[dpdk-dev] [PATCH v5 10/18] net/axgbe: add transmit and receive data path apis

[Ravi Kumar <Ravi1.kumar@amd.com>]

Supported scalar implementation for RX data path.
Supported scalar and vector implementation for TX data path.

Signed-off-by: Ravi Kumar <Ravi1.kumar@amd.com>
---
 drivers/net/axgbe/Makefile             |   3 +
 drivers/net/axgbe/axgbe_ethdev.c       |  22 +-
 drivers/net/axgbe/axgbe_rxtx.c         | 433 +++++++++++++++++++++++++++++++++
 drivers/net/axgbe/axgbe_rxtx.h         |  19 ++
 drivers/net/axgbe/axgbe_rxtx_vec_sse.c |  93 +++++++
 5 files changed, 569 insertions(+), 1 deletion(-)
 create mode 100644 drivers/net/axgbe/axgbe_rxtx_vec_sse.c

diff --git a/drivers/net/axgbe/Makefile b/drivers/net/axgbe/Makefile
index e1e5f53..72215ae 100644
--- a/drivers/net/axgbe/Makefile
+++ b/drivers/net/axgbe/Makefile
@@ -28,5 +28,8 @@ SRCS-$(CONFIG_RTE_LIBRTE_AXGBE_PMD) += axgbe_mdio.c
 SRCS-$(CONFIG_RTE_LIBRTE_AXGBE_PMD) += axgbe_phy_impl.c
 SRCS-$(CONFIG_RTE_LIBRTE_AXGBE_PMD) += axgbe_i2c.c
 SRCS-$(CONFIG_RTE_LIBRTE_AXGBE_PMD) += axgbe_rxtx.c
+ifeq ($(CONFIG_RTE_ARCH_X86),y)
+SRCS-$(CONFIG_RTE_LIBRTE_AXGBE_PMD) += axgbe_rxtx_vec_sse.c
+endif
 
 include $(RTE_SDK)/mk/rte.lib.mk
diff --git a/drivers/net/axgbe/axgbe_ethdev.c b/drivers/net/axgbe/axgbe_ethdev.c
index f8cfbd8..a293058 100644
--- a/drivers/net/axgbe/axgbe_ethdev.c
+++ b/drivers/net/axgbe/axgbe_ethdev.c
@@ -102,9 +102,22 @@ axgbe_dev_interrupt_handler(void *param)
 {
 	struct rte_eth_dev *dev = (struct rte_eth_dev *)param;
 	struct axgbe_port *pdata = dev->data->dev_private;
+	unsigned int dma_isr, dma_ch_isr;
 
 	pdata->phy_if.an_isr(pdata);
-
+	/*DMA related interrupts*/
+	dma_isr = AXGMAC_IOREAD(pdata, DMA_ISR);
+	if (dma_isr) {
+		if (dma_isr & 1) {
+			dma_ch_isr =
+				AXGMAC_DMA_IOREAD((struct axgbe_rx_queue *)
+						  pdata->rx_queues[0],
+						  DMA_CH_SR);
+			AXGMAC_DMA_IOWRITE((struct axgbe_rx_queue *)
+					   pdata->rx_queues[0],
+					   DMA_CH_SR, dma_ch_isr);
+		}
+	}
 	/* Enable interrupts since disabled after generation*/
 	rte_intr_enable(&pdata->pci_dev->intr_handle);
 }
@@ -166,6 +179,8 @@ axgbe_dev_start(struct rte_eth_dev *dev)
 
