DPDK patches and discussions
 help / color / mirror / Atom feed
From: Ziyang Xuan <xuanziyang2@huawei.com>
To: <dev@dpdk.org>
Cc: <ferruh.yigit@intel.com>, <cloud.wangxiaoyun@huawei.com>,
	<shahar.belkar@huawei.com>, <tanya.brokhman@huawei.com>,
	<luoxianjun@huawei.com>, Ziyang Xuan <xuanziyang2@huawei.com>
Subject: [dpdk-dev] [PATCH v5 14/15] net/hinic: add tx/rx package burst
Date: Thu, 20 Jun 2019 00:25:12 +0800	[thread overview]
Message-ID: <691dcc833654d7e8a666e88ea07dae95a4951230.1560958308.git.xuanziyang2@huawei.com> (raw)
Message-ID: <20190619162512.Qf09MQkfiUCMI7Kxa6lwJf9njaq29NVa-RoEgj957m0@z> (raw)
In-Reply-To: <cover.1560958308.git.xuanziyang2@huawei.com>

This patch add package sending and receiving function codes.

Signed-off-by: Ziyang Xuan <xuanziyang2@huawei.com>
---
 drivers/net/hinic/hinic_pmd_ethdev.c |    4 +
 drivers/net/hinic/hinic_pmd_rx.c     |  439 +++++++++++
 drivers/net/hinic/hinic_pmd_tx.c     | 1082 ++++++++++++++++++++++++++
 3 files changed, 1525 insertions(+)

diff --git a/drivers/net/hinic/hinic_pmd_ethdev.c b/drivers/net/hinic/hinic_pmd_ethdev.c
index eeb4227fa..6bd35a227 100644
--- a/drivers/net/hinic/hinic_pmd_ethdev.c
+++ b/drivers/net/hinic/hinic_pmd_ethdev.c
@@ -1543,6 +1543,8 @@ static int hinic_dev_init(struct rte_eth_dev *eth_dev)
 
 	/* rte_eth_dev ops, rx_burst and tx_burst */
 	eth_dev->dev_ops = &hinic_pmd_ops;
+	eth_dev->rx_pkt_burst = hinic_recv_pkts;
+	eth_dev->tx_pkt_burst = hinic_xmit_pkts;
 
 	return hinic_func_init(eth_dev);
 }
@@ -1560,6 +1562,8 @@ static int hinic_dev_uninit(struct rte_eth_dev *dev)
 	hinic_dev_close(dev);
 
 	dev->dev_ops = NULL;
+	dev->rx_pkt_burst = NULL;
+	dev->tx_pkt_burst = NULL;
 
 	rte_free(dev->data->mac_addrs);
 	dev->data->mac_addrs = NULL;
diff --git a/drivers/net/hinic/hinic_pmd_rx.c b/drivers/net/hinic/hinic_pmd_rx.c
index 592a889f0..e474deefe 100644
--- a/drivers/net/hinic/hinic_pmd_rx.c
+++ b/drivers/net/hinic/hinic_pmd_rx.c
@@ -4,6 +4,9 @@
 
 #include <rte_ether.h>
 #include <rte_mbuf.h>
+#ifdef __ARM64_NEON__
+#include <arm_neon.h>
+#endif
 
 #include "base/hinic_compat.h"
 #include "base/hinic_pmd_hwdev.h"
@@ -35,8 +38,69 @@
 
 #define HINIC_GET_RQ_FREE_WQEBBS(rxq)	((rxq)->wq->delta - 1)
 
+/* rxq cqe done and status bit */
+#define HINIC_GET_RX_DONE_BE(status)	\
+	((status) & 0x80U)
+
 #define HINIC_RX_CSUM_OFFLOAD_EN	0xFFF
 
+#define RQ_CQE_SGE_VLAN_SHIFT			0
+#define RQ_CQE_SGE_LEN_SHIFT			16
+
+#define RQ_CQE_SGE_VLAN_MASK			0xFFFFU
+#define RQ_CQE_SGE_LEN_MASK			0xFFFFU
+
+#define RQ_CQE_SGE_GET(val, member)		\
+	(((val) >> RQ_CQE_SGE_##member##_SHIFT) & RQ_CQE_SGE_##member##_MASK)
+
+#define HINIC_GET_RX_VLAN_TAG(vlan_len)	\
+		RQ_CQE_SGE_GET(vlan_len, VLAN)
+
+#define HINIC_GET_RX_PKT_LEN(vlan_len)	\
+		RQ_CQE_SGE_GET(vlan_len, LEN)
+
+#define RQ_CQE_STATUS_CSUM_ERR_SHIFT		0
+#define RQ_CQE_STATUS_NUM_LRO_SHIFT		16
+#define RQ_CQE_STATUS_LRO_PUSH_SHIFT		25
+#define RQ_CQE_STATUS_LRO_ENTER_SHIFT		26
+#define RQ_CQE_STATUS_LRO_INTR_SHIFT		27
+
+#define RQ_CQE_STATUS_BP_EN_SHIFT		30
+#define RQ_CQE_STATUS_RXDONE_SHIFT		31
+#define RQ_CQE_STATUS_FLUSH_SHIFT		28
+
+#define RQ_CQE_STATUS_CSUM_ERR_MASK		0xFFFFU
+#define RQ_CQE_STATUS_NUM_LRO_MASK		0xFFU
+#define RQ_CQE_STATUS_LRO_PUSH_MASK		0X1U
+#define RQ_CQE_STATUS_LRO_ENTER_MASK		0X1U
+#define RQ_CQE_STATUS_LRO_INTR_MASK		0X1U
+#define RQ_CQE_STATUS_BP_EN_MASK		0X1U
+#define RQ_CQE_STATUS_RXDONE_MASK		0x1U
+#define RQ_CQE_STATUS_FLUSH_MASK		0x1U
+
+#define RQ_CQE_STATUS_GET(val, member)		\
+		(((val) >> RQ_CQE_STATUS_##member##_SHIFT) & \
+				RQ_CQE_STATUS_##member##_MASK)
+
+#define RQ_CQE_STATUS_CLEAR(val, member)	\
+		((val) & (~(RQ_CQE_STATUS_##member##_MASK << \
+				RQ_CQE_STATUS_##member##_SHIFT)))
+
+#define HINIC_GET_RX_CSUM_ERR(status)	\
+		RQ_CQE_STATUS_GET(status, CSUM_ERR)
+
+#define HINIC_GET_RX_DONE(status)	\
+		RQ_CQE_STATUS_GET(status, RXDONE)
+
+#define HINIC_GET_RX_FLUSH(status)	\
+		RQ_CQE_STATUS_GET(status, FLUSH)
+
+#define HINIC_GET_RX_BP_EN(status)	\
+		RQ_CQE_STATUS_GET(status, BP_EN)
+
+#define HINIC_GET_RX_NUM_LRO(status)	\
+		RQ_CQE_STATUS_GET(status, NUM_LRO)
+
 /* RQ_CTRL */
 #define	RQ_CTRL_BUFDESC_SECT_LEN_SHIFT		0
 #define	RQ_CTRL_COMPLETE_FORMAT_SHIFT		15
@@ -57,6 +121,72 @@
 #define RQ_CTRL_CLEAR(val, member)		\
 	((val) & (~(RQ_CTRL_##member##_MASK << RQ_CTRL_##member##_SHIFT)))
 
+#define RQ_CQE_PKT_NUM_SHIFT			1
+#define RQ_CQE_PKT_FIRST_LEN_SHIFT		19
+#define RQ_CQE_PKT_LAST_LEN_SHIFT		6
+#define RQ_CQE_SUPER_CQE_EN_SHIFT		0
+
+#define RQ_CQE_PKT_FIRST_LEN_MASK		0x1FFFU
+#define RQ_CQE_PKT_LAST_LEN_MASK		0x1FFFU
+#define RQ_CQE_PKT_NUM_MASK			0x1FU
+#define RQ_CQE_SUPER_CQE_EN_MASK		0x1
+
+#define RQ_CQE_PKT_NUM_GET(val, member)		\
+	(((val) >> RQ_CQE_PKT_##member##_SHIFT) & RQ_CQE_PKT_##member##_MASK)
+
+#define HINIC_GET_RQ_CQE_PKT_NUM(pkt_info) RQ_CQE_PKT_NUM_GET(pkt_info, NUM)
+
+#define RQ_CQE_SUPER_CQE_EN_GET(val, member)	\
+	(((val) >> RQ_CQE_##member##_SHIFT) & RQ_CQE_##member##_MASK)
+
+#define HINIC_GET_SUPER_CQE_EN(pkt_info)	\
+	RQ_CQE_SUPER_CQE_EN_GET(pkt_info, SUPER_CQE_EN)
+
+#define RQ_CQE_OFFOLAD_TYPE_VLAN_EN_SHIFT		21
+#define RQ_CQE_OFFOLAD_TYPE_VLAN_EN_MASK		0x1U
+
+#define RQ_CQE_OFFOLAD_TYPE_PKT_TYPE_SHIFT		0
+#define RQ_CQE_OFFOLAD_TYPE_PKT_TYPE_MASK		0xFFFU
+
+#define RQ_CQE_OFFOLAD_TYPE_PKT_UMBCAST_SHIFT		19
+#define RQ_CQE_OFFOLAD_TYPE_PKT_UMBCAST_MASK		0x3U
+
+#define RQ_CQE_OFFOLAD_TYPE_RSS_TYPE_SHIFT		24
+#define RQ_CQE_OFFOLAD_TYPE_RSS_TYPE_MASK		0xFFU
+
+#define RQ_CQE_OFFOLAD_TYPE_GET(val, member)		(((val) >> \
+				RQ_CQE_OFFOLAD_TYPE_##member##_SHIFT) & \
+				RQ_CQE_OFFOLAD_TYPE_##member##_MASK)
+
+#define HINIC_GET_RX_VLAN_OFFLOAD_EN(offload_type)	\
+		RQ_CQE_OFFOLAD_TYPE_GET(offload_type, VLAN_EN)
+
+#define HINIC_GET_RSS_TYPES(offload_type)	\
+		RQ_CQE_OFFOLAD_TYPE_GET(offload_type, RSS_TYPE)
+
+#define HINIC_GET_RX_PKT_TYPE(offload_type)	\
+		RQ_CQE_OFFOLAD_TYPE_GET(offload_type, PKT_TYPE)
+
+#define HINIC_GET_RX_PKT_UMBCAST(offload_type)	\
+		RQ_CQE_OFFOLAD_TYPE_GET(offload_type, PKT_UMBCAST)
+
+#define RQ_CQE_STATUS_CSUM_BYPASS_VAL			0x80U
+#define RQ_CQE_STATUS_CSUM_ERR_IP_MASK			0x39U
+#define RQ_CQE_STATUS_CSUM_ERR_L4_MASK			0x46U
+#define RQ_CQE_STATUS_CSUM_ERR_OTHER			0x100U
+
+#define HINIC_CSUM_ERR_BYPASSED(csum_err)	 \
+	((csum_err) == RQ_CQE_STATUS_CSUM_BYPASS_VAL)
+
+#define HINIC_CSUM_ERR_IP(csum_err)	 \
+	((csum_err) & RQ_CQE_STATUS_CSUM_ERR_IP_MASK)
+
+#define HINIC_CSUM_ERR_L4(csum_err)	 \
+	((csum_err) & RQ_CQE_STATUS_CSUM_ERR_L4_MASK)
+
+#define HINIC_CSUM_ERR_OTHER(csum_err)	 \
+	((csum_err) == RQ_CQE_STATUS_CSUM_ERR_OTHER)
+
 
