Re: [PATCH v5 03/11] tpm: Allow PCR 23 to be restricted to kernel-only use

From: Jarkko Sakkinen
Date: Thu Jan 26 2023 - 16:31:02 EST


On Thu, Jan 26, 2023 at 11:32:22AM -0600, William Roberts wrote:
> On Thu, Jan 26, 2023 at 11:21 AM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
> >
> > On Tue, Jan 24, 2023 at 07:38:04AM -0500, James Bottomley wrote:
> > > On Mon, 2023-01-23 at 11:48 -0600, William Roberts wrote:
> > > > On Fri, Jan 20, 2023 at 9:29 PM Jarkko Sakkinen <jarkko@xxxxxxxxxx>
> > > > wrote:
> > > > >
> > > > > On Sat, Jan 14, 2023 at 09:55:37AM -0500, James Bottomley wrote:
> > > > > > On Tue, 2023-01-03 at 13:10 -0800, Matthew Garrett wrote:
> > > > > > > On Tue, Jan 3, 2023 at 1:05 PM William Roberts
> > > > > > > <bill.c.roberts@xxxxxxxxx> wrote:
> > > > > > >
> > > > > > > > What's the use case of using the creation data and ticket in
> > > > > > > > this context? Who gets the creationData and the ticket?
> > > > > > > > Could a user supplied outsideInfo work? IIRC I saw some
> > > > > > > > patches flying around where the sessions will get encrypted
> > > > > > > > and presumably correctly as well. This would allow the
> > > > > > > > transfer of that outsideInfo, like the NV Index PCR value to
> > > > > > > > be included and integrity protected by the session HMAC.
> > > > > > >
> > > > > > > The goal is to ensure that the key was generated by the kernel.
> > > > > > > In the absence of the creation data, an attacker could generate
> > > > > > > a hibernation image using their own key and trick the kernel
> > > > > > > into resuming arbitrary code. We don't have any way to pass
> > > > > > > secret data from the hibernate kernel to the resume kernel, so
> > > > > > > I don't think there's any easy way to do it with outsideinfo.
> > > > > >
> > > > > > Can we go back again to why you can't use locality? It's exactly
> > > > > > designed for this since locality is part of creation data.
> > > > > > Currently everything only uses locality 0, so it's impossible for
> > > > > > anyone on Linux to produce a key with anything other than 0 in
> > > > > > the creation data for locality. However, the dynamic launch
> > > > > > people are proposing that the Kernel should use Locality 2 for
> > > > > > all its operations, which would allow you to distinguish a key
> > > > > > created by the kernel from one created by a user by locality.
> > > > > >
> > > > > > I think the previous objection was that not all TPMs implement
> > > > > > locality, but then not all laptops have TPMs either, so if you
> > > > > > ever come across one which has a TPM but no locality, it's in a
> > > > > > very similar security boat to one which has no TPM.
> > > > >
> > > > > Kernel could try to use locality 2 and use locality 0 as fallback.
> > > >
> > > > I don't think that would work for Matthew, they need something
> > > > reliable to indicate key provenance.
> > >
> > > No, I think it would be good enough: locality 0 means anyone (including
> > > the kernel on a machine which doesn't function correctly) could have
> > > created this key. Locality 2 would mean only the kernel could have
> > > created this key.
> > >
> > > By the time the kernel boots and before it loads the hibernation image
> > > it will know the answer to the question "does my TPM support locality
> > > 2", so it can use that in its security assessment: if the kernel
> > > supports locality 2 and the key wasn't created in locality 2 then
> > > assume an attack. Obviously, if the kernel doesn't support locality 2
> > > then the hibernation resume has to accept any old key, but that's the
> > > same as the situation today.
> >
> > This sounds otherwise great to me but why bother even allowing a
> > machine with no-locality TPM to be involved with hibernate? Simply
> > detect locality support during driver initialization and disallow
> > sealed hibernation (or whatever the feature was called) if localities
> > were not detected.
> >
> > I get supporting old hardware with old features but it does not make
> > sense to maintain new features with hardware, which clearly does not
> > scale, right?
> >
> > BR, Jarkko
>
> Here's a thought, what if we had a static/cmd line configurable
> no-auth NV Index and writelocked it with the expected key information,
> name or something. I guess the problem is atomicity with write/lock,
> but can't the kernel lock out all other users?
>
> An attacker would need to issue tpm2_startup, which in this case would DOS
> the kernel in both scenarios. If an attacker already wrote and locked the NV
> index, that would also be a DOS. If they already wrote it, the kernel simply
> writes whatever they want. Is there an attack I am missing?
>
> I guess the issue here would be setup, since creating the NV index requires
> hierarchy auth, does the kernel have platform auth or is that already shut down
> by firmware (I can't recall)? A null hierarchy volatile lockable index would be
> nice for this, too bad that doesn't exist.

How do you see this would better when compared to finding a way to use
locality, which could potentially be made to somewhat simple to setup
(practically zero config)?

BR, Jarkko