On Tue, May 12, 2020 at 10:27:33PM +0200, Daniel Borkmann wrote:
> On 5/8/20 11:53 PM, Alexei Starovoitov wrote:
> [...]
> > @@ -2880,8 +2933,6 @@ static int bpf_prog_test_run(const union bpf_attr *attr,
> > struct bpf_prog *prog;
> > int ret = -ENOTSUPP;
> > - if (!capable(CAP_SYS_ADMIN))
> > - return -EPERM;
>
> Should above be under bpf_capable() as well or is the intention to really let
> (fully) unpriv users run sk_filter test progs here? I would assume only progs
> that have prior been loaded under bpf_capable() should suffice, so no need to
> lower the bar for now, no?
Unpriv can load sk_filter and attach to a socket. Then send data through
the socket to trigger execution.
bpf_prog_test_run is doing the same prog execution without creating a socket.
What is the concern?
> > if (CHECK_ATTR(BPF_PROG_TEST_RUN))
> > return -EINVAL;
> > @@ -3163,7 +3214,7 @@ static int bpf_prog_get_info_by_fd(struct bpf_prog *prog,
> > info.run_time_ns = stats.nsecs;
> > info.run_cnt = stats.cnt;
> > - if (!capable(CAP_SYS_ADMIN)) {
> > + if (!bpf_capable()) {
>
> Given the JIT dump this also exposes addresses when bpf_dump_raw_ok() passes.
> I presume okay, but should probably be documented given CAP_SYS_ADMIN isn't
> required anymore?
Exactly. dump_raw_ok() is there. I'm not even sure why this cap_sys_admin
check is there. It looks like it can be completely removed, but I didn't
want to go that far in this set.
