Jann Horn <j...@thejh.net> writes:
> On Tue, Oct 18, 2016 at 09:56:53AM -0500, Eric W. Biederman wrote:
>> Michal Hocko <mho...@kernel.org> writes:
>> > On Mon 17-10-16 11:39:49, Eric W. Biederman wrote:
>> >> During exec dumpable is cleared if the file that is being executed is
>> >> not readable by the user executing the file. A bug in
>> >> ptrace_may_access allows reading the file if the executable happens to
>> >> enter into a subordinate user namespace (aka clone(CLONE_NEWUSER),
>> >> unshare(CLONE_NEWUSER), or setns(fd, CLONE_NEWUSER).
>> >> This problem is fixed with only necessary userspace breakage by adding
>> >> a user namespace owner to mm_struct, captured at the time of exec,
>> >> so it is clear in which user namespace CAP_SYS_PTRACE must be present
>> >> in to be able to safely give read permission to the executable.
>> >> The function ptrace_may_access is modified to verify that the ptracer
>> >> has CAP_SYS_ADMIN in task->mm->user_ns instead of task->cred->user_ns.
>> >> This ensures that if the task changes it's cred into a subordinate
>> >> user namespace it does not become ptraceable.
>> > I haven't studied your patch too deeply but one thing that immediately
>> > raised a red flag was that mm might be shared between processes (aka
>> > thread groups). What prevents those two to sit in different user
>> > namespaces?
>> > I am primarily asking because this generated a lot of headache for the
>> > memcg handling as those processes might sit in different cgroups while
>> > there is only one correct memcg for them which can disagree with the
>> > cgroup associated with one of the processes.
>> That is a legitimate concern, but I do not see any of those kinds of
>> issues here.
>> Part of the memcg pain comes from the fact that control groups are
>> process centric, and part of the pain comes from the fact that it is
>> possible to change control groups. What I am doing is making the mm
>> owned by a user namespace (at creation time), and I am not allowing
>> changes to that ownership. The credentials of the tasks that use that mm
>> may be in the same user namespace or descendent user namespaces.
>> The core goal is to enforce the unreadability of an mm when an
>> non-readable file is executed. This is a time of mm creation property.
>> The enforcement of which fits very well with the security/permission
>> checking role of the user namespace.
> How is that going to work? I thought the core goal was better security for
> entering containers.
The better security when entering containers came from fixing the the
check for unreadable files. Because that is fundamentally what
the mm dumpable settings are for.
> If I want to dump a non-readable file, afaik, I can just make a new user
> namespace, then run the file in there and dump its memory.
> I guess you could fix that by entirely prohibiting the execution of a
> non-readable file whose owner UID is not mapped. (Adding more dumping
> restrictions wouldn't help much because you could still e.g. supply a
> malicious dynamic linker if you control the mount namespace.)
That seems to be a part of this puzzle I have incompletely addressed,
It looks like I need to change either the owning user namespace or
fail the exec. Malicious dynamic linkers are doubly interesting.
As mount name spaces are also owned if I have privileges I can address
the possibility of a malicious dynamic linker that way. AKA who cares
about the link if the owner of the mount namespace has permissions to
read the file.
I am going to look at failing the exec if the owning user namespace
of the mm would not have permissions to read the file. That should just
be a couple of lines of code and easy to maintain. Plus it does not
appear that non-readable executables are particularly common.