[dpdk-dev] [PATCH v5 2/9] nfp: rx/tx functionality

["Alejandro.Lucero" <alejandro.lucero@netronome.com>]

From: "Alejandro.Lucero" <alejandro.lucero@netronome.com>

Signed-off-by: Alejandro.Lucero <alejandro.lucero@netronome.com>
Signed-off-by: Rolf.Neugebauer <rolf.neugebauer@netronome.com>
---
 drivers/net/nfp/nfp_net.c |  921 +++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 921 insertions(+)

diff --git a/drivers/net/nfp/nfp_net.c b/drivers/net/nfp/nfp_net.c
index a33ed37..8339c2d 100644
--- a/drivers/net/nfp/nfp_net.c
+++ b/drivers/net/nfp/nfp_net.c
@@ -212,6 +212,100 @@ nn_cfg_writeq(struct nfp_net_hw *hw, int off, uint64_t val)
 	nn_writeq(rte_cpu_to_le_64(val), hw->ctrl_bar + off);
 }
 
+/* Creating memzone for hardware rings. */
+static const struct rte_memzone *
+ring_dma_zone_reserve(struct rte_eth_dev *dev, const char *ring_name,
+		      uint16_t queue_id, uint32_t ring_size, int socket_id)
+{
+	char z_name[RTE_MEMZONE_NAMESIZE];
+	const struct rte_memzone *mz;
+
+	snprintf(z_name, sizeof(z_name), "%s_%s_%d_%d",
+		 dev->driver->pci_drv.name,
+		 ring_name, dev->data->port_id, queue_id);
+
+	mz = rte_memzone_lookup(z_name);
+	if (mz)
+		return mz;
+
+	return rte_memzone_reserve_aligned(z_name, ring_size, socket_id, 0,
+					   NFP_MEMZONE_ALIGN);
+}
+
+static void
+nfp_net_rx_queue_release_mbufs(struct nfp_net_rxq *rxq)
+{
+	unsigned i;
+
+	if (rxq->rxbufs == NULL)
+		return;
+
+	for (i = 0; i < rxq->rx_count; i++) {
+		if (rxq->rxbufs[i].mbuf) {
+			rte_pktmbuf_free_seg(rxq->rxbufs[i].mbuf);
+			rxq->rxbufs[i].mbuf = NULL;
+		}
+	}
+}
+
+static void
+nfp_net_rx_queue_release(void *rx_queue)
+{
+	struct nfp_net_rxq *rxq = rx_queue;
+
+	if (rxq) {
+		nfp_net_rx_queue_release_mbufs(rxq);
+		rte_free(rxq->rxbufs);
+		rte_free(rxq);
+	}
+}
+
+static void
+nfp_net_reset_rx_queue(struct nfp_net_rxq *rxq)
+{
+	nfp_net_rx_queue_release_mbufs(rxq);
+	rxq->wr_p = 0;
+	rxq->rd_p = 0;
+	rxq->nb_rx_hold = 0;
+}
+
+static void
+nfp_net_tx_queue_release_mbufs(struct nfp_net_txq *txq)
+{
+	unsigned i;
+
+	if (txq->txbufs == NULL)
+		return;
+
+	for (i = 0; i < txq->tx_count; i++) {
+		if (txq->txbufs[i].mbuf) {
+			rte_pktmbuf_free_seg(txq->txbufs[i].mbuf);
+			txq->txbufs[i].mbuf = NULL;
+		}
+	}
+}
+
+static void
+nfp_net_tx_queue_release(void *tx_queue)
+{
+	struct nfp_net_txq *txq = tx_queue;
+
+	if (txq) {
+		nfp_net_tx_queue_release_mbufs(txq);
+		rte_free(txq->txbufs);
+		rte_free(txq);
+	}
+}
+
+static void
+nfp_net_reset_tx_queue(struct nfp_net_txq *txq)
+{
+	nfp_net_tx_queue_release_mbufs(txq);
+	txq->wr_p = 0;
+	txq->rd_p = 0;
+	txq->tail = 0;
+}
+
 static int
 __nfp_net_reconfig(struct nfp_net_hw *hw, uint32_t update)
 {
@@ -449,6 +543,18 @@ nfp_net_disable_queues(struct rte_eth_dev *dev)
 	hw->ctrl = new_ctrl;
 }
 
+static int
+nfp_net_rx_freelist_setup(struct rte_eth_dev *dev)
+{
+	int i;
+
+	for (i = 0; i < dev->data->nb_rx_queues; i++) {
+		if (nfp_net_rx_fill_freelist(dev->data->rx_queues[i]) < 0)
+			return -1;
+	}
+	return 0;
+}
+
 static void
 nfp_net_params_setup(struct nfp_net_hw *hw)
 {
@@ -582,6 +688,821 @@ nfp_net_close(struct rte_eth_dev *dev)