 void hinic_get_func_rx_buf_size(struct hinic_nic_dev *nic_dev)
 {
@@ -156,6 +286,25 @@ hinic_prepare_rq_wqe(void *wqe, __rte_unused u16 pi, dma_addr_t buf_addr,
 	buf_desc->addr_low = lower_32_bits(buf_addr);
 }
 
+void hinic_rxq_get_stats(struct hinic_rxq *rxq, struct hinic_rxq_stats *stats)
+{
+	if (!rxq || !stats)
+		return;
+
+	memcpy(stats, &rxq->rxq_stats, sizeof(rxq->rxq_stats));
+}
+
+void hinic_rxq_stats_reset(struct hinic_rxq *rxq)
+{
+	struct hinic_rxq_stats *rxq_stats;
+
+	if (rxq == NULL)
+		return;
+
+	rxq_stats = &rxq->rxq_stats;
+	memset(rxq_stats, 0, sizeof(*rxq_stats));
+}
+
 static int hinic_rx_alloc_cqe(struct hinic_rxq *rxq)
 {
 	size_t cqe_mem_size;
@@ -289,6 +438,42 @@ void hinic_free_all_rx_mbuf(struct rte_eth_dev *eth_dev)
 		hinic_free_all_rx_skbs(nic_dev->rxqs[q_id]);
 }
 
+static void hinic_recv_jumbo_pkt(struct hinic_rxq *rxq,
+				 struct rte_mbuf *head_skb,
+				 u32 remain_pkt_len)
+{
+	struct hinic_nic_dev *nic_dev = rxq->nic_dev;
+	struct rte_mbuf *cur_mbuf, *rxm = NULL;
+	struct hinic_rx_info *rx_info;
+	u16 sw_ci, rx_buf_len = rxq->buf_len;
+	u32 pkt_len;
+
+	while (remain_pkt_len > 0) {
+		sw_ci = hinic_get_rq_local_ci(nic_dev->hwdev, rxq->q_id);
+		rx_info = &rxq->rx_info[sw_ci];
+
+		hinic_update_rq_local_ci(nic_dev->hwdev, rxq->q_id, 1);
+
+		pkt_len = remain_pkt_len > rx_buf_len ?
+			rx_buf_len : remain_pkt_len;
+		remain_pkt_len -= pkt_len;
+
+		cur_mbuf = rx_info->mbuf;
+		cur_mbuf->data_len = (u16)pkt_len;
+		cur_mbuf->next = NULL;
+
+		head_skb->pkt_len += cur_mbuf->data_len;
+		head_skb->nb_segs++;
+
+		if (!rxm)
+			head_skb->next = cur_mbuf;
+		else
+			rxm->next = cur_mbuf;
+
+		rxm = cur_mbuf;
+	}
+}
+
 static void hinic_rss_deinit(struct hinic_nic_dev *nic_dev)
 {
 	u8 prio_tc[HINIC_DCB_UP_MAX] = {0};
@@ -543,6 +728,113 @@ void hinic_free_all_rx_skbs(struct hinic_rxq *rxq)
 	}
 }
 
+static inline void hinic_rq_cqe_be_to_cpu32(void *dst_le32,
+					    volatile void *src_be32)
+{
+#ifndef __ARM64_NEON__
+	volatile __m128i *wqe_be = (volatile __m128i *)src_be32;
+	__m128i *wqe_le = (__m128i *)dst_le32;
+	__m128i shuf_mask =  _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
+					  11, 4, 5, 6, 7, 0, 1, 2, 3);
+
+	/* l2nic just use first 128 bits */
+	wqe_le[0] = _mm_shuffle_epi8(wqe_be[0], shuf_mask);
+#else
+	volatile uint8x16_t *wqe_be = (volatile uint8x16_t *)src_be32;
+	uint8x16_t *wqe_le = (uint8x16_t *)dst_le32;
+	const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
+					9, 8, 15, 14, 13, 12};
+
+	/* l2nic just use first 128 bits */
+	wqe_le[0] = vqtbl1q_u8(wqe_be[0], shuf_mask);
+#endif
+}
+
+static inline uint64_t hinic_rx_rss_hash(uint32_t offload_type,
+					 uint32_t cqe_hass_val,
+					 uint32_t *rss_hash)
+{
+	uint32_t rss_type;
+
+	rss_type = HINIC_GET_RSS_TYPES(offload_type);
+	if (likely(rss_type != 0)) {
+		*rss_hash = cqe_hass_val;
+		return PKT_RX_RSS_HASH;
+	}
+
+	return 0;
+}
+
+static inline uint64_t hinic_rx_csum(uint32_t status, struct hinic_rxq *rxq)
+{
+	uint32_t checksum_err;
+	uint64_t flags;
+
+	/* most case checksum is ok */
+	checksum_err = HINIC_GET_RX_CSUM_ERR(status);
+	if (likely(checksum_err == 0))
+		return (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD);
+
+	/* If BYPASS bit set, all other status indications should be ignored */
+	if (unlikely(HINIC_CSUM_ERR_BYPASSED(checksum_err)))
+		return PKT_RX_IP_CKSUM_UNKNOWN;
+
+	flags = 0;
+
+	/* IP checksum error */
+	if (HINIC_CSUM_ERR_IP(checksum_err))
+		flags |= PKT_RX_IP_CKSUM_BAD;
+	else
+		flags |= PKT_RX_IP_CKSUM_GOOD;
+
+	/* L4 checksum error */
+	if (HINIC_CSUM_ERR_L4(checksum_err))
+		flags |= PKT_RX_L4_CKSUM_BAD;
+	else
+		flags |= PKT_RX_L4_CKSUM_GOOD;
+
+	if (unlikely(HINIC_CSUM_ERR_OTHER(checksum_err)))
+		flags = PKT_RX_L4_CKSUM_NONE;
+
+	rxq->rxq_stats.errors++;
+
+	return flags;
+}
+
+static inline uint64_t hinic_rx_vlan(uint32_t offload_type, uint32_t vlan_len,
+				     uint16_t *vlan_tci)
+{
+	uint16_t vlan_tag;
+
+	vlan_tag = HINIC_GET_RX_VLAN_TAG(vlan_len);
+	if (!HINIC_GET_RX_VLAN_OFFLOAD_EN(offload_type) || 0 == vlan_tag) {
+		*vlan_tci = 0;
+		return 0;
+	}
+
+	*vlan_tci = vlan_tag;
+
+	return PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED;
+}
+
+static inline u32 hinic_rx_alloc_mbuf_bulk(struct hinic_rxq *rxq,
+					   struct rte_mbuf **mbufs,
+					   u32 exp_mbuf_cnt)
+{
+	int rc;
+	u32 avail_cnt;
+
+	rc = rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, exp_mbuf_cnt);
+	if (likely(rc == HINIC_OK)) {
+		avail_cnt = exp_mbuf_cnt;
+	} else {
+		avail_cnt = 0;
+		rxq->rxq_stats.rx_nombuf += exp_mbuf_cnt;
+	}
+
+	return avail_cnt;
+}
+
 static struct rte_mbuf *hinic_rx_alloc_mbuf(struct hinic_rxq *rxq,
 					dma_addr_t *dma_addr)
 {
@@ -557,6 +849,51 @@ static struct rte_mbuf *hinic_rx_alloc_mbuf(struct hinic_rxq *rxq,
 	return mbuf;
 }
 
