Re: [RFC PATCH v5 7/9] target/avocado: Pass parameters to migration test on aarch64

[Fabiano Rosas]

Philippe Mathieu-Daudé <philmd@linaro.org> writes:

> On 20/1/23 19:48, Fabiano Rosas wrote:
>> The migration tests are currently broken for an aarch64 host because
>> the tests pass no 'machine' and 'cpu' options on the QEMU command
>> line. Most other architectures define a default value in QEMU for
>> these options, but arm does not.
>
> There was some discussions around that in the past:
> https://lore.kernel.org/qemu-devel/20190621153806.13489-1-wainersm@redhat.com/
> https://lore.kernel.org/qemu-devel/CAFEAcA9NBu+L4wHfkLTv93wy90wjnV05EZ12PT6PmLjdZ5h_YA@mail.gmail.com/

There's more than one topic being discussed, specially in this last
thread, but here's my two cents.

About defaults: It's probably best to be explicit in tests. And if we
wanted, have a separate test to make sure the lack of an option still
does what it's expected, either outputting a message or behaving the
same as the explicit version.

About host architecture-specific tests: Unless we're talking about KVM,
I see no point. Having to change hosts to test agnostic features makes
no sense (the migration test is one example).

About generic tests: If a feature is required to behave the same for all
architectures/machines/cpus then sure. But most low level stuff would be
quite dependent on specifics.

>> Add these options to the test class in case the test is being executed
>> in an aarch64 host.
>
> I'm not sure what we are aiming to test here.
>
> Migration in general? If so, any random machine should work.
> By hardcoding the 'virt' machine, at least this test is reproducible.

Yeah, I cannot say for sure there isn't some machine property that gets
transferred during migration. It seemed more conservative to define a
specific one.

> I'd rather fix that generically as "if a test requires a default
> machine and the target doesn't provide any default, then SKIP the
> test". Then adding machine-specific tests. Can be done on top, so

I agree, but the only tests that should *require* a default are the ones
that test the command line parsing or adjacent features. We could always
test "-machine foo" and then separately test that the lack of a machine
option still gives the Foo machine.

The fact that we sometimes use defaults to be able to have the same-ish
command line for every case is more of a limitation of our testing
infrastructure in my opinion.