Re: [xen-unstable test] 33399: regressions - FAIL

[Andrew Cooper <andrew.cooper3@citrix.com>]

On 15/01/15 14:53, Jan Beulich wrote:
>>>> On 15.01.15 at 15:14, <JBeulich@suse.com> wrote:
>> But I think I made a wrong assumption above regarding the
>> guest size: test-amd64-i386-xl-win7-amd64 produces a 64-bit
>> guest with a 32-bit tool stack, so the crucial part of all the
>> tests failing is not the guest's bitness, but tool stack's. So I'll
>> next look into which of the three feature flags might be off
>> when inspected from a 32-bit Dom0, as I now suspect that the
>> guest simply doesn't get its CPUID bits correctly set up by a
>> 32-bit Dom0 (i.e. the patch might just have uncovered a latent
>> bug).
> And there you go: The hypervisor deliberately clears the
> syscall feature flag for 32-bit PV guests on non-AMD CPUs, and
> hardware appears to do so too when CPUID gets executed from
> a non-64-bit CS (i.e. no matter whether you execute raw or
> "Xen-ified" CPUID there, you won't see that flag set). Yet 64-bit
> guests won't be bothered to check whether the flag is enabled,
> as x86-64 requires the feature to be there.

As AMD had supported syscall in 32bit systems for a long time, I presume
it is only Intel where the feature bit in cpuid changes depending on cs.L

>
> But I think even if this could be taken care of in the tool stack,
> it's better to have hvm_cpuid() mimic that behavior, i.e. force
> FEATURE_SYSCALL on when hvm_guest_x86_mode() == 8.

This seems like the least bad of the available options, but this is
going to be another level of complication for my domain cpuid changes.

~Andrew