net/nfp: add NFDk Tx

Implement NFP3800 card packet transmit function for firmware
with NFDk.

Signed-off-by: Jin Liu <jin.liu@corigine.com>
Signed-off-by: Diana Wang <na.wang@corigine.com>
Signed-off-by: Peng Zhang <peng.zhang@corigine.com>
Signed-off-by: Chaoyong He <chaoyong.he@corigine.com>
Signed-off-by: Niklas Söderlund <niklas.soderlund@corigine.com>

drivers/net/nfp/nfp_ethdev.c

@@ -377,6 +377,7 @@ nfp_net_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 	switch (NFD_CFG_CLASS_VER_of(hw->ver)) {
 	case NFP_NET_CFG_VERSION_DP_NFD3:
 		eth_dev->dev_ops = &nfp_net_nfd3_eth_dev_ops;
+		eth_dev->tx_pkt_burst = &nfp_net_nfd3_xmit_pkts;
 		break;
 	case NFP_NET_CFG_VERSION_DP_NFDK:
 		if (NFD_CFG_MAJOR_VERSION_of(hw->ver) < 5) {
@@ -385,6 +386,7 @@ nfp_net_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 			return -EINVAL;
 		}
 		eth_dev->dev_ops = &nfp_net_nfdk_eth_dev_ops;
+		eth_dev->tx_pkt_burst = &nfp_net_nfdk_xmit_pkts;
 		break;
 	default:
 		PMD_DRV_LOG(ERR, "The version of firmware is not correct.");
@@ -393,7 +395,6 @@ nfp_net_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 
 	eth_dev->rx_queue_count = nfp_net_rx_queue_count;
 	eth_dev->rx_pkt_burst = &nfp_net_recv_pkts;
-	eth_dev->tx_pkt_burst = &nfp_net_nfd3_xmit_pkts;
 
 	return 0;
 }

drivers/net/nfp/nfp_ethdev_vf.c

@@ -282,6 +282,7 @@ nfp_netvf_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 	switch (NFD_CFG_CLASS_VER_of(hw->ver)) {
 	case NFP_NET_CFG_VERSION_DP_NFD3:
 		eth_dev->dev_ops = &nfp_netvf_nfd3_eth_dev_ops;
+		eth_dev->tx_pkt_burst = &nfp_net_nfd3_xmit_pkts;
 		break;
 	case NFP_NET_CFG_VERSION_DP_NFDK:
 		if (NFD_CFG_MAJOR_VERSION_of(hw->ver) < 5) {
@@ -290,6 +291,7 @@ nfp_netvf_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 			return -EINVAL;
 		}
 		eth_dev->dev_ops = &nfp_netvf_nfdk_eth_dev_ops;
+		eth_dev->tx_pkt_burst = &nfp_net_nfdk_xmit_pkts;
 		break;
 	default:
 		PMD_DRV_LOG(ERR, "The version of firmware is not correct.");
@@ -298,7 +300,6 @@ nfp_netvf_ethdev_ops_mount(struct nfp_net_hw *hw, struct rte_eth_dev *eth_dev)
 
 	eth_dev->rx_queue_count = nfp_net_rx_queue_count;
 	eth_dev->rx_pkt_burst = &nfp_net_recv_pkts;
-	eth_dev->tx_pkt_burst = &nfp_net_nfd3_xmit_pkts;
 
 	return 0;
 }

drivers/net/nfp/nfp_rxtx.c

@@ -20,6 +20,9 @@
 #include "nfp_rxtx.h"
 #include "nfp_logs.h"
 #include "nfp_ctrl.h"
+#include "nfpcore/nfp_mip.h"
+#include "nfpcore/nfp_rtsym.h"
+#include "nfpcore/nfp-common/nfp_platform.h"
 
 static int
 nfp_net_rx_fill_freelist(struct nfp_net_rxq *rxq)
@@ -1104,3 +1107,283 @@ nfp_net_nfdk_tx_queue_setup(struct rte_eth_dev *dev,
 
