On Wed, 2025-02-12 at 12:15 -0800, Ihor Solodrai wrote:
[...]
> diff --git a/btf_encoder.c b/btf_encoder.c
> index 965e8f0..3cec106 100644
> --- a/btf_encoder.c
> +++ b/btf_encoder.c
[...]
> +static int btf__tag_bpf_arena_ptr(struct btf *btf, int ptr_id)
> +{
> + const struct btf_type *ptr;
> + int tagged_type_id;
> +
> + ptr = btf__type_by_id(btf, ptr_id);
> + if (!btf_is_ptr(ptr))
> + return -EINVAL;
> +
> + tagged_type_id = btf__add_type_attr(btf, BPF_ARENA_ATTR, ptr->type);
> + if (tagged_type_id < 0)
> + return tagged_type_id;
> +
> + return btf__add_ptr(btf, tagged_type_id);
> +}
I might be confused, but this is a bit strange.
The type constructed here is: ptr -> type_tag -> t.
However, address_space is an attribute of a pointer, not a pointed type.
I think that correct sequence should be: type_tag -> ptr -> t.
This would make libbpf emit C declaration as follows:
void * __attribute__((address_space(1))) ptr;
Instead of current:
void __attribute__((address_space(1))) * ptr;
clang generates identical IR for both declarations:
@ptr = dso_local global ptr addrspace(1) null, align 8
Thus, imo, this function should be simplified as below:
static int btf__tag_bpf_arena_ptr(struct btf *btf, int ptr_id)
{
const struct btf_type *ptr;
ptr = btf__type_by_id(btf, ptr_id);
if (!btf_is_ptr(ptr))
return -EINVAL;
return btf__add_type_attr(btf, BPF_ARENA_ATTR, ptr_id);
}
[...]
