DPDK patches and discussions
 help / color / mirror / Atom feed
From: Yanling Song <songyl@ramaxel.com>
To: <dev@dpdk.org>
Cc: <songyl@ramaxel.com>, <yanling.song@linux.dev>,
	<yanggan@ramaxel.com>, <ferruh.yigit@intel.com>
Subject: [PATCH v1 12/25] net/spnic: support mbuf handling of Tx/Rx
Date: Sat, 18 Dec 2021 10:51:39 +0800	[thread overview]
Message-ID: <600043103a745391c369c703b38251ad1cd0d54f.1639636621.git.songyl@ramaxel.com> (raw)
In-Reply-To: <cover.1639636621.git.songyl@ramaxel.com>

This patch defines a wqe data structure for hardware to
learn the sge info and offload info of packet. Furthermore,
this commit implements the interfaces to fill wqe with DPDK mbuf.

Signed-off-by: Yanling Song <songyl@ramaxel.com>
---
 drivers/net/spnic/base/spnic_nic_cfg.c |  23 ++
 drivers/net/spnic/base/spnic_nic_cfg.h |  23 ++
 drivers/net/spnic/meson.build          |   2 +
 drivers/net/spnic/spnic_ethdev.c       | 502 ++++++++++++++++++++++++-
 drivers/net/spnic/spnic_rx.c           | 302 +++++++++++++++
 drivers/net/spnic/spnic_rx.h           |  41 ++
 drivers/net/spnic/spnic_tx.c           | 334 ++++++++++++++++
 drivers/net/spnic/spnic_tx.h           | 228 +++++++++++
 8 files changed, 1454 insertions(+), 1 deletion(-)
 create mode 100644 drivers/net/spnic/spnic_rx.c
 create mode 100644 drivers/net/spnic/spnic_tx.c

diff --git a/drivers/net/spnic/base/spnic_nic_cfg.c b/drivers/net/spnic/base/spnic_nic_cfg.c
index 25d98d67dd..f6914f6f6d 100644
--- a/drivers/net/spnic/base/spnic_nic_cfg.c
+++ b/drivers/net/spnic/base/spnic_nic_cfg.c
@@ -378,6 +378,29 @@ int spnic_set_port_enable(void *hwdev, bool enable)
 	return 0;
 }
 
+int spnic_flush_qps_res(void *hwdev)
+{
+	struct spnic_cmd_clear_qp_resource sq_res;
+	u16 out_size = sizeof(sq_res);
+	int err;
+
+	if (!hwdev)
+		return -EINVAL;
+
+	memset(&sq_res, 0, sizeof(sq_res));
+	sq_res.func_id = spnic_global_func_id(hwdev);
+
+	err = l2nic_msg_to_mgmt_sync(hwdev, SPNIC_CMD_CLEAR_QP_RESOURCE, &sq_res,
+				     sizeof(sq_res), &sq_res, &out_size);
+	if (err || !out_size || sq_res.msg_head.status) {
+		PMD_DRV_LOG(ERR, "Clear sq resources failed, err: %d, status: 0x%x, out size: 0x%x",
+			    err, sq_res.msg_head.status, out_size);
+		return -EIO;
+	}
+
+	return 0;
+}
+
 static int spnic_set_function_table(void *hwdev, u32 cfg_bitmap,
 				     struct spnic_func_tbl_cfg *cfg)
 {
diff --git a/drivers/net/spnic/base/spnic_nic_cfg.h b/drivers/net/spnic/base/spnic_nic_cfg.h
index ce9792f8ee..746c1c342d 100644
--- a/drivers/net/spnic/base/spnic_nic_cfg.h
+++ b/drivers/net/spnic/base/spnic_nic_cfg.h
@@ -255,6 +255,17 @@ struct spnic_cmd_register_vf {
 	u8 rsvd[39];
 };
 
+
+struct spnic_cmd_set_rq_flush {
+	union {
+		struct {
+			u16 global_rq_id;
+			u16 local_rq_id;
+		};
+		u32 value;
+	};
+};
+
 int l2nic_msg_to_mgmt_sync(void *hwdev, u16 cmd, void *buf_in, u16 in_size,
 			   void *buf_out, u16 *out_size);
 
@@ -381,6 +392,18 @@ int spnic_set_port_enable(void *hwdev, bool enable);
  */
 int spnic_get_link_state(void *hwdev, u8 *link_state);
 
+/**
+ * Flush queue pairs resource in hardware
+ *
+ * @param[in] hwdev
+ *   Device pointer to hwdev
+ *
+ * @retval zero : Success
+ * @retval non-zero : Failure
+ */
+int spnic_flush_qps_res(void *hwdev);
+
+
 /**
  * Init nic hwdev
  *
diff --git a/drivers/net/spnic/meson.build b/drivers/net/spnic/meson.build
index 20d5151a8d..cd8f316366 100644
--- a/drivers/net/spnic/meson.build
+++ b/drivers/net/spnic/meson.build
@@ -7,6 +7,8 @@ objs = [base_objs]
 sources = files(
 	'spnic_ethdev.c',
 	'spnic_io.c',
+	'spnic_rx.c',
+	'spnic_tx.c'
 )
 
 includes += include_directories('base')
diff --git a/drivers/net/spnic/spnic_ethdev.c b/drivers/net/spnic/spnic_ethdev.c
index 4205ab43a4..27942e5d68 100644
--- a/drivers/net/spnic/spnic_ethdev.c
+++ b/drivers/net/spnic/spnic_ethdev.c
@@ -139,6 +139,468 @@ static int spnic_link_update(struct rte_eth_dev *dev, int wait_to_complete)
 	return rte_eth_linkstatus_set(dev, &link);
 }
 
