clusters[i] != 0) {
+			bs_release_cluster(bs, cluster_num);
+		}
+	}
+
+	if (blob->active.num_clusters == 0) {
+		free(blob->active.clusters);
+		blob->active.clusters = NULL;
+		blob->active.cluster_array_size = 0;
+	} else if (blob->active.num_clusters != blob->active.cluster_array_size) {
+#ifndef __clang_analyzer__
+		void *tmp;
+
+		/* scan-build really can't figure reallocs, workaround it */
+		tmp = realloc(blob->active.clusters, sizeof(*blob->active.clusters) * blob->active.num_clusters);
+		assert(tmp != NULL);
+		blob->active.clusters = tmp;
+
+		tmp = realloc(blob->active.extent_pages, sizeof(uint32_t) * blob->active.num_extent_pages);
+		assert(tmp != NULL);
+		blob->active.extent_pages = tmp;
+#endif
+		blob->active.extent_pages_array_size = blob->active.num_extent_pages;
+		blob->active.cluster_array_size = blob->active.num_clusters;
+	}
+
+	/* TODO: Add path to persist clear extent pages. */
+	blob_persist_complete(seq, ctx, bserrno);
+}
+
+static void
+blob_persist_clear_clusters(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	spdk_bs_batch_t			*batch;
+	size_t				i;
+	uint64_t			lba;
+	uint32_t			lba_count;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	/* Clusters don't move around in blobs. The list shrinks or grows
+	 * at the end, but no changes ever occur in the middle of the list.
+	 */
+
+	batch = bs_sequence_to_batch(seq, blob_persist_clear_clusters_cpl, ctx);
+
+	/* Clear all clusters that were truncated */
+	lba = 0;
+	lba_count = 0;
+	for (i = blob->active.num_clusters; i < blob->active.cluster_array_size; i++) {
+		uint64_t next_lba = blob->active.clusters[i];
+		uint32_t next_lba_count = bs_cluster_to_lba(bs, 1);
+
+		if (next_lba > 0 && (lba + lba_count) == next_lba) {
+			/* This cluster is contiguous with the previous one. */
+			lba_count += next_lba_count;
+			continue;
+		}
+
+		/* This cluster is not contiguous with the previous one. */
+
+		/* If a run of LBAs previously existing, clear them now */
+		if (lba_count > 0) {
+			bs_batch_clear_dev(ctx, batch, lba, lba_count);
+		}
+
+		/* Start building the next batch */
+		lba = next_lba;
+		if (next_lba > 0) {
+			lba_count = next_lba_count;
+		} else {
+			lba_count = 0;
+		}
+	}
+
+	/* If we ended with a contiguous set of LBAs, clear them now */
+	if (lba_count > 0) {
+		bs_batch_clear_dev(ctx, batch, lba, lba_count);
+	}
+
+	bs_batch_close(batch);
+}
+
+static void
+blob_persist_zero_pages_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	size_t				i;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	/* This loop starts at 1 because the first page is special and handled
+	 * below. The pages (except the first) are never written in place,
+	 * so any pages in the clean list must be zeroed.
+	 */
+	for (i = 1; i < blob->clean.num_pages; i++) {
+		bs_release_md_page(bs, blob->clean.pages[i]);
+	}
+
+	if (blob->active.num_pages == 0) {
+		uint32_t page_num;
+
+		page_num = bs_blobid_to_page(blob->id);
+		bs_release_md_page(bs, page_num);
+	}
+
+	/* Move on to clearing clusters */
+	blob_persist_clear_clusters(seq, ctx, 0);
+}
+
+static void
+blob_persist_zero_pages(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	uint64_t			lba;
+	uint32_t			lba_count;
+	spdk_bs_batch_t			*batch;
+	size_t				i;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	batch = bs_sequence_to_batch(seq, blob_persist_zero_pages_cpl, ctx);
+
+	lba_count = bs_byte_to_lba(bs, SPDK_BS_PAGE_SIZE);
+
+	/* This loop starts at 1 because the first page is special and handled
+	 * below. The pages (except the first) are never written in place,
+	 * so any pages in the clean list must be zeroed.
+	 */
+	for (i = 1; i < blob->clean.num_pages; i++) {
+		lba = bs_md_page_to_lba(bs, blob->clean.pages[i]);
+
+		bs_batch_write_zeroes_dev(batch, lba, lba_count);
+	}
+
+	/* The first page will only be zeroed if this is a delete. */
+	if (blob->active.num_pages == 0) {
+		uint32_t page_num;
+
+		/* The first page in the metadata goes where the blobid indicates */
+		page_num = bs_blobid_to_page(blob->id);
+		lba = bs_md_page_to_lba(bs, page_num);
+
+		bs_batch_write_zeroes_dev(batch, lba, lba_count);
+	}
+
+	bs_batch_close(batch);
+}
+
+static void
+blob_persist_write_page_root(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	uint64_t			lba;
+	uint32_t			lba_count;
+	struct spdk_blob_md_page	*page;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	if (blob->active.num_pages == 0) {
+		/* Move on to the next step */
+		blob_persist_zero_pages(seq, ctx, 0);
+		return;
+	}
+
+	lba_count = bs_byte_to_lba(bs, sizeof(*page));
+
+	page = &ctx->pages[0];
+	/* The first page in the metadata goes where the blobid indicates */
+	lba = bs_md_page_to_lba(bs, bs_blobid_to_page(blob->id));
+
+	bs_sequence_write_dev(seq, page, lba, lba_count,
+			      blob_persist_zero_pages, ctx);
+}
+
+static void
+blob_persist_write_page_chain(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	struct spdk_blob_store		*bs = blob->bs;
+	uint64_t			lba;
+	uint32_t			lba_count;
+	struct spdk_blob_md_page	*page;
+	spdk_bs_batch_t			*batch;
+	size_t				i;
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	/* Clusters don't move around in blobs. The list shrinks or grows
+	 * at the end, but no changes ever occur in the middle of the list.
+	 */
+
+	lba_count = bs_byte_to_lba(bs, sizeof(*page));
+
+	batch = bs_sequence_to_batch(seq, blob_persist_write_page_root, ctx);
+
+	/* This starts at 1. The root page is not written until
+	 * all of the others are finished
+	 */
+	for (i = 1; i < blob->active.num_pages; i++) {
+		page = &ctx->pages[i];
+		assert(page->sequence_num == i);
+
+		lba = bs_md_page_to_lba(bs, blob->active.pages[i]);
+
+		bs_batch_write_dev(batch, page, lba, lba_count);
+	}
+
+	bs_batch_close(batch);
+}
+
+static int
+blob_resize(struct spdk_blob *blob, uint64_t sz)
+{
+	uint64_t	i;
+	uint64_t	*tmp;
+	uint64_t	lfc; /* lowest free cluster */
+	uint32_t	lfmd; /*  lowest free md page */
+	uint64_t	num_clusters;
+	uint32_t	*ep_tmp;
+	uint64_t	new_num_ep = 0, current_num_ep = 0;
+	struct spdk_blob_store *bs;
+
+	bs = blob->bs;
+
+	blob_verify_md_op(blob);
+
+	if (blob->active.num_clusters == sz) {
+		return 0;
+	}
+
+	if (blob->active.num_clusters < blob->active.cluster_array_size) {
+		/* If this blob was resized to be larger, then smaller, then
+		 * larger without syncing, then the cluster array already
+		 * contains spare assigned clusters we can use.
+		 */
+		num_clusters = spdk_min(blob->active.cluster_array_size,
+					sz);
+	} else {
+		num_clusters = blob->active.num_clusters;
+	}
+
+	if (blob->use_extent_table) {
+		/* Round up since every cluster beyond current Extent Table size,
+		 * requires new extent page. */
+		new_num_ep = spdk_divide_round_up(sz, SPDK_EXTENTS_PER_EP);
+		current_num_ep = spdk_divide_round_up(num_clusters, SPDK_EXTENTS_PER_EP);
+	}
+
+	/* Do two passes - one to verify that we can obtain enough clusters
+	 * and md pages, another to actually claim them.
+	 */
+
+	if (spdk_blob_is_thin_provisioned(blob) == false) {
+		lfc = 0;
+		for (i = num_clusters; i < sz; i++) {
+			lfc = spdk_bit_array_find_first_clear(bs->used_clusters, lfc);
+			if (lfc == UINT32_MAX) {
+				/* No more free clusters. Cannot satisfy the request */
+				return -ENOSPC;
+			}
+			lfc++;
+		}
+		lfmd = 0;
+		for (i = current_num_ep; i < new_num_ep ; i++) {
+			lfmd = spdk_bit_array_find_first_clear(blob->bs->used_md_pages, lfmd);
+			if (lfmd == UINT32_MAX) {
+				/* No more free md pages. Cannot satisfy the request */
+				return -ENOSPC;
+			}
+		}
+	}
+
+	if (sz > num_clusters) {
+		/* Expand the cluster array if necessary.
+		 * We only shrink the array when persisting.
+		 */
+		tmp = realloc(blob->active.clusters, sizeof(*blob->active.clusters) * sz);
+		if (sz > 0 && tmp == NULL) {
+			return -ENOMEM;
+		}
+		memset(tmp + blob->active.cluster_array_size, 0,
+		       sizeof(*blob->active.clusters) * (sz - blob->active.cluster_array_size));
+		blob->active.clusters = tmp;
+		blob->active.cluster_array_size = sz;
+
+		/* Expand the extents table, only if enough clusters were added */
+		if (new_num_ep > current_num_ep && blob->use_extent_table) {
+			ep_tmp = realloc(blob->active.extent_pages, sizeof(*blob->active.extent_pages) * new_num_ep);
+			if (new_num_ep > 0 && ep_tmp == NULL) {
+				return -ENOMEM;
+			}
+			memset(ep_tmp + blob->active.extent_pages_array_size, 0,
+			       sizeof(*blob->active.extent_pages) * (new_num_ep - blob->active.extent_pages_array_size));
+			blob->active.extent_pages = ep_tmp;
+			blob->active.extent_pages_array_size = new_num_ep;
+		}
+	}
+
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+
+	if (spdk_blob_is_thin_provisioned(blob) == false) {
+		lfc = 0;
+		lfmd = 0;
+		for (i = num_clusters; i < sz; i++) {
+			bs_allocate_cluster(blob, i, &lfc, &lfmd, true);
+			lfc++;
+			lfmd++;
+		}
+	}
+
+	blob->active.num_clusters = sz;
+	blob->active.num_extent_pages = new_num_ep;
+
+	return 0;
+}
+
+static void
+blob_persist_generate_new_md(struct spdk_blob_persist_ctx *ctx)
+{
+	spdk_bs_sequence_t *seq = ctx->seq;
+	struct spdk_blob *blob = ctx->blob;
+	struct spdk_blob_store *bs = blob->bs;
+	uint64_t i;
+	uint32_t page_num;
+	void *tmp;
+	int rc;
+
+	/* Generate the new metadata */
+	rc = blob_serialize(blob, &ctx->pages, &blob->active.num_pages);
+	if (rc < 0) {
+		blob_persist_complete(seq, ctx, rc);
+		return;
+	}
+
+	assert(blob->active.num_pages >= 1);
+
+	/* Resize the cache of page indices */
+	tmp = realloc(blob->active.pages, blob->active.num_pages * sizeof(*blob->active.pages));
+	if (!tmp) {
+		blob_persist_complete(seq, ctx, -ENOMEM);
+		return;
+	}
+	blob->active.pages = tmp;
+
+	/* Assign this metadata to pages. This requires two passes -
+	 * one to verify that there are enough pages and a second
+	 * to actually claim them. */
+	page_num = 0;
+	/* Note that this loop starts at one. The first page location is fixed by the blobid. */
+	for (i = 1; i < blob->active.num_pages; i++) {
+		page_num = spdk_bit_array_find_first_clear(bs->used_md_pages, page_num);
+		if (page_num == UINT32_MAX) {
+			blob_persist_complete(seq, ctx, -ENOMEM);
+			return;
+		}
+		page_num++;
+	}
+
+	page_num = 0;
+	blob->active.pages[0] = bs_blobid_to_page(blob->id);
+	for (i = 1; i < blob->active.num_pages; i++) {
+		page_num = spdk_bit_array_find_first_clear(bs->used_md_pages, page_num);
+		ctx->pages[i - 1].next = page_num;
+		/* Now that previous metadata page is complete, calculate the crc for it. */
+		ctx->pages[i - 1].crc = blob_md_page_calc_crc(&ctx->pages[i - 1]);
+		blob->active.pages[i] = page_num;
+		bs_claim_md_page(bs, page_num);
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Claiming page %u for blob %lu\n", page_num, blob->id);
+		page_num++;
+	}
+	ctx->pages[i - 1].crc = blob_md_page_calc_crc(&ctx->pages[i - 1]);
+	/* Start writing the metadata from last page to first */
+	blob->state = SPDK_BLOB_STATE_CLEAN;
+	blob_persist_write_page_chain(seq, ctx, 0);
+}
+
+static void
+blob_persist_write_extent_pages(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx	*ctx = cb_arg;
+	struct spdk_blob		*blob = ctx->blob;
+	size_t				i;
+	uint32_t			extent_page_id;
+	uint32_t                        page_count = 0;
+	int				rc;
+
+	if (ctx->extent_page != NULL) {
+		spdk_free(ctx->extent_page);
+		ctx->extent_page = NULL;
+	}
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	/* Only write out changed extent pages */
+	for (i = ctx->next_extent_page; i < blob->active.num_extent_pages; i++) {
+		extent_page_id = blob->active.extent_pages[i];
+		if (extent_page_id == 0) {
+			/* No Extent Page to persist */
+			assert(spdk_blob_is_thin_provisioned(blob));
+			continue;
+		}
+		/* Writing out new extent page for the first time. Either active extent pages is larger
+		 * than clean extent pages or there was no extent page assigned due to thin provisioning. */
+		if (i >= blob->clean.extent_pages_array_size || blob->clean.extent_pages[i] == 0) {
+			blob->state = SPDK_BLOB_STATE_DIRTY;
+			assert(spdk_bit_array_get(blob->bs->used_md_pages, extent_page_id));
+			ctx->next_extent_page = i + 1;
+			rc = blob_serialize_add_page(ctx->blob, &ctx->extent_page, &page_count, &ctx->extent_page);
+			if (rc < 0) {
+				blob_persist_complete(seq, ctx, rc);
+				return;
+			}
+
+			blob_serialize_extent_page(blob, i * SPDK_EXTENTS_PER_EP, ctx->extent_page);
+
+			ctx->extent_page->crc = blob_md_page_calc_crc(ctx->extent_page);
+
+			bs_sequence_write_dev(seq, ctx->extent_page, bs_md_page_to_lba(blob->bs, extent_page_id),
+					      bs_byte_to_lba(blob->bs, SPDK_BS_PAGE_SIZE),
+					      blob_persist_write_extent_pages, ctx);
+			return;
+		}
+		assert(blob->clean.extent_pages[i] != 0);
+	}
+
+	blob_persist_generate_new_md(ctx);
+}
+
+static void
+blob_persist_start(struct spdk_blob_persist_ctx *ctx)
+{
+	spdk_bs_sequence_t *seq = ctx->seq;
+	struct spdk_blob *blob = ctx->blob;
+
+	if (blob->active.num_pages == 0) {
+		/* This is the signal that the blob should be deleted.
+		 * Immediately jump to the clean up routine. */
+		assert(blob->clean.num_pages > 0);
+		blob->state = SPDK_BLOB_STATE_CLEAN;
+		blob_persist_zero_pages(seq, ctx, 0);
+		return;
+
+	}
+
+	blob_persist_write_extent_pages(seq, ctx, 0);
+}
+
+static void
+blob_persist_dirty_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx *ctx = cb_arg;
+
+	spdk_free(ctx->super);
+
+	if (bserrno != 0) {
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	ctx->blob->bs->clean = 0;
+
+	blob_persist_start(ctx);
+}
+
+static void
+bs_write_super(spdk_bs_sequence_t *seq, struct spdk_blob_store *bs,
+	       struct spdk_bs_super_block *super, spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+
+static void
+blob_persist_dirty(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_persist_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		spdk_free(ctx->super);
+		blob_persist_complete(seq, ctx, bserrno);
+		return;
+	}
+
+	ctx->super->clean = 0;
+	if (ctx->super->size == 0) {
+		ctx->super->size = ctx->blob->bs->dev->blockcnt * ctx->blob->bs->dev->blocklen;
+	}
+
+	bs_write_super(seq, ctx->blob->bs, ctx->super, blob_persist_dirty_cpl, ctx);
+}
+
+static void
+blob_persist_check_dirty(struct spdk_blob_persist_ctx *ctx)
+{
+	if (ctx->blob->bs->clean) {
+		ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+					  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+		if (!ctx->super) {
+			blob_persist_complete(ctx->seq, ctx, -ENOMEM);
+			return;
+		}
+
+		bs_sequence_read_dev(ctx->seq, ctx->super, bs_page_to_lba(ctx->blob->bs, 0),
+				     bs_byte_to_lba(ctx->blob->bs, sizeof(*ctx->super)),
+				     blob_persist_dirty, ctx);
+	} else {
+		blob_persist_start(ctx);
+	}
+}
+
+/* Write a blob to disk */
+static void
+blob_persist(spdk_bs_sequence_t *seq, struct spdk_blob *blob,
+	     spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_blob_persist_ctx *ctx;
+
+	blob_verify_md_op(blob);
+
+	if (blob->state == SPDK_BLOB_STATE_CLEAN && TAILQ_EMPTY(&blob->pending_persists)) {
+		cb_fn(seq, cb_arg, 0);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(seq, cb_arg, -ENOMEM);
+		return;
+	}
+	ctx->blob = blob;
+	ctx->seq = seq;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+	ctx->next_extent_page = 0;
+
+	/* Multiple blob persists can affect one another, via blob->state or
+	 * blob mutable data changes. To prevent it, queue up the persists. */
+	if (!TAILQ_EMPTY(&blob->pending_persists)) {
+		TAILQ_INSERT_TAIL(&blob->pending_persists, ctx, link);
+		return;
+	}
+	TAILQ_INSERT_HEAD(&blob->pending_persists, ctx, link);
+
+	blob_persist_check_dirty(ctx);
+}
+
+struct spdk_blob_copy_cluster_ctx {
+	struct spdk_blob *blob;
+	uint8_t *buf;
+	uint64_t page;
+	uint64_t new_cluster;
+	uint32_t new_extent_page;
+	spdk_bs_sequence_t *seq;
+};
+
+static void
+blob_allocate_and_copy_cluster_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_blob_copy_cluster_ctx *ctx = cb_arg;
+	struct spdk_bs_request_set *set = (struct spdk_bs_request_set *)ctx->seq;
+	TAILQ_HEAD(, spdk_bs_request_set) requests;
+	spdk_bs_user_op_t *op;
+
+	TAILQ_INIT(&requests);
+	TAILQ_SWAP(&set->channel->need_cluster_alloc, &requests, spdk_bs_request_set, link);
+
+	while (!TAILQ_EMPTY(&requests)) {
+		op = TAILQ_FIRST(&requests);
+		TAILQ_REMOVE(&requests, op, link);
+		if (bserrno == 0) {
+			bs_user_op_execute(op);
+		} else {
+			bs_user_op_abort(op);
+		}
+	}
+
+	spdk_free(ctx->buf);
+	free(ctx);
+}
+
+static void
+blob_insert_cluster_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_blob_copy_cluster_ctx *ctx = cb_arg;
+
+	if (bserrno) {
+		if (bserrno == -EEXIST) {
+			/* The metadata insert failed because another thread
+			 * allocated the cluster first. Free our cluster
+			 * but continue without error. */
+			bserrno = 0;
+		}
+		bs_release_cluster(ctx->blob->bs, ctx->new_cluster);
+		if (ctx->new_extent_page != 0) {
+			bs_release_md_page(ctx->blob->bs, ctx->new_extent_page);
+		}
+	}
+
+	bs_sequence_finish(ctx->seq, bserrno);
+}
+
+static void
+blob_write_copy_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_copy_cluster_ctx *ctx = cb_arg;
+	uint32_t cluster_number;
+
+	if (bserrno) {
+		/* The write failed, so jump to the final completion handler */
+		bs_sequence_finish(seq, bserrno);
+		return;
+	}
+
+	cluster_number = bs_page_to_cluster(ctx->blob->bs, ctx->page);
+
+	blob_insert_cluster_on_md_thread(ctx->blob, cluster_number, ctx->new_cluster,
+					 ctx->new_extent_page, blob_insert_cluster_cpl, ctx);
+}
+
+static void
+blob_write_copy(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_copy_cluster_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		/* The read failed, so jump to the final completion handler */
+		bs_sequence_finish(seq, bserrno);
+		return;
+	}
+
+	/* Write whole cluster */
+	bs_sequence_write_dev(seq, ctx->buf,
+			      bs_cluster_to_lba(ctx->blob->bs, ctx->new_cluster),
+			      bs_cluster_to_lba(ctx->blob->bs, 1),
+			      blob_write_copy_cpl, ctx);
+}
+
+static void
+bs_allocate_and_copy_cluster(struct spdk_blob *blob,
+			     struct spdk_io_channel *_ch,
+			     uint64_t io_unit, spdk_bs_user_op_t *op)
+{
+	struct spdk_bs_cpl cpl;
+	struct spdk_bs_channel *ch;
+	struct spdk_blob_copy_cluster_ctx *ctx;
+	uint32_t cluster_start_page;
+	uint32_t cluster_number;
+	int rc;
+
+	ch = spdk_io_channel_get_ctx(_ch);
+
+	if (!TAILQ_EMPTY(&ch->need_cluster_alloc)) {
+		/* There are already operations pending. Queue this user op
+		 * and return because it will be re-executed when the outstanding
+		 * cluster allocation completes. */
+		TAILQ_INSERT_TAIL(&ch->need_cluster_alloc, op, link);
+		return;
+	}
+
+	/* Round the io_unit offset down to the first page in the cluster */
+	cluster_start_page = bs_io_unit_to_cluster_start(blob, io_unit);
+
+	/* Calculate which index in the metadata cluster array the corresponding
+	 * cluster is supposed to be at. */
+	cluster_number = bs_io_unit_to_cluster_number(blob, io_unit);
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		bs_user_op_abort(op);
+		return;
+	}
+
+	assert(blob->bs->cluster_sz % blob->back_bs_dev->blocklen == 0);
+
+	ctx->blob = blob;
+	ctx->page = cluster_start_page;
+
+	if (blob->parent_id != SPDK_BLOBID_INVALID) {
+		ctx->buf = spdk_malloc(blob->bs->cluster_sz, blob->back_bs_dev->blocklen,
+				       NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+		if (!ctx->buf) {
+			SPDK_ERRLOG("DMA allocation for cluster of size = %" PRIu32 " failed.\n",
+				    blob->bs->cluster_sz);
+			free(ctx);
+			bs_user_op_abort(op);
+			return;
+		}
+	}
+
+	rc = bs_allocate_cluster(blob, cluster_number, &ctx->new_cluster, &ctx->new_extent_page,
+				 false);
+	if (rc != 0) {
+		spdk_free(ctx->buf);
+		free(ctx);
+		bs_user_op_abort(op);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = blob_allocate_and_copy_cluster_cpl;
+	cpl.u.blob_basic.cb_arg = ctx;
+
+	ctx->seq = bs_sequence_start(_ch, &cpl);
+	if (!ctx->seq) {
+		bs_release_cluster(blob->bs, ctx->new_cluster);
+		spdk_free(ctx->buf);
+		free(ctx);
+		bs_user_op_abort(op);
+		return;
+	}
+
+	/* Queue the user op to block other incoming operations */
+	TAILQ_INSERT_TAIL(&ch->need_cluster_alloc, op, link);
+
+	if (blob->parent_id != SPDK_BLOBID_INVALID) {
+		/* Read cluster from backing device */
+		bs_sequence_read_bs_dev(ctx->seq, blob->back_bs_dev, ctx->buf,
+					bs_dev_page_to_lba(blob->back_bs_dev, cluster_start_page),
+					bs_dev_byte_to_lba(blob->back_bs_dev, blob->bs->cluster_sz),
+					blob_write_copy, ctx);
+	} else {
+		blob_insert_cluster_on_md_thread(ctx->blob, cluster_number, ctx->new_cluster,
+						 ctx->new_extent_page, blob_insert_cluster_cpl, ctx);
+	}
+}
+
+static inline void
+blob_calculate_lba_and_lba_count(struct spdk_blob *blob, uint64_t io_unit, uint64_t length,
+				 uint64_t *lba,	uint32_t *lba_count)
+{
+	*lba_count = length;
+
+	if (!bs_io_unit_is_allocated(blob, io_unit)) {
+		assert(blob->back_bs_dev != NULL);
+		*lba = bs_io_unit_to_back_dev_lba(blob, io_unit);
+		*lba_count = bs_io_unit_to_back_dev_lba(blob, *lba_count);
+	} else {
+		*lba = bs_blob_io_unit_to_lba(blob, io_unit);
+	}
+}
+
+struct op_split_ctx {
+	struct spdk_blob *blob;
+	struct spdk_io_channel *channel;
+	uint64_t io_unit_offset;
+	uint64_t io_units_remaining;
+	void *curr_payload;
+	enum spdk_blob_op_type op_type;
+	spdk_bs_sequence_t *seq;
+};
+
+static void
+blob_request_submit_op_split_next(void *cb_arg, int bserrno)
+{
+	struct op_split_ctx	*ctx = cb_arg;
+	struct spdk_blob	*blob = ctx->blob;
+	struct spdk_io_channel	*ch = ctx->channel;
+	enum spdk_blob_op_type	op_type = ctx->op_type;
+	uint8_t			*buf = ctx->curr_payload;
+	uint64_t		offset = ctx->io_unit_offset;
+	uint64_t		length = ctx->io_units_remaining;
+	uint64_t		op_length;
+
+	if (bserrno != 0 || ctx->io_units_remaining == 0) {
+		bs_sequence_finish(ctx->seq, bserrno);
+		free(ctx);
+		return;
+	}
+
+	op_length = spdk_min(length, bs_num_io_units_to_cluster_boundary(blob,
+			     offset));
+
+	/* Update length and payload for next operation */
+	ctx->io_units_remaining -= op_length;
+	ctx->io_unit_offset += op_length;
+	if (op_type == SPDK_BLOB_WRITE || op_type == SPDK_BLOB_READ) {
+		ctx->curr_payload += op_length * blob->bs->io_unit_size;
+	}
+
+	switch (op_type) {
+	case SPDK_BLOB_READ:
+		spdk_blob_io_read(blob, ch, buf, offset, op_length,
+				  blob_request_submit_op_split_next, ctx);
+		break;
+	case SPDK_BLOB_WRITE:
+		spdk_blob_io_write(blob, ch, buf, offset, op_length,
+				   blob_request_submit_op_split_next, ctx);
+		break;
+	case SPDK_BLOB_UNMAP:
+		spdk_blob_io_unmap(blob, ch, offset, op_length,
+				   blob_request_submit_op_split_next, ctx);
+		break;
+	case SPDK_BLOB_WRITE_ZEROES:
+		spdk_blob_io_write_zeroes(blob, ch, offset, op_length,
+					  blob_request_submit_op_split_next, ctx);
+		break;
+	case SPDK_BLOB_READV:
+	case SPDK_BLOB_WRITEV:
+		SPDK_ERRLOG("readv/write not valid\n");
+		bs_sequence_finish(ctx->seq, -EINVAL);
+		free(ctx);
+		break;
+	}
+}
+
+static void
+blob_request_submit_op_split(struct spdk_io_channel *ch, struct spdk_blob *blob,
+			     void *payload, uint64_t offset, uint64_t length,
+			     spdk_blob_op_complete cb_fn, void *cb_arg, enum spdk_blob_op_type op_type)
+{
+	struct op_split_ctx *ctx;
+	spdk_bs_sequence_t *seq;
+	struct spdk_bs_cpl cpl;
+
+	assert(blob != NULL);
+
+	ctx = calloc(1, sizeof(struct op_split_ctx));
+	if (ctx == NULL) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(ch, &cpl);
+	if (!seq) {
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->blob = blob;
+	ctx->channel = ch;
+	ctx->curr_payload = payload;
+	ctx->io_unit_offset = offset;
+	ctx->io_units_remaining = length;
+	ctx->op_type = op_type;
+	ctx->seq = seq;
+
+	blob_request_submit_op_split_next(ctx, 0);
+}
+
+static void
+blob_request_submit_op_single(struct spdk_io_channel *_ch, struct spdk_blob *blob,
+			      void *payload, uint64_t offset, uint64_t length,
+			      spdk_blob_op_complete cb_fn, void *cb_arg, enum spdk_blob_op_type op_type)
+{
+	struct spdk_bs_cpl cpl;
+	uint64_t lba;
+	uint32_t lba_count;
+
+	assert(blob != NULL);
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	blob_calculate_lba_and_lba_count(blob, offset, length, &lba, &lba_count);
+
+	if (blob->frozen_refcnt) {
+		/* This blob I/O is frozen */
+		spdk_bs_user_op_t *op;
+		struct spdk_bs_channel *bs_channel = spdk_io_channel_get_ctx(_ch);
+
+		op = bs_user_op_alloc(_ch, &cpl, op_type, blob, payload, 0, offset, length);
+		if (!op) {
+			cb_fn(cb_arg, -ENOMEM);
+			return;
+		}
+
+		TAILQ_INSERT_TAIL(&bs_channel->queued_io, op, link);
+
+		return;
+	}
+
+	switch (op_type) {
+	case SPDK_BLOB_READ: {
+		spdk_bs_batch_t *batch;
+
+		batch = bs_batch_open(_ch, &cpl);
+		if (!batch) {
+			cb_fn(cb_arg, -ENOMEM);
+			return;
+		}
+
+		if (bs_io_unit_is_allocated(blob, offset)) {
+			/* Read from the blob */
+			bs_batch_read_dev(batch, payload, lba, lba_count);
+		} else {
+			/* Read from the backing block device */
+			bs_batch_read_bs_dev(batch, blob->back_bs_dev, payload, lba, lba_count);
+		}
+
+		bs_batch_close(batch);
+		break;
+	}
+	case SPDK_BLOB_WRITE:
+	case SPDK_BLOB_WRITE_ZEROES: {
+		if (bs_io_unit_is_allocated(blob, offset)) {
+			/* Write to the blob */
+			spdk_bs_batch_t *batch;
+
+			if (lba_count == 0) {
+				cb_fn(cb_arg, 0);
+				return;
+			}
+
+			batch = bs_batch_open(_ch, &cpl);
+			if (!batch) {
+				cb_fn(cb_arg, -ENOMEM);
+				return;
+			}
+
+			if (op_type == SPDK_BLOB_WRITE) {
+				bs_batch_write_dev(batch, payload, lba, lba_count);
+			} else {
+				bs_batch_write_zeroes_dev(batch, lba, lba_count);
+			}
+
+			bs_batch_close(batch);
+		} else {
+			/* Queue this operation and allocate the cluster */
+			spdk_bs_user_op_t *op;
+
+			op = bs_user_op_alloc(_ch, &cpl, op_type, blob, payload, 0, offset, length);
+			if (!op) {
+				cb_fn(cb_arg, -ENOMEM);
+				return;
+			}
+
+			bs_allocate_and_copy_cluster(blob, _ch, offset, op);
+		}
+		break;
+	}
+	case SPDK_BLOB_UNMAP: {
+		spdk_bs_batch_t *batch;
+
+		batch = bs_batch_open(_ch, &cpl);
+		if (!batch) {
+			cb_fn(cb_arg, -ENOMEM);
+			return;
+		}
+
+		if (bs_io_unit_is_allocated(blob, offset)) {
+			bs_batch_unmap_dev(batch, lba, lba_count);
+		}
+
+		bs_batch_close(batch);
+		break;
+	}
+	case SPDK_BLOB_READV:
+	case SPDK_BLOB_WRITEV:
+		SPDK_ERRLOG("readv/write not valid\n");
+		cb_fn(cb_arg, -EINVAL);
+		break;
+	}
+}
+
+static void
+blob_request_submit_op(struct spdk_blob *blob, struct spdk_io_channel *_channel,
+		       void *payload, uint64_t offset, uint64_t length,
+		       spdk_blob_op_complete cb_fn, void *cb_arg, enum spdk_blob_op_type op_type)
+{
+	assert(blob != NULL);
+
+	if (blob->data_ro && op_type != SPDK_BLOB_READ) {
+		cb_fn(cb_arg, -EPERM);
+		return;
+	}
+
+	if (offset + length > bs_cluster_to_lba(blob->bs, blob->active.num_clusters)) {
+		cb_fn(cb_arg, -EINVAL);
+		return;
+	}
+	if (length <= bs_num_io_units_to_cluster_boundary(blob, offset)) {
+		blob_request_submit_op_single(_channel, blob, payload, offset, length,
+					      cb_fn, cb_arg, op_type);
+	} else {
+		blob_request_submit_op_split(_channel, blob, payload, offset, length,
+					     cb_fn, cb_arg, op_type);
+	}
+}
+
+struct rw_iov_ctx {
+	struct spdk_blob *blob;
+	struct spdk_io_channel *channel;
+	spdk_blob_op_complete cb_fn;
+	void *cb_arg;
+	bool read;
+	int iovcnt;
+	struct iovec *orig_iov;
+	uint64_t io_unit_offset;
+	uint64_t io_units_remaining;
+	uint64_t io_units_done;
+	struct iovec iov[0];
+};
+
+static void
+rw_iov_done(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	assert(cb_arg == NULL);
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+rw_iov_split_next(void *cb_arg, int bserrno)
+{
+	struct rw_iov_ctx *ctx = cb_arg;
+	struct spdk_blob *blob = ctx->blob;
+	struct iovec *iov, *orig_iov;
+	int iovcnt;
+	size_t orig_iovoff;
+	uint64_t io_units_count, io_units_to_boundary, io_unit_offset;
+	uint64_t byte_count;
+
+	if (bserrno != 0 || ctx->io_units_remaining == 0) {
+		ctx->cb_fn(ctx->cb_arg, bserrno);
+		free(ctx);
+		return;
+	}
+
+	io_unit_offset = ctx->io_unit_offset;
+	io_units_to_boundary = bs_num_io_units_to_cluster_boundary(blob, io_unit_offset);
+	io_units_count = spdk_min(ctx->io_units_remaining, io_units_to_boundary);
+	/*
+	 * Get index and offset into the original iov array for our current position in the I/O sequence.
+	 *  byte_count will keep track of how many bytes remaining until orig_iov and orig_iovoff will
+	 *  point to the current position in the I/O sequence.
+	 */
+	byte_count = ctx->io_units_done * blob->bs->io_unit_size;
+	orig_iov = &ctx->orig_iov[0];
+	orig_iovoff = 0;
+	while (byte_count > 0) {
+		if (byte_count >= orig_iov->iov_len) {
+			byte_count -= orig_iov->iov_len;
+			orig_iov++;
+		} else {
+			orig_iovoff = byte_count;
+			byte_count = 0;
+		}
+	}
+
+	/*
+	 * Build an iov array for the next I/O in the sequence.  byte_count will keep track of how many
+	 *  bytes of this next I/O remain to be accounted for in the new iov array.
+	 */
+	byte_count = io_units_count * blob->bs->io_unit_size;
+	iov = &ctx->iov[0];
+	iovcnt = 0;
+	while (byte_count > 0) {
+		assert(iovcnt < ctx->iovcnt);
+		iov->iov_len = spdk_min(byte_count, orig_iov->iov_len - orig_iovoff);
+		iov->iov_base = orig_iov->iov_base + orig_iovoff;
+		byte_count -= iov->iov_len;
+		orig_iovoff = 0;
+		orig_iov++;
+		iov++;
+		iovcnt++;
+	}
+
+	ctx->io_unit_offset += io_units_count;
+	ctx->io_units_remaining -= io_units_count;
+	ctx->io_units_done += io_units_count;
+	iov = &ctx->iov[0];
+
+	if (ctx->read) {
+		spdk_blob_io_readv(ctx->blob, ctx->channel, iov, iovcnt, io_unit_offset,
+				   io_units_count, rw_iov_split_next, ctx);
+	} else {
+		spdk_blob_io_writev(ctx->blob, ctx->channel, iov, iovcnt, io_unit_offset,
+				    io_units_count, rw_iov_split_next, ctx);
+	}
+}
+
+static void
+blob_request_submit_rw_iov(struct spdk_blob *blob, struct spdk_io_channel *_channel,
+			   struct iovec *iov, int iovcnt, uint64_t offset, uint64_t length,
+			   spdk_blob_op_complete cb_fn, void *cb_arg, bool read)
+{
+	struct spdk_bs_cpl	cpl;
+
+	assert(blob != NULL);
+
+	if (!read && blob->data_ro) {
+		cb_fn(cb_arg, -EPERM);
+		return;
+	}
+
+	if (length == 0) {
+		cb_fn(cb_arg, 0);
+		return;
+	}
+
+	if (offset + length > bs_cluster_to_lba(blob->bs, blob->active.num_clusters)) {
+		cb_fn(cb_arg, -EINVAL);
+		return;
+	}
+
+	/*
+	 * For now, we implement readv/writev using a sequence (instead of a batch) to account for having
+	 *  to split a request that spans a cluster boundary.  For I/O that do not span a cluster boundary,
+	 *  there will be no noticeable difference compared to using a batch.  For I/O that do span a cluster
+	 *  boundary, the target LBAs (after blob offset to LBA translation) may not be contiguous, so we need
+	 *  to allocate a separate iov array and split the I/O such that none of the resulting
+	 *  smaller I/O cross a cluster boundary.  These smaller I/O will be issued in sequence (not in parallel)
+	 *  but since this case happens very infrequently, any performance impact will be negligible.
+	 *
+	 * This could be optimized in the future to allocate a big enough iov array to account for all of the iovs
+	 *  for all of the smaller I/Os, pre-build all of the iov arrays for the smaller I/Os, then issue them
+	 *  in a batch.  That would also require creating an intermediate spdk_bs_cpl that would get called
+	 *  when the batch was completed, to allow for freeing the memory for the iov arrays.
+	 */
+	if (spdk_likely(length <= bs_num_io_units_to_cluster_boundary(blob, offset))) {
+		uint32_t lba_count;
+		uint64_t lba;
+
+		cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+		cpl.u.blob_basic.cb_fn = cb_fn;
+		cpl.u.blob_basic.cb_arg = cb_arg;
+
+		if (blob->frozen_refcnt) {
+			/* This blob I/O is frozen */
+			enum spdk_blob_op_type op_type;
+			spdk_bs_user_op_t *op;
+			struct spdk_bs_channel *bs_channel = spdk_io_channel_get_ctx(_channel);
+
+			op_type = read ? SPDK_BLOB_READV : SPDK_BLOB_WRITEV;
+			op = bs_user_op_alloc(_channel, &cpl, op_type, blob, iov, iovcnt, offset, length);
+			if (!op) {
+				cb_fn(cb_arg, -ENOMEM);
+				return;
+			}
+
+			TAILQ_INSERT_TAIL(&bs_channel->queued_io, op, link);
+
+			return;
+		}
+
+		blob_calculate_lba_and_lba_count(blob, offset, length, &lba, &lba_count);
+
+		if (read) {
+			spdk_bs_sequence_t *seq;
+
+			seq = bs_sequence_start(_channel, &cpl);
+			if (!seq) {
+				cb_fn(cb_arg, -ENOMEM);
+				return;
+			}
+
+			if (bs_io_unit_is_allocated(blob, offset)) {
+				bs_sequence_readv_dev(seq, iov, iovcnt, lba, lba_count, rw_iov_done, NULL);
+			} else {
+				bs_sequence_readv_bs_dev(seq, blob->back_bs_dev, iov, iovcnt, lba, lba_count,
+							 rw_iov_done, NULL);
+			}
+		} else {
+			if (bs_io_unit_is_allocated(blob, offset)) {
+				spdk_bs_sequence_t *seq;
+
+				seq = bs_sequence_start(_channel, &cpl);
+				if (!seq) {
+					cb_fn(cb_arg, -ENOMEM);
+					return;
+				}
+
+				bs_sequence_writev_dev(seq, iov, iovcnt, lba, lba_count, rw_iov_done, NULL);
+			} else {
+				/* Queue this operation and allocate the cluster */
+				spdk_bs_user_op_t *op;
+
+				op = bs_user_op_alloc(_channel, &cpl, SPDK_BLOB_WRITEV, blob, iov, iovcnt, offset,
+						      length);
+				if (!op) {
+					cb_fn(cb_arg, -ENOMEM);
+					return;
+				}
+
+				bs_allocate_and_copy_cluster(blob, _channel, offset, op);
+			}
+		}
+	} else {
+		struct rw_iov_ctx *ctx;
+
+		ctx = calloc(1, sizeof(struct rw_iov_ctx) + iovcnt * sizeof(struct iovec));
+		if (ctx == NULL) {
+			cb_fn(cb_arg, -ENOMEM);
+			return;
+		}
+
+		ctx->blob = blob;
+		ctx->channel = _channel;
+		ctx->cb_fn = cb_fn;
+		ctx->cb_arg = cb_arg;
+		ctx->read = read;
+		ctx->orig_iov = iov;
+		ctx->iovcnt = iovcnt;
+		ctx->io_unit_offset = offset;
+		ctx->io_units_remaining = length;
+		ctx->io_units_done = 0;
+
+		rw_iov_split_next(ctx, 0);
+	}
+}
+
+static struct spdk_blob *
+blob_lookup(struct spdk_blob_store *bs, spdk_blob_id blobid)
+{
+	struct spdk_blob *blob;
+
+	if (spdk_bit_array_get(bs->open_blobids, blobid) == 0) {
+		return NULL;
+	}
+
+	TAILQ_FOREACH(blob, &bs->blobs, link) {
+		if (blob->id == blobid) {
+			return blob;
+		}
+	}
+
+	return NULL;
+}
+
+static void
+blob_get_snapshot_and_clone_entries(struct spdk_blob *blob,
+				    struct spdk_blob_list **snapshot_entry, struct spdk_blob_list **clone_entry)
+{
+	assert(blob != NULL);
+	*snapshot_entry = NULL;
+	*clone_entry = NULL;
+
+	if (blob->parent_id == SPDK_BLOBID_INVALID) {
+		return;
+	}
+
+	TAILQ_FOREACH(*snapshot_entry, &blob->bs->snapshots, link) {
+		if ((*snapshot_entry)->id == blob->parent_id) {
+			break;
+		}
+	}
+
+	if (*snapshot_entry != NULL) {
+		TAILQ_FOREACH(*clone_entry, &(*snapshot_entry)->clones, link) {
+			if ((*clone_entry)->id == blob->id) {
+				break;
+			}
+		}
+
+		assert(clone_entry != NULL);
+	}
+}
+
+static int
+bs_channel_create(void *io_device, void *ctx_buf)
+{
+	struct spdk_blob_store		*bs = io_device;
+	struct spdk_bs_channel		*channel = ctx_buf;
+	struct spdk_bs_dev		*dev;
+	uint32_t			max_ops = bs->max_channel_ops;
+	uint32_t			i;
+
+	dev = bs->dev;
+
+	channel->req_mem = calloc(max_ops, sizeof(struct spdk_bs_request_set));
+	if (!channel->req_mem) {
+		return -1;
+	}
+
+	TAILQ_INIT(&channel->reqs);
+
+	for (i = 0; i < max_ops; i++) {
+		TAILQ_INSERT_TAIL(&channel->reqs, &channel->req_mem[i], link);
+	}
+
+	channel->bs = bs;
+	channel->dev = dev;
+	channel->dev_channel = dev->create_channel(dev);
+
+	if (!channel->dev_channel) {
+		SPDK_ERRLOG("Failed to create device channel.\n");
+		free(channel->req_mem);
+		return -1;
+	}
+
+	TAILQ_INIT(&channel->need_cluster_alloc);
+	TAILQ_INIT(&channel->queued_io);
+
+	return 0;
+}
+
+static void
+bs_channel_destroy(void *io_device, void *ctx_buf)
+{
+	struct spdk_bs_channel *channel = ctx_buf;
+	spdk_bs_user_op_t *op;
+
+	while (!TAILQ_EMPTY(&channel->need_cluster_alloc)) {
+		op = TAILQ_FIRST(&channel->need_cluster_alloc);
+		TAILQ_REMOVE(&channel->need_cluster_alloc, op, link);
+		bs_user_op_abort(op);
+	}
+
+	while (!TAILQ_EMPTY(&channel->queued_io)) {
+		op = TAILQ_FIRST(&channel->queued_io);
+		TAILQ_REMOVE(&channel->queued_io, op, link);
+		bs_user_op_abort(op);
+	}
+
+	free(channel->req_mem);
+	channel->dev->destroy_channel(channel->dev, channel->dev_channel);
+}
+
+static void
+bs_dev_destroy(void *io_device)
+{
+	struct spdk_blob_store *bs = io_device;
+	struct spdk_blob	*blob, *blob_tmp;
+
+	bs->dev->destroy(bs->dev);
+
+	TAILQ_FOREACH_SAFE(blob, &bs->blobs, link, blob_tmp) {
+		TAILQ_REMOVE(&bs->blobs, blob, link);
+		spdk_bit_array_clear(bs->open_blobids, blob->id);
+		blob_free(blob);
+	}
+
+	pthread_mutex_destroy(&bs->used_clusters_mutex);
+
+	spdk_bit_array_free(&bs->open_blobids);
+	spdk_bit_array_free(&bs->used_blobids);
+	spdk_bit_array_free(&bs->used_md_pages);
+	spdk_bit_array_free(&bs->used_clusters);
+	/*
+	 * If this function is called for any reason except a successful unload,
+	 * the unload_cpl type will be NONE and this will be a nop.
