Re: [PATCH, RFC 45/62] mm: Add the encrypt_mprotect() system call for MKTME

[Andy Lutomirski]

On Mon, Jun 17, 2019 at 7:11 PM Kai Huang <kai.huang@xxxxxxxxxxxxxxx> wrote:
>
> On Mon, 2019-06-17 at 18:50 -0700, Andy Lutomirski wrote:
> > On Mon, Jun 17, 2019 at 5:48 PM Kai Huang <kai.huang@xxxxxxxxxxxxxxx> wrote:
> > >
> > >
> > > >
> > > > > And another silly argument: if we had /dev/mktme, then we could
> > > > > possibly get away with avoiding all the keyring stuff entirely.
> > > > > Instead, you open /dev/mktme and you get your own key under the hook.
> > > > > If you want two keys, you open /dev/mktme twice.  If you want some
> > > > > other program to be able to see your memory, you pass it the fd.
> > > >
> > > > We still like the keyring because it's one-stop-shopping as the place
> > > > that *owns* the hardware KeyID slots.  Those are global resources and
> > > > scream for a single global place to allocate and manage them.  The
> > > > hardware slots also need to be shared between any anonymous and
> > > > file-based users, no matter what the APIs for the anonymous side.
> > >
> > > MKTME driver (who creates /dev/mktme) can also be the one-stop-shopping. I think whether to
> > > choose
> > > keyring to manage MKTME key should be based on whether we need/should take advantage of existing
> > > key
> > > retention service functionalities. For example, with key retention service we can
> > > revoke/invalidate/set expiry for a key (not sure whether MKTME needs those although), and we
> > > have
> > > several keyrings -- thread specific keyring, process specific keyring, user specific keyring,
> > > etc,
> > > thus we can control who can/cannot find the key, etc. I think managing MKTME key in MKTME driver
> > > doesn't have those advantages.
> > >
> >
> > Trying to evaluate this with the current proposed code is a bit odd, I
> > think.  Suppose you create a thread-specific key and then fork().  The
> > child can presumably still use the key regardless of whether the child
> > can nominally access the key in the keyring because the PTEs are still
> > there.
>
> Right. This is a little bit odd, although virtualization (Qemu, which is the main use case of MKTME
> at least so far) doesn't use fork().
>
> >
> > More fundamentally, in some sense, the current code has no semantics.
> > Associating a key with memory and "encrypting" it doesn't actually do
> > anything unless you are attacking the memory bus but you haven't
> > compromised the kernel.  There's no protection against a guest that
> > can corrupt its EPT tables, there's no protection against kernel bugs
> > (*especially* if the duplicate direct map design stays), and there
> > isn't even any fd or other object around by which you can only access
> > the data if you can see the key.
>
> I am not saying managing MKTME key/keyID in key retention service is definitely better, but it seems
> all those you mentioned are not related to whether to choose key retention service to manage MKTME
> key/keyID? Or you are saying it doesn't matter we manage key/keyID in key retention service or in
> MKTME driver, since MKTME barely have any security benefits (besides physical attack)?

Mostly the latter.  I think it's very hard to evaluate whether a given
key allocation model makes sense given that MKTME provides such weak
security benefits.  TME has obvious security benefits, as does
encryption of persistent memory, but this giant patch set isn't needed
for plain TME and it doesn't help with persistent memory.


>
> >
> > I'm also wondering whether the kernel will always be able to be a
> > one-stop shop for key allocation -- if the MKTME hardware gains
> > interesting new uses down the road, who knows how key allocation will
> > work?
>
> I by now don't have any use case which requires to manage key/keyID specifically for its own use,
> rather than letting kernel to manage keyID allocation. Please inspire us if you have any potential.
>

Other than compliance, I can't think of much reason that using
multiple keys is useful, regardless of how their allocated.  The only
thing I've thought of is that, with multiple keys, you can use PCONFIG
to remove one and flush caches and the data is most definitely gone.
On the other hand, you can just zero the memory and the data is just
as gone even without any encryption.