[PATCH v2] fscrypt: support trusted keys

[Eric Biggers]

On Wed, Aug 11, 2021 at 07:34:18AM -0400, Mimi Zohar wrote:
> On Wed, 2021-08-11 at 03:17 +0300, Jarkko Sakkinen wrote:
> > On Tue, Aug 10, 2021 at 02:27:24PM -0700, Eric Biggers wrote:
> > > On Wed, Aug 11, 2021 at 12:21:40AM +0300, Jarkko Sakkinen wrote:
> > > > On Tue, Aug 10, 2021 at 11:46:49AM -0700, Eric Biggers wrote:
> > > > > On Tue, Aug 10, 2021 at 09:06:36PM +0300, Jarkko Sakkinen wrote:
> > > > > > > > 
> > > > > > > > I don't think this is right, or at least it does not follow the pattern
> > > > > > > > in [*]. I.e. you should rather use trusted key to seal your fscrypt key.
> > > > > > > 
> > > > > > > What's the benefit of the extra layer of indirection over just using a "trusted"
> > > > > > > key directly?  The use case for "encrypted" keys is not at all clear to me.
> > > > > > 
> > > > > > Because it is more robust to be able to use small amount of trusted keys,
> > > > > > which are not entirely software based.
> > > > > > 
> > > > > > And since it's also a pattern on existing kernel features utilizing trusted
> > > > > > keys, the pressure here to explain why break the pattern, should be on the
> > > > > > side of the one who breaks it.
> > > > > 
> > > > > This is a new feature, so it's on the person proposing the feature to explain
> > > > > why it's useful.  The purpose of "encrypted" keys is not at all clear, and the
> > > > > documentation for them is heavily misleading.  E.g.:
> > > > > 
> > > > >     "user space sees, stores, and loads only encrypted blobs"
> > > > >     (Not necessarily true, as I've explained previously.)
> > > > > 
> > > > >     "Encrypted keys do not depend on a trust source" ...  "The main disadvantage
> > > > >     of encrypted keys is that if they are not rooted in a trusted key"
> > > > >     (Not necessarily true, and in fact it seems they're only useful when they
> > > > >     *do* depend on a trust source.  At least that's the use case that is being
> > > > >     proposed here, IIUC.)
> > > > > 
> > > > > I do see a possible use for the layer of indirection that "encrypted" keys are,
> > > > > which is that it would reduce the overhead of having lots of keys be directly
> > > > > encrypted by the TPM/TEE/CAAM.  Is this the use case?  If so, it needs to be
> > > > > explained.
> > > > 
> > > > If trusted keys are used directly, it's an introduction of a bottleneck.
> > > > If they are used indirectly, you can still choose to have one trusted
> > > > key per fscrypt key.
> > > > 
> > > > Thus, it's obviously a bad idea to use them directly.
> > > 
> > > So actually explain that in the documentation.  It's not obvious at all.  And
> > > does this imply that the support for trusted keys in dm-crypt is a mistake?
> > 
> > Looking at dm-crypt implementation, you can choose to use 'encrypted' key
> > type, which you can seal with a trusted key.
> > 
> > Note: I have not been involved when the feature was added to dm-crypt.
> 
> At least for TPM 1.2,  "trusted" keys may be sealed to a PCR and then
> extended to prevent subsequent usage.  For example, in the initramfs
> all of the "encrypted" keys could be decrypted by a single "trusted"
> key, before extending the PCR.
> 
> Mimi
> 

Neither of you actually answered my question, which is whether the support for
trusted keys in dm-crypt is a mistake.  I think you're saying that it is?  That
would imply that fscrypt shouldn't support trusted keys, but rather encrypted
keys -- which conflicts with Ahmad's patch which is adding support for trusted
keys.  Note that your reasoning for this is not documented at all in the
trusted-encrypted keys documentation; it needs to be (email threads don't really
matter), otherwise how would anyone know when/how to use this feature?

- Eric