Files
nx9-wg/docs/ARCHITECTURE.md
2026-09-02 15:19:19 +05:30

160 lines
12 KiB
Markdown

# NX9 WireGuard — System Architecture & Workspace Structure
`nx9-wg` is structured as a modular six-crate Rust workspace designed for separation of concerns, strict failure domains, and testability.
```
nx9-wg (Workspace Root & Binary Executable)
├── crates/nx9-wg-core (Domain models, validation, cryptography, config)
├── crates/nx9-wg-db (SQLite store, migrations, repositories, WAL)
├── crates/nx9-wireguard (WireGuard Netlink execution, config builder, QR engine)
├── crates/nx9-wg-network (RTNETLINK networking, route management, nftables, procfs)
├── crates/nx9-wg-api (Axum REST API, WebSockets, auth, reconciliation, allocator)
└── crates/nx9-wg-ui (Design tokens, CSS engine, view models, SPA integration)
```
---
## 1. Crate Inventory & Layer Responsibilities
### `nx9-wg-core`
- **Domain Entities**: Typed domain structures (`Interface`, `Peer`, `Network`, `Route`, `FirewallRule`, `Setting`, `Admin`, `Session`, `ApiToken`, `BackupMeta`, `AuditEvent`).
- **Validation**: Strict RFC-compliant validators for CIDRs, IP addresses, MTU ranges, listen ports, interface names, peer names, and password strength.
- **Cryptography**: Argon2id password hashing, SHA-256 token digest generation, X25519 keypair generation, and secure random string generation.
- **Configuration**: Hierarchical TOML and `NX9_WG_*` environment variable configuration loader.
### `nx9-wg-db`
- **Authoritative Persistence**: Migration-driven SQLite storage using `sqlx` in Write-Ahead Logging (WAL) mode.
- **Single Admin Invariant**: SQL constraint `CHECK (id = 1)` enforcing single-administrator identity.
- **Repositories**: Isolated data access layers for admin, auth, audit, backups, client profiles, interfaces, login attempts, networks, peers, routes, sessions, settings, and tokens.
- **Transactional Consistency**: Foreign key cascades (`ON DELETE CASCADE`) between interfaces and peers.
### `nx9-wireguard`
- **Native Netlink Engine**: `NativeLinuxWireGuardEngine` interacting with Linux RTNETLINK for link lifecycle and Generic Netlink family `wireguard` (`SET_DEVICE`, `GET_DEVICE`, `ReplacePeers`).
- **Client Config Generation**: Pure Rust `.conf` generation supporting full-tunnel and split-tunnel topologies.
- **QR Code Engine**: High-resolution SVG, PNG byte streams, and UTF-8 ASCII terminal QR code rendering.
- **Simulation Engine**: `SimulatedWireGuardEngine` for non-Linux platform development and testing.
### `nx9-wg-network`
- **Native Network Engine**: `NativeLinuxNetworkEngine` managing RTNETLINK interfaces, IPv4/IPv6 addresses, and route tables.
- **nftables Netfilter Engine**: `NativeLinuxNftablesEngine` utilizing `libnftables.so.1` for transactional rule generation in `table inet nx9_wg`.
- **Kernel IP Forwarding**: Direct `/proc/sys/net/ipv4/ip_forward` and `/proc/sys/net/ipv6/conf/all/forwarding` mutation.
- **Simulation Engine**: `SimulatedNetworkEngine` for local platform testing.
### `nx9-wg-api`
- **REST Router**: Axum HTTP handlers for auth, interfaces, peers, networks, routes, firewall, NAT, forwarding, settings, backups, diagnostics, and audit logs.
- **WebSocket Event Bus**: Tokio broadcast channel broadcasting real-time system events (`AuditEvent`, `InterfaceChanged`, `PeerChanged`, `PeerHandshake`, `SettingsChanged`).
- **Reconciliation Engine**: Mutex-serialized continuous drift detection and convergence controller.
- **Deterministic IP Allocator**: Collision-resistant next-IP allocation engine for managed subnets.
- **Backup & Restore Service**: Online SQLite atomic snapshot (`VACUUM INTO`) and verification engine.
### `nx9-wg-ui`
- **Design Tokens**: Structured CSS variable tokens for Dark and Light themes.
- **Stylesheet Generator**: In-process CSS compiler producing responsive layouts (`@media (max-width: 768px)`).
- **View Models**: Strongly-typed Rust UI state representations for dashboards, diagnostics, client export modals, and peer management tables.
