[dpdk-dev] [PATCH v2 06/20] net/dpaa2: support parallel Rx mode

Hemant Agrawal hemant.agrawal at nxp.com
Mon May 15 14:37:30 CEST 2017


Typically when the PMD issues a RX command to DPAA2 hardware,
the hw writes the available descriptors into the given memory.
The RX function then processes the frames and prepare them as
mbufs.

This patch adds support to issue another pull request to hardware
in another memory location, before we start processing the output
of the first request. This help in controlling the cpu cycles
wasted during the wait for the hardware to write the descriptors.

Signed-off-by: Hemant Agrawal <hemant.agrawal at nxp.com>
---
 drivers/net/dpaa2/dpaa2_ethdev.c |   4 +-
 drivers/net/dpaa2/dpaa2_ethdev.h |   3 +-
 drivers/net/dpaa2/dpaa2_rxtx.c   | 119 ++++++++++++++++++++++++---------------
 3 files changed, 79 insertions(+), 47 deletions(-)

diff --git a/drivers/net/dpaa2/dpaa2_ethdev.c b/drivers/net/dpaa2/dpaa2_ethdev.c
index 698fb84..797e71d 100644
--- a/drivers/net/dpaa2/dpaa2_ethdev.c
+++ b/drivers/net/dpaa2/dpaa2_ethdev.c
@@ -391,7 +391,7 @@
 		RTE_PTYPE_UNKNOWN
 	};
 
-	if (dev->rx_pkt_burst == dpaa2_dev_rx)
+	if (dev->rx_pkt_burst == dpaa2_dev_prefetch_rx)
 		return ptypes;
 	return NULL;
 }
@@ -886,7 +886,7 @@ void dpaa2_dev_stats_reset(struct rte_eth_dev *dev)
 	eth_dev->dev_ops = &dpaa2_ethdev_ops;
 	eth_dev->data->drv_name = rte_dpaa2_pmd.driver.name;
 
-	eth_dev->rx_pkt_burst = dpaa2_dev_rx;
+	eth_dev->rx_pkt_burst = dpaa2_dev_prefetch_rx;
 	eth_dev->tx_pkt_burst = dpaa2_dev_tx;
 	rte_fslmc_vfio_dmamap();
 
diff --git a/drivers/net/dpaa2/dpaa2_ethdev.h b/drivers/net/dpaa2/dpaa2_ethdev.h
index 7196398..7fa7e7d 100644
--- a/drivers/net/dpaa2/dpaa2_ethdev.h
+++ b/drivers/net/dpaa2/dpaa2_ethdev.h
@@ -77,7 +77,8 @@ int dpaa2_remove_flow_dist(struct rte_eth_dev *eth_dev,
 
 int dpaa2_attach_bp_list(struct dpaa2_dev_priv *priv, void *blist);
 
-uint16_t dpaa2_dev_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts);
+uint16_t dpaa2_dev_prefetch_rx(void *queue, struct rte_mbuf **bufs,
+			       uint16_t nb_pkts);
 uint16_t dpaa2_dev_tx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts);
 
 #endif /* _DPAA2_ETHDEV_H */
diff --git a/drivers/net/dpaa2/dpaa2_rxtx.c b/drivers/net/dpaa2/dpaa2_rxtx.c
index c5d49cb..9b7539a 100644
--- a/drivers/net/dpaa2/dpaa2_rxtx.c
+++ b/drivers/net/dpaa2/dpaa2_rxtx.c
@@ -242,17 +242,18 @@ static inline int __attribute__((hot))
 }
 
 uint16_t
-dpaa2_dev_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
+dpaa2_dev_prefetch_rx(void *queue, struct rte_mbuf **bufs, uint16_t nb_pkts)
 {
-	/* Function is responsible to receive frames for a given device and VQ*/
+	/* Function receive frames for a given device and VQ*/
 	struct dpaa2_queue *dpaa2_q = (struct dpaa2_queue *)queue;
 	struct qbman_result *dq_storage;
 	uint32_t fqid = dpaa2_q->fqid;
 	int ret, num_rx = 0;
 	uint8_t is_last = 0, status;
 	struct qbman_swp *swp;
-	const struct qbman_fd *fd;
+	const struct qbman_fd *fd[DPAA2_DQRR_RING_SIZE];
 	struct qbman_pull_desc pulldesc;
+	struct queue_storage_info_t *q_storage = dpaa2_q->q_storage;
 	struct rte_eth_dev *dev = dpaa2_q->dev;
 