+static void spnic_reset_rx_queue(struct rte_eth_dev *dev)
+{
+	struct spnic_rxq *rxq = NULL;
+	struct spnic_nic_dev *nic_dev;
+	int q_id = 0;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(dev);
+
+	for (q_id = 0; q_id < nic_dev->num_rqs; q_id++) {
+		rxq = nic_dev->rxqs[q_id];
+
+		rxq->cons_idx = 0;
+		rxq->prod_idx = 0;
+		rxq->delta = rxq->q_depth;
+		rxq->next_to_update = 0;
+	}
+}
+
+static void spnic_reset_tx_queue(struct rte_eth_dev *dev)
+{
+	struct spnic_nic_dev *nic_dev;
+	struct spnic_txq *txq = NULL;
+	int q_id = 0;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(dev);
+
+	for (q_id = 0; q_id < nic_dev->num_sqs; q_id++) {
+		txq = nic_dev->txqs[q_id];
+
+		txq->cons_idx = 0;
+		txq->prod_idx = 0;
+		txq->owner = 1;
+
+		/* Clear hardware ci */
+		*(u16 *)txq->ci_vaddr_base = 0;
+	}
+}
+
+/**
+ * Create the receive queue.
+ *
+ * @param[in] dev
+ *   Pointer to ethernet device structure.
+ * @param[in] qid
+ *   Receive queue index.
+ * @param[in] nb_desc
+ *   Number of descriptors for receive queue.
+ * @param[in] socket_id
+ *   Socket index on which memory must be allocated.
+ * @param rx_conf
+ *   Thresholds parameters (unused_).
+ * @param mp
+ *   Memory pool for buffer allocations.
+ *
+ * @retval zero : Success
+ * @retval non-zero : Failure
+ */
+static int spnic_rx_queue_setup(struct rte_eth_dev *dev, uint16_t qid,
+			uint16_t nb_desc, unsigned int socket_id,
+			__rte_unused const struct rte_eth_rxconf *rx_conf,
+			struct rte_mempool *mp)
+{
+	struct spnic_nic_dev *nic_dev;
+	struct spnic_rxq *rxq = NULL;
+	const struct rte_memzone *rq_mz = NULL;
+	const struct rte_memzone *cqe_mz = NULL;
+	const struct rte_memzone *pi_mz = NULL;
+	u16 rq_depth, rx_free_thresh;
+	u32 queue_buf_size, mb_buf_size;
+	void *db_addr = NULL;
+	int wqe_count;
+	u32 buf_size;
+	int err;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(dev);
+
+	/* Queue depth must be power of 2, otherwise will be aligned up */
+	rq_depth = (nb_desc & (nb_desc - 1)) ?
+		   ((u16)(1U << (ilog2(nb_desc) + 1))) : nb_desc;
+
+	/*
+	 * Validate number of receive descriptors.
+	 * It must not exceed hardware maximum and minimum.
+	 */
+	if (rq_depth > SPNIC_MAX_QUEUE_DEPTH ||
+	    rq_depth < SPNIC_MIN_QUEUE_DEPTH) {
+		PMD_DRV_LOG(ERR, "RX queue depth is out of range from %d to %d,"
+			    "(nb_desc: %d, q_depth: %d, port: %d queue: %d)",
+			    SPNIC_MIN_QUEUE_DEPTH, SPNIC_MAX_QUEUE_DEPTH,
+			    (int)nb_desc, (int)rq_depth,
+			    (int)dev->data->port_id, (int)qid);
+		return -EINVAL;
+	}
+
+	/*
+	 * The RX descriptor ring will be cleaned after rxq->rx_free_thresh
+	 * descriptors are used or if the number of descriptors required
+	 * to transmit a packet is greater than the number of free RX
+	 * descriptors.
+	 * The following constraints must be satisfied:
+	 *  -rx_free_thresh must be greater than 0.
+	 *  -rx_free_thresh must be less than the size of the ring minus 1.
+	 * When set to zero use default values.
+	 */
+	rx_free_thresh = (u16)((rx_conf->rx_free_thresh) ?
+			rx_conf->rx_free_thresh : SPNIC_DEFAULT_RX_FREE_THRESH);
+	if (rx_free_thresh >= (rq_depth - 1)) {
+		PMD_DRV_LOG(ERR, "rx_free_thresh must be less than the number "
+			    "of RX descriptors minus 1, rx_free_thresh: %u port: %d queue: %d)",
+			    (unsigned int)rx_free_thresh,
+			    (int)dev->data->port_id, (int)qid);
+		return -EINVAL;
+	}
+
+	rxq = rte_zmalloc_socket("spnic_rq", sizeof(struct spnic_rxq),
+				 RTE_CACHE_LINE_SIZE, socket_id);
+	if (!rxq) {
+		PMD_DRV_LOG(ERR, "Allocate rxq[%d] failed, dev_name: %s",
+			    qid, dev->data->name);
+		return -ENOMEM;
+	}
+
+	/* Init rq parameters */
+	rxq->nic_dev = nic_dev;
+	nic_dev->rxqs[qid] = rxq;
+	rxq->mb_pool = mp;
+	rxq->q_id = qid;
+	rxq->next_to_update = 0;
+	rxq->q_depth = rq_depth;
+	rxq->q_mask = rq_depth - 1;
+	rxq->delta = rq_depth;
+	rxq->cons_idx = 0;
+	rxq->prod_idx = 0;
+	rxq->wqe_type = SPNIC_NORMAL_RQ_WQE;
+	rxq->wqebb_shift = SPNIC_RQ_WQEBB_SHIFT + rxq->wqe_type;
+	rxq->wqebb_size = (u16)BIT(rxq->wqebb_shift);
+	rxq->rx_free_thresh = rx_free_thresh;
+	rxq->rxinfo_align_end = rxq->q_depth - rxq->rx_free_thresh;
+	rxq->port_id = dev->data->port_id;
+
+	/* If buf_len used for function table, need to translated */
+	mb_buf_size = rte_pktmbuf_data_room_size(rxq->mb_pool) -
+		      RTE_PKTMBUF_HEADROOM;
+	err = spnic_convert_rx_buf_size(mb_buf_size, &buf_size);
+	if (err) {
+		PMD_DRV_LOG(ERR, "Adjust buf size failed, dev_name: %s",
+			    dev->data->name);
+		goto adjust_bufsize_fail;
+	}
+
+	rxq->buf_len = buf_size;
+	rxq->rx_buff_shift = ilog2(rxq->buf_len);
+
+	pi_mz = rte_eth_dma_zone_reserve(dev, "spnic_rq_pi", qid,
+					 RTE_PGSIZE_4K, RTE_CACHE_LINE_SIZE,
+					 socket_id);
+	if (!pi_mz) {
+		PMD_DRV_LOG(ERR, "Allocate rxq[%d] pi_mz failed, dev_name: %s",
+			    qid, dev->data->name);
+		err = -ENOMEM;
+		goto alloc_pi_mz_fail;
+	}
+	rxq->pi_mz = pi_mz;
+	rxq->pi_dma_addr = pi_mz->iova;
+	rxq->pi_virt_addr = pi_mz->addr;
+
+	/* Rxq doesn't use direct wqe */
+	err = spnic_alloc_db_addr(nic_dev->hwdev, &db_addr, NULL);
+	if (err) {
+		PMD_DRV_LOG(ERR, "Alloc rq doorbell addr failed");
+		goto alloc_db_err_fail;
+	}
+	rxq->db_addr = db_addr;
+
+	queue_buf_size = BIT(rxq->wqebb_shift) * rq_depth;
+	rq_mz = rte_eth_dma_zone_reserve(dev, "spnic_rq_mz", qid,
+					 queue_buf_size, RTE_PGSIZE_256K,
+					 socket_id);
+	if (!rq_mz) {
+		PMD_DRV_LOG(ERR, "Allocate rxq[%d] rq_mz failed, dev_name: %s",
+			    qid, dev->data->name);
+		err = -ENOMEM;
+		goto alloc_rq_mz_fail;
+	}
+
+	memset(rq_mz->addr, 0, queue_buf_size);
+	rxq->rq_mz = rq_mz;
+	rxq->queue_buf_paddr = rq_mz->iova;
+	rxq->queue_buf_vaddr = rq_mz->addr;
+
+	rxq->rx_info = rte_zmalloc_socket("rx_info",
+					  rq_depth * sizeof(*rxq->rx_info),
+					  RTE_CACHE_LINE_SIZE, socket_id);
+	if (!rxq->rx_info) {
+		PMD_DRV_LOG(ERR, "Allocate rx_info failed, dev_name: %s",
+			dev->data->name);
+		err = -ENOMEM;
+		goto alloc_rx_info_fail;
+	}
+
+	cqe_mz = rte_eth_dma_zone_reserve(dev, "spnic_cqe_mz", qid,
+					  rq_depth * sizeof(*rxq->rx_cqe),
+					  RTE_CACHE_LINE_SIZE, socket_id);
+	if (!cqe_mz) {
+		PMD_DRV_LOG(ERR, "Allocate cqe mem zone failed, dev_name: %s",
+			    dev->data->name);
+		err = -ENOMEM;
+		goto alloc_cqe_mz_fail;
+	}
+	memset(cqe_mz->addr, 0, rq_depth * sizeof(*rxq->rx_cqe));
+	rxq->cqe_mz = cqe_mz;
+	rxq->cqe_start_paddr = cqe_mz->iova;
+	rxq->cqe_start_vaddr = cqe_mz->addr;
+	rxq->rx_cqe = (struct spnic_rq_cqe *)rxq->cqe_start_vaddr;
+
+	wqe_count = spnic_rx_fill_wqe(rxq);
+	if (wqe_count != rq_depth) {
+		PMD_DRV_LOG(ERR, "Fill rx wqe failed, wqe_count: %d, dev_name: %s",