 	 */
 }
 
+static uint32_t
+nfp_net_rx_queue_count(struct rte_eth_dev *dev, uint16_t queue_idx)
+{
+	struct nfp_net_rxq *rxq;
+	struct nfp_net_rx_desc *rxds;
+	uint32_t idx;
+	uint32_t count;
+
+	rxq = (struct nfp_net_rxq *)dev->data->rx_queues[queue_idx];
+
+	if (rxq == NULL) {
+		PMD_INIT_LOG(ERR, "Bad queue: %u\n", queue_idx);
+		return 0;
+	}
+
+	idx = rxq->rd_p % rxq->rx_count;
+	rxds = &rxq->rxds[idx];
+
+	count = 0;
+
+	/*
+	 * Other PMDs are just checking the DD bit in intervals of 4
+	 * descriptors and counting all four if the first has the DD
+	 * bit on. Of course, this is not accurate but can be good for
+	 * perfomance. But ideally that should be done in descriptors
+	 * chunks belonging to the same cache line
+	 */
+
+	while (count < rxq->rx_count) {
+		rxds = &rxq->rxds[idx];
+		if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
+			break;
+
+		count++;
+		idx++;
+
+		/* Wrapping? */
+		if ((idx) == rxq->rx_count)
+			idx = 0;
+	}
+
+	return count;
+}
+
+static int
+nfp_net_rx_queue_setup(struct rte_eth_dev *dev,
+		       uint16_t queue_idx, uint16_t nb_desc,
+		       unsigned int socket_id,
+		       const struct rte_eth_rxconf *rx_conf,
+		       struct rte_mempool *mp)
+{
+	const struct rte_memzone *tz;
+	struct nfp_net_rxq *rxq;
+	struct nfp_net_hw *hw;
+
+	hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+
+	PMD_INIT_FUNC_TRACE();
+
+	/* Validating number of descriptors */
+	if (((nb_desc * sizeof(struct nfp_net_rx_desc)) % 128) != 0 ||
+	    (nb_desc > NFP_NET_MAX_RX_DESC) ||
+	    (nb_desc < NFP_NET_MIN_RX_DESC)) {
+		RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
+		return (-EINVAL);
+	}
+
+	/*
+	 * Free memory prior to re-allocation if needed. This is the case after
+	 * calling nfp_net_stop
+	 */
+	if (dev->data->rx_queues[queue_idx]) {
+		nfp_net_rx_queue_release(dev->data->rx_queues[queue_idx]);
+		dev->data->rx_queues[queue_idx] = NULL;
+	}
+
+	/* Allocating rx queue data structure */
+	rxq = rte_zmalloc_socket("ethdev RX queue", sizeof(struct nfp_net_rxq),
+				 RTE_CACHE_LINE_SIZE, socket_id);
+	if (rxq == NULL)
+		return (-ENOMEM);
+
+	/* Hw queues mapping based on firmware confifguration */
+	rxq->qidx = queue_idx;
+	rxq->fl_qcidx = queue_idx * hw->stride_rx;
+	rxq->rx_qcidx = rxq->fl_qcidx + (hw->stride_rx - 1);
+	rxq->qcp_fl = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->fl_qcidx);
+	rxq->qcp_rx = hw->rx_bar + NFP_QCP_QUEUE_OFF(rxq->rx_qcidx);
+
+	/*
+	 * Tracking mbuf size for detecting a potential mbuf overflow due to
+	 * NFP_NET_RX_OFFSET
+	 */
+	rxq->mem_pool = mp;
+	rxq->mbuf_size = rxq->mem_pool->elt_size;
+	rxq->mbuf_size -= (sizeof(struct rte_mbuf) + RTE_PKTMBUF_HEADROOM);
+	hw->flbufsz = rxq->mbuf_size;
+
+	rxq->rx_count = nb_desc;
+	rxq->port_id = dev->data->port_id;
+	rxq->rx_free_thresh = rx_conf->rx_free_thresh;
+	rxq->drop_en = rx_conf->rx_drop_en;
+
+	/*
+	 * Allocate RX ring hardware descriptors. A memzone large enough to
+	 * handle the maximum ring size is allocated in order to allow for
+	 * resizing in later calls to the queue setup function.
