Hi Sakari, On Wed, Apr 15, 2020 at 07:27:22PM +0300, Sakari Ailus wrote: > On Tue, Apr 14, 2020 at 11:55:52PM +0300, Laurent Pinchart wrote: > > 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 ? > > No. In a general case there is no reliable way to come up with an external > clock frequency based on another, different if related, frequency. > > > In any case, this policy needs to be carefully documented. > > I thought after ten or so years this would be already an established > practice. :-) > > I agree it should be documented. We don't seem to have specific > documentation for camera sensor drivers at the moment. I can submit a > patch... > > >>>> 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. > > The external clock frequency is significantly lower than the link frequency > (usually), but it still comes out of the SoC (or a PMIC chip). The clock > signal track on PCB as well as wiring may also be rather long, depending on > where the camera sensor is --- quite possibly tens of centimetres. > Therefore I wouldn't categorically rule out possible EMC issues with that > one either. That's a valid point. > The bottom line is: use a known-good, safe frequency. What if different input clock frequencies are needed to achieve different link frequencies ? -- Regards, Laurent Pinchart