[dpdk-dev] [PATCH v3] crypto/scheduler: add multicore scheduling mode

[Kirill Rybalchenko <kirill.rybalchenko@intel.com>]

Multi-core scheduling mode is a mode where scheduler distributes
crypto operations in a round-robin base, between several core
assigned as workers.

Signed-off-by: Kirill Rybalchenko <kirill.rybalchenko@intel.com>
---
Changes in v3:
- Rebased against dpdk-next-crypto
- Fixed op status check in the reordering ring dequeuing
- Modified size of the ring between scheduler and workers
- Modified macro names as per Fan's comment
- Changed type of variable from 'usigned int' to uint32_t

 doc/guides/cryptodevs/scheduler.rst                |  23 ++
 doc/guides/rel_notes/release_17_08.rst             |   7 +-
 drivers/crypto/scheduler/Makefile                  |   1 +
 drivers/crypto/scheduler/rte_cryptodev_scheduler.c |   7 +
 drivers/crypto/scheduler/rte_cryptodev_scheduler.h |   9 +
 drivers/crypto/scheduler/scheduler_multicore.c     | 412 +++++++++++++++++++++
 drivers/crypto/scheduler/scheduler_pmd.c           |  78 +++-
 drivers/crypto/scheduler/scheduler_pmd_private.h   |   5 +-
 8 files changed, 538 insertions(+), 4 deletions(-)
 create mode 100644 drivers/crypto/scheduler/scheduler_multicore.c

diff --git a/doc/guides/cryptodevs/scheduler.rst b/doc/guides/cryptodevs/scheduler.rst
index 32e5653..d393ae8 100644
--- a/doc/guides/cryptodevs/scheduler.rst
+++ b/doc/guides/cryptodevs/scheduler.rst
@@ -170,3 +170,26 @@ operation:
    crypto operation burst to the primary slave. When one or more crypto
    operations fail to be enqueued, then they will be enqueued to the secondary
    slave.
+
+*   **CDEV_SCHED_MODE_MULTICORE:**
+
+   *Initialization mode parameter*: **multi-core**
+
+   Multi-core mode, which distributes the workload with several (up to eight)
+   worker cores. The enqueued bursts are distributed among the worker cores in a
+   round-robin manner. If scheduler cannot enqueue entire burst to the same worker,
+   it will enqueue the remaining operations to the next available worker.
+   For pure small packet size (64 bytes) traffic however the multi-core mode is not
+   an optimal solution, as it doesn't give significant per-core performance improvement.
+   For mixed traffic (IMIX) the optimal number of worker cores is around 2-3.
+   For large packets (1.5 Kbytes) scheduler shows linear scaling in performance
+   up to eight cores.
+   Each worker uses its own slave cryptodev. Only software cryptodevs
+   are supported. Only the same type of cryptodevs should be used concurrently.
+
+   The multi-core mode uses one extra parameter:
+   * corelist: Semicolon-separated list of logical cores to be used as workers.
+     The number of worker cores should be equal to the number of slave cryptodevs.
+
+   Example:
+    ... --vdev "crypto_aesni_mb_pmd,name=aesni_mb_1" --vdev "crypto_aesni_mb_pmd,name=aesni_mb_2" --vdev "crypto_scheduler_pmd,slave=aesni_mb_1,slave=aesni_mb_2,mode=multi-core,corelist=23;24" ...
diff --git a/doc/guides/rel_notes/release_17_08.rst b/doc/guides/rel_notes/release_17_08.rst
index f7aca8e..4dea05f 100644
--- a/doc/guides/rel_notes/release_17_08.rst
+++ b/doc/guides/rel_notes/release_17_08.rst
@@ -92,7 +92,7 @@ New Features
 
     * 12-byte IV on AES Counter Mode, apart from the previous 16-byte IV.
 
-**Updated the AES-NI GCM PMD.**
+* **Updated the AES-NI GCM PMD.**
 
   The AES-NI GCM PMD was migrated from the ISA-L library to the Multi Buffer
   library, as the latter library has Scatter Gather List support
@@ -100,6 +100,11 @@ New Features
 
   * 192-bit key.
 