+	 */
+	bs_call_cpl(&bs->unload_cpl, bs->unload_err);
+
+	free(bs);
+}
+
+static int
+bs_blob_list_add(struct spdk_blob *blob)
+{
+	spdk_blob_id snapshot_id;
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+
+	assert(blob != NULL);
+
+	snapshot_id = blob->parent_id;
+	if (snapshot_id == SPDK_BLOBID_INVALID) {
+		return 0;
+	}
+
+	snapshot_entry = bs_get_snapshot_entry(blob->bs, snapshot_id);
+	if (snapshot_entry == NULL) {
+		/* Snapshot not found */
+		snapshot_entry = calloc(1, sizeof(struct spdk_blob_list));
+		if (snapshot_entry == NULL) {
+			return -ENOMEM;
+		}
+		snapshot_entry->id = snapshot_id;
+		TAILQ_INIT(&snapshot_entry->clones);
+		TAILQ_INSERT_TAIL(&blob->bs->snapshots, snapshot_entry, link);
+	} else {
+		TAILQ_FOREACH(clone_entry, &snapshot_entry->clones, link) {
+			if (clone_entry->id == blob->id) {
+				break;
+			}
+		}
+	}
+
+	if (clone_entry == NULL) {
+		/* Clone not found */
+		clone_entry = calloc(1, sizeof(struct spdk_blob_list));
+		if (clone_entry == NULL) {
+			return -ENOMEM;
+		}
+		clone_entry->id = blob->id;
+		TAILQ_INIT(&clone_entry->clones);
+		TAILQ_INSERT_TAIL(&snapshot_entry->clones, clone_entry, link);
+		snapshot_entry->clone_count++;
+	}
+
+	return 0;
+}
+
+static void
+bs_blob_list_remove(struct spdk_blob *blob)
+{
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+
+	blob_get_snapshot_and_clone_entries(blob, &snapshot_entry, &clone_entry);
+
+	if (snapshot_entry == NULL) {
+		return;
+	}
+
+	blob->parent_id = SPDK_BLOBID_INVALID;
+	TAILQ_REMOVE(&snapshot_entry->clones, clone_entry, link);
+	free(clone_entry);
+
+	snapshot_entry->clone_count--;
+}
+
+static int
+bs_blob_list_free(struct spdk_blob_store *bs)
+{
+	struct spdk_blob_list *snapshot_entry;
+	struct spdk_blob_list *snapshot_entry_tmp;
+	struct spdk_blob_list *clone_entry;
+	struct spdk_blob_list *clone_entry_tmp;
+
+	TAILQ_FOREACH_SAFE(snapshot_entry, &bs->snapshots, link, snapshot_entry_tmp) {
+		TAILQ_FOREACH_SAFE(clone_entry, &snapshot_entry->clones, link, clone_entry_tmp) {
+			TAILQ_REMOVE(&snapshot_entry->clones, clone_entry, link);
+			free(clone_entry);
+		}
+		TAILQ_REMOVE(&bs->snapshots, snapshot_entry, link);
+		free(snapshot_entry);
+	}
+
+	return 0;
+}
+
+static void
+bs_free(struct spdk_blob_store *bs)
+{
+	bs_blob_list_free(bs);
+
+	bs_unregister_md_thread(bs);
+	spdk_io_device_unregister(bs, bs_dev_destroy);
+}
+
+void
+spdk_bs_opts_init(struct spdk_bs_opts *opts)
+{
+	opts->cluster_sz = SPDK_BLOB_OPTS_CLUSTER_SZ;
+	opts->num_md_pages = SPDK_BLOB_OPTS_NUM_MD_PAGES;
+	opts->max_md_ops = SPDK_BLOB_OPTS_MAX_MD_OPS;
+	opts->max_channel_ops = SPDK_BLOB_OPTS_DEFAULT_CHANNEL_OPS;
+	opts->clear_method = BS_CLEAR_WITH_UNMAP;
+	memset(&opts->bstype, 0, sizeof(opts->bstype));
+	opts->iter_cb_fn = NULL;
+	opts->iter_cb_arg = NULL;
+}
+
+static int
+bs_opts_verify(struct spdk_bs_opts *opts)
+{
+	if (opts->cluster_sz == 0 || opts->num_md_pages == 0 || opts->max_md_ops == 0 ||
+	    opts->max_channel_ops == 0) {
+		SPDK_ERRLOG("Blobstore options cannot be set to 0\n");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int
+bs_alloc(struct spdk_bs_dev *dev, struct spdk_bs_opts *opts, struct spdk_blob_store **_bs)
+{
+	struct spdk_blob_store	*bs;
+	uint64_t dev_size;
+	int rc;
+
+	dev_size = dev->blocklen * dev->blockcnt;
+	if (dev_size < opts->cluster_sz) {
+		/* Device size cannot be smaller than cluster size of blobstore */
+		SPDK_INFOLOG(SPDK_LOG_BLOB, "Device size %" PRIu64 " is smaller than cluster size %" PRIu32 "\n",
+			     dev_size, opts->cluster_sz);
+		return -ENOSPC;
+	}
+	if (opts->cluster_sz < SPDK_BS_PAGE_SIZE) {
+		/* Cluster size cannot be smaller than page size */
+		SPDK_ERRLOG("Cluster size %" PRIu32 " is smaller than page size %d\n",
+			    opts->cluster_sz, SPDK_BS_PAGE_SIZE);
+		return -EINVAL;
+	}
+	bs = calloc(1, sizeof(struct spdk_blob_store));
+	if (!bs) {
+		return -ENOMEM;
+	}
+
+	TAILQ_INIT(&bs->blobs);
+	TAILQ_INIT(&bs->snapshots);
+	bs->dev = dev;
+	bs->md_thread = spdk_get_thread();
+	assert(bs->md_thread != NULL);
+
+	/*
+	 * Do not use bs_lba_to_cluster() here since blockcnt may not be an
+	 *  even multiple of the cluster size.
+	 */
+	bs->cluster_sz = opts->cluster_sz;
+	bs->total_clusters = dev->blockcnt / (bs->cluster_sz / dev->blocklen);
+	bs->pages_per_cluster = bs->cluster_sz / SPDK_BS_PAGE_SIZE;
+	if (spdk_u32_is_pow2(bs->pages_per_cluster)) {
+		bs->pages_per_cluster_shift = spdk_u32log2(bs->pages_per_cluster);
+	}
+	bs->num_free_clusters = bs->total_clusters;
+	bs->used_clusters = spdk_bit_array_create(bs->total_clusters);
+	bs->io_unit_size = dev->blocklen;
+	if (bs->used_clusters == NULL) {
+		free(bs);
+		return -ENOMEM;
+	}
+
+	bs->max_channel_ops = opts->max_channel_ops;
+	bs->super_blob = SPDK_BLOBID_INVALID;
+	memcpy(&bs->bstype, &opts->bstype, sizeof(opts->bstype));
+
+	/* The metadata is assumed to be at least 1 page */
+	bs->used_md_pages = spdk_bit_array_create(1);
+	bs->used_blobids = spdk_bit_array_create(0);
+	bs->open_blobids = spdk_bit_array_create(0);
+
+	pthread_mutex_init(&bs->used_clusters_mutex, NULL);
+
+	spdk_io_device_register(bs, bs_channel_create, bs_channel_destroy,
+				sizeof(struct spdk_bs_channel), "blobstore");
+	rc = bs_register_md_thread(bs);
+	if (rc == -1) {
+		spdk_io_device_unregister(bs, NULL);
+		pthread_mutex_destroy(&bs->used_clusters_mutex);
+		spdk_bit_array_free(&bs->open_blobids);
+		spdk_bit_array_free(&bs->used_blobids);
+		spdk_bit_array_free(&bs->used_md_pages);
+		spdk_bit_array_free(&bs->used_clusters);
+		free(bs);
+		/* FIXME: this is a lie but don't know how to get a proper error code here */
+		return -ENOMEM;
+	}
+
+	*_bs = bs;
+	return 0;
+}
+
+/* START spdk_bs_load, spdk_bs_load_ctx will used for both load and unload. */
+
+struct spdk_bs_load_ctx {
+	struct spdk_blob_store		*bs;
+	struct spdk_bs_super_block	*super;
+
+	struct spdk_bs_md_mask		*mask;
+	bool				in_page_chain;
+	uint32_t			page_index;
+	uint32_t			cur_page;
+	struct spdk_blob_md_page	*page;
+
+	uint64_t			num_extent_pages;
+	uint32_t			*extent_page_num;
+	struct spdk_blob_md_page	*extent_pages;
+
+	spdk_bs_sequence_t			*seq;
+	spdk_blob_op_with_handle_complete	iter_cb_fn;
+	void					*iter_cb_arg;
+	struct spdk_blob			*blob;
+	spdk_blob_id				blobid;
+};
+
+static void
+bs_load_ctx_fail(struct spdk_bs_load_ctx *ctx, int bserrno)
+{
+	assert(bserrno != 0);
+
+	spdk_free(ctx->super);
+	bs_sequence_finish(ctx->seq, bserrno);
+	bs_free(ctx->bs);
+	free(ctx);
+}
+
+static void
+bs_set_mask(struct spdk_bit_array *array, struct spdk_bs_md_mask *mask)
+{
+	uint32_t i = 0;
+
+	while (true) {
+		i = spdk_bit_array_find_first_set(array, i);
+		if (i >= mask->length) {
+			break;
+		}
+		mask->mask[i / 8] |= 1U << (i % 8);
+		i++;
+	}
+}
+
+static int
+bs_load_mask(struct spdk_bit_array **array_ptr, struct spdk_bs_md_mask *mask)
+{
+	struct spdk_bit_array *array;
+	uint32_t i;
+
+	if (spdk_bit_array_resize(array_ptr, mask->length) < 0) {
+		return -ENOMEM;
+	}
+
+	array = *array_ptr;
+	for (i = 0; i < mask->length; i++) {
+		if (mask->mask[i / 8] & (1U << (i % 8))) {
+			spdk_bit_array_set(array, i);
+		}
+	}
+
+	return 0;
+}
+
+static void
+bs_write_super(spdk_bs_sequence_t *seq, struct spdk_blob_store *bs,
+	       struct spdk_bs_super_block *super, spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	/* Update the values in the super block */
+	super->super_blob = bs->super_blob;
+	memcpy(&super->bstype, &bs->bstype, sizeof(bs->bstype));
+	super->crc = blob_md_page_calc_crc(super);
+	bs_sequence_write_dev(seq, super, bs_page_to_lba(bs, 0),
+			      bs_byte_to_lba(bs, sizeof(*super)),
+			      cb_fn, cb_arg);
+}
+
+static void
+bs_write_used_clusters(spdk_bs_sequence_t *seq, void *arg, spdk_bs_sequence_cpl cb_fn)
+{
+	struct spdk_bs_load_ctx	*ctx = arg;
+	uint64_t	mask_size, lba, lba_count;
+
+	/* Write out the used clusters mask */
+	mask_size = ctx->super->used_cluster_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL,
+				 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	ctx->mask->type = SPDK_MD_MASK_TYPE_USED_CLUSTERS;
+	ctx->mask->length = ctx->bs->total_clusters;
+	assert(ctx->mask->length == spdk_bit_array_capacity(ctx->bs->used_clusters));
+
+	bs_set_mask(ctx->bs->used_clusters, ctx->mask);
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_cluster_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_cluster_mask_len);
+	bs_sequence_write_dev(seq, ctx->mask, lba, lba_count, cb_fn, arg);
+}
+
+static void
+bs_write_used_md(spdk_bs_sequence_t *seq, void *arg, spdk_bs_sequence_cpl cb_fn)
+{
+	struct spdk_bs_load_ctx	*ctx = arg;
+	uint64_t	mask_size, lba, lba_count;
+
+	mask_size = ctx->super->used_page_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL,
+				 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	ctx->mask->type = SPDK_MD_MASK_TYPE_USED_PAGES;
+	ctx->mask->length = ctx->super->md_len;
+	assert(ctx->mask->length == spdk_bit_array_capacity(ctx->bs->used_md_pages));
+
+	bs_set_mask(ctx->bs->used_md_pages, ctx->mask);
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_page_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_page_mask_len);
+	bs_sequence_write_dev(seq, ctx->mask, lba, lba_count, cb_fn, arg);
+}
+
+static void
+bs_write_used_blobids(spdk_bs_sequence_t *seq, void *arg, spdk_bs_sequence_cpl cb_fn)
+{
+	struct spdk_bs_load_ctx	*ctx = arg;
+	uint64_t	mask_size, lba, lba_count;
+
+	if (ctx->super->used_blobid_mask_len == 0) {
+		/*
+		 * This is a pre-v3 on-disk format where the blobid mask does not get
+		 *  written to disk.
+		 */
+		cb_fn(seq, arg, 0);
+		return;
+	}
+
+	mask_size = ctx->super->used_blobid_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL, SPDK_ENV_SOCKET_ID_ANY,
+				 SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	ctx->mask->type = SPDK_MD_MASK_TYPE_USED_BLOBIDS;
+	ctx->mask->length = ctx->super->md_len;
+	assert(ctx->mask->length == spdk_bit_array_capacity(ctx->bs->used_blobids));
+
+	bs_set_mask(ctx->bs->used_blobids, ctx->mask);
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_blobid_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_blobid_mask_len);
+	bs_sequence_write_dev(seq, ctx->mask, lba, lba_count, cb_fn, arg);
+}
+
+static void
+blob_set_thin_provision(struct spdk_blob *blob)
+{
+	blob_verify_md_op(blob);
+	blob->invalid_flags |= SPDK_BLOB_THIN_PROV;
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+}
+
+static void
+blob_set_clear_method(struct spdk_blob *blob, enum blob_clear_method clear_method)
+{
+	blob_verify_md_op(blob);
+	blob->clear_method = clear_method;
+	blob->md_ro_flags |= (clear_method << SPDK_BLOB_CLEAR_METHOD_SHIFT);
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+}
+
+static void bs_load_iter(void *arg, struct spdk_blob *blob, int bserrno);
+
+static void
+bs_delete_corrupted_blob_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	spdk_blob_id id;
+	int64_t page_num;
+
+	/* Iterate to next blob (we can't use spdk_bs_iter_next function as our
+	 * last blob has been removed */
+	page_num = bs_blobid_to_page(ctx->blobid);
+	page_num++;
+	page_num = spdk_bit_array_find_first_set(ctx->bs->used_blobids, page_num);
+	if (page_num >= spdk_bit_array_capacity(ctx->bs->used_blobids)) {
+		bs_load_iter(ctx, NULL, -ENOENT);
+		return;
+	}
+
+	id = bs_page_to_blobid(page_num);
+
+	spdk_bs_open_blob(ctx->bs, id, bs_load_iter, ctx);
+}
+
+static void
+bs_delete_corrupted_close_cb(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Failed to close corrupted blob\n");
+		spdk_bs_iter_next(ctx->bs, ctx->blob, bs_load_iter, ctx);
+		return;
+	}
+
+	spdk_bs_delete_blob(ctx->bs, ctx->blobid, bs_delete_corrupted_blob_cpl, ctx);
+}
+
+static void
+bs_delete_corrupted_blob(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint64_t i;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Failed to close clone of a corrupted blob\n");
+		spdk_bs_iter_next(ctx->bs, ctx->blob, bs_load_iter, ctx);
+		return;
+	}
+
+	/* Snapshot and clone have the same copy of cluster map and extent pages
+	 * at this point. Let's clear both for snpashot now,
+	 * so that it won't be cleared for clone later when we remove snapshot.
+	 * Also set thin provision to pass data corruption check */
+	for (i = 0; i < ctx->blob->active.num_clusters; i++) {
+		ctx->blob->active.clusters[i] = 0;
+	}
+	for (i = 0; i < ctx->blob->active.num_extent_pages; i++) {
+		ctx->blob->active.extent_pages[i] = 0;
+	}
+
+	ctx->blob->md_ro = false;
+
+	blob_set_thin_provision(ctx->blob);
+
+	ctx->blobid = ctx->blob->id;
+
+	spdk_blob_close(ctx->blob, bs_delete_corrupted_close_cb, ctx);
+}
+
+static void
+bs_update_corrupted_blob(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Failed to close clone of a corrupted blob\n");
+		spdk_bs_iter_next(ctx->bs, ctx->blob, bs_load_iter, ctx);
+		return;
+	}
+
+	ctx->blob->md_ro = false;
+	blob_remove_xattr(ctx->blob, SNAPSHOT_PENDING_REMOVAL, true);
+	blob_remove_xattr(ctx->blob, SNAPSHOT_IN_PROGRESS, true);
+	spdk_blob_set_read_only(ctx->blob);
+
+	if (ctx->iter_cb_fn) {
+		ctx->iter_cb_fn(ctx->iter_cb_arg, ctx->blob, 0);
+	}
+	bs_blob_list_add(ctx->blob);
+
+	spdk_bs_iter_next(ctx->bs, ctx->blob, bs_load_iter, ctx);
+}
+
+static void
+bs_examine_clone(void *cb_arg, struct spdk_blob *blob, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Failed to open clone of a corrupted blob\n");
+		spdk_bs_iter_next(ctx->bs, ctx->blob, bs_load_iter, ctx);
+		return;
+	}
+
+	if (blob->parent_id == ctx->blob->id) {
+		/* Power failure occured before updating clone (snapshot delete case)
+		 * or after updating clone (creating snapshot case) - keep snapshot */
+		spdk_blob_close(blob, bs_update_corrupted_blob, ctx);
+	} else {
+		/* Power failure occured after updating clone (snapshot delete case)
+		 * or before updating clone (creating snapshot case) - remove snapshot */
+		spdk_blob_close(blob, bs_delete_corrupted_blob, ctx);
+	}
+}
+
+static void
+bs_load_iter(void *arg, struct spdk_blob *blob, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = arg;
+	const void *value;
+	size_t len;
+	int rc = 0;
+
+	if (bserrno == 0) {
+		/* Examine blob if it is corrupted after power failure. Fix
+		 * the ones that can be fixed and remove any other corrupted
+		 * ones. If it is not corrupted just process it */
+		rc = blob_get_xattr_value(blob, SNAPSHOT_PENDING_REMOVAL, &value, &len, true);
+		if (rc != 0) {
+			rc = blob_get_xattr_value(blob, SNAPSHOT_IN_PROGRESS, &value, &len, true);
+			if (rc != 0) {
+				/* Not corrupted - process it and continue with iterating through blobs */
+				if (ctx->iter_cb_fn) {
+					ctx->iter_cb_fn(ctx->iter_cb_arg, blob, 0);
+				}
+				bs_blob_list_add(blob);
+				spdk_bs_iter_next(ctx->bs, blob, bs_load_iter, ctx);
+				return;
+			}
+
+		}
+
+		assert(len == sizeof(spdk_blob_id));
+
+		ctx->blob = blob;
+
+		/* Open clone to check if we are able to fix this blob or should we remove it */
+		spdk_bs_open_blob(ctx->bs, *(spdk_blob_id *)value, bs_examine_clone, ctx);
+		return;
+	} else if (bserrno == -ENOENT) {
+		bserrno = 0;
+	} else {
+		/*
+		 * This case needs to be looked at further.  Same problem
+		 *  exists with applications that rely on explicit blob
+		 *  iteration.  We should just skip the blob that failed
+		 *  to load and continue on to the next one.
+		 */
+		SPDK_ERRLOG("Error in iterating blobs\n");
+	}
+
+	ctx->iter_cb_fn = NULL;
+
+	spdk_free(ctx->super);
+	spdk_free(ctx->mask);
+	bs_sequence_finish(ctx->seq, bserrno);
+	free(ctx);
+}
+
+static void
+bs_load_complete(struct spdk_bs_load_ctx *ctx)
+{
+	spdk_bs_iter_first(ctx->bs, bs_load_iter, ctx);
+}
+
+static void
+bs_load_used_blobids_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	int rc;
+
+	/* The type must be correct */
+	assert(ctx->mask->type == SPDK_MD_MASK_TYPE_USED_BLOBIDS);
+
+	/* The length of the mask (in bits) must not be greater than
+	 * the length of the buffer (converted to bits) */
+	assert(ctx->mask->length <= (ctx->super->used_blobid_mask_len * SPDK_BS_PAGE_SIZE * 8));
+
+	/* The length of the mask must be exactly equal to the size
+	 * (in pages) of the metadata region */
+	assert(ctx->mask->length == ctx->super->md_len);
+
+	rc = bs_load_mask(&ctx->bs->used_blobids, ctx->mask);
+	if (rc < 0) {
+		spdk_free(ctx->mask);
+		bs_load_ctx_fail(ctx, rc);
+		return;
+	}
+
+	bs_load_complete(ctx);
+}
+
+static void
+bs_load_used_clusters_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint64_t		lba, lba_count, mask_size;
+	int			rc;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	/* The type must be correct */
+	assert(ctx->mask->type == SPDK_MD_MASK_TYPE_USED_CLUSTERS);
+	/* The length of the mask (in bits) must not be greater than the length of the buffer (converted to bits) */
+	assert(ctx->mask->length <= (ctx->super->used_cluster_mask_len * sizeof(
+					     struct spdk_blob_md_page) * 8));
+	/* The length of the mask must be exactly equal to the total number of clusters */
+	assert(ctx->mask->length == ctx->bs->total_clusters);
+
+	rc = bs_load_mask(&ctx->bs->used_clusters, ctx->mask);
+	if (rc < 0) {
+		spdk_free(ctx->mask);
+		bs_load_ctx_fail(ctx, rc);
+		return;
+	}
+
+	ctx->bs->num_free_clusters = spdk_bit_array_count_clear(ctx->bs->used_clusters);
+	assert(ctx->bs->num_free_clusters <= ctx->bs->total_clusters);
+
+	spdk_free(ctx->mask);
+
+	/* Read the used blobids mask */
+	mask_size = ctx->super->used_blobid_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL, SPDK_ENV_SOCKET_ID_ANY,
+				 SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_blobid_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_blobid_mask_len);
+	bs_sequence_read_dev(seq, ctx->mask, lba, lba_count,
+			     bs_load_used_blobids_cpl, ctx);
+}
+
+static void
+bs_load_used_pages_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint64_t		lba, lba_count, mask_size;
+	int			rc;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	/* The type must be correct */
+	assert(ctx->mask->type == SPDK_MD_MASK_TYPE_USED_PAGES);
+	/* The length of the mask (in bits) must not be greater than the length of the buffer (converted to bits) */
+	assert(ctx->mask->length <= (ctx->super->used_page_mask_len * SPDK_BS_PAGE_SIZE *
+				     8));
+	/* The length of the mask must be exactly equal to the size (in pages) of the metadata region */
+	assert(ctx->mask->length == ctx->super->md_len);
+
+	rc = bs_load_mask(&ctx->bs->used_md_pages, ctx->mask);
+	if (rc < 0) {
+		spdk_free(ctx->mask);
+		bs_load_ctx_fail(ctx, rc);
+		return;
+	}
+
+	spdk_free(ctx->mask);
+
+	/* Read the used clusters mask */
+	mask_size = ctx->super->used_cluster_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL, SPDK_ENV_SOCKET_ID_ANY,
+				 SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_cluster_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_cluster_mask_len);
+	bs_sequence_read_dev(seq, ctx->mask, lba, lba_count,
+			     bs_load_used_clusters_cpl, ctx);
+}
+
+static void
+bs_load_read_used_pages(struct spdk_bs_load_ctx *ctx)
+{
+	uint64_t lba, lba_count, mask_size;
+
+	/* Read the used pages mask */
+	mask_size = ctx->super->used_page_mask_len * SPDK_BS_PAGE_SIZE;
+	ctx->mask = spdk_zmalloc(mask_size, 0x1000, NULL,
+				 SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->mask) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	lba = bs_page_to_lba(ctx->bs, ctx->super->used_page_mask_start);
+	lba_count = bs_page_to_lba(ctx->bs, ctx->super->used_page_mask_len);
+	bs_sequence_read_dev(ctx->seq, ctx->mask, lba, lba_count,
+			     bs_load_used_pages_cpl, ctx);
+}
+
+static int
+bs_load_replay_md_parse_page(struct spdk_bs_load_ctx *ctx, struct spdk_blob_md_page *page)
+{
+	struct spdk_blob_store *bs = ctx->bs;
+	struct spdk_blob_md_descriptor *desc;
+	size_t	cur_desc = 0;
+
+	desc = (struct spdk_blob_md_descriptor *)page->descriptors;
+	while (cur_desc < sizeof(page->descriptors)) {
+		if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_PADDING) {
+			if (desc->length == 0) {
+				/* If padding and length are 0, this terminates the page */
+				break;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_RLE) {
+			struct spdk_blob_md_descriptor_extent_rle	*desc_extent_rle;
+			unsigned int				i, j;
+			unsigned int				cluster_count = 0;
+			uint32_t				cluster_idx;
+
+			desc_extent_rle = (struct spdk_blob_md_descriptor_extent_rle *)desc;
+
+			for (i = 0; i < desc_extent_rle->length / sizeof(desc_extent_rle->extents[0]); i++) {
+				for (j = 0; j < desc_extent_rle->extents[i].length; j++) {
+					cluster_idx = desc_extent_rle->extents[i].cluster_idx;
+					/*
+					 * cluster_idx = 0 means an unallocated cluster - don't mark that
+					 * in the used cluster map.
+					 */
+					if (cluster_idx != 0) {
+						spdk_bit_array_set(bs->used_clusters, cluster_idx + j);
+						if (bs->num_free_clusters == 0) {
+							return -ENOSPC;
+						}
+						bs->num_free_clusters--;
+					}
+					cluster_count++;
+				}
+			}
+			if (cluster_count == 0) {
+				return -EINVAL;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE) {
+			struct spdk_blob_md_descriptor_extent_page	*desc_extent;
+			uint32_t					i;
+			uint32_t					cluster_count = 0;
+			uint32_t					cluster_idx;
+			size_t						cluster_idx_length;
+
+			desc_extent = (struct spdk_blob_md_descriptor_extent_page *)desc;
+			cluster_idx_length = desc_extent->length - sizeof(desc_extent->start_cluster_idx);
+
+			if (desc_extent->length <= sizeof(desc_extent->start_cluster_idx) ||
+			    (cluster_idx_length % sizeof(desc_extent->cluster_idx[0]) != 0)) {
+				return -EINVAL;
+			}
+
+			for (i = 0; i < cluster_idx_length / sizeof(desc_extent->cluster_idx[0]); i++) {
+				cluster_idx = desc_extent->cluster_idx[i];
+				/*
+				 * cluster_idx = 0 means an unallocated cluster - don't mark that
+				 * in the used cluster map.
+				 */
+				if (cluster_idx != 0) {
+					if (cluster_idx < desc_extent->start_cluster_idx &&
+					    cluster_idx >= desc_extent->start_cluster_idx + cluster_count) {
+						return -EINVAL;
+					}
+					spdk_bit_array_set(bs->used_clusters, cluster_idx);
+					if (bs->num_free_clusters == 0) {
+						return -ENOSPC;
+					}
+					bs->num_free_clusters--;
+				}
+				cluster_count++;
+			}
+
+			if (cluster_count == 0) {
+				return -EINVAL;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR) {
+			/* Skip this item */
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR_INTERNAL) {
+			/* Skip this item */
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_FLAGS) {
+			/* Skip this item */
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_TABLE) {
+			struct spdk_blob_md_descriptor_extent_table *desc_extent_table;
+			uint32_t num_extent_pages = ctx->num_extent_pages;
+			uint32_t i;
+			size_t extent_pages_length;
+			void *tmp;
+
+			desc_extent_table = (struct spdk_blob_md_descriptor_extent_table *)desc;
+			extent_pages_length = desc_extent_table->length - sizeof(desc_extent_table->num_clusters);
+
+			if (desc_extent_table->length == 0 ||
+			    (extent_pages_length % sizeof(desc_extent_table->extent_page[0]) != 0)) {
+				return -EINVAL;
+			}
+
+			for (i = 0; i < extent_pages_length / sizeof(desc_extent_table->extent_page[0]); i++) {
+				if (desc_extent_table->extent_page[i].page_idx != 0) {
+					if (desc_extent_table->extent_page[i].num_pages != 1) {
+						return -EINVAL;
+					}
+					num_extent_pages += 1;
+				}
+			}
+
+			if (num_extent_pages > 0) {
+				tmp = realloc(ctx->extent_page_num, num_extent_pages * sizeof(uint32_t));
+				if (tmp == NULL) {
+					return -ENOMEM;
+				}
+				ctx->extent_page_num = tmp;
+
+				/* Extent table entries contain md page numbers for extent pages.
+				 * Zeroes represent unallocated extent pages, those are run-length-encoded.
+				 */
+				for (i = 0; i < extent_pages_length / sizeof(desc_extent_table->extent_page[0]); i++) {
+					if (desc_extent_table->extent_page[i].page_idx != 0) {
+						ctx->extent_page_num[ctx->num_extent_pages] = desc_extent_table->extent_page[i].page_idx;
+						ctx->num_extent_pages += 1;
+					}
+				}
+			}
+		} else {
+			/* Error */
+			return -EINVAL;
+		}
+		/* Advance to the next descriptor */
+		cur_desc += sizeof(*desc) + desc->length;
+		if (cur_desc + sizeof(*desc) > sizeof(page->descriptors)) {
+			break;
+		}
+		desc = (struct spdk_blob_md_descriptor *)((uintptr_t)page->descriptors + cur_desc);
+	}
+	return 0;
+}
+
+static bool bs_load_cur_extent_page_valid(struct spdk_blob_md_page *page)
+{
+	uint32_t crc;
+	struct spdk_blob_md_descriptor *desc = (struct spdk_blob_md_descriptor *)page->descriptors;
+	size_t desc_len;
+
+	crc = blob_md_page_calc_crc(page);
+	if (crc != page->crc) {
+		return false;
+	}
+
+	/* Extent page should always be of sequence num 0. */
+	if (page->sequence_num != 0) {
+		return false;
+	}
+
+	/* Descriptor type must be EXTENT_PAGE. */
+	if (desc->type != SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE) {
+		return false;
+	}
+
+	/* Descriptor length cannot exceed the page. */
+	desc_len = sizeof(*desc) + desc->length;
+	if (desc_len > sizeof(page->descriptors)) {
+		return false;
+	}
+
+	/* It has to be the only descriptor in the page. */
+	if (desc_len + sizeof(*desc) <= sizeof(page->descriptors)) {
+		desc = (struct spdk_blob_md_descriptor *)((uintptr_t)page->descriptors + desc_len);
+		if (desc->length != 0) {
+			return false;
+		}
+	}
+
+	return true;
+}
+
+static bool bs_load_cur_md_page_valid(struct spdk_bs_load_ctx *ctx)
+{
+	uint32_t crc;
+	struct spdk_blob_md_page *page = ctx->page;
+
+	crc = blob_md_page_calc_crc(page);
+	if (crc != page->crc) {
+		return false;
+	}
+
+	/* First page of a sequence should match the blobid. */
+	if (page->sequence_num == 0 &&
+	    bs_page_to_blobid(ctx->cur_page) != page->id) {
+		return false;
+	}
+	assert(bs_load_cur_extent_page_valid(page) == false);
+
+	return true;
+}
+
+static void
+bs_load_replay_cur_md_page(struct spdk_bs_load_ctx *ctx);
+
+static void
+bs_load_write_used_clusters_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	bs_load_complete(ctx);
+}
+
+static void
+bs_load_write_used_blobids_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	spdk_free(ctx->mask);
+	ctx->mask = NULL;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	bs_write_used_clusters(seq, ctx, bs_load_write_used_clusters_cpl);
+}
+
+static void
+bs_load_write_used_pages_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	spdk_free(ctx->mask);
+	ctx->mask = NULL;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	bs_write_used_blobids(seq, ctx, bs_load_write_used_blobids_cpl);
+}
+
+static void
+bs_load_write_used_md(struct spdk_bs_load_ctx *ctx)
+{
+	bs_write_used_md(ctx->seq, ctx, bs_load_write_used_pages_cpl);
+}
+
+static void
+bs_load_replay_md_chain_cpl(struct spdk_bs_load_ctx *ctx)
+{
+	uint64_t num_md_clusters;
+	uint64_t i;
+
+	ctx->in_page_chain = false;
+
+	do {
+		ctx->page_index++;
+	} while (spdk_bit_array_get(ctx->bs->used_md_pages, ctx->page_index) == true);
+
+	if (ctx->page_index < ctx->super->md_len) {
+		ctx->cur_page = ctx->page_index;
+		bs_load_replay_cur_md_page(ctx);
+	} else {
+		/* Claim all of the clusters used by the metadata */
+		num_md_clusters = spdk_divide_round_up(ctx->super->md_len, ctx->bs->pages_per_cluster);
+		for (i = 0; i < num_md_clusters; i++) {
+			bs_claim_cluster(ctx->bs, i);
+		}
+		spdk_free(ctx->page);
+		bs_load_write_used_md(ctx);
+	}
+}
+
+static void
+bs_load_replay_extent_page_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint32_t page_num;
+	uint64_t i;
+
+	if (bserrno != 0) {
+		spdk_free(ctx->extent_pages);
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	for (i = 0; i < ctx->num_extent_pages; i++) {
+		/* Extent pages are only read when present within in chain md.
+		 * Integrity of md is not right if that page was not a valid extent page. */
+		if (bs_load_cur_extent_page_valid(&ctx->extent_pages[i]) != true) {
+			spdk_free(ctx->extent_pages);
+			bs_load_ctx_fail(ctx, -EILSEQ);
+			return;
+		}
+
+		page_num = ctx->extent_page_num[i];
+		spdk_bit_array_set(ctx->bs->used_md_pages, page_num);
+		if (bs_load_replay_md_parse_page(ctx, &ctx->extent_pages[i])) {
+			spdk_free(ctx->extent_pages);
+			bs_load_ctx_fail(ctx, -EILSEQ);
+			return;
+		}
+	}
+
+	spdk_free(ctx->extent_pages);
+	free(ctx->extent_page_num);
+	ctx->extent_page_num = NULL;
+	ctx->num_extent_pages = 0;
+
+	bs_load_replay_md_chain_cpl(ctx);
+}
+
+static void
+bs_load_replay_extent_pages(struct spdk_bs_load_ctx *ctx)
+{
+	spdk_bs_batch_t *batch;
+	uint32_t page;
+	uint64_t lba;
+	uint64_t i;
+
+	ctx->extent_pages = spdk_zmalloc(SPDK_BS_PAGE_SIZE * ctx->num_extent_pages, SPDK_BS_PAGE_SIZE,
+					 NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->extent_pages) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	batch = bs_sequence_to_batch(ctx->seq, bs_load_replay_extent_page_cpl, ctx);
+
+	for (i = 0; i < ctx->num_extent_pages; i++) {
+		page = ctx->extent_page_num[i];
+		assert(page < ctx->super->md_len);
+		lba = bs_md_page_to_lba(ctx->bs, page);
+		bs_batch_read_dev(batch, &ctx->extent_pages[i], lba,
+				  bs_byte_to_lba(ctx->bs, SPDK_BS_PAGE_SIZE));
+	}
+
+	bs_batch_close(batch);
+}
+
+static void
+bs_load_replay_md_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint32_t page_num;
+	struct spdk_blob_md_page *page;
+
+	if (bserrno != 0) {
+		bs_load_ctx_fail(ctx, bserrno);
+		return;
+	}
+
+	page_num = ctx->cur_page;
+	page = ctx->page;
+	if (bs_load_cur_md_page_valid(ctx) == true) {
+		if (page->sequence_num == 0 || ctx->in_page_chain == true) {
+			bs_claim_md_page(ctx->bs, page_num);
+			if (page->sequence_num == 0) {
+				spdk_bit_array_set(ctx->bs->used_blobids, page_num);
+			}
+			if (bs_load_replay_md_parse_page(ctx, page)) {
+				bs_load_ctx_fail(ctx, -EILSEQ);
+				return;
+			}
+			if (page->next != SPDK_INVALID_MD_PAGE) {
+				ctx->in_page_chain = true;
+				ctx->cur_page = page->next;
+				bs_load_replay_cur_md_page(ctx);
+				return;
+			}
+			if (ctx->num_extent_pages != 0) {
+				bs_load_replay_extent_pages(ctx);
+				return;
+			}
+		}
+	}
+	bs_load_replay_md_chain_cpl(ctx);
+}
+
+static void
+bs_load_replay_cur_md_page(struct spdk_bs_load_ctx *ctx)
+{
+	uint64_t lba;
+
+	assert(ctx->cur_page < ctx->super->md_len);
+	lba = bs_md_page_to_lba(ctx->bs, ctx->cur_page);
+	bs_sequence_read_dev(ctx->seq, ctx->page, lba,
+			     bs_byte_to_lba(ctx->bs, SPDK_BS_PAGE_SIZE),
+			     bs_load_replay_md_cpl, ctx);
+}
+
+static void
+bs_load_replay_md(struct spdk_bs_load_ctx *ctx)
+{
+	ctx->page_index = 0;
+	ctx->cur_page = 0;
+	ctx->page = spdk_zmalloc(SPDK_BS_PAGE_SIZE, SPDK_BS_PAGE_SIZE,
+				 NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->page) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+	bs_load_replay_cur_md_page(ctx);
+}
+
+static void
+bs_recover(struct spdk_bs_load_ctx *ctx)
+{
+	int		rc;
+
+	rc = spdk_bit_array_resize(&ctx->bs->used_md_pages, ctx->super->md_len);
+	if (rc < 0) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	rc = spdk_bit_array_resize(&ctx->bs->used_blobids, ctx->super->md_len);
+	if (rc < 0) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	rc = spdk_bit_array_resize(&ctx->bs->used_clusters, ctx->bs->total_clusters);
+	if (rc < 0) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	rc = spdk_bit_array_resize(&ctx->bs->open_blobids, ctx->super->md_len);
+	if (rc < 0) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+
+	ctx->bs->num_free_clusters = ctx->bs->total_clusters;
+	bs_load_replay_md(ctx);
+}
+
+static void
+bs_load_super_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx *ctx = cb_arg;
+	uint32_t	crc;
+	int		rc;
+	static const char zeros[SPDK_BLOBSTORE_TYPE_LENGTH];
+
+	if (ctx->super->version > SPDK_BS_VERSION ||
+	    ctx->super->version < SPDK_BS_INITIAL_VERSION) {
+		bs_load_ctx_fail(ctx, -EILSEQ);
+		return;
+	}
+
+	if (memcmp(ctx->super->signature, SPDK_BS_SUPER_BLOCK_SIG,
+		   sizeof(ctx->super->signature)) != 0) {
+		bs_load_ctx_fail(ctx, -EILSEQ);
+		return;
+	}
+
+	crc = blob_md_page_calc_crc(ctx->super);
+	if (crc != ctx->super->crc) {
+		bs_load_ctx_fail(ctx, -EILSEQ);
+		return;
+	}
+
+	if (memcmp(&ctx->bs->bstype, &ctx->super->bstype, SPDK_BLOBSTORE_TYPE_LENGTH) == 0) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Bstype matched - loading blobstore\n");
+	} else if (memcmp(&ctx->bs->bstype, zeros, SPDK_BLOBSTORE_TYPE_LENGTH) == 0) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Bstype wildcard used - loading blobstore regardless bstype\n");
+	} else {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Unexpected bstype\n");
+		SPDK_LOGDUMP(SPDK_LOG_BLOB, "Expected:", ctx->bs->bstype.bstype, SPDK_BLOBSTORE_TYPE_LENGTH);
+		SPDK_LOGDUMP(SPDK_LOG_BLOB, "Found:", ctx->super->bstype.bstype, SPDK_BLOBSTORE_TYPE_LENGTH);
+		bs_load_ctx_fail(ctx, -ENXIO);
+		return;
+	}
+
+	if (ctx->super->size > ctx->bs->dev->blockcnt * ctx->bs->dev->blocklen) {
+		SPDK_NOTICELOG("Size mismatch, dev size: %lu, blobstore size: %lu\n",
+			       ctx->bs->dev->blockcnt * ctx->bs->dev->blocklen, ctx->super->size);
+		bs_load_ctx_fail(ctx, -EILSEQ);
+		return;
+	}
+
+	if (ctx->super->size == 0) {
+		ctx->super->size = ctx->bs->dev->blockcnt * ctx->bs->dev->blocklen;
+	}
+
+	if (ctx->super->io_unit_size == 0) {
+		ctx->super->io_unit_size = SPDK_BS_PAGE_SIZE;
+	}
+
+	/* Parse the super block */
+	ctx->bs->clean = 1;
+	ctx->bs->cluster_sz = ctx->super->cluster_size;
+	ctx->bs->total_clusters = ctx->super->size / ctx->super->cluster_size;
+	ctx->bs->pages_per_cluster = ctx->bs->cluster_sz / SPDK_BS_PAGE_SIZE;
+	if (spdk_u32_is_pow2(ctx->bs->pages_per_cluster)) {
+		ctx->bs->pages_per_cluster_shift = spdk_u32log2(ctx->bs->pages_per_cluster);
+	}
+	ctx->bs->io_unit_size = ctx->super->io_unit_size;
+	rc = spdk_bit_array_resize(&ctx->bs->used_clusters, ctx->bs->total_clusters);
+	if (rc < 0) {
+		bs_load_ctx_fail(ctx, -ENOMEM);
+		return;
+	}
+	ctx->bs->md_start = ctx->super->md_start;
+	ctx->bs->md_len = ctx->super->md_len;
+	ctx->bs->total_data_clusters = ctx->bs->total_clusters - spdk_divide_round_up(
+					       ctx->bs->md_start + ctx->bs->md_len, ctx->bs->pages_per_cluster);
+	ctx->bs->super_blob = ctx->super->super_blob;
+	memcpy(&ctx->bs->bstype, &ctx->super->bstype, sizeof(ctx->super->bstype));
+
+	if (ctx->super->used_blobid_mask_len == 0 || ctx->super->clean == 0) {
+		bs_recover(ctx);
+	} else {
+		bs_load_read_used_pages(ctx);
+	}
+}
+
+void
+spdk_bs_load(struct spdk_bs_dev *dev, struct spdk_bs_opts *o,
+	     spdk_bs_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_blob_store	*bs;
+	struct spdk_bs_cpl	cpl;
+	struct spdk_bs_load_ctx *ctx;
+	struct spdk_bs_opts	opts = {};
+	int err;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Loading blobstore from dev %p\n", dev);
+
+	if ((SPDK_BS_PAGE_SIZE % dev->blocklen) != 0) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "unsupported dev block length of %d\n", dev->blocklen);
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, -EINVAL);
+		return;
+	}
+
+	if (o) {
+		opts = *o;
+	} else {
+		spdk_bs_opts_init(&opts);
+	}
+
+	if (opts.max_md_ops == 0 || opts.max_channel_ops == 0) {
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, -EINVAL);
+		return;
+	}
+
+	err = bs_alloc(dev, &opts, &bs);
+	if (err) {
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, err);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+	ctx->iter_cb_fn = opts.iter_cb_fn;
+	ctx->iter_cb_arg = opts.iter_cb_arg;
+
+	/* Allocate memory for the super block */
+	ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+				  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->super) {
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_HANDLE;
+	cpl.u.bs_handle.cb_fn = cb_fn;
+	cpl.u.bs_handle.cb_arg = cb_arg;
+	cpl.u.bs_handle.bs = bs;
+
+	ctx->seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!ctx->seq) {
+		spdk_free(ctx->super);
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	/* Read the super block */
+	bs_sequence_read_dev(ctx->seq, ctx->super, bs_page_to_lba(bs, 0),
+			     bs_byte_to_lba(bs, sizeof(*ctx->super)),
+			     bs_load_super_cpl, ctx);
+}
+
+/* END spdk_bs_load */
+
+/* START spdk_bs_dump */
+
+struct spdk_bs_dump_ctx {
+	struct spdk_blob_store		*bs;
+	struct spdk_bs_super_block	*super;
+	uint32_t			cur_page;
+	struct spdk_blob_md_page	*page;
+	spdk_bs_sequence_t		*seq;
+	FILE				*fp;
+	spdk_bs_dump_print_xattr	print_xattr_fn;
+	char				xattr_name[4096];
+};
+
+static void
+bs_dump_finish(spdk_bs_sequence_t *seq, struct spdk_bs_dump_ctx *ctx, int bserrno)
+{
+	spdk_free(ctx->super);
+
+	/*
+	 * We need to defer calling bs_call_cpl() until after
+	 * dev destruction, so tuck these away for later use.
