@braedonsaunders/appkit-desk 0.7.0 → 0.9.0

package/vz/Sources/DeskNet/NetworkStack.swift

@@ -0,0 +1,547 @@
+import Darwin
+import Dispatch
+
+public struct NetworkConfiguration: Sendable {
+  public var guestAddress: IPv4
+  public var gatewayAddress: IPv4
+  public var prefixLength: Int
+  /// Resolvers handed to the guest over DHCP. The first also answers DNS sent
+  /// to the gateway address, for guests configured to use it.
+  public var dns: [IPv4]
+  /// The guest NIC's address, when known up front; otherwise learned.
+  public var guestMAC: MAC?
+  public var allowLan: Bool
+
+  public init(
+    guestAddress: IPv4 = IPv4(10, 0, 2, 15), gatewayAddress: IPv4 = IPv4(10, 0, 2, 2),
+    prefixLength: Int = 24, dns: [IPv4] = [IPv4(1, 1, 1, 1), IPv4(9, 9, 9, 9)], guestMAC: MAC? = nil,
+    allowLan: Bool = false
+  ) {
+    self.guestAddress = guestAddress
+    self.gatewayAddress = gatewayAddress
+    self.prefixLength = prefixLength
+    self.dns = dns
+    self.guestMAC = guestMAC
+    self.allowLan = allowLan
+  }
+}
+
+public func monotonicMilliseconds() -> UInt64 {
+  DispatchTime.now().uptimeNanoseconds / 1_000_000
+}
+
+/// A NAT between one guest and the Mac's own sockets.
+///
+/// Frames from the guest arrive through `receiveFrame`; frames for the guest
+/// leave through `sendFrame`. In between, the gateway answers ARP and DHCP on
+/// the link, and every TCP connection, UDP exchange and ICMP echo the guest
+/// starts is re-originated from a socket on the Mac — after the fence has
+/// approved its destination. Nothing the guest sends is ever forwarded as a
+/// packet, so nothing it sends can reach a destination the fence refused.
+///
+/// The Mac reaches guest TCP ports through `connectToGuest`, which opens a
+/// connection from the gateway address to the guest.
+@MainActor
+public final class NetworkStack {
+  public let configuration: NetworkConfiguration
+  public var onGuestAddress: ((IPv4) -> Void)?
+  public let now: () -> UInt64
+
+  private let dialer: Dialer
+  private let fence: EgressFence
+  private let hostAddresses: () -> Set<IPv4>
+  private let sendFrame: ([UInt8]) -> Void
+  private let dhcp: DHCPServer
+  private var guestMAC: MAC?
+  private var announcedGuest = false
+  private var tcpFlows: [FlowKey: TCPFlow] = [:]
+  private var udpFlows: [FlowKey: UDPFlow] = [:]
+  private var echoFlows: [EchoKey: EchoFlow] = [:]
+  private var nextEphemeralPort: UInt16 = 49_152
+  private var timer: DispatchSourceTimer?
+  private var identification: UInt16 = 0
+
+  static let maximumTCPFlows = 4_096
+  static let maximumUDPFlows = 1_024
+  static let maximumEchoFlows = 64
+
+  public init(
+    configuration: NetworkConfiguration, dialer: Dialer = SystemDialer(),
+    now: @escaping () -> UInt64 = monotonicMilliseconds,
+    hostAddresses: @escaping () -> Set<IPv4> = Socket.hostAddresses,
+    sendFrame: @escaping ([UInt8]) -> Void
+  ) {
+    self.configuration = configuration
+    self.dialer = dialer
+    self.now = now
+    self.hostAddresses = hostAddresses
+    self.sendFrame = sendFrame
+    fence = EgressFence(allowLan: configuration.allowLan)
+    guestMAC = configuration.guestMAC
+    dhcp = DHCPServer(
+      guest: configuration.guestAddress, gateway: configuration.gatewayAddress,
+      prefixLength: configuration.prefixLength, dns: configuration.dns)
+  }
+
+  public var flowCount: Int { tcpFlows.count + udpFlows.count + echoFlows.count }
+
+  // MARK: Inbound frames
+
+  public func receiveFrame(_ bytes: [UInt8]) {
+    guard let frame = EthernetFrame(parsing: bytes) else { return }
+    guard frame.destination == .gateway || frame.destination.isBroadcast else { return }
+    if guestMAC == nil { guestMAC = frame.source }
+    switch frame.type {
+    case EtherType.arp:
+      if let arp = ARPPacket(parsing: frame.payload) { receiveARP(arp, from: frame.source) }
+    case EtherType.ipv4:
+      if let packet = IPv4Packet(parsing: frame.payload) { receiveIPv4(packet, from: frame.source) }
+    default:
+      // IPv6 and everything else stop here: the guest gets no route for it.