 	/* phy start*/
 	pdata->phy_if.phy_start(pdata);
+	axgbe_dev_enable_tx(dev);
+	axgbe_dev_enable_rx(dev);
 
 	axgbe_clear_bit(AXGBE_STOPPED, &pdata->dev_state);
 	axgbe_clear_bit(AXGBE_DOWN, &pdata->dev_state);
@@ -185,6 +200,8 @@ axgbe_dev_stop(struct rte_eth_dev *dev)
 		return;
 
 	axgbe_set_bit(AXGBE_STOPPED, &pdata->dev_state);
+	axgbe_dev_disable_tx(dev);
+	axgbe_dev_disable_rx(dev);
 
 	pdata->phy_if.phy_stop(pdata);
 	pdata->hw_if.exit(pdata);
@@ -423,6 +440,7 @@ eth_axgbe_dev_init(struct rte_eth_dev *eth_dev)
 	int ret;
 
 	eth_dev->dev_ops = &axgbe_eth_dev_ops;
+	eth_dev->rx_pkt_burst = &axgbe_recv_pkts;
 
 	/*
 	 * For secondary processes, we don't initialise any further as primary
@@ -573,6 +591,8 @@ eth_axgbe_dev_uninit(struct rte_eth_dev *eth_dev)
 	rte_free(eth_dev->data->mac_addrs);
 	eth_dev->data->mac_addrs = NULL;
 	eth_dev->dev_ops = NULL;
+	eth_dev->rx_pkt_burst = NULL;
+	eth_dev->tx_pkt_burst = NULL;
 	axgbe_dev_clear_queues(eth_dev);
 
 	/* disable uio intr before callback unregister */
diff --git a/drivers/net/axgbe/axgbe_rxtx.c b/drivers/net/axgbe/axgbe_rxtx.c
index 1dff7c8..e96e2be 100644
--- a/drivers/net/axgbe/axgbe_rxtx.c
+++ b/drivers/net/axgbe/axgbe_rxtx.c
@@ -113,6 +113,197 @@ int axgbe_dev_rx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	return 0;
 }
 
