[dpdk-dev] [PATCH 1/5] net/qede: fix scatter-gather issue

[dpdk-dev] [PATCH 1/5] net/qede: fix scatter-gather issue

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

 - Make qede_process_sg_pkts() inline and add unlikely check
 - Fix mbuf segment chaining logic in qede_process_sg_pkts()
 - Change qede_encode_sg_bd() to return total segments required
 - Fix first TX buffer descriptor's length
 - Replace repeatitive code using a macro

Fixes: bec0228816c0 ("net/qede: support scatter gather")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_rxtx.c |  139 ++++++++++++++++++++----------------------
 drivers/net/qede/qede_rxtx.h |    4 --
 2 files changed, 65 insertions(+), 78 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 814d384..ecff5bc 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -810,39 +810,28 @@ static inline uint32_t qede_rx_cqe_to_tunn_pkt_type(uint16_t flags)
 		return RTE_PTYPE_UNKNOWN;
 }
 
-
-int qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
-			 int num_segs, uint16_t pkt_len)
+static inline int
+qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
+		     uint8_t num_segs, uint16_t pkt_len)
 {
 	struct qede_rx_queue *rxq = p_rxq;
 	struct qede_dev *qdev = rxq->qdev;
 	struct ecore_dev *edev = &qdev->edev;
-	uint16_t sw_rx_index, cur_size;
-
 	register struct rte_mbuf *seg1 = NULL;
 	register struct rte_mbuf *seg2 = NULL;
+	uint16_t sw_rx_index;
+	uint16_t cur_size;
 
 	seg1 = rx_mb;
 	while (num_segs) {
-		cur_size = pkt_len > rxq->rx_buf_size ?
-				rxq->rx_buf_size : pkt_len;
-		if (!cur_size) {
-			PMD_RX_LOG(DEBUG, rxq,
-				   "SG packet, len and num BD mismatch\n");
+		cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
+							pkt_len;
+		if (unlikely(!cur_size)) {
+			PMD_RX_LOG(ERR, rxq, "Length is 0 while %u BDs"
+				   " left for mapping jumbo\n", num_segs);
 			qede_recycle_rx_bd_ring(rxq, qdev, num_segs);
 			return -EINVAL;
 		}
-
-		if (qede_alloc_rx_buffer(rxq)) {
-			uint8_t index;
-
-			PMD_RX_LOG(DEBUG, rxq, "Buffer allocation failed\n");
-			index = rxq->port_id;
-			rte_eth_devices[index].data->rx_mbuf_alloc_failed++;
-			rxq->rx_alloc_errors++;
-			return -ENOMEM;
-		}
-
 		sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
 		seg2 = rxq->sw_rx_ring[sw_rx_index].mbuf;
 		qede_rx_bd_ring_consume(rxq);
@@ -852,16 +841,9 @@ int qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
 		seg1 = seg1->next;
 		num_segs--;
 		rxq->rx_segs++;
-		continue;
 	}
-	seg1 = NULL;
-
-	if (pkt_len)
-		PMD_RX_LOG(DEBUG, rxq,
-			   "Mapped all BDs of jumbo, but still have %d bytes\n",
-			   pkt_len);
 
-	return ECORE_SUCCESS;
+	return 0;
 }
 
 uint16_t
@@ -878,11 +860,16 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 	register struct rte_mbuf *rx_mb = NULL;
 	register struct rte_mbuf *seg1 = NULL;
 	enum eth_rx_cqe_type cqe_type;
-	uint16_t len, pad, preload_idx, pkt_len, parse_flag;
-	uint8_t csum_flag, num_segs;
+	uint16_t pkt_len; /* Sum of all BD segments */
+	uint16_t len; /* Length of first BD */
+	uint8_t num_segs = 1;
+	uint16_t pad;
+	uint16_t preload_idx;
+	uint8_t csum_flag;
+	uint16_t parse_flag;
 	enum rss_hash_type htype;
 	uint8_t tunn_parse_flag;
-	int ret;
+	uint8_t j;
 
 	hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
 	sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
@@ -915,6 +902,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		fp_cqe = &cqe->fast_path_regular;
 
 		len = rte_le_to_cpu_16(fp_cqe->len_on_first_bd);
+		pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
 		pad = fp_cqe->placement_offset;
 		assert((len + pad) <= rx_mb->buf_len);
 
@@ -979,25 +967,29 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 			rxq->rx_alloc_errors++;
 			break;
 		}
-
 		qede_rx_bd_ring_consume(rxq);
-
 		if (fp_cqe->bd_num > 1) {
-			pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
+			PMD_RX_LOG(DEBUG, rxq, "Jumbo-over-BD packet: %02x BDs"
+				   " len on first: %04x Total Len: %04x\n",
+				   fp_cqe->bd_num, len, pkt_len);
 			num_segs = fp_cqe->bd_num - 1;
-
-			rxq->rx_segs++;
-
-			pkt_len -= len;
 			seg1 = rx_mb;
-			ret = qede_process_sg_pkts(p_rxq, seg1, num_segs,
-						   pkt_len);
-			if (ret != ECORE_SUCCESS) {
-				qede_recycle_rx_bd_ring(rxq, qdev,
-							fp_cqe->bd_num);
+			if (qede_process_sg_pkts(p_rxq, seg1, num_segs,
+						 pkt_len - len))
 				goto next_cqe;
+			for (j = 0; j < num_segs; j++) {
+				if (qede_alloc_rx_buffer(rxq)) {
+					PMD_RX_LOG(ERR, rxq,
+						"Buffer allocation failed\n");
+					rte_eth_devices[rxq->port_id].
+						data->rx_mbuf_alloc_failed++;
+					rxq->rx_alloc_errors++;
+					break;
+				}
+				rxq->rx_segs++;
 			}
 		}
+		rxq->rx_segs++; /* for the first segment */
 
 		/* Prefetch next mbuf while processing current one. */
 		preload_idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
@@ -1007,7 +999,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		rx_mb->data_off = pad + RTE_PKTMBUF_HEADROOM;
 		rx_mb->nb_segs = fp_cqe->bd_num;
 		rx_mb->data_len = len;
-		rx_mb->pkt_len = fp_cqe->pkt_len;
+		rx_mb->pkt_len = pkt_len;
 		rx_mb->port = rxq->port_id;
 
 		htype = (uint8_t)GET_FIELD(fp_cqe->bitfields,
@@ -1114,17 +1106,16 @@ qede_process_tx_compl(struct ecore_dev *edev, struct qede_tx_queue *txq)
 }
 
 /* Populate scatter gather buffer descriptor fields */
-static inline uint16_t qede_encode_sg_bd(struct qede_tx_queue *p_txq,
-					 struct rte_mbuf *m_seg,
-					 uint16_t count,
-					 struct eth_tx_1st_bd *bd1)
+static inline uint8_t
+qede_encode_sg_bd(struct qede_tx_queue *p_txq, struct rte_mbuf *m_seg,
+		  struct eth_tx_1st_bd *bd1)
 {
 	struct qede_tx_queue *txq = p_txq;
 	struct eth_tx_2nd_bd *bd2 = NULL;
 	struct eth_tx_3rd_bd *bd3 = NULL;
 	struct eth_tx_bd *tx_bd = NULL;
-	uint16_t nb_segs = count;
 	dma_addr_t mapping;
+	uint8_t nb_segs = 1; /* min one segment per packet */
 
