Re: [Qemu-devel] [PATCH 00/10] xen: pv domain support.

[Avi Kivity <avi@redhat.com>]

Alexander Graf wrote:
>
> On 05.04.2009, at 10:48, Avi Kivity wrote:
>
>> Alexander Graf wrote:
>>>
>>> One idea I had for full virtualization in a Xen environment would be 
>>> an PV vmenter/vmexit framework - either by implementing a completely 
>>> new abstraction or simple traps for privileged operations like VMRUN.
>>>
>>> That way we could have a kvm that talks to xen for the VM, rendering 
>>> kvm useful on Xen dom0s, giving people the best of both worlds.
>>>
>>> That was only one of the ideas that came up while talking to people 
>>> why running kvm on xen isn't as easy as just recompiling :-). Would 
>>> you think of such a thing as useful?
>>>
>>>
>>
>> Why would anyone want to do that?  If you've got Xen running, just 
>> start up a Xen guest.
>
> I'm not saying it's a great idea - that's why I didn't even consider 
> to develop it yet :-).
>
> Basically it would solve two problems:
>
> 1) Migration path. If you could already use KVM on a Xen host, you 
> could have Xen PV guests and KVM guests in parallel, easing migration 
> to KVM for customers.

I like this, of course, but we have a path through Xenner.  Maybe this 
(kvm-on-xen) path will be easier to take.

>
> 2) Alternative to HVM. That's how this came up from Gerd's mail. We do 
> have KVM support in upstream qemu, but we don't have Xen HVM support. 
> That way you could use the same binary for all your needs. Admittedly, 
> it might make more sense to just implement HVM support :-).

I was under the impression that this is underway.

>
> Again, I just like talking to others about random ideas I have and 
> this was one. I don't think it's worth it - IMHO it'd be more useful 
> to create an in-kernel xen-like module that exposes Xen PV 
> functionality, so you get all the PV benefits without the performance 
> hit from full virtualization and duplication of code.
>

With npt/ept pv performance might be higher running under kvm+xenner 
than with software-only Xen by letting the guest kernel access 
pagetables directly.  Though Gerd had some issues with 64-bit guests 
IIRC, which is a pity since it's there that the pv performance hit is 
greatest.


-- 
error compiling committee.c: too many arguments to function