+static void axgbe_prepare_rx_stop(struct axgbe_port *pdata,
+				  unsigned int queue)
+{
+	unsigned int rx_status;
+	unsigned long rx_timeout;
+
+	/* The Rx engine cannot be stopped if it is actively processing
+	 * packets. Wait for the Rx queue to empty the Rx fifo.  Don't
+	 * wait forever though...
+	 */
+	rx_timeout = rte_get_timer_cycles() + (AXGBE_DMA_STOP_TIMEOUT *
+					       rte_get_timer_hz());
+
+	while (time_before(rte_get_timer_cycles(), rx_timeout)) {
+		rx_status = AXGMAC_MTL_IOREAD(pdata, queue, MTL_Q_RQDR);
+		if ((AXGMAC_GET_BITS(rx_status, MTL_Q_RQDR, PRXQ) == 0) &&
+		    (AXGMAC_GET_BITS(rx_status, MTL_Q_RQDR, RXQSTS) == 0))
+			break;
+
+		rte_delay_us(900);
+	}
+
+	if (!time_before(rte_get_timer_cycles(), rx_timeout))
+		PMD_DRV_LOG(ERR,
+			    "timed out waiting for Rx queue %u to empty\n",
+			    queue);
+}
+
+void axgbe_dev_disable_rx(struct rte_eth_dev *dev)
+{
+	struct axgbe_rx_queue *rxq;
+	struct axgbe_port *pdata = dev->data->dev_private;
+	unsigned int i;
+
+	/* Disable MAC Rx */
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, DCRCC, 0);
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, CST, 0);
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, ACS, 0);
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, RE, 0);
+
+	/* Prepare for Rx DMA channel stop */
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		rxq = dev->data->rx_queues[i];
+		axgbe_prepare_rx_stop(pdata, i);
+	}
+	/* Disable each Rx queue */
+	AXGMAC_IOWRITE(pdata, MAC_RQC0R, 0);
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		rxq = dev->data->rx_queues[i];
+		/* Disable Rx DMA channel */
+		AXGMAC_DMA_IOWRITE_BITS(rxq, DMA_CH_RCR, SR, 0);
+	}
+}
+
+void axgbe_dev_enable_rx(struct rte_eth_dev *dev)
+{
+	struct axgbe_rx_queue *rxq;
+	struct axgbe_port *pdata = dev->data->dev_private;
+	unsigned int i;
+	unsigned int reg_val = 0;
+
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		rxq = dev->data->rx_queues[i];
+		/* Enable Rx DMA channel */
+		AXGMAC_DMA_IOWRITE_BITS(rxq, DMA_CH_RCR, SR, 1);
+	}
+
+	reg_val = 0;
+	for (i = 0; i < pdata->rx_q_count; i++)
+		reg_val |= (0x02 << (i << 1));
+	AXGMAC_IOWRITE(pdata, MAC_RQC0R, reg_val);
+
+	/* Enable MAC Rx */
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, DCRCC, 1);
+	/* Frame is forwarded after stripping CRC to application*/
+	if (pdata->crc_strip_enable) {
+		AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, CST, 1);
+		AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, ACS, 1);
+	}
+	AXGMAC_IOWRITE_BITS(pdata, MAC_RCR, RE, 1);
+}
+
+/* Rx function one to one refresh */
+uint16_t
+axgbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
+		uint16_t nb_pkts)
+{
+	PMD_INIT_FUNC_TRACE();
+	uint16_t nb_rx = 0;
+	struct axgbe_rx_queue *rxq = rx_queue;
+	volatile union axgbe_rx_desc *desc;
+	uint64_t old_dirty = rxq->dirty;
+	struct rte_mbuf *mbuf, *tmbuf;
+	unsigned int err;
+	uint32_t error_status;
+	uint16_t idx, pidx, pkt_len;
+
+	idx = AXGBE_GET_DESC_IDX(rxq, rxq->cur);
+	while (nb_rx < nb_pkts) {
