On Wed, Jan 15, 2020 at 9:15 AM KP Singh [off-list ref] wrote:

From: KP Singh <redacted>

# Changes since v1 (https://lore.kernel.org/bpf/20191220154208.15895-1-kpsingh@chromium.org/ (local)):

* Eliminate the requirement to maintain LSM hooks separately in
  security/bpf/hooks.h Use BPF trampolines to dynamically allocate
  security hooks
* Drop the use of securityfs as bpftool provides the required
  introspection capabilities.  Update the tests to use the bpf_skeleton
  and global variables
* Use O_CLOEXEC anonymous fds to represent BPF attachment in line with
  the other BPF programs with the possibility to use bpf program pinning
  in the future to provide "permanent attachment".
* Drop the logic based on prog names for handling re-attachment.
* Drop bpf_lsm_event_output from this series and send it as a separate
  patch.

# Motivation

Google does analysis of rich runtime security data to detect and thwart
threats in real-time. Currently, this is done in custom kernel modules
but we would like to replace this with something that's upstream and
useful to others.

The current kernel infrastructure for providing telemetry (Audit, Perf
etc.) is disjoint from access enforcement (i.e. LSMs).  Augmenting the
information provided by audit requires kernel changes to audit, its
policy language and user-space components. Furthermore, building a MAC
policy based on the newly added telemetry data requires changes to
various LSMs and their respective policy languages.

This patchset proposes a new stackable and privileged LSM which allows
the LSM hooks to be implemented using eBPF. This facilitates a unified
and dynamic (not requiring re-compilation of the kernel) audit and MAC
policy.

# Why an LSM?

Linux Security Modules target security behaviours rather than the
kernel's API. For example, it's easy to miss out a newly added system
call for executing processes (eg. execve, execveat etc.) but the LSM
framework ensures that all process executions trigger the relevant hooks
irrespective of how the process was executed.

Allowing users to implement LSM hooks at runtime also benefits the LSM
eco-system by enabling a quick feedback loop from the security community
about the kind of behaviours that the LSM Framework should be targeting.

# How does it work?

The LSM introduces a new eBPF (https://docs.cilium.io/en/v1.6/bpf/)
program type BPF_PROG_TYPE_LSM which can only be attached to LSM hooks.
Attachment requires CAP_SYS_ADMIN for loading eBPF programs and
CAP_MAC_ADMIN for modifying MAC policies.

The eBPF programs are attached to a separate security_hook_heads
maintained by the BPF LSM for mutable hooks and executed after all the
statically defined hooks (i.e. the ones declared by SELinux, AppArmor,
Smack etc). This also ensures that statically defined LSM hooks retain
the behaviour of "being read-only after init", i.e. __lsm_ro_after_init.

Upon attachment, a security hook is dynamically allocated with
arch_bpf_prepare_trampoline which generates code to handle the
conversion from the signature of the hook to the BPF context and allows
the JIT'ed BPF program to be called as a C function with the same
arguments as the LSM hooks. If any of the attached eBPF programs returns
an error (like ENOPERM), the behaviour represented by the hook is
denied.

Audit logs can be written using a format chosen by the eBPF program to
the perf events buffer or to global eBPF variables or maps and can be
further processed in user-space.

# BTF Based Design

The current design uses BTF
(https://facebookmicrosites.github.io/bpf/blog/2018/11/14/btf-enhancement.html,
https://lwn.net/Articles/803258/) which allows verifiable read-only
structure accesses by field names rather than fixed offsets. This allows
accessing the hook parameters using a dynamically created context which
provides a certain degree of ABI stability:


// Only declare the structure and fields intended to be used
// in the program
struct vm_area_struct {
        unsigned long vm_start;
} __attribute__((preserve_access_index));

It would be nice to also mention that you don't even have to
"re-define" these structs if you use vmlinux.h generated with `bpftool
btf dump file <path-to-vm-linux-or-/sys/kernel/btf/vmlinux> format c`.
Its output will contain all types of the kernel, including internal
ones not exposed through any public headers. And it will also
automatically have __attribute__((preserve_access_index)) applied to
all structs/unions. It can be pre-generated and checked in somewhere
along the application or generated on the fly, if environment and use
case allows.

// Declare the eBPF program mprotect_audit which attaches to
// to the file_mprotect LSM hook and accepts three arguments.
SEC("lsm/file_mprotect")
int BPF_PROG(mprotect_audit, struct vm_area_struct *vma,
             unsigned long reqprot, unsigned long prot)
{
        unsigned long vm_start = vma->vm_start;

        return 0;
}

By relocating field offsets, BTF makes a large portion of kernel data
structures readily accessible across kernel versions without requiring a
large corpus of BPF helper functions and requiring recompilation with
every kernel version. The BTF type information is also used by the BPF
verifier to validate memory accesses within the BPF program and also
prevents arbitrary writes to the kernel memory.

The limitations of BTF compatibility are described in BPF Co-Re
(http://vger.kernel.org/bpfconf2019_talks/bpf-core.pdf, i.e. field
renames, #defines and changes to the signature of LSM hooks).

This design imposes that the MAC policy (eBPF programs) be updated when
the inspected kernel structures change outside of BTF compatibility
guarantees. In practice, this is only required when a structure field
used by a current policy is removed (or renamed) or when the used LSM
hooks change. We expect the maintenance cost of these changes to be
acceptable as compared to the previous design
(https://lore.kernel.org/bpf/20190910115527.5235-1-kpsingh@chromium.org/ (local)).

[...]