On Fri, Sep 28, 2018 at 1:43 AM James Morris [off-list ref] wrote:
On Thu, 27 Sep 2018, Schaufler, Casey wrote:
quoted
quoted
quoted
On 9/27/2018 2:45 PM, James Morris wrote:
quoted
On Wed, 26 Sep 2018, Casey Schaufler wrote:
quoted
+ /*
+ * Namespace checks. Considered safe if:
+ * cgroup namespace is the same
+ * User namespace is the same
+ * PID namespace is the same
+ */
+ if (current->nsproxy)
+ ccgn = current->nsproxy->cgroup_ns;
+ if (p->nsproxy)
+ pcgn = p->nsproxy->cgroup_ns;
+ if (ccgn != pcgn)
+ return -EACCES;
+ if (current->cred->user_ns != p->cred->user_ns)
+ return -EACCES;
+ if (task_active_pid_ns(current) != task_active_pid_ns(p))
+ return -EACCES;
+ return 0;
I really don't like the idea of hard-coding namespace security semantics
in an LSM. Also, I'm not sure if these semantics make any sense.
Checks on namespaces where explicitly requested.
By whom and what is the rationale?
The rationale is to protect containers. Since those closest thing
there is to a definition of containers is "uses namespaces" that
becomes the focus. Separating them out does not make too much
sense as I would expect someone concerned with one to be concerned
with all.
A lot of people will not be using user namespaces due to security
concerns,
Ugh.
so with this hard-coded logic, you are saying this case is
'safe' in a sidechannel context.
Which hints at the deeper issue that containers are a userland
abstraction. Protection of containers needs to be defined by userland
policy.
Or just compare mount namespaces additionally/instead. I think that
containers will always use those, because AFAIK nobody uses chroot()
for containers, given that the kernel makes absolutely no security
guarantees about chroot().