DPDK patches and discussions
 help / color / mirror / Atom feed
* [dpdk-dev] [PATCH 1/2] net/qede: fix ovs-dpdk failure when using odd number of queues on 100Gb mode
@ 2019-09-04 11:30 Shahed Shaikh
  2019-09-04 11:30 ` [dpdk-dev] [PATCH 2/2] net/qede: implement rte_flow drop action Shahed Shaikh
  2019-09-04 13:30 ` [dpdk-dev] [PATCH 1/2] net/qede: fix ovs-dpdk failure when using odd number of queues on 100Gb mode Jerin Jacob Kollanukkaran
  0 siblings, 2 replies; 5+ messages in thread
From: Shahed Shaikh @ 2019-09-04 11:30 UTC (permalink / raw)
  To: dev; +Cc: rmody, ferruh.yigit, GR-Everest-DPDK-Dev, stable

As per HW design of 100Gb mode, device internally uses 2 engines
(eng0 and eng1), and both engines need to be configured symmetrically.
Based on this requirement, driver design chose an approach
to allow user to allocate only even number of queues and split
those queues on both engines equally.

This approach puts a limitation on number of queues to be allocated -
i.e. user can't configure odd number of queues on 100Gb mode.
OVS configures DPDK port with 1 rxq and 1 txq, which causes initialization
of qede port to fail.

This patch changes the implementation of queue allocation method
for 100Gb devices by removing above mentioned limitation and allowing
user to configure odd number of queues.

Key changes in this patch -
 - Allocate requested queue count on both engines, so that
   actual hardware queue count will be double of what user requested.
 - Create a pair of queues from both engines and provide it to
   rte_ethdev queue structure. So ethdev will see only one queue for
   underlying queue pair created for hw engine pair.
 - Rx and Tx methods from ethdev will provide that queue pair
   object and PMD will internally split Rx and Tx packet processing across
   both engines in separately installed Rx and Tx handlers.
 - Consolidate statistics of both HW queues while reporting to application.
 - Report engine wise queue statistics in xstats flow.
   e.g. - rx_q<hw_eng_id>.<qid>_xxxxxxx

Fixes: 2af14ca79c0a ("net/qede: support 100G")
Cc: stable@dpdk.org

Signed-off-by: Shahed Shaikh <shshaikh@marvell.com>
---
 drivers/net/qede/qede_ethdev.c | 326 +++++++++++++----------------
 drivers/net/qede/qede_ethdev.h |   5 +-
 drivers/net/qede/qede_filter.c |   5 +-
 drivers/net/qede/qede_rxtx.c   | 368 ++++++++++++++++++++++++---------
 drivers/net/qede/qede_rxtx.h   |  26 ++-
 5 files changed, 440 insertions(+), 290 deletions(-)

diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index 528b33e8c..a033807e7 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -304,6 +304,7 @@ static void qede_print_adapter_info(struct qede_dev *qdev)
 
 static void qede_reset_queue_stats(struct qede_dev *qdev, bool xstats)
 {
+	struct rte_eth_dev *dev = (struct rte_eth_dev *)qdev->ethdev;
 	struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
 	unsigned int i = 0, j = 0, qid;
 	unsigned int rxq_stat_cntrs, txq_stat_cntrs;
@@ -311,12 +312,12 @@ static void qede_reset_queue_stats(struct qede_dev *qdev, bool xstats)
 
 	DP_VERBOSE(edev, ECORE_MSG_DEBUG, "Clearing queue stats\n");
 
-	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(dev),
 			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
-	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(qdev),
+	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(dev),
 			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
 
-	for_each_rss(qid) {
+	for (qid = 0; qid < qdev->num_rx_queues; qid++) {
 		OSAL_MEMSET(((char *)(qdev->fp_array[qid].rxq)) +
 			     offsetof(struct qede_rx_queue, rcv_pkts), 0,
 			    sizeof(uint64_t));
@@ -342,7 +343,7 @@ static void qede_reset_queue_stats(struct qede_dev *qdev, bool xstats)
 
 	i = 0;
 
-	for_each_tss(qid) {
+	for (qid = 0; qid < qdev->num_tx_queues; qid++) {
 		txq = qdev->fp_array[qid].txq;
 
 		OSAL_MEMSET((uint64_t *)(uintptr_t)
@@ -991,7 +992,7 @@ int qede_config_rss(struct rte_eth_dev *eth_dev)
 	for (i = 0; i < ECORE_RSS_IND_TABLE_SIZE; i++) {
 		id = i / RTE_RETA_GROUP_SIZE;
 		pos = i % RTE_RETA_GROUP_SIZE;
-		q = i % QEDE_RSS_COUNT(qdev);
+		q = i % QEDE_RSS_COUNT(eth_dev);
 		reta_conf[id].reta[pos] = q;
 	}
 	if (qede_rss_reta_update(eth_dev, &reta_conf[0],
@@ -1165,22 +1166,6 @@ static int qede_dev_configure(struct rte_eth_dev *eth_dev)
 
 	PMD_INIT_FUNC_TRACE(edev);
 
-	/* Check requirements for 100G mode */
-	if (ECORE_IS_CMT(edev)) {
-		if (eth_dev->data->nb_rx_queues < 2 ||
-		    eth_dev->data->nb_tx_queues < 2) {
-			DP_ERR(edev, "100G mode needs min. 2 RX/TX queues\n");
-			return -EINVAL;
-		}
-
-		if ((eth_dev->data->nb_rx_queues % 2 != 0) ||
-		    (eth_dev->data->nb_tx_queues % 2 != 0)) {
-			DP_ERR(edev,
-			       "100G mode needs even no. of RX/TX queues\n");
-			return -EINVAL;
-		}
-	}
-
 	/* We need to have min 1 RX queue.There is no min check in
 	 * rte_eth_dev_configure(), so we are checking it here.
 	 */
@@ -1207,8 +1192,9 @@ static int qede_dev_configure(struct rte_eth_dev *eth_dev)
 		return -ENOTSUP;
 
 	qede_dealloc_fp_resc(eth_dev);
-	qdev->num_tx_queues = eth_dev->data->nb_tx_queues;
-	qdev->num_rx_queues = eth_dev->data->nb_rx_queues;
+	qdev->num_tx_queues = eth_dev->data->nb_tx_queues * edev->num_hwfns;
+	qdev->num_rx_queues = eth_dev->data->nb_rx_queues * edev->num_hwfns;
+
 	if (qede_alloc_fp_resc(qdev))
 		return -ENOMEM;
 
@@ -1233,7 +1219,12 @@ static int qede_dev_configure(struct rte_eth_dev *eth_dev)
 		return ret;
 