+* **Updated the Cryptodev Scheduler PMD.**
+
+  Added a multicore based distribution mode, which distributes the enqueued
+  crypto operations among several slaves, running on different logical cores.
+
 
 Resolved Issues
 ---------------
diff --git a/drivers/crypto/scheduler/Makefile b/drivers/crypto/scheduler/Makefile
index c273e78..b045410 100644
--- a/drivers/crypto/scheduler/Makefile
+++ b/drivers/crypto/scheduler/Makefile
@@ -56,5 +56,6 @@ SRCS-$(CONFIG_RTE_LIBRTE_PMD_CRYPTO_SCHEDULER) += rte_cryptodev_scheduler.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_CRYPTO_SCHEDULER) += scheduler_roundrobin.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_CRYPTO_SCHEDULER) += scheduler_pkt_size_distr.c
 SRCS-$(CONFIG_RTE_LIBRTE_PMD_CRYPTO_SCHEDULER) += scheduler_failover.c
+SRCS-$(CONFIG_RTE_LIBRTE_PMD_CRYPTO_SCHEDULER) += scheduler_multicore.c
 
 include $(RTE_SDK)/mk/rte.lib.mk
diff --git a/drivers/crypto/scheduler/rte_cryptodev_scheduler.c b/drivers/crypto/scheduler/rte_cryptodev_scheduler.c
index 9c364c2..11d0d81 100644
--- a/drivers/crypto/scheduler/rte_cryptodev_scheduler.c
+++ b/drivers/crypto/scheduler/rte_cryptodev_scheduler.c
@@ -351,6 +351,13 @@ rte_cryptodev_scheduler_mode_set(uint8_t scheduler_id,
 			return -1;
 		}
 		break;
+	case CDEV_SCHED_MODE_MULTICORE:
+		if (rte_cryptodev_scheduler_load_user_scheduler(scheduler_id,
+				multicore_scheduler) < 0) {
+			CS_LOG_ERR("Failed to load scheduler");
+			return -1;
+		}
+		break;
 	default:
 		CS_LOG_ERR("Not yet supported");
 		return -ENOTSUP;
diff --git a/drivers/crypto/scheduler/rte_cryptodev_scheduler.h b/drivers/crypto/scheduler/rte_cryptodev_scheduler.h
index 33570ec..e8d3c45 100644
--- a/drivers/crypto/scheduler/rte_cryptodev_scheduler.h
+++ b/drivers/crypto/scheduler/rte_cryptodev_scheduler.h
@@ -58,12 +58,17 @@ extern "C" {
 #define RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES	(8)
 #endif
 
+/** Maximum number of multi-core worker cores */
+#define RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES	(64)
+
 /** Round-robin scheduling mode string */
 #define SCHEDULER_MODE_NAME_ROUND_ROBIN		round-robin
 /** Packet-size based distribution scheduling mode string */
 #define SCHEDULER_MODE_NAME_PKT_SIZE_DISTR	packet-size-distr
 /** Fail-over scheduling mode string */
 #define SCHEDULER_MODE_NAME_FAIL_OVER		fail-over
+/** multi-core scheduling mode string */
+#define SCHEDULER_MODE_NAME_MULTI_CORE		multi-core
 
 /**
  * Crypto scheduler PMD operation modes
@@ -78,6 +83,8 @@ enum rte_cryptodev_scheduler_mode {
 	CDEV_SCHED_MODE_PKT_SIZE_DISTR,
 	/** Fail-over mode */
 	CDEV_SCHED_MODE_FAILOVER,
+	/** multi-core mode */
+	CDEV_SCHED_MODE_MULTICORE,
 
 	CDEV_SCHED_MODE_COUNT /**< number of modes */
 };
@@ -295,6 +302,8 @@ extern struct rte_cryptodev_scheduler *roundrobin_scheduler;
 extern struct rte_cryptodev_scheduler *pkt_size_based_distr_scheduler;
 /** Fail-over mode scheduler */
 extern struct rte_cryptodev_scheduler *failover_scheduler;
+/** multi-core mode scheduler */
+extern struct rte_cryptodev_scheduler *multicore_scheduler;
 