+      break
+    }
+  }
+
+  private func receiveARP(_ arp: ARPPacket, from source: MAC) {
+    guard arp.operation == ARPPacket.request else { return }
+    // Answer for every address but the guest's own. The guest then sends all
+    // its traffic to the gateway whether it thinks a destination is on-link
+    // or not, and the fence judges it there. Probes from 0.0.0.0 are the
+    // guest checking its own address is free, and must go unanswered.
+    guard arp.senderIP != .zero, arp.targetIP != arp.senderIP,
+      arp.targetIP != configuration.guestAddress
+    else { return }
+    let reply = ARPPacket(
+      operation: ARPPacket.reply, senderMAC: .gateway, senderIP: arp.targetIP,
+      targetMAC: arp.senderMAC, targetIP: arp.senderIP)
+    sendFrame(EthernetFrame(destination: source, source: .gateway, type: EtherType.arp, payload: reply.encoded).encoded)
+  }
+
+  private func receiveIPv4(_ packet: IPv4Packet, from source: MAC) {
+    if packet.proto == IPProtocol.udp, let udp = UDPDatagram(parsing: packet.payload),
+      udp.destinationPort == 67, udp.sourcePort == 68
+    {
+      receiveDHCP(udp, from: source)
+      return
+    }
+    // One guest, one address: anything else is spoofed or misconfigured.
+    guard packet.source == configuration.guestAddress, !packet.isFragment else { return }
+    announceGuest()
+    let destination = packet.destination
+    if destination == configuration.gatewayAddress {
+      receiveForGateway(packet)
+      return
+    }
+    switch fence.verdict(for: destination, hostAddresses: hostAddresses()) {
+    case .allowed:
+      break
+    case .invalid:
+      return
+    case .fenced:
+      sendUnreachable(ICMPMessage.administrativelyProhibited, quoting: packet)
+      return
+    }
+    switch packet.proto {
+    case IPProtocol.tcp:
+      if let segment = TCPSegment(parsing: packet.payload) { receiveTCP(segment, packet: packet) }
+    case IPProtocol.udp:
+      if let datagram = UDPDatagram(parsing: packet.payload) { receiveUDP(datagram, packet: packet, upstream: destination) }
+    case IPProtocol.icmp:
+      if let message = ICMPMessage(parsing: packet.payload), message.type == ICMPMessage.echoRequest {
+        receiveEcho(message, destination: destination)
+      }
+    default:
+      sendUnreachable(ICMPMessage.administrativelyProhibited, quoting: packet)
+    }
+  }
+
+  private func announceGuest() {
+    guard !announcedGuest else { return }
+    announcedGuest = true
+    onGuestAddress?(configuration.guestAddress)
+  }
+
+  private func receiveDHCP(_ datagram: UDPDatagram, from source: MAC) {
+    guard let reply = dhcp.respond(to: datagram.payload) else { return }
+    let udp = UDPDatagram(sourcePort: 67, destinationPort: 68, payload: reply.payload)
+    let ip = IPv4Packet.encode(
+      source: configuration.gatewayAddress, destination: .broadcast, proto: IPProtocol.udp,
+      payload: udp.encoded(source: configuration.gatewayAddress, destination: .broadcast),
+      identification: nextIdentification())
+    sendFrame(EthernetFrame(destination: reply.clientMAC, source: .gateway, type: EtherType.ipv4, payload: ip).encoded)
+    if reply.assigned { announceGuest() }
+  }
+
+  /// The gateway is the Mac's face on the link, so it offers nothing but a
+  /// ping reply and a DNS forwarder: no port on it reaches a Mac service.
