On Wed, Sep 16, 2020 at 12:59 PM Andrii Nakryiko
<andrii.nakryiko@xxxxxxxxx> wrote:
>
> On Tue, Sep 15, 2020 at 5:50 PM Toke Høiland-Jørgensen <toke@xxxxxxxxxx> wrote:
> >
> > From: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
> >
> > Eelco reported we can't properly access arguments if the tracing
> > program is attached to extension program.
> >
> > Having following program:
> >
> > SEC("classifier/test_pkt_md_access")
> > int test_pkt_md_access(struct __sk_buff *skb)
> >
> > with its extension:
> >
> > SEC("freplace/test_pkt_md_access")
> > int test_pkt_md_access_new(struct __sk_buff *skb)
> >
> > and tracing that extension with:
> >
> > SEC("fentry/test_pkt_md_access_new")
> > int BPF_PROG(fentry, struct sk_buff *skb)
> >
> > It's not possible to access skb argument in the fentry program,
> > with following error from verifier:
> >
> > ; int BPF_PROG(fentry, struct sk_buff *skb)
> > 0: (79) r1 = *(u64 *)(r1 +0)
> > invalid bpf_context access off=0 size=8
> >
> > The problem is that btf_ctx_access gets the context type for the
> > traced program, which is in this case the extension.
> >
> > But when we trace extension program, we want to get the context
> > type of the program that the extension is attached to, so we can
> > access the argument properly in the trace program.
> >
> > This version of the patch is tweaked slightly from Jiri's original one,
> > since the refactoring in the previous patches means we have to get the
> > target prog type from the new variable in prog->aux instead of directly
> > from the target prog.
> >
> > Reported-by: Eelco Chaudron <echaudro@xxxxxxxxxx>
> > Suggested-by: Jiri Olsa <jolsa@xxxxxxxxxx>
> > Signed-off-by: Toke Høiland-Jørgensen <toke@xxxxxxxxxx>
> > ---
> > kernel/bpf/btf.c | 9 ++++++++-
> > 1 file changed, 8 insertions(+), 1 deletion(-)
> >
> > diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
> > index 9228af9917a8..55f7b2ba1cbd 100644
> > --- a/kernel/bpf/btf.c
> > +++ b/kernel/bpf/btf.c
> > @@ -3860,7 +3860,14 @@ bool btf_ctx_access(int off, int size, enum bpf_access_type type,
> >
> > info->reg_type = PTR_TO_BTF_ID;
> > if (tgt_prog) {
> > - ret = btf_translate_to_vmlinux(log, btf, t, tgt_prog->type, arg);
> > + enum bpf_prog_type tgt_type;
> > +
> > + if (tgt_prog->type == BPF_PROG_TYPE_EXT)
> > + tgt_type = tgt_prog->aux->tgt_prog_type;
>
> what if tgt_prog->aux->tgt_prog_type is also BPF_PROG_TYPE_EXT? Should
> this be a loop?
ok, never mind this specifically. there is an explicit check
if (tgt_prog->type == prog->type) {
verbose(env, "Cannot recursively attach\n");
return -EINVAL;
}
that will prevent this.
But, I think we still will be able to construct a long chain of
fmod_ret -> freplace -> fmod_ret -> freplace -> and so on ad
infinitum. Can you please construct such a selftest? And then we
should probably fix those checks to also disallow FMOD_RET, in
addition to BPF_TRACE_FENTRY/FEXIT (and someone more familiar with LSM
prog type should check if that can cause any problems).
>
> Which also brings up a few follow up questions. Now that we allow same
> PROG_EXT program to be attached to multiple other programs:
>
> 1. what prevents us from attaching PROG_EXT to itself?
> 2. How do we prevent long chain of EXT programs or even loops?
>
> Can you please add a few selftests testing such cases? I have a
> feeling that with your changes in this patch set now it's possible to
> break the kernel very easily. I don't know what the proper solution
> is, but let's at least have a test that does show breakage, then try
> to figure out the solution. See also comment in check_attach_btf_id()
> about fentry/fexit and freplace interactions. That might not be
> enough.
>
>
> > + else
> > + tgt_type = tgt_prog->type;
> > +
> > + ret = btf_translate_to_vmlinux(log, btf, t, tgt_type, arg);
> > if (ret > 0) {
> > info->btf_id = ret;
> > return true;
> >
