mc-proxy/ARCHITECTURE.md

# ARCHITECTURE.md

mc-proxy is a Layer 4 TLS proxy and router for Metacircular Dynamics
services. It inspects the SNI field of incoming TLS ClientHello messages to
determine the target backend, then proxies raw TCP between the client and
the appropriate container. A global firewall evaluates every connection
before routing.

## Table of Contents

1. [System Overview](#system-overview)
2. [Connection Lifecycle](#connection-lifecycle)
3. [Firewall](#firewall)
4. [Routing](#routing)
5. [Configuration](#configuration)
6. [Storage](#storage)
7. [Deployment](#deployment)
8. [Security Model](#security-model)
9. [Future Work](#future-work)

---

## System Overview

```
                        ┌─────────────────────────────────────┐
                        │             mc-proxy                │
    Clients ──────┐     │                                     │
                  │     │  ┌──────────┐   ┌───────┐   ┌─────┐ │     ┌────────────┐
                  ├────▶│  │ Firewall │──▶│  SNI  │──▶│Route│─│────▶│ Backend A  │
                  │     │  │ (global) │   │Extract│   │Table│ │     │ :8443      │
                  ├────▶│  └──────────┘   └───────┘   └─────┘ │     ├────────────┤
                  │     │       │ RST                    │    │     │ Backend B  │
    Clients ──────┘     │       ▼                        └────│────▶│ :9443      │
                        │   (blocked)                         │     └────────────┘
                        └─────────────────────────────────────┘

    Listener 1 (:443)  ─┐
    Listener 2 (:8443) ─┼─ Each listener runs the same pipeline
    Listener N (:9443) ─┘
```

Key properties:

- **Layer 4 only.** mc-proxy never terminates TLS. It reads just enough of
  the ClientHello to extract the SNI hostname, then proxies the raw TCP
  stream to the matched backend. The backend handles TLS termination.
- **TLS-only.** Non-TLS connections are not supported. If the first bytes of
  a connection are not a TLS ClientHello, the connection is reset.
- **Multiple listeners.** A single mc-proxy instance binds to one or more
  ports. Each listener runs the same firewall → SNI → route pipeline.
- **Global firewall.** Firewall rules apply to all listeners uniformly.
  There are no per-route firewall rules.
- **No authentication.** mc-proxy is pre-auth infrastructure. It sits in
  front of services that handle their own authentication via MCIAS.

---

## Connection Lifecycle

Every inbound connection follows this sequence:

```
1. ACCEPT        Listener accepts TCP connection.
2. FIREWALL      Check source IP against blocklists:
                   a. IP/CIDR block check.
                   b. GeoIP country block check.
                 If blocked → RST, done.
3. SNI EXTRACT   Read the TLS ClientHello (without consuming it).
                 Extract the SNI hostname.
                 If no valid ClientHello or no SNI → RST, done.
4. ROUTE LOOKUP  Match SNI hostname against the route table.
                 If no match → RST, done.
5. BACKEND DIAL  Open TCP connection to the matched backend address.
                 If dial fails → RST, done.
6. PROXY         Bidirectional byte copy: client ↔ backend.
                 The buffered ClientHello bytes are forwarded first,
                 then both directions copy concurrently.
7. CLOSE         Either side closes → half-close propagation → done.
```

### SNI Extraction

The proxy peeks at the initial bytes of the connection without consuming
them. It parses just enough of the TLS record layer and ClientHello to
extract the `server_name` extension. The full ClientHello (including the
SNI) is then forwarded to the backend so the backend's TLS handshake
proceeds normally.

If the ClientHello spans multiple TCP segments, the proxy buffers up to
16 KiB (the maximum TLS record size) before giving up.

### Bidirectional Copy

After the backend connection is established, the proxy runs two concurrent
copy loops (client→backend and backend→client). When either direction
encounters an EOF or error:

1. The write side of the opposite direction is half-closed.
2. The remaining direction drains to completion.
3. Both connections are closed.

Timeouts apply to the copy phase to prevent idle connections from
accumulating indefinitely (see [Configuration](#configuration)).

---

## Firewall

The firewall is a global, ordered rule set evaluated on every new
connection before SNI extraction. Rules are evaluated in definition order;
the first matching rule determines the outcome. If no rule matches, the
connection is **allowed** (default allow — the firewall is an explicit
blocklist, not an allowlist).

### Rule Types

| Config Field | Match Field | Example |
|--------------|-------------|---------|
| `blocked_ips` | Source IP address | `"192.0.2.1"` |
| `blocked_cidrs` | Source IP prefix | `"198.51.100.0/24"` |
| `blocked_countries` | Source country (ISO 3166-1 alpha-2) | `"KP"`, `"CN"`, `"IN"`, `"IL"` |

### Blocked Connection Handling

Blocked connections receive a TCP RST. No TLS alert, no HTTP error page, no
indication of why the connection was refused. This is intentional — blocked
sources should receive minimal information.