+static inline void hinic_rearm_rxq_mbuf(struct hinic_rxq *rxq)
+{
+	u16 pi;
+	u32 i, free_wqebbs, rearm_wqebbs, exp_wqebbs;
+	dma_addr_t dma_addr;
+	struct hinic_rq_wqe *rq_wqe;
+	struct rte_mbuf **rearm_mbufs;
+
+	/* check free wqebb fo rearm */
+	free_wqebbs = HINIC_GET_RQ_FREE_WQEBBS(rxq);
+	if (unlikely(free_wqebbs < rxq->rx_free_thresh))
+		return;
+
+	/* get rearm mbuf array */
+	pi = HINIC_GET_RQ_LOCAL_PI(rxq);
+	rearm_mbufs = (struct rte_mbuf **)(&rxq->rx_info[pi]);
+
+	/* check rxq free wqebbs turn around */
+	exp_wqebbs = rxq->q_depth - pi;
+	if (free_wqebbs < exp_wqebbs)
+		exp_wqebbs = free_wqebbs;
+
+	/* alloc mbuf in bulk */
+	rearm_wqebbs = hinic_rx_alloc_mbuf_bulk(rxq, rearm_mbufs, exp_wqebbs);
+	if (unlikely(rearm_wqebbs == 0))
+		return;
+
+	/* rearm rx mbuf */
+	rq_wqe = WQ_WQE_ADDR(rxq->wq, (u32)pi);
+	for (i = 0; i < rearm_wqebbs; i++) {
+		dma_addr = rte_mbuf_data_iova_default(rearm_mbufs[i]);
+		rq_wqe->buf_desc.addr_high =
+					cpu_to_be32(upper_32_bits(dma_addr));
+		rq_wqe->buf_desc.addr_low =
+					cpu_to_be32(lower_32_bits(dma_addr));
+		rq_wqe++;
+	}
+	rxq->wq->prod_idx += rearm_wqebbs;
+	rxq->wq->delta -= rearm_wqebbs;
+
+	/* update rq hw_pi */
+	rte_wmb();
+	HINIC_UPDATE_RQ_HW_PI(rxq, pi + rearm_wqebbs);
+}
+
 void hinic_rx_alloc_pkts(struct hinic_rxq *rxq)
 {
 	struct hinic_nic_dev *nic_dev = rxq->nic_dev;
@@ -596,3 +933,105 @@ void hinic_rx_alloc_pkts(struct hinic_rxq *rxq)
 		HINIC_UPDATE_RQ_HW_PI(rxq, pi + 1);
 	}
 }
+
+u16 hinic_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts, u16 nb_pkts)
+{
+	struct rte_mbuf *rxm;
+	struct hinic_rxq *rxq = rx_queue;
+	struct hinic_rx_info *rx_info;
+	volatile struct hinic_rq_cqe *rx_cqe;
+	u16 rx_buf_len, pkts = 0;
+	u16 sw_ci, ci_mask, wqebb_cnt = 0;
+	u32 pkt_len, status, vlan_len;
+	u64 rx_bytes = 0;
+	struct hinic_rq_cqe cqe;
+	u32 offload_type, rss_hash;
+
+	rx_buf_len = rxq->buf_len;
+
+	/* 1. get polling start ci */
+	ci_mask = HINIC_GET_RQ_WQE_MASK(rxq);
+	sw_ci = HINIC_GET_RQ_LOCAL_CI(rxq);
+
+	while (pkts < nb_pkts) {
+		 /* 2. current ci is done */
+		rx_cqe = &rxq->rx_cqe[sw_ci];
+		status = rx_cqe->status;
+		if (!HINIC_GET_RX_DONE_BE(status))
+			break;
+
+		/* read other cqe member after status */
+		rte_rmb();
+
+		/* convert cqe and get packet length */
+		hinic_rq_cqe_be_to_cpu32(&cqe, (volatile void *)rx_cqe);
+		vlan_len = cqe.vlan_len;
+
+		rx_info = &rxq->rx_info[sw_ci];
+		rxm = rx_info->mbuf;
+
+		/* 3. next ci point and prefetch */
+		sw_ci++;
+		sw_ci &= ci_mask;
+
+		/* prefetch next mbuf first 64B */
+		rte_prefetch0(rxq->rx_info[sw_ci].mbuf);
+
+		/* 4. jumbo frame process */
+		pkt_len = HINIC_GET_RX_PKT_LEN(vlan_len);
+		if (likely(pkt_len <= rx_buf_len)) {
+			rxm->data_len = pkt_len;
+			rxm->pkt_len = pkt_len;
+			wqebb_cnt++;
+		} else {
+			rxm->data_len = rx_buf_len;
+			rxm->pkt_len = rx_buf_len;
+
+			/* if jumbo use multi-wqebb update ci,
+			 * recv_jumbo_pkt will also update ci
+			 */
+			HINIC_UPDATE_RQ_LOCAL_CI(rxq, wqebb_cnt + 1);
+			wqebb_cnt = 0;
+			hinic_recv_jumbo_pkt(rxq, rxm, pkt_len - rx_buf_len);
+			sw_ci = HINIC_GET_RQ_LOCAL_CI(rxq);
+		}
+
+		/* 5. vlan/checksum/rss/pkt_type/gro offload */
+		rxm->data_off = RTE_PKTMBUF_HEADROOM;
+		rxm->port = rxq->port_id;
+		offload_type = cqe.offload_type;
+
+		/* vlan offload */
+		rxm->ol_flags |= hinic_rx_vlan(offload_type, vlan_len,
+					       &rxm->vlan_tci);
+
+		/* checksum offload */
+		rxm->ol_flags |= hinic_rx_csum(cqe.status, rxq);
+
+		/* rss hash offload */
+		rss_hash = cqe.rss_hash;
+		rxm->ol_flags |= hinic_rx_rss_hash(offload_type, rss_hash,
+						   &rxm->hash.rss);
+
+		/* 6. clear done bit */
+		rx_cqe->status = 0;
+
+		rx_bytes += pkt_len;
+		rx_pkts[pkts++] = rxm;
+	}
+
+	if (pkts) {
+		/* 7. update ci */
+		HINIC_UPDATE_RQ_LOCAL_CI(rxq, wqebb_cnt);
+
+		/* do packet stats */
+		rxq->rxq_stats.packets += pkts;
+		rxq->rxq_stats.bytes += rx_bytes;
+	}
+	rxq->rxq_stats.burst_pkts = pkts;
+
+	/* 8. rearm mbuf to rxq */
+	hinic_rearm_rxq_mbuf(rxq);
+
+	return pkts;
+}
diff --git a/drivers/net/hinic/hinic_pmd_tx.c b/drivers/net/hinic/hinic_pmd_tx.c
index 854b94dc5..84101aebb 100644
--- a/drivers/net/hinic/hinic_pmd_tx.c
+++ b/drivers/net/hinic/hinic_pmd_tx.c
@@ -7,6 +7,9 @@
 #include <rte_sctp.h>
 #include <rte_udp.h>
 #include <rte_ip.h>
+#ifdef __ARM64_NEON__
+#include <arm_neon.h>
+#endif
 
 #include "base/hinic_compat.h"
 #include "base/hinic_pmd_hwdev.h"
@@ -16,6 +19,1085 @@
 #include "hinic_pmd_ethdev.h"
 #include "hinic_pmd_tx.h"
 