+	 */
+	ctx->bs->unload_err = bserrno;
+	memcpy(&ctx->bs->unload_cpl, &seq->cpl, sizeof(struct spdk_bs_cpl));
+	seq->cpl.type = SPDK_BS_CPL_TYPE_NONE;
+
+	bs_sequence_finish(seq, 0);
+	bs_free(ctx->bs);
+	free(ctx);
+}
+
+static void bs_dump_read_md_page(spdk_bs_sequence_t *seq, void *cb_arg);
+
+static void
+bs_dump_print_md_page(struct spdk_bs_dump_ctx *ctx)
+{
+	uint32_t page_idx = ctx->cur_page;
+	struct spdk_blob_md_page *page = ctx->page;
+	struct spdk_blob_md_descriptor *desc;
+	size_t cur_desc = 0;
+	uint32_t crc;
+
+	fprintf(ctx->fp, "=========\n");
+	fprintf(ctx->fp, "Metadata Page Index: %" PRIu32 " (0x%" PRIx32 ")\n", page_idx, page_idx);
+	fprintf(ctx->fp, "Blob ID: 0x%" PRIx64 "\n", page->id);
+
+	crc = blob_md_page_calc_crc(page);
+	fprintf(ctx->fp, "CRC: 0x%" PRIx32 " (%s)\n", page->crc, crc == page->crc ? "OK" : "Mismatch");
+
+	desc = (struct spdk_blob_md_descriptor *)page->descriptors;
+	while (cur_desc < sizeof(page->descriptors)) {
+		if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_PADDING) {
+			if (desc->length == 0) {
+				/* If padding and length are 0, this terminates the page */
+				break;
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_RLE) {
+			struct spdk_blob_md_descriptor_extent_rle	*desc_extent_rle;
+			unsigned int				i;
+
+			desc_extent_rle = (struct spdk_blob_md_descriptor_extent_rle *)desc;
+
+			for (i = 0; i < desc_extent_rle->length / sizeof(desc_extent_rle->extents[0]); i++) {
+				if (desc_extent_rle->extents[i].cluster_idx != 0) {
+					fprintf(ctx->fp, "Allocated Extent - Start: %" PRIu32,
+						desc_extent_rle->extents[i].cluster_idx);
+				} else {
+					fprintf(ctx->fp, "Unallocated Extent - ");
+				}
+				fprintf(ctx->fp, " Length: %" PRIu32, desc_extent_rle->extents[i].length);
+				fprintf(ctx->fp, "\n");
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE) {
+			struct spdk_blob_md_descriptor_extent_page	*desc_extent;
+			unsigned int					i;
+
+			desc_extent = (struct spdk_blob_md_descriptor_extent_page *)desc;
+
+			for (i = 0; i < desc_extent->length / sizeof(desc_extent->cluster_idx[0]); i++) {
+				if (desc_extent->cluster_idx[i] != 0) {
+					fprintf(ctx->fp, "Allocated Extent - Start: %" PRIu32,
+						desc_extent->cluster_idx[i]);
+				} else {
+					fprintf(ctx->fp, "Unallocated Extent");
+				}
+				fprintf(ctx->fp, "\n");
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR) {
+			struct spdk_blob_md_descriptor_xattr *desc_xattr;
+			uint32_t i;
+
+			desc_xattr = (struct spdk_blob_md_descriptor_xattr *)desc;
+
+			if (desc_xattr->length !=
+			    sizeof(desc_xattr->name_length) + sizeof(desc_xattr->value_length) +
+			    desc_xattr->name_length + desc_xattr->value_length) {
+			}
+
+			memcpy(ctx->xattr_name, desc_xattr->name, desc_xattr->name_length);
+			ctx->xattr_name[desc_xattr->name_length] = '\0';
+			fprintf(ctx->fp, "XATTR: name = \"%s\"\n", ctx->xattr_name);
+			fprintf(ctx->fp, "       value = \"");
+			ctx->print_xattr_fn(ctx->fp, ctx->super->bstype.bstype, ctx->xattr_name,
+					    (void *)((uintptr_t)desc_xattr->name + desc_xattr->name_length),
+					    desc_xattr->value_length);
+			fprintf(ctx->fp, "\"\n");
+			for (i = 0; i < desc_xattr->value_length; i++) {
+				if (i % 16 == 0) {
+					fprintf(ctx->fp, "               ");
+				}
+				fprintf(ctx->fp, "%02" PRIx8 " ", *((uint8_t *)desc_xattr->name + desc_xattr->name_length + i));
+				if ((i + 1) % 16 == 0) {
+					fprintf(ctx->fp, "\n");
+				}
+			}
+			if (i % 16 != 0) {
+				fprintf(ctx->fp, "\n");
+			}
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_XATTR_INTERNAL) {
+			/* TODO */
+		} else if (desc->type == SPDK_MD_DESCRIPTOR_TYPE_FLAGS) {
+			/* TODO */
+		} else {
+			/* Error */
+		}
+		/* Advance to the next descriptor */
+		cur_desc += sizeof(*desc) + desc->length;
+		if (cur_desc + sizeof(*desc) > sizeof(page->descriptors)) {
+			break;
+		}
+		desc = (struct spdk_blob_md_descriptor *)((uintptr_t)page->descriptors + cur_desc);
+	}
+}
+
+static void
+bs_dump_read_md_page_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_dump_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		bs_dump_finish(seq, ctx, bserrno);
+		return;
+	}
+
+	if (ctx->page->id != 0) {
+		bs_dump_print_md_page(ctx);
+	}
+
+	ctx->cur_page++;
+
+	if (ctx->cur_page < ctx->super->md_len) {
+		bs_dump_read_md_page(seq, ctx);
+	} else {
+		spdk_free(ctx->page);
+		bs_dump_finish(seq, ctx, 0);
+	}
+}
+
+static void
+bs_dump_read_md_page(spdk_bs_sequence_t *seq, void *cb_arg)
+{
+	struct spdk_bs_dump_ctx *ctx = cb_arg;
+	uint64_t lba;
+
+	assert(ctx->cur_page < ctx->super->md_len);
+	lba = bs_page_to_lba(ctx->bs, ctx->super->md_start + ctx->cur_page);
+	bs_sequence_read_dev(seq, ctx->page, lba,
+			     bs_byte_to_lba(ctx->bs, SPDK_BS_PAGE_SIZE),
+			     bs_dump_read_md_page_cpl, ctx);
+}
+
+static void
+bs_dump_super_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_dump_ctx *ctx = cb_arg;
+
+	fprintf(ctx->fp, "Signature: \"%.8s\" ", ctx->super->signature);
+	if (memcmp(ctx->super->signature, SPDK_BS_SUPER_BLOCK_SIG,
+		   sizeof(ctx->super->signature)) != 0) {
+		fprintf(ctx->fp, "(Mismatch)\n");
+		bs_dump_finish(seq, ctx, bserrno);
+		return;
+	} else {
+		fprintf(ctx->fp, "(OK)\n");
+	}
+	fprintf(ctx->fp, "Version: %" PRIu32 "\n", ctx->super->version);
+	fprintf(ctx->fp, "CRC: 0x%x (%s)\n", ctx->super->crc,
+		(ctx->super->crc == blob_md_page_calc_crc(ctx->super)) ? "OK" : "Mismatch");
+	fprintf(ctx->fp, "Blobstore Type: %.*s\n", SPDK_BLOBSTORE_TYPE_LENGTH, ctx->super->bstype.bstype);
+	fprintf(ctx->fp, "Cluster Size: %" PRIu32 "\n", ctx->super->cluster_size);
+	fprintf(ctx->fp, "Super Blob ID: ");
+	if (ctx->super->super_blob == SPDK_BLOBID_INVALID) {
+		fprintf(ctx->fp, "(None)\n");
+	} else {
+		fprintf(ctx->fp, "%" PRIu64 "\n", ctx->super->super_blob);
+	}
+	fprintf(ctx->fp, "Clean: %" PRIu32 "\n", ctx->super->clean);
+	fprintf(ctx->fp, "Used Metadata Page Mask Start: %" PRIu32 "\n", ctx->super->used_page_mask_start);
+	fprintf(ctx->fp, "Used Metadata Page Mask Length: %" PRIu32 "\n", ctx->super->used_page_mask_len);
+	fprintf(ctx->fp, "Used Cluster Mask Start: %" PRIu32 "\n", ctx->super->used_cluster_mask_start);
+	fprintf(ctx->fp, "Used Cluster Mask Length: %" PRIu32 "\n", ctx->super->used_cluster_mask_len);
+	fprintf(ctx->fp, "Used Blob ID Mask Start: %" PRIu32 "\n", ctx->super->used_blobid_mask_start);
+	fprintf(ctx->fp, "Used Blob ID Mask Length: %" PRIu32 "\n", ctx->super->used_blobid_mask_len);
+	fprintf(ctx->fp, "Metadata Start: %" PRIu32 "\n", ctx->super->md_start);
+	fprintf(ctx->fp, "Metadata Length: %" PRIu32 "\n", ctx->super->md_len);
+
+	ctx->cur_page = 0;
+	ctx->page = spdk_zmalloc(SPDK_BS_PAGE_SIZE, SPDK_BS_PAGE_SIZE,
+				 NULL, SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->page) {
+		bs_dump_finish(seq, ctx, -ENOMEM);
+		return;
+	}
+	bs_dump_read_md_page(seq, ctx);
+}
+
+void
+spdk_bs_dump(struct spdk_bs_dev *dev, FILE *fp, spdk_bs_dump_print_xattr print_xattr_fn,
+	     spdk_bs_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_blob_store	*bs;
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+	struct spdk_bs_dump_ctx *ctx;
+	struct spdk_bs_opts	opts = {};
+	int err;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Dumping blobstore from dev %p\n", dev);
+
+	spdk_bs_opts_init(&opts);
+
+	err = bs_alloc(dev, &opts, &bs);
+	if (err) {
+		dev->destroy(dev);
+		cb_fn(cb_arg, err);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		bs_free(bs);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+	ctx->fp = fp;
+	ctx->print_xattr_fn = print_xattr_fn;
+
+	/* Allocate memory for the super block */
+	ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+				  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->super) {
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	cpl.u.bs_basic.cb_fn = cb_fn;
+	cpl.u.bs_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		spdk_free(ctx->super);
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	/* Read the super block */
+	bs_sequence_read_dev(seq, ctx->super, bs_page_to_lba(bs, 0),
+			     bs_byte_to_lba(bs, sizeof(*ctx->super)),
+			     bs_dump_super_cpl, ctx);
+}
+
+/* END spdk_bs_dump */
+
+/* START spdk_bs_init */
+
+struct spdk_bs_init_ctx {
+	struct spdk_blob_store		*bs;
+	struct spdk_bs_super_block	*super;
+};
+
+static void
+bs_init_persist_super_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_init_ctx *ctx = cb_arg;
+
+	spdk_free(ctx->super);
+	free(ctx);
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+bs_init_trim_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_init_ctx *ctx = cb_arg;
+
+	/* Write super block */
+	bs_sequence_write_dev(seq, ctx->super, bs_page_to_lba(ctx->bs, 0),
+			      bs_byte_to_lba(ctx->bs, sizeof(*ctx->super)),
+			      bs_init_persist_super_cpl, ctx);
+}
+
+void
+spdk_bs_init(struct spdk_bs_dev *dev, struct spdk_bs_opts *o,
+	     spdk_bs_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_init_ctx *ctx;
+	struct spdk_blob_store	*bs;
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+	spdk_bs_batch_t		*batch;
+	uint64_t		num_md_lba;
+	uint64_t		num_md_pages;
+	uint64_t		num_md_clusters;
+	uint32_t		i;
+	struct spdk_bs_opts	opts = {};
+	int			rc;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Initializing blobstore on dev %p\n", dev);
+
+	if ((SPDK_BS_PAGE_SIZE % dev->blocklen) != 0) {
+		SPDK_ERRLOG("unsupported dev block length of %d\n",
+			    dev->blocklen);
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, -EINVAL);
+		return;
+	}
+
+	if (o) {
+		opts = *o;
+	} else {
+		spdk_bs_opts_init(&opts);
+	}
+
+	if (bs_opts_verify(&opts) != 0) {
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, -EINVAL);
+		return;
+	}
+
+	rc = bs_alloc(dev, &opts, &bs);
+	if (rc) {
+		dev->destroy(dev);
+		cb_fn(cb_arg, NULL, rc);
+		return;
+	}
+
+	if (opts.num_md_pages == SPDK_BLOB_OPTS_NUM_MD_PAGES) {
+		/* By default, allocate 1 page per cluster.
+		 * Technically, this over-allocates metadata
+		 * because more metadata will reduce the number
+		 * of usable clusters. This can be addressed with
+		 * more complex math in the future.
+		 */
+		bs->md_len = bs->total_clusters;
+	} else {
+		bs->md_len = opts.num_md_pages;
+	}
+	rc = spdk_bit_array_resize(&bs->used_md_pages, bs->md_len);
+	if (rc < 0) {
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	rc = spdk_bit_array_resize(&bs->used_blobids, bs->md_len);
+	if (rc < 0) {
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	rc = spdk_bit_array_resize(&bs->open_blobids, bs->md_len);
+	if (rc < 0) {
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+
+	/* Allocate memory for the super block */
+	ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+				  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->super) {
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+	memcpy(ctx->super->signature, SPDK_BS_SUPER_BLOCK_SIG,
+	       sizeof(ctx->super->signature));
+	ctx->super->version = SPDK_BS_VERSION;
+	ctx->super->length = sizeof(*ctx->super);
+	ctx->super->super_blob = bs->super_blob;
+	ctx->super->clean = 0;
+	ctx->super->cluster_size = bs->cluster_sz;
+	ctx->super->io_unit_size = bs->io_unit_size;
+	memcpy(&ctx->super->bstype, &bs->bstype, sizeof(bs->bstype));
+
+	/* Calculate how many pages the metadata consumes at the front
+	 * of the disk.
+	 */
+
+	/* The super block uses 1 page */
+	num_md_pages = 1;
+
+	/* The used_md_pages mask requires 1 bit per metadata page, rounded
+	 * up to the nearest page, plus a header.
+	 */
+	ctx->super->used_page_mask_start = num_md_pages;
+	ctx->super->used_page_mask_len = spdk_divide_round_up(sizeof(struct spdk_bs_md_mask) +
+					 spdk_divide_round_up(bs->md_len, 8),
+					 SPDK_BS_PAGE_SIZE);
+	num_md_pages += ctx->super->used_page_mask_len;
+
+	/* The used_clusters mask requires 1 bit per cluster, rounded
+	 * up to the nearest page, plus a header.
+	 */
+	ctx->super->used_cluster_mask_start = num_md_pages;
+	ctx->super->used_cluster_mask_len = spdk_divide_round_up(sizeof(struct spdk_bs_md_mask) +
+					    spdk_divide_round_up(bs->total_clusters, 8),
+					    SPDK_BS_PAGE_SIZE);
+	num_md_pages += ctx->super->used_cluster_mask_len;
+
+	/* The used_blobids mask requires 1 bit per metadata page, rounded
+	 * up to the nearest page, plus a header.
+	 */
+	ctx->super->used_blobid_mask_start = num_md_pages;
+	ctx->super->used_blobid_mask_len = spdk_divide_round_up(sizeof(struct spdk_bs_md_mask) +
+					   spdk_divide_round_up(bs->md_len, 8),
+					   SPDK_BS_PAGE_SIZE);
+	num_md_pages += ctx->super->used_blobid_mask_len;
+
+	/* The metadata region size was chosen above */
+	ctx->super->md_start = bs->md_start = num_md_pages;
+	ctx->super->md_len = bs->md_len;
+	num_md_pages += bs->md_len;
+
+	num_md_lba = bs_page_to_lba(bs, num_md_pages);
+
+	ctx->super->size = dev->blockcnt * dev->blocklen;
+
+	ctx->super->crc = blob_md_page_calc_crc(ctx->super);
+
+	num_md_clusters = spdk_divide_round_up(num_md_pages, bs->pages_per_cluster);
+	if (num_md_clusters > bs->total_clusters) {
+		SPDK_ERRLOG("Blobstore metadata cannot use more clusters than is available, "
+			    "please decrease number of pages reserved for metadata "
+			    "or increase cluster size.\n");
+		spdk_free(ctx->super);
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+	/* Claim all of the clusters used by the metadata */
+	for (i = 0; i < num_md_clusters; i++) {
+		bs_claim_cluster(bs, i);
+	}
+
+	bs->total_data_clusters = bs->num_free_clusters;
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_HANDLE;
+	cpl.u.bs_handle.cb_fn = cb_fn;
+	cpl.u.bs_handle.cb_arg = cb_arg;
+	cpl.u.bs_handle.bs = bs;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		spdk_free(ctx->super);
+		free(ctx);
+		bs_free(bs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	batch = bs_sequence_to_batch(seq, bs_init_trim_cpl, ctx);
+
+	/* Clear metadata space */
+	bs_batch_write_zeroes_dev(batch, 0, num_md_lba);
+
+	switch (opts.clear_method) {
+	case BS_CLEAR_WITH_UNMAP:
+		/* Trim data clusters */
+		bs_batch_unmap_dev(batch, num_md_lba, ctx->bs->dev->blockcnt - num_md_lba);
+		break;
+	case BS_CLEAR_WITH_WRITE_ZEROES:
+		/* Write_zeroes to data clusters */
+		bs_batch_write_zeroes_dev(batch, num_md_lba, ctx->bs->dev->blockcnt - num_md_lba);
+		break;
+	case BS_CLEAR_WITH_NONE:
+	default:
+		break;
+	}
+
+	bs_batch_close(batch);
+}
+
+/* END spdk_bs_init */
+
+/* START spdk_bs_destroy */
+
+static void
+bs_destroy_trim_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_init_ctx *ctx = cb_arg;
+	struct spdk_blob_store *bs = ctx->bs;
+
+	/*
+	 * We need to defer calling bs_call_cpl() until after
+	 * dev destruction, so tuck these away for later use.
+	 */
+	bs->unload_err = bserrno;
+	memcpy(&bs->unload_cpl, &seq->cpl, sizeof(struct spdk_bs_cpl));
+	seq->cpl.type = SPDK_BS_CPL_TYPE_NONE;
+
+	bs_sequence_finish(seq, bserrno);
+
+	bs_free(bs);
+	free(ctx);
+}
+
+void
+spdk_bs_destroy(struct spdk_blob_store *bs, spdk_bs_op_complete cb_fn,
+		void *cb_arg)
+{
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+	struct spdk_bs_init_ctx *ctx;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Destroying blobstore\n");
+
+	if (!TAILQ_EMPTY(&bs->blobs)) {
+		SPDK_ERRLOG("Blobstore still has open blobs\n");
+		cb_fn(cb_arg, -EBUSY);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	cpl.u.bs_basic.cb_fn = cb_fn;
+	cpl.u.bs_basic.cb_arg = cb_arg;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	/* Write zeroes to the super block */
+	bs_sequence_write_zeroes_dev(seq,
+				     bs_page_to_lba(bs, 0),
+				     bs_byte_to_lba(bs, sizeof(struct spdk_bs_super_block)),
+				     bs_destroy_trim_cpl, ctx);
+}
+
+/* END spdk_bs_destroy */
+
+/* START spdk_bs_unload */
+
+static void
+bs_unload_finish(struct spdk_bs_load_ctx *ctx, int bserrno)
+{
+	spdk_bs_sequence_t *seq = ctx->seq;
+
+	spdk_free(ctx->super);
+
+	/*
+	 * We need to defer calling bs_call_cpl() until after
+	 * dev destruction, so tuck these away for later use.
+	 */
+	ctx->bs->unload_err = bserrno;
+	memcpy(&ctx->bs->unload_cpl, &seq->cpl, sizeof(struct spdk_bs_cpl));
+	seq->cpl.type = SPDK_BS_CPL_TYPE_NONE;
+
+	bs_sequence_finish(seq, bserrno);
+
+	bs_free(ctx->bs);
+	free(ctx);
+}
+
+static void
+bs_unload_write_super_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	bs_unload_finish(ctx, bserrno);
+}
+
+static void
+bs_unload_write_used_clusters_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	spdk_free(ctx->mask);
+
+	if (bserrno != 0) {
+		bs_unload_finish(ctx, bserrno);
+		return;
+	}
+
+	ctx->super->clean = 1;
+
+	bs_write_super(seq, ctx->bs, ctx->super, bs_unload_write_super_cpl, ctx);
+}
+
+static void
+bs_unload_write_used_blobids_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	spdk_free(ctx->mask);
+	ctx->mask = NULL;
+
+	if (bserrno != 0) {
+		bs_unload_finish(ctx, bserrno);
+		return;
+	}
+
+	bs_write_used_clusters(seq, ctx, bs_unload_write_used_clusters_cpl);
+}
+
+static void
+bs_unload_write_used_pages_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	spdk_free(ctx->mask);
+	ctx->mask = NULL;
+
+	if (bserrno != 0) {
+		bs_unload_finish(ctx, bserrno);
+		return;
+	}
+
+	bs_write_used_blobids(seq, ctx, bs_unload_write_used_blobids_cpl);
+}
+
+static void
+bs_unload_read_super_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_load_ctx	*ctx = cb_arg;
+
+	if (bserrno != 0) {
+		bs_unload_finish(ctx, bserrno);
+		return;
+	}
+
+	bs_write_used_md(seq, cb_arg, bs_unload_write_used_pages_cpl);
+}
+
+void
+spdk_bs_unload(struct spdk_blob_store *bs, spdk_bs_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_cpl	cpl;
+	struct spdk_bs_load_ctx *ctx;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Syncing blobstore\n");
+
+	if (!TAILQ_EMPTY(&bs->blobs)) {
+		SPDK_ERRLOG("Blobstore still has open blobs\n");
+		cb_fn(cb_arg, -EBUSY);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+
+	ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+				  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->super) {
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	cpl.u.bs_basic.cb_fn = cb_fn;
+	cpl.u.bs_basic.cb_arg = cb_arg;
+
+	ctx->seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!ctx->seq) {
+		spdk_free(ctx->super);
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	/* Read super block */
+	bs_sequence_read_dev(ctx->seq, ctx->super, bs_page_to_lba(bs, 0),
+			     bs_byte_to_lba(bs, sizeof(*ctx->super)),
+			     bs_unload_read_super_cpl, ctx);
+}
+
+/* END spdk_bs_unload */
+
+/* START spdk_bs_set_super */
+
+struct spdk_bs_set_super_ctx {
+	struct spdk_blob_store		*bs;
+	struct spdk_bs_super_block	*super;
+};
+
+static void
+bs_set_super_write_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_set_super_ctx	*ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Unable to write to super block of blobstore\n");
+	}
+
+	spdk_free(ctx->super);
+
+	bs_sequence_finish(seq, bserrno);
+
+	free(ctx);
+}
+
+static void
+bs_set_super_read_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_set_super_ctx	*ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Unable to read super block of blobstore\n");
+		spdk_free(ctx->super);
+		bs_sequence_finish(seq, bserrno);
+		free(ctx);
+		return;
+	}
+
+	bs_write_super(seq, ctx->bs, ctx->super, bs_set_super_write_cpl, ctx);
+}
+
+void
+spdk_bs_set_super(struct spdk_blob_store *bs, spdk_blob_id blobid,
+		  spdk_bs_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_cpl		cpl;
+	spdk_bs_sequence_t		*seq;
+	struct spdk_bs_set_super_ctx	*ctx;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Setting super blob id on blobstore\n");
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->bs = bs;
+
+	ctx->super = spdk_zmalloc(sizeof(*ctx->super), 0x1000, NULL,
+				  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (!ctx->super) {
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BS_BASIC;
+	cpl.u.bs_basic.cb_fn = cb_fn;
+	cpl.u.bs_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		spdk_free(ctx->super);
+		free(ctx);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	bs->super_blob = blobid;
+
+	/* Read super block */
+	bs_sequence_read_dev(seq, ctx->super, bs_page_to_lba(bs, 0),
+			     bs_byte_to_lba(bs, sizeof(*ctx->super)),
+			     bs_set_super_read_cpl, ctx);
+}
+
+/* END spdk_bs_set_super */
+
+void
+spdk_bs_get_super(struct spdk_blob_store *bs,
+		  spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	if (bs->super_blob == SPDK_BLOBID_INVALID) {
+		cb_fn(cb_arg, SPDK_BLOBID_INVALID, -ENOENT);
+	} else {
+		cb_fn(cb_arg, bs->super_blob, 0);
+	}
+}
+
+uint64_t
+spdk_bs_get_cluster_size(struct spdk_blob_store *bs)
+{
+	return bs->cluster_sz;
+}
+
+uint64_t
+spdk_bs_get_page_size(struct spdk_blob_store *bs)
+{
+	return SPDK_BS_PAGE_SIZE;
+}
+
+uint64_t
+spdk_bs_get_io_unit_size(struct spdk_blob_store *bs)
+{
+	return bs->io_unit_size;
+}
+
+uint64_t
+spdk_bs_free_cluster_count(struct spdk_blob_store *bs)
+{
+	return bs->num_free_clusters;
+}
+
+uint64_t
+spdk_bs_total_data_cluster_count(struct spdk_blob_store *bs)
+{
+	return bs->total_data_clusters;
+}
+
+static int
+bs_register_md_thread(struct spdk_blob_store *bs)
+{
+	bs->md_channel = spdk_get_io_channel(bs);
+	if (!bs->md_channel) {
+		SPDK_ERRLOG("Failed to get IO channel.\n");
+		return -1;
+	}
+
+	return 0;
+}
+
+static int
+bs_unregister_md_thread(struct spdk_blob_store *bs)
+{
+	spdk_put_io_channel(bs->md_channel);
+
+	return 0;
+}
+
+spdk_blob_id spdk_blob_get_id(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+
+	return blob->id;
+}
+
+uint64_t spdk_blob_get_num_pages(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+
+	return bs_cluster_to_page(blob->bs, blob->active.num_clusters);
+}
+
+uint64_t spdk_blob_get_num_io_units(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+
+	return spdk_blob_get_num_pages(blob) * bs_io_unit_per_page(blob->bs);
+}
+
+uint64_t spdk_blob_get_num_clusters(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+
+	return blob->active.num_clusters;
+}
+
+/* START spdk_bs_create_blob */
+
+static void
+bs_create_blob_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob *blob = cb_arg;
+	uint32_t page_idx = bs_blobid_to_page(blob->id);
+
+	if (bserrno != 0) {
+		spdk_bit_array_clear(blob->bs->used_blobids, page_idx);
+		bs_release_md_page(blob->bs, page_idx);
+	}
+
+	blob_free(blob);
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static int
+blob_set_xattrs(struct spdk_blob *blob, const struct spdk_blob_xattr_opts *xattrs,
+		bool internal)
+{
+	uint64_t i;
+	size_t value_len = 0;
+	int rc;
+	const void *value = NULL;
+	if (xattrs->count > 0 && xattrs->get_value == NULL) {
+		return -EINVAL;
+	}
+	for (i = 0; i < xattrs->count; i++) {
+		xattrs->get_value(xattrs->ctx, xattrs->names[i], &value, &value_len);
+		if (value == NULL || value_len == 0) {
+			return -EINVAL;
+		}
+		rc = blob_set_xattr(blob, xattrs->names[i], value, value_len, internal);
+		if (rc < 0) {
+			return rc;
+		}
+	}
+	return 0;
+}
+
+static void
+bs_create_blob(struct spdk_blob_store *bs,
+	       const struct spdk_blob_opts *opts,
+	       const struct spdk_blob_xattr_opts *internal_xattrs,
+	       spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	struct spdk_blob	*blob;
+	uint32_t		page_idx;
+	struct spdk_bs_cpl	cpl;
+	struct spdk_blob_opts	opts_default;
+	struct spdk_blob_xattr_opts internal_xattrs_default;
+	spdk_bs_sequence_t	*seq;
+	spdk_blob_id		id;
+	int rc;
+
+	assert(spdk_get_thread() == bs->md_thread);
+
+	page_idx = spdk_bit_array_find_first_clear(bs->used_md_pages, 0);
+	if (page_idx == UINT32_MAX) {
+		cb_fn(cb_arg, 0, -ENOMEM);
+		return;
+	}
+	spdk_bit_array_set(bs->used_blobids, page_idx);
+	bs_claim_md_page(bs, page_idx);
+
+	id = bs_page_to_blobid(page_idx);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Creating blob with id %lu at page %u\n", id, page_idx);
+
+	blob = blob_alloc(bs, id);
+	if (!blob) {
+		spdk_bit_array_clear(bs->used_blobids, page_idx);
+		bs_release_md_page(bs, page_idx);
+		cb_fn(cb_arg, 0, -ENOMEM);
+		return;
+	}
+
+	if (!opts) {
+		spdk_blob_opts_init(&opts_default);
+		opts = &opts_default;
+	}
+
+	blob->use_extent_table = opts->use_extent_table;
+	if (blob->use_extent_table) {
+		blob->invalid_flags |= SPDK_BLOB_EXTENT_TABLE;
+	}
+
+	if (!internal_xattrs) {
+		blob_xattrs_init(&internal_xattrs_default);
+		internal_xattrs = &internal_xattrs_default;
+	}
+
+	rc = blob_set_xattrs(blob, &opts->xattrs, false);
+	if (rc < 0) {
+		blob_free(blob);
+		spdk_bit_array_clear(bs->used_blobids, page_idx);
+		bs_release_md_page(bs, page_idx);
+		cb_fn(cb_arg, 0, rc);
+		return;
+	}
+
+	rc = blob_set_xattrs(blob, internal_xattrs, true);
+	if (rc < 0) {
+		blob_free(blob);
+		spdk_bit_array_clear(bs->used_blobids, page_idx);
+		bs_release_md_page(bs, page_idx);
+		cb_fn(cb_arg, 0, rc);
+		return;
+	}
+
+	if (opts->thin_provision) {
+		blob_set_thin_provision(blob);
+	}
+
+	blob_set_clear_method(blob, opts->clear_method);
+
+	rc = blob_resize(blob, opts->num_clusters);
+	if (rc < 0) {
+		blob_free(blob);
+		spdk_bit_array_clear(bs->used_blobids, page_idx);
+		bs_release_md_page(bs, page_idx);
+		cb_fn(cb_arg, 0, rc);
+		return;
+	}
+	cpl.type = SPDK_BS_CPL_TYPE_BLOBID;
+	cpl.u.blobid.cb_fn = cb_fn;
+	cpl.u.blobid.cb_arg = cb_arg;
+	cpl.u.blobid.blobid = blob->id;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		blob_free(blob);
+		spdk_bit_array_clear(bs->used_blobids, page_idx);
+		bs_release_md_page(bs, page_idx);
+		cb_fn(cb_arg, 0, -ENOMEM);
+		return;
+	}
+
+	blob_persist(seq, blob, bs_create_blob_cpl, blob);
+}
+
+void spdk_bs_create_blob(struct spdk_blob_store *bs,
+			 spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	bs_create_blob(bs, NULL, NULL, cb_fn, cb_arg);
+}
+
+void spdk_bs_create_blob_ext(struct spdk_blob_store *bs, const struct spdk_blob_opts *opts,
+			     spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	bs_create_blob(bs, opts, NULL, cb_fn, cb_arg);
+}
+
+/* END spdk_bs_create_blob */
+
+/* START blob_cleanup */
+
+struct spdk_clone_snapshot_ctx {
+	struct spdk_bs_cpl      cpl;
+	int bserrno;
+	bool frozen;
+
+	struct spdk_io_channel *channel;
+
+	/* Current cluster for inflate operation */
+	uint64_t cluster;
+
+	/* For inflation force allocation of all unallocated clusters and remove
+	 * thin-provisioning. Otherwise only decouple parent and keep clone thin. */
+	bool allocate_all;
+
+	struct {
+		spdk_blob_id id;
+		struct spdk_blob *blob;
+	} original;
+	struct {
+		spdk_blob_id id;
+		struct spdk_blob *blob;
+	} new;
+
+	/* xattrs specified for snapshot/clones only. They have no impact on
+	 * the original blobs xattrs. */
+	const struct spdk_blob_xattr_opts *xattrs;
+};
+
+static void
+bs_clone_snapshot_cleanup_finish(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = cb_arg;
+	struct spdk_bs_cpl *cpl = &ctx->cpl;
+
+	if (bserrno != 0) {
+		if (ctx->bserrno != 0) {
+			SPDK_ERRLOG("Cleanup error %d\n", bserrno);
+		} else {
+			ctx->bserrno = bserrno;
+		}
+	}
+
+	switch (cpl->type) {
+	case SPDK_BS_CPL_TYPE_BLOBID:
+		cpl->u.blobid.cb_fn(cpl->u.blobid.cb_arg, cpl->u.blobid.blobid, ctx->bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_BLOB_BASIC:
+		cpl->u.blob_basic.cb_fn(cpl->u.blob_basic.cb_arg, ctx->bserrno);
+		break;
+	default:
+		SPDK_UNREACHABLE();
+		break;
+	}
+
+	free(ctx);
+}
+
+static void
+bs_snapshot_unfreeze_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+
+	if (bserrno != 0) {
+		if (ctx->bserrno != 0) {
+			SPDK_ERRLOG("Unfreeze error %d\n", bserrno);
+		} else {
+			ctx->bserrno = bserrno;
+		}
+	}
+
+	ctx->original.id = origblob->id;
+	origblob->locked_operation_in_progress = false;
+
+	spdk_blob_close(origblob, bs_clone_snapshot_cleanup_finish, ctx);
+}
+
+static void
+bs_clone_snapshot_origblob_cleanup(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+
+	if (bserrno != 0) {
+		if (ctx->bserrno != 0) {
+			SPDK_ERRLOG("Cleanup error %d\n", bserrno);
+		} else {
+			ctx->bserrno = bserrno;
+		}
+	}
+
+	if (ctx->frozen) {
+		/* Unfreeze any outstanding I/O */
+		blob_unfreeze_io(origblob, bs_snapshot_unfreeze_cpl, ctx);
+	} else {
+		bs_snapshot_unfreeze_cpl(ctx, 0);
+	}
+
+}
+
+static void
+bs_clone_snapshot_newblob_cleanup(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *newblob = ctx->new.blob;
+
+	if (bserrno != 0) {
+		if (ctx->bserrno != 0) {
+			SPDK_ERRLOG("Cleanup error %d\n", bserrno);
+		} else {
+			ctx->bserrno = bserrno;
+		}
+	}
+
+	ctx->new.id = newblob->id;
+	spdk_blob_close(newblob, bs_clone_snapshot_origblob_cleanup, ctx);
+}
+
+/* END blob_cleanup */
+
+/* START spdk_bs_create_snapshot */
+
+static void
+bs_snapshot_swap_cluster_maps(struct spdk_blob *blob1, struct spdk_blob *blob2)
+{
+	uint64_t *cluster_temp;
+	uint32_t *extent_page_temp;
+
+	cluster_temp = blob1->active.clusters;
+	blob1->active.clusters = blob2->active.clusters;
+	blob2->active.clusters = cluster_temp;
+
+	extent_page_temp = blob1->active.extent_pages;
+	blob1->active.extent_pages = blob2->active.extent_pages;
+	blob2->active.extent_pages = extent_page_temp;
+}
+
+static void
+bs_snapshot_origblob_sync_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+	struct spdk_blob *newblob = ctx->new.blob;
+
+	if (bserrno != 0) {
+		bs_snapshot_swap_cluster_maps(newblob, origblob);
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	/* Remove metadata descriptor SNAPSHOT_IN_PROGRESS */
+	bserrno = blob_remove_xattr(newblob, SNAPSHOT_IN_PROGRESS, true);
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	bs_blob_list_add(ctx->original.blob);
+
+	spdk_blob_set_read_only(newblob);
+
+	/* sync snapshot metadata */
+	spdk_blob_sync_md(newblob, bs_clone_snapshot_origblob_cleanup, ctx);
+}
+
+static void
+bs_snapshot_newblob_sync_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+	struct spdk_blob *newblob = ctx->new.blob;
+
+	if (bserrno != 0) {
+		/* return cluster map back to original */
+		bs_snapshot_swap_cluster_maps(newblob, origblob);
+
+		/* Newblob md sync failed. Valid clusters are only present in origblob.
+		 * Since I/O is frozen on origblob, not changes to zeroed out cluster map should have occured.
