I use C# iterators as a replacement for coroutines, and it has been working great. I want to switch to async/await as I think the syntax is cleaner and it gives me type safety.
In this (outdated) blog post, Jon Skeet shows a possible way to implement it.
I chose to go a slightly different way (by implementing my own SynchronizationContext and using Task.Yield). This worked fine.
Then I realized there would be a problem; currently a coroutine doesn't have to finish running. It can be stopped gracefully at any point where it yields. We might have code like this:
private IEnumerator Sleep(int milliseconds)
{
Stopwatch timer = Stopwatch.StartNew();
do
{
yield return null;
}
while (timer.ElapsedMilliseconds < milliseconds);
}
private IEnumerator CoroutineMain()
{
try
{
// Do something that runs over several frames
yield return Coroutine.Sleep(5000);
}
finally
{
Log("Coroutine finished, either after 5 seconds, or because it was stopped");
}
}
The coroutine works by keeping track of all enumerators in a stack. The C# compiler generates a Dispose function which can be called to ensure that the 'finally' block is correctly invoked in CoroutineMain, even if the enumeration isn't finished. This way we can stop a coroutine gracefully, and still ensure finally blocks are invoked, by calling Dispose on all the IEnumerator objects on the stack. This is basically manually unwinding.
When I wrote my implementation with async/await I realized that we would lose this feature, unless I'm mistaken. I then looked up other coroutine solutions, and it doesn't look like Jon Skeet's version handles it in any way either.
The only way I can think of to handle this would be to have our own custom 'Yield' function, which would check if the coroutine was stopped, and then raise an exception that indicated this. This would propagate up, executing finally blocks, and then be caught somewhere near the root. I don't find this pretty though, as 3rd party code could potentially catch the exception.
Am I misunderstanding something, and is this possible to do in an easier way? Or do I need to go the exception way to do this?
EDIT: More information/code has been requested, so here's some. I can guarantee this is going to be running on only a single thread, so there's no threading involved here.
Our current coroutine implementation looks a bit like this (this is simplified, but it works in this simple case):
public sealed class Coroutine : IDisposable
{
private class RoutineState
{
public RoutineState(IEnumerator enumerator)
{
Enumerator = enumerator;
}
public IEnumerator Enumerator { get; private set; }
}
private readonly Stack<RoutineState> _enumStack = new Stack<RoutineState>();
public Coroutine(IEnumerator enumerator)
{
_enumStack.Push(new RoutineState(enumerator));
}
public bool IsDisposed { get; private set; }
public void Dispose()
{
if (IsDisposed)
return;
while (_enumStack.Count > 0)
{
DisposeEnumerator(_enumStack.Pop().Enumerator);
}
IsDisposed = true;
}
public bool Resume()
{
while (true)
{
RoutineState top = _enumStack.Peek();
bool movedNext;
try
{
movedNext = top.Enumerator.MoveNext();
}
catch (Exception ex)
{
// Handle exception thrown by coroutine
throw;
}
if (!movedNext)
{
// We finished this (sub-)routine, so remove it from the stack
_enumStack.Pop();
// Clean up..
DisposeEnumerator(top.Enumerator);
if (_enumStack.Count <= 0)
{
// This was the outer routine, so coroutine is finished.
return false;
}
// Go back and execute the parent.
continue;
}
// We executed a step in this coroutine. Check if a subroutine is supposed to run..
object value = top.Enumerator.Current;
IEnumerator newEnum = value as IEnumerator;
if (newEnum != null)
{
// Our current enumerator yielded a new enumerator, which is a subroutine.
// Push our new subroutine and run the first iteration immediately
RoutineState newState = new RoutineState(newEnum);
_enumStack.Push(newState);
continue;
}
// An actual result was yielded, so we've completed an iteration/step.
return true;
}
}
private static void DisposeEnumerator(IEnumerator enumerator)
{
IDisposable disposable = enumerator as IDisposable;
if (disposable != null)
disposable.Dispose();
}
}
Assume we have code like the following:
private IEnumerator MoveToPlayer()
{
try
{
while (!AtPlayer())
{
yield return Sleep(500); // Move towards player twice every second
CalculatePosition();
}
}
finally
{
Log("MoveTo Finally");
}
}
private IEnumerator OrbLogic()
{
try
{
yield return MoveToPlayer();
yield return MakeExplosion();
}
finally
{
Log("OrbLogic Finally");
}
}
This would be created by passing an instance of the OrbLogic enumerator to a Coroutine, and then running it. This allows us to tick the coroutine every frame. If the player kills the orb, the coroutine doesn't finish running; Dispose is simply called on the coroutine. If MoveTo was logically in the 'try' block, then calling Dispose on the top IEnumerator will, semantically, make the finally block in MoveTo execute. Then afterwards the finally block in OrbLogic will execute.
Note that this is a simple case and the cases are much more complex.
I am struggling to implement similar behavior in the async/await version. The code for this version looks like this (error checking omitted):
public class Coroutine
{
private readonly CoroutineSynchronizationContext _syncContext = new CoroutineSynchronizationContext();
public Coroutine(Action action)
{
if (action == null)
throw new ArgumentNullException("action");
_syncContext.Next = new CoroutineSynchronizationContext.Continuation(state => action(), null);
}
public bool IsFinished { get { return !_syncContext.Next.HasValue; } }
public void Tick()
{
if (IsFinished)
throw new InvalidOperationException("Cannot resume Coroutine that has finished");
SynchronizationContext curContext = SynchronizationContext.Current;
try
{
SynchronizationContext.SetSynchronizationContext(_syncContext);
// Next is guaranteed to have value because of the IsFinished check
Debug.Assert(_syncContext.Next.HasValue);
// Invoke next continuation
var next = _syncContext.Next.Value;
_syncContext.Next = null;
next.Invoke();
}
finally
{
SynchronizationContext.SetSynchronizationContext(curContext);
}
}
}
public class CoroutineSynchronizationContext : SynchronizationContext
{
internal struct Continuation
{
public Continuation(SendOrPostCallback callback, object state)
{
Callback = callback;
State = state;
}
public SendOrPostCallback Callback;
public object State;
public void Invoke()
{
Callback(State);
}
}
internal Continuation? Next { get; set; }
public override void Post(SendOrPostCallback callback, object state)
{
if (callback == null)
throw new ArgumentNullException("callback");
if (Current != this)
throw new InvalidOperationException("Cannot Post to CoroutineSynchronizationContext from different thread!");
Next = new Continuation(callback, state);
}
public override void Send(SendOrPostCallback d, object state)
{
throw new NotSupportedException();
}
public override int Wait(IntPtr[] waitHandles, bool waitAll, int millisecondsTimeout)
{
throw new NotSupportedException();
}
public override SynchronizationContext CreateCopy()
{
throw new NotSupportedException();
}
}
I don't see how to implement similar behavior to the iterator version using this.
Apologies in advance for the lengthy code!
EDIT 2: The new method seems to be working. It allows me to do stuff like:
private static async Task Test()
{
// Second resume
await Sleep(1000);
// Unknown how many resumes
}
private static async Task Main()
{
// First resume
await Coroutine.Yield();
// Second resume
await Test();
}
Which provides a very nice way of building AI for games.
See Question&Answers more detail:
os 