+/* packet header and tx offload info */
+#define VXLANLEN			8
+#define MAX_PLD_OFFSET			221
+#define MAX_SINGLE_SGE_SIZE		65536
+#define TSO_ENABLE			1
+#define TX_MSS_DEFAULT			0x3E00
+#define TX_MSS_MIN			0x50
+
+#define HINIC_NONTSO_PKT_MAX_SGE		17	/* non-tso max sge 17 */
+#define HINIC_NONTSO_SEG_NUM_INVALID(num)	\
+			((num) > HINIC_NONTSO_PKT_MAX_SGE)
+
+#define HINIC_TSO_PKT_MAX_SGE			127	/* tso max sge 127 */
+#define HINIC_TSO_SEG_NUM_INVALID(num)		((num) > HINIC_TSO_PKT_MAX_SGE)
+
+/* sizeof(struct hinic_sq_bufdesc) == 16, shift 4 */
+#define HINIC_BUF_DESC_SIZE(nr_descs)	(SIZE_8BYTES(((u32)nr_descs) << 4))
+
+#define MASKED_SQ_IDX(sq, idx)		((idx) & (sq)->wq->mask)
+
+/* SQ_CTRL */
+#define SQ_CTRL_BUFDESC_SECT_LEN_SHIFT		0
+#define SQ_CTRL_TASKSECT_LEN_SHIFT		16
+#define SQ_CTRL_DATA_FORMAT_SHIFT		22
+#define SQ_CTRL_LEN_SHIFT			29
+#define SQ_CTRL_OWNER_SHIFT			31
+
+#define SQ_CTRL_BUFDESC_SECT_LEN_MASK		0xFFU
+#define SQ_CTRL_TASKSECT_LEN_MASK		0x1FU
+#define SQ_CTRL_DATA_FORMAT_MASK		0x1U
+#define SQ_CTRL_LEN_MASK			0x3U
+#define SQ_CTRL_OWNER_MASK			0x1U
+
+#define SQ_CTRL_SET(val, member)	\
+	(((val) & SQ_CTRL_##member##_MASK) << SQ_CTRL_##member##_SHIFT)
+
+#define SQ_CTRL_QUEUE_INFO_PLDOFF_SHIFT		2
+#define SQ_CTRL_QUEUE_INFO_UFO_SHIFT		10
+#define SQ_CTRL_QUEUE_INFO_TSO_SHIFT		11
+#define SQ_CTRL_QUEUE_INFO_TCPUDP_CS_SHIFT	12
+#define SQ_CTRL_QUEUE_INFO_MSS_SHIFT		13
+#define SQ_CTRL_QUEUE_INFO_SCTP_SHIFT		27
+#define SQ_CTRL_QUEUE_INFO_UC_SHIFT		28
+#define SQ_CTRL_QUEUE_INFO_PRI_SHIFT		29
+
+#define SQ_CTRL_QUEUE_INFO_PLDOFF_MASK		0xFFU
+#define SQ_CTRL_QUEUE_INFO_UFO_MASK		0x1U
+#define SQ_CTRL_QUEUE_INFO_TSO_MASK		0x1U
+#define SQ_CTRL_QUEUE_INFO_TCPUDP_CS_MASK	0x1U
+#define SQ_CTRL_QUEUE_INFO_MSS_MASK		0x3FFFU
+#define SQ_CTRL_QUEUE_INFO_SCTP_MASK		0x1U
+#define SQ_CTRL_QUEUE_INFO_UC_MASK		0x1U
+#define SQ_CTRL_QUEUE_INFO_PRI_MASK		0x7U
+
+#define SQ_CTRL_QUEUE_INFO_SET(val, member)	\
+	(((u32)(val) & SQ_CTRL_QUEUE_INFO_##member##_MASK) <<	\
+			SQ_CTRL_QUEUE_INFO_##member##_SHIFT)
+
+#define SQ_CTRL_QUEUE_INFO_GET(val, member)	\
+	(((val) >> SQ_CTRL_QUEUE_INFO_##member##_SHIFT) &	\
+			SQ_CTRL_QUEUE_INFO_##member##_MASK)
+
+#define SQ_CTRL_QUEUE_INFO_CLEAR(val, member)	\
+	((val) & (~(SQ_CTRL_QUEUE_INFO_##member##_MASK << \
+			SQ_CTRL_QUEUE_INFO_##member##_SHIFT)))
+
+#define	SQ_TASK_INFO0_L2HDR_LEN_SHIFT		0
+#define	SQ_TASK_INFO0_L4OFFLOAD_SHIFT		8
+#define	SQ_TASK_INFO0_INNER_L3TYPE_SHIFT	10
+#define	SQ_TASK_INFO0_VLAN_OFFLOAD_SHIFT	12
+#define	SQ_TASK_INFO0_PARSE_FLAG_SHIFT		13
+#define	SQ_TASK_INFO0_UFO_AVD_SHIFT		14
+#define	SQ_TASK_INFO0_TSO_UFO_SHIFT		15
+#define SQ_TASK_INFO0_VLAN_TAG_SHIFT		16
+
+#define	SQ_TASK_INFO0_L2HDR_LEN_MASK		0xFFU
+#define	SQ_TASK_INFO0_L4OFFLOAD_MASK		0x3U
+#define	SQ_TASK_INFO0_INNER_L3TYPE_MASK		0x3U
+#define	SQ_TASK_INFO0_VLAN_OFFLOAD_MASK		0x1U
+#define	SQ_TASK_INFO0_PARSE_FLAG_MASK		0x1U
+#define	SQ_TASK_INFO0_UFO_AVD_MASK		0x1U
+#define SQ_TASK_INFO0_TSO_UFO_MASK		0x1U
+#define SQ_TASK_INFO0_VLAN_TAG_MASK		0xFFFFU
+
+#define SQ_TASK_INFO0_SET(val, member)			\
+	(((u32)(val) & SQ_TASK_INFO0_##member##_MASK) <<	\
+			SQ_TASK_INFO0_##member##_SHIFT)
+
+#define	SQ_TASK_INFO1_MD_TYPE_SHIFT		8
+#define SQ_TASK_INFO1_INNER_L4LEN_SHIFT		16
+#define SQ_TASK_INFO1_INNER_L3LEN_SHIFT		24
+
+#define	SQ_TASK_INFO1_MD_TYPE_MASK		0xFFU
+#define SQ_TASK_INFO1_INNER_L4LEN_MASK		0xFFU
+#define SQ_TASK_INFO1_INNER_L3LEN_MASK		0xFFU
+
+#define SQ_TASK_INFO1_SET(val, member)			\
+	(((val) & SQ_TASK_INFO1_##member##_MASK) <<	\
+			SQ_TASK_INFO1_##member##_SHIFT)
+
+#define SQ_TASK_INFO2_TUNNEL_L4LEN_SHIFT	0
+#define SQ_TASK_INFO2_OUTER_L3LEN_SHIFT		8
+#define SQ_TASK_INFO2_TUNNEL_L4TYPE_SHIFT	16
+#define SQ_TASK_INFO2_OUTER_L3TYPE_SHIFT	24
+
+#define SQ_TASK_INFO2_TUNNEL_L4LEN_MASK		0xFFU
+#define SQ_TASK_INFO2_OUTER_L3LEN_MASK		0xFFU
+#define SQ_TASK_INFO2_TUNNEL_L4TYPE_MASK	0x7U
+#define SQ_TASK_INFO2_OUTER_L3TYPE_MASK		0x3U
+
+#define SQ_TASK_INFO2_SET(val, member)			\
+	(((val) & SQ_TASK_INFO2_##member##_MASK) <<	\
+			SQ_TASK_INFO2_##member##_SHIFT)
+
+#define	SQ_TASK_INFO4_L2TYPE_SHIFT		31
+
+#define	SQ_TASK_INFO4_L2TYPE_MASK		0x1U
+
+#define SQ_TASK_INFO4_SET(val, member)		\
+	(((u32)(val) & SQ_TASK_INFO4_##member##_MASK) << \
+			SQ_TASK_INFO4_##member##_SHIFT)
+
+/* SQ_DB */
+#define SQ_DB_OFF				0x00000800
+#define SQ_DB_INFO_HI_PI_SHIFT			0
+#define SQ_DB_INFO_QID_SHIFT			8
+#define SQ_DB_INFO_CFLAG_SHIFT			23
+#define SQ_DB_INFO_COS_SHIFT			24
+#define SQ_DB_INFO_TYPE_SHIFT			27
+
+#define SQ_DB_INFO_HI_PI_MASK			0xFFU
+#define SQ_DB_INFO_QID_MASK			0x3FFU
+#define SQ_DB_INFO_CFLAG_MASK			0x1U
+#define SQ_DB_INFO_COS_MASK			0x7U
+#define SQ_DB_INFO_TYPE_MASK			0x1FU
+#define SQ_DB_INFO_SET(val, member)		\
+	(((u32)(val) & SQ_DB_INFO_##member##_MASK) <<	\
+			SQ_DB_INFO_##member##_SHIFT)
+
+#define SQ_DB					1
+#define SQ_CFLAG_DP				0	/* CFLAG_DATA_PATH */
+
+#define SQ_DB_PI_LOW_MASK			0xFF
+#define SQ_DB_PI_LOW(pi)			((pi) & SQ_DB_PI_LOW_MASK)
+#define SQ_DB_PI_HI_SHIFT			8
+#define SQ_DB_PI_HIGH(pi)			((pi) >> SQ_DB_PI_HI_SHIFT)
+#define SQ_DB_ADDR(sq, pi)		\
+	((u64 *)((u8 __iomem *)((sq)->db_addr) + SQ_DB_OFF) + SQ_DB_PI_LOW(pi))
+
+/* txq wq operations */
+#define HINIC_GET_SQ_WQE_MASK(txq)		((txq)->wq->mask)
+
+#define HINIC_GET_SQ_HW_CI(txq)	\
+	((be16_to_cpu(*(txq)->cons_idx_addr)) & HINIC_GET_SQ_WQE_MASK(txq))
+
+#define HINIC_GET_SQ_LOCAL_CI(txq)	\
+	(((txq)->wq->cons_idx) & HINIC_GET_SQ_WQE_MASK(txq))
+
+#define HINIC_UPDATE_SQ_LOCAL_CI(txq, wqebb_cnt)	\
+	do {						\
+		(txq)->wq->cons_idx += wqebb_cnt;	\
+		(txq)->wq->delta += wqebb_cnt;		\
+	} while (0)
+
+#define HINIC_GET_SQ_FREE_WQEBBS(txq)	((txq)->wq->delta - 1)
+
+#define HINIC_IS_SQ_EMPTY(txq)	(((txq)->wq->delta) == ((txq)->q_depth))
+
+#define BUF_DESC_SIZE_SHIFT		4
+
+#define HINIC_SQ_WQE_SIZE(num_sge)		\
+	(sizeof(struct hinic_sq_ctrl) + sizeof(struct hinic_sq_task) +  \
+			(unsigned int)((num_sge) << BUF_DESC_SIZE_SHIFT))
+
+#define HINIC_SQ_WQEBB_CNT(num_sge)	\
+	(int)(ALIGN(HINIC_SQ_WQE_SIZE((u32)num_sge), \
+			HINIC_SQ_WQEBB_SIZE) >> HINIC_SQ_WQEBB_SHIFT)
+
+
+static inline void hinic_sq_wqe_cpu_to_be32(void *data, int nr_wqebb)
+{
+	int i;
+#ifndef __ARM64_NEON__
+	__m128i *wqe_line = (__m128i *)data;
+	__m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
+					 11, 4, 5, 6, 7, 0, 1, 2, 3);
+
+	for (i = 0; i < nr_wqebb; i++) {
+		/* convert 64B wqebb using 4 SSE instructions */
+		wqe_line[0] = _mm_shuffle_epi8(wqe_line[0], shuf_mask);
+		wqe_line[1] = _mm_shuffle_epi8(wqe_line[1], shuf_mask);
+		wqe_line[2] = _mm_shuffle_epi8(wqe_line[2], shuf_mask);
+		wqe_line[3] = _mm_shuffle_epi8(wqe_line[3], shuf_mask);
+		wqe_line += 4;
+	}
+#else
+	uint8x16_t *wqe_line = (uint8x16_t *)data;
+	const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
+					9, 8, 15, 14, 13, 12};
+
+	for (i = 0; i < nr_wqebb; i++) {
+		wqe_line[0] = vqtbl1q_u8(wqe_line[0], shuf_mask);
+		wqe_line[1] = vqtbl1q_u8(wqe_line[1], shuf_mask);
+		wqe_line[2] = vqtbl1q_u8(wqe_line[2], shuf_mask);
+		wqe_line[3] = vqtbl1q_u8(wqe_line[3], shuf_mask);
+		wqe_line += 4;
+	}
+#endif
+}
+
+static inline void hinic_sge_cpu_to_be32(void *data, int nr_sge)
+{
+	int i;
+#ifndef __ARM64_NEON__
+	__m128i *sge_line = (__m128i *)data;
+	__m128i shuf_mask = _mm_set_epi8(12, 13, 14, 15, 8, 9, 10,
+					 11, 4, 5, 6, 7, 0, 1, 2, 3);
+
+	for (i = 0; i < nr_sge; i++) {
+		/* convert 16B sge using 1 SSE instructions */
+		*sge_line = _mm_shuffle_epi8(*sge_line, shuf_mask);
+		sge_line++;
+	}
+#else
+	uint8x16_t *sge_line = (uint8x16_t *)data;
+	const uint8x16_t shuf_mask = {3, 2, 1, 0, 7, 6, 5, 4, 11, 10,
+					9, 8, 15, 14, 13, 12};
+
+	for (i = 0; i < nr_sge; i++) {
+		*sge_line = vqtbl1q_u8(*sge_line, shuf_mask);
+		sge_line++;
+	}
+#endif
+}
+
+void hinic_txq_get_stats(struct hinic_txq *txq, struct hinic_txq_stats *stats)
+{
+	if (!txq || !stats) {
+		PMD_DRV_LOG(ERR, "Txq or stats is NULL");
+		return;
+	}
+
+	memcpy(stats, &txq->txq_stats, sizeof(txq->txq_stats));
+}
+
