[v2] net/hns3: support Tx push quick doorbell to improve perf
Checks
Commit Message
From: Chengwen Feng <fengchengwen@huawei.com>
Kunpeng 930 support Tx push mode which could improve performance, It
works like below:
1. Add PCIe bar45 which support driver direct write the Tx descriptor
or tail reg to it.
2. Support three operations: a) direct write one Tx descriptor, b)
direct write two Tx descriptors, c) direct write tail reg.
3. The original tail reg located at bar23, the above bar45 tail reg
could provide better bandwidth from the hardware perspective.
The hns3 driver only support direct write tail reg (also have the name
of quick doorbell), the detail:
Considering compatibility, firmware will report Tx push capa if the
hardware support it.
Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
---
v2:
* change abbreviation style.
---
drivers/net/hns3/hns3_ethdev.c | 4 +-
drivers/net/hns3/hns3_ethdev.h | 4 ++
drivers/net/hns3/hns3_ethdev_vf.c | 4 +-
drivers/net/hns3/hns3_rxtx.c | 73 ++++++++++++++++++++++++++++++++++-
drivers/net/hns3/hns3_rxtx.h | 19 +++++++++
drivers/net/hns3/hns3_rxtx_vec_neon.h | 2 +-
drivers/net/hns3/hns3_rxtx_vec_sve.c | 2 +-
7 files changed, 102 insertions(+), 6 deletions(-)
Comments
On Tue, 15 Jun 2021 09:34:29 +0800
"Min Hu (Connor)" <humin29@huawei.com> wrote:
> + return (void *)((char *)pci_dev->mem_resource[bar_id].addr +
> + (pci_dev->mem_resource[bar_id].len >> 1) +
> + HNS3_TX_PUSH_TQP_REGION_SIZE * queue_id +
> + HNS3_TX_PUSH_QUICK_DOORBELL_OFFSET);
Minor nit: cast here (in C) is unnecessary. C will automatically
cast the return any pointer type to void *
On Tue, 15 Jun 2021 09:34:29 +0800
"Min Hu (Connor)" <humin29@huawei.com> wrote:
> +void
> +hns3_tx_push_init(struct rte_eth_dev *dev)
> +{
> + struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
> + volatile uint32_t *reg;
> + uint32_t val;
> +
> + if (!hns3_dev_tx_push_supported(hw))
> + return;
> +
> + reg = (volatile uint32_t *)hns3_tx_push_get_queue_tail_reg(dev, 0);
Better to use proper barrier function than using volatile.
Hi, Stephen,
Volatile is used here to prevent compiler optimization (deleting the reg
write operation).
By the way, this is an initialization process and does not involve
multi-thread synchronization.
Therefore no need to use the barrier.
在 2021/6/15 10:37, Stephen Hemminger 写道:
> On Tue, 15 Jun 2021 09:34:29 +0800
> "Min Hu (Connor)" <humin29@huawei.com> wrote:
>
>> +void
>> +hns3_tx_push_init(struct rte_eth_dev *dev)
>> +{
>> + struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
>> + volatile uint32_t *reg;
>> + uint32_t val;
>> +
>> + if (!hns3_dev_tx_push_supported(hw))
>> + return;
>> +
>> + reg = (volatile uint32_t *)hns3_tx_push_get_queue_tail_reg(dev, 0);
>
> Better to use proper barrier function than using volatile.
> .
>
On 6/15/21 4:34 AM, Min Hu (Connor) wrote:
> From: Chengwen Feng <fengchengwen@huawei.com>
>
> Kunpeng 930 support Tx push mode which could improve performance, It
> works like below:
> 1. Add PCIe bar45 which support driver direct write the Tx descriptor
> or tail reg to it.
> 2. Support three operations: a) direct write one Tx descriptor, b)
> direct write two Tx descriptors, c) direct write tail reg.
> 3. The original tail reg located at bar23, the above bar45 tail reg
> could provide better bandwidth from the hardware perspective.
