feat: M1 lumen-core (FEC/crypto/packet/session + C ABI) and workspace scaffold

Ground-up low-latency streaming stack per docs/implementation-plan.md. M1 is
complete and tested; Linux host backends are cfg-gated stubs to be filled in on
real hardware (M0/M2).

lumen-core (built + tested on macOS/aarch64 — 21 tests):
- fec: ErasureCoder over GF(2^8) (reed-solomon-erasure, Moonlight-compatible)
  and GF(2^16) Leopard-RS (reed-solomon-simd, the >1 Gbps wall-breaker); proptested
- packet: zero-copy #[repr(C)] framing, multi-block, FEC-aware reassembly
- crypto: AES-128-GCM with per-direction nonce salts + sequence-as-AAD
- session: host submit / client poll hot paths + input; loopback & UDP transports
- abi: opaque handles, versioned LumenConfig, panic guards; cbindgen-generated header
- acceptance: Rust loopback+proptest and a C harness that links the staticlib

Scaffold (compiles green on all platforms): lumen-host (vdisplay/capture/encode/
inject/web/pipeline seams under cfg(linux)), lumen-client-rs, tools/{loss-harness,
latency-probe}, Apple/Android client stubs, Gitea CI, docs.

Hardened against a multi-agent adversarial review (13 verified findings fixed,
regression-tested): reassembler memory-DoS bounds + block-consistency validation,
GCM nonce-reuse direction separation, ABI struct_size guard + range checks, FEC
shard-length guards, shard_payload datagram bound, key zeroization + Debug redaction.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
2026-06-09 00:02:52 +02:00
parent 4a1e3cd2fd
commit a913042367
47 changed files with 6015 additions and 0 deletions
+32
View File
@@ -0,0 +1,32 @@
//! Frame capture (plan §7). On Linux: a PipeWire ScreenCast portal stream delivering
//! dmabuf frames with no copy to the CPU. The encoder imports the dmabuf directly.
use anyhow::Result;
/// A captured frame. For zero-copy the real type wraps a dmabuf fd + modifier; the CPU
/// buffer is only a fallback path (plan §9 risk: per-GPU dmabuf import quirks).
pub struct CapturedFrame {
pub width: u32,
pub height: u32,
pub pts_ns: u64,
/// Fallback CPU pixels (empty when a dmabuf is used).
pub cpu_bytes: Vec<u8>,
}
/// Produces frames from a captured output. Lives on its own thread, feeding the encoder
/// over a bounded drop-oldest channel (never block the compositor).
pub trait Capturer: Send {
fn next_frame(&mut self) -> Result<CapturedFrame>;
}
/// Open a capturer for a PipeWire node id (from the ScreenCast portal).
pub fn open_pipewire(_node_id: u32) -> Result<Box<dyn Capturer>> {
#[cfg(target_os = "linux")]
{
anyhow::bail!("pipewire capture not yet implemented (M0)")
}
#[cfg(not(target_os = "linux"))]
{
anyhow::bail!("capture requires Linux + PipeWire")
}
}
+40
View File
@@ -0,0 +1,40 @@
//! Hardware video encode (plan §7). Binds FFmpeg (VAAPI / NVENC); never rewrites codecs.
//! Low-latency preset, lookahead off, dmabuf import for zero-copy from [`crate::capture`].
use crate::capture::CapturedFrame;
use anyhow::Result;
/// An encoded access unit (one NAL/AU) to hand to `lumen_core` for FEC + packetization.
pub struct EncodedFrame {
pub data: Vec<u8>,
pub pts_ns: u64,
/// True for IDR/keyframes (sets the SOF/keyframe wire flags).
pub keyframe: bool,
}
/// Codec selection negotiated with the client.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Codec {
H264,
H265,
Av1,
}
/// A hardware encoder. One per session; runs on the encode thread.
pub trait Encoder: Send {
fn submit(&mut self, frame: &CapturedFrame) -> Result<()>;
/// Pull the next encoded AU if one is ready.
fn poll(&mut self) -> Result<Option<EncodedFrame>>;
}
/// Open an encoder. `bitrate_bps` and `codec` come from session negotiation.
pub fn open(_codec: Codec, _bitrate_bps: u64) -> Result<Box<dyn Encoder>> {
#[cfg(target_os = "linux")]
{
anyhow::bail!("VAAPI/NVENC encode not yet implemented (M0)")
}
#[cfg(not(target_os = "linux"))]
{
anyhow::bail!("encode requires Linux (VAAPI/NVENC via FFmpeg)")
}
}
+30
View File
@@ -0,0 +1,30 @@
//! Input injection (plan §4): turn client [`lumen_core::input::InputEvent`]s into host
//! input. Wayland-native via libei (`reis`) first; uinput as the universal fallback.
use anyhow::Result;
use lumen_core::input::InputEvent;
/// Injects input events into the host session.
pub trait InputInjector: Send {
fn inject(&mut self, event: &InputEvent) -> Result<()>;
}
/// Preferred injection backend.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Backend {
/// libei via `reis` — Wayland-native (RemoteDesktop portal).
