[RFC 1/1] Use ioctl selinux callback io_uring commands that implement the ioctl op convention

[Ming Lei]

On Mon, Nov 28, 2022 at 11:13:29AM +0100, Joel Granados wrote:
> On Tue, Nov 22, 2022 at 10:04:03PM +0800, Ming Lei wrote:
> > On Mon, Nov 21, 2022 at 04:05:37PM -0500, Paul Moore wrote:
> > > On Mon, Nov 21, 2022 at 2:53 PM Luis Chamberlain <mcgrof at kernel.org> wrote:
> > > > On Thu, Nov 17, 2022 at 05:10:07PM -0500, Paul Moore wrote:
> > > > > On Thu, Nov 17, 2022 at 4:40 AM Joel Granados <j.granados at samsung.com> wrote:
> > > > > > On Wed, Nov 16, 2022 at 02:21:14PM -0500, Paul Moore wrote:
> > > > >
> > > > > ...
> > > > >
> > > > > > > * As we discussed previously, the real problem is the fact that we are
> > > > > > > missing the necessary context in the LSM hook to separate the
> > > > > > > different types of command targets.  With traditional ioctls we can
> > > > > > > look at the ioctl number and determine both the type of
> > > > > > > device/subsystem/etc. as well as the operation being requested; there
> > > > > > > is no such information available with the io_uring command
> > > > > > > passthrough.  In this sense, the io_uring command passthrough is
> > > > > > > actually worse than traditional ioctls from an access control
> > > > > > > perspective.  Until we have an easy(ish)[1] way to determine the
> > > > > > > io_uring command target type, changes like the one suggested here are
> > > > > > > going to be doomed as each target type is free to define their own
> > > > > > > io_uring commands.
> > > > > >
> > > > > > The only thing that comes immediately to mind is that we can have
> > > > > > io_uring users define a function that is then passed to the LSM
> > > > > > infrastructure. This function will have all the logic to give relative
> > > > > > context to LSM. It would be general enough to fit all the possible commands
> > > > > > and the logic would be implemented in the "drivers" side so there is no
> > > > > > need for LSM folks to know all io_uring users.
> > > > >
> > > > > Passing a function pointer to the LSM to fetch, what will likely be
> > > > > just a constant value, seems kinda ugly, but I guess we only have ugly
> > > > > options at this point.
> > > >
> > > > I am not sure if this helps yet, but queued on modules-next we now have
> > > > an improvement in speed of about 1500x for kallsyms_lookup_name(), and
> > > > so symbol lookups are now fast. Makes me wonder if a type of special
> > > > export could be drawn up for specific calls which follow a structure
> > > > and so the respective lsm could be inferred by a prefix instead of
> > > > placing the calls in-place. Then it would not mattter where a call is
> > > > used, so long as it would follow a specific pattern / structure with
> > > > all the crap you need on it.
> > > 
> > > I suspect we may be talking about different things here, I don't think
> > > the issue is which LSM(s) may be enabled, as the call is to
> > > security_uring_cmd() regardless.  I believe the issue is more of how
> > > do the LSMs determine the target of the io_uring command, e.g. nvme or
> > > ublk.
> > > 
> > > My understanding is that Joel was suggesting a change to the LSM hook
> > > to add a function specific pointer (presumably defined as part of the
> > > file_operations struct) that could be called by the LSM to determine
> > > the target.
> > > 
> > > Although now that I'm looking again at the file_operations struct, I
> > > wonder if we would be better off having the LSMs inspect the
> > > file_operations::owner field, potentially checking the module::name
> > > field.  It's a little painful in the sense that it is potentially
> > > multiple strcmp() calls for each security_uring_cmd() call, but I'm
> > > not sure the passed function approach would be much better.  Do we
> > > have a consistent per-module scalar value we can use instead of a
> > > character string?
> > 
> > In future there may be more uring_cmd kernel users, maybe we need to
> > consider to use one reserved field in io_uring_sqe for describing the
> > target type if it is one must for security subsystem, and this way
> > will be similar with traditional ioctl which encodes this kind of
> > info in command type.
> This is of course another option. I was a bit reluctant to start the
> discussion with this implementation, but now that you have brought it up
> I can come up with an initial RFC and we can add it to the mix of
> possibilities.
> 
> Would you just add it to the end of the struct? or what reserved field
> are you referring to?

io_uring_sqe is uapi, so you can't add any field to sqe, and '__pad1'
could be best field for carrying this info, given it is close to 'cmd_op',
and 'u8' should be enough for storing target type info.


thanks,
Ming