>
> The hns3 driver only support direct write tail reg (also have the name
> of quick doorbell), the detail:
> Considering compatibility, firmware will report Tx push capa if the
> hardware support it.
>
> Signed-off-by: Chengwen Feng <fengchengwen@huawei.com>
> Signed-off-by: Min Hu (Connor) <humin29@huawei.com>
With description mangled a bit and few minor fixes
described below.
Applied, thanks.
[snip]
> diff --git a/drivers/net/hns3/hns3_rxtx.c b/drivers/net/hns3/hns3_rxtx.c
> index 1d7a769..1fb16cd 100644
> --- a/drivers/net/hns3/hns3_rxtx.c
> +++ b/drivers/net/hns3/hns3_rxtx.c
> @@ -2892,6 +2892,69 @@ hns3_tx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_txconf *conf,
> return 0;
> }
>
> +static void *
> +hns3_tx_push_get_queue_tail_reg(struct rte_eth_dev *dev, uint16_t queue_id)
> +{
> +#define HNS3_TX_PUSH_TQP_REGION_SIZE 0x10000
> +#define HNS3_TX_PUSH_QUICK_DOORBELL_OFFSET 64
> +#define HNS3_TX_PUSH_PCI_BAR_INDEX 4
> +
> + struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
> + uint8_t bar_id = HNS3_TX_PUSH_PCI_BAR_INDEX;
> +
> + /*
> + * If device support Tx push then its PCIe bar45 must exist, and DPDK
> + * framework will mmap the bar45 default in pci probe stage.
pci -> PCI
> + *
> + * In the bar45, the first half is for roce(RDMA over Converged
roce -> RoCE
> + * Ethernet), and the second half is for NIC, every TQP occupy 64KB.
> + *
> + * The quick doorbell located at 64B offset in the TQP region.
> + */
> + return (void *)((char *)pci_dev->mem_resource[bar_id].addr +
> + (pci_dev->mem_resource[bar_id].len >> 1) +
> + HNS3_TX_PUSH_TQP_REGION_SIZE * queue_id +
> + HNS3_TX_PUSH_QUICK_DOORBELL_OFFSET);
Remove unnecessary type cast to 'void *'.
[snip]
@@ -5177,6 +5177,8 @@ hns3_init_pf(struct rte_eth_dev *eth_dev)
goto err_cmd_init;
}
+ hns3_tx_push_init(eth_dev);
+
/*
* To ensure that the hardware environment is clean during
* initialization, the driver actively clear the hardware environment
@@ -7422,8 +7424,8 @@ hns3_dev_init(struct rte_eth_dev *eth_dev)
"process, ret = %d", ret);
goto err_mp_init_secondary;
}
-
hw->secondary_cnt++;
+ hns3_tx_push_init(eth_dev);
return 0;
}
@@ -862,6 +862,7 @@ enum {
HNS3_DEV_SUPPORT_COPPER_B,
HNS3_DEV_SUPPORT_FD_QUEUE_REGION_B,
HNS3_DEV_SUPPORT_PTP_B,
+ HNS3_DEV_SUPPORT_TX_PUSH_B,
HNS3_DEV_SUPPORT_INDEP_TXRX_B,
HNS3_DEV_SUPPORT_STASH_B,
HNS3_DEV_SUPPORT_RXD_ADV_LAYOUT_B,
@@ -900,6 +901,9 @@ enum {
#define hns3_dev_ras_imp_supported(hw) \
hns3_get_bit((hw)->capability, HNS3_DEV_SUPPORT_RAS_IMP_B)
+#define hns3_dev_tx_push_supported(hw) \
+ hns3_get_bit((hw)->capability, HNS3_DEV_SUPPORT_TX_PUSH_B)
+
#define HNS3_DEV_PRIVATE_TO_HW(adapter) \
(&((struct hns3_adapter *)adapter)->hw)
#define HNS3_DEV_PRIVATE_TO_PF(adapter) \
@@ -1921,6 +1921,8 @@ hns3vf_init_vf(struct rte_eth_dev *eth_dev)
