What happens to Tasks that are never completed? Ar

Say I have the following class:

class SomeClass
{
    private TaskCompletionSource<string> _someTask;

    public Task<string> WaitForThing()
    {
        _someTask = new TaskCompletionSource<string>();
        return _someTask.Task;
    }

    //Other code which calls _someTask.SetResult(..);
}

Then elsewhere, I call

//Some code..
await someClassInstance.WaitForThing();
//Some more code

The //Some more code won't be called until _someTask.SetResult(..) is called. The calling-context is waiting around in memory somewhere.

However, let's say SetResult(..) is never called, and someClassInstance stops being referenced and is garbage collected. Does this create a memory leak? Or does .Net auto-magically know the calling-context needs to be disposed?

回答1:

Updated, a good point by @SriramSakthivel, it turns out I've already answered a very similar question:

Why does GC collects my object when I have a reference to it?

So I'm marking this one as a community wiki.

However, let's say SetResult(..) is never called, and someClassInstance stops being referenced and is garbage collected. Does this create a memory leak? Or does .Net auto-magically know the calling-context needs to be disposed?

If by the calling-context you mean the compiler-generated state machine object (which represents the state of the async method), then yes, it will indeed be finalized.

Example:

static void Main(string[] args)
{
    var task = TestSomethingAsync();
    Console.WriteLine("Press enter to GC");
    Console.ReadLine();
    GC.Collect(GC.MaxGeneration, GCCollectionMode.Forced, true);
    GC.WaitForFullGCComplete();
    GC.WaitForPendingFinalizers();
    Console.WriteLine("Press enter to exit");
    Console.ReadLine();
}

static async Task TestSomethingAsync()
{
    using (var something = new SomeDisposable())
    {
        await something.WaitForThingAsync();
    }
}

class SomeDisposable : IDisposable
{
    readonly TaskCompletionSource<string> _tcs = new TaskCompletionSource<string>();

    ~SomeDisposable()
    {
        Console.WriteLine("~SomeDisposable");
    }

    public Task<string> WaitForThingAsync()
    {
        return _tcs.Task;
    }

    public void Dispose()
    {
        Console.WriteLine("SomeDisposable.Dispose");
        GC.SuppressFinalize(this);
    }
}

Output:

Press enter to GC

~SomeDisposable
Press enter to exit

IMO, this behavior is logical, but it still might be a bit unexpected that something gets finalized despite the fact that the using scope for it has never ended (and hence its SomeDisposable.Dispose has never been called) and that the Task returned by TestSomethingAsync is still alive and referenced in Main.

This could lead to some obscure bugs when coding system-level asynchronous stuff. It's really important to use GCHandle.Alloc(callback) on any OS interop callbacks which are not referenced outside async methods. Doing GC.KeepAlive(callback) alone at the end of the async method is not effective. I wrote about this in details here:

Async/await, custom awaiter and garbage collector

On a side note, there's another kind of C# state machine: a method with return yield. Interestingly, along with IEnumerable or IEnumerator, it also implements IDisposable. Invoking its Dispose will unwind any using and finally statements (even in the case of incomplete enumerable sequence):

static IEnumerator SomethingEnumerable()
{
    using (var disposable = new SomeDisposable())
    {
        try
        {
            Console.WriteLine("Step 1");
            yield return null;
            Console.WriteLine("Step 2");
            yield return null;
            Console.WriteLine("Step 3");
            yield return null;
        }
        finally
        {
            Console.WriteLine("Finally");
        }
    }
}
// ...
var something = SomethingEnumerable();
something.MoveNext(); // prints "Step 1"
var disposable = (IDisposable)something;
disposable.Dispose(); // prints "Finally", "SomeDisposable.Dispose"

Unlike this, with async methods there's no direct way of controlling the unwiding of using and finally.

回答2:

You should ensure your tasks are always completed.

In the usual case, the "Other code which calls SetResult" is registered as a callback somewhere. E.g., if it's using unmanaged overlapped I/O, then that callback method is a GC root. Then that callback explicitly keeps _someTask alive, which keeps its Task alive, which keeps the delegate for //Some more code alive.

If the "Other code which calls SetResult" is not (directly or indirectly) registered as a callback, then I don't think there will be a leak. Note that this is not a supported use case, so this is not guaranteed. But I did create a memory profiling test using the code in your question and it does not appear to leak.