 	return 0;
 }
+
+static inline uint32_t
+nfp_net_nfdk_free_tx_desc(struct nfp_net_txq *txq)
+{
+	uint32_t free_desc;
+
+	if (txq->wr_p >= txq->rd_p)
+		free_desc = txq->tx_count - (txq->wr_p - txq->rd_p);
+	else
+		free_desc = txq->rd_p - txq->wr_p;
+
+	return (free_desc > NFDK_TX_DESC_STOP_CNT) ?
+		(free_desc - NFDK_TX_DESC_STOP_CNT) : 0;
+}
+
+static inline uint32_t
+nfp_net_nfdk_txq_full(struct nfp_net_txq *txq)
+{
+	return (nfp_net_nfdk_free_tx_desc(txq) < txq->tx_free_thresh);
+}
+
+static inline int
+nfp_net_nfdk_headlen_to_segs(unsigned int headlen)
+{
+	return DIV_ROUND_UP(headlen +
+			NFDK_TX_MAX_DATA_PER_DESC -
+			NFDK_TX_MAX_DATA_PER_HEAD,
+			NFDK_TX_MAX_DATA_PER_DESC);
+}
+
+static int
+nfp_net_nfdk_tx_maybe_close_block(struct nfp_net_txq *txq, struct rte_mbuf *pkt)
+{
+	unsigned int n_descs, wr_p, i, nop_slots;
+	struct rte_mbuf *pkt_temp;
+
+	pkt_temp = pkt;
+	n_descs = nfp_net_nfdk_headlen_to_segs(pkt_temp->data_len);
+	while (pkt_temp->next) {
+		pkt_temp = pkt_temp->next;
+		n_descs += DIV_ROUND_UP(pkt_temp->data_len, NFDK_TX_MAX_DATA_PER_DESC);
+	}
+
+	if (unlikely(n_descs > NFDK_TX_DESC_GATHER_MAX))
+		return -EINVAL;
+
+	n_descs += !!(pkt->ol_flags & RTE_MBUF_F_TX_TCP_SEG);
+
+	if (round_down(txq->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
+			round_down(txq->wr_p + n_descs, NFDK_TX_DESC_BLOCK_CNT))
+		goto close_block;
+
+	if ((uint32_t)txq->data_pending + pkt->pkt_len > NFDK_TX_MAX_DATA_PER_BLOCK)
+		goto close_block;
+
+	return 0;
+
+close_block:
+	wr_p = txq->wr_p;
+	nop_slots = D_BLOCK_CPL(wr_p);
+
+	memset(&txq->ktxds[wr_p], 0, nop_slots * sizeof(struct nfp_net_nfdk_tx_desc));
+	for (i = wr_p; i < nop_slots + wr_p; i++) {
+		if (txq->txbufs[i].mbuf) {
+			rte_pktmbuf_free_seg(txq->txbufs[i].mbuf);
+			txq->txbufs[i].mbuf = NULL;
+		}
+	}
+	txq->data_pending = 0;
+	txq->wr_p = D_IDX(txq, txq->wr_p + nop_slots);
+
+	return nop_slots;
+}
+
+static inline uint64_t
+nfp_net_nfdk_tx_cksum(struct nfp_net_txq *txq, struct rte_mbuf *mb,
+		uint64_t flags)
+{
+	uint64_t ol_flags;
+	struct nfp_net_hw *hw = txq->hw;
+
+	if (!(hw->cap & NFP_NET_CFG_CTRL_TXCSUM))
+		return flags;
+
+	ol_flags = mb->ol_flags;
+
+	/* IPv6 does not need checksum */
+	if (ol_flags & RTE_MBUF_F_TX_IP_CKSUM)
+		flags |= NFDK_DESC_TX_L3_CSUM;
+
+	if (ol_flags & RTE_MBUF_F_TX_L4_MASK)
+		flags |= NFDK_DESC_TX_L4_CSUM;
+
+	return flags;
+}
+
+static inline uint64_t
+nfp_net_nfdk_tx_tso(struct nfp_net_txq *txq, struct rte_mbuf *mb)
+{
+	uint64_t ol_flags;
+	struct nfp_net_nfdk_tx_desc txd;
+	struct nfp_net_hw *hw = txq->hw;
+
+	if (!(hw->cap & NFP_NET_CFG_CTRL_LSO_ANY))
