Re: RFC: [0/2] Remove netloop by lazy copying in netback

[Keir Fraser <Keir.Fraser@cl.cam.ac.uk>]

> On Tue, Mar 20, 2007 at 03:18:54AM -0700, Keir Fraser wrote:
>> 
>> That's quite a valid concern, but I think that the required addition to the
>> #PF handler (certainly for i386 and x86/64) will be clean and small, and it
>> will not affect #PF critical-path latencies. I'd be fairly optimistic about
>> it getting accepted upstream, perhaps modulo concerns over whether we'd need
>> to implement it for *every* architecture.
> 
> OK I've given it a spin and it's pretty straight-forward for
> x86 as you said.  However, we'll need a bit more work for
> ia64/ppc either on the kernel side or on the hypervisor side
> as there is no way currently to swap entries in the P2M table.
> 
> Any better suggestions for dealing with those two?

I didn't look super closely at the precise set of steps on x86 either. Do
you take a normal page of memory in the guest's address space and simply
give it an extra mapping in the netback area? Or do you take a page with no
pseudophysical address and assign it a pseudophysical address corresponding
to its place in the netback area?

I certainly assumed (a), and I think that would work fine on ia64 and
powerpc, as the page would already have a pseudophysical address?

Maybe I'm misunderstanding something... :-)

 -- Keir