[dpdk-dev] [PATCH 03/11] net/bnxt: implement vector mode driver

[Ajit Khaparde <ajit.khaparde@broadcom.com>]

From: Lance Richardson <lance.richardson@broadcom.com>

Introduce vector mode support for the bnxt pmd.

Signed-off-by: Lance Richardson <lance.richardson@broadcom.com>
Signed-off-by: Ajit Khaparde <ajit.khaparde@broadcom.com>
---
 config/common_base                   |   1 +
 drivers/net/bnxt/Makefile            |   1 +
 drivers/net/bnxt/bnxt_ethdev.c       |  92 ++++-
 drivers/net/bnxt/bnxt_ring.h         |   3 +-
 drivers/net/bnxt/bnxt_rxq.c          |   5 +
 drivers/net/bnxt/bnxt_rxq.h          |   4 +
 drivers/net/bnxt/bnxt_rxr.h          |   9 +-
 drivers/net/bnxt/bnxt_rxtx_vec_sse.c | 481 +++++++++++++++++++++++++++
 drivers/net/bnxt/bnxt_txr.h          |  11 +-
 9 files changed, 598 insertions(+), 9 deletions(-)
 create mode 100644 drivers/net/bnxt/bnxt_rxtx_vec_sse.c

diff --git a/config/common_base b/config/common_base
index 6b96e0e80..1bbb7c10b 100644
--- a/config/common_base
+++ b/config/common_base
@@ -212,6 +212,7 @@ CONFIG_RTE_LIBRTE_BNX2X_DEBUG_PERIODIC=n
 # Compile burst-oriented Broadcom BNXT PMD driver
 #
 CONFIG_RTE_LIBRTE_BNXT_PMD=y
+CONFIG_RTE_LIBRTE_BNXT_INC_VECTOR=n
 
 #
 # Compile burst-oriented Chelsio Terminator (CXGBE) PMD
diff --git a/drivers/net/bnxt/Makefile b/drivers/net/bnxt/Makefile
index 8be3cb0e4..9e006b5d1 100644
--- a/drivers/net/bnxt/Makefile
+++ b/drivers/net/bnxt/Makefile
@@ -41,6 +41,7 @@ SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += bnxt_vnic.c
 SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += bnxt_irq.c
 SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += bnxt_util.c
 SRCS-$(CONFIG_RTE_LIBRTE_BNXT_PMD) += rte_pmd_bnxt.c
+SRCS-$(CONFIG_RTE_LIBRTE_BNXT_INC_VECTOR) += bnxt_rxtx_vec_sse.c
 
 #
 # Export include files
diff --git a/drivers/net/bnxt/bnxt_ethdev.c b/drivers/net/bnxt/bnxt_ethdev.c
index e0e0b72c6..064a153ec 100644
--- a/drivers/net/bnxt/bnxt_ethdev.c
+++ b/drivers/net/bnxt/bnxt_ethdev.c
@@ -645,6 +645,67 @@ static int bnxt_scattered_rx(struct rte_eth_dev *eth_dev)
 	return 0;
 }
 