+	 */
+	tz = ring_dma_zone_reserve(dev, "rx_ring", queue_idx,
+				   sizeof(struct nfp_net_rx_desc) *
+				   NFP_NET_MAX_RX_DESC, socket_id);
+
+	if (tz == NULL) {
+		RTE_LOG(ERR, PMD, "Error allocatig rx dma\n");
+		nfp_net_rx_queue_release(rxq);
+		return (-ENOMEM);
+	}
+
+	/* Saving physical and virtual addresses for the RX ring */
+	rxq->dma = (uint64_t)tz->phys_addr;
+	rxq->rxds = (struct nfp_net_rx_desc *)tz->addr;
+
+	/* mbuf pointers array for referencing mbufs linked to RX descriptors */
+	rxq->rxbufs = rte_zmalloc_socket("rxq->rxbufs",
+					 sizeof(*rxq->rxbufs) * nb_desc,
+					 RTE_CACHE_LINE_SIZE, socket_id);
+	if (rxq->rxbufs == NULL) {
+		nfp_net_rx_queue_release(rxq);
+		return (-ENOMEM);
+	}
+
+	PMD_RX_LOG(DEBUG, "rxbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
+		   rxq->rxbufs, rxq->rxds, (unsigned long int)rxq->dma);
+
+	nfp_net_reset_rx_queue(rxq);
+
+	dev->data->rx_queues[queue_idx] = rxq;
+	rxq->hw = hw;
+
+	/*
+	 * Telling the HW about the physical address of the RX ring and number
+	 * of descriptors in log2 format
+	 */
+	nn_cfg_writeq(hw, NFP_NET_CFG_RXR_ADDR(queue_idx), rxq->dma);
+	nn_cfg_writeb(hw, NFP_NET_CFG_RXR_SZ(queue_idx), log2(nb_desc));
+
+	return 0;
+}
+
+static int
+nfp_net_rx_fill_freelist(struct nfp_net_rxq *rxq)
+{
+	struct nfp_net_rx_buff *rxe = rxq->rxbufs;
+	uint64_t dma_addr;
+	unsigned i;
+
+	PMD_RX_LOG(DEBUG, "nfp_net_rx_fill_freelist for %u descriptors\n",
+		   rxq->rx_count);
+
+	for (i = 0; i < rxq->rx_count; i++) {
+		struct nfp_net_rx_desc *rxd;
+		struct rte_mbuf *mbuf = rte_pktmbuf_alloc(rxq->mem_pool);
+
+		if (mbuf == NULL) {
+			RTE_LOG(ERR, PMD, "RX mbuf alloc failed queue_id=%u\n",
+				(unsigned)rxq->qidx);
+			return (-ENOMEM);
+		}
+
+		dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(mbuf));
+
+		rxd = &rxq->rxds[i];
+		rxd->fld.dd = 0;
+		rxd->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
+		rxd->fld.dma_addr_lo = dma_addr & 0xffffffff;
+		rxe[i].mbuf = mbuf;
+		PMD_RX_LOG(DEBUG, "[%d]: %" PRIx64 "\n", i, dma_addr);
+
+		rxq->wr_p++;
+	}
+
+	/* Make sure all writes are flushed before telling the hardware */
+	rte_wmb();
+
+	/* Not advertising the whole ring as the firmware gets confused if so */
+	PMD_RX_LOG(DEBUG, "Increment FL write pointer in %u\n",
+		   rxq->rx_count - 1);
+
+	nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, rxq->rx_count - 1);
+
+	return 0;
+}
+
+static int
+nfp_net_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)
+{
+	const struct rte_memzone *tz;
+	struct nfp_net_txq *txq;
+	uint16_t tx_free_thresh;
+	struct nfp_net_hw *hw;
+
+	hw = NFP_NET_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+
+	PMD_INIT_FUNC_TRACE();
+
+	/* Validating number of descriptors */
+	if (((nb_desc * sizeof(struct nfp_net_tx_desc)) % 128) != 0 ||
+	    (nb_desc > NFP_NET_MAX_TX_DESC) ||
+	    (nb_desc < NFP_NET_MIN_TX_DESC)) {
+		RTE_LOG(ERR, PMD, "Wrong nb_desc value\n");
+		return -EINVAL;
+	}
+
+	tx_free_thresh = (uint16_t)((tx_conf->tx_free_thresh) ?