 #ifdef __cplusplus
 }
diff --git a/drivers/crypto/scheduler/scheduler_multicore.c b/drivers/crypto/scheduler/scheduler_multicore.c
new file mode 100644
index 0000000..8757552
--- /dev/null
+++ b/drivers/crypto/scheduler/scheduler_multicore.c
@@ -0,0 +1,412 @@
+/*-
+ *   BSD LICENSE
+ *
+ *   Copyright(c) 2017 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 <unistd.h>
+
+#include <rte_cryptodev.h>
+#include <rte_malloc.h>
+
+#include "rte_cryptodev_scheduler_operations.h"
+#include "scheduler_pmd_private.h"
+
+#define MC_SCHED_ENQ_RING_NAME_PREFIX	"MCS_ENQR_"
+#define MC_SCHED_DEQ_RING_NAME_PREFIX	"MCS_DEQR_"
+
+#define MC_SCHED_BUFFER_SIZE 32
+
+/** multi-core scheduler context */
+struct mc_scheduler_ctx {
+	uint32_t num_workers;             /**< Number of workers polling */
+	uint32_t stop_signal;
+
+	struct rte_ring *sched_enq_ring[RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES];
+	struct rte_ring *sched_deq_ring[RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES];
+};
+
+struct mc_scheduler_qp_ctx {
+	struct scheduler_slave slaves[RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES];
+	uint32_t nb_slaves;
+
+	uint32_t last_enq_worker_idx;
+	uint32_t last_deq_worker_idx;
+
+	struct mc_scheduler_ctx *mc_private_ctx;
+};
+
+static uint16_t
+schedule_enqueue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
+{
+	struct mc_scheduler_qp_ctx *mc_qp_ctx =
+			((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
+	struct mc_scheduler_ctx *mc_ctx = mc_qp_ctx->mc_private_ctx;
+	uint32_t worker_idx = mc_qp_ctx->last_enq_worker_idx;
+	uint16_t i, processed_ops = 0;
+
+	if (unlikely(nb_ops == 0))
+		return 0;
+
+	for (i = 0; i <  mc_ctx->num_workers && nb_ops != 0; i++) {
+		struct rte_ring *enq_ring = mc_ctx->sched_enq_ring[worker_idx];
+		uint16_t nb_queue_ops = rte_ring_enqueue_burst(enq_ring,
+			(void *)(&ops[processed_ops]), nb_ops, NULL);
+
+		nb_ops -= nb_queue_ops;
+		processed_ops += nb_queue_ops;
+
+		if (++worker_idx == mc_ctx->num_workers)
+			worker_idx = 0;
+	}
+	mc_qp_ctx->last_enq_worker_idx = worker_idx;
+
+	return processed_ops;
+}
+
+static uint16_t
+schedule_enqueue_ordering(void *qp, struct rte_crypto_op **ops,
+		uint16_t nb_ops)
+{
+	struct rte_ring *order_ring =
+			((struct scheduler_qp_ctx *)qp)->order_ring;
+	uint16_t nb_ops_to_enq = get_max_enqueue_order_count(order_ring,
+			nb_ops);
+	uint16_t nb_ops_enqd = schedule_enqueue(qp, ops,
+			nb_ops_to_enq);
+
+	scheduler_order_insert(order_ring, ops, nb_ops_enqd);
+
+	return nb_ops_enqd;
+}
+
+
+static uint16_t
+schedule_dequeue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
+{
+	struct mc_scheduler_qp_ctx *mc_qp_ctx =
+			((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
+	struct mc_scheduler_ctx *mc_ctx = mc_qp_ctx->mc_private_ctx;
+	uint32_t worker_idx = mc_qp_ctx->last_deq_worker_idx;
+	uint16_t i, processed_ops = 0;
+
+	for (i = 0; i < mc_ctx->num_workers && nb_ops != 0; i++) {
+		struct rte_ring *deq_ring = mc_ctx->sched_deq_ring[worker_idx];
+		uint16_t nb_deq_ops = rte_ring_dequeue_burst(deq_ring,
+			(void *)(&ops[processed_ops]), nb_ops, NULL);
+
+		nb_ops -= nb_deq_ops;
+		processed_ops += nb_deq_ops;
+		if (++worker_idx == mc_ctx->num_workers)
+			worker_idx = 0;
+	}
+
+	mc_qp_ctx->last_deq_worker_idx = worker_idx;
+
+	return processed_ops;
+
+}
+
+static uint16_t
+schedule_dequeue_ordering(void *qp, struct rte_crypto_op **ops,
+		uint16_t nb_ops)
+{
+	struct rte_ring *order_ring =
+			((struct scheduler_qp_ctx *)qp)->order_ring;
+
+	return scheduler_order_drain(order_ring, ops, nb_ops);
+}
+
+static int
+slave_attach(__rte_unused struct rte_cryptodev *dev,
+		__rte_unused uint8_t slave_id)
+{
+	return 0;
+}
+
+static int
+slave_detach(__rte_unused struct rte_cryptodev *dev,
+		__rte_unused uint8_t slave_id)
+{
+	return 0;
+}
+
+static int
+mc_scheduler_worker(struct rte_cryptodev *dev)
+{
+	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
+	struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