+		 * Newblob needs to be reverted to thin_provisioned state at creation to properly close. */
+		blob_set_thin_provision(newblob);
+		assert(spdk_mem_all_zero(newblob->active.clusters,
+					 newblob->active.num_clusters * sizeof(*newblob->active.clusters)));
+		assert(spdk_mem_all_zero(newblob->active.extent_pages,
+					 newblob->active.num_extent_pages * sizeof(*newblob->active.extent_pages)));
+
+		bs_clone_snapshot_newblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	/* Set internal xattr for snapshot id */
+	bserrno = blob_set_xattr(origblob, BLOB_SNAPSHOT, &newblob->id, sizeof(spdk_blob_id), true);
+	if (bserrno != 0) {
+		/* return cluster map back to original */
+		bs_snapshot_swap_cluster_maps(newblob, origblob);
+		bs_clone_snapshot_newblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	bs_blob_list_remove(origblob);
+	origblob->parent_id = newblob->id;
+
+	/* Create new back_bs_dev for snapshot */
+	origblob->back_bs_dev = bs_create_blob_bs_dev(newblob);
+	if (origblob->back_bs_dev == NULL) {
+		/* return cluster map back to original */
+		bs_snapshot_swap_cluster_maps(newblob, origblob);
+		bs_clone_snapshot_newblob_cleanup(ctx, -EINVAL);
+		return;
+	}
+
+	/* set clone blob as thin provisioned */
+	blob_set_thin_provision(origblob);
+
+	bs_blob_list_add(newblob);
+
+	/* sync clone metadata */
+	spdk_blob_sync_md(origblob, bs_snapshot_origblob_sync_cpl, ctx);
+}
+
+static void
+bs_snapshot_freeze_cpl(void *cb_arg, int rc)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+	struct spdk_blob *newblob = ctx->new.blob;
+	int bserrno;
+
+	if (rc != 0) {
+		bs_clone_snapshot_newblob_cleanup(ctx, rc);
+		return;
+	}
+
+	ctx->frozen = true;
+
+	/* set new back_bs_dev for snapshot */
+	newblob->back_bs_dev = origblob->back_bs_dev;
+	/* Set invalid flags from origblob */
+	newblob->invalid_flags = origblob->invalid_flags;
+
+	/* inherit parent from original blob if set */
+	newblob->parent_id = origblob->parent_id;
+	if (origblob->parent_id != SPDK_BLOBID_INVALID) {
+		/* Set internal xattr for snapshot id */
+		bserrno = blob_set_xattr(newblob, BLOB_SNAPSHOT,
+					 &origblob->parent_id, sizeof(spdk_blob_id), true);
+		if (bserrno != 0) {
+			bs_clone_snapshot_newblob_cleanup(ctx, bserrno);
+			return;
+		}
+	}
+
+	/* swap cluster maps */
+	bs_snapshot_swap_cluster_maps(newblob, origblob);
+
+	/* Set the clear method on the new blob to match the original. */
+	blob_set_clear_method(newblob, origblob->clear_method);
+
+	/* sync snapshot metadata */
+	spdk_blob_sync_md(newblob, bs_snapshot_newblob_sync_cpl, ctx);
+}
+
+static void
+bs_snapshot_newblob_open_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+	struct spdk_blob *newblob = _blob;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	ctx->new.blob = newblob;
+	assert(spdk_blob_is_thin_provisioned(newblob));
+	assert(spdk_mem_all_zero(newblob->active.clusters,
+				 newblob->active.num_clusters * sizeof(*newblob->active.clusters)));
+	assert(spdk_mem_all_zero(newblob->active.extent_pages,
+				 newblob->active.num_extent_pages * sizeof(*newblob->active.extent_pages)));
+
+	blob_freeze_io(origblob, bs_snapshot_freeze_cpl, ctx);
+}
+
+static void
+bs_snapshot_newblob_create_cpl(void *cb_arg, spdk_blob_id blobid, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *origblob = ctx->original.blob;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	ctx->new.id = blobid;
+	ctx->cpl.u.blobid.blobid = blobid;
+
+	spdk_bs_open_blob(origblob->bs, ctx->new.id, bs_snapshot_newblob_open_cpl, ctx);
+}
+
+
+static void
+bs_xattr_snapshot(void *arg, const char *name,
+		  const void **value, size_t *value_len)
+{
+	assert(strncmp(name, SNAPSHOT_IN_PROGRESS, sizeof(SNAPSHOT_IN_PROGRESS)) == 0);
+
+	struct spdk_blob *blob = (struct spdk_blob *)arg;
+	*value = &blob->id;
+	*value_len = sizeof(blob->id);
+}
+
+static void
+bs_snapshot_origblob_open_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob_opts opts;
+	struct spdk_blob_xattr_opts internal_xattrs;
+	char *xattrs_names[] = { SNAPSHOT_IN_PROGRESS };
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_cleanup_finish(ctx, bserrno);
+		return;
+	}
+
+	ctx->original.blob = _blob;
+
+	if (_blob->data_ro || _blob->md_ro) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot create snapshot from read only blob with id %lu\n",
+			      _blob->id);
+		ctx->bserrno = -EINVAL;
+		spdk_blob_close(_blob, bs_clone_snapshot_cleanup_finish, ctx);
+		return;
+	}
+
+	if (_blob->locked_operation_in_progress) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot create snapshot - another operation in progress\n");
+		ctx->bserrno = -EBUSY;
+		spdk_blob_close(_blob, bs_clone_snapshot_cleanup_finish, ctx);
+		return;
+	}
+
+	_blob->locked_operation_in_progress = true;
+
+	spdk_blob_opts_init(&opts);
+	blob_xattrs_init(&internal_xattrs);
+
+	/* Change the size of new blob to the same as in original blob,
+	 * but do not allocate clusters */
+	opts.thin_provision = true;
+	opts.num_clusters = spdk_blob_get_num_clusters(_blob);
+	opts.use_extent_table = _blob->use_extent_table;
+
+	/* If there are any xattrs specified for snapshot, set them now */
+	if (ctx->xattrs) {
+		memcpy(&opts.xattrs, ctx->xattrs, sizeof(*ctx->xattrs));
+	}
+	/* Set internal xattr SNAPSHOT_IN_PROGRESS */
+	internal_xattrs.count = 1;
+	internal_xattrs.ctx = _blob;
+	internal_xattrs.names = xattrs_names;
+	internal_xattrs.get_value = bs_xattr_snapshot;
+
+	bs_create_blob(_blob->bs, &opts, &internal_xattrs,
+		       bs_snapshot_newblob_create_cpl, ctx);
+}
+
+void spdk_bs_create_snapshot(struct spdk_blob_store *bs, spdk_blob_id blobid,
+			     const struct spdk_blob_xattr_opts *snapshot_xattrs,
+			     spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	struct spdk_clone_snapshot_ctx *ctx = calloc(1, sizeof(*ctx));
+
+	if (!ctx) {
+		cb_fn(cb_arg, SPDK_BLOBID_INVALID, -ENOMEM);
+		return;
+	}
+	ctx->cpl.type = SPDK_BS_CPL_TYPE_BLOBID;
+	ctx->cpl.u.blobid.cb_fn = cb_fn;
+	ctx->cpl.u.blobid.cb_arg = cb_arg;
+	ctx->cpl.u.blobid.blobid = SPDK_BLOBID_INVALID;
+	ctx->bserrno = 0;
+	ctx->frozen = false;
+	ctx->original.id = blobid;
+	ctx->xattrs = snapshot_xattrs;
+
+	spdk_bs_open_blob(bs, ctx->original.id, bs_snapshot_origblob_open_cpl, ctx);
+}
+/* END spdk_bs_create_snapshot */
+
+/* START spdk_bs_create_clone */
+
+static void
+bs_xattr_clone(void *arg, const char *name,
+	       const void **value, size_t *value_len)
+{
+	assert(strncmp(name, BLOB_SNAPSHOT, sizeof(BLOB_SNAPSHOT)) == 0);
+
+	struct spdk_blob *blob = (struct spdk_blob *)arg;
+	*value = &blob->id;
+	*value_len = sizeof(blob->id);
+}
+
+static void
+bs_clone_newblob_open_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *clone = _blob;
+
+	ctx->new.blob = clone;
+	bs_blob_list_add(clone);
+
+	spdk_blob_close(clone, bs_clone_snapshot_origblob_cleanup, ctx);
+}
+
+static void
+bs_clone_newblob_create_cpl(void *cb_arg, spdk_blob_id blobid, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+
+	ctx->cpl.u.blobid.blobid = blobid;
+	spdk_bs_open_blob(ctx->original.blob->bs, blobid, bs_clone_newblob_open_cpl, ctx);
+}
+
+static void
+bs_clone_origblob_open_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx	*ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob_opts		opts;
+	struct spdk_blob_xattr_opts internal_xattrs;
+	char *xattr_names[] = { BLOB_SNAPSHOT };
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_cleanup_finish(ctx, bserrno);
+		return;
+	}
+
+	ctx->original.blob = _blob;
+
+	if (!_blob->data_ro || !_blob->md_ro) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Clone not from read-only blob\n");
+		ctx->bserrno = -EINVAL;
+		spdk_blob_close(_blob, bs_clone_snapshot_cleanup_finish, ctx);
+		return;
+	}
+
+	if (_blob->locked_operation_in_progress) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot create clone - another operation in progress\n");
+		ctx->bserrno = -EBUSY;
+		spdk_blob_close(_blob, bs_clone_snapshot_cleanup_finish, ctx);
+		return;
+	}
+
+	_blob->locked_operation_in_progress = true;
+
+	spdk_blob_opts_init(&opts);
+	blob_xattrs_init(&internal_xattrs);
+
+	opts.thin_provision = true;
+	opts.num_clusters = spdk_blob_get_num_clusters(_blob);
+	opts.use_extent_table = _blob->use_extent_table;
+	if (ctx->xattrs) {
+		memcpy(&opts.xattrs, ctx->xattrs, sizeof(*ctx->xattrs));
+	}
+
+	/* Set internal xattr BLOB_SNAPSHOT */
+	internal_xattrs.count = 1;
+	internal_xattrs.ctx = _blob;
+	internal_xattrs.names = xattr_names;
+	internal_xattrs.get_value = bs_xattr_clone;
+
+	bs_create_blob(_blob->bs, &opts, &internal_xattrs,
+		       bs_clone_newblob_create_cpl, ctx);
+}
+
+void spdk_bs_create_clone(struct spdk_blob_store *bs, spdk_blob_id blobid,
+			  const struct spdk_blob_xattr_opts *clone_xattrs,
+			  spdk_blob_op_with_id_complete cb_fn, void *cb_arg)
+{
+	struct spdk_clone_snapshot_ctx	*ctx = calloc(1, sizeof(*ctx));
+
+	if (!ctx) {
+		cb_fn(cb_arg, SPDK_BLOBID_INVALID, -ENOMEM);
+		return;
+	}
+
+	ctx->cpl.type = SPDK_BS_CPL_TYPE_BLOBID;
+	ctx->cpl.u.blobid.cb_fn = cb_fn;
+	ctx->cpl.u.blobid.cb_arg = cb_arg;
+	ctx->cpl.u.blobid.blobid = SPDK_BLOBID_INVALID;
+	ctx->bserrno = 0;
+	ctx->xattrs = clone_xattrs;
+	ctx->original.id = blobid;
+
+	spdk_bs_open_blob(bs, ctx->original.id, bs_clone_origblob_open_cpl, ctx);
+}
+
+/* END spdk_bs_create_clone */
+
+/* START spdk_bs_inflate_blob */
+
+static void
+bs_inflate_blob_set_parent_cpl(void *cb_arg, struct spdk_blob *_parent, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *_blob = ctx->original.blob;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	assert(_parent != NULL);
+
+	bs_blob_list_remove(_blob);
+	_blob->parent_id = _parent->id;
+	blob_set_xattr(_blob, BLOB_SNAPSHOT, &_blob->parent_id,
+		       sizeof(spdk_blob_id), true);
+
+	_blob->back_bs_dev->destroy(_blob->back_bs_dev);
+	_blob->back_bs_dev = bs_create_blob_bs_dev(_parent);
+	bs_blob_list_add(_blob);
+
+	spdk_blob_sync_md(_blob, bs_clone_snapshot_origblob_cleanup, ctx);
+}
+
+static void
+bs_inflate_blob_done(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *_blob = ctx->original.blob;
+	struct spdk_blob *_parent;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	if (ctx->allocate_all) {
+		/* remove thin provisioning */
+		bs_blob_list_remove(_blob);
+		blob_remove_xattr(_blob, BLOB_SNAPSHOT, true);
+		_blob->invalid_flags = _blob->invalid_flags & ~SPDK_BLOB_THIN_PROV;
+		_blob->back_bs_dev->destroy(_blob->back_bs_dev);
+		_blob->back_bs_dev = NULL;
+		_blob->parent_id = SPDK_BLOBID_INVALID;
+	} else {
+		_parent = ((struct spdk_blob_bs_dev *)(_blob->back_bs_dev))->blob;
+		if (_parent->parent_id != SPDK_BLOBID_INVALID) {
+			/* We must change the parent of the inflated blob */
+			spdk_bs_open_blob(_blob->bs, _parent->parent_id,
+					  bs_inflate_blob_set_parent_cpl, ctx);
+			return;
+		}
+
+		bs_blob_list_remove(_blob);
+		blob_remove_xattr(_blob, BLOB_SNAPSHOT, true);
+		_blob->parent_id = SPDK_BLOBID_INVALID;
+		_blob->back_bs_dev->destroy(_blob->back_bs_dev);
+		_blob->back_bs_dev = bs_create_zeroes_dev();
+	}
+
+	_blob->state = SPDK_BLOB_STATE_DIRTY;
+	spdk_blob_sync_md(_blob, bs_clone_snapshot_origblob_cleanup, ctx);
+}
+
+/* Check if cluster needs allocation */
+static inline bool
+bs_cluster_needs_allocation(struct spdk_blob *blob, uint64_t cluster, bool allocate_all)
+{
+	struct spdk_blob_bs_dev *b;
+
+	assert(blob != NULL);
+
+	if (blob->active.clusters[cluster] != 0) {
+		/* Cluster is already allocated */
+		return false;
+	}
+
+	if (blob->parent_id == SPDK_BLOBID_INVALID) {
+		/* Blob have no parent blob */
+		return allocate_all;
+	}
+
+	b = (struct spdk_blob_bs_dev *)blob->back_bs_dev;
+	return (allocate_all || b->blob->active.clusters[cluster] != 0);
+}
+
+static void
+bs_inflate_blob_touch_next(void *cb_arg, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	struct spdk_blob *_blob = ctx->original.blob;
+	uint64_t offset;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_origblob_cleanup(ctx, bserrno);
+		return;
+	}
+
+	for (; ctx->cluster < _blob->active.num_clusters; ctx->cluster++) {
+		if (bs_cluster_needs_allocation(_blob, ctx->cluster, ctx->allocate_all)) {
+			break;
+		}
+	}
+
+	if (ctx->cluster < _blob->active.num_clusters) {
+		offset = bs_cluster_to_lba(_blob->bs, ctx->cluster);
+
+		/* We may safely increment a cluster before write */
+		ctx->cluster++;
+
+		/* Use zero length write to touch a cluster */
+		spdk_blob_io_write(_blob, ctx->channel, NULL, offset, 0,
+				   bs_inflate_blob_touch_next, ctx);
+	} else {
+		bs_inflate_blob_done(cb_arg, bserrno);
+	}
+}
+
+static void
+bs_inflate_blob_open_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_clone_snapshot_ctx *ctx = (struct spdk_clone_snapshot_ctx *)cb_arg;
+	uint64_t lfc; /* lowest free cluster */
+	uint64_t i;
+
+	if (bserrno != 0) {
+		bs_clone_snapshot_cleanup_finish(ctx, bserrno);
+		return;
+	}
+
+	ctx->original.blob = _blob;
+
+	if (_blob->locked_operation_in_progress) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot inflate blob - another operation in progress\n");
+		ctx->bserrno = -EBUSY;
+		spdk_blob_close(_blob, bs_clone_snapshot_cleanup_finish, ctx);
+		return;
+	}
+
+	_blob->locked_operation_in_progress = true;
+
+	if (!ctx->allocate_all && _blob->parent_id == SPDK_BLOBID_INVALID) {
+		/* This blob have no parent, so we cannot decouple it. */
+		SPDK_ERRLOG("Cannot decouple parent of blob with no parent.\n");
+		bs_clone_snapshot_origblob_cleanup(ctx, -EINVAL);
+		return;
+	}
+
+	if (spdk_blob_is_thin_provisioned(_blob) == false) {
+		/* This is not thin provisioned blob. No need to inflate. */
+		bs_clone_snapshot_origblob_cleanup(ctx, 0);
+		return;
+	}
+
+	/* Do two passes - one to verify that we can obtain enough clusters
+	 * and another to actually claim them.
+	 */
+	lfc = 0;
+	for (i = 0; i < _blob->active.num_clusters; i++) {
+		if (bs_cluster_needs_allocation(_blob, i, ctx->allocate_all)) {
+			lfc = spdk_bit_array_find_first_clear(_blob->bs->used_clusters, lfc);
+			if (lfc == UINT32_MAX) {
+				/* No more free clusters. Cannot satisfy the request */
+				bs_clone_snapshot_origblob_cleanup(ctx, -ENOSPC);
+				return;
+			}
+			lfc++;
+		}
+	}
+
+	ctx->cluster = 0;
+	bs_inflate_blob_touch_next(ctx, 0);
+}
+
+static void
+bs_inflate_blob(struct spdk_blob_store *bs, struct spdk_io_channel *channel,
+		spdk_blob_id blobid, bool allocate_all, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_clone_snapshot_ctx *ctx = calloc(1, sizeof(*ctx));
+
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+	ctx->cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	ctx->cpl.u.bs_basic.cb_fn = cb_fn;
+	ctx->cpl.u.bs_basic.cb_arg = cb_arg;
+	ctx->bserrno = 0;
+	ctx->original.id = blobid;
+	ctx->channel = channel;
+	ctx->allocate_all = allocate_all;
+
+	spdk_bs_open_blob(bs, ctx->original.id, bs_inflate_blob_open_cpl, ctx);
+}
+
+void
+spdk_bs_inflate_blob(struct spdk_blob_store *bs, struct spdk_io_channel *channel,
+		     spdk_blob_id blobid, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	bs_inflate_blob(bs, channel, blobid, true, cb_fn, cb_arg);
+}
+
+void
+spdk_bs_blob_decouple_parent(struct spdk_blob_store *bs, struct spdk_io_channel *channel,
+			     spdk_blob_id blobid, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	bs_inflate_blob(bs, channel, blobid, false, cb_fn, cb_arg);
+}
+/* END spdk_bs_inflate_blob */
+
+/* START spdk_blob_resize */
+struct spdk_bs_resize_ctx {
+	spdk_blob_op_complete cb_fn;
+	void *cb_arg;
+	struct spdk_blob *blob;
+	uint64_t sz;
+	int rc;
+};
+
+static void
+bs_resize_unfreeze_cpl(void *cb_arg, int rc)
+{
+	struct spdk_bs_resize_ctx *ctx = (struct spdk_bs_resize_ctx *)cb_arg;
+
+	if (rc != 0) {
+		SPDK_ERRLOG("Unfreeze failed, rc=%d\n", rc);
+	}
+
+	if (ctx->rc != 0) {
+		SPDK_ERRLOG("Unfreeze failed, ctx->rc=%d\n", ctx->rc);
+		rc = ctx->rc;
+	}
+
+	ctx->blob->locked_operation_in_progress = false;
+
+	ctx->cb_fn(ctx->cb_arg, rc);
+	free(ctx);
+}
+
+static void
+bs_resize_freeze_cpl(void *cb_arg, int rc)
+{
+	struct spdk_bs_resize_ctx *ctx = (struct spdk_bs_resize_ctx *)cb_arg;
+
+	if (rc != 0) {
+		ctx->blob->locked_operation_in_progress = false;
+		ctx->cb_fn(ctx->cb_arg, rc);
+		free(ctx);
+		return;
+	}
+
+	ctx->rc = blob_resize(ctx->blob, ctx->sz);
+
+	blob_unfreeze_io(ctx->blob, bs_resize_unfreeze_cpl, ctx);
+}
+
+void
+spdk_blob_resize(struct spdk_blob *blob, uint64_t sz, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_resize_ctx *ctx;
+
+	blob_verify_md_op(blob);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Resizing blob %lu to %lu clusters\n", blob->id, sz);
+
+	if (blob->md_ro) {
+		cb_fn(cb_arg, -EPERM);
+		return;
+	}
+
+	if (sz == blob->active.num_clusters) {
+		cb_fn(cb_arg, 0);
+		return;
+	}
+
+	if (blob->locked_operation_in_progress) {
+		cb_fn(cb_arg, -EBUSY);
+		return;
+	}
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	blob->locked_operation_in_progress = true;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+	ctx->blob = blob;
+	ctx->sz = sz;
+	blob_freeze_io(blob, bs_resize_freeze_cpl, ctx);
+}
+
+/* END spdk_blob_resize */
+
+
+/* START spdk_bs_delete_blob */
+
+static void
+bs_delete_close_cpl(void *cb_arg, int bserrno)
+{
+	spdk_bs_sequence_t *seq = cb_arg;
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+bs_delete_persist_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob *blob = cb_arg;
+
+	if (bserrno != 0) {
+		/*
+		 * We already removed this blob from the blobstore tailq, so
+		 *  we need to free it here since this is the last reference
+		 *  to it.
+		 */
+		blob_free(blob);
+		bs_delete_close_cpl(seq, bserrno);
+		return;
+	}
+
+	/*
+	 * This will immediately decrement the ref_count and call
+	 *  the completion routine since the metadata state is clean.
+	 *  By calling spdk_blob_close, we reduce the number of call
+	 *  points into code that touches the blob->open_ref count
+	 *  and the blobstore's blob list.
+	 */
+	spdk_blob_close(blob, bs_delete_close_cpl, seq);
+}
+
+struct delete_snapshot_ctx {
+	struct spdk_blob_list *parent_snapshot_entry;
+	struct spdk_blob *snapshot;
+	bool snapshot_md_ro;
+	struct spdk_blob *clone;
+	bool clone_md_ro;
+	spdk_blob_op_with_handle_complete cb_fn;
+	void *cb_arg;
+	int bserrno;
+};
+
+static void
+delete_blob_cleanup_finish(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		SPDK_ERRLOG("Snapshot cleanup error %d\n", bserrno);
+	}
+
+	assert(ctx != NULL);
+
+	if (bserrno != 0 && ctx->bserrno == 0) {
+		ctx->bserrno = bserrno;
+	}
+
+	ctx->cb_fn(ctx->cb_arg, ctx->snapshot, ctx->bserrno);
+	free(ctx);
+}
+
+static void
+delete_snapshot_cleanup_snapshot(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	if (bserrno != 0) {
+		ctx->bserrno = bserrno;
+		SPDK_ERRLOG("Clone cleanup error %d\n", bserrno);
+	}
+
+	if (ctx->bserrno != 0) {
+		assert(blob_lookup(ctx->snapshot->bs, ctx->snapshot->id) == NULL);
+		TAILQ_INSERT_HEAD(&ctx->snapshot->bs->blobs, ctx->snapshot, link);
+		spdk_bit_array_set(ctx->snapshot->bs->open_blobids, ctx->snapshot->id);
+	}
+
+	ctx->snapshot->locked_operation_in_progress = false;
+	ctx->snapshot->md_ro = ctx->snapshot_md_ro;
+
+	spdk_blob_close(ctx->snapshot, delete_blob_cleanup_finish, ctx);
+}
+
+static void
+delete_snapshot_cleanup_clone(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	ctx->clone->locked_operation_in_progress = false;
+	ctx->clone->md_ro = ctx->clone_md_ro;
+
+	spdk_blob_close(ctx->clone, delete_snapshot_cleanup_snapshot, ctx);
+}
+
+static void
+delete_snapshot_unfreeze_cpl(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	if (bserrno) {
+		ctx->bserrno = bserrno;
+		delete_snapshot_cleanup_clone(ctx, 0);
+		return;
+	}
+
+	ctx->clone->locked_operation_in_progress = false;
+	spdk_blob_close(ctx->clone, delete_blob_cleanup_finish, ctx);
+}
+
+static void
+delete_snapshot_sync_snapshot_cpl(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+	struct spdk_blob_list *parent_snapshot_entry = NULL;
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+	struct spdk_blob_list *snapshot_clone_entry = NULL;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to sync MD on blob\n");
+		ctx->bserrno = bserrno;
+		delete_snapshot_cleanup_clone(ctx, 0);
+		return;
+	}
+
+	/* Get snapshot entry for the snapshot we want to remove */
+	snapshot_entry = bs_get_snapshot_entry(ctx->snapshot->bs, ctx->snapshot->id);
+
+	assert(snapshot_entry != NULL);
+
+	/* Remove clone entry in this snapshot (at this point there can be only one clone) */
+	clone_entry = TAILQ_FIRST(&snapshot_entry->clones);
+	assert(clone_entry != NULL);
+	TAILQ_REMOVE(&snapshot_entry->clones, clone_entry, link);
+	snapshot_entry->clone_count--;
+	assert(TAILQ_EMPTY(&snapshot_entry->clones));
+
+	if (ctx->snapshot->parent_id != SPDK_BLOBID_INVALID) {
+		/* This snapshot is at the same time a clone of another snapshot - we need to
+		 * update parent snapshot (remove current clone, add new one inherited from
+		 * the snapshot that is being removed) */
+
+		/* Get snapshot entry for parent snapshot and clone entry within that snapshot for
+		 * snapshot that we are removing */
+		blob_get_snapshot_and_clone_entries(ctx->snapshot, &parent_snapshot_entry,
+						    &snapshot_clone_entry);
+
+		/* Switch clone entry in parent snapshot */
+		TAILQ_INSERT_TAIL(&parent_snapshot_entry->clones, clone_entry, link);
+		TAILQ_REMOVE(&parent_snapshot_entry->clones, snapshot_clone_entry, link);
+		free(snapshot_clone_entry);
+	} else {
+		/* No parent snapshot - just remove clone entry */
+		free(clone_entry);
+	}
+
+	/* Restore md_ro flags */
+	ctx->clone->md_ro = ctx->clone_md_ro;
+	ctx->snapshot->md_ro = ctx->snapshot_md_ro;
+
+	blob_unfreeze_io(ctx->clone, delete_snapshot_unfreeze_cpl, ctx);
+}
+
+static void
+delete_snapshot_sync_clone_cpl(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+	uint64_t i;
+
+	ctx->snapshot->md_ro = false;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to sync MD on clone\n");
+		ctx->bserrno = bserrno;
+
+		/* Restore snapshot to previous state */
+		bserrno = blob_remove_xattr(ctx->snapshot, SNAPSHOT_PENDING_REMOVAL, true);
+		if (bserrno != 0) {
+			delete_snapshot_cleanup_clone(ctx, bserrno);
+			return;
+		}
+
+		spdk_blob_sync_md(ctx->snapshot, delete_snapshot_cleanup_clone, ctx);
+		return;
+	}
+
+	/* Clear cluster map entries for snapshot */
+	for (i = 0; i < ctx->snapshot->active.num_clusters && i < ctx->clone->active.num_clusters; i++) {
+		if (ctx->clone->active.clusters[i] == ctx->snapshot->active.clusters[i]) {
+			ctx->snapshot->active.clusters[i] = 0;
+		}
+	}
+	for (i = 0; i < ctx->snapshot->active.num_extent_pages &&
+	     i < ctx->clone->active.num_extent_pages; i++) {
+		if (ctx->clone->active.extent_pages[i] == ctx->snapshot->active.extent_pages[i]) {
+			ctx->snapshot->active.extent_pages[i] = 0;
+		}
+	}
+
+	blob_set_thin_provision(ctx->snapshot);
+	ctx->snapshot->state = SPDK_BLOB_STATE_DIRTY;
+
+	if (ctx->parent_snapshot_entry != NULL) {
+		ctx->snapshot->back_bs_dev = NULL;
+	}
+
+	spdk_blob_sync_md(ctx->snapshot, delete_snapshot_sync_snapshot_cpl, ctx);
+}
+
+static void
+delete_snapshot_sync_snapshot_xattr_cpl(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+	uint64_t i;
+
+	/* Temporarily override md_ro flag for clone for MD modification */
+	ctx->clone_md_ro = ctx->clone->md_ro;
+	ctx->clone->md_ro = false;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to sync MD with xattr on blob\n");
+		ctx->bserrno = bserrno;
+		delete_snapshot_cleanup_clone(ctx, 0);
+		return;
+	}
+
+	/* Copy snapshot map to clone map (only unallocated clusters in clone) */
+	for (i = 0; i < ctx->snapshot->active.num_clusters && i < ctx->clone->active.num_clusters; i++) {
+		if (ctx->clone->active.clusters[i] == 0) {
+			ctx->clone->active.clusters[i] = ctx->snapshot->active.clusters[i];
+		}
+	}
+	for (i = 0; i < ctx->snapshot->active.num_extent_pages &&
+	     i < ctx->clone->active.num_extent_pages; i++) {
+		if (ctx->clone->active.extent_pages[i] == 0) {
+			ctx->clone->active.extent_pages[i] = ctx->snapshot->active.extent_pages[i];
+		}
+	}
+
+	/* Delete old backing bs_dev from clone (related to snapshot that will be removed) */
+	ctx->clone->back_bs_dev->destroy(ctx->clone->back_bs_dev);
+
+	/* Set/remove snapshot xattr and switch parent ID and backing bs_dev on clone... */
+	if (ctx->parent_snapshot_entry != NULL) {
+		/* ...to parent snapshot */
+		ctx->clone->parent_id = ctx->parent_snapshot_entry->id;
+		ctx->clone->back_bs_dev = ctx->snapshot->back_bs_dev;
+		blob_set_xattr(ctx->clone, BLOB_SNAPSHOT, &ctx->parent_snapshot_entry->id,
+			       sizeof(spdk_blob_id),
+			       true);
+	} else {
+		/* ...to blobid invalid and zeroes dev */
+		ctx->clone->parent_id = SPDK_BLOBID_INVALID;
+		ctx->clone->back_bs_dev = bs_create_zeroes_dev();
+		blob_remove_xattr(ctx->clone, BLOB_SNAPSHOT, true);
+	}
+
+	spdk_blob_sync_md(ctx->clone, delete_snapshot_sync_clone_cpl, ctx);
+}
+
+static void
+delete_snapshot_freeze_io_cb(void *cb_arg, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to freeze I/O on clone\n");
+		ctx->bserrno = bserrno;
+		delete_snapshot_cleanup_clone(ctx, 0);
+		return;
+	}
+
+	/* Temporarily override md_ro flag for snapshot for MD modification */
+	ctx->snapshot_md_ro = ctx->snapshot->md_ro;
+	ctx->snapshot->md_ro = false;
+
+	/* Mark blob as pending for removal for power failure safety, use clone id for recovery */
+	ctx->bserrno = blob_set_xattr(ctx->snapshot, SNAPSHOT_PENDING_REMOVAL, &ctx->clone->id,
+				      sizeof(spdk_blob_id), true);
+	if (ctx->bserrno != 0) {
+		delete_snapshot_cleanup_clone(ctx, 0);
+		return;
+	}
+
+	spdk_blob_sync_md(ctx->snapshot, delete_snapshot_sync_snapshot_xattr_cpl, ctx);
+}
+
+static void
+delete_snapshot_open_clone_cb(void *cb_arg, struct spdk_blob *clone, int bserrno)
+{
+	struct delete_snapshot_ctx *ctx = cb_arg;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to open clone\n");
+		ctx->bserrno = bserrno;
+		delete_snapshot_cleanup_snapshot(ctx, 0);
+		return;
+	}
+
+	ctx->clone = clone;
+
+	if (clone->locked_operation_in_progress) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot remove blob - another operation in progress on its clone\n");
+		ctx->bserrno = -EBUSY;
+		spdk_blob_close(ctx->clone, delete_snapshot_cleanup_snapshot, ctx);
+		return;
+	}
+
+	clone->locked_operation_in_progress = true;
+
+	blob_freeze_io(clone, delete_snapshot_freeze_io_cb, ctx);
+}
+
+static void
+update_clone_on_snapshot_deletion(struct spdk_blob *snapshot, struct delete_snapshot_ctx *ctx)
+{
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+	struct spdk_blob_list *snapshot_clone_entry = NULL;
+
+	/* Get snapshot entry for the snapshot we want to remove */
+	snapshot_entry = bs_get_snapshot_entry(snapshot->bs, snapshot->id);
+
+	assert(snapshot_entry != NULL);
+
+	/* Get clone of the snapshot (at this point there can be only one clone) */
+	clone_entry = TAILQ_FIRST(&snapshot_entry->clones);
+	assert(snapshot_entry->clone_count == 1);
+	assert(clone_entry != NULL);
+
+	/* Get snapshot entry for parent snapshot and clone entry within that snapshot for
+	 * snapshot that we are removing */
+	blob_get_snapshot_and_clone_entries(snapshot, &ctx->parent_snapshot_entry,
+					    &snapshot_clone_entry);
+
+	spdk_bs_open_blob(snapshot->bs, clone_entry->id, delete_snapshot_open_clone_cb, ctx);
+}
+
+static void
+bs_delete_blob_finish(void *cb_arg, struct spdk_blob *blob, int bserrno)
+{
+	spdk_bs_sequence_t *seq = cb_arg;
+	struct spdk_blob_list *snapshot_entry = NULL;
+	uint32_t page_num;
+
+	if (bserrno) {
+		SPDK_ERRLOG("Failed to remove blob\n");
+		bs_sequence_finish(seq, bserrno);
+		return;
+	}
+
+	/* Remove snapshot from the list */
+	snapshot_entry = bs_get_snapshot_entry(blob->bs, blob->id);
+	if (snapshot_entry != NULL) {
+		TAILQ_REMOVE(&blob->bs->snapshots, snapshot_entry, link);
+		free(snapshot_entry);
+	}
+
+	page_num = bs_blobid_to_page(blob->id);
+	spdk_bit_array_clear(blob->bs->used_blobids, page_num);
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+	blob->active.num_pages = 0;
+	blob_resize(blob, 0);
+
+	blob_persist(seq, blob, bs_delete_persist_cpl, blob);
+}
+
+static int
+bs_is_blob_deletable(struct spdk_blob *blob, bool *update_clone)
+{
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+	struct spdk_blob *clone = NULL;
+	bool has_one_clone = false;
+
+	/* Check if this is a snapshot with clones */
+	snapshot_entry = bs_get_snapshot_entry(blob->bs, blob->id);
+	if (snapshot_entry != NULL) {
+		if (snapshot_entry->clone_count > 1) {
+			SPDK_ERRLOG("Cannot remove snapshot with more than one clone\n");
+			return -EBUSY;
+		} else if (snapshot_entry->clone_count == 1) {
+			has_one_clone = true;
+		}
+	}
+
+	/* Check if someone has this blob open (besides this delete context):
+	 * - open_ref = 1 - only this context opened blob, so it is ok to remove it
+	 * - open_ref <= 2 && has_one_clone = true - clone is holding snapshot
+	 *	and that is ok, because we will update it accordingly */
+	if (blob->open_ref <= 2 && has_one_clone) {
+		clone_entry = TAILQ_FIRST(&snapshot_entry->clones);
+		assert(clone_entry != NULL);
+		clone = blob_lookup(blob->bs, clone_entry->id);
+
+		if (blob->open_ref == 2 && clone == NULL) {
+			/* Clone is closed and someone else opened this blob */
+			SPDK_ERRLOG("Cannot remove snapshot because it is open\n");
+			return -EBUSY;
+		}
+
+		*update_clone = true;
+		return 0;
+	}
+
+	if (blob->open_ref > 1) {
+		SPDK_ERRLOG("Cannot remove snapshot because it is open\n");
+		return -EBUSY;
+	}
+
+	assert(has_one_clone == false);
+	*update_clone = false;
+	return 0;
+}
+
+static void
+bs_delete_enomem_close_cpl(void *cb_arg, int bserrno)
+{
+	spdk_bs_sequence_t *seq = cb_arg;
+
+	bs_sequence_finish(seq, -ENOMEM);
+}
+
+static void
+bs_delete_open_cpl(void *cb_arg, struct spdk_blob *blob, int bserrno)
+{
+	spdk_bs_sequence_t *seq = cb_arg;
+	struct delete_snapshot_ctx *ctx;
+	bool update_clone = false;
+
+	if (bserrno != 0) {
+		bs_sequence_finish(seq, bserrno);
+		return;
+	}
+
+	blob_verify_md_op(blob);
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (ctx == NULL) {
+		spdk_blob_close(blob, bs_delete_enomem_close_cpl, seq);
+		return;
+	}
+
+	ctx->snapshot = blob;
+	ctx->cb_fn = bs_delete_blob_finish;
+	ctx->cb_arg = seq;
+
+	/* Check if blob can be removed and if it is a snapshot with clone on top of it */
+	ctx->bserrno = bs_is_blob_deletable(blob, &update_clone);
+	if (ctx->bserrno) {
+		spdk_blob_close(blob, delete_blob_cleanup_finish, ctx);
+		return;
+	}
+
+	if (blob->locked_operation_in_progress) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Cannot remove blob - another operation in progress\n");
+		ctx->bserrno = -EBUSY;
+		spdk_blob_close(blob, delete_blob_cleanup_finish, ctx);
+		return;
+	}
+
+	blob->locked_operation_in_progress = true;
+
+	/*
+	 * Remove the blob from the blob_store list now, to ensure it does not
+	 *  get returned after this point by blob_lookup().
+	 */
+	spdk_bit_array_clear(blob->bs->open_blobids, blob->id);
+	TAILQ_REMOVE(&blob->bs->blobs, blob, link);
+
+	if (update_clone) {
+		/* This blob is a snapshot with active clone - update clone first */
+		update_clone_on_snapshot_deletion(blob, ctx);
+	} else {
+		/* This blob does not have any clones - just remove it */
+		bs_blob_list_remove(blob);
+		bs_delete_blob_finish(seq, blob, 0);
+		free(ctx);
+	}
+}
+
+void
+spdk_bs_delete_blob(struct spdk_blob_store *bs, spdk_blob_id blobid,
+		    spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Deleting blob %lu\n", blobid);
+
+	assert(spdk_get_thread() == bs->md_thread);
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	spdk_bs_open_blob(bs, blobid, bs_delete_open_cpl, seq);
+}
+
+/* END spdk_bs_delete_blob */
+
+/* START spdk_bs_open_blob */
+
+static void
+bs_open_blob_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob *blob = cb_arg;
+
+	if (bserrno != 0) {
+		blob_free(blob);
+		seq->cpl.u.blob_handle.blob = NULL;
+		bs_sequence_finish(seq, bserrno);
+		return;
+	}
+
+	blob->open_ref++;
+
+	spdk_bit_array_set(blob->bs->open_blobids, blob->id);
+	TAILQ_INSERT_HEAD(&blob->bs->blobs, blob, link);
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+bs_open_blob(struct spdk_blob_store *bs,
+	     spdk_blob_id blobid,
+	     struct spdk_blob_open_opts *opts,
+	     spdk_blob_op_with_handle_complete cb_fn,
+	     void *cb_arg)
+{
+	struct spdk_blob		*blob;
+	struct spdk_bs_cpl		cpl;
+	struct spdk_blob_open_opts	opts_default;
+	spdk_bs_sequence_t		*seq;
+	uint32_t			page_num;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Opening blob %lu\n", blobid);
+	assert(spdk_get_thread() == bs->md_thread);
+
+	page_num = bs_blobid_to_page(blobid);
+	if (spdk_bit_array_get(bs->used_blobids, page_num) == false) {
+		/* Invalid blobid */
+		cb_fn(cb_arg, NULL, -ENOENT);
+		return;
+	}
+
+	blob = blob_lookup(bs, blobid);
+	if (blob) {
+		blob->open_ref++;
+		cb_fn(cb_arg, blob, 0);
+		return;
+	}
+
+	blob = blob_alloc(bs, blobid);
+	if (!blob) {
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	if (!opts) {
+		spdk_blob_open_opts_init(&opts_default);
+		opts = &opts_default;
+	}
+
+	blob->clear_method = opts->clear_method;
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_HANDLE;
+	cpl.u.blob_handle.cb_fn = cb_fn;
+	cpl.u.blob_handle.cb_arg = cb_arg;
+	cpl.u.blob_handle.blob = blob;
+
+	seq = bs_sequence_start(bs->md_channel, &cpl);
+	if (!seq) {
+		blob_free(blob);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	blob_load(seq, blob, bs_open_blob_cpl, blob);
+}
+
+void spdk_bs_open_blob(struct spdk_blob_store *bs, spdk_blob_id blobid,
+		       spdk_blob_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	bs_open_blob(bs, blobid, NULL, cb_fn, cb_arg);
+}
+
+void spdk_bs_open_blob_ext(struct spdk_blob_store *bs, spdk_blob_id blobid,
+			   struct spdk_blob_open_opts *opts, spdk_blob_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	bs_open_blob(bs, blobid, opts, cb_fn, cb_arg);
+}
+
+/* END spdk_bs_open_blob */
+
+/* START spdk_blob_set_read_only */
+int spdk_blob_set_read_only(struct spdk_blob *blob)
+{
+	blob_verify_md_op(blob);
+
+	blob->data_ro_flags |= SPDK_BLOB_READ_ONLY;
+
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+	return 0;
+}
+/* END spdk_blob_set_read_only */
+
+/* START spdk_blob_sync_md */
+
+static void
+blob_sync_md_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob *blob = cb_arg;
+
+	if (bserrno == 0 && (blob->data_ro_flags & SPDK_BLOB_READ_ONLY)) {
+		blob->data_ro = true;
+		blob->md_ro = true;
+	}
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+blob_sync_md(struct spdk_blob *blob, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(blob->bs->md_channel, &cpl);
+	if (!seq) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	blob_persist(seq, blob, blob_sync_md_cpl, blob);
+}
+
+void
+spdk_blob_sync_md(struct spdk_blob *blob, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_verify_md_op(blob);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Syncing blob %lu\n", blob->id);
+
+	if (blob->md_ro) {
+		assert(blob->state == SPDK_BLOB_STATE_CLEAN);
+		cb_fn(cb_arg, 0);
+		return;
+	}
+
+	blob_sync_md(blob, cb_fn, cb_arg);
+}
+
+/* END spdk_blob_sync_md */
+
+struct spdk_blob_insert_cluster_ctx {
+	struct spdk_thread	*thread;
+	struct spdk_blob	*blob;
+	uint32_t		cluster_num;	/* cluster index in blob */
+	uint32_t		cluster;	/* cluster on disk */
+	uint32_t		extent_page;	/* extent page on disk */
+	int			rc;
+	spdk_blob_op_complete	cb_fn;
+	void			*cb_arg;
+};
+
+static void
+blob_insert_cluster_msg_cpl(void *arg)
+{
+	struct spdk_blob_insert_cluster_ctx *ctx = arg;
+
+	ctx->cb_fn(ctx->cb_arg, ctx->rc);
+	free(ctx);
+}
+
+static void
+blob_insert_cluster_msg_cb(void *arg, int bserrno)
+{
+	struct spdk_blob_insert_cluster_ctx *ctx = arg;
+
+	ctx->rc = bserrno;
+	spdk_thread_send_msg(ctx->thread, blob_insert_cluster_msg_cpl, ctx);
+}
+
+static void
+blob_persist_extent_page_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob_md_page        *page = cb_arg;
+
+	bs_sequence_finish(seq, bserrno);
+	spdk_free(page);
+}
+
+static void
+blob_insert_extent(struct spdk_blob *blob, uint32_t extent, uint64_t cluster_num,
+		   spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	spdk_bs_sequence_t		*seq;
+	struct spdk_bs_cpl		cpl;
+	struct spdk_blob_md_page	*page = NULL;
+	uint32_t			page_count = 0;
+	int				rc;
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(blob->bs->md_channel, &cpl);
+	if (!seq) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+	rc = blob_serialize_add_page(blob, &page, &page_count, &page);
+	if (rc < 0) {
+		bs_sequence_finish(seq, rc);
+		return;
+	}
+
+	blob_serialize_extent_page(blob, cluster_num, page);
+
+	page->crc = blob_md_page_calc_crc(page);
+
+	assert(spdk_bit_array_get(blob->bs->used_md_pages, extent) == true);
+
+	bs_sequence_write_dev(seq, page, bs_md_page_to_lba(blob->bs, extent),
+			      bs_byte_to_lba(blob->bs, SPDK_BS_PAGE_SIZE),
+			      blob_persist_extent_page_cpl, page);
+}
+
+static void
+blob_insert_cluster_msg(void *arg)
+{
+	struct spdk_blob_insert_cluster_ctx *ctx = arg;
+	uint32_t *extent_page;
+
+	ctx->rc = blob_insert_cluster(ctx->blob, ctx->cluster_num, ctx->cluster);
+	if (ctx->rc != 0) {
+		spdk_thread_send_msg(ctx->thread, blob_insert_cluster_msg_cpl, ctx);
+		return;
+	}
+
+	if (ctx->blob->use_extent_table == false) {
+		/* Extent table is not used, proceed with sync of md that will only use extents_rle. */
+		ctx->blob->state = SPDK_BLOB_STATE_DIRTY;
+		blob_sync_md(ctx->blob, blob_insert_cluster_msg_cb, ctx);
+		return;
+	}
+
+	extent_page = bs_cluster_to_extent_page(ctx->blob, ctx->cluster_num);
+	if (*extent_page == 0) {
+		/* Extent page requires allocation.
