Re: [Qemu-devel] [RFC PATCH v1 10/22] sev: add SEV debug decrypt command

[Michael S. Tsirkin]

On Wed, Sep 14, 2016 at 02:35:41PM -0300, Eduardo Habkost wrote:
> On Wed, Sep 14, 2016 at 06:46:20PM +0300, Michael S. Tsirkin wrote:
> > On Wed, Sep 14, 2016 at 04:06:33PM +0100, Daniel P. Berrange wrote:
> > > On Wed, Sep 14, 2016 at 05:48:17PM +0300, Michael S. Tsirkin wrote:
> > > > On Wed, Sep 14, 2016 at 03:15:07PM +0100, Daniel P. Berrange wrote:
> > > > > On Wed, Sep 14, 2016 at 04:50:51PM +0300, Michael S. Tsirkin wrote:
> > > > > > On Wed, Sep 14, 2016 at 02:37:49PM +0100, Daniel P. Berrange wrote:
> > > > > > > On Wed, Sep 14, 2016 at 04:32:44PM +0300, Michael S. Tsirkin 
> > > > > > > wrote:
> > > > > > > > On Wed, Sep 14, 2016 at 02:23:14PM +0100, Daniel P. Berrange 
> > > > > > > > wrote:
> > > > > > > > > On Wed, Sep 14, 2016 at 03:07:58PM +0200, Paolo Bonzini wrote:
> > > > > > > > > > 
> > > > > > > > > > 
> > > > > > > > > > On 14/09/2016 15:05, Michael S. Tsirkin wrote:
> > > > > > > > > > > I assumed that with debug on, memory is still encrypted 
> > > > > > > > > > > but the
> > > > > > > > > > > hypervisor can break encryption, and as the cover letter 
> > > > > > > > > > > states, the
> > > > > > > > > > > hypervisor is assumed benign. If true I don't see a need 
> > > > > > > > > > > to
> > > > > > > > > > > give users more rope.
> > > > > > > > > > 
> > > > > > > > > > The hypervisor is assumed benign but vulnerable.
> > > > > > > > > > 
> > > > > > > > > > So, if somebody breaks the hypervisor, you would like to 
> > > > > > > > > > make it as hard
> > > > > > > > > > as possible for the attacker to do evil stuff to the 
> > > > > > > > > > guests.  If the
> > > > > > > > > > attacker can just ask the secure processor "decrypt some 
> > > > > > > > > > memory for me",
> > > > > > > > > > then the encryption is effectively broken.
> > > > > > > > > 
> > > > > > > > > So there's going to be a tradeoff here between use of SEV and 
> > > > > > > > > use of
> > > > > > > > > certain other features. eg, it seems that if you're using 
> > > > > > > > > SEV, then
> > > > > > > > > any concept of creating & analysing guest core dumps from the 
> > > > > > > > > host
> > > > > > > > > is out.
> > > > > > > > 
> > > > > > > > I don't see why - as long as we don't trigger dumps, there's no 
> > > > > > > > leak :)
> > > > > > > 
> > > > > > > If the facility to trigger dumps is available, then the memory
> > > > > > > encryption feature of SEV is as useful as a chocolate teapot,
> > > > > > > as the would be attacker can simply trigger a dump
> > > > > > 
> > > > > > If attacker can trigger things, IOW execute code in hypervisor,
> > > > > > then encrypting memory is not useful anyway.
> > > > > 
> > > > > The presentation at KVM forum claimed it *would* protect against
> > > > > this, and that things like core dump of unencrypted memory would
> > > > > not be permitted, so there's a disconnect between that preso and
> > > > > what you're saying.
> > > > > 
> > > > > Regards,
> > > > > Daniel
> > > > 
> > > > You mean presentation claimed protection against leaks to a malicious
> > > > active attacker within a hypervisor?  I guess the presentation covers
> > > > more than this patchset does then.  And the disconnect would be with
> > > > what the patchset cover letter says, not just with what I say.  Clearly
> > > > encrypting memory is not enough to protect against a malicious
> > > > hypervisor. E.g. just running info cpus is enough to leak information
> > > > from guest.
> > > 
> > > It was explicit about the fact that the host admin would not have any
> > > way to get access to the full contents of guest memory, without the
> > > guest admin granting it. Only those non-encrypted pages used for I/O
> > > transfer between host & guest would be accessible.
> > > 
> > > Regards,
> > > Daniel
> > 
> > If you like, that's the vision. I'd rather discuss the patchset in
> > question though. It encrypts all memory but this does not protect against
> > all attackers, only passive ones. If you disable debugging,
> > it seems to additionally reduce the amount of information that can be
> > leaked to an active attacker in the hypervisor at one go.
> > 
> > Paolo seems to think it's useful, but it's a far cry from a deal
> > breaker, and your email just makes me worry that it has been oversold to
> > the point where everyone will start disabling debugging everywhere in
> > production and claim that otherwise it's a security problem.  IMO a much
> > better in-tree documentation is needed so people know what they are
> > getting in return.
> > 
> > Attestation seems mostly unrelated. The whitepaper says
> >     With this attestation, a guest owner can ensure that the hypervisor did
> >     not interfere with the initialization of SEV before transmitting
> >     confidential information to the guest.
> > which seems to imply an active attacker that is able to interfere
> > with the hypervisor during guest initialization but not afterwards.
> 
> I believe this assumes a compromised hypervisor both before and
> after guest launch, but this assumes the hypervisor:
> 1) Won't be able to change guest memory before attestation
>    without being detected.
> 2) Won't be able to attack the guest after memory is encrypted.

Why would you need to measure things then?  If you assume this, at what
point *can* attacker change memory?

> > So I have no idea why that's useful at the moment - I suspect
> > it's part of the future vision when there are protections
> > against all active attackers in place, but for now it seems to extend the
> > firmware/software interface unnecessarily.
> 
> "Protection against all active attackers" is a very broad
> requirement. Effective protection against a given subset of
> attacks would be reasonable enough to me.

Well selecting a random point in time and saying "I protect against
attacks at this point only" would be a very weak protection, akin to
just adding an assert statement at a random place in code - even though
yes, if you hit that assert you are protected.

This is not to say this is what this patchset does, merely
that it should include a bit more information about the
motivation for the measurement part than
"this is what we can easily implement".


> -- 
> Eduardo