From ee7e31f9ae53dfba18bba37acb4462abeb3a95c2 Mon Sep 17 00:00:00 2001
From: paul luse <paul.e.luse@intel.com>
Date: Fri, 3 Jul 2020 10:08:47 -0400
Subject: [PATCH] lib/accel: remove the need for the app to allocate an
 accel_task

This was sort of a clunky interface requiring a couple of inline
functions in every app that wants to use the accel_fw moving
forward. By having the accel_fw public API accept a callback arg
instead of an accel_task combined with adding a pool of accel_tasks
in the accel_fw engine we can eliminate this.

After changing the parm to a cb_arg, changes were made to all accel_fw
interfaces to put cb_fn and cb_arg as the last parms in public and
private function calls.

Related bdev_malloc changes need to be in this patch in order to pass CI.

Signed-off-by: paul luse <paul.e.luse@intel.com>
Change-Id: I2b75764e534562d91484a094c3352266156d8425
Reviewed-on: https://review.spdk.io/gerrit/c/spdk/spdk/+/3209
Community-CI: Mellanox Build Bot
Tested-by: SPDK CI Jenkins <sys_sgci@intel.com>
Reviewed-by: Jim Harris <james.r.harris@intel.com>
Reviewed-by: Ben Walker <benjamin.walker@intel.com>
---
 CHANGELOG.md                          |   4 +
 examples/accel/perf/accel_perf.c      |  60 +++--
 include/spdk/accel_engine.h           | 118 +++++----
 include/spdk_internal/accel_engine.h  |  46 ++--
 lib/accel/accel_engine.c              | 341 ++++++++++++++++----------
 module/accel/idxd/accel_engine_idxd.c |  67 ++---
 module/accel/ioat/accel_engine_ioat.c |  55 ++---
 module/bdev/malloc/bdev_malloc.c      |  34 +--
 8 files changed, 403 insertions(+), 322 deletions(-)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 2839e8d05..e6b9b57e6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,10 @@
 
 ### accel_fw
 
+The accel_fw was updated to no longer require the app to allocate an
+accel_task on its behalf. All public APIs now take a callback arg as
+the parameter that used to be the accel_task.
+
 The accel_fw was updated to support compare, dualcast, crc32c.
 
 The accel_fw introduced batching support for all commands in all plug-ins.
diff --git a/examples/accel/perf/accel_perf.c b/examples/accel/perf/accel_perf.c
index 5144c78df..920b9b26c 100644
--- a/examples/accel/perf/accel_perf.c
+++ b/examples/accel/perf/accel_perf.c
@@ -210,20 +210,18 @@ _submit_single(void *arg1, void *arg2)
 	task->worker->current_queue_depth++;
 	switch (g_workload_selection) {
 	case ACCEL_COPY:
-		rc = spdk_accel_submit_copy(__accel_task_from_ap_task(task),
-					    worker->ch, task->dst,
-					    task->src, g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_submit_copy(worker->ch, task->dst,
+					    task->src, g_xfer_size_bytes, accel_done,
+					    __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_FILL:
 		/* For fill use the first byte of the task->dst buffer */
-		rc = spdk_accel_submit_fill(__accel_task_from_ap_task(task),
-					    worker->ch, task->dst, *(uint8_t *)task->src,
-					    g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_submit_fill(worker->ch, task->dst, *(uint8_t *)task->src,
+					    g_xfer_size_bytes, accel_done, __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_CRC32C:
-		rc = spdk_accel_submit_crc32c(__accel_task_from_ap_task(task),
-					      worker->ch, (uint32_t *)task->dst, task->src, g_crc32c_seed,
-					      g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_submit_crc32c(worker->ch, (uint32_t *)task->dst, task->src, g_crc32c_seed,
+					      g_xfer_size_bytes, accel_done, __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_COMPARE:
 		random_num = rand() % 100;
@@ -234,14 +232,13 @@ _submit_single(void *arg1, void *arg2)
 			task->expected_status = 0;
 			*(uint8_t *)task->dst = DATA_PATTERN;
 		}
-		rc = spdk_accel_submit_compare(__accel_task_from_ap_task(task),
-					       worker->ch, task->dst, task->src,
-					       g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_submit_compare(worker->ch, task->dst, task->src,
+					       g_xfer_size_bytes, accel_done, __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_DUALCAST:
-		rc = spdk_accel_submit_dualcast(__accel_task_from_ap_task(task),
-						worker->ch, task->dst, task->dst2,
-						task->src, g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_submit_dualcast(worker->ch, task->dst, task->dst2,
+						task->src, g_xfer_size_bytes, accel_done,
+						__accel_task_from_ap_task(task));
 		break;
 	default:
 		assert(false);
@@ -470,29 +467,29 @@ _batch_prep_cmd(struct worker_thread *worker, struct ap_task *task, struct spdk_
 
 	switch (g_workload_selection) {
 	case ACCEL_COPY:
-		rc = spdk_accel_batch_prep_copy(__accel_task_from_ap_task(task),
-						worker->ch, batch, task->dst,
-						task->src, g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_batch_prep_copy(worker->ch, batch, task->dst,
+						task->src, g_xfer_size_bytes, accel_done,
+						__accel_task_from_ap_task(task));
 		break;
 	case ACCEL_DUALCAST:
-		rc = spdk_accel_batch_prep_dualcast(__accel_task_from_ap_task(task),
-						    worker->ch, batch, task->dst, task->dst2,
-						    task->src, g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_batch_prep_dualcast(worker->ch, batch, task->dst, task->dst2,
+						    task->src, g_xfer_size_bytes, accel_done,
+						    __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_COMPARE:
-		rc = spdk_accel_batch_prep_compare(__accel_task_from_ap_task(task),
-						   worker->ch, batch, task->dst, task->src,
-						   g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_batch_prep_compare(worker->ch, batch, task->dst, task->src,
+						   g_xfer_size_bytes, accel_done,
+						   __accel_task_from_ap_task(task));
 		break;
 	case ACCEL_FILL:
-		rc = spdk_accel_batch_prep_fill(__accel_task_from_ap_task(task),
-						worker->ch, batch, task->dst, *(uint8_t *)task->src,
-						g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_batch_prep_fill(worker->ch, batch, task->dst, *(uint8_t *)task->src,
+						g_xfer_size_bytes, accel_done,
+						__accel_task_from_ap_task(task));
 		break;
 	case ACCEL_CRC32C:
-		rc = spdk_accel_batch_prep_crc32c(__accel_task_from_ap_task(task),
-						  worker->ch, batch, (uint32_t *)task->dst, task->src,
-						  g_crc32c_seed, g_xfer_size_bytes, accel_done);
+		rc = spdk_accel_batch_prep_crc32c(worker->ch, batch, (uint32_t *)task->dst, task->src,
+						  g_crc32c_seed, g_xfer_size_bytes, accel_done,
+						  __accel_task_from_ap_task(task));
 		break;
 	default:
 		assert(false);
@@ -594,8 +591,7 @@ _init_thread(void *arg1)
 			task->worker = worker;
 			task->worker->current_queue_depth++;
 
