patches for DPDK stable branches
 help / color / Atom feed
* [dpdk-stable] [PATCH v2 4/5] crypto/ccp: fix support scheduling CCP's on burst
@ 2019-11-22  6:47 asomalap
  0 siblings, 0 replies; only message in thread
From: asomalap @ 2019-11-22  6:47 UTC (permalink / raw)
  To: dev; +Cc: stable

From: Amaranath Somalapuram <asomalap@amd.com>

CCP driver was scheduling only one CCP in a single burst(enqueue).
Effective throughput was limited to 1 CCP performance.
Scheduling multiple ccp within one burst will increase the ccp performance.
this changes will divide the enqueue packets equally among the multiple CCP

Fixes: e0d88a394e ("crypto/ccp: support run-time CPU based auth")
Cc: stable@dpdk.org

Signed-off-by: Amaranath Somalapuram <asomalap@amd.com>
---
 drivers/crypto/ccp/ccp_crypto.c      | 22 +++++++----
 drivers/crypto/ccp/ccp_crypto.h      |  7 +++-
 drivers/crypto/ccp/ccp_dev.h         |  2 +-
 drivers/crypto/ccp/ccp_pmd_private.h |  2 +
 drivers/crypto/ccp/rte_ccp_pmd.c     | 57 +++++++++++++++++++---------
 5 files changed, 63 insertions(+), 27 deletions(-)

diff --git a/drivers/crypto/ccp/ccp_crypto.c b/drivers/crypto/ccp/ccp_crypto.c
index 1837c8543..4256734d1 100644
--- a/drivers/crypto/ccp/ccp_crypto.c
+++ b/drivers/crypto/ccp/ccp_crypto.c
@@ -2680,7 +2680,9 @@ process_ops_to_enqueue(struct ccp_qp *qp,
 		       struct rte_crypto_op **op,
 		       struct ccp_queue *cmd_q,
 		       uint16_t nb_ops,
-		       int slots_req)
+		       uint16_t total_nb_ops,
+		       int slots_req,
+		       uint16_t b_idx)
 {
 	int i, result = 0;
 	struct ccp_batch_info *b_info;
@@ -2701,6 +2703,7 @@ process_ops_to_enqueue(struct ccp_qp *qp,
 
 	/* populate batch info necessary for dequeue */
 	b_info->op_idx = 0;
+	b_info->b_idx = 0;
 	b_info->lsb_buf_idx = 0;
 	b_info->desccnt = 0;
 	b_info->cmd_q = cmd_q;
@@ -2710,7 +2713,7 @@ process_ops_to_enqueue(struct ccp_qp *qp,
 
 	b_info->head_offset = (uint32_t)(cmd_q->qbase_phys_addr + cmd_q->qidx *
 					 Q_DESC_SIZE);
-	for (i = 0; i < nb_ops; i++) {
+	for (i = b_idx; i < (nb_ops+b_idx); i++) {
 		session = (struct ccp_session *)get_sym_session_private_data(
 						 op[i]->sym->session,
 						 ccp_cryptodev_driver_id);
@@ -2762,6 +2765,8 @@ process_ops_to_enqueue(struct ccp_qp *qp,
 	}
 
 	b_info->opcnt = i;
+	b_info->b_idx = b_idx;
+	b_info->total_nb_ops = total_nb_ops;
 	b_info->tail_offset = (uint32_t)(cmd_q->qbase_phys_addr + cmd_q->qidx *
 					 Q_DESC_SIZE);
 
@@ -2776,7 +2781,7 @@ process_ops_to_enqueue(struct ccp_qp *qp,
 	rte_ring_enqueue(qp->processed_pkts, (void *)b_info);
 
 	EVP_MD_CTX_destroy(auth_ctx);
-	return i;
+	return i-b_idx;
 }
 
 static inline void ccp_auth_dq_prepare(struct rte_crypto_op *op)
@@ -2861,8 +2866,8 @@ ccp_prepare_ops(struct ccp_qp *qp,
 	}
 	min_ops = RTE_MIN(nb_ops, b_info->opcnt);
 
-	for (i = 0; i < min_ops; i++) {
-		op_d[i] = b_info->op[b_info->op_idx++];
+	for (i =  b_info->b_idx; i < min_ops; i++) {
+		op_d[i] = b_info->op[b_info->b_idx + b_info->op_idx++];
 		session = (struct ccp_session *)get_sym_session_private_data(
 						 op_d[i]->sym->session,
 						ccp_cryptodev_driver_id);
@@ -2903,7 +2908,8 @@ ccp_prepare_ops(struct ccp_qp *qp,
 int
 process_ops_to_dequeue(struct ccp_qp *qp,
 		       struct rte_crypto_op **op,
-		       uint16_t nb_ops)
+		       uint16_t nb_ops,
+		       uint16_t *total_nb_ops)
 {
 	struct ccp_batch_info *b_info;
 	uint32_t cur_head_offset;
@@ -2918,6 +2924,7 @@ process_ops_to_dequeue(struct ccp_qp *qp,
 
 	if (b_info->auth_ctr == b_info->opcnt)
 		goto success;
+	*total_nb_ops = b_info->total_nb_ops;
 	cur_head_offset = CCP_READ_REG(b_info->cmd_q->reg_base,
 				       CMD_Q_HEAD_LO_BASE);
 