+static eth_rx_burst_t
+bnxt_receive_function(__rte_unused struct rte_eth_dev *eth_dev)
+{
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+	/*
+	 * Vector mode receive can be enabled only if scatter rx is not
+	 * in use and rx offloads are limited to VLAN stripping and
+	 * CRC stripping.
+	 */
+	if (!eth_dev->data->scattered_rx &&
+	    !(eth_dev->data->dev_conf.rxmode.offloads &
+	      ~(DEV_RX_OFFLOAD_VLAN_STRIP |
+		DEV_RX_OFFLOAD_KEEP_CRC |
+		DEV_RX_OFFLOAD_JUMBO_FRAME |
+		DEV_RX_OFFLOAD_IPV4_CKSUM |
+		DEV_RX_OFFLOAD_UDP_CKSUM |
+		DEV_RX_OFFLOAD_TCP_CKSUM |
+		DEV_RX_OFFLOAD_OUTER_IPV4_CKSUM |
+		DEV_RX_OFFLOAD_VLAN_FILTER))) {
+		PMD_DRV_LOG(INFO, "Using vector mode receive for port %d\n",
+			    eth_dev->data->port_id);
+		return bnxt_recv_pkts_vec;
+	}
+	PMD_DRV_LOG(INFO, "Vector mode receive disabled for port %d\n",
+		    eth_dev->data->port_id);
+	PMD_DRV_LOG(INFO,
+		    "Port %d scatter: %d rx offload: %" PRIX64 "\n",
+		    eth_dev->data->port_id,
+		    eth_dev->data->scattered_rx,
+		    eth_dev->data->dev_conf.rxmode.offloads);
+#endif
+	return bnxt_recv_pkts;
+}
+
+static eth_tx_burst_t
+bnxt_transmit_function(__rte_unused struct rte_eth_dev *eth_dev)
+{
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+	/*
+	 * Vector mode receive can be enabled only if scatter tx is not
+	 * in use and tx offloads other than VLAN insertion are not
+	 * in use.
+	 */
+	if (!eth_dev->data->scattered_rx &&
+	    !(eth_dev->data->dev_conf.txmode.offloads &
+	      ~DEV_TX_OFFLOAD_VLAN_INSERT)) {
+		PMD_DRV_LOG(INFO, "Using vector mode transmit for port %d\n",
+			    eth_dev->data->port_id);
+		return bnxt_xmit_pkts_vec;
+	}
+	PMD_DRV_LOG(INFO, "Vector mode transmit disabled for port %d\n",
+		    eth_dev->data->port_id);
+	PMD_DRV_LOG(INFO,
+		    "Port %d scatter: %d tx offload: %" PRIX64 "\n",
+		    eth_dev->data->port_id,
+		    eth_dev->data->scattered_rx,
+		    eth_dev->data->dev_conf.txmode.offloads);
+#endif
+	return bnxt_xmit_pkts;
+}
+
 static int bnxt_dev_start_op(struct rte_eth_dev *eth_dev)
 {
 	struct bnxt *bp = (struct bnxt *)eth_dev->data->dev_private;
@@ -675,6 +736,8 @@ static int bnxt_dev_start_op(struct rte_eth_dev *eth_dev)
 	if (rc)
 		goto error;
 
+	eth_dev->rx_pkt_burst = bnxt_receive_function(eth_dev);
+	eth_dev->tx_pkt_burst = bnxt_transmit_function(eth_dev);
 	bp->flags |= BNXT_FLAG_INIT_DONE;
 	return 0;
 
@@ -1597,6 +1660,8 @@ bnxt_txq_info_get_op(struct rte_eth_dev *dev, uint16_t queue_id,
 
 static int bnxt_mtu_set_op(struct rte_eth_dev *eth_dev, uint16_t new_mtu)
 {
+	uint32_t new_pkt_size = new_mtu + ETHER_HDR_LEN + ETHER_CRC_LEN +
+				VLAN_TAG_SIZE * BNXT_NUM_VLANS;
 	struct bnxt *bp = eth_dev->data->dev_private;
 	struct rte_eth_dev_info dev_info;
 	uint32_t rc = 0;
@@ -1610,6 +1675,23 @@ static int bnxt_mtu_set_op(struct rte_eth_dev *eth_dev, uint16_t new_mtu)
 		return -EINVAL;
 	}
 
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+	/*
+	 * If vector-mode tx/rx is active, disallow any MTU change that would
+	 * require scattered receive support.
+	 */
+	if (eth_dev->data->dev_started &&
+	    (eth_dev->rx_pkt_burst == bnxt_recv_pkts_vec ||
+	     eth_dev->tx_pkt_burst == bnxt_xmit_pkts_vec) &&
+	    (new_pkt_size >
+	     eth_dev->data->min_rx_buf_size - RTE_PKTMBUF_HEADROOM)) {
+		PMD_DRV_LOG(ERR,
+			    "MTU change would require scattered rx support. ");
+		PMD_DRV_LOG(ERR, "Stop port before changing MTU.\n");
+		return -EINVAL;
+	}
+#endif
+
 	if (new_mtu > ETHER_MTU) {
 		bp->flags |= BNXT_FLAG_JUMBO;
 		bp->eth_dev->data->dev_conf.rxmode.offloads |=
@@ -1620,8 +1702,7 @@ static int bnxt_mtu_set_op(struct rte_eth_dev *eth_dev, uint16_t new_mtu)
 		bp->flags &= ~BNXT_FLAG_JUMBO;
 	}
 
-	eth_dev->data->dev_conf.rxmode.max_rx_pkt_len =
-		new_mtu + ETHER_HDR_LEN + ETHER_CRC_LEN + VLAN_TAG_SIZE * 2;
+	eth_dev->data->dev_conf.rxmode.max_rx_pkt_len = new_pkt_size;
 
 	eth_dev->data->mtu = new_mtu;
 	PMD_DRV_LOG(INFO, "New MTU is %d\n", eth_dev->data->mtu);
@@ -2655,9 +2736,10 @@ bnxt_dev_supported_ptypes_get_op(struct rte_eth_dev *dev)
 		RTE_PTYPE_UNKNOWN
 	};
 
