This patch adds NEON vector instructions to optimize Tx burst process.
Signed-off-by: Huisong Li <lihuisong@huawei.com>
Signed-off-by: Wei Hu (Xavier) <xavier.huwei@huawei.com>
Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
hns->rx_simple_allowed = true;
hns->tx_simple_allowed = true;
+ hns->tx_vec_allowed = true;
+
hns3_init_rx_ptype_tble(dev);
hw->adapter_state = HNS3_NIC_CONFIGURED;
bool rx_simple_allowed;
bool tx_simple_allowed;
+ bool tx_vec_allowed;
+
struct hns3_ptype_table ptype_tbl __rte_cache_min_aligned;
};
hns->rx_simple_allowed = true;
hns->tx_simple_allowed = true;
+ hns->tx_vec_allowed = true;
+
hns3_init_rx_ptype_tble(dev);
hw->adapter_state = HNS3_NIC_CONFIGURED;
rte_memzone_free(txq->mz);
if (txq->sw_ring)
rte_free(txq->sw_ring);
+ if (txq->free)
+ rte_free(txq->free);
rte_free(txq);
}
}
/* Don't need alloc sw_ring, because upper applications don't use it */
txq->sw_ring = NULL;
+ txq->free = NULL;
txq->hns = hns;
txq->tx_deferred_start = false;
txq->tx_bd_ready = txq->nb_tx_desc - 1;
txq->tx_free_thresh = tx_free_thresh;
txq->tx_rs_thresh = tx_rs_thresh;
+ txq->free = rte_zmalloc_socket("hns3 TX mbuf free array",
+ sizeof(struct rte_mbuf *) * txq->tx_rs_thresh,
+ RTE_CACHE_LINE_SIZE, socket_id);
+ if (!txq->free) {
+ hns3_err(hw, "failed to allocate tx mbuf free array!");
+ hns3_tx_queue_release(txq);
+ return -ENOMEM;
+ }
+
txq->port_id = dev->data->port_id;
txq->pvid_state = hw->port_base_vlan_cfg.state;
txq->configured = true;
return nb_tx;
}
+int __rte_weak
+hns3_tx_check_vec_support(__rte_unused struct rte_eth_dev *dev)
+{
+ return -ENOTSUP;
+}
+
+uint16_t __rte_weak
+hns3_xmit_pkts_vec(__rte_unused void *tx_queue,
+ __rte_unused struct rte_mbuf **tx_pkts,
+ __rte_unused uint16_t nb_pkts)
+{
+ return 0;
+}
+
int
hns3_tx_burst_mode_get(struct rte_eth_dev *dev, __rte_unused uint16_t queue_id,
struct rte_eth_burst_mode *mode)
info = "Scalar Simple";
else if (pkt_burst == hns3_xmit_pkts)
info = "Scalar";
+ else if (pkt_burst == hns3_xmit_pkts_vec)
+ info = "Vector Neon";
if (info == NULL)
return -EINVAL;
uint64_t offloads = dev->data->dev_conf.txmode.offloads;
struct hns3_adapter *hns = dev->data->dev_private;
+ if (hns->tx_vec_allowed && hns3_tx_check_vec_support(dev) == 0) {
+ *prep = NULL;
+ return hns3_xmit_pkts_vec;
+ }
+
if (hns->tx_simple_allowed &&
offloads == (offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE)) {
*prep = NULL;
#define HNS3_DEFAULT_TX_RS_THRESH 32
#define HNS3_TX_FAST_FREE_AHEAD 64
+#define HNS3_UINT8_BIT 8
+#define HNS3_UINT16_BIT 16
+#define HNS3_UINT32_BIT 32
+
#define HNS3_512_BD_BUF_SIZE 512
#define HNS3_1K_BD_BUF_SIZE 1024
#define HNS3_2K_BD_BUF_SIZE 2048
#define HNS3_L3_LEN_UNIT 2UL
#define HNS3_L4_LEN_UNIT 2UL
+#define HNS3_TXD_DEFAULT_BDTYPE 0
+#define HNS3_TXD_VLD_CMD (0x1 << HNS3_TXD_VLD_B)
+#define HNS3_TXD_FE_CMD (0x1 << HNS3_TXD_FE_B)
+#define HNS3_TXD_DEFAULT_VLD_FE_BDTYPE \
+ (HNS3_TXD_VLD_CMD | HNS3_TXD_FE_CMD | HNS3_TXD_DEFAULT_BDTYPE)
+#define HNS3_TXD_SEND_SIZE_SHIFT 16
+
enum hns3_pkt_l2t_type {
HNS3_L2_TYPE_UNICAST,
HNS3_L2_TYPE_MULTICAST,
* all descriptors are cleared. and then free all mbufs in the batch.
