Files
RouterOS/RouterOSAssistant/Core/Networking/SSHTransport.swift
T
KayandClaude Sonnet 5 cae59db39e Fix Haenger beim Anlegen eines DHCP-Pools (bugs.md #11)
SSHTransport.connect()/run() hatten keinen Timeout - im Gegensatz zu
RestTransport (timeoutInterval=5). Jeder erste Schreibvorgang einer
Session loest ueber ensureSessionBackup() eine dedizierte SSH-
Verbindung aus; haengt die, blieb isApplying unbegrenzt aktiv (kein
Fehler, kein Recovery, nur Force-Quit). Live vom Nutzer bestaetigt
(Experte-Tab, dauerhaft haengend) bevor der Fix geschrieben wurde.

Neuer genererischer SSHTransport.withTimeout(_:operation:) (Task-
Group-Race gegen eine Deadline), angewendet auf connect() (10s) und
run() (30s). 3 neue Regressionstests fuer die Race-Logik isoliert.

Nebenbefund: xcodegen generate ueberschreibt Info.plist komplett aus
project.yml (kein Merge) - ein Regenerieren fuer die neue Testdatei
setzte die Version stillschweigend von 1.1.0 auf 1.0 zurueck. Version
jetzt explizit in project.yml verankert.

Build + alle 106 Unit-Tests gruen.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
2026-09-17 21:51:11 +02:00

420 lines
25 KiB
Swift

import Foundation
import Citadel
import NIOCore
import NIOSSH
/// SSH+CLI transport fallback for RouterOS devices/firmware without the REST API (pre-7.1).
final class SSHTransport: RouterOSTransport {
let kind: RouterOSTransportKind = .ssh
private let credentials: RouterOSCredentials
private let hostKeyTrust: SSHHostKeyTrustStore
private var client: SSHClient?
init(credentials: RouterOSCredentials, hostKeyTrust: SSHHostKeyTrustStore = SSHHostKeyTrustStore()) {
self.credentials = credentials
self.hostKeyTrust = hostKeyTrust
}
/// Unlike `RestTransport` (explicit `request.timeoutInterval = 5` on every request), neither
/// this nor `run(_:)` used to bound how long they'd wait at all — `Citadel.SSHClient.connect`
/// has no built-in timeout. Live-confirmed as a real, reproducible bug (bugs.md #11,
/// 2026-09-17): creating an `/ip pool` entry in the Experte tab hung the app permanently (no
/// error, no recovery, force-quit needed) — traced to this connect call, reached via
/// `ExpertViewModel.saveEditingItem()`'s mandatory `ensureSessionBackup()`, which opens a
/// fresh, dedicated SSH connection (`BackupService`) before every session's first write. A
/// connection attempt that stalls (transient network hiccup, or the hAP-lite test router
/// itself being slow under load — MIPS 24Kc/650MHz/1 core, seen at 80% CPU this session) had
/// no way to ever resolve, so `isApplying` never cleared. `withTimeout` below races the real
/// operation against a deadline and cancels whichever loses.
private static let connectTimeout: Duration = .seconds(10)
private static let commandTimeout: Duration = .seconds(30)
func connect() async throws {
do {
client = try await Self.withTimeout(Self.connectTimeout) {
try await SSHClient.connect(
host: self.credentials.host,
port: self.credentials.sshPort,
authenticationMethod: .passwordBased(username: self.credentials.username, password: self.credentials.password),
hostKeyValidator: .custom(self),
reconnect: .never,
// RouterOS' SSH server typically only offers legacy algorithms
// (diffie-hellman-group14-sha1 key exchange, RSA host keys) that
// Citadel's defaults don't include — `.all` adds them.
algorithms: .all
)
}
} catch let error as RouterOSError {
throw error
} catch {
// NIOSSHError's `.localizedDescription` is a useless generic NSError-bridged string
// ("The operation couldn't be completed."); its real diagnostics only surface via
// CustomStringConvertible, which `String(describing:)` picks up.