 	DP_INFO(edev, "Device configured with RSS=%d TSS=%d\n",
-			QEDE_RSS_COUNT(qdev), QEDE_TSS_COUNT(qdev));
+			QEDE_RSS_COUNT(eth_dev), QEDE_TSS_COUNT(eth_dev));
+
+	if (ECORE_IS_CMT(edev))
+		DP_INFO(edev, "Actual HW queues for CMT mode - RX = %d TX = %d\n",
+			qdev->num_rx_queues, qdev->num_tx_queues);
+
 
 	return 0;
 }
@@ -1275,6 +1266,10 @@ qede_dev_info_get(struct rte_eth_dev *eth_dev,
 	else
 		dev_info->max_rx_queues = (uint16_t)RTE_MIN(
 			QEDE_MAX_RSS_CNT(qdev), ECORE_MAX_VF_CHAINS_PER_PF);
+	/* Since CMT mode internally doubles the number of queues */
+	if (ECORE_IS_CMT(edev))
+		dev_info->max_rx_queues  = dev_info->max_rx_queues / 2;
+
 	dev_info->max_tx_queues = dev_info->max_rx_queues;
 
 	dev_info->max_mac_addrs = qdev->dev_info.num_mac_filters;
@@ -1482,7 +1477,7 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	struct qede_dev *qdev = eth_dev->data->dev_private;
 	struct ecore_dev *edev = &qdev->edev;
 	struct ecore_eth_stats stats;
-	unsigned int i = 0, j = 0, qid;
+	unsigned int i = 0, j = 0, qid, idx, hw_fn;
 	unsigned int rxq_stat_cntrs, txq_stat_cntrs;
 	struct qede_tx_queue *txq;
 
@@ -1518,44 +1513,59 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	eth_stats->oerrors = stats.common.tx_err_drop_pkts;
 
 	/* Queue stats */
-	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(eth_dev),
 			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
-	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(qdev),
+	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(eth_dev),
 			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
-	if ((rxq_stat_cntrs != (unsigned int)QEDE_RSS_COUNT(qdev)) ||
-	    (txq_stat_cntrs != (unsigned int)QEDE_TSS_COUNT(qdev)))
+	if (rxq_stat_cntrs != (unsigned int)QEDE_RSS_COUNT(eth_dev) ||
+	    txq_stat_cntrs != (unsigned int)QEDE_TSS_COUNT(eth_dev))
 		DP_VERBOSE(edev, ECORE_MSG_DEBUG,
 		       "Not all the queue stats will be displayed. Set"
 		       " RTE_ETHDEV_QUEUE_STAT_CNTRS config param"
 		       " appropriately and retry.\n");
 
-	for_each_rss(qid) {
-		eth_stats->q_ipackets[i] =
-			*(uint64_t *)(
-				((char *)(qdev->fp_array[qid].rxq)) +
-				offsetof(struct qede_rx_queue,
-				rcv_pkts));
-		eth_stats->q_errors[i] =
-			*(uint64_t *)(
-				((char *)(qdev->fp_array[qid].rxq)) +
-				offsetof(struct qede_rx_queue,
-				rx_hw_errors)) +
-			*(uint64_t *)(
-				((char *)(qdev->fp_array[qid].rxq)) +
-				offsetof(struct qede_rx_queue,
-				rx_alloc_errors));
+	for (qid = 0; qid < eth_dev->data->nb_rx_queues; qid++) {
+		eth_stats->q_ipackets[i] = 0;
+		eth_stats->q_errors[i] = 0;
+
+		for_each_hwfn(edev, hw_fn) {
+			idx = qid * edev->num_hwfns + hw_fn;
+
+			eth_stats->q_ipackets[i] +=
+				*(uint64_t *)
+					(((char *)(qdev->fp_array[idx].rxq)) +
+					 offsetof(struct qede_rx_queue,
+					 rcv_pkts));
+			eth_stats->q_errors[i] +=
+				*(uint64_t *)
+					(((char *)(qdev->fp_array[idx].rxq)) +
+					 offsetof(struct qede_rx_queue,
+					 rx_hw_errors)) +
+				*(uint64_t *)
+					(((char *)(qdev->fp_array[idx].rxq)) +
+					 offsetof(struct qede_rx_queue,
+					 rx_alloc_errors));
+		}
+
 		i++;
 		if (i == rxq_stat_cntrs)
 			break;
 	}
 
-	for_each_tss(qid) {
-		txq = qdev->fp_array[qid].txq;
-		eth_stats->q_opackets[j] =
-			*((uint64_t *)(uintptr_t)
-				(((uint64_t)(uintptr_t)(txq)) +
-				 offsetof(struct qede_tx_queue,
-					  xmit_pkts)));
+	for (qid = 0; qid < eth_dev->data->nb_tx_queues; qid++) {
+		eth_stats->q_opackets[j] = 0;
+
+		for_each_hwfn(edev, hw_fn) {
+			idx = qid * edev->num_hwfns + hw_fn;
+
+			txq = qdev->fp_array[idx].txq;
+			eth_stats->q_opackets[j] +=
+				*((uint64_t *)(uintptr_t)
+					(((uint64_t)(uintptr_t)(txq)) +
+					 offsetof(struct qede_tx_queue,
+						  xmit_pkts)));
+		}
+
 		j++;
 		if (j == txq_stat_cntrs)
 			break;
@@ -1566,18 +1576,18 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 
 static unsigned
 qede_get_xstats_count(struct qede_dev *qdev) {
+	struct rte_eth_dev *dev = (struct rte_eth_dev *)qdev->ethdev;
+
 	if (ECORE_IS_BB(&qdev->edev))
 		return RTE_DIM(qede_xstats_strings) +
 		       RTE_DIM(qede_bb_xstats_strings) +
 		       (RTE_DIM(qede_rxq_xstats_strings) *
-			RTE_MIN(QEDE_RSS_COUNT(qdev),
-				RTE_ETHDEV_QUEUE_STAT_CNTRS));
+			QEDE_RSS_COUNT(dev) * qdev->edev.num_hwfns);
 	else
 		return RTE_DIM(qede_xstats_strings) +
 		       RTE_DIM(qede_ah_xstats_strings) +
 		       (RTE_DIM(qede_rxq_xstats_strings) *
-			RTE_MIN(QEDE_RSS_COUNT(qdev),
-				RTE_ETHDEV_QUEUE_STAT_CNTRS));
+			QEDE_RSS_COUNT(dev));
 }
 
 static int
@@ -1588,42 +1598,43 @@ qede_get_xstats_names(struct rte_eth_dev *dev,
 	struct qede_dev *qdev = dev->data->dev_private;
 	struct ecore_dev *edev = &qdev->edev;
 	const unsigned int stat_cnt = qede_get_xstats_count(qdev);
-	unsigned int i, qid, stat_idx = 0;
-	unsigned int rxq_stat_cntrs;
+	unsigned int i, qid, hw_fn, stat_idx = 0;
 
-	if (xstats_names != NULL) {
-		for (i = 0; i < RTE_DIM(qede_xstats_strings); i++) {
+	if (xstats_names == NULL)
+		return stat_cnt;
+
+	for (i = 0; i < RTE_DIM(qede_xstats_strings); i++) {
+		strlcpy(xstats_names[stat_idx].name,
+			qede_xstats_strings[i].name,
+			sizeof(xstats_names[stat_idx].name));
+		stat_idx++;
+	}
+
+	if (ECORE_IS_BB(edev)) {
+		for (i = 0; i < RTE_DIM(qede_bb_xstats_strings); i++) {
 			strlcpy(xstats_names[stat_idx].name,
-				qede_xstats_strings[i].name,
+				qede_bb_xstats_strings[i].name,
 				sizeof(xstats_names[stat_idx].name));
 			stat_idx++;
 		}
-
-		if (ECORE_IS_BB(edev)) {
-			for (i = 0; i < RTE_DIM(qede_bb_xstats_strings); i++) {
-				strlcpy(xstats_names[stat_idx].name,
-					qede_bb_xstats_strings[i].name,
-					sizeof(xstats_names[stat_idx].name));
-				stat_idx++;
-			}
-		} else {
-			for (i = 0; i < RTE_DIM(qede_ah_xstats_strings); i++) {
-				strlcpy(xstats_names[stat_idx].name,
-					qede_ah_xstats_strings[i].name,
-					sizeof(xstats_names[stat_idx].name));
-				stat_idx++;
-			}
+	} else {
+		for (i = 0; i < RTE_DIM(qede_ah_xstats_strings); i++) {
+			strlcpy(xstats_names[stat_idx].name,
+				qede_ah_xstats_strings[i].name,
+				sizeof(xstats_names[stat_idx].name));
+			stat_idx++;
 		}
+	}
 