+		if (unlikely(idx == rxq->nb_desc))
+			idx = 0;
+
+		desc = &rxq->desc[idx];
+
+		if (AXGMAC_GET_BITS_LE(desc->write.desc3, RX_NORMAL_DESC3, OWN))
+			break;
+		tmbuf = rte_mbuf_raw_alloc(rxq->mb_pool);
+		if (unlikely(!tmbuf)) {
+			PMD_DRV_LOG(ERR, "RX mbuf alloc failed port_id = %u"
+				    " queue_id = %u\n",
+				    (unsigned int)rxq->port_id,
+				    (unsigned int)rxq->queue_id);
+			rte_eth_devices[
+				rxq->port_id].data->rx_mbuf_alloc_failed++;
+			break;
+		}
+		pidx = idx + 1;
+		if (unlikely(pidx == rxq->nb_desc))
+			pidx = 0;
+
+		rte_prefetch0(rxq->sw_ring[pidx]);
+		if ((pidx & 0x3) == 0) {
+			rte_prefetch0(&rxq->desc[pidx]);
+			rte_prefetch0(&rxq->sw_ring[pidx]);
+		}
+
+		mbuf = rxq->sw_ring[idx];
+		/* Check for any errors and free mbuf*/
+		err = AXGMAC_GET_BITS_LE(desc->write.desc3,
+					 RX_NORMAL_DESC3, ES);
+		error_status = 0;
+		if (unlikely(err)) {
+			error_status = desc->write.desc3 & AXGBE_ERR_STATUS;
+			if ((error_status != AXGBE_L3_CSUM_ERR) &&
+			    (error_status != AXGBE_L4_CSUM_ERR)) {
+				rxq->errors++;
+				rte_pktmbuf_free(mbuf);
+				goto err_set;
+			}
+		}
+		if (rxq->pdata->rx_csum_enable) {
+			mbuf->ol_flags = 0;
+			mbuf->ol_flags |= PKT_RX_IP_CKSUM_GOOD;
+			mbuf->ol_flags |= PKT_RX_L4_CKSUM_GOOD;
+			if (unlikely(error_status == AXGBE_L3_CSUM_ERR)) {
+				mbuf->ol_flags &= ~PKT_RX_IP_CKSUM_GOOD;
+				mbuf->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+				mbuf->ol_flags &= ~PKT_RX_L4_CKSUM_GOOD;
+				mbuf->ol_flags |= PKT_RX_L4_CKSUM_UNKNOWN;
+			} else if (
+				unlikely(error_status == AXGBE_L4_CSUM_ERR)) {
+				mbuf->ol_flags &= ~PKT_RX_L4_CKSUM_GOOD;
+				mbuf->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+			}
+		}
+		rte_prefetch1(rte_pktmbuf_mtod(mbuf, void *));
+		/* Get the RSS hash */
+		if (AXGMAC_GET_BITS_LE(desc->write.desc3, RX_NORMAL_DESC3, RSV))
+			mbuf->hash.rss = rte_le_to_cpu_32(desc->write.desc1);
+		pkt_len = AXGMAC_GET_BITS_LE(desc->write.desc3, RX_NORMAL_DESC3,
+					     PL) - rxq->crc_len;
+		/* Mbuf populate */
+		mbuf->next = NULL;
+		mbuf->data_off = RTE_PKTMBUF_HEADROOM;
+		mbuf->nb_segs = 1;
+		mbuf->port = rxq->port_id;
+		mbuf->pkt_len = pkt_len;
+		mbuf->data_len = pkt_len;
+		rxq->bytes += pkt_len;
+		rx_pkts[nb_rx++] = mbuf;
+err_set:
+		rxq->cur++;
+		rxq->sw_ring[idx++] = tmbuf;
+		desc->read.baddr =
+			rte_cpu_to_le_64(rte_mbuf_data_iova_default(tmbuf));
+		memset((void *)(&desc->read.desc2), 0, 8);
+		AXGMAC_SET_BITS_LE(desc->read.desc3, RX_NORMAL_DESC3, OWN, 1);
+		rxq->dirty++;
+	}
+	rxq->pkts += nb_rx;
+	if (rxq->dirty != old_dirty) {
+		rte_wmb();
+		idx = AXGBE_GET_DESC_IDX(rxq, rxq->dirty - 1);
+		AXGMAC_DMA_IOWRITE(rxq, DMA_CH_RDTR_LO,
+				   low32_value(rxq->ring_phys_addr +
+				   (idx * sizeof(union axgbe_rx_desc))));
+	}
+
+	return nb_rx;
+}
+
 /* Tx Apis */
 static void axgbe_tx_queue_release(struct axgbe_tx_queue *tx_queue)
 {
@@ -174,6 +365,10 @@ int axgbe_dev_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 		txq->free_thresh = (txq->nb_desc >> 1);
 	txq->free_batch_cnt = txq->free_thresh;
 