throw RouterOSError.transportUnavailable("SSH-Verbindung fehlgeschlagen: \(String(describing: error))")
}
}
/// Races `operation` against `duration`, cancelling whichever loses — the generic mechanism
/// behind both `connect()`'s and `run(_:)`'s timeouts (bugs.md #11). `operation` must be
/// `@Sendable`: it runs inside a detached task group child, not on the caller's isolation.
static func withTimeout<T: Sendable>(_ duration: Duration, operation: @escaping @Sendable () async throws -> T) async throws -> T {
try await withThrowingTaskGroup(of: T.self) { group in
group.addTask { try await operation() }
group.addTask {
try await Task.sleep(for: duration)
throw RouterOSError.transportUnavailable("Zeitüberschreitung (\(Int(duration.components.seconds))s) — Router antwortet nicht.")
}
defer { group.cancelAll() }
return try await group.next()!
}
}
func fetchDeviceInfo() async throws -> RouterDeviceInfo {
let output = try await run("/system resource print without-paging")
return RouterOSCliParser.parseDeviceInfo(output)
}
func fetchInterfaces() async throws -> [NetworkInterface] {
let output = try await run("/interface print without-paging terse")
return RouterOSCliParser.parseInterfaces(output)
}
/// Live throughput for one interface — RouterOS' `/interface monitor-traffic <name> once` is
/// a standard, stable CLI command (documented single-interface usage), not a menu item, so
/// it's implemented directly here rather than through the generic `fetchMenuItems` machinery.
/// Not exposed over REST: this is a CLI-only command with no documented REST equivalent, so
/// callers needing it (see `InterfaceTrafficMonitor`) always use a dedicated SSH connection,
/// same reasoning as `BackupService`/`UpdateService`.
/// Runs an arbitrary RouterOS CLI command and returns its raw text output — the one
/// deliberate escape hatch out of the otherwise-private `run(_:)`, for `NetworkToolsService`'s
/// on-demand diagnostics (ping/traceroute/DNS lookup have no menu-item/REST shape to go
/// through the generic `fetchMenuItems` machinery). Callers are responsible for sanitizing any
/// untrusted value (e.g. a DHCP-supplied hostname) before interpolating it into `command` —
/// RouterOS' console treats ";" as a command separator, so an unsanitized value could inject
/// a second command.
func runDiagnosticCommand(_ command: String) async throws -> String {
try await run(command)
}
func fetchInterfaceTraffic(interfaceName: String) async throws -> InterfaceTraffic {
let output = try await run("/interface monitor-traffic \(interfaceName) once")
let fields = RouterOSCliParser.parseSingletonItem(output).fields
return InterfaceTraffic(
rxBitsPerSecond: Self.parseBitsPerSecond(fields["rx-bits-per-second"]),
txBitsPerSecond: Self.parseBitsPerSecond(fields["tx-bits-per-second"])
)
}
/// RouterOS reports these as human-formatted strings with a unit suffix (e.g. "50.7kbps",
/// "34.0kbps", or plain "0" when idle), never a bare integer — confirmed live (2026-09-15,
/// hEX/RouterOS 7.x). Longer/more specific suffixes are checked before shorter ones ("kbps"
/// before "bps") since "kbps" itself ends with "bps" too.
static func parseBitsPerSecond(_ raw: String?) -> Int {
guard let trimmed = raw?.trimmingCharacters(in: .whitespaces), !trimmed.isEmpty else { return 0 }
let unitsBySpecificity: [(suffix: String, multiplier: Double)] = [
("Gbps", 1_000_000_000), ("Mbps", 1_000_000), ("kbps", 1_000), ("bps", 1)
]
for unit in unitsBySpecificity where trimmed.hasSuffix(unit.suffix) {
let numberPart = trimmed.dropLast(unit.suffix.count)
if let value = Double(numberPart) {
return Int(value * unit.multiplier)
}
}
return Int(trimmed) ?? 0
}
func disconnect() async {
try? await client?.close()
client = nil
}
/// Full human-readable config export (`/export terse`), used for local backups.