Libei,
/// `/dev/uinput` — universal fallback, always available.
Uinput,
}
pub fn open(_backend: Backend) -> Result<Box<dyn InputInjector>> {
#[cfg(target_os = "linux")]
{
anyhow::bail!("libei/uinput injection not yet implemented (M2)")
}
#[cfg(not(target_os = "linux"))]
{
anyhow::bail!("input injection requires Linux (libei/uinput)")
}
}
+61
View File
@@ -0,0 +1,61 @@
//! `lumen-host` — the Linux streaming host (plan §2, §6, §7).
//!
//! Creates a client-sized virtual display, captures it via PipeWire, encodes with
//! VAAPI/NVENC, and hands encoded access units to `lumen_core` for FEC + packetization +
//! pacing + send. Input flows back via libei/uinput. The platform backends are
//! `#[cfg(target_os = "linux")]`; the crate compiles everywhere so the workspace builds
//! on non-Linux dev machines — it just can't run the pipeline there.
//!
//! Status: scaffold. M0 wires capture→encode→file; M2 wires the full P1 host that a
//! stock Moonlight client connects to.
// Scaffold: trait methods and config paths are defined ahead of their backends.
#![allow(dead_code)]
mod capture;
mod encode;
mod inject;
mod pipeline;
mod vdisplay;
mod web;
use vdisplay::{Compositor, Mode};
fn main() {
tracing_subscriber::fmt()
.with_env_filter(
tracing_subscriber::EnvFilter::try_from_default_env().unwrap_or_else(|_| "info".into()),
)
.init();
tracing::info!(
"lumen-host scaffold (lumen_core ABI v{})",
lumen_core::ABI_VERSION
);
// The intended startup sequence (each step is a separate, pluggable subsystem):
// 1. negotiate mode + codec + FEC scheme over the control plane (web::WebConfig)
// 2. vdisplay::open(compositor).create(mode) -> client-sized virtual output
// 3. capture::open_pipewire(node) ; encode::open(codec, bitrate)
// 4. build a lumen_core::Session (host role) over a UDP transport to the client
// 5. loop pipeline::pump_once(..) until disconnect, then destroy the output
let target_mode = Mode {
width: 2560,
height: 1440,
refresh_hz: 240,
};
let compositor = Compositor::Kwin; // MVP target
if cfg!(target_os = "linux") {
tracing::info!(
?compositor,
?target_mode,
"would create a virtual output and start streaming (backends pending M0/M2)"
);
} else {
tracing::warn!(
"this is a Linux host; on {} only the shared lumen_core builds and is testable",
std::env::consts::OS
);
}
}
+39
View File
@@ -0,0 +1,39 @@
//! The host hot path (plan §7), wiring the platform stages to `lumen_core`:
//!
//! ```text
//! capture(dmabuf) → encode(NVENC/VAAPI) → core[FEC+packetize+pace+send]
//! ```
//!
//! Each stage runs on its own native OS thread, connected by bounded SPSC channels with
//! drop-oldest on overflow so the encoder is never blocked. No async runtime here.
use crate::capture::Capturer;
use crate::encode::{EncodedFrame, Encoder};
use anyhow::Result;
use lumen_core::packet::{FLAG_PIC, FLAG_SOF};
use lumen_core::Session;
/// Drive one capture→encode→submit step. The real pipeline spawns threads and uses
/// bounded channels; this documents the data flow and the `lumen_core` submit contract.
pub fn pump_once(
capturer: &mut dyn Capturer,
encoder: &mut dyn Encoder,
session: &mut Session,
) -> Result<()> {
let frame = capturer.next_frame()?;
encoder.submit(&frame)?;
while let Some(EncodedFrame {
data,
pts_ns,
keyframe,
}) = encoder.poll()?
{
let mut flags = FLAG_PIC as u32;
if keyframe {
flags |= FLAG_SOF as u32;
}
// core does FEC + packetize + pace + send.
session.submit_frame(&data, pts_ns, flags)?;
}
Ok(())
}
+96
View File
@@ -0,0 +1,96 @@
//! Virtual display orchestration (plan §6) — the project's differentiator.