+		 * It was already claimed in the used_md_pages map and placed in ctx.
+		 * Blob persist will take care of writing out new extent page on disk. */
+		assert(ctx->extent_page != 0);
+		assert(spdk_bit_array_get(ctx->blob->bs->used_md_pages, ctx->extent_page) == true);
+		*extent_page = ctx->extent_page;
+		ctx->blob->state = SPDK_BLOB_STATE_DIRTY;
+		blob_sync_md(ctx->blob, blob_insert_cluster_msg_cb, ctx);
+	} else {
+		/* It is possible for original thread to allocate extent page for
+		 * different cluster in the same extent page. In such case proceed with
+		 * updating the existing extent page, but release the additional one. */
+		if (ctx->extent_page != 0) {
+			assert(spdk_bit_array_get(ctx->blob->bs->used_md_pages, ctx->extent_page) == true);
+			bs_release_md_page(ctx->blob->bs, ctx->extent_page);
+			ctx->extent_page = 0;
+		}
+		/* Extent page already allocated.
+		 * Every cluster allocation, requires just an update of single extent page. */
+		blob_insert_extent(ctx->blob, *extent_page, ctx->cluster_num,
+				   blob_insert_cluster_msg_cb, ctx);
+	}
+}
+
+static void
+blob_insert_cluster_on_md_thread(struct spdk_blob *blob, uint32_t cluster_num,
+				 uint64_t cluster, uint32_t extent_page, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_blob_insert_cluster_ctx *ctx;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (ctx == NULL) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	ctx->thread = spdk_get_thread();
+	ctx->blob = blob;
+	ctx->cluster_num = cluster_num;
+	ctx->cluster = cluster;
+	ctx->extent_page = extent_page;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+
+	spdk_thread_send_msg(blob->bs->md_thread, blob_insert_cluster_msg, ctx);
+}
+
+/* START spdk_blob_close */
+
+static void
+blob_close_cpl(spdk_bs_sequence_t *seq, void *cb_arg, int bserrno)
+{
+	struct spdk_blob *blob = cb_arg;
+
+	if (bserrno == 0) {
+		blob->open_ref--;
+		if (blob->open_ref == 0) {
+			/*
+			 * Blobs with active.num_pages == 0 are deleted blobs.
+			 *  these blobs are removed from the blob_store list
+			 *  when the deletion process starts - so don't try to
+			 *  remove them again.
+			 */
+			if (blob->active.num_pages > 0) {
+				spdk_bit_array_clear(blob->bs->open_blobids, blob->id);
+				TAILQ_REMOVE(&blob->bs->blobs, blob, link);
+			}
+			blob_free(blob);
+		}
+	}
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+void spdk_blob_close(struct spdk_blob *blob, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_cpl	cpl;
+	spdk_bs_sequence_t	*seq;
+
+	blob_verify_md_op(blob);
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Closing blob %lu\n", blob->id);
+
+	if (blob->open_ref == 0) {
+		cb_fn(cb_arg, -EBADF);
+		return;
+	}
+
+	cpl.type = SPDK_BS_CPL_TYPE_BLOB_BASIC;
+	cpl.u.blob_basic.cb_fn = cb_fn;
+	cpl.u.blob_basic.cb_arg = cb_arg;
+
+	seq = bs_sequence_start(blob->bs->md_channel, &cpl);
+	if (!seq) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	/* Sync metadata */
+	blob_persist(seq, blob, blob_close_cpl, blob);
+}
+
+/* END spdk_blob_close */
+
+struct spdk_io_channel *spdk_bs_alloc_io_channel(struct spdk_blob_store *bs)
+{
+	return spdk_get_io_channel(bs);
+}
+
+void spdk_bs_free_io_channel(struct spdk_io_channel *channel)
+{
+	spdk_put_io_channel(channel);
+}
+
+void spdk_blob_io_unmap(struct spdk_blob *blob, struct spdk_io_channel *channel,
+			uint64_t offset, uint64_t length, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_op(blob, channel, NULL, offset, length, cb_fn, cb_arg,
+			       SPDK_BLOB_UNMAP);
+}
+
+void spdk_blob_io_write_zeroes(struct spdk_blob *blob, struct spdk_io_channel *channel,
+			       uint64_t offset, uint64_t length, spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_op(blob, channel, NULL, offset, length, cb_fn, cb_arg,
+			       SPDK_BLOB_WRITE_ZEROES);
+}
+
+void spdk_blob_io_write(struct spdk_blob *blob, struct spdk_io_channel *channel,
+			void *payload, uint64_t offset, uint64_t length,
+			spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_op(blob, channel, payload, offset, length, cb_fn, cb_arg,
+			       SPDK_BLOB_WRITE);
+}
+
+void spdk_blob_io_read(struct spdk_blob *blob, struct spdk_io_channel *channel,
+		       void *payload, uint64_t offset, uint64_t length,
+		       spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_op(blob, channel, payload, offset, length, cb_fn, cb_arg,
+			       SPDK_BLOB_READ);
+}
+
+void spdk_blob_io_writev(struct spdk_blob *blob, struct spdk_io_channel *channel,
+			 struct iovec *iov, int iovcnt, uint64_t offset, uint64_t length,
+			 spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_rw_iov(blob, channel, iov, iovcnt, offset, length, cb_fn, cb_arg, false);
+}
+
+void spdk_blob_io_readv(struct spdk_blob *blob, struct spdk_io_channel *channel,
+			struct iovec *iov, int iovcnt, uint64_t offset, uint64_t length,
+			spdk_blob_op_complete cb_fn, void *cb_arg)
+{
+	blob_request_submit_rw_iov(blob, channel, iov, iovcnt, offset, length, cb_fn, cb_arg, true);
+}
+
+struct spdk_bs_iter_ctx {
+	int64_t page_num;
+	struct spdk_blob_store *bs;
+
+	spdk_blob_op_with_handle_complete cb_fn;
+	void *cb_arg;
+};
+
+static void
+bs_iter_cpl(void *cb_arg, struct spdk_blob *_blob, int bserrno)
+{
+	struct spdk_bs_iter_ctx *ctx = cb_arg;
+	struct spdk_blob_store *bs = ctx->bs;
+	spdk_blob_id id;
+
+	if (bserrno == 0) {
+		ctx->cb_fn(ctx->cb_arg, _blob, bserrno);
+		free(ctx);
+		return;
+	}
+
+	ctx->page_num++;
+	ctx->page_num = spdk_bit_array_find_first_set(bs->used_blobids, ctx->page_num);
+	if (ctx->page_num >= spdk_bit_array_capacity(bs->used_blobids)) {
+		ctx->cb_fn(ctx->cb_arg, NULL, -ENOENT);
+		free(ctx);
+		return;
+	}
+
+	id = bs_page_to_blobid(ctx->page_num);
+
+	spdk_bs_open_blob(bs, id, bs_iter_cpl, ctx);
+}
+
+void
+spdk_bs_iter_first(struct spdk_blob_store *bs,
+		   spdk_blob_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_iter_ctx *ctx;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	ctx->page_num = -1;
+	ctx->bs = bs;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+
+	bs_iter_cpl(ctx, NULL, -1);
+}
+
+static void
+bs_iter_close_cpl(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_iter_ctx *ctx = cb_arg;
+
+	bs_iter_cpl(ctx, NULL, -1);
+}
+
+void
+spdk_bs_iter_next(struct spdk_blob_store *bs, struct spdk_blob *blob,
+		  spdk_blob_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_bs_iter_ctx *ctx;
+
+	assert(blob != NULL);
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	ctx->page_num = bs_blobid_to_page(blob->id);
+	ctx->bs = bs;
+	ctx->cb_fn = cb_fn;
+	ctx->cb_arg = cb_arg;
+
+	/* Close the existing blob */
+	spdk_blob_close(blob, bs_iter_close_cpl, ctx);
+}
+
+static int
+blob_set_xattr(struct spdk_blob *blob, const char *name, const void *value,
+	       uint16_t value_len, bool internal)
+{
+	struct spdk_xattr_tailq *xattrs;
+	struct spdk_xattr	*xattr;
+	size_t			desc_size;
+	void			*tmp;
+
+	blob_verify_md_op(blob);
+
+	if (blob->md_ro) {
+		return -EPERM;
+	}
+
+	desc_size = sizeof(struct spdk_blob_md_descriptor_xattr) + strlen(name) + value_len;
+	if (desc_size > SPDK_BS_MAX_DESC_SIZE) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOB, "Xattr '%s' of size %ld does not fix into single page %ld\n", name,
+			      desc_size, SPDK_BS_MAX_DESC_SIZE);
+		return -ENOMEM;
+	}
+
+	if (internal) {
+		xattrs = &blob->xattrs_internal;
+		blob->invalid_flags |= SPDK_BLOB_INTERNAL_XATTR;
+	} else {
+		xattrs = &blob->xattrs;
+	}
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		if (!strcmp(name, xattr->name)) {
+			tmp = malloc(value_len);
+			if (!tmp) {
+				return -ENOMEM;
+			}
+
+			free(xattr->value);
+			xattr->value_len = value_len;
+			xattr->value = tmp;
+			memcpy(xattr->value, value, value_len);
+
+			blob->state = SPDK_BLOB_STATE_DIRTY;
+
+			return 0;
+		}
+	}
+
+	xattr = calloc(1, sizeof(*xattr));
+	if (!xattr) {
+		return -ENOMEM;
+	}
+
+	xattr->name = strdup(name);
+	if (!xattr->name) {
+		free(xattr);
+		return -ENOMEM;
+	}
+
+	xattr->value_len = value_len;
+	xattr->value = malloc(value_len);
+	if (!xattr->value) {
+		free(xattr->name);
+		free(xattr);
+		return -ENOMEM;
+	}
+	memcpy(xattr->value, value, value_len);
+	TAILQ_INSERT_TAIL(xattrs, xattr, link);
+
+	blob->state = SPDK_BLOB_STATE_DIRTY;
+
+	return 0;
+}
+
+int
+spdk_blob_set_xattr(struct spdk_blob *blob, const char *name, const void *value,
+		    uint16_t value_len)
+{
+	return blob_set_xattr(blob, name, value, value_len, false);
+}
+
+static int
+blob_remove_xattr(struct spdk_blob *blob, const char *name, bool internal)
+{
+	struct spdk_xattr_tailq *xattrs;
+	struct spdk_xattr	*xattr;
+
+	blob_verify_md_op(blob);
+
+	if (blob->md_ro) {
+		return -EPERM;
+	}
+	xattrs = internal ? &blob->xattrs_internal : &blob->xattrs;
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		if (!strcmp(name, xattr->name)) {
+			TAILQ_REMOVE(xattrs, xattr, link);
+			free(xattr->value);
+			free(xattr->name);
+			free(xattr);
+
+			if (internal && TAILQ_EMPTY(&blob->xattrs_internal)) {
+				blob->invalid_flags &= ~SPDK_BLOB_INTERNAL_XATTR;
+			}
+			blob->state = SPDK_BLOB_STATE_DIRTY;
+
+			return 0;
+		}
+	}
+
+	return -ENOENT;
+}
+
+int
+spdk_blob_remove_xattr(struct spdk_blob *blob, const char *name)
+{
+	return blob_remove_xattr(blob, name, false);
+}
+
+static int
+blob_get_xattr_value(struct spdk_blob *blob, const char *name,
+		     const void **value, size_t *value_len, bool internal)
+{
+	struct spdk_xattr	*xattr;
+	struct spdk_xattr_tailq *xattrs;
+
+	xattrs = internal ? &blob->xattrs_internal : &blob->xattrs;
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		if (!strcmp(name, xattr->name)) {
+			*value = xattr->value;
+			*value_len = xattr->value_len;
+			return 0;
+		}
+	}
+	return -ENOENT;
+}
+
+int
+spdk_blob_get_xattr_value(struct spdk_blob *blob, const char *name,
+			  const void **value, size_t *value_len)
+{
+	blob_verify_md_op(blob);
+
+	return blob_get_xattr_value(blob, name, value, value_len, false);
+}
+
+struct spdk_xattr_names {
+	uint32_t	count;
+	const char	*names[0];
+};
+
+static int
+blob_get_xattr_names(struct spdk_xattr_tailq *xattrs, struct spdk_xattr_names **names)
+{
+	struct spdk_xattr	*xattr;
+	int			count = 0;
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		count++;
+	}
+
+	*names = calloc(1, sizeof(struct spdk_xattr_names) + count * sizeof(char *));
+	if (*names == NULL) {
+		return -ENOMEM;
+	}
+
+	TAILQ_FOREACH(xattr, xattrs, link) {
+		(*names)->names[(*names)->count++] = xattr->name;
+	}
+
+	return 0;
+}
+
+int
+spdk_blob_get_xattr_names(struct spdk_blob *blob, struct spdk_xattr_names **names)
+{
+	blob_verify_md_op(blob);
+
+	return blob_get_xattr_names(&blob->xattrs, names);
+}
+
+uint32_t
+spdk_xattr_names_get_count(struct spdk_xattr_names *names)
+{
+	assert(names != NULL);
+
+	return names->count;
+}
+
+const char *
+spdk_xattr_names_get_name(struct spdk_xattr_names *names, uint32_t index)
+{
+	if (index >= names->count) {
+		return NULL;
+	}
+
+	return names->names[index];
+}
+
+void
+spdk_xattr_names_free(struct spdk_xattr_names *names)
+{
+	free(names);
+}
+
+struct spdk_bs_type
+spdk_bs_get_bstype(struct spdk_blob_store *bs)
+{
+	return bs->bstype;
+}
+
+void
+spdk_bs_set_bstype(struct spdk_blob_store *bs, struct spdk_bs_type bstype)
+{
+	memcpy(&bs->bstype, &bstype, sizeof(bstype));
+}
+
+bool
+spdk_blob_is_read_only(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+	return (blob->data_ro || blob->md_ro);
+}
+
+bool
+spdk_blob_is_snapshot(struct spdk_blob *blob)
+{
+	struct spdk_blob_list *snapshot_entry;
+
+	assert(blob != NULL);
+
+	snapshot_entry = bs_get_snapshot_entry(blob->bs, blob->id);
+	if (snapshot_entry == NULL) {
+		return false;
+	}
+
+	return true;
+}
+
+bool
+spdk_blob_is_clone(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+
+	if (blob->parent_id != SPDK_BLOBID_INVALID) {
+		assert(spdk_blob_is_thin_provisioned(blob));
+		return true;
+	}
+
+	return false;
+}
+
+bool
+spdk_blob_is_thin_provisioned(struct spdk_blob *blob)
+{
+	assert(blob != NULL);
+	return !!(blob->invalid_flags & SPDK_BLOB_THIN_PROV);
+}
+
+static void
+blob_update_clear_method(struct spdk_blob *blob)
+{
+	enum blob_clear_method stored_cm;
+
+	assert(blob != NULL);
+
+	/* If BLOB_CLEAR_WITH_DEFAULT was passed in, use the setting stored
+	 * in metadata previously.  If something other than the default was
+	 * specified, ignore stored value and used what was passed in.
+	 */
+	stored_cm = ((blob->md_ro_flags & SPDK_BLOB_CLEAR_METHOD) >> SPDK_BLOB_CLEAR_METHOD_SHIFT);
+
+	if (blob->clear_method == BLOB_CLEAR_WITH_DEFAULT) {
+		blob->clear_method = stored_cm;
+	} else if (blob->clear_method != stored_cm) {
+		SPDK_WARNLOG("Using passed in clear method 0x%x instead of stored value of 0x%x\n",
+			     blob->clear_method, stored_cm);
+	}
+}
+
+spdk_blob_id
+spdk_blob_get_parent_snapshot(struct spdk_blob_store *bs, spdk_blob_id blob_id)
+{
+	struct spdk_blob_list *snapshot_entry = NULL;
+	struct spdk_blob_list *clone_entry = NULL;
+
+	TAILQ_FOREACH(snapshot_entry, &bs->snapshots, link) {
+		TAILQ_FOREACH(clone_entry, &snapshot_entry->clones, link) {
+			if (clone_entry->id == blob_id) {
+				return snapshot_entry->id;
+			}
+		}
+	}
+
+	return SPDK_BLOBID_INVALID;
+}
+
+int
+spdk_blob_get_clones(struct spdk_blob_store *bs, spdk_blob_id blobid, spdk_blob_id *ids,
+		     size_t *count)
+{
+	struct spdk_blob_list *snapshot_entry, *clone_entry;
+	size_t n;
+
+	snapshot_entry = bs_get_snapshot_entry(bs, blobid);
+	if (snapshot_entry == NULL) {
+		*count = 0;
+		return 0;
+	}
+
+	if (ids == NULL || *count < snapshot_entry->clone_count) {
+		*count = snapshot_entry->clone_count;
+		return -ENOMEM;
+	}
+	*count = snapshot_entry->clone_count;
+
+	n = 0;
+	TAILQ_FOREACH(clone_entry, &snapshot_entry->clones, link) {
+		ids[n++] = clone_entry->id;
+	}
+
+	return 0;
+}
+
+SPDK_LOG_REGISTER_COMPONENT("blob", SPDK_LOG_BLOB)
diff --git a/src/spdk/lib/blob/blobstore.h b/src/spdk/lib/blob/blobstore.h
new file mode 100644
index 000000000..5e93bd6ad
--- /dev/null
+++ b/src/spdk/lib/blob/blobstore.h
@@ -0,0 +1,702 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef SPDK_BLOBSTORE_H
+#define SPDK_BLOBSTORE_H
+
+#include "spdk/assert.h"
+#include "spdk/blob.h"
+#include "spdk/queue.h"
+#include "spdk/util.h"
+
+#include "request.h"
+
+/* In Memory Data Structures
+ *
+ * The following data structures exist only in memory.
+ */
+
+#define SPDK_BLOB_OPTS_CLUSTER_SZ (1024 * 1024)
+#define SPDK_BLOB_OPTS_NUM_MD_PAGES UINT32_MAX
+#define SPDK_BLOB_OPTS_MAX_MD_OPS 32
+#define SPDK_BLOB_OPTS_DEFAULT_CHANNEL_OPS 512
+#define SPDK_BLOB_BLOBID_HIGH_BIT (1ULL << 32)
+
+struct spdk_xattr {
+	uint32_t	index;
+	uint16_t	value_len;
+	char		*name;
+	void		*value;
+	TAILQ_ENTRY(spdk_xattr)	link;
+};
+
+/* The mutable part of the blob data that is sync'd to
+ * disk. The data in here is both mutable and persistent.
+ */
+struct spdk_blob_mut_data {
+	/* Number of data clusters in the blob */
+	uint64_t	num_clusters;
+
+	/* Array LBAs that are the beginning of a cluster, in
+	 * the order they appear in the blob.
+	 */
+	uint64_t	*clusters;
+
+	/* The size of the clusters array. This is greater than or
+	 * equal to 'num_clusters'.
+	 */
+	size_t		cluster_array_size;
+
+	/* Number of extent pages */
+	uint64_t	num_extent_pages;
+
+	/* Array of page offsets into the metadata region,
+	 * containing extents. Can contain entries for not yet
+	 * allocated pages. */
+	uint32_t	*extent_pages;
+
+	/* The size of the extent page array. This is greater than or
+	 * equal to 'num_extent_pages'. */
+	size_t		extent_pages_array_size;
+
+	/* Number of metadata pages */
+	uint32_t	num_pages;
+
+	/* Array of page offsets into the metadata region, in
+	 * the order of the metadata page sequence.
+	 */
+	uint32_t	*pages;
+};
+
+enum spdk_blob_state {
+	/* The blob in-memory version does not match the on-disk
+	 * version.
+	 */
+	SPDK_BLOB_STATE_DIRTY,
+
+	/* The blob in memory version of the blob matches the on disk
+	 * version.
+	 */
+	SPDK_BLOB_STATE_CLEAN,
+
+	/* The in-memory state being synchronized with the on-disk
+	 * blob state. */
+	SPDK_BLOB_STATE_LOADING,
+};
+
+TAILQ_HEAD(spdk_xattr_tailq, spdk_xattr);
+
+struct spdk_blob_list {
+	spdk_blob_id id;
+	size_t clone_count;
+	TAILQ_HEAD(, spdk_blob_list) clones;
+	TAILQ_ENTRY(spdk_blob_list) link;
+};
+
+struct spdk_blob {
+	struct spdk_blob_store *bs;
+
+	uint32_t	open_ref;
+
+	spdk_blob_id	id;
+	spdk_blob_id	parent_id;
+
+	enum spdk_blob_state		state;
+
+	/* Two copies of the mutable data. One is a version
+	 * that matches the last known data on disk (clean).
+	 * The other (active) is the current data. Syncing
+	 * a blob makes the clean match the active.
+	 */
+	struct spdk_blob_mut_data	clean;
+	struct spdk_blob_mut_data	active;
+
+	bool		invalid;
+	bool		data_ro;
+	bool		md_ro;
+
+	uint64_t	invalid_flags;
+	uint64_t	data_ro_flags;
+	uint64_t	md_ro_flags;
+
+	struct spdk_bs_dev *back_bs_dev;
+
+	/* TODO: The xattrs are mutable, but we don't want to be
+	 * copying them unnecessarily. Figure this out.
+	 */
+	struct spdk_xattr_tailq xattrs;
+	struct spdk_xattr_tailq xattrs_internal;
+
+	TAILQ_ENTRY(spdk_blob) link;
+
+	uint32_t frozen_refcnt;
+	bool locked_operation_in_progress;
+	enum blob_clear_method clear_method;
+	bool extent_rle_found;
+	bool extent_table_found;
+	bool use_extent_table;
+
+	/* A list of pending metadata pending_persists */
+	TAILQ_HEAD(, spdk_blob_persist_ctx) pending_persists;
+
+	/* Number of data clusters retrived from extent table,
+	 * that many have to be read from extent pages. */
+	uint64_t	remaining_clusters_in_et;
+};
+
+struct spdk_blob_store {
+	uint64_t			md_start; /* Offset from beginning of disk, in pages */
+	uint32_t			md_len; /* Count, in pages */
+
+	struct spdk_io_channel		*md_channel;
+	uint32_t			max_channel_ops;
+
+	struct spdk_thread		*md_thread;
+
+	struct spdk_bs_dev		*dev;
+
+	struct spdk_bit_array		*used_md_pages;
+	struct spdk_bit_array		*used_clusters;
+	struct spdk_bit_array		*used_blobids;
+	struct spdk_bit_array		*open_blobids;
+
+	pthread_mutex_t			used_clusters_mutex;
+
+	uint32_t			cluster_sz;
+	uint64_t			total_clusters;
+	uint64_t			total_data_clusters;
+	uint64_t			num_free_clusters;
+	uint64_t			pages_per_cluster;
+	uint8_t				pages_per_cluster_shift;
+	uint32_t			io_unit_size;
+
+	spdk_blob_id			super_blob;
+	struct spdk_bs_type		bstype;
+
+	struct spdk_bs_cpl		unload_cpl;
+	int				unload_err;
+
+	TAILQ_HEAD(, spdk_blob)		blobs;
+	TAILQ_HEAD(, spdk_blob_list)	snapshots;
+
+	bool                            clean;
+};
+
+struct spdk_bs_channel {
+	struct spdk_bs_request_set	*req_mem;
+	TAILQ_HEAD(, spdk_bs_request_set) reqs;
+
+	struct spdk_blob_store		*bs;
+
+	struct spdk_bs_dev		*dev;
+	struct spdk_io_channel		*dev_channel;
+
+	TAILQ_HEAD(, spdk_bs_request_set) need_cluster_alloc;
+	TAILQ_HEAD(, spdk_bs_request_set) queued_io;
+};
+
+/** operation type */
+enum spdk_blob_op_type {
+	SPDK_BLOB_WRITE,
+	SPDK_BLOB_READ,
+	SPDK_BLOB_UNMAP,
+	SPDK_BLOB_WRITE_ZEROES,
+	SPDK_BLOB_WRITEV,
+	SPDK_BLOB_READV,
+};
+
+/* back bs_dev */
+
+#define BLOB_SNAPSHOT "SNAP"
+#define SNAPSHOT_IN_PROGRESS "SNAPTMP"
+#define SNAPSHOT_PENDING_REMOVAL "SNAPRM"
+
+struct spdk_blob_bs_dev {
+	struct spdk_bs_dev bs_dev;
+	struct spdk_blob *blob;
+};
+
+/* On-Disk Data Structures
+ *
+ * The following data structures exist on disk.
+ */
+#define SPDK_BS_INITIAL_VERSION 1
+#define SPDK_BS_VERSION 3 /* current version */
+
+#pragma pack(push, 1)
+
+#define SPDK_MD_MASK_TYPE_USED_PAGES 0
+#define SPDK_MD_MASK_TYPE_USED_CLUSTERS 1
+#define SPDK_MD_MASK_TYPE_USED_BLOBIDS 2
+
+struct spdk_bs_md_mask {
+	uint8_t		type;
+	uint32_t	length; /* In bits */
+	uint8_t		mask[0];
+};
+
+#define SPDK_MD_DESCRIPTOR_TYPE_PADDING 0
+#define SPDK_MD_DESCRIPTOR_TYPE_XATTR 2
+#define SPDK_MD_DESCRIPTOR_TYPE_FLAGS 3
+#define SPDK_MD_DESCRIPTOR_TYPE_XATTR_INTERNAL 4
+
+/* Following descriptors define cluster layout in a blob.
+ * EXTENT_RLE cannot be present in blobs metadata,
+ * at the same time as EXTENT_TABLE and EXTENT_PAGE descriptors. */
+
+/* EXTENT_RLE descriptor holds an array of LBA that points to
+ * beginning of allocated clusters. The array is run-length encoded,
+ * with 0's being unallocated clusters. It is part of serialized
+ * metadata chain for a blob. */
+#define SPDK_MD_DESCRIPTOR_TYPE_EXTENT_RLE 1
+/* EXTENT_TABLE descriptor holds array of md page offsets that
+ * point to pages with EXTENT_PAGE descriptor. The 0's in the array
+ * are run-length encoded, non-zero values are unallocated pages.
+ * It is part of serialized metadata chain for a blob. */
+#define SPDK_MD_DESCRIPTOR_TYPE_EXTENT_TABLE 5
+/* EXTENT_PAGE descriptor holds an array of LBAs that point to
+ * beginning of allocated clusters. The array is run-length encoded,
+ * with 0's being unallocated clusters. It is NOT part of
+ * serialized metadata chain for a blob. */
+#define SPDK_MD_DESCRIPTOR_TYPE_EXTENT_PAGE 6
+
+struct spdk_blob_md_descriptor_xattr {
+	uint8_t		type;
+	uint32_t	length;
+
+	uint16_t	name_length;
+	uint16_t	value_length;
+
+	char		name[0];
+	/* String name immediately followed by string value. */
+};
+
+struct spdk_blob_md_descriptor_extent_rle {
+	uint8_t		type;
+	uint32_t	length;
+
+	struct {
+		uint32_t        cluster_idx;
+		uint32_t        length; /* In units of clusters */
+	} extents[0];
+};
+
+struct spdk_blob_md_descriptor_extent_table {
+	uint8_t		type;
+	uint32_t	length;
+
+	/* Number of data clusters in the blob */
+	uint64_t	num_clusters;
+
+	struct {
+		uint32_t	page_idx;
+		uint32_t	num_pages; /* In units of pages */
+	} extent_page[0];
+};
+
+struct spdk_blob_md_descriptor_extent_page {
+	uint8_t		type;
+	uint32_t	length;
+
+	/* First cluster index in this extent page */
+	uint32_t	start_cluster_idx;
+
+	uint32_t        cluster_idx[0];
+};
+
+#define SPDK_BLOB_THIN_PROV (1ULL << 0)
+#define SPDK_BLOB_INTERNAL_XATTR (1ULL << 1)
+#define SPDK_BLOB_EXTENT_TABLE (1ULL << 2)
+#define SPDK_BLOB_INVALID_FLAGS_MASK	(SPDK_BLOB_THIN_PROV | SPDK_BLOB_INTERNAL_XATTR | SPDK_BLOB_EXTENT_TABLE)
+
+#define SPDK_BLOB_READ_ONLY (1ULL << 0)
+#define SPDK_BLOB_DATA_RO_FLAGS_MASK	SPDK_BLOB_READ_ONLY
+
+#define SPDK_BLOB_CLEAR_METHOD_SHIFT 0
+#define SPDK_BLOB_CLEAR_METHOD (3ULL << SPDK_BLOB_CLEAR_METHOD_SHIFT)
+#define SPDK_BLOB_MD_RO_FLAGS_MASK	SPDK_BLOB_CLEAR_METHOD
+
+struct spdk_blob_md_descriptor_flags {
+	uint8_t		type;
+	uint32_t	length;
+
+	/*
+	 * If a flag in invalid_flags is set that the application is not aware of,
+	 *  it will not allow the blob to be opened.
+	 */
+	uint64_t	invalid_flags;
+
+	/*
+	 * If a flag in data_ro_flags is set that the application is not aware of,
+	 *  allow the blob to be opened in data_read_only and md_read_only mode.
+	 */
+	uint64_t	data_ro_flags;
+
+	/*
+	 * If a flag in md_ro_flags is set the the application is not aware of,
+	 *  allow the blob to be opened in md_read_only mode.
+	 */
+	uint64_t	md_ro_flags;
+};
+
+struct spdk_blob_md_descriptor {
+	uint8_t		type;
+	uint32_t	length;
+};
+
+#define SPDK_INVALID_MD_PAGE UINT32_MAX
+
+struct spdk_blob_md_page {
+	spdk_blob_id     id;
+
+	uint32_t        sequence_num;
+	uint32_t	reserved0;
+
+	/* Descriptors here */
+	uint8_t		descriptors[4072];
+
+	uint32_t	next;
+	uint32_t	crc;
+};
+#define SPDK_BS_PAGE_SIZE 0x1000
+SPDK_STATIC_ASSERT(SPDK_BS_PAGE_SIZE == sizeof(struct spdk_blob_md_page), "Invalid md page size");
+
+#define SPDK_BS_MAX_DESC_SIZE sizeof(((struct spdk_blob_md_page*)0)->descriptors)
+
+/* Maximum number of extents a single Extent Page can fit.
+ * For an SPDK_BS_PAGE_SIZE of 4K SPDK_EXTENTS_PER_EP would be 512. */
+#define SPDK_EXTENTS_PER_EP_MAX ((SPDK_BS_MAX_DESC_SIZE - sizeof(struct spdk_blob_md_descriptor_extent_page)) / sizeof(uint32_t))
+#define SPDK_EXTENTS_PER_EP (spdk_align64pow2(SPDK_EXTENTS_PER_EP_MAX + 1) >> 1u)
+
+#define SPDK_BS_SUPER_BLOCK_SIG "SPDKBLOB"
+
+struct spdk_bs_super_block {
+	uint8_t		signature[8];
+	uint32_t        version;
+	uint32_t        length;
+	uint32_t	clean; /* If there was a clean shutdown, this is 1. */
+	spdk_blob_id	super_blob;
+
+	uint32_t	cluster_size; /* In bytes */
+
+	uint32_t	used_page_mask_start; /* Offset from beginning of disk, in pages */
+	uint32_t	used_page_mask_len; /* Count, in pages */
+
+	uint32_t	used_cluster_mask_start; /* Offset from beginning of disk, in pages */
+	uint32_t	used_cluster_mask_len; /* Count, in pages */
+
+	uint32_t	md_start; /* Offset from beginning of disk, in pages */
+	uint32_t	md_len; /* Count, in pages */
+
+	struct spdk_bs_type	bstype; /* blobstore type */
+
+	uint32_t	used_blobid_mask_start; /* Offset from beginning of disk, in pages */
+	uint32_t	used_blobid_mask_len; /* Count, in pages */
+
+	uint64_t        size; /* size of blobstore in bytes */
+	uint32_t        io_unit_size; /* Size of io unit in bytes */
+
+	uint8_t         reserved[4000];
+	uint32_t	crc;
+};
+SPDK_STATIC_ASSERT(sizeof(struct spdk_bs_super_block) == 0x1000, "Invalid super block size");
+
+#pragma pack(pop)
+
+struct spdk_bs_dev *bs_create_zeroes_dev(void);
+struct spdk_bs_dev *bs_create_blob_bs_dev(struct spdk_blob *blob);
+
+/* Unit Conversions
+ *
+ * The blobstore works with several different units:
+ * - Byte: Self explanatory
+ * - LBA: The logical blocks on the backing storage device.
+ * - Page: The read/write units of blobs and metadata. This is
+ *         an offset into a blob in units of 4KiB.
+ * - Cluster Index: The disk is broken into a sequential list of
+ *		    clusters. This is the offset from the beginning.
+ *
+ * NOTE: These conversions all act on simple magnitudes, not with any sort
+ *        of knowledge about the blobs themselves. For instance, converting
+ *        a page to an lba with the conversion function below simply converts
+ *        a number of pages to an equivalent number of lbas, but that
+ *        lba certainly isn't the right lba that corresponds to a page offset
+ *        for a particular blob.
+ */
+static inline uint64_t
+bs_byte_to_lba(struct spdk_blob_store *bs, uint64_t length)
+{
+	assert(length % bs->dev->blocklen == 0);
+
+	return length / bs->dev->blocklen;
+}
+
+static inline uint64_t
+bs_dev_byte_to_lba(struct spdk_bs_dev *bs_dev, uint64_t length)
+{
+	assert(length % bs_dev->blocklen == 0);
+
+	return length / bs_dev->blocklen;
+}
+
+static inline uint64_t
+bs_page_to_lba(struct spdk_blob_store *bs, uint64_t page)
+{
+	return page * SPDK_BS_PAGE_SIZE / bs->dev->blocklen;
+}
+
+static inline uint64_t
+bs_md_page_to_lba(struct spdk_blob_store *bs, uint32_t page)
+{
+	assert(page < bs->md_len);
+	return bs_page_to_lba(bs, page + bs->md_start);
+}
+
+static inline uint64_t
+bs_dev_page_to_lba(struct spdk_bs_dev *bs_dev, uint64_t page)
+{
+	return page * SPDK_BS_PAGE_SIZE / bs_dev->blocklen;
+}
+
+static inline uint64_t
+bs_io_unit_per_page(struct spdk_blob_store *bs)
+{
+	return SPDK_BS_PAGE_SIZE / bs->io_unit_size;
+}
+
+static inline uint64_t
+bs_io_unit_to_page(struct spdk_blob_store *bs, uint64_t io_unit)
+{
+	return io_unit / bs_io_unit_per_page(bs);
+}
+
+static inline uint64_t
+bs_cluster_to_page(struct spdk_blob_store *bs, uint32_t cluster)
+{
+	return (uint64_t)cluster * bs->pages_per_cluster;
+}
+
+static inline uint32_t
+bs_page_to_cluster(struct spdk_blob_store *bs, uint64_t page)
+{
+	assert(page % bs->pages_per_cluster == 0);
+
+	return page / bs->pages_per_cluster;
+}
+
+static inline uint64_t
+bs_cluster_to_lba(struct spdk_blob_store *bs, uint32_t cluster)
+{
+	return (uint64_t)cluster * (bs->cluster_sz / bs->dev->blocklen);
+}
+
+static inline uint32_t
+bs_lba_to_cluster(struct spdk_blob_store *bs, uint64_t lba)
+{
+	assert(lba % (bs->cluster_sz / bs->dev->blocklen) == 0);
+
+	return lba / (bs->cluster_sz / bs->dev->blocklen);
+}
+
+static inline uint64_t
+bs_io_unit_to_back_dev_lba(struct spdk_blob *blob, uint64_t io_unit)
+{
+	return io_unit * (blob->bs->io_unit_size / blob->back_bs_dev->blocklen);
+}
+
+static inline uint64_t
+bs_back_dev_lba_to_io_unit(struct spdk_blob *blob, uint64_t lba)
+{
+	return lba * (blob->back_bs_dev->blocklen / blob->bs->io_unit_size);
+}
+
+static inline uint64_t
+bs_cluster_to_extent_table_id(uint64_t cluster_num)
+{
+	return cluster_num / SPDK_EXTENTS_PER_EP;
+}
+
+static inline uint32_t *
+bs_cluster_to_extent_page(struct spdk_blob *blob, uint64_t cluster_num)
+{
+	uint64_t extent_table_id = bs_cluster_to_extent_table_id(cluster_num);
+
+	assert(blob->use_extent_table);
+	assert(extent_table_id < blob->active.extent_pages_array_size);
+
+	return &blob->active.extent_pages[extent_table_id];
+}
+
+/* End basic conversions */
+
+static inline uint64_t
+bs_blobid_to_page(spdk_blob_id id)
+{
+	return id & 0xFFFFFFFF;
+}
+
+/* The blob id is a 64 bit number. The lower 32 bits are the page_idx. The upper
+ * 32 bits are not currently used. Stick a 1 there just to catch bugs where the
+ * code assumes blob id == page_idx.
+ */
+static inline spdk_blob_id
+bs_page_to_blobid(uint64_t page_idx)
+{
+	if (page_idx > UINT32_MAX) {
+		return SPDK_BLOBID_INVALID;
+	}
+	return SPDK_BLOB_BLOBID_HIGH_BIT | page_idx;
+}
+
+/* Given an io unit offset into a blob, look up the LBA for the
+ * start of that io unit.
+ */
+static inline uint64_t
+bs_blob_io_unit_to_lba(struct spdk_blob *blob, uint64_t io_unit)
+{
+	uint64_t	lba;
+	uint64_t	pages_per_cluster;
+	uint8_t		shift;
+	uint64_t	io_units_per_cluster;
+	uint64_t	io_units_per_page;
+	uint64_t	page;
+
+	page = bs_io_unit_to_page(blob->bs, io_unit);
+
+	pages_per_cluster = blob->bs->pages_per_cluster;
+	shift = blob->bs->pages_per_cluster_shift;
+	io_units_per_page = bs_io_unit_per_page(blob->bs);
+
+	assert(page < blob->active.num_clusters * pages_per_cluster);
+
+	if (shift != 0) {
+		io_units_per_cluster = io_units_per_page << shift;
+		lba = blob->active.clusters[page >> shift];
+	} else {
+		io_units_per_cluster = io_units_per_page * pages_per_cluster;
+		lba = blob->active.clusters[page / pages_per_cluster];
+	}
+	lba += io_unit % io_units_per_cluster;
+	return lba;
+}
+
+/* Given an io_unit offset into a blob, look up the number of io_units until the
+ * next cluster boundary.
+ */
+static inline uint32_t
+bs_num_io_units_to_cluster_boundary(struct spdk_blob *blob, uint64_t io_unit)
+{
+	uint64_t	io_units_per_cluster;
+	uint8_t         shift = blob->bs->pages_per_cluster_shift;
+
+	if (shift != 0) {
+		io_units_per_cluster = bs_io_unit_per_page(blob->bs) << shift;
+	} else {
+		io_units_per_cluster = bs_io_unit_per_page(blob->bs) * blob->bs->pages_per_cluster;
+	}
+
+	return io_units_per_cluster - (io_unit % io_units_per_cluster);
+}
+
+/* Given a page offset into a blob, look up the number of pages until the
+ * next cluster boundary.