func exportConfiguration() async throws -> String {
try await run("/export terse")
}
func fetchFirewallRuleCounts() async throws -> FirewallRuleCounts {
let filterOutput = try await run("/ip firewall filter print count-only")
let natOutput = try await run("/ip firewall nat print count-only")
let filterCount = Int(filterOutput.trimmingCharacters(in: .whitespacesAndNewlines)) ?? 0
let natCount = Int(natOutput.trimmingCharacters(in: .whitespacesAndNewlines)) ?? 0
return FirewallRuleCounts(filterRuleCount: filterCount, natRuleCount: natCount)
}
/// RouterOS 7.24.2 (verified live against a hEX test device) does *not* include `.id` in
/// `print terse` CLI output, unlike REST's JSON, which always does — every field parses
/// fine, but every item's `id` would otherwise fall back to a synthetic, unusable "row-N".
/// `:put [<menuPath> find]` returns the real internal IDs (e.g. "*C;*1;*2") in the same
/// order as `print terse` (also verified live: position 0 → first ID, position 1 → second),
/// so they're overlaid onto the parsed items by position here.
func fetchMenuItems(menuPath: String, restPath: String) async throws -> [RouterOSMenuItem] {
let output: String
do {
output = try await run("\(menuPath) print without-paging terse")
} catch RouterOSError.invalidResponse(let detail) where Self.rejectsTerse(detail) {
// Singleton menu (e.g. "/ip dns", "/system identity") — no list, no "terse" support.
let plain = try await run("\(menuPath) print without-paging")
return [RouterOSCliParser.parseSingletonItem(plain)]
}
// Same singleton rejection, but confirmed live (for "/system routerboard") that RouterOS
// doesn't always report it as a command failure Citadel would throw on — it can come
// back as plain output with exit code 0 instead, exactly the "exit code isn't reliable"
// lesson from Bug 10 applying to a *read* command here, not just add/set/remove. Without
// this, such a menu's data silently comes back empty rather than throwing or falling
// back — no crash, no error, just nothing, which is worse than either.
if Self.rejectsTerse(output) {
let plain = try await run("\(menuPath) print without-paging")
return [RouterOSCliParser.parseSingletonItem(plain)]
}
var items = RouterOSCliParser.parseGenericItems(output)
let idOutput = try await run(":put [\(menuPath) find]")
let ids = idOutput
.trimmingCharacters(in: .whitespacesAndNewlines)
.split(separator: ";")
.map(String.init)
for index in items.indices where index < ids.count {
items[index] = RouterOSMenuItem(id: ids[index], fields: items[index].fields)
}
return items
}
/// RouterOS rejects `terse` on a singleton menu in more than one wording, confirmed live
/// across two different devices/RouterOS versions (2026-09-16): "bad parameter terse"
/// (original hEX test device) and a harder parser error, "expected end of command (line 1
/// column N)" (a second, different router) — both were seen on `/system routerboard`, the
/// second one also on `/system package update`, silently leaving Routerboard info (including
/// the serial number "Bekannte Router" needs to tell two same-address devices apart) and the
/// software-update check both empty with no visible error. Scoped safely to this one call
/// site: `detail`/`output` here only ever come from running `"<menuPath> print
/// without-paging terse"`, so any parser error on that exact line is — by construction —
/// about the trailing "terse" token, not some unrelated syntax problem elsewhere.
private static func rejectsTerse(_ text: String) -> Bool {
text.contains("bad parameter terse") || text.contains("expected end of command")
}
/// `:foreach i in=[<menuPath> find whereField=whereValue] do={:put [<menuPath> get $i
/// returnField]}` — built from two independently confirmed-live primitives only: a bare
/// `find` with one condition (verified repeatedly this session, e.g.