+	/* In vector_tx path threshold should be multiple of queue_size*/
+	if (txq->nb_desc % txq->free_thresh != 0)
+		txq->vector_disable = 1;
+
 	if ((tx_conf->txq_flags & (uint32_t)ETH_TXQ_FLAGS_NOOFFLOADS) !=
 	    ETH_TXQ_FLAGS_NOOFFLOADS) {
 		txq->vector_disable = 1;
@@ -211,9 +406,247 @@ int axgbe_dev_tx_queue_setup(struct rte_eth_dev *dev, uint16_t queue_idx,
 	if (!pdata->tx_queues)
 		pdata->tx_queues = dev->data->tx_queues;
 
+	if (txq->vector_disable)
+		dev->tx_pkt_burst = &axgbe_xmit_pkts;
+	else
+#ifdef RTE_ARCH_X86
+		dev->tx_pkt_burst = &axgbe_xmit_pkts_vec;
+#else
+		dev->tx_pkt_burst = &axgbe_xmit_pkts;
+#endif
+
 	return 0;
 }
 
+static void axgbe_txq_prepare_tx_stop(struct axgbe_port *pdata,
+				      unsigned int queue)
+{
+	unsigned int tx_status;
+	unsigned long tx_timeout;
+
+	/* The Tx engine cannot be stopped if it is actively processing
+	 * packets. Wait for the Tx queue to empty the Tx fifo.  Don't
+	 * wait forever though...
+	 */
+	tx_timeout = rte_get_timer_cycles() + (AXGBE_DMA_STOP_TIMEOUT *
+					       rte_get_timer_hz());
+	while (time_before(rte_get_timer_cycles(), tx_timeout)) {
+		tx_status = AXGMAC_MTL_IOREAD(pdata, queue, MTL_Q_TQDR);
+		if ((AXGMAC_GET_BITS(tx_status, MTL_Q_TQDR, TRCSTS) != 1) &&
+		    (AXGMAC_GET_BITS(tx_status, MTL_Q_TQDR, TXQSTS) == 0))
+			break;
+
+		rte_delay_us(900);
+	}
+
+	if (!time_before(rte_get_timer_cycles(), tx_timeout))
+		PMD_DRV_LOG(ERR,
+			    "timed out waiting for Tx queue %u to empty\n",
+			    queue);
+}
+
+static void axgbe_prepare_tx_stop(struct axgbe_port *pdata,
+				  unsigned int queue)
+{
+	unsigned int tx_dsr, tx_pos, tx_qidx;
+	unsigned int tx_status;
+	unsigned long tx_timeout;
+
+	if (AXGMAC_GET_BITS(pdata->hw_feat.version, MAC_VR, SNPSVER) > 0x20)
+		return axgbe_txq_prepare_tx_stop(pdata, queue);
+
+	/* Calculate the status register to read and the position within */
+	if (queue < DMA_DSRX_FIRST_QUEUE) {
+		tx_dsr = DMA_DSR0;
+		tx_pos = (queue * DMA_DSR_Q_WIDTH) + DMA_DSR0_TPS_START;
+	} else {
+		tx_qidx = queue - DMA_DSRX_FIRST_QUEUE;
+
+		tx_dsr = DMA_DSR1 + ((tx_qidx / DMA_DSRX_QPR) * DMA_DSRX_INC);
+		tx_pos = ((tx_qidx % DMA_DSRX_QPR) * DMA_DSR_Q_WIDTH) +
+			DMA_DSRX_TPS_START;
+	}
+
+	/* The Tx engine cannot be stopped if it is actively processing
+	 * descriptors. Wait for the Tx engine to enter the stopped or
+	 * suspended state.  Don't wait forever though...
+	 */
+	tx_timeout = rte_get_timer_cycles() + (AXGBE_DMA_STOP_TIMEOUT *
+					       rte_get_timer_hz());
+	while (time_before(rte_get_timer_cycles(), tx_timeout)) {
+		tx_status = AXGMAC_IOREAD(pdata, tx_dsr);
+		tx_status = GET_BITS(tx_status, tx_pos, DMA_DSR_TPS_WIDTH);
+		if ((tx_status == DMA_TPS_STOPPED) ||
+		    (tx_status == DMA_TPS_SUSPENDED))
+			break;
+
