On Tue, 2017-10-31 at 14:11 +1100, James Morris wrote: > On Mon, 30 Oct 2017, Stephen Smalley wrote: > > > Thanks, interesting approach. One drawback is that it doesn't > > presently > > support any form of inheritance of labels from the parent > > namespace, so > > files that are shared read-only from the init namespace will show > > up as > > unlabeled in the child namespace until they are assigned the > > namespaced > > attributes. This for example breaks running the selinux-testsuite > > with > > this patch applied (unless perhaps you run restorecon -R / after > > unsharing); otherwise just trying to invoke /usr/bin/runcon will > > fail > > since it is unlabeled in the child. It seems like we should > > provide > > some form of inheritance from the parent when there is no xattr for > > the > > namespace itself. > > I was assuming that practical use of this would involve doing a > filesystem > relabel under the newly loaded policy, on first instantiation at > least. I think that would be problematic, e.g. if you want to share /usr read- only to all of the containers or similar, then they cannot relabel it themselves and even if you provided a way to set the xattrs from the init namespace, in the case where you are using the same or very similar policies, it seems wasteful to set the same xattr values for N xattr names for N containers. So I think inheritance support will be necessary. This will be easier if we make the xattr names themselves hierarchical (e.g. if vm8 writes vm1 to /sys/fs/selinux/unshare, then the child xattr name would be security.selinux.ns.vm8.vm1, and we would inherit from either security.selinux.ns.vm8 or security.selinux if security.selinux.ns.vm8.vm1 is not set.) > > We could try adding an selinuxfs node to specify default handling of > unlabeled files in a child namespace, and write to that after > mounting > selinuxfs in that namespace. > > e.g. echo inherit > /sys/fs/selinux/parent_ns_labels > > or something. > > > > > > Another potential concern is that files created in a non-init > > namespace > > are left completely unlabeled in the init namespace (or in any > > parent). > > As long as access to unlabeled is tightly controlled, this > > might > > not be a problem, but I'm not sure that's guaranteed by the > > refpolicy > > or Fedora/RHEL policies. We might want to initialize an xattr at > > every > > level of the namespace hierarchy when a new file is created based > > on > > the process and parent directory labels and policy at that level. > > Otherwise, we lose all provenance information for the file outside > > of > > the namespace. > > Ok. > > > > For example, suppose I want to leak information out of > > my category set; I unshare and create files with the information, > > and > > they appear in the init namespace with no categories. > > Nice :) >