* - tx_rs_thresh
* Number of mbufs released at a time.
-
+ *
+ * - free
+ * Tx mbuf free array used for preserving temporarily address of mbuf
+ * released back to mempool, when releasing mbuf in batches.
*/
uint16_t tx_rs_thresh;
+ struct rte_mbuf **free;
/*
* port based vlan configuration state.
uint16_t nb_pkts);
uint16_t hns3_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts,
uint16_t nb_pkts);
+uint16_t hns3_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts,
+ uint16_t nb_pkts);
int hns3_tx_burst_mode_get(struct rte_eth_dev *dev,
__rte_unused uint16_t queue_id,
struct rte_eth_burst_mode *mode);
void hns3_update_all_queues_pvid_state(struct hns3_hw *hw);
void hns3_rx_scattered_reset(struct rte_eth_dev *dev);
void hns3_rx_scattered_calc(struct rte_eth_dev *dev);
+int hns3_tx_check_vec_support(struct rte_eth_dev *dev);
void hns3_rxq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
struct rte_eth_rxq_info *qinfo);
void hns3_txq_info_get(struct rte_eth_dev *dev, uint16_t queue_id,
--- /dev/null
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Hisilicon Limited.
+ */
+
+#include <rte_io.h>
+#include <rte_ethdev_driver.h>
+
+#include "hns3_ethdev.h"
+#include "hns3_rxtx.h"
+#include "hns3_rxtx_vec.h"
+
+#if defined RTE_ARCH_ARM64
+#include "hns3_rxtx_vec_neon.h"
+#endif
+
+int
+hns3_tx_check_vec_support(struct rte_eth_dev *dev)
+{
+ struct rte_eth_txmode *txmode = &dev->data->dev_conf.txmode;
+
+ /* Only support DEV_TX_OFFLOAD_MBUF_FAST_FREE */
+ if (txmode->offloads != DEV_TX_OFFLOAD_MBUF_FAST_FREE)
+ return -ENOTSUP;
+
+ return 0;
+}
+
+uint16_t
+hns3_xmit_pkts_vec(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
+{
+ struct hns3_tx_queue *txq = (struct hns3_tx_queue *)tx_queue;
+ uint16_t nb_tx = 0;
+
+ while (nb_pkts) {
+ uint16_t ret, new_burst;
+
+ new_burst = RTE_MIN(nb_pkts, txq->tx_rs_thresh);
+ ret = hns3_xmit_fixed_burst_vec(tx_queue, &tx_pkts[nb_tx],
+ new_burst);
+ nb_tx += ret;
+ nb_pkts -= ret;
+ if (ret < new_burst)
+ break;
+ }
+
+ return nb_tx;
+}
--- /dev/null
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Hisilicon Limited.
+ */
+
+#ifndef _HNS3_RXTX_VEC_H_
+#define _HNS3_RXTX_VEC_H_
+
+#include "hns3_rxtx.h"
+#include "hns3_ethdev.h"
+
+static inline void
+hns3_tx_free_buffers(struct hns3_tx_queue *txq)
+{
+ struct rte_mbuf **free = txq->free;
+ struct hns3_entry *tx_entry;
+ struct hns3_desc *tx_desc;
+ struct rte_mbuf *m;
+ int nb_free = 0;
+ int i;
+
+ /*
+ * All mbufs can be released only when the VLD bits of all
+ * descriptors in a batch are cleared.