+		rte_delay_us(900);
+	}
+
+	if (!time_before(rte_get_timer_cycles(), tx_timeout))
+		PMD_DRV_LOG(ERR,
+			    "timed out waiting for Tx DMA channel %u to stop\n",
+			    queue);
+}
+
+void axgbe_dev_disable_tx(struct rte_eth_dev *dev)
+{
+	struct axgbe_tx_queue *txq;
+	struct axgbe_port *pdata = dev->data->dev_private;
+	unsigned int i;
+
+	/* Prepare for stopping DMA channel */
+	for (i = 0; i < pdata->tx_q_count; i++) {
+		txq = dev->data->tx_queues[i];
+		axgbe_prepare_tx_stop(pdata, i);
+	}
+	/* Disable MAC Tx */
+	AXGMAC_IOWRITE_BITS(pdata, MAC_TCR, TE, 0);
+	/* Disable each Tx queue*/
+	for (i = 0; i < pdata->tx_q_count; i++)
+		AXGMAC_MTL_IOWRITE_BITS(pdata, i, MTL_Q_TQOMR, TXQEN,
+					0);
+	/* Disable each  Tx DMA channel */
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		AXGMAC_DMA_IOWRITE_BITS(txq, DMA_CH_TCR, ST, 0);
+	}
+}
+
+void axgbe_dev_enable_tx(struct rte_eth_dev *dev)
+{
+	struct axgbe_tx_queue *txq;
+	struct axgbe_port *pdata = dev->data->dev_private;
+	unsigned int i;
+
+	for (i = 0; i < dev->data->nb_tx_queues; i++) {
+		txq = dev->data->tx_queues[i];
+		/* Enable Tx DMA channel */
+		AXGMAC_DMA_IOWRITE_BITS(txq, DMA_CH_TCR, ST, 1);
+	}
+	/* Enable Tx queue*/
+	for (i = 0; i < pdata->tx_q_count; i++)
+		AXGMAC_MTL_IOWRITE_BITS(pdata, i, MTL_Q_TQOMR, TXQEN,
+					MTL_Q_ENABLED);
+	/* Enable MAC Tx */
+	AXGMAC_IOWRITE_BITS(pdata, MAC_TCR, TE, 1);
+}
+
+/* Free Tx conformed mbufs */
+static void axgbe_xmit_cleanup(struct axgbe_tx_queue *txq)
+{
+	volatile struct axgbe_tx_desc *desc;
+	uint16_t idx;
+
+	idx = AXGBE_GET_DESC_IDX(txq, txq->dirty);
+	while (txq->cur != txq->dirty) {
+		if (unlikely(idx == txq->nb_desc))
+			idx = 0;
+		desc = &txq->desc[idx];
+		/* Check for ownership */
+		if (AXGMAC_GET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, OWN))
+			return;
+		memset((void *)&desc->desc2, 0, 8);
+		/* Free mbuf */
+		rte_pktmbuf_free(txq->sw_ring[idx]);
+		txq->sw_ring[idx++] = NULL;
+		txq->dirty++;
+	}
+}
+
+/* Tx Descriptor formation
+ * Considering each mbuf requires one desc
+ * mbuf is linear
+ */
+static int axgbe_xmit_hw(struct axgbe_tx_queue *txq,
+			 struct rte_mbuf *mbuf)
+{
+	volatile struct axgbe_tx_desc *desc;
+	uint16_t idx;
+	uint64_t mask;
+
+	idx = AXGBE_GET_DESC_IDX(txq, txq->cur);
+	desc = &txq->desc[idx];
+
+	/* Update buffer address  and length */
+	desc->baddr = rte_mbuf_data_iova(mbuf);
+	AXGMAC_SET_BITS_LE(desc->desc2, TX_NORMAL_DESC2, HL_B1L,
+			   mbuf->pkt_len);
+	/* Total msg length to transmit */
+	AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, FL,
+			   mbuf->pkt_len);
+	/* Mark it as First and Last Descriptor */
+	AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, FD, 1);
+	AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, LD, 1);
