Re: [CFT PATCH v2 2/2] KVM: x86: support XSAVES usage in the host

[Paolo Bonzini <pbonzini@redhat.com>]

On 26/11/2014 15:42, Radim Krčmář wrote:
> 2014-11-26 14:57+0100, Paolo Bonzini:
>>
>>
>> On 26/11/2014 14:53, Radim Krčmář wrote:
>>>>>>> get_xsave = native_xrstor(guest_xsave);  xsave(aligned_userspace_buffer)
>>>>>>> set_xsave = xrstor(aligned_userspace_buffer);  native_xsave(guest_xsave)
>>>>>>>
>>>>>>> Could that work?
>>>>>
>>>>> It could, though it is more like
>>>>>
>>>>>    get_fpu()
>>>>>    native_xrstor(guest_xsave)
>>>>>    xsave(buffer)
>>>>>    put_fpu()
>>>>>
>>>>> and vice versa.  Also, the userspace buffer is mos likely not aligned,
>>>>> so you need some kind of bounce buffer.  It can be done if the CPUID
>>>>> turns out to be a bottleneck, apart from that it'd most likely be slower.
>>> Yeah, it was mostly making this code more future-proof ... it is easier
>>> to convince xsave.h to export its structures if CPUID is the problem.
>>> (I still see some hope for Linux, so performance isn't my primary goal.)
>>>
>>> I'm quite interested in CPUID now though, so I'll try to benchmark it,
>>> someday.
> 
> (Sorry, I don't fully understand your thoughts and I just talk more of
>  the same in those scenarios.)
> 
>> I'm not sure what is more future proof. :)  I wonder if native_xrstor
>> could be a problem the day XRSTORS actually sets/restores MSRs as the
>> processor documentation promises.
> 
> XRSTORS won't affect the guest in any way, we are just going to use it
> to convert the xsave, so any side-effects are going to stay in the host.
> (This could break the host though.)

Yes, that's the problem. :)

> My main presumption is that XSAVE*->XRSTOR*->XSAVE->XRSTOR has the same
> result as XSAVE->XRSTOR, because we are only interested in the state,
> not in any metadata.
> (If it isn't possible to combine intructions, like XSAVE after XRSTORS,
>  this solution won't work.)

Yes, that should be right.  But actually what KVM would do it is
XRSTOR->XSAVE*->XRSTOR*->XSAVE.  The problem here is the side effects of
doing XRSTORS far from a guest entry... though that would likely be
handled by load_guest_fpu/put_guest_fpu.