+ */
+static inline uint32_t
+bs_num_pages_to_cluster_boundary(struct spdk_blob *blob, uint64_t page)
+{
+	uint64_t	pages_per_cluster;
+
+	pages_per_cluster = blob->bs->pages_per_cluster;
+
+	return pages_per_cluster - (page % pages_per_cluster);
+}
+
+/* Given an io_unit offset into a blob, look up the number of pages into blob to beginning of current cluster */
+static inline uint32_t
+bs_io_unit_to_cluster_start(struct spdk_blob *blob, uint64_t io_unit)
+{
+	uint64_t	pages_per_cluster;
+	uint64_t	page;
+
+	pages_per_cluster = blob->bs->pages_per_cluster;
+	page = bs_io_unit_to_page(blob->bs, io_unit);
+
+	return page - (page % pages_per_cluster);
+}
+
+/* Given an io_unit offset into a blob, look up the number of pages into blob to beginning of current cluster */
+static inline uint32_t
+bs_io_unit_to_cluster_number(struct spdk_blob *blob, uint64_t io_unit)
+{
+	uint64_t	pages_per_cluster = blob->bs->pages_per_cluster;
+	uint8_t		shift = blob->bs->pages_per_cluster_shift;
+	uint32_t	page_offset;
+
+	page_offset = io_unit / bs_io_unit_per_page(blob->bs);
+	if (shift != 0) {
+		return page_offset >> shift;
+	} else {
+		return page_offset / pages_per_cluster;
+	}
+}
+
+/* Given an io unit offset into a blob, look up if it is from allocated cluster. */
+static inline bool
+bs_io_unit_is_allocated(struct spdk_blob *blob, uint64_t io_unit)
+{
+	uint64_t	lba;
+	uint64_t	page;
+	uint64_t	pages_per_cluster;
+	uint8_t		shift;
+
+	shift = blob->bs->pages_per_cluster_shift;
+	pages_per_cluster = blob->bs->pages_per_cluster;
+	page = bs_io_unit_to_page(blob->bs, io_unit);
+
+	assert(page < blob->active.num_clusters * pages_per_cluster);
+
+	if (shift != 0) {
+		lba = blob->active.clusters[page >> shift];
+	} else {
+		lba = blob->active.clusters[page / pages_per_cluster];
+	}
+
+	if (lba == 0) {
+		assert(spdk_blob_is_thin_provisioned(blob));
+		return false;
+	} else {
+		return true;
+	}
+}
+
+#endif
diff --git a/src/spdk/lib/blob/request.c b/src/spdk/lib/blob/request.c
new file mode 100644
index 000000000..0975bcf24
--- /dev/null
+++ b/src/spdk/lib/blob/request.c
@@ -0,0 +1,521 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+
+#include "blobstore.h"
+#include "request.h"
+
+#include "spdk/thread.h"
+#include "spdk/queue.h"
+
+#include "spdk_internal/log.h"
+
+void
+bs_call_cpl(struct spdk_bs_cpl *cpl, int bserrno)
+{
+	switch (cpl->type) {
+	case SPDK_BS_CPL_TYPE_BS_BASIC:
+		cpl->u.bs_basic.cb_fn(cpl->u.bs_basic.cb_arg,
+				      bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_BS_HANDLE:
+		cpl->u.bs_handle.cb_fn(cpl->u.bs_handle.cb_arg,
+				       bserrno == 0 ? cpl->u.bs_handle.bs : NULL,
+				       bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_BLOB_BASIC:
+		cpl->u.blob_basic.cb_fn(cpl->u.blob_basic.cb_arg,
+					bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_BLOBID:
+		cpl->u.blobid.cb_fn(cpl->u.blobid.cb_arg,
+				    bserrno == 0 ? cpl->u.blobid.blobid : SPDK_BLOBID_INVALID,
+				    bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_BLOB_HANDLE:
+		cpl->u.blob_handle.cb_fn(cpl->u.blob_handle.cb_arg,
+					 bserrno == 0 ? cpl->u.blob_handle.blob : NULL,
+					 bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_NESTED_SEQUENCE:
+		cpl->u.nested_seq.cb_fn(cpl->u.nested_seq.cb_arg,
+					cpl->u.nested_seq.parent,
+					bserrno);
+		break;
+	case SPDK_BS_CPL_TYPE_NONE:
+		/* this completion's callback is handled elsewhere */
+		break;
+	}
+}
+
+static void
+bs_request_set_complete(struct spdk_bs_request_set *set)
+{
+	struct spdk_bs_cpl cpl = set->cpl;
+	int bserrno = set->bserrno;
+
+	TAILQ_INSERT_TAIL(&set->channel->reqs, set, link);
+
+	bs_call_cpl(&cpl, bserrno);
+}
+
+static void
+bs_sequence_completion(struct spdk_io_channel *channel, void *cb_arg, int bserrno)
+{
+	struct spdk_bs_request_set *set = cb_arg;
+
+	set->bserrno = bserrno;
+	set->u.sequence.cb_fn((spdk_bs_sequence_t *)set, set->u.sequence.cb_arg, bserrno);
+}
+
+spdk_bs_sequence_t *
+bs_sequence_start(struct spdk_io_channel *_channel,
+		  struct spdk_bs_cpl *cpl)
+{
+	struct spdk_bs_channel		*channel;
+	struct spdk_bs_request_set	*set;
+
+	channel = spdk_io_channel_get_ctx(_channel);
+	assert(channel != NULL);
+	set = TAILQ_FIRST(&channel->reqs);
+	if (!set) {
+		return NULL;
+	}
+	TAILQ_REMOVE(&channel->reqs, set, link);
+
+	set->cpl = *cpl;
+	set->bserrno = 0;
+	set->channel = channel;
+
+	set->cb_args.cb_fn = bs_sequence_completion;
+	set->cb_args.cb_arg = set;
+	set->cb_args.channel = channel->dev_channel;
+
+	return (spdk_bs_sequence_t *)set;
+}
+
+void
+bs_sequence_read_bs_dev(spdk_bs_sequence_t *seq, struct spdk_bs_dev *bs_dev,
+			void *payload, uint64_t lba, uint32_t lba_count,
+			spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	bs_dev->read(bs_dev, spdk_io_channel_from_ctx(channel), payload, lba, lba_count, &set->cb_args);
+}
+
+void
+bs_sequence_read_dev(spdk_bs_sequence_t *seq, void *payload,
+		     uint64_t lba, uint32_t lba_count,
+		     spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	channel->dev->read(channel->dev, channel->dev_channel, payload, lba, lba_count, &set->cb_args);
+}
+
+void
+bs_sequence_write_dev(spdk_bs_sequence_t *seq, void *payload,
+		      uint64_t lba, uint32_t lba_count,
+		      spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Writing %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	channel->dev->write(channel->dev, channel->dev_channel, payload, lba, lba_count,
+			    &set->cb_args);
+}
+
+void
+bs_sequence_readv_bs_dev(spdk_bs_sequence_t *seq, struct spdk_bs_dev *bs_dev,
+			 struct iovec *iov, int iovcnt, uint64_t lba, uint32_t lba_count,
+			 spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	bs_dev->readv(bs_dev, spdk_io_channel_from_ctx(channel), iov, iovcnt, lba, lba_count,
+		      &set->cb_args);
+}
+
+void
+bs_sequence_readv_dev(spdk_bs_sequence_t *seq, struct iovec *iov, int iovcnt,
+		      uint64_t lba, uint32_t lba_count, spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+	channel->dev->readv(channel->dev, channel->dev_channel, iov, iovcnt, lba, lba_count,
+			    &set->cb_args);
+}
+
+void
+bs_sequence_writev_dev(spdk_bs_sequence_t *seq, struct iovec *iov, int iovcnt,
+		       uint64_t lba, uint32_t lba_count,
+		       spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Writing %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	channel->dev->writev(channel->dev, channel->dev_channel, iov, iovcnt, lba, lba_count,
+			     &set->cb_args);
+}
+
+void
+bs_sequence_write_zeroes_dev(spdk_bs_sequence_t *seq,
+			     uint64_t lba, uint32_t lba_count,
+			     spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set      *set = (struct spdk_bs_request_set *)seq;
+	struct spdk_bs_channel       *channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "writing zeroes to %" PRIu32 " blocks at LBA %" PRIu64 "\n",
+		      lba_count, lba);
+
+	set->u.sequence.cb_fn = cb_fn;
+	set->u.sequence.cb_arg = cb_arg;
+
+	channel->dev->write_zeroes(channel->dev, channel->dev_channel, lba, lba_count,
+				   &set->cb_args);
+}
+
+void
+bs_sequence_finish(spdk_bs_sequence_t *seq, int bserrno)
+{
+	if (bserrno != 0) {
+		seq->bserrno = bserrno;
+	}
+	bs_request_set_complete((struct spdk_bs_request_set *)seq);
+}
+
+void
+bs_user_op_sequence_finish(void *cb_arg, int bserrno)
+{
+	spdk_bs_sequence_t *seq = cb_arg;
+
+	bs_sequence_finish(seq, bserrno);
+}
+
+static void
+bs_batch_completion(struct spdk_io_channel *_channel,
+		    void *cb_arg, int bserrno)
+{
+	struct spdk_bs_request_set	*set = cb_arg;
+
+	set->u.batch.outstanding_ops--;
+	if (bserrno != 0) {
+		set->bserrno = bserrno;
+	}
+
+	if (set->u.batch.outstanding_ops == 0 && set->u.batch.batch_closed) {
+		if (set->u.batch.cb_fn) {
+			set->cb_args.cb_fn = bs_sequence_completion;
+			set->u.batch.cb_fn((spdk_bs_sequence_t *)set, set->u.batch.cb_arg, bserrno);
+		} else {
+			bs_request_set_complete(set);
+		}
+	}
+}
+
+spdk_bs_batch_t *
+bs_batch_open(struct spdk_io_channel *_channel,
+	      struct spdk_bs_cpl *cpl)
+{
+	struct spdk_bs_channel		*channel;
+	struct spdk_bs_request_set	*set;
+
+	channel = spdk_io_channel_get_ctx(_channel);
+	assert(channel != NULL);
+	set = TAILQ_FIRST(&channel->reqs);
+	if (!set) {
+		return NULL;
+	}
+	TAILQ_REMOVE(&channel->reqs, set, link);
+
+	set->cpl = *cpl;
+	set->bserrno = 0;
+	set->channel = channel;
+
+	set->u.batch.cb_fn = NULL;
+	set->u.batch.cb_arg = NULL;
+	set->u.batch.outstanding_ops = 0;
+	set->u.batch.batch_closed = 0;
+
+	set->cb_args.cb_fn = bs_batch_completion;
+	set->cb_args.cb_arg = set;
+	set->cb_args.channel = channel->dev_channel;
+
+	return (spdk_bs_batch_t *)set;
+}
+
+void
+bs_batch_read_bs_dev(spdk_bs_batch_t *batch, struct spdk_bs_dev *bs_dev,
+		     void *payload, uint64_t lba, uint32_t lba_count)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+	struct spdk_bs_channel		*channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.batch.outstanding_ops++;
+	bs_dev->read(bs_dev, spdk_io_channel_from_ctx(channel), payload, lba, lba_count, &set->cb_args);
+}
+
+void
+bs_batch_read_dev(spdk_bs_batch_t *batch, void *payload,
+		  uint64_t lba, uint32_t lba_count)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+	struct spdk_bs_channel		*channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Reading %" PRIu32 " blocks from LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.batch.outstanding_ops++;
+	channel->dev->read(channel->dev, channel->dev_channel, payload, lba, lba_count, &set->cb_args);
+}
+
+void
+bs_batch_write_dev(spdk_bs_batch_t *batch, void *payload,
+		   uint64_t lba, uint32_t lba_count)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+	struct spdk_bs_channel		*channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Writing %" PRIu32 " blocks to LBA %" PRIu64 "\n", lba_count, lba);
+
+	set->u.batch.outstanding_ops++;
+	channel->dev->write(channel->dev, channel->dev_channel, payload, lba, lba_count,
+			    &set->cb_args);
+}
+
+void
+bs_batch_unmap_dev(spdk_bs_batch_t *batch,
+		   uint64_t lba, uint32_t lba_count)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+	struct spdk_bs_channel		*channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Unmapping %" PRIu32 " blocks at LBA %" PRIu64 "\n", lba_count,
+		      lba);
+
+	set->u.batch.outstanding_ops++;
+	channel->dev->unmap(channel->dev, channel->dev_channel, lba, lba_count,
+			    &set->cb_args);
+}
+
+void
+bs_batch_write_zeroes_dev(spdk_bs_batch_t *batch,
+			  uint64_t lba, uint32_t lba_count)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+	struct spdk_bs_channel		*channel = set->channel;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOB_RW, "Zeroing %" PRIu32 " blocks at LBA %" PRIu64 "\n", lba_count, lba);
+
+	set->u.batch.outstanding_ops++;
+	channel->dev->write_zeroes(channel->dev, channel->dev_channel, lba, lba_count,
+				   &set->cb_args);
+}
+
+void
+bs_batch_close(spdk_bs_batch_t *batch)
+{
+	struct spdk_bs_request_set	*set = (struct spdk_bs_request_set *)batch;
+
+	set->u.batch.batch_closed = 1;
+
+	if (set->u.batch.outstanding_ops == 0) {
+		if (set->u.batch.cb_fn) {
+			set->cb_args.cb_fn = bs_sequence_completion;
+			set->u.batch.cb_fn((spdk_bs_sequence_t *)set, set->u.batch.cb_arg, set->bserrno);
+		} else {
+			bs_request_set_complete(set);
+		}
+	}
+}
+
+spdk_bs_batch_t *
+bs_sequence_to_batch(spdk_bs_sequence_t *seq, spdk_bs_sequence_cpl cb_fn, void *cb_arg)
+{
+	struct spdk_bs_request_set *set = (struct spdk_bs_request_set *)seq;
+
+	set->u.batch.cb_fn = cb_fn;
+	set->u.batch.cb_arg = cb_arg;
+	set->u.batch.outstanding_ops = 0;
+	set->u.batch.batch_closed = 0;
+
+	set->cb_args.cb_fn = bs_batch_completion;
+
+	return set;
+}
+
+spdk_bs_user_op_t *
+bs_user_op_alloc(struct spdk_io_channel *_channel, struct spdk_bs_cpl *cpl,
+		 enum spdk_blob_op_type op_type, struct spdk_blob *blob,
+		 void *payload, int iovcnt, uint64_t offset, uint64_t length)
+{
+	struct spdk_bs_channel		*channel;
+	struct spdk_bs_request_set	*set;
+	struct spdk_bs_user_op_args	*args;
+
+	channel = spdk_io_channel_get_ctx(_channel);
+	assert(channel != NULL);
+	set = TAILQ_FIRST(&channel->reqs);
+	if (!set) {
+		return NULL;
+	}
+	TAILQ_REMOVE(&channel->reqs, set, link);
+
+	set->cpl = *cpl;
+	set->channel = channel;
+
+	args = &set->u.user_op;
+
+	args->type = op_type;
+	args->iovcnt = iovcnt;
+	args->blob = blob;
+	args->offset = offset;
+	args->length = length;
+	args->payload = payload;
+
+	return (spdk_bs_user_op_t *)set;
+}
+
+void
+bs_user_op_execute(spdk_bs_user_op_t *op)
+{
+	struct spdk_bs_request_set	*set;
+	struct spdk_bs_user_op_args	*args;
+	struct spdk_io_channel		*ch;
+
+	set = (struct spdk_bs_request_set *)op;
+	args = &set->u.user_op;
+	ch = spdk_io_channel_from_ctx(set->channel);
+
+	switch (args->type) {
+	case SPDK_BLOB_READ:
+		spdk_blob_io_read(args->blob, ch, args->payload, args->offset, args->length,
+				  set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	case SPDK_BLOB_WRITE:
+		spdk_blob_io_write(args->blob, ch, args->payload, args->offset, args->length,
+				   set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	case SPDK_BLOB_UNMAP:
+		spdk_blob_io_unmap(args->blob, ch, args->offset, args->length,
+				   set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	case SPDK_BLOB_WRITE_ZEROES:
+		spdk_blob_io_write_zeroes(args->blob, ch, args->offset, args->length,
+					  set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	case SPDK_BLOB_READV:
+		spdk_blob_io_readv(args->blob, ch, args->payload, args->iovcnt,
+				   args->offset, args->length,
+				   set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	case SPDK_BLOB_WRITEV:
+		spdk_blob_io_writev(args->blob, ch, args->payload, args->iovcnt,
+				    args->offset, args->length,
+				    set->cpl.u.blob_basic.cb_fn, set->cpl.u.blob_basic.cb_arg);
+		break;
+	}
+	TAILQ_INSERT_TAIL(&set->channel->reqs, set, link);
+}
+
+void
+bs_user_op_abort(spdk_bs_user_op_t *op)
+{
+	struct spdk_bs_request_set	*set;
+
+	set = (struct spdk_bs_request_set *)op;
+
+	set->cpl.u.blob_basic.cb_fn(set->cpl.u.blob_basic.cb_arg, -EIO);
+	TAILQ_INSERT_TAIL(&set->channel->reqs, set, link);
+}
+
+void
+bs_sequence_to_batch_completion(void *cb_arg, int bserrno)
+{
+	struct spdk_bs_request_set *set = (struct spdk_bs_request_set *)cb_arg;
+
+	set->u.batch.outstanding_ops--;
+
+	if (set->u.batch.outstanding_ops == 0 && set->u.batch.batch_closed) {
+		if (set->cb_args.cb_fn) {
+			set->cb_args.cb_fn(set->cb_args.channel, set->cb_args.cb_arg, bserrno);
+		}
+	}
+}
+
+SPDK_LOG_REGISTER_COMPONENT("blob_rw", SPDK_LOG_BLOB_RW)
diff --git a/src/spdk/lib/blob/request.h b/src/spdk/lib/blob/request.h
new file mode 100644
index 000000000..81dc161db
--- /dev/null
+++ b/src/spdk/lib/blob/request.h
@@ -0,0 +1,217 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef SPDK_BS_REQUEST_H
+#define SPDK_BS_REQUEST_H
+
+#include "spdk/stdinc.h"
+
+#include "spdk/blob.h"
+
+enum spdk_bs_cpl_type {
+	SPDK_BS_CPL_TYPE_NONE,
+	SPDK_BS_CPL_TYPE_BS_BASIC,
+	SPDK_BS_CPL_TYPE_BS_HANDLE,
+	SPDK_BS_CPL_TYPE_BLOB_BASIC,
+	SPDK_BS_CPL_TYPE_BLOBID,
+	SPDK_BS_CPL_TYPE_BLOB_HANDLE,
+	SPDK_BS_CPL_TYPE_NESTED_SEQUENCE,
+};
+
+enum spdk_blob_op_type;
+
+struct spdk_bs_request_set;
+
+/* Use a sequence to submit a set of requests serially */
+typedef struct spdk_bs_request_set spdk_bs_sequence_t;
+
+/* Use a batch to submit a set of requests in parallel */
+typedef struct spdk_bs_request_set spdk_bs_batch_t;
+
+/* Use a user_op to queue a user operation for later execution */
+typedef struct spdk_bs_request_set spdk_bs_user_op_t;
+
+typedef void (*spdk_bs_nested_seq_complete)(void *cb_arg, spdk_bs_sequence_t *parent, int bserrno);
+
+struct spdk_bs_cpl {
+	enum spdk_bs_cpl_type type;
+	union {
+		struct {
+			spdk_bs_op_complete     cb_fn;
+			void                    *cb_arg;
+		} bs_basic;
+
+		struct {
+			spdk_bs_op_with_handle_complete cb_fn;
+			void                            *cb_arg;
+			struct spdk_blob_store          *bs;
+		} bs_handle;
+
+		struct {
+			spdk_blob_op_complete   cb_fn;
+			void                    *cb_arg;
+		} blob_basic;
+
+		struct {
+			spdk_blob_op_with_id_complete   cb_fn;
+			void                            *cb_arg;
+			spdk_blob_id                     blobid;
+		} blobid;
+
+		struct {
+			spdk_blob_op_with_handle_complete       cb_fn;
+			void                                    *cb_arg;
+			struct spdk_blob                        *blob;
+		} blob_handle;
+
+		struct {
+			spdk_bs_nested_seq_complete	cb_fn;
+			void				*cb_arg;
+			spdk_bs_sequence_t		*parent;
+		} nested_seq;
+	} u;
+};
+
+typedef void (*spdk_bs_sequence_cpl)(spdk_bs_sequence_t *sequence,
+				     void *cb_arg, int bserrno);
+
+/* A generic request set. Can be a sequence, batch or a user_op. */
+struct spdk_bs_request_set {
+	struct spdk_bs_cpl      cpl;
+
+	int                     bserrno;
+
+	struct spdk_bs_channel		*channel;
+
+	struct spdk_bs_dev_cb_args	cb_args;
+
+	union {
+		struct {
+			spdk_bs_sequence_cpl    cb_fn;
+			void                    *cb_arg;
+		} sequence;
+
+		struct {
+			uint32_t		outstanding_ops;
+			uint32_t		batch_closed;
+			spdk_bs_sequence_cpl	cb_fn;
+			void			*cb_arg;
+		} batch;
+
+		struct spdk_bs_user_op_args {
+			int			type;
+			int			iovcnt;
+			struct spdk_blob	*blob;
+			uint64_t		offset;
+			uint64_t		length;
+			spdk_blob_op_complete	cb_fn;
+			void			*cb_arg;
+			void			*payload; /* cast to iov for readv/writev */
+		} user_op;
+	} u;
+
+	TAILQ_ENTRY(spdk_bs_request_set) link;
+};
+
+void bs_call_cpl(struct spdk_bs_cpl *cpl, int bserrno);
+
+spdk_bs_sequence_t *bs_sequence_start(struct spdk_io_channel *channel,
+				      struct spdk_bs_cpl *cpl);
+
+void bs_sequence_read_bs_dev(spdk_bs_sequence_t *seq, struct spdk_bs_dev *bs_dev,
+			     void *payload, uint64_t lba, uint32_t lba_count,
+			     spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_read_dev(spdk_bs_sequence_t *seq, void *payload,
+			  uint64_t lba, uint32_t lba_count,
+			  spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_write_dev(spdk_bs_sequence_t *seq, void *payload,
+			   uint64_t lba, uint32_t lba_count,
+			   spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_readv_bs_dev(spdk_bs_batch_t *batch, struct spdk_bs_dev *bs_dev,
+			      struct iovec *iov, int iovcnt, uint64_t lba, uint32_t lba_count,
+			      spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_readv_dev(spdk_bs_batch_t *batch, struct iovec *iov, int iovcnt,
+			   uint64_t lba, uint32_t lba_count,
+			   spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_writev_dev(spdk_bs_batch_t *batch, struct iovec *iov, int iovcnt,
+			    uint64_t lba, uint32_t lba_count,
+			    spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_write_zeroes_dev(spdk_bs_sequence_t *seq,
+				  uint64_t lba, uint32_t lba_count,
+				  spdk_bs_sequence_cpl cb_fn, void *cb_arg);
+
+void bs_sequence_finish(spdk_bs_sequence_t *seq, int bserrno);
+
+void bs_user_op_sequence_finish(void *cb_arg, int bserrno);
+
+spdk_bs_batch_t *bs_batch_open(struct spdk_io_channel *channel,
+			       struct spdk_bs_cpl *cpl);
+
+void bs_batch_read_bs_dev(spdk_bs_batch_t *batch, struct spdk_bs_dev *bs_dev,
+			  void *payload, uint64_t lba, uint32_t lba_count);
+
+void bs_batch_read_dev(spdk_bs_batch_t *batch, void *payload,
+		       uint64_t lba, uint32_t lba_count);
+
+void bs_batch_write_dev(spdk_bs_batch_t *batch, void *payload,
+			uint64_t lba, uint32_t lba_count);
+
+void bs_batch_unmap_dev(spdk_bs_batch_t *batch,
+			uint64_t lba, uint32_t lba_count);
+
+void bs_batch_write_zeroes_dev(spdk_bs_batch_t *batch,
+			       uint64_t lba, uint32_t lba_count);
+
+void bs_batch_close(spdk_bs_batch_t *batch);
+
+spdk_bs_batch_t *bs_sequence_to_batch(spdk_bs_sequence_t *seq,
+				      spdk_bs_sequence_cpl cb_fn,
+				      void *cb_arg);
+
+spdk_bs_user_op_t *bs_user_op_alloc(struct spdk_io_channel *channel, struct spdk_bs_cpl *cpl,
+				    enum spdk_blob_op_type op_type, struct spdk_blob *blob,
+				    void *payload, int iovcnt, uint64_t offset, uint64_t length);
+
+void bs_user_op_execute(spdk_bs_user_op_t *op);
+
+void bs_user_op_abort(spdk_bs_user_op_t *op);
+
+void bs_sequence_to_batch_completion(void *cb_arg, int bserrno);
+
+#endif
diff --git a/src/spdk/lib/blob/spdk_blob.map b/src/spdk/lib/blob/spdk_blob.map
new file mode 100644
index 000000000..7c1bc473f
--- /dev/null
+++ b/src/spdk/lib/blob/spdk_blob.map
@@ -0,0 +1,64 @@
+{
+	global:
+
+	# Public functions
+	spdk_bs_opts_init;
+	spdk_bs_load;
+	spdk_bs_init;
+	spdk_bs_dump;
+	spdk_bs_destroy;
+	spdk_bs_unload;
+	spdk_bs_set_super;
+	spdk_bs_get_super;
+	spdk_bs_get_cluster_size;
+	spdk_bs_get_page_size;
+	spdk_bs_get_io_unit_size;
+	spdk_bs_free_cluster_count;
+	spdk_bs_total_data_cluster_count;
+	spdk_blob_get_id;
+	spdk_blob_get_num_pages;
+	spdk_blob_get_num_io_units;
+	spdk_blob_get_num_clusters;
+	spdk_blob_opts_init;
+	spdk_bs_create_blob_ext;
+	spdk_bs_create_blob;
+	spdk_bs_create_snapshot;
+	spdk_bs_create_clone;
+	spdk_blob_get_clones;
+	spdk_blob_get_parent_snapshot;
+	spdk_blob_is_read_only;
+	spdk_blob_is_snapshot;
+	spdk_blob_is_clone;
+	spdk_blob_is_thin_provisioned;
+	spdk_bs_delete_blob;
+	spdk_bs_inflate_blob;
+	spdk_bs_blob_decouple_parent;
+	spdk_blob_open_opts_init;
+	spdk_bs_open_blob;
+	spdk_bs_open_blob_ext;
+	spdk_blob_resize;
+	spdk_blob_set_read_only;
+	spdk_blob_sync_md;
+	spdk_blob_close;
+	spdk_bs_alloc_io_channel;
+	spdk_bs_free_io_channel;
+	spdk_blob_io_write;
+	spdk_blob_io_read;
+	spdk_blob_io_writev;
+	spdk_blob_io_readv;
+	spdk_blob_io_unmap;
+	spdk_blob_io_write_zeroes;
+	spdk_bs_iter_first;
+	spdk_bs_iter_next;
+	spdk_blob_set_xattr;
+	spdk_blob_remove_xattr;
+	spdk_blob_get_xattr_value;
+	spdk_blob_get_xattr_names;
+	spdk_xattr_names_get_count;
+	spdk_xattr_names_get_name;
+	spdk_xattr_names_free;
+	spdk_bs_get_bstype;
+	spdk_bs_set_bstype;
+
+	local: *;
+};
diff --git a/src/spdk/lib/blob/zeroes.c b/src/spdk/lib/blob/zeroes.c
new file mode 100644
index 000000000..5e8d70545
--- /dev/null
+++ b/src/spdk/lib/blob/zeroes.c
@@ -0,0 +1,122 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+#include "spdk/blob.h"
+
+#include "blobstore.h"
+
+static void
+zeroes_destroy(struct spdk_bs_dev *bs_dev)
+{
+	return;
+}
+
+static void
+zeroes_read(struct spdk_bs_dev *dev, struct spdk_io_channel *channel, void *payload,
+	    uint64_t lba, uint32_t lba_count, struct spdk_bs_dev_cb_args *cb_args)
+{
+	memset(payload, 0, dev->blocklen * lba_count);
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, 0);
+}
+
+static void
+zeroes_write(struct spdk_bs_dev *dev, struct spdk_io_channel *channel, void *payload,
+	     uint64_t lba, uint32_t lba_count,
+	     struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+zeroes_readv(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+	     struct iovec *iov, int iovcnt,
+	     uint64_t lba, uint32_t lba_count, struct spdk_bs_dev_cb_args *cb_args)
+{
+	int i;
+
+	for (i = 0; i < iovcnt; i++) {
+		memset(iov[i].iov_base, 0, iov[i].iov_len);
+	}
+
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, 0);
+}
+
+static void
+zeroes_writev(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+	      struct iovec *iov, int iovcnt,
+	      uint64_t lba, uint32_t lba_count,
+	      struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+zeroes_write_zeroes(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+		    uint64_t lba, uint32_t lba_count,
+		    struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static void
+zeroes_unmap(struct spdk_bs_dev *dev, struct spdk_io_channel *channel,
+	     uint64_t lba, uint32_t lba_count,
+	     struct spdk_bs_dev_cb_args *cb_args)
+{
+	cb_args->cb_fn(cb_args->channel, cb_args->cb_arg, -EPERM);
+	assert(false);
+}
+
+static struct spdk_bs_dev g_zeroes_bs_dev = {
+	.blockcnt = UINT64_MAX,
+	.blocklen = 512,
+	.create_channel = NULL,
+	.destroy_channel = NULL,
+	.destroy = zeroes_destroy,
+	.read = zeroes_read,
+	.write = zeroes_write,
+	.readv = zeroes_readv,
+	.writev = zeroes_writev,
+	.write_zeroes = zeroes_write_zeroes,
+	.unmap = zeroes_unmap,
+};
+
+struct spdk_bs_dev *
+bs_create_zeroes_dev(void)
+{
+	return &g_zeroes_bs_dev;
+}
diff --git a/src/spdk/lib/blobfs/Makefile b/src/spdk/lib/blobfs/Makefile
new file mode 100644
index 000000000..d0c46de02
--- /dev/null
+++ b/src/spdk/lib/blobfs/Makefile
@@ -0,0 +1,45 @@
+#
+#  BSD LICENSE
+#
+#  Copyright (c) Intel Corporation.
+#  All rights reserved.
+#
+#  Redistribution and use in source and binary forms, with or without
+#  modification, are permitted provided that the following conditions
+#  are met:
+#
+#    * Redistributions of source code must retain the above copyright
+#      notice, this list of conditions and the following disclaimer.
+#    * Redistributions in binary form must reproduce the above copyright
+#      notice, this list of conditions and the following disclaimer in
+#      the documentation and/or other materials provided with the
+#      distribution.
+#    * Neither the name of Intel Corporation nor the names of its
+#      contributors may be used to endorse or promote products derived
+#      from this software without specific prior written permission.
+#
+#  THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+#  "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+#  LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+#  A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+#  OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+#  SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+#  LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+#  DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+#  THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+#  (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+#  OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+#
+
+SPDK_ROOT_DIR := $(abspath $(CURDIR)/../..)
+include $(SPDK_ROOT_DIR)/mk/spdk.common.mk
+
+SO_VER := 3
+SO_MINOR := 0
+
+C_SRCS = blobfs.c tree.c
+LIBNAME = blobfs
+
+SPDK_MAP_FILE = $(abspath $(CURDIR)/spdk_blobfs.map)
+
+include $(SPDK_ROOT_DIR)/mk/spdk.lib.mk
diff --git a/src/spdk/lib/blobfs/blobfs.c b/src/spdk/lib/blobfs/blobfs.c
new file mode 100644
index 000000000..3af6b0639
--- /dev/null
+++ b/src/spdk/lib/blobfs/blobfs.c
@@ -0,0 +1,2980 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+
+#include "spdk/blobfs.h"
+#include "spdk/conf.h"
+#include "tree.h"
+
+#include "spdk/queue.h"
+#include "spdk/thread.h"
+#include "spdk/assert.h"
+#include "spdk/env.h"
+#include "spdk/util.h"
+#include "spdk_internal/log.h"
+#include "spdk/trace.h"
+
+#define BLOBFS_TRACE(file, str, args...) \
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s " str, file->name, ##args)
+
+#define BLOBFS_TRACE_RW(file, str, args...) \
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS_RW, "file=%s " str, file->name, ##args)
+
+#define BLOBFS_DEFAULT_CACHE_SIZE (4ULL * 1024 * 1024 * 1024)
+#define SPDK_BLOBFS_DEFAULT_OPTS_CLUSTER_SZ (1024 * 1024)
+
+#define SPDK_BLOBFS_SIGNATURE	"BLOBFS"
+
+static uint64_t g_fs_cache_size = BLOBFS_DEFAULT_CACHE_SIZE;
+static struct spdk_mempool *g_cache_pool;
+static TAILQ_HEAD(, spdk_file) g_caches;
+static struct spdk_poller *g_cache_pool_mgmt_poller;
+static struct spdk_thread *g_cache_pool_thread;
+#define BLOBFS_CACHE_POOL_POLL_PERIOD_IN_US 1000ULL
+static int g_fs_count = 0;
+static pthread_mutex_t g_cache_init_lock = PTHREAD_MUTEX_INITIALIZER;
+
+#define TRACE_GROUP_BLOBFS	0x7
+#define TRACE_BLOBFS_XATTR_START	SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x0)
+#define TRACE_BLOBFS_XATTR_END		SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x1)
+#define TRACE_BLOBFS_OPEN		SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x2)
+#define TRACE_BLOBFS_CLOSE		SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x3)
+#define TRACE_BLOBFS_DELETE_START	SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x4)
+#define TRACE_BLOBFS_DELETE_DONE	SPDK_TPOINT_ID(TRACE_GROUP_BLOBFS, 0x5)
+
+SPDK_TRACE_REGISTER_FN(blobfs_trace, "blobfs", TRACE_GROUP_BLOBFS)
+{
+	spdk_trace_register_description("BLOBFS_XATTR_START",
+					TRACE_BLOBFS_XATTR_START,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+	spdk_trace_register_description("BLOBFS_XATTR_END",
+					TRACE_BLOBFS_XATTR_END,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+	spdk_trace_register_description("BLOBFS_OPEN",
+					TRACE_BLOBFS_OPEN,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+	spdk_trace_register_description("BLOBFS_CLOSE",
+					TRACE_BLOBFS_CLOSE,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+	spdk_trace_register_description("BLOBFS_DELETE_START",
+					TRACE_BLOBFS_DELETE_START,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+	spdk_trace_register_description("BLOBFS_DELETE_DONE",
+					TRACE_BLOBFS_DELETE_DONE,
+					OWNER_NONE, OBJECT_NONE, 0,
+					SPDK_TRACE_ARG_TYPE_STR,
+					"file:    ");
+}
+
+void
+cache_buffer_free(struct cache_buffer *cache_buffer)
+{
+	spdk_mempool_put(g_cache_pool, cache_buffer->buf);
+	free(cache_buffer);
+}
+
+#define CACHE_READAHEAD_THRESHOLD	(128 * 1024)
+
+struct spdk_file {
+	struct spdk_filesystem	*fs;
+	struct spdk_blob	*blob;
+	char			*name;
+	uint64_t		trace_arg_name;
+	uint64_t		length;
+	bool                    is_deleted;
+	bool			open_for_writing;
+	uint64_t		length_flushed;
+	uint64_t		length_xattr;
+	uint64_t		append_pos;
+	uint64_t		seq_byte_count;
+	uint64_t		next_seq_offset;
+	uint32_t		priority;
+	TAILQ_ENTRY(spdk_file)	tailq;
+	spdk_blob_id		blobid;
+	uint32_t		ref_count;
+	pthread_spinlock_t	lock;
+	struct cache_buffer	*last;
+	struct cache_tree	*tree;
+	TAILQ_HEAD(open_requests_head, spdk_fs_request) open_requests;
+	TAILQ_HEAD(sync_requests_head, spdk_fs_request) sync_requests;
+	TAILQ_ENTRY(spdk_file)	cache_tailq;
+};
+
+struct spdk_deleted_file {
+	spdk_blob_id	id;
+	TAILQ_ENTRY(spdk_deleted_file)	tailq;
+};
+
+struct spdk_filesystem {
+	struct spdk_blob_store	*bs;
+	TAILQ_HEAD(, spdk_file)	files;
+	struct spdk_bs_opts	bs_opts;
+	struct spdk_bs_dev	*bdev;
+	fs_send_request_fn	send_request;
+
+	struct {
+		uint32_t		max_ops;
+		struct spdk_io_channel	*sync_io_channel;
+		struct spdk_fs_channel	*sync_fs_channel;
+	} sync_target;
+
+	struct {
+		uint32_t		max_ops;
+		struct spdk_io_channel	*md_io_channel;
+		struct spdk_fs_channel	*md_fs_channel;
+	} md_target;
+
+	struct {
+		uint32_t		max_ops;
+	} io_target;
+};
+
+struct spdk_fs_cb_args {
+	union {
+		spdk_fs_op_with_handle_complete		fs_op_with_handle;
+		spdk_fs_op_complete			fs_op;
+		spdk_file_op_with_handle_complete	file_op_with_handle;
+		spdk_file_op_complete			file_op;
+		spdk_file_stat_op_complete		stat_op;
+	} fn;
+	void *arg;
+	sem_t *sem;
+	struct spdk_filesystem *fs;
+	struct spdk_file *file;
+	int rc;
+	struct iovec *iovs;
+	uint32_t iovcnt;
+	struct iovec iov;
+	union {
+		struct {
+			TAILQ_HEAD(, spdk_deleted_file)	deleted_files;
+		} fs_load;
+		struct {
+			uint64_t	length;
+		} truncate;
+		struct {
+			struct spdk_io_channel	*channel;
+			void		*pin_buf;
+			int		is_read;
+			off_t		offset;
+			size_t		length;
+			uint64_t	start_lba;
+			uint64_t	num_lba;
+			uint32_t	blocklen;
+		} rw;
+		struct {
+			const char	*old_name;
+			const char	*new_name;
+		} rename;
+		struct {
+			struct cache_buffer	*cache_buffer;
+			uint64_t		length;
+		} flush;
+		struct {
+			struct cache_buffer	*cache_buffer;
+			uint64_t		length;
+			uint64_t		offset;
+		} readahead;
+		struct {
+			/* offset of the file when the sync request was made */
+			uint64_t			offset;
+			TAILQ_ENTRY(spdk_fs_request)	tailq;
+			bool				xattr_in_progress;
+			/* length written to the xattr for this file - this should
+			 * always be the same as the offset if only one thread is
+			 * writing to the file, but could differ if multiple threads
+			 * are appending
+			 */
+			uint64_t			length;
+		} sync;
+		struct {
+			uint32_t			num_clusters;
+		} resize;
+		struct {
+			const char	*name;
+			uint32_t	flags;
+			TAILQ_ENTRY(spdk_fs_request)	tailq;
+		} open;
+		struct {
+			const char		*name;
+			struct spdk_blob	*blob;
+		} create;
+		struct {
+			const char	*name;
+		} delete;
+		struct {
+			const char	*name;
+		} stat;
+	} op;
+};
+
+static void file_free(struct spdk_file *file);
+static void fs_io_device_unregister(struct spdk_filesystem *fs);
+static void fs_free_io_channels(struct spdk_filesystem *fs);
+
+void
+spdk_fs_opts_init(struct spdk_blobfs_opts *opts)
+{
+	opts->cluster_sz = SPDK_BLOBFS_DEFAULT_OPTS_CLUSTER_SZ;
+}
+
+static int _blobfs_cache_pool_reclaim(void *arg);
+
+static bool
+blobfs_cache_pool_need_reclaim(void)
+{
+	size_t count;
+
+	count = spdk_mempool_count(g_cache_pool);
+	/* We define a aggressive policy here as the requirements from db_bench are batched, so start the poller
+	 *  when the number of available cache buffer is less than 1/5 of total buffers.