+				    tx_conf->tx_free_thresh :
+				    DEFAULT_TX_FREE_THRESH);
+
+	if (tx_free_thresh > (nb_desc)) {
+		RTE_LOG(ERR, PMD,
+			"tx_free_thresh must be less than the number of TX "
+			"descriptors. (tx_free_thresh=%u port=%d "
+			"queue=%d)\n", (unsigned int)tx_free_thresh,
+			(int)dev->data->port_id, (int)queue_idx);
+		return -(EINVAL);
+	}
+
+	/*
+	 * Free memory prior to re-allocation if needed. This is the case after
+	 * calling nfp_net_stop
+	 */
+	if (dev->data->tx_queues[queue_idx]) {
+		PMD_TX_LOG(DEBUG, "Freeing memory prior to re-allocation %d\n",
+			   queue_idx);
+		nfp_net_tx_queue_release(dev->data->tx_queues[queue_idx]);
+		dev->data->tx_queues[queue_idx] = NULL;
+	}
+
+	/* Allocating tx queue data structure */
+	txq = rte_zmalloc_socket("ethdev TX queue", sizeof(struct nfp_net_txq),
+				 RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq == NULL) {
+		RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
+		return (-ENOMEM);
+	}
+
+	/*
+	 * Allocate TX ring hardware descriptors. A memzone large enough to
+	 * handle the maximum ring size is allocated in order to allow for
+	 * resizing in later calls to the queue setup function.
+	 */
+	tz = ring_dma_zone_reserve(dev, "tx_ring", queue_idx,
+				   sizeof(struct nfp_net_tx_desc) *
+				   NFP_NET_MAX_TX_DESC, socket_id);
+	if (tz == NULL) {
+		RTE_LOG(ERR, PMD, "Error allocating tx dma\n");
+		nfp_net_tx_queue_release(txq);
+		return (-ENOMEM);
+	}
+
+	txq->tx_count = nb_desc;
+	txq->tx_free_thresh = tx_free_thresh;
+	txq->tx_pthresh = tx_conf->tx_thresh.pthresh;
+	txq->tx_hthresh = tx_conf->tx_thresh.hthresh;
+	txq->tx_wthresh = tx_conf->tx_thresh.wthresh;
+
+	/* queue mapping based on firmware configuration */
+	txq->qidx = queue_idx;
+	txq->tx_qcidx = queue_idx * hw->stride_tx;
+	txq->qcp_q = hw->tx_bar + NFP_QCP_QUEUE_OFF(txq->tx_qcidx);
+
+	txq->port_id = dev->data->port_id;
+	txq->txq_flags = tx_conf->txq_flags;
+
+	/* Saving physical and virtual addresses for the TX ring */
+	txq->dma = (uint64_t)tz->phys_addr;
+	txq->txds = (struct nfp_net_tx_desc *)tz->addr;
+
+	/* mbuf pointers array for referencing mbufs linked to TX descriptors */
+	txq->txbufs = rte_zmalloc_socket("txq->txbufs",
+					 sizeof(*txq->txbufs) * nb_desc,
+					 RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq->txbufs == NULL) {
+		nfp_net_tx_queue_release(txq);
+		return (-ENOMEM);
+	}
+	PMD_TX_LOG(DEBUG, "txbufs=%p hw_ring=%p dma_addr=0x%" PRIx64 "\n",
+		   txq->txbufs, txq->txds, (unsigned long int)txq->dma);
+
+	nfp_net_reset_tx_queue(txq);
+
+	dev->data->tx_queues[queue_idx] = txq;
+	txq->hw = hw;
+
+	/*
+	 * Telling the HW about the physical address of the TX ring and number
+	 * of descriptors in log2 format
+	 */
+	nn_cfg_writeq(hw, NFP_NET_CFG_TXR_ADDR(queue_idx), txq->dma);
+	nn_cfg_writeb(hw, NFP_NET_CFG_TXR_SZ(queue_idx), log2(nb_desc));
+
+	return 0;
+}
+
+/* nfp_net_tx_cksum - Set TX CSUM offload flags in TX descriptor */
+static inline void
+nfp_net_tx_cksum(struct nfp_net_txq *txq, struct nfp_net_tx_desc *txd,
+		 struct rte_mbuf *mb)
+{
+	uint16_t ol_flags;
+	struct nfp_net_hw *hw = txq->hw;
+
+	if (!(hw->cap & NFP_NET_CFG_CTRL_TXCSUM))
+		return;
+
+	ol_flags = mb->ol_flags;
+
+	/* IPv6 does not need checksum */
+	if (ol_flags & PKT_TX_IP_CKSUM)
+		txd->flags |= PCIE_DESC_TX_IP4_CSUM;
+
+	switch (ol_flags & PKT_TX_L4_MASK) {
+	case PKT_TX_UDP_CKSUM:
+		txd->flags |= PCIE_DESC_TX_UDP_CSUM;
+		break;
+	case PKT_TX_TCP_CKSUM:
+		txd->flags |= PCIE_DESC_TX_TCP_CSUM;
+		break;
+	}
+
+	txd->flags |= PCIE_DESC_TX_CSUM;
+}
+
+/* nfp_net_rx_cksum - set mbuf checksum flags based on RX descriptor flags */
+static inline void
+nfp_net_rx_cksum(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
+		 struct rte_mbuf *mb)
+{
+	struct nfp_net_hw *hw = rxq->hw;
+
+	if (!(hw->ctrl & NFP_NET_CFG_CTRL_RXCSUM))
+		return;
+
+	/* If IPv4 and IP checksum error, fail */
+	if ((rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM) &&
+	    !(rxd->rxd.flags & PCIE_DESC_RX_IP4_CSUM_OK))
+		mb->ol_flags |= PKT_RX_IP_CKSUM_BAD;
+
+	/* If neither UDP nor TCP return */
+	if (!(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
+	    !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM))
+		return;
+
+	if ((rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM) &&
+	    !(rxd->rxd.flags & PCIE_DESC_RX_TCP_CSUM_OK))
+		mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+
+	if ((rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM) &&
+	    !(rxd->rxd.flags & PCIE_DESC_RX_UDP_CSUM_OK))
+		mb->ol_flags |= PKT_RX_L4_CKSUM_BAD;
+}
+
+#define NFP_HASH_OFFSET      ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 4)
+#define NFP_HASH_TYPE_OFFSET ((uint8_t *)mbuf->buf_addr + mbuf->data_off - 8)
+
+/*
+ * nfp_net_set_hash - Set mbuf hash data
+ *
+ * The RSS hash and hash-type are pre-pended to the packet data.
