Re: [PATCH] binderfs: implement sysctls

[Christian Brauner]

On Fri, Dec 21, 2018 at 05:33:16PM +0100, Greg KH wrote:
> On Fri, Dec 21, 2018 at 04:59:19PM +0100, Christian Brauner wrote:
> > On Fri, Dec 21, 2018 at 04:37:58PM +0100, Greg KH wrote:
> > > On Fri, Dec 21, 2018 at 03:12:42PM +0100, Christian Brauner wrote:
> > > > On Fri, Dec 21, 2018 at 02:55:09PM +0100, Greg KH wrote:
> > > > > On Fri, Dec 21, 2018 at 02:39:09PM +0100, Christian Brauner wrote:
> > > > > > This implements three sysctls that have very specific goals:
> > > > > 
> > > > > Ick, why?
> > > > > 
> > > > > What are these going to be used for?  Who will "control" them?  As you
> > > > 
> > > > Only global root in the initial user namespace. See the reasons below. :)
> > > > 
> > > > > are putting them in the "global" namespace, that feels like something
> > > > > that binderfs was trying to avoid in the first place.
> > > > 
> > > > There are a couple of reason imho:
> > > > - Global root needs a way to restrict how many binder devices can be
> > > >   allocated across all user + ipc namespace pairs.
> > > >   One obvious reason is that otherwise userns root in a non-initial user
> > > >   namespace can allocate a huge number of binder devices (pick a random
> > > >   number say 10.000) and use up a lot of kernel memory.
> > > 
> > > Root can do tons of other bad things too, why are you picking on
> > 
> > That's global root not userns root though. :) These sysctls are about
> > global root gaining the ability to proactively restrict binder device
> > delegation.
> > 
> > > binderfs here?  :)
> > > 
> > > >   In addition they can pound on the binder.c code causing a lot of
> > > >   contention for the remaining global lock in there.
> > > 
> > > That's the problem of that container, don't let it do that.  Or remove
> > > the global lock :)
> > > 
> > > >   We should let global root explicitly restrict non-initial namespaces
> > > >   in this respect. Imho, that's just good security design. :)
> > > 
> > > If you do not trust your container enough to have it properly allocate
> > > the correct binder resources, then perhaps you shouldn't be allowing it
> > > to allocate any resources at all?
> > 
> > Containers just like VMs get delegated and you might not have control
> > over what is running in there. That's AWS in a nutshell. :) Restricting
> > it by saying "just don't do that" seems not something that is
> > appropriate given the workloads out there in the wild.
> > In general, I do *understand* the reasoning but I think the premise is
> > flawed if we can somewhat trivially make this safe.
> > 
> > > 
> > > > - The reason for having a number of reserved devices is when the initial
> > > >   binderfs mount needs to bump the number of binder devices after the
> > > >   initial allocation done during say boot (e.g. it could've removed
> > > >   devices and wants to reallocate new ones but all binder minor numbers
> > > >   have been given out or just needs additional devices). By reserving an
> > > >   initial pool of binder devices this can be easily accounted for and
> > > >   future proofs userspace. This is to say: global root in the initial
> > > >   userns + ipcns gets dibs on however many devices it wants. :)
> > > 
> > > binder devices do not "come and go" at runtime, you need to set them up
> > > initially and then all is fine.  So there should never be a need for the
> > > "global" instance to need "more" binder devices once it is up and
> > > running.  So I don't see what you are really trying to solve here.
> > 
> > That's dismissing a whole range of use-cases where you might allocate
> > and deallocate devices on the fly which this is somewhat designed for.
> > But I guess ok for now.
> > 
> > > 
> > > You seem to be trying to protect the system from the container you just
> > > gave root to and trusted it with creating its own binder instances.
> > > If you do not trust it to create binder instances then do not allow it
> > > to create binder instances!  :)
> > 
> > Again, I get the reasoning but think that this dismisses major
> > real-world use-cases not just for binderfs but for all instances where
> > untrusted workloads are run which both containers and VMs aim to make
> > sure are possible.
> > Note, I mean untrusted not in the sense of necessarily being malicious
> > but just "can't guarantee that things don't blow up in your face".
> > 
> > > 
> > > > - The fact that we have a single shared pool of binder device minor
> > > >   numbers for all namespaces imho makes it necessary for the global root
> > > >   user in the initial ipc + user namespace to manage device allocation
> > > >   and delegation.
> > > 
> > > You are managing the allocation, you are giving who ever asks for one a
> > > device.  If you run out of devices, oops, you run out of devices, that's
> > > it.  Are you really ever going to run out of a major's number of binder
> > > devices?
> > 
> > The point is more about someone intentionally trying to do that.
> > 
> > > 
> > > > The binderfs sysctl stuff is really small code-wise and adds a lot of
> > > > security without any performance impact on the code itself. So we
> > > > actually very strictly adhere to the requirement to not blindly
> > > > sacrifice performance for security. :)
> > > 
> > > But you are adding a brand new user/kernel api by emulating one that is
> > > very old and not the best at all, to try to protect from something that
> > > seems like you can't really "protect" from in the first place.
> > 
> > Of course we can protect from that. It's about init root never running
> > out of devices. We don't care about non-init-userns running out of
> > devices at all.
> > 
> > > 
> > > You now have a mis-match of sysctls, ioctls and file operations all
> > > working on the same logical thing.  And all interacting in different and
> > > uncertian ways.  Are you sure that's wise?
> > 
> > The sysctl approach is present in other pseudo-filesystems apart from
> > devpts. For example, mqueue.
> > 
> > > 
> > > If the binderfs code as-is isn't "safe enough" to use without this, then
> > > we need to revisit it before someone starts to use it...
> > 
> > *It is safe.* I just don't see a good argument against additional
> > hardening. *But I'm definitely not going to push this patch if it's
> > additional hardening features are used to push the unsound argument that
> > binderfs isn't safe.* :)
> 
> Ok, so what you really want is just "limits" to prevent a container from
> doing something really crazy, right?  So, how about a limit of 10 binder
> nodes per container?  Make it a kernel build option so it can be changed
> by a vendor if they really find that is a problem.
> 
> Would that solve the issue you are thinking might be here?

Actually you brought me to reason through this with a couple of people
and one thing I totally forgot and which was pointed out to me is that
we can limit the container's memory consumption via the memory cgroup.
So consuming all kernel memory wouldn't be possible with this. And if
you don't use the memory cgroup you're hosed in a variety of other ways
anyway. :) So that attack is gone which means we can do away with a
global limit. Which I very much prefer! Thanks for making me look harder
into the attack surface!

What still makes sense is a per-instance limit, i.e. a max=<count> mount
option. So the initial root user can mount binderfs:

mkdir /dev/binderfs
mount -t binder binder -o max=1024 /dev/binderfs

grant ACLs or whatever and bind-mount the whole binderfs host instance
into another user + ipc namespace. This way there's a hard-cap at "max"
since the "max" mount option cannot be changed by root in a non-initial
user namespace. That secures the use-case where users don't want to have
multiple binderfs mounts but rather use a single one which some
requested afair.

Thanks!
Christian