2013-04-11

CancellationToken Behavior with AsyncEx Coordination Primitives

Today I'm going to talk a little bit about some behavior that I standardized on for the asynchronous coordination primitives in my AsyncEx library. I'll first review what's currently available in the BCL and then describe my approach.

Waiting on SemaphoreSlim

The only built-in .NET asynchronous-compatible coordination primitive (as of .NET 4.5) is SemaphoreSlim. This type can be used by both synchronous and asynchronous code, which is a very interesting aspect. What I'm looking at today is all the different ways to wait for a semaphore to be available (and I'm just going to consider asynchronous code).

First, there's the obvious Task WaitAsync() method for an unconditional wait. This is the most commonly-used type of wait: the code knows it needs to acquire the semaphore and it will wait however long it takes until the semaphore is available.

Next, there's a couple of overloads for timeouts. Task<bool> WaitAsync(int) and Task<bool> WaitAsync(TimeSpan) both do an (asynchronous) timed wait, which will return false if the wait ran out of time before it was granted access.

There's an overload taking a cancellation token, Task WaitAsync(CancellationToken). This performs a cancelable wait, where some event can interrupt the wait if it determines that the code doesn't need that semaphore anymore. If a wait is canceled, the wait task is canceled (instead of returning false).

You can also have cancelable timed waits, via the Task<bool> WaitAsync(int, CancellationToken) and Task<bool> WaitAsync(TimeSpan, CancellationToken). These overloads will cancel their tasks if the cancellation token fires, and return false if they hit the timeout.

Finally, there's one other important kind of wait you can do: an atomic wait, where you immediately (synchronously) acquire the semaphore if it is available. Asynchronous code can use bool Wait(int) and pass zero to perform an atomic wait. This is logically similar to Monitor.TryEnter.

Waiting on AsyncSemaphore

I'll skip right to the punchline. There are only two wait overloads in the AsyncSemaphore class: Task WaitAsync