+			    wqe_count, dev->data->name);
+		err = -ENOMEM;
+		goto fill_rx_wqe_fail;
+	}
+
+	/* Record rxq pointer in rte_eth rx_queues */
+	dev->data->rx_queues[qid] = rxq;
+
+	return 0;
+
+fill_rx_wqe_fail:
+	rte_memzone_free(rxq->cqe_mz);
+alloc_cqe_mz_fail:
+	rte_free(rxq->rx_info);
+
+alloc_rx_info_fail:
+	rte_memzone_free(rxq->rq_mz);
+
+alloc_rq_mz_fail:
+	spnic_free_db_addr(nic_dev->hwdev, rxq->db_addr, NULL);
+
+alloc_db_err_fail:
+	rte_memzone_free(rxq->pi_mz);
+
+alloc_pi_mz_fail:
+adjust_bufsize_fail:
+
+	rte_free(rxq);
+	nic_dev->rxqs[qid] = NULL;
+
+	return err;
+}
+
+/**
+ * Create the transmit queue.
+ *
+ * @param[in] dev
+ *   Pointer to ethernet device structure.
+ * @param[in] queue_idx
+ *   Transmit queue index.
+ * @param[in] nb_desc
+ *   Number of descriptors for transmit queue.
+ * @param[in] socket_id
+ *   Socket index on which memory must be allocated.
+ * @param[in] tx_conf
+ *   Tx queue configuration parameters (unused_).
+ *
+ * @retval zero : Success
+ * @retval non-zero : Failure
+ */
+static int spnic_tx_queue_setup(struct rte_eth_dev *dev, uint16_t qid,
+			 uint16_t nb_desc, unsigned int socket_id,
+			 __rte_unused const struct rte_eth_txconf *tx_conf)
+{
+	struct spnic_nic_dev *nic_dev;
+	struct spnic_hwdev *hwdev;
+	struct spnic_txq *txq = NULL;
+	const struct rte_memzone *sq_mz = NULL;
+	const struct rte_memzone *ci_mz = NULL;
+	void *db_addr = NULL;
+	u16 sq_depth, tx_free_thresh;
+	u32 queue_buf_size;
+	int err;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(dev);
+	hwdev = nic_dev->hwdev;
+
+	/* Queue depth must be power of 2, otherwise will be aligned up */
+	sq_depth = (nb_desc & (nb_desc - 1)) ?
+		   ((u16)(1U << (ilog2(nb_desc) + 1))) : nb_desc;
+
+	/*
+	 * Validate number of transmit descriptors.
+	 * It must not exceed hardware maximum and minimum.
+	 */
+	if (sq_depth > SPNIC_MAX_QUEUE_DEPTH ||
+		sq_depth < SPNIC_MIN_QUEUE_DEPTH) {
+		PMD_DRV_LOG(ERR, "TX queue depth is out of range from %d to %d,"
+			    "(nb_desc: %d, q_depth: %d, port: %d queue: %d)",
+			    SPNIC_MIN_QUEUE_DEPTH, SPNIC_MAX_QUEUE_DEPTH,
+			    (int)nb_desc, (int)sq_depth,
+			    (int)dev->data->port_id, (int)qid);
+		return -EINVAL;
+	}
+
+	/*
+	 * The TX descriptor ring will be cleaned after txq->tx_free_thresh
+	 * descriptors are used or if the number of descriptors required
+	 * to transmit a packet is greater than the number of free TX
+	 * descriptors.
+	 * The following constraints must be satisfied:
+	 *  -tx_free_thresh must be greater than 0.
+	 *  -tx_free_thresh must be less than the size of the ring minus 1.
+	 * When set to zero use default values.
+	 */
+	tx_free_thresh = (u16)((tx_conf->tx_free_thresh) ?
+		tx_conf->tx_free_thresh : SPNIC_DEFAULT_TX_FREE_THRESH);
+	if (tx_free_thresh >= (sq_depth - 1)) {
+		PMD_DRV_LOG(ERR, "tx_free_thresh must be less than the number of tx "
+			    "descriptors minus 1, tx_free_thresh: %u port: %d queue: %d",
+			    (unsigned int)tx_free_thresh,
+			    (int)dev->data->port_id, (int)qid);
+		return -EINVAL;
+	}
+
+	txq = rte_zmalloc_socket("spnic_tx_queue", sizeof(struct spnic_txq),
+				 RTE_CACHE_LINE_SIZE, socket_id);
+	if (!txq) {
+		PMD_DRV_LOG(ERR, "Allocate txq[%d] failed, dev_name: %s",
+			    qid, dev->data->name);
+		return -ENOMEM;
+	}
+	nic_dev->txqs[qid] = txq;
+	txq->nic_dev = nic_dev;
+	txq->q_id = qid;
+	txq->q_depth = sq_depth;
+	txq->q_mask = sq_depth - 1;
+	txq->cons_idx = 0;
+	txq->prod_idx = 0;
+	txq->wqebb_shift = SPNIC_SQ_WQEBB_SHIFT;
+	txq->wqebb_size = (u16)BIT(txq->wqebb_shift);
+	txq->tx_free_thresh = tx_free_thresh;
+	txq->owner = 1;
+	txq->cos = nic_dev->default_cos;
+
+	ci_mz = rte_eth_dma_zone_reserve(dev, "spnic_sq_ci", qid,
+					 SPNIC_CI_Q_ADDR_SIZE,
+					 SPNIC_CI_Q_ADDR_SIZE, socket_id);
+	if (!ci_mz) {
+		PMD_DRV_LOG(ERR, "Allocate txq[%d] ci_mz failed, dev_name: %s",
+			    qid, dev->data->name);
+		err = -ENOMEM;
+		goto alloc_ci_mz_fail;
+	}
+	txq->ci_mz = ci_mz;
+	txq->ci_dma_base = ci_mz->iova;
+	txq->ci_vaddr_base = ci_mz->addr;
+
+	queue_buf_size = BIT(txq->wqebb_shift) * sq_depth;
+	sq_mz = rte_eth_dma_zone_reserve(dev, "spnic_sq_mz", qid,
+					 queue_buf_size, RTE_PGSIZE_256K,
+					 socket_id);
+	if (!sq_mz) {
+		PMD_DRV_LOG(ERR, "Allocate txq[%d] sq_mz failed, dev_name: %s",
+			    qid, dev->data->name);
+		err = -ENOMEM;
+		goto alloc_sq_mz_fail;
+	}
+	memset(sq_mz->addr, 0, queue_buf_size);
+	txq->sq_mz = sq_mz;
+	txq->queue_buf_paddr = sq_mz->iova;
+	txq->queue_buf_vaddr = sq_mz->addr;
+	txq->sq_head_addr = (u64)txq->queue_buf_vaddr;
+	txq->sq_bot_sge_addr = txq->sq_head_addr + queue_buf_size;
+
+	/* Sq doesn't use direct wqe */
+	err = spnic_alloc_db_addr(hwdev, &db_addr, NULL);
+	if (err) {
+		PMD_DRV_LOG(ERR, "Alloc sq doorbell addr failed");
+		goto alloc_db_err_fail;
+	}
+	txq->db_addr = db_addr;
+
+	txq->tx_info = rte_zmalloc_socket("tx_info",
+					  sq_depth * sizeof(*txq->tx_info),
+					  RTE_CACHE_LINE_SIZE, socket_id);
+	if (!txq->tx_info) {
+		PMD_DRV_LOG(ERR, "Allocate tx_info failed, dev_name: %s",
+			    dev->data->name);
+		err = -ENOMEM;
+		goto alloc_tx_info_fail;
+	}
+
+	/* Record txq pointer in rte_eth tx_queues */
+	dev->data->tx_queues[qid] = txq;
+
+	return 0;
+
+alloc_tx_info_fail:
+	spnic_free_db_addr(hwdev, txq->db_addr, NULL);
+
+alloc_db_err_fail:
+	rte_memzone_free(txq->sq_mz);
+
+alloc_sq_mz_fail:
+	rte_memzone_free(txq->ci_mz);
+
+alloc_ci_mz_fail:
+	rte_free(txq);
+
+	return err;
+}
+
+static void spnic_rx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
+{
+	struct spnic_rxq *rxq = dev->data->rx_queues[qid];
+	struct spnic_nic_dev *nic_dev;
+
+	if (!rxq) {
+		PMD_DRV_LOG(WARNING, "Rxq is null when release");
+		return;
+	}
+	nic_dev = rxq->nic_dev;
+
+	spnic_free_rxq_mbufs(rxq);
+
+	rte_memzone_free(rxq->cqe_mz);
+
+	rte_free(rxq->rx_info);
+
+	rte_memzone_free(rxq->rq_mz);
+
+	rte_memzone_free(rxq->pi_mz);
+
+	nic_dev->rxqs[rxq->q_id] = NULL;
+	rte_free(rxq);
+}
+
+static void spnic_tx_queue_release(struct rte_eth_dev *dev, uint16_t qid)
+{
+	struct spnic_txq *txq = dev->data->tx_queues[qid];
+	struct spnic_nic_dev *nic_dev;
+
+	if (!txq) {
+		PMD_DRV_LOG(WARNING, "Txq is null when release");
+		return;
+	}
+	nic_dev = txq->nic_dev;
+
+	spnic_free_txq_mbufs(txq);
+
+	rte_free(txq->tx_info);
+	txq->tx_info = NULL;
+
+	spnic_free_db_addr(nic_dev->hwdev, txq->db_addr, NULL);
+
+	rte_memzone_free(txq->sq_mz);
+
+	rte_memzone_free(txq->ci_mz);
+
+	nic_dev->txqs[txq->q_id] = NULL;
+	rte_free(txq);
+}
+
 static void spnic_delete_mc_addr_list(struct spnic_nic_dev *nic_dev);
 