-	if (dev->rx_pkt_burst == bnxt_recv_pkts)
-		return ptypes;
-	return NULL;
+	if (!dev->rx_pkt_burst)
+		return NULL;
+
+	return ptypes;
 }
 
 static int bnxt_map_regs(struct bnxt *bp, uint32_t *reg_arr, int count,
diff --git a/drivers/net/bnxt/bnxt_ring.h b/drivers/net/bnxt/bnxt_ring.h
index 1446d784f..fd3d0bd73 100644
--- a/drivers/net/bnxt/bnxt_ring.h
+++ b/drivers/net/bnxt/bnxt_ring.h
@@ -10,7 +10,8 @@
 
 #include <rte_memory.h>
 
-#define RING_NEXT(ring, idx)		(((idx) + 1) & (ring)->ring_mask)
+#define RING_ADV(ring, idx, n)		(((idx) + (n)) & (ring)->ring_mask)
+#define RING_NEXT(ring, idx)		RING_ADV(ring, idx, 1)
 
 #define DB_IDX_MASK						0xffffff
 #define DB_IDX_VALID						(0x1 << 26)
diff --git a/drivers/net/bnxt/bnxt_rxq.c b/drivers/net/bnxt/bnxt_rxq.c
index 17e2909a7..ff219f4b8 100644
--- a/drivers/net/bnxt/bnxt_rxq.c
+++ b/drivers/net/bnxt/bnxt_rxq.c
@@ -355,6 +355,11 @@ int bnxt_rx_queue_setup_op(struct rte_eth_dev *eth_dev,
 						RTE_ETH_QUEUE_STATE_STARTED;
 	eth_dev->data->rx_queue_state[queue_idx] = queue_state;
 	rte_spinlock_init(&rxq->lock);
+
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+	bnxt_rxq_vec_setup(rxq);
+#endif
+
 out:
 	return rc;
 }
diff --git a/drivers/net/bnxt/bnxt_rxq.h b/drivers/net/bnxt/bnxt_rxq.h
index e5d6001d3..057a59e2c 100644
--- a/drivers/net/bnxt/bnxt_rxq.h
+++ b/drivers/net/bnxt/bnxt_rxq.h
@@ -22,6 +22,10 @@ struct bnxt_rx_queue {
 	uint16_t		nb_rx_hold; /* num held free RX desc */
 	uint16_t		rx_free_thresh; /* max free RX desc to hold */
 	uint16_t		queue_id; /* RX queue index */
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+	uint16_t		rxrearm_nb; /* number of descs to reinit. */
+	uint16_t		rxrearm_start; /* next desc index to reinit. */
+#endif
 	uint16_t		reg_idx; /* RX queue register index */
 	uint16_t		port_id; /* Device port identifier */
 	uint8_t			crc_len; /* 0 if CRC stripped, 4 otherwise */
diff --git a/drivers/net/bnxt/bnxt_rxr.h b/drivers/net/bnxt/bnxt_rxr.h
index 3815a2199..cf1c7e5cf 100644
--- a/drivers/net/bnxt/bnxt_rxr.h
+++ b/drivers/net/bnxt/bnxt_rxr.h
@@ -7,7 +7,7 @@
 #define _BNXT_RXR_H_
 
 #define B_RX_DB(db, prod)						\
-		(*(uint32_t *)db = (DB_KEY_RX | prod))
+		(*(uint32_t *)db = (DB_KEY_RX | (prod)))
 
