Re: [PATCH 18/18] ipu3: Add driver for dummy INT3472 ACPI device

From: Laurent Pinchart
Date: Tue Dec 01 2020 - 07:41:33 EST


Hi Sakari,

On Tue, Dec 01, 2020 at 02:32:44PM +0200, Sakari Ailus wrote:
> On Tue, Dec 01, 2020 at 08:08:26AM +0000, Dan Scally wrote:
> > On 01/12/2020 06:44, Sakari Ailus wrote:
> > > On Mon, Nov 30, 2020 at 11:06:03PM +0000, Dan Scally wrote:
> > >> On 30/11/2020 20:52, Sakari Ailus wrote:
> > >>>> +static const struct acpi_device_id int3472_device_id[] = {
> > >>>> + { "INT3472", 0 },
> > >>>
> > >>> The INT3472 _HID is really allocated for the tps68470 PMIC chip. It may not
> > >>> be used by other drivers; people will want to build kernels where both of
> > >>> these ACPI table layouts are functional.
> > >>>
> > >>> Instead, I propose, that you add this as an option to the tps68470 driver
> > >>> that figures out whether the ACPI device for the tps68470 device actually
> > >>> describes something else, in a similar fashion you do with the cio2-bridge
> > >>> driver. I think it may need a separate Kconfig option albeit this and
> > >>> cio2-bridge cannot be used separately.
> > >>
> > >> It actually occurs to me that that may not work (I know I called that
> > >> out as an option we considered, but that was a while ago actually). The
> > >> reason I wasn't worried about the existing tps68470 driver binding to
> > >> these devices is that it's an i2c driver, and these dummy devices don't
> > >> have an I2cSerialBusV2, so no I2C device is created by them the kernel.
> > >>
> > >> Won't that mean the tps68470 driver won't ever be probed for these devices?
> > >
> > > Oops. I missed this indeed was not an I²C driver. So please ignore the
> > > comment.
> > >
> > > So I guess this wouldn't be an actual problem. I'd still like to test this
> > > on a system with tps68470, as the rest of the set.
> >
> > On my Go2, it .probes() for the actual tps68740 (that machine has both
> > types of INT3472 device) but fails with EINVAL when it can't find the
> > CLDB buffer that these discrete type devices have. My understanding is
> > that means it's free for the actual tps68470 driver to grab the device;
> > although that's not happening because I had to blacklist that driver or
> > it stops the machine from booting at the moment - haven't gotten round
> > to investigating yet.
>
> Oh, then the problem is actually there. If it probes the tps68470 driver on
> the systems with Windows ACPI tables, then it should be that driver which
> works with the Windows ACPI tables, too.
>
> Checking for random objects such as CLDB in multiple drivers and returning
> an error based on them being there or not wouldn't be exactly neat.
> Although I'm not sure thare are options that are obviosly pretty here. I
> wouldn't two separate drivers checking for e.g. CLDB (tps68470 + this one).
>
> The tps68470 driver is an MFD driver that instantiates a number of platform
> devices. Alternatively, if you make this one a platform device, you can, in
> case the CLDB (or whatever object) is present, in the tps68470 driver
> instantiate a device for this driver instead of the rest.
>
> So I'd think what matters is that both drivers can be selected at the same
> time but the user does not need to manually select them. Both ways could
> work I guess?

Let's make it simpler instead of creating lots of devices. Here's what
I've proposed in a different e-mail in this thread.

> Given that INT3472 means Intel camera power management device (that's
> more or less the wording in Windows, I can double-check), would the
> following make sense ?
>
> A top-level module named intel-camera-pmic (or int3472, or ...) would
> register two drivers, a platform driver and an I2C driver, to
> accommodate for both cases ("discrete PMIC" that doesn't have an
> I2cSerialBusV2, and TPS64870 or uP6641Q that are I2C devices). The probe
> function would perform the following:
>
> - If there's no CLDB, then the device uses the Chrome OS "ACPI
> bindings", and refers to a TPS64870. The code that exists in the
> kernel today (registering GPIOs, and registering an OpRegion to
> communicate with the power management code in the DSDT) would be
> activated.
>
> - If there's a CLDB, then the device type would be retrieved from it:
>
> - If the device is a "discrete PMIC", the driver would register clocks
> and regulators controlled by GPIOs, and create clock, regulator and
> GPIO lookup entries for the sensor device that references the PMIC.
>
> - If the device is a TPS64870, the code that exists in the kernel
> today to register GPIOs would be activated, and new code would need
> to be written to register regulators and clocks.
>
> - If the device is a uP6641Q, a new driver will need to be written (I
> don't know on which devices this PMIC is used, so this can probably
> be deferred).
>
> We can split this in multiple files and/or modules.

Could you reply to 20201130233232.GD25713@xxxxxxxxxxxxxxxxxxxxxxxxxx to
avoid splitting the conversation ?

--
Regards,

Laurent Pinchart