+ * Extract and decode it and set the mbuf fields.
+ */
+static inline void
+nfp_net_set_hash(struct nfp_net_rxq *rxq, struct nfp_net_rx_desc *rxd,
+		 struct rte_mbuf *mbuf)
+{
+	uint32_t hash;
+	uint32_t hash_type;
+	struct nfp_net_hw *hw = rxq->hw;
+
+	if (!(hw->ctrl & NFP_NET_CFG_CTRL_RSS))
+		return;
+
+	if (!(rxd->rxd.flags & PCIE_DESC_RX_RSS))
+		return;
+
+	hash = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_OFFSET);
+	hash_type = rte_be_to_cpu_32(*(uint32_t *)NFP_HASH_TYPE_OFFSET);
+
+	/*
+	 * hash type is sharing the same word with input port info
+	 * 31-8: input port
+	 * 7:0: hash type
+	 */
+	hash_type &= 0xff;
+	mbuf->hash.rss = hash;
+	mbuf->ol_flags |= PKT_RX_RSS_HASH;
+
+	switch (hash_type) {
+	case NFP_NET_RSS_IPV4:
+		mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV4;
+		break;
+	case NFP_NET_RSS_IPV6:
+		mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6;
+		break;
+	case NFP_NET_RSS_IPV6_EX:
+		mbuf->packet_type |= RTE_PTYPE_INNER_L3_IPV6_EXT;
+		break;
+	default:
+		mbuf->packet_type |= RTE_PTYPE_INNER_L4_MASK;
+	}
+}
+
+/* nfp_net_check_port - Set mbuf in_port field */
+static void
+nfp_net_check_port(struct nfp_net_rx_desc *rxd, struct rte_mbuf *mbuf)
+{
+	uint32_t port;
+
+	if (!(rxd->rxd.flags & PCIE_DESC_RX_INGRESS_PORT)) {
+		mbuf->port = 0;
+		return;
+	}
+
+	port = rte_be_to_cpu_32(*(uint32_t *)((uint8_t *)mbuf->buf_addr +
+					      mbuf->data_off - 8));
+
+	/*
+	 * hash type is sharing the same word with input port info
+	 * 31-8: input port
+	 * 7:0: hash type
+	 */
+	port = (uint8_t)(port >> 8);
+	mbuf->port = port;
+}
+
+static inline void
+nfp_net_mbuf_alloc_failed(struct nfp_net_rxq *rxq)
+{
+	rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
+}
+
+#define NFP_DESC_META_LEN(d) (d->rxd.meta_len_dd & PCIE_DESC_RX_META_LEN_MASK)
+
+/*
+ * RX path design:
+ *
+ * There are some decissions to take:
+ * 1) How to check DD RX descriptors bit
+ * 2) How and when to allocate new mbufs
+ *
+ * Current implementation checks just one single DD bit each loop. As each
+ * descriptor is 8 bytes, it is likely a good idea to check descriptors in
+ * a single cache line instead. Tests with this change have not shown any
+ * performance improvement but it requires further investigation. For example,
+ * depending on which descriptor is next, the number of descriptors could be
+ * less than 8 for just checking those in the same cache line. This implies
+ * extra work which could be counterproductive by itself. Indeed, last firmware
+ * changes are just doing this: writing several descriptors with the DD bit
+ * for saving PCIe bandwidth and DMA operations from the NFP.