-		rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
-					 RTE_ETHDEV_QUEUE_STAT_CNTRS);
-		for (qid = 0; qid < rxq_stat_cntrs; qid++) {
+	for (qid = 0; qid < QEDE_RSS_COUNT(dev); qid++) {
+		for_each_hwfn(edev, hw_fn) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
 				snprintf(xstats_names[stat_idx].name,
-					sizeof(xstats_names[stat_idx].name),
-					"%.4s%d%s",
-					qede_rxq_xstats_strings[i].name, qid,
-					qede_rxq_xstats_strings[i].name + 4);
+					 RTE_ETH_XSTATS_NAME_SIZE,
+					 "%.4s%d.%d%s",
+					 qede_rxq_xstats_strings[i].name,
+					 hw_fn, qid,
+					 qede_rxq_xstats_strings[i].name + 4);
 				stat_idx++;
 			}
 		}
@@ -1640,8 +1651,7 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 	struct ecore_dev *edev = &qdev->edev;
 	struct ecore_eth_stats stats;
 	const unsigned int num = qede_get_xstats_count(qdev);
-	unsigned int i, qid, stat_idx = 0;
-	unsigned int rxq_stat_cntrs;
+	unsigned int i, qid, hw_fn, fpidx, stat_idx = 0;
 
 	if (n < num)
 		return num;
@@ -1673,13 +1683,12 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 		}
 	}
 
-	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
-				 RTE_ETHDEV_QUEUE_STAT_CNTRS);
-	for (qid = 0; qid < rxq_stat_cntrs; qid++) {
-		for_each_rss(qid) {
+	for (qid = 0; qid < dev->data->nb_rx_queues; qid++) {
+		for_each_hwfn(edev, hw_fn) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
+				fpidx = qid * edev->num_hwfns + hw_fn;
 				xstats[stat_idx].value = *(uint64_t *)(
-					((char *)(qdev->fp_array[qid].rxq)) +
+					((char *)(qdev->fp_array[fpidx].rxq)) +
 					 qede_rxq_xstats_strings[i].offset);
 				xstats[stat_idx].id = stat_idx;
 				stat_idx++;
@@ -1938,7 +1947,8 @@ qede_dev_supported_ptypes_get(struct rte_eth_dev *eth_dev)
 		RTE_PTYPE_UNKNOWN
 	};
 
-	if (eth_dev->rx_pkt_burst == qede_recv_pkts)
+	if (eth_dev->rx_pkt_burst == qede_recv_pkts ||
+	    eth_dev->rx_pkt_burst == qede_recv_pkts_cmt)
 		return ptypes;
 
 	return NULL;
@@ -1968,8 +1978,7 @@ int qede_rss_hash_update(struct rte_eth_dev *eth_dev,
 	uint32_t *key = (uint32_t *)rss_conf->rss_key;
 	uint64_t hf = rss_conf->rss_hf;
 	uint8_t len = rss_conf->rss_key_len;
-	uint8_t idx;
-	uint8_t i;
+	uint8_t idx, i, j, fpidx;
 	int rc;
 
 	memset(&vport_update_params, 0, sizeof(vport_update_params));
@@ -2003,14 +2012,18 @@ int qede_rss_hash_update(struct rte_eth_dev *eth_dev,
 	/* tbl_size has to be set with capabilities */
 	rss_params.rss_table_size_log = 7;
 	vport_update_params.vport_id = 0;
-	/* pass the L2 handles instead of qids */
-	for (i = 0 ; i < ECORE_RSS_IND_TABLE_SIZE ; i++) {
-		idx = i % QEDE_RSS_COUNT(qdev);
-		rss_params.rss_ind_table[i] = qdev->fp_array[idx].rxq->handle;
-	}
-	vport_update_params.rss_params = &rss_params;
 
 	for_each_hwfn(edev, i) {
+		/* pass the L2 handles instead of qids */
+		for (j = 0 ; j < ECORE_RSS_IND_TABLE_SIZE ; j++) {
+			idx = j % QEDE_RSS_COUNT(eth_dev);
+			fpidx = idx * edev->num_hwfns + i;
+			rss_params.rss_ind_table[j] =
+				qdev->fp_array[fpidx].rxq->handle;
+		}
+
+		vport_update_params.rss_params = &rss_params;
+
 		p_hwfn = &edev->hwfns[i];
 		vport_update_params.opaque_fid = p_hwfn->hw_info.opaque_fid;
 		rc = ecore_sp_vport_update(p_hwfn, &vport_update_params,
@@ -2062,61 +2075,6 @@ static int qede_rss_hash_conf_get(struct rte_eth_dev *eth_dev,
 	return 0;
 }
 
-static bool qede_update_rss_parm_cmt(struct ecore_dev *edev,
-				    struct ecore_rss_params *rss)
-{
-	int i, fn;
-	bool rss_mode = 1; /* enable */
-	struct ecore_queue_cid *cid;
-	struct ecore_rss_params *t_rss;
-
-	/* In regular scenario, we'd simply need to take input handlers.
-	 * But in CMT, we'd have to split the handlers according to the
-	 * engine they were configured on. We'd then have to understand
-	 * whether RSS is really required, since 2-queues on CMT doesn't
-	 * require RSS.
-	 */
-
-	/* CMT should be round-robin */
-	for (i = 0; i < ECORE_RSS_IND_TABLE_SIZE; i++) {
-		cid = rss->rss_ind_table[i];
-
-		if (cid->p_owner == ECORE_LEADING_HWFN(edev))
-			t_rss = &rss[0];
-		else
-			t_rss = &rss[1];
-
-		t_rss->rss_ind_table[i / edev->num_hwfns] = cid;
-	}
-
-	t_rss = &rss[1];
-	t_rss->update_rss_ind_table = 1;
-	t_rss->rss_table_size_log = 7;
-	t_rss->update_rss_config = 1;
-
-	/* Make sure RSS is actually required */
-	for_each_hwfn(edev, fn) {
-		for (i = 1; i < ECORE_RSS_IND_TABLE_SIZE / edev->num_hwfns;
-		     i++) {
-			if (rss[fn].rss_ind_table[i] !=
-			    rss[fn].rss_ind_table[0])
-				break;
-		}
-
-		if (i == ECORE_RSS_IND_TABLE_SIZE / edev->num_hwfns) {
-			DP_INFO(edev,
-				"CMT - 1 queue per-hwfn; Disabling RSS\n");
-			rss_mode = 0;
-			goto out;
-		}
-	}
-
-out:
-	t_rss->rss_enable = rss_mode;
-
-	return rss_mode;
-}
-
 int qede_rss_reta_update(struct rte_eth_dev *eth_dev,
 			 struct rte_eth_rss_reta_entry64 *reta_conf,
 			 uint16_t reta_size)
@@ -2125,8 +2083,8 @@ int qede_rss_reta_update(struct rte_eth_dev *eth_dev,
 	struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
 	struct ecore_sp_vport_update_params vport_update_params;
 	struct ecore_rss_params *params;
+	uint16_t i, j, idx, fid, shift;
 	struct ecore_hwfn *p_hwfn;
-	uint16_t i, idx, shift;
 	uint8_t entry;
 	int rc = 0;
 
