lkml.org 
[lkml]   [2018]   [Dec]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: [PATCH] binderfs: implement sysctls
On Fri, Dec 21, 2018 at 8:33 AM Greg KH <gregkh@linuxfoundation.org> 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?

We already have CONFIG_ANDROID_BINDER_DEVICES which specifies 3
devices (binder,hwbinder,vndbinder). How about limiting each container
to no more than the number of devices specified there?

>
> thanks,
>
> greg k-h

\
 
 \ /
  Last update: 2018-12-21 17:40    [W:0.044 / U:0.236 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site