 #define BNXT_TPA_L4_SIZE(x)	\
 	{ \
@@ -110,4 +110,11 @@ int bnxt_init_rx_ring_struct(struct bnxt_rx_queue *rxq, unsigned int socket_id);
 int bnxt_init_one_rx_ring(struct bnxt_rx_queue *rxq);
 int bnxt_rx_queue_start(struct rte_eth_dev *dev, uint16_t rx_queue_id);
 int bnxt_rx_queue_stop(struct rte_eth_dev *dev, uint16_t rx_queue_id);
+
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+uint16_t bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+			    uint16_t nb_pkts);
+int bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq);
+#endif
+
 #endif
diff --git a/drivers/net/bnxt/bnxt_rxtx_vec_sse.c b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
new file mode 100644
index 000000000..1c32c986c
--- /dev/null
+++ b/drivers/net/bnxt/bnxt_rxtx_vec_sse.c
@@ -0,0 +1,481 @@
+// SPDX-License-Identifier: BSD-3-Clause
+/* Copyright(c) 2019 Broadcom All rights reserved. */
+
+#include <inttypes.h>
+#include <stdbool.h>
+
+#include <rte_bitmap.h>
+#include <rte_byteorder.h>
+#include <rte_malloc.h>
+#include <rte_memory.h>
+#if defined(RTE_ARCH_X86)
+#include <tmmintrin.h>
+#else
+#error "bnxt vector pmd: unsupported target."
+#endif
+
+#include "bnxt.h"
+#include "bnxt_cpr.h"
+#include "bnxt_ring.h"
+#include "bnxt_rxr.h"
+#include "bnxt_rxq.h"
+#include "hsi_struct_def_dpdk.h"
+
+#include "bnxt_txq.h"
+#include "bnxt_txr.h"
+
+/*
+ * RX Ring handling
+ */
+
+#define RTE_BNXT_MAX_RX_BURST		32
+#define RTE_BNXT_MAX_TX_BURST		32
+#define RTE_BNXT_RXQ_REARM_THRESH	32
+#define RTE_BNXT_DESCS_PER_LOOP		4
+
+static inline void
+bnxt_rxq_rearm(struct bnxt_rx_queue *rxq, struct bnxt_rx_ring_info *rxr)
+{
+	struct rx_prod_pkt_bd *rxbds = &rxr->rx_desc_ring[rxq->rxrearm_start];
+	struct bnxt_sw_rx_bd *rx_bufs = &rxr->rx_buf_ring[rxq->rxrearm_start];
+	struct rte_mbuf *mb0, *mb1;
+	int i;
+
+	const __m128i hdr_room = _mm_set_epi64x(RTE_PKTMBUF_HEADROOM, 0);
+	const __m128i addrmask = _mm_set_epi64x(UINT64_MAX, 0);
+
+	/* Pull RTE_BNXT_RXQ_REARM_THRESH more mbufs into the software ring */
+	if (rte_mempool_get_bulk(rxq->mb_pool,
+				 (void *)rx_bufs,
+				 RTE_BNXT_RXQ_REARM_THRESH) < 0) {
+		rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed +=
+			RTE_BNXT_RXQ_REARM_THRESH;
+
+		return;
+	}
+
+	/* Initialize the mbufs in vector, process 2 mbufs in one loop */
+	for (i = 0; i < RTE_BNXT_RXQ_REARM_THRESH; i += 2, rx_bufs += 2) {
+		__m128i buf_addr0, buf_addr1;
+		__m128i rxbd0, rxbd1;
+
+		mb0 = rx_bufs[0].mbuf;
+		mb1 = rx_bufs[1].mbuf;
+
+		/* Load address fields from both mbufs */
+		buf_addr0 = _mm_loadu_si128((__m128i *)&mb0->buf_addr);
+		buf_addr1 = _mm_loadu_si128((__m128i *)&mb1->buf_addr);
+
+		/* Load both rx descriptors (preserving some existing fields) */
+		rxbd0 = _mm_loadu_si128((__m128i *)(rxbds + 0));
+		rxbd1 = _mm_loadu_si128((__m128i *)(rxbds + 1));
+
+		/* Add default offset to buffer address. */
+		buf_addr0 = _mm_add_epi64(buf_addr0, hdr_room);
+		buf_addr1 = _mm_add_epi64(buf_addr1, hdr_room);
+
+		/* Clear all fields except address. */
+		buf_addr0 =  _mm_and_si128(buf_addr0, addrmask);
+		buf_addr1 =  _mm_and_si128(buf_addr1, addrmask);
+
+		/* Clear address field in descriptor. */
+		rxbd0 = _mm_andnot_si128(addrmask, rxbd0);
+		rxbd1 = _mm_andnot_si128(addrmask, rxbd1);
+
+		/* Set address field in descriptor. */
+		rxbd0 = _mm_add_epi64(rxbd0, buf_addr0);
+		rxbd1 = _mm_add_epi64(rxbd1, buf_addr1);
+
+		/* Store descriptors to memory. */
+		_mm_store_si128((__m128i *)(rxbds++), rxbd0);
+		_mm_store_si128((__m128i *)(rxbds++), rxbd1);
+	}
+
