Re: Debugging eBPF was: Re: [PATCH bpf-next 0/8] tools: bpftool: add probes for system and device

From: Arnaldo Carvalho de Melo <acme@kernel.org>
To: Alexei Starovoitov <alexei.starovoitov@gmail.com>
Cc: Quentin Monnet <quentin.monnet@netronome.com>,
	Alexei Starovoitov <ast@kernel.org>,
	Daniel Borkmann <daniel@iogearbox.net>,
	netdev@vger.kernel.org, oss-drivers@netronome.com,
	Jesper Dangaard Brouer <brouer@redhat.com>,
	Stanislav Fomichev <sdf@google.com>
Subject: Re: Debugging eBPF was: Re: [PATCH bpf-next 0/8] tools: bpftool: add probes for system and device
Date: Fri, 14 Dec 2018 10:39:27 -0300	[thread overview]
Message-ID: <20181214133927.GP21027@kernel.org> (raw)
In-Reply-To: <20181213205508.u6izf4ugnpincyps@ast-mbp.dhcp.thefacebook.com>

Em Thu, Dec 13, 2018 at 12:55:10PM -0800, Alexei Starovoitov escreveu:
> On Thu, Dec 13, 2018 at 10:49:09AM -0300, Arnaldo Carvalho de Melo wrote:
> > Only root can attach eBPF programs to tracepoints.

> > Would be really good if we could have a more restricted program type to
> > attach to tracepoints, one that would be able to run only in the context
> > of their threads and access only the pointers in the tracepoints, that
> > way the 'perf trace' augmented syscalls code would be usable for
> > non-root users just like the other 'perf' commands are, allowing us to,
> > as with root, to copy the pointer arguments, like:

> I don't think there is a clean way of doing non-root with tracepoints or syscalls.
> The kernel side would need to start filtering the progs.
> Like current uid == uid of loaded prog. But then there are tail_calls.

Yeah, that program would run only for threads owned by the prog owner.

> they would need to be disabled. 

I think if that is not possible, then would be an acceptable limitation
in a first implementation. I.e. my understanding is that eBPF started
with some limited scope, then as it goes maturing, more features were
added as its security/performance implications were understood.

> tracepoints args can be pointers. _all_ of them in the kernel would need to
> be annotated to make sure pointers don't leak into unpriv user space.
> and so on and so forth.

Yes, I thought about heavily restricting them, i.e. points would be
allowed just for some very special cases, like the arguments to
raw_syscalls.sys_{sys_enter,sys_exit}, as a starting point.

> I think better way forward would be to introduce something in the middle.
> Between root and unpriv. Something that tracing bpf progs can use.
> May be new capability?

Well, that would be interesting too, I think, would make go a bit
forward, for a class of applications where trusting the tracer is
possible.

- Arnaldo