 /**
@@ -235,6 +697,7 @@ static int spnic_dev_start(struct rte_eth_dev *eth_dev)
 
 	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
 
+	spnic_get_func_rx_buf_size(nic_dev);
 	err = spnic_init_function_table(nic_dev->hwdev, nic_dev->rx_buff_len);
 	if (err) {
 		PMD_DRV_LOG(ERR, "Init function table failed, dev_name: %s",
@@ -253,6 +716,9 @@ static int spnic_dev_start(struct rte_eth_dev *eth_dev)
 		goto get_feature_err;
 	}
 
+	/* reset rx and tx queue */
+	spnic_reset_rx_queue(eth_dev);
+	spnic_reset_tx_queue(eth_dev);
 
 	/* Init txq and rxq context */
 	err = spnic_init_qp_ctxts(nic_dev);
@@ -270,6 +736,15 @@ static int spnic_dev_start(struct rte_eth_dev *eth_dev)
 		goto set_mtu_fail;
 	}
 
+	err = spnic_start_all_rqs(eth_dev);
+	if (err) {
+		PMD_DRV_LOG(ERR, "Set rx config failed, dev_name: %s",
+			    eth_dev->data->name);
+		goto start_rqs_fail;
+	}
+
+	spnic_start_all_sqs(eth_dev);
+
 	/* Update eth_dev link status */
 	if (eth_dev->data->dev_conf.intr_conf.lsc != 0)
 		(void)spnic_link_update(eth_dev, 0);
@@ -278,6 +753,7 @@ static int spnic_dev_start(struct rte_eth_dev *eth_dev)
 
 	return 0;
 
+start_rqs_fail:
 set_mtu_fail:
 	spnic_free_qp_ctxts(nic_dev->hwdev);
 
@@ -313,9 +789,17 @@ static int spnic_dev_stop(struct rte_eth_dev *dev)
 	memset(&link, 0, sizeof(link));
 	(void)rte_eth_linkstatus_set(dev, &link);
 
+	/* Flush pending io request */
+	spnic_flush_txqs(nic_dev);
+
+	spnic_flush_qps_res(nic_dev->hwdev);
 	/* Clean root context */
 	spnic_free_qp_ctxts(nic_dev->hwdev);
 
+	/* Free all tx and rx mbufs */
+	spnic_free_all_txq_mbufs(nic_dev);
+	spnic_free_all_rxq_mbufs(nic_dev);
+
 	return 0;
 }
 
@@ -329,6 +813,7 @@ static int spnic_dev_close(struct rte_eth_dev *eth_dev)
 {
 	struct spnic_nic_dev *nic_dev =
 	SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
+	int qid;
 
 	if (rte_bit_relaxed_test_and_set32(SPNIC_DEV_CLOSE, &nic_dev->dev_status)) {
 		PMD_DRV_LOG(WARNING, "Device %s already closed",
@@ -338,6 +823,13 @@ static int spnic_dev_close(struct rte_eth_dev *eth_dev)
 
 	spnic_dev_stop(eth_dev);
 
+	/* Release io resource */
+	for (qid = 0; qid < nic_dev->num_sqs; qid++)
+		spnic_tx_queue_release(eth_dev, qid);
+
+	for (qid = 0; qid < nic_dev->num_rqs; qid++)
+		spnic_rx_queue_release(eth_dev, qid);
+
 	spnic_deinit_sw_rxtxqs(nic_dev);
 	spnic_deinit_mac_addr(eth_dev);
 	rte_free(nic_dev->mc_list);
@@ -581,6 +1073,10 @@ static const struct eth_dev_ops spnic_pmd_ops = {
 	.dev_set_link_up               = spnic_dev_set_link_up,
 	.dev_set_link_down             = spnic_dev_set_link_down,
 	.link_update                   = spnic_link_update,
+	.rx_queue_setup                = spnic_rx_queue_setup,
+	.tx_queue_setup                = spnic_tx_queue_setup,
+	.rx_queue_release              = spnic_rx_queue_release,
+	.tx_queue_release              = spnic_tx_queue_release,
 	.dev_start                     = spnic_dev_start,
 	.dev_stop                      = spnic_dev_stop,
 	.dev_close                     = spnic_dev_close,
@@ -592,8 +1088,12 @@ static const struct eth_dev_ops spnic_pmd_ops = {
 };
 