@@ -2137,40 +2095,36 @@ int qede_rss_reta_update(struct rte_eth_dev *eth_dev,
 	}
 
 	memset(&vport_update_params, 0, sizeof(vport_update_params));
-	params = rte_zmalloc("qede_rss", sizeof(*params) * edev->num_hwfns,
-			     RTE_CACHE_LINE_SIZE);
+	params = rte_zmalloc("qede_rss", sizeof(*params), RTE_CACHE_LINE_SIZE);
 	if (params == NULL) {
 		DP_ERR(edev, "failed to allocate memory\n");
 		return -ENOMEM;
 	}
 
-	for (i = 0; i < reta_size; i++) {
-		idx = i / RTE_RETA_GROUP_SIZE;
-		shift = i % RTE_RETA_GROUP_SIZE;
-		if (reta_conf[idx].mask & (1ULL << shift)) {
-			entry = reta_conf[idx].reta[shift];
-			/* Pass rxq handles to ecore */
-			params->rss_ind_table[i] =
-					qdev->fp_array[entry].rxq->handle;
-			/* Update the local copy for RETA query command */
-			qdev->rss_ind_table[i] = entry;
-		}
-	}
-
 	params->update_rss_ind_table = 1;
 	params->rss_table_size_log = 7;
 	params->update_rss_config = 1;
 
-	/* Fix up RETA for CMT mode device */
-	if (ECORE_IS_CMT(edev))
-		qdev->rss_enable = qede_update_rss_parm_cmt(edev,
-							    params);
 	vport_update_params.vport_id = 0;
 	/* Use the current value of rss_enable */
 	params->rss_enable = qdev->rss_enable;
 	vport_update_params.rss_params = params;
 
 	for_each_hwfn(edev, i) {
+		for (j = 0; j < reta_size; j++) {
+			idx = j / RTE_RETA_GROUP_SIZE;
+			shift = j % RTE_RETA_GROUP_SIZE;
+			if (reta_conf[idx].mask & (1ULL << shift)) {
+				entry = reta_conf[idx].reta[shift];
+				fid = entry * edev->num_hwfns + i;
+				/* Pass rxq handles to ecore */
+				params->rss_ind_table[j] =
+						qdev->fp_array[fid].rxq->handle;
+				/* Update the local copy for RETA query cmd */
+				qdev->rss_ind_table[j] = entry;
+			}
+		}
+
 		p_hwfn = &edev->hwfns[i];
 		vport_update_params.opaque_fid = p_hwfn->hw_info.opaque_fid;
 		rc = ecore_sp_vport_update(p_hwfn, &vport_update_params,
@@ -2257,7 +2211,7 @@ static int qede_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 	qdev->mtu = mtu;
 
 	/* Fix up RX buf size for all queues of the port */
-	for_each_rss(i) {
+	for (i = 0; i < qdev->num_rx_queues; i++) {
 		fp = &qdev->fp_array[i];
 		if (fp->rxq != NULL) {
 			bufsz = (uint16_t)rte_pktmbuf_data_room_size(
@@ -2286,9 +2240,13 @@ static int qede_set_mtu(struct rte_eth_dev *dev, uint16_t mtu)
 	/* update max frame size */
 	dev->data->dev_conf.rxmode.max_rx_pkt_len = max_rx_pkt_len;
 	/* Reassign back */
-	dev->rx_pkt_burst = qede_recv_pkts;
-	dev->tx_pkt_burst = qede_xmit_pkts;
-
+	if (ECORE_IS_CMT(edev)) {
+		dev->rx_pkt_burst = qede_recv_pkts_cmt;
+		dev->tx_pkt_burst = qede_xmit_pkts_cmt;
+	} else {
+		dev->rx_pkt_burst = qede_recv_pkts;
+		dev->tx_pkt_burst = qede_xmit_pkts;
+	}
 	return 0;
 }
 
@@ -2429,10 +2387,6 @@ static int qede_common_dev_init(struct rte_eth_dev *eth_dev, bool is_vf)
 		 pci_addr.bus, pci_addr.devid, pci_addr.function,
 		 eth_dev->data->port_id);
 
-	eth_dev->rx_pkt_burst = qede_recv_pkts;
-	eth_dev->tx_pkt_burst = qede_xmit_pkts;
-	eth_dev->tx_pkt_prepare = qede_xmit_prep_pkts;
-
 	if (rte_eal_process_type() != RTE_PROC_PRIMARY) {
 		DP_ERR(edev, "Skipping device init from secondary process\n");
 		return 0;
@@ -2490,6 +2444,16 @@ static int qede_common_dev_init(struct rte_eth_dev *eth_dev, bool is_vf)
 	strncpy((char *)params.name, QEDE_PMD_VER_PREFIX,
 		QEDE_PMD_DRV_VER_STR_SIZE);
 
+	if (ECORE_IS_CMT(edev)) {
+		eth_dev->rx_pkt_burst = qede_recv_pkts_cmt;
+		eth_dev->tx_pkt_burst = qede_xmit_pkts_cmt;
+	} else {
+		eth_dev->rx_pkt_burst = qede_recv_pkts;
+		eth_dev->tx_pkt_burst = qede_xmit_pkts;
+	}
+
+	eth_dev->tx_pkt_prepare = qede_xmit_prep_pkts;
+
 	/* For CMT mode device do periodic polling for slowpath events.
 	 * This is required since uio device uses only one MSI-x
 	 * interrupt vector but we need one for each engine.
diff --git a/drivers/net/qede/qede_ethdev.h b/drivers/net/qede/qede_ethdev.h
index d0e7c70be..5549d0bf3 100644
--- a/drivers/net/qede/qede_ethdev.h
+++ b/drivers/net/qede/qede_ethdev.h
@@ -66,8 +66,8 @@
 					(edev)->dev_info.num_tc)
 
 #define QEDE_QUEUE_CNT(qdev) ((qdev)->num_queues)
-#define QEDE_RSS_COUNT(qdev) ((qdev)->num_rx_queues)
-#define QEDE_TSS_COUNT(qdev) ((qdev)->num_tx_queues)
+#define QEDE_RSS_COUNT(dev) ((dev)->data->nb_rx_queues)
+#define QEDE_TSS_COUNT(dev) ((dev)->data->nb_tx_queues)
 
 #define QEDE_DUPLEX_FULL	1
 #define QEDE_DUPLEX_HALF	2
@@ -215,6 +215,7 @@ struct qede_dev {
 	struct qed_dev_eth_info dev_info;
 	struct ecore_sb_info *sb_array;
 	struct qede_fastpath *fp_array;
+	struct qede_fastpath_cmt *fp_array_cmt;
 	uint16_t mtu;
 	bool enable_tx_switching;
 	bool rss_enable;
diff --git a/drivers/net/qede/qede_filter.c b/drivers/net/qede/qede_filter.c
index b3f62e0dd..81509f04b 100644
--- a/drivers/net/qede/qede_filter.c
+++ b/drivers/net/qede/qede_filter.c
@@ -431,7 +431,7 @@ qede_fdir_filter_add(struct rte_eth_dev *eth_dev,
 		return -EINVAL;
 	}
 
