Re: [PATCH net-next 2/5] net: ethernet: oa_tc6: Allow custom mii_bus

[Andrew Lunn <andrew@lunn.ch>]

> > This all seems pretty invasive and ugly. Please could you think what
> > happens if instead of passing in an mdiobus, you pass a phydev. Is the
> > change to the core simpler and cleaner?
> > 
> > Andrew
> 

> Kind of agree. Initially we were thinking about changing the
> existing code (Microchip's vendor code) to alloc mii_bus so that
> code would be same across multiple vendors. Either way, it would be
> invasive changes. So, we decide to go with minimal change to other
> vendor's code.

That would be wrong. The standard defines this, so it should be in the
core. Anything which the standard defines should be in the core, so
that drivers for hardware which actually follow the standard are
minimal. Also, we try to keep workarounds for broken hardware out of
the core, hide it in the driver. That is not always possible, but the
aim should be to make the core clean. We don't want to penalise
vendors which got the implementation correct because of vendors who
got is wrong.

> Trying to understand your suggestion. Are you suggesting to move
> entire mii_bus allocation/APIs implementation to vendor side and
> keep only phy dev usage in oa_tc6.c?

No. I'm thinking maybe extend oa_tc6_init, similar to what you
did. Add a quirks flag, and define TC6_QUIRK_BROKEN_PHY. And allow a
phydev to be passed as well.

If the quirk is set, don't call oa_tc6_mdiobus_register() or
phy_find_first(), nor oa_tc6_mdiobus_unregister().

You probably want to start with a patch which breaks oa_tc6_phy_init()
into two, since you still need the phy_connect_direct() and
phy_attached_info(). Then add the quirk, and lastly your driver making
use of the quirk.

The quirks flag could also be used for devices which have MMD 30
mapped into a vendor reserved MMS.

	Andrew