Re: [PATCH v2] tests/functional/aarch64: add tests for FEAT_RME

[Alex Bennée]

Gustavo Romero <gustavo.romero@linaro.org> writes:

> Hi Pierrick,
>
> On 12/2/24 14:33, Pierrick Bouvier wrote:
>> On 12/2/24 03:04, Peter Maydell wrote:
>>> On Mon, 2 Dec 2024 at 10:58, Marcin Juszkiewicz
>>> <marcin.juszkiewicz@linaro.org> wrote:
>>>>
>>>> W dniu 1.12.2024 o 19:09, Pierrick Bouvier pisze:
>>>>> Hi Marcin,
>>>>>
>>>>> On 12/1/24 05:34, Marcin Juszkiewicz wrote:
>>>>
>>>>>> So RME tests probably need longer timeouts or would not run at all.
>>>>>>
>>>>>
>>>>> By any chance, are you running those tests in debug mode?
>>>>
>>>> zOMG, thanks! Turned out that I left "--enable-debug" in my local build
>>>> and forgot to remove it.
>>>>
>>>> 90s vs 2974s is a difference ;D
>>>
>>> That is a particularly awful debug-to-non-debug performance
>>> difference; it might be worth looking into what exactly
>>> is causing it.
>>>
>> Indeed.
>> This test boots a nested VM (using kvm, not tcg), do you any idea
>> why such use case could be so slow with QEMU?
>
> It's not "really" using kvm acceleration right?
>
> It's using kvm but the kvm module in the host kernel is actually running
> fully emulated in QEMU system mode (-accel tcg is being passed to it).
>
> So I understand it's "all TCG" in the end and no HW acceleration is
> ever used in the tests.

Right, but from the internal QEMU's point of view its just real hardware
so we don't have nested translation. All the overhead of that is in the
"host" QEMU.

In theory we could also use kvmtool to launch a realm but I doubt in
this case the internal QEMU is actually taking up that much of the
runtime.

>
>
> Cheers,
> Gustavo

-- 
Alex Bennée
Virtualisation Tech Lead @ Linaro