-	if (fdir->action.rx_queue >= QEDE_RSS_COUNT(qdev)) {
+	if (fdir->action.rx_queue >= QEDE_RSS_COUNT(eth_dev)) {
 		DP_ERR(edev, "invalid queue number %u\n",
 		       fdir->action.rx_queue);
 		return -EINVAL;
@@ -1345,7 +1345,6 @@ qede_flow_parse_actions(struct rte_eth_dev *dev,
 			struct rte_flow_error *error,
 			struct rte_flow *flow)
 {
-	struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
 	const struct rte_flow_action_queue *queue;
 
 	if (actions == NULL) {
@@ -1360,7 +1359,7 @@ qede_flow_parse_actions(struct rte_eth_dev *dev,
 		case RTE_FLOW_ACTION_TYPE_QUEUE:
 			queue = actions->conf;
 
-			if (queue->index >= QEDE_RSS_COUNT(qdev)) {
+			if (queue->index >= QEDE_RSS_COUNT(dev)) {
 				rte_flow_error_set(error, EINVAL,
 						   RTE_FLOW_ERROR_TYPE_ACTION,
 						   actions,
diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index c38cbb905..a882ba238 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -124,36 +124,20 @@ qede_calc_rx_buf_size(struct rte_eth_dev *dev, uint16_t mbufsz,
 	return QEDE_FLOOR_TO_CACHE_LINE_SIZE(rx_buf_size);
 }
 
-int
-qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
-		    uint16_t nb_desc, unsigned int socket_id,
-		    __rte_unused const struct rte_eth_rxconf *rx_conf,
-		    struct rte_mempool *mp)
+static struct qede_rx_queue *
+qede_alloc_rx_queue_mem(struct rte_eth_dev *dev,
+			uint16_t queue_idx,
+			uint16_t nb_desc,
+			unsigned int socket_id,
+			struct rte_mempool *mp,
+			uint16_t bufsz)
 {
 	struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
 	struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
-	struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
 	struct qede_rx_queue *rxq;
-	uint16_t max_rx_pkt_len;
-	uint16_t bufsz;
 	size_t size;
 	int rc;
 
-	PMD_INIT_FUNC_TRACE(edev);
-
-	/* Note: Ring size/align is controlled by struct rte_eth_desc_lim */
-	if (!rte_is_power_of_2(nb_desc)) {
-		DP_ERR(edev, "Ring size %u is not power of 2\n",
-			  nb_desc);
-		return -EINVAL;
-	}
-
-	/* Free memory prior to re-allocation if needed... */
-	if (dev->data->rx_queues[queue_idx] != NULL) {
-		qede_rx_queue_release(dev->data->rx_queues[queue_idx]);
-		dev->data->rx_queues[queue_idx] = NULL;
-	}
-
 	/* First allocate the rx queue data structure */
 	rxq = rte_zmalloc_socket("qede_rx_queue", sizeof(struct qede_rx_queue),
 				 RTE_CACHE_LINE_SIZE, socket_id);
@@ -161,7 +145,7 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	if (!rxq) {
 		DP_ERR(edev, "Unable to allocate memory for rxq on socket %u",
 			  socket_id);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	rxq->qdev = qdev;
@@ -170,27 +154,8 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	rxq->queue_id = queue_idx;
 	rxq->port_id = dev->data->port_id;
 
-	max_rx_pkt_len = (uint16_t)rxmode->max_rx_pkt_len;
-
-	/* Fix up RX buffer size */
-	bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
-	/* cache align the mbuf size to simplfy rx_buf_size calculation */
-	bufsz = QEDE_FLOOR_TO_CACHE_LINE_SIZE(bufsz);
-	if ((rxmode->offloads & DEV_RX_OFFLOAD_SCATTER)	||
-	    (max_rx_pkt_len + QEDE_ETH_OVERHEAD) > bufsz) {
-		if (!dev->data->scattered_rx) {
-			DP_INFO(edev, "Forcing scatter-gather mode\n");
-			dev->data->scattered_rx = 1;
-		}
-	}
-
-	rc = qede_calc_rx_buf_size(dev, bufsz, max_rx_pkt_len);
-	if (rc < 0) {
-		rte_free(rxq);
-		return rc;
-	}
 
-	rxq->rx_buf_size = rc;
+	rxq->rx_buf_size = bufsz;
 
 	DP_INFO(edev, "mtu %u mbufsz %u bd_max_bytes %u scatter_mode %d\n",
 		qdev->mtu, bufsz, rxq->rx_buf_size, dev->data->scattered_rx);
@@ -203,7 +168,7 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 		DP_ERR(edev, "Memory allocation fails for sw_rx_ring on"
 		       " socket %u\n", socket_id);
 		rte_free(rxq);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	/* Allocate FW Rx ring  */
@@ -221,7 +186,7 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 		       " on socket %u\n", socket_id);
 		rte_free(rxq->sw_rx_ring);
 		rte_free(rxq);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	/* Allocate FW completion ring */
@@ -240,14 +205,88 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 		qdev->ops->common->chain_free(edev, &rxq->rx_bd_ring);
 		rte_free(rxq->sw_rx_ring);
 		rte_free(rxq);
-		return -ENOMEM;
+		return NULL;
+	}
+
+	return rxq;
+}
+
+int
+qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qid,
+		    uint16_t nb_desc, unsigned int socket_id,
+		    __rte_unused const struct rte_eth_rxconf *rx_conf,
+		    struct rte_mempool *mp)
+{
+	struct qede_dev *qdev = QEDE_INIT_QDEV(dev);
+	struct ecore_dev *edev = QEDE_INIT_EDEV(qdev);
+	struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
+	struct qede_rx_queue *rxq;
+	uint16_t max_rx_pkt_len;
+	uint16_t bufsz;
+	int rc;
+
+	PMD_INIT_FUNC_TRACE(edev);
+
+	/* Note: Ring size/align is controlled by struct rte_eth_desc_lim */
+	if (!rte_is_power_of_2(nb_desc)) {
+		DP_ERR(edev, "Ring size %u is not power of 2\n",
+			  nb_desc);
+		return -EINVAL;
+	}
+
+	/* Free memory prior to re-allocation if needed... */
+	if (dev->data->rx_queues[qid] != NULL) {
+		qede_rx_queue_release(dev->data->rx_queues[qid]);
+		dev->data->rx_queues[qid] = NULL;
+	}
+
+	max_rx_pkt_len = (uint16_t)rxmode->max_rx_pkt_len;
+
+	/* Fix up RX buffer size */
+	bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
+	/* cache align the mbuf size to simplfy rx_buf_size calculation */
+	bufsz = QEDE_FLOOR_TO_CACHE_LINE_SIZE(bufsz);
+	if ((rxmode->offloads & DEV_RX_OFFLOAD_SCATTER)	||
+	    (max_rx_pkt_len + QEDE_ETH_OVERHEAD) > bufsz) {
+		if (!dev->data->scattered_rx) {
+			DP_INFO(edev, "Forcing scatter-gather mode\n");
+			dev->data->scattered_rx = 1;
+		}
 	}
 