 	/* Check for scattered buffers */
 	while (m_seg) {
@@ -1133,28 +1124,27 @@ static inline uint16_t qede_encode_sg_bd(struct qede_tx_queue *p_txq,
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(bd2, 0, sizeof(*bd2));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			bd2->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			bd2->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			bd2->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(bd2, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD2 len %04x\n",
+				   m_seg->data_len);
 		} else if (nb_segs == 2) {
 			bd3 = (struct eth_tx_3rd_bd *)
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(bd3, 0, sizeof(*bd3));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			bd3->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			bd3->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			bd3->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(bd3, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD3 len %04x\n",
+				   m_seg->data_len);
 		} else {
 			tx_bd = (struct eth_tx_bd *)
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(tx_bd, 0, sizeof(*tx_bd));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			tx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			tx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			tx_bd->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(tx_bd, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD len %04x\n",
+				   m_seg->data_len);
 		}
 		nb_segs++;
-		bd1->data.nbds = nb_segs;
 		m_seg = m_seg->next;
 	}
 
@@ -1170,13 +1160,14 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	struct ecore_dev *edev = &qdev->edev;
 	struct qede_fastpath *fp;
 	struct eth_tx_1st_bd *bd1;
+	struct rte_mbuf *mbuf;
 	struct rte_mbuf *m_seg = NULL;
 	uint16_t nb_tx_pkts;
-	uint16_t nb_pkt_sent = 0;
 	uint16_t bd_prod;
 	uint16_t idx;
 	uint16_t tx_count;
-	uint16_t nb_segs = 0;
+	uint16_t nb_frags;
+	uint16_t nb_pkt_sent = 0;
 
 	fp = &qdev->fp_array[QEDE_RSS_COUNT(qdev) + txq->queue_id];
 
@@ -1198,19 +1189,19 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	while (nb_tx_pkts--) {
 		/* Fill the entry in the SW ring and the BDs in the FW ring */
 		idx = TX_PROD(txq);
-		struct rte_mbuf *mbuf = *tx_pkts++;
-
+		mbuf = *tx_pkts++;
 		txq->sw_tx_ring[idx].mbuf = mbuf;
 		bd1 = (struct eth_tx_1st_bd *)ecore_chain_produce(&txq->tx_pbl);
-		/* Zero init struct fields */
-		bd1->data.bd_flags.bitfields = 0;
-		bd1->data.bitfields = 0;
-
 		bd1->data.bd_flags.bitfields =
 			1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
+		/* FW 8.10.x specific change */
+		bd1->data.bitfields =
+			(mbuf->pkt_len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK)
+				<< ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
 		/* Map MBUF linear data for DMA and set in the first BD */
 		QEDE_BD_SET_ADDR_LEN(bd1, rte_mbuf_data_dma_addr(mbuf),
-				     mbuf->pkt_len);
+				     mbuf->data_len);
+		PMD_TX_LOG(INFO, txq, "BD1 len %04x\n", mbuf->data_len);
 
 		if (RTE_ETH_IS_TUNNEL_PKT(mbuf->packet_type)) {
 			PMD_TX_LOG(INFO, txq, "Tx tunnel packet\n");
@@ -1267,18 +1258,18 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 		/* Handle fragmented MBUF */
 		m_seg = mbuf->next;
-		nb_segs++;
-		bd1->data.nbds = nb_segs;
 		/* Encode scatter gather buffer descriptors if required */
-		nb_segs = qede_encode_sg_bd(txq, m_seg, nb_segs, bd1);
-		txq->nb_tx_avail = txq->nb_tx_avail - nb_segs;
-		nb_segs = 0;
+		nb_frags = qede_encode_sg_bd(txq, m_seg, bd1);
+		bd1->data.nbds = nb_frags;
+		txq->nb_tx_avail -= nb_frags;
 		txq->sw_tx_prod++;
 		rte_prefetch0(txq->sw_tx_ring[TX_PROD(txq)].mbuf);
 		bd_prod =
 		    rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
 		nb_pkt_sent++;
 		txq->xmit_pkts++;
+		PMD_TX_LOG(INFO, txq, "nbds = %d pkt_len = %04x\n",
+			   bd1->data.nbds, mbuf->pkt_len);
 	}
 
 	/* Write value of prod idx into bd_prod */
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index 2a8645a..a95b4ab 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -41,10 +41,6 @@
 		(bd)->addr.hi = rte_cpu_to_le_32(U64_HI(maddr)); \
 		(bd)->addr.lo = rte_cpu_to_le_32(U64_LO(maddr)); \
 		(bd)->nbytes = rte_cpu_to_le_16(len); \
-		/* FW 8.10.x specific change */  \
-		(bd)->data.bitfields = ((len) & \
-					 ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) \
-					<< ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; \
 	} while (0)
 
 #define CQE_HAS_VLAN(flags) \
-- 
1.7.10.3

[dpdk-dev] [PATCH 2/5] net/qede: fix minimum buffer size and scatter Rx check

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

 - Fix minimum RX buffer size to 1024B
 - Force enable scatter/gather mode if given RX buf size is lesser than MTU
 - Adjust RX buffer size to cache-line size with overhead included

Fixes: bec0228816c0 ("net/qede: support scatter gather")
Fixes: 2ea6f76aff40 ("qede: add core driver")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_ethdev.c |    3 +--
 drivers/net/qede/qede_rxtx.c   |   47 +++++++++++++++++-----------------------
 drivers/net/qede/qede_rxtx.h   |   11 ++++++++--
 3 files changed, 30 insertions(+), 31 deletions(-)

diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index c8581d8..b7606c8 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -968,8 +968,7 @@ qede_dev_info_get(struct rte_eth_dev *eth_dev,
 
 	PMD_INIT_FUNC_TRACE(edev);
 
-	dev_info->min_rx_bufsize = (uint32_t)(ETHER_MIN_MTU +
-					      QEDE_ETH_OVERHEAD);
+	dev_info->min_rx_bufsize = (uint32_t)QEDE_MIN_RX_BUFF_SIZE;
 	dev_info->max_rx_pktlen = (uint32_t)ETH_TX_MAX_NON_LSO_PKT_LEN;
 	dev_info->rx_desc_lim = qede_rx_desc_lim;
 	dev_info->tx_desc_lim = qede_tx_desc_lim;
diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index ecff5bc..aebe8cb 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -89,11 +89,11 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 {
 	struct qede_dev *qdev = dev->data->dev_private;
 	struct ecore_dev *edev = &qdev->edev;
-	struct rte_eth_dev_data *eth_data = dev->data;
+	struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
 	struct qede_rx_queue *rxq;
-	uint16_t pkt_len = (uint16_t)dev->data->dev_conf.rxmode.max_rx_pkt_len;
+	uint16_t max_rx_pkt_len;
+	uint16_t bufsz;
 	size_t size;
-	uint16_t data_size;
 	int rc;
 	int i;
 
