Re: [PATCH v7 5/7] media: i2c: add DS90UB960 driver

[Laurent Pinchart]

Hi Tomi,

On Wed, Feb 08, 2023 at 05:10:05PM +0200, Tomi Valkeinen wrote:
> On 27/01/2023 11:15, Andy Shevchenko wrote:
> > On Fri, Jan 27, 2023 at 10:24:04AM +0200, Tomi Valkeinen wrote:
> >> On 26/01/2023 12:51, Laurent Pinchart wrote:
> >>> On Thu, Jan 26, 2023 at 12:21:06PM +0200, Andy Shevchenko wrote:
> >>>> On Thu, Jan 26, 2023 at 10:41:47AM +0200, Tomi Valkeinen wrote:
> >>>>> On 25/01/2023 17:27, Andy Shevchenko wrote:
> > 
> > ...
> > 
> >>>>>> But I probably don't understand the ATR structure and what exactly we need to
> >>>>>> pass to it, perhaps it also can be replaced with properties (note, that we have
> >>>>>> some interesting ones that called references, which is an alternative to DT
> >>>>>> phandle).
> >>>>>
> >>>>> Well, maybe this needs a Linux bus implementation. I'm not that familiar
> >>>>> with implementing a bus, but I think that would make it easier to share data
> >>>>> between the deserializer and the serializer. A bus sounds a bit like an
> >>>>> overkill for a 1-to-1 connection, used by a few drivers, but maybe it
> >>>>> wouldn't be too much code.
> >>>>
> >>>> Have you looked at auxiliary bus (appeared a few releases ago in kernel)?
> >>>
> >>> As far as I understand, the auxiliary bus infrastructure is meant for
> >>> use cases where a single hardware device needs to be split into multiple
> >>> logical devices (as in struct device). Platform devices were
> >>> historically (ab)used for this, and the auxiliary bus is meant as a
> >>> cleaner solution. I'm not sure if it would be a good match here, or if
> >>> it would be considered an abuse of the auxiliary bus API.
> >>
> >> The aux bus docs say "A key requirement for utilizing the auxiliary bus is
> >> that there is no dependency on a physical bus, device, register accesses or
> >> regmap support. These individual devices split from the core cannot live on
> >> the platform bus as they are not physical devices that are controlled by
> >> DT/ACPI.", which doesn't sound like a good fit.
> > 
> > Thanks for checking!
> > 
> >> The deserializer and serializers are currently independent devices and
> >> drivers (the pdata is the only shared thing), but I think we may need
> >> something better here. The devices are more tightly tied together than
> >> "normal" video devices, in my opinion, as the serializer is fully controlled
> >> by the deserializer (including power).
> >>
> >> And if we ever want to implement something like power management, we
> >> probably need something more than what we have now. Although I don't know
> >> how that would be done, as all the peripherals behind the serializer would
> >> also lose power...
> > 
> > I believe you have to create a power domain for them and when such device
> > is added, the power domain of it should belong to the serialized.
> 
> I was testing this, and got something working.

As discussed offline, I'm not sure power domains are the right tool for
this. I would model the power supplies as regulators, provided by the
deserializer, and acquired by the serializers. If the devices on the
remote side are all children of the serializer (which I think they
should be), then enabling the regulator in the PM resume handler of the
serializer should be all you need.

> I have the deserializer introducing a separate power-domain for each RX 
> port, and the serializer and the sensor both refer to their port's 
> domain. I can see that the deserializer gets power on/off callbacks 
> correctly when either serializer or sensor resumes.
> 
> The problem I have now is that while the power comes from the 
> deserializer and is thus covered with the power domain, the sensor uses 
> services from the serializer (gpios, clocks, i2c bus), and the 
> serializer is not woken up when the sensor does runtime-pm resume (the 
> power domain is powered up correctly when the sensor resumes).

Is the sensor not a child of the serializer ?

> The serializer creates the i2c adapter to which the sensor is added, so, 
> afaics, there should be a child-parent relationship there. But maybe I 
> have something wrong there, or it just doesn't work as I imagine it 
> would work.

You can check the parent/child relationships fairly easily in sysfs.

-- 
Regards,

Laurent Pinchart