Re: [PATCH 1/2] irqchip/qcom-pdc: don't read version register if it is not available

[Marc Zyngier <maz@kernel.org>]

On Mon, 28 Aug 2023 11:18:10 +0100,
Dmitry Baryshkov <dmitry.baryshkov@linaro.org> wrote:
> 
> On Mon, 28 Aug 2023 at 13:04, Marc Zyngier <maz@kernel.org> wrote:
> >
> > On Mon, 28 Aug 2023 10:46:10 +0100,
> > Dmitry Baryshkov <dmitry.baryshkov@linaro.org> wrote:
> > >
> > > On Mon, 28 Aug 2023 at 12:36, Maulik Shah (mkshah)
> > > <quic_mkshah@quicinc.com> wrote:
> > > >
> > > > Hi Dmitry,
> > > >
> > > > This patch may be useful if there was a case where some PDCs don't have
> > > > version register populated/available,
> > > > In all PDC versions, version register is always available but due to reg
> > > > size not good enough in device tree for SM8150 it failed to read.
> > > >
> > > > reg size in device node must be expanded if its too small to access all
> > > > registers and i think
> > > > additional check in driver to check if size is good enough would not be
> > > > of much use.
> > >
> > > Unfortunately, it doesn't work this way. DT files are ABI. Even if we
> > > change the DT, the kernel should continue working with the older
> > > version.
> > > Thus, we have to add such bandaid code, which will keep the kernel
> > > from crashing if old DT was used.
> >
> > You're missing the point: all existing PDC HW have version register.
> > The fact that the DT is crap doesn't invalidate this simple fact. It
> > is thus perfectly possible for the driver to *ignore* the crap and do
> > the right thing by expanding the size of the mapping, rather than
> > falling back to the non-versioned code.
> 
> Ah. Interesting idea. If that's the overall consensus I can send v2
> doing this. Not sure what is better though.

Given that DT files are mostly generated using copy-paste by people
making a point not to read specifications, odds are that your current
patch would end-up applying the v0 behaviour to v3.2 HW.

What could possibly go wrong?

	M.

-- 
Without deviation from the norm, progress is not possible.