 	if (unlikely(!DPAA2_PER_LCORE_DPIO)) {
@@ -263,44 +264,51 @@ static inline int __attribute__((hot))
 		}
 	}
 	swp = DPAA2_PER_LCORE_PORTAL;
-	dq_storage = dpaa2_q->q_storage->dq_storage[0];
-
-	qbman_pull_desc_clear(&pulldesc);
-	qbman_pull_desc_set_numframes(&pulldesc,
-				      (nb_pkts > DPAA2_DQRR_RING_SIZE) ?
-				       DPAA2_DQRR_RING_SIZE : nb_pkts);
-	qbman_pull_desc_set_fq(&pulldesc, fqid);
-	/* todo optimization - we can have dq_storage_phys available*/
-	qbman_pull_desc_set_storage(&pulldesc, dq_storage,
+	if (!q_storage->active_dqs) {
+		q_storage->toggle = 0;
+		dq_storage = q_storage->dq_storage[q_storage->toggle];
+		qbman_pull_desc_clear(&pulldesc);
+		qbman_pull_desc_set_numframes(&pulldesc,
+					      (nb_pkts > DPAA2_DQRR_RING_SIZE) ?
+					       DPAA2_DQRR_RING_SIZE : nb_pkts);
+		qbman_pull_desc_set_fq(&pulldesc, fqid);
+		qbman_pull_desc_set_storage(&pulldesc, dq_storage,
 			(dma_addr_t)(DPAA2_VADDR_TO_IOVA(dq_storage)), 1);
-
-	/*Issue a volatile dequeue command. */
-	while (1) {
-		if (qbman_swp_pull(swp, &pulldesc)) {
-			PMD_RX_LOG(ERR, "VDQ command is not issued."
-				   "QBMAN is busy\n");
-			/* Portal was busy, try again */
-			continue;
+		if (check_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index)) {
+			while (!qbman_check_command_complete(swp,
+			       get_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index)))
+				;
+			clear_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index);
 		}
-		break;
-	};
-
-	/* Receive the packets till Last Dequeue entry is found with
-	 * respect to the above issues PULL command.
+		while (1) {
+			if (qbman_swp_pull(swp, &pulldesc)) {
+				PMD_RX_LOG(WARNING, "VDQ command is not issued."
+					   "QBMAN is busy\n");
+				/* Portal was busy, try again */
+				continue;
+			}
+			break;
+		}
+		q_storage->active_dqs = dq_storage;
+		q_storage->active_dpio_id = DPAA2_PER_LCORE_DPIO->index;
+		set_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index, dq_storage);
+	}
+	dq_storage = q_storage->active_dqs;
+	/* Check if the previous issued command is completed.
+	 * Also seems like the SWP is shared between the Ethernet Driver
+	 * and the SEC driver.
 	 */
+	while (!qbman_check_command_complete(swp, dq_storage))
+		;
+	if (dq_storage == get_swp_active_dqs(q_storage->active_dpio_id))
+		clear_swp_active_dqs(q_storage->active_dpio_id);
 	while (!is_last) {
-		struct rte_mbuf *mbuf;
-		/*Check if the previous issued command is completed.
-		 * Also seems like the SWP is shared between the
-		 * Ethernet Driver and the SEC driver.
-		 */
-		while (!qbman_check_command_complete(swp, dq_storage))
-			;
 		/* Loop until the dq_storage is updated with
 		 * new token by QBMAN
 		 */
 		while (!qbman_result_has_new_result(swp, dq_storage))
 			;
+		rte_prefetch0((void *)((uint64_t)(dq_storage + 1)));
 		/* Check whether Last Pull command is Expired and
 		 * setting Condition for Loop termination
 		 */
@@ -311,27 +319,50 @@ static inline int __attribute__((hot))
 			if (unlikely((status & QBMAN_DQ_STAT_VALIDFRAME) == 0))
 				continue;
 		}
+		fd[num_rx] = qbman_result_DQ_fd(dq_storage);
 
-		fd = qbman_result_DQ_fd(dq_storage);
-		mbuf = (struct rte_mbuf *)DPAA2_IOVA_TO_VADDR(
-		   DPAA2_GET_FD_ADDR(fd)
-		   - rte_dpaa2_bpid_info[DPAA2_GET_FD_BPID(fd)].meta_data_size);
-		/* Prefeth mbuf */
-		rte_prefetch0(mbuf);
 		/* Prefetch Annotation address for the parse results */
-		rte_prefetch0((void *)((uint64_t)DPAA2_GET_FD_ADDR(fd)
-						+ DPAA2_FD_PTA_SIZE + 16));
+		rte_prefetch0((void *)((uint64_t)DPAA2_GET_FD_ADDR(fd[num_rx])
+				+ DPAA2_FD_PTA_SIZE + 16));
 
-		bufs[num_rx] = eth_fd_to_mbuf(fd);
+		bufs[num_rx] = eth_fd_to_mbuf(fd[num_rx]);
 		bufs[num_rx]->port = dev->data->port_id;
+		if (dev->data->dev_conf.rxmode.hw_vlan_strip)
+			rte_vlan_strip(bufs[num_rx]);
 
-		num_rx++;
 		dq_storage++;
-	} /* End of Packet Rx loop */
+		num_rx++;
+	}
+
+	if (check_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index)) {
+		while (!qbman_check_command_complete(swp,
+		       get_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index)))
+			;
+		clear_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index);
+	}
+	q_storage->toggle ^= 1;
+	dq_storage = q_storage->dq_storage[q_storage->toggle];
+	qbman_pull_desc_clear(&pulldesc);
+	qbman_pull_desc_set_numframes(&pulldesc, DPAA2_DQRR_RING_SIZE);
+	qbman_pull_desc_set_fq(&pulldesc, fqid);
+	qbman_pull_desc_set_storage(&pulldesc, dq_storage,
+			(dma_addr_t)(DPAA2_VADDR_TO_IOVA(dq_storage)), 1);
+	/* Issue a volatile dequeue command. */
+	while (1) {
+		if (qbman_swp_pull(swp, &pulldesc)) {
+			PMD_RX_LOG(WARNING, "VDQ command is not issued."
+				   "QBMAN is busy\n");
+			continue;
+		}
+		break;
+	}
+	q_storage->active_dqs = dq_storage;
+	q_storage->active_dpio_id = DPAA2_PER_LCORE_DPIO->index;
+	set_swp_active_dqs(DPAA2_PER_LCORE_DPIO->index, dq_storage);
 
 	dpaa2_q->rx_pkts += num_rx;
 
-	/*Return the total number of packets received to DPAA2 app*/
+	/* Return the total number of packets received to DPAA2 app */
 	return num_rx;
 }
 
-- 
1.9.1



More information about the dev mailing list