+void hinic_txq_stats_reset(struct hinic_txq *txq)
+{
+	struct hinic_txq_stats *txq_stats;
+
+	if (txq == NULL)
+		return;
+
+	txq_stats = &txq->txq_stats;
+	memset(txq_stats, 0, sizeof(*txq_stats));
+}
+
+static inline struct rte_mbuf *hinic_copy_tx_mbuf(struct hinic_nic_dev *nic_dev,
+						  struct rte_mbuf *mbuf,
+						  u16 sge_cnt)
+{
+	struct rte_mbuf *dst_mbuf;
+	u32 offset = 0;
+	u16 i;
+
+	if (unlikely(!nic_dev->cpy_mpool))
+		return NULL;
+
+	dst_mbuf = rte_pktmbuf_alloc(nic_dev->cpy_mpool);
+	if (unlikely(!dst_mbuf))
+		return NULL;
+
+	dst_mbuf->data_off = 0;
+	for (i = 0; i < sge_cnt; i++) {
+		rte_memcpy((char *)dst_mbuf->buf_addr + offset,
+			   (char *)mbuf->buf_addr + mbuf->data_off,
+			   mbuf->data_len);
+		dst_mbuf->data_len += mbuf->data_len;
+		offset += mbuf->data_len;
+		mbuf = mbuf->next;
+	}
+
+	return dst_mbuf;
+}
+
+static inline bool hinic_mbuf_dma_map_sge(struct hinic_txq *txq,
+					  struct rte_mbuf *mbuf,
+					  struct hinic_sq_bufdesc *sges,
+					  struct hinic_wqe_info *sqe_info)
+{
+	dma_addr_t dma_addr;
+	u16 i, around_sges;
+	u16 nb_segs = sqe_info->sge_cnt - sqe_info->cpy_mbuf_cnt;
+	u16 real_nb_segs = mbuf->nb_segs;
+	struct hinic_sq_bufdesc *sge_idx = sges;
+
+	if (unlikely(sqe_info->around)) {
+		/* parts of wqe is in sq bottom while parts
+		 * of wqe is in sq head
+		 */
+		i = 0;
+		for (sge_idx = sges; (u64)sge_idx <= txq->sq_bot_sge_addr;
+		     sge_idx++) {
+			dma_addr = rte_mbuf_data_iova(mbuf);
+			hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
+				      mbuf->data_len);
+			mbuf = mbuf->next;
+			i++;
+		}
+
+		around_sges = nb_segs - i;
+		sge_idx = (struct hinic_sq_bufdesc *)
+				((void *)txq->sq_head_addr);
+		for (; i < nb_segs; i++) {
+			dma_addr = rte_mbuf_data_iova(mbuf);
+			hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
+				      mbuf->data_len);
+			mbuf = mbuf->next;
+			sge_idx++;
+		}
+
+		/* covert sges at head to big endian */
+		hinic_sge_cpu_to_be32((void *)txq->sq_head_addr, around_sges);
+	} else {
+		/* wqe is in continuous space */
+		for (i = 0; i < nb_segs; i++) {
+			dma_addr = rte_mbuf_data_iova(mbuf);
+			hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
+				      mbuf->data_len);
+			mbuf = mbuf->next;
+			sge_idx++;
+		}
+	}
+
+	/* for now: support non-tso over 17 sge, copy the last 2 mbuf */
+	if (unlikely(sqe_info->cpy_mbuf_cnt != 0)) {
+		/* copy invalid mbuf segs to a valid buffer, lost performance */
+		txq->txq_stats.cpy_pkts += 1;
+		mbuf = hinic_copy_tx_mbuf(txq->nic_dev, mbuf,
+					  real_nb_segs - nb_segs);
+		if (unlikely(!mbuf))
+			return false;
+
+		txq->tx_info[sqe_info->pi].cpy_mbuf = mbuf;
+
+		/* deal with the last mbuf */
+		dma_addr = rte_mbuf_data_iova(mbuf);
+		hinic_set_sge((struct hinic_sge *)sge_idx, dma_addr,
+			      mbuf->data_len);
+		if (unlikely(sqe_info->around))
+			hinic_sge_cpu_to_be32((void *)sge_idx, 1);
+	}
+
+	return true;
+}
+
+static inline void hinic_fill_sq_wqe_header(struct hinic_sq_ctrl *ctrl,
+					    u32 queue_info, int nr_descs,
+					    u8 owner)
+{
+	u32 ctrl_size, task_size, bufdesc_size;
+
+	ctrl_size = SIZE_8BYTES(sizeof(struct hinic_sq_ctrl));
+	task_size = SIZE_8BYTES(sizeof(struct hinic_sq_task));
+	bufdesc_size = HINIC_BUF_DESC_SIZE(nr_descs);
+
+	ctrl->ctrl_fmt = SQ_CTRL_SET(bufdesc_size, BUFDESC_SECT_LEN) |
+			SQ_CTRL_SET(task_size, TASKSECT_LEN)	|
+			SQ_CTRL_SET(SQ_NORMAL_WQE, DATA_FORMAT)	|
+			SQ_CTRL_SET(ctrl_size, LEN)		|
+			SQ_CTRL_SET(owner, OWNER);
+
+	ctrl->queue_info = queue_info;
+	ctrl->queue_info |= SQ_CTRL_QUEUE_INFO_SET(1U, UC);
+
+	if (!SQ_CTRL_QUEUE_INFO_GET(ctrl->queue_info, MSS)) {
+		ctrl->queue_info |=
+			SQ_CTRL_QUEUE_INFO_SET(TX_MSS_DEFAULT, MSS);
+	} else if (SQ_CTRL_QUEUE_INFO_GET(ctrl->queue_info, MSS) < TX_MSS_MIN) {
+		/* mss should not be less than 80 */
+		ctrl->queue_info =
+				SQ_CTRL_QUEUE_INFO_CLEAR(ctrl->queue_info, MSS);
+		ctrl->queue_info |= SQ_CTRL_QUEUE_INFO_SET(TX_MSS_MIN, MSS);
+	}
+}
+
+static inline bool hinic_is_tso_sge_valid(struct rte_mbuf *mbuf,
+					  struct hinic_tx_offload_info
+					  *poff_info,
+					  struct hinic_wqe_info *sqe_info)
+{
+	u32 total_len, limit_len, checked_len, left_len;
+	u32 i, first_mss_sges, left_sges;
+	struct rte_mbuf *mbuf_head, *mbuf_pre;
+
+	left_sges = mbuf->nb_segs;
+	mbuf_head = mbuf;
+
+	/* tso sge number validation */
+	if (unlikely(left_sges >= HINIC_NONTSO_PKT_MAX_SGE)) {
+		checked_len = 0;
+		limit_len = mbuf->tso_segsz + poff_info->payload_offset;
+		first_mss_sges = HINIC_NONTSO_PKT_MAX_SGE;
+
+		/* each continues 17 mbufs segmust do one check */
+		while (left_sges >= HINIC_NONTSO_PKT_MAX_SGE) {
+			/* total len of first 16 mbufs must equal
+			 * or more than limit_len
+			 */
+			total_len = 0;
+			for (i = 0; i < first_mss_sges; i++) {
+				total_len += mbuf->data_len;
+				mbuf_pre = mbuf;
+				mbuf = mbuf->next;
+				if (total_len >= limit_len) {
+					limit_len = mbuf_head->tso_segsz;
+					break;
+				}
+			}
+
+			checked_len += total_len;
+
+			/* try to copy if not valid */
+			if (unlikely(first_mss_sges == i)) {
+				left_sges -= first_mss_sges;
+				checked_len -= mbuf_pre->data_len;
+
+				left_len = mbuf_head->pkt_len - checked_len;
+				if (left_len > HINIC_COPY_MBUF_SIZE)
+					return false;
+
+				sqe_info->sge_cnt = mbuf_head->nb_segs -
+							left_sges;
+				sqe_info->cpy_mbuf_cnt = 1;
+
+				return true;
+			}
+			first_mss_sges = (HINIC_NONTSO_PKT_MAX_SGE - 1);
+
+			/* continue next 16 mbufs */
+			left_sges -= (i + 1);
+		} /* end of while */
+	}
+
+	sqe_info->sge_cnt = mbuf_head->nb_segs;
+	return true;
+}
+
+static inline void
+hinic_set_l4_csum_info(struct hinic_sq_task *task,
+		u32 *queue_info, struct hinic_tx_offload_info *poff_info)
+{
+	u32 tcp_udp_cs, sctp;
+	u16 l2hdr_len;
+
+	sctp = 0;
+	if (unlikely(poff_info->inner_l4_type == SCTP_OFFLOAD_ENABLE))
+		sctp = 1;
+
+	tcp_udp_cs = poff_info->inner_l4_tcp_udp;
+
+	if (poff_info->tunnel_type == TUNNEL_UDP_NO_CSUM) {
+		l2hdr_len =  poff_info->outer_l2_len;
+
+		task->pkt_info2 |=
+		SQ_TASK_INFO2_SET(poff_info->outer_l3_type, OUTER_L3TYPE) |
+		SQ_TASK_INFO2_SET(poff_info->outer_l3_len, OUTER_L3LEN);
+		task->pkt_info2 |=
+		SQ_TASK_INFO2_SET(poff_info->tunnel_type, TUNNEL_L4TYPE) |
+		SQ_TASK_INFO2_SET(poff_info->tunnel_length, TUNNEL_L4LEN);
+	} else {
+		l2hdr_len = poff_info->inner_l2_len;
+	}
+
+	task->pkt_info0 |= SQ_TASK_INFO0_SET(l2hdr_len, L2HDR_LEN);
+	task->pkt_info1 |=
+		SQ_TASK_INFO1_SET(poff_info->inner_l3_len, INNER_L3LEN);
+	task->pkt_info0 |=
+		SQ_TASK_INFO0_SET(poff_info->inner_l3_type, INNER_L3TYPE);
+	task->pkt_info1 |=
+		SQ_TASK_INFO1_SET(poff_info->inner_l4_len, INNER_L4LEN);
+	task->pkt_info0 |=
+		SQ_TASK_INFO0_SET(poff_info->inner_l4_type, L4OFFLOAD);
+	*queue_info |=
+		SQ_CTRL_QUEUE_INFO_SET(poff_info->payload_offset, PLDOFF) |
+		SQ_CTRL_QUEUE_INFO_SET(tcp_udp_cs, TCPUDP_CS) |
+		SQ_CTRL_QUEUE_INFO_SET(sctp, SCTP);
+}
+
+static inline void
+hinic_set_tso_info(struct hinic_sq_task *task,
+		u32 *queue_info, struct rte_mbuf *mbuf,
+		struct hinic_tx_offload_info *poff_info)
+{
+	hinic_set_l4_csum_info(task, queue_info, poff_info);
+
+	/* wqe for tso */
+	task->pkt_info0 |=
+		SQ_TASK_INFO0_SET(poff_info->inner_l3_type, INNER_L3TYPE);
+	task->pkt_info0 |= SQ_TASK_INFO0_SET(TSO_ENABLE, TSO_UFO);
+	*queue_info |= SQ_CTRL_QUEUE_INFO_SET(TSO_ENABLE, TSO);
+	/* qsf was initialized in prepare_sq_wqe */
+	*queue_info = SQ_CTRL_QUEUE_INFO_CLEAR(*queue_info, MSS);
+	*queue_info |= SQ_CTRL_QUEUE_INFO_SET(mbuf->tso_segsz, MSS);
+}
+
+static inline void
+hinic_set_vlan_tx_offload(struct hinic_sq_task *task,
+			u32 *queue_info, u16 vlan_tag, u16 vlan_pri)
+{
+	task->pkt_info0 |= SQ_TASK_INFO0_SET(vlan_tag, VLAN_TAG) |
+				SQ_TASK_INFO0_SET(1U, VLAN_OFFLOAD);
+
+	*queue_info |= SQ_CTRL_QUEUE_INFO_SET(vlan_pri, PRI);
+}
+
+static inline void
+hinic_fill_tx_offload_info(struct rte_mbuf *mbuf,
+		struct hinic_sq_task *task, u32 *queue_info,
+		struct hinic_tx_offload_info *tx_off_info)
+{
+	u16 vlan_tag;
+	uint64_t ol_flags = mbuf->ol_flags;
+