goto err_cmd_init;
}
+ hns3_tx_push_init(eth_dev);
+
/* Get VF resource */
ret = hns3_query_vf_resource(hw);
if (ret)
@@ -2928,8 +2930,8 @@ hns3vf_dev_init(struct rte_eth_dev *eth_dev)
"process, ret = %d", ret);
goto err_mp_init_secondary;
}
-
hw->secondary_cnt++;
+ hns3_tx_push_init(eth_dev);
return 0;
}
@@ -2892,6 +2892,69 @@ hns3_tx_queue_conf_check(struct hns3_hw *hw, const struct rte_eth_txconf *conf,
return 0;
}
+static void *
+hns3_tx_push_get_queue_tail_reg(struct rte_eth_dev *dev, uint16_t queue_id)
+{
+#define HNS3_TX_PUSH_TQP_REGION_SIZE 0x10000
+#define HNS3_TX_PUSH_QUICK_DOORBELL_OFFSET 64
+#define HNS3_TX_PUSH_PCI_BAR_INDEX 4
+
+ struct rte_pci_device *pci_dev = RTE_DEV_TO_PCI(dev->device);
+ uint8_t bar_id = HNS3_TX_PUSH_PCI_BAR_INDEX;
+
+ /*
+ * If device support Tx push then its PCIe bar45 must exist, and DPDK
+ * framework will mmap the bar45 default in pci probe stage.
+ *
+ * In the bar45, the first half is for roce(RDMA over Converged
+ * Ethernet), and the second half is for NIC, every TQP occupy 64KB.
+ *
+ * The quick doorbell located at 64B offset in the TQP region.
+ */
+ return (void *)((char *)pci_dev->mem_resource[bar_id].addr +
+ (pci_dev->mem_resource[bar_id].len >> 1) +
+ HNS3_TX_PUSH_TQP_REGION_SIZE * queue_id +
+ HNS3_TX_PUSH_QUICK_DOORBELL_OFFSET);
+}
+
+void
+hns3_tx_push_init(struct rte_eth_dev *dev)
+{
+ struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+ volatile uint32_t *reg;
+ uint32_t val;
+
+ if (!hns3_dev_tx_push_supported(hw))
+ return;
+
+ reg = (volatile uint32_t *)hns3_tx_push_get_queue_tail_reg(dev, 0);
+ /*
+ * Because the size of bar45 is about 8GB size, it may take a long time
+ * to do the page fault in Tx process when work with vfio-pci, so use
+ * one read operation to make kernel setup page table mapping for bar45
+ * in the init stage.
+ * Note: the bar45 is readable but the result is all 1.
+ */
+ val = *reg;
+ RTE_SET_USED(val);
+}
+
+static void
+hns3_tx_push_queue_init(struct rte_eth_dev *dev,
+ uint16_t queue_id,
+ struct hns3_tx_queue *txq)
+{
+ struct hns3_hw *hw = HNS3_DEV_PRIVATE_TO_HW(dev->data->dev_private);
+ if (!hns3_dev_tx_push_supported(hw)) {
+ txq->tx_push_enable = false;
+ return;
+ }
+
+ txq->io_tail_reg = (volatile void *)hns3_tx_push_get_queue_tail_reg(dev,
+ queue_id);
+ txq->tx_push_enable = true;
+}
+
int
hns3_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
unsigned int socket_id, const struct rte_eth_txconf *conf)
@@ -2983,6 +3046,12 @@ hns3_tx_queue_setup(struct rte_eth_dev *dev, uint16_t idx, uint16_t nb_desc,
memset(&txq->basic_stats, 0, sizeof(struct hns3_tx_basic_stats));
memset(&txq->dfx_stats, 0, sizeof(struct hns3_tx_dfx_stats));
+ /*
+ * Call hns3_tx_push_queue_init after assigned io_tail_reg field because
+ * it may overwrite the io_tail_reg field.