+		goto clean_txd;
+
+	ol_flags = mb->ol_flags;
+
+	if (!(ol_flags & RTE_MBUF_F_TX_TCP_SEG))
+		goto clean_txd;
+
+	txd.l3_offset = mb->l2_len;
+	txd.l4_offset = mb->l2_len + mb->l3_len;
+	txd.lso_meta_res = 0;
+	txd.mss = rte_cpu_to_le_16(mb->tso_segsz);
+	txd.lso_hdrlen = mb->l2_len + mb->l3_len + mb->l4_len;
+	txd.lso_totsegs = (mb->pkt_len + mb->tso_segsz) / mb->tso_segsz;
+
+clean_txd:
+	txd.l3_offset = 0;
+	txd.l4_offset = 0;
+	txd.lso_hdrlen = 0;
+	txd.mss = 0;
+	txd.lso_totsegs = 0;
+	txd.lso_meta_res = 0;
+
+	return txd.raw;
+}
+
+uint16_t
+nfp_net_nfdk_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+	uint32_t buf_idx;
+	uint64_t dma_addr;
+	uint16_t free_descs;
+	uint32_t npkts = 0;
+	uint64_t metadata = 0;
+	uint16_t issued_descs = 0;
+	struct nfp_net_txq *txq;
+	struct nfp_net_hw *hw;
+	struct nfp_net_nfdk_tx_desc *ktxds;
+	struct rte_mbuf *pkt, *temp_pkt;
+	struct rte_mbuf **lmbuf;
+
+	txq = tx_queue;
+	hw = txq->hw;
+
+	PMD_TX_LOG(DEBUG, "working for queue %u at pos %d and %u packets",
+		txq->qidx, txq->wr_p, nb_pkts);
+
+	if ((nfp_net_nfdk_free_tx_desc(txq) < NFDK_TX_DESC_PER_SIMPLE_PKT *
+			nb_pkts) || (nfp_net_nfdk_txq_full(txq)))
+		nfp_net_tx_free_bufs(txq);
+
+	free_descs = (uint16_t)nfp_net_nfdk_free_tx_desc(txq);
+	if (unlikely(free_descs == 0))
+		return 0;
+
+	PMD_TX_LOG(DEBUG, "queue: %u. Sending %u packets", txq->qidx, nb_pkts);
+	/* Sending packets */
+	while ((npkts < nb_pkts) && free_descs) {
+		uint32_t type, dma_len, dlen_type, tmp_dlen;
+		int nop_descs, used_descs;
+
+		pkt = *(tx_pkts + npkts);
+		nop_descs = nfp_net_nfdk_tx_maybe_close_block(txq, pkt);
+		if (nop_descs < 0)
+			goto xmit_end;
+
+		issued_descs += nop_descs;
+		ktxds = &txq->ktxds[txq->wr_p];
+		/* Grabbing the mbuf linked to the current descriptor */
+		buf_idx = txq->wr_p;
+		lmbuf = &txq->txbufs[buf_idx++].mbuf;
+		/* Warming the cache for releasing the mbuf later on */
+		RTE_MBUF_PREFETCH_TO_FREE(*lmbuf);
+
+		temp_pkt = pkt;
+
+		if (unlikely(pkt->nb_segs > 1 &&
+				!(hw->cap & NFP_NET_CFG_CTRL_GATHER))) {
+			PMD_INIT_LOG(INFO, "NFP_NET_CFG_CTRL_GATHER not set");
+			PMD_INIT_LOG(INFO, "Multisegment packet unsupported");
+			goto xmit_end;
+		}
+
+		/*
+		 * Checksum and VLAN flags just in the first descriptor for a
+		 * multisegment packet, but TSO info needs to be in all of them.
+		 */
+
+		dma_len = pkt->data_len;
+		if ((hw->cap & NFP_NET_CFG_CTRL_LSO_ANY) &&
+				(pkt->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) {
+			type = NFDK_DESC_TX_TYPE_TSO;
+		} else if (!pkt->next && dma_len < NFDK_TX_MAX_DATA_PER_HEAD) {
+			type = NFDK_DESC_TX_TYPE_SIMPLE;