### GeoIP Database

mc-proxy uses the [MaxMind GeoLite2](https://dev.maxmind.com/geoip/geolite2-free-geolocation-data)
free database for country-level IP geolocation. The database file is
distributed separately from the binary and referenced by path in the
configuration.

The GeoIP database is loaded into memory at startup and can be reloaded
via `SIGHUP` without restarting the proxy. If the database file is missing
or unreadable at startup and GeoIP rules are configured, the proxy refuses
to start.

---

## Routing

Each listener has its own route table mapping SNI hostnames to backend
addresses. A route entry consists of:

| Field | Type | Description |
|-------|------|-------------|
| `hostname` | string | Exact SNI hostname to match (e.g. `metacrypt.metacircular.net`) |
| `backend` | string | Backend address as `host:port` (e.g. `127.0.0.1:8443`) |

Routes are scoped to the listener that accepted the connection. The same
hostname can appear on different listeners with different backends, allowing
the proxy to route the same service name to different backends depending
on which port the client connected to.

### Match Semantics

- Hostname matching is **exact** and **case-insensitive** (per RFC 6066,
  SNI hostnames are DNS names and compared case-insensitively).
- Wildcard matching is not supported in the initial implementation.
- If duplicate hostnames appear within the same listener, the proxy refuses
  to start.
- If no route matches an incoming SNI hostname, the connection is reset.

### Route Table Source

Route tables are persisted in the SQLite database. On first run, they are
seeded from the TOML configuration. On subsequent runs, the database is
the source of truth. Routes can be added or removed at runtime via the
gRPC admin API.

---

## Configuration

TOML configuration file, loaded at startup. The proxy refuses to start if
required fields are missing or invalid.

```toml
# Database. Required.
[database]
path = "/srv/mc-proxy/mc-proxy.db"

# Listeners. Each has its own route table (seeds DB on first run).
[[listeners]]
addr = ":443"

  [[listeners.routes]]
  hostname = "metacrypt.metacircular.net"
  backend  = "127.0.0.1:18443"

  [[listeners.routes]]
  hostname = "mcias.metacircular.net"
  backend  = "127.0.0.1:28443"

[[listeners]]
addr = ":8443"

  [[listeners.routes]]
  hostname = "metacrypt.metacircular.net"
  backend  = "127.0.0.1:18443"

[[listeners]]
addr = ":9443"

  [[listeners.routes]]
  hostname = "mcias.metacircular.net"
  backend  = "127.0.0.1:28443"

# gRPC admin API. Optional — omit addr to disable.
[grpc]
addr      = "127.0.0.1:9090"
tls_cert  = "/srv/mc-proxy/certs/cert.pem"
tls_key   = "/srv/mc-proxy/certs/key.pem"
client_ca = "/srv/mc-proxy/certs/ca.pem"

# Firewall. Global blocklist, evaluated before routing. Default allow.
[firewall]
geoip_db          = "/srv/mc-proxy/GeoLite2-Country.mmdb"
blocked_ips       = ["192.0.2.1"]
blocked_cidrs     = ["198.51.100.0/24"]
blocked_countries = ["KP", "CN", "IN", "IL"]

# Proxy behavior.
[proxy]
connect_timeout  = "5s"     # Timeout for dialing backend
idle_timeout     = "300s"   # Close connections idle longer than this
shutdown_timeout = "30s"    # Graceful shutdown drain period

# Logging.
[log]
level = "info"              # debug, info, warn, error
```

### Environment Variable Overrides

Configuration values can be overridden via environment variables using the
prefix `MCPROXY_` with underscore-separated paths:

```
MCPROXY_LOG_LEVEL=debug
MCPROXY_PROXY_IDLE_TIMEOUT=600s
```

Environment variables cannot define listeners, routes, or firewall rules —
these are structural and must be in the TOML file.

---

## Storage

### SQLite Database

Listeners, routes, and firewall rules are persisted in a SQLite database
(WAL mode, foreign keys enabled, busy timeout 5000ms). The pure-Go driver
`modernc.org/sqlite` is used (no CGo).

**Startup behavior:**

1. Open the database at the configured path. Run migrations.
2. If the database is empty (first run): seed from the TOML config.
3. If the database has data: load from it. TOML listener/route/firewall
   fields are ignored.

The TOML config continues to own operational settings: proxy timeouts,
log level, gRPC config, GeoIP database path.

**Write-through pattern:** The gRPC admin API writes to the database first,
then updates in-memory state. If the database write fails, the in-memory
state is not modified.

