Re: [PATCH] selftests: vDSO: Do not rely on $ARCH for vdso_test_getrandom && vdso_test_chacha

[Christophe Leroy <christophe.leroy@csgroup.eu>]

Le 02/09/2024 à 16:18, Christophe Leroy a écrit :
> 
> 
> Le 02/09/2024 à 15:57, Jason A. Donenfeld a écrit :
>> On Mon, Sep 02, 2024 at 03:23:47PM +0200, Christophe Leroy wrote:
>>>
>>>
>>> Le 02/09/2024 à 14:37, Mark Brown a écrit :
>>>> On Mon, Sep 02, 2024 at 02:22:38PM +0200, Christophe Leroy wrote:
>>>>
>>>>> When vdso_test_getcpu doesn't find the vDSO entry point, it prints 
>>>>> an error
>>>>> text and returns KSFT_SKIP
>>>>
>>>>> I thought it would be more correct to have the same behaviour on
>>>>> vdso_test_getrandom instead of trying to build it only when the 
>>>>> underlying
>>>>> kernel supports it.
>>>>
>>>> The problem is that the test incorporates assembler code so it simply
>>>> won't build for architectures without explicit porting, the issue isn't
>>>> if the target kernel supports it but rather that the test won't compile
>>>> in the first place.
>>>
>>> Yes indeed and that was the purpose of my patch, have a macro in
>>> vdso_config.h to tell where the assembler code is:
>>>
>>> diff --git a/tools/testing/selftests/vDSO/vdso_config.h
>>> b/tools/testing/selftests/vDSO/vdso_config.h
>>> index 740ce8c98d2e..693920471160 100644
>>> --- a/tools/testing/selftests/vDSO/vdso_config.h
>>> +++ b/tools/testing/selftests/vDSO/vdso_config.h
>>> @@ -47,6 +47,7 @@
>>>    #elif defined(__x86_64__)
>>>    #define VDSO_VERSION        0
>>>    #define VDSO_NAMES        1
>>> +#define VDSO_GETRANDOM
>>> "../../../../arch/x86/entry/vdso/vgetrandom-chacha.S"
>>>    #elif defined(__riscv__) || defined(__riscv)
>>>    #define VDSO_VERSION        5
>>>    #define VDSO_NAMES        1
>>>
>>>
>>> And then:
>>>
>>> diff --git a/tools/testing/selftests/vDSO/vdso_test_chacha-asm.S
>>> b/tools/testing/selftests/vDSO/vdso_test_chacha-asm.S
>>> new file mode 100644
>>> index 000000000000..8e704165f6f2
>>> --- /dev/null
>>> +++ b/tools/testing/selftests/vDSO/vdso_test_chacha-asm.S
>>> @@ -0,0 +1,7 @@
>>> +#include "vdso_config.h"
>>> +
>>> +#ifdef VDSO_GETRANDOM
>>> +
>>> +#include VDSO_GETRANDOM
>>> +
>>> +#endif
>>>
>>> I thought it was a lot easier to handle if through necessary #ifdefs in
>>> vdso_config.h that implementing an additional logic in Makefiles.
>>
>> Yet it still tripped up the test robot, right?
> 
> Yes I need to look at that.
> 
>>
>> In general I'm not crazy about this approach.
> 
> I have the feeling I get things done easier with that approach. But if 
> you feel better playing up with the makefile, I incline.

Also I thing that one day or another someone will want to implement it a 
more performant way on power10 which is one of the latest powerpc CPU, 
something similar to arch/powerpc/crypto/chacha-p10le-8x.S

When that happens, we will need a way to tell vdso_test_chacha to build 
another vgetrandom-chacha.S and I feel that doing it in the Makefile 
will become really tricky.