[dpdk-dev] [PATCH v2 1/3] crypto/scheduler: add packet size based mode code

Declan Doherty declan.doherty at intel.com
Tue Mar 28 12:37:10 CEST 2017


On 23/03/17 11:03, Fan Zhang wrote:
> This patch adds the packet size based distribution mode main source
> file.
>
A little bit of detail on how the fail over scheduling will work would 
be good in the commit comment.

> Signed-off-by: Fan Zhang <roy.fan.zhang at intel.com>
> ---
>  .../crypto/scheduler/scheduler_pkt_size_distr.c    | 427 +++++++++++++++++++++
>  1 file changed, 427 insertions(+)
>  create mode 100644 drivers/crypto/scheduler/scheduler_pkt_size_distr.c
>
> diff --git a/drivers/crypto/scheduler/scheduler_pkt_size_distr.c b/drivers/crypto/scheduler/scheduler_pkt_size_distr.c
> new file mode 100644
> index 0000000..d1e8b7c
> --- /dev/null
> +++ b/drivers/crypto/scheduler/scheduler_pkt_size_distr.c
> @@ -0,0 +1,427 @@
> +/*-
> + *   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 <rte_cryptodev.h>
> +#include <rte_malloc.h>
> +
> +#include "rte_cryptodev_scheduler_operations.h"
> +#include "scheduler_pmd_private.h"
> +
> +#define PKT_SIZE_THRESHOLD			(0xff80)
> +#define SLAVE_IDX_SWITCH_MASK			(0x01)
> +#define PRIMARY_SLAVE_IDX			0
> +#define SECONDARY_SLAVE_IDX			1
> +#define NB_PKT_SIZE_SLAVES			2
> +
> +struct psd_scheduler_ctx {
> +	uint16_t threshold;
> +};
> +
> +struct psd_scheduler_qp_ctx {
> +	struct scheduler_slave primary_slave;
> +	struct scheduler_slave secondary_slave;
> +	uint16_t threshold;
> +	uint8_t deq_idx;
> +} __rte_cache_aligned;
> +
> +/** scheduling operation variables' wrapping */
> +struct psd_schedule_op {
> +	uint16_t slave_idx;
> +	int pos;
> +	int pos_shift;
> +};
> +
> +static uint16_t
> +schedule_enqueue(void *qp, struct rte_crypto_op **ops, uint16_t nb_ops)
> +{
> +	struct psd_scheduler_qp_ctx *qp_ctx =
> +			((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
> +	struct rte_crypto_op *sched_ops[nb_ops];
> +	struct rte_cryptodev_sym_session *sessions[nb_ops];
> +	struct scheduler_session *sess;
> +	struct psd_schedule_op enq_ops[NB_PKT_SIZE_SLAVES] = {
> +		{PRIMARY_SLAVE_IDX, 0, 1},
> +		{SECONDARY_SLAVE_IDX, (int)(nb_ops - 1), -1}
> +	};
> +	struct psd_schedule_op *p_enq_op;
> +	uint16_t i, processed_ops = 0, processed_ops2 = 0, nb_ops_to_enq;
> +	uint32_t job_len;
> +
> +	if (unlikely(nb_ops == 0))
> +		return 0;
> +
> +	for (i = 0; i < nb_ops && i < 4; i++) {
> +		rte_prefetch0(ops[i]->sym);
> +		rte_prefetch0(ops[i]->sym->session);
> +	}
> +
> +	for (i = 0; (i < (nb_ops - 8)) && (nb_ops > 8); i += 4) {
> +		rte_prefetch0(ops[i + 4]->sym);
> +		rte_prefetch0(ops[i + 4]->sym->session);
> +		rte_prefetch0(ops[i + 5]->sym);
> +		rte_prefetch0(ops[i + 5]->sym->session);
> +		rte_prefetch0(ops[i + 6]->sym);
> +		rte_prefetch0(ops[i + 6]->sym->session);
> +		rte_prefetch0(ops[i + 7]->sym);
> +		rte_prefetch0(ops[i + 7]->sym->session);
> +
> +		sess = (struct scheduler_session *)
> +				ops[i]->sym->session->_private;
> +		job_len = ops[i]->sym->cipher.data.length;
> +		job_len += (ops[i]->sym->auth.data.length == 0) *
> +				ops[i]->sym->auth.data.length;
> +		/* decide the target op based on the job length */
> +		p_enq_op = &enq_ops[!(job_len & qp_ctx->threshold)];
> +		sched_ops[p_enq_op->pos] = ops[i];
> +		sessions[p_enq_op->pos] = ops[i]->sym->session;
> +		ops[i]->sym->session = sess->sessions[p_enq_op->slave_idx];
> +		/* update position */
> +		p_enq_op->pos += p_enq_op->pos_shift;
> +
> +		sess = (struct scheduler_session *)
> +				ops[i+1]->sym->session->_private;
> +		job_len = ops[i+1]->sym->cipher.data.length;
> +		job_len += (ops[i+1]->sym->auth.data.length == 0) *
> +				ops[i+1]->sym->auth.data.length;
> +		p_enq_op = &enq_ops[!(job_len & qp_ctx->threshold)];

