[RESEND PATCH v18 2/4] overlayfs: handle XATTR_NOSECURITY flag for get xattr method

[Tyler Hicks]

On 2021-02-05 12:01:55, Tyler Hicks wrote:
> On 2020-10-30 09:00:35, Mark Salyzyn wrote:
> > On 10/30/20 8:07 AM, Miklos Szeredi wrote:
> > > On Wed, Oct 21, 2020 at 5:19 PM Mark Salyzyn <salyzyn at android.com> wrote:
> > > > Because of the overlayfs getxattr recursion, the incoming inode fails
> > > > to update the selinux sid resulting in avc denials being reported
> > > > against a target context of u:object_r:unlabeled:s0.
> > > > 
> > > > Solution is to respond to the XATTR_NOSECURITY flag in get xattr
> > > > method that calls the __vfs_getxattr handler instead so that the
> > > > context can be read in, rather than being denied with an -EACCES
> > > > when vfs_getxattr handler is called.
> > > > 
> > > > For the use case where access is to be blocked by the security layer.
> > > > 
> > > > The path then would be security(dentry) ->
> > > > __vfs_getxattr({dentry...XATTR_NOSECURITY}) ->
> > > > handler->get({dentry...XATTR_NOSECURITY}) ->
> > > > __vfs_getxattr({realdentry...XATTR_NOSECURITY}) ->
> > > > lower_handler->get({realdentry...XATTR_NOSECURITY}) which
> > > > would report back through the chain data and success as expected,
> > > > the logging security layer at the top would have the data to
> > > > determine the access permissions and report back to the logs and
> > > > the caller that the target context was blocked.
> > > > 
> > > > For selinux this would solve the cosmetic issue of the selinux log
> > > > and allow audit2allow to correctly report the rule needed to address
> > > > the access problem.
> > > > 
> > > > Check impure, opaque, origin & meta xattr with no sepolicy audit
> > > > (using __vfs_getxattr) since these operations are internal to
> > > > overlayfs operations and do not disclose any data.  This became
> > > > an issue for credential override off since sys_admin would have
> > > > been required by the caller; whereas would have been inherently
> > > > present for the creator since it performed the mount.
> > > > 
> > > > This is a change in operations since we do not check in the new
> > > > ovl_do_getxattr function if the credential override is off or not.
> > > > Reasoning is that the sepolicy check is unnecessary overhead,
> > > > especially since the check can be expensive.
> > > > 
> > > > Because for override credentials off, this affects _everyone_ that
> > > > underneath performs private xattr calls without the appropriate
> > > > sepolicy permissions and sys_admin capability.  Providing blanket
> > > > support for sys_admin would be bad for all possible callers.
> > > > 
> > > > For the override credentials on, this will affect only the mounter,
> > > > should it lack sepolicy permissions. Not considered a security
> > > > problem since mounting by definition has sys_admin capabilities,
> > > > but sepolicy contexts would still need to be crafted.
> > > This would be a problem when unprivileged mounting of overlay is
> > > introduced.  I'd really like to avoid weakening the current security
> > > model.
> > 
> > The current security model does not deal with non-overlapping security
> > contexts between init (which on android has MAC permissions only when
> > necessary, only enough permissions to perform the mount and other mundane
> > operations, missing exec and read permissions in key spots) and user calls.
> > 
> > We are only weakening (that is actually an incorrect statement, security is
> > there, just not double security of both mounter and caller) the security
> > around calls that retrieve the xattr for administrative and internal
> > purposes. No data is exposed to the caller that it would not otherwise have
> > permissions for.
> 
> We've ran into the same issues that Mark is trying to solve with this
> series. I came across Mark's series while searching around before I
> wrote up a similar patch to Mark's patch #3.
> 
> We have a confined process that sets up Overlayfs mounts, then that process
> starts a service confined by another security context, then that service
> may execute binaries that run under a third security context. In this
> case, I'm talking about SELinux security contexts but it could be
> AppArmor or anything else that you use to separate out
> privileges/permissions at fine-grained detail.
> 
> We don't want to grant all the privileges/permissions required by the
> service (and its helper utilities) to the process that sets up the
> Overlayfs mounts because we've been very careful in separating them
> apart with security policy. However, we want to make use of Overlayfs
> and adding a mount option to bypass the check on the mounter's cred
> seems like a safe way of using Overlayfs without violating our principle
> of least privilege.

I just realized that I missed one noteworthy aspect of our use case. We
would be alright if this functionality to bypass the mounter's cred
checks (conditional, based on a mount option) was only allowed for
read-only overlays[1].

I'm not sure if that would meet the needs of Android. Can you comment on
that, Mark?

Miklos, would that restriction make this series any more acceptable to
you?

Thanks!

Tyler

[1] https://www.kernel.org/doc/html/latest/filesystems/overlayfs.html#multiple-lower-layers

> 
> Tyler
> 
> > 
> > This patch becomes necessary when matched with the PATCH v18 3/4 of the
> > series which fixes the user space break introduced in ~4.6 that formerly
> > used the callers credentials for all accesses in all places. Security is
> > weakened already as-is in overlayfs with all the overriding of the
> > credentials for internal accesses to overlayfs mechanics based on the
> > mounter credentials. Using the mounter credentials as a wider security hole
> > is the problem, at least with PATCH v18 3/4 of the series we go back
> > optionally to only using the caller's credentials to perform the operations.
> > Admittedly some of the internal operations like mknod are privileged, but at
> > least in Android's use case we are not using them with callers without the
> > necessary credentials.
> > 
> > Android does not give the mounter more credentials than the callers, there
> > is very little overlap in the MAC security.
> > 
> > > The big API churn in the 1/4 patch also seems excessive considering
> > > that this seems to be mostly a cosmetic issue for android.  Am I
> > > missing something?
> > 
> > Breaks sepolicy, it no longer has access to the context data at the
> > overlayfs security boundary.
> > 
> > unknown is a symptom of being denied based on the denial to xattr data from
> > the underlying filesystem layer. Being denied the security context of the
> > target is not a good thing within the sepolicy security layer.
> > 
> > > 
> > > Thanks,
> > > Miklos
> > 
> >