Bound egress canary probes with a watchdog and cancel them on invalidate

The Tor egress self-check probe had no outer bound: the proxied session
sets waitsForConnectivity = true, which can defer the per-request timer
indefinitely, leaving a canary request bounded only by URLSession's
default 7-day resource timeout. Because verify() joins concurrent
callers on the in-flight task and invalidate() neither cancelled nor
cleared it, one hung probe wedged every subsequent verify() caller --
even across a Tor restart -- parking awaitingTorForConnections in
NostrRelayManager until process restart.

Fixes (liveness only; the fail-closed policy is unchanged):
- Race every probe against an independent async watchdog
  (probeTimeout, default 20s = the live probe's request timeout, via
  TorEgressVerifier.defaultProbeTimeout). On timeout the probe task is
  cancelled (cooperatively cancelling the underlying URLSessionTask),
  the verdict is the fail-closed .unreachable, and the in-flight slot
  is cleared so the next verify() starts fresh.
- invalidate() now cancels the in-flight probe and clears it (plus the
  throttle timestamp it already cleared), so Tor restart/dormant/
  shutdown genuinely resets the verifier.
- A probe generation counter keeps actor reentrancy safe: a cancelled/
  superseded probe cannot re-seed the throttle/cache that invalidate()
  just cleared or clobber a fresh probe's in-flight slot; its awaiting
  callers resolve promptly as false.

Concurrent-caller join semantics are preserved (one shared probe), and
the race resolves exactly once behind an NSLock never held across an
await. Tests cover the hung-probe timeout bound, invalidate-cancels-
in-flight, post-restart recovery, and the shared-probe invariant, all
with the injected probe/clock harness (no real network).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This commit is contained in:
jack
2026-07-02 11:01:15 +02:00
co-authored by Claude Fable 5
parent 74f2cd98ab
commit 6346870250
2 changed files with 330 additions and 10 deletions
+187 -4
View File
@@ -19,20 +19,72 @@ struct TorEgressVerifierTests {
private let lock = NSLock()
private var _now = Date(timeIntervalSince1970: 1_000_000)
private var _result: TorEgressVerifier.ProbeResult = .verifiedTor
private(set) var probeCount = 0
private var _hanging = false
private var _gated = false
private var _released = false
private var _probeCount = 0
private var _cancelledCount = 0
var now: Date {
lock.lock(); defer { lock.unlock() }; return _now
}
var probeCount: Int {
lock.lock(); defer { lock.unlock() }; return _probeCount
}
/// Number of hung probes that observed cooperative cancellation.
var cancelledCount: Int {
lock.lock(); defer { lock.unlock() }; return _cancelledCount
}
func advance(_ seconds: TimeInterval) {
lock.lock(); _now = _now.addingTimeInterval(seconds); lock.unlock()
}
func setResult(_ r: TorEgressVerifier.ProbeResult) {
lock.lock(); _result = r; lock.unlock()
}
/// When `true`, probes park forever and only exit via cooperative
/// cancellation models a canary request wedged by
/// `waitsForConnectivity` deferring the request timer.
func setHanging(_ hanging: Bool) {
lock.lock(); _hanging = hanging; lock.unlock()
}
/// When `true`, probes wait for `release()` before returning models
/// a slow-but-completing canary for join-semantics tests.
func setGated(_ gated: Bool) {
lock.lock(); _gated = gated; lock.unlock()
}
func release() {
lock.lock(); _released = true; lock.unlock()
}
/// Suspends until at least `n` probes have started.
func waitUntilProbeCount(atLeast n: Int) async {
while probeCount < n {
try? await Task.sleep(nanoseconds: 1_000_000)
}
}
func makeProbe() -> @Sendable () async -> TorEgressVerifier.ProbeResult {
return { [self] in
lock.lock(); probeCount += 1; let r = _result; lock.unlock()
lock.lock()
_probeCount += 1
let r = _result
let hang = _hanging
let gated = _gated
lock.unlock()
if hang {
// Park until cancelled (verifier watchdog or invalidate());
// cancellation-responsive so no task outlives the test.
