On Tue, Mar 27, 2007 at 05:53:18PM +1000, Herbert Xu wrote:
> On Tue, Mar 27, 2007 at 08:51:15AM +0100, Keir Fraser wrote:
> > Also bear in mind the plan to avoid the grant mapping at all in the
> > direct-to-wire transmit case. This also will rely on a #PF (or some other
> > fault, potentially) if the guest tries to access that pseudophysical page of
> > memory. If we can work out a way to make this work on ia64 then the
> > delayed-copy implementation problems are solved too.
> That's to avoid the TLB flush costs, right?
Xen/IA64 has optimization there. Xen/ia64 tracks tlb inserts
on grant table mapped page.
If there was no tlb insert(i.e. direct-to-wire transmit case),
it doesn't issue tlb flush, but only clears the p2m entry
when grant table unmap.
On Xen/IA64 case, something like copy-on-demand-access can be implemented
as copy-on-tlb-insert in xen, I suppose.
Xen-devel mailing list