+	 */
+	if (count > (size_t)g_fs_cache_size / CACHE_BUFFER_SIZE / 5) {
+		return false;
+	}
+
+	return true;
+}
+
+static void
+__start_cache_pool_mgmt(void *ctx)
+{
+	assert(g_cache_pool == NULL);
+
+	g_cache_pool = spdk_mempool_create("spdk_fs_cache",
+					   g_fs_cache_size / CACHE_BUFFER_SIZE,
+					   CACHE_BUFFER_SIZE,
+					   SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
+					   SPDK_ENV_SOCKET_ID_ANY);
+	if (!g_cache_pool) {
+		SPDK_ERRLOG("Create mempool failed, you may "
+			    "increase the memory and try again\n");
+		assert(false);
+	}
+	TAILQ_INIT(&g_caches);
+
+	assert(g_cache_pool_mgmt_poller == NULL);
+	g_cache_pool_mgmt_poller = SPDK_POLLER_REGISTER(_blobfs_cache_pool_reclaim, NULL,
+				   BLOBFS_CACHE_POOL_POLL_PERIOD_IN_US);
+}
+
+static void
+__stop_cache_pool_mgmt(void *ctx)
+{
+	spdk_poller_unregister(&g_cache_pool_mgmt_poller);
+
+	assert(g_cache_pool != NULL);
+	assert(spdk_mempool_count(g_cache_pool) == g_fs_cache_size / CACHE_BUFFER_SIZE);
+	spdk_mempool_free(g_cache_pool);
+	g_cache_pool = NULL;
+
+	spdk_thread_exit(g_cache_pool_thread);
+}
+
+static void
+initialize_global_cache(void)
+{
+	pthread_mutex_lock(&g_cache_init_lock);
+	if (g_fs_count == 0) {
+		g_cache_pool_thread = spdk_thread_create("cache_pool_mgmt", NULL);
+		assert(g_cache_pool_thread != NULL);
+		spdk_thread_send_msg(g_cache_pool_thread, __start_cache_pool_mgmt, NULL);
+	}
+	g_fs_count++;
+	pthread_mutex_unlock(&g_cache_init_lock);
+}
+
+static void
+free_global_cache(void)
+{
+	pthread_mutex_lock(&g_cache_init_lock);
+	g_fs_count--;
+	if (g_fs_count == 0) {
+		spdk_thread_send_msg(g_cache_pool_thread, __stop_cache_pool_mgmt, NULL);
+	}
+	pthread_mutex_unlock(&g_cache_init_lock);
+}
+
+static uint64_t
+__file_get_blob_size(struct spdk_file *file)
+{
+	uint64_t cluster_sz;
+
+	cluster_sz = file->fs->bs_opts.cluster_sz;
+	return cluster_sz * spdk_blob_get_num_clusters(file->blob);
+}
+
+struct spdk_fs_request {
+	struct spdk_fs_cb_args		args;
+	TAILQ_ENTRY(spdk_fs_request)	link;
+	struct spdk_fs_channel		*channel;
+};
+
+struct spdk_fs_channel {
+	struct spdk_fs_request		*req_mem;
+	TAILQ_HEAD(, spdk_fs_request)	reqs;
+	sem_t				sem;
+	struct spdk_filesystem		*fs;
+	struct spdk_io_channel		*bs_channel;
+	fs_send_request_fn		send_request;
+	bool				sync;
+	uint32_t			outstanding_reqs;
+	pthread_spinlock_t		lock;
+};
+
+/* For now, this is effectively an alias. But eventually we'll shift
+ * some data members over. */
+struct spdk_fs_thread_ctx {
+	struct spdk_fs_channel	ch;
+};
+
+static struct spdk_fs_request *
+alloc_fs_request_with_iov(struct spdk_fs_channel *channel, uint32_t iovcnt)
+{
+	struct spdk_fs_request *req;
+	struct iovec *iovs = NULL;
+
+	if (iovcnt > 1) {
+		iovs = calloc(iovcnt, sizeof(struct iovec));
+		if (!iovs) {
+			return NULL;
+		}
+	}
+
+	if (channel->sync) {
+		pthread_spin_lock(&channel->lock);
+	}
+
+	req = TAILQ_FIRST(&channel->reqs);
+	if (req) {
+		channel->outstanding_reqs++;
+		TAILQ_REMOVE(&channel->reqs, req, link);
+	}
+
+	if (channel->sync) {
+		pthread_spin_unlock(&channel->lock);
+	}
+
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate req on spdk_fs_channel =%p\n", channel);
+		free(iovs);
+		return NULL;
+	}
+	memset(req, 0, sizeof(*req));
+	req->channel = channel;
+	if (iovcnt > 1) {
+		req->args.iovs = iovs;
+	} else {
+		req->args.iovs = &req->args.iov;
+	}
+	req->args.iovcnt = iovcnt;
+
+	return req;
+}
+
+static struct spdk_fs_request *
+alloc_fs_request(struct spdk_fs_channel *channel)
+{
+	return alloc_fs_request_with_iov(channel, 0);
+}
+
+static void
+free_fs_request(struct spdk_fs_request *req)
+{
+	struct spdk_fs_channel *channel = req->channel;
+
+	if (req->args.iovcnt > 1) {
+		free(req->args.iovs);
+	}
+
+	if (channel->sync) {
+		pthread_spin_lock(&channel->lock);
+	}
+
+	TAILQ_INSERT_HEAD(&req->channel->reqs, req, link);
+	channel->outstanding_reqs--;
+
+	if (channel->sync) {
+		pthread_spin_unlock(&channel->lock);
+	}
+}
+
+static int
+fs_channel_create(struct spdk_filesystem *fs, struct spdk_fs_channel *channel,
+		  uint32_t max_ops)
+{
+	uint32_t i;
+
+	channel->req_mem = calloc(max_ops, sizeof(struct spdk_fs_request));
+	if (!channel->req_mem) {
+		return -1;
+	}
+
+	channel->outstanding_reqs = 0;
+	TAILQ_INIT(&channel->reqs);
+	sem_init(&channel->sem, 0, 0);
+
+	for (i = 0; i < max_ops; i++) {
+		TAILQ_INSERT_TAIL(&channel->reqs, &channel->req_mem[i], link);
+	}
+
+	channel->fs = fs;
+
+	return 0;
+}
+
+static int
+fs_md_channel_create(void *io_device, void *ctx_buf)
+{
+	struct spdk_filesystem		*fs;
+	struct spdk_fs_channel		*channel = ctx_buf;
+
+	fs = SPDK_CONTAINEROF(io_device, struct spdk_filesystem, md_target);
+
+	return fs_channel_create(fs, channel, fs->md_target.max_ops);
+}
+
+static int
+fs_sync_channel_create(void *io_device, void *ctx_buf)
+{
+	struct spdk_filesystem		*fs;
+	struct spdk_fs_channel		*channel = ctx_buf;
+
+	fs = SPDK_CONTAINEROF(io_device, struct spdk_filesystem, sync_target);
+
+	return fs_channel_create(fs, channel, fs->sync_target.max_ops);
+}
+
+static int
+fs_io_channel_create(void *io_device, void *ctx_buf)
+{
+	struct spdk_filesystem		*fs;
+	struct spdk_fs_channel		*channel = ctx_buf;
+
+	fs = SPDK_CONTAINEROF(io_device, struct spdk_filesystem, io_target);
+
+	return fs_channel_create(fs, channel, fs->io_target.max_ops);
+}
+
+static void
+fs_channel_destroy(void *io_device, void *ctx_buf)
+{
+	struct spdk_fs_channel *channel = ctx_buf;
+
+	if (channel->outstanding_reqs > 0) {
+		SPDK_ERRLOG("channel freed with %" PRIu32 " outstanding requests!\n",
+			    channel->outstanding_reqs);
+	}
+
+	free(channel->req_mem);
+	if (channel->bs_channel != NULL) {
+		spdk_bs_free_io_channel(channel->bs_channel);
+	}
+}
+
+static void
+__send_request_direct(fs_request_fn fn, void *arg)
+{
+	fn(arg);
+}
+
+static void
+common_fs_bs_init(struct spdk_filesystem *fs, struct spdk_blob_store *bs)
+{
+	fs->bs = bs;
+	fs->bs_opts.cluster_sz = spdk_bs_get_cluster_size(bs);
+	fs->md_target.md_fs_channel->bs_channel = spdk_bs_alloc_io_channel(fs->bs);
+	fs->md_target.md_fs_channel->send_request = __send_request_direct;
+	fs->sync_target.sync_fs_channel->bs_channel = spdk_bs_alloc_io_channel(fs->bs);
+	fs->sync_target.sync_fs_channel->send_request = __send_request_direct;
+
+	initialize_global_cache();
+}
+
+static void
+init_cb(void *ctx, struct spdk_blob_store *bs, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+
+	if (bserrno == 0) {
+		common_fs_bs_init(fs, bs);
+	} else {
+		free(fs);
+		fs = NULL;
+	}
+
+	args->fn.fs_op_with_handle(args->arg, fs, bserrno);
+	free_fs_request(req);
+}
+
+static void
+fs_conf_parse(void)
+{
+	struct spdk_conf_section *sp;
+	int cache_buffer_shift;
+
+	sp = spdk_conf_find_section(NULL, "Blobfs");
+	if (sp == NULL) {
+		g_fs_cache_buffer_shift = CACHE_BUFFER_SHIFT_DEFAULT;
+		return;
+	}
+
+	cache_buffer_shift = spdk_conf_section_get_intval(sp, "CacheBufferShift");
+	if (cache_buffer_shift <= 0) {
+		g_fs_cache_buffer_shift = CACHE_BUFFER_SHIFT_DEFAULT;
+	} else {
+		g_fs_cache_buffer_shift = cache_buffer_shift;
+	}
+}
+
+static struct spdk_filesystem *
+fs_alloc(struct spdk_bs_dev *dev, fs_send_request_fn send_request_fn)
+{
+	struct spdk_filesystem *fs;
+
+	fs = calloc(1, sizeof(*fs));
+	if (fs == NULL) {
+		return NULL;
+	}
+
+	fs->bdev = dev;
+	fs->send_request = send_request_fn;
+	TAILQ_INIT(&fs->files);
+
+	fs->md_target.max_ops = 512;
+	spdk_io_device_register(&fs->md_target, fs_md_channel_create, fs_channel_destroy,
+				sizeof(struct spdk_fs_channel), "blobfs_md");
+	fs->md_target.md_io_channel = spdk_get_io_channel(&fs->md_target);
+	fs->md_target.md_fs_channel = spdk_io_channel_get_ctx(fs->md_target.md_io_channel);
+
+	fs->sync_target.max_ops = 512;
+	spdk_io_device_register(&fs->sync_target, fs_sync_channel_create, fs_channel_destroy,
+				sizeof(struct spdk_fs_channel), "blobfs_sync");
+	fs->sync_target.sync_io_channel = spdk_get_io_channel(&fs->sync_target);
+	fs->sync_target.sync_fs_channel = spdk_io_channel_get_ctx(fs->sync_target.sync_io_channel);
+
+	fs->io_target.max_ops = 512;
+	spdk_io_device_register(&fs->io_target, fs_io_channel_create, fs_channel_destroy,
+				sizeof(struct spdk_fs_channel), "blobfs_io");
+
+	return fs;
+}
+
+static void
+__wake_caller(void *arg, int fserrno)
+{
+	struct spdk_fs_cb_args *args = arg;
+
+	args->rc = fserrno;
+	sem_post(args->sem);
+}
+
+void
+spdk_fs_init(struct spdk_bs_dev *dev, struct spdk_blobfs_opts *opt,
+	     fs_send_request_fn send_request_fn,
+	     spdk_fs_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_filesystem *fs;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	struct spdk_bs_opts opts = {};
+
+	fs = fs_alloc(dev, send_request_fn);
+	if (fs == NULL) {
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	fs_conf_parse();
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		fs_free_io_channels(fs);
+		fs_io_device_unregister(fs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.fs_op_with_handle = cb_fn;
+	args->arg = cb_arg;
+	args->fs = fs;
+
+	spdk_bs_opts_init(&opts);
+	snprintf(opts.bstype.bstype, sizeof(opts.bstype.bstype), SPDK_BLOBFS_SIGNATURE);
+	if (opt) {
+		opts.cluster_sz = opt->cluster_sz;
+	}
+	spdk_bs_init(dev, &opts, init_cb, req);
+}
+
+static struct spdk_file *
+file_alloc(struct spdk_filesystem *fs)
+{
+	struct spdk_file *file;
+
+	file = calloc(1, sizeof(*file));
+	if (file == NULL) {
+		return NULL;
+	}
+
+	file->tree = calloc(1, sizeof(*file->tree));
+	if (file->tree == NULL) {
+		free(file);
+		return NULL;
+	}
+
+	if (pthread_spin_init(&file->lock, 0)) {
+		free(file->tree);
+		free(file);
+		return NULL;
+	}
+
+	file->fs = fs;
+	TAILQ_INIT(&file->open_requests);
+	TAILQ_INIT(&file->sync_requests);
+	TAILQ_INSERT_TAIL(&fs->files, file, tailq);
+	file->priority = SPDK_FILE_PRIORITY_LOW;
+	return file;
+}
+
+static void fs_load_done(void *ctx, int bserrno);
+
+static int
+_handle_deleted_files(struct spdk_fs_request *req)
+{
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+
+	if (!TAILQ_EMPTY(&args->op.fs_load.deleted_files)) {
+		struct spdk_deleted_file *deleted_file;
+
+		deleted_file = TAILQ_FIRST(&args->op.fs_load.deleted_files);
+		TAILQ_REMOVE(&args->op.fs_load.deleted_files, deleted_file, tailq);
+		spdk_bs_delete_blob(fs->bs, deleted_file->id, fs_load_done, req);
+		free(deleted_file);
+		return 0;
+	}
+
+	return 1;
+}
+
+static void
+fs_load_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+
+	/* The filesystem has been loaded.  Now check if there are any files that
+	 *  were marked for deletion before last unload.  Do not complete the
+	 *  fs_load callback until all of them have been deleted on disk.
+	 */
+	if (_handle_deleted_files(req) == 0) {
+		/* We found a file that's been marked for deleting but not actually
+		 *  deleted yet.  This function will get called again once the delete
+		 *  operation is completed.
+		 */
+		return;
+	}
+
+	args->fn.fs_op_with_handle(args->arg, fs, 0);
+	free_fs_request(req);
+
+}
+
+static void
+_file_build_trace_arg_name(struct spdk_file *f)
+{
+	f->trace_arg_name = 0;
+	memcpy(&f->trace_arg_name, f->name,
+	       spdk_min(sizeof(f->trace_arg_name), strlen(f->name)));
+}
+
+static void
+iter_cb(void *ctx, struct spdk_blob *blob, int rc)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+	uint64_t *length;
+	const char *name;
+	uint32_t *is_deleted;
+	size_t value_len;
+
+	if (rc < 0) {
+		args->fn.fs_op_with_handle(args->arg, fs, rc);
+		free_fs_request(req);
+		return;
+	}
+
+	rc = spdk_blob_get_xattr_value(blob, "name", (const void **)&name, &value_len);
+	if (rc < 0) {
+		args->fn.fs_op_with_handle(args->arg, fs, rc);
+		free_fs_request(req);
+		return;
+	}
+
+	rc = spdk_blob_get_xattr_value(blob, "length", (const void **)&length, &value_len);
+	if (rc < 0) {
+		args->fn.fs_op_with_handle(args->arg, fs, rc);
+		free_fs_request(req);
+		return;
+	}
+
+	assert(value_len == 8);
+
+	/* This file could be deleted last time without close it, then app crashed, so we delete it now */
+	rc = spdk_blob_get_xattr_value(blob, "is_deleted", (const void **)&is_deleted, &value_len);
+	if (rc < 0) {
+		struct spdk_file *f;
+
+		f = file_alloc(fs);
+		if (f == NULL) {
+			SPDK_ERRLOG("Cannot allocate file to handle deleted file on disk\n");
+			args->fn.fs_op_with_handle(args->arg, fs, -ENOMEM);
+			free_fs_request(req);
+			return;
+		}
+
+		f->name = strdup(name);
+		_file_build_trace_arg_name(f);
+		f->blobid = spdk_blob_get_id(blob);
+		f->length = *length;
+		f->length_flushed = *length;
+		f->length_xattr = *length;
+		f->append_pos = *length;
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "added file %s length=%ju\n", f->name, f->length);
+	} else {
+		struct spdk_deleted_file *deleted_file;
+
+		deleted_file = calloc(1, sizeof(*deleted_file));
+		if (deleted_file == NULL) {
+			args->fn.fs_op_with_handle(args->arg, fs, -ENOMEM);
+			free_fs_request(req);
+			return;
+		}
+		deleted_file->id = spdk_blob_get_id(blob);
+		TAILQ_INSERT_TAIL(&args->op.fs_load.deleted_files, deleted_file, tailq);
+	}
+}
+
+static void
+load_cb(void *ctx, struct spdk_blob_store *bs, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+	struct spdk_bs_type bstype;
+	static const struct spdk_bs_type blobfs_type = {SPDK_BLOBFS_SIGNATURE};
+	static const struct spdk_bs_type zeros;
+
+	if (bserrno != 0) {
+		args->fn.fs_op_with_handle(args->arg, NULL, bserrno);
+		free_fs_request(req);
+		fs_free_io_channels(fs);
+		fs_io_device_unregister(fs);
+		return;
+	}
+
+	bstype = spdk_bs_get_bstype(bs);
+
+	if (!memcmp(&bstype, &zeros, sizeof(bstype))) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "assigning bstype\n");
+		spdk_bs_set_bstype(bs, blobfs_type);
+	} else if (memcmp(&bstype, &blobfs_type, sizeof(bstype))) {
+		SPDK_ERRLOG("not blobfs\n");
+		SPDK_LOGDUMP(SPDK_LOG_BLOBFS, "bstype", &bstype, sizeof(bstype));
+		args->fn.fs_op_with_handle(args->arg, NULL, -EINVAL);
+		free_fs_request(req);
+		fs_free_io_channels(fs);
+		fs_io_device_unregister(fs);
+		return;
+	}
+
+	common_fs_bs_init(fs, bs);
+	fs_load_done(req, 0);
+}
+
+static void
+fs_io_device_unregister(struct spdk_filesystem *fs)
+{
+	assert(fs != NULL);
+	spdk_io_device_unregister(&fs->md_target, NULL);
+	spdk_io_device_unregister(&fs->sync_target, NULL);
+	spdk_io_device_unregister(&fs->io_target, NULL);
+	free(fs);
+}
+
+static void
+fs_free_io_channels(struct spdk_filesystem *fs)
+{
+	assert(fs != NULL);
+	spdk_fs_free_io_channel(fs->md_target.md_io_channel);
+	spdk_fs_free_io_channel(fs->sync_target.sync_io_channel);
+}
+
+void
+spdk_fs_load(struct spdk_bs_dev *dev, fs_send_request_fn send_request_fn,
+	     spdk_fs_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_filesystem *fs;
+	struct spdk_fs_cb_args *args;
+	struct spdk_fs_request *req;
+	struct spdk_bs_opts	bs_opts;
+
+	fs = fs_alloc(dev, send_request_fn);
+	if (fs == NULL) {
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	fs_conf_parse();
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		fs_free_io_channels(fs);
+		fs_io_device_unregister(fs);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.fs_op_with_handle = cb_fn;
+	args->arg = cb_arg;
+	args->fs = fs;
+	TAILQ_INIT(&args->op.fs_load.deleted_files);
+	spdk_bs_opts_init(&bs_opts);
+	bs_opts.iter_cb_fn = iter_cb;
+	bs_opts.iter_cb_arg = req;
+	spdk_bs_load(dev, &bs_opts, load_cb, req);
+}
+
+static void
+unload_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_filesystem *fs = args->fs;
+	struct spdk_file *file, *tmp;
+
+	TAILQ_FOREACH_SAFE(file, &fs->files, tailq, tmp) {
+		TAILQ_REMOVE(&fs->files, file, tailq);
+		file_free(file);
+	}
+
+	free_global_cache();
+
+	args->fn.fs_op(args->arg, bserrno);
+	free(req);
+
+	fs_io_device_unregister(fs);
+}
+
+void
+spdk_fs_unload(struct spdk_filesystem *fs, spdk_fs_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	/*
+	 * We must free the md_channel before unloading the blobstore, so just
+	 *  allocate this request from the general heap.
+	 */
+	req = calloc(1, sizeof(*req));
+	if (req == NULL) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.fs_op = cb_fn;
+	args->arg = cb_arg;
+	args->fs = fs;
+
+	fs_free_io_channels(fs);
+	spdk_bs_unload(fs->bs, unload_cb, req);
+}
+
+static struct spdk_file *
+fs_find_file(struct spdk_filesystem *fs, const char *name)
+{
+	struct spdk_file *file;
+
+	TAILQ_FOREACH(file, &fs->files, tailq) {
+		if (!strncmp(name, file->name, SPDK_FILE_NAME_MAX)) {
+			return file;
+		}
+	}
+
+	return NULL;
+}
+
+void
+spdk_fs_file_stat_async(struct spdk_filesystem *fs, const char *name,
+			spdk_file_stat_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_file_stat stat;
+	struct spdk_file *f = NULL;
+
+	if (strnlen(name, SPDK_FILE_NAME_MAX + 1) == SPDK_FILE_NAME_MAX + 1) {
+		cb_fn(cb_arg, NULL, -ENAMETOOLONG);
+		return;
+	}
+
+	f = fs_find_file(fs, name);
+	if (f != NULL) {
+		stat.blobid = f->blobid;
+		stat.size = f->append_pos >= f->length ? f->append_pos : f->length;
+		cb_fn(cb_arg, &stat, 0);
+		return;
+	}
+
+	cb_fn(cb_arg, NULL, -ENOENT);
+}
+
+static void
+__copy_stat(void *arg, struct spdk_file_stat *stat, int fserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	args->rc = fserrno;
+	if (fserrno == 0) {
+		memcpy(args->arg, stat, sizeof(*stat));
+	}
+	sem_post(args->sem);
+}
+
+static void
+__file_stat(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_fs_file_stat_async(args->fs, args->op.stat.name,
+				args->fn.stat_op, req);
+}
+
+int
+spdk_fs_file_stat(struct spdk_filesystem *fs, struct spdk_fs_thread_ctx *ctx,
+		  const char *name, struct spdk_file_stat *stat)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	int rc;
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate stat req on file=%s\n", name);
+		return -ENOMEM;
+	}
+
+	req->args.fs = fs;
+	req->args.op.stat.name = name;
+	req->args.fn.stat_op = __copy_stat;
+	req->args.arg = stat;
+	req->args.sem = &channel->sem;
+	channel->send_request(__file_stat, req);
+	sem_wait(&channel->sem);
+
+	rc = req->args.rc;
+	free_fs_request(req);
+
+	return rc;
+}
+
+static void
+fs_create_blob_close_cb(void *ctx, int bserrno)
+{
+	int rc;
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	rc = args->rc ? args->rc : bserrno;
+	args->fn.file_op(args->arg, rc);
+	free_fs_request(req);
+}
+
+static void
+fs_create_blob_resize_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *f = args->file;
+	struct spdk_blob *blob = args->op.create.blob;
+	uint64_t length = 0;
+
+	args->rc = bserrno;
+	if (bserrno) {
+		spdk_blob_close(blob, fs_create_blob_close_cb, args);
+		return;
+	}
+
+	spdk_blob_set_xattr(blob, "name", f->name, strlen(f->name) + 1);
+	spdk_blob_set_xattr(blob, "length", &length, sizeof(length));
+
+	spdk_blob_close(blob, fs_create_blob_close_cb, args);
+}
+
+static void
+fs_create_blob_open_cb(void *ctx, struct spdk_blob *blob, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	if (bserrno) {
+		args->fn.file_op(args->arg, bserrno);
+		free_fs_request(req);
+		return;
+	}
+
+	args->op.create.blob = blob;
+	spdk_blob_resize(blob, 1, fs_create_blob_resize_cb, req);
+}
+
+static void
+fs_create_blob_create_cb(void *ctx, spdk_blob_id blobid, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *f = args->file;
+
+	if (bserrno) {
+		args->fn.file_op(args->arg, bserrno);
+		free_fs_request(req);
+		return;
+	}
+
+	f->blobid = blobid;
+	spdk_bs_open_blob(f->fs->bs, blobid, fs_create_blob_open_cb, req);
+}
+
+void
+spdk_fs_create_file_async(struct spdk_filesystem *fs, const char *name,
+			  spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_file *file;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	if (strnlen(name, SPDK_FILE_NAME_MAX + 1) == SPDK_FILE_NAME_MAX + 1) {
+		cb_fn(cb_arg, -ENAMETOOLONG);
+		return;
+	}
+
+	file = fs_find_file(fs, name);
+	if (file != NULL) {
+		cb_fn(cb_arg, -EEXIST);
+		return;
+	}
+
+	file = file_alloc(fs);
+	if (file == NULL) {
+		SPDK_ERRLOG("Cannot allocate new file for creation\n");
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate create async req for file=%s\n", name);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->file = file;
+	args->fn.file_op = cb_fn;
+	args->arg = cb_arg;
+
+	file->name = strdup(name);
+	_file_build_trace_arg_name(file);
+	spdk_bs_create_blob(fs->bs, fs_create_blob_create_cb, args);
+}
+
+static void
+__fs_create_file_done(void *arg, int fserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	__wake_caller(args, fserrno);
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", args->op.create.name);
+}
+
+static void
+__fs_create_file(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", args->op.create.name);
+	spdk_fs_create_file_async(args->fs, args->op.create.name, __fs_create_file_done, req);
+}
+
+int
+spdk_fs_create_file(struct spdk_filesystem *fs, struct spdk_fs_thread_ctx *ctx, const char *name)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	int rc;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", name);
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate req to create file=%s\n", name);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+	args->fs = fs;
+	args->op.create.name = name;
+	args->sem = &channel->sem;
+	fs->send_request(__fs_create_file, req);
+	sem_wait(&channel->sem);
+	rc = args->rc;
+	free_fs_request(req);
+
+	return rc;
+}
+
+static void
+fs_open_blob_done(void *ctx, struct spdk_blob *blob, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *f = args->file;
+
+	f->blob = blob;
+	while (!TAILQ_EMPTY(&f->open_requests)) {
+		req = TAILQ_FIRST(&f->open_requests);
+		args = &req->args;
+		TAILQ_REMOVE(&f->open_requests, req, args.op.open.tailq);
+		spdk_trace_record(TRACE_BLOBFS_OPEN, 0, 0, 0, f->trace_arg_name);
+		args->fn.file_op_with_handle(args->arg, f, bserrno);
+		free_fs_request(req);
+	}
+}
+
+static void
+fs_open_blob_create_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+	struct spdk_filesystem *fs = args->fs;
+
+	if (file == NULL) {
+		/*
+		 * This is from an open with CREATE flag - the file
+		 *  is now created so look it up in the file list for this
+		 *  filesystem.
+		 */
+		file = fs_find_file(fs, args->op.open.name);
+		assert(file != NULL);
+		args->file = file;
+	}
+
+	file->ref_count++;
+	TAILQ_INSERT_TAIL(&file->open_requests, req, args.op.open.tailq);
+	if (file->ref_count == 1) {
+		assert(file->blob == NULL);
+		spdk_bs_open_blob(fs->bs, file->blobid, fs_open_blob_done, req);
+	} else if (file->blob != NULL) {
+		fs_open_blob_done(req, file->blob, 0);
+	} else {
+		/*
+		 * The blob open for this file is in progress due to a previous
+		 *  open request.  When that open completes, it will invoke the
+		 *  open callback for this request.
+		 */
+	}
+}
+
+void
+spdk_fs_open_file_async(struct spdk_filesystem *fs, const char *name, uint32_t flags,
+			spdk_file_op_with_handle_complete cb_fn, void *cb_arg)
+{
+	struct spdk_file *f = NULL;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	if (strnlen(name, SPDK_FILE_NAME_MAX + 1) == SPDK_FILE_NAME_MAX + 1) {
+		cb_fn(cb_arg, NULL, -ENAMETOOLONG);
+		return;
+	}
+
+	f = fs_find_file(fs, name);
+	if (f == NULL && !(flags & SPDK_BLOBFS_OPEN_CREATE)) {
+		cb_fn(cb_arg, NULL, -ENOENT);
+		return;
+	}
+
+	if (f != NULL && f->is_deleted == true) {
+		cb_fn(cb_arg, NULL, -ENOENT);
+		return;
+	}
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate async open req for file=%s\n", name);
+		cb_fn(cb_arg, NULL, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.file_op_with_handle = cb_fn;
+	args->arg = cb_arg;
+	args->file = f;
+	args->fs = fs;
+	args->op.open.name = name;
+
+	if (f == NULL) {
+		spdk_fs_create_file_async(fs, name, fs_open_blob_create_cb, req);
+	} else {
+		fs_open_blob_create_cb(req, 0);
+	}
+}
+
+static void
+__fs_open_file_done(void *arg, struct spdk_file *file, int bserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	args->file = file;
+	__wake_caller(args, bserrno);
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", args->op.open.name);
+}
+
+static void
+__fs_open_file(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", args->op.open.name);
+	spdk_fs_open_file_async(args->fs, args->op.open.name, args->op.open.flags,
+				__fs_open_file_done, req);
+}
+
+int
+spdk_fs_open_file(struct spdk_filesystem *fs, struct spdk_fs_thread_ctx *ctx,
+		  const char *name, uint32_t flags, struct spdk_file **file)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	int rc;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", name);
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate req for opening file=%s\n", name);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+	args->fs = fs;
+	args->op.open.name = name;
+	args->op.open.flags = flags;
+	args->sem = &channel->sem;
+	fs->send_request(__fs_open_file, req);
+	sem_wait(&channel->sem);
+	rc = args->rc;
+	if (rc == 0) {
+		*file = args->file;
+	} else {
+		*file = NULL;
+	}
+	free_fs_request(req);
+
+	return rc;
+}
+
+static void
+fs_rename_blob_close_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	args->fn.fs_op(args->arg, bserrno);
+	free_fs_request(req);
+}
+
+static void
+fs_rename_blob_open_cb(void *ctx, struct spdk_blob *blob, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	const char *new_name = args->op.rename.new_name;
+
+	spdk_blob_set_xattr(blob, "name", new_name, strlen(new_name) + 1);
+	spdk_blob_close(blob, fs_rename_blob_close_cb, req);
+}
+
+static void
+_fs_md_rename_file(struct spdk_fs_request *req)
+{
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *f;
+
+	f = fs_find_file(args->fs, args->op.rename.old_name);
+	if (f == NULL) {
+		args->fn.fs_op(args->arg, -ENOENT);
+		free_fs_request(req);
+		return;
+	}
+
+	free(f->name);
+	f->name = strdup(args->op.rename.new_name);
+	_file_build_trace_arg_name(f);
+	args->file = f;
+	spdk_bs_open_blob(args->fs->bs, f->blobid, fs_rename_blob_open_cb, req);
+}
+
+static void
+fs_rename_delete_done(void *arg, int fserrno)
+{
+	_fs_md_rename_file(arg);
+}
+
+void
+spdk_fs_rename_file_async(struct spdk_filesystem *fs,
+			  const char *old_name, const char *new_name,
+			  spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_file *f;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "old=%s new=%s\n", old_name, new_name);
+	if (strnlen(new_name, SPDK_FILE_NAME_MAX + 1) == SPDK_FILE_NAME_MAX + 1) {
+		cb_fn(cb_arg, -ENAMETOOLONG);
+		return;
+	}
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate rename async req for renaming file from %s to %s\n", old_name,
+			    new_name);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.fs_op = cb_fn;
+	args->fs = fs;
+	args->arg = cb_arg;
+	args->op.rename.old_name = old_name;
+	args->op.rename.new_name = new_name;
+
+	f = fs_find_file(fs, new_name);
+	if (f == NULL) {
+		_fs_md_rename_file(req);
+		return;
+	}
+
+	/*
+	 * The rename overwrites an existing file.  So delete the existing file, then
+	 *  do the actual rename.
+	 */
+	spdk_fs_delete_file_async(fs, new_name, fs_rename_delete_done, req);
+}
+
+static void
+__fs_rename_file_done(void *arg, int fserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	__wake_caller(args, fserrno);
+}
+
+static void
+__fs_rename_file(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_fs_rename_file_async(args->fs, args->op.rename.old_name, args->op.rename.new_name,
+				  __fs_rename_file_done, req);
+}
+
+int
+spdk_fs_rename_file(struct spdk_filesystem *fs, struct spdk_fs_thread_ctx *ctx,
+		    const char *old_name, const char *new_name)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	int rc;
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate rename req for file=%s\n", old_name);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+
+	args->fs = fs;
+	args->op.rename.old_name = old_name;
+	args->op.rename.new_name = new_name;
+	args->sem = &channel->sem;
+	fs->send_request(__fs_rename_file, req);
+	sem_wait(&channel->sem);
+	rc = args->rc;
+	free_fs_request(req);
+	return rc;
+}
+
+static void
+blob_delete_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	args->fn.file_op(args->arg, bserrno);
+	free_fs_request(req);
+}
+
+void
+spdk_fs_delete_file_async(struct spdk_filesystem *fs, const char *name,
+			  spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_file *f;
+	spdk_blob_id blobid;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s\n", name);
+
+	if (strnlen(name, SPDK_FILE_NAME_MAX + 1) == SPDK_FILE_NAME_MAX + 1) {
+		cb_fn(cb_arg, -ENAMETOOLONG);
+		return;
+	}
+
+	f = fs_find_file(fs, name);
+	if (f == NULL) {
+		SPDK_ERRLOG("Cannot find the file=%s to deleted\n", name);
+		cb_fn(cb_arg, -ENOENT);
+		return;
+	}
+
+	req = alloc_fs_request(fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate the req for the file=%s to deleted\n", name);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.file_op = cb_fn;
+	args->arg = cb_arg;
+
+	if (f->ref_count > 0) {
+		/* If the ref > 0, we mark the file as deleted and delete it when we close it. */
+		f->is_deleted = true;
+		spdk_blob_set_xattr(f->blob, "is_deleted", &f->is_deleted, sizeof(bool));
+		spdk_blob_sync_md(f->blob, blob_delete_cb, req);
+		return;
+	}
+
+	blobid = f->blobid;
+	TAILQ_REMOVE(&fs->files, f, tailq);
+
+	file_free(f);
+
+	spdk_bs_delete_blob(fs->bs, blobid, blob_delete_cb, req);
+}
+
+static uint64_t
+fs_name_to_uint64(const char *name)
+{
+	uint64_t result = 0;
+	memcpy(&result, name, spdk_min(sizeof(result), strlen(name)));
+	return result;
+}
+
+static void
+__fs_delete_file_done(void *arg, int fserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_trace_record(TRACE_BLOBFS_DELETE_DONE, 0, 0, 0, fs_name_to_uint64(args->op.delete.name));
+	__wake_caller(args, fserrno);
+}
+
+static void
+__fs_delete_file(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_trace_record(TRACE_BLOBFS_DELETE_START, 0, 0, 0, fs_name_to_uint64(args->op.delete.name));
+	spdk_fs_delete_file_async(args->fs, args->op.delete.name, __fs_delete_file_done, req);
+}
+
+int
+spdk_fs_delete_file(struct spdk_filesystem *fs, struct spdk_fs_thread_ctx *ctx,
+		    const char *name)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	int rc;
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "Cannot allocate req to delete file=%s\n", name);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+	args->fs = fs;
+	args->op.delete.name = name;
+	args->sem = &channel->sem;
+	fs->send_request(__fs_delete_file, req);
+	sem_wait(&channel->sem);
+	rc = args->rc;
+	free_fs_request(req);
+
+	return rc;
+}
+
+spdk_fs_iter
+spdk_fs_iter_first(struct spdk_filesystem *fs)
+{
+	struct spdk_file *f;
+
+	f = TAILQ_FIRST(&fs->files);
+	return f;
+}
+
+spdk_fs_iter
+spdk_fs_iter_next(spdk_fs_iter iter)
+{
+	struct spdk_file *f = iter;
+
+	if (f == NULL) {
+		return NULL;
+	}
+
+	f = TAILQ_NEXT(f, tailq);
+	return f;
+}
+
+const char *
+spdk_file_get_name(struct spdk_file *file)
+{
+	return file->name;
+}
+
+uint64_t
+spdk_file_get_length(struct spdk_file *file)
+{
+	uint64_t length;
+
+	assert(file != NULL);
+
+	length = file->append_pos >= file->length ? file->append_pos : file->length;
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s length=0x%jx\n", file->name, length);
+	return length;
+}
+
+static void
+fs_truncate_complete_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	args->fn.file_op(args->arg, bserrno);
+	free_fs_request(req);
+}
+
+static void
+fs_truncate_resize_cb(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+	uint64_t *length = &args->op.truncate.length;
+
+	if (bserrno) {
+		args->fn.file_op(args->arg, bserrno);
+		free_fs_request(req);
+		return;
+	}
+
+	spdk_blob_set_xattr(file->blob, "length", length, sizeof(*length));
+
+	file->length = *length;
+	if (file->append_pos > file->length) {
+		file->append_pos = file->length;
+	}
+
+	spdk_blob_sync_md(file->blob, fs_truncate_complete_cb, req);
+}
+
+static uint64_t
+__bytes_to_clusters(uint64_t length, uint64_t cluster_sz)
+{
+	return (length + cluster_sz - 1) / cluster_sz;
+}
+
+void
+spdk_file_truncate_async(struct spdk_file *file, uint64_t length,
+			 spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_filesystem *fs;
+	size_t num_clusters;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s old=0x%jx new=0x%jx\n", file->name, file->length, length);
+	if (length == file->length) {
+		cb_fn(cb_arg, 0);
+		return;
+	}
+
+	req = alloc_fs_request(file->fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->fn.file_op = cb_fn;
+	args->arg = cb_arg;
+	args->file = file;
+	args->op.truncate.length = length;
+	fs = file->fs;
+
+	num_clusters = __bytes_to_clusters(length, fs->bs_opts.cluster_sz);
+
+	spdk_blob_resize(file->blob, num_clusters, fs_truncate_resize_cb, req);
+}
+
+static void
+__truncate(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_file_truncate_async(args->file, args->op.truncate.length,
+				 args->fn.file_op, args);
+}
+
+int
+spdk_file_truncate(struct spdk_file *file, struct spdk_fs_thread_ctx *ctx,
+		   uint64_t length)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	int rc;
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+
+	args->file = file;
+	args->op.truncate.length = length;
+	args->fn.file_op = __wake_caller;
+	args->sem = &channel->sem;
+
+	channel->send_request(__truncate, req);
+	sem_wait(&channel->sem);
+	rc = args->rc;
+	free_fs_request(req);
+
+	return rc;
+}
+
+static void
+__rw_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	spdk_free(args->op.rw.pin_buf);
+	args->fn.file_op(args->arg, bserrno);
+	free_fs_request(req);
+}
+
+static void
+_copy_iovs_to_buf(void *buf, size_t buf_len, struct iovec *iovs, int iovcnt)
+{
+	int i;
+	size_t len;
+
+	for (i = 0; i < iovcnt; i++) {
+		len = spdk_min(iovs[i].iov_len, buf_len);
+		memcpy(buf, iovs[i].iov_base, len);
+		buf += len;
+		assert(buf_len >= len);
+		buf_len -= len;
+	}
+}
+
+static void
+_copy_buf_to_iovs(struct iovec *iovs, int iovcnt, void *buf, size_t buf_len)
+{
+	int i;
+	size_t len;
+
+	for (i = 0; i < iovcnt; i++) {
+		len = spdk_min(iovs[i].iov_len, buf_len);
+		memcpy(iovs[i].iov_base, buf, len);
+		buf += len;
+		assert(buf_len >= len);
+		buf_len -= len;
+	}
+}
+
+static void
+__read_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	void *buf;
+
+	assert(req != NULL);
+	buf = (void *)((uintptr_t)args->op.rw.pin_buf + (args->op.rw.offset & (args->op.rw.blocklen - 1)));
+	if (args->op.rw.is_read) {
+		_copy_buf_to_iovs(args->iovs, args->iovcnt, buf, args->op.rw.length);
+		__rw_done(req, 0);
+	} else {
+		_copy_iovs_to_buf(buf, args->op.rw.length, args->iovs, args->iovcnt);
+		spdk_blob_io_write(args->file->blob, args->op.rw.channel,
+				   args->op.rw.pin_buf,
+				   args->op.rw.start_lba, args->op.rw.num_lba,
+				   __rw_done, req);
+	}
+}
+
+static void
+__do_blob_read(void *ctx, int fserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	if (fserrno) {
+		__rw_done(req, fserrno);
+		return;
+	}
+	spdk_blob_io_read(args->file->blob, args->op.rw.channel,
+			  args->op.rw.pin_buf,
+			  args->op.rw.start_lba, args->op.rw.num_lba,
+			  __read_done, req);
+}
+
+static void
+__get_page_parameters(struct spdk_file *file, uint64_t offset, uint64_t length,
+		      uint64_t *start_lba, uint32_t *lba_size, uint64_t *num_lba)
+{
+	uint64_t end_lba;
+
+	*lba_size = spdk_bs_get_io_unit_size(file->fs->bs);
+	*start_lba = offset / *lba_size;
+	end_lba = (offset + length - 1) / *lba_size;
+	*num_lba = (end_lba - *start_lba + 1);
+}
+
+static bool
+__is_lba_aligned(struct spdk_file *file, uint64_t offset, uint64_t length)
+{
+	uint32_t lba_size = spdk_bs_get_io_unit_size(file->fs->bs);
+
+	if ((offset % lba_size == 0) && (length % lba_size == 0)) {
+		return true;
+	}
+
+	return false;
+}
+
+static void
+_fs_request_setup_iovs(struct spdk_fs_request *req, struct iovec *iovs, uint32_t iovcnt)
+{
+	uint32_t i;
+
+	for (i = 0; i < iovcnt; i++) {
+		req->args.iovs[i].iov_base = iovs[i].iov_base;
+		req->args.iovs[i].iov_len = iovs[i].iov_len;
+	}
+}
+
+static void
+__readvwritev(struct spdk_file *file, struct spdk_io_channel *_channel,
+	      struct iovec *iovs, uint32_t iovcnt, uint64_t offset, uint64_t length,
+	      spdk_file_op_complete cb_fn, void *cb_arg, int is_read)
+{
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+	struct spdk_fs_channel *channel = spdk_io_channel_get_ctx(_channel);
+	uint64_t start_lba, num_lba, pin_buf_length;
+	uint32_t lba_size;
+
+	if (is_read && offset + length > file->length) {
+		cb_fn(cb_arg, -EINVAL);
+		return;
+	}
+
+	req = alloc_fs_request_with_iov(channel, iovcnt);
+	if (req == NULL) {
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	__get_page_parameters(file, offset, length, &start_lba, &lba_size, &num_lba);
+
+	args = &req->args;
+	args->fn.file_op = cb_fn;
+	args->arg = cb_arg;
+	args->file = file;
+	args->op.rw.channel = channel->bs_channel;
+	_fs_request_setup_iovs(req, iovs, iovcnt);
+	args->op.rw.is_read = is_read;
+	args->op.rw.offset = offset;
+	args->op.rw.blocklen = lba_size;
+
+	pin_buf_length = num_lba * lba_size;
+	args->op.rw.length = pin_buf_length;
+	args->op.rw.pin_buf = spdk_malloc(pin_buf_length, lba_size, NULL,
+					  SPDK_ENV_SOCKET_ID_ANY, SPDK_MALLOC_DMA);
+	if (args->op.rw.pin_buf == NULL) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "Failed to allocate buf for: file=%s offset=%jx length=%jx\n",
+			      file->name, offset, length);
+		free_fs_request(req);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args->op.rw.start_lba = start_lba;
+	args->op.rw.num_lba = num_lba;
+
+	if (!is_read && file->length < offset + length) {
+		spdk_file_truncate_async(file, offset + length, __do_blob_read, req);
+	} else if (!is_read && __is_lba_aligned(file, offset, length)) {
+		_copy_iovs_to_buf(args->op.rw.pin_buf, args->op.rw.length, args->iovs, args->iovcnt);
+		spdk_blob_io_write(args->file->blob, args->op.rw.channel,
+				   args->op.rw.pin_buf,
+				   args->op.rw.start_lba, args->op.rw.num_lba,
+				   __rw_done, req);
+	} else {
+		__do_blob_read(req, 0);
+	}
+}
+
+static void
+__readwrite(struct spdk_file *file, struct spdk_io_channel *channel,
+	    void *payload, uint64_t offset, uint64_t length,
+	    spdk_file_op_complete cb_fn, void *cb_arg, int is_read)
+{
+	struct iovec iov;
+
+	iov.iov_base = payload;
+	iov.iov_len = (size_t)length;
+
+	__readvwritev(file, channel, &iov, 1, offset, length, cb_fn, cb_arg, is_read);
+}
+
+void
+spdk_file_write_async(struct spdk_file *file, struct spdk_io_channel *channel,
+		      void *payload, uint64_t offset, uint64_t length,
+		      spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	__readwrite(file, channel, payload, offset, length, cb_fn, cb_arg, 0);
+}
+
+void
+spdk_file_writev_async(struct spdk_file *file, struct spdk_io_channel *channel,
+		       struct iovec *iovs, uint32_t iovcnt, uint64_t offset, uint64_t length,
+		       spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s offset=%jx length=%jx\n",
+		      file->name, offset, length);
+
+	__readvwritev(file, channel, iovs, iovcnt, offset, length, cb_fn, cb_arg, 0);
+}
+
+void
+spdk_file_read_async(struct spdk_file *file, struct spdk_io_channel *channel,
+		     void *payload, uint64_t offset, uint64_t length,
+		     spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s offset=%jx length=%jx\n",
+		      file->name, offset, length);
+	__readwrite(file, channel, payload, offset, length, cb_fn, cb_arg, 1);
+}
+
+void
+spdk_file_readv_async(struct spdk_file *file, struct spdk_io_channel *channel,
+		      struct iovec *iovs, uint32_t iovcnt, uint64_t offset, uint64_t length,
+		      spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "file=%s offset=%jx length=%jx\n",
+		      file->name, offset, length);
+
+	__readvwritev(file, channel, iovs, iovcnt, offset, length, cb_fn, cb_arg, 1);
+}
+
+struct spdk_io_channel *
+spdk_fs_alloc_io_channel(struct spdk_filesystem *fs)
+{
+	struct spdk_io_channel *io_channel;
+	struct spdk_fs_channel *fs_channel;
+
+	io_channel = spdk_get_io_channel(&fs->io_target);
+	fs_channel = spdk_io_channel_get_ctx(io_channel);
+	fs_channel->bs_channel = spdk_bs_alloc_io_channel(fs->bs);
+	fs_channel->send_request = __send_request_direct;
+
+	return io_channel;
+}
+
+void
+spdk_fs_free_io_channel(struct spdk_io_channel *channel)
+{
+	spdk_put_io_channel(channel);
+}
+
+struct spdk_fs_thread_ctx *
+spdk_fs_alloc_thread_ctx(struct spdk_filesystem *fs)
+{
+	struct spdk_fs_thread_ctx *ctx;
+
+	ctx = calloc(1, sizeof(*ctx));
+	if (!ctx) {
+		return NULL;
+	}
+
+	if (pthread_spin_init(&ctx->ch.lock, 0)) {
+		free(ctx);
+		return NULL;
+	}
+
+	fs_channel_create(fs, &ctx->ch, 512);
+
+	ctx->ch.send_request = fs->send_request;
+	ctx->ch.sync = 1;
+
+	return ctx;
+}
+
+
+void
+spdk_fs_free_thread_ctx(struct spdk_fs_thread_ctx *ctx)
+{
+	assert(ctx->ch.sync == 1);
+
+	while (true) {
+		pthread_spin_lock(&ctx->ch.lock);
+		if (ctx->ch.outstanding_reqs == 0) {
+			pthread_spin_unlock(&ctx->ch.lock);
+			break;
+		}
+		pthread_spin_unlock(&ctx->ch.lock);
+		usleep(1000);
+	}
+
+	fs_channel_destroy(NULL, &ctx->ch);
+	free(ctx);
+}
+
+int
+spdk_fs_set_cache_size(uint64_t size_in_mb)
+{
+	/* setting g_fs_cache_size is only permitted if cache pool
+	 * is already freed or hasn't been initialized
+	 */
+	if (g_cache_pool != NULL) {
+		return -EPERM;
+	}
+
+	g_fs_cache_size = size_in_mb * 1024 * 1024;
+
+	return 0;
+}
+
+uint64_t
+spdk_fs_get_cache_size(void)
+{
+	return g_fs_cache_size / (1024 * 1024);
+}
+
+static void __file_flush(void *ctx);
+
+/* Try to free some cache buffers from this file.
