RE: [Xen-devel] Windows SMP

["Tian, Kevin" <kevin.tian@xxxxxxxxx>]

>From: James Harper [mailto:james.harper@xxxxxxxxxxxxxxxx] 
>Sent: Wednesday, December 31, 2008 4:35 PM
>
>> >Is there a similar approach that would work on an Intel system?
>> 
>> On Intel CPU with FlexPriority support, you don't need patching
>> guest since TPR accesses would be recognized  by hardware
>> for acceleration automatically.
>> 
>> But on CPUs without h/w acceleration support, you may expect
>> borrow that overall idea, but instead of patching with LOCK MOV
>> CR0, you would replace it with a piece of code lines to emulate
>> similar acceleration as what h/w is assumed to do.
>> 
>
>Do you have an example :)
>
>One thing Keir suggested would be to install the patch to jump to some
>code which compared the value being written to the TPR 
>register with the
>value last written, and only perform the actual write if the values are

That's basically what I meant, and also what KVM does today. 
VM-exit in such case is only proactively requested by vmcall
in inserted lines if Xen emulation logic has to be involved.

>different. I can do that without too much fuss but if there is 
>something
>faster then even better.
>

If you compare to VM-exit overhead for every TPR access, above
is already far far faster. Of course fewer memory accesses used
in inserted lines, less overhead you'll see then.:-)

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-devel