Re: [RFC PATCH 00/11] Removing Calxeda platform support

From: Rob Herring
Date: Tue Feb 18 2020 - 13:40:26 EST

On Tue, Feb 18, 2020 at 12:14 PM Andre Przywara <andre.przywara@xxxxxxx> wrote:
> On Tue, 18 Feb 2020 11:13:10 -0600
> Rob Herring <robh@xxxxxxxxxx> wrote:
> Hi,
> > Calxeda has been defunct for 6 years now. Use of Calxeda servers carried
> > on for some time afterwards primarily as distro builders for 32-bit ARM.
> > AFAIK, those systems have been retired in favor of 32-bit VMs on 64-bit
> > hosts.
> >
> > The other use of Calxeda Midway I'm aware of was testing 32-bit ARM KVM
> > support as there are few or no other systems with enough RAM and LPAE. Now
> > 32-bit KVM host support is getting removed[1].
> >
> > While it's not much maintenance to support, I don't care to convert the
> > Calxeda DT bindings to schema nor fix any resulting errors in the dts files
> > (which already don't exactly match what's shipping in firmware).
> While every kernel maintainer seems always happy to take patches with a negative diffstat, I wonder if this is really justification enough to remove a perfectly working platform. I don't really know about any active users, but experience tells that some platforms really are used for quite a long time, even if they are somewhat obscure. N900 or Netwinder, anyone?
> So to not give the impression that actually *everyone* (from that small subset of people actively reading the kernel list) is happy with that, I think that having support for at least Midway would be useful. On the one hand it's a decent LPAE platform (with memory actually exceeding 4GB), and on the other hand it's something with capable I/O (SATA) and networking, so one can actually stress test the system. Which is the reason I was using that for KVM testing, but even with that probably going away now there remain still some use cases, and be it for general ARM(32) testing.

Does LPAE with more than 4GB actually need to work if there's not
another platform out there?

> I don't particularly care about the more optional parts like EDAC, cpuidle, or cpufreq, but I wonder if keeping in at least the rather small SATA and XGMAC drivers and basic platform support is feasible.

cpuidle isn't actually stable from what I remember. I think without
cpufreq, we default to 1.1GHz instead of 1.4.

> If YAML DT bindings are used as an excuse, I am more than happy to convert those over.


> And if anyone has any particular gripes with some code, maybe there is a way to fix that instead of removing it? I was always wondering if we could get rid of the mach-highbank directory, for instance. I think most of it is Highbank (Cortex-A9) related.

All the reset/suspend/poweroff and coherency parts are shared. The SCU
and L2 parts could be removed, but not really worth the surgery IMO.