On Tue, 14 Nov 2023, Andrew Lunn wrote:
> > +static inline
> > +int phy_start_calibration(struct phy_device *phydev)
> > +{
> > + if (!(phydev->drv &&
> > + phydev->drv->calibration_start &&
> > + phydev->drv->calibration_stop))
> > + return -EOPNOTSUPP;
> > +
> > + return phydev->drv->calibration_start(phydev);
> > +}
> > +
> > +static inline
> > +int phy_stop_calibration(struct phy_device *phydev)
> > +{
> > + if (!(phydev->drv &&
> > + phydev->drv->calibration_stop))
> > + return -EOPNOTSUPP;
> > +
> > + return phydev->drv->calibration_stop(phydev);
> > +}
> > +
>
> What is the locking model?
>
> Andrew
>
This driver currently uses an atomic flag to make sure that the calibration
doesn't run twice. It doesn't acquire any locks before calling
phy_start_calibration(), which is a mistake.
I think a good locking model for this would be similar to the one used for
phy_cable_test. The phy_start_calibration() and phy_stop_calibration() wrappers
would acquire a lock on the PHY device and then test phydev->state, to check for
an ongoing calibration. A new enum member such as PHY_CALIB could be defined for
this purpose. The lock would be released by the phylib wrapper once the
phy_driver callback returns.
The problem with this is that one calibration run can access multiple
phy_device instances at the same time, e.g. if a switch is linked to a multiport
PHY via a PSGMII link.
So acquiring a lock on a single phy device isn't enough. Ideally, these
calls could somehow acquire one lock on all the hardware resources of a
multiport PHY simultaneously. From what I've seen, there is no standard kernel
interface that allows MAC drivers to know about link-sharing between phy
devices. I'll have to do more research on this but if you know of an existing
interface that I can use for this, please tell me.
Best,
--
Romain Gantois, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com
On Wed, Nov 15, 2023 at 04:31:07PM +0100, Romain Gantois wrote:
> On Tue, 14 Nov 2023, Andrew Lunn wrote:
>
> > > +static inline
> > > +int phy_start_calibration(struct phy_device *phydev)
> > > +{
> > > + if (!(phydev->drv &&
> > > + phydev->drv->calibration_start &&
> > > + phydev->drv->calibration_stop))
> > > + return -EOPNOTSUPP;
> > > +
> > > + return phydev->drv->calibration_start(phydev);
> > > +}
> > > +
> > > +static inline
> > > +int phy_stop_calibration(struct phy_device *phydev)
> > > +{
> > > + if (!(phydev->drv &&
> > > + phydev->drv->calibration_stop))
> > > + return -EOPNOTSUPP;
> > > +
> > > + return phydev->drv->calibration_stop(phydev);
> > > +}
> > > +
> >
> > What is the locking model?
> >
> > Andrew
> >
> This driver currently uses an atomic flag to make sure that the calibration
> doesn't run twice. It doesn't acquire any locks before calling
> phy_start_calibration(), which is a mistake.
>
> I think a good locking model for this would be similar to the one used for
> phy_cable_test. The phy_start_calibration() and phy_stop_calibration() wrappers
> would acquire a lock on the PHY device and then test phydev->state, to check for
> an ongoing calibration. A new enum member such as PHY_CALIB could be defined for
> this purpose. The lock would be released by the phylib wrapper once the
> phy_driver callback returns.
>
> The problem with this is that one calibration run can access multiple
> phy_device instances at the same time, e.g. if a switch is linked to a multiport
> PHY via a PSGMII link.
>
> So acquiring a lock on a single phy device isn't enough. Ideally, these
> calls could somehow acquire one lock on all the hardware resources of a
> multiport PHY simultaneously. From what I've seen, there is no standard kernel
> interface that allows MAC drivers to know about link-sharing between phy
> devices. I'll have to do more research on this but if you know of an existing
> interface that I can use for this, please tell me.
Lets get the switch parts merged first, then we can think about this
calibration problem. I need a better understanding of the requirements
before i can suggest something.
Andrew