@@ -127,34 +127,27 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	rxq->nb_rx_desc = nb_desc;
 	rxq->queue_id = queue_idx;
 	rxq->port_id = dev->data->port_id;
-
-	/* Sanity check */
-	data_size = (uint16_t)rte_pktmbuf_data_room_size(mp) -
-				RTE_PKTMBUF_HEADROOM;
-
-	if (pkt_len > data_size && !dev->data->scattered_rx) {
-		DP_ERR(edev, "MTU %u should not exceed dataroom %u\n",
-		       pkt_len, data_size);
-		rte_free(rxq);
-		return -EINVAL;
+	max_rx_pkt_len = (uint16_t)rxmode->max_rx_pkt_len;
+	qdev->mtu = max_rx_pkt_len;
+
+	/* Fix up RX buffer size */
+	bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
+	if ((rxmode->enable_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;
+		}
 	}
-
 	if (dev->data->scattered_rx)
-		rxq->rx_buf_size = data_size;
+		rxq->rx_buf_size = bufsz + QEDE_ETH_OVERHEAD;
 	else
-		rxq->rx_buf_size = pkt_len + QEDE_ETH_OVERHEAD;
-
-	qdev->mtu = pkt_len;
+		rxq->rx_buf_size = qdev->mtu + QEDE_ETH_OVERHEAD;
+	/* Align to cache-line size if needed */
+	rxq->rx_buf_size = QEDE_CEIL_TO_CACHE_LINE_SIZE(rxq->rx_buf_size);
 
-	DP_INFO(edev, "MTU = %u ; RX buffer = %u\n",
-		qdev->mtu, rxq->rx_buf_size);
-
-	if (pkt_len > ETHER_MAX_LEN) {
-		dev->data->dev_conf.rxmode.jumbo_frame = 1;
-		DP_NOTICE(edev, false, "jumbo frame enabled\n");
-	} else {
-		dev->data->dev_conf.rxmode.jumbo_frame = 0;
-	}
+	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);
 
 	/* Allocate the parallel driver ring for Rx buffers */
 	size = sizeof(*rxq->sw_rx_ring) * rxq->nb_rx_desc;
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index a95b4ab..9a393e9 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -51,14 +51,21 @@
 	((flags) & (PARSING_AND_ERR_FLAGS_TUNNEL8021QTAGEXIST_MASK \
 		<< PARSING_AND_ERR_FLAGS_TUNNEL8021QTAGEXIST_SHIFT))
 
+#define QEDE_MIN_RX_BUFF_SIZE		(1024)
+#define QEDE_VLAN_TAG_SIZE		(4)
+#define QEDE_LLC_SNAP_HDR_LEN		(8)
+
 /* Max supported alignment is 256 (8 shift)
  * minimal alignment shift 6 is optimal for 57xxx HW performance
  */
 #define QEDE_L1_CACHE_SHIFT	6
 #define QEDE_RX_ALIGN_SHIFT	(RTE_MAX(6, RTE_MIN(8, QEDE_L1_CACHE_SHIFT)))
 #define QEDE_FW_RX_ALIGN_END	(1UL << QEDE_RX_ALIGN_SHIFT)
-
-#define QEDE_ETH_OVERHEAD       (ETHER_HDR_LEN + 8 + 8 + QEDE_FW_RX_ALIGN_END)
+#define QEDE_CEIL_TO_CACHE_LINE_SIZE(n) (((n) + (QEDE_FW_RX_ALIGN_END - 1)) & \
+					~(QEDE_FW_RX_ALIGN_END - 1))
+/* Note: QEDE_LLC_SNAP_HDR_LEN is optional */
+#define QEDE_ETH_OVERHEAD	((ETHER_HDR_LEN) + ((2 * QEDE_VLAN_TAG_SIZE)) \
+				+ (QEDE_LLC_SNAP_HDR_LEN))
 
 #define QEDE_RSS_OFFLOAD_ALL    (ETH_RSS_IPV4			|\
 				 ETH_RSS_NONFRAG_IPV4_TCP	|\
-- 
1.7.10.3

[dpdk-dev] [PATCH 3/5] net/qede: fix PF fastpath status block index

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

Allocate double the number of fastpath status block index
since the PF RX/TX queues are not sharing the status block.
This is an interim solution till other parts of the code
is modified to handle the same.

Fixes: f1e4b6c0acee ("net/qede: fix status block index for VF queues")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_rxtx.c |   14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index aebe8cb..f20881c 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -431,13 +431,15 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 	struct ecore_dev *edev = &qdev->edev;
 	struct qede_fastpath *fp;
 	uint32_t num_sbs;
-	int rc, i;
+	uint16_t i;
+	uint16_t sb_idx;
+	int rc;
 
 	if (IS_VF(edev))
 		ecore_vf_get_num_sbs(ECORE_LEADING_HWFN(edev), &num_sbs);
 	else
-		num_sbs = (ecore_cxt_get_proto_cid_count
-			  (ECORE_LEADING_HWFN(edev), PROTOCOLID_ETH, NULL)) / 2;
+		num_sbs = ecore_cxt_get_proto_cid_count
+			  (ECORE_LEADING_HWFN(edev), PROTOCOLID_ETH, NULL);
 
 	if (num_sbs == 0) {
 		DP_ERR(edev, "No status blocks available\n");
@@ -455,7 +457,11 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 
 	for (i = 0; i < QEDE_QUEUE_CNT(qdev); i++) {
 		fp = &qdev->fp_array[i];
-		if (qede_alloc_mem_sb(qdev, fp->sb_info, i % num_sbs)) {
+		if (IS_VF(edev))
+			sb_idx = i % num_sbs;
+		else
+			sb_idx = i;
+		if (qede_alloc_mem_sb(qdev, fp->sb_info, sb_idx)) {
 			qede_free_fp_arrays(qdev);
 			return -ENOMEM;
 		}
-- 
1.7.10.3

[dpdk-dev] [PATCH 4/5] net/qede: fix per queue stats/xstats

[Rasesh Mody]

From: Rasesh Mody <Rasesh.Mody@cavium.com>

If value of number of rxq/txq is diffrent than
RTE_ETHDEV_QUEUE_STAT_CNTRS, limit per queue
stats/xstats to minimum of the two.

Fixes: 7634c5f91569 ("net/qede: add queue statistics")

Signed-off-by: Rasesh Mody <Rasesh.Mody@cavium.com>
---
 drivers/net/qede/qede_ethdev.c |   32 +++++++++++++++++++++++++++++---
 1 file changed, 29 insertions(+), 3 deletions(-)

diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index b7606c8..de8286c 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -1160,6 +1160,7 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	struct ecore_dev *edev = &qdev->edev;
 	struct ecore_eth_stats stats;
 	unsigned int i = 0, j = 0, qid;
+	unsigned int rxq_stat_cntrs, txq_stat_cntrs;
 	struct qede_tx_queue *txq;
 
 	qdev->ops->get_vport_stats(edev, &stats);
@@ -1193,6 +1194,17 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	eth_stats->oerrors = stats.tx_err_drop_pkts;
 
 	/* Queue stats */
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(qdev),
+			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	if ((rxq_stat_cntrs != QEDE_RSS_COUNT(qdev)) ||
+	    (txq_stat_cntrs != QEDE_TSS_COUNT(qdev)))
+		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 (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_RX) {
 			eth_stats->q_ipackets[i] =
@@ -1211,7 +1223,11 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 					rx_alloc_errors));
 			i++;
 		}
+		if (i == rxq_stat_cntrs)
+			break;
+	}
 