/// `find dynamic=no` reliably returning exactly the right id), and `get <id> field` on a
/// single, already-known id (standard, unambiguous RouterOS syntax). An earlier version tried
/// `get [find ...] returnField` as one combined call to do this in a single round trip —
/// that specific combined form was never actually verified and was confirmed live to be
/// wrong (a real make-static conversion — confirmed via Winbox — wasn't found by it). `:put`
/// inside `:foreach` prints one value per line, so this splits on newlines, not ";" (the
/// semicolon-joined shape only applies to a single `:put [<menuPath> find ...]` list).
/// `whereValue` is quoted/escaped the same way as `RouterOSCommand`'s CLI rendering (see its
/// `quoteIfNeeded` doc comment for the live-confirmed injection this prevents) — both current
/// callers only ever pass the hardcoded literal `"no"`, but this is a generic, reusable
/// `RouterOSTransport` method, so a future caller passing user/device-controlled text (e.g. a
/// hostname) must not silently reopen the same command-injection class this app has already
/// had to fix once.
func fetchFieldValues(menuPath: String, restPath: String, whereField: String, whereValue: String, returnField: String) async throws -> Set<String> {
let escapedValue = whereValue
.replacingOccurrences(of: "\\", with: "\\\\")
.replacingOccurrences(of: "\"", with: "\\\"")
let script = ":foreach i in=[\(menuPath) find \(whereField)=\"\(escapedValue)\"] do={:put [\(menuPath) get $i \(returnField)]}"
let output = try await run(script)
let values = output
.split(whereSeparator: \.isNewline)
.map { $0.trimmingCharacters(in: .whitespaces) }
.filter { !$0.isEmpty }
return Set(values)
}
/// RouterOS' SSH CLI exits 0 even when a command fails — confirmed live: both
/// `/ip dhcp-server add ...` on an interface that already has one ("failure: server or
/// relay with such interface already exists") and an invalid action ("syntax error (line 1
/// column 46)") returned exit status 0, meaning `run()`'s exit-code check alone silently
/// treats every such failure as success. A mutating command (add/set/remove) is always
/// silent on success in every case observed live this session (15+ menu families) — so any
/// non-empty output here is treated as the error text, since there is no more reliable
/// signal available over this transport.
func apply(_ command: RouterOSCommand) async throws {
let output = try await run(command.cliLine)
let trimmed = output.trimmingCharacters(in: .whitespacesAndNewlines)
guard trimmed.isEmpty else {
throw RouterOSError.invalidResponse(trimmed)
}
}
/// Forcibly ends already-tracked connections between two networks (bugs.md #7 / Gitea #19) —
/// RouterOS' firewall rules only affect *new* connections, so a connection already open
/// between two networks at the moment they're marked isolated would otherwise keep flowing
/// through the pre-existing "forward established/related -> accept" rule indefinitely.
/// `/ip firewall connection` has no interface field, only address fields, so this matches by
/// CIDR-range membership instead. No REST equivalent exists for this CIDR-membership query
/// language (REST mirrors CRUD over menu paths, not scripting), so — same reasoning as
/// `BackupService`/`NetworkToolsService`/`UpdateService` — this always runs over SSH.
///
/// **Verification status, honestly (2026-09-17):** `print count-only where src-address in
/// <cidr>` is live-confirmed to filter by real subnet containment (40 matches for a
/// populated /24 vs. 0 for an empty one, and `remove [find where dst-address=<exact-ip>]` —
/// single equality condition, no `in` — is live-confirmed to actually delete an entry (a
/// disposable ICMP ping's tracked connection genuinely disappeared). The compound
/// `remove [find where (src-address in A) and (dst-address in B)]` form this function
/// actually uses could NOT be cleanly proven live: a first attempt looked successful, but
/// that was later traced to ICMP conntrack's very short natural timeout (a few seconds)
/// coinciding with the multi-second gap between separate manual SSH round-trips, not the
/// `remove` itself. A cleaner retest against a still-*active* long-lived TCP connection (an
/// open SSH session) showed the tracked entry reappearing immediately after `remove` — which
/// is expected/correct behavior for connection tracking in general (removing the tracking
/// state doesn't RST the socket; the very next packet on an actively-flowing connection just
/// gets re-tracked as "new"), not proof the `remove` itself is a no-op, but this app has no
/// two genuinely separate test networks available to observe the one behavior that actually
/// matters here: whether a fresh isolation drop rule catches that re-tracked "new" packet
/// instead of silently re-admitting it. The `remove [find where ... in ...]` compound-CIDR
/// technique itself is real and community-documented (MikroTik forum), just not end-to-end
/// live-verified against this app's specific isolation scenario. Best-effort by design either
/// way (see `ConnectionService.flushConnections`'s caller in `SetupViewModel`) — a live
/// multi-network test is the natural next verification step.