-	dev->data->rx_queues[queue_idx] = rxq;
-	qdev->fp_array[queue_idx].rxq = rxq;
+	rc = qede_calc_rx_buf_size(dev, bufsz, max_rx_pkt_len);
+	if (rc < 0)
+		return rc;
+
+	bufsz = rc;
+
+	if (ECORE_IS_CMT(edev)) {
+		rxq = qede_alloc_rx_queue_mem(dev, qid * 2, nb_desc,
+					      socket_id, mp, bufsz);
+		if (!rxq)
+			return -ENOMEM;
+
+		qdev->fp_array[qid * 2].rxq = rxq;
+		rxq = qede_alloc_rx_queue_mem(dev, qid * 2 + 1, nb_desc,
+					      socket_id, mp, bufsz);
+		if (!rxq)
+			return -ENOMEM;
+
+		qdev->fp_array[qid * 2 + 1].rxq = rxq;
+		/* provide per engine fp struct as rx queue */
+		dev->data->rx_queues[qid] = &qdev->fp_array_cmt[qid];
+	} else {
+		rxq = qede_alloc_rx_queue_mem(dev, qid, nb_desc,
+					      socket_id, mp, bufsz);
+		if (!rxq)
+			return -ENOMEM;
+
+		dev->data->rx_queues[qid] = rxq;
+		qdev->fp_array[qid].rxq = rxq;
+	}
 
 	DP_INFO(edev, "rxq %d num_desc %u rx_buf_size=%u socket %u\n",
-		  queue_idx, nb_desc, rxq->rx_buf_size, socket_id);
+		  qid, nb_desc, rxq->rx_buf_size, socket_id);
 
 	return 0;
 }
@@ -278,21 +317,43 @@ static void qede_rx_queue_release_mbufs(struct qede_rx_queue *rxq)
 	}
 }
 
+static void _qede_rx_queue_release(struct qede_dev *qdev,
+				   struct ecore_dev *edev,
+				   struct qede_rx_queue *rxq)
+{
+	qede_rx_queue_release_mbufs(rxq);
+	qdev->ops->common->chain_free(edev, &rxq->rx_bd_ring);
+	qdev->ops->common->chain_free(edev, &rxq->rx_comp_ring);
+	rte_free(rxq->sw_rx_ring);
+	rte_free(rxq);
+}
+
 void qede_rx_queue_release(void *rx_queue)
 {
 	struct qede_rx_queue *rxq = rx_queue;
+	struct qede_fastpath_cmt *fp_cmt;
 	struct qede_dev *qdev;
 	struct ecore_dev *edev;
 
 	if (rxq) {
+		/* This function is shared for CMT and regular mode,
+		 * and we provide struct qede_fastpath_cmt in CMT mode and
+		 * struct qede_rx_queue in regular mode to DPDK app.
+		 * Only way to detect if we are in CMT mode is by reading
+		 * qdev element, hence both structures have qdev as first
+		 * element so dereferencing becomes easy using rxq structure in
+		 * even if we are in CMT mode.
+		 */
 		qdev = rxq->qdev;
 		edev = QEDE_INIT_EDEV(qdev);
 		PMD_INIT_FUNC_TRACE(edev);
-		qede_rx_queue_release_mbufs(rxq);
-		qdev->ops->common->chain_free(edev, &rxq->rx_bd_ring);
-		qdev->ops->common->chain_free(edev, &rxq->rx_comp_ring);
-		rte_free(rxq->sw_rx_ring);
-		rte_free(rxq);
+		if (ECORE_IS_CMT(edev)) {
+			fp_cmt = rx_queue;
+			_qede_rx_queue_release(qdev, edev, fp_cmt->fp0->rxq);
+			_qede_rx_queue_release(qdev, edev, fp_cmt->fp1->rxq);
+		} else {
+			_qede_rx_queue_release(qdev, edev, rxq);
+		}
 	}
 }
 
@@ -306,8 +367,8 @@ static int qede_rx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
 	int hwfn_index;
 	int rc;
 
-	if (rx_queue_id < eth_dev->data->nb_rx_queues) {
-		rxq = eth_dev->data->rx_queues[rx_queue_id];
+	if (rx_queue_id < qdev->num_rx_queues) {
+		rxq = qdev->fp_array[rx_queue_id].rxq;
 		hwfn_index = rx_queue_id % edev->num_hwfns;
 		p_hwfn = &edev->hwfns[hwfn_index];
 		rc = ecore_eth_rx_queue_stop(p_hwfn, rxq->handle,
@@ -329,32 +390,18 @@ static int qede_rx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
 	return rc;
 }
 
-int
-qede_tx_queue_setup(struct rte_eth_dev *dev,
-		    uint16_t queue_idx,
-		    uint16_t nb_desc,
-		    unsigned int socket_id,
-		    const struct rte_eth_txconf *tx_conf)
+static struct qede_tx_queue *
+qede_alloc_tx_queue_mem(struct rte_eth_dev *dev,
+			uint16_t queue_idx,
+			uint16_t nb_desc,
+			unsigned int socket_id,
+			const struct rte_eth_txconf *tx_conf)
 {
 	struct qede_dev *qdev = dev->data->dev_private;
 	struct ecore_dev *edev = &qdev->edev;
 	struct qede_tx_queue *txq;
 	int rc;
 
-	PMD_INIT_FUNC_TRACE(edev);
-
-	if (!rte_is_power_of_2(nb_desc)) {
-		DP_ERR(edev, "Ring size %u is not power of 2\n",
-		       nb_desc);
-		return -EINVAL;
-	}
-
-	/* Free memory prior to re-allocation if needed... */
-	if (dev->data->tx_queues[queue_idx] != NULL) {
-		qede_tx_queue_release(dev->data->tx_queues[queue_idx]);
-		dev->data->tx_queues[queue_idx] = NULL;
-	}
-
 	txq = rte_zmalloc_socket("qede_tx_queue", sizeof(struct qede_tx_queue),
 				 RTE_CACHE_LINE_SIZE, socket_id);
 
@@ -362,7 +409,7 @@ qede_tx_queue_setup(struct rte_eth_dev *dev,
 		DP_ERR(edev,
 		       "Unable to allocate memory for txq on socket %u",
 		       socket_id);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	txq->nb_tx_desc = nb_desc;
@@ -382,7 +429,7 @@ qede_tx_queue_setup(struct rte_eth_dev *dev,
 		       "Unable to allocate memory for txbd ring on socket %u",
 		       socket_id);
 		qede_tx_queue_release(txq);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	/* Allocate software ring */
@@ -397,7 +444,7 @@ qede_tx_queue_setup(struct rte_eth_dev *dev,
 		       socket_id);
 		qdev->ops->common->chain_free(edev, &txq->tx_pbl);
 		qede_tx_queue_release(txq);
-		return -ENOMEM;
+		return NULL;
 	}
 
 	txq->queue_id = queue_idx;
@@ -408,12 +455,61 @@ qede_tx_queue_setup(struct rte_eth_dev *dev,
 	    tx_conf->tx_free_thresh ? tx_conf->tx_free_thresh :
 	    (txq->nb_tx_desc - QEDE_DEFAULT_TX_FREE_THRESH);
 