+	for (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_TX) {
 			txq = qdev->fp_array[(qid)].txqs[0];
 			eth_stats->q_opackets[j] =
@@ -1221,13 +1237,17 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 						  xmit_pkts)));
 			j++;
 		}
+		if (j == txq_stat_cntrs)
+			break;
 	}
 }
 
 static unsigned
 qede_get_xstats_count(struct qede_dev *qdev) {
 	return RTE_DIM(qede_xstats_strings) +
-		(RTE_DIM(qede_rxq_xstats_strings) * QEDE_RSS_COUNT(qdev));
+		(RTE_DIM(qede_rxq_xstats_strings) *
+		 RTE_MIN(QEDE_RSS_COUNT(qdev),
+			 RTE_ETHDEV_QUEUE_STAT_CNTRS));
 }
 
 static int
@@ -1237,6 +1257,7 @@ qede_get_xstats_names(__rte_unused struct rte_eth_dev *dev,
 	struct qede_dev *qdev = dev->data->dev_private;
 	const unsigned int stat_cnt = qede_get_xstats_count(qdev);
 	unsigned int i, qid, stat_idx = 0;
+	unsigned int rxq_stat_cntrs;
 
 	if (xstats_names != NULL) {
 		for (i = 0; i < RTE_DIM(qede_xstats_strings); i++) {
@@ -1247,7 +1268,9 @@ qede_get_xstats_names(__rte_unused struct rte_eth_dev *dev,
 			stat_idx++;
 		}
 
-		for (qid = 0; qid < QEDE_RSS_COUNT(qdev); qid++) {
+		rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+					 RTE_ETHDEV_QUEUE_STAT_CNTRS);
+		for (qid = 0; qid < rxq_stat_cntrs; qid++) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
 				snprintf(xstats_names[stat_idx].name,
 					sizeof(xstats_names[stat_idx].name),
@@ -1271,6 +1294,7 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 	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;
 
 	if (n < num)
 		return num;
@@ -1283,7 +1307,9 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 		stat_idx++;
 	}
 
-	for (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+				 RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	for (qid = 0; qid < rxq_stat_cntrs; qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_RX) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
 				xstats[stat_idx].value = *(uint64_t *)(
-- 
1.7.10.3

Re: [dpdk-dev] [dpdk-stable] [PATCH 1/5] net/qede: fix scatter-gather issue

[Ferruh Yigit]

On 12/31/2016 8:16 AM, Rasesh Mody wrote:
> From: Harish Patil <harish.patil@qlogic.com>
> 
>  - Make qede_process_sg_pkts() inline and add unlikely check
>  - Fix mbuf segment chaining logic in qede_process_sg_pkts()
>  - Change qede_encode_sg_bd() to return total segments required
>  - Fix first TX buffer descriptor's length
>  - Replace repeatitive code using a macro
> 
> Fixes: bec0228816c0 ("net/qede: support scatter gather")
> 
> Signed-off-by: Harish Patil <harish.patil@qlogic.com>
> ---

Hi Harish,

This patch doesn't apply cleanly on top next-net, after pci_dev changes
rebased into next-net tree.

Can you please send the new version of the patchset rebased on top of
latest next-net?

Thanks,
ferruh

<...>

Re: [dpdk-dev] [dpdk-stable] [PATCH 1/5] net/qede: fix scatter-gather issue

[Harish Patil]

>
>On 12/31/2016 8:16 AM, Rasesh Mody wrote:
>> From: Harish Patil <harish.patil@qlogic.com>
>> 
>>  - Make qede_process_sg_pkts() inline and add unlikely check
>>  - Fix mbuf segment chaining logic in qede_process_sg_pkts()
>>  - Change qede_encode_sg_bd() to return total segments required
>>  - Fix first TX buffer descriptor's length
>>  - Replace repeatitive code using a macro
>> 
>> Fixes: bec0228816c0 ("net/qede: support scatter gather")
>> 
>> Signed-off-by: Harish Patil <harish.patil@qlogic.com>
>> ---
>
>Hi Harish,
>
>This patch doesn't apply cleanly on top next-net, after pci_dev changes
>rebased into next-net tree.
>
>Can you please send the new version of the patchset rebased on top of
>latest next-net?
>
>Thanks,
>ferruh
>
><...>
>
Hi Ferruh,
Sure will do that.
Thanks,

Harish

Re: [dpdk-dev] [dpdk-stable] [PATCH 1/5] net/qede: fix scatter-gather issue

[Mody, Rasesh]

> From: Harish Patil [mailto:harish.patil@qlogic.com]
> Sent: Thursday, January 05, 2017 1:12 PM
> 
> >
> >On 12/31/2016 8:16 AM, Rasesh Mody wrote:
> >> From: Harish Patil <harish.patil@qlogic.com>
> >>
> >>  - Make qede_process_sg_pkts() inline and add unlikely check
> >>  - Fix mbuf segment chaining logic in qede_process_sg_pkts()
> >>  - Change qede_encode_sg_bd() to return total segments required
> >>  - Fix first TX buffer descriptor's length
> >>  - Replace repeatitive code using a macro
> >>
> >> Fixes: bec0228816c0 ("net/qede: support scatter gather")
> >>
> >> Signed-off-by: Harish Patil <harish.patil@qlogic.com>
> >> ---
> >
> >Hi Harish,
> >
> >This patch doesn't apply cleanly on top next-net, after pci_dev changes
> >rebased into next-net tree.

We are not seeing failure when applying [PATCH 1/5] net/qede: fix scatter-gather issue on latest dpdk-next-net. However, we do see a failure when trying to apply [PATCH 2/5] net/qede: fix minimum buffer size and scatter Rx check due to pci_dev changes rebased into next-net tree. We'll submit a v2 series rebased on latest next-net.
Thanks!
-Rasesh

[dpdk-dev] [PATCH v2 1/5] net/qede: fix scatter-gather issue

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

 - Make qede_process_sg_pkts() inline and add unlikely check
 - Fix mbuf segment chaining logic in qede_process_sg_pkts()
 - Change qede_encode_sg_bd() to return total segments required
 - Fix first TX buffer descriptor's length
 - Replace repeatitive code using a macro

Fixes: bec0228816c0 ("net/qede: support scatter gather")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_rxtx.c |  139 ++++++++++++++++++++----------------------
 drivers/net/qede/qede_rxtx.h |    4 --
 2 files changed, 65 insertions(+), 78 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index 814d384..ecff5bc 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -810,39 +810,28 @@ static inline uint32_t qede_rx_cqe_to_tunn_pkt_type(uint16_t flags)
 		return RTE_PTYPE_UNKNOWN;
 }
 
-
-int qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
-			 int num_segs, uint16_t pkt_len)
+static inline int
+qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
+		     uint8_t num_segs, uint16_t pkt_len)
 {
 	struct qede_rx_queue *rxq = p_rxq;
 	struct qede_dev *qdev = rxq->qdev;
 	struct ecore_dev *edev = &qdev->edev;
-	uint16_t sw_rx_index, cur_size;
-
 	register struct rte_mbuf *seg1 = NULL;
 	register struct rte_mbuf *seg2 = NULL;
+	uint16_t sw_rx_index;
+	uint16_t cur_size;
 
