[PATCH] evm: Correct inode_init_security hooks behaviors

[Mimi Zohar]

Hi Nicolas,

On Fri, 2022-10-21 at 15:47 +0200, Nicolas Bouchinet wrote:
> Hi Mimi,
> 
> Thanks for the IMA/EVM project which I enjoy very much.
> 
> On Thu, Oct 20, 2022 at 03:51:38PM -0400, Mimi Zohar wrote:
> > On Thu, 2022-10-20 at 15:55 +0200, Nicolas Bouchinet wrote:
> > > From: Nicolas Bouchinet <nicolas.bouchinet at ssi.gouv.fr>
> > > 
> > > Fixes a NULL pointer dereference occuring in the
> > > `evm_protected_xattr_common` function of the EVM LSM. The bug is
> > > triggered if a `inode_init_security` hook returns 0 without initializing
> > > the given `struct xattr` fields (which is the case of BPF) and if no
> > > other LSM overrides thoses fields after. This also leads to memory
> > > leaks.
> > > 
> > > Adds a `call_int_hook_xattr` macro that fetches and feed the
> > > `new_xattrs` array with every called hook xattr values.
> > > 
> > > Adds a `evm_init_hmacs` function which init the EVM hmac using every
> > > entry of the array contrary to `evm_init_hmac`.
> >   
> > Only EVM portable digital signatures include all of the protected
> > xattrs.   Refer to commit 8c7a703ec978 ("evm: Verify portable
> > signatures against all protected xattrs").
> > 
> Sorry, maybe I was not clear enough, the proposed patch does not change the
> set of the protected security xattrs initialized by the LSMs and processed by EVM.
> 
> As I explained to Paul, based on my understanding, the `security_inode_init_security()`
> hook is supposed to initialize every hooked LSM security xattr and next,
> if evm is enabled, protect them using a HMAC algorithm.
> However, in the current implementation, the use of the `call_int_hook()` macro by
> `security_inode_init_security()` overwrites the previously initialized xattr for
> each iteration of the `hlist_for_each_entry()` loop.
> 
> I have noticed that more than one LSM may initialize a security xattr at a time,
> eg. SELinux + BPF.

Does BPF have a security xattr and, if so, does it need to be
protected?   It would need to be defined and included in the list of
evm_config_xattrnames[].  If it doesn't define a security bpf xattr,
then bpf should not be on the security_inode_init_security() hook.  (I
assume Roberto's patch is going in this direction.)

Before the EVM hmac is updated, the existing EVM hmac is verified.  I
would be concerned if bpf defined a protected security xattr.   Could
the same guarantees, that security.evm isn't updated without first
being verified, be enforced with bpf?

> 
> IMHO my supplementary `evm_init_hmacs()` function name is a bit confusing, I would
> enjoy if you have a better proposition. Note that `evm_init_hmacs()` have the same
> behavior as `evm_init_hmac()` if only one security xattr is given as a parameter.

I'm missing something here.  As evm_inode_init_security() is the only
caller of evm_init_hmac(), why is a new function defined instead of
updating the existing one?   If there is a valid reason, then one
function should be a wrapper for the other.

> > > 
> > > Fixes the `evm_inode_init_security` function to use `evm_init_hmacs`.
> > 
> > Won't this break existing EVM hmac usage?
> I might be wrong, but as far as I understand it, the only working condition for
> EVM now is when only one security xattr is involved, otherwise there will have
> a mismatch between the initialization and the verification.
> Indeed, the verification takes into account every security xattr written in its
> refering dentry.

Agreed, independently as to whether BPF defines a security xattr, if
two LSMs initialize security xattrs, then this change is needed.  Are
there any other examples?

(nit: I understand the line size has generally been relaxed, but for
IMA/EVM I would prefer it to be remain as 80 chars.)

Mimi