On Mon, Oct 31, 2016 at 1:11 AM, Saeed Mahameed
<sae...@dev.mellanox.co.il> wrote:
> On Mon, Oct 31, 2016 at 12:44 AM, Alexei Starovoitov
> <alexei.starovoi...@gmail.com> wrote:
>> On Sun, Oct 30, 2016 at 06:03:06PM +0200, Tariq Toukan wrote:
>>>
>>> Note that the XDP TX rings are no longer shown in ethtool -S.
>>
>> ouch. Can you make it to show them as some large TX numbers instead?
>> It would really sux to lose stats on them.
>>
>
> Right, Tariq, how did we miss this ?
>
> FYI, I don't think we need the whole TX queue stats for XDP tx rings,
> it is just an overkill, there are only two active counters for XDP TX
> ring (XDP_TX_FWD/XDP_TX_DROP).
>
> XDP_TX_FWD or currently "tx_packets" will count successfully forwarded packets
> XDP_TX_DROP or currently "tx_dropped" will count TX dropped packets
> due to full ring.
>
> do we need tx_bytes as well ? I think yes.
>
> The whole idea of this refactoring i.e. differentiating between TXQ
> netdev rings and XDP TX rings, that XDP is a fast path with minimal
> system overhead, we don't need to have the full set of regular TXQ
> counters for XDP TX rings.
BTW in mlx5 we have the required xdp stats as a part of the rx ring,
static const struct counter_desc rq_stats_desc[] = {
[...]
{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, xdp_drop) },
{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, xdp_tx) },
{ MLX5E_DECLARE_RX_STAT(struct mlx5e_rq_stats, xdp_tx_full) },
[...]
We should do the same here.
