On Tue, May 10, 2022 at 12:18 PM Martin KaFai Lau <kafai@xxxxxx> wrote:
>
> On Tue, May 10, 2022 at 10:30:57AM -0700, Stanislav Fomichev wrote:
> > On Tue, May 10, 2022 at 12:13 AM Martin KaFai Lau <kafai@xxxxxx> wrote:
> > >
> > > On Mon, May 09, 2022 at 04:38:36PM -0700, Stanislav Fomichev wrote:
> > > > > > +unsigned int __cgroup_bpf_run_lsm_current(const void *ctx,
> > > > > > + const struct bpf_insn *insn)
> > > > > > +{
> > > > > > + const struct bpf_prog *shim_prog;
> > > > > > + struct cgroup *cgrp;
> > > > > > + int ret = 0;
> > > > > From lsm_hook_defs.h, there are some default return values that are not 0.
> > > > > Is it ok to always return 0 in cases like the cgroup array is empty ?
> > > >
> > > > That's a good point, I haven't thought about it. You're right, it
> > > > seems like attaching to this hook for some LSMs will change the
> > > > default from some error to zero.
> > > > Let's start by prohibiting those hooks for now? I guess in theory,
> > > > when we generate a trampoline, we can put this default value as an
> > > > input arg to these new __cgroup_bpf_run_lsm_xxx helpers (in the
> > > > future)?
> > > After looking at arch_prepare_bpf_trampoline, return 0 here should be fine.
> > > If I read it correctly, when the shim_prog returns 0, the trampoline
> > > will call the original kernel function which is the bpf_lsm_##NAME()
> > > defined in bpf_lsm.c and it will then return the zero/-ve DEFAULT.
> >
> > Not sure I read the same :-/ I'm assuming that for those cases we
> > actually end up generating fmod_ret trampoline which seems to be
> > unconditionally saving r0 into fp-8 ?
> invoke_bpf_mod_ret() calls invoke_bpf_prog(..., true) that saves the r0.
>
> Later, the "if (flags & BPF_TRAMP_F_CALL_ORIG)" will still
> "/* call the original function */" and then stores the r0 retval
> from the original function, no? or I mis-read something ?
I was under the wrong assumption this whole time that fmod_ret
programs run after the original one and the first bpf program sees the
output of the original one.
Turns out it's not the case; agreed that we already do the right
thing; thanks for pointing it out!
> > > > Another thing that seems to be related: there are a bunch of hooks
> > > > that return void, so returning EPERM from the cgroup programs won't
> > > > work as expected.
> > > > I can probably record, at verification time, whether lsm_cgroup
> > > > programs return any "non-success" return codes and prohibit attaching
> > > > these progs to the void hooks?
> > > hmm...yeah, BPF_LSM_CGROUP can be enforced to return either 0 or 1 as
> > > most other cgroup-progs do.
> > >
> > > Do you have a use case that needs to return something other than -EPERM ?
> >
> > We do already enforce 0/1 for cgroup progs (and we have helpers to
> > expose custom errno). What I want to avoid is letting users attach
> > programs that try to return the error for the void hooks. And it seems
> > like we record that return range for a particular cgroup program and
> > verify it at attach time, WDYT?
> Make sense. Do that in check_return_code() at load time instead of
> attach time?
> To be specific, meaning enforce BPF_LSM_CGROUP to 0/1 for int return type
> and always 1 for void return type?
Yeah, let's try to enforce the following at load time:
- return 0 or call to bpf_set_retval should happen only for the hooks
that return int
- for the void ones, only 'return 1' should be accepted
> Ah, I forgot there is a bpf_set_retval(). I assume we eventually want
> to allow that for BPF_LSM_CGROUP later? Once it is allowed,
> the verifier should also reject bpf_set_retval() when the
> attach_btf_id has a void return type?
Right, let me actually try to add bpf_set_retval to the set of allowed
helpers from the start, shouldn't be hard..