 	seg1 = rx_mb;
 	while (num_segs) {
-		cur_size = pkt_len > rxq->rx_buf_size ?
-				rxq->rx_buf_size : pkt_len;
-		if (!cur_size) {
-			PMD_RX_LOG(DEBUG, rxq,
-				   "SG packet, len and num BD mismatch\n");
+		cur_size = pkt_len > rxq->rx_buf_size ? rxq->rx_buf_size :
+							pkt_len;
+		if (unlikely(!cur_size)) {
+			PMD_RX_LOG(ERR, rxq, "Length is 0 while %u BDs"
+				   " left for mapping jumbo\n", num_segs);
 			qede_recycle_rx_bd_ring(rxq, qdev, num_segs);
 			return -EINVAL;
 		}
-
-		if (qede_alloc_rx_buffer(rxq)) {
-			uint8_t index;
-
-			PMD_RX_LOG(DEBUG, rxq, "Buffer allocation failed\n");
-			index = rxq->port_id;
-			rte_eth_devices[index].data->rx_mbuf_alloc_failed++;
-			rxq->rx_alloc_errors++;
-			return -ENOMEM;
-		}
-
 		sw_rx_index = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
 		seg2 = rxq->sw_rx_ring[sw_rx_index].mbuf;
 		qede_rx_bd_ring_consume(rxq);
@@ -852,16 +841,9 @@ int qede_process_sg_pkts(void *p_rxq,  struct rte_mbuf *rx_mb,
 		seg1 = seg1->next;
 		num_segs--;
 		rxq->rx_segs++;
-		continue;
 	}
-	seg1 = NULL;
-
-	if (pkt_len)
-		PMD_RX_LOG(DEBUG, rxq,
-			   "Mapped all BDs of jumbo, but still have %d bytes\n",
-			   pkt_len);
 
-	return ECORE_SUCCESS;
+	return 0;
 }
 
 uint16_t
@@ -878,11 +860,16 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 	register struct rte_mbuf *rx_mb = NULL;
 	register struct rte_mbuf *seg1 = NULL;
 	enum eth_rx_cqe_type cqe_type;
-	uint16_t len, pad, preload_idx, pkt_len, parse_flag;
-	uint8_t csum_flag, num_segs;
+	uint16_t pkt_len; /* Sum of all BD segments */
+	uint16_t len; /* Length of first BD */
+	uint8_t num_segs = 1;
+	uint16_t pad;
+	uint16_t preload_idx;
+	uint8_t csum_flag;
+	uint16_t parse_flag;
 	enum rss_hash_type htype;
 	uint8_t tunn_parse_flag;
-	int ret;
+	uint8_t j;
 
 	hw_comp_cons = rte_le_to_cpu_16(*rxq->hw_cons_ptr);
 	sw_comp_cons = ecore_chain_get_cons_idx(&rxq->rx_comp_ring);
@@ -915,6 +902,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		fp_cqe = &cqe->fast_path_regular;
 
 		len = rte_le_to_cpu_16(fp_cqe->len_on_first_bd);
+		pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
 		pad = fp_cqe->placement_offset;
 		assert((len + pad) <= rx_mb->buf_len);
 
@@ -979,25 +967,29 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 			rxq->rx_alloc_errors++;
 			break;
 		}
-
 		qede_rx_bd_ring_consume(rxq);
-
 		if (fp_cqe->bd_num > 1) {
-			pkt_len = rte_le_to_cpu_16(fp_cqe->pkt_len);
+			PMD_RX_LOG(DEBUG, rxq, "Jumbo-over-BD packet: %02x BDs"
+				   " len on first: %04x Total Len: %04x\n",
+				   fp_cqe->bd_num, len, pkt_len);
 			num_segs = fp_cqe->bd_num - 1;
-
-			rxq->rx_segs++;
-
-			pkt_len -= len;
 			seg1 = rx_mb;
-			ret = qede_process_sg_pkts(p_rxq, seg1, num_segs,
-						   pkt_len);
-			if (ret != ECORE_SUCCESS) {
-				qede_recycle_rx_bd_ring(rxq, qdev,
-							fp_cqe->bd_num);
+			if (qede_process_sg_pkts(p_rxq, seg1, num_segs,
+						 pkt_len - len))
 				goto next_cqe;
+			for (j = 0; j < num_segs; j++) {
+				if (qede_alloc_rx_buffer(rxq)) {
+					PMD_RX_LOG(ERR, rxq,
+						"Buffer allocation failed\n");
+					rte_eth_devices[rxq->port_id].
+						data->rx_mbuf_alloc_failed++;
+					rxq->rx_alloc_errors++;
+					break;
+				}
+				rxq->rx_segs++;
 			}
 		}
+		rxq->rx_segs++; /* for the first segment */
 
 		/* Prefetch next mbuf while processing current one. */
 		preload_idx = rxq->sw_rx_cons & NUM_RX_BDS(rxq);
@@ -1007,7 +999,7 @@ qede_recv_pkts(void *p_rxq, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
 		rx_mb->data_off = pad + RTE_PKTMBUF_HEADROOM;
 		rx_mb->nb_segs = fp_cqe->bd_num;
 		rx_mb->data_len = len;
-		rx_mb->pkt_len = fp_cqe->pkt_len;
+		rx_mb->pkt_len = pkt_len;
 		rx_mb->port = rxq->port_id;
 
 		htype = (uint8_t)GET_FIELD(fp_cqe->bitfields,
@@ -1114,17 +1106,16 @@ qede_process_tx_compl(struct ecore_dev *edev, struct qede_tx_queue *txq)
 }
 
 /* Populate scatter gather buffer descriptor fields */
-static inline uint16_t qede_encode_sg_bd(struct qede_tx_queue *p_txq,
-					 struct rte_mbuf *m_seg,
-					 uint16_t count,
-					 struct eth_tx_1st_bd *bd1)
+static inline uint8_t
+qede_encode_sg_bd(struct qede_tx_queue *p_txq, struct rte_mbuf *m_seg,
+		  struct eth_tx_1st_bd *bd1)
 {
 	struct qede_tx_queue *txq = p_txq;
 	struct eth_tx_2nd_bd *bd2 = NULL;
 	struct eth_tx_3rd_bd *bd3 = NULL;
 	struct eth_tx_bd *tx_bd = NULL;
-	uint16_t nb_segs = count;
 	dma_addr_t mapping;
+	uint8_t nb_segs = 1; /* min one segment per packet */
 
 	/* Check for scattered buffers */
 	while (m_seg) {
@@ -1133,28 +1124,27 @@ static inline uint16_t qede_encode_sg_bd(struct qede_tx_queue *p_txq,
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(bd2, 0, sizeof(*bd2));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			bd2->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			bd2->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			bd2->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(bd2, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD2 len %04x\n",
+				   m_seg->data_len);
 		} else if (nb_segs == 2) {
 			bd3 = (struct eth_tx_3rd_bd *)
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(bd3, 0, sizeof(*bd3));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			bd3->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			bd3->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			bd3->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(bd3, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD3 len %04x\n",
+				   m_seg->data_len);
 		} else {
 			tx_bd = (struct eth_tx_bd *)
 				ecore_chain_produce(&txq->tx_pbl);
 			memset(tx_bd, 0, sizeof(*tx_bd));
 			mapping = rte_mbuf_data_dma_addr(m_seg);
-			tx_bd->addr.hi = rte_cpu_to_le_32(U64_HI(mapping));
-			tx_bd->addr.lo = rte_cpu_to_le_32(U64_LO(mapping));
-			tx_bd->nbytes = rte_cpu_to_le_16(m_seg->data_len);
+			QEDE_BD_SET_ADDR_LEN(tx_bd, mapping, m_seg->data_len);
+			PMD_TX_LOG(DEBUG, txq, "BD len %04x\n",
+				   m_seg->data_len);
 		}
 		nb_segs++;
-		bd1->data.nbds = nb_segs;
 		m_seg = m_seg->next;
 	}
 
