Re: [PATCH net-next 1/2] bpf: enable non-root eBPF programs

[Alexei Starovoitov <ast@plumgrid.com>]

On 10/7/15 3:22 PM, Kees Cook wrote:
>> Yes, I agree with you that there would be a CVE regardless. I still
>> >like the option of configurable access, not a big fan of the sysctl
>> >either. Thinking out loudly, what about a Kconfig option? We started
>> >out like this on bpf(2) itself (initially under expert settings, now
>> >afaik not anymore), and depending on usage scenarios, a requirement
>> >could be to have immutable cap_sys_admin-only, for other use-cases a
>> >requirement on the kernel might instead be to have unprivileged users
>> >as well.
> It'd be nice to have it just be a Kconfig, but this shoots
> distro-users in the foot if a distro decides to include unpriv bpf and
> the user doesn't want it. I think it's probably a good idea to keep
> the sysctl.

I don't like introducing Kconfig for no clear reason. It only adds
to the testing matrix and makes it harder to hack around.
Paranoid distros can disable bpf via single config already,
there is no reason to go more fine grained here.
Unpriv checks add minimal amount of code, so even for tinification
purpose there is no need to chop of few bytes. tiny kernels would
disable bpf all together.

As far as sysctl we can look at two with similar purpose:
sysctl_perf_event_paranoid and modules_disabled.
First one is indeed multi level, but not because of the fear of bugs,
but because of real security implications. Like raw events on
hyperthreaded cpu or uncore events can extract data from other
user processes. So it controls these extra privileges.
For bpf there are no hw implications to deal with.
If we make seccomp+bpf in the future it shouldn't need another knob
or extra bit. There are no extra privileges to grant, so not needed.

modules_disabled is off by default and can be toggled on once.
I think for paranoid distro users that "don't want bpf" that is
the better model.
So I'm thinking to do sysctl_unprivileged_bpf_disabled that will be
0=off by default (meaning that users can load unpriv socket filter
programs and seccomp in the future) and that can be switched
to 1=on once and stay that way until reboot.
I think that's the best balance that avoids adding checks to all
apps that want to use bpf and admins can still act on it.
 From app point of view it's no different than bpf syscall
was not compiled in. So single feature test for bpf syscall will
be enough.