//!
//! A [`VirtualDisplay`] creates a client-sized output on demand, to be captured and
//! streamed, then torn down on disconnect. Two deployment models exist (Model A: attach
//! to the running session; Model B: dedicated headless session); both sit behind this
//! trait so compositors are pluggable and a stuck one never blocks the project.
//!
//! Backends are `#[cfg(target_os = "linux")]` and currently stubs (see the per-backend
//! modules). The MVP target is KWin; a wlroots spike validates the pipeline first.
use anyhow::Result;
pub use lumen_core::Mode;
/// Opaque handle to a created virtual output, returned by [`VirtualDisplay::create`].
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub struct OutputHandle(pub u64);
/// Pluggable virtual-output creation, per compositor.
pub trait VirtualDisplay: Send {
/// Human-readable backend name (e.g. `"kwin"`, `"wlroots"`, `"mutter"`).
fn name(&self) -> &'static str;
/// Create a virtual output of the given mode.
fn create(&mut self, mode: Mode) -> Result<OutputHandle>;
/// Destroy a previously created output.
fn destroy(&mut self, handle: OutputHandle) -> Result<()>;
}
/// Compositors lumen knows how to drive (plan §6).
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub enum Compositor {
/// KWin / Plasma 6 — MVP target (matches the CachyOS/KDE daily driver).
Kwin,
/// wlroots (Sway/Hyprland) — fastest to prototype the pipeline.
Wlroots,
/// Mutter / GNOME — headless backend + Mutter DBus.
Mutter,
}
/// Detect or select a backend and return its driver.
pub fn open(compositor: Compositor) -> Result<Box<dyn VirtualDisplay>> {
#[cfg(target_os = "linux")]
{
match compositor {
Compositor::Kwin => Ok(Box::new(linux::kwin::KwinDisplay::new()?)),
Compositor::Wlroots => Ok(Box::new(linux::wlroots::WlrootsDisplay::new()?)),
Compositor::Mutter => Ok(Box::new(linux::mutter::MutterDisplay::new()?)),
}
}
#[cfg(not(target_os = "linux"))]
{
let _ = compositor;
anyhow::bail!("virtual displays require Linux (Wayland compositor)")
}
}
#[cfg(target_os = "linux")]
mod linux {
//! Linux backends. TODO(M2): drive KWin via DBus (study KRdp's source for the
//! virtual-output path); wlroots via `create_output` on the headless backend;
//! Mutter via `org.gnome.Mutter.*`.
macro_rules! stub_backend {
($modname:ident, $ty:ident, $name:literal) => {
pub mod $modname {
use super::super::{Mode, OutputHandle, VirtualDisplay};
use anyhow::Result;
pub struct $ty;
impl $ty {
pub fn new() -> Result<Self> {
Ok($ty)
}
}
impl VirtualDisplay for $ty {
fn name(&self) -> &'static str {
$name
}
fn create(&mut self, _mode: Mode) -> Result<OutputHandle> {
anyhow::bail!(concat!(
$name,
" virtual-output creation not yet implemented"
))
}
fn destroy(&mut self, _handle: OutputHandle) -> Result<()> {
anyhow::bail!(concat!(
$name,
" virtual-output destroy not yet implemented"
))
}
}
}
};
}
stub_backend!(kwin, KwinDisplay, "kwin");
stub_backend!(wlroots, WlrootsDisplay, "wlroots");
stub_backend!(mutter, MutterDisplay, "mutter");
}
+24
View File
@@ -0,0 +1,24 @@
//! Web config / pairing API (plan §4) — control plane only. This is where `tokio`/`axum`
//! are permitted; the per-frame pipeline never touches them. Serves pairing, client
//! identity/permissions, and surfaces [`lumen_core::Stats`] for the measurement UI (M3).
use anyhow::Result;
/// Control-plane configuration server. Stub until the pairing/RTSP surface is scoped
/// (plan §12 action 4: confirm exactly which serverinfo/RTSP/pairing messages a current
/// Moonlight client needs for P1).
pub struct WebConfig {
pub bind: String,
}
impl WebConfig {
pub fn new(bind: impl Into<String>) -> Self {
WebConfig { bind: bind.into() }
}
/// Run the control-plane server. TODO(M2): axum + tokio; GameStream `serverinfo`,
/// pairing handshake, RTSP SETUP with the `lumen/1` capability flag for negotiation.
pub fn run(&self) -> Result<()> {
anyhow::bail!("web/pairing control plane not yet implemented (M2)")
}
}