Re: [kvm-unit-tests PATCH] arm/arm64: introduce is_aarch32

[Wei Huang <wei@redhat.com>]

On 11/16/2016 04:02 PM, Christopher Covington wrote:
> On 11/16/2016 12:46 PM, Marc Zyngier wrote:
>> On 16/11/16 14:38, Andrew Jones wrote:
>>> ARMv7-A isn't exactly the same as ARMv8-A32 (AArch32). This
>>> function allows unit tests to make the distinction.
>>
>> Hi Drew,
>>
>> Overall, having to find out about the architecture is a bad idea most of
>> the time. We have feature registers for most things, and it definitely
>> makes more sense to check for those than trying to cast a wider net.
>>
>>>
>>> Signed-off-by: Andrew Jones <drjones@redhat.com>
>>>
>>> ---
>>> I'm actually unsure if there's a feature bit or not that I could
>>> probe instead. It'd be nice if somebody can confirm. Thanks, drew
> 
> I'd be happy to settle with the hard-coded CPU list.
> 
> But if you're curious about alternatives, I've taken a look through some
> documentation. ID_ISAR0.coproc describes whether mrrc is available but
> I think it is generally available on v7 and above. I think ID_ISAR5 will
> be zero on v7 and nonzero on v8-A32. But PMCR.LC seems like the best bit
> to check.
> 
>>> diff --git a/lib/arm64/asm/processor.h b/lib/arm64/asm/processor.h
>>> index 84d5c7ce752b..b602e1fbbc2d 100644
>>> --- a/lib/arm64/asm/processor.h
>>> +++ b/lib/arm64/asm/processor.h
>>> @@ -81,5 +81,10 @@ DEFINE_GET_SYSREG32(mpidr)
>>>  extern void start_usr(void (*func)(void *arg), void *arg, unsigned long sp_usr);
>>>  extern bool is_user(void);
>>>  
>>> +static inline bool is_aarch32(void)
>>> +{
>>> +	return false;
>>> +}
>>> +
>>>  #endif /* !__ASSEMBLY__ */
>>>  #endif /* _ASMARM64_PROCESSOR_H_ */
>>>
>>
>> So the real question is: what are you trying to check for?
> 
> The question is "how many bits wide is pmccntr?" I think we
> can test whether writing PMCR.LC = 1 sticks. Based on the
> documentation, it seems to me like it wouldn't for v7 and
> would for v8-A32.
> 
> uint8_t size_pmccntr(void) {
>   uint32_t pmcr = get_pmcr();
>   if (pmcr & PMU_PMCR_LC_MASK)
>     return 64;
>   set_pmcr(pmcr | (1 << PMU_PMCR_LC_SHIFT));
>   if (get_pmcr() & PMU_PMCR_LC_MASK)
>     return 64;
>   return 32;
> }

This might actually be the solution if we can't find a more reliable
detection approach. I briefly tested it and it seemed to work.

> 
> Thanks,
> Cov
>