+  private func receiveForGateway(_ packet: IPv4Packet) {
+    switch packet.proto {
+    case IPProtocol.icmp:
+      guard let message = ICMPMessage(parsing: packet.payload), message.type == ICMPMessage.echoRequest else { return }
+      let reply = ICMPMessage(type: ICMPMessage.echoReply, code: 0, rest: message.rest, payload: message.payload)
+      sendIPv4(reply.encoded, proto: IPProtocol.icmp, from: configuration.gatewayAddress)
+    case IPProtocol.tcp:
+      if let segment = TCPSegment(parsing: packet.payload) {
+        let key = FlowKey(guestPort: segment.sourcePort, remote: packet.destination, remotePort: segment.destinationPort)
+        if let flow = tcpFlows[key] {
+          flow.receive(segment)
+        } else {
+          sendReset(answering: segment, key: key)
+        }
+      }
+    case IPProtocol.udp:
+      guard let datagram = UDPDatagram(parsing: packet.payload) else { return }
+      if datagram.destinationPort == 53, let resolver = configuration.dns.first {
+        guard fence.verdict(for: resolver, hostAddresses: hostAddresses()) == .allowed else {
+          sendUnreachable(ICMPMessage.administrativelyProhibited, quoting: packet)
+          return
+        }
+        receiveUDP(datagram, packet: packet, upstream: resolver)
+      } else {
+        sendUnreachable(ICMPMessage.portUnreachable, quoting: packet)
+      }
+    default:
+      return
+    }
+  }
+
+  // MARK: TCP
+
+  private func receiveTCP(_ segment: TCPSegment, packet: IPv4Packet) {
+    let key = FlowKey(guestPort: segment.sourcePort, remote: packet.destination, remotePort: segment.destinationPort)
+    if let flow = tcpFlows[key] {
+      flow.receive(segment)
+      return
+    }
+    guard segment.flags.contains(.syn), !segment.flags.contains(.ack), !segment.flags.contains(.rst) else {
+      sendReset(answering: segment, key: key)
+      return
+    }
+    guard tcpFlows.count < Self.maximumTCPFlows else {
+      sendReset(answering: segment, key: key)
+      return
+    }
+    let fd: Int32
+    do {
+      fd = try dialer.tcp(to: packet.destination, port: segment.destinationPort)
+    } catch let failure as POSIXFailure {
+      reportUpstreamFailure(failure.code, key: key, quoting: packet, segment: segment)
+      return
+    } catch {
+      sendUnreachable(ICMPMessage.hostUnreachable, quoting: packet)
+      return
+    }
+    let flow = TCPFlow(stack: self, key: key, guest: configuration.guestAddress, syn: segment, hostFD: fd)
+    pendingSYNs[ObjectIdentifier(flow)] = packet
+    tcpFlows[key] = flow
+    startTimer()
+  }
+
+  /// The SYN each connecting flow came from, for quoting in an ICMP error.
