On Fri, Jul 14, 2023 at 10:56:00AM -0700, Alexei Starovoitov wrote:
> On Fri, Jul 14, 2023 at 7:20 AM Anton Protopopov <aspsk@xxxxxxxxxxxxx> wrote:
> >
> > Patch verifier to regard values of type CONST_PTR_TO_MAP as trusted
> > pointers to struct bpf_map. This allows kfuncs to work with `struct
> > bpf_map *` arguments.
> >
> > Save some bytes by defining btf_bpf_map_id as BTF_ID_LIST_GLOBAL_SINGLE
> > (which is u32[1]), not as BTF_ID_LIST (which is u32[64]).
> >
> > Signed-off-by: Anton Protopopov <aspsk@xxxxxxxxxxxxx>
> > ---
> > include/linux/btf_ids.h | 1 +
> > kernel/bpf/map_iter.c | 3 +--
> > kernel/bpf/verifier.c | 5 ++++-
> > 3 files changed, 6 insertions(+), 3 deletions(-)
> >
> > diff --git a/include/linux/btf_ids.h b/include/linux/btf_ids.h
> > index 00950cc03bff..a3462a9b8e18 100644
> > --- a/include/linux/btf_ids.h
> > +++ b/include/linux/btf_ids.h
> > @@ -267,5 +267,6 @@ MAX_BTF_TRACING_TYPE,
> > extern u32 btf_tracing_ids[];
> > extern u32 bpf_cgroup_btf_id[];
> > extern u32 bpf_local_storage_map_btf_id[];
> > +extern u32 btf_bpf_map_id[];
> >
> > #endif
> > diff --git a/kernel/bpf/map_iter.c b/kernel/bpf/map_iter.c
> > index d06d3b7150e5..b67996147895 100644
> > --- a/kernel/bpf/map_iter.c
> > +++ b/kernel/bpf/map_iter.c
> > @@ -78,8 +78,7 @@ static const struct seq_operations bpf_map_seq_ops = {
> > .show = bpf_map_seq_show,
> > };
> >
> > -BTF_ID_LIST(btf_bpf_map_id)
> > -BTF_ID(struct, bpf_map)
> > +BTF_ID_LIST_GLOBAL_SINGLE(btf_bpf_map_id, struct, bpf_map)
> >
> > static const struct bpf_iter_seq_info bpf_map_seq_info = {
> > .seq_ops = &bpf_map_seq_ops,
> > diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> > index 0b9da95331d7..5663f97ef292 100644
> > --- a/kernel/bpf/verifier.c
> > +++ b/kernel/bpf/verifier.c
> > @@ -5419,6 +5419,9 @@ static bool is_trusted_reg(const struct bpf_reg_state *reg)
> > if (reg->ref_obj_id)
> > return true;
> >
> > + if (reg->type == CONST_PTR_TO_MAP)
> > + return true;
> > +
>
> Overall it looks great.
> Instead of above, how about the following instead:
>
> diff --git a/kernel/bpf/verifier.c b/kernel/bpf/verifier.c
> index 0b9da95331d7..cd08167dc347 100644
> --- a/kernel/bpf/verifier.c
> +++ b/kernel/bpf/verifier.c
> @@ -10775,7 +10775,7 @@ static int check_kfunc_args(struct
> bpf_verifier_env *env, struct bpf_kfunc_call_
> if (!is_kfunc_trusted_args(meta) && !is_kfunc_rcu(meta))
> break;
>
> - if (!is_trusted_reg(reg)) {
> + if (!is_trusted_reg(reg) &&
> !reg2btf_ids[base_type(reg->type)]) {
>
>
> This way we won't need to list every convertible type in is_trusted_reg.
>
> I'm a bit hesitant to put reg2btf_ids[] check directly into is_trusted_reg().
> Maybe it's ok, but it needs more analysis.
I am not sure I see a difference in adding a check you proposed above and
adding the reg2btf_ids[] check directly into the is_trusted_reg() function.
Basically, we say "if type is in reg2btf_ids[], then consider it trusted" in
both cases. AFAIS, currently the reg2btf_ids[] contains only trusted types,
however, could it happen that we add a non-trusted type there?
So, I would leave the patch as is (which also makes sense because the
const-ptr-to-map is a special case), or add the "reg2btf_ids[] check"
directly into the is_trusted_reg() function.
