Re: [PATCH v2] arm64: dts: foundation-v8: Enable PSCI mode

From: Daniel Thompson
Date: Tue Oct 03 2017 - 05:12:12 EST

On 02/10/17 18:26, Sudeep Holla wrote:
Sorry for late response, I thought I had sent this mail out long back
but was sitting in my draft :(

No worries. I've been at Linaro connect this last week anyway.


On 20/09/17 12:17, Daniel Thompson wrote:
On 20/09/17 10:42, Sudeep Holla wrote:


On 19/09/17 19:32, Daniel Thompson wrote:
Currently if the Foundation model is running ARM Trusted Firmware then
the kernel, which is configured to use spin tables, cannot start
secondary
processors or "power off" the simulation.

After adding a couple of labels to the include file and splitting out
the
spin-table configuration into a header, we add a couple of new headers
together with two new DTs (GICv2+PSCI and GICv3+PSCI).

The new GICv3+PSCI DT has been boot tested, the remaining three (two of
which existed prior to this patch) have been "tested" by decompiling the
blobs and comparing them against a reference.


How different are these from the ones hosted in [1] ?

They look like they were either independently written or diverged a long
time ago. The existing kernel DTs describe hardware absent from the ARM
TF ones and vice versa.


OK.

With specific reference to PSCI it looks like my patches could perhaps
be improved by adding idle-state support.

Yes I know.

You want a v3 with it added?


On argument is that we want to take the DTS out of device tree as
firmware is responsible for generating them. Alternatively, we may be
duplicating resulting in discrepancies over time by coping it into
kernel.

The general problem is copying from where?

The kernel DTs are a well maintained centralized repository which is
*really* useful. git grep across the kernel DTs is a hugely powerful
tool when trying to better understand an ecosystem as sprawling and
diverse as ARMs. In fact I've even seen those sort of searchs used as a
basis to clean up unused code. Seeing that centralized repository
splinter into separate per-vendor silos would be a huge loss for kernel
developers.


Agreed. But models are configurable and last time this discussion came
up, some argued that the DTs must be modified based on the configuration
automatically by models or some external scripts.

Indeed. I can definitely understand why the *models* might want
to be bundled with DTs (or a DT generator, or a
--just-give-me-a-working-dt-for-my-command-line-options-option).


<snip>
In other words, whilst people could discuss alternative ways to manage
DTs[1], I can't see any universe where ARM TF would be a logical place
to keep them.

[1] ... and I'd further suggest that only perhaps people who are
ÂÂÂ prepared to put resources into fixing it should convene such a
ÂÂÂ discussion.

While I agree with you, I am worried on the scalability. Models provide
tons of options(may not be foundation platform but for sure the base AEM
ones). It may so get unmanageable if we keep adding one DTS per
configuration.

If we are OK restricting the set of configurations to what you have
proposed, then it seems fine.

I will try to push this for v4.15, but I still want to hear more
opinions on the scalability part here.

I'm sure you mean other peoples opinions ;-) (especially so, since after reading things back, I realized I expressed mine rather more forefully than intended) but I think there is a reason why PSCI is special... basically its to do with the docs.

I figured out how to use foundation model by reading the documentation. That documentation says "to run Linux go read the Linaro stuff". As it happens the Linaro "stuff" loops you through to a different ARM doc which tells you how to download ARM TF and bootloader binaries from Linaro. Thus if you follow the docs (and are patient enough) you come out the other side with a model running ARM TF and ready for PSCI.

At that point you're ready for mainline kernel development... and you wonder why there is only one CPU running ;-).


Daniel.