Re: [PATCH v3 13/13] platform/x86: intel_cht_int33fe: Replacing the old connections with references

[Hans de Goede]

Hi,

On 17-04-19 08:39, Heikki Krogerus wrote:
> On Tue, Apr 16, 2019 at 11:35:36PM +0200, Hans de Goede wrote:
> > Hi,
> >
> > On 12-04-19 15:41, Heikki Krogerus wrote:
> > > Now that the software nodes support references, and the
> > > device connection API support parsing fwnode references,
> > > replacing the old connection descriptions with software node
> > > references. Relying on device names when matching the
> > > connection would not have been possible to link the USB
> > > Type-C connector and the DisplayPort connector together, but
> > > with real references it's not problem.
> > >
> > > The DisplayPort ACPI node is dag up, and the drivers own
> > > software node for the DisplayPort is set as the secondary
> > > node for it. The USB Type-C connector refers the software
> > > node, but it is now tied to the ACPI node, and therefore any
> > > device entry (struct drm_connector in practice) that the
> > > node combo is assigned to.
> > >
> > > The USB role switch device does not have ACPI node, so we
> > > have to wait for the device to appear. Then we can simply
> > > assign our software node for the to the device.
> > >
> > > Reviewed-by: Andy Shevchenko <andy.shevchenko@xxxxxxxxx>
> > > Signed-off-by: Heikki Krogerus <heikki.krogerus@xxxxxxxxxxxxxxx>
> >
> > So as promised I've been testing this series and this commit
> > breaks type-c functionality on devices using this driver.
> >
> > The problem is that typec_switch_get() and  typec_mux_get()
> > after this both return the same pointer, which is pointing
> > to the switch, so typec_mux_get() is returning the wrong
> > pointer.
> >
> > This is not surprising since the references for both are
> > both pointing to the fwnode attached to the piusb30532 devices:
> >
> > 	args[0].fwnode = data->node[INT33FE_NODE_PI3USB30532];
> >
> > So the class_find_device here:
> >
> > static void *typec_switch_match(struct device_connection *con, int ep,
> >                                  void *data)
> > {
> >          struct device *dev;
> >
> >          if (con->fwnode) {
> >                  if (con->id && !fwnode_property_present(con->fwnode, con->id))
> >                          return NULL;
> >
> >                  dev = class_find_device(&typec_mux_class, NULL, con->fwnode,
> >                                          fwnode_match);
> >          } else {
> >                  dev = class_find_device(&typec_mux_class, NULL,
> >                                          con->endpoint[ep], name_match);
> >          }
> >
> >          return dev ? to_typec_switch(dev) : ERR_PTR(-EPROBE_DEFER);
> > }
> >
> > Simply returns the first typec_mux_class device registered.
> >
> > I see 2 possible solutions to this problem:
> >
> > 1) Use separate typec_mux_class and typec_orientation_switch_class-es
> >
> > 2) Merge struct typec_switch and struct typec_mux into a single struct,
> > so that all typec_mux_class devices have the same memory layout, add
> > a subclass enum to this new merged struct and use that to identify
> > which of the typec_mux_class devices with the same fwnode pointer we
> > want.
> >
> > Any other suggestions?
>
> I think the correct fix is that we supply separate nodes for both
> device entries.

That is not going to work since the (virtual) mux / orientation-switch
devices are only registered once the driver binds to the piusb30532 i2c
device, so when creating the nodes we only have the piusb30532 i2c device.

I've been thinking some more about this and an easy fix is to have separate
fwnode_match functions for typec_switch_match and typec_mux_match and have
them check that the dev_name ends in "-mux" resp. "-switch" that requires
only a very minimal change to "usb: typec: Registering real device entries for the muxes"
and then everything should be fine.

Note that another problem with this series which I noticed while testing
is that the usb-role-switch is not being found at all anymore after
this ("Replacing the old connections with references") patch. I still need
start debugging that.

Regards,

Hans