• API documentation

ocaml-libbpf

Libbpf C-bindings for loading eBPF ELF files into the kernel with OCaml.

Writing eBPF programs consist of two distinct parts. Implementing the code that executes in-kernel and user-level code responsible for loading/initializing/linking/teardown of the in-kernel code. This OCaml library provides the latter via binding the C libbpf library. It exposes both the raw low-level bindings as well as a set of high-level API's for handling your eBPF objects. As of now, the kernel part must still be written in restricted C and compiled with llvm to eBPF bytecode.

The full API set of Libbpf is quite large, see supported for the list of currently bound API's. Contributions are welcome.

External dependencies

ocaml-libbpf depends on the system package of libbpf.

Usage

⚠️ Disambiguation: The name of this repository and references to it will be "ocaml-libbpf". However, the library's entry module and package name is Libbpf. To install it, you would use opam install libbpf. To access it's High-level API's use Libbpf.<api>. To use the raw bindings, they are exposed in Libbpf.C.<api> namespace.

See examples directory on how ocaml-libbpf can be used to load eBPF ELF files into the kernel and interact with the loaded kernel program. The eBPF kernel programs are defined in *.bpf.c source files and are compiled with clang as specified in the dune rules. ocaml-libbpf exposes some high-level API's exposed by the toplevel Libbpf module to make it easy to perform repetitive tasks such as open/load/linking/initializing/teardown of bpf programs.

To run these examples, clone this repository and set up the package with

git clone git@github.com:koonwen/ocaml-libbpf.git
cd ocaml-libbpf
opam install . --deps-only
eval $(opam env)

then run make < minimal | kprobe | bootstrap | tc > to try out the different bpf programs. These examples are all taken from libbpf-bootstrap repository and rewritten in OCaml.

Open/Load/Link

Now let's run through an example of how we would use ocaml-libbpf. This usage tutorial assumes some knowledge of how to write eBPF kernel programs in C compile them to ELF files. If not, you can check out this resource. ocaml-libbpf provides an easy API to install your eBPF program into the kernel. Say your eBPF kernel program looks like this where we print something whenever the syscall write event occurs. We also want to implement a filtering mechanism to only print on write calls for our process. To do this, we initialize a BPF array map with a single entry that works like a holder for our global variable. The BPF map is neccessary to because it allows us to communicate values between user and kernel space.

The libbpf C library in fact already supports declarations of global variables in the usual form with the ability to manage them in user space. However for various technical reasons, ocaml-libbpf does not enable that feature yet. So we use the old style of working with global variables here.

// SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause
/* Copyright (c) 2020 Facebook */
#include <linux/bpf.h>
#include "bpf/bpf_helpers.h" /* This is from our libbpf library */

char LICENSE[] SEC("license") = "Dual BSD/GPL";

/* Globals implemented as an array */
struct {
  __uint(type, BPF_MAP_TYPE_ARRAY);
  __uint(max_entries, 1);
  __type(key, int);
  __type(value, long);
} globals SEC(".maps");

int my_pid_index = 0;

SEC("tp/syscalls/sys_enter_write")
int handle_tp(void *ctx) {
  int pid = bpf_get_current_pid_tgid() >> 32;

  long *my_pid;
  my_pid = bpf_map_lookup_elem(&globals, &my_pid_index);
  if (my_pid == NULL) {
    bpf_printk("Error got NULL");
    return 1;
  };

  if (pid != *my_pid)
    return 0;

  bpf_printk("Hello, BPF triggered from PID %d", pid);

  return 0;
}

After compilation to eBPF ELF file as minimal.o. Users just need to provide the path to this ELF file along with the name of the program and optionally an initialization function. Note that the name of the program refers to the function identifier under the SEC(...) attribute, in this case it is "handle_tp".

open Libbpf

let obj_path = "minimal.bpf.o"
let program_names = [ "handle_tp" ]

let () =
  with_bpf_object_open_load_link ~obj_path ~program_names ~before_link
    (fun obj link ->

	< user code to interact with bpf program running in kernel >

	)

The API provided by ocaml-libbpf with_bpf_object_open_load_link is a context manager that ensures the proper cleanup of resources if a failure is encountered. Right now our loaded kernel program is attached to the kernel and then immediately unloaded, users are responsible for keeping the bpf program alive by looping within the function block.

Users may also pin the bpf program to persist after user code exits. Do note that if pinning is desired, users should not use the with_bpf_object_open_load_link API and instead manually load and attach their bpf program since the context manager shutdowns all resources on exit.

Now let's add some looping logic to keep the loaded bpf program alive.

let obj_path = "minimal.bpf.o"
let program_names = [ "handle_tp" ]

let () =
  with_bpf_object_open_load_link ~obj_path ~program_names ~before_link
    (fun obj link ->

	(* Set up signal handlers *)
      let exitting = ref true in
      let sig_handler = Sys.Signal_handle (fun _ -> exitting := false) in
      Sys.(set_signal sigint sig_handler);
      Sys.(set_signal sigterm sig_handler);

      Printf.printf
        "Successfully started! Please run `sudo cat \
         /sys/kernel/debug/tracing/trace_pipe` to see output of the BPF \
         programs.\n\
         %!"

    (* Loop until Ctrl-C is called *)
      while !exitting do
        Printf.eprintf ".%!";
        Unix.sleepf 1.0
      done)

Our bpf program is now running in the kernel until we decide to interrupt it. However, it doesn't do exactly what we want. In particular, it doesn't filter for our process PID. This is because we haven't loaded our process PID into the BPF map. To do this, we need the name of the map we declared in the minimal.bpf.c program. In this case, our BPF array map was named globals.

let map = "globals"

(* Load PID into BPF map *)
let before_link obj =
  let pid = Unix.getpid () |> Signed.Long.of_int in
  let global_map = bpf_object_find_map_by_name obj map in
  (* When updating an element, users need to specify the type of the key and value
     declared by the map which checks that the key and value size are consistent. *)
  bpf_map_update_elem ~key_ty:Ctypes.int ~val_ty:Ctypes.long global_map 0 pid

Put together in minimal.ml, your bpf program runs in kernel and print to the trace pipe every second.

Maps

libbpf_maps is an optional convenience package that provides wrappers for BPF maps. Currently only Ringbuffer maps are added. An example usage of them can be found in examples/bootstrap.ml. This has been packaged separately since it drags in libffi dependency.

Notes on compatibility

The libbpf C library is designed to be kernel-agnostic and work across multitude of kernel versions. It has built-in mechanisms to gracefully handle older kernels, that are missing some of the features, by working around or gracefully degrading functionality.

Vendoring libbpf was a option. However, since bpf programs require writing the kernel components that may use libbpf, we made the choice to use the system's package versioned instead. This avoids users from knowingly/unknowingly using libbpf API's from two different versions. As a consequence, this library support operating systems that package libbpf.v.1.1 and up. Check ocaml-ci for the list of operating systems that successfully builds.

If so desired, you can also checkout the vendored branch in this repo which builds ocaml-libbpf with the latest libbpf package.