Re: [patch 4/5] KVM: MMU: reload request from GET_DIRTY_LOG path

[Marcelo Tosatti <mtosatti@redhat.com>]

On Thu, Jun 19, 2014 at 11:17:19AM +0300, Gleb Natapov wrote:
> On Wed, Jun 18, 2014 at 08:12:07PM -0300, mtosatti@redhat.com wrote:
> > Reload remote vcpus MMU from GET_DIRTY_LOG codepath, before
> > deleting a pinned spte.
> > 
> > Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
> > 
> > ---
> >  arch/x86/kvm/mmu.c |    3 +++
> >  1 file changed, 3 insertions(+)
> > 
> > Index: kvm.pinned-sptes/arch/x86/kvm/mmu.c
> > ===================================================================
> > --- kvm.pinned-sptes.orig/arch/x86/kvm/mmu.c	2014-06-13 16:50:50.040140594 -0300
> > +++ kvm.pinned-sptes/arch/x86/kvm/mmu.c	2014-06-13 16:51:05.620104451 -0300
> > @@ -1247,6 +1247,9 @@
> >  		spte &= ~SPTE_MMU_WRITEABLE;
> >  	spte = spte & ~PT_WRITABLE_MASK;
> >  
> > +	if (is_pinned_spte(spte))
> > +		mmu_reload_pinned_vcpus(kvm);
> > +
> Why write protect it at all? mmu_reload_pinned_vcpus() will unprotected it anyway
> on the next vmentry. Isn't it better to just report all pinned pages as dirty alway.

That was the initial plan, however its awkward to stop vcpus, execute
get_dirty_log twice, and have pages marked as dirty on the second
execution.

That is, it is in "incorrect" to report pages as dirty when they are
clean.

Moreover, if the number of pinned pages is larger than the dirty
threshold to stop VM and migrate, you'll never migrate. If vcpus are
in HLT and don't VM-enter immediately, the pages should not be refaulted
right away.

Do you think the optimization is worthwhile ?