@@ -1170,13 +1160,14 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	struct ecore_dev *edev = &qdev->edev;
 	struct qede_fastpath *fp;
 	struct eth_tx_1st_bd *bd1;
+	struct rte_mbuf *mbuf;
 	struct rte_mbuf *m_seg = NULL;
 	uint16_t nb_tx_pkts;
-	uint16_t nb_pkt_sent = 0;
 	uint16_t bd_prod;
 	uint16_t idx;
 	uint16_t tx_count;
-	uint16_t nb_segs = 0;
+	uint16_t nb_frags;
+	uint16_t nb_pkt_sent = 0;
 
 	fp = &qdev->fp_array[QEDE_RSS_COUNT(qdev) + txq->queue_id];
 
@@ -1198,19 +1189,19 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 	while (nb_tx_pkts--) {
 		/* Fill the entry in the SW ring and the BDs in the FW ring */
 		idx = TX_PROD(txq);
-		struct rte_mbuf *mbuf = *tx_pkts++;
-
+		mbuf = *tx_pkts++;
 		txq->sw_tx_ring[idx].mbuf = mbuf;
 		bd1 = (struct eth_tx_1st_bd *)ecore_chain_produce(&txq->tx_pbl);
-		/* Zero init struct fields */
-		bd1->data.bd_flags.bitfields = 0;
-		bd1->data.bitfields = 0;
-
 		bd1->data.bd_flags.bitfields =
 			1 << ETH_TX_1ST_BD_FLAGS_START_BD_SHIFT;
+		/* FW 8.10.x specific change */
+		bd1->data.bitfields =
+			(mbuf->pkt_len & ETH_TX_DATA_1ST_BD_PKT_LEN_MASK)
+				<< ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT;
 		/* Map MBUF linear data for DMA and set in the first BD */
 		QEDE_BD_SET_ADDR_LEN(bd1, rte_mbuf_data_dma_addr(mbuf),
-				     mbuf->pkt_len);
+				     mbuf->data_len);
+		PMD_TX_LOG(INFO, txq, "BD1 len %04x\n", mbuf->data_len);
 
 		if (RTE_ETH_IS_TUNNEL_PKT(mbuf->packet_type)) {
 			PMD_TX_LOG(INFO, txq, "Tx tunnel packet\n");
@@ -1267,18 +1258,18 @@ qede_xmit_pkts(void *p_txq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 
 		/* Handle fragmented MBUF */
 		m_seg = mbuf->next;
-		nb_segs++;
-		bd1->data.nbds = nb_segs;
 		/* Encode scatter gather buffer descriptors if required */
-		nb_segs = qede_encode_sg_bd(txq, m_seg, nb_segs, bd1);
-		txq->nb_tx_avail = txq->nb_tx_avail - nb_segs;
-		nb_segs = 0;
+		nb_frags = qede_encode_sg_bd(txq, m_seg, bd1);
+		bd1->data.nbds = nb_frags;
+		txq->nb_tx_avail -= nb_frags;
 		txq->sw_tx_prod++;
 		rte_prefetch0(txq->sw_tx_ring[TX_PROD(txq)].mbuf);
 		bd_prod =
 		    rte_cpu_to_le_16(ecore_chain_get_prod_idx(&txq->tx_pbl));
 		nb_pkt_sent++;
 		txq->xmit_pkts++;
+		PMD_TX_LOG(INFO, txq, "nbds = %d pkt_len = %04x\n",
+			   bd1->data.nbds, mbuf->pkt_len);
 	}
 
 	/* Write value of prod idx into bd_prod */
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index 2a8645a..a95b4ab 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -41,10 +41,6 @@
 		(bd)->addr.hi = rte_cpu_to_le_32(U64_HI(maddr)); \
 		(bd)->addr.lo = rte_cpu_to_le_32(U64_LO(maddr)); \
 		(bd)->nbytes = rte_cpu_to_le_16(len); \
-		/* FW 8.10.x specific change */  \
-		(bd)->data.bitfields = ((len) & \
-					 ETH_TX_DATA_1ST_BD_PKT_LEN_MASK) \
-					<< ETH_TX_DATA_1ST_BD_PKT_LEN_SHIFT; \
 	} while (0)
 
 #define CQE_HAS_VLAN(flags) \
-- 
1.7.10.3

[dpdk-dev] [PATCH v2 2/5] net/qede: fix minimum buffer size and scatter Rx check

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

 - Fix minimum RX buffer size to 1024B
 - Force enable scatter/gather mode if given RX buf size is lesser than MTU
 - Adjust RX buffer size to cache-line size with overhead included

Fixes: bec0228816c0 ("net/qede: support scatter gather")
Fixes: 2ea6f76aff40 ("qede: add core driver")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_ethdev.c |    3 +--
 drivers/net/qede/qede_rxtx.c   |   47 +++++++++++++++++-----------------------
 drivers/net/qede/qede_rxtx.h   |   11 ++++++++--
 3 files changed, 30 insertions(+), 31 deletions(-)

diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index b7886f4..0b40d1b 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -969,8 +969,7 @@ qede_dev_info_get(struct rte_eth_dev *eth_dev,
 	PMD_INIT_FUNC_TRACE(edev);
 
 	dev_info->pci_dev = RTE_DEV_TO_PCI(eth_dev->device);
-	dev_info->min_rx_bufsize = (uint32_t)(ETHER_MIN_MTU +
-					      QEDE_ETH_OVERHEAD);
+	dev_info->min_rx_bufsize = (uint32_t)QEDE_MIN_RX_BUFF_SIZE;
 	dev_info->max_rx_pktlen = (uint32_t)ETH_TX_MAX_NON_LSO_PKT_LEN;
 	dev_info->rx_desc_lim = qede_rx_desc_lim;
 	dev_info->tx_desc_lim = qede_tx_desc_lim;
diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index ecff5bc..aebe8cb 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -89,11 +89,11 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 {
 	struct qede_dev *qdev = dev->data->dev_private;
 	struct ecore_dev *edev = &qdev->edev;
-	struct rte_eth_dev_data *eth_data = dev->data;
+	struct rte_eth_rxmode *rxmode = &dev->data->dev_conf.rxmode;
 	struct qede_rx_queue *rxq;
-	uint16_t pkt_len = (uint16_t)dev->data->dev_conf.rxmode.max_rx_pkt_len;
+	uint16_t max_rx_pkt_len;
+	uint16_t bufsz;
 	size_t size;
-	uint16_t data_size;
 	int rc;
 	int i;
 
@@ -127,34 +127,27 @@ qede_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	rxq->nb_rx_desc = nb_desc;
 	rxq->queue_id = queue_idx;
 	rxq->port_id = dev->data->port_id;
-
-	/* Sanity check */
-	data_size = (uint16_t)rte_pktmbuf_data_room_size(mp) -
-				RTE_PKTMBUF_HEADROOM;
-
-	if (pkt_len > data_size && !dev->data->scattered_rx) {
-		DP_ERR(edev, "MTU %u should not exceed dataroom %u\n",
-		       pkt_len, data_size);
-		rte_free(rxq);
-		return -EINVAL;
+	max_rx_pkt_len = (uint16_t)rxmode->max_rx_pkt_len;
+	qdev->mtu = max_rx_pkt_len;
+
+	/* Fix up RX buffer size */
+	bufsz = (uint16_t)rte_pktmbuf_data_room_size(mp) - RTE_PKTMBUF_HEADROOM;
+	if ((rxmode->enable_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;
+		}
 	}
-
 	if (dev->data->scattered_rx)
-		rxq->rx_buf_size = data_size;
+		rxq->rx_buf_size = bufsz + QEDE_ETH_OVERHEAD;
 	else
-		rxq->rx_buf_size = pkt_len + QEDE_ETH_OVERHEAD;
-
-	qdev->mtu = pkt_len;
+		rxq->rx_buf_size = qdev->mtu + QEDE_ETH_OVERHEAD;
+	/* Align to cache-line size if needed */
+	rxq->rx_buf_size = QEDE_CEIL_TO_CACHE_LINE_SIZE(rxq->rx_buf_size);
 
