Re: [PATCH v5 2/5] media: i2c: ov5645: Drop reading clock-frequency dt-property

[Laurent Pinchart]

Hi Sakari,

On Tue, Apr 07, 2020 at 06:14:01PM +0300, Sakari Ailus wrote:
> On Tue, Apr 07, 2020 at 03:21:06PM +0300, Laurent Pinchart wrote:
> > On Tue, Apr 07, 2020 at 09:22:41AM +0300, Sakari Ailus wrote:
> >> On Mon, Apr 06, 2020 at 08:32:34PM +0300, Laurent Pinchart wrote:
> >>> On Mon, Apr 06, 2020 at 07:51:08PM +0300, Sakari Ailus wrote:
> >>>> On Mon, Apr 06, 2020 at 05:42:38PM +0100, Lad Prabhakar wrote:
> >>>>> Modes in the driver are based on xvclk frequency fixed to 24MHz, but where
> >>>>> as the OV5645 sensor can support the xvclk frequency ranging from 6MHz to
> >>>>> 24MHz. So instead making clock-frequency as dt-property just let the
> >>>>> driver enforce the required clock frequency.
> >>>> 
> >>>> Even if some current systems where the driver is used are using 24 MHz
> >>>> clock, that doesn't mean there wouldn't be systems using another frequency
> >>>> that the driver does not support right now.
> >>>> 
> >>>> The driver really should not set the frequency unless it gets it from DT,
> >>>> but I think the preferred means is to use assigned-clock-rates instead, and
> >>>> not to involve the driver with setting the frequency.
> >>>> 
> >>>> Otherwise we'll make it impossible to support other frequencies, at least
> >>>> without more or less random defaults.
> >>> 
> >>> We're running in circles here.
> >>> 
> >>> As the driver only supports 24MHz at the moment, the frequency should be
> >>> set by the driver, as it's a driver limitation. We can then work on
> >>> supporting additional frequencies, which will require DT to provide a
> >>> list of supported frequencies for the system, but that can be done on
> >>> top.
> >> 
> >> I guess it would be possible to use different external clock frequencies on
> >> a sensor in a given system but that seems to be a bit far fetched, to the
> >> extent I've never seen anyone doing that in practice.
> >> 
> >> Originally, the driver set the frequency based on the clock-frequency
> >> property. If we're removing that but use a fixed frequency instead, then
> >> how is that going to work going forward when someone adds support for other
> >> frequencies in the driver and has a system requiring that, while there are
> >> some other platforms relying on the driver setting a particular frequency?
> > 
> > The standard property for this is link-frequencies, not clock-frequency.
> > Deprecating clock-frequency now paves the way to use the standard
> > property later when/if someone implements support for additional
> > frequencies.
> 
> The external clock frequency and link frequency are different indeed, but
> they are related. The link frequency has been selected in a way that it is
> possible to generate that exact frequency using the chosen external clock
> frequency. If you change the external clock frequency, chances are good
> there is no PLL configuration to generate that link frequency.

But aren't we supposed to pick the clock frequency based on the link
frequency specified in DT ?

In any case, this policy needs to be carefully documented.

> >> Although, if you're saying that this driver only needs to work with DT that
> >> comes with the kernel and you don't care about DT binary compatibility,
> >> this would be fine.
> > 
> > I believe this series to not break backward compatibility, as the driver
> > only works with a 24MHz clock, so I expect all DTs to specify that.
> 
> What you're still doing here is defining the DT bindings based on the
> current driver implementation, not the device properties.

Quite the contrary, the device doesn't require any particular input
clock frequency, so we're removing that from DT :-) Specifying the clock
frequency in DT is in my opinion a manual workaround for not computing
it at runtime based on the desired link frequency, while the link
frequency is a property of the system as it specifies the range of link
frequencies that are safe to use from an EMC point of view.

-- 
Regards,

Laurent Pinchart