+ *
+ * Mbuf allocation is done when a new packet is received. Then the descriptor
+ * is automatically linked with the new mbuf and the old one is given to the
+ * user. The main drawback with this design is mbuf allocation is heavier than
+ * using bulk allocations allowed by DPDK with rte_mempool_get_bulk. From the
+ * cache point of view it does not seem allocating the mbuf early on as we are
+ * doing now have any benefit at all. Again, tests with this change have not
+ * shown any improvement. Also, rte_mempool_get_bulk returns all or nothing
+ * so looking at the implications of this type of allocation should be studied
+ * deeply
+ */
+
+static uint16_t
+nfp_net_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts)
+{
+	struct nfp_net_rxq *rxq;
+	struct nfp_net_rx_desc *rxds;
+	struct nfp_net_rx_buff *rxb;
+	struct nfp_net_hw *hw;
+	struct rte_mbuf *mb;
+	struct rte_mbuf *new_mb;
+	int idx;
+	uint16_t nb_hold;
+	uint64_t dma_addr;
+	int avail;
+
+	rxq = rx_queue;
+	if (unlikely(rxq == NULL)) {
+		/*
+		 * DPDK just checks the queue is lower than max queues
+		 * enabled. But the queue needs to be configured
+		 */
+		RTE_LOG(ERR, PMD, "RX Bad queue\n");
+		return -EINVAL;
+	}
+
+	hw = rxq->hw;
+	avail = 0;
+	nb_hold = 0;
+
+	while (avail < nb_pkts) {
+		idx = rxq->rd_p % rxq->rx_count;
+
+		rxb = &rxq->rxbufs[idx];
+		if (unlikely(rxb == NULL)) {
+			RTE_LOG(ERR, PMD, "rxb does not exist!\n");
+			break;
+		}
+
+		/*
+		 * Memory barrier to ensure that we won't do other
+		 * reads before the DD bit.
+		 */
+		rte_rmb();
+
+		rxds = &rxq->rxds[idx];
+		if ((rxds->rxd.meta_len_dd & PCIE_DESC_RX_DD) == 0)
+			break;
+
+		/*
+		 * We got a packet. Let's alloc a new mbuff for refilling the
+		 * free descriptor ring as soon as possible
+		 */
+		new_mb = rte_pktmbuf_alloc(rxq->mem_pool);
+		if (unlikely(new_mb == NULL)) {
+			RTE_LOG(DEBUG, PMD, "RX mbuf alloc failed port_id=%u "
+				"queue_id=%u\n", (unsigned)rxq->port_id,
+				(unsigned)rxq->qidx);
+			nfp_net_mbuf_alloc_failed(rxq);
+			break;
+		}
+
+		nb_hold++;
+
+		/*
+		 * Grab the mbuff and refill the descriptor with the
+		 * previously allocated mbuff
+		 */
+		mb = rxb->mbuf;
+		rxb->mbuf = new_mb;
+
+		PMD_RX_LOG(DEBUG, "Packet len: %u, mbuf_size: %u\n",
+			   rxds->rxd.data_len, rxq->mbuf_size);
+
+		/* Size of this segment */
+		mb->data_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
+		/* Size of the whole packet. We just support 1 segment */
+		mb->pkt_len = rxds->rxd.data_len - NFP_DESC_META_LEN(rxds);
+
+		if (unlikely((mb->data_len + NFP_NET_RX_OFFSET) >
+			     rxq->mbuf_size)) {
+			/*
+			 * This should not happen and the user has the
+			 * responsibility of avoiding it. But we have
+			 * to give some info about the error
+			 */
+			RTE_LOG(ERR, PMD,
+				"mbuf overflow likely due to NFP_NET_RX_OFFSET\n"
+				"\t\tYour mbuf size should have extra space for"
+				" NFP_NET_RX_OFFSET=%u bytes.\n"
+				"\t\tCurrently you just have %u bytes available"
+				" but the received packet is %u bytes long",
+				NFP_NET_RX_OFFSET,
+				rxq->mbuf_size - NFP_NET_RX_OFFSET,
+				mb->data_len);
+			return -EINVAL;
+		}
+
+		/* Filling the received mbuff with packet info */
+		mb->data_off = RTE_PKTMBUF_HEADROOM + NFP_NET_RX_OFFSET;