while !Task.isCancelled {
try? await Task.sleep(nanoseconds: 2_000_000)
}
lock.lock(); _cancelledCount += 1; lock.unlock()
return .unreachable("hung probe cancelled")
}
if gated {
while !Task.isCancelled {
lock.lock(); let released = _released; lock.unlock()
if released { break }
try? await Task.sleep(nanoseconds: 1_000_000)
}
}
return r
}
}
@@ -41,8 +93,19 @@ struct TorEgressVerifierTests {
}
}
private func makeVerifier(_ h: Harness, ttl: TimeInterval = 300, minRetry: TimeInterval = 5) -> TorEgressVerifier {
TorEgressVerifier(ttl: ttl, minRetryInterval: minRetry, now: h.nowProvider(), probe: h.makeProbe())
private func makeVerifier(
_ h: Harness,
ttl: TimeInterval = 300,
minRetry: TimeInterval = 5,
probeTimeout: TimeInterval = TorEgressVerifier.defaultProbeTimeout
) -> TorEgressVerifier {
TorEgressVerifier(
ttl: ttl,
minRetryInterval: minRetry,
probeTimeout: probeTimeout,
now: h.nowProvider(),
probe: h.makeProbe()
)
}
@Test("verifiedTor allows and is cached within TTL (single probe)")
@@ -217,4 +280,124 @@ struct TorEgressVerifierTests {
_ = await v.verify()
#expect(await v.lastProbeResult() == .notTor)
}
// MARK: - Liveness (probe timeout watchdog + invalidate cancellation)
@Test("a probe that never completes is bounded by probeTimeout and fails closed")
func hungProbeIsBoundedByTimeout() async {
let h = Harness()
h.setHanging(true)
let v = makeVerifier(h, ttl: 300, minRetry: 0, probeTimeout: 0.05)
// Without the watchdog this would wedge: waitsForConnectivity can
// defer the request timer, leaving the canary bounded only by the
// 7-day resource timeout.
#expect(await v.verify() == false)
let last = await v.lastProbeResult()
switch last {
case .unreachable:
break // fail-closed timeout verdict recorded
default:
Issue.record("expected .unreachable after probe timeout, got \(String(describing: last))")
}
#expect(v.hasFreshVerification == false)
// The hung probe task itself was cancelled (URLSession task would be
// torn down), not abandoned.
while h.cancelledCount < 1 {
try? await Task.sleep(nanoseconds: 1_000_000)
}
#expect(h.cancelledCount == 1)
// The in-flight slot was cleared: the next verify() starts a fresh
// probe (does not join the hung one) and recovers.
h.setHanging(false)
h.setResult(.verifiedTor)
#expect(await v.verify() == true)
#expect(h.probeCount == 2)
#expect(v.hasFreshVerification == true)
}
@Test("invalidate() cancels the in-flight probe and the awaiting caller fails closed")
func invalidateCancelsInFlightProbe() async {
let h = Harness()
h.setHanging(true)
// Long (real-time) timeout and throttle: only invalidate() can
// unblock the caller, and only invalidate() clearing the throttle
// lets the follow-up probe run without advancing the clock.
let v = makeVerifier(h, ttl: 300, minRetry: 600, probeTimeout: 600)
let first = Task { await v.verify() }
await h.waitUntilProbeCount(atLeast: 1)
await v.invalidate()
// The awaiting caller resolves promptly (no 600s wait) and refuses.
#expect(await first.value == false)
#expect(v.hasFreshVerification == false)
// The hung probe observed cancellation (a Tor restart genuinely
// shakes the wedged canary request).
while h.cancelledCount < 1 {
try? await Task.sleep(nanoseconds: 1_000_000)
}
// Recovery: a fresh verify() runs a NEW probe it neither joins the
// cancelled one nor inherits its throttle/last-result state.
h.setHanging(false)
h.setResult(.verifiedTor)
#expect(await v.verify() == true)
#expect(h.probeCount == 2)
#expect(v.hasFreshVerification == true)
}
@Test("recovery after a hung probe survives invalidate + Tor restart cycle")
func hungThenInvalidatedThenRecovers() async {
let h = Harness()
h.setHanging(true)
let v = makeVerifier(h, ttl: 300, minRetry: 5, probeTimeout: 600)
// Wedge one probe, then simulate a Tor restart mid-flight.
let wedged = Task { await v.verify() }
await h.waitUntilProbeCount(atLeast: 1)
await v.invalidate()
#expect(await wedged.value == false)
// Canary still down right after restart: fresh probe, fail closed
// and the throttle applies to the FRESH result (bounded retry intact).
h.setHanging(false)
h.setResult(.unreachable("circuit not built"))
#expect(await v.verify() == false)
#expect(h.probeCount == 2)
h.advance(1)
#expect(await v.verify() == false)
#expect(h.probeCount == 2) // throttled, no hammering
// Canary returns after the retry window: verification recovers.
h.setResult(.verifiedTor)
h.advance(5)
#expect(await v.verify() == true)
#expect(v.hasFreshVerification == true)
}
@Test("concurrent verify() callers still share a single in-flight probe")
func concurrentCallersShareOneProbe() async {
let h = Harness()
h.setResult(.verifiedTor)
h.setGated(true)
let v = makeVerifier(h, ttl: 300, minRetry: 0)
let t1 = Task { await v.verify() }
await h.waitUntilProbeCount(atLeast: 1)
let t2 = Task { await v.verify() }
// Give t2 a chance to join the in-flight probe before releasing it.
