The logic is the following (see also the last patch for some more
documentation):
- hid-bpf first preloads a BPF program in the kernel that does a few
  things:
   * find out which attach_btf_id are associated with our trace points
   * adds a bpf_tail_call() BPF program that I can use to "call" any
     other BPF program stored into a jump table
   * monitors the releases of struct bpf_prog, and when there are no
     other users than us, detach the bpf progs from the HID devices
- users then declare their tracepoints and then call
  hid_bpf_attach_prog() in a SEC("syscall") program
- hid-bpf then calls multiple time the bpf_tail_call() program with a
  different index in the jump table whenever there is an event coming
  from a matching HID device

So driver abstractions like UDI are now perfectly fine as long as they
are written using a hip new VM?

This whole idea seems like a bad idea, against the Linux spirit and
now actually useful - it is totally trivial to write a new HID
driver alreay, and if it isn't in some cases we need to fix that.

So a big fat NAK to the idea of using eBPF for actual driver logic.