-			rc = spdk_accel_batch_submit(__accel_task_from_ap_task(task),
-						     worker->ch, batch, batch_done);
+			rc = spdk_accel_batch_submit(worker->ch, batch, batch_done, __accel_task_from_ap_task(task));
 			if (rc) {
 				fprintf(stderr, "error ending batch %d\n", rc);
 				goto error;
diff --git a/include/spdk/accel_engine.h b/include/spdk/accel_engine.h
index dcf0ab130..b7ab59112 100644
--- a/include/spdk/accel_engine.h
+++ b/include/spdk/accel_engine.h
@@ -123,20 +123,20 @@ uint64_t spdk_accel_get_capabilities(struct spdk_io_channel *ch);
 /**
  * Submit a copy request.
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param dst Destination to copy to.
  * \param src Source to copy from.
  * \param nbytes Length in bytes to copy.
- * \param cb Called when this copy operation completes.
+ * \param cb_fn Called when this copy operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_submit_copy(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch, void *dst,
-			   void *src, uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_submit_copy(struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
+			   spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
- * Synchronous call to get batch size.  This is the maximum number of
+ * Synchronous call to get batch size. This is the maximum number of
  *  descriptors that a batch can contain. Once this limit is reached the batch
  *  should be processed with spdk_accel_batch_submit().
  *
@@ -158,15 +158,15 @@ struct spdk_accel_batch *spdk_accel_batch_create(struct spdk_io_channel *ch);
 /**
  * Asynchronous call to submit a batch sequence.
  *
- * \param accel_req Accel request task.
  * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_submit(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			    struct spdk_accel_batch *batch, spdk_accel_completion_cb cb);
+int spdk_accel_batch_submit(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			    spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Synchronous call to prepare a copy request into a previously initialized batch
@@ -174,19 +174,19 @@ int spdk_accel_batch_submit(struct spdk_accel_task *accel_req, struct spdk_io_ch
  *  completes after the batch has been submitted by an asynchronous call to
  *  spdk_accel_batch_submit().
  *
- * \param accel_req Accel request task.
  * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
  * \param dst Destination to copy to.
  * \param src Source to copy from.
  * \param nbytes Length in bytes to copy.
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_prep_copy(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			       struct spdk_accel_batch *batch, void *dst, void *src,
-			       uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_batch_prep_copy(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			       void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn,
+			       void *cb_arg);
 
 /**
  * Synchronous call to prepare a dualcast request into a previously initialized batch
@@ -194,37 +194,36 @@ int spdk_accel_batch_prep_copy(struct spdk_accel_task *accel_req, struct spdk_io
  *  completes after the batch has been submitted by an asynchronous call to
  *  spdk_accel_batch_submit().
  *
- * \param accel_req Accel request task.
  * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
  * \param dst1 First destination to copy to (must be 4K aligned).
  * \param dst2 Second destination to copy to (must be 4K aligned).
  * \param src Source to copy from.
  * \param nbytes Length in bytes to copy.
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_prep_dualcast(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-				   struct spdk_accel_batch *batch, void *dst1, void *dst2, void *src,
-				   uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_batch_prep_dualcast(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+				   void *dst1, void *dst2, void *src, uint64_t nbytes,
+				   spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Submit a dual cast copy request.
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param dst1 First destination to copy to (must be 4K aligned).
  * \param dst2 Second destination to copy to (must be 4K aligned).
  * \param src Source to copy from.
  * \param nbytes Length in bytes to copy.
- * \param cb Called when this copy operation completes.
+ * \param cb_fn Called when this copy operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_submit_dualcast(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			       void *dst1, void *dst2, void *src, uint64_t nbytes,
-			       spdk_accel_completion_cb cb);
+int spdk_accel_submit_dualcast(struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
+			       uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Synchronous call to prepare a compare request into a previously initialized batch
@@ -232,35 +231,34 @@ int spdk_accel_submit_dualcast(struct spdk_accel_task *accel_req, struct spdk_io
  *  completes after the batch has been submitted by an asynchronous call to
  *  spdk_accel_batch_submit().
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
  * \param src1 First location to perform compare on.
  * \param src2 Second location to perform compare on.
  * \param nbytes Length in bytes to compare.
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_prep_compare(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-				  struct spdk_accel_batch *batch, void *src1, void *src2,
-				  uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_batch_prep_compare(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+				  void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn,
+				  void *cb_arg);
 