+  private var pendingSYNs: [ObjectIdentifier: IPv4Packet] = [:]
+
+  func upstreamFailed(_ flow: TCPFlow, errno code: Int32) {
+    guard let packet = pendingSYNs.removeValue(forKey: ObjectIdentifier(flow)),
+      let segment = TCPSegment(parsing: packet.payload)
+    else { return }
+    reportUpstreamFailure(code, key: flow.key, quoting: packet, segment: segment)
+  }
+
+  private func reportUpstreamFailure(_ code: Int32, key: FlowKey, quoting packet: IPv4Packet, segment: TCPSegment) {
+    if code == ECONNREFUSED {
+      sendReset(answering: segment, key: key)
+    } else {
+      sendUnreachable(ICMPMessage.hostUnreachable, quoting: packet)
+    }
+  }
+
+  func sendReset(answering segment: TCPSegment, key: FlowKey) {
+    guard !segment.flags.contains(.rst) else { return }
+    let reset: TCPSegment
+    if segment.flags.contains(.ack) {
+      reset = TCPSegment(
+        sourcePort: key.remotePort, destinationPort: key.guestPort, sequence: segment.acknowledgement,
+        acknowledgement: 0, flags: [.rst], window: 0)
+    } else {
+      reset = TCPSegment(
+        sourcePort: key.remotePort, destinationPort: key.guestPort, sequence: 0,
+        acknowledgement: segment.sequence &+ segment.sequenceLength, flags: [.rst, .ack], window: 0)
+    }
+    sendTCP(reset, from: key.remote, to: configuration.guestAddress)
+  }
+
+  func sendTCP(_ segment: TCPSegment, from source: IPv4, to destination: IPv4) {
+    sendIPv4(segment.encoded(source: source, destination: destination), proto: IPProtocol.tcp, from: source)
+  }
+
+  func flowClosed(_ flow: TCPFlow) {
+    pendingSYNs.removeValue(forKey: ObjectIdentifier(flow))
+    if tcpFlows[flow.key] === flow { tcpFlows.removeValue(forKey: flow.key) }
+  }
+
+  /// Open a TCP connection from the gateway to `port` on the guest, relayed to
+  /// `clientFD`. Once the guest accepts, `OK <local port>\n` is written to the
+  /// client and the descriptor carries the connection; a refusal or a guest
+  /// that never answers closes it with nothing written. `early` is client data
+  /// that arrived with the request and is delivered first.
+  public func connectToGuest(port: UInt16, clientFD: Int32, early: [UInt8] = []) {
+    guard tcpFlows.count < Self.maximumTCPFlows, let localPort = allocateEphemeralPort() else {
+      Darwin.close(clientFD)
+      return
+    }
+    let key = FlowKey(guestPort: port, remote: configuration.gatewayAddress, remotePort: localPort)
+    Socket.setNonBlocking(clientFD)
+    Socket.setNoSigPipe(clientFD)
+    let flow = TCPFlow(
+      stack: self, key: key, guest: configuration.guestAddress, clientFD: clientFD,
+      ready: Array("OK \(localPort)\n".utf8), early: early)
+    if flow.state != .closed { tcpFlows[key] = flow }
+    startTimer()
+  }
+
+  private func allocateEphemeralPort() -> UInt16? {
+    for _ in 0..<16_384 {
+      let port = nextEphemeralPort
+      nextEphemeralPort = nextEphemeralPort == 65_535 ? 49_152 : nextEphemeralPort + 1
+      let inUse = tcpFlows.keys.contains { $0.remote == configuration.gatewayAddress && $0.remotePort == port }
+      if !inUse { return port }
+    }
+    return nil
+  }
+
+  // MARK: UDP
+
+  private func receiveUDP(_ datagram: UDPDatagram, packet: IPv4Packet, upstream: IPv4) {
+    let key = FlowKey(guestPort: datagram.sourcePort, remote: packet.destination, remotePort: datagram.destinationPort)
+    let flow: UDPFlow
+    if let existing = udpFlows[key] {
+      flow = existing
+    } else {
+      if udpFlows.count >= Self.maximumUDPFlows, let oldest = udpFlows.values.min(by: { $0.lastActive < $1.lastActive }) {
+        oldest.close()
+      }
+      let fd: Int32
+      do {
+        fd = try dialer.udp(to: upstream, port: datagram.destinationPort)
+      } catch {
+        sendUnreachable(ICMPMessage.hostUnreachable, quoting: packet)
+        return
+      }
