RE: [PATCH v2 net 1/6] net: enetc: increment error counter in enetc_xdp_xmit() on DMA mapping errors

[Wei Fang]

> diff --git a/drivers/net/ethernet/freescale/enetc/enetc.c
> b/drivers/net/ethernet/freescale/enetc/enetc.c
> index aa8a87124b10..ece10a58d44e 100644
> --- a/drivers/net/ethernet/freescale/enetc/enetc.c
> +++ b/drivers/net/ethernet/freescale/enetc/enetc.c
> @@ -1801,8 +1801,10 @@ int enetc_xdp_xmit(struct net_device *ndev, int
> num_frames,
>  		xdp_tx_bd_cnt = enetc_xdp_frame_to_xdp_tx_swbd(tx_ring,
>  							       xdp_redirect_arr,
>  							       frames[k]);
> -		if (unlikely(xdp_tx_bd_cnt < 0))
> +		if (unlikely(xdp_tx_bd_cnt < 0)) {
> +			tx_ring->stats.xdp_tx_drops++;

xdp_tx_drops is also used for XDP_TX statistics, so it is difficult to distinguish
how many packets are dropped by XDP_TX and how many are dropped by
ndo_xdp_xmit(). Would it be better to add another statistic variable to track
the packets dropped by ndo_xdp_xmit()?

Additionally, I checked the code of bq_xmit_all(), when ndo_xdp_xmit()
returns due to failure, packets after 'sent' will be dropped by
xdp_return_frame_rx_napi(). Therefore, 'xdp_tx_drops' here seems more
like a statistic of the number of enetc_xdp_xmit() failures than a packet loss
statistic.

>  			break;
> +		}
> 
>  		if (unlikely(!enetc_xdp_tx(tx_ring, xdp_redirect_arr,
>  					   xdp_tx_bd_cnt))) {
> --
> 2.43.0