The threshold mask is only a uint16_t whereas the the job_len is a uint32_t

> +		sched_ops[p_enq_op->pos] = ops[i+1];
> +		sessions[p_enq_op->pos] = ops[i+1]->sym->session;
> +		ops[i+1]->sym->session = sess->sessions[p_enq_op->slave_idx];
> +		p_enq_op->pos += p_enq_op->pos_shift;
> +
> +		sess = (struct scheduler_session *)
> +				ops[i+2]->sym->session->_private;
> +		job_len = ops[i+2]->sym->cipher.data.length;
> +		job_len += (ops[i+2]->sym->auth.data.length == 0) *
> +				ops[i+2]->sym->auth.data.length;
> +		p_enq_op = &enq_ops[!(job_len & qp_ctx->threshold)];
> +		sched_ops[p_enq_op->pos] = ops[i+2];
> +		sessions[p_enq_op->pos] = ops[i+2]->sym->session;
> +		ops[i+2]->sym->session = sess->sessions[p_enq_op->slave_idx];
> +		p_enq_op->pos += p_enq_op->pos_shift;
> +
> +		sess = (struct scheduler_session *)
> +				ops[i+3]->sym->session->_private;
> +
> +		job_len = ops[i+3]->sym->cipher.data.length;
> +		job_len += (ops[i+3]->sym->auth.data.length == 0) *
> +				ops[i+3]->sym->auth.data.length;
> +		p_enq_op = &enq_ops[!(job_len & qp_ctx->threshold)];
> +		sched_ops[p_enq_op->pos] = ops[i+3];
> +		sessions[p_enq_op->pos] = ops[i+3]->sym->session;
> +		ops[i+3]->sym->session = sess->sessions[p_enq_op->slave_idx];
> +		p_enq_op->pos += p_enq_op->pos_shift;
> +	}
> +
> +	for (; i < nb_ops; i++) {
> +		sess = (struct scheduler_session *)
> +				ops[i]->sym->session->_private;
> +
> +		job_len = ops[i]->sym->cipher.data.length;
> +		job_len += (ops[i]->sym->auth.data.length == 0) *
> +				ops[i]->sym->auth.data.length;
> +		p_enq_op = &enq_ops[!(job_len & qp_ctx->threshold)];
> +		sched_ops[p_enq_op->pos] = ops[i];
> +		sessions[p_enq_op->pos] = ops[i]->sym->session;
> +		ops[i]->sym->session = sess->sessions[p_enq_op->slave_idx];
> +		p_enq_op->pos += p_enq_op->pos_shift;
> +	}
> +
Unless there is a measurable different in performance I think the intent 
of this code would be much clearer if you just kept two independent 
crypto op arrays for the primary and secondary slave

