On Wed, 29 Sep 2021 18:44:30 +0200 Olivier MOYSAN <olivier.moysan@xxxxxxxxxxx> wrote: > Hi Jonathan, > > >>>> > >>>> If 'backend' option turns out to be the most appropriated to match DFSDM > >>>> constraints, I can prepare some patches to support it. > >>>> Would you have some guidelines or requirements for the implementation of > >>>> such feature, in this case ? > >>> > >>> Closest example is that rcar-gyroadc but in this case we'd want to define > >>> something standard to support the modulators so that if we have other filters > >>> in future we can reuse them. > >>> > >>> That means implementing them as child devices of the filter - probably put > >>> the on the IIO bus, but as different device type. Take a look at how > >>> triggers are done in industrialio-trigger.c > >>> You need struct device_type sd_modulator > >>> and a suitable device struct (burred in an iio_sd_modulator struct probably). > >>> > >>> Also needed would be a bunch of standard callbacks to allow you to query things > >>> like scaling. Keep that interface simple. Until we have a lot of modulator > >>> drivers it will be hard to know exactly what is needed. Also whilst we don't > >>> have many it is easy to modify the interface. > >>> > >>> Then have your filter driver walk it's own dt children and instantiate > >>> appropriate new elements and register them on the iio_bus. They will have > >>> the filter as their parent. > >>> > >>> There are various examples of this sort of thing in tree. > >>> If you want a good one, drivers/cxl does a lot of this sort magic to manage > >>> a fairly complex graph of devices including some nice registration stuff to > >>> cause the correct device drivers to load automatically. > >>> > >>> Hmm. Thinking more on this, there is an ordering issue for driver load. > >>> Instead of making the modulator nodes children of the modulator, you may need > >>> to give them their own existence and use a phandle to reference them. > >>> That will let you defer probe in the filter driver until those > >>> modulator drivers are ready. > >>> > >>> This isn't going to be particularly simple, so you may want to have a look > >>> at how various other subsystems do similar things and mock up the dependencies > >>> to make sure you have something that doesn't end up with a loop of dependencies. > >>> In some ways the modulators are on a bus below the filter, but the filter driver > >>> needs them to be in place to do the rest. > >>> You may end up with some sort of delayed load. > >>> 1. Initial filter driver load + parsing of the modulator dt children (if done that way). > >>> 2. Filter driver goes to sleep until... > >>> 3. Modulator drivers call something on the filter driver to say they are ready. > >>> 4. Filter driver finishes loading and create the IIO device etc. > >>> You'll need some reference counting etc in there to make removal safe etc but it > >>> shouldn't be 'too bad'. > >>> > >>> Good luck! > >>> > >>> Jonathan > >>> > I'am on the way to prototype this proposal for DFSDM. > Looking at your advices, I see that the current topolgy based on > hardware consumer, already meets most of the requirements. > > - SD modulators are described in DT with their own nodes and are > referred in DFSDM nodes through their phandle. > - Dependencies at probe are managed (defer probe through > devm_iio_hw_consumer_alloc()) > - SD modulator scaling is retrieved through iio_read_channel_scale() ABI. > > So, it seems that the current implementation is not so far from this > solution. > It remains the unwanted sysfs interface for SD modulator. Or more than > that, if I missed something ? > Instead of introducing a new device type for SD modulator, could the > mode field be used to identify devices not requesting an IIO sysfs ? > (A dedicated mode may be used to skip sysfs register in device registration) > Otherwise let's go for a new type. I'd rather see them as a new device type than overload the IIO device. We want to be able to control what can 'connect' to the DFSDM afterall and device type is a convenient route to doing this. Obviously if there is infrastructure that can be factored out and used for both this and a normal IIO device we can do that to save on duplication. Thanks, Jonathan > > Regards > Olivier