+ */
+ tx_desc = &txq->tx_ring[txq->next_to_clean];
+ for (i = 0; i < txq->tx_rs_thresh; i++, tx_desc++) {
+ if (tx_desc->tx.tp_fe_sc_vld_ra_ri &
+ rte_le_to_cpu_16(BIT(HNS3_TXD_VLD_B)))
+ return;
+ }
+
+ tx_entry = &txq->sw_ring[txq->next_to_clean];
+ for (i = 0; i < txq->tx_rs_thresh; i++, tx_entry++) {
+ m = rte_pktmbuf_prefree_seg(tx_entry->mbuf);
+ tx_entry->mbuf = NULL;
+
+ if (m == NULL)
+ continue;
+
+ if (nb_free && m->pool != free[0]->pool) {
+ rte_mempool_put_bulk(free[0]->pool, (void **)free,
+ nb_free);
+ nb_free = 0;
+ }
+ free[nb_free++] = m;
+ }
+
+ if (nb_free)
+ rte_mempool_put_bulk(free[0]->pool, (void **)free, nb_free);
+
+ /* Update numbers of available descriptor due to buffer freed */
+ txq->tx_bd_ready += txq->tx_rs_thresh;
+ txq->next_to_clean += txq->tx_rs_thresh;
+ if (txq->next_to_clean >= txq->nb_tx_desc)
+ txq->next_to_clean = 0;
+}
+#endif /* _HNS3_RXTX_VEC_H_ */
--- /dev/null
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(c) 2020 Hisilicon Limited.
+ */
+
+#ifndef _HNS3_RXTX_VEC_NEON_H_
+#define _HNS3_RXTX_VEC_NEON_H_
+
+#include <arm_neon.h>
+
+#pragma GCC diagnostic ignored "-Wcast-qual"
+
+static inline void
+hns3_vec_tx(volatile struct hns3_desc *desc, struct rte_mbuf *pkt)
+{
+ uint64x2_t val1 = {
+ pkt->buf_iova + pkt->data_off,
+ ((uint64_t)pkt->data_len) << HNS3_TXD_SEND_SIZE_SHIFT
+ };
+ uint64x2_t val2 = {
+ 0,
+ ((uint64_t)HNS3_TXD_DEFAULT_VLD_FE_BDTYPE) << HNS3_UINT32_BIT
+ };
+ vst1q_u64((uint64_t *)&desc->addr, val1);
+ vst1q_u64((uint64_t *)&desc->tx.outer_vlan_tag, val2);
+}
+
+static uint16_t
+hns3_xmit_fixed_burst_vec(void *__restrict tx_queue,
+ struct rte_mbuf **__restrict tx_pkts,
+ uint16_t nb_pkts)
+{
+ struct hns3_tx_queue *txq = (struct hns3_tx_queue *)tx_queue;
+ volatile struct hns3_desc *tx_desc;
+ struct hns3_entry *tx_entry;
+ uint16_t next_to_use;
+ uint16_t nb_commit;
+ uint16_t nb_tx;
+ uint16_t n, i;
+
+ if (txq->tx_bd_ready < txq->tx_free_thresh)
+ hns3_tx_free_buffers(txq);
+
+ nb_commit = RTE_MIN(txq->tx_bd_ready, nb_pkts);
+ if (unlikely(nb_commit == 0)) {
+ txq->queue_full_cnt++;
+ return 0;
+ }
+ nb_tx = nb_commit;
+
+ next_to_use = txq->next_to_use;
+ tx_desc = &txq->tx_ring[next_to_use];
+ tx_entry = &txq->sw_ring[next_to_use];
+
+ /*
+ * We need to deal with n descriptors first for better performance,
+ * if nb_commit is greater than the difference between txq->nb_tx_desc
+ * and next_to_use in sw_ring and tx_ring.
+ */
+ n = txq->nb_tx_desc - next_to_use;
+ if (nb_commit >= n) {
+ for (i = 0; i < n; i++, tx_pkts++, tx_desc++) {
+ hns3_vec_tx(tx_desc, *tx_pkts);
+ tx_entry[i].mbuf = *tx_pkts;
+ }
+
+ nb_commit -= n;
+ next_to_use = 0;
+ tx_desc = &txq->tx_ring[next_to_use];
+ tx_entry = &txq->sw_ring[next_to_use];
+ }
+
+ for (i = 0; i < nb_commit; i++, tx_pkts++, tx_desc++) {
+ hns3_vec_tx(tx_desc, *tx_pkts);
+ tx_entry[i].mbuf = *tx_pkts;
+ }
+
+ next_to_use += nb_commit;
+ txq->next_to_use = next_to_use;
+ txq->tx_bd_ready -= nb_tx;
+
+ hns3_write_reg_opt(txq->io_tail_reg, nb_tx);
+
+ return nb_tx;
+}
+#endif /* _HNS3_RXTX_VEC_NEON_H_ */
'hns3_mp.c')
deps += ['hash']
+
+if arch_subdir == 'arm' and dpdk_conf.get('RTE_ARCH_64')
+ sources += files('hns3_rxtx_vec.c')
+endif