Re: [PATCH v15 00/11] LSM: Three basic syscalls

[Roberto Sassu]

On Wed, 2023-10-18 at 16:14 +0200, Roberto Sassu wrote:
> On 10/18/2023 3:09 PM, Mimi Zohar wrote:
> > On Wed, 2023-10-18 at 11:31 +0200, Roberto Sassu wrote:
> > > On Tue, 2023-10-17 at 18:07 +0200, Roberto Sassu wrote:
> > > > On Tue, 2023-10-17 at 11:58 -0400, Paul Moore wrote:
> > > > > On Tue, Oct 17, 2023 at 3:01 AM Roberto Sassu
> > > > > <roberto.sassu@xxxxxxxxxxxxxxx> wrote:
> > > > > > On Mon, 2023-10-16 at 11:06 -0400, Paul Moore wrote:
> > > > > > > On Mon, Oct 16, 2023 at 8:05 AM Roberto Sassu
> > > > > > > <roberto.sassu@xxxxxxxxxxxxxxx> wrote:
> > > > > > > > 
> > > > > > > > Sorry, I just noticed LSM_ID_IMA. Since we have the 'integrity' LSM, I
> > > > > > > > think it should be LSM_ID_INTEGRITY.
> > > > > > > > 
> > > > > > > > Mimi, all, do you agree? If yes, I send a patch shortly.
> > > > > > > 
> > > > > > > I believe LSM_ID_IMA is the better option, despite "integrity" already
> > > > > > > being present in Kconfig and possibly other areas.  "IMA" is a
> > > > > > > specific thing/LSM whereas "integrity" is a property, principle, or
> > > > > > > quality.  Especially as we move forward with promoting IMA as a full
> > > > > > > and proper LSM, we should work towards referring to it as "IMA" and
> > > > > > > not "integrity".
> > > > > > > 
> > > > > > > If anything we should be working to support "IMA" in places where we
> > > > > > > currently have "integrity" so that we can eventually deprecate
> > > > > > > "integrity".
> > > > > > 
> > > > > > Hi Paul
> > > > > > 
> > > > > > I fully understand your argument. However, 'integrity' has been the
> > > > > > word to identify the integrity subsystem since long time ago.
> > > > > > 
> > > > > > Reducing the scope to 'ima' would create some confusion since, while
> > > > > > 'ima' is associated to integrity, it would not encompass EVM.
> > > > > 
> > > > > Using LSM_ID_IMA to reference the combination of IMA+EVM makes much
> > > > > more sense to me than using LSM_ID_INTEGRITY, especially as we move
> > > > > towards promoting IMA+EVM and adopting LSM hooks for integrity
> > > > > verification, opening the door for other integrity focused LSMs.
> > > > 
> > > > + Mimi, linux-integrity
> > > > 
> > > > Ok, just to understand before posting v4, the code looks like this:
> > > 
> > > I worked on a new proposal. Let me know what you think. It is available
> > > here:
> > > 
> > > https://github.com/robertosassu/linux/tree/ima-evm-lsms-v4-devel-v6
> > > 
> > > 
> > > I made IMA and EVM as standalone LSMs and removed 'integrity'. They
> > > maintain the same properties of 'integrity', i.e. they are the last and
> > > always enabled.
> > > 
> > > During initialization, 'ima' and 'evm' call integrity_iintcache_init(),
> > > so that they can get integrity metadata. I added a check to ensure that
> > > this function is called only once. I also added the lsmid parameter so
> > > that the integrity-specific functions are added under the LSM ID of the
> > > caller.
> > > 
> > > I added a new LSM ID for EVM, does not look good that IMA and EVM are
> > > represented by LSM_ID_IMA.
> > > 
> > > Finally, I had to drop the patch to remove the rbtree, because without
> > > the 'integrity' LSM, space in the security blob cannot be reserved.
> > > Since integrity metadata is shared, it cannot be reserved by 'ima' or
> > > 'evm'.
> > > 
> > > An intermediate solution would be to keep the 'integrity' LSM just to
> > > reserve space in the security blob. Or, we remove the rbtree if/when
> > > IMA and EVM use disjoint integrity metadata.
> > 
> > One of the major benefits for making IMA and EVM LSMs was removing the
> > rbtree and replacing it with the ability of using i_security.
> > 
> > I agree with Roberto.  All three should be defined: LSM_ID_INTEGRITY,
> > LSM_ID_IMA, LSM_ID_EVM.
> 
> I did not try yet, but the 'integrity' LSM does not need an LSM ID. With 
> the last version adding hooks to 'ima' or 'evm', it should be sufficient 
> to keep DEFINE_LSM(integrity) with the request to store a pointer in the 
> security blob (even the init function can be a dummy function).

Ok, I rebased on top of Paul's 'lsm: drop LSM_ID_IMA', made IMA and EVM
as standalone LSMs, and assigned LSM IDs to them in the correct
chronological order.

Still left DEFINE_LSM(integrity) to reserve space for the
integrity_iint_cache pointer, but it does not register any LSM hook,
then no LSM ID required.

In the future, we might be able to make IMA and EVM use disjoint
metadata, so they will reserve space in the security blob and register
appropriate hooks for the metadata management.

With this intermediate solution, I can keep the rbtree patch, which
will provide performance improvements due to searching metadata in
constant time.

If you want to have a look before I send the patch set, the code is
available here:

https://github.com/robertosassu/linux/commits/ima-evm-lsms-v4-devel-v7

It passes all IMA tests:

https://github.com/robertosassu/ima-evm-utils/actions/runs/6571587570

Roberto