+
+		/* No scatter mode supported */
+		mb->nb_segs = 1;
+		mb->next = NULL;
+
+		/* Checking the RSS flag */
+		nfp_net_set_hash(rxq, rxds, mb);
+
+		/* Checking the checksum flag */
+		nfp_net_rx_cksum(rxq, rxds, mb);
+
+		/* Checking the port flag */
+		nfp_net_check_port(rxds, mb);
+
+		if ((rxds->rxd.flags & PCIE_DESC_RX_VLAN) &&
+		    (hw->ctrl & NFP_NET_CFG_CTRL_RXVLAN)) {
+			mb->vlan_tci = rte_cpu_to_le_32(rxds->rxd.vlan);
+			mb->ol_flags |= PKT_RX_VLAN_PKT;
+		}
+
+		/* Adding the mbuff to the mbuff array passed by the app */
+		rx_pkts[avail++] = mb;
+
+		/* Now resetting and updating the descriptor */
+		rxds->vals[0] = 0;
+		rxds->vals[1] = 0;
+		dma_addr = rte_cpu_to_le_64(RTE_MBUF_DMA_ADDR_DEFAULT(new_mb));
+		rxds->fld.dd = 0;
+		rxds->fld.dma_addr_hi = (dma_addr >> 32) & 0xff;
+		rxds->fld.dma_addr_lo = dma_addr & 0xffffffff;
+
+		rxq->rd_p++;
+	}
+
+	if (nb_hold == 0)
+		return nb_hold;
+
+	PMD_RX_LOG(DEBUG, "RX  port_id=%u queue_id=%u, %d packets received\n",
+		   (unsigned)rxq->port_id, (unsigned)rxq->qidx, nb_hold);
+
+	nb_hold += rxq->nb_rx_hold;
+
+	/*
+	 * FL descriptors needs to be written before incrementing the
+	 * FL queue WR pointer
+	 */
+	rte_wmb();
+	if (nb_hold > rxq->rx_free_thresh) {
+		PMD_RX_LOG(DEBUG, "port=%u queue=%u nb_hold=%u avail=%u\n",
+			   (unsigned)rxq->port_id, (unsigned)rxq->qidx,
+			   (unsigned)nb_hold, (unsigned)avail);
+		nfp_qcp_ptr_add(rxq->qcp_fl, NFP_QCP_WRITE_PTR, nb_hold);
+		nb_hold = 0;
+	}
+	rxq->nb_rx_hold = nb_hold;
+
+	return avail;
+}
+
+/*
+ * nfp_net_tx_free_bufs - Check for descriptors with a complete
+ * status
+ * @txq: TX queue to work with
+ * Returns number of descriptors freed
+ */
+int
+nfp_net_tx_free_bufs(struct nfp_net_txq *txq)
+{
+	uint32_t qcp_rd_p;
+	int todo;
+
+	PMD_TX_LOG(DEBUG, "queue %u. Check for descriptor with a complete"
+		   " status\n", txq->qidx);
+
+	/* Work out how many packets have been sent */
+	qcp_rd_p = nfp_qcp_read(txq->qcp_q, NFP_QCP_READ_PTR);
+
+	if (qcp_rd_p == txq->qcp_rd_p) {
+		PMD_TX_LOG(DEBUG, "queue %u: It seems harrier is not sending "
+			   "packets (%u, %u)\n", txq->qidx,
+			   qcp_rd_p, txq->qcp_rd_p);
+		return 0;
+	}
+
+	if (qcp_rd_p > txq->qcp_rd_p)
+		todo = qcp_rd_p - txq->qcp_rd_p;
+	else
+		todo = qcp_rd_p + txq->tx_count - txq->qcp_rd_p;
+
+	PMD_TX_LOG(DEBUG, "qcp_rd_p %u, txq->qcp_rd_p: %u, qcp->rd_p: %u\n",
+		   qcp_rd_p, txq->qcp_rd_p, txq->rd_p);
+
+	if (todo == 0)
+		return todo;
+
+	txq->qcp_rd_p += todo;
+	txq->qcp_rd_p %= txq->tx_count;
+	txq->rd_p += todo;
+
+	return todo;
+}
+
+/* Leaving always free descriptors for avoiding wrapping confusion */
+#define NFP_FREE_TX_DESC(t) (t->tx_count - (t->wr_p - t->rd_p) - 8)
+
+/*
+ * nfp_net_txq_full - Check if the TX queue free descriptors
+ * is below tx_free_threshold
+ *
+ * @txq: TX queue to check
+ *
+ * This function uses the host copy* of read/write pointers
+ */
+static inline
+int nfp_net_txq_full(struct nfp_net_txq *txq)
+{
+	return NFP_FREE_TX_DESC(txq) < txq->tx_free_thresh;
+}
+
+static uint16_t
+nfp_net_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	struct nfp_net_txq *txq;
+	struct nfp_net_hw *hw;
+	struct nfp_net_tx_desc *txds;
+	struct rte_mbuf *pkt;
+	uint64_t dma_addr;