### Schema

```sql
CREATE TABLE listeners (
    id   INTEGER PRIMARY KEY,
    addr TEXT NOT NULL UNIQUE
);

CREATE TABLE routes (
    id          INTEGER PRIMARY KEY,
    listener_id INTEGER NOT NULL REFERENCES listeners(id) ON DELETE CASCADE,
    hostname    TEXT NOT NULL,
    backend     TEXT NOT NULL,
    UNIQUE(listener_id, hostname)
);

CREATE TABLE firewall_rules (
    id    INTEGER PRIMARY KEY,
    type  TEXT NOT NULL CHECK(type IN ('ip', 'cidr', 'country')),
    value TEXT NOT NULL,
    UNIQUE(type, value)
);
```

### Data Directory

```
/srv/mc-proxy/
├── mc-proxy.toml               Configuration
├── mc-proxy.db                 SQLite database
├── certs/                      TLS certificates (for gRPC admin API)
├── GeoLite2-Country.mmdb       GeoIP database (if using country blocks)
└── backups/                    Database snapshots
```

mc-proxy does not terminate TLS on the proxy listeners, so no proxy
certificates are needed. The `certs/` directory is for the gRPC admin
API's TLS and optional mTLS keypair.

---

## Deployment

### Binary

Single static binary, built with `CGO_ENABLED=0`. No runtime dependencies
beyond the configuration file and optional GeoIP database.

### Container

Multi-stage Docker build:

1. **Builder**: `golang:<version>-alpine`, static compilation.
2. **Runtime**: `alpine`, non-root user (`mc-proxy`), port exposure
   determined by configuration.

### systemd

| File | Purpose |
|------|---------|
| `mc-proxy.service` | Main proxy service |

The proxy binds to privileged ports (443) and should use `AmbientCapabilities=CAP_NET_BIND_SERVICE`
in the systemd unit rather than running as root.

Standard security hardening directives apply per engineering standards
(`NoNewPrivileges=true`, `ProtectSystem=strict`, etc.).

### Graceful Shutdown

On `SIGINT` or `SIGTERM`:

1. Stop accepting new connections on all listeners.
2. Wait for in-flight connections to complete (up to `shutdown_timeout`).
3. Force-close remaining connections.
4. Exit.

On `SIGHUP`:

1. Reload the GeoIP database from disk.
2. Continue serving with the updated database.

Configuration changes (routes, listeners, firewall rules) require a full
restart. Hot reload of routing rules is deferred to the future SQLite-backed
implementation.

---

## Security Model

mc-proxy is infrastructure that sits in front of authenticated services.
It has no authentication or authorization of its own.

### Threat Mitigations

| Threat | Mitigation |
|--------|------------|
| SNI spoofing | Backend performs its own TLS handshake — a spoofed SNI will fail certificate validation at the backend. mc-proxy does not trust SNI for security decisions beyond routing. |
| Resource exhaustion (connection flood) | Idle timeout closes stale connections. Per-listener connection limits (future). Rate limiting (future). |
| GeoIP evasion via IPv6 | GeoLite2 database includes IPv6 mappings. Both IPv4 and IPv6 source addresses are checked. |
| GeoIP evasion via VPN/proxy | Accepted risk. GeoIP blocking is a compliance measure, not a security boundary. Determined adversaries will bypass it. |
| Slowloris / slow ClientHello | Timeout on the SNI extraction phase. If a complete ClientHello is not received within a reasonable window (e.g. 10s), the connection is reset. |
| Backend unavailability | Connect timeout prevents indefinite hangs. Connection is reset if the backend is unreachable. |
| Information leakage | Blocked connections receive only a TCP RST. No version strings, no error messages, no TLS alerts. |

### Security Invariants

1. mc-proxy never terminates TLS. It cannot read application-layer traffic.
2. mc-proxy never modifies the byte stream between client and backend.
3. Firewall rules are always evaluated before any routing decision.
4. The proxy never logs connection content — only metadata (source IP,
   SNI hostname, backend, timestamps, bytes transferred).

---

## Future Work

Items are listed roughly in priority order:

| Item | Description |
|------|-------------|
| **MCP integration** | Wire the gRPC admin API into the Metacircular Control Plane for centralized management. |
| **L7 HTTPS support** | TLS-terminating mode for selected routes, enabling HTTP-level features (user-agent blocking, header inspection, request routing). |
| **ACME integration** | Automatic certificate provisioning via Let's Encrypt for L7 routes. |
| **User-agent blocking** | Block connections based on user-agent string (requires L7 mode). |
| **Connection rate limiting** | Per-source-IP rate limits to mitigate connection floods. |
| **Per-listener connection limits** | Cap maximum concurrent connections per listener. |
| **Health check endpoint** | Lightweight TCP or HTTP health check for load balancers and monitoring. |
| **Metrics** | Prometheus-compatible metrics: connections per listener, firewall blocks by rule, backend dial latency, active connections. |