Re: [RFC] io_uring: add restrictions to support untrusted applications and guests

[Jens Axboe <axboe@kernel.dk>]

On 6/15/20 7:33 AM, Stefano Garzarella wrote:
> On Mon, Jun 15, 2020 at 11:04:06AM +0200, Jann Horn wrote:
>> +Kees, Christian, Sargun, Aleksa, kernel-hardening for their opinions
>> on seccomp-related aspects
>>
>> On Tue, Jun 9, 2020 at 4:24 PM Stefano Garzarella <sgarzare@redhat.com> wrote:
>>> Hi Jens,
>>> Stefan and I have a proposal to share with io_uring community.
>>> Before implementing it we would like to discuss it to receive feedbacks and
>>> to see if it could be accepted:
>>>
>>> Adding restrictions to io_uring
>>> =====================================
>>> The io_uring API provides submission and completion queues for performing
>>> asynchronous I/O operations. The queues are located in memory that is
>>> accessible to both the host userspace application and the kernel, making it
>>> possible to monitor for activity through polling instead of system calls. This
>>> design offers good performance and this makes exposing io_uring to guests an
>>> attractive idea for improving I/O performance in virtualization.
>> [...]
>>> Restrictions
>>> ------------
>>> This document proposes io_uring API changes that safely allow untrusted
>>> applications or guests to use io_uring. io_uring's existing security model is
>>> that of kernel system call handler code. It is designed to reject invalid
>>> inputs from host userspace applications. Supporting guests as io_uring API
>>> clients adds a new trust domain with access to even fewer resources than host
>>> userspace applications.
>>>
>>> Guests do not have direct access to host userspace application file descriptors
>>> or memory. The host userspace application, a Virtual Machine Monitor (VMM) such
>>> as QEMU, grants access to a subset of its file descriptors and memory. The
>>> allowed file descriptors are typically the disk image files belonging to the
>>> guest. The memory is typically the virtual machine's RAM that the VMM has
>>> allocated on behalf of the guest.
>>>
>>> The following extensions to the io_uring API allow the host application to
>>> grant access to some of its file descriptors.
>>>
>>> These extensions are designed to be applicable to other use cases besides
>>> untrusted guests and are not virtualization-specific. For example, the
>>> restrictions can be used to allow only a subset of sqe operations available to
>>> an application similar to seccomp syscall whitelisting.
>>>
>>> An address translation and memory restriction mechanism would also be
>>> necessary, but we can discuss this later.
>>>
>>> The IOURING_REGISTER_RESTRICTIONS opcode
>>> ----------------------------------------
>>> The new io_uring_register(2) IOURING_REGISTER_RESTRICTIONS opcode permanently
>>> installs a feature whitelist on an io_ring_ctx. The io_ring_ctx can then be
>>> passed to untrusted code with the knowledge that only operations present in the
>>> whitelist can be executed.
>>
>> This approach of first creating a normal io_uring instance and then
>> installing restrictions separately in a second syscall means that it
>> won't be possible to use seccomp to restrict newly created io_uring
>> instances; code that should be subject to seccomp restrictions and
>> uring restrictions would only be able to use preexisting io_uring
>> instances that have already been configured by trusted code.
>>
>> So I think that from the seccomp perspective, it might be preferable
>> to set up these restrictions in the io_uring_setup() syscall. It might
>> also be a bit nicer from a code cleanliness perspective, since you
>> won't have to worry about concurrently changing restrictions.
>>
> 
> Thank you for these details!
> 
> It seems feasible to include the restrictions during io_uring_setup().
> 
> The only doubt concerns the possibility of allowing the trusted code to
> do some operations, before passing queues to the untrusted code, for
> example registering file descriptors, buffers, eventfds, etc.
> 
> To avoid this, I should include these operations in io_uring_setup(),
> adding some code that I wanted to avoid by reusing io_uring_register().
> 
> If I add restrictions in io_uring_setup() and then add an operation to
> go into safe mode (e.g. a flag in io_uring_enter()), we would have the same
> problem, right?
> 
> Just to be clear, I mean something like this:
> 
>     /* params will include restrictions */
>     fd = io_uring_setup(entries, params);
> 
>     /* trusted code */
>     io_uring_register_files(fd, ...);
>     io_uring_register_buffers(fd, ...);
>     io_uring_register_eventfd(fd, ...);
> 
>     /* enable safe mode */
>     io_uring_enter(fd, ..., IORING_ENTER_ENABLE_RESTRICTIONS);
> 
> 
> Anyway, including a list of things to register in the 'params', passed
> to io_uring_setup(), should be feasible, if Jens agree :-)

I wonder how best to deal with this, in terms of ring visibility vs
registering restrictions. We could potentially start the ring in a
disabled mode, if asked to. It'd still be visible in terms of having
the fd installed, but it'd just error requests. That'd leave you with
time to do the various setup routines needed before then flagging it
as enabled. My only worry on that would be adding overhead for doing
that. It'd be cheap enough to check for IORING_SETUP_DISABLED in
ctx->flags in io_uring_enter(), and return -EBADFD or something if
that's the case. That doesn't cover the SQPOLL case though, but maybe we
just don't start the sq thread if IORING_SETUP_DISABLED is set.

We'd need a way to clear IORING_SETUP_DISABLED through
io_uring_register(). When clearing, that could then start the sq thread
as well, when SQPOLL is set.

-- 
Jens Axboe