On Thu, May 15, 2014 at 2:58 AM, Arnd Bergmann [off-list ref] wrote:

On Wednesday 14 May 2014 14:26:12 Stephen Boyd wrote:

quoted

On 05/14, Maxime Ripard wrote:

quoted

On Tue, May 13, 2014 at 03:47:32PM -0700, Stephen Boyd wrote:

quoted

On 05/13, Matthias Brugger wrote:

quoted

+       "mediatek,mt6589",
+       NULL,
+};
+
+DT_MACHINE_START(MEDIATEK_DT, "Mediatek Cortex-A7 (Device Tree)")
+       .dt_compat      = mediatek_board_dt_compat,
+MACHINE_END

You shouldn't need this file at all if the platform is part of
the multi-platform kernel.

From a technical point of view, you don't. But it's interesting to
keep it mostly for two things:
  - You get to see the platform name in /proc/cpuinfo
  - If you ever need to add platform quirks, it's already there

We had a similar discussion two weeks ago for mach-sunxi with Olof and
Arnd, and ended up keeping this minimal machine.

It looks like it's only useful to make /proc/cpuinfo have the
platform name because it really isn't that hard to add this file
if we need to add platform quirks. The downside is we have to
keep adding compatibles when we support new SoCs.

We also still add Kconfig entries for each new platform, and I'd like
to leave it at that for the time being. In a lot of cases we end
up adding stuff to the machine descriptor later, e.g. for SMP support
(hopefully no more thanks to your work though).

Once we have a significant number of machines that are actually
usable rather than stubs and that we are confident about never
needing any additional pointers, we can revisit this discussion.

At that point, we should also discuss how to avoid adding a Kconfig
entry for each new platform, which e.g. involves making the
clocksource drivers user selectable. That part has been surprisingly
controversial in the past.

The kconfig part doesn't really worry me. The cpuinfo string is
something we should decide now because it leaks to userspace. There
are 3 possible positions I can think of:

1) The name is part of the ABI and we can never remove/change the string.

2) The name is not an ABI and either:
a) we don't care if it changes with matching a machine desc or not.
b) we still want it to be consistent independent of matching

2a is what we have today. I'm in favor of 2b and therefore should get
the name from DT. It is similar to names in /sys/bus/platform/ which
are not considered part of the ABI. The string already changed moving
platforms to DT and that did not seem to cause any issues that I heard
of.

Rob