Re: [PATCH v3] net: ethernet: mediatek: ppe: add support for flow accounting

[Felix Fietkau]

On 02.11.22 02:26, Andrew Lunn wrote:
> On Wed, Nov 02, 2022 at 01:03:08AM +0000, Daniel Golle wrote:
> > Hi Andrew,
> >
> > On Wed, Nov 02, 2022 at 01:51:21AM +0100, Andrew Lunn wrote:
> > > On Wed, Nov 02, 2022 at 12:42:40AM +0000, Daniel Golle wrote:
> > > > The PPE units found in MT7622 and newer support packet and byte
> > > > accounting of hw-offloaded flows. Add support for reading those
> > > > counters as found in MediaTek's SDK[1].
> > > > 
> > > > [1]: https://git01.mediatek.com/plugins/gitiles/openwrt/feeds/mtk-openwrt-feeds/+/bc6a6a375c800dc2b80e1a325a2c732d1737df92
> > > > Signed-off-by: Daniel Golle <daniel@xxxxxxxxxxxxxx>
> > > 
> > > Sorry, but NACK.
> > > 
> > > You have not explained why doing this correctly via ethtool -S cannot
> > > be done. debugfs is a vendor crap way of doing this.
> >
> > The debugfs interface is pre-existing and **in addition** to the
> > standard Linux interfaces which are also provided. It is true that
> > the debugfs interface in this case doesn't provide much additional
> > value apart from having the counter listed next to the hardware-
> > specific hashtable keys. As the debugfs interface for now aims to
> > be as complete as possible, naturally there is some redundance of
> > things which can also be accessed using other (standard) interfaces.
>
> debugfs is by definition unstable. It is not ABI. Anything using it is
> expected to break in the near future when it changes its layout. It is
> also totally option, you cannot expect it to be mounted.
>
> I hope you don't have any user space code using it.
>
> Maybe i should submit a patch which just for the fun of it rearranged
> the order in debugfs and change the file name?

I believe that OpenWrt is still the main user of the PPE offloading 
code, since most vendors of devices with these SoC still use different 
out-of-tree implementations.
OpenWrt does not ship or contain any user space code that relies on 
these debugfs files.

Whenever I'm debugging PPE related issues, I rely heavily on the debugfs 
API, so keeping it as complete as possible is important to me as well.

Aside from that, exposing per-flow statistics (which is what this patch 
does) via ethtool -S API makes no sense to me at all. I'm not aware of 
any other offload capable driver that does this.

- Felix