@@ -2927,7 +2934,7 @@ process_ops_to_dequeue(struct ccp_qp *qp,
 			qp->b_info = b_info;
 			return 0;
 		}
-	} else {
+	} else if (b_info->tail_offset != b_info->head_offset) {
 		if ((cur_head_offset >= b_info->head_offset) ||
 		    (cur_head_offset < b_info->tail_offset)) {
 			qp->b_info = b_info;
@@ -2937,6 +2944,7 @@ process_ops_to_dequeue(struct ccp_qp *qp,
 
 
 success:
+	*total_nb_ops = b_info->total_nb_ops;
 	nb_ops = ccp_prepare_ops(qp, op, b_info, nb_ops);
 	rte_atomic64_add(&b_info->cmd_q->free_slots, b_info->desccnt);
 	b_info->desccnt = 0;
diff --git a/drivers/crypto/ccp/ccp_crypto.h b/drivers/crypto/ccp/ccp_crypto.h
index 882b398ac..8e6d03efc 100644
--- a/drivers/crypto/ccp/ccp_crypto.h
+++ b/drivers/crypto/ccp/ccp_crypto.h
@@ -353,7 +353,9 @@ int process_ops_to_enqueue(struct ccp_qp *qp,
 			   struct rte_crypto_op **op,
 			   struct ccp_queue *cmd_q,
 			   uint16_t nb_ops,
-			   int slots_req);
+			   uint16_t total_nb_ops,
+			   int slots_req,
+			   uint16_t b_idx);
 
 /**
  * process crypto ops to be dequeued
@@ -365,7 +367,8 @@ int process_ops_to_enqueue(struct ccp_qp *qp,
  */
 int process_ops_to_dequeue(struct ccp_qp *qp,
 			   struct rte_crypto_op **op,
-			   uint16_t nb_ops);
+			   uint16_t nb_ops,
+			   uint16_t *total_nb_ops);
 
 
 /**
diff --git a/drivers/crypto/ccp/ccp_dev.h b/drivers/crypto/ccp/ccp_dev.h
index de3e4bcc6..f4ad9eafd 100644
--- a/drivers/crypto/ccp/ccp_dev.h
+++ b/drivers/crypto/ccp/ccp_dev.h
@@ -59,7 +59,7 @@
 #define CMD_Q_RUN			0x1
 #define CMD_Q_SIZE			0x1F
 #define CMD_Q_SHIFT			3
-#define COMMANDS_PER_QUEUE		2048
+#define COMMANDS_PER_QUEUE		8192
 
 #define QUEUE_SIZE_VAL                  ((ffs(COMMANDS_PER_QUEUE) - 2) & \
 					 CMD_Q_SIZE)
diff --git a/drivers/crypto/ccp/ccp_pmd_private.h b/drivers/crypto/ccp/ccp_pmd_private.h
index 781050c31..1c4118ee3 100644
--- a/drivers/crypto/ccp/ccp_pmd_private.h
+++ b/drivers/crypto/ccp/ccp_pmd_private.h
@@ -50,8 +50,10 @@ struct ccp_batch_info {
 	struct rte_crypto_op *op[CCP_MAX_BURST];
 	/**< optable populated at enque time from app*/
 	int op_idx;
+	uint16_t b_idx;
 	struct ccp_queue *cmd_q;
 	uint16_t opcnt;
+	uint16_t total_nb_ops;
 	/**< no. of crypto ops in batch*/
 	int desccnt;
 	/**< no. of ccp queue descriptors*/
diff --git a/drivers/crypto/ccp/rte_ccp_pmd.c b/drivers/crypto/ccp/rte_ccp_pmd.c
index 4810d799c..38cb1fe3d 100644
--- a/drivers/crypto/ccp/rte_ccp_pmd.c
+++ b/drivers/crypto/ccp/rte_ccp_pmd.c
@@ -22,6 +22,7 @@
  */
 static unsigned int ccp_pmd_init_done;
 uint8_t ccp_cryptodev_driver_id;
+uint8_t cryptodev_cnt;
 
