Re: [PATCH v5 2/4] dt-bindings: cpufreq: apple,soc-cpufreq: Add binding for Apple SoC cpufreq

[Hector Martin <marcan@marcan.st>]

On 29/11/2022 20.36, Ulf Hansson wrote:
> On Mon, 28 Nov 2022 at 15:29, Hector Martin <marcan@marcan.st> wrote:
>> +examples:
>> +  - |
>> +    // This example shows a single CPU per domain and 2 domains,
>> +    // with two p-states per domain.
>> +    // Shipping hardware has 2-4 CPUs per domain and 2-6 domains.
>> +    cpus {
>> +      #address-cells = <2>;
>> +      #size-cells = <0>;
>> +
>> +      cpu@0 {
>> +        compatible = "apple,icestorm";
>> +        device_type = "cpu";
>> +        reg = <0x0 0x0>;
>> +        operating-points-v2 = <&ecluster_opp>;
> 
> To me, it looks like the operating-points-v2 phandle better belongs in
> the performance-domains provider node. I mean, isn't the OPPs really a
> description of the performance-domain provider?
> 
> That said, I suggest we try to extend the generic performance-domain
> binding [1] with an "operating-points-v2". In that way, we should
> instead be able to reference it from this binding.
> 
> In fact, that would be very similar to what already exists for the
> generic power-domain binding [2]. I think it would be rather nice to
> follow a similar pattern for the performance-domain binding.

While I agree with the technical rationale and the proposed approach
being better in principle...

We're at v5 of bikeshedding this trivial driver's DT binding, and the
comment could've been made at v3. To quote IRC just now:

> this way the machines will be obsolete before things are fully upstreamed

I think it's long overdue for the kernel community to take a deep look
at itself and its development and review process, because it is quite
honestly insane how pathologically inefficient it is compared to,
basically, every other large and healthy open source project of similar
or even greater impact and scope.

Cc Linus, because this is for your Mac and I assume you care. We're at
v5 here for this silly driver. Meanwhile, rmk recently threw the towel
on upstreaming macsmc for us. We're trying, and I'll keep trying because
I actually get paid (by very generous donors) to do this, but if I
weren't I'd have given up a long time ago. And while I won't give up, I
can't deny this situation affects my morale and willingness to keep
pushing on upstreaming on a regular basis.

Meanwhile, OpenBSD has been *shipping* full M1 support for a while now
in official release images (and since Linux is the source of truth for
DT bindings, every time we re-bikeshed it we break their users because
they, quite reasonably, aren't interested in waiting for us Linux
slowpokes to figure it out first).

Please, let's introspect about this for a moment. Something is deeply
broken if people with 25+ years being an arch maintainer can't get a
700-line mfd driver upstreamed before giving up. I don't know how we
expect to ever get a Rust GPU driver merged if it takes 6+ versions to
upstream the world's easiest cpufreq hardware.

- Hector