Re: [PATCH 0/3] BPF signature verification

[Blaise Boscaccy <bboscaccy@linux.microsoft.com>]

KP Singh <kpsingh@kernel.org> writes:

> On Fri, May 30, 2025 at 11:19 PM Blaise Boscaccy
> <bboscaccy@linux.microsoft.com> wrote:
>>
>> KP Singh <kpsingh@kernel.org> writes:
>>
>
> [...]
>
>> >
>>
>> And that isn't at odds with the kernel being able to do it nor is it
>> with what I posted.
>>
>> > If your build environment that signs the BPF program is compromised
>> > and can inject arbitrary code, then signing does not help.  Can you
>> > explain what a supply chain attack would look like here?
>> >
>>
>> Most people here can read C code. The number of people that can read
>> ebpf assembly metaprogramming code is much smaller. Compromising clang
>> is one thing, compromising libbpf is another. Your proposal increases
>> the attack surface with no observable benefit. If I was going to leave a
>> hard-to-find backdoor into ring0, gen.c would be a fun place to explore
>> doing it. Module and UEFI signature verification code doesn't live
>> inside of GCC or Clang as set of meta-instructions that get emitted, and
>> there are very good reasons for that.
>>
>> Further, since the signature verification code is unique for each and
>> every program it needs to be verified/proved/tested for each and every
>> program. Additionally, since all these checks are being forced outside
>> of the kernel proper, with the insistence of keeping the LSM layer in
>> the dark of the ultimate result, the only way to test that a program
>> will fail if the map is corrupted is to physically corrupt each and
>> every program and test that individually. That isn't "elegant" nor "user
>> friendly" in any way, shape or form.
>>
>> >> subsystem.  Additionally, it is impossible to verify the code
>> >> performing the signature verification, as it is uniquely regenerated
>> >
>> > The LSM needs to ensure that it allows trusted LOADER programs i.e.
>> > with signatures and potentially trusted signed user-space binaries
>> > with unsigned or delegated signing (this will be needed for Cilium and
>> > bpftrace that dynamically generate BPF programs), that's a more
>> > important aspect of the LSM policy from a BPF perspective.
>> >
>>
>> I would like to be able to sign my programs please and have the kernel
>> verify it was done correctly. Why are you insisting that I *don't* do
>> that?  I'm yet to see any technical objection to doing that. Do you have
>> one that you'd like to share at this point?
>
> The kernel allows a trusted loader that's signed with your private
> key, that runs in the kernel context to delegate the verification.
> This pattern of a trusted / delegated loader is going to be required
> for many of the BPF use-cases that are out there (Cilium, bpftrace)
> that dynamically generate eBPF programs.
>
> The technical objection is that:
>
> * It does not align with most BPF use-cases out there as most
> use-cases need a trusted loader.

No, it's definitely a use case. It's trivial to support both a trusted
loader and a signature over the hash chain of supplied assets.

> * Locks us into a UAPI, whereas a signed LOADER allows us to
> incrementally build signing for all use-cases without compromising the
> security properties.
>

Your proposal locks us into a UAPI as well. There is no way to make to
do this via UAPI without making a UAPI design choice.

> BPF's philosophy is that of flexibility and not locking the users into
> a rigid in-kernel implementation and UAPI.
>

Then why are you locking us into a rigid
only-signing-the-loader-is-allowed implementation?

