Date:   Wed, 13 May 2020 11:09:02 -0400
From:   Mimi Zohar <zohar@...ux.ibm.com>
To:     Roberto Sassu <roberto.sassu@...wei.com>,
        "david.safford@...il.com" <david.safford@...il.com>,
        "viro@...iv.linux.org.uk" <viro@...iv.linux.org.uk>,
        "jmorris@...ei.org" <jmorris@...ei.org>,
        John Johansen <john.johansen@...onical.com>,
        "matthewgarrett@...gle.com" <matthewgarrett@...gle.com>
Cc:     "linux-fsdevel@...r.kernel.org" <linux-fsdevel@...r.kernel.org>,
        "linux-integrity@...r.kernel.org" <linux-integrity@...r.kernel.org>,
        "linux-security-module@...r.kernel.org" 
        <linux-security-module@...r.kernel.org>,
        "linux-kernel@...r.kernel.org" <linux-kernel@...r.kernel.org>,
        Silviu Vlasceanu <Silviu.Vlasceanu@...wei.com>
Subject: Re: [RFC][PATCH 1/3] evm: Move hooks outside LSM infrastructure

On Wed, 2020-05-13 at 07:21 +0000, Roberto Sassu wrote:
> > From: Mimi Zohar [mailto:zohar@...ux.ibm.com]
> > Sent: Tuesday, May 12, 2020 9:38 PM
> > On Tue, 2020-05-12 at 16:31 +0000, Roberto Sassu wrote:
> > > > From: Mimi Zohar [mailto:zohar@...ux.ibm.com]
> > 
> > > > > > Each time the EVM protected file metadata is updated, the EVM
> > HMAC
> > > > is
> > > > > > updated, assuming the existing EVM HMAC is valid.  Userspace
> > should
> > > > > > not have access to the HMAC key, so we only allow writing EVM
> > > > > > signatures.
> > > > > >
> > > > > > The only difference between writing the original EVM signature and
> > the
> > > > > > new portable and immutable signature is the security.ima xattr
> > > > > > requirement.  Since the new EVM signature does not include the
> > > > > > filesystem specific data, something else needs to bind the file
> > > > > > metadata to the file data.  Thus the IMA xattr requirement.
> > > > > >
> > > > > > Assuming that the new EVM signature is written last, as long as there
> > > > > > is an IMA xattr, there shouldn't be a problem writing the new EVM
> > > > > > signature.
> > > > >
> > > > >         /* first need to know the sig type */
> > > > >         rc = vfs_getxattr_alloc(dentry, XATTR_NAME_EVM, (char
> > > > **)&xattr_data, 0,
> > > > >                                 GFP_NOFS);
> > > > >         if (rc <= 0) {
> > > > >                 evm_status = INTEGRITY_FAIL;
> > > > >                 if (rc == -ENODATA) {
> > > > >                         rc = evm_find_protected_xattrs(dentry);
> > > > >                         if (rc > 0)
> > > > >                                 evm_status = INTEGRITY_NOLABEL;
> > > > >                         else if (rc == 0)
> > > > >                                 evm_status = INTEGRITY_NOXATTRS; /* new file */
> > > > >
> > > > > If EVM_ALLOW_METADATA_WRITES is cleared, only the first xattr
> > > > > can be written (status INTEGRITY_NOXATTRS is ok). After,
> > > > > evm_find_protected_xattrs() returns rc > 0, so the status is
> > > > > INTEGRITY_NOLABEL, which is not ignored by evm_protect_xattr().
> > > >
> > > > With EVM HMAC enabled, as a result of writing the first protected
> > > > xattr, an EVM HMAC should be calculated and written in
> > > > evm_inode_post_setxattr().
> > >
> > > To solve the ordering issue, wouldn't allowing setxattr() on a file
> > > with portable signature that does not yet pass verification be safe?
> > > evm_update_evmxattr() checks if the signature is portable and
> > > if yes, does not calculate the HMAC.
> > 
> > Before agreeing to allowing the protected xattrs to be written on a
> > file with a portable signature that does not yet pass verification are
> > safe, would we be introducing any new types of attacks?
> 
> Allowing xattr/attr update means that someone can do:
> 
> setxattr(path, "security.evm", ...);	with type=5
> 
> all subsequent setxattr()/setattr() succeed until the correct
> combination is set.
> 
> At that point, any xattr/attr operation fails, even if one tries to set
> an xattr with the same value. If we still deny the operation when the
> verification succeeds, we have to fix that.
> 
> It is common that the signature passes verification before user space
> tools finish to set xattrs/attrs. For example, if a file is created with
> mode 644 and this was the mode at the time of signing, any attempt
> by tar for example to set again the same mode fails.

We have a couple of options: always fail the write, differentiate the
reason for the failure, or check the value before failing the write.
 The easiest obviously would be to always fail the write once it is
valid.  Whatever you decide, please keep it simple.

> 
> If allowing a change of xattrs/attrs for portable signatures is safe or
> not, I would say yes. Portable signatures cannot be modified even
> if __vfs_setxattr_noperm() is called directly.
> 
> > For example, would we differentiate between portable signatures that
> > don't pass verification and ones that do?  If we don't differentiate,
> > could it be used for DoS?  Should it be limited to new files?
> 
> I would prefer to lock files when signatures pass the verification to
> avoid accidental changes.
> 
> Unless we find a better way to identify new file, without depending
> on the appraisal policy, I would allow the operation even for existing
> files.

The existing code verifies the EVM xattr before allowing the
xattr/attr change.  It sounds like for EVM_XATTR_PORTABLE_DIGSIG we
need to do exactly the reverse.

Mimi

Powered by blists - more mailing lists