 /**
  * Submit a compare request.
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param src1 First location to perform compare on.
  * \param src2 Second location to perform compare on.
  * \param nbytes Length in bytes to compare.
- * \param cb Called when this compare operation completes.
+ * \param cb_fn Called when this compare operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, any other value means there was a miscompare.
  */
-int spdk_accel_submit_compare(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			      void *src1, void *src2, uint64_t nbytes,
-			      spdk_accel_completion_cb cb);
+int spdk_accel_submit_compare(struct spdk_io_channel *ch, void *src1, void *src2, uint64_t nbytes,
+			      spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Synchronous call to prepare a fill request into a previously initialized batch
@@ -268,36 +266,36 @@ int spdk_accel_submit_compare(struct spdk_accel_task *accel_req, struct spdk_io_
  *  completes after the batch has been submitted by an asynchronous call to
  *  spdk_accel_batch_submit().
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
  * \param dst Destination to fill.
  * \param fill Constant byte to fill to the destination.
  * \param nbytes Length in bytes to fill.
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_prep_fill(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			       struct spdk_accel_batch *batch, void *dst, uint8_t fill,
-			       uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_batch_prep_fill(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			       void *dst, uint8_t fill, uint64_t nbytes,
+			       spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Submit a fill request.
  *
  * This operation will fill the destination buffer with the specified value.
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param dst Destination to fill.
  * \param fill Constant byte to fill to the destination.
  * \param nbytes Length in bytes to fill.
- * \param cb Called when this fill operation completes.
+ * \param cb_fn Called when this fill operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_submit_fill(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			   void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill, uint64_t nbytes,
+			   spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Synchronous call to prepare a crc32c request into a previously initialized batch
@@ -305,38 +303,38 @@ int spdk_accel_submit_fill(struct spdk_accel_task *accel_req, struct spdk_io_cha
  *  completes after the batch has been submitted by an asynchronous call to
  *  spdk_accel_batch_submit().
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param batch Handle provided when the batch was started with spdk_accel_batch_create().
  * \param dst Destination to write the CRC-32C to.
  * \param src The source address for the data.
  * \param seed Four byte seed value.
  * \param nbytes Length in bytes.
- * \param cb Called when this operation completes.
+ * \param cb_fn Called when this operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_batch_prep_crc32c(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-				 struct spdk_accel_batch *batch, uint32_t *dst, void *src, uint32_t seed,
-				 uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_batch_prep_crc32c(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+				 uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes,
+				 spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Submit a CRC-32C calculation request.
  *
  * This operation will calculate the 4 byte CRC32-C for the given data.
  *
- * \param accel_req Accel request task.
- * \param ch I/O channel to submit request to the accel engine.
+ * \param ch I/O channel associated with this call.
  * \param dst Destination to write the CRC-32C to.
  * \param src The source address for the data.
  * \param seed Four byte seed value.
  * \param nbytes Length in bytes.
- * \param cb Called when this CRC-32C operation completes.
+ * \param cb_fn Called when this CRC-32C operation completes.
+ * \param cb_arg Callback argument.
  *
  * \return 0 on success, negative errno on failure.
  */
-int spdk_accel_submit_crc32c(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			     uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb);
+int spdk_accel_submit_crc32c(struct spdk_io_channel *ch, uint32_t *dst, void *src, uint32_t seed,
+			     uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
 
 /**
  * Get the size of an acceleration task.
diff --git a/include/spdk_internal/accel_engine.h b/include/spdk_internal/accel_engine.h
index 89f2117b5..3e047cf4a 100644
--- a/include/spdk_internal/accel_engine.h
+++ b/include/spdk_internal/accel_engine.h
@@ -41,36 +41,38 @@
 
 struct spdk_accel_task {
 	spdk_accel_completion_cb	cb;
+	void				*cb_arg;
 	uint8_t				offload_ctx[0];
 };
 
 struct spdk_accel_engine {
 	uint64_t (*get_capabilities)(void);
-	int (*copy)(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src,
-		    uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*dualcast)(void *cb_arg, struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
-			uint64_t nbytes, spdk_accel_completion_cb cb);
+	int (*copy)(struct spdk_io_channel *ch, void *dst, void *src,
+		    uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*dualcast)(struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
+			uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
 	uint32_t (*batch_get_max)(void);
 	struct spdk_accel_batch *(*batch_create)(struct spdk_io_channel *ch);
-	int (*batch_prep_copy)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-			       void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*batch_prep_dualcast)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-				   void *dst1, void *dst2, void *src, uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*batch_prep_compare)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-				  void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*batch_prep_fill)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-			       void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*batch_prep_crc32c)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+	int (*batch_prep_copy)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			       void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*batch_prep_dualcast)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+				   void *dst1, void *dst2, void *src, uint64_t nbytes,
+				   spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*batch_prep_compare)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+				  void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*batch_prep_fill)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			       void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*batch_prep_crc32c)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
 				 uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes,
-				 spdk_accel_completion_cb cb);
-	int (*batch_submit)(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-			    spdk_accel_completion_cb cb);
-	int (*compare)(void *cb_arg, struct spdk_io_channel *ch, void *src1, void *src2,
-		       uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*fill)(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fill,
-		    uint64_t nbytes, spdk_accel_completion_cb cb);
-	int (*crc32c)(void *cb_arg, struct spdk_io_channel *ch, uint32_t *dst, void *src,
-		      uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb);
+				 spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*batch_submit)(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			    spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*compare)(struct spdk_io_channel *ch, void *src1, void *src2,
+		       uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*fill)(struct spdk_io_channel *ch, void *dst, uint8_t fill,
+		    uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+	int (*crc32c)(struct spdk_io_channel *ch, uint32_t *dst, void *src,
+		      uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
 	struct spdk_io_channel *(*get_io_channel)(void);
 };
 
diff --git a/lib/accel/accel_engine.c b/lib/accel/accel_engine.c
index 1f7973224..15c4d7d31 100644
--- a/lib/accel/accel_engine.c
+++ b/lib/accel/accel_engine.c
@@ -48,7 +48,10 @@
  * later in this file.
  */
 
-#define ALIGN_4K 0x1000
+#define ALIGN_4K		0x1000
+#define SPDK_ACCEL_NUM_TASKS	0x4000
+
+static struct spdk_mempool *g_accel_task_pool;
 
 /* Largest context size for all accel modules */
 static size_t g_max_accel_module_size = 0;
@@ -71,16 +74,17 @@ struct accel_io_channel {
 /* Forward declarations of software implementations used when an
  * engine has not implemented the capability.
  */
-static int sw_accel_submit_dualcast(void *cb_arg, struct spdk_io_channel *ch, void *dst1,
-				    void *dst2, void *src, uint64_t nbytes, spdk_accel_completion_cb cb);
-static int sw_accel_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src,
-				uint64_t nbytes, spdk_accel_completion_cb cb);
-static int sw_accel_submit_compare(void *cb_arg, struct spdk_io_channel *ch, void *src1, void *src2,
-				   uint64_t nbytes, spdk_accel_completion_cb cb);
-static int sw_accel_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fill,
-				uint64_t nbytes, spdk_accel_completion_cb cb);
-static int sw_accel_submit_crc32c(void *cb_arg, struct spdk_io_channel *ch, uint32_t *dst,
-				  void *src, uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb);
+static int sw_accel_submit_dualcast(struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
+				    uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+static int sw_accel_submit_copy(struct spdk_io_channel *ch, void *dst, void *src,
+				uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+static int sw_accel_submit_compare(struct spdk_io_channel *ch, void *src1, void *src2,
+				   uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+static int sw_accel_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill,
+				uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg);
+static int sw_accel_submit_crc32c(struct spdk_io_channel *ch, uint32_t *dst, void *src,
+				  uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn,
+				  void *cb_arg);
 