///
/// `networkA`/`networkB` are interpolated directly into the script, so they're validated as
/// plain CIDR notation first (digits/dots/slash only) — an unvalidated value here would
/// reopen the exact command-injection class already fixed once in
/// `RouterOSCommand.quoteIfNeeded` (bugs.md #5). Both directions are removed since either
/// isolated network could be the connection's source or destination.
func flushConnections(between networkA: String, and networkB: String) async throws {
guard Self.isPlainCIDR(networkA), Self.isPlainCIDR(networkB) else {
throw RouterOSError.invalidResponse("Ungültiges Netzwerkformat: \(networkA) / \(networkB)")
}
let script = """
/ip firewall connection remove [find where (src-address in \(networkA)) and (dst-address in \(networkB))]
/ip firewall connection remove [find where (src-address in \(networkB)) and (dst-address in \(networkA))]
"""
_ = try await run(script)
}
private static func isPlainCIDR(_ value: String) -> Bool {
let allowed = CharacterSet(charactersIn: "0123456789./")
return !value.isEmpty && value.unicodeScalars.allSatisfy(allowed.contains)
}
/// Restores RouterOS' own vendor-default configuration and reboots the device. See
/// FactoryResetService for why this bypasses the RouterOSCommand add/set model entirely.
func resetToFactoryDefaults() async throws {
_ = try await run("/system reset-configuration no-defaults=no skip-backup=no")
}
/// Uploads a `.rsc` script's text content to the router's own file storage via SFTP — the
/// only documented way found to get a file from this Mac onto the router (RouterOS' SSH
/// server has no documented SCP support, but DOES accept SFTP: confirmed live against a real
/// hEX with a plain `sftp` CLI session; Citadel, already a dependency, ships an SFTP client).
/// `remoteName` must be the full `flash/...`-prefixed form — confirmed live that RouterOS'
/// `/file` and `/import` reject a bare filename ("file does not exist") even though `/file
/// print` lists the same file that way too.
func uploadScript(remoteName: String, contents: String) async throws {
guard let client else { throw RouterOSError.notConnected }
let sftp = try await client.openSFTP()
do {
try await sftp.withFile(filePath: remoteName, flags: [.write, .create, .truncate]) { file in
try await file.write(ByteBuffer(string: contents))
}
} catch {
try? await sftp.close()
throw error
}
try await sftp.close()
}
/// The RouterOS-documented restore workflow in one command: wipe the entire configuration
/// (`no-defaults=yes`, not even the vendor defaults — a genuinely blank slate) and
/// immediately re-apply the given already-uploaded script. Safer than `/import`-ing straight
/// over a live, different configuration, which MikroTik's own docs describe as needing a
/// reset first (every `add`-type line in the script would otherwise risk colliding with
/// whatever's already there). Reboots the device; like `resetToFactoryDefaults()`, the
/// connection dying mid-command is the expected outcome, not a failure — ignored here the
/// same way.
func applyRestoreScript(remoteName: String) async throws {
_ = try await run("/system reset-configuration no-defaults=yes run-after-reset=\(remoteName)")
}
/// Reboots immediately — needed after `upgradeRouterboardFirmware()`, which (confirmed via
/// docs) does not reboot on its own. Confirmed live: RouterOS' normally-interactive "Reboot,
/// yes? [y/n]" console prompt (same class of prompt `resetToFactoryDefaults()` and
/// `upgradeRouterboardFirmware()` also have) doesn't block this app's non-interactive SSH
/// exec — both of those ran fine without any special handling, this follows the same
/// established precedent.
