[PATCH v1 0/2] Add LSM access controls for io_uring_setup

[Paul Moore]

On Thu, Nov 10, 2022 at 12:54 PM Jeffrey Vander Stoep <jeffv at google.com> wrote:
> On Mon, Nov 7, 2022 at 10:17 PM Paul Moore <paul at paul-moore.com> wrote:
> >
> > On Mon, Nov 7, 2022 at 3:58 PM Gil Cukierman <cukie at google.com> wrote:
> > >
> > > This patchset provides the changes required for controlling access to
> > > the io_uring_setup system call by LSMs. It does this by adding a new
> > > hook to io_uring. It also provides the SELinux implementation for a new
> > > permission, io_uring { setup }, using the new hook.
> > >
> > > This is important because existing io_uring hooks only support limiting
> > > the sharing of credentials and access to the sensitive uring_cmd file
> > > op. Users of LSMs may also want the ability to tightly control which
> > > callers can retrieve an io_uring capable fd from the kernel, which is
> > > needed for all subsequent io_uring operations.
> >
> > It isn't immediately obvious to me why simply obtaining a io_uring fd
> > from io_uring_setup() would present a problem, as the security
> > relevant operations that are possible with that io_uring fd *should*
> > still be controlled by other LSM hooks.  Can you help me understand
> > what security issue you are trying to resolve with this control?
>
> I think there are a few reasons why we want this particular hook.
>
> 1.  It aligns well with how other resources are managed by selinux
> where access to the resource is the first control point (e.g. "create"
> for files, sockets, or bpf_maps, "prog_load" for bpf programs, and
> "open" for perf_event) and then additional functionality or
> capabilities require additional permissions.

[NOTE: there were two reply sections in your email, and while similar,
they were not identical; I've trimmed the other for the sake of
clarity]

The resources you mention are all objects which contain some type of
information (either user data, configuration, or program
instructions), with the resulting fd being a handle to those objects.
In the case of io_uring the fd is a handle to the io_uring
interface/rings, which by itself does not contain any information
which is not already controlled by other permissions.

I/O operations which transfer data between the io_uring buffers and
other system objects, e.g. IORING_OP_READV, are still subject to the
same file access controls as those done by the application using
syscalls.  Even the IORING_OP_OPENAT command goes through the standard
VFS code path which means it will trigger the same access control
checks as an open*() done by the application normally.

The 'interesting' scenarios are those where the io_uring operation
servicing credentials, aka personalities, differ from the task
controlling the io_uring.  However in those cases we have the new
io_uring controls to gate these delegated operations.  Passing an
io_uring fd is subject to the fd/use permission like any other fd.

Although perhaps the most relevant to your request is the fact that
the io_uring inode is created using the new(ish) secure anon inode
interface which ensures that the creating task has permission to
create an io_uring.  This io_uring inode label also comes into play
when a task attempts to mmap() the io_uring rings, a critical part of
the io_uring API.

If I'm missing something you believe to be important, please share the details.

> 2. It aligns well with how resources are managed on Android. We often
> do not grant direct access to resources (like memory buffers).

Accessing the io_uring buffers requires a task to mmap() the io_uring
fd which is controlled by the normal SELinux mmap() access controls.

> 3. Attack surface management. One of the primary uses of selinux on
> Android is to assess and limit attack surface (e.g.
> https://twitter.com/jeffvanderstoep/status/1422771606309335043) . As
> io_uring vulnerabilities have made their way through our vulnerability
> management system, it's become apparent that it's complicated to
> assess the impact. Is a use-after-free reachable? Creating
> proof-of-concept exploits takes a lot of time, and often functionality
> can be reached by multiple paths. How many of the known io_uring
> vulnerabilities would be gated by the existing checks? How many future
> ones will be gated by the existing checks? I don't know the answer to
> either of these questions and it's not obvious. This hook makes that
> initial assessment simple and effective.

It should be possible to deny access to io_uring via the anonymous
inode labels, the mmap() controls, and the fd/use permission.  If you
find a way to do meaningful work with an io_uring fd that can't be
controlled via an existing permission check please let me know.

--
paul-moore.com