// Either way the invariant holds: t2 joins the shared probe, or (if
// scheduled after completion) hits the fresh TTL cache exactly one
// probe runs.
try? await Task.sleep(nanoseconds: 20_000_000)
h.release()
#expect(await t1.value == true)
#expect(await t2.value == true)
#expect(h.probeCount == 1)
}
}
@@ -44,6 +44,20 @@ import Foundation
/// gate, GeoRelayDirectory backoff) re-probe and succeed once the canary
/// is reachable again.
///
/// Liveness:
/// - Every probe is hard-bounded by `probeTimeout` via an independent async
/// watchdog. The proxied session sets `waitsForConnectivity = true`, which
/// can defer the per-request timer indefinitely, leaving the request
/// bounded only by the default 7-day resource timeout without the
/// watchdog a hung canary would wedge every subsequent `verify()` caller.
/// On timeout the probe task is cancelled (cooperatively cancelling the
/// underlying `URLSessionTask`), the verdict is the fail-closed
/// `.unreachable`, and the in-flight slot is cleared so the next
/// `verify()` (after the retry throttle) starts a fresh probe.
/// - `invalidate()` cancels any in-flight probe and clears all cached state
/// (including the retry throttle), so a Tor restart/dormant/shutdown
/// genuinely resets the verifier: a hung probe cannot survive it.
///
/// The probe is injectable so the policy/caching logic is unit-tested without a
/// live network (see `TorEgressVerifierTests`).
public actor TorEgressVerifier {
@@ -56,6 +70,11 @@ public actor TorEgressVerifier {
case unreachable(String)
}
/// Hard upper bound for a single canary probe, enforced independently of
/// URLSession timers (see the "Liveness" section of the type doc). Matches
/// the live probe's per-request timeout.
public static let defaultProbeTimeout: TimeInterval = 20
private let probe: @Sendable () async -> ProbeResult
private let now: @Sendable () -> Date
private let ttl: TimeInterval
@@ -63,11 +82,18 @@ public actor TorEgressVerifier {
/// persistent `.unreachable` cannot hammer the canary endpoint on every
/// reconnect burst.
private let minRetryInterval: TimeInterval
/// Outer wall-clock bound on a single probe (watchdog; fail-closed).
private let probeTimeout: TimeInterval
private var lastVerifiedAt: Date?
private var lastProbeAt: Date?
private var lastResult: ProbeResult?
private var inFlight: Task<Bool, Never>?
/// Bumped whenever a new probe starts or `invalidate()` runs. A completing
/// probe only records its outcome (cache/throttle) and clears `inFlight`
/// if its generation is still current, so a cancelled/superseded probe
/// cannot clobber state owned by a fresh one (actor-reentrancy safety).
private var probeGeneration = 0
/// Lock-protected mirror of "verified within TTL" so synchronous gates
/// (e.g. `NostrRelayManager`'s connect path) can consult the cache without
@@ -95,18 +121,26 @@ public actor TorEgressVerifier {
public init(
ttl: TimeInterval,
minRetryInterval: TimeInterval = 5.0,
probeTimeout: TimeInterval = TorEgressVerifier.defaultProbeTimeout,
now: @escaping @Sendable () -> Date = Date.init,
probe: @escaping @Sendable () async -> ProbeResult
) {
self.ttl = ttl
self.minRetryInterval = minRetryInterval
self.probeTimeout = probeTimeout
self.now = now
self.probe = probe
}
/// Drop any cached verification (e.g. after a Tor restart or when the
/// network path changes). The next `verify()` re-probes.
/// network path changes) AND cancel any in-flight probe. The next
/// `verify()` starts a fresh probe it neither joins the cancelled one
/// nor is throttled by its outcome, so a probe hung from before a Tor
/// restart cannot wedge callers after it.
public func invalidate() {
probeGeneration += 1
inFlight?.cancel()
inFlight = nil
lastVerifiedAt = nil
lastProbeAt = nil
lastResult = nil
@@ -137,10 +171,15 @@ public actor TorEgressVerifier {
}
if let inFlight { return await inFlight.value }
let task = Task<Bool, Never> { await self.runProbe() }
probeGeneration += 1
let generation = probeGeneration
let task = Task<Bool, Never> { await self.runProbe(generation: generation) }
inFlight = task
let allowed = await task.value
inFlight = nil
// Only clear the slot if this probe is still the current one: an
// `invalidate()` while we were suspended has already cleared it and a
// newer probe may occupy it (do not clobber the fresh task).
if probeGeneration == generation { inFlight = nil }
return allowed
}
@@ -158,8 +197,13 @@ public actor TorEgressVerifier {
}
}
private func runProbe() async -> Bool {
let result = await probe()
private func runProbe(generation: Int) async -> Bool {
let result = await boundedProbe()
// Superseded by `invalidate()` (Tor restart/dormant/shutdown) while the
// probe ran: its verdict predates the reset, so discard it recording
// it would re-seed the throttle/cache that invalidate() just cleared.
