Re: [PATCH bpf-next 0/2] libbpf: Implement BTF Generator API

On Thu, Nov 4, 2021 at 7:58 AM Mauricio Vásquez Bernal
[off-list ref] wrote:

quoted

quoted

/* reduced version of struct bpf_core_spec */
struct bpf_core_spec_pub {
const struct btf *btf;
__u32 root_type_id;
enum bpf_core_relo_kind kind;
/* raw, low-level spec: 1-to-1 with accessor spec string */ --> we can
also use access_str_off and let the user parse it
int raw_spec[BPF_CORE_SPEC_MAX_LEN];

string might be a more "extensible" way, but we'll need to construct
that string for each relocation

quoted

/* raw spec length */
int raw_len;

using string would eliminate the need for this

quoted

};

struct bpf_core_relo_pub {
const char *prog_name; --> if we expose it by program then it's not needed.

yep, not sure about per-program yet, but that's minor

quoted

int insn_idx;

bool poison; --> allows the user to understand if the relocation
succeeded or not.

/* new field offset for field based core relos */
__u32 new_offset;

// TODO: fields for type and enum-based relos

isn't it always just u64 new_value for all types of relos? We can also
expose old_value just for completeness

Oh right. We can expose new_val, orig_val and let the user interpret
their meaning based on the relo_kind.

yep

quoted

quoted

struct bpf_core_spec_pub local_spec, targ_spec; --> BTFGen only needs
targ_spec, I suppose local spec would be useful for other use cases.

targ_spec doesn't seem necessary given we have root_type_id, relo
kind, access_string (or raw_spec). What am I missing?

Not sure I follow. root_type, relo kind and access_string are all part
of bpf_core_spec_pub, there are two instances of this structure,
targ_spec and local_spec.

Ah, ok, I got a bit confused by the formatting of your response. I got
the impression that we are exposing the same thing twice (and I'm not
talking about local vs target). So never mind.