On Mon, Oct 23, 2023 at 3:00 PM Dave Marchevsky <davemarchevsky@xxxxxx> wrote:
>
> If a struct has a bpf_refcount field, the refcount controls lifetime of
> the entire struct. Currently there's no usecase or support for multiple
> bpf_refcount fields in a struct.
>
> bpf_spin_lock and bpf_timer fields don't support multiples either, but
> with better error behavior. Parsing BTF w/ a struct containing multiple
> {bpf_spin_lock, bpf_timer} fields fails in btf_get_field_type, while
> multiple bpf_refcount fields doesn't fail BTF parsing at all, instead
> triggering a WARN_ON_ONCE in btf_parse_fields, with the verifier using
> the last bpf_refcount field to actually do refcounting.
>
> This patch changes bpf_refcount handling in btf_get_field_type to use
> same error logic as bpf_spin_lock and bpf_timer. Since it's being used
> 3x and is boilerplatey, the logic is shoved into
> field_mask_test_name_check_seen helper macro.
>
> Signed-off-by: Dave Marchevsky <davemarchevsky@xxxxxx>
> Fixes: d54730b50bae ("bpf: Introduce opaque bpf_refcount struct and add btf_record plumbing")
> ---
> kernel/bpf/btf.c | 37 ++++++++++++++++---------------------
> 1 file changed, 16 insertions(+), 21 deletions(-)
>
LGTM
Acked-by: Andrii Nakryiko <andrii@xxxxxxxxxx>
> diff --git a/kernel/bpf/btf.c b/kernel/bpf/btf.c
> index 15d71d2986d3..975ef8e73393 100644
> --- a/kernel/bpf/btf.c
> +++ b/kernel/bpf/btf.c
> @@ -3374,8 +3374,17 @@ btf_find_graph_root(const struct btf *btf, const struct btf_type *pt,
> return BTF_FIELD_FOUND;
> }
>
[...]