 /* Registration of hw modules (currently supports only 1 at a time) */
 void
@@ -113,7 +117,8 @@ _accel_engine_done(void *ref, int status)
 {
 	struct spdk_accel_task *req = (struct spdk_accel_task *)ref;
 
-	req->cb(req, status);
+	req->cb(req->cb_arg, status);
+	spdk_mempool_put(g_accel_task_pool, req);
 }
 
 uint64_t
@@ -127,45 +132,58 @@ spdk_accel_get_capabilities(struct spdk_io_channel *ch)
 
 /* Accel framework public API for copy function */
 int
-spdk_accel_submit_copy(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-		       void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+spdk_accel_submit_copy(struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
+		       spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
 
 	/* If the engine does not support it, fallback to the sw implementation. */
 	if (accel_ch->engine->copy) {
-		return accel_ch->engine->copy(accel_req->offload_ctx, accel_ch->ch, dst, src, nbytes,
-					      _accel_engine_done);
+		return accel_ch->engine->copy(accel_ch->ch, dst, src, nbytes,
+					      _accel_engine_done, accel_req->offload_ctx);
 	} else {
-		return sw_accel_submit_copy(accel_req->offload_ctx, accel_ch->ch, dst, src, nbytes,
-					    _accel_engine_done);
+		return sw_accel_submit_copy(accel_ch->ch, dst, src, nbytes,
+					    _accel_engine_done, accel_req->offload_ctx);
 	}
 }
 
 /* Accel framework public API for dual cast copy function */
 int
-spdk_accel_submit_dualcast(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			   void *dst1, void *dst2, void *src, uint64_t nbytes,
-			   spdk_accel_completion_cb cb)
+spdk_accel_submit_dualcast(struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
+			   uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
+
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
 
 	if ((uintptr_t)dst1 & (ALIGN_4K - 1) || (uintptr_t)dst2 & (ALIGN_4K - 1)) {
 		SPDK_ERRLOG("Dualcast requires 4K alignment on dst addresses\n");
 		return -EINVAL;
 	}
 
-	accel_req->cb = cb;
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
 
 	/* If the engine does not support it, fallback to the sw implementation. */
 	if (accel_ch->engine->dualcast) {
-		return accel_ch->engine->dualcast(accel_req->offload_ctx, accel_ch->ch, dst1, dst2, src, nbytes,
-						  _accel_engine_done);
+		return accel_ch->engine->dualcast(accel_ch->ch, dst1, dst2, src, nbytes,
+						  _accel_engine_done, accel_req->offload_ctx);
 	} else {
-		return sw_accel_submit_dualcast(accel_req->offload_ctx, accel_ch->ch, dst1, dst2, src, nbytes,
-						_accel_engine_done);
+		return sw_accel_submit_dualcast(accel_ch->ch, dst1, dst2, src, nbytes,
+						_accel_engine_done, accel_req->offload_ctx);
 	}
 }
 
@@ -184,14 +202,22 @@ spdk_accel_batch_create(struct spdk_io_channel *ch)
  * required to implement this API.
  */
 int
-spdk_accel_batch_submit(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			struct spdk_accel_batch *batch, spdk_accel_completion_cb cb)
+spdk_accel_batch_submit(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_submit(accel_req->offload_ctx, accel_ch->ch, batch,
-					      _accel_engine_done);
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_submit(accel_ch->ch, batch, _accel_engine_done,
+					      accel_req->offload_ctx);
 }
 
 /* Accel framework public API for getting max batch. All engines are
@@ -209,136 +235,195 @@ spdk_accel_batch_get_max(struct spdk_io_channel *ch)
  * required to implement this API.
  */
 int
-spdk_accel_batch_prep_copy(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			   struct spdk_accel_batch *batch, void *dst, void *src, uint64_t nbytes,
-			   spdk_accel_completion_cb cb)
+spdk_accel_batch_prep_copy(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *dst,
+			   void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_prep_copy(accel_req->offload_ctx, accel_ch->ch, batch, dst, src,
-			nbytes, _accel_engine_done);
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_prep_copy(accel_ch->ch, batch, dst, src, nbytes,
+			_accel_engine_done, accel_req->offload_ctx);
 }
 
 /* Accel framework public API for batch prep_dualcast function.  All engines are
  * required to implement this API.
  */
 int
-spdk_accel_batch_prep_dualcast(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			       struct spdk_accel_batch *batch, void *dst1, void *dst2, void *src, uint64_t nbytes,
-			       spdk_accel_completion_cb cb)
+spdk_accel_batch_prep_dualcast(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			       void *dst1, void *dst2, void *src, uint64_t nbytes,
+			       spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
+
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
 
 	if ((uintptr_t)dst1 & (ALIGN_4K - 1) || (uintptr_t)dst2 & (ALIGN_4K - 1)) {
 		SPDK_ERRLOG("Dualcast requires 4K alignment on dst addresses\n");
 		return -EINVAL;
 	}
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_prep_dualcast(accel_req->offload_ctx, accel_ch->ch,
-			batch, dst1, dst2, src,	nbytes, _accel_engine_done);
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_prep_dualcast(accel_ch->ch, batch, dst1, dst2, src,
+			nbytes, _accel_engine_done, accel_req->offload_ctx);
 }
 
 /* Accel framework public API for batch prep_compare function.  All engines are
  * required to implement this API.
  */
 int
-spdk_accel_batch_prep_compare(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			      struct spdk_accel_batch *batch, void *src1, void *src2, uint64_t nbytes,
-			      spdk_accel_completion_cb cb)
+spdk_accel_batch_prep_compare(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			      void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn,
+			      void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_prep_compare(accel_req->offload_ctx, accel_ch->ch,
-			batch, src1, src2, nbytes, _accel_engine_done);
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_prep_compare(accel_ch->ch, batch, src1, src2, nbytes,
+			_accel_engine_done, accel_req->offload_ctx);
 }
 
 /* Accel framework public API for batch prep_fill function.  All engines are
  * required to implement this API.
  */
 int
-spdk_accel_batch_prep_fill(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			   struct spdk_accel_batch *batch, void *dst, uint8_t fill, uint64_t nbytes,
-			   spdk_accel_completion_cb cb)
+spdk_accel_batch_prep_fill(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *dst,
+			   uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_prep_fill(accel_req->offload_ctx, accel_ch->ch,
-			batch, dst, fill, nbytes, _accel_engine_done);
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_prep_fill(accel_ch->ch, batch, dst, fill, nbytes,
+			_accel_engine_done, accel_req->offload_ctx);
 }
 