---
## 2. End-to-End Control and Execution Plane
```mermaid
flowchart TD
subgraph ControlPlane["Control Plane (User & Administration)"]
User["Administrator (Browser / CLI)"]
SPA["Embedded SPA Web UI"]
CLI["Native CLI (nx9-wg)"]
API["Axum REST API / WebSockets"]
DB[("Authoritative SQLite Database")]
end
subgraph ReconciliationLayer["Reconciliation & Engine Layer"]
Reconciler["Reconciliation Engine (Serialized Mutex)"]
WGEngine["WireGuard Engine (Genl / RTNETLINK)"]
NetEngine["Network Engine (RTNETLINK & procfs)"]
NftEngine["Nftables Engine (libnftables Netfilter)"]
end
subgraph LinuxKernel["Linux Kernel Execution Plane"]
WGSocket["WireGuard Kernel Module (wireguard.ko)"]
RTNL["Kernel RTNETLINK (Links, Addrs, Routes)"]
Procfs["/proc/sys/net (IP Forwarding)"]
Netfilter["Netfilter Subsystem (table inet nx9_wg)"]
end
User -->|HTTPS / WSS| SPA
User -->|CLI Invocations| CLI
SPA -->|REST API Calls| API
CLI -->|In-Process Service Calls| API
API -->|Authoritative Mutations| DB
DB -->|Desired State Snapshot| Reconciler
Reconciler -->|Read Live Telemetry| WGEngine
Reconciler -->|Read Live Routes/Forwarding| NetEngine
Reconciler -->|Read Live Ruleset| NftEngine
WGEngine -->|Generic Netlink Messages| WGSocket
NetEngine -->|RTM_NEWLINK / NEWROUTE| RTNL
NetEngine -->|Write 1/0| Procfs
NftEngine -->|Atomic Netfilter Transactions| Netfilter
```
---
## 3. Subsystem Invariants & Security Boundaries
1. **Subprocess Isolation**: Zero invocations of `std::process::Command` or shell scripts across the entire production codebase.
2. **Persistence Authority**: SQLite remains the single authoritative source of truth. Kernel state is continuously reconciled to match database state.
3. **Firewall Isolation**: All nftables operations are confined to `table inet nx9_wg`. Unmanaged host tables are untouched.
4. **Route Safety**: Default gateway routes and physical host networking routes are protected against accidental deletion or flushing.
5. **Secret Redaction**: Private keys, preshared keys, password hashes, and token hashes are masked in `Debug` formatters, CLI outputs, and API responses.
---
## 4. Interface Roles & Upstream Architecture
`nx9-wg` implements explicit `InterfaceRole` categorization across domain models, Netlink device configuration, and reconciliation:
```
┌────────────────────────────────────────────────────────────────────────────────────────┐
│ Linux Host Network │
│ │
│ ┌────────────────────────────────┐ ┌──────────────────────────────────┐ │
│ │ Primary Overlay (wg0) │ │ Optional Upstream (proton0) │ │
│ │ Role: Overlay │ │ Role: Upstream │ │
│ │ Local Listen Port: 51820 │ │ Local Listen Port: Auto (Dyn) │ │
│ │ Peers: 1..N Clients (Mobile) │ │ Peers: Exactly 1 Provider Peer │ │
│ │ Cryptokey AllowedIPs: /32 │ │ Cryptokey AllowedIPs: 0/0, ::0 │ │
│ └────────────────────────────────┘ └──────────────────────────────────┘ │
│ │ │ │
│ ▼ ▼ │
│ ┌────────────────────────────────┐ ┌──────────────────────────────────┐ │
│ │ Private Overlay Clients │ │ Remote Provider Endpoint │ │
│ │ (10.100.0.0/24 Subnet) │ │ (37.19.199.155:51820) │ │
│ └────────────────────────────────┘ └──────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌────────────────────────────────────────┐ │
│ │ Physical WAN Default Route (eno2) │ │
│ │ Gateway: 192.168.1.1 (FIB Unchanged) │ │
│ └────────────────────────────────────────┘ │
└────────────────────────────────────────────────────────────────────────────────────────┘
```
### A. Role Discriminator & Invariants
- **`InterfaceRole::Overlay`**: The primary private WireGuard overlay network. Exactly one instance exists (`wg0`). It binds to an explicit listen port (`51820`), hosts enrolled client peers, and is protected from deletion or disabling.
- **`InterfaceRole::Upstream`**: Optional third-party WireGuard VPN interfaces (e.g. `proton0`). Zero or more instances may exist concurrently. Each Upstream interface connects NX9-WG to an external service provider through exactly one provider peer.
### B. Optional Local Listen Port & Ephemeral Kernel Binding
- `Interface.listen_port` is modeled as `Option<u16>`.
- Standard third-party `.conf` imports (e.g. ProtonVPN) omit `[Interface] ListenPort`. NX9-WG preserves `listen_port = None` without defaulting to `51820`.
- In `configure_device`, omitting the `WireguardAttribute::ListenPort` Netlink attribute signals the Linux kernel to assign an ephemeral dynamic UDP port automatically.
- This prevents local UDP port contention and allows `wg0` (51820) and `proton0` (dynamic) to coexist without `-EADDRINUSE` errors.
- Dynamic kernel ports produce 0 false drift actions in the reconciliation engine when desired `listen_port` is `None`.
### C. Cryptokey Routing vs. Linux FIB Default Routes
- An Upstream provider peer often specifies `AllowedIPs = 0.0.0.0/0, ::/0` in its `.conf`.
- In WireGuard, `AllowedIPs` defines the device-level cryptokey packet routing filter; it does **not** install a Linux kernel route.
- NX9-WG preserves `0.0.0.0/0, ::/0` on the `proton0` WireGuard device without modifying the server's Linux FIB default gateway (`192.168.1.1`).
- The provider endpoint (`37.19.199.155:51820`) remains reachable via the host physical WAN interface.
### D. NAT Masquerade Isolation
- Outbound NAT masquerading compiled into `table inet nx9_wg` is strictly scoped to Overlay client subnets (`10.100.0.0/24`).
- Upstream tunnel addresses (`10.2.0.2/32`) are not treated as client subnets and do not trigger unsolicited global masquerading.