-	dev->data->tx_queues[queue_idx] = txq;
-	qdev->fp_array[queue_idx].txq = txq;
-
 	DP_INFO(edev,
 		  "txq %u num_desc %u tx_free_thresh %u socket %u\n",
 		  queue_idx, nb_desc, txq->tx_free_thresh, socket_id);
+	return txq;
+}
+
+int
+qede_tx_queue_setup(struct rte_eth_dev *dev,
+		    uint16_t queue_idx,
+		    uint16_t nb_desc,
+		    unsigned int socket_id,
+		    const struct rte_eth_txconf *tx_conf)
+{
+	struct qede_dev *qdev = dev->data->dev_private;
+	struct ecore_dev *edev = &qdev->edev;
+	struct qede_tx_queue *txq;
+
+	PMD_INIT_FUNC_TRACE(edev);
+
+	if (!rte_is_power_of_2(nb_desc)) {
+		DP_ERR(edev, "Ring size %u is not power of 2\n",
+		       nb_desc);
+		return -EINVAL;
+	}
+
+	/* Free memory prior to re-allocation if needed... */
+	if (dev->data->tx_queues[queue_idx] != NULL) {
+		qede_tx_queue_release(dev->data->tx_queues[queue_idx]);
+		dev->data->tx_queues[queue_idx] = NULL;
+	}
+
+	if (ECORE_IS_CMT(edev)) {
+		txq = qede_alloc_tx_queue_mem(dev, queue_idx * 2, nb_desc,
+					      socket_id, tx_conf);
+		if (!txq)
+			return -ENOMEM;
+
+		qdev->fp_array[queue_idx * 2].txq = txq;
+		txq = qede_alloc_tx_queue_mem(dev, (queue_idx * 2) + 1, nb_desc,
+					      socket_id, tx_conf);
+		if (!txq)
+			return -ENOMEM;
+
+		qdev->fp_array[(queue_idx * 2) + 1].txq = txq;
+		dev->data->tx_queues[queue_idx] =
+					&qdev->fp_array_cmt[queue_idx];
+	} else {
+		txq = qede_alloc_tx_queue_mem(dev, queue_idx, nb_desc,
+					      socket_id, tx_conf);
+		if (!txq)
+			return -ENOMEM;
+
+		dev->data->tx_queues[queue_idx] = txq;
+		qdev->fp_array[queue_idx].txq = txq;
+	}
 
 	return 0;
 }
@@ -443,9 +539,20 @@ static void qede_tx_queue_release_mbufs(struct qede_tx_queue *txq)
 	}
 }
 
+static void _qede_tx_queue_release(struct qede_dev *qdev,
+				   struct ecore_dev *edev,
+				   struct qede_tx_queue *txq)
+{
+	qede_tx_queue_release_mbufs(txq);
+	qdev->ops->common->chain_free(edev, &txq->tx_pbl);
+	rte_free(txq->sw_tx_ring);
+	rte_free(txq);
+}
+
 void qede_tx_queue_release(void *tx_queue)
 {
 	struct qede_tx_queue *txq = tx_queue;
+	struct qede_fastpath_cmt *fp_cmt;
 	struct qede_dev *qdev;
 	struct ecore_dev *edev;
 
@@ -453,10 +560,13 @@ void qede_tx_queue_release(void *tx_queue)
 		qdev = txq->qdev;
 		edev = QEDE_INIT_EDEV(qdev);
 		PMD_INIT_FUNC_TRACE(edev);
-		qede_tx_queue_release_mbufs(txq);
-		qdev->ops->common->chain_free(edev, &txq->tx_pbl);
-		rte_free(txq->sw_tx_ring);
-		rte_free(txq);
+		if (ECORE_IS_CMT(edev)) {
+			fp_cmt = tx_queue;
+			_qede_tx_queue_release(qdev, edev, fp_cmt->fp0->txq);
+			_qede_tx_queue_release(qdev, edev, fp_cmt->fp1->txq);
+		} else {
+			_qede_tx_queue_release(qdev, edev, txq);
+		}
 	}
 }
 
@@ -494,6 +604,7 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 	struct qede_fastpath *fp;
 	uint32_t num_sbs;
 	uint16_t sb_idx;
+	uint16_t i;
 
 	if (IS_VF(edev))
 		ecore_vf_get_num_sbs(ECORE_LEADING_HWFN(edev), &num_sbs);
@@ -517,6 +628,27 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 	memset((void *)qdev->fp_array, 0, QEDE_RXTX_MAX(qdev) *
 			sizeof(*qdev->fp_array));
 
+	if (ECORE_IS_CMT(edev)) {
+		qdev->fp_array_cmt = rte_calloc("fp_cmt",
+						QEDE_RXTX_MAX(qdev) / 2,
+						sizeof(*qdev->fp_array_cmt),
+						RTE_CACHE_LINE_SIZE);
+
+		if (!qdev->fp_array_cmt) {
+			DP_ERR(edev, "fp array for CMT allocation failed\n");
+			return -ENOMEM;
+		}
+
+		memset((void *)qdev->fp_array_cmt, 0,
+		       (QEDE_RXTX_MAX(qdev) / 2) * sizeof(*qdev->fp_array_cmt));
+
+		for (i = 0; i < QEDE_RXTX_MAX(qdev) / 2; i++) {
+			qdev->fp_array_cmt[i].qdev = qdev;
+			qdev->fp_array_cmt[i].fp0 = &qdev->fp_array[i * 2];
+			qdev->fp_array_cmt[i].fp1 = &qdev->fp_array[i * 2 + 1];
+		}
+	}
+
 	for (sb_idx = 0; sb_idx < QEDE_RXTX_MAX(qdev); sb_idx++) {
 		fp = &qdev->fp_array[sb_idx];
 		if (!fp)
@@ -581,6 +713,10 @@ void qede_dealloc_fp_resc(struct rte_eth_dev *eth_dev)
 	if (qdev->fp_array)
 		rte_free(qdev->fp_array);
 	qdev->fp_array = NULL;
+
+	if (qdev->fp_array_cmt)
+		rte_free(qdev->fp_array_cmt);
+	qdev->fp_array_cmt = NULL;
 }
 
 static inline void
@@ -632,9 +768,9 @@ qede_rx_queue_start(struct rte_eth_dev *eth_dev, uint16_t rx_queue_id)
 	int hwfn_index;
 	int rc;
 
-	if (rx_queue_id < eth_dev->data->nb_rx_queues) {
+	if (rx_queue_id < qdev->num_rx_queues) {
 		fp = &qdev->fp_array[rx_queue_id];
-		rxq = eth_dev->data->rx_queues[rx_queue_id];
+		rxq = fp->rxq;
 		/* Allocate buffers for the Rx ring */
 		for (j = 0; j < rxq->nb_rx_desc; j++) {
 			rc = qede_alloc_rx_buffer(rxq);
@@ -703,9 +839,9 @@ qede_tx_queue_start(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
 	int hwfn_index;
 	int rc;
 
-	if (tx_queue_id < eth_dev->data->nb_tx_queues) {
-		txq = eth_dev->data->tx_queues[tx_queue_id];
+	if (tx_queue_id < qdev->num_tx_queues) {
 		fp = &qdev->fp_array[tx_queue_id];
+		txq = fp->txq;
 		memset(&params, 0, sizeof(params));
 		params.queue_id = tx_queue_id / edev->num_hwfns;
 		params.vport_id = 0;
@@ -846,8 +982,8 @@ static int qede_tx_queue_stop(struct rte_eth_dev *eth_dev, uint16_t tx_queue_id)
 	int hwfn_index;
 	int rc;
 
-	if (tx_queue_id < eth_dev->data->nb_tx_queues) {
-		txq = eth_dev->data->tx_queues[tx_queue_id];
+	if (tx_queue_id < qdev->num_tx_queues) {
+		txq = qdev->fp_array[tx_queue_id].txq;
 		/* Drain txq */
 		if (qede_drain_txq(qdev, txq, true))
 			return -1; /* For the lack of retcodes */
@@ -878,13 +1014,13 @@ int qede_start_queues(struct rte_eth_dev *eth_dev)
 	uint8_t id;
 	int rc = -1;
 