func reboot() async throws {
_ = try? await run("/system reboot")
}
/// Triggers a fresh check against MikroTik's update servers — confirmed via docs that this
/// is one of two required steps (the other, reading the result, happens separately via
/// `fetchMenuItems("/system package update", ...)`, which now correctly falls back to the
/// singleton-item parser for this menu too). Requires the router to have working internet
/// access; a failure surfaces as the "status" field becoming an "ERROR: ..." string, not a
/// thrown exception here.
func checkForPackageUpdates() async throws {
_ = try? await run("/system package update check-for-updates")
}
/// Downloads and installs the checked update. Confirmed via docs: this reboots the router
/// automatically on success, no separate reboot step — the connection dying mid-command is
/// the expected outcome, not a failure.
func installPackageUpdate() async throws {
_ = try? await run("/system package update install")
}
/// Applies a newer RouterBOARD bootloader firmware — only becomes available after installing
/// a newer RouterOS package first (the firmware file ships bundled inside RouterOS packages,
/// confirmed via docs). Unlike this app's other mutating commands, success here is NOT
/// silent — RouterOS returns real confirmation text ("Firmware upgraded successfully, please
/// reboot for changes to take effect!"), so the raw output is returned rather than treated as
/// an error on any non-empty response. Confirmed via docs this command is normally
/// interactive in Winbox/console ("Do you really want to upgrade firmware? [y/n]") — not
/// verified whether that prompt applies over this app's non-interactive SSH exec too, though
/// `/system reset-configuration` (also normally interactive) has run fine this way all
/// session, so it's likely safe; flagged to the user in the UI regardless. Requires a manual
/// `/system reboot` afterward — not automatic (confirmed via docs).
func upgradeRouterboardFirmware() async throws -> String {
try await run("/system routerboard upgrade")
}
/// Runs a command via `executeCommandStream` (not the simpler `executeCommand`), because
/// `executeCommand` discards whatever output it already collected the moment the command
/// exits non-zero — exactly the RouterOS error text we need. Collecting the stream ourselves
/// keeps that text available even when the command fails.
/// Wrapped in `withTimeout` for the same reason as `connect()` (bugs.md #11) — a command that
/// never returns (router hangs mid-execution, connection drops without a clean error) used to
/// block forever with no recovery. The `CommandFailed` handling stays inside the timed
/// closure so `output` (partial text collected before the failure) is still in scope to build
/// the error detail — the already-established domain error (`RouterOSError.invalidResponse`)
/// is what actually crosses the timeout race, not the raw Citadel type.
private func run(_ command: String) async throws -> String {
guard let client else { throw RouterOSError.notConnected }
return try await Self.withTimeout(Self.commandTimeout) {
var output = ""
do {
let stream = try await client.executeCommandStream(command)
for try await chunk in stream {
switch chunk {
case .stdout(let buffer), .stderr(let buffer):
output += String(buffer: buffer)
}
}
return output
} catch let failure as SSHClient.CommandFailed {
let detail = output.trimmingCharacters(in: .whitespacesAndNewlines)
throw RouterOSError.invalidResponse(
"RouterOS meldete Fehler (Exit-Code \(failure.exitCode)) für \"\(command)\""
+ (detail.isEmpty ? "" : ": \(detail)")
)
}
}
}
}
extension SSHTransport: NIOSSHClientServerAuthenticationDelegate {
func validateHostKey(hostKey: NIOSSHPublicKey, validationCompletePromise: EventLoopPromise<Void>) {
let fingerprint = SSHHostKeyFingerprint.sha256(of: hostKey)
if hostKeyTrust.isTrusted(host: credentials.host, fingerprint: fingerprint) {
validationCompletePromise.succeed(())
} else {
validationCompletePromise.fail(RouterOSError.untrustedSSHHostKey(fingerprint: fingerprint))
}
}
}