On Wed, May 13, 2020 at 11:50:42AM +0100, Marek Majkowski wrote:

On Wed, May 13, 2020 at 4:19 AM Alexei Starovoitov
[off-list ref] wrote:

quoted

CAP_BPF solves three main goals:
1. provides isolation to user space processes that drop CAP_SYS_ADMIN and switch to CAP_BPF.
   More on this below. This is the major difference vs v4 set back from Sep 2019.
2. makes networking BPF progs more secure, since CAP_BPF + CAP_NET_ADMIN
   prevents pointer leaks and arbitrary kernel memory access.
3. enables fuzzers to exercise all of the verifier logic. Eventually finding bugs
   and making BPF infra more secure. Currently fuzzers run in unpriv.
   They will be able to run with CAP_BPF.

Alexei, looking at this from a user point of view, this looks fine.

I'm slightly worried about REUSEPORT_EBPF. Currently without your
patch, as far as I understand it:

- You can load SOCKET_FILTER and SO_ATTACH_REUSEPORT_EBPF without any
permissions

correct.

- For loading BPF_PROG_TYPE_SK_REUSEPORT program and for SOCKARRAY map
creation CAP_SYS_ADMIN is needed. But again, no permissions check for
SO_ATTACH_REUSEPORT_EBPF later.

correct. With clarification that attaching process needs to own
FD of prog and FD of socket.

If I read the patchset correctly, the former SOCKET_FILTER case
remains as it is and is not affected in any way by presence or absence
of CAP_BPF.

correct. As commit log says:
"Existing unprivileged BPF operations are not affected."

The latter case is different. Presence of CAP_BPF is sufficient for
map creation, but not sufficient for loading SK_REUSEPORT program. It
still requires CAP_SYS_ADMIN. 

Not quite.
The patch will allow BPF_PROG_TYPE_SK_REUSEPORT progs to be loaded
with CAP_BPF + CAP_NET_ADMIN.
Since this type of progs is clearly networking type I figured it's
better to be consistent with the rest of networking types.
Two unpriv types SOCKET_FILTER and CGROUP_SKB is the only exception.

I think it's a good opportunity to relax
this CAP_SYS_ADMIN requirement. I think the presence of CAP_BPF should
be sufficient for loading BPF_PROG_TYPE_SK_REUSEPORT.

Our specific use case is simple - we want an application program -
like nginx - to control REUSEPORT programs. We will grant it CAP_BPF,
but we don't want to grant it CAP_SYS_ADMIN.

You'll be able to grant nginx CAP_BPF + CAP_NET_ADMIN to load SK_REUSEPORT
and unpriv child process will be able to attach just like before if
it has right FDs.
I suspect your load balancer needs CAP_NET_ADMIN already anyway due to
use of XDP and TC progs.
So granting CAP_BPF + CAP_NET_ADMIN should cover all bpf prog needs.
Does it address your concern?