+		} else {
+			type = NFDK_DESC_TX_TYPE_GATHER;
+		}
+		dma_len -= 1;
+		dlen_type = (NFDK_DESC_TX_DMA_LEN_HEAD & dma_len) |
+			(NFDK_DESC_TX_TYPE_HEAD & (type << 12));
+		ktxds->dma_len_type = rte_cpu_to_le_16(dlen_type);
+		dma_addr = rte_mbuf_data_iova(pkt);
+		PMD_TX_LOG(DEBUG, "Working with mbuf at dma address:"
+				"%" PRIx64 "", dma_addr);
+		ktxds->dma_addr_hi = rte_cpu_to_le_16(dma_addr >> 32);
+		ktxds->dma_addr_lo = rte_cpu_to_le_32(dma_addr & 0xffffffff);
+		ktxds++;
+
+		tmp_dlen = dlen_type & NFDK_DESC_TX_DMA_LEN_HEAD;
+		dma_len -= tmp_dlen;
+		dma_addr += tmp_dlen + 1;
+
+		while (pkt) {
+			if (*lmbuf)
+				rte_pktmbuf_free_seg(*lmbuf);
+			*lmbuf = pkt;
+			while (dma_len > 0) {
+				dma_len -= 1;
+				dlen_type = NFDK_DESC_TX_DMA_LEN & dma_len;
+
+				ktxds->dma_len_type = rte_cpu_to_le_16(dlen_type);
+				ktxds->dma_addr_hi = rte_cpu_to_le_16(dma_addr >> 32);
+				ktxds->dma_addr_lo = rte_cpu_to_le_32(dma_addr & 0xffffffff);
+				ktxds++;
+
+				dma_len -= dlen_type;
+				dma_addr += dlen_type + 1;
+			}
+
+			if (!pkt->next)
+				break;
+
+			pkt = pkt->next;
+			dma_len = pkt->data_len;
+			dma_addr = rte_mbuf_data_iova(pkt);
+			PMD_TX_LOG(DEBUG, "Working with mbuf at dma address:"
+				"%" PRIx64 "", dma_addr);
+
+			lmbuf = &txq->txbufs[buf_idx++].mbuf;
+		}
+
+		(ktxds - 1)->dma_len_type = rte_cpu_to_le_16(dlen_type | NFDK_DESC_TX_EOP);
+
+		ktxds->raw = rte_cpu_to_le_64(nfp_net_nfdk_tx_cksum(txq, temp_pkt, metadata));
+		ktxds++;
+
+		if ((hw->cap & NFP_NET_CFG_CTRL_LSO_ANY) &&
+				(temp_pkt->ol_flags & RTE_MBUF_F_TX_TCP_SEG)) {
+			ktxds->raw = rte_cpu_to_le_64(nfp_net_nfdk_tx_tso(txq, temp_pkt));
+			ktxds++;
+		}
+
+		used_descs = ktxds - txq->ktxds - txq->wr_p;
+		if (round_down(txq->wr_p, NFDK_TX_DESC_BLOCK_CNT) !=
+			round_down(txq->wr_p + used_descs - 1, NFDK_TX_DESC_BLOCK_CNT)) {
+			PMD_INIT_LOG(INFO, "Used descs cross block boundary");
+			goto xmit_end;
+		}
+
+		txq->wr_p = D_IDX(txq, txq->wr_p + used_descs);
+		if (txq->wr_p % NFDK_TX_DESC_BLOCK_CNT)
+			txq->data_pending += temp_pkt->pkt_len;
+		else
+			txq->data_pending = 0;
+
+		issued_descs += used_descs;
+		npkts++;
+		free_descs = (uint16_t)nfp_net_nfdk_free_tx_desc(txq);
+	}
+
+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 npkts;
+}

drivers/net/nfp/nfp_rxtx.h

@@ -352,6 +352,9 @@ int nfp_net_nfdk_tx_queue_setup(struct rte_eth_dev *dev,
 		uint16_t nb_desc,
 		unsigned int socket_id,
 		const struct rte_eth_txconf *tx_conf);
+uint16_t nfp_net_nfdk_xmit_pkts(void *tx_queue,
+		struct rte_mbuf **tx_pkts,
+		uint16_t nb_pkts);
 
 #endif /* _NFP_RXTX_H_ */
 /*