Re: [PATCH V2 3/4] IMA: Optionally make use of filesystem-provided hashes

[Matthew Garrett <mjg59@google.com>]

On Wed, Mar 6, 2019 at 2:39 PM Mimi Zohar <zohar@linux.ibm.com> wrote:
>
> On Wed, 2019-03-06 at 10:31 -0800, Matthew Garrett wrote:
> > Ok. Would annotating the audit message to indicate that the hash was
> > provided directly by the filesystem be sufficient?
>
> The audit log doesn't have the same security properties as the TPM
> quote and IMA measurement list.  Nor does the attestation server
> necessarily have access to it.

Ok.

> > I'm not clear on
> > why an admin would set this flag without having read the documentation
> > for it - like many security features, enabling an inappropriate
> > combination of them may result in bad things happening.  I'm not keen
> > on tying it to signing because:
> >
> > a) There are multiple configurations where requiring signed policy
> > doesn't give a security benefit - if the IMA policy is part of a
> > verified or measured initramfs, we already have integrity guarantees
> > and adding an additional layer of signing doesn't win us anything (eg,
> > in this configuration the IMA key may be loaded from the initramfs as
> > well, so an attacker able to modify policy could add an additional
> > signing key).
>
> Agreed, as long as there is no possibility of additional files being
> installed/downloaded to the rootfs or files on other filesystems being
> accessed before the IMA keyring is locked or a custom policy is
> installed, a verified or measured initramfs might be enough.
>
> This implies that both the custom policy and the keys loaded onto the
> IMA keyring are included in the initramfs, which isn't what is
> currently being done today.  My preference would be to remove the
> original method of loading unsigned IMA policies, but that could/would
> break existing systems.

It's the way we handle IMA configuration - the policy is loaded and
further policy loads are disabled due to IMA_WRITE_POLICY being
default n. We're not currently making use of signing, but longer term
plan is for the keys to be loaded at the same time.

> > b) Users who are already using signed policy won't get the additional
> > hint that you think is necessary.
>
> But they would have to knowingly add "get_hash" to the IMA policy and
> have signed it.

But in the non-signed case they'd still need to knowingly add
"get_hash" to the IMA policy. Why does signing indicate stronger
understanding of policy? If my understanding of ima_match_policy()
correct, if there's already a measurement rule that applies to a
filesystem then adding an additional trust_vfs rule will be ignored,
so once the initial policy is loaded it's not possible for someone to
transition a filesystem from a full read to using the vfs call. IE, a
policy like:

measure
measure fsmagic=0x46555345 trust_vfs

is still going to perform the full measurement even on FUSE.

> > I'm happy to add this if there's a real threat model around it, but
> > requiring signing for something other than security reasons seems like
> > it's conflating unrelated issues.
>
> A colleague said, relying on the filesystem to provide the file hash
> extends the TCB to include filesystems.

The TCB already includes filesystems - IMA's measurements are only
accurate if the filesystem returns the same data on subsequent reads
(assuming i_version hasn't been updated). We assert that this is true,
but it the filesystem is outside the TCB then that assertion is
invalid.