Re: [PATCH v1 03/10] drm/panthor: Introduce framework for architecture-specific features

[Boris Brezillon <boris.brezillon@collabora.com>]

On Fri, 24 Oct 2025 10:26:16 +0100
Karunika Choo <karunika.choo@arm.com> wrote:

> On 24/10/2025 07:43, Boris Brezillon wrote:
> > On Tue, 14 Oct 2025 10:43:30 +0100
> > Karunika Choo <karunika.choo@arm.com> wrote:
> >   
> >> Add a framework to support architecture-specific features. This allows
> >> other parts of the driver to adjust their behaviour based on the feature
> >> bits enabled for a given architecture.  
> > 
> > I'm not convinced we need this just yet. AFAICT, the only feature flag
> > being added in this patchset is PANTHOR_HW_FEATURE_PWR_CONTROL, and
> > most of this is abstracted away with function pointers already. The
> > only part that tests this FEATURE_PWR_CONTROL flag is the
> > initialization, which could very much be abstracted away with a
> > function pointer (NULL meaning no PWR block present). There might be
> > other use cases you're planning to use this for, so I'd like to hear
> > about them to make my final opinion on that.
> >   
> 
> I see your point — the intent here is mainly to have the feature flag
> reflect hardware-level changes. In this series, for example, it
> corresponds to the addition of the new PWR_CONTROL block.

Yes, but those are not really optional features. Those are functional
changes that are usually done on major version changes. But let's say
it was something done on a minor version change, it's still something
that I think would be better off abstracted using a vtable of some
sort, and have this vtable forked everytime a version changes requires
something new.

> 
> Another use case would be arch v11, where a new PRFCNT_FEATURES register
> was introduced. In that case, we might want to adjust the
> counters_per_block [1] value depending on that register’s value.

Again, it looks like a property that can be determined at init time. For
v10 it'd be hardcoded to X, and on v11+, you'd extract that from
PERFCNT_FEATURES. I'm really not a huge fan of this feature flag
pattern because it's very easy to forget to add/propagate one flag when
adding support for new HW/flags. So I'd much rather rely on ">= X.Y"
version checks in the init path, and for anything more involved or
happening in some hot path, function based pointer specialization.

> 
> I would also expect this mechanism to remain useful for future hardware
> revisions, as it provides a clean way to describe architectural
> differences without scattering version-specific checks throughout the
> code, while still being lighter-weight than function pointers.

Well, that's questionable. What I usually see in practice is the
following pattern spreading over the code base:

	if (SUPPORTS(OBSCURE_FEATURE_NAME)) {
		// do stuff that are not obviously related to the
		// feature flag name
	}

whereas, if we're having a model where the specialization is done high
enough, you'd just end up with functions calling more specialized
helpers:

void do_something_for_v12()
{
	hw_block_a_do_y()
	hw_block_b_do_x()
	...
}