+	/* clear DW0~2 of task section for offload */
+	task->pkt_info0 = 0;
+	task->pkt_info1 = 0;
+	task->pkt_info2 = 0;
+
+	/* Base VLAN */
+	if (unlikely(ol_flags & PKT_TX_VLAN_PKT)) {
+		vlan_tag = mbuf->vlan_tci;
+		hinic_set_vlan_tx_offload(task, queue_info, vlan_tag,
+					  vlan_tag >> VLAN_PRIO_SHIFT);
+	}
+
+	/* non checksum or tso */
+	if (unlikely(!(ol_flags & HINIC_TX_CKSUM_OFFLOAD_MASK)))
+		return;
+
+	if ((ol_flags & PKT_TX_TCP_SEG))
+		/* set tso info for task and qsf */
+		hinic_set_tso_info(task, queue_info, mbuf, tx_off_info);
+	else /* just support l4 checksum offload */
+		hinic_set_l4_csum_info(task, queue_info, tx_off_info);
+}
+
+static inline void hinic_xmit_mbuf_cleanup(struct hinic_txq *txq)
+{
+	struct hinic_tx_info *tx_info;
+	struct rte_mbuf *mbuf, *m, *mbuf_free[HINIC_MAX_TX_FREE_BULK];
+	int i, nb_free = 0;
+	u16 hw_ci, sw_ci, sq_mask;
+	int wqebb_cnt = 0;
+
+	hw_ci = HINIC_GET_SQ_HW_CI(txq);
+	sw_ci = HINIC_GET_SQ_LOCAL_CI(txq);
+	sq_mask = HINIC_GET_SQ_WQE_MASK(txq);
+
+	for (i = 0; i < txq->tx_free_thresh; ++i) {
+		tx_info = &txq->tx_info[sw_ci];
+		if (hw_ci == sw_ci ||
+			(((hw_ci - sw_ci) & sq_mask) < tx_info->wqebb_cnt))
+			break;
+
+		sw_ci = (sw_ci + tx_info->wqebb_cnt) & sq_mask;
+
+		if (unlikely(tx_info->cpy_mbuf != NULL)) {
+			rte_pktmbuf_free(tx_info->cpy_mbuf);
+			tx_info->cpy_mbuf = NULL;
+		}
+
+		wqebb_cnt += tx_info->wqebb_cnt;
+		mbuf = tx_info->mbuf;
+
+		if (likely(mbuf->nb_segs == 1)) {
+			m = rte_pktmbuf_prefree_seg(mbuf);
+			tx_info->mbuf = NULL;
+
+			if (unlikely(m == NULL))
+				continue;
+
+			mbuf_free[nb_free++] = m;
+			if (unlikely(m->pool != mbuf_free[0]->pool ||
+				nb_free >= HINIC_MAX_TX_FREE_BULK)) {
+				rte_mempool_put_bulk(mbuf_free[0]->pool,
+					(void **)mbuf_free, (nb_free - 1));
+				nb_free = 0;
+				mbuf_free[nb_free++] = m;
+			}
+		} else {
+			rte_pktmbuf_free(mbuf);
+			tx_info->mbuf = NULL;
+		}
+	}
+
+	if (nb_free > 0)
+		rte_mempool_put_bulk(mbuf_free[0]->pool, (void **)mbuf_free,
+				     nb_free);
+
+	HINIC_UPDATE_SQ_LOCAL_CI(txq, wqebb_cnt);
+}
+
+static inline struct hinic_sq_wqe *
+hinic_get_sq_wqe(struct hinic_txq *txq, int wqebb_cnt,
+		struct hinic_wqe_info *wqe_info)
+{
+	u32 cur_pi, end_pi;
+	u16 remain_wqebbs;
+	struct hinic_sq *sq = txq->sq;
+	struct hinic_wq *wq = txq->wq;
+
+	/* record current pi */
+	cur_pi = MASKED_WQE_IDX(wq, wq->prod_idx);
+	end_pi = cur_pi + wqebb_cnt;
+
+	/* update next pi and delta */
+	wq->prod_idx += wqebb_cnt;
+	wq->delta -= wqebb_cnt;
+
+	/* return current pi and owner */
+	wqe_info->pi = cur_pi;
+	wqe_info->owner = sq->owner;
+	wqe_info->around = 0;
+	wqe_info->seq_wqebbs = wqebb_cnt;
+
+	if (unlikely(end_pi >= txq->q_depth)) {
+		/* update owner of next prod_idx */
+		sq->owner = !sq->owner;
+
+		/* turn around to head */
+		if (unlikely(end_pi > txq->q_depth)) {
+			wqe_info->around = 1;
+			remain_wqebbs = txq->q_depth - cur_pi;
+			wqe_info->seq_wqebbs = remain_wqebbs;
+		}
+	}
+
+	return (struct hinic_sq_wqe *)WQ_WQE_ADDR(wq, cur_pi);
+}
+
+static inline int
+hinic_validate_tx_offload(const struct rte_mbuf *m)
+{
+	uint64_t ol_flags = m->ol_flags;
+	uint64_t inner_l3_offset = m->l2_len;
+
+	/* just support vxlan offload */
+	if ((ol_flags & PKT_TX_TUNNEL_MASK) &&
+	    !(ol_flags & PKT_TX_TUNNEL_VXLAN))
+		return -ENOTSUP;
+
+	if (ol_flags & PKT_TX_OUTER_IP_CKSUM)
+		inner_l3_offset += m->outer_l2_len + m->outer_l3_len;
+
+	/* Headers are fragmented */
+	if (rte_pktmbuf_data_len(m) < inner_l3_offset + m->l3_len + m->l4_len)
+		return -ENOTSUP;
+
+	/* IP checksum can be counted only for IPv4 packet */
+	if ((ol_flags & PKT_TX_IP_CKSUM) && (ol_flags & PKT_TX_IPV6))
+		return -EINVAL;
+
+	/* IP type not set when required */
+	if (ol_flags & (PKT_TX_L4_MASK | PKT_TX_TCP_SEG)) {
+		if (!(ol_flags & (PKT_TX_IPV4 | PKT_TX_IPV6)))
+			return -EINVAL;
+	}
+
+	/* Check requirements for TSO packet */
+	if (ol_flags & PKT_TX_TCP_SEG) {
+		if (m->tso_segsz == 0 ||
+			((ol_flags & PKT_TX_IPV4) &&
+			!(ol_flags & PKT_TX_IP_CKSUM)))
+			return -EINVAL;
+	}
+
+	/* PKT_TX_OUTER_IP_CKSUM set for non outer IPv4 packet. */
+	if ((ol_flags & PKT_TX_OUTER_IP_CKSUM) &&
+		!(ol_flags & PKT_TX_OUTER_IPV4))
+		return -EINVAL;
+
+	return 0;
+}
+
+static inline uint16_t
+hinic_ipv4_phdr_cksum(const struct rte_ipv4_hdr *ipv4_hdr, uint64_t ol_flags)
+{
+	struct ipv4_psd_header {
+		uint32_t src_addr; /* IP address of source host. */
+		uint32_t dst_addr; /* IP address of destination host. */
+		uint8_t  zero;     /* zero. */
+		uint8_t  proto;    /* L4 protocol type. */
+		uint16_t len;      /* L4 length. */
+	} psd_hdr;
+	uint8_t ihl;
+
+	psd_hdr.src_addr = ipv4_hdr->src_addr;
+	psd_hdr.dst_addr = ipv4_hdr->dst_addr;
+	psd_hdr.zero = 0;
+	psd_hdr.proto = ipv4_hdr->next_proto_id;
+	if (ol_flags & PKT_TX_TCP_SEG) {
+		psd_hdr.len = 0;
+	} else {
+		/* ipv4_hdr->version_ihl is uint8_t big endian, ihl locates
+		 * lower 4 bits and unit is 4 bytes
+		 */
+		ihl = (ipv4_hdr->version_ihl & 0xF) << 2;
+		psd_hdr.len =
+		rte_cpu_to_be_16(rte_be_to_cpu_16(ipv4_hdr->total_length) -
+				 ihl);
+	}
+	return rte_raw_cksum(&psd_hdr, sizeof(psd_hdr));
+}
+
+static inline uint16_t
+hinic_ipv6_phdr_cksum(const struct rte_ipv6_hdr *ipv6_hdr, uint64_t ol_flags)
+{
+	uint32_t sum;
+	struct {
+		uint32_t len;   /* L4 length. */
+		uint32_t proto; /* L4 protocol - top 3 bytes must be zero */
+	} psd_hdr;
+
+	psd_hdr.proto = (ipv6_hdr->proto << 24);
+	if (ol_flags & PKT_TX_TCP_SEG)
+		psd_hdr.len = 0;
+	else
+		psd_hdr.len = ipv6_hdr->payload_len;
+
+	sum = __rte_raw_cksum(ipv6_hdr->src_addr,
+		sizeof(ipv6_hdr->src_addr) + sizeof(ipv6_hdr->dst_addr), 0);
+	sum = __rte_raw_cksum(&psd_hdr, sizeof(psd_hdr), sum);
+	return __rte_raw_cksum_reduce(sum);
+}
+
+static inline int
+hinic_tx_offload_pkt_prepare(struct rte_mbuf *m,
+				struct hinic_tx_offload_info *off_info)
+{
+	struct rte_ipv4_hdr *ipv4_hdr;
+	struct rte_ipv6_hdr *ipv6_hdr;
+	struct rte_tcp_hdr *tcp_hdr;
+	struct rte_udp_hdr *udp_hdr;
+	struct rte_ether_hdr *eth_hdr;
+	struct rte_vlan_hdr *vlan_hdr;
+	u16 eth_type = 0;
+	uint64_t inner_l3_offset = m->l2_len;
+	uint64_t ol_flags = m->ol_flags;
+
+	/* Does packet set any of available offloads */
+	if (!(ol_flags & HINIC_TX_CKSUM_OFFLOAD_MASK))
+		return 0;
+
+	if (unlikely(hinic_validate_tx_offload(m)))
+		return -EINVAL;
+
+	if ((ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
+			(ol_flags & PKT_TX_OUTER_IPV6) ||
+			(ol_flags & PKT_TX_TUNNEL_VXLAN)) {
+		inner_l3_offset += m->outer_l2_len + m->outer_l3_len;
+		off_info->outer_l2_len = m->outer_l2_len;
+		off_info->outer_l3_len = m->outer_l3_len;
+		/* just support vxlan tunneling pkt */
+		off_info->inner_l2_len = m->l2_len - VXLANLEN -
+						sizeof(struct rte_udp_hdr);
+		off_info->inner_l3_len = m->l3_len;
+		off_info->inner_l4_len = m->l4_len;
+		off_info->tunnel_length = m->l2_len;
+		off_info->payload_offset = m->outer_l2_len +
+				m->outer_l3_len + m->l2_len + m->l3_len;
+		off_info->tunnel_type = TUNNEL_UDP_NO_CSUM;
+	} else {
+		off_info->inner_l2_len = m->l2_len;
+		off_info->inner_l3_len = m->l3_len;
+		off_info->inner_l4_len = m->l4_len;
+		off_info->tunnel_type = NOT_TUNNEL;
+		off_info->payload_offset = m->l2_len + m->l3_len;
+	}
+
+	if (((ol_flags & PKT_TX_L4_MASK) != PKT_TX_SCTP_CKSUM) &&
+	    ((ol_flags & PKT_TX_L4_MASK) != PKT_TX_UDP_CKSUM))
+		off_info->payload_offset += m->l4_len;
+
+	/* invalid udp or tcp header */
+	if (unlikely(off_info->payload_offset > MAX_PLD_OFFSET))
+		return -EINVAL;
+
+	/* Process outter udp pseudo-header checksum */
+	if ((ol_flags & PKT_TX_TUNNEL_VXLAN) && ((ol_flags & PKT_TX_TCP_SEG) ||
+			(ol_flags & PKT_TX_OUTER_IP_CKSUM) ||
+			(ol_flags & PKT_TX_OUTER_IPV6))) {
+		off_info->tunnel_type = TUNNEL_UDP_CSUM;
+
+		/* inner_l4_tcp_udp csum should be setted to calculate outter
+		 * udp checksum when vxlan packets without inner l3 and l4
+		 */
+		off_info->inner_l4_tcp_udp = 1;
+
+		eth_hdr = rte_pktmbuf_mtod(m, struct rte_ether_hdr *);
+		eth_type = rte_be_to_cpu_16(eth_hdr->ether_type);