-	DP_INFO(edev, "MTU = %u ; RX buffer = %u\n",
-		qdev->mtu, rxq->rx_buf_size);
-
-	if (pkt_len > ETHER_MAX_LEN) {
-		dev->data->dev_conf.rxmode.jumbo_frame = 1;
-		DP_NOTICE(edev, false, "jumbo frame enabled\n");
-	} else {
-		dev->data->dev_conf.rxmode.jumbo_frame = 0;
-	}
+	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);
 
 	/* Allocate the parallel driver ring for Rx buffers */
 	size = sizeof(*rxq->sw_rx_ring) * rxq->nb_rx_desc;
diff --git a/drivers/net/qede/qede_rxtx.h b/drivers/net/qede/qede_rxtx.h
index a95b4ab..9a393e9 100644
--- a/drivers/net/qede/qede_rxtx.h
+++ b/drivers/net/qede/qede_rxtx.h
@@ -51,14 +51,21 @@
 	((flags) & (PARSING_AND_ERR_FLAGS_TUNNEL8021QTAGEXIST_MASK \
 		<< PARSING_AND_ERR_FLAGS_TUNNEL8021QTAGEXIST_SHIFT))
 
+#define QEDE_MIN_RX_BUFF_SIZE		(1024)
+#define QEDE_VLAN_TAG_SIZE		(4)
+#define QEDE_LLC_SNAP_HDR_LEN		(8)
+
 /* Max supported alignment is 256 (8 shift)
  * minimal alignment shift 6 is optimal for 57xxx HW performance
  */
 #define QEDE_L1_CACHE_SHIFT	6
 #define QEDE_RX_ALIGN_SHIFT	(RTE_MAX(6, RTE_MIN(8, QEDE_L1_CACHE_SHIFT)))
 #define QEDE_FW_RX_ALIGN_END	(1UL << QEDE_RX_ALIGN_SHIFT)
-
-#define QEDE_ETH_OVERHEAD       (ETHER_HDR_LEN + 8 + 8 + QEDE_FW_RX_ALIGN_END)
+#define QEDE_CEIL_TO_CACHE_LINE_SIZE(n) (((n) + (QEDE_FW_RX_ALIGN_END - 1)) & \
+					~(QEDE_FW_RX_ALIGN_END - 1))
+/* Note: QEDE_LLC_SNAP_HDR_LEN is optional */
+#define QEDE_ETH_OVERHEAD	((ETHER_HDR_LEN) + ((2 * QEDE_VLAN_TAG_SIZE)) \
+				+ (QEDE_LLC_SNAP_HDR_LEN))
 
 #define QEDE_RSS_OFFLOAD_ALL    (ETH_RSS_IPV4			|\
 				 ETH_RSS_NONFRAG_IPV4_TCP	|\
-- 
1.7.10.3

[dpdk-dev] [PATCH v2 3/5] net/qede: fix PF fastpath status block index

[Rasesh Mody]

From: Harish Patil <harish.patil@qlogic.com>

Allocate double the number of fastpath status block index
since the PF RX/TX queues are not sharing the status block.
This is an interim solution till other parts of the code
is modified to handle the same.

Fixes: f1e4b6c0acee ("net/qede: fix status block index for VF queues")

Signed-off-by: Harish Patil <harish.patil@qlogic.com>
---
 drivers/net/qede/qede_rxtx.c |   14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/drivers/net/qede/qede_rxtx.c b/drivers/net/qede/qede_rxtx.c
index aebe8cb..f20881c 100644
--- a/drivers/net/qede/qede_rxtx.c
+++ b/drivers/net/qede/qede_rxtx.c
@@ -431,13 +431,15 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 	struct ecore_dev *edev = &qdev->edev;
 	struct qede_fastpath *fp;
 	uint32_t num_sbs;
-	int rc, i;
+	uint16_t i;
+	uint16_t sb_idx;
+	int rc;
 
 	if (IS_VF(edev))
 		ecore_vf_get_num_sbs(ECORE_LEADING_HWFN(edev), &num_sbs);
 	else
-		num_sbs = (ecore_cxt_get_proto_cid_count
-			  (ECORE_LEADING_HWFN(edev), PROTOCOLID_ETH, NULL)) / 2;
+		num_sbs = ecore_cxt_get_proto_cid_count
+			  (ECORE_LEADING_HWFN(edev), PROTOCOLID_ETH, NULL);
 
 	if (num_sbs == 0) {
 		DP_ERR(edev, "No status blocks available\n");
@@ -455,7 +457,11 @@ int qede_alloc_fp_resc(struct qede_dev *qdev)
 
 	for (i = 0; i < QEDE_QUEUE_CNT(qdev); i++) {
 		fp = &qdev->fp_array[i];
-		if (qede_alloc_mem_sb(qdev, fp->sb_info, i % num_sbs)) {
+		if (IS_VF(edev))
+			sb_idx = i % num_sbs;
+		else
+			sb_idx = i;
+		if (qede_alloc_mem_sb(qdev, fp->sb_info, sb_idx)) {
 			qede_free_fp_arrays(qdev);
 			return -ENOMEM;
 		}
-- 
1.7.10.3

[dpdk-dev] [PATCH v2 4/5] net/qede: fix per queue stats/xstats

[Rasesh Mody]

From: Rasesh Mody <Rasesh.Mody@cavium.com>

If value of number of rxq/txq is diffrent than
RTE_ETHDEV_QUEUE_STAT_CNTRS, limit per queue
stats/xstats to minimum of the two.

Fixes: 7634c5f91569 ("net/qede: add queue statistics")

Signed-off-by: Rasesh Mody <Rasesh.Mody@cavium.com>
---
 drivers/net/qede/qede_ethdev.c |   32 +++++++++++++++++++++++++++++---
 1 file changed, 29 insertions(+), 3 deletions(-)

diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index 0b40d1b..6d90c46 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -1162,6 +1162,7 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	struct ecore_dev *edev = &qdev->edev;
 	struct ecore_eth_stats stats;
 	unsigned int i = 0, j = 0, qid;
+	unsigned int rxq_stat_cntrs, txq_stat_cntrs;
 	struct qede_tx_queue *txq;
 
 	qdev->ops->get_vport_stats(edev, &stats);
@@ -1195,6 +1196,17 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 	eth_stats->oerrors = stats.tx_err_drop_pkts;
 
 	/* Queue stats */
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	txq_stat_cntrs = RTE_MIN(QEDE_TSS_COUNT(qdev),
+			       RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	if ((rxq_stat_cntrs != QEDE_RSS_COUNT(qdev)) ||
+	    (txq_stat_cntrs != QEDE_TSS_COUNT(qdev)))
+		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 (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_RX) {
 			eth_stats->q_ipackets[i] =
@@ -1213,7 +1225,11 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 					rx_alloc_errors));
 			i++;
 		}
+		if (i == rxq_stat_cntrs)
+			break;
+	}
 