+	rxq->rxrearm_start += RTE_BNXT_RXQ_REARM_THRESH;
+	B_RX_DB(rxr->rx_doorbell, rxq->rxrearm_start - 1);
+	if (rxq->rxrearm_start >= rxq->nb_rx_desc)
+		rxq->rxrearm_start = 0;
+
+	rxq->rxrearm_nb -= RTE_BNXT_RXQ_REARM_THRESH;
+}
+
+static uint32_t
+bnxt_parse_pkt_type(struct rx_pkt_cmpl *rxcmp, struct rx_pkt_cmpl_hi *rxcmp1)
+{
+	uint32_t l3, pkt_type = 0;
+	uint32_t t_ipcs = 0, ip6 = 0, vlan = 0;
+	uint32_t flags_type;
+
+	vlan = !!(rxcmp1->flags2 &
+		rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN));
+	pkt_type |= vlan ? RTE_PTYPE_L2_ETHER_VLAN : RTE_PTYPE_L2_ETHER;
+
+	t_ipcs = !!(rxcmp1->flags2 &
+		rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_T_IP_CS_CALC));
+	ip6 = !!(rxcmp1->flags2 &
+		 rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS2_IP_TYPE));
+
+	flags_type = rxcmp->flags_type &
+		rte_cpu_to_le_32(RX_PKT_CMPL_FLAGS_ITYPE_MASK);
+
+	if (!t_ipcs && !ip6)
+		l3 = RTE_PTYPE_L3_IPV4_EXT_UNKNOWN;
+	else if (!t_ipcs && ip6)
+		l3 = RTE_PTYPE_L3_IPV6_EXT_UNKNOWN;
+	else if (t_ipcs && !ip6)
+		l3 = RTE_PTYPE_INNER_L3_IPV4_EXT_UNKNOWN;
+	else
+		l3 = RTE_PTYPE_INNER_L3_IPV6_EXT_UNKNOWN;
+
+	switch (flags_type) {
+	case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_ICMP):
+		if (!t_ipcs)
+			pkt_type |= l3 | RTE_PTYPE_L4_ICMP;
+		else
+			pkt_type |= l3 | RTE_PTYPE_INNER_L4_ICMP;
+		break;
+
+	case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_TCP):
+		if (!t_ipcs)
+			pkt_type |= l3 | RTE_PTYPE_L4_TCP;
+		else
+			pkt_type |= l3 | RTE_PTYPE_INNER_L4_TCP;
+		break;
+
+	case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_UDP):
+		if (!t_ipcs)
+			pkt_type |= l3 | RTE_PTYPE_L4_UDP;
+		else
+			pkt_type |= l3 | RTE_PTYPE_INNER_L4_UDP;
+		break;
+
+	case RTE_LE32(RX_PKT_CMPL_FLAGS_ITYPE_IP):
+		pkt_type |= l3;
+		break;
+	}
+
+	return pkt_type;
+}
+
+uint16_t
+bnxt_recv_pkts_vec(void *rx_queue, struct rte_mbuf **rx_pkts,
+		   uint16_t nb_pkts)
+{
+	struct bnxt_rx_queue *rxq = rx_queue;
+	struct bnxt_cp_ring_info *cpr = rxq->cp_ring;
+	struct bnxt_rx_ring_info *rxr = rxq->rx_ring;
+	uint32_t raw_cons = cpr->cp_raw_cons;
+	uint32_t cons;
+	int nb_rx_pkts = 0;
+	struct rx_pkt_cmpl *rxcmp;
+	bool evt = false;
+	const __m128i mbuf_init = _mm_set_epi64x(0, rxq->mbuf_initializer);
+	const __m128i shuf_msk =
+		_mm_set_epi8(15, 14, 13, 12,          /* rss */
+			     0xFF, 0xFF,              /* vlan_tci (zeroes) */
+			     3, 2,                    /* data_len */
+			     0xFF, 0xFF, 3, 2,        /* pkt_len */
+			     0xFF, 0xFF, 0xFF, 0xFF); /* pkt_type (zeroes) */
+
+	/* If Rx Q was stopped return */
+	if (rxq->rx_deferred_start)
+		return 0;
+
+	if (rxq->rxrearm_nb >= RTE_BNXT_RXQ_REARM_THRESH)
+		bnxt_rxq_rearm(rxq, rxr);
+
+	/* Return no more than RTE_BNXT_MAX_RX_BURST per call. */
+	nb_pkts = RTE_MIN(nb_pkts, RTE_BNXT_MAX_RX_BURST);
+
+	/* Make nb_pkts an integer multiple of RTE_BNXT_DESCS_PER_LOOP */
+	nb_pkts = RTE_ALIGN_FLOOR(nb_pkts, RTE_BNXT_DESCS_PER_LOOP);
+
+	/* Handle RX burst request */
+	while (1) {
+		cons = RING_CMP(cpr->cp_ring_struct, raw_cons);
+
+		rxcmp = (struct rx_pkt_cmpl *)&cpr->cp_desc_ring[cons];
+
+		if (!CMP_VALID(rxcmp, raw_cons, cpr->cp_ring_struct))
+			break;
+
+		cpr->valid = FLIP_VALID(cons,
+					cpr->cp_ring_struct->ring_mask,
+					cpr->valid);
+
+		if (likely(CMP_TYPE(rxcmp) == RX_PKT_CMPL_TYPE_RX_L2)) {
+			struct rx_pkt_cmpl_hi *rxcmp1;
+			uint32_t tmp_raw_cons;
+			uint16_t cp_cons;
+			struct rte_mbuf *mbuf;
+			__m128i mm_rxcmp, pkt_mb;
+
+			tmp_raw_cons = NEXT_RAW_CMP(raw_cons);
