Re: TPM: hibernate with IMA PCR 10

From: Jarkko Sakkinen
Date: Fri Sep 23 2022 - 09:31:07 EST


On Wed, Sep 21, 2022 at 04:15:20PM -0400, Mimi Zohar wrote:
> On Tue, 2022-09-20 at 07:36 +0300, Jarkko Sakkinen wrote:
> > On Sat, Sep 10, 2022 at 10:40:05PM -0400, Mimi Zohar wrote:
> > > On Thu, 2022-09-08 at 08:25 +0300, Jarkko Sakkinen wrote:
> > > > On Wed, Sep 07, 2022 at 07:57:27PM -0400, Mimi Zohar wrote:
> > > > > On Wed, 2022-09-07 at 13:47 -0700, Evan Green wrote:
> > > > > > On Tue, Aug 30, 2022 at 7:48 PM Jarkko Sakkinen <jarkko@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Mon, Aug 29, 2022 at 02:51:50PM -0700, Matthew Garrett wrote:
> > > > > > > > On Mon, Aug 29, 2022 at 2:45 PM Ken Goldman <kgold@xxxxxxxxxxxxx> wrote:
> > > > > > > > >
> > > > > > > > > On 5/4/2022 7:20 PM, Evan Green wrote:
> > > > > > > > > > Enabling the kernel to be able to do encryption and integrity checks on
> > > > > > > > > > the hibernate image prevents a malicious userspace from escalating to
> > > > > > > > > > kernel execution via hibernation resume. [snip]
> > > > > > > > >
> > > > > > > > > I have a related question.
> > > > > > > > >
> > > > > > > > > When a TPM powers up from hibernation, PCR 10 is reset. When a
> > > > > > > > > hibernate image is restored:
> > > > > > > > >
> > > > > > > > > 1. Is there a design for how PCR 10 is restored?
> > > > > > > >
> > > > > > > > I don't see anything that does that at present.
> > > > > > > >
> > > > > > > > > 2. How are /sys/kernel/security/ima/[pseudofiles] saved and
> > > > > > > > > restored?
> > > > > > > >
> > > > > > > > They're part of the running kernel state, so should re-appear without
> > > > > > > > any special casing. However, in the absence of anything repopulating
> > > > > > > > PCR 10, they'll no longer match the in-TPM value.
> > > > > > >
> > > > > > > This feature could still be supported, if IMA is disabled
> > > > > > > in the kernel configuration, which I see a non-issue as
> > > > > > > long as config flag checks are there.
> > > > > >
> > > > > > Right, from what I understand about IMA, the TPM's PCR getting out of
> > > > > > sync with the in-kernel measurement list across a hibernate (because
> > > > > > TPM is reset) or kexec() (because in-memory list gets reset) is
> > > > > > already a problem. This series doesn't really address that, in that it
> > > > > > doesn't really make that situation better or worse.
> > > > >
> > > > > For kexec, the PCRs are not reset, so the IMA measurment list needs to
> > > > > be carried across kexec and restored. This is now being done on most
> > > > > architectures. Afterwards, the IMA measurement list does match the
> > > > > PCRs.
> > > > >
> > > > > Hibernation introduces a different situation, where the the PCRs are
> > > > > reset, but the measurement list is restored, resulting in their not
> > > > > matching.
> > > >
> > > > As I said earlier the feature still can be supported if
> > > > kernel does not use IMA but obviously needs to be flagged.
> > >
> > > Jumping to the conclusion that "hibernate" is acceptable for non-IMA
> > > enabled kernels misses the security implications of mixing (kexec) non-
> > > IMA and IMA enabled kernels.
> > > I would prefer some sort of hibernate marker, the equivalent of a
> > > "boot_aggregate" record.
> >
> > Not sure if this matters. If you run a kernel, which is not aware
> > of IMA, it's your choice. I don't undestand why here is so important
> > to protect user from doing illogical decisions.
> >
> > If you want non-IMA kernels to support IMA, CONFIG_IMA should not
> > probably even exist because you are essentially saying that any
> > kernel play well with IMA.
>
> That will never happen, nor am I suggesting it should.
>
> Enabling hibernate or IMA shouldn't be an either-or decision, if at all
> possible. The main concern is that attestation servers be able to
> detect hibernation and possibly the loss of measurement
> history. Luckily, although the PCRs are reset, the TPM
> pcrUpdateCounter is not.
>
> I would appreciate including a "hibernate" marker, similar to the
> "boot_aggregate".

Yeah, I guess that would not do harm.

BR, Jarkko