Re: [Xen-devel] Re: [GIT PULL] xen /proc/mtrr implementation

[Ingo Molnar <mingo@elte.hu>]

* Gerd Hoffmann <kraxel@redhat.com> wrote:

> On 05/19/09 14:26, Ingo Molnar wrote:
>> * Gerd Hoffmann<kraxel@redhat.com>  wrote:
>>
>>> On 05/19/09 13:08, Ingo Molnar wrote:
>>>> Or, alternatively, the hypervisor can expose its own administrative
>>>> interface to manage MTRRs.
>>> Guess what?  Xen does exactly that.  And the xen mtrr_ops
>>> implementation uses that interface ...
>>
>> No, that is not an 'administrative interface' - that is a guest
>> kernel level hack that complicates Linux, extends its effective ABI
>> dependencies and which has to be maintained there from that point
>> on.
>>
>> There's really just three proper technical solutions here:
>>
>> - either catch the lowlevel CPU hw ops (the MSR modifications, which
>>    isnt really all that different from the mtrr_ops approach so it
>>    shouldnt pose undue difficulties to the Xen hypervisor).
>
> Devil is in the details.
>
> The dom0 kernel might not see all physical cpus on the system.  So 
> Xen can't leave the job of looping over all cpus to the dom0 
> kernel, Xen has to apply the changes made by the (priviledged) 
> guest kernel on any (virtual) cpu to all (physical) cpus in the 
> machine.

Applying MTRR changes to only part of the CPUs is utter madness.

> Which in turn means the "lowlevel cpu hw op" would work in a 
> slightly different way on Xen and native.  Nasty.
>
>>    That will
>>    be maximally transparent and compatible, with zero changes needed
>>    to the Linux kernel.
>
> No, the linux kernel probably should do the wrmsr on one cpu only then.

Why?

>> - or introduce its own hypercall API based administration API,
>>    without bothering the guest kernel with crap. Trivially patch Xorg
>>    to make use of it and that's it.
>
> I have serious doubts that this is going to fly with KMS.
>
> Oops, the third "proper technical solutions" is missing.

Yeah, the third one is to not touch MTRRs after bootup and use PAT.

	Ingo