+ */
+ hns3_tx_push_queue_init(dev, idx, txq);
+
rte_spinlock_lock(&hw->lock);
dev->data->tx_queues[idx] = txq;
rte_spinlock_unlock(&hw->lock);
@@ -4029,7 +4098,7 @@ hns3_xmit_pkts_simple(void *tx_queue,
hns3_tx_fill_hw_ring(txq, tx_pkts + nb_tx, nb_pkts - nb_tx);
txq->next_to_use += nb_pkts - nb_tx;
- hns3_write_reg_opt(txq->io_tail_reg, nb_pkts);
+ hns3_write_txq_tail_reg(txq, nb_pkts);
return nb_pkts;
}
@@ -4146,7 +4215,7 @@ hns3_xmit_pkts(void *tx_queue, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
end_of_tx:
if (likely(nb_tx))
- hns3_write_reg_opt(txq->io_tail_reg, nb_hold);
+ hns3_write_txq_tail_reg(txq, nb_hold);
return nb_tx;
}
@@ -419,6 +419,7 @@ struct hns3_tx_dfx_stats {
};
struct hns3_tx_queue {
+ /* The io_tail_reg is write-only if working in tx push mode */
volatile void *io_tail_reg;
struct hns3_desc *tx_ring;
struct hns3_entry *sw_ring;
@@ -659,6 +660,23 @@ hns3_rx_calc_ptype(struct hns3_rx_queue *rxq, const uint32_t l234_info,
return ptype_tbl->l3table[l3id] | ptype_tbl->l4table[l4id];
}
+/*
+ * If enable using Tx push feature and also device support it, then use quick
+ * doorbell (bar45) to inform the hardware.
+ *
+ * The other cases (such as: device don't support or user don't enable using)
+ * then use normal doorbell (bar23) to inform the hardware.
+ */
+static inline void
+hns3_write_txq_tail_reg(struct hns3_tx_queue *txq, uint32_t value)
+{
+ rte_io_wmb();
+ if (txq->tx_push_enable)
+ rte_write64_relaxed(rte_cpu_to_le_32(value), txq->io_tail_reg);
+ else
+ rte_write32_relaxed(rte_cpu_to_le_32(value), txq->io_tail_reg);
+}
+
void hns3_dev_rx_queue_release(void *queue);
void hns3_dev_tx_queue_release(void *queue);
void hns3_free_all_queues(struct rte_eth_dev *dev);
@@ -741,5 +759,6 @@ int hns3_tx_done_cleanup(void *txq, uint32_t free_cnt);
void hns3_enable_rxd_adv_layout(struct hns3_hw *hw);
int hns3_dev_rx_descriptor_status(void *rx_queue, uint16_t offset);
int hns3_dev_tx_descriptor_status(void *tx_queue, uint16_t offset);
+void hns3_tx_push_init(struct rte_eth_dev *dev);
#endif /* _HNS3_RXTX_H_ */
@@ -84,7 +84,7 @@ hns3_xmit_fixed_burst_vec(void *__restrict tx_queue,
txq->next_to_use = next_to_use;
txq->tx_bd_ready -= nb_tx;
- hns3_write_reg_opt(txq->io_tail_reg, nb_tx);
+ hns3_write_txq_tail_reg(txq, nb_tx);
return nb_tx;
}
@@ -475,7 +475,7 @@ hns3_xmit_fixed_burst_vec_sve(void *__restrict tx_queue,
txq->next_to_use += nb_pkts - nb_tx;
txq->tx_bd_ready -= nb_pkts;
- hns3_write_reg_opt(txq->io_tail_reg, nb_pkts);
+ hns3_write_txq_tail_reg(txq, nb_pkts);
return nb_pkts;
}