GitHub - boratanrikulu/gobee: Write your BPF programs in Go, not C. gobee transpiles a Go subset to BPF C and generates typed cilium/ebpf bindings.

Write your BPF programs in Go, not C. gobee transpiles a strict subset of Go into BPF C, generates typed Go bindings for the userspace side, and gates loads against the running kernel. The Go ecosystem has solid userspace tooling for BPF. The kernel side has always ended with "now write your program in C." Aya brought eBPF to Rust by writing a new BPF backend in rustc. gobee gets there a different way: by transpiling to C and reusing clang's mature backend. A Go file in, a BPF program out A tracepoint that streams every execve to userspace via a ringbuf:

Your input (Go) What gobee emits (BPF C)

//go:build ignore

package main

import "github.com/boratanrikulu/gobee/bpf"

//bpf:license GPL

type Event struct { Pid uint32 Comm [16]byte }

var Events = bpf.RingBuf[Event]{ MaxEntries: 4096, }

//bpf:section tracepoint/syscalls/sys_enter_execve func OnExec(ctx *bpf.ExecveEnterCtx) bpf.TpReturn { e, ok := Events.Reserve() if !ok { return bpf.TpOk } e.Pid = bpf.GetCurrentPid() bpf.GetTaskComm(&e.Comm) Events.Submit(e) return bpf.TpOk }

func main() {}

// Code generated by gobee. DO NOT EDIT.

#include "vmlinux.h" #include <bpf/bpf_helpers.h> #include <bpf/bpf_core_read.h>

char _license[] SEC("license") = "GPL";

struct Event { __u32 Pid; __u8 Comm[16]; };

struct { __uint(type, BPF_MAP_TYPE_RINGBUF); __uint(max_entries, 4096); } Events SEC(".maps");

SEC("tracepoint/syscalls/sys_enter_execve") int OnExec(struct trace_event_raw_sys_enter *ctx) { struct Event *e = bpf_ringbuf_reserve( &Events, sizeof(struct Event), 0); if (!

e) { return 0; } e->Pid = (__u32)( bpf_get_current_pid_tgid() >> 32); bpf_get_current_comm(&e->Comm, 16); bpf_ringbuf_submit(e, 0); return 0; }

gobee translate --bindings-dir ./bpf ./bpf/src produces both files, plus a sourcemap (events.bpf.c.map) so verifier errors map back to Go lines and a typed bindings file (bpf/events_bindings.go) so the userspace driver writes objs.Events, objs.AttachOnExec(), and decodes ringbuf payloads straight into bpf.Event (the same struct you see above, re-published in Go) instead of stringly-typed coll.Programs["..."] lookups. The C is readable on purpose. If gobee emits something weird, you can see it. For tracepoints + kprobes + XDP combined into one binary, see example/sysmon/. How it compares

gobee C + clang + bpf2go Aya (Rust) bpftrace BCC

Kernel-side language Go subset C Rust DSL C

Userspace integration typed Go bindings + cilium/ebpf bpf2go aya-runtime none python

CO-RE ✅ via clang ✅ ✅ via LLVM ✅ ✅

Helper coverage 200 typed Go wrappers full (write C) full limited full (write C)

Verifier error → source ✅ Go file:line:col ❌ raw C ✅ Rust file:line ❌ partial

Kernel-version gate at load ✅ via bpfvet manual manual n/a runtime

Toolchain deps Go + clang clang + bpf2go rustc + LLVM bpftrace python + bcc

Generated artifact .bpf.o + Go binary .bpf.o + Go binary .bpf.o + Rust binary JIT JIT

If you're already in a C / libbpf workflow, gobee is not trying to replace it wholesale.

It's for cases where you want the kernel side, the userspace side, and the build pipeline all in one Go module. What's supported today See docs/status.md for the full matrix (Go subset, statements, expressions, every helper, every map type, every directive). Quick view:

Surface Coverage

Program types (8) XDP, tracepoint, kprobe / kretprobe, uprobe / uretprobe, sock_ops, TC, cgroup_skb, LSM

Map types (19) array, hash, lru_hash, per-CPU variants, bloom_filter, lpm_trie, ringbuf, perf_event_array, prog_array, queue, stack, sk/task/inode storage, devmap/cpumap/xskmap

BPF helpers ~200 typed Go stubs auto-generated from libbpf v1.5.0 headers. The ones exercised by example/helloworld/ and example/sysmon/ are tested in real-kernel CI; the rest are unverified. File an issue if a stub doesn't match the kernel signature

CO-RE ✅ auto-detected. BPF_CORE_READ for kernel-internal struct fields (task_struct, sock, inode); direct ctx->field for UAPI BPF context structs (xdp_md, __sk_buff, bpf_sock_ops). Exercised on Linux 6.x (Ubuntu 24.04 CI); older kernels not yet in the CI matrix

BTF-ready output ✅ emitted C includes vmlinux.h and uses BPF_CORE_READ for kernel-internal field reads, so the BTF clang generates from clang -g carries the right relocations. clang itself stays your responsibility (the example Makefiles show the canonical invocation)

User-defined helpers ✅ top-level Go funcs without //bpf:section are emitted as static __always_inline C functions

Typed Go bindings ✅ Load<Stem>, Close, per-program Attach<Name>, AttachAll, plus your kernel-side struct types and constants re-published in Go

Kernel-version gate ✅ bpfvet runs at load time.