+
+		if (eth_type == RTE_ETHER_TYPE_VLAN) {
+			vlan_hdr = (struct rte_vlan_hdr *)(eth_hdr + 1);
+			eth_type = rte_be_to_cpu_16(vlan_hdr->eth_proto);
+		}
+
+		if (eth_type == RTE_ETHER_TYPE_IPV4) {
+			ipv4_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
+						m->outer_l2_len);
+			off_info->outer_l3_type = IPV4_PKT_WITH_CHKSUM_OFFLOAD;
+			ipv4_hdr->hdr_checksum = 0;
+
+			udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr +
+							m->outer_l3_len);
+			udp_hdr->dgram_cksum =
+				hinic_ipv4_phdr_cksum(ipv4_hdr, ol_flags);
+		} else if (eth_type == RTE_ETHER_TYPE_IPV6) {
+			off_info->outer_l3_type = IPV6_PKT;
+			ipv6_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
+						m->outer_l2_len);
+
+			udp_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
+						(m->outer_l2_len +
+						m->outer_l3_len));
+			udp_hdr->dgram_cksum =
+				hinic_ipv6_phdr_cksum(ipv6_hdr, ol_flags);
+		}
+	}
+
+	if (ol_flags & PKT_TX_IPV4)
+		off_info->inner_l3_type = (ol_flags & PKT_TX_IP_CKSUM) ?
+					IPV4_PKT_WITH_CHKSUM_OFFLOAD :
+					IPV4_PKT_NO_CHKSUM_OFFLOAD;
+	else if (ol_flags & PKT_TX_IPV6)
+		off_info->inner_l3_type = IPV6_PKT;
+
+	/* Process the pseudo-header checksum */
+	if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_UDP_CKSUM) {
+		if (ol_flags & PKT_TX_IPV4) {
+			ipv4_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
+						inner_l3_offset);
+
+			if (ol_flags & PKT_TX_IP_CKSUM)
+				ipv4_hdr->hdr_checksum = 0;
+
+			udp_hdr = (struct rte_udp_hdr *)((char *)ipv4_hdr +
+								m->l3_len);
+			udp_hdr->dgram_cksum =
+				hinic_ipv4_phdr_cksum(ipv4_hdr, ol_flags);
+		} else {
+			ipv6_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
+						inner_l3_offset);
+
+			udp_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_udp_hdr *,
+						(inner_l3_offset + m->l3_len));
+			udp_hdr->dgram_cksum =
+				hinic_ipv6_phdr_cksum(ipv6_hdr, ol_flags);
+		}
+
+		off_info->inner_l4_type = UDP_OFFLOAD_ENABLE;
+		off_info->inner_l4_tcp_udp = 1;
+		off_info->inner_l4_len = sizeof(struct rte_udp_hdr);
+	} else if (((ol_flags & PKT_TX_L4_MASK) == PKT_TX_TCP_CKSUM) ||
+			(ol_flags & PKT_TX_TCP_SEG)) {
+		if (ol_flags & PKT_TX_IPV4) {
+			ipv4_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv4_hdr *,
+						inner_l3_offset);
+
+			if (ol_flags & PKT_TX_IP_CKSUM)
+				ipv4_hdr->hdr_checksum = 0;
+
+			/* non-TSO tcp */
+			tcp_hdr = (struct rte_tcp_hdr *)((char *)ipv4_hdr +
+								m->l3_len);
+			tcp_hdr->cksum =
+				hinic_ipv4_phdr_cksum(ipv4_hdr, ol_flags);
+		} else {
+			ipv6_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_ipv6_hdr *,
+						inner_l3_offset);
+			/* non-TSO tcp */
+			tcp_hdr =
+			rte_pktmbuf_mtod_offset(m, struct rte_tcp_hdr *,
+						(inner_l3_offset + m->l3_len));
+			tcp_hdr->cksum =
+				hinic_ipv6_phdr_cksum(ipv6_hdr, ol_flags);
+		}
+
+		off_info->inner_l4_type = TCP_OFFLOAD_ENABLE;
+		off_info->inner_l4_tcp_udp = 1;
+	} else if ((ol_flags & PKT_TX_L4_MASK) == PKT_TX_SCTP_CKSUM) {
+		off_info->inner_l4_type = SCTP_OFFLOAD_ENABLE;
+		off_info->inner_l4_tcp_udp = 0;
+		off_info->inner_l4_len = sizeof(struct rte_sctp_hdr);
+	}
+
+	return 0;
+}
+
+static inline bool hinic_get_sge_txoff_info(struct rte_mbuf *mbuf_pkt,
+					    struct hinic_wqe_info *sqe_info,
+					    struct hinic_tx_offload_info
+					    *off_info)
+{
+	u16  i, total_len, sge_cnt = mbuf_pkt->nb_segs;
+	struct rte_mbuf *mbuf;
+	int ret;
+
+	memset(off_info, 0, sizeof(*off_info));
+
+	ret = hinic_tx_offload_pkt_prepare(mbuf_pkt, off_info);
+	if (unlikely(ret))
+		return false;
+
+	sqe_info->cpy_mbuf_cnt = 0;
+
+	/* non tso mbuf */
+	if (likely(!(mbuf_pkt->ol_flags & PKT_TX_TCP_SEG))) {
+		if (unlikely(mbuf_pkt->pkt_len > MAX_SINGLE_SGE_SIZE)) {
+			/* non tso packet len must less than 64KB */
+			return false;
+		} else if (unlikely(HINIC_NONTSO_SEG_NUM_INVALID(sge_cnt))) {
+			/* non tso packet buffer number must less than 17
+			 * the mbuf segs more than 17 must copy to one buffer
+			 */
+			total_len = 0;
+			mbuf = mbuf_pkt;
+			for (i = 0; i < (HINIC_NONTSO_PKT_MAX_SGE - 1) ; i++) {
+				total_len += mbuf->data_len;
+				mbuf = mbuf->next;
+			}
+
+			/* default support copy total 4k mbuf segs */
+			if ((u32)(total_len + (u16)HINIC_COPY_MBUF_SIZE) <
+				  mbuf_pkt->pkt_len)
+				return false;
+
+			sqe_info->sge_cnt = HINIC_NONTSO_PKT_MAX_SGE;
+			sqe_info->cpy_mbuf_cnt = 1;
+			return true;
+		}
+
+		/* valid non tso mbuf */
+		sqe_info->sge_cnt = sge_cnt;
+	} else {
+		/* tso mbuf */
+		if (unlikely(HINIC_TSO_SEG_NUM_INVALID(sge_cnt)))
+			/* too many mbuf segs */
+			return false;
+
+		/* check tso mbuf segs are valid or not */
+		if (unlikely(!hinic_is_tso_sge_valid(mbuf_pkt,
+			     off_info, sqe_info)))
+			return false;
+	}
+
+	return true;
+}
+
+static inline void hinic_sq_write_db(struct hinic_sq *sq, int cos)
+{
+	u16 prod_idx;
+	u32 hi_prod_idx;
+	struct hinic_sq_db sq_db;
+
+	prod_idx = MASKED_SQ_IDX(sq, sq->wq->prod_idx);
+	hi_prod_idx = SQ_DB_PI_HIGH(prod_idx);
+
+	sq_db.db_info = SQ_DB_INFO_SET(hi_prod_idx, HI_PI) |
+			SQ_DB_INFO_SET(SQ_DB, TYPE) |
+			SQ_DB_INFO_SET(SQ_CFLAG_DP, CFLAG) |
+			SQ_DB_INFO_SET(cos, COS) |
+			SQ_DB_INFO_SET(sq->q_id, QID);
+
+	/* Data should be written to HW in Big Endian Format */
+	sq_db.db_info = cpu_to_be32(sq_db.db_info);
+
+	/* Write all before the doorbell */
+	rte_wmb();
+	writel(sq_db.db_info, SQ_DB_ADDR(sq, prod_idx));
+}
+
+u16 hinic_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, u16 nb_pkts)
+{
+	int free_wqebb_cnt, wqe_wqebb_cnt;
+	u32 queue_info, tx_bytes = 0;
+	u16 nb_tx;
+	struct hinic_wqe_info sqe_info;
+	struct hinic_tx_offload_info off_info;
+	struct rte_mbuf *mbuf_pkt;
+	struct hinic_txq *txq = tx_queue;
+	struct hinic_tx_info *tx_info;
+	struct hinic_sq_wqe *sq_wqe;
+	struct hinic_sq_task *task;
+
+	/* reclaim tx mbuf before xmit new packet */
+	if (HINIC_GET_SQ_FREE_WQEBBS(txq) < txq->tx_free_thresh)
+		hinic_xmit_mbuf_cleanup(txq);
+
+	/* tx loop routine */
+	for (nb_tx = 0; nb_tx < nb_pkts; nb_tx++) {
+		mbuf_pkt = *tx_pkts++;
+		queue_info = 0;
+
+		/* 1. parse sge and tx offlod info from mbuf */
+		if (unlikely(!hinic_get_sge_txoff_info(mbuf_pkt,
+						       &sqe_info, &off_info))) {
+			txq->txq_stats.off_errs++;
+			break;
+		}
+
+		/* 2. try to get enough wqebb */
+		wqe_wqebb_cnt = HINIC_SQ_WQEBB_CNT(sqe_info.sge_cnt);
+		free_wqebb_cnt = HINIC_GET_SQ_FREE_WQEBBS(txq);
+		if (unlikely(wqe_wqebb_cnt > free_wqebb_cnt)) {
+			/* reclaim again */
+			hinic_xmit_mbuf_cleanup(txq);
+			free_wqebb_cnt = HINIC_GET_SQ_FREE_WQEBBS(txq);
+			if (unlikely(wqe_wqebb_cnt > free_wqebb_cnt)) {
+				txq->txq_stats.tx_busy += (nb_pkts - nb_tx);
+				break;
+			}
+		}
+
+		/* 3. get sq tail wqe address from wqe_page,
+		 * sq have enough wqebb for this packet
+		 */
+		sq_wqe = hinic_get_sq_wqe(txq, wqe_wqebb_cnt, &sqe_info);
+
+		/* 4. fill sq wqe sge section */
+		if (unlikely(!hinic_mbuf_dma_map_sge(txq, mbuf_pkt,
+						     sq_wqe->buf_descs,
+						     &sqe_info))) {
+			hinic_return_sq_wqe(txq->nic_dev->hwdev, txq->q_id,
+					    wqe_wqebb_cnt, sqe_info.owner);
+			txq->txq_stats.off_errs++;
+			break;
+		}
+
+		/* 5. fill sq wqe task section and queue info */
+		task = &sq_wqe->task;
+
+		/* tx packet offload configure */
+		hinic_fill_tx_offload_info(mbuf_pkt, task, &queue_info,
+					   &off_info);
+
+		/* 6. record tx info */
+		tx_info = &txq->tx_info[sqe_info.pi];
+		tx_info->mbuf = mbuf_pkt;
+		tx_info->wqebb_cnt = wqe_wqebb_cnt;
+
+		/* 7. fill sq wqe header section */
+		hinic_fill_sq_wqe_header(&sq_wqe->ctrl, queue_info,
+					 sqe_info.sge_cnt, sqe_info.owner);
+
+		/* 8.convert continue or bottom wqe byteorder to big endian */
+		hinic_sq_wqe_cpu_to_be32(sq_wqe, sqe_info.seq_wqebbs);
+
+		tx_bytes += mbuf_pkt->pkt_len;
+	}
+
+	/* 9. write sq doorbell in burst mode */
+	if (nb_tx) {
+		hinic_sq_write_db(txq->sq, txq->cos);
+
+		txq->txq_stats.packets += nb_tx;
+		txq->txq_stats.bytes += tx_bytes;
+	}
+	txq->txq_stats.burst_pkts = nb_tx;
+
+	return nb_tx;
+}
 
