[lxc-devel] [RFC PATCH 00/11] Add support for devtmpfs in user namespaces

[Seth Forshee]

On Fri, May 16, 2014 at 12:28:35PM -0700, James Bottomley wrote:
> On Fri, 2014-05-16 at 11:57 -0700, Greg Kroah-Hartman wrote:
> > On Fri, May 16, 2014 at 09:06:07AM -0500, Seth Forshee wrote:
> > > On Thu, May 15, 2014 at 09:35:32PM -0700, Greg Kroah-Hartman wrote:
> > > > On Fri, May 16, 2014 at 01:49:59AM +0000, Serge Hallyn wrote:
> > > > > > I think having to pick and choose what device nodes you want in a
> > > > > > container is a good thing.  Becides, you would have to do the same thing
> > > > > > in the kernel anyway, what's wrong with userspace making the decision
> > > > > > here, especially as it knows exactly what it wants to do much more so
> > > > > > than the kernel ever can.
> > > > > 
> > > > > For 'real' devices that sounds sensible.  The thing about loop devices
> > > > > is that we simply want to allow a container to say "give me a loop
> > > > > device to use" and have it receive a unique loop device (or 3), without
> > > > > having to pre-assign them.  I think that would be cleaner to do using
> > > > > a pseudofs and loop-control device, rather than having to have a
> > > > > daemon in userspace on the host farming those out in response to
> > > > > some, I don't know, dbus request?
> > > > 
> > > > I agree that loop devices would be nice to have in a container, and that
> > > > the existing loop interface doesn't really lend itself to that.  So
> > > > create a new type of thing that acts like a loop device in a container.
> > > > But don't try to mess with the whole driver core just for a single type
> > > > of device.
> > > 
> > > No matter what I don't think we get out of this without driver core
> > > changes, whether this was done in loop or by creating something new.
> > > Not unless the whole thing is punted to userspace, anyway.
> > > 
> > > The first problem is that many block device ioctls check for
> > > CAP_SYS_ADMIN. Most of these might not ever be used on loop devices, I'm
> > > not really sure. But loop does at minimum support partitions, and to get
> > > that functionality in an unprivileged container at least the block layer
> > > needs to know the namespace which has privileges for that device.
> > 
> > That's fine, you should have those permissions in a container if you
> > want to do something like that on a loop device, right?
> 
> Really, no.  CAP_SYS_ADMIN is effectively a pseudo root security hole.
> Any user possessing CAP_SYS_ADMIN can do about as much damage as real
> root can, whether or not you use user namespaces, so it would compromise
> a lot of the security we're just bringing to containers.
> 
> > > The second is that all block devices automatically appear in devtmpfs.
> > > The scenario I'm concerned about is that the host could unknowingly use
> > > a loop device exposed to a container, then the container could see data
> > > from the host.
> > 
> > I don't think that's a real issue, the host should know not to do that.
> > 
> > > So we either need a flag to tell the driver core not to create a node
> > > in devtmpfs, or we need a privileged manager in userspace to remove
> > > them (which kind of defeats the purpose). And it gets more complicated
> > > when partition block devs are mixed in, because they can be created
> > > without involvement from the driver - they would need to inherit the
> > > "no devtmpfs node" property from their parent, and if the driver uses
> > > a psuedo fs to create device nodes for userspace then it needs to be
> > > informed about the partitions too so it can create those nodes.
> > 
> > I don't think that will be needed.  Root in a host can do whatever it
> > wants in the containers, so mixing up block devices is the least of the
> > issues involved :)
> > 
> > > So maybe we could get by without the privileged ioctls, as long as it
> > > was understood that unprivileged containers can't do partitioning. But I
> > > do think the devtmpfs problem would need to be addressed.
> > 
> > I don't think unpriviliged containers should be able to do partitioning.
> > An unpriviliged user can't do that, so why should a container be any
> > different?
> 
> To make sure we're on the same page with terminology, there's an
> unprivileged container and a secure container.  In the former, there's
> no root user (all the processes run as non-root), so the container isn't
> expected to perform any actions root would ... that's easy.  In a secure
> container, root is mapped to a nobody user in the host, so is
> effectively unprivileged, but root in the container expects to look like
> a real root within the VPS (and thus may expect to partition things,
> depending on how they've been given access to the block device).  The
> big problem is giving back capabilities to the container root such that
> a) it loses them if it escapes the container and b) it doesn't get
> sufficient capabilities to damage the system.

Based on your description what I was talking about is a secure
container. Thanks for clearing that up, and sorry for misusing the
terminology.

What I set out for was feature parity between loop devices in a secure
container and loop devices on the host. Since some operations currently
check for system-wide CAP_SYS_ADMIN, the only way I see to accomplish
this is to push knowledge of the user namespace farther down into the
driver stack so the check can instead be for CAP_SYS_ADMIN in the user
namespace associated with the device.

That said, I suspect our current use cases can get by without these
capabilities. Really though I suspect this is just deferring the
discussion rather than settling it, and what we'll end up with is little
more than a fancy way for userspace to ask the kernel to run mknod on
its behalf.

Thanks,
Seth