+	struct rte_ring *enq_ring;
+	struct rte_ring *deq_ring;
+	uint32_t core_id = rte_lcore_id();
+	int i, worker_idx = -1;
+	struct scheduler_slave *slave;
+	struct rte_crypto_op *enq_ops[MC_SCHED_BUFFER_SIZE];
+	struct rte_crypto_op *deq_ops[MC_SCHED_BUFFER_SIZE];
+	struct scheduler_session *sess0, *sess1, *sess2, *sess3;
+	uint16_t processed_ops;
+	uint16_t left_op = 0;
+	uint16_t left_op_idx = 0;
+	uint16_t inflight_ops = 0;
+
+	for (i = 0; i < (int)sched_ctx->nb_wc; i++) {
+		if (sched_ctx->wc_pool[i] == core_id) {
+			worker_idx = i;
+			break;
+		}
+	}
+	if (worker_idx == -1) {
+		CS_LOG_ERR("worker on core %u:cannot find worker index!\n", core_id);
+		return -1;
+	}
+
+	slave = &sched_ctx->slaves[worker_idx];
+	enq_ring = mc_ctx->sched_enq_ring[worker_idx];
+	deq_ring = mc_ctx->sched_deq_ring[worker_idx];
+
+	while (!mc_ctx->stop_signal) {
+		if (left_op) {
+			processed_ops =
+				rte_cryptodev_enqueue_burst(slave->dev_id,
+						slave->qp_id,
+						&enq_ops[left_op_idx], left_op);
+
+			left_op -= processed_ops;
+			left_op_idx += processed_ops;
+		} else {
+			uint16_t nb_deq_ops = rte_ring_dequeue_burst(enq_ring,
+				(void *)enq_ops, MC_SCHED_BUFFER_SIZE, NULL);
+			if (nb_deq_ops) {
+				uint16_t i;
+
+				for (i = 0; i < nb_deq_ops && i < 4; i++)
+					rte_prefetch0(enq_ops[i]->sym->session);
+
+				for (i = 0; (i < (nb_deq_ops - 8))
+							&& (nb_deq_ops > 8); i += 4) {
+					sess0 = (struct scheduler_session *)
+						enq_ops[i]->sym->session->_private;
+					sess1 = (struct scheduler_session *)
+						enq_ops[i+1]->sym->session->_private;
+					sess2 = (struct scheduler_session *)
+						enq_ops[i+2]->sym->session->_private;
+					sess3 = (struct scheduler_session *)
+						enq_ops[i+3]->sym->session->_private;
+
+					enq_ops[i]->sym->session =
+							sess0->sessions[worker_idx];
+					enq_ops[i + 1]->sym->session =
+							sess1->sessions[worker_idx];
+					enq_ops[i + 2]->sym->session =
+							sess2->sessions[worker_idx];
+					enq_ops[i + 3]->sym->session =
+							sess3->sessions[worker_idx];
+
+					rte_prefetch0(enq_ops[i + 4]->sym->session);
+					rte_prefetch0(enq_ops[i + 5]->sym->session);
+					rte_prefetch0(enq_ops[i + 6]->sym->session);
+					rte_prefetch0(enq_ops[i + 7]->sym->session);
+				}
+
+				for (; i < nb_deq_ops; i++) {
+					sess0 = (struct scheduler_session *)
+						enq_ops[i]->sym->session->_private;
+					enq_ops[i]->sym->session =
+							sess0->sessions[worker_idx];
+				}
+
+				processed_ops = rte_cryptodev_enqueue_burst(slave->dev_id,
+						slave->qp_id, enq_ops, nb_deq_ops);
+
+				if (unlikely(processed_ops < nb_deq_ops)) {
+					left_op = nb_deq_ops - processed_ops;
+					left_op_idx = processed_ops;
+				}
+
+				inflight_ops += processed_ops;
+			}
+		}
+
+		if (inflight_ops > 0) {
+			processed_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
+					slave->qp_id, deq_ops, MC_SCHED_BUFFER_SIZE);
+			if (processed_ops) {
+				uint16_t nb_enq_ops = rte_ring_enqueue_burst(deq_ring,
+					(void *)deq_ops, processed_ops, NULL);
+				inflight_ops -= nb_enq_ops;
+			}
+		}
+
+		rte_pause();
+	}
+
+	return 0;
+}
+
+static int
+scheduler_start(struct rte_cryptodev *dev)
+{
+	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
+	struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
+	uint16_t i;
+
+	mc_ctx->stop_signal = 0;
+
+	for (i = 0; i < sched_ctx->nb_wc; i++)
+		rte_eal_remote_launch(
+			(lcore_function_t *)mc_scheduler_worker, dev,
+					sched_ctx->wc_pool[i]);
+
+	if (sched_ctx->reordering_enabled) {
+		dev->enqueue_burst = &schedule_enqueue_ordering;
+		dev->dequeue_burst = &schedule_dequeue_ordering;
+	} else {
+		dev->enqueue_burst = &schedule_enqueue;
+		dev->dequeue_burst = &schedule_dequeue;
+	}
+
+	for (i = 0; i < dev->data->nb_queue_pairs; i++) {
+		struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
+		struct mc_scheduler_qp_ctx *mc_qp_ctx =
+				qp_ctx->private_qp_ctx;
+		uint32_t j;
+
