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

[Ingo Molnar <mingo@elte.hu>]

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

>>>>> No, the linux kernel probably should do the wrmsr on one cpu only then.
>>>> Why?
>
>> | The change of MTRR's on _any_ of the guest CPUs in a dom0 context
>> | should immediately be refected on all CPUs. Assymetric MTRR
>> | settings are madness.
>
> Exactly.  And thats why it is pointless to let the dom0 kernel 
> write the mtrr msrs on more than one cpu.

How does this negate my claim that the Linux kernel needs no 
modifications? (which i think your point is - let me know if you 
have some other point here.)

the Xen hypervisor can simply repeat all requests (i.e. not care at 
all about the fact that a guest does these modifications on all CPUs 
it sees), or realize that the modification has already been done and 
skip it. This is in no way a performance sensitive codepath - mtrrs 
are only modified in init sequences - and setting mtrrs is slow and 
globally serialized to begin with.

>>>>> Oops, the third "proper technical solutions" is missing.
>>>> Yeah, the third one is to not touch MTRRs after bootup and use PAT.
>>> Works only in case the CPU has PAT support.
>>
>> Which specific CPU without PAT support do you worry about?
>
> For example: I have a Notebook here, with MTRR and without PAT 
> according to the boot log.  "Pentium III (Coppermine)" says 
> /proc/cpuinfo.

That's a really old CPU, but even Coppermine has PAT support in the 
CPU. You need to go back to things like P5 200 MHz CPUs to find 
PAT-less CPUs.

On the Linux side it's easy to enable it (and _such_ a patch would 
make sense indeed) - it's just that nobody has yet gone through the 
effort of testing and validating the PAT code on that CPU. If you 
care enough, you can do it, send a patch and ping the Intel folks 
about it.

Once the issues are framed correctly, Linux is helped for real.

	Ingo