+			cp_cons = RING_CMP(cpr->cp_ring_struct, tmp_raw_cons);
+			rxcmp1 = (struct rx_pkt_cmpl_hi *)
+						&cpr->cp_desc_ring[cp_cons];
+
+			if (!CMP_VALID(rxcmp1, tmp_raw_cons,
+				       cpr->cp_ring_struct))
+				break;
+
+			raw_cons = tmp_raw_cons;
+			cons = rxcmp->opaque;
+
+			mbuf = rxr->rx_buf_ring[cons].mbuf;
+			rte_prefetch0(mbuf);
+			rxr->rx_buf_ring[cons].mbuf = NULL;
+
+			cpr->valid = FLIP_VALID(cp_cons,
+						cpr->cp_ring_struct->ring_mask,
+						cpr->valid);
+
+			/* Set constant fields from mbuf initializer. */
+			_mm_store_si128((__m128i *)&mbuf->rearm_data,
+					mbuf_init);
+
+			/* Set mbuf pkt_len, data_len, and rss_hash fields. */
+			mm_rxcmp = _mm_load_si128((__m128i *)rxcmp);
+			pkt_mb = _mm_shuffle_epi8(mm_rxcmp, shuf_msk);
+			_mm_storeu_si128((void *)&mbuf->rx_descriptor_fields1,
+					 pkt_mb);
+
+			rte_compiler_barrier();
+
+			if (rxcmp->flags_type & RX_PKT_CMPL_FLAGS_RSS_VALID)
+				mbuf->ol_flags |= PKT_RX_RSS_HASH;
+
+			if (rxcmp1->flags2 &
+			    RX_PKT_CMPL_FLAGS2_META_FORMAT_VLAN) {
+				mbuf->vlan_tci = rxcmp1->metadata &
+					(RX_PKT_CMPL_METADATA_VID_MASK |
+					RX_PKT_CMPL_METADATA_DE |
+					RX_PKT_CMPL_METADATA_PRI_MASK);
+				mbuf->ol_flags |= PKT_RX_VLAN;
+			}
+
+			mbuf->packet_type = bnxt_parse_pkt_type(rxcmp, rxcmp1);
+
+			rx_pkts[nb_rx_pkts++] = mbuf;
+		} else {
+			evt =
+			bnxt_event_hwrm_resp_handler(rxq->bp,
+						     (struct cmpl_base *)rxcmp);
+		}
+
+		raw_cons = NEXT_RAW_CMP(raw_cons);
+		if (nb_rx_pkts == nb_pkts || evt)
+			break;
+	}
+	rxr->rx_prod = RING_ADV(rxr->rx_ring_struct, rxr->rx_prod, nb_rx_pkts);
+
+	rxq->rxrearm_nb += nb_rx_pkts;
+	cpr->cp_raw_cons = raw_cons;
+	if (nb_rx_pkts || evt)
+		B_CP_DIS_DB(cpr, cpr->cp_raw_cons);
+
+	return nb_rx_pkts;
+}
+
+static inline void bnxt_next_cmpl(struct bnxt_cp_ring_info *cpr, uint32_t *idx,
+				  bool *v, uint32_t inc)
+{
+	*idx += inc;
+	if (unlikely(*idx == cpr->cp_ring_struct->ring_size)) {
+		*v = !*v;
+		*idx = 0;
+	}
+}
+
+static void
+bnxt_tx_cmp_vec(struct bnxt_tx_queue *txq, int nr_pkts)
+{
+	struct bnxt_tx_ring_info *txr = txq->tx_ring;
+	struct rte_mbuf **free = txq->free;
+	uint16_t cons = txr->tx_cons;
+	unsigned int blk = 0;
+
+	while (nr_pkts--) {
+		struct bnxt_sw_tx_bd *tx_buf;
+		struct rte_mbuf *mbuf;
+
+		tx_buf = &txr->tx_buf_ring[cons];
+		cons = RING_NEXT(txr->tx_ring_struct, cons);
+		mbuf = rte_pktmbuf_prefree_seg(tx_buf->mbuf);
+		tx_buf->mbuf = NULL;
+
+		if (blk && mbuf->pool != free[0]->pool) {
+			rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
+			blk = 0;
+		}
+		free[blk++] = mbuf;
+	}
+	if (blk)
+		rte_mempool_put_bulk(free[0]->pool, (void **)free, blk);
+
+	txr->tx_cons = cons;
+}
+
+static void
+bnxt_handle_tx_cp_vec(struct bnxt_tx_queue *txq)
+{
+	struct bnxt_cp_ring_info *cpr = txq->cp_ring;
+	uint32_t raw_cons = cpr->cp_raw_cons;
+	uint32_t cons;
+	uint32_t nb_tx_pkts = 0;
+	struct tx_cmpl *txcmp;
+	struct cmpl_base *cp_desc_ring = cpr->cp_desc_ring;
+	struct bnxt_ring *cp_ring_struct = cpr->cp_ring_struct;
+	uint32_t ring_mask = cp_ring_struct->ring_mask;
+
+	do {
+		cons = RING_CMPL(ring_mask, raw_cons);
+		txcmp = (struct tx_cmpl *)&cp_desc_ring[cons];
+
+		if (!CMPL_VALID(txcmp, cpr->valid))
+			break;
+		bnxt_next_cmpl(cpr, &cons, &cpr->valid, 1);
+		rte_prefetch0(&cp_desc_ring[cons]);
+
+		if (likely(CMP_TYPE(txcmp) == TX_CMPL_TYPE_TX_L2))
+			nb_tx_pkts += txcmp->opaque;
+		else
+			RTE_LOG_DP(ERR, PMD,
+				   "Unhandled CMP type %02x\n",
+				   CMP_TYPE(txcmp));
+		raw_cons = cons;