+	/* Mark it as a NORMAL descriptor */
+	AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, CTXT, 0);
+	/* configure h/w Offload */
+	mask = mbuf->ol_flags & PKT_TX_L4_MASK;
+	if ((mask == PKT_TX_TCP_CKSUM) || (mask == PKT_TX_UDP_CKSUM))
+		AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, CIC, 0x3);
+	else if (mbuf->ol_flags & PKT_TX_IP_CKSUM)
+		AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, CIC, 0x1);
+	rte_wmb();
+
+	/* Set OWN bit */
+	AXGMAC_SET_BITS_LE(desc->desc3, TX_NORMAL_DESC3, OWN, 1);
+	rte_wmb();
+
+	/* Save mbuf */
+	txq->sw_ring[idx] = mbuf;
+	/* Update current index*/
+	txq->cur++;
+	/* Update stats */
+	txq->bytes += mbuf->pkt_len;
+
+	return 0;
+}
+
+/* Eal supported tx wrapper*/
+uint16_t
+axgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
+		uint16_t nb_pkts)
+{
+	PMD_INIT_FUNC_TRACE();
+
+	if (unlikely(nb_pkts == 0))
+		return nb_pkts;
+
+	struct axgbe_tx_queue *txq;
+	uint16_t nb_desc_free;
+	uint16_t nb_pkt_sent = 0;
+	uint16_t idx;
+	uint32_t tail_addr;
+	struct rte_mbuf *mbuf;
+
+	txq  = (struct axgbe_tx_queue *)tx_queue;
+	nb_desc_free = txq->nb_desc - (txq->cur - txq->dirty);
+
+	if (unlikely(nb_desc_free <= txq->free_thresh)) {
+		axgbe_xmit_cleanup(txq);
+		nb_desc_free = txq->nb_desc - (txq->cur - txq->dirty);
+		if (unlikely(nb_desc_free == 0))
+			return 0;
+	}
+	nb_pkts = RTE_MIN(nb_desc_free, nb_pkts);
+	while (nb_pkts--) {
+		mbuf = *tx_pkts++;
+		if (axgbe_xmit_hw(txq, mbuf))
+			goto out;
+		nb_pkt_sent++;
+	}
+out:
+	/* Sync read and write */
+	rte_mb();
+	idx = AXGBE_GET_DESC_IDX(txq, txq->cur);
+	tail_addr = low32_value(txq->ring_phys_addr +
+				idx * sizeof(struct axgbe_tx_desc));
+	/* Update tail reg with next immediate address to kick Tx DMA channel*/
+	AXGMAC_DMA_IOWRITE(txq, DMA_CH_TDTR_LO, tail_addr);
+	txq->pkts += nb_pkt_sent;
+	return nb_pkt_sent;
+}
+
 void axgbe_dev_clear_queues(struct rte_eth_dev *dev)
 {
 	PMD_INIT_FUNC_TRACE();
diff --git a/drivers/net/axgbe/axgbe_rxtx.h b/drivers/net/axgbe/axgbe_rxtx.h
index 1b88d7a..f221cc3 100644
--- a/drivers/net/axgbe/axgbe_rxtx.h
+++ b/drivers/net/axgbe/axgbe_rxtx.h
@@ -156,12 +156,31 @@ void axgbe_dev_tx_queue_release(void *txq);
 int  axgbe_dev_tx_queue_setup(struct rte_eth_dev *dev, uint16_t tx_queue_id,
 			      uint16_t nb_tx_desc, unsigned int socket_id,
 			      const struct rte_eth_txconf *tx_conf);
+void axgbe_dev_enable_tx(struct rte_eth_dev *dev);
+void axgbe_dev_disable_tx(struct rte_eth_dev *dev);
+int axgbe_dev_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
+int axgbe_dev_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
+
+uint16_t axgbe_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts);
+uint16_t axgbe_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+			 uint16_t nb_pkts);
+
 