 /* Accel framework public API for batch prep_crc32c function.  All engines are
  * required to implement this API.
  */
 int
-spdk_accel_batch_prep_crc32c(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			     struct spdk_accel_batch *batch, uint32_t *dst, void *src, uint32_t seed,
-			     uint64_t nbytes, spdk_accel_completion_cb cb)
+spdk_accel_batch_prep_crc32c(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			     uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes,
+			     spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
-	return accel_ch->engine->batch_prep_crc32c(accel_req->offload_ctx, accel_ch->ch,
-			batch, dst, src, seed, nbytes, _accel_engine_done);
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
+
+	return accel_ch->engine->batch_prep_crc32c(accel_ch->ch, batch, dst, src, seed, nbytes,
+			_accel_engine_done, accel_req->offload_ctx);
 }
 
 /* Accel framework public API for compare function */
 int
-spdk_accel_submit_compare(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			  void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb)
+spdk_accel_submit_compare(struct spdk_io_channel *ch, void *src1, void *src2, uint64_t nbytes,
+			  spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
 
 	/* If the engine does not support it, fallback to the sw implementation. */
 	if (accel_ch->engine->compare) {
-		return accel_ch->engine->compare(accel_req->offload_ctx, accel_ch->ch, src1, src2, nbytes,
-						 _accel_engine_done);
+		return accel_ch->engine->compare(accel_ch->ch, src1, src2, nbytes,
+						 _accel_engine_done, accel_req->offload_ctx);
 	} else {
-		return sw_accel_submit_compare(accel_req->offload_ctx, accel_ch->ch, src1, src2, nbytes,
-					       _accel_engine_done);
+		return sw_accel_submit_compare(accel_ch->ch, src1, src2, nbytes,
+					       _accel_engine_done, accel_req->offload_ctx);
 	}
 }
 
 /* Accel framework public API for fill function */
 int
-spdk_accel_submit_fill(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-		       void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb)
+spdk_accel_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill, uint64_t nbytes,
+		       spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
 
 	/* If the engine does not support it, fallback to the sw implementation. */
 	if (accel_ch->engine->fill) {
-		return accel_ch->engine->fill(accel_req->offload_ctx, accel_ch->ch, dst, fill, nbytes,
-					      _accel_engine_done);
+		return accel_ch->engine->fill(accel_ch->ch, dst, fill, nbytes,
+					      _accel_engine_done, accel_req->offload_ctx);
 	} else {
-		return sw_accel_submit_fill(accel_req->offload_ctx, accel_ch->ch, dst, fill, nbytes,
-					    _accel_engine_done);
+		return sw_accel_submit_fill(accel_ch->ch, dst, fill, nbytes,
+					    _accel_engine_done, accel_req->offload_ctx);
 	}
 }
 
 /* Accel framework public API for CRC-32C function */
 int
-spdk_accel_submit_crc32c(struct spdk_accel_task *accel_req, struct spdk_io_channel *ch,
-			 uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb)
+spdk_accel_submit_crc32c(struct spdk_io_channel *ch, uint32_t *dst, void *src, uint32_t seed,
+			 uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct accel_io_channel *accel_ch = spdk_io_channel_get_ctx(ch);
+	struct spdk_accel_task *accel_req = spdk_mempool_get(g_accel_task_pool);
 
-	accel_req->cb = cb;
+	if (accel_req == NULL) {
+		SPDK_ERRLOG("Unable to get an accel task.\n");
+		return -ENOMEM;
+	}
+
+	accel_req->cb = cb_fn;
+	accel_req->cb_arg = cb_arg;
 
 	/* If the engine does not support it, fallback to the sw implementation. */
 	if (accel_ch->engine->crc32c) {
-		return accel_ch->engine->crc32c(accel_req->offload_ctx, accel_ch->ch, dst, src,
-						seed, nbytes, _accel_engine_done);
+		return accel_ch->engine->crc32c(accel_ch->ch, dst, src,	seed, nbytes,
+						_accel_engine_done, accel_req->offload_ctx);
 	} else {
-		return sw_accel_submit_crc32c(accel_req->offload_ctx, accel_ch->ch, dst, src,
-					      seed, nbytes, _accel_engine_done);
+		return sw_accel_submit_crc32c(accel_ch->ch, dst, src, seed, nbytes,
+					      _accel_engine_done, accel_req->offload_ctx);
 	}
 }
 
@@ -398,10 +483,20 @@ static void
 accel_engine_module_initialize(void)
 {
 	struct spdk_accel_module_if *accel_engine_module;
+	char task_pool_name[30];
 
 	TAILQ_FOREACH(accel_engine_module, &spdk_accel_module_list, tailq) {
 		accel_engine_module->module_init();
 	}
+
+	snprintf(task_pool_name, sizeof(task_pool_name), "accel_task_pool");
+	g_accel_task_pool = spdk_mempool_create(task_pool_name,
+						SPDK_ACCEL_NUM_TASKS,
+						g_max_accel_module_size,
+						SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
+						SPDK_ENV_SOCKET_ID_ANY);
+	assert(g_accel_task_pool);
+
 }
 
 int
@@ -479,6 +574,7 @@ spdk_accel_engine_finish(spdk_accel_fini_cb cb_fn, void *cb_arg)
 
 	spdk_io_device_unregister(&spdk_accel_module_list, NULL);
 	spdk_accel_engine_module_finish();
+	spdk_mempool_free(g_accel_task_pool);
 }
 
 void
@@ -562,8 +658,8 @@ sw_accel_batch_start(struct spdk_io_channel *ch)
 }
 
 static struct sw_accel_op *
-_prep_op(void *cb_arg, struct sw_accel_io_channel *sw_ch, struct spdk_accel_batch *batch,
-	 spdk_accel_completion_cb cb)
+_prep_op(struct sw_accel_io_channel *sw_ch, struct spdk_accel_batch *batch,
+	 spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 
@@ -581,20 +677,20 @@ _prep_op(void *cb_arg, struct sw_accel_io_channel *sw_ch, struct spdk_accel_batc
 	}
 
 	op->cb_arg = cb_arg;
-	op->cb_fn = cb;
+	op->cb_fn = cb_fn;
 	op->sw_ch = sw_ch;
 
