Re: [Qemu-devel] Re: [PATCH] qemu/virtio: move features to an inline function

[Michael S. Tsirkin]

On Thu, Aug 13, 2009 at 12:28:52PM +0300, Avi Kivity wrote:
> On 08/13/2009 09:15 AM, Michael S. Tsirkin wrote:
>>
>>> I wonder though if mmu notifiers can be used to make it transparent...
>>>      
>>
>> Maybe they can, but that decision belongs to KVM.
>> Avi, what do you think?
>>
>>    
>
> I don't see how mmu notifiers help.  You can use mmu notifiers to sync  
> an external mmu to the linux pagetables, but that's not the case here.
>
> I see the following options:
>
> - mprotect() guest memory, trap faults, mprotect() back
>
> Will work, but exceedingly slowly and wastefully

I think this is what Anthony had in mind.

> - add a generic user visible dirty bit tracking based on linux ptes
>
> A lot of work, not sure if useful beyond kvm
>
> - implement vhost dirty bit tracking
>
> Not too difficult; not sure if it's worth the effort though
>
> - reuse the kvm dirty bitmap
>
> Not too difficult but introduces tight coupling between vhost and kvm  
> which might slow down development
>
> - drop to userspace for live migration
>
> Reuse qemu code, lose some performance

This is what I planned.  Note that ability to drop to userspace is
required for non-MSI mode, anyway.

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