Fails fast with bpf program needs kernel >= 5.8, host is 5.4 instead of opaque EINVAL

Verifier error → Go source ✅ auto-annotated inside Load<Stem>. No manual pipe to gobee diagnose; *ebpf.VerifierError comes back with → counter.go:18:5 markers

Sourcemap sidecar ✅ <stem>.bpf.c.map written next to every .bpf.c for offline gobee diagnose use too

Cross-arch ✅ Linux arm64 + amd64

What gobee does

Transpiles a Go subset to BPF C (and runs go/types over your input first, so misuses surface at file:line:col). Generates a typed <Stem>_bindings.go next to the .bpf.c: bpf.LoadCounter(spec), objs.PerIface.Lookup(...), objs.AttachAll(ifindex), plus your kernel-side struct types and constants re-published in Go. Auto-annotates *ebpf.VerifierError from LoadAndAssign with Go source positions, no manual gobee diagnose pipe needed. Runs bpfvet inside Load<Stem> so old kernels fail fast with bpf program needs kernel >= 5.8, host is 5.4. Surfaces ~200 typed Go stubs for the libbpf v1.5.0 helper set, plus user-defined helper functions emitted as static __always_inline.

What gobee won't do

Replace clang. clang's BPF backend gives us CO-RE, BTF, and verifier-friendly codegen for free. Reimplementing that costs years and gains nothing. Replace cilium/ebpf. The generated bindings sit on top of it. Hide BPF. The Go subset maps 1:1 to BPF C idioms. If you know BPF, gobee is thin sugar. If you don't, the manual is still required reading. Run clang for you. Compile, embed, and load remain user-owned. Same pattern as bpf2go.

Why transpile, not generate BPF directly gc, the Go compiler, has no LLVM-based BPF backend. Adding one is a multi-year compiler project. rustc is built on LLVM and that's why Aya works.

So gobee emits C and reuses clang's BPF backend, which gives us mature codegen, BTF, and CO-RE relocations for free. Quickstart go install github.com/boratanrikulu/gobee/cmd/gobee@latest

cd example/helloworld make build # gobee translate, clang, go build sudo ./helloworld eth0 You'll need clang with the BPF target. On Linux that's the distro package; on macOS, brew install llvm. The transpiler itself is pure Go and runs anywhere. Project layout (typical) yourproject/ ├── bpf/ # Go package, importable from anywhere in your project │ ├── embed_amd64.go # //go:embed bin/x86/your.bpf.o │ ├── embed_arm64.go │ ├── your_bindings.go # generated by gobee │ ├── bin/{x86,arm64}/your.bpf.o │ └── src/ # not a Go package; clang lives here │ ├── your.go # //go:build ignore: BPF source │ ├── your.bpf.c # generated │ ├── Makefile # clang per arch │ └── vmlinux.h # vendored BTF dump ├── main.go # imports yourproject/bpf └── Makefile

The split keeps bpf/ a clean importable Go package (Go rejects .c files in non-cgo packages). Kernel sources and clang artifacts live one level down in bpf/src/. Examples

example/helloworld/: the canonical XDP packet counter, ~40 lines BPF, ~80 lines userspace. example/sysmon/: XDP, two tracepoints, and a kprobe in one binary, sharing a ringbuf for events. Demonstrates per-syscall typed contexts, user-defined helper functions, and the AttachAll shortcut.

CI GitHub Actions runs four layers on every push:

go test, go vet, transpiler golden tests Coverage matrix: every map type and //bpf:section kind has at least one example clang compile of every curated example, then bpfvet portability report Real-kernel verifier acceptance: ebpf.

NewCollectionWithOptions on each .bpf.o (Ubuntu 24.04 runner, kernel 6.x)

Docs

docs/design.md: architecture and rationale docs/go-subset.md: accepted Go syntax in BPF source files docs/directives.md: //bpf:* reference docs/status.md: support matrix (single source of truth)

Toolchain

The gobee binary builds anywhere (pure Go, no CGO). Compiling .bpf.o needs clang with the BPF target. Apple's bundled clang doesn't ship with it; on macOS use brew install llvm or build inside a Linux VM. Running the artifact needs Linux on arm64 or amd64.

Inspirations

Solod: the Go-to-C transpiler that proved this pattern works. Aya: the Rust eBPF framework whose ergonomics gobee chases.

License MIT. See LICENSE. Copyright (c) 2026 Bora Tanrikulu <me@bora.sh>