Re: [net-next PATCH v2] net: dsa: qca8k: pack driver struct and improve cache use

[Vladimir Oltean]

On Mon, Mar 21, 2022 at 07:06:07PM +0100, Ansuel Smith wrote:
> > If you provide your own ds->slave_mii_bus there should be no reason to
> > need a ds->ops->phy_read, since your slave_mii_bus already has one.
> > 
> > Bottom line, phy_read is just there in case they it is helpful.
> > Whereas ds->slave_mii_bus is there if you want to have a non-OF based
> > phy_connect, with the implicit assumption that the phy_addr is equal to
> > the port, and it can't be in any other way.
> > 
> > So if ds->ops->phy_read / ds->ops->phy_write aren't useful, don't use them.
> > I suppose one of the drivers you saw with "custom mdiobus" was sja1105.
> > I have no interest whatsoever in converting that driver to use
> > ds->slave_mii_bus or ds->phy_read.
> 
> My idea was to provide some type of API where the switch insert some
> type of data and DSA decide to declare a dsa mdiobus or a
> dedicated one based on the config provided. Think I will have to provide
> an RFC to better describe this idea. The idea would be to drop phy_read
> and phy_write and replace them with a single op that provide a
> configuration.

Just register a non-OF based MDIO bus for the legacy_phy_port_mapping = true
case using plain mdiobus_register() and provide it to ds->slave_mii_bus,
and delete the ds->ops->phy_read and ds->ops->phy_write, instead
adapting qca8k_phy_read and qca8k_phy_write to be your bus->read and
bus->write. Problem solved. No need to touch the existing logic and
implicitly the other drivers that use it.