+		memset(mc_qp_ctx->slaves, 0,
+				RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES *
+				sizeof(struct scheduler_slave));
+		for (j = 0; j < sched_ctx->nb_slaves; j++) {
+			mc_qp_ctx->slaves[j].dev_id =
+					sched_ctx->slaves[j].dev_id;
+			mc_qp_ctx->slaves[j].qp_id = i;
+		}
+
+		mc_qp_ctx->nb_slaves = sched_ctx->nb_slaves;
+
+		mc_qp_ctx->last_enq_worker_idx = 0;
+		mc_qp_ctx->last_deq_worker_idx = 0;
+	}
+
+	return 0;
+}
+
+static int
+scheduler_stop(struct rte_cryptodev *dev)
+{
+	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
+	struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
+
+	mc_ctx->stop_signal = 1;
+
+	for (uint16_t i = 0; i < sched_ctx->nb_wc; i++)
+		rte_eal_wait_lcore(sched_ctx->wc_pool[i]);
+
+	return 0;
+}
+
+static int
+scheduler_config_qp(struct rte_cryptodev *dev, uint16_t qp_id)
+{
+	struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[qp_id];
+	struct mc_scheduler_qp_ctx *mc_qp_ctx;
+	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
+	struct mc_scheduler_ctx *mc_ctx = sched_ctx->private_ctx;
+
+	mc_qp_ctx = rte_zmalloc_socket(NULL, sizeof(*mc_qp_ctx), 0,
+			rte_socket_id());
+	if (!mc_qp_ctx) {
+		CS_LOG_ERR("failed allocate memory for private queue pair");
+		return -ENOMEM;
+	}
+
+	mc_qp_ctx->mc_private_ctx = mc_ctx;
+	qp_ctx->private_qp_ctx = (void *)mc_qp_ctx;
+
+
+	return 0;
+}
+
+static int
+scheduler_create_private_ctx(struct rte_cryptodev *dev)
+{
+	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
+	struct mc_scheduler_ctx *mc_ctx;
+
+	if (sched_ctx->private_ctx)
+		rte_free(sched_ctx->private_ctx);
+
+	mc_ctx = rte_zmalloc_socket(NULL, sizeof(struct mc_scheduler_ctx), 0,
+			rte_socket_id());
+	if (!mc_ctx) {
+		CS_LOG_ERR("failed allocate memory");
+		return -ENOMEM;
+	}
+
+	mc_ctx->num_workers = sched_ctx->nb_wc;
+	for (uint16_t i = 0; i < sched_ctx->nb_wc; i++) {
+		char r_name[16];
+
+		snprintf(r_name, sizeof(r_name), MC_SCHED_ENQ_RING_NAME_PREFIX "%u", i);
+		mc_ctx->sched_enq_ring[i] = rte_ring_create(r_name, PER_SLAVE_BUFF_SIZE,
+					rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
+		if (!mc_ctx->sched_enq_ring[i]) {
+			CS_LOG_ERR("Cannot create ring for worker %u", i);
+			return -1;
+		}
+		snprintf(r_name, sizeof(r_name), MC_SCHED_DEQ_RING_NAME_PREFIX "%u", i);
+		mc_ctx->sched_deq_ring[i] = rte_ring_create(r_name, PER_SLAVE_BUFF_SIZE,
+					rte_socket_id(), RING_F_SC_DEQ | RING_F_SP_ENQ);
+		if (!mc_ctx->sched_deq_ring[i]) {
+			CS_LOG_ERR("Cannot create ring for worker %u", i);
+			return -1;
+		}
+	}
+
+	sched_ctx->private_ctx = (void *)mc_ctx;
+
+	return 0;
+}
+
+struct rte_cryptodev_scheduler_ops scheduler_mc_ops = {
+	slave_attach,
+	slave_detach,
+	scheduler_start,
+	scheduler_stop,
+	scheduler_config_qp,
+	scheduler_create_private_ctx,
+	NULL,	/* option_set */
+	NULL	/* option_get */
+};
+
+struct rte_cryptodev_scheduler mc_scheduler = {
+		.name = "multicore-scheduler",
+		.description = "scheduler which will run burst across multiple cpu cores",
+		.mode = CDEV_SCHED_MODE_MULTICORE,
+		.ops = &scheduler_mc_ops
+};
+
+struct rte_cryptodev_scheduler *multicore_scheduler = &mc_scheduler;
diff --git a/drivers/crypto/scheduler/scheduler_pmd.c b/drivers/crypto/scheduler/scheduler_pmd.c
index b385851..67269d7 100644
--- a/drivers/crypto/scheduler/scheduler_pmd.c
+++ b/drivers/crypto/scheduler/scheduler_pmd.c
@@ -49,6 +49,7 @@ struct scheduler_init_params {
 	uint32_t nb_slaves;
 	enum rte_cryptodev_scheduler_mode mode;
 	uint32_t enable_ordering;
+	uint64_t wcmask;
 	char slave_names[RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES]
 			[RTE_CRYPTODEV_SCHEDULER_NAME_MAX_LEN];
 };
@@ -60,6 +61,8 @@ struct scheduler_init_params {
 #define RTE_CRYPTODEV_VDEV_MAX_NB_QP_ARG	("max_nb_queue_pairs")
 #define RTE_CRYPTODEV_VDEV_MAX_NB_SESS_ARG	("max_nb_sessions")
 #define RTE_CRYPTODEV_VDEV_SOCKET_ID		("socket_id")
+#define RTE_CRYPTODEV_VDEV_COREMASK		("coremask")
+#define RTE_CRYPTODEV_VDEV_CORELIST		("corelist")
 
