On 09/28/2010 11:19 AM, H. Peter Anvin wrote:
> On 09/28/2010 11:13 AM, Jeremy Fitzhardinge wrote:
>> On 09/28/2010 10:56 AM, H. Peter Anvin wrote:
>>> On 09/28/2010 10:13 AM, Jeremy Fitzhardinge wrote:
>>>> Yes, we could just mask out the MTRR CPU feature and rely entirely on PAT.
>>>> The alternative would be to use the wrmsr hooks to emulate the Intel
>>>> MTRR registers by mapping them to hypercalls, but that seems needlessly
>>> Indeed. Relying on pure PAT is the Right Thing[TM].
>> Is there a plan to formally deprecate /proc/mtrr and the kernel
>> infrastructure behind it?
> No, and we really can't do it for a couple of reasons:
> a) Pre-PAT hardware;
> b) MTRRs and PAT interact on hardware;
> c) MTRRs, but not PAT, interact with SMM.
What about pre-PAT software (ie, X servers which still use /proc/mtrr)?
> However, since a virtual machine like Xen doesn't have these issues, it
> doesn't apply
Well, we're specifically talking about a virtual machine which has
direct access to hardware, so it is concerned about the real physical
memory properties of real physical pages. If we can assume that
BIOS/Xen will always set up MTRR correctly then there shouldn't be any
need for the kernel to modify the MTRR itself. How true is that in
general? I don't know, but if we could rely on BIOS then there'd never
be a need to touch MTRR, would there?
Xen-devel mailing list