On Sat, Mar 4, 2023 at 9:39 AM Frank Rowand <frowand.list@xxxxxxxxx> wrote: > > On 3/2/23 17:57, Stephen Boyd wrote: > > Quoting Rob Herring (2023-03-02 12:18:34) > >> On Thu, Mar 2, 2023 at 1:44 PM Stephen Boyd <sboyd@xxxxxxxxxx> wrote: > >>> > >>> Quoting Rob Herring (2023-03-02 09:13:59) > >>>> > >>>> Good to see bindings for this. I've been meaning to do something about > >>>> the DT unittest ones being undocumented, but I hadn't really decided > >>>> whether it was worth writing schemas for them. The compatibles at > >>>> least show up with 'make dt_compatible_check'. Perhaps we want to just > >>>> define some vendor (not 'linux') that's an exception rather than > >>>> requiring schemas (actually, that already works for 'foo'). > >>> > >>> Sure. Maybe "kunit" should be the vendor prefix? Or "dtbunit"? > >> > >> We'd want to use the same thing on the DT unittests or anything else > >> potentially. How about just 'test'? > > > > Sounds good. > > > >> > >>>> It's > >>>> likely that we want test DTs that fail normal checks and schemas get > >>>> in the way of that as we don't have a way to turn off checks. > >>> > >>> Having the schemas is nice to make sure tests that are expecting some > >>> binding are actually getting that. But supporting broken bindings is > >>> also important to test any error paths in functions that parse > >>> properties. Maybe we keep the schema and have it enforce that incorrect > >>> properties are being set? > >> > >> I wasn't suggesting throwing them out. More why I hadn't written any I guess. > >> > >>> Do we really need to test incorrect bindings? Doesn't the > >>> dt_bindings_check catch these problems so we don't have to write DTB > >>> verifiers in the kernel? > >> > >> Fair enough. Using my frequently stated position against me. :) > >> > >> I do have a secret plan to implement (debug) type checks into the > >> of_property_* APIs by extracting the type information from schemas > >> into C. > >> > > > > Ok. I suspect we may want to test error paths though so I don't know > > Yes, exactly. > > > what to do here. For now I'll just leave the bindings in place and > > change the prefix to "test". > > > >> > >>>> We already have GPIO tests in the DT unittests, so why is clocks > >>>> different? Or should the GPIO tests be moved out (yes, please!)? > >>> > >>> Ah I didn't notice the GPIO tests in there. There are i2c tests too, > >>> right? All I can say is clks are using kunit, that's the difference ;-) > >> > >> Yeah, they should perhaps all move to the subsystems. > > > > Got it. > > > >> > >>>> What happens when/if the DT unittest is converted to kunit? I think > >>>> that would look confusing from the naming. My initial thought is > >>>> 'kunit' should be dropped from the naming of a lot of this. Note that > >>>> the original kunit submission converted the DT unittests. I would > >>>> still like to see that happen. Frank disagreed over what's a unit test > >>>> or not, then agreed, then didn't... I don't really care. If there's a > >>>> framework to use, then we should use it IMO. > >>> > >>> Honestly I don't want to get involved in migrating the existing DT > >>> unittest code to kunit. I'm aware that it was attempted years ago when > >>> kunit was introduced. Maybe if the overlay route works well enough I can > >>> completely sidestep introducing any code in drivers/of/ besides some > >>> kunit wrappers for this. I'll cross my fingers! > >> > >> Yeah, I wasn't expecting you to. I just want to make sure this meshes > >> with any future conversion to kunit. > > > > Phew! > > > >> > >> There's also some plans to always populate the DT root node if not > >> present. That may help here. Or not. There's been a few versions > >> posted with Frank's in the last week or 2. > >> > > > > Ok. I think I have some time to try this overlay approach so let me see > > what is needed. > > Please avoid overlays. See my other replies in this thread for why. If overlays work for the constrained environment of unit tests, then use them. If overlays are not to be used, then remove the support from the kernel. Putting issues in a todo list is not going to get them done. Having users will. Rob
