Re: [RFC net-next 0/2] Introduce MDIO probe order C45 over C22

[Wong Vee Khee]

On Tue, Jun 01, 2021 at 03:04:51PM +0200, Andrew Lunn wrote:
> On Tue, Jun 01, 2021 at 06:47:34PM +0800, Wong Vee Khee wrote:
> > On Tue, May 25, 2021 at 03:34:34PM +0200, Andrew Lunn wrote:
> > > On Tue, May 25, 2021 at 01:58:03PM +0800, Wong Vee Khee wrote:
> > > > Synopsys MAC controller is capable of pairing with external PHY devices
> > > > that accessible via Clause-22 and Clause-45.
> > > > 
> > > > There is a problem when it is paired with Marvell 88E2110 which returns
> > > > PHY ID of 0 using get_phy_c22_id(). We can add this check in that
> > > > function, but this will break swphy, as swphy_reg_reg() return 0. [1]
> > > 
> > > Is it possible to identify it is a Marvell PHY? Do any of the other
> > > C22 registers return anything unique? I'm wondering if adding
> > > .match_phy_device to genphy would work to identify it is a Marvell PHY
> > > and not bind to it. Or we can turn it around, make the
> > > .match_phy_device specifically look for the fixed-link device by
> > > putting a magic number in one of the vendor registers.
> > >
> > 
> > I checked the Marvell and did not see any unique register values.
> > Also, since get_phy_c22_id() returns a *phy_id== 0, it is not bind to
> > any PHY driver, so I don't think adding the match_phy_device check in
> > getphy would help.
> 
> It has a Marvell ID in C45 space. So maybe we need to special case for
> ID 0. If we get that, go look in C45 space. If we find a valid ID, use
> it. If we get EOPNOTSUP because the MDIO bus is not C45 capable, or we
> don't find a vendor ID in C45 space, keep with id == 0 and load
> genphy?
>

Make sense for me.
Let me what you think of adding the checks in *get_phy_device():

diff --git a/drivers/net/phy/phy_device.c b/drivers/net/phy/phy_device.c
index 1539ea021ac0..ad9a87fadea1 100644
--- a/drivers/net/phy/phy_device.c
+++ b/drivers/net/phy/phy_device.c
@@ -862,11 +862,21 @@ struct phy_device *get_phy_device(struct mii_bus *bus, int addr, bool is_c45)
        c45_ids.mmds_present = 0;
        memset(c45_ids.device_ids, 0xff, sizeof(c45_ids.device_ids));

-       if (is_c45)
+       if (is_c45) {
                r = get_phy_c45_ids(bus, addr, &c45_ids);
-       else
+       } else {
                r = get_phy_c22_id(bus, addr, &phy_id);

+               if (phy_id == 0) {
+                       r = get_phy_c45_ids(bus, addr, &c45_ids);
+                       if (r == -ENOTSUPP || r == -ENODEV)
+                               return 0;
+                       else
+                               return phy_device_create(bus, addr, phy_id,
+                                                        true, &c45_ids);
+               }
+       }
+
        if (r)
                return ERR_PTR(r);


> The other option is consider the PHY broken, and require that you put
> the correct ID in DT as the compatible string. The correct driver will
> then be loaded, based on the compatible string, rather than what is
> found by probing.

Unfortunately all Intel platforms (ElkhartLake/TigerLake/AlderLake) are non-DT.


VK