On Thu, Dec 05, 2024 at 03:31:05PM -0800, James Houghton wrote:
> > > @@ -2062,6 +2069,20 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
> > > enum kvm_mr_change change)
> > > {
> > > bool log_dirty_pages = new && new->flags & KVM_MEM_LOG_DIRTY_PAGES;
> > > + u32 changed_flags = (new ? new->flags : 0) ^ (old ? old->flags : 0);
> > > +
> > > + /*
> > > + * If KVM_MEM_USERFAULT changed, drop all the stage-2 mappings so that
> > > + * we can (1) respect userfault-ness or (2) create block mappings.
> > > + */
> > > + if ((changed_flags & KVM_MEM_USERFAULT) && change == KVM_MR_FLAGS_ONLY)
> > > + kvm_arch_flush_shadow_memslot(kvm, old);
> >
> > I'd strongly prefer that we make (2) a userspace problem and don't
> > eagerly invalidate stage-2 mappings on the USERFAULT -> !USERFAULT
> > change.
> >
> > Having implied user-visible behaviors on ioctls is never good, and for
> > systems without FEAT_S2FWB you might be better off avoiding the unmap in
> > the first place.
> >
> > So, if userspace decides there's a benefit to invalidating the stage-2
> > MMU, it can just delete + recreate the memslot.
>
> Ok I think that's reasonable. So for USERFAULT -> !USERFAULT, I'll
> just follow the precedent set by dirty logging. For x86 today, we
> collapse the mappings, and for arm64 we do not.
>
> Is arm64 ever going to support collapsing back to huge mappings after
> dirty logging is disabled?
Patches on list is always a good place to start :)
What I'd expect on FEAT_S2FWB hardware is that invalidating the whole
stage-2 and faulting back in block entries would give the best
experience.
Only in the case of !FWB would a literal table -> block collapse be
beneficial, as the MMU could potentially elide CMOs when remapping. But
that assumes you're starting with a fully-mapped table and there are no
holes that are "out of sync" with the guest.
--
Thanks,
Oliver
