On Wed, Nov 29, 2017 at 07:35:31PM -0500, Theodore Ts'o wrote:
On Wed, Nov 29, 2017 at 11:28:52AM -0600, Serge E. Hallyn wrote:
quoted
Just to be clear, module loading requires - and must always continue to
require - CAP_SYS_MODULE against the initial user namespace. Containers
in user namespaces do not have that.
I don't believe anyone has ever claimed that containers which are not in
a user namespace are in any way secure.
Unless the container performs some action which causes the kernel to
call request_module(), which then loads some kernel module,
A local unprivileged user can do the same thing. I reject the popular
notion that linux is a single user operating system. More interesting
are the (very real) cases where root in a container can do something
which a local unprivileged user could not do. Since a local unprivileged
user can always create a new namespace, *those* constitute a real and
interesting problem.
potentially containing cr*p unmaintained code which was included when
the distro compiled the world, into the host kernel.
This is an attack vector that doesn't exist if you are using VM's.
Until the vm tenant uses a trivial vm escape.
And in general, the attack surface of the entire Linux
kernel<->userspace API is far larger than that which is exposed by the
guest<->host interface.
For that reason, containers are *far* more insecure than VM's, since
once the attacker gets root on the guest VM, they then have to attack
the hypervisor interface. And if you compare the attack surface of
the two, it's pretty clear which is larger, and it's not the
hypervisor interface.
Any time anyone spends a day looking at either the black hole that is
the hardware emulators or the xen and kvm code itself they walk away
with a set of cve's. It *should* be more secure, it's not. You're
telling me your house is safe because you put up a no tresspassing
sign.
-serge