Re: [dpdk-dev] [PATCH V4 4/4] net/i40e: FDIR update rate optimization

["Wu, Jingjing" <jingjing.wu@intel.com>]

[...]

> +static inline unsigned char *
> +i40e_find_available_buffer(struct rte_eth_dev *dev)
> +{
> +	struct i40e_pf *pf = I40E_DEV_PRIVATE_TO_PF(dev->data->dev_private);
> +	struct i40e_fdir_info *fdir_info = &pf->fdir;
> +	struct i40e_tx_queue *txq = pf->fdir.txq;
> +	volatile struct i40e_tx_desc *txdp = &txq->tx_ring[txq->tx_tail + 1];
> +	uint32_t i;
> +
> +	/* no available buffer
> +	 * search for more available buffers from the current
> +	 * descriptor, until an unavailable one
> +	 */
> +	if (fdir_info->txq_available_buf_count <= 0) {
> +		uint16_t tmp_tail;
> +		volatile struct i40e_tx_desc *tmp_txdp;
> +
> +		tmp_tail = txq->tx_tail;
> +		tmp_txdp = &txq->tx_ring[tmp_tail + 1];
> +
> +		do {
> +			if ((tmp_txdp->cmd_type_offset_bsz &
> +
> 	rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) ==
> +
> 	rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
> +				fdir_info->txq_available_buf_count++;
> +			else
> +				break;
> +
> +			tmp_tail += 2;
> +			if (tmp_tail >= txq->nb_tx_desc)
> +				tmp_tail = 0;
> +		} while (tmp_tail != txq->tx_tail);
> +	}
> +
> +	/*
> +	 * if txq_available_buf_count > 0, just use the next one is ok,
> +	 * else wait for the next DD until it's set to make sure the data
> +	 * had been fetched by hardware
> +	 */
> +	if (fdir_info->txq_available_buf_count > 0) {
> +		fdir_info->txq_available_buf_count--;
> +	} else {
> +		/* wait until the tx descriptor is ready */
> +		for (i = 0; i < I40E_FDIR_MAX_WAIT_US; i++) {
> +			if ((txdp->cmd_type_offset_bsz &
> +
> 	rte_cpu_to_le_64(I40E_TXD_QW1_DTYPE_MASK)) ==
> +
> 	rte_cpu_to_le_64(I40E_TX_DESC_DTYPE_DESC_DONE))
> +				break;
> +			rte_delay_us(1);
> +		}
> +		if (i >= I40E_FDIR_MAX_WAIT_US) {
> +			PMD_DRV_LOG(ERR,
> +			    "Failed to program FDIR filter: time out to get DD on tx
> queue.");
> +			return NULL;
> +		}
> +	}
Why wait for I40E_FDIR_MAX_WAIT_US but not return NULL immediately?

[...]


>  i40e_flow_fdir_filter_programming(struct i40e_pf *pf,
>  				  enum i40e_filter_pctype pctype,
>  				  const struct i40e_fdir_filter_conf *filter,
> -				  bool add)
> +				  bool add, bool wait_status)
>  {
>  	struct i40e_tx_queue *txq = pf->fdir.txq;
>  	struct i40e_rx_queue *rxq = pf->fdir.rxq;
> @@ -2011,8 +2092,10 @@ i40e_flow_fdir_filter_programming(struct i40e_pf *pf,
>  	volatile struct i40e_tx_desc *txdp;
>  	volatile struct i40e_filter_program_desc *fdirdp;
>  	uint32_t td_cmd;
> -	uint16_t vsi_id, i;
> +	uint16_t vsi_id;
>  	uint8_t dest;
> +	uint32_t i;
> +	uint8_t retry_count = 0;
> 
>  	PMD_DRV_LOG(INFO, "filling filter programming descriptor.");
>  	fdirdp = (volatile struct i40e_filter_program_desc *)
> @@ -2087,7 +2170,8 @@ i40e_flow_fdir_filter_programming(struct i40e_pf *pf,
> 
>  	PMD_DRV_LOG(INFO, "filling transmit descriptor.");
>  	txdp = &txq->tx_ring[txq->tx_tail + 1];
> -	txdp->buffer_addr = rte_cpu_to_le_64(pf->fdir.dma_addr);
> +	txdp->buffer_addr = rte_cpu_to_le_64(pf->fdir.dma_addr[txq->tx_tail / 2]);
> +
[txq->tx_tail / 2] is not readable, how about use the avail pkt you get directly? Or another index to identify it?
 
>  	td_cmd = I40E_TX_DESC_CMD_EOP |
>  		 I40E_TX_DESC_CMD_RS  |
>  		 I40E_TX_DESC_CMD_DUMMY;
> @@ -2100,25 +2184,34 @@ i40e_flow_fdir_filter_programming(struct i40e_pf *pf,
>  		txq->tx_tail = 0;
>  	/* Update the tx tail register */
>  	rte_wmb();
> +
> +	/* capture the previous error report(if any) from rx ring */
> +	while ((i40e_check_fdir_programming_status(rxq) < 0) &&
> +		(++retry_count < 100))
> +		PMD_DRV_LOG(INFO, "previous error report captured.");
> +
Why check FDIR ring for 100 times? And "&&" is used here, the log is only print if the 100th check fails? 

> 
> --
> 2.17.1