 static const struct eth_dev_ops spnic_pmd_vf_ops = {
-	.link_update                   = spnic_link_update,
+	.rx_queue_setup                = spnic_rx_queue_setup,
+	.tx_queue_setup                = spnic_tx_queue_setup,
 	.dev_start                     = spnic_dev_start,
+	.link_update                   = spnic_link_update,
+	.rx_queue_release              = spnic_rx_queue_release,
+	.tx_queue_release              = spnic_tx_queue_release,
 	.dev_stop                      = spnic_dev_stop,
 	.dev_close                     = spnic_dev_close,
 	.mtu_set                       = spnic_dev_set_mtu,
diff --git a/drivers/net/spnic/spnic_rx.c b/drivers/net/spnic/spnic_rx.c
new file mode 100644
index 0000000000..20cd50c0c4
--- /dev/null
+++ b/drivers/net/spnic/spnic_rx.c
@@ -0,0 +1,302 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2021 Ramaxel Memory Technology, Ltd
+ */
+
+#include <rte_ether.h>
+#include <rte_mbuf.h>
+#include <ethdev_driver.h>
+
+#include "base/spnic_compat.h"
+#include "base/spnic_cmd.h"
+#include "base/spnic_hwif.h"
+#include "base/spnic_hwdev.h"
+#include "base/spnic_wq.h"
+#include "base/spnic_mgmt.h"
+#include "base/spnic_nic_cfg.h"
+#include "spnic_io.h"
+#include "spnic_rx.h"
+#include "spnic_ethdev.h"
+
+/**
+ * Get receive queue wqe
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @param[out] pi
+ *   Return current pi
+ * @return
+ *   RX wqe base address
+ */
+static inline void *spnic_get_rq_wqe(struct spnic_rxq *rxq, u16 *pi)
+{
+	*pi = MASKED_QUEUE_IDX(rxq, rxq->prod_idx);
+
+	/* Get only one rq wqe for once */
+	rxq->prod_idx++;
+	rxq->delta--;
+
+	return NIC_WQE_ADDR(rxq, *pi);
+}
+
+/**
+ * Put receive queue wqe
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @param[in] wqe_cnt
+ *   Wqebb counters
+ */
+static inline void spnic_put_rq_wqe(struct spnic_rxq *rxq, u16 wqe_cnt)
+{
+	rxq->delta += wqe_cnt;
+	rxq->prod_idx -= wqe_cnt;
+}
+
+/**
+ * Get receive queue local pi
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @return
+ *   Receive queue local pi
+ */
+static inline u16 spnic_get_rq_local_pi(struct spnic_rxq *rxq)
+{
+	return MASKED_QUEUE_IDX(rxq, rxq->prod_idx);
+}
+
+/**
+ * Update receive queue hardware pi
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @param[in] pi
+ *   Receive queue pi to update
+ */
+static inline void spnic_update_rq_hw_pi(struct spnic_rxq *rxq, u16 pi)
+{
+	*rxq->pi_virt_addr = cpu_to_be16((pi & rxq->q_mask) << rxq->wqe_type);
+}
+
+int spnic_rx_fill_wqe(struct spnic_rxq *rxq)
+{
+	struct spnic_rq_wqe *rq_wqe = NULL;
+	struct spnic_nic_dev *nic_dev = rxq->nic_dev;
+	rte_iova_t cqe_dma;
+	u16 pi = 0;
+	int i;
+
+	cqe_dma = rxq->cqe_start_paddr;
+	for (i = 0; i < rxq->q_depth; i++) {
+		rq_wqe = spnic_get_rq_wqe(rxq, &pi);
+		if (!rq_wqe) {
+			PMD_DRV_LOG(ERR, "Get rq wqe failed, rxq id: %d, wqe id: %d",
+				    rxq->q_id, i);
+			break;
+		}
+
+		if (rxq->wqe_type == SPNIC_EXTEND_RQ_WQE) {
+			/* Unit of cqe length is 16B */
+			spnic_set_sge(&rq_wqe->extend_wqe.cqe_sect.sge,
+				       cqe_dma,
+				       sizeof(struct spnic_rq_cqe) >>
+				       SPNIC_CQE_SIZE_SHIFT);
+			/* Use fixed len */
+			rq_wqe->extend_wqe.buf_desc.sge.len =
+							nic_dev->rx_buff_len;
+		} else {
+			rq_wqe->normal_wqe.cqe_hi_addr = upper_32_bits(cqe_dma);
+			rq_wqe->normal_wqe.cqe_lo_addr = lower_32_bits(cqe_dma);
+		}
+
+		cqe_dma += sizeof(struct spnic_rq_cqe);
+	}
+
+	spnic_put_rq_wqe(rxq, (u16)i);
+
+	return i;
+}
+
+static struct rte_mbuf *spnic_rx_alloc_mbuf(struct spnic_rxq *rxq,
+					    rte_iova_t *dma_addr)
+{
+	struct rte_mbuf *mbuf = NULL;
+
+	if (unlikely(rte_pktmbuf_alloc_bulk(rxq->mb_pool, &mbuf, 1) != 0))
+		return NULL;
+
+	*dma_addr = rte_mbuf_data_iova_default(mbuf);
+
+	return mbuf;
+}
+
+u32 spnic_rx_fill_buffers(struct spnic_rxq *rxq)
+{
+	struct spnic_rq_wqe *rq_wqe = NULL;
+	struct spnic_rx_info *rx_info = NULL;
+	struct rte_mbuf *mb = NULL;
+	rte_iova_t dma_addr;
+	int i, free_wqebbs;
+
+	free_wqebbs = rxq->delta - 1;
+	for (i = 0; i < free_wqebbs; i++) {
+		rx_info = &rxq->rx_info[rxq->next_to_update];
+
+		mb = spnic_rx_alloc_mbuf(rxq, &dma_addr);
+		if (!mb) {
+			PMD_DRV_LOG(ERR, "Alloc mbuf failed");
+			break;
+		}
+
+		rx_info->mbuf = mb;
+
+		rq_wqe = NIC_WQE_ADDR(rxq, rxq->next_to_update);
+
+		/* Fill buffer address only */
+		if (rxq->wqe_type == SPNIC_EXTEND_RQ_WQE) {
+			rq_wqe->extend_wqe.buf_desc.sge.hi_addr = upper_32_bits(dma_addr);
+			rq_wqe->extend_wqe.buf_desc.sge.lo_addr = lower_32_bits(dma_addr);
+		} else {
+			rq_wqe->normal_wqe.buf_hi_addr = upper_32_bits(dma_addr);
+			rq_wqe->normal_wqe.buf_lo_addr = lower_32_bits(dma_addr);
+		}
+
+		rxq->next_to_update = (rxq->next_to_update + 1) & rxq->q_mask;
+	}
+
+	if (likely(i > 0)) {
+		spnic_write_db(rxq->db_addr, rxq->q_id, 0, RQ_CFLAG_DP,
+				rxq->next_to_update << rxq->wqe_type);
+		/* Init rq contxet used, need to optimization */
+		rxq->prod_idx = rxq->next_to_update;
+		rxq->delta -= i;
+	} else {
+		PMD_DRV_LOG(ERR, "Alloc rx buffers failed, rxq_id: %d",
+			    rxq->q_id);
+	}
+
+	return i;
+}
+
+void spnic_free_rxq_mbufs(struct spnic_rxq *rxq)
+{
+	struct spnic_rx_info *rx_info = NULL;
+	int free_wqebbs = spnic_get_rq_free_wqebb(rxq) + 1;
+	volatile struct spnic_rq_cqe *rx_cqe = NULL;
+	u16 ci;
+
+	while (free_wqebbs++ < rxq->q_depth) {
+		ci = spnic_get_rq_local_ci(rxq);
+
+		rx_cqe = &rxq->rx_cqe[ci];
+
+		/* Clear done bit */
+		rx_cqe->status = 0;
+
+		rx_info = &rxq->rx_info[ci];
+		rte_pktmbuf_free(rx_info->mbuf);
+		rx_info->mbuf = NULL;
+
+		spnic_update_rq_local_ci(rxq, 1);
+	}
+}
+
+void spnic_free_all_rxq_mbufs(struct spnic_nic_dev *nic_dev)
+{
+	u16 qid;
+
+	for (qid = 0; qid < nic_dev->num_rqs; qid++)
+		spnic_free_rxq_mbufs(nic_dev->rxqs[qid]);
+}
+
+static inline u32 spnic_rx_alloc_mbuf_bulk(struct spnic_rxq *rxq,
+					   struct rte_mbuf **mbufs,
+					   u32 exp_mbuf_cnt)
+{
+	u32 avail_cnt;
+	int err;
+
+	err = rte_pktmbuf_alloc_bulk(rxq->mb_pool, mbufs, exp_mbuf_cnt);
+	if (likely(err == 0)) {
+		avail_cnt = exp_mbuf_cnt;
+	} else {
+		avail_cnt = 0;
+		rxq->rxq_stats.rx_nombuf += exp_mbuf_cnt;
+	}
+
+	return avail_cnt;
+}
+
+static inline void spnic_rearm_rxq_mbuf(struct spnic_rxq *rxq)
+{
+	struct spnic_rq_wqe *rq_wqe = NULL;
+	struct rte_mbuf **rearm_mbufs;
+	u32 i, free_wqebbs, rearm_wqebbs, exp_wqebbs;
+	rte_iova_t dma_addr;
+	u16 pi;
+
+	/* Check free wqebb cnt fo rearm */
+	free_wqebbs = spnic_get_rq_free_wqebb(rxq);
+	if (unlikely(free_wqebbs < rxq->rx_free_thresh))
+		return;
+
+	/* Get rearm mbuf array */
+	pi = spnic_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 = spnic_rx_alloc_mbuf_bulk(rxq, rearm_mbufs, exp_wqebbs);
+	if (unlikely(rearm_wqebbs == 0))
+		return;
+
+	/* Rearm rx mbuf */
+	rq_wqe = NIC_WQE_ADDR(rxq, pi);
+	for (i = 0; i < rearm_wqebbs; i++) {
+		dma_addr = rte_mbuf_data_iova_default(rearm_mbufs[i]);
+
+		/* Fill buffer address only */
+		if (rxq->wqe_type == SPNIC_EXTEND_RQ_WQE) {
+			rq_wqe->extend_wqe.buf_desc.sge.hi_addr = upper_32_bits(dma_addr);
+			rq_wqe->extend_wqe.buf_desc.sge.lo_addr = lower_32_bits(dma_addr);
+		} else {
+			rq_wqe->normal_wqe.buf_hi_addr = upper_32_bits(dma_addr);
+			rq_wqe->normal_wqe.buf_lo_addr = lower_32_bits(dma_addr);
+		}
+
+		rq_wqe = (struct spnic_rq_wqe *)((u64)rq_wqe +
+			 rxq->wqebb_size);
+	}
+	rxq->prod_idx += rearm_wqebbs;
+	rxq->delta -= rearm_wqebbs;
+
+#ifndef SPNIC_RQ_DB
+	spnic_write_db(rxq->db_addr, rxq->q_id, 0, RQ_CFLAG_DP,
+			((pi + rearm_wqebbs) & rxq->q_mask) << rxq->wqe_type);
+#else
+	/* Update rq hw_pi */
+	rte_wmb();
+	spnic_update_rq_hw_pi(rxq, pi + rearm_wqebbs);
+#endif
+}
+
+int spnic_start_all_rqs(struct rte_eth_dev *eth_dev)
+{
+	struct spnic_nic_dev *nic_dev = NULL;
+	struct spnic_rxq *rxq = NULL;
+	int i;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
+
+	for (i = 0; i < nic_dev->num_rqs; i++) {
+		rxq = eth_dev->data->rx_queues[i];
+		spnic_rearm_rxq_mbuf(rxq);
+		eth_dev->data->rx_queue_state[i] = RTE_ETH_QUEUE_STATE_STARTED;
+	}
+
+	return 0;
+}
diff --git a/drivers/net/spnic/spnic_rx.h b/drivers/net/spnic/spnic_rx.h
index b2f0052533..46f4e1276d 100644
--- a/drivers/net/spnic/spnic_rx.h
+++ b/drivers/net/spnic/spnic_rx.h
@@ -110,4 +110,45 @@ struct spnic_rxq {
 	struct spnic_rxq_stats	rxq_stats;
 } __rte_cache_aligned;
 