 const char *scheduler_valid_params[] = {
 	RTE_CRYPTODEV_VDEV_NAME,
@@ -68,7 +71,9 @@ const char *scheduler_valid_params[] = {
 	RTE_CRYPTODEV_VDEV_ORDERING,
 	RTE_CRYPTODEV_VDEV_MAX_NB_QP_ARG,
 	RTE_CRYPTODEV_VDEV_MAX_NB_SESS_ARG,
-	RTE_CRYPTODEV_VDEV_SOCKET_ID
+	RTE_CRYPTODEV_VDEV_SOCKET_ID,
+	RTE_CRYPTODEV_VDEV_COREMASK,
+	RTE_CRYPTODEV_VDEV_CORELIST
 };
 
 struct scheduler_parse_map {
@@ -82,7 +87,9 @@ const struct scheduler_parse_map scheduler_mode_map[] = {
 	{RTE_STR(SCHEDULER_MODE_NAME_PKT_SIZE_DISTR),
 			CDEV_SCHED_MODE_PKT_SIZE_DISTR},
 	{RTE_STR(SCHEDULER_MODE_NAME_FAIL_OVER),
-			CDEV_SCHED_MODE_FAILOVER}
+			CDEV_SCHED_MODE_FAILOVER},
+	{RTE_STR(SCHEDULER_MODE_NAME_MULTI_CORE),
+			CDEV_SCHED_MODE_MULTICORE}
 };
 