 	return op;
 }
 
 static int
-sw_accel_batch_prep_copy(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-			 void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_batch_prep_copy(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			 void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, sw_ch, batch, cb);
+	op = _prep_op(sw_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -610,14 +706,14 @@ sw_accel_batch_prep_copy(void *cb_arg, struct spdk_io_channel *ch, struct spdk_a
 }
 
 static int
-sw_accel_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch,
-			     struct spdk_accel_batch *batch, void *dst1, void *dst2,
-			     void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_batch_prep_dualcast(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *dst1,
+			     void *dst2,
+			     void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, sw_ch, batch, cb);
+	op = _prep_op(sw_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -634,14 +730,13 @@ sw_accel_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-sw_accel_batch_prep_compare(void *cb_arg, struct spdk_io_channel *ch,
-			    struct spdk_accel_batch *batch, void *src1,
-			    void *src2, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_batch_prep_compare(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *src1,
+			    void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, sw_ch, batch, cb);
+	op = _prep_op(sw_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -657,14 +752,14 @@ sw_accel_batch_prep_compare(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-sw_accel_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch,
-			 struct spdk_accel_batch *batch, void *dst, uint8_t fill,
-			 uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_batch_prep_fill(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *dst,
+			 uint8_t fill,
+			 uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, sw_ch, batch, cb);
+	op = _prep_op(sw_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -680,14 +775,14 @@ sw_accel_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-sw_accel_batch_prep_crc32c(void *cb_arg, struct spdk_io_channel *ch,
-			   struct spdk_accel_batch *batch, uint32_t *dst, void *src,
-			   uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_batch_prep_crc32c(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+			   uint32_t *dst,
+			   void *src, uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, sw_ch, batch, cb);
+	op = _prep_op(sw_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -704,8 +799,8 @@ sw_accel_batch_prep_crc32c(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-sw_accel_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-		      spdk_accel_completion_cb cb)
+sw_accel_batch_submit(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+		      spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct sw_accel_op *op;
 	struct sw_accel_io_channel *sw_ch = spdk_io_channel_get_ctx(ch);
@@ -753,14 +848,14 @@ sw_accel_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_acce
 	/* Now complete the batch request itself. */
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, batch_status);
+	cb_fn(accel_req, batch_status);
 
 	return 0;
 }
 
 static int
-sw_accel_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src,
-		     uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_submit_copy(struct spdk_io_channel *ch, void *dst, void *src,
+		     uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct spdk_accel_task *accel_req;
 
@@ -768,13 +863,13 @@ sw_accel_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *
 
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, 0);
+	cb_fn(accel_req, 0);
 	return 0;
 }
 
 static int
-sw_accel_submit_dualcast(void *cb_arg, struct spdk_io_channel *ch, void *dst1, void *dst2,
-			 void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_submit_dualcast(struct spdk_io_channel *ch, void *dst1, void *dst2,
+			 void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct spdk_accel_task *accel_req;
 
@@ -783,13 +878,13 @@ sw_accel_submit_dualcast(void *cb_arg, struct spdk_io_channel *ch, void *dst1, v
 
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, 0);
+	cb_fn(accel_req, 0);
 	return 0;
 }
 
 static int
-sw_accel_submit_compare(void *cb_arg, struct spdk_io_channel *ch, void *src1, void *src2,
-			uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_submit_compare(struct spdk_io_channel *ch, void *src1, void *src2,
+			uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct spdk_accel_task *accel_req;
 	int result;
@@ -798,35 +893,35 @@ sw_accel_submit_compare(void *cb_arg, struct spdk_io_channel *ch, void *src1, vo
 
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, result);
+	cb_fn(accel_req, result);
 
 	return 0;
 }
 
 static int
-sw_accel_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fill,
-		     uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill,
+		     uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct spdk_accel_task *accel_req;
 
 	memset(dst, fill, nbytes);
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, 0);
+	cb_fn(accel_req, 0);
 
 	return 0;
 }
 
 static int
-sw_accel_submit_crc32c(void *cb_arg, struct spdk_io_channel *ch, uint32_t *dst, void *src,
-		       uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb)
+sw_accel_submit_crc32c(struct spdk_io_channel *ch, uint32_t *dst, void *src,
+		       uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct spdk_accel_task *accel_req;
 
 	*dst = spdk_crc32c_update(src, nbytes, ~seed);
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, 0);
+	cb_fn(accel_req, 0);
 
 	return 0;
 }
diff --git a/module/accel/idxd/accel_engine_idxd.c b/module/accel/idxd/accel_engine_idxd.c
index 546cb67fb..e9d92eab5 100644
--- a/module/accel/idxd/accel_engine_idxd.c
+++ b/module/accel/idxd/accel_engine_idxd.c
@@ -219,14 +219,14 @@ _prep_queue_command(struct idxd_io_channel *chan, spdk_accel_completion_cb cb_fn
 }
 
 static int
-idxd_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
-		 spdk_accel_completion_cb cb)
+idxd_submit_copy(struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
+		 spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	int rc = 0;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
 		rc = spdk_idxd_submit_copy(chan->chan, dst, src, nbytes, idxd_done, idxd_task);
@@ -259,14 +259,14 @@ idxd_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src,
 }
 
 static int
-idxd_submit_dualcast(void *cb_arg, struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
-		     uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_submit_dualcast(struct spdk_io_channel *ch, void *dst1, void *dst2, void *src,
+		     uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	int rc = 0;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
 		rc = spdk_idxd_submit_dualcast(chan->chan, dst1, dst2, src, nbytes, idxd_done, idxd_task);
@@ -300,14 +300,14 @@ idxd_submit_dualcast(void *cb_arg, struct spdk_io_channel *ch, void *dst1, void
 }
 
 static int
-idxd_submit_compare(void *cb_arg, struct spdk_io_channel *ch, void *src1, void *src2,
-		    uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_submit_compare(struct spdk_io_channel *ch, void *src1, void *src2,
+		    uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	int rc = 0;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
 		rc = spdk_idxd_submit_compare(chan->chan, src1, src2, nbytes, idxd_done, idxd_task);
@@ -340,15 +340,15 @@ idxd_submit_compare(void *cb_arg, struct spdk_io_channel *ch, void *src1, void *
 }
 
 static int
-idxd_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fill,
-		 uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill,
+		 uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	int rc = 0;
 	uint64_t fill_pattern;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 	memset(&fill_pattern, fill, sizeof(uint64_t));
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
@@ -382,14 +382,14 @@ idxd_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fi
 }
 
