Re: [RFC PATCH net-next 1/5] net: dsa: realtek-smi: fix mdio_free bug on module unload

[Vladimir Oltean]

On Sun, Aug 22, 2021 at 10:42:23PM +0000, Alvin Šipraga wrote:
> > Objection: dsa_switch_teardown has:
> >
> > 	if (ds->slave_mii_bus && ds->ops->phy_read)
> > 		mdiobus_unregister(ds->slave_mii_bus);
>
> This is unregistering an mdiobus registered in dsa_switch_setup:
>
> 	if (!ds->slave_mii_bus && ds->ops->phy_read) {
> 		ds->slave_mii_bus = devm_mdiobus_alloc(ds->dev);
> 		if (!ds->slave_mii_bus) {
> 			err = -ENOMEM;
> 			goto teardown;
> 		}
>
> 		dsa_slave_mii_bus_init(ds);
>
> 		err = mdiobus_register(ds->slave_mii_bus);
> 		if (err < 0)
> 			goto teardown;
> 	}
>
> However, we don't enter this codepath because:
>
> - ds->slave_mii_bus is already set in the call to ds->ops->setup()
> before the code snippet above;
> - ds->ops->phy_read is not set.
>
> We don't want to either, since we want to use of_mdiobus_register().
>
> >
> > The realtek_smi_setup_mdio function does:
> >
> > 	smi->ds->slave_mii_bus = smi->slave_mii_bus;
> >
> > so I would expect that this would result in a double unregister on some
> > systems.
> >
> > I haven't went through your new driver, but I wonder whether you have
> > the phy_read and phy_write methods implemented? Maybe that is the
> > difference?
>
> Right, phy_read/phy_write are not set in the dsa_switch_ops of
> rtl8365mb. So we should be safe.

Correct, DSA only unregisters the ds->slave_mii_bus it has registered,
which is provided when the driver implements phy_read and/or phy_write
but does not set/register ds->slave_mii_bus itself. The patch is fine.

>
> It did get me thinking that it would be nice if dsa_register_switch()
> could call of_mdiobus_register() when necessary, since the snippet above
> (and its call to dsa_slave_mii_bus_init()) is almost same as
> realtek_smi_setup_mdio(). It would simplify some logic in realtek-smi
> drivers and obviate the need for this patch. I am not sure what the
> right approach to this would be but with some pointers I can give it a shot.

I don't think DSA could call of_mdiobus_register, since you would need
to pass it the OF node you want and the read/write ops for the bus and
its name and a place to store it (one DSA switch might have more than
one MDIO bus), and I just fail to see the point of doing that.

The whole point of having ds->slave_mii_bus (either allocated by the
driver or by DSA) is to access the PHY registers of a port under a very
narrow set of assumptions: it implicitly assumes that the port N has a
PHY at MDIO address N, as opposed to doing the usual which is to follow
the phy-handle, and that there is a single internal MDIO bus. DSA will
do this as last resort in dsa_slave_phy_setup. But if you use
of_mdiobus_register, just put a phy-handle in the device tree and be
done, you don't need ds->ops->phy_read or ds->ops->phy_write, nor can
you/should you overload these pointers for DSA to do the
of_mdiobus_register for you.

> >
> >>   static enum dsa_tag_protocol rtl8366_get_tag_protocol(struct dsa_switch *ds,
> >>   						      int port,
> >>   						      enum dsa_tag_protocol mp)
> >> @@ -1505,6 +1512,7 @@ static int rtl8366rb_detect(struct realtek_smi *smi)
> >>   static const struct dsa_switch_ops rtl8366rb_switch_ops = {
> >>   	.get_tag_protocol = rtl8366_get_tag_protocol,
> >>   	.setup = rtl8366rb_setup,
> >> +	.teardown = rtl8366rb_teardown,
> >>   	.phylink_mac_link_up = rtl8366rb_mac_link_up,
> >>   	.phylink_mac_link_down = rtl8366rb_mac_link_down,
> >>   	.get_strings = rtl8366_get_strings,
> >> --
> >> 2.32.0
> >>