+	} while (nb_tx_pkts < ring_mask);
+
+	if (nb_tx_pkts) {
+		bnxt_tx_cmp_vec(txq, nb_tx_pkts);
+		cpr->cp_raw_cons = raw_cons;
+		B_CP_DB(cpr, raw_cons, ring_mask);
+	}
+}
+
+#define TX_BD_FLAGS_CMPL ((1 << TX_BD_LONG_FLAGS_BD_CNT_SFT) | \
+			  TX_BD_SHORT_FLAGS_COAL_NOW | \
+			  TX_BD_SHORT_TYPE_TX_BD_SHORT | \
+			  TX_BD_LONG_FLAGS_PACKET_END)
+
+#define TX_BD_FLAGS_NOCMPL (TX_BD_FLAGS_CMPL | TX_BD_LONG_FLAGS_NO_CMPL)
+
+static inline uint32_t
+bnxt_xmit_flags_len(uint16_t len, uint16_t flags)
+{
+	switch (len >> 9) {
+	case 0:
+		return flags | TX_BD_LONG_FLAGS_LHINT_LT512;
+	case 1:
+		return flags | TX_BD_LONG_FLAGS_LHINT_LT1K;
+	case 2:
+		return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
+	case 3:
+		return flags | TX_BD_LONG_FLAGS_LHINT_LT2K;
+	default:
+		return flags | TX_BD_LONG_FLAGS_LHINT_GTE2K;
+	}
+}
+
+static uint16_t
+bnxt_xmit_fixed_burst_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+			  uint16_t nb_pkts)
+{
+	struct bnxt_tx_queue *txq = tx_queue;
+	struct bnxt_tx_ring_info *txr = txq->tx_ring;
+	uint16_t prod = txr->tx_prod;
+	struct rte_mbuf *tx_mbuf;
+	struct tx_bd_long *txbd = NULL;
+	struct bnxt_sw_tx_bd *tx_buf;
+	uint16_t to_send;
+
+	nb_pkts = RTE_MIN(nb_pkts, bnxt_tx_avail(txq));
+
+	if (unlikely(nb_pkts == 0))
+		return 0;
+
+	/* Handle TX burst request */
+	to_send = nb_pkts;
+	while (to_send) {
+		tx_mbuf = *tx_pkts++;
+		rte_prefetch0(tx_mbuf);
+
+		tx_buf = &txr->tx_buf_ring[prod];
+		tx_buf->mbuf = tx_mbuf;
+		tx_buf->nr_bds = 1;
+
+		txbd = &txr->tx_desc_ring[prod];
+		txbd->address = tx_mbuf->buf_iova + tx_mbuf->data_off;
+		txbd->len = tx_mbuf->data_len;
+		txbd->flags_type = bnxt_xmit_flags_len(tx_mbuf->data_len,
+						       TX_BD_FLAGS_NOCMPL);
+		prod = RING_NEXT(txr->tx_ring_struct, prod);
+		to_send--;
+	}
+
+	/* Request a completion for last packet in burst */
+	if (txbd) {
+		txbd->opaque = nb_pkts;
+		txbd->flags_type &= ~TX_BD_LONG_FLAGS_NO_CMPL;
+	}
+
+	rte_compiler_barrier();
+	B_TX_DB(txr->tx_doorbell, prod);
+
+	txr->tx_prod = prod;
+
+	return nb_pkts;
+}
+
+uint16_t
+bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+		   uint16_t nb_pkts)
+{
+	int nb_sent = 0;
+	struct bnxt_tx_queue *txq = tx_queue;
+
+	/* Tx queue was stopped; wait for it to be restarted */
+	if (unlikely(txq->tx_deferred_start)) {
+		PMD_DRV_LOG(DEBUG, "Tx q stopped;return\n");
+		return 0;
+	}
+
+	/* Handle TX completions */
+	if (bnxt_tx_bds_in_hw(txq) >= txq->tx_free_thresh)
+		bnxt_handle_tx_cp_vec(txq);
+
+	while (nb_pkts) {
+		uint16_t ret, num;
+
+		num = RTE_MIN(nb_pkts, RTE_BNXT_MAX_TX_BURST);
+		ret = bnxt_xmit_fixed_burst_vec(tx_queue,
+						&tx_pkts[nb_sent],
+						num);
+		nb_sent += ret;
+		nb_pkts -= ret;
+		if (ret < num)
+			break;
+	}
+
+	return nb_sent;
+}
+
+int __attribute__((cold))
+bnxt_rxq_vec_setup(struct bnxt_rx_queue *rxq)
+{
+	uintptr_t p;
+	struct rte_mbuf mb_def = { .buf_addr = 0 }; /* zeroed mbuf */
+
+	mb_def.nb_segs = 1;
+	mb_def.data_off = RTE_PKTMBUF_HEADROOM;
+	mb_def.port = rxq->port_id;
+	rte_mbuf_refcnt_set(&mb_def, 1);
+
+	/* prevent compiler reordering: rearm_data covers previous fields */
+	rte_compiler_barrier();
+	p = (uintptr_t)&mb_def.rearm_data;
+	rxq->mbuf_initializer = *(uint64_t *)p;
+	rxq->rxrearm_nb = 0;
+	rxq->rxrearm_start = 0;
+	return 0;
+}
diff --git a/drivers/net/bnxt/bnxt_txr.h b/drivers/net/bnxt/bnxt_txr.h
index 13ca04676..c28bd9c18 100644
--- a/drivers/net/bnxt/bnxt_txr.h
+++ b/drivers/net/bnxt/bnxt_txr.h
@@ -37,13 +37,15 @@ struct bnxt_sw_tx_bd {
 	unsigned short		nr_bds;
 };
 