 void hinic_free_all_tx_skbs(struct hinic_txq *txq)
 {
-- 
2.18.0


  parent reply	other threads:[~2019-06-19 16:23 UTC|newest]

Thread overview: 29+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2019-06-19 15:45 [dpdk-dev] [PATCH v5 00/15] A new net PMD - hinic Ziyang Xuan
2019-06-19 15:46 ` [dpdk-dev] [PATCH v5 01/15] net/hinic/base: add HW registers definition Ziyang Xuan
2019-06-19 15:50 ` [dpdk-dev] [PATCH v5 02/15] net/hinic/base: add HW interfaces of bar operation Ziyang Xuan
2019-06-19 15:47   ` Ziyang Xuan
2019-06-19 15:54 ` [dpdk-dev] [PATCH v5 03/15] net/hinic/base: add api command channel code Ziyang Xuan
2019-06-19 15:55 ` [dpdk-dev] [PATCH v5 04/15] net/hinic/base: add support for cmdq mechanism Ziyang Xuan
2019-06-19 15:56 ` [dpdk-dev] [PATCH v5 05/15] net/hinic/base: add eq mechanism function code Ziyang Xuan
2019-06-19 15:57 ` [dpdk-dev] [PATCH v5 06/15] net/hinic/base: add mgmt module " Ziyang Xuan
2019-06-19 16:02 ` [dpdk-dev] [PATCH v5 07/15] net/hinic/base: add code about hardware operation Ziyang Xuan
2019-06-19 15:58   ` Ziyang Xuan
2019-06-19 16:04 ` [dpdk-dev] [PATCH v5 08/15] net/hinic/base: add nic business configurations Ziyang Xuan
2019-06-19 16:05 ` [dpdk-dev] [PATCH v5 09/15] net/hinic/base: add context and work queue support Ziyang Xuan
2019-06-19 16:08 ` [dpdk-dev] [PATCH v5 10/15] net/hinic: add various headers Ziyang Xuan
2019-06-26 11:54   ` Ferruh Yigit
2019-06-19 16:09 ` [dpdk-dev] [PATCH v5 11/15] net/hinic: add hinic PMD build and doc files Ziyang Xuan
2019-06-19 16:13   ` Ziyang Xuan
2019-06-26 11:55   ` Ferruh Yigit
2019-06-19 16:16 ` [dpdk-dev] [PATCH v5 12/15] net/hinic: add device initailization Ziyang Xuan
2019-06-19 16:14   ` Ziyang Xuan
2019-06-19 16:30   ` Ziyang Xuan
2019-06-19 16:18 ` [dpdk-dev] [PATCH v5 13/15] net/hinic: add start stop close queue ops Ziyang Xuan
2019-06-19 16:32   ` Ziyang Xuan
2019-06-19 16:20 ` Ziyang Xuan [this message]
2019-06-19 16:25   ` [dpdk-dev] [PATCH v5 14/15] net/hinic: add tx/rx package burst Ziyang Xuan
2019-06-26 11:54   ` Ferruh Yigit
2019-06-26 15:58     ` [dpdk-dev] 答复: " Xuanziyang (William, Chip Application Design Logic and Hardware Development Dept IT_Products & Solutions)
2019-06-26 16:05       ` Ferruh Yigit
2019-06-19 16:23 ` [dpdk-dev] [PATCH v5 15/15] net/hinic: add rss stats promisc ops Ziyang Xuan
2019-06-26 11:56 ` [dpdk-dev] [PATCH v5 00/15] A new net PMD - hinic Ferruh Yigit

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=691dcc833654d7e8a666e88ea07dae95a4951230.1560958308.git.xuanziyang2@huawei.com \
    --to=xuanziyang2@huawei.com \
    --cc=cloud.wangxiaoyun@huawei.com \
    --cc=dev@dpdk.org \
    --cc=ferruh.yigit@intel.com \
    --cc=luoxianjun@huawei.com \
    --cc=shahar.belkar@huawei.com \
    --cc=tanya.brokhman@huawei.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).