+      flow = UDPFlow(stack: self, key: key, fd: fd)
+      udpFlows[key] = flow
+      startTimer()
+    }
+    flow.send(datagram.payload, quoting: packet)
+  }
+
+  func sendUDP(_ payload: [UInt8], key: FlowKey) {
+    let datagram = UDPDatagram(sourcePort: key.remotePort, destinationPort: key.guestPort, payload: payload)
+    sendIPv4(
+      datagram.encoded(source: key.remote, destination: configuration.guestAddress), proto: IPProtocol.udp,
+      from: key.remote)
+  }
+
+  func udpFlowClosed(_ flow: UDPFlow) {
+    if udpFlows[flow.key] === flow { udpFlows.removeValue(forKey: flow.key) }
+  }
+
+  // MARK: ICMP
+
+  private func receiveEcho(_ message: ICMPMessage, destination: IPv4) {
+    let key = EchoKey(remote: destination, identifier: message.identifier)
+    let flow: EchoFlow
+    if let existing = echoFlows[key] {
+      flow = existing
+    } else {
+      guard echoFlows.count < Self.maximumEchoFlows else { return }
+      let fd: Int32
+      do {
+        fd = try dialer.icmp()
+      } catch {
+        return
+      }
+      flow = EchoFlow(stack: self, key: key, fd: fd)
+      echoFlows[key] = flow
+      startTimer()
+    }
+    flow.send(message)
+  }
+
+  func sendEchoReply(_ message: ICMPMessage, key: EchoKey) {
+    sendIPv4(message.encoded, proto: IPProtocol.icmp, from: key.remote)
+  }
+
+  func echoFlowClosed(_ flow: EchoFlow) {
+    if echoFlows[flow.key] === flow { echoFlows.removeValue(forKey: flow.key) }
+  }
+
+  func sendUnreachable(_ code: UInt8, quoting packet: IPv4Packet) {
+    let quoted = packet.header + Array(packet.payload.prefix(8))
+    let message = ICMPMessage(type: ICMPMessage.destinationUnreachable, code: code, rest: 0, payload: quoted)
+    sendIPv4(message.encoded, proto: IPProtocol.icmp, from: configuration.gatewayAddress)
+  }
+
+  // MARK: Output
+
+  private func nextIdentification() -> UInt16 {
+    identification &+= 1
+    return identification
+  }
+
+  private func sendIPv4(_ payload: [UInt8], proto: UInt8, from source: IPv4) {
+    let packet = IPv4Packet.encode(
+      source: source, destination: configuration.guestAddress, proto: proto, payload: payload,
+      identification: nextIdentification())
+    let destination = guestMAC ?? .broadcast
+    sendFrame(EthernetFrame(destination: destination, source: .gateway, type: EtherType.ipv4, payload: packet).encoded)
+  }
+
+  // MARK: Timers
+
+  /// Retransmissions and idle expiry. Runs only while there are flows, so an
+  /// idle guest costs the Mac no wakeups.
+  public func tick() {
+    let current = now()
+    for flow in Array(tcpFlows.values) { flow.tick(now: current) }
+    for flow in Array(udpFlows.values) where flow.expired(now: current) { flow.close() }
+    for flow in Array(echoFlows.values) where flow.expired(now: current) { flow.close() }
+    if flowCount == 0 { stopTimer() }
+  }
+
+  private func startTimer() {
+    guard timer == nil else { return }
+    let source = DispatchSource.makeTimerSource(queue: .main)
+    source.schedule(deadline: .now() + .milliseconds(50), repeating: .milliseconds(50), leeway: .milliseconds(10))
+    source.setEventHandler { @Sendable [weak self] in Task { @MainActor in self?.tick() } }
+    source.resume()
+    timer = source
+  }
+
+  private func stopTimer() {
+    timer?.cancel()
+    timer = nil
+  }
+
+  /// Drop every connection, resetting the host side of each.
+  public func shutdown() {
+    for flow in Array(tcpFlows.values) { flow.abort() }
+    for flow in Array(udpFlows.values) { flow.close() }
+    for flow in Array(echoFlows.values) { flow.close() }
+    stopTimer()
+  }
+}
+
+// MARK: - UDP and ICMP flows
+
+@MainActor
+final class UDPFlow {
+  let key: FlowKey
+  private(set) var lastActive: UInt64
+  private unowned let stack: NetworkStack
+  private let socket: FDWatcher
+  private var lastQuote: IPv4Packet?
