[PATCH 00/18] bpf: Secure and authenticated preloading of eBPF programs

[Roberto Sassu]

> From: Casey Schaufler [mailto:casey at schaufler-ca.com]
> Sent: Tuesday, April 5, 2022 4:50 PM
> On 4/4/2022 10:20 AM, Roberto Sassu wrote:
> >> From: Djalal Harouni [mailto:tixxdz at gmail.com]
> >> Sent: Monday, April 4, 2022 9:45 AM
> >> On Sun, Apr 3, 2022 at 5:42 PM KP Singh <kpsingh at kernel.org> wrote:
> >>> On Sat, Apr 2, 2022 at 1:55 AM Alexei Starovoitov
> >>> <alexei.starovoitov at gmail.com> wrote:
> >> ...
> >>>>> Pinning
> >>>>> them to unreachable inodes intuitively looked the
> >>>>> way to go for achieving the stated goal.
> >>>> We can consider inodes in bpffs that are not unlinkable by root
> >>>> in the future, but certainly not for this use case.
> >>> Can this not be already done by adding a BPF_LSM program to the
> >>> inode_unlink LSM hook?
> >>>
> >> Also, beside of the inode_unlink... and out of curiosity: making
> sysfs/bpffs/
> >> readonly after pinning, then using bpf LSM hooks
> >> sb_mount|remount|unmount...
> >> family combining bpf() LSM hook... isn't this enough to:
> >> 1. Restrict who can pin to bpffs without using a full MAC
> >> 2. Restrict who can delete or unmount bpf filesystem
> >>
> >> ?
> > I'm thinking to implement something like this.
> >
> > First, I add a new program flag called
> > BPF_F_STOP_ONCONFIRM, which causes the ref count
> > of the link to increase twice at creation time. In this way,
> > user space cannot make the link disappear, unless a
> > confirmation is explicitly sent via the bpf() system call.
> >
> > Another advantage is that other LSMs can decide
> > whether or not they allow a program with this flag
> > (in the bpf security hook).
> >
> > This would work regardless of the method used to
> > load the eBPF program (user space or kernel space).
> >
> > Second, I extend the bpf() system call with a new
> > subcommand, BPF_LINK_CONFIRM_STOP, which
> > decreasres the ref count for the link of the programs
> > with the BPF_F_STOP_ONCONFIRM flag. I will also
> > introduce a new security hook (something like
> > security_link_confirm_stop), so that an LSM has the
> > opportunity to deny the stop (the bpf security hook
> > would not be sufficient to determine exactly for
> > which link the confirmation is given, an LSM should
> > be able to deny the stop for its own programs).
> 
> Would you please stop referring to a set of eBPF programs
> loaded into the BPF LSM as an LSM? Call it a BPF security
> module (BSM) if you must use an abbreviation. An LSM is a
> provider of security_ hooks. In your case that is BPF. When
> you call the set of eBPF programs an LSM it is like calling
> an SELinux policy an LSM.

An eBPF program could be a provider of security_ hooks
too. The bpf LSM is an aggregator, similarly to your
infrastructure to manage built-in LSMs. Maybe, calling
it second-level LSM or secondary LSM would better
represent this new class.

The only differences are the registration method, (SEC
directive instead of DEFINE_LSM), and what the hook
implementation can access.

The implementation of a security_ hook via eBPF can
follow the same structure of built-in LSMs, i.e. it can be
uniquely responsible for enforcing and be policy-agnostic,
and can retrieve the decisions based on a policy from a
component implemented somewhere else.

Hopefully, I understood the basic principles correctly.
I let the eBPF maintainers comment on this.

Thanks

Roberto

HUAWEI TECHNOLOGIES Duesseldorf GmbH, HRB 56063
Managing Director: Li Peng, Zhong Ronghua

> > What do you think?
> >
> > Thanks
> >
> > Roberto
> >
> > HUAWEI TECHNOLOGIES Duesseldorf GmbH, HRB 56063
> > Managing Director: Li Peng, Zhong Ronghua