Re: [PATCH 03/16] KVM: arm64: Turn kvm_pgtable_stage2_set_owner into kvm_pgtable_stage2_annotate

From: Marc Zyngier
Date: Tue Jul 20 2021 - 09:16:56 EST


On Tue, 20 Jul 2021 12:36:21 +0100,
Quentin Perret <qperret@xxxxxxxxxx> wrote:
>
> On Tuesday 20 Jul 2021 at 12:20:58 (+0100), Marc Zyngier wrote:
> > On Tue, 20 Jul 2021 11:38:17 +0100,
> > Quentin Perret <qperret@xxxxxxxxxx> wrote:
> > >
> > > On Tuesday 20 Jul 2021 at 11:21:17 (+0100), Marc Zyngier wrote:
> > > > On Tue, 20 Jul 2021 11:09:21 +0100,
> > > > Quentin Perret <qperret@xxxxxxxxxx> wrote:
> > > > >
> > > > > On Thursday 15 Jul 2021 at 17:31:46 (+0100), Marc Zyngier wrote:
> > > > > > @@ -815,7 +807,7 @@ int kvm_pgtable_stage2_set_owner(struct kvm_pgtable *pgt, u64 addr, u64 size,
> > > > > > .arg = &map_data,
> > > > > > };
> > > > > >
> > > > > > - if (owner_id > KVM_MAX_OWNER_ID)
> > > > > > + if (!annotation || (annotation & PTE_VALID))
> > > > > > return -EINVAL;
> > > > >
> > > > > Why do you consider annotation==0 invalid? The assumption so far has
> > > > > been that the owner_id for the host is 0, so annotating a range with 0s
> > > > > should be a valid operation -- this will be required when e.g.
> > > > > transferring ownership of a page back to the host.
> > > >
> > > > How do you then distinguish it from an empty entry that doesn't map to
> > > > anything at all?
> > >
> > > You don't, but that's beauty of it :)
> > >
> > > The host starts with a PGD full of zeroes, which in terms of ownership
> > > means that it owns the entire (I)PA space. And it loses ownership of a
> > > page only when we explicitly annotate it with an owner id != 0.
> >
> > Right. But this scheme doesn't apply to the guests, does it?
>
> Right, the meaning of a NULL PTE in guests will clearly be something
> different, but I guess the interpretation of what invalid mappings mean
> is up to the caller.
>
> > Don't we
> > need something that is non-null to preserve the table refcounting?
>
> Sure, but do we care? If the table entry gets zeroed we're then
> basically using an 'invalid block' mapping to annotate the entire block
> range with '0', whatever that means. For guests it won't mean much, but
> for the host that would mean sole ownership of the entire range.

I see. You let the refcount drop to 0, unmap the table and let
transfer the 0 annotation one level up, covering the whole block.

I guess I'll revert back to allowing 0, but I'd like to make sure we
don't do that for guests unless we actually tear down the address
space (checking for KVM_PGTABLE_S2_IDMAP should work).

Thanks,

M.

--
Without deviation from the norm, progress is not possible.