 void axgbe_dev_rx_queue_release(void *rxq);
 int  axgbe_dev_rx_queue_setup(struct rte_eth_dev *dev, uint16_t rx_queue_id,
 			      uint16_t nb_rx_desc, unsigned int socket_id,
 			      const struct rte_eth_rxconf *rx_conf,
 			      struct rte_mempool *mb_pool);
+void axgbe_dev_enable_rx(struct rte_eth_dev *dev);
+void axgbe_dev_disable_rx(struct rte_eth_dev *dev);
+int axgbe_dev_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id);
+int axgbe_dev_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id);
+uint16_t axgbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
+			 uint16_t nb_pkts);
+uint16_t axgbe_recv_pkts_threshold_refresh(void *rx_queue,
+					   struct rte_mbuf **rx_pkts,
+					   uint16_t nb_pkts);
 void axgbe_dev_clear_queues(struct rte_eth_dev *dev);
 
 #endif /* _AXGBE_RXTX_H_ */
diff --git a/drivers/net/axgbe/axgbe_rxtx_vec_sse.c b/drivers/net/axgbe/axgbe_rxtx_vec_sse.c
new file mode 100644
index 0000000..9be7037
--- /dev/null
+++ b/drivers/net/axgbe/axgbe_rxtx_vec_sse.c
@@ -0,0 +1,93 @@
+/*   SPDX-License-Identifier: BSD-3-Clause
+ *   Copyright(c) 2018 Advanced Micro Devices, Inc. All rights reserved.
+ *   Copyright(c) 2018 Synopsys, Inc. All rights reserved.
+ */
+
+#include "axgbe_ethdev.h"
+#include "axgbe_rxtx.h"
+#include "axgbe_phy.h"
+
+#include <rte_time.h>
+#include <rte_mempool.h>
+#include <rte_mbuf.h>
+
+/* Useful to avoid shifting for every descriptor prepration*/
+#define TX_DESC_CTRL_FLAGS 0xb000000000000000
+#define TX_FREE_BULK	   8
+#define TX_FREE_BULK_CHECK (TX_FREE_BULK - 1)
+
+static inline void
+axgbe_vec_tx(volatile struct axgbe_tx_desc *desc,
+	     struct rte_mbuf *mbuf)
+{
+	__m128i descriptor = _mm_set_epi64x((uint64_t)mbuf->pkt_len << 32 |
+					    TX_DESC_CTRL_FLAGS | mbuf->data_len,
+					    mbuf->buf_iova
+					    + mbuf->data_off);
+	_mm_store_si128((__m128i *)desc, descriptor);
+}
+
+static void
+axgbe_xmit_cleanup_vec(struct axgbe_tx_queue *txq)
+{
+	volatile struct axgbe_tx_desc *desc;
+	int idx, i;
+
+	idx = AXGBE_GET_DESC_IDX(txq, txq->dirty + txq->free_batch_cnt
+				 - 1);
+	desc = &txq->desc[idx];
+	if (desc->desc3 & AXGBE_DESC_OWN)
+		return;
+	/* memset avoided for desc ctrl fields since in vec_tx path
+	 * all 128 bits are populated
+	 */
+	for (i = 0; i < txq->free_batch_cnt; i++, idx--)
+		rte_pktmbuf_free_seg(txq->sw_ring[idx]);
+
+
+	txq->dirty += txq->free_batch_cnt;
+	txq->nb_desc_free += txq->free_batch_cnt;
+}
+
+uint16_t
+axgbe_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+		    uint16_t nb_pkts)
+{
+	PMD_INIT_FUNC_TRACE();
+
+	struct axgbe_tx_queue *txq;
+	uint16_t idx, nb_commit, loop, i;
+	uint32_t tail_addr;
+
+	txq  = (struct axgbe_tx_queue *)tx_queue;
+	if (txq->nb_desc_free < txq->free_thresh) {
+		axgbe_xmit_cleanup_vec(txq);
+		if (unlikely(txq->nb_desc_free == 0))
+			return 0;
+	}
+	nb_pkts = RTE_MIN(txq->nb_desc_free, nb_pkts);
+	nb_commit = nb_pkts;
+	idx = AXGBE_GET_DESC_IDX(txq, txq->cur);
+	loop = txq->nb_desc - idx;
+	if (nb_commit >= loop) {
+		for (i = 0; i < loop; ++i, ++idx, ++tx_pkts) {
+			axgbe_vec_tx(&txq->desc[idx], *tx_pkts);
+			txq->sw_ring[idx] = *tx_pkts;
+		}
+		nb_commit -= loop;
+		idx = 0;
+	}
+	for (i = 0; i < nb_commit; ++i, ++idx, ++tx_pkts) {
+		axgbe_vec_tx(&txq->desc[idx], *tx_pkts);
+		txq->sw_ring[idx] = *tx_pkts;
+	}
+	txq->cur += nb_pkts;
+	tail_addr = (uint32_t)(txq->ring_phys_addr +
+			       idx * sizeof(struct axgbe_tx_desc));
+	/* Update tail reg with next immediate address to kick Tx DMA channel*/
+	rte_write32(tail_addr, (void *)txq->dma_tail_reg);
+	txq->pkts += nb_pkts;
+	txq->nb_desc_free -= nb_pkts;
+
+	return nb_pkts;
+}
-- 
2.7.4