+	int pkt_size, dma_size;
+	uint16_t free_descs, issued_descs;
+	struct rte_mbuf **lmbuf;
+	int i;
+
+	txq = tx_queue;
+	hw = txq->hw;
+	txds = &txq->txds[txq->tail];
+
+	PMD_TX_LOG(DEBUG, "working for queue %u at pos %d and %u packets\n",
+		   txq->qidx, wr_idx, nb_pkts);
+
+	if ((NFP_FREE_TX_DESC(txq) < nb_pkts) || (nfp_net_txq_full(txq)))
+		nfp_net_tx_free_bufs(txq);
+
+	free_descs = (uint16_t)NFP_FREE_TX_DESC(txq);
+	if (unlikely(free_descs == 0))
+		return 0;
+
+	pkt = *tx_pkts;
+
+	i = 0;
+	issued_descs = 0;
+	PMD_TX_LOG(DEBUG, "queue: %u. Sending %u packets\n",
+		   txq->qidx, nb_pkts);
+	/* Sending packets */
+	while ((i < nb_pkts) && free_descs) {
+		/* Grabbing the mbuf linked to the current descriptor */
+		lmbuf = &txq->txbufs[txq->tail].mbuf;
+		/* Warming the cache for releasing the mbuf later on */
+		RTE_MBUF_PREFETCH_TO_FREE(*lmbuf);
+
+		pkt = *(tx_pkts + i);
+
+		if (unlikely((pkt->nb_segs > 1) &&
+			     !(hw->cap & NFP_NET_CFG_CTRL_GATHER))) {
+			PMD_INIT_LOG(INFO, "NFP_NET_CFG_CTRL_GATHER not set\n");
+			rte_panic("Multisegment packet unsupported\n");
+		}
+
+		/* Checking if we have enough descriptors */
+		if (unlikely(pkt->nb_segs > free_descs))
+			goto xmit_end;
+
+		/*
+		 * Checksum and VLAN flags just in the first descriptor for a
+		 * multisegment packet
+		 */
+		nfp_net_tx_cksum(txq, txds, pkt);
+
+		if ((pkt->ol_flags & PKT_TX_VLAN_PKT) &&
+		    (hw->cap & NFP_NET_CFG_CTRL_TXVLAN)) {
+			txds->flags |= PCIE_DESC_TX_VLAN;
+			txds->vlan = pkt->vlan_tci;
+		}
+
+		if (pkt->ol_flags & PKT_TX_TCP_SEG)
+			rte_panic("TSO is not supported\n");
+
+		/*
+		 * mbuf data_len is the data in one segment and pkt_len data
+		 * in the whole packet. When the packet is just one segment,
+		 * then data_len = pkt_len
+		 */
+		pkt_size = pkt->pkt_len;
+
+		while (pkt_size) {
+			/* Releasing mbuf which was prefetched above */
+			if (*lmbuf)
+				rte_pktmbuf_free_seg(*lmbuf);
+
+			dma_size = pkt->data_len;
+			dma_addr = RTE_MBUF_DATA_DMA_ADDR(pkt);
+			PMD_TX_LOG(DEBUG, "Working with mbuf at dma address:"
+				   "%" PRIx64 "\n", dma_addr);
+
+			/* Filling descriptors fields */
+			txds->dma_len = dma_size;
+			txds->data_len = pkt->pkt_len;
+			txds->dma_addr_hi = (dma_addr >> 32) & 0xff;
+			txds->dma_addr_lo = (dma_addr & 0xffffffff);
+			ASSERT(free_descs > 0);
+			free_descs--;
+
+			/*
+			 * Linking mbuf with descriptor for being released
+			 * next time descriptor is used
+			 */
+			*lmbuf = pkt;
+
+			txq->wr_p++;
+			txq->tail++;
+			if (unlikely(txq->tail == txq->tx_count)) /* wrapping?*/
+				txq->tail = 0;
+
+			pkt_size -= dma_size;
+			if (!pkt_size) {
+				/* End of packet */
+				txds->offset_eop |= PCIE_DESC_TX_EOP;
+			} else {
+				txds->offset_eop &= PCIE_DESC_TX_OFFSET_MASK;
+				pkt = pkt->next;
+			}
+			/* Referencing next free TX descriptor */
+			txds = &txq->txds[txq->tail];
+			issued_descs++;
+		}
+		i++;
+	}
+
+xmit_end:
+	/* Increment write pointers. Force memory write before we let HW know */
+	rte_wmb();
+	nfp_qcp_ptr_add(txq->qcp_q, NFP_QCP_WRITE_PTR, issued_descs);
+
+	return i;
+}
+
 /* Initialise and register driver with DPDK Application */
 static struct eth_dev_ops nfp_net_eth_dev_ops = {
 	.dev_configure		= nfp_net_configure,
-- 
1.7.9.5