On 6/4/2021 5:08 PM, Alexei Starovoitov wrote:

On Fri, Jun 4, 2021 at 4:34 PM Paul Moore [off-list ref] wrote:

quoted

quoted

Again, the problem is not limited to BPF at all. kprobes is doing register-
the hooks which are equivalent to the one of BPF. Anything in run-time
trying to prevent probe_read_kernel by kprobes or BPF is broken by design.

Not being an expert on kprobes I can't really comment on that, but
right now I'm focused on trying to make things work for the BPF
helpers.  I suspect that if we can get the SELinux lockdown
implementation working properly for BPF the solution for kprobes won't
be far off.

Paul,

Both kprobe and bpf can call probe_read_kernel==copy_from_kernel_nofault
from all contexts.
Including NMI. Most of audit_log_* is not acceptable.
Just removing a wakeup is not solving anything.
Audit hooks don't belong in NMI.
Audit design needs memory allocation. Hence it's not suitable
for NMI and hardirq. But kprobes and bpf progs do run just fine there.
BPF, for example, only uses pre-allocated memory.

You have fallen into a common fallacy. The fact that the "code runs"
does not assure that the "system works right". In the security world
we face this all the time, often with performance expectations. In this
case the BPF design has failed to accommodate the long standing needs
of audit and SELinux. Shifting the responsibility for these design flaws
to SELinux is inappropriate. Integration of sub-systems is usually the
burden of the newcomer, which in this case is BPF. Paul is doing the
bulk of your work for you. Maybe you could step up to your responsibility
and work with him, not against him.