-	for_each_rss(id) {
+	for (id = 0; id < qdev->num_rx_queues; id++) {
 		rc = qede_rx_queue_start(eth_dev, id);
 		if (rc != ECORE_SUCCESS)
 			return -1;
 	}
 
-	for_each_tss(id) {
+	for (id = 0; id < qdev->num_tx_queues; id++) {
 		rc = qede_tx_queue_start(eth_dev, id);
 		if (rc != ECORE_SUCCESS)
 			return -1;
@@ -899,13 +1035,11 @@ void qede_stop_queues(struct rte_eth_dev *eth_dev)
 	uint8_t id;
 
 	/* Stopping RX/TX queues */
-	for_each_tss(id) {
+	for (id = 0; id < qdev->num_tx_queues; id++)
 		qede_tx_queue_stop(eth_dev, id);
-	}
 
-	for_each_rss(id) {
+	for (id = 0; id < qdev->num_rx_queues; id++)
 		qede_rx_queue_stop(eth_dev, id);
-	}
 }
 
 static inline bool qede_tunn_exist(uint16_t flag)
@@ -1687,6 +1821,24 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 	return rx_pkt;
 }
 
+uint16_t
+qede_recv_pkts_cmt(void *p_fp_cmt, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+	struct qede_fastpath_cmt *fp_cmt = p_fp_cmt;
+	uint16_t eng0_pkts, eng1_pkts;
+
+	eng0_pkts = nb_pkts / 2;
+
+	eng0_pkts = qede_recv_pkts(fp_cmt->fp0->rxq, rx_pkts, eng0_pkts);
+
+	eng1_pkts = nb_pkts - eng0_pkts;
+
+	eng1_pkts = qede_recv_pkts(fp_cmt->fp1->rxq, rx_pkts + eng0_pkts,
+				   eng1_pkts);
+
+	return eng0_pkts + eng1_pkts;
+}
+
 
 /* Populate scatter gather buffer descriptor fields */
 static inline uint16_t
@@ -2209,6 +2361,24 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	return nb_pkt_sent;
 }
 
+uint16_t
+qede_xmit_pkts_cmt(void *p_fp_cmt, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	struct qede_fastpath_cmt *fp_cmt = p_fp_cmt;
+	uint16_t eng0_pkts, eng1_pkts;
+
+	eng0_pkts = nb_pkts / 2;
+
+	eng0_pkts = qede_xmit_pkts(fp_cmt->fp0->txq, tx_pkts, eng0_pkts);
+
+	eng1_pkts = nb_pkts - eng0_pkts;
+
+	eng1_pkts = qede_xmit_pkts(fp_cmt->fp1->txq, tx_pkts + eng0_pkts,
+				   eng1_pkts);
+
+	return eng0_pkts + eng1_pkts;
+}
+
 uint16_t
 qede_rxtx_pkts_dummy(__rte_unused void *p_rxq,
 		     __rte_unused struct rte_mbuf **pkts,
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index 41a5f0f5c..75cc930fd 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -81,10 +81,8 @@
 				 ETH_RSS_VXLAN			|\
 				 ETH_RSS_GENEVE)
 
-#define for_each_rss(i)		for (i = 0; i < qdev->num_rx_queues; i++)
-#define for_each_tss(i)		for (i = 0; i < qdev->num_tx_queues; i++)
 #define QEDE_RXTX_MAX(qdev) \
-	(RTE_MAX(QEDE_RSS_COUNT(qdev), QEDE_TSS_COUNT(qdev)))
+	(RTE_MAX(qdev->num_rx_queues, qdev->num_tx_queues))
 
 /* Macros for non-tunnel packet types lkup table */
 #define QEDE_PKT_TYPE_UNKNOWN				0x0
@@ -179,6 +177,8 @@ struct qede_agg_info {
  * Structure associated with each RX queue.
  */
 struct qede_rx_queue {
+	/* Always keep qdev as first member */
+	struct qede_dev *qdev;
 	struct rte_mempool *mb_pool;
 	struct ecore_chain rx_bd_ring;
 	struct ecore_chain rx_comp_ring;
@@ -199,7 +199,6 @@ struct qede_rx_queue {
 	uint64_t rx_hw_errors;
 	uint64_t rx_alloc_errors;
 	struct qede_agg_info tpa_info[ETH_TPA_MAX_AGGS_NUM];
-	struct qede_dev *qdev;
 	void *handle;
 };
 
@@ -217,6 +216,8 @@ union db_prod {
 };
 
 struct qede_tx_queue {
+	/* Always keep qdev as first member */
+	struct qede_dev *qdev;
 	struct ecore_chain tx_pbl;
 	struct qede_tx_entry *sw_tx_ring;
 	uint16_t nb_tx_desc;
@@ -231,7 +232,6 @@ struct qede_tx_queue {
 	uint16_t port_id;
 	uint64_t xmit_pkts;
 	bool is_legacy;
-	struct qede_dev *qdev;
 	void *handle;
 };
 
@@ -241,6 +241,18 @@ struct qede_fastpath {
 	struct qede_tx_queue *txq;
 };
 
+/* This structure holds the inforation of fast path queues
+ * belonging to individual engines in CMT mode.
+ */
+struct qede_fastpath_cmt {
+	/* Always keep this a first element */
+	struct qede_dev *qdev;
+	/* fastpath info of engine 0 */
+	struct qede_fastpath *fp0;
+	/* fastpath info of engine 1 */
+	struct qede_fastpath *fp1;
+};
+
 /*
  * RX/TX function prototypes
  */
@@ -261,12 +273,16 @@ void qede_tx_queue_release(void *tx_queue);
 
 uint16_t qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts,
 			uint16_t nb_pkts);
+uint16_t qede_xmit_pkts_cmt(void *p_txq, struct rte_mbuf **tx_pkts,
+			    uint16_t nb_pkts);
 
 uint16_t qede_xmit_prep_pkts(void *p_txq, struct rte_mbuf **tx_pkts,
 			     uint16_t nb_pkts);
 
 uint16_t qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts,
 			uint16_t nb_pkts);
+uint16_t qede_recv_pkts_cmt(void *p_rxq, struct rte_mbuf **rx_pkts,
+			    uint16_t nb_pkts);
 
 uint16_t qede_rxtx_pkts_dummy(void *p_rxq,
 			      struct rte_mbuf **pkts,
-- 
2.17.1


^ permalink raw reply	[flat|nested] 5+ messages in thread
end of thread, other threads:[~2019-09-04 17:52 UTC | newest]

Thread overview: 5+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2019-09-04 11:30 [dpdk-dev] [PATCH 1/2] net/qede: fix ovs-dpdk failure when using odd number of queues on 100Gb mode Shahed Shaikh
2019-09-04 11:30 ` [dpdk-dev] [PATCH 2/2] net/qede: implement rte_flow drop action Shahed Shaikh
2019-09-04 13:28   ` Jerin Jacob Kollanukkaran
2019-09-04 13:30 ` [dpdk-dev] [PATCH 1/2] net/qede: fix ovs-dpdk failure when using odd number of queues on 100Gb mode Jerin Jacob Kollanukkaran
2019-09-04 17:52   ` Shahed Shaikh

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).