-static inline uint32_t bnxt_tx_bds_in_hw(struct bnxt_tx_queue *txq)
+static inline uint32_t
+bnxt_tx_bds_in_hw(struct bnxt_tx_queue *txq)
 {
 	return ((txq->tx_ring->tx_prod - txq->tx_ring->tx_cons) &
 		txq->tx_ring->tx_ring_struct->ring_mask);
 }
 
-static inline uint32_t bnxt_tx_avail(struct bnxt_tx_queue *txq)
+static inline uint32_t
+bnxt_tx_avail(struct bnxt_tx_queue *txq)
 {
 	/* Tell compiler to fetch tx indices from memory. */
 	rte_compiler_barrier();
@@ -57,6 +59,11 @@ int bnxt_init_one_tx_ring(struct bnxt_tx_queue *txq);
 int bnxt_init_tx_ring_struct(struct bnxt_tx_queue *txq, unsigned int socket_id);
 uint16_t bnxt_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
 			       uint16_t nb_pkts);
+#ifdef RTE_LIBRTE_BNXT_INC_VECTOR
+uint16_t bnxt_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+			    uint16_t nb_pkts);
+#endif
+
 int bnxt_tx_queue_start(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 int bnxt_tx_queue_stop(struct rte_eth_dev *dev, uint16_t tx_queue_id);
 
-- 
2.20.1 (Apple Git-117)