> +	processed_ops = rte_cryptodev_enqueue_burst(
> +			qp_ctx->primary_slave.dev_id,
> +			qp_ctx->primary_slave.qp_id,
> +			sched_ops,
> +			enq_ops[PRIMARY_SLAVE_IDX].pos);
> +	qp_ctx->primary_slave.nb_inflight_cops += processed_ops;
> +
> +	/* for cops failed to enqueue to primary slave, enqueue to 2nd slave */
> +	if (processed_ops < enq_ops[PRIMARY_SLAVE_IDX].pos)
> +		for (i = processed_ops;
> +				i < enq_ops[PRIMARY_SLAVE_IDX].pos; i++) {
> +			sess = (struct scheduler_session *)
> +				sessions[i]->_private;
> +			sched_ops[i]->sym->session = sess->sessions[
> +					SECONDARY_SLAVE_IDX];
> +		}
> +
> +	nb_ops_to_enq = nb_ops - processed_ops;
> +
> +	processed_ops2 = rte_cryptodev_enqueue_burst(
> +			qp_ctx->secondary_slave.dev_id,
> +			qp_ctx->secondary_slave.qp_id,
> +			&sched_ops[processed_ops],
> +			nb_ops_to_enq);
> +	qp_ctx->secondary_slave.nb_inflight_cops += processed_ops2;
> +
> +	processed_ops += processed_ops2;
> +
> +	/* for cops failed to enqueue in the end, recover session */
> +	if (unlikely(processed_ops < nb_ops))
> +		for (i = processed_ops; i < nb_ops; i++)
> +			sched_ops[i]->sym->session = sessions[i];
> +
> +	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 psd_scheduler_qp_ctx *qp_ctx =
> +			((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
> +	struct scheduler_slave *slaves[NB_PKT_SIZE_SLAVES] = {
> +			&qp_ctx->primary_slave, &qp_ctx->secondary_slave};
> +	struct scheduler_slave *slave = slaves[qp_ctx->deq_idx];
> +	uint16_t nb_deq_ops = 0, nb_deq_ops2 = 0;
> +
> +	if (slave->nb_inflight_cops) {
> +		nb_deq_ops = rte_cryptodev_dequeue_burst(slave->dev_id,
> +			slave->qp_id, ops, nb_ops);
> +		slave->nb_inflight_cops -= nb_deq_ops;
> +
> +		/* force a flush */
> +		if (unlikely(nb_deq_ops == 0))
> +			rte_cryptodev_enqueue_burst(slave->dev_id, slave->qp_id,
> +					NULL, 0);
If your forcing a flush here, you should probably called dequeue 
afterwards to get the flushed packets

> +	}
> +
> +	qp_ctx->deq_idx = (~qp_ctx->deq_idx) & SLAVE_IDX_SWITCH_MASK;
> +
> +	if (nb_deq_ops == nb_ops)
> +		return nb_deq_ops;
> +
> +	slave = slaves[qp_ctx->deq_idx];
> +
> +	if (slave->nb_inflight_cops) {
> +		nb_deq_ops2 = rte_cryptodev_dequeue_burst(slave->dev_id,
> +			slave->qp_id, &ops[nb_deq_ops], nb_ops - nb_deq_ops);
> +		slave->nb_inflight_cops -= nb_deq_ops2;
> +
> +		/* force a flush */
> +		if (unlikely(nb_deq_ops == 0))
> +			rte_cryptodev_enqueue_burst(slave->dev_id, slave->qp_id,
> +					NULL, 0);
> +	}
> +
> +	return nb_deq_ops + nb_deq_ops2;
> +}
> +
> +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;
> +	struct psd_scheduler_qp_ctx *qp_ctx =
> +			((struct scheduler_qp_ctx *)qp)->private_qp_ctx;
> +	uint16_t nb_deq_ops = 0;
> +
> +	if (qp_ctx->primary_slave.nb_inflight_cops) {
> +		nb_deq_ops = rte_cryptodev_dequeue_burst(
> +				qp_ctx->primary_slave.dev_id,
> +				qp_ctx->primary_slave.qp_id, ops, nb_ops);
> +		qp_ctx->primary_slave.nb_inflight_cops -= nb_deq_ops;
> +
> +		/* force a flush */
> +		if (unlikely(nb_deq_ops == 0))
> +			rte_cryptodev_enqueue_burst(
> +					qp_ctx->primary_slave.dev_id,
> +					qp_ctx->primary_slave.qp_id,
> +					NULL, 0);
> +	}
> +
> +	if (qp_ctx->secondary_slave.nb_inflight_cops) {
> +		nb_deq_ops = rte_cryptodev_dequeue_burst(
> +				qp_ctx->secondary_slave.dev_id,
> +				qp_ctx->secondary_slave.qp_id, ops, nb_ops);
> +		qp_ctx->secondary_slave.nb_inflight_cops -= nb_deq_ops;
> +
> +		/* force a flush */
> +		if (unlikely(nb_deq_ops == 0))
> +			rte_cryptodev_enqueue_burst(
> +					qp_ctx->secondary_slave.dev_id,
> +					qp_ctx->secondary_slave.qp_id,
> +					NULL, 0);
> +	}

Can you not just call the schedule_dequeue function above it looks to be 
mostly the same