+
+  init(stack: NetworkStack, key: FlowKey, fd: Int32) {
+    self.stack = stack
+    self.key = key
+    socket = FDWatcher(fd: fd)
+    lastActive = stack.now()
+    socket.onReadable = { [weak self] in self?.readable() }
+    socket.wantRead(true)
+  }
+
+  /// DNS answers come back at once or not at all; anything else may be a
+  /// long-lived exchange such as QUIC.
+  private var idleLimit: UInt64 { key.remotePort == 53 ? 15_000 : 120_000 }
+
+  func expired(now: UInt64) -> Bool { now - lastActive > idleLimit }
+
+  func send(_ payload: [UInt8], quoting packet: IPv4Packet) {
+    lastActive = stack.now()
+    lastQuote = packet
+    _ = payload.withUnsafeBytes { Darwin.send(socket.fd, $0.baseAddress, payload.count, 0) }
+  }
+
+  private func readable() {
+    var buffer = [UInt8](repeating: 0, count: 65_536)
+    while true {
+      let count = buffer.withUnsafeMutableBytes { Darwin.recv(socket.fd, $0.baseAddress, $0.count, 0) }
+      if count < 0 {
+        if errno == ECONNREFUSED, let quote = lastQuote {
+          stack.sendUnreachable(ICMPMessage.portUnreachable, quoting: quote)
+          continue
+        }
+        return
+      }
+      lastActive = stack.now()
+      stack.sendUDP(Array(buffer[0..<count]), key: key)
+    }
+  }
+
+  func close() {
+    socket.close()
+    stack.udpFlowClosed(self)
+  }
+}
+
+struct EchoKey: Hashable {
+  let remote: IPv4
+  let identifier: UInt16
+}
+
+@MainActor
+final class EchoFlow {
+  let key: EchoKey
+  private var lastActive: UInt64
+  private unowned let stack: NetworkStack
+  private let socket: FDWatcher
+
+  init(stack: NetworkStack, key: EchoKey, fd: Int32) {
+    self.stack = stack
+    self.key = key
+    socket = FDWatcher(fd: fd)
+    lastActive = stack.now()
+    socket.onReadable = { [weak self] in self?.readable() }
+    socket.wantRead(true)
+  }
+
+  func expired(now: UInt64) -> Bool { now - lastActive > 10_000 }
+
+  func send(_ message: ICMPMessage) {
+    lastActive = stack.now()
+    let bytes = message.encoded
+    var address = Socket.address(key.remote, 0)
+    _ = withUnsafePointer(to: &address) { pointer in
+      pointer.withMemoryRebound(to: sockaddr.self, capacity: 1) { sockaddrPointer in
+        bytes.withUnsafeBytes {
+          Darwin.sendto(socket.fd, $0.baseAddress, bytes.count, 0, sockaddrPointer, socklen_t(MemoryLayout<sockaddr_in>.size))
+        }
+      }
+    }
+  }
+
+  private func readable() {
+    var buffer = [UInt8](repeating: 0, count: 65_536)
+    while true {
+      let count = buffer.withUnsafeMutableBytes { Darwin.recv(socket.fd, $0.baseAddress, $0.count, 0) }
+      if count <= 0 { return }
+      var bytes = Array(buffer[0..<count])
+      // Darwin hands datagram ICMP sockets the IP header too.
+      if bytes.count >= 20, bytes[0] >> 4 == 4 {
+        let headerLength = Int(bytes[0] & 0x0f) * 4
+        guard bytes.count >= headerLength + 8 else { continue }
+        bytes = Array(bytes[headerLength...])
+      }
+      guard let reply = ICMPMessage(parsing: bytes), reply.type == ICMPMessage.echoReply else { continue }
+      lastActive = stack.now()
+      // The kernel may have chosen its own identifier; the guest gets its own back.
+      let rest = UInt32(key.identifier) << 16 | UInt32(reply.sequenceNumber)
+      stack.sendEchoReply(ICMPMessage(type: ICMPMessage.echoReply, code: 0, rest: rest, payload: reply.payload), key: key)
+    }
+  }
+
+  func close() {
+    socket.close()
+    stack.echoFlowClosed(self)
+  }
+}