Re: [RFC] Expose request_module via syscall

[Andy Lutomirski]

On Wed, Sep 22, 2021, at 8:52 AM, Christian Brauner wrote:
> On Wed, Sep 22, 2021 at 08:34:23AM -0700, Andy Lutomirski wrote:
>> On Wed, Sep 22, 2021, at 5:25 AM, Christian Brauner wrote:
>> > On Mon, Sep 20, 2021 at 11:36:47AM -0700, Andy Lutomirski wrote:
>> >> On Mon, Sep 20, 2021 at 11:16 AM Luis Chamberlain <mcgrof@xxxxxxxxxx> wrote:
>> >> >
>> >> > On Mon, Sep 20, 2021 at 04:51:19PM +0200, Thomas Weißschuh wrote:
>> >> 
>> >> > > > Do you mean it literally invokes /sbin/modprobe?  If so, hooking this
>> >> > > > at /sbin/modprobe and calling out to the container manager seems like
>> >> > > > a decent solution.
>> >> > >
>> >> > > Yes it does. Thanks for the idea, I'll see how this works out.
>> >> >
>> >> > Would documentation guiding you in that way have helped? If so
>> >> > I welcome a patch that does just that.
>> >> 
>> >> If someone wants to make this classy, we should probably have the
>> >> container counterpart of a standardized paravirt interface.  There
>> >> should be a way for a container to, in a runtime-agnostic way, issue
>> >> requests to its manager, and requesting a module by (name, Linux
>> >> kernel version for which that name makes sense) seems like an
>> >> excellent use of such an interface.
>> >
>> > I always thought of this in two ways we currently do this:
>> >
>> > 1. Caller transparent container manager requests.
>> >    This is the seccomp notifier where we transparently handle syscalls
>> >    including intercepting init_module() where we parse out the module to
>> >    be loaded from the syscall args of the container and if it is
>> >    allow-listed load it for the container otherwise continue the syscall
>> >    letting it fail or failing directly through seccomp return value.
>> 
>> Specific problems here include aliases and dependencies.  My modules.alias file, for example, has:
>> 
>> alias net-pf-16-proto-16-family-wireguard wireguard
>> 
>> If I do modprobe net-pf-16-proto-16-family-wireguard, modprobe parses some files in /lib/modules/`uname -r` and issues init_module() asking for 'wireguard'.  So hooking init_module() is at the wrong layer -- for that to work, the container's /sbin/modprobe needs to already have figured out that the desired module is wireguard and have a .ko for it.
>
> You can't use the container's .ko module. For this you would need to
> trust the image that the container wants you to load. The container
> manager should always load a host module.
>

Agreed. 

>> 
>> >
>> > 2. A process in the container explicitly calling out to the container
>> >    manager.
>> >    One example how this happens is systemd-nspawn via dbus messages
>> >    between systemd in the container and systemd outside the container to
>> >    e.g. allocate a new terminal in the container (kinda insecure but
>> >    that's another issue) or other stuff.
>> >
>> > So what was your idea: would it be like a device file that could be
>> > exposed to the container where it writes requestes to the container
>> > manager? What would be the advantage to just standardizing a socket
>> > protocol which is what we do for example (it doesn't do module loading
>> > of course as we handle that differently):
>> 
>> My idea is standardizing *something*.  I think it would be nice if, for example, distros could ship a /sbin/modprobe that would do the right thing inside any compliant container runtime as well as when running outside a container.
>> 
>> I suppose container managers could also bind-mount over /sbin/modprobe, but that's more intrusive.
>
> I don't see this is a big issue because that is fairly trivial.
> I think we never want to trust the container's modules.
> What probably should be happening is that the manager exposes a list of
> modules the container can request in some form. We have precedence for
> doing something like this.
> So now modprobe and similar tools can be made aware that if they are in
> a container they should request that module from the container manager
> be it via a socket request or something else.
> Nesting will be a bit funny but can probably be made to work by just
> bind-mounting the outermost socket into the container or relaying the
> request.

Why bother with a list?  I think it should be sufficient for the container to ask for a module and either get it or not get it.