Communication in x86

Communication in x86

[James Edouard]

Hi,

The hypercall KVM_HC_PPC_MAP_MAGIC_PAGE is only implemented for PPC,
but I am wondering why there is not a similar hypercall for x86? Is it
not needed (I'm a complete newbie...)?

James

Re: Communication in x86

[Radim Krčmář]

Disclaimer: I have very little knowledge about PPC.

2017-03-16 11:43-0400, James Edouard:
> Hi,
> 
> The hypercall KVM_HC_PPC_MAP_MAGIC_PAGE is only implemented for PPC,
> but I am wondering why there is not a similar hypercall for x86?

PPC traps all privileged instructions while emulating privileged guest
mode and the magic page accelerates instructions where the trap would be
trivial.  By patching instructions to access the page instead.

x86 allows execution of most performance sensitive privileged
instructions without a trap from guest mode.

>                                                                  Is it
> not needed (I'm a complete newbie...)?

Yes, it is optional at best.  I'm not aware of any x86 instruction where
adding a magic page for its emulation would be a clearly good idea.

While there is no magic page in the PPC sense on x86, there are areas of
guest memory where the host writes some data for the guest to read
without a trap, but they are also usually setup using MSR interfaces:
kvm clock, steal time, ...

Re: Communication in x86

[James Edouard]

Thank you the response earlier. Would you know why MSR is more
suitable for implementing these particular features over the hypercall
interface? To me, they seem like alternate methods for guests to send
directives to the host. I'm wondering when one would be preferable
over the other.

On Tue, Mar 28, 2017 at 10:51 AM, Radim Krčmář <rkrcmar@redhat.com> wrote:
> Disclaimer: I have very little knowledge about PPC.
>
> 2017-03-16 11:43-0400, James Edouard:
>> Hi,
>>
>> The hypercall KVM_HC_PPC_MAP_MAGIC_PAGE is only implemented for PPC,
>> but I am wondering why there is not a similar hypercall for x86?
>
> PPC traps all privileged instructions while emulating privileged guest
> mode and the magic page accelerates instructions where the trap would be
> trivial.  By patching instructions to access the page instead.
>
> x86 allows execution of most performance sensitive privileged
> instructions without a trap from guest mode.
>
>>                                                                  Is it
>> not needed (I'm a complete newbie...)?
>
> Yes, it is optional at best.  I'm not aware of any x86 instruction where
> adding a magic page for its emulation would be a clearly good idea.
>
> While there is no magic page in the PPC sense on x86, there are areas of
> guest memory where the host writes some data for the guest to read
> without a trap, but they are also usually setup using MSR interfaces:
> kvm clock, steal time, ...