Re: [PATCH 0/6] ArmVirtPkg: don't forward the DT to the OS if QEMU provides ACPI

[Laszlo Ersek <lersek@redhat.com>]

On 03/09/17 13:26, Ard Biesheuvel wrote:
> On 9 March 2017 at 12:01, Laszlo Ersek <lersek@redhat.com> wrote:
>> On 03/09/17 09:16, Ard Biesheuvel wrote:
>>> On 8 March 2017 at 20:05, Laszlo Ersek <lersek@redhat.com> wrote:
>>>> This series replaces the PURE_ACPI_BOOT_ENABLE build option with dynamic
>>>> behavior, matching QEMU's (inverse sense) "-no-acpi" switch. In
>>>> particular, DT and ACPI are no longer exposed to the guest at the same
>>>> time. (DT is exposed with "-no-acpi", or else ACPI is exposed without
>>>> "-no-acpi".)
>>>>
>>>> Repo:   https://github.com/lersek/edk2.git
>>>> Branch: dynamic_pure_acpi
>>>> RHBZ:   https://bugzilla.redhat.com/show_bug.cgi?id=1430262
>>>>
>>>> Tested with RHEL-7.3 for ARM and Fedora 24 guests (DT vs. ACPI).
>>>>
>>>> Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
>>>>
>>>
>>> Hi Laszlo,
>>>
>>> This looks complicated to me. Given that it is arguably a policy to
>>> only expose on h/w description or the other, couldn't we simply remove
>>> the FDT config table in BDS if an ACPI/ACPI2.0 config table is
>>> present?
>>
>> Technically we could do that, but I dislike it for two reasons:
>>
>> - BDS is often the first victim found when looking for a driver to add
>> new code to that doesn't seem to fit very well elsewhere. That doesn't
>> make BDS any better a recipient, however. "For lack of a better driver"
>> is not a strong enough argument to dump code into BDS. If there's really
>> no better "topical" driver, then the code usually goes to PlatformDxe.
>>
>> - Installing a sysconfig table (or any other system-wide resource) in
>> one driver, then undoing it in another driver, should be avoided as much
>> as possible, because it leads to non-trivial lifecycles and boggles our
>> minds over the longer term. If we can come to a decision that the table
>> shouldn't be installed in the first place, we should pursue that.
>>
>> Another approach we could look into is: move the installation of the
>> sysconfig table from FdtClientDxe to AcpiPlatformDxe. Look for the ACPI
>> payload first, and fall back to installing DT (from within
>> AcpiPlatformDxe). However, DT should be installed even in builds (like
>> ARM32) that don't contain AcpiPlatformDxe at all.
>>
> 
> Or we could hook to the ReadyToBoot event in FdtClientDxe, and install
> the DT config table if there is no ACPI/ACPI2.0 table registered.

Yes, that's doable in our case, because we control the full platform.

Installing tables (any kinds of tables) in ReadyToBoot and similar event
handlers is generally a bad idea, because everyone thinks, "okay I'll
wait until the rest of the system is done setting up, and I'll just add
my stuff afterwards". Obviously, this results in much of the logic being
simply moved to such event callbacks, and the invocation order of
callbacks remains unspecified.

In more precise terms, if the ACPI tables too were installed in a
ReadyToBoot callback, your suggestion above would not work. And our ACPI
tables are not installed in a ReadyToBoot callback partly because I
ultimately introduced a separate event group for "PCI bridges have been
connected", and we signal it now explicitly from BDS (device connections
are BDS jurisdiction, and QEMU's ACPI generation depends on PCI state).

I just want to point out that we have a kind of "capital" here. By
carefully coding stuff we build capital, and by hooking stuff into
ReadyToBoot callbacks we spend (hopefully not "squander") that capital.

Originally, the APM Mustang firmware (open source, so I can talk about
it) would first install ACPI tables with constant, platform-tailored
contents (built from *.asl / *.aslc files), but reusing the stock
AcpiPlatformDxe C code without customization. Then it would install a
ReadyToBoot callback which looked up the right DSDT or SSDT, by walking
the table tree manually, and then it would poke data into the installed
table (DSDT or SSDT) in-place, using the ACPI SDT protocol.

Of course it was completely bogus and unreliable, and I changed the
constant table to contain external references, and I provided those
external references in a minimal, hand- and runtime- built SSDT, right
in AcpiPlatformDxe.

I'm not trying to carefully compose a strawman argument here, just
presenting why I'm nervous about ReadyToBoot callbacks that try to rely
on ordering between system-wide resources.


Also, please don't forget about the other (current) consumer of the
feature PCD, ArmVirtPL031FdtClientLib, which is plugged into
RealTimeClockRuntimeDxe. How do you suggest to rewrite the PCD test in
that driver under the ReadyToBoot callback scenario?

RealTimeClockRuntimeDxe's dispatch order is unspecified relative to the
installation of ACPI tables, so you couldn't look at the latter's
presence to see if the DTB needs an update. (In fact, because
AcpiPlatformDxe's main actions are cued in from BDS in practice, the
tables would be guaranteed not to exist when RealTimeClockRuntimeDxe looks.)

So ArmVirtPL031FdtClientLib would have to install another ReadyToBoot
callback for modifying the DTB. And this callback could be invoked
before or after the callback to FdtClientDxe. (I guess it would be okay,
but not very intuitive.)

> 
>> This series indeed turned out a bit more complex than I had expected,
>> but it was the one I could post with a good conscience. Can you perhaps
>> identify the part(s) in more detail that seem overly complex to you?
>>
> 
> Building the same library in two different ways, having to call a
> library constructor explicitly in some cases and muck about with TPL
> levels to prevent a protocol notify from triggering are all things I
> would really like to avoid tbh

Alright; can you please post the alternative patch set? (With the
ReadyToBoot callback(s), that is, not the BDS hack.)

Thanks!
Laszlo