> +
> +	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
> +scheduler_start(struct rte_cryptodev *dev)
> +{
> +	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
> +	struct psd_scheduler_ctx *psd_ctx = sched_ctx->private_ctx;
> +	uint16_t i;
> +
> +	/* for packet size based scheduler, nb_slaves have to >= 2 */
> +	if (sched_ctx->nb_slaves < NB_PKT_SIZE_SLAVES) {
> +		CS_LOG_ERR("not enough slaves to start");
> +		return -1;
> +	}
> +
> +	for (i = 0; i < dev->data->nb_queue_pairs; i++) {
> +		struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
> +		struct psd_scheduler_qp_ctx *ps_qp_ctx =
> +				qp_ctx->private_qp_ctx;
> +
> +		ps_qp_ctx->primary_slave.dev_id =
> +				sched_ctx->slaves[PRIMARY_SLAVE_IDX].dev_id;
> +		ps_qp_ctx->primary_slave.qp_id = i;
> +		ps_qp_ctx->primary_slave.nb_inflight_cops = 0;
> +
> +		ps_qp_ctx->secondary_slave.dev_id =
> +				sched_ctx->slaves[SECONDARY_SLAVE_IDX].dev_id;
> +		ps_qp_ctx->secondary_slave.qp_id = i;
> +		ps_qp_ctx->secondary_slave.nb_inflight_cops = 0;
> +
> +		ps_qp_ctx->threshold = psd_ctx->threshold;
> +	}
> +
> +	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;
> +	}
> +
> +	return 0;
> +}
> +
> +static int
> +scheduler_stop(struct rte_cryptodev *dev)
> +{
> +	uint16_t i;
> +
> +	for (i = 0; i < dev->data->nb_queue_pairs; i++) {
> +		struct scheduler_qp_ctx *qp_ctx = dev->data->queue_pairs[i];
> +		struct psd_scheduler_qp_ctx *ps_qp_ctx = qp_ctx->private_qp_ctx;
> +
> +		if (ps_qp_ctx->primary_slave.nb_inflight_cops +
> +				ps_qp_ctx->secondary_slave.nb_inflight_cops) {
> +			CS_LOG_ERR("Some crypto ops left in slave queue");
> +			return -1;
> +		}
> +	}
> +
> +	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 psd_scheduler_qp_ctx *ps_qp_ctx;
> +
> +	ps_qp_ctx = rte_zmalloc_socket(NULL, sizeof(*ps_qp_ctx), 0,
> +			rte_socket_id());
> +	if (!ps_qp_ctx) {
> +		CS_LOG_ERR("failed allocate memory for private queue pair");
> +		return -ENOMEM;
> +	}
> +
> +	qp_ctx->private_qp_ctx = (void *)ps_qp_ctx;
> +
> +	return 0;
> +}
> +
> +static int
> +scheduler_create_private_ctx(struct rte_cryptodev *dev)
> +{
> +	struct scheduler_ctx *sched_ctx = dev->data->dev_private;
> +	struct psd_scheduler_ctx *psd_ctx;
> +
> +	if (sched_ctx->private_ctx)
> +		rte_free(sched_ctx->private_ctx);
> +
> +	psd_ctx = rte_zmalloc_socket(NULL, sizeof(struct psd_scheduler_ctx), 0,
> +			rte_socket_id());
> +	if (!psd_ctx) {
> +		CS_LOG_ERR("failed allocate memory");
> +		return -ENOMEM;
> +	}
> +
> +	psd_ctx->threshold = PKT_SIZE_THRESHOLD;
> +
> +	sched_ctx->private_ctx = (void *)psd_ctx;
> +
> +	return 0;
> +}
> +
> +struct rte_cryptodev_scheduler_ops scheduler_ps_ops = {
> +	slave_attach,
> +	slave_detach,
> +	scheduler_start,
> +	scheduler_stop,
> +	scheduler_config_qp,
> +	scheduler_create_private_ctx,
> +};
> +
> +struct rte_cryptodev_scheduler psd_scheduler = {
> +		.name = "packet-size-based-scheduler",
> +		.description = "scheduler which will distribute crypto op "
> +				"burst based on the packet size",
> +		.mode = CDEV_SCHED_MODE_PKT_SIZE_DISTR,
> +		.ops = &scheduler_ps_ops
> +};
> +
> +struct rte_cryptodev_scheduler *pkt_size_based_distr_scheduler = &psd_scheduler;
>



More information about the dev mailing list