+int spnic_rx_fill_wqe(struct spnic_rxq *rxq);
+
+u32 spnic_rx_fill_buffers(struct spnic_rxq *rxq);
+
+void spnic_free_rxq_mbufs(struct spnic_rxq *rxq);
+
+void spnic_free_all_rxq_mbufs(struct spnic_nic_dev *nic_dev);
+
+int spnic_start_all_rqs(struct rte_eth_dev *eth_dev);
+/**
+ * Get receive queue local ci
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @return
+ *   Receive queue local ci
+ */
+static inline u16 spnic_get_rq_local_ci(struct spnic_rxq *rxq)
+{
+	return MASKED_QUEUE_IDX(rxq, rxq->cons_idx);
+}
+
+static inline u16 spnic_get_rq_free_wqebb(struct spnic_rxq *rxq)
+{
+	return rxq->delta - 1;
+}
+
+/**
+ * Update receive queue local ci
+ *
+ * @param[in] rxq
+ *   Receive queue
+ * @param[in] wqe_cnt
+ *   Wqebb counters
+ */
+static inline void spnic_update_rq_local_ci(struct spnic_rxq *rxq,
+					     u16 wqe_cnt)
+{
+	rxq->cons_idx += wqe_cnt;
+	rxq->delta += wqe_cnt;
+}
 #endif /* _SPNIC_RX_H_ */
diff --git a/drivers/net/spnic/spnic_tx.c b/drivers/net/spnic/spnic_tx.c
new file mode 100644
index 0000000000..d905879412
--- /dev/null
+++ b/drivers/net/spnic/spnic_tx.c
@@ -0,0 +1,334 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2021 Ramaxel Memory Technology, Ltd
+ */
+
+#include <rte_ether.h>
+#include <rte_mbuf.h>
+#include <rte_io.h>
+#include <ethdev_driver.h>
+
+#include "base/spnic_compat.h"
+#include "base/spnic_cmd.h"
+#include "base/spnic_wq.h"
+#include "base/spnic_mgmt.h"
+#include "base/spnic_hwdev.h"
+#include "base/spnic_nic_cfg.h"
+#include "spnic_io.h"
+#include "spnic_tx.h"
+#include "spnic_ethdev.h"
+
+#define SPNIC_TX_TASK_WRAPPED		1
+#define SPNIC_TX_BD_DESC_WRAPPED	2
+
+#define TX_MSS_DEFAULT			0x3E00
+#define TX_MSS_MIN			0x50
+
+#define SPNIC_MAX_TX_FREE_BULK		64
+
+#define	MAX_PAYLOAD_OFFSET		221
+
+#define SPNIC_TX_OUTER_CHECKSUM_FLAG_SET       1
+#define SPNIC_TX_OUTER_CHECKSUM_FLAG_NO_SET    0
+
+/**
+ * Get send queue free wqebb cnt
+ *
+ * @param[in] sq
+ *   Send queue
+ * @return
+ *   Number of free wqebb
+ */
+static inline u16 spnic_get_sq_free_wqebbs(struct spnic_txq *sq)
+{
+	return (sq->q_depth -
+		((sq->prod_idx - sq->cons_idx + sq->q_depth) & sq->q_mask) - 1);
+}
+
+/**
+ * Update send queue local ci
+ *
+ * @param[in] sq
+ *   Send queue
+ * @param[in] wqe_cnt
+ *   Number of wqebb
+ */
+static inline void spnic_update_sq_local_ci(struct spnic_txq *sq, u16 wqe_cnt)
+{
+	sq->cons_idx += wqe_cnt;
+}
+
+/**
+ * Get send queue local ci
+ *
+ * @param[in] sq
+ *   Send queue
+ * @return
+ *   Local ci
+ */
+static inline u16 spnic_get_sq_local_ci(struct spnic_txq *sq)
+{
+	return MASKED_QUEUE_IDX(sq, sq->cons_idx);
+}
+
+/**
+ * Get send queue hardware ci
+ *
+ * @param[in] sq
+ *   Send queue
+ * @return
+ *   Hardware ci
+ */
+static inline u16 spnic_get_sq_hw_ci(struct spnic_txq *sq)
+{
+	return MASKED_QUEUE_IDX(sq, *(u16 *)sq->ci_vaddr_base);
+}
+
+/**
+ * Get send queue wqe
+ *
+ * @param[in] sq
+ *   Send queue
+ * @param[in] wqebb_cnt
+ *   Num of wqebb counter
+ * @param[out] pi
+ *   Return current pi
+ * @param[out] owner
+ *   Owner bit for hardware
+ * @param[out] wrapped
+ *   Indicate whether wqe is wrapped
+ * @return
+ *   Send queue wqe base address
+ */
+static inline void *spnic_get_sq_wqe(struct spnic_txq *sq,
+				     struct spnic_wqe_info *wqe_info)
+{
+	u16 cur_pi = MASKED_QUEUE_IDX(sq, sq->prod_idx);
+	u32 end_pi;
+
+	end_pi = cur_pi + wqe_info->wqebb_cnt;
+	sq->prod_idx += wqe_info->wqebb_cnt;
+
+	wqe_info->owner = sq->owner;
+	wqe_info->pi = cur_pi;
+	wqe_info->wrapped = 0;
+
+	if (unlikely(end_pi >= sq->q_depth)) {
+		sq->owner = !sq->owner;
+
+		if (likely(end_pi > sq->q_depth))
+			wqe_info->wrapped = sq->q_depth - cur_pi;
+	}
+
+	return NIC_WQE_ADDR(sq, cur_pi);
+}
+
+/**
+ * Put send queue wqe
+ *
+ * @param[in] sq
+ *   Send queue
+ * @param[in] wqebb_cnt
+ *   Num of wqebb counter
+ * @param[out] owner
+ *   Owner bit for hardware
+ */
+static inline void spnic_put_sq_wqe(struct spnic_txq *sq,
+				    struct spnic_wqe_info *wqe_info)
+{
+	if (wqe_info->owner != sq->owner)
+		sq->owner = wqe_info->owner;
+
+	sq->prod_idx -= wqe_info->wqebb_cnt;
+}
+
+static inline void spnic_set_wqe_combo(struct spnic_txq *txq,
+				    struct spnic_sq_wqe_combo *wqe_combo,
+					struct spnic_sq_wqe *wqe,
+					struct spnic_wqe_info *wqe_info)
+{
+	wqe_combo->hdr = &wqe->compact_wqe.wqe_desc;
+
+	if (wqe_info->offload) {
+		if (wqe_info->wrapped == SPNIC_TX_TASK_WRAPPED) {
+			wqe_combo->task = (struct spnic_sq_task *)
+				(void *)txq->sq_head_addr;
+			wqe_combo->bds_head = (struct spnic_sq_bufdesc *)
+				(void *)(txq->sq_head_addr + txq->wqebb_size);
+		} else if (wqe_info->wrapped == SPNIC_TX_BD_DESC_WRAPPED) {
+			wqe_combo->task = &wqe->extend_wqe.task;
+			wqe_combo->bds_head = (struct spnic_sq_bufdesc *)
+				(void *)(txq->sq_head_addr);
+		} else {
+			wqe_combo->task = &wqe->extend_wqe.task;
+			wqe_combo->bds_head = wqe->extend_wqe.buf_desc;
+		}
+
+		wqe_combo->wqe_type = SQ_WQE_EXTENDED_TYPE;
+		wqe_combo->task_type = SQ_WQE_TASKSECT_16BYTES;
+		return;
+	}
+
+	if (wqe_info->wrapped == SPNIC_TX_TASK_WRAPPED) {
+		wqe_combo->bds_head = (struct spnic_sq_bufdesc *)
+				(void *)(txq->sq_head_addr);
+	} else {
+		wqe_combo->bds_head =
+			(struct spnic_sq_bufdesc *)(&wqe->extend_wqe.task);
+	}
+
+	if (wqe_info->wqebb_cnt > 1) {
+		wqe_combo->wqe_type = SQ_WQE_EXTENDED_TYPE;
+		wqe_combo->task_type = SQ_WQE_TASKSECT_46BITS;
+		/* This section used as vlan insert, needs to clear */
+		wqe_combo->bds_head->rsvd = 0;
+	} else {
+		wqe_combo->wqe_type = SQ_WQE_COMPACT_TYPE;
+	}
+}
+
+void spnic_free_txq_mbufs(struct spnic_txq *txq)
+{
+	struct spnic_tx_info *tx_info = NULL;
+	u16 free_wqebbs;
+	u16 ci;
+
+	free_wqebbs = spnic_get_sq_free_wqebbs(txq) + 1;
+
+	while (free_wqebbs < txq->q_depth) {
+		ci = spnic_get_sq_local_ci(txq);
+
+		tx_info = &txq->tx_info[ci];
+
+		rte_pktmbuf_free(tx_info->mbuf);
+		spnic_update_sq_local_ci(txq, tx_info->wqebb_cnt);
+
+		free_wqebbs += tx_info->wqebb_cnt;
+		tx_info->mbuf = NULL;
+	}
+}
+
+void spnic_free_all_txq_mbufs(struct spnic_nic_dev *nic_dev)
+{
+	u16 qid;
+
+	for (qid = 0; qid < nic_dev->num_sqs; qid++)
+		spnic_free_txq_mbufs(nic_dev->txqs[qid]);
+}
+
+int spnic_start_all_sqs(struct rte_eth_dev *eth_dev)
+{
+	struct spnic_nic_dev *nic_dev = NULL;
+	int i;
+
+	nic_dev = SPNIC_ETH_DEV_TO_PRIVATE_NIC_DEV(eth_dev);
+
+	for (i = 0; i < nic_dev->num_rqs; i++)
+		eth_dev->data->tx_queue_state[i] = RTE_ETH_QUEUE_STATE_STARTED;
+
+	return 0;
+}
+static inline int spnic_xmit_mbuf_cleanup(struct spnic_txq *txq, u32 free_cnt)
+{
+	struct spnic_tx_info *tx_info = NULL;
+	struct rte_mbuf *mbuf = NULL;
+	struct rte_mbuf *mbuf_temp = NULL;
+	struct rte_mbuf *mbuf_free[SPNIC_MAX_TX_FREE_BULK];
+	int nb_free = 0;
+	int wqebb_cnt = 0;
+	u16 hw_ci, sw_ci, sq_mask;
+	u32 i;
+
+	hw_ci = spnic_get_sq_hw_ci(txq);
+	sw_ci = spnic_get_sq_local_ci(txq);
+	sq_mask = txq->q_mask;
+
+	for (i = 0; i < free_cnt; ++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;
+
+		wqebb_cnt += tx_info->wqebb_cnt;
+		mbuf = tx_info->mbuf;
+
+		if (likely(mbuf->nb_segs == 1)) {
+			mbuf_temp = rte_pktmbuf_prefree_seg(mbuf);
+			tx_info->mbuf = NULL;
+
+			if (unlikely(mbuf_temp == NULL))
+				continue;
+
+			mbuf_free[nb_free++] = mbuf_temp;
+			if (unlikely(mbuf_temp->pool != mbuf_free[0]->pool ||
+			    nb_free >= SPNIC_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++] = mbuf_temp;
+			}
+		} 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);
+
+	spnic_update_sq_local_ci(txq, wqebb_cnt);
+	return i;
+}
+
+static int spnic_tx_done_cleanup(void *txq, u32 free_cnt)
+{
+	struct spnic_txq *tx_queue = txq;
+	u32 try_free_cnt = !free_cnt ? tx_queue->q_depth : free_cnt;
+
+	return spnic_xmit_mbuf_cleanup(tx_queue, try_free_cnt);
+}
+int spnic_stop_sq(struct spnic_txq *txq)
+{
+	struct spnic_nic_dev *nic_dev = txq->nic_dev;
+	unsigned long timeout;
+	int err = -EFAULT;
+	int free_wqebbs;
+
+	timeout = msecs_to_jiffies(SPNIC_FLUSH_QUEUE_TIMEOUT) + jiffies;
+	do {
+		spnic_tx_done_cleanup(txq, 0);
+		free_wqebbs = spnic_get_sq_free_wqebbs(txq) + 1;
+		if (free_wqebbs == txq->q_depth) {
+			err = 0;
+			break;
+		}
+
+		rte_delay_us(1);
+	} while (time_before(jiffies, timeout));
+
+	if (err)
+		PMD_DRV_LOG(WARNING, "%s Wait sq empty timeout, queue_idx: %u, sw_ci: %u, "
+			    "hw_ci: %u, sw_pi: %u, free_wqebbs: %u, q_depth:%u\n",
+			    nic_dev->dev_name, txq->q_id,
+			    spnic_get_sq_local_ci(txq),
+			    spnic_get_sq_hw_ci(txq),
+			    MASKED_QUEUE_IDX(txq, txq->prod_idx),
+			    free_wqebbs, txq->q_depth);
+
+	return err;
+}
+
+/* Should stop transmiting any packets before calling this function */
+void spnic_flush_txqs(struct spnic_nic_dev *nic_dev)
+{
+	u16 qid;
+	int err;
+
+	for (qid = 0; qid < nic_dev->num_sqs; qid++) {
+		err = spnic_stop_sq(nic_dev->txqs[qid]);
+		if (err)
+			PMD_DRV_LOG(ERR, "Stop sq%d failed", qid);
+	}
+}
diff --git a/drivers/net/spnic/spnic_tx.h b/drivers/net/spnic/spnic_tx.h
index 7528b27bd9..d770b15c21 100644
--- a/drivers/net/spnic/spnic_tx.h
+++ b/drivers/net/spnic/spnic_tx.h
@@ -4,6 +4,224 @@
 
