[PATCH net-next v2 4/6] net: mdio: scan bus based on bus capabilities for C22 and C45

Andrew Lunn andrew at lunn.ch
Sun Mar 12 16:57:36 PDT 2023


On Sun, Mar 12, 2023 at 04:15:41PM +0100, Klaus Kudielka wrote:
> On Sun, 2023-03-12 at 10:04 +0100, Klaus Kudielka wrote:
> > On Sun, 2023-03-12 at 03:53 +0100, Andrew Lunn wrote:
> > > 
> > > Correct. But their also should not of been any noticeable slow down,
> > > because there should not be any additional scanning when everything is
> > > described in DT. And the move of the MDIO bus registration from probe
> > > to setup should actually make it faster than before.
> > > 
> > 
> > But then, why *do* I see such a big difference on the Omnia?
> > 
> > mdiobus_scan_bus_c45() takes:
> > ~2.7 seconds without phy_mask patch
> > ~0.2 seconds with phy_mask patch
> 
> Following up myself, the answer is in the call path
> mv88e6xxx_mdios_register()
> 	 -> mv88e6xxx_mdio_register()
> 		-> of_mdiobus_register()
> 
> A child node "mdio" would be needed for the scan to be limited by
> the device tree. And this one is *not* in armada-385-turris-omnia.dts.
> 
> My (incorrect) understanding was, the child node "ports" would trigger
> that behaviour.

Yes, of_mdiobus_register() calls mdiobus_register() if there is no
MDIO node in DT. And that will result in a full bus scan, limited by
phy_mask.

And for completeness, there is one additional case. When there is a DT
description, reg = <> is optional for a PHY. Most cases, it is used,
but if you have a board designs which can take different pin
compatible PHYs, the address of the PHY might not be known. After
probing PHYs which are listed in DT with reg properties, it will scan
the bus for additional PHYs and assign them to entries which do not
have reg properties.

	Andrew



More information about the Linux-mediatek mailing list