On 01.08.2010, at 16:02, Avi Kivity wrote:
> On 07/29/2010 03:47 PM, Alexander Graf wrote:
>> On PPC we run PR=0 (kernel mode) code in PR=1 (user mode) and don't use the
>> hypervisor extensions.
>>
>> While that is all great to show that virtualization is possible, there are
>> quite some cases where the emulation overhead of privileged instructions is
>> killing performance.
>>
>> This patchset tackles exactly that issue. It introduces a paravirtual framework
>> using which KVM and Linux share a page to exchange register state with. That
>> way we don't have to switch to the hypervisor just to change a value of a
>> privileged register.
>>
>> To prove my point, I ran the same test I did for the MMU optimizations against
>> the PV framework. Here are the results:
>>
>> [without]
>>
>> debian-powerpc:~# time for i in {1..1000}; do /bin/echo hello> /dev/null; done
>>
>> real 0m14.659s
>> user 0m8.967s
>> sys 0m5.688s
>>
>> [with]
>>
>> debian-powerpc:~# time for i in {1..1000}; do /bin/echo hello> /dev/null; done
>>
>> real 0m7.557s
>> user 0m4.121s
>> sys 0m3.426s
>>
>>
>> So this is a significant performance improvement! I'm quite happy how fast this
>> whole thing becomes :)
>>
>> I tried to take all comments I've heard from people so far about such a PV
>> framework into account. In case you told me something before that is a no-go
>> and I still did it, please just tell me again.
>>
>> To make use of this whole thing you also need patches to qemu and openbios. I
>> have them in my queue, but want to see this set upstream first before I start
>> sending patches to the other projects.
>>
>> Now go and have fun with fast VMs on PPC! Get yourself a G5 on ebay and start
>> experiencing the power yourself. - heh
>>
>> v1 -> v2:
>>
>> - change hypervisor calls to use r0 and r3
>> - make crit detection only trigger in supervisor mode
>> - RMO -> PAM
>> - introduce kvm_patch_ins
>> - only flush icache when patching
>> - introduce kvm_patch_ins_b
>> - update documentation
>>
>> v2 -> v3:
>>
>> - use pPAPR conventions for hypercall interface
>> - only use r0 as magic sc number
>> - remove PVR detection
>> - remove BookE shared page mapping support
>> - combine book3s-64 and -32 magic page ra override
>> - add self-test check if the mapping works to guest code
>> - add safety check for relocatable kernels
>>
>
> Looks reasonable. Since it's fair to say I understand nothing about powerpc, I'd like someone who does to review it and ack, please, with an emphasis on the interfaces.
Sounds good. Preferably someone with access to the ePAPR spec :).
Alex
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
