Re: [PATCH] Ability to read the MKTME status from userspace (patch v2)

[Andy Lutomirski]

On Thu, Jun 25, 2020 at 2:38 PM Dave Hansen <dave.hansen@xxxxxxxxx> wrote:
>
> On 6/25/20 2:27 PM, Borislav Petkov wrote:
> > On Thu, Jun 25, 2020 at 06:16:12PM -0300, Daniel Gutson wrote:
> >> What didn't become clear from the thread last time is the direction to
> >> proceed. Concrete suggestion?
> > Here are two:
> >
> > https://lkml.kernel.org/r/20200619161752.GG32683@xxxxxxx
> > https://lkml.kernel.org/r/20200619161026.GF32683@xxxxxxx
> >
> > but before that happens, I'd like to hear Dave confirm that when we
> > expose all that information to userspace, it will actually be true and
> > show the necessary bits which *actually* tell you that encryption is
> > enabled.
> >
> > If you're still unclear, go over the thread again pls.
>
> It boils down to this: we shouldn't expose low-level, vendor-specific
> implementation details if we can avoid it.  Let's expose something that
> app can actually use.
>
> Something that will work for all of the TME, MKTME and SEV platforms
> that I know of and continue to work for a while would be to have a
> per-numa-node (/sys/devices/system/node[X]/file) that says: "user data
> on this node is protected by memory encryption".
>
> SEV guests would always have a 1 in all nodes.
>
> TME systems with no platform screwiness like PMEM would always have a 1.
>
> Old systems would have a 0 in there.
>
> TME systems which also have PMEM-only nodes would set 0 in PMEM nodes
> and 1 on DRAM nodes.
>
> Systems with screwy EFI_MEMORY_CPU_CRYPTO mixing within NUMA nodes would
> turn it off for the screwy nodes.
>
> Is that concrete enough?

What about MKTME platforms that (using hypothetical future kernel
support) have encryption enabled for a node but have disabled it for
specific pages using madvise()?  Or that have any other nontrivial
policy like that?