Re: [PATCH RFC 2/3] rust: Add bindings for device properties

[Dirk Behme <dirk.behme@gmail.com>]

On 30.10.24 17:47, Rob Herring wrote:
> On Wed, Oct 30, 2024 at 11:03 AM Dirk Behme <dirk.behme@gmail.com> wrote:
>>
>> On 30.10.24 15:05, Rob Herring wrote:
>>> On Wed, Oct 30, 2024 at 3:15 AM Alice Ryhl <aliceryhl@google.com> wrote:
>>>>
>>>> On Tue, Oct 29, 2024 at 8:35 PM Rob Herring <robh@kernel.org> wrote:
>>>>>
>>>>> On Tue, Oct 29, 2024 at 1:57 PM Miguel Ojeda
>>>>> <miguel.ojeda.sandonis@gmail.com> wrote:
>>>>>>
>>>>>> On Tue, Oct 29, 2024 at 7:48 PM Alice Ryhl <aliceryhl@google.com> wrote:
>>>>>>>
>>>>>>> One option is to define a trait for integers:
>>>>>
>>>>> Yeah, but that doesn't feel like something I should do here. I imagine
>>>>> other things might need the same thing. Perhaps the bindings for
>>>>> readb/readw/readl for example. And essentially the crate:num already
>>>>> has the trait I need. Shouldn't the kernel mirror that? I recall
>>>>> seeing some topic of including crates in the kernel?
>>>>
>>>> You can design the trait to look similar to traits in external crates.
>>>> We did that for FromBytes/AsBytes.
>>>>
>>>> I assume you're referring to the PrimInt trait [1]? That trait doesn't
>>>> really let you get rid of the catch-all case, and it's not even
>>>> unreachable due to the u128 type.
>>>
>>> It was num::Integer which seems to be similar.
>>>
>>>>
>>>> [1]: https://docs.rs/num-traits/0.2.19/num_traits/int/trait.PrimInt.html
>>>>
>>>>>> +1, one more thing to consider is whether it makes sense to define a
>>>>>> DT-only trait that holds all the types that can be a device property
>>>>>> (like `bool` too, not just the `Integer`s).
>>>>>>
>>>>>> Then we can avoid e.g. `property_read_bool` and simply do it in `property_read`.
>>>>>
>>>>> Is there no way to say must have traitA or traitB?
>>>>
>>>> No. What should it do if you pass it something that implements both traits?
>>>>
>>>> If you want a single function name, you'll need one trait.
>>>
>>> I'm not sure I want that actually.
>>>
>>> DT boolean is a bit special. A property not present is false.
>>> Everything else is true. For example, 'prop = <0>' or 'prop =
>>> "string"' are both true. I'm moving things in the kernel to be
>>> stricter so that those cases are errors. I recently introduced
>>> (of|device)_property_present() for that reason. There's no type
>>> information stored in DT.  At the DT level, it's all just byte arrays.
>>> However, we now have all the type information for properties within
>>> the schema. So eventually, I want to use that to warn on accessing
>>> properties with the wrong type.
>>>
>>> For example, I think I don't want this to work:
>>>
>>> if dev.property_read(c_str!("test,i16-array"))? {
>>>     // do something
>>> }
>>>
>>> But instead have:
>>>
>>> if dev.property_present(c_str!("test,i16-array")) {
>>>     // do something
>>> }
>>
>> I think we have "optional" properties which can be there (== true) or
>> not (== false). Let's assume for this example "test,i16-array" is such
>> kind of "optional" property. With what you gave above we need two
>> device tree accesses, then? One to check if it is there and one to
>> read the data:
> 
> Yes, lots of properties are optional especially since any new property
> added has to be because the DT is an ABI.
> 
>> let mut array = <empty_marker>;
>> if dev.property_present(c_str!("test,i16-array")) {
>>     array = dev.property_read(c_str!("test,i16-array"))?;
>> }
>>
>> ?
>>
>> Instead of these two accesses, I was thinking to use the error
>> property_read() will return if the optional property is not there to
>> just do one access:
>>
>> let mut array = <empty_marker>;
>> if let Ok(val) = dev.property_read(c_str!("test,i16-array")) {
>>        array = val;
>> }
>>
>> (and ignore the error case as its irrelvant in the optional case)
>>
>> Have I missed anything?
> 
> If you grep "_property_present", most if not all calls never need the
> data. When you need the data, you read it and test for EINVAL if you
> want to handle "not present". The overhead of parsing the data is not
> nothing, so I think it is better to provide both.
> 
> The typical pattern in the C code is:
> 
> u32 val = DEFAULT_VALUE;
> of_property_read_u32(node, "a-property", &val);
> 
> // val is now either the read property or the default. If the property
> is required, then the error code needs to be checked.

Yes :)

> Maybe we should have:
> 
> let val: u32 = dev.property_read_optional(c_str!("test,i16-array"),
> DEFAULT_VALUE);
> 
> Or looks like Option<> could be used here?:
> 
> let val: u32 = dev.property_read(c_str!("test,i16-array"),
> Option<DEFAULT_VALUE>);


In the success case we will get back Some(val)? In the error case
'val' will get Some(DEFAULT_VALUE)? But where would we get the error
value itself (e.g. EINVAL)? Or is the idea to not care about that any
more then? When would we use None?

Best regards

Dirk

> One thing I'd like to improve is having fewer driver error messages
> and a printk for a missing required property is a common one. We have
> APIs like clk_get and clk_get_optional (which parse firmware
> properties). The difference is the former prints an error message on
> error case and the latter is silent.
> 
> Rob