On Fri, Nov 15, 2024 at 12:32 PM Jakub Kicinski <kuba@xxxxxxxxxx> wrote:
>
> On Fri, 15 Nov 2024 11:56:38 +0800 Yafang Shao wrote:
> > > On Thu, 14 Nov 2024 10:17:11 +0800 Yafang Shao wrote:
> > > > - * Not recommended for use in drivers for high speed interfaces.
> > >
> > > I thought I suggested we provide clear guidance on this counter being
> > > related to processing pipeline being to slow, vs host backpressure.
> > > Just deleting the line that says "don't use" is not going to cut it :|
> >
> > Hello Jakub,
> >
> > After investigating other network drivers, I found that they all
> > report this metric to rx_missed_errors:
> >
> > - i40e
> > The corresponding ethtool metric is port.rx_discards, which was
> > mapped to rx_missed_errors in commit 5337d2949733 ("i40e: Add
> > rx_missed_errors for buffer exhaustion").
> >
> > - broadcom
> > The equivalent metric is rx_total_discard_pkts, reported as
> > rx_missed_errors in commit c0c050c58d84 ("bnxt_en: New Broadcom
> > ethernet driver")
> >
> > Given this, it seems we should align with the standard practice and
> > report this metric to rx_missed_errors.
> >
> > Tariq, what are your thoughts?
>
> mlx5 already reports rx_missed_errors and AFAIU rx_discards_phy are very
> different kind of drops than the drops reported as 'missed'.
> The distinction is useful in production in my experience working with
> mlx5 devices.