// Fail closed for the callers that were awaiting this probe.
guard generation == probeGeneration else { return false }
lastProbeAt = now()
lastResult = result
switch result {
@@ -185,6 +229,95 @@ public actor TorEgressVerifier {
return false
}
}
/// Runs the injected probe raced against `probeTimeout`, guaranteeing a
/// result in bounded time regardless of URLSession timer behavior (the
/// proxied session's `waitsForConnectivity` can defer the per-request
/// timeout indefinitely). Whichever side loses the race is cancelled:
/// - on timeout, the probe task is cancelled (URLSession's async APIs
/// cancel the underlying `URLSessionTask` cooperatively) and the result
/// is the fail-closed `.unreachable`;
/// - on completion, the watchdog's sleep is cancelled so no timer lingers.
/// Cancelling the enclosing task (`invalidate()`) resolves immediately as
/// `.unreachable` and cancels both sides.
///
/// `nonisolated` so the race body never re-enters the actor; it touches
/// only immutable `Sendable` state.
nonisolated private func boundedProbe() async -> ProbeResult {
let probe = self.probe
let timeout = self.probeTimeout
let race = ProbeRace()
return await withTaskCancellationHandler {
await withCheckedContinuation { (continuation: CheckedContinuation<ProbeResult, Never>) in
race.install(continuation)
let probeTask = Task { race.finish(await probe()) }
let watchdog = Task {
try? await Task.sleep(nanoseconds: UInt64(max(0, timeout) * 1_000_000_000))
guard !Task.isCancelled else { return }
race.finish(.unreachable("probe timed out after \(Int(timeout))s"))
}
race.register(probeTask: probeTask, watchdog: watchdog)
}
} onCancel: {
race.finish(.unreachable("probe cancelled"))
}
}
/// Resolve-once rendezvous for the probe/watchdog race. Lock-protected
/// (never held across an await); the first `finish()` wins, resumes the
/// continuation exactly once, and cancels both tasks.
private final class ProbeRace: @unchecked Sendable {
private let lock = NSLock()
private var continuation: CheckedContinuation<ProbeResult, Never>?
private var pendingResult: ProbeResult?
private var resolved = false
private var probeTask: Task<Void, Never>?
private var watchdog: Task<Void, Never>?
func install(_ continuation: CheckedContinuation<ProbeResult, Never>) {
lock.lock()
if let result = pendingResult {
// finish() ran before the continuation existed (e.g. the
// enclosing task was already cancelled): resolve immediately.
pendingResult = nil
lock.unlock()
continuation.resume(returning: result)
return
}
self.continuation = continuation
lock.unlock()
}
func register(probeTask: Task<Void, Never>, watchdog: Task<Void, Never>) {
lock.lock()
if resolved {
lock.unlock()
probeTask.cancel()
watchdog.cancel()
return
}
self.probeTask = probeTask
self.watchdog = watchdog
lock.unlock()
}
func finish(_ result: ProbeResult) {
lock.lock()
guard !resolved else { lock.unlock(); return }
resolved = true
let continuation = self.continuation
self.continuation = nil
if continuation == nil { pendingResult = result }
let probeTask = self.probeTask
let watchdog = self.watchdog
self.probeTask = nil
self.watchdog = nil
lock.unlock()
probeTask?.cancel()
watchdog?.cancel()
continuation?.resume(returning: result)
}
}
}
// MARK: - Live probe
@@ -198,9 +331,13 @@ public extension TorEgressVerifier {
///
/// Follow-up (see PR): make the canary endpoint configurable and add an
/// onion-service canary so verification does not depend on a single host.
/// Note: the per-request `timeoutInterval` below is best-effort only the
/// proxied session's `waitsForConnectivity` can defer it. The authoritative
/// bound is the verifier's `probeTimeout` watchdog, whose cancellation
/// propagates into `session.data(for:)` and cancels the URLSessionTask.
static func liveProbe(
endpoint: URL = URL(string: "https://check.torproject.org/api/ip")!,
timeout: TimeInterval = 20
timeout: TimeInterval = TorEgressVerifier.defaultProbeTimeout
) -> @Sendable () async -> ProbeResult {
return {
var request = URLRequest(url: endpoint)