+	for (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_TX) {
 			txq = qdev->fp_array[(qid)].txqs[0];
 			eth_stats->q_opackets[j] =
@@ -1223,13 +1239,17 @@ qede_get_stats(struct rte_eth_dev *eth_dev, struct rte_eth_stats *eth_stats)
 						  xmit_pkts)));
 			j++;
 		}
+		if (j == txq_stat_cntrs)
+			break;
 	}
 }
 
 static unsigned
 qede_get_xstats_count(struct qede_dev *qdev) {
 	return RTE_DIM(qede_xstats_strings) +
-		(RTE_DIM(qede_rxq_xstats_strings) * QEDE_RSS_COUNT(qdev));
+		(RTE_DIM(qede_rxq_xstats_strings) *
+		 RTE_MIN(QEDE_RSS_COUNT(qdev),
+			 RTE_ETHDEV_QUEUE_STAT_CNTRS));
 }
 
 static int
@@ -1239,6 +1259,7 @@ qede_get_xstats_names(__rte_unused struct rte_eth_dev *dev,
 	struct qede_dev *qdev = dev->data->dev_private;
 	const unsigned int stat_cnt = qede_get_xstats_count(qdev);
 	unsigned int i, qid, stat_idx = 0;
+	unsigned int rxq_stat_cntrs;
 
 	if (xstats_names != NULL) {
 		for (i = 0; i < RTE_DIM(qede_xstats_strings); i++) {
@@ -1249,7 +1270,9 @@ qede_get_xstats_names(__rte_unused struct rte_eth_dev *dev,
 			stat_idx++;
 		}
 
-		for (qid = 0; qid < QEDE_RSS_COUNT(qdev); qid++) {
+		rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+					 RTE_ETHDEV_QUEUE_STAT_CNTRS);
+		for (qid = 0; qid < rxq_stat_cntrs; qid++) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
 				snprintf(xstats_names[stat_idx].name,
 					sizeof(xstats_names[stat_idx].name),
@@ -1273,6 +1296,7 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 	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;
 
 	if (n < num)
 		return num;
@@ -1285,7 +1309,9 @@ qede_get_xstats(struct rte_eth_dev *dev, struct rte_eth_xstat *xstats,
 		stat_idx++;
 	}
 
-	for (qid = 0; qid < QEDE_QUEUE_CNT(qdev); qid++) {
+	rxq_stat_cntrs = RTE_MIN(QEDE_RSS_COUNT(qdev),
+				 RTE_ETHDEV_QUEUE_STAT_CNTRS);
+	for (qid = 0; qid < rxq_stat_cntrs; qid++) {
 		if (qdev->fp_array[qid].type & QEDE_FASTPATH_RX) {
 			for (i = 0; i < RTE_DIM(qede_rxq_xstats_strings); i++) {
 				xstats[stat_idx].value = *(uint64_t *)(
-- 
1.7.10.3

[dpdk-dev] [PATCH v2 5/5] net/qede: convert few DP_NOTICE and DP_INFO to DP_ERR

[Rasesh Mody]

From: Rasesh Mody <Rasesh.Mody@cavium.com>

Signed-off-by: Rasesh Mody <Rasesh.Mody@cavium.com>
---
 drivers/net/qede/base/ecore_mcp.c |    2 +-
 drivers/net/qede/qede_ethdev.c    |   11 +++++------
 2 files changed, 6 insertions(+), 7 deletions(-)

diff --git a/drivers/net/qede/base/ecore_mcp.c b/drivers/net/qede/base/ecore_mcp.c
index bb13828..f634d98 100644
--- a/drivers/net/qede/base/ecore_mcp.c
+++ b/drivers/net/qede/base/ecore_mcp.c
@@ -931,7 +931,7 @@ static void ecore_mcp_send_protocol_stats(struct ecore_hwfn *p_hwfn,
 		hsi_param = DRV_MSG_CODE_STATS_TYPE_LAN;
 		break;
 	default:
-		DP_NOTICE(p_hwfn, false, "Invalid protocol type %d\n", type);
+		DP_INFO(p_hwfn, "Invalid protocol type %d\n", type);
 		return;
 	}
 
diff --git a/drivers/net/qede/qede_ethdev.c b/drivers/net/qede/qede_ethdev.c
index 6d90c46..c67fbb6 100644
--- a/drivers/net/qede/qede_ethdev.c
+++ b/drivers/net/qede/qede_ethdev.c
@@ -650,7 +650,7 @@ static void qede_vlan_offload_set(struct rte_eth_dev *eth_dev, int mask)
 			qede_vlan_filter_set(eth_dev, 0, 1);
 		} else {
 			if (qdev->configured_vlans > 1) { /* Excluding VLAN0 */
-				DP_NOTICE(edev, false,
+				DP_ERR(edev,
 				  " Please remove existing VLAN filters"
 				  " before disabling VLAN filtering\n");
 				/* Signal app that VLAN filtering is still
@@ -684,7 +684,7 @@ static int qede_vlan_filter_set(struct rte_eth_dev *eth_dev,
 
 	if (on) {
 		if (qdev->configured_vlans == dev_info->num_vlan_filters) {
-			DP_INFO(edev, "Reached max VLAN filter limit"
+			DP_ERR(edev, "Reached max VLAN filter limit"
 				      " enabling accept_any_vlan\n");
 			qede_config_accept_any_vlan(qdev, true);
 			return 0;
@@ -849,14 +849,13 @@ static int qede_dev_configure(struct rte_eth_dev *eth_dev)
 	if (edev->num_hwfns > 1) {
 		if (eth_dev->data->nb_rx_queues < 2 ||
 		    eth_dev->data->nb_tx_queues < 2) {
-			DP_NOTICE(edev, false,
-				  "100G mode needs min. 2 RX/TX queues\n");
+			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_NOTICE(edev, false,
+			DP_ERR(edev,
 				  "100G mode needs even no. of RX/TX queues\n");
 			return -EINVAL;
 		}
@@ -867,7 +866,7 @@ static int qede_dev_configure(struct rte_eth_dev *eth_dev)
 		eth_dev->data->scattered_rx = 1;
 
 	if (rxmode->enable_lro == 1) {
-		DP_INFO(edev, "LRO is not supported\n");
+		DP_ERR(edev, "LRO is not supported\n");
 		return -EINVAL;
 	}
 
-- 
1.7.10.3

Re: [dpdk-dev] [PATCH v2 1/5] net/qede: fix scatter-gather issue

[Ferruh Yigit]

On 1/6/2017 8:16 AM, Rasesh Mody wrote:
> From: Harish Patil <harish.patil@qlogic.com>
> 
>  - Make qede_process_sg_pkts() inline and add unlikely check
>  - Fix mbuf segment chaining logic in qede_process_sg_pkts()
>  - Change qede_encode_sg_bd() to return total segments required
>  - Fix first TX buffer descriptor's length
>  - Replace repeatitive code using a macro
> 
> Fixes: bec0228816c0 ("net/qede: support scatter gather")
> 
> Signed-off-by: Harish Patil <harish.patil@qlogic.com>

Series applied to dpdk-next-net/master, thanks.