where is the entry point in kernel for an eBPF msg_verdict program?

Question

From the git log of linux at 174a7, I know the verdict flow is given as below

• recv_sock -> str_parser (parse_prog) -> verdict_prog -> skb_send_sock

I wonder what the flow is when there is only msg_verdict program.

tracing tcp_bpf_sendmsg(), I can see socket get parsed, but how does it pass to verdict part?

tcp_bpf_sendmsg
  tcp_bpf_send_verdict
    sk_psock_msg_verdict
      ...
      prog = READ_ONCE(psock->progs.msg_parser)
      ...
      ret = bpf_prog_run_pin_on_cpu(prog, msg)

Answer 1

This progs.msg_parser BPF program is actually the program you attached with BPF_SK_MSG_VERDICT, i.e. the program that will return a verdict (ex. __SK_PASS).

That can be observed by looking at the kernel code used to lookup and update programs attached to the sockmap:

static int sock_map_prog_lookup(struct bpf_map *map, struct bpf_prog ***pprog,
                u32 which)
{
    struct sk_psock_progs *progs = sock_map_progs(map);

    ...

    switch (which) {
    case BPF_SK_MSG_VERDICT:
        *pprog = &progs->msg_parser;
        break;

Although that may seem surprising, it is confirmed by the documentation: https://docs.kernel.org/bpf/map_sockmap.html (search for BPF_SK_MSG_VERDICT).