 static int
-idxd_submit_crc32c(void *cb_arg, struct spdk_io_channel *ch, uint32_t *dst, void *src,
-		   uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_submit_crc32c(struct spdk_io_channel *ch, uint32_t *dst, void *src,
+		   uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	int rc = 0;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
 		rc = spdk_idxd_submit_crc32c(chan->chan, dst, src, seed, nbytes, idxd_done, idxd_task);
@@ -444,15 +444,15 @@ idxd_batch_start(struct spdk_io_channel *ch)
 }
 
 static int
-idxd_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
-		  spdk_accel_completion_cb cb)
+idxd_batch_submit(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+		  spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 	int rc = 0;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	if (chan->state == IDXD_CHANNEL_ACTIVE) {
 		rc = spdk_idxd_batch_submit(chan->chan, batch, idxd_done, idxd_task);
@@ -483,29 +483,29 @@ idxd_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_ba
 }
 
 static int
-idxd_batch_prep_copy(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
-		     void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_batch_prep_copy(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+		     void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	return spdk_idxd_batch_prep_copy(chan->chan, batch, dst, src, nbytes,
 					 idxd_done, idxd_task);
 }
 
 static int
-idxd_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
-		     void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_batch_prep_fill(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+		     void *dst, uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	uint64_t fill_pattern;
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 	memset(&fill_pattern, fill, sizeof(uint64_t));
 
 	return spdk_idxd_batch_prep_fill(chan->chan, batch, dst, fill_pattern, nbytes, idxd_done,
@@ -513,43 +513,44 @@ idxd_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel
 }
 
 static int
-idxd_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
-			 void *dst1, void *dst2, void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_batch_prep_dualcast(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+			 void *dst1, void *dst2, void *src, uint64_t nbytes,
+			 spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	return spdk_idxd_batch_prep_dualcast(chan->chan, batch, dst1, dst2, src, nbytes, idxd_done,
 					     idxd_task);
 }
 
 static int
-idxd_batch_prep_crc32c(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+idxd_batch_prep_crc32c(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
 		       uint32_t *dst, void *src, uint32_t seed, uint64_t nbytes,
-		       spdk_accel_completion_cb cb)
+		       spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	return spdk_idxd_batch_prep_crc32c(chan->chan, batch, dst, src, seed, nbytes, idxd_done,
 					   idxd_task);
 }
 
 static int
-idxd_batch_prep_compare(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
-			void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb)
+idxd_batch_prep_compare(struct spdk_io_channel *ch, struct spdk_accel_batch *_batch,
+			void *src1, void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct idxd_task *idxd_task = (struct idxd_task *)cb_arg;
 	struct idxd_io_channel *chan = spdk_io_channel_get_ctx(ch);
 	struct idxd_batch *batch = (struct idxd_batch *)_batch;
 
-	idxd_task->cb = cb;
+	idxd_task->cb = cb_fn;
 
 	return spdk_idxd_batch_prep_compare(chan->chan, batch, src1, src2, nbytes, idxd_done,
 					    idxd_task);
diff --git a/module/accel/ioat/accel_engine_ioat.c b/module/accel/ioat/accel_engine_ioat.c
index 2ecfad5e6..7368d60e6 100644
--- a/module/accel/ioat/accel_engine_ioat.c
+++ b/module/accel/ioat/accel_engine_ioat.c
@@ -182,22 +182,22 @@ ioat_done(void *cb_arg)
 }
 
 static int
-ioat_submit_copy(void *cb_arg, struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
-		 spdk_accel_completion_cb cb)
+ioat_submit_copy(struct spdk_io_channel *ch, void *dst, void *src, uint64_t nbytes,
+		 spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_task *ioat_task = (struct ioat_task *)cb_arg;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
 
 	assert(ioat_ch->ioat_ch != NULL);
 
-	ioat_task->cb = cb;
+	ioat_task->cb = cb_fn;
 
 	return spdk_ioat_submit_copy(ioat_ch->ioat_ch, ioat_task, ioat_done, dst, src, nbytes);
 }
 
 static int
-ioat_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fill,
-		 uint64_t nbytes, spdk_accel_completion_cb cb)
+ioat_submit_fill(struct spdk_io_channel *ch, void *dst, uint8_t fill,
+		 uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_task *ioat_task = (struct ioat_task *)cb_arg;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
@@ -205,7 +205,7 @@ ioat_submit_fill(void *cb_arg, struct spdk_io_channel *ch, void *dst, uint8_t fi
 
 	assert(ioat_ch->ioat_ch != NULL);
 
-	ioat_task->cb = cb;
+	ioat_task->cb = cb_fn;
 
 	return spdk_ioat_submit_fill(ioat_ch->ioat_ch, ioat_task, ioat_done, dst, fill64, nbytes);
 }
@@ -259,8 +259,8 @@ ioat_batch_create(struct spdk_io_channel *ch)
 }
 
 static struct ioat_accel_op *
-_prep_op(void *cb_arg, struct ioat_io_channel *ioat_ch, struct spdk_accel_batch *batch,
-	 spdk_accel_completion_cb cb)
+_prep_op(struct ioat_io_channel *ioat_ch, struct spdk_accel_batch *batch,
+	 spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_accel_op *op;
 
@@ -278,20 +278,20 @@ _prep_op(void *cb_arg, struct ioat_io_channel *ioat_ch, struct spdk_accel_batch
 	}
 
 	op->cb_arg = cb_arg;
-	op->cb_fn = cb;
+	op->cb_fn = cb_fn;
 	op->ioat_ch = ioat_ch;
 
 	return op;
 }
 
 static int
-ioat_batch_prep_copy(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-		     void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+ioat_batch_prep_copy(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+		     void *dst, void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
 	struct ioat_task *ioat_task = (struct ioat_task *)cb_arg;
 
-	ioat_task->cb = cb;
+	ioat_task->cb = cb_fn;
 	ioat_ch->hw_batch = true;
 
 	/* Call the IOAT library prep function. */
@@ -299,14 +299,13 @@ ioat_batch_prep_copy(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel
 }
 
 static int
-ioat_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch,
-		     struct spdk_accel_batch *batch, void *dst, uint8_t fill,
-		     uint64_t nbytes, spdk_accel_completion_cb cb)
+ioat_batch_prep_fill(struct spdk_io_channel *ch, struct spdk_accel_batch *batch, void *dst,
+		     uint8_t fill, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
 	struct ioat_task *ioat_task = (struct ioat_task *)cb_arg;
 