+ */
+static int
+reclaim_cache_buffers(struct spdk_file *file)
+{
+	int rc;
+
+	BLOBFS_TRACE(file, "free=%s\n", file->name);
+
+	/* The function is safe to be called with any threads, while the file
+	 * lock maybe locked by other thread for now, so try to get the file
+	 * lock here.
+	 */
+	rc = pthread_spin_trylock(&file->lock);
+	if (rc != 0) {
+		return -1;
+	}
+
+	if (file->tree->present_mask == 0) {
+		pthread_spin_unlock(&file->lock);
+		return -1;
+	}
+	tree_free_buffers(file->tree);
+
+	TAILQ_REMOVE(&g_caches, file, cache_tailq);
+	/* If not freed, put it in the end of the queue */
+	if (file->tree->present_mask != 0) {
+		TAILQ_INSERT_TAIL(&g_caches, file, cache_tailq);
+	} else {
+		file->last = NULL;
+	}
+	pthread_spin_unlock(&file->lock);
+
+	return 0;
+}
+
+static int
+_blobfs_cache_pool_reclaim(void *arg)
+{
+	struct spdk_file *file, *tmp;
+	int rc;
+
+	if (!blobfs_cache_pool_need_reclaim()) {
+		return SPDK_POLLER_IDLE;
+	}
+
+	TAILQ_FOREACH_SAFE(file, &g_caches, cache_tailq, tmp) {
+		if (!file->open_for_writing &&
+		    file->priority == SPDK_FILE_PRIORITY_LOW) {
+			rc = reclaim_cache_buffers(file);
+			if (rc < 0) {
+				continue;
+			}
+			if (!blobfs_cache_pool_need_reclaim()) {
+				return SPDK_POLLER_BUSY;
+			}
+			break;
+		}
+	}
+
+	TAILQ_FOREACH_SAFE(file, &g_caches, cache_tailq, tmp) {
+		if (!file->open_for_writing) {
+			rc = reclaim_cache_buffers(file);
+			if (rc < 0) {
+				continue;
+			}
+			if (!blobfs_cache_pool_need_reclaim()) {
+				return SPDK_POLLER_BUSY;
+			}
+			break;
+		}
+	}
+
+	TAILQ_FOREACH_SAFE(file, &g_caches, cache_tailq, tmp) {
+		rc = reclaim_cache_buffers(file);
+		if (rc < 0) {
+			continue;
+		}
+		break;
+	}
+
+	return SPDK_POLLER_BUSY;
+}
+
+static void
+_add_file_to_cache_pool(void *ctx)
+{
+	struct spdk_file *file = ctx;
+
+	TAILQ_INSERT_TAIL(&g_caches, file, cache_tailq);
+}
+
+static void
+_remove_file_from_cache_pool(void *ctx)
+{
+	struct spdk_file *file = ctx;
+
+	TAILQ_REMOVE(&g_caches, file, cache_tailq);
+}
+
+static struct cache_buffer *
+cache_insert_buffer(struct spdk_file *file, uint64_t offset)
+{
+	struct cache_buffer *buf;
+	int count = 0;
+	bool need_update = false;
+
+	buf = calloc(1, sizeof(*buf));
+	if (buf == NULL) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "calloc failed\n");
+		return NULL;
+	}
+
+	do {
+		buf->buf = spdk_mempool_get(g_cache_pool);
+		if (buf->buf) {
+			break;
+		}
+		if (count++ == 100) {
+			SPDK_ERRLOG("Could not allocate cache buffer for file=%p on offset=%jx\n",
+				    file, offset);
+			free(buf);
+			return NULL;
+		}
+		usleep(BLOBFS_CACHE_POOL_POLL_PERIOD_IN_US);
+	} while (true);
+
+	buf->buf_size = CACHE_BUFFER_SIZE;
+	buf->offset = offset;
+
+	if (file->tree->present_mask == 0) {
+		need_update = true;
+	}
+	file->tree = tree_insert_buffer(file->tree, buf);
+
+	if (need_update) {
+		spdk_thread_send_msg(g_cache_pool_thread, _add_file_to_cache_pool, file);
+	}
+
+	return buf;
+}
+
+static struct cache_buffer *
+cache_append_buffer(struct spdk_file *file)
+{
+	struct cache_buffer *last;
+
+	assert(file->last == NULL || file->last->bytes_filled == file->last->buf_size);
+	assert((file->append_pos % CACHE_BUFFER_SIZE) == 0);
+
+	last = cache_insert_buffer(file, file->append_pos);
+	if (last == NULL) {
+		SPDK_DEBUGLOG(SPDK_LOG_BLOBFS, "cache_insert_buffer failed\n");
+		return NULL;
+	}
+
+	file->last = last;
+
+	return last;
+}
+
+static void __check_sync_reqs(struct spdk_file *file);
+
+static void
+__file_cache_finish_sync(void *ctx, int bserrno)
+{
+	struct spdk_file *file;
+	struct spdk_fs_request *sync_req = ctx;
+	struct spdk_fs_cb_args *sync_args;
+
+	sync_args = &sync_req->args;
+	file = sync_args->file;
+	pthread_spin_lock(&file->lock);
+	file->length_xattr = sync_args->op.sync.length;
+	assert(sync_args->op.sync.offset <= file->length_flushed);
+	spdk_trace_record(TRACE_BLOBFS_XATTR_END, 0, sync_args->op.sync.offset,
+			  0, file->trace_arg_name);
+	BLOBFS_TRACE(file, "sync done offset=%jx\n", sync_args->op.sync.offset);
+	TAILQ_REMOVE(&file->sync_requests, sync_req, args.op.sync.tailq);
+	pthread_spin_unlock(&file->lock);
+
+	sync_args->fn.file_op(sync_args->arg, bserrno);
+
+	free_fs_request(sync_req);
+	__check_sync_reqs(file);
+}
+
+static void
+__check_sync_reqs(struct spdk_file *file)
+{
+	struct spdk_fs_request *sync_req;
+
+	pthread_spin_lock(&file->lock);
+
+	TAILQ_FOREACH(sync_req, &file->sync_requests, args.op.sync.tailq) {
+		if (sync_req->args.op.sync.offset <= file->length_flushed) {
+			break;
+		}
+	}
+
+	if (sync_req != NULL && !sync_req->args.op.sync.xattr_in_progress) {
+		BLOBFS_TRACE(file, "set xattr length 0x%jx\n", file->length_flushed);
+		sync_req->args.op.sync.xattr_in_progress = true;
+		sync_req->args.op.sync.length = file->length_flushed;
+		spdk_blob_set_xattr(file->blob, "length", &file->length_flushed,
+				    sizeof(file->length_flushed));
+
+		pthread_spin_unlock(&file->lock);
+		spdk_trace_record(TRACE_BLOBFS_XATTR_START, 0, file->length_flushed,
+				  0, file->trace_arg_name);
+		spdk_blob_sync_md(file->blob, __file_cache_finish_sync, sync_req);
+	} else {
+		pthread_spin_unlock(&file->lock);
+	}
+}
+
+static void
+__file_flush_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+	struct cache_buffer *next = args->op.flush.cache_buffer;
+
+	BLOBFS_TRACE(file, "length=%jx\n", args->op.flush.length);
+
+	pthread_spin_lock(&file->lock);
+	next->in_progress = false;
+	next->bytes_flushed += args->op.flush.length;
+	file->length_flushed += args->op.flush.length;
+	if (file->length_flushed > file->length) {
+		file->length = file->length_flushed;
+	}
+	if (next->bytes_flushed == next->buf_size) {
+		BLOBFS_TRACE(file, "write buffer fully flushed 0x%jx\n", file->length_flushed);
+		next = tree_find_buffer(file->tree, file->length_flushed);
+	}
+
+	/*
+	 * Assert that there is no cached data that extends past the end of the underlying
+	 *  blob.
+	 */
+	assert(next == NULL || next->offset < __file_get_blob_size(file) ||
+	       next->bytes_filled == 0);
+
+	pthread_spin_unlock(&file->lock);
+
+	__check_sync_reqs(file);
+
+	__file_flush(req);
+}
+
+static void
+__file_flush(void *ctx)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+	struct cache_buffer *next;
+	uint64_t offset, length, start_lba, num_lba;
+	uint32_t lba_size;
+
+	pthread_spin_lock(&file->lock);
+	next = tree_find_buffer(file->tree, file->length_flushed);
+	if (next == NULL || next->in_progress ||
+	    ((next->bytes_filled < next->buf_size) && TAILQ_EMPTY(&file->sync_requests))) {
+		/*
+		 * There is either no data to flush, a flush I/O is already in
+		 *  progress, or the next buffer is partially filled but there's no
+		 *  outstanding request to sync it.
+		 * So return immediately - if a flush I/O is in progress we will flush
+		 *  more data after that is completed, or a partial buffer will get flushed
+		 *  when it is either filled or the file is synced.
+		 */
+		free_fs_request(req);
+		if (next == NULL) {
+			/*
+			 * For cases where a file's cache was evicted, and then the
+			 *  file was later appended, we will write the data directly
+			 *  to disk and bypass cache.  So just update length_flushed
+			 *  here to reflect that all data was already written to disk.
+			 */
+			file->length_flushed = file->append_pos;
+		}
+		pthread_spin_unlock(&file->lock);
+		if (next == NULL) {
+			/*
+			 * There is no data to flush, but we still need to check for any
+			 *  outstanding sync requests to make sure metadata gets updated.
+			 */
+			__check_sync_reqs(file);
+		}
+		return;
+	}
+
+	offset = next->offset + next->bytes_flushed;
+	length = next->bytes_filled - next->bytes_flushed;
+	if (length == 0) {
+		free_fs_request(req);
+		pthread_spin_unlock(&file->lock);
+		/*
+		 * There is no data to flush, but we still need to check for any
+		 *  outstanding sync requests to make sure metadata gets updated.
+		 */
+		__check_sync_reqs(file);
+		return;
+	}
+	args->op.flush.length = length;
+	args->op.flush.cache_buffer = next;
+
+	__get_page_parameters(file, offset, length, &start_lba, &lba_size, &num_lba);
+
+	next->in_progress = true;
+	BLOBFS_TRACE(file, "offset=0x%jx length=0x%jx page start=0x%jx num=0x%jx\n",
+		     offset, length, start_lba, num_lba);
+	pthread_spin_unlock(&file->lock);
+	spdk_blob_io_write(file->blob, file->fs->sync_target.sync_fs_channel->bs_channel,
+			   next->buf + (start_lba * lba_size) - next->offset,
+			   start_lba, num_lba, __file_flush_done, req);
+}
+
+static void
+__file_extend_done(void *arg, int bserrno)
+{
+	struct spdk_fs_cb_args *args = arg;
+
+	__wake_caller(args, bserrno);
+}
+
+static void
+__file_extend_resize_cb(void *_args, int bserrno)
+{
+	struct spdk_fs_cb_args *args = _args;
+	struct spdk_file *file = args->file;
+
+	if (bserrno) {
+		__wake_caller(args, bserrno);
+		return;
+	}
+
+	spdk_blob_sync_md(file->blob, __file_extend_done, args);
+}
+
+static void
+__file_extend_blob(void *_args)
+{
+	struct spdk_fs_cb_args *args = _args;
+	struct spdk_file *file = args->file;
+
+	spdk_blob_resize(file->blob, args->op.resize.num_clusters, __file_extend_resize_cb, args);
+}
+
+static void
+__rw_from_file_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+
+	__wake_caller(&req->args, bserrno);
+	free_fs_request(req);
+}
+
+static void
+__rw_from_file(void *ctx)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+
+	if (args->op.rw.is_read) {
+		spdk_file_read_async(file, file->fs->sync_target.sync_io_channel, args->iovs[0].iov_base,
+				     args->op.rw.offset, (uint64_t)args->iovs[0].iov_len,
+				     __rw_from_file_done, req);
+	} else {
+		spdk_file_write_async(file, file->fs->sync_target.sync_io_channel, args->iovs[0].iov_base,
+				      args->op.rw.offset, (uint64_t)args->iovs[0].iov_len,
+				      __rw_from_file_done, req);
+	}
+}
+
+static int
+__send_rw_from_file(struct spdk_file *file, void *payload,
+		    uint64_t offset, uint64_t length, bool is_read,
+		    struct spdk_fs_channel *channel)
+{
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	req = alloc_fs_request_with_iov(channel, 1);
+	if (req == NULL) {
+		sem_post(&channel->sem);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+	args->file = file;
+	args->sem = &channel->sem;
+	args->iovs[0].iov_base = payload;
+	args->iovs[0].iov_len = (size_t)length;
+	args->op.rw.offset = offset;
+	args->op.rw.is_read = is_read;
+	file->fs->send_request(__rw_from_file, req);
+	return 0;
+}
+
+int
+spdk_file_write(struct spdk_file *file, struct spdk_fs_thread_ctx *ctx,
+		void *payload, uint64_t offset, uint64_t length)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *flush_req;
+	uint64_t rem_length, copy, blob_size, cluster_sz;
+	uint32_t cache_buffers_filled = 0;
+	uint8_t *cur_payload;
+	struct cache_buffer *last;
+
+	BLOBFS_TRACE_RW(file, "offset=%jx length=%jx\n", offset, length);
+
+	if (length == 0) {
+		return 0;
+	}
+
+	if (offset != file->append_pos) {
+		BLOBFS_TRACE(file, " error offset=%jx append_pos=%jx\n", offset, file->append_pos);
+		return -EINVAL;
+	}
+
+	pthread_spin_lock(&file->lock);
+	file->open_for_writing = true;
+
+	if ((file->last == NULL) && (file->append_pos % CACHE_BUFFER_SIZE == 0)) {
+		cache_append_buffer(file);
+	}
+
+	if (file->last == NULL) {
+		int rc;
+
+		file->append_pos += length;
+		pthread_spin_unlock(&file->lock);
+		rc = __send_rw_from_file(file, payload, offset, length, false, channel);
+		sem_wait(&channel->sem);
+		return rc;
+	}
+
+	blob_size = __file_get_blob_size(file);
+
+	if ((offset + length) > blob_size) {
+		struct spdk_fs_cb_args extend_args = {};
+
+		cluster_sz = file->fs->bs_opts.cluster_sz;
+		extend_args.sem = &channel->sem;
+		extend_args.op.resize.num_clusters = __bytes_to_clusters((offset + length), cluster_sz);
+		extend_args.file = file;
+		BLOBFS_TRACE(file, "start resize to %u clusters\n", extend_args.op.resize.num_clusters);
+		pthread_spin_unlock(&file->lock);
+		file->fs->send_request(__file_extend_blob, &extend_args);
+		sem_wait(&channel->sem);
+		if (extend_args.rc) {
+			return extend_args.rc;
+		}
+	}
+
+	flush_req = alloc_fs_request(channel);
+	if (flush_req == NULL) {
+		pthread_spin_unlock(&file->lock);
+		return -ENOMEM;
+	}
+
+	last = file->last;
+	rem_length = length;
+	cur_payload = payload;
+	while (rem_length > 0) {
+		copy = last->buf_size - last->bytes_filled;
+		if (copy > rem_length) {
+			copy = rem_length;
+		}
+		BLOBFS_TRACE_RW(file, "  fill offset=%jx length=%jx\n", file->append_pos, copy);
+		memcpy(&last->buf[last->bytes_filled], cur_payload, copy);
+		file->append_pos += copy;
+		if (file->length < file->append_pos) {
+			file->length = file->append_pos;
+		}
+		cur_payload += copy;
+		last->bytes_filled += copy;
+		rem_length -= copy;
+		if (last->bytes_filled == last->buf_size) {
+			cache_buffers_filled++;
+			last = cache_append_buffer(file);
+			if (last == NULL) {
+				BLOBFS_TRACE(file, "nomem\n");
+				free_fs_request(flush_req);
+				pthread_spin_unlock(&file->lock);
+				return -ENOMEM;
+			}
+		}
+	}
+
+	pthread_spin_unlock(&file->lock);
+
+	if (cache_buffers_filled == 0) {
+		free_fs_request(flush_req);
+		return 0;
+	}
+
+	flush_req->args.file = file;
+	file->fs->send_request(__file_flush, flush_req);
+	return 0;
+}
+
+static void
+__readahead_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct cache_buffer *cache_buffer = args->op.readahead.cache_buffer;
+	struct spdk_file *file = args->file;
+
+	BLOBFS_TRACE(file, "offset=%jx\n", cache_buffer->offset);
+
+	pthread_spin_lock(&file->lock);
+	cache_buffer->bytes_filled = args->op.readahead.length;
+	cache_buffer->bytes_flushed = args->op.readahead.length;
+	cache_buffer->in_progress = false;
+	pthread_spin_unlock(&file->lock);
+
+	free_fs_request(req);
+}
+
+static void
+__readahead(void *ctx)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+	uint64_t offset, length, start_lba, num_lba;
+	uint32_t lba_size;
+
+	offset = args->op.readahead.offset;
+	length = args->op.readahead.length;
+	assert(length > 0);
+
+	__get_page_parameters(file, offset, length, &start_lba, &lba_size, &num_lba);
+
+	BLOBFS_TRACE(file, "offset=%jx length=%jx page start=%jx num=%jx\n",
+		     offset, length, start_lba, num_lba);
+	spdk_blob_io_read(file->blob, file->fs->sync_target.sync_fs_channel->bs_channel,
+			  args->op.readahead.cache_buffer->buf,
+			  start_lba, num_lba, __readahead_done, req);
+}
+
+static uint64_t
+__next_cache_buffer_offset(uint64_t offset)
+{
+	return (offset + CACHE_BUFFER_SIZE) & ~(CACHE_TREE_LEVEL_MASK(0));
+}
+
+static void
+check_readahead(struct spdk_file *file, uint64_t offset,
+		struct spdk_fs_channel *channel)
+{
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	offset = __next_cache_buffer_offset(offset);
+	if (tree_find_buffer(file->tree, offset) != NULL || file->length <= offset) {
+		return;
+	}
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		return;
+	}
+	args = &req->args;
+
+	BLOBFS_TRACE(file, "offset=%jx\n", offset);
+
+	args->file = file;
+	args->op.readahead.offset = offset;
+	args->op.readahead.cache_buffer = cache_insert_buffer(file, offset);
+	if (!args->op.readahead.cache_buffer) {
+		BLOBFS_TRACE(file, "Cannot allocate buf for offset=%jx\n", offset);
+		free_fs_request(req);
+		return;
+	}
+
+	args->op.readahead.cache_buffer->in_progress = true;
+	if (file->length < (offset + CACHE_BUFFER_SIZE)) {
+		args->op.readahead.length = file->length & (CACHE_BUFFER_SIZE - 1);
+	} else {
+		args->op.readahead.length = CACHE_BUFFER_SIZE;
+	}
+	file->fs->send_request(__readahead, req);
+}
+
+int64_t
+spdk_file_read(struct spdk_file *file, struct spdk_fs_thread_ctx *ctx,
+	       void *payload, uint64_t offset, uint64_t length)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	uint64_t final_offset, final_length;
+	uint32_t sub_reads = 0;
+	struct cache_buffer *buf;
+	uint64_t read_len;
+	int rc = 0;
+
+	pthread_spin_lock(&file->lock);
+
+	BLOBFS_TRACE_RW(file, "offset=%ju length=%ju\n", offset, length);
+
+	file->open_for_writing = false;
+
+	if (length == 0 || offset >= file->append_pos) {
+		pthread_spin_unlock(&file->lock);
+		return 0;
+	}
+
+	if (offset + length > file->append_pos) {
+		length = file->append_pos - offset;
+	}
+
+	if (offset != file->next_seq_offset) {
+		file->seq_byte_count = 0;
+	}
+	file->seq_byte_count += length;
+	file->next_seq_offset = offset + length;
+	if (file->seq_byte_count >= CACHE_READAHEAD_THRESHOLD) {
+		check_readahead(file, offset, channel);
+		check_readahead(file, offset + CACHE_BUFFER_SIZE, channel);
+	}
+
+	final_length = 0;
+	final_offset = offset + length;
+	while (offset < final_offset) {
+		length = NEXT_CACHE_BUFFER_OFFSET(offset) - offset;
+		if (length > (final_offset - offset)) {
+			length = final_offset - offset;
+		}
+
+		buf = tree_find_filled_buffer(file->tree, offset);
+		if (buf == NULL) {
+			pthread_spin_unlock(&file->lock);
+			rc = __send_rw_from_file(file, payload, offset, length, true, channel);
+			pthread_spin_lock(&file->lock);
+			if (rc == 0) {
+				sub_reads++;
+			}
+		} else {
+			read_len = length;
+			if ((offset + length) > (buf->offset + buf->bytes_filled)) {
+				read_len = buf->offset + buf->bytes_filled - offset;
+			}
+			BLOBFS_TRACE(file, "read %p offset=%ju length=%ju\n", payload, offset, read_len);
+			memcpy(payload, &buf->buf[offset - buf->offset], read_len);
+			if ((offset + read_len) % CACHE_BUFFER_SIZE == 0) {
+				tree_remove_buffer(file->tree, buf);
+				if (file->tree->present_mask == 0) {
+					spdk_thread_send_msg(g_cache_pool_thread, _remove_file_from_cache_pool, file);
+				}
+			}
+		}
+
+		if (rc == 0) {
+			final_length += length;
+		} else {
+			break;
+		}
+		payload += length;
+		offset += length;
+	}
+	pthread_spin_unlock(&file->lock);
+	while (sub_reads > 0) {
+		sem_wait(&channel->sem);
+		sub_reads--;
+	}
+	if (rc == 0) {
+		return final_length;
+	} else {
+		return rc;
+	}
+}
+
+static void
+_file_sync(struct spdk_file *file, struct spdk_fs_channel *channel,
+	   spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_fs_request *sync_req;
+	struct spdk_fs_request *flush_req;
+	struct spdk_fs_cb_args *sync_args;
+	struct spdk_fs_cb_args *flush_args;
+
+	BLOBFS_TRACE(file, "offset=%jx\n", file->append_pos);
+
+	pthread_spin_lock(&file->lock);
+	if (file->append_pos <= file->length_xattr) {
+		BLOBFS_TRACE(file, "done - file already synced\n");
+		pthread_spin_unlock(&file->lock);
+		cb_fn(cb_arg, 0);
+		return;
+	}
+
+	sync_req = alloc_fs_request(channel);
+	if (!sync_req) {
+		SPDK_ERRLOG("Cannot allocate sync req for file=%s\n", file->name);
+		pthread_spin_unlock(&file->lock);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+	sync_args = &sync_req->args;
+
+	flush_req = alloc_fs_request(channel);
+	if (!flush_req) {
+		SPDK_ERRLOG("Cannot allocate flush req for file=%s\n", file->name);
+		free_fs_request(sync_req);
+		pthread_spin_unlock(&file->lock);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+	flush_args = &flush_req->args;
+
+	sync_args->file = file;
+	sync_args->fn.file_op = cb_fn;
+	sync_args->arg = cb_arg;
+	sync_args->op.sync.offset = file->append_pos;
+	sync_args->op.sync.xattr_in_progress = false;
+	TAILQ_INSERT_TAIL(&file->sync_requests, sync_req, args.op.sync.tailq);
+	pthread_spin_unlock(&file->lock);
+
+	flush_args->file = file;
+	channel->send_request(__file_flush, flush_req);
+}
+
+int
+spdk_file_sync(struct spdk_file *file, struct spdk_fs_thread_ctx *ctx)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_cb_args args = {};
+
+	args.sem = &channel->sem;
+	_file_sync(file, channel, __wake_caller, &args);
+	sem_wait(&channel->sem);
+
+	return args.rc;
+}
+
+void
+spdk_file_sync_async(struct spdk_file *file, struct spdk_io_channel *_channel,
+		     spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_fs_channel *channel = spdk_io_channel_get_ctx(_channel);
+
+	_file_sync(file, channel, cb_fn, cb_arg);
+}
+
+void
+spdk_file_set_priority(struct spdk_file *file, uint32_t priority)
+{
+	BLOBFS_TRACE(file, "priority=%u\n", priority);
+	file->priority = priority;
+
+}
+
+/*
+ * Close routines
+ */
+
+static void
+__file_close_async_done(void *ctx, int bserrno)
+{
+	struct spdk_fs_request *req = ctx;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+
+	spdk_trace_record(TRACE_BLOBFS_CLOSE, 0, 0, 0, file->trace_arg_name);
+
+	if (file->is_deleted) {
+		spdk_fs_delete_file_async(file->fs, file->name, blob_delete_cb, ctx);
+		return;
+	}
+
+	args->fn.file_op(args->arg, bserrno);
+	free_fs_request(req);
+}
+
+static void
+__file_close_async(struct spdk_file *file, struct spdk_fs_request *req)
+{
+	struct spdk_blob *blob;
+
+	pthread_spin_lock(&file->lock);
+	if (file->ref_count == 0) {
+		pthread_spin_unlock(&file->lock);
+		__file_close_async_done(req, -EBADF);
+		return;
+	}
+
+	file->ref_count--;
+	if (file->ref_count > 0) {
+		pthread_spin_unlock(&file->lock);
+		req->args.fn.file_op(req->args.arg, 0);
+		free_fs_request(req);
+		return;
+	}
+
+	pthread_spin_unlock(&file->lock);
+
+	blob = file->blob;
+	file->blob = NULL;
+	spdk_blob_close(blob, __file_close_async_done, req);
+}
+
+static void
+__file_close_async__sync_done(void *arg, int fserrno)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+
+	__file_close_async(args->file, req);
+}
+
+void
+spdk_file_close_async(struct spdk_file *file, spdk_file_op_complete cb_fn, void *cb_arg)
+{
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	req = alloc_fs_request(file->fs->md_target.md_fs_channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate close async req for file=%s\n", file->name);
+		cb_fn(cb_arg, -ENOMEM);
+		return;
+	}
+
+	args = &req->args;
+	args->file = file;
+	args->fn.file_op = cb_fn;
+	args->arg = cb_arg;
+
+	spdk_file_sync_async(file, file->fs->md_target.md_io_channel, __file_close_async__sync_done, req);
+}
+
+static void
+__file_close(void *arg)
+{
+	struct spdk_fs_request *req = arg;
+	struct spdk_fs_cb_args *args = &req->args;
+	struct spdk_file *file = args->file;
+
+	__file_close_async(file, req);
+}
+
+int
+spdk_file_close(struct spdk_file *file, struct spdk_fs_thread_ctx *ctx)
+{
+	struct spdk_fs_channel *channel = (struct spdk_fs_channel *)ctx;
+	struct spdk_fs_request *req;
+	struct spdk_fs_cb_args *args;
+
+	req = alloc_fs_request(channel);
+	if (req == NULL) {
+		SPDK_ERRLOG("Cannot allocate close req for file=%s\n", file->name);
+		return -ENOMEM;
+	}
+
+	args = &req->args;
+
+	spdk_file_sync(file, ctx);
+	BLOBFS_TRACE(file, "name=%s\n", file->name);
+	args->file = file;
+	args->sem = &channel->sem;
+	args->fn.file_op = __wake_caller;
+	args->arg = args;
+	channel->send_request(__file_close, req);
+	sem_wait(&channel->sem);
+
+	return args->rc;
+}
+
+int
+spdk_file_get_id(struct spdk_file *file, void *id, size_t size)
+{
+	if (size < sizeof(spdk_blob_id)) {
+		return -EINVAL;
+	}
+
+	memcpy(id, &file->blobid, sizeof(spdk_blob_id));
+
+	return sizeof(spdk_blob_id);
+}
+
+static void
+_file_free(void *ctx)
+{
+	struct spdk_file *file = ctx;
+
+	TAILQ_REMOVE(&g_caches, file, cache_tailq);
+
+	free(file->name);
+	free(file->tree);
+	free(file);
+}
+
+static void
+file_free(struct spdk_file *file)
+{
+	BLOBFS_TRACE(file, "free=%s\n", file->name);
+	pthread_spin_lock(&file->lock);
+	if (file->tree->present_mask == 0) {
+		pthread_spin_unlock(&file->lock);
+		free(file->name);
+		free(file->tree);
+		free(file);
+		return;
+	}
+
+	tree_free_buffers(file->tree);
+	assert(file->tree->present_mask == 0);
+	spdk_thread_send_msg(g_cache_pool_thread, _file_free, file);
+	pthread_spin_unlock(&file->lock);
+}
+
+SPDK_LOG_REGISTER_COMPONENT("blobfs", SPDK_LOG_BLOBFS)
+SPDK_LOG_REGISTER_COMPONENT("blobfs_rw", SPDK_LOG_BLOBFS_RW)
diff --git a/src/spdk/lib/blobfs/spdk_blobfs.map b/src/spdk/lib/blobfs/spdk_blobfs.map
new file mode 100644
index 000000000..91c02f61e
--- /dev/null
+++ b/src/spdk/lib/blobfs/spdk_blobfs.map
@@ -0,0 +1,45 @@
+{
+	global:
+
+	# public functions
+	spdk_fs_opts_init;
+	spdk_fs_init;
+	spdk_fs_load;
+	spdk_fs_unload;
+	spdk_fs_alloc_io_channel;
+	spdk_fs_free_io_channel;
+	spdk_fs_alloc_thread_ctx;
+	spdk_fs_free_thread_ctx;
+	spdk_fs_file_stat;
+	spdk_fs_create_file;
+	spdk_fs_open_file;
+	spdk_file_close;
+	spdk_fs_rename_file;
+	spdk_fs_delete_file;
+	spdk_fs_iter_first;
+	spdk_fs_iter_next;
+	spdk_file_truncate;
+	spdk_file_get_name;
+	spdk_file_get_length;
+	spdk_file_write;
+	spdk_file_read;
+	spdk_fs_set_cache_size;
+	spdk_fs_get_cache_size;
+	spdk_file_set_priority;
+	spdk_file_sync;
+	spdk_file_get_id;
+	spdk_file_readv_async;
+	spdk_file_writev_async;
+	spdk_fs_file_stat_async;
+	spdk_fs_create_file_async;
+	spdk_fs_open_file_async;
+	spdk_file_close_async;
+	spdk_fs_rename_file_async;
+	spdk_fs_delete_file_async;
+	spdk_file_truncate_async;
+	spdk_file_write_async;
+	spdk_file_read_async;
+	spdk_file_sync_async;
+
+	local: *;
+};
diff --git a/src/spdk/lib/blobfs/tree.c b/src/spdk/lib/blobfs/tree.c
new file mode 100644
index 000000000..32779766f
--- /dev/null
+++ b/src/spdk/lib/blobfs/tree.c
@@ -0,0 +1,181 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "spdk/stdinc.h"
+
+#include "spdk/blobfs.h"
+#include "tree.h"
+
+#include "spdk/queue.h"
+#include "spdk/assert.h"
+#include "spdk/env.h"
+#include "spdk_internal/log.h"
+
+uint32_t g_fs_cache_buffer_shift = CACHE_BUFFER_SHIFT_DEFAULT;
+
+struct cache_buffer *
+tree_find_buffer(struct cache_tree *tree, uint64_t offset)
+{
+	uint64_t index;
+
+	while (tree != NULL) {
+		index = offset / CACHE_TREE_LEVEL_SIZE(tree->level);
+		if (index >= CACHE_TREE_WIDTH) {
+			return NULL;
+		}
+		if (tree->level == 0) {
+			return tree->u.buffer[index];
+		} else {
+			offset &= CACHE_TREE_LEVEL_MASK(tree->level);
+			tree = tree->u.tree[index];
+		}
+	}
+
+	return NULL;
+}
+
+struct cache_buffer *
+tree_find_filled_buffer(struct cache_tree *tree, uint64_t offset)
+{
+	struct cache_buffer *buf;
+
+	buf = tree_find_buffer(tree, offset);
+	if (buf != NULL && buf->bytes_filled > 0) {
+		return buf;
+	} else {
+		return NULL;
+	}
+}
+
+struct cache_tree *
+tree_insert_buffer(struct cache_tree *root, struct cache_buffer *buffer)
+{
+	struct cache_tree *tree;
+	uint64_t index, offset;
+
+	offset = buffer->offset;
+	while (offset >= CACHE_TREE_LEVEL_SIZE(root->level + 1)) {
+		if (root->present_mask != 0) {
+			tree = calloc(1, sizeof(*tree));
+			tree->level = root->level + 1;
+			tree->u.tree[0] = root;
+			root = tree;
+			root->present_mask = 0x1ULL;
+		} else {
+			root->level++;
+		}
+	}
+
+	tree = root;
+	while (tree->level > 0) {
+		index = offset / CACHE_TREE_LEVEL_SIZE(tree->level);
+		assert(index < CACHE_TREE_WIDTH);
+		offset &= CACHE_TREE_LEVEL_MASK(tree->level);
+		if (tree->u.tree[index] == NULL) {
+			tree->u.tree[index] = calloc(1, sizeof(*tree));
+			tree->u.tree[index]->level = tree->level - 1;
+			tree->present_mask |= (1ULL << index);
+		}
+		tree = tree->u.tree[index];
+	}
+
+	index = offset / CACHE_BUFFER_SIZE;
+	assert(index < CACHE_TREE_WIDTH);
+	assert(tree->u.buffer[index] == NULL);
+	tree->u.buffer[index] = buffer;
+	tree->present_mask |= (1ULL << index);
+	return root;
+}
+
+void
+tree_remove_buffer(struct cache_tree *tree, struct cache_buffer *buffer)
+{
+	struct cache_tree *child;
+	uint64_t index;
+
+	index = CACHE_TREE_INDEX(tree->level, buffer->offset);
+
+	if (tree->level == 0) {
+		assert(tree->u.buffer[index] != NULL);
+		assert(buffer == tree->u.buffer[index]);
+		tree->present_mask &= ~(1ULL << index);
+		tree->u.buffer[index] = NULL;
+		cache_buffer_free(buffer);
+		return;
+	}
+
+	child = tree->u.tree[index];
+	assert(child != NULL);
+	tree_remove_buffer(child, buffer);
+	if (child->present_mask == 0) {
+		tree->present_mask &= ~(1ULL << index);
+		tree->u.tree[index] = NULL;
+		free(child);
+	}
+}
+
+void
+tree_free_buffers(struct cache_tree *tree)
+{
+	struct cache_buffer *buffer;
+	struct cache_tree *child;
+	uint32_t i;
+
+	if (tree->present_mask == 0) {
+		return;
+	}
+
+	if (tree->level == 0) {
+		for (i = 0; i < CACHE_TREE_WIDTH; i++) {
+			buffer = tree->u.buffer[i];
+			if (buffer != NULL && buffer->in_progress == false &&
+			    buffer->bytes_filled == buffer->bytes_flushed) {
+				cache_buffer_free(buffer);
+				tree->u.buffer[i] = NULL;
+				tree->present_mask &= ~(1ULL << i);
+			}
+		}
+	} else {
+		for (i = 0; i < CACHE_TREE_WIDTH; i++) {
+			child = tree->u.tree[i];
+			if (child != NULL) {
+				tree_free_buffers(child);
+				if (child->present_mask == 0) {
+					free(child);
+					tree->u.tree[i] = NULL;
+					tree->present_mask &= ~(1ULL << i);
+				}
+			}
+		}
+	}
+}
diff --git a/src/spdk/lib/blobfs/tree.h b/src/spdk/lib/blobfs/tree.h
new file mode 100644
index 000000000..71df71090
--- /dev/null
+++ b/src/spdk/lib/blobfs/tree.h
@@ -0,0 +1,77 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright (c) Intel Corporation.
+ *   All rights reserved.
+ *
+ *   Redistribution and use in source and binary forms, with or without
+ *   modification, are permitted provided that the following conditions
+ *   are met:
+ *
+ *     * Redistributions of source code must retain the above copyright
+ *       notice, this list of conditions and the following disclaimer.
+ *     * Redistributions in binary form must reproduce the above copyright
+ *       notice, this list of conditions and the following disclaimer in
+ *       the documentation and/or other materials provided with the
+ *       distribution.
+ *     * Neither the name of Intel Corporation nor the names of its
+ *       contributors may be used to endorse or promote products derived
+ *       from this software without specific prior written permission.
+ *
+ *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+ *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+ *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+ *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+ *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+ *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+ *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+ *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#ifndef SPDK_TREE_H_
+#define SPDK_TREE_H_
+
+struct cache_buffer {
+	uint8_t			*buf;
+	uint64_t		offset;
+	uint32_t		buf_size;
+	uint32_t		bytes_filled;
+	uint32_t		bytes_flushed;
+	bool			in_progress;
+};
+
+extern uint32_t g_fs_cache_buffer_shift;
+
+#define CACHE_BUFFER_SHIFT_DEFAULT 18
+#define CACHE_BUFFER_SIZE (1U << g_fs_cache_buffer_shift)
+#define NEXT_CACHE_BUFFER_OFFSET(offset)	\
+	(((offset + CACHE_BUFFER_SIZE) >> g_fs_cache_buffer_shift) << g_fs_cache_buffer_shift)
+
+#define CACHE_TREE_SHIFT 6
+#define CACHE_TREE_WIDTH (1U << CACHE_TREE_SHIFT)
+#define CACHE_TREE_LEVEL_SHIFT(level)	(g_fs_cache_buffer_shift + (level) * CACHE_TREE_SHIFT)
+#define CACHE_TREE_LEVEL_SIZE(level)	(1ULL << CACHE_TREE_LEVEL_SHIFT(level))
+#define CACHE_TREE_LEVEL_MASK(level)	(CACHE_TREE_LEVEL_SIZE(level) - 1)
+#define CACHE_TREE_INDEX(level, offset)	((offset >> CACHE_TREE_LEVEL_SHIFT(level)) & (CACHE_TREE_WIDTH - 1))
+
+struct cache_tree {
+	uint8_t			level;
+	uint64_t		present_mask;
+	union {
+		struct cache_buffer	*buffer[CACHE_TREE_WIDTH];
+		struct cache_tree	*tree[CACHE_TREE_WIDTH];
+	} u;
+};
+
+void cache_buffer_free(struct cache_buffer *cache_buffer);
+
+struct cache_tree *tree_insert_buffer(struct cache_tree *root, struct cache_buffer *buffer);
+void tree_free_buffers(struct cache_tree *tree);
+struct cache_buffer *tree_find_buffer(struct cache_tree *tree, uint64_t offset);
+struct cache_buffer *tree_find_filled_buffer(struct cache_tree *tree, uint64_t offset);
+void tree_remove_buffer(struct cache_tree *tree, struct cache_buffer *buffer);
+
+#endif /* SPDK_TREE_H_ */