Re: [PATCH v3 1/1] KVM: arm64: Allow cacheable stage 2 mapping using VMA flags

[Catalin Marinas]

On Tue, Mar 18, 2025 at 08:17:36PM -0300, Jason Gunthorpe wrote:
> On Tue, Mar 18, 2025 at 07:27:27PM +0000, Catalin Marinas wrote:
> > Thinking some more about a KVM capability that the VMM can check, I'm
> > not sure what it can do with this. The VMM simply maps something from a
> > device and cannot probe the cacheability 
> 
> KVM is mirroring the MM's PTEs to the S2's PTEs.
> 
> You can think about this differently - KVM currently has a little bug
> where the MM's PTE's can say cachable but KVM will mirror it to a S2
> PTE that is forced non-cachable. KVM will not do any cache flushing to
> make this safe.
> 
> Fundamentally that discrepancy is what is being addressed
> here. Cachable PTEs in the MM should be mirrored to cachable PTEs in
> the S2.

If we treat this as a bug (and I agree it's a potential problem if we
assign devices to the guest that have cacheable VMM mappings), the first
step would be to reject such memory slots, report some error to the VMM.
I don't think it's worth attempting to do any cache maintenance to sync
the aliases.

Subsequently, this can be allowed with on FWB-capable hardware. Whether
we want to expose such capability to the VMM, I guess it's harmless but
I doubt the VMM would do anything with it since it doesn't know the
device mapping attributes anyway.

-- 
Catalin