-	ioat_task->cb = cb;
+	ioat_task->cb = cb_fn;
 	ioat_ch->hw_batch = true;
 
 	/* Call the IOAT library prep function. */
@@ -314,9 +313,9 @@ ioat_batch_prep_fill(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-ioat_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch,
+ioat_batch_prep_dualcast(struct spdk_io_channel *ch,
 			 struct spdk_accel_batch *batch, void *dst1, void *dst2,
-			 void *src, uint64_t nbytes, spdk_accel_completion_cb cb)
+			 void *src, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_accel_op *op;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
@@ -326,7 +325,7 @@ ioat_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch,
 		return -EINVAL;
 	}
 
-	op = _prep_op(cb_arg, ioat_ch, batch, cb);
+	op = _prep_op(ioat_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -343,14 +342,14 @@ ioat_batch_prep_dualcast(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-ioat_batch_prep_compare(void *cb_arg, struct spdk_io_channel *ch,
+ioat_batch_prep_compare(struct spdk_io_channel *ch,
 			struct spdk_accel_batch *batch, void *src1,
-			void *src2, uint64_t nbytes, spdk_accel_completion_cb cb)
+			void *src2, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_accel_op *op;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, ioat_ch, batch, cb);
+	op = _prep_op(ioat_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -366,14 +365,14 @@ ioat_batch_prep_compare(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-ioat_batch_prep_crc32c(void *cb_arg, struct spdk_io_channel *ch,
+ioat_batch_prep_crc32c(struct spdk_io_channel *ch,
 		       struct spdk_accel_batch *batch, uint32_t *dst, void *src,
-		       uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb)
+		       uint32_t seed, uint64_t nbytes, spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_accel_op *op;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
 
-	op = _prep_op(cb_arg, ioat_ch, batch, cb);
+	op = _prep_op(ioat_ch, batch, cb_fn, cb_arg);
 	if (op == NULL) {
 		return -EINVAL;
 	}
@@ -390,8 +389,8 @@ ioat_batch_prep_crc32c(void *cb_arg, struct spdk_io_channel *ch,
 }
 
 static int
-ioat_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
-		  spdk_accel_completion_cb cb)
+ioat_batch_submit(struct spdk_io_channel *ch, struct spdk_accel_batch *batch,
+		  spdk_accel_completion_cb cb_fn, void *cb_arg)
 {
 	struct ioat_accel_op *op;
 	struct ioat_io_channel *ioat_ch = spdk_io_channel_get_ctx(ch);
@@ -437,7 +436,7 @@ ioat_batch_submit(void *cb_arg, struct spdk_io_channel *ch, struct spdk_accel_ba
 	/* Now complete the batch request itself. */
 	accel_req = (struct spdk_accel_task *)((uintptr_t)cb_arg -
 					       offsetof(struct spdk_accel_task, offload_ctx));
-	cb(accel_req, batch_status);
+	cb_fn(accel_req, batch_status);
 
 	return 0;
 }
diff --git a/module/bdev/malloc/bdev_malloc.c b/module/bdev/malloc/bdev_malloc.c
index ed7e94afc..ce0403153 100644
--- a/module/bdev/malloc/bdev_malloc.c
+++ b/module/bdev/malloc/bdev_malloc.c
@@ -58,22 +58,10 @@ struct malloc_task {
 	enum spdk_bdev_io_status	status;
 };
 
-static struct malloc_task *
-__malloc_task_from_accel_task(struct spdk_accel_task *ct)
-{
-	return (struct malloc_task *)((uintptr_t)ct - sizeof(struct malloc_task));
-}
-
-static struct spdk_accel_task *
-__accel_task_from_malloc_task(struct malloc_task *mt)
-{
-	return (struct spdk_accel_task *)((uintptr_t)mt + sizeof(struct malloc_task));
-}
-
 static void
 malloc_done(void *ref, int status)
 {
-	struct malloc_task *task = __malloc_task_from_accel_task(ref);
+	struct malloc_task *task = (struct malloc_task *)ref;
 
 	if (status != 0) {
 		if (status == -ENOMEM) {
@@ -98,7 +86,7 @@ static void bdev_malloc_get_spdk_running_config(FILE *fp);
 static int
 bdev_malloc_get_ctx_size(void)
 {
-	return sizeof(struct malloc_task) + spdk_accel_task_size();
+	return sizeof(struct malloc_task);
 }
 
 static struct spdk_bdev_module malloc_if = {
@@ -171,12 +159,11 @@ bdev_malloc_readv(struct malloc_disk *mdisk, struct spdk_io_channel *ch,
 	task->num_outstanding = iovcnt;
 
 	for (i = 0; i < iovcnt; i++) {
-		res = spdk_accel_submit_copy(__accel_task_from_malloc_task(task),
-					     ch, iov[i].iov_base,
-					     src, iov[i].iov_len, malloc_done);
+		res = spdk_accel_submit_copy(ch, iov[i].iov_base,
+					     src, iov[i].iov_len, malloc_done, task);
 
 		if (res != 0) {
-			malloc_done(__accel_task_from_malloc_task(task), res);
+			malloc_done(task, res);
 		}
 
 		src += iov[i].iov_len;
@@ -206,12 +193,11 @@ bdev_malloc_writev(struct malloc_disk *mdisk, struct spdk_io_channel *ch,
 	task->num_outstanding = iovcnt;
 
 	for (i = 0; i < iovcnt; i++) {
-		res = spdk_accel_submit_copy(__accel_task_from_malloc_task(task),
-					     ch, dst, iov[i].iov_base,
-					     iov[i].iov_len, malloc_done);
+		res = spdk_accel_submit_copy(ch, dst, iov[i].iov_base,
+					     iov[i].iov_len, malloc_done, task);
 
 		if (res != 0) {
-			malloc_done(__accel_task_from_malloc_task(task), res);
+			malloc_done(task, res);
 		}
 
 		dst += iov[i].iov_len;
@@ -228,8 +214,8 @@ bdev_malloc_unmap(struct malloc_disk *mdisk,
 	task->status = SPDK_BDEV_IO_STATUS_SUCCESS;
 	task->num_outstanding = 1;
 
-	return spdk_accel_submit_fill(__accel_task_from_malloc_task(task), ch,
-				      mdisk->malloc_buf + offset, 0, byte_count, malloc_done);
+	return spdk_accel_submit_fill(ch, mdisk->malloc_buf + offset, 0,
+				      byte_count, malloc_done, task);
 }
 
 static int64_t