 #ifndef _SPNIC_TX_H_
 #define _SPNIC_TX_H_
+/* Tx offload info */
+struct spnic_tx_offload_info {
+	u8 outer_l2_len;
+	u8 outer_l3_type;
+	u16 outer_l3_len;
+
+	u8 inner_l2_len;
+	u8 inner_l3_type;
+	u16 inner_l3_len;
+
+	u8 tunnel_length;
+	u8 tunnel_type;
+	u8 inner_l4_type;
+	u8 inner_l4_len;
+
+	u16 payload_offset;
+	u8 inner_l4_tcp_udp;
+	u8 rsvd0;
+};
+
+/* tx wqe ctx */
+struct spnic_wqe_info {
+	u8 around;
+	u8 cpy_mbuf_cnt;
+	u16 sge_cnt;
+
+	u8 offload;
+	u8 rsvd0;
+	u16 payload_offset;
+
+	u8 wrapped;
+	u8 owner;
+	u16 pi;
+
+	u16 wqebb_cnt;
+	u16 rsvd1;
+
+	u32 queue_info;
+};
+struct spnic_sq_wqe_desc {
+	u32 ctrl_len;
+	u32 queue_info;
+	u32 hi_addr;
+	u32 lo_addr;
+};
+
+/*
+ * Engine only pass first 12B TS field directly to uCode through metadata,
+ * vlan_offoad is used for hardware when vlan insert in tx
+ */
+struct spnic_sq_task {
+	u32 pkt_info0;
+	u32 ip_identify;
+	u32 pkt_info2; /* Rsvd for ipsec spi */
+	u32 vlan_offload;
+};
+
+struct spnic_sq_bufdesc {
+	u32 len; /* 31-bits Length, L2NIC only use length[17:0] */
+	u32 rsvd;
+	u32 hi_addr;
+	u32 lo_addr;
+};
+
+struct spnic_sq_compact_wqe {
+	struct spnic_sq_wqe_desc wqe_desc;
+};
+
+struct spnic_sq_extend_wqe {
+	struct spnic_sq_wqe_desc wqe_desc;
+	struct spnic_sq_task task;
+	struct spnic_sq_bufdesc buf_desc[0];
+};
+
+struct spnic_sq_wqe {
+	union {
+		struct spnic_sq_compact_wqe compact_wqe;
+		struct spnic_sq_extend_wqe extend_wqe;
+	};
+};
+
+/* Use section pointer to support non continuous wqe */
+struct spnic_sq_wqe_combo {
+	struct spnic_sq_wqe_desc *hdr;
+	struct spnic_sq_task *task;
+	struct spnic_sq_bufdesc *bds_head;
+	u32 wqe_type;
+	u32 task_type;
+};
+
+/* SQ ctrl info */
+enum sq_wqe_data_format {
+	SQ_NORMAL_WQE = 0,
+};
+
+enum sq_wqe_ec_type {
+	SQ_WQE_COMPACT_TYPE = 0,
+	SQ_WQE_EXTENDED_TYPE = 1,
+};
+
+#define COMPACT_WQE_MAX_CTRL_LEN 0x3FFF
+
+enum sq_wqe_tasksect_len_type {
+	SQ_WQE_TASKSECT_46BITS = 0,
+	SQ_WQE_TASKSECT_16BYTES = 1,
+};
+
+#define SQ_CTRL_BD0_LEN_SHIFT			0
+#define SQ_CTRL_RSVD_SHIFT			18
+#define SQ_CTRL_BUFDESC_NUM_SHIFT		19
+#define SQ_CTRL_TASKSECT_LEN_SHIFT		27
+#define SQ_CTRL_DATA_FORMAT_SHIFT		28
+#define SQ_CTRL_DIRECT_SHIFT			29
+#define SQ_CTRL_EXTENDED_SHIFT			30
+#define SQ_CTRL_OWNER_SHIFT			31
+
+#define SQ_CTRL_BD0_LEN_MASK			0x3FFFFU
+#define SQ_CTRL_RSVD_MASK			0x1U
+#define SQ_CTRL_BUFDESC_NUM_MASK		0xFFU
+#define SQ_CTRL_TASKSECT_LEN_MASK		0x1U
+#define SQ_CTRL_DATA_FORMAT_MASK		0x1U
+#define SQ_CTRL_DIRECT_MASK			0x1U
+#define SQ_CTRL_EXTENDED_MASK			0x1U
+#define SQ_CTRL_OWNER_MASK			0x1U
+
+#define SQ_CTRL_SET(val, member)	(((u32)(val) & \
+					 SQ_CTRL_##member##_MASK) << \
+					 SQ_CTRL_##member##_SHIFT)
+
+#define SQ_CTRL_GET(val, member)	(((val) >> SQ_CTRL_##member##_SHIFT) \
+					& SQ_CTRL_##member##_MASK)
+
+#define SQ_CTRL_CLEAR(val, member)	((val) & \
+					(~(SQ_CTRL_##member##_MASK << \
+					SQ_CTRL_##member##_SHIFT)))
+
+#define SQ_CTRL_QUEUE_INFO_PKT_TYPE_SHIFT		0
+#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_PKT_TYPE_MASK		0x3U
+#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_TUNNEL_FLAG_SHIFT		19
+#define	SQ_TASK_INFO0_ESP_NEXT_PROTO_SHIFT	22
+#define	SQ_TASK_INFO0_INNER_L4_EN_SHIFT		24
+#define	SQ_TASK_INFO0_INNER_L3_EN_SHIFT		25
+#define	SQ_TASK_INFO0_INNER_L4_PSEUDO_SHIFT	26
+#define	SQ_TASK_INFO0_OUT_L4_EN_SHIFT		27
+#define	SQ_TASK_INFO0_OUT_L3_EN_SHIFT		28
+#define	SQ_TASK_INFO0_OUT_L4_PSEUDO_SHIFT	29
+#define	SQ_TASK_INFO0_ESP_OFFLOAD_SHIFT		30
+#define	SQ_TASK_INFO0_IPSEC_PROTO_SHIFT		31
+
+#define	SQ_TASK_INFO0_TUNNEL_FLAG_MASK		0x1U
+#define	SQ_TASK_INFO0_ESP_NEXT_PROTO_MASK	0x3U
+#define	SQ_TASK_INFO0_INNER_L4_EN_MASK		0x1U
+#define	SQ_TASK_INFO0_INNER_L3_EN_MASK		0x1U
+#define	SQ_TASK_INFO0_INNER_L4_PSEUDO_MASK	0x1U
+#define	SQ_TASK_INFO0_OUT_L4_EN_MASK		0x1U
+#define	SQ_TASK_INFO0_OUT_L3_EN_MASK		0x1U
+#define	SQ_TASK_INFO0_OUT_L4_PSEUDO_MASK	0x1U
+#define	SQ_TASK_INFO0_ESP_OFFLOAD_MASK		0x1U
+#define	SQ_TASK_INFO0_IPSEC_PROTO_MASK		0x1U
+
+#define SQ_TASK_INFO0_SET(val, member)			\
+		(((u32)(val) & SQ_TASK_INFO0_##member##_MASK) <<	\
+		SQ_TASK_INFO0_##member##_SHIFT)
+#define SQ_TASK_INFO0_GET(val, member)			\
+		(((val) >> SQ_TASK_INFO0_##member##_SHIFT) &	\
+		SQ_TASK_INFO0_##member##_MASK)
+
+#define SQ_TASK_INFO1_SET(val, member)			\
+		(((val) & SQ_TASK_INFO1_##member##_MASK) <<	\
+		SQ_TASK_INFO1_##member##_SHIFT)
+#define SQ_TASK_INFO1_GET(val, member)			\
+		(((val) >> SQ_TASK_INFO1_##member##_SHIFT) &	\
+		SQ_TASK_INFO1_##member##_MASK)
+
+#define	SQ_TASK_INFO3_VLAN_TAG_SHIFT		0
+#define	SQ_TASK_INFO3_VLAN_TYPE_SHIFT		16
+#define	SQ_TASK_INFO3_VLAN_TAG_VALID_SHIFT	19
+
+#define	SQ_TASK_INFO3_VLAN_TAG_MASK		0xFFFFU
+#define	SQ_TASK_INFO3_VLAN_TYPE_MASK		0x7U
+#define	SQ_TASK_INFO3_VLAN_TAG_VALID_MASK	0x1U
+
+#define SQ_TASK_INFO3_SET(val, member)			\
+		(((val) & SQ_TASK_INFO3_##member##_MASK) <<	\
+		SQ_TASK_INFO3_##member##_SHIFT)
+#define SQ_TASK_INFO3_GET(val, member)			\
+		(((val) >> SQ_TASK_INFO3_##member##_SHIFT) &	\
+		SQ_TASK_INFO3_##member##_MASK)
 
