On Mon, 15 Sep 2014 10:15:16 +0100, Catalin Marinas [off-list ref] wrote:

On Mon, Sep 15, 2014 at 05:31:16AM +0100, Grant Likely wrote:

quoted

On Thu, 11 Sep 2014 16:37:39 +0100, Catalin Marinas [off-list ref] wrote:

quoted

On Thu, Sep 11, 2014 at 02:29:34PM +0100, Grant Likely wrote:

quoted

Regarding the requests to refactor ACPICA to work better for ARM. I
completely agree that it should be done, but I do not think it should be
a prerequisite to getting this core support merged. That kind of
refactoring is far easier to justify when it has immediate improvement
on the mainline codebase, and it gives us a working baseline to test
against. Doing it the other way around just makes things harder.

I have to disagree here. As I said, I'm perfectly fine with refactoring
happening later but I'm not happy with compiling in code with undefined
behaviour on ARM that may actually be executed at run-time.

I'm being told one of the main advantages of ACPI is forward
compatibility: running older kernels on newer hardware (potentially with
newer ACPI version tables). ACPI 5.1 includes partial support for ARM
but the S and C states are not defined yet. We therefore assume that
hardware vendors deploying servers using ACPI would not provide such
yet to be defined information in ACPI 5.1 tables.

We're good on this front. ACPI-future platforms aren't allowed to use
new features when booting an older kernel.

Do you mean ACPI-future firmware should not provide new information to
older kernels?

I mean that ACPI has the mechanism so that the platform (via AML)
doesn't try to do things that the OS doesn't understand.  AML methods are
to use the \_REV object to determine whether or not it can use a
feature. If, for example, the _REV is 5, then the AML must switch itself
to code paths that don't use ACPI 6 features.

The firmware must not care which kernel version (not even
which OS) runs on top. That's the whole point of an OS-agnostic
standardisation body).

It shouldn't care about the OS, but it absolutely needs to care about
which revision of ACPI spec that the OS implements, and the ACPI spec
accounts for exactly that.

quoted

ACPI has a mechanism to check version numbers.

I assume you are referring to the kernel checking of the ACPI version
the tables are compatible with (acpi_parse_fadt() in patch 7/17 in the
v3 series). With these patches, it currently assumes that it is
compatible with anything 5.1 and later.

No, that's the other side. That's how the OS determines which version of
the spec the ACPI tables implement. That's how the OS decides if it
needs to use the behaviour from an older revision of the spec.

For the kernel to accept anything 5.1 or later is the correct behavour
here. ACPI compatibility is driven by the lowest common denominator. The
kernel and the platform (AML) choose the most recent version of the spec
that is supported by both.

quoted

Similarly, when ACPI-future defines those states, the kernel shouldn't
expose them to an older platform.

When the kernel implements ACPI-future, all is fine, it should still be
backwards compatible with 5.1.

quoted

quoted

What if in a year time we get ACPI 5.2 or 6 (or an errata update)
covering the S and C states for ARM? I would expect hardware vendors
to take advantage and provide such information in ACPI tables. Can we
guarantee that a kernel with the current ACPI patches wouldn't blow up
when it tries to interpret the new tables?

Yes - at least as much as we can guarantee anything in systems
programming.

No, we don't. Once you define some sleep states in ACPI-future and
the firmware provides such tables to an older kernel, it is possible
that the current (x86-centric) ACPI sleep code will fault. Just look at
the acpi_suspend_enter() (and you also seemed OK to define
ACPI_FLUSH_CPU_CACHE as BUG).

Even if v4 of these patches removed the sleep code, given uncertainties
like above I'm thinking more and more that we should actually make the
acpi_parse_fadt() version check succeed only on ACPI 5.1 and no later
version (apart from ACPI sleep, is there anything else we would expect
to blow up when additional information is provided via ACPI-future
tables?).

I was about to agree with you that we could do that temporarily, but
after thinking about it more, this is actually an actively dangerous
thing to do because it will encourage firmware vendors to lie about
which version of the ACPI spec it conforms to. ie. A vendor may build a
platform and use the ACPI6.0 spec, but they want it to boot on an older
kernel, and so they make their tables lie and report 5.1.  This would be
a horrible situation. We would loose the ability to use reported ACPI
spec comformance to determine which features to enable.

Selecting on v5.1 or later really is the right choice here. Vendors who
do the wrong thing and produce broken firmware will get rightly dumped
on by Red Hat and Canonical for shipping something that doesn't work.
We have a lot less risk here than ACPI had historically in the x86 world
because these systems are primarily built to run Linux instead of
Windows. That means the systems will be tested and validated with
released Linux distros, and they will fix the firmware if it is broken.

g.