Re: [PATCH bpf-next 2/4] bpf: verifier: explain opcode check in check_ld_imm()

[Yonghong Song]

On 5/20/22 4:50 PM, Yonghong Song wrote:
> On 5/20/22 4:37 AM, Shung-Hsi Yu wrote:
> > The BPF_SIZE check in the beginning of check_ld_imm() actually guard
> > against program with JMP instructions that goes to the second
> > instruction of BPF_LD_IMM64, but may be easily dismissed as an simple
> > opcode check that's duplicating the effort of bpf_opcode_in_insntable().
> >
> > Add comment to better reflect the importance of the check.
> >
> > Signed-off-by: Shung-Hsi Yu <shung-hsi.yu@xxxxxxxx>
> > ---
> >   kernel/bpf/verifier.c | 4 ++++
> >   1 file changed, 4 insertions(+)
> >
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index 79a2695ee2e2..133929751f80 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> > @@ -9921,6 +9921,10 @@ static int check_ld_imm(struct bpf_verifier_env 
> > *env, struct bpf_insn *insn)
> >       struct bpf_map *map;
> >       int err;
> > +    /* checks that this is not the second part of BPF_LD_IMM64, which is
> > +     * skipped over during opcode check, but a JMP with invalid 
> > offset may
> > +     * cause check_ld_imm() to be called upon it.
> > +     */
>
> The check_ld_imm() call context is:
>
>                  } else if (class == BPF_LD) {
>                          u8 mode = BPF_MODE(insn->code);
>
>                          if (mode == BPF_ABS || mode == BPF_IND) {
>                                  err = check_ld_abs(env, insn);
>                                  if (err)
>                                          return err;
>
>                          } else if (mode == BPF_IMM) {
>                                  err = check_ld_imm(env, insn);
>                                  if (err)
>                                          return err;
>
>                                  env->insn_idx++;
>                                  sanitize_mark_insn_seen(env);
>                          } else {
>                                  verbose(env, "invalid BPF_LD mode\n");
>                                  return -EINVAL;
>                          }
>                  }
>
> which is a normal checking of LD_imm64 insn.
>
> I think the to-be-added comment is incorrect and unnecessary.

Okay, double check again and now I understand what happens
when hitting the second insn of ldimm64 with a branch target.
Here we have BPF_LD = 0 and BPF_IMM = 0, so for a branch
target to the 2nd part of ldimm64, it will come to
check_ld_imm() and have error "invalid BPF_LD_IMM insn"

So check_ld_imm() is to check whether the insn is a
*legal* insn for the first part of ldimm64.

So the comment may be rewritten as below.

This is to verify whether an insn is a BPF_LD_IMM64
or not. But since BPF_LD = 0 and BPF_IMM = 0, if the branch
target comes to the second part of BPF_LD_IMM64,
the control may come here as well.

> >       if (BPF_SIZE(insn->code) != BPF_DW) {
> >           verbose(env, "invalid BPF_LD_IMM insn\n");
> >           return -EINVAL;