Re: [PATCH 1/2] clk: qocm: add qcom_cc_map_norequest

[Dmitry Baryshkov]

On Thu, 11 Jul 2024 at 03:04, Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
>
> Quoting Dmitry Baryshkov (2024-07-10 16:32:18)
> > On Tue, 9 Jul 2024 at 01:30, Stephen Boyd <sboyd@xxxxxxxxxx> wrote:
> > >
> > > Quoting Dmitry Baryshkov (2024-06-27 22:20:22)
> > > > The GPU clock controllers use memory region that is a part of the GMU's
> > > > memory region. Add qcom_cc_map_norequest() to be used by GPUCC, so that
> > > > GPU driver can use devm_ioremap_resource for GMU resources.
> > >
> > > Why does GMU map the gpu clk controller? Does it use those registers? We
> > > don't want to allow two different drivers to map the same region because
> > > then they don't coordinate and write over things.
> >
> > It's not that GMU maps gpu CC separately. It looks more like gpucc is
> > a part of the GMU address space. I think GMU manages some of the
> > clocks or GDSCs directly.
> >
>
> I imagine GMU is a collection of stuff, so the register range is large
> because it's basically a subsystem unto itself. Can the range in DT be
> split up, or changed so that different devices within GMU are split out?

No, we have to remain compatible with existing DT. It's not a problem
of a single new platform, the issue has always been present there.

> Or maybe the gpu clk controller can be made into a child of some GMU
> node, where the GMU node has a driver that populates devices that match
> drivers in different subsystems.

Well... Technically yes, but this brings another pack of issues. There
is no separate GMU driver, so we will likely have a chicken-and-egg
problem, as probing of the GPU driver will also create the gpucc
device which is further used by the GPU.

-- 
With best wishes
Dmitry