 struct ccp_pmd_init_params {
 	struct rte_cryptodev_pmd_init_params def_p;
@@ -201,30 +202,46 @@ ccp_pmd_enqueue_burst(void *queue_pair, struct rte_crypto_op **ops,
 	struct ccp_queue *cmd_q;
 	struct rte_cryptodev *dev = qp->dev;
 	uint16_t i, enq_cnt = 0, slots_req = 0;
+	uint16_t tmp_ops = nb_ops, b_idx, cur_ops = 0;
 
 	if (nb_ops == 0)
 		return 0;
 
 	if (unlikely(rte_ring_full(qp->processed_pkts) != 0))
 		return 0;
+	if (tmp_ops >= cryptodev_cnt)
+		cur_ops = nb_ops / cryptodev_cnt + (nb_ops)%cryptodev_cnt;
+	else
+		cur_ops = tmp_ops;
+	while (tmp_ops)	{
+		b_idx = nb_ops - tmp_ops;
+		slots_req = 0;
+		if (cur_ops <= tmp_ops) {
+			tmp_ops -= cur_ops;
+		} else {
+			cur_ops = tmp_ops;
+			tmp_ops = 0;
+		}
+		for (i = 0; i < cur_ops; i++) {
+			sess = get_ccp_session(qp, ops[i + b_idx]);
+			if (unlikely(sess == NULL) && (i == 0)) {
+				qp->qp_stats.enqueue_err_count++;
+				return 0;
+			} else if (sess == NULL) {
+				cur_ops = i;
+				break;
+			}
+			slots_req += ccp_compute_slot_count(sess);
+		}
 
-	for (i = 0; i < nb_ops; i++) {
-		sess = get_ccp_session(qp, ops[i]);
-		if (unlikely(sess == NULL) && (i == 0)) {
-			qp->qp_stats.enqueue_err_count++;
+		cmd_q = ccp_allot_queue(dev, slots_req);
+		if (unlikely(cmd_q == NULL))
 			return 0;
-		} else if (sess == NULL) {
-			nb_ops = i;
-			break;
-		}
-		slots_req += ccp_compute_slot_count(sess);
+		enq_cnt += process_ops_to_enqueue(qp, ops, cmd_q, cur_ops,
+				nb_ops, slots_req, b_idx);
+		i++;
 	}
 
-	cmd_q = ccp_allot_queue(dev, slots_req);
-	if (unlikely(cmd_q == NULL))
-		return 0;
-
-	enq_cnt = process_ops_to_enqueue(qp, ops, cmd_q, nb_ops, slots_req);
 	qp->qp_stats.enqueued_count += enq_cnt;
 	return enq_cnt;
 }
@@ -234,9 +251,16 @@ ccp_pmd_dequeue_burst(void *queue_pair, struct rte_crypto_op **ops,
 		uint16_t nb_ops)
 {
 	struct ccp_qp *qp = queue_pair;
-	uint16_t nb_dequeued = 0, i;
+	uint16_t nb_dequeued = 0, i, total_nb_ops;
+
+	nb_dequeued = process_ops_to_dequeue(qp, ops, nb_ops, &total_nb_ops);
 
-	nb_dequeued = process_ops_to_dequeue(qp, ops, nb_ops);
+	if (total_nb_ops) {
+		while (nb_dequeued != total_nb_ops) {
+			nb_dequeued = process_ops_to_dequeue(qp,
+					ops, nb_ops, &total_nb_ops);
+		}
+	}
 
 	/* Free session if a session-less crypto op */
 	for (i = 0; i < nb_dequeued; i++)
@@ -296,7 +320,6 @@ cryptodev_ccp_create(const char *name,
 {
 	struct rte_cryptodev *dev;
 	struct ccp_private *internals;
-	uint8_t cryptodev_cnt = 0;
 
 	if (init_params->def_p.name[0] == '\0')
 		strlcpy(init_params->def_p.name, name,
-- 
2.17.1


^ permalink raw reply	[flat|nested] only message in thread

only message in thread, back to index

Thread overview: (only message) (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-11-22  6:47 [dpdk-stable] [PATCH v2 4/5] crypto/ccp: fix support scheduling CCP's on burst asomalap

patches for DPDK stable branches

Archives are clonable:
	git clone --mirror http://inbox.dpdk.org/stable/0 stable/git/0.git

	# If you have public-inbox 1.1+ installed, you may
	# initialize and index your mirror using the following commands:
	public-inbox-init -V2 stable stable/ http://inbox.dpdk.org/stable \
		stable@dpdk.org
	public-inbox-index stable


Newsgroup available over NNTP:
	nntp://inbox.dpdk.org/inbox.dpdk.stable


AGPL code for this site: git clone https://public-inbox.org/ public-inbox