> - KP
>
>>
>> > MAP_EXCLUSIVE is missing and is required which prevents maps from
>> > being accessed by other programs as explained in the proposal.
>> >
>> > Please hold off on further iterations, I am working on a series and
>> > will share these patches based on the design that was proposed.
>> >
>>
>> So the premise here seems to be that people should only be allowed to
>> sign trusted loaders, and that trusted loaders must additionally be
>> authored by you, correct?
>>
>> When can we expect to see your patchset posted?
>>
>> >>
>> >> for every program.
>> >>
>> >>
>> >>
>> >> 2. Timing of Signature Check
>> >>
>> >> This patchset moves the signature check to a point before
>> >> security_bpf_prog_load is invoked, due to an unresolved discussion
>> >> here:
>> >
>> > This is fine and what I had in mind, signature verification does not
>> > need to happen in the verifier and the existing hooks are good enough.
>> > I did not reply to Paul's comment since this is a fairly trivial
>> > detail and would be obvious in the implementation that the verifier is
>> > not the right place to check the signature anyways as the instruction
>> > buffer is only stable pre-verification.
>> >
>> >> https://lore.kernel.org/linux-security-module/CAHC9VhTj3=ZXgrYMNA+G64zsOyZO+78uDs1g=kh91=GR5KypYg@mail.gmail.com/
>> >> This change allows the LSM subsystem to be informed of the signature
>> >> verification result—if it occurred—and the method used, all without
>> >> introducing a new hook. It improves visibility and auditability,
>> >> reducing the “trust me, friend” aspect of the original design.
>> >
>> >
>> > On Wed, May 28, 2025 at 11:50 PM Blaise Boscaccy
>> > <bboscaccy@linux.microsoft.com> wrote:
>> >>
>> >> As suggested or mandated by KP Singh
>> >> https://lore.kernel.org/linux-security-module/CACYkzJ6VQUExfyt0=-FmXz46GHJh3d=FXh5j4KfexcEFbHV-vg@mail.gmail.com/,
>> >> this patchset proposes and implements an alternative hash-chain
>> >> algorithm for signature verification of BPF programs.
>> >>
>> >> This design diverges in two key ways:
>> >>
>> >> 1. Signature Strategy
>> >>
>> >> Two different signature strategies are
>> >> implemented. One verifies only the signature of the loader program in
>> >> the kernel, as described in the link above. The other verifies the
>> >> program’s maps in-kernel via a hash chain.  The original design
>> >> required loader programs to be “self-aborting” and embedded the
>> >> terminal hash verification logic as metaprogramming code generation
>> >> routines inside libbpf. While this patchset supports that scheme, it
>> >> is considered undesirable in certain environments due to the potential
>> >> for supply-chain attack vectors and the lack of visibility for the LSM
>> >> subsystem.  Additionally, it is impossible to verify the code
>> >> performing the signature verification, as it is uniquely regenerated
>> >> for every program.
>> >>
>> >> 2. Timing of Signature Check
>> >>
>> >> This patchset moves the signature check to a point before
>> >> security_bpf_prog_load is invoked, due to an unresolved discussion
>> >> here:
>> >> https://lore.kernel.org/linux-security-module/CAHC9VhTj3=ZXgrYMNA+G64zsOyZO+78uDs1g=kh91=GR5KypYg@mail.gmail.com/
>> >> This change allows the LSM subsystem to be informed of the signature
>> >> verification result—if it occurred—and the method used, all without
>> >> introducing a new hook. It improves visibility and auditability,
>> >> reducing the “trust me, friend” aspect of the original design.
>> >>
>> >>
>> >> Blaise Boscaccy (3):
>> >>   bpf: Add bpf_check_signature
>> >>   bpf: Support light-skeleton signatures in autogenerated code
>> >>   bpftool: Allow signing of light-skeleton programs
>> >>
>> >>  include/linux/bpf.h            |   2 +
>> >>  include/linux/verification.h   |   1 +
>> >>  include/uapi/linux/bpf.h       |   4 +
>> >>  kernel/bpf/arraymap.c          |  11 +-
>> >>  kernel/bpf/syscall.c           | 123 +++++++++++++++++++-
>> >>  tools/bpf/bpftool/Makefile     |   4 +-
>> >>  tools/bpf/bpftool/common.c     | 204 +++++++++++++++++++++++++++++++++
>> >>  tools/bpf/bpftool/gen.c        |  66 ++++++++++-
>> >>  tools/bpf/bpftool/main.c       |  24 +++-
>> >>  tools/bpf/bpftool/main.h       |  23 ++++
>> >>  tools/include/uapi/linux/bpf.h |   4 +
>> >>  tools/lib/bpf/libbpf.h         |   4 +
>> >>  tools/lib/bpf/skel_internal.h  |  28 ++++-
>> >>  13 files changed, 491 insertions(+), 7 deletions(-)
>> >>
>> >> --
>> >> 2.48.1
>> >>