 /* Txq info */
 struct spnic_txq_stats {
@@ -59,4 +277,14 @@ struct spnic_txq {
 	struct spnic_txq_stats txq_stats;
 } __rte_cache_aligned;
 
+void spnic_flush_txqs(struct spnic_nic_dev *nic_dev);
+
+void spnic_free_txq_mbufs(struct spnic_txq *txq);
+
+void spnic_free_all_txq_mbufs(struct spnic_nic_dev *nic_dev);
+
+u16 spnic_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, u16 nb_pkts);
+
+int spnic_stop_sq(struct spnic_txq *txq);
+int spnic_start_all_sqs(struct rte_eth_dev *eth_dev);
 #endif /* _SPNIC_TX_H_ */
-- 
2.27.0


  parent reply	other threads:[~2021-12-18  2:54 UTC|newest]

Thread overview: 32+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-12-18  2:51 [PATCH v1 00/25] Net/SPNIC: support SPNIC into DPDK 22.03 Yanling Song
2021-12-18  2:51 ` [PATCH v1 01/25] drivers/net: introduce a new PMD driver Yanling Song
2021-12-19 19:40   ` Stephen Hemminger
2021-12-22  0:54     ` Yanling Song
2021-12-22 16:55       ` Stephen Hemminger
2021-12-23  8:10         ` Yanling Song
2021-12-18  2:51 ` [PATCH v1 02/25] net/spnic: initialize the HW interface Yanling Song
2021-12-18  2:51 ` [PATCH v1 03/25] net/spnic: add mbox message channel Yanling Song
2021-12-18  2:51 ` [PATCH v1 04/25] net/spnic: introduce event queue Yanling Song
2021-12-18  2:51 ` [PATCH v1 05/25] net/spnic: add mgmt module Yanling Song
2021-12-18  2:51 ` [PATCH v1 06/25] net/spnic: add cmdq and work queue Yanling Song
2021-12-18  2:51 ` [PATCH v1 07/25] net/spnic: add interface handling cmdq message Yanling Song
2021-12-18  2:51 ` [PATCH v1 08/25] net/spnic: add hardware info initialization Yanling Song
2021-12-18  2:51 ` [PATCH v1 09/25] net/spnic: support MAC and link event handling Yanling Song
2021-12-18  2:51 ` [PATCH v1 10/25] net/spnic: add function info initialization Yanling Song
2021-12-18  2:51 ` [PATCH v1 11/25] net/spnic: add queue pairs context initialization Yanling Song
2021-12-18  2:51 ` Yanling Song [this message]
2021-12-18  2:51 ` [PATCH v1 13/25] net/spnic: support Rx congfiguration Yanling Song
2021-12-18  2:51 ` [PATCH v1 14/25] net/spnic: add port/vport enable Yanling Song
2021-12-18  2:51 ` [PATCH v1 15/25] net/spnic: support IO packets handling Yanling Song
2021-12-18  2:51 ` [PATCH v1 16/25] net/spnic: add device configure/version/info Yanling Song
2021-12-20  0:23   ` Stephen Hemminger
2021-12-22  0:56     ` Yanling Song
2021-12-18  2:51 ` [PATCH v1 17/25] net/spnic: support RSS configuration update and get Yanling Song
2021-12-18  2:51 ` [PATCH v1 18/25] net/spnic: support VLAN filtering and offloading Yanling Song
2021-12-18  2:51 ` [PATCH v1 19/25] net/spnic: support promiscuous and allmulticast Rx modes Yanling Song
2021-12-18  2:51 ` [PATCH v1 20/25] net/spnic: support flow control Yanling Song
2021-12-18  2:51 ` [PATCH v1 21/25] net/spnic: support getting Tx/Rx queues info Yanling Song
2021-12-18  2:51 ` [PATCH v1 22/25] net/spnic: net/spnic: support xstats statistics Yanling Song
2021-12-18  2:51 ` [PATCH v1 23/25] net/spnic: support VFIO interrupt Yanling Song
2021-12-18  2:51 ` [PATCH v1 24/25] net/spnic: support Tx/Rx queue start/stop Yanling Song
2021-12-18  2:51 ` [PATCH v1 25/25] net/spnic: add doc infrastructure Yanling Song

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=600043103a745391c369c703b38251ad1cd0d54f.1639636621.git.songyl@ramaxel.com \
    --to=songyl@ramaxel.com \
    --cc=dev@dpdk.org \
    --cc=ferruh.yigit@intel.com \
    --cc=yanggan@ramaxel.com \
    --cc=yanling.song@linux.dev \
    /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).