 const struct scheduler_parse_map scheduler_ordering_map[] = {
@@ -122,6 +129,21 @@ cryptodev_scheduler_create(const char *name,
 	sched_ctx->max_nb_queue_pairs =
 			init_params->def_p.max_nb_queue_pairs;
 
+	if (init_params->mode == CDEV_SCHED_MODE_MULTICORE) {
+		uint16_t i;
+
+		sched_ctx->nb_wc = 0;
+
+		for (i = 0; i < RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES; i++) {
+			if (init_params->wcmask & (1ULL << i)) {
+				sched_ctx->wc_pool[sched_ctx->nb_wc++] = i;
+				RTE_LOG(INFO, PMD,
+					"  Worker core[%u]=%u added\n",
+					sched_ctx->nb_wc-1, i);
+			}
+		}
+	}
+
 	if (init_params->mode > CDEV_SCHED_MODE_USERDEFINED &&
 			init_params->mode < CDEV_SCHED_MODE_COUNT) {
 		ret = rte_cryptodev_scheduler_mode_set(dev->data->dev_id,
@@ -240,6 +262,43 @@ parse_integer_arg(const char *key __rte_unused,
 	return 0;
 }
 
+/** Parse integer from hexadecimal integer argument */
+static int
+parse_coremask_arg(const char *key __rte_unused,
+		const char *value, void *extra_args)
+{
+	struct scheduler_init_params *params = extra_args;
+
+	params->wcmask = strtoull(value, NULL, 16);
+
+	return 0;
+}
+
+/** Parse integer from list of integers argument */
+static int
+parse_corelist_arg(const char *key __rte_unused,
+		const char *value, void *extra_args)
+{
+	struct scheduler_init_params *params = extra_args;
+
+	params->wcmask = 0ULL;
+
+	const char *token = value;
+
+	while (isdigit(token[0])) {
+		char *rval;
+		unsigned int core = strtoul(token, &rval, 10);
+
+		params->wcmask |= 1ULL << core;
+		token = (const char *)rval;
+		if (token[0] == '\0')
+			break;
+		token++;
+	}
+
+	return 0;
+}
+
 /** Parse name */
 static int
 parse_name_arg(const char *key __rte_unused,
@@ -359,6 +418,18 @@ scheduler_parse_init_params(struct scheduler_init_params *params,
 		if (ret < 0)
 			goto free_kvlist;
 
+		ret = rte_kvargs_process(kvlist, RTE_CRYPTODEV_VDEV_COREMASK,
+				&parse_coremask_arg,
+				params);
+		if (ret < 0)
+			goto free_kvlist;
+
+		ret = rte_kvargs_process(kvlist, RTE_CRYPTODEV_VDEV_CORELIST,
+				&parse_corelist_arg,
+				params);
+		if (ret < 0)
+			goto free_kvlist;
+
 		ret = rte_kvargs_process(kvlist, RTE_CRYPTODEV_VDEV_NAME,
 				&parse_name_arg,
 				&params->def_p);
@@ -420,6 +491,9 @@ cryptodev_scheduler_probe(struct rte_vdev_device *vdev)
 	if (init_params.def_p.name[0] != '\0')
 		RTE_LOG(INFO, PMD, "  User defined name = %s\n",
 			init_params.def_p.name);
+	if (init_params.wcmask != 0)
+		RTE_LOG(INFO, PMD, "  workers core mask = %lx\n",
+			init_params.wcmask);
 
 	return cryptodev_scheduler_create(name,
 					vdev,
diff --git a/drivers/crypto/scheduler/scheduler_pmd_private.h b/drivers/crypto/scheduler/scheduler_pmd_private.h
index a78e9a6..b38567c 100644
--- a/drivers/crypto/scheduler/scheduler_pmd_private.h
+++ b/drivers/crypto/scheduler/scheduler_pmd_private.h
@@ -89,6 +89,8 @@ struct scheduler_ctx {
 
 	char name[RTE_CRYPTODEV_SCHEDULER_NAME_MAX_LEN];
 	char description[RTE_CRYPTODEV_SCHEDULER_DESC_MAX_LEN];
+	uint16_t wc_pool[RTE_CRYPTODEV_SCHEDULER_MAX_NB_WORKER_CORES];
+	uint16_t nb_wc;
 
 	char *init_slave_names[RTE_CRYPTODEV_SCHEDULER_MAX_NB_SLAVES];
 	int nb_init_slaves;
@@ -144,7 +146,8 @@ scheduler_order_drain(struct rte_ring *order_ring,
 
 	while (nb_ops_to_deq < nb_objs) {
 		SCHEDULER_GET_RING_OBJ(order_ring, nb_ops_to_deq, op);
-		if (op->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED)
+		if (op->status == RTE_CRYPTO_OP_STATUS_NOT_PROCESSED ||
+					op->status == RTE_CRYPTO_OP_STATUS_ENQUEUED)
 			break;
 		nb_ops_to_deq++;
 	}
-- 
2.5.5