Combines withTaskCancellationHandler with a CheckedContinuation.

TaskCancellationHandler.swift

import Synchronization

func untilCancelled() async {
	print("1")
	let mutex: Mutex<CheckedContinuation<Void, Never>?> = .init(nil)
	
	await withTaskCancellationHandler {
		print("2")
		await withCheckedContinuation { continuation in
			print("3")
			mutex.withLock {
				if Task.isCancelled {
					print("abort")
					continuation.resume()
					return
				}
				
				print("4")
				$0 = continuation
			}
		}
	} onCancel: {
		let continuation = mutex.withLock {
			print("5")
			return $0.take()
		}
		continuation?.resume()
	}
	
	print("6")
}

func main() async throws {
	let hangsForever = Task {
		await untilCancelled()
	}
	
	// Comment out this line to test early cancellation.
	try await Task.sleep(for: .milliseconds(10))
	
	print("will cancel")
	hangsForever.cancel()
	print("did cancel")
	
	try await Task.sleep(for: .milliseconds(10))
}

try await main()

@Noobish1 even if it was, you could not throw there.

The body of even withCheckedThrowingContinuation is always non-throwing. It's defined as _ body: (CheckedContinuation<T, any Error>) -> Void.

I guess this makes sense, also keeps things simpler and the semantics cleaner. There already is a way to throw out of this continuation by calling .resume(throwing: CancellationError()). Also when trowing an error there that resumes the continuation, you'd need to make sure the continuation isn't already stored somewhere else and might be over-resumed. Just too many pitfalls, I guess.

@macguru yup that makes total sense, you wouldn't want to throw inside the continuation as you said, you are given one way of doing that.

I was thinking of this for a throwing version:

func untilCancelled() async throws {
    print("1")
    let mutex: Mutex<CheckedContinuation<Void, Error>?> = .init(nil)

    try await withTaskCancellationHandler {
        print("2")

        try Task.checkCancellation() // Note this cancellation check here

        try await withCheckedThrowingContinuation { continuation in
            print("3")
            mutex.withLock {
                print("4")
                $0 = continuation
            }

            doSomeAsyncThing { result in
                let continuation = mutex.withLock { $0.take() }
                continuation?.resume(with: result)
            }
        }
    } onCancel: {
        let continuation = mutex.withLock {
            print("5")
            return $0.take()
        }
        continuation?.resume(throwing: CancellationError())
    }

    print("6")
}

@mattmassicotte / @macguru is this an abomination? xD

func cancellableContinuation<T: Sendable>(_ perform: (@escaping @Sendable (Result<T, Error & Sendable>) -> Void) -> Void) async throws -> T {
    let continuationRef: Mutex<CheckedContinuation<T, Error>?> = .init(nil)

    return try await withTaskCancellationHandler {
        try Task.checkCancellation()

        return try await withCheckedThrowingContinuation { continuation in
            continuationRef.withLock {
                $0 = continuation
            }

            perform({ result in
                let continuation = continuationRef.withLock { $0.take() }
                continuation?.resume(with: result)
            })
        }
    } onCancel: {
        let continuation = continuationRef.withLock { $0.take() }
        continuation?.resume(throwing: CancellationError())
    }
}

I liked the article too, it is very insightful to be honest. I was actually thinking about creating something like what @Noobish1 created to handle the cancellation all over the app I am working on. But I wanted to ask a small question a found a small solution on swift forums that makes use share properties created in the function between the operation closure and the cancellation closure by doing so:

        let onCancel = { dataTask?.cancel() }
        return try await withTaskCancellationHandler {
            // ..........
        } onCancel: {
            onCancel()
        }

And what about replacing mutex with this legacy mutex for app's that supports iOS 16+:

final class LegacyMutex<Wrapped>: @unchecked Sendable {
    private var mutex = pthread_mutex_t()
    private var wrapped: Wrapped
    
    init(_ initialValue: Wrapped) {
        pthread_mutex_init(&mutex, nil)
        self.wrapped = initialValue
    }
    
    deinit {
        pthread_mutex_destroy(&mutex)
    }
    
    func withLock<R>(_ body: @Sendable (inout Wrapped) throws -> R) rethrows -> R {
        pthread_mutex_lock(&mutex); defer { pthread_mutex_unlock(&mutex) }
        return try body(&wrapped)
    }
}

I think OSAllocatedUnfairLock is a great option if you cannot use Mutex. It's not as safe, but pretty good.

I'm also not 100% sure about that withLock implementation. I think you may need either R: Sendable or maybe a sending in there.

I think you may need either R: Sendable or maybe a sending in there.

why should I do this ?

This is locking pattern known to be easy to use wrong in subtle ways. It can be used safely, but is very easy to use wrong. The core problem is as implemented LegacyMutex actually isn't fully Sendable. Here's an example:

class NonSendable {}

struct UsesLegacyMutex: Sendable {
	private let locked = LegacyMutex(NonSendable())

	var thisIsNotActuallySafe: NonSendable {
		locked.withLock { $0 }
	}
}

@EngOmarElsayed we’ve been using this package for a legacy mutex: https://github.com/swhitty/swift-mutex

@Noobish1 About your example… I'm wondering what the perform() would be doing. Because for anything "heavy" you'd probably want an asynchronous context, so you'd need to make a task there. And then, you could leave out the continuation handling entirely and just wrap what you're doing in a cancellation handler.

@EngOmarElsayed I'm not entirely sure what you mean in the reference to the forum post you liked. But if I understand you correctly, then in those cases where you have a Task, you don't need any of the dance here. I talked about this in my post under The simple example.

@macguru perform would be some API that uses closure callbacks that can't be easily converted to async/await.

	var thisIsNotActuallySafe: NonSendable {
		locked.withLock { $0 }
	}

But why this isn't safe, I think there is something I am missing but what is it 😅 ?

I'm not entirely sure what you mean in the reference to the forum post you liked. But if I understand you correctly, then in those cases where you have a Task, you don't need any of the dance here. I talked about this in my post under The simple example.

You are right, I was sharing something I found with you 😅

But why this isn't safe, I think there is something I am missing but what is it 😅 ?

This type, as implemented, makes it possible to send non-Sendable data. That doesn't mean it actually will happen, just that the compiler will not be able to prevent it.

But why this isn't safe, I think there is something I am missing but what is it 😅 ?

This type, as implemented, makes it possible to send non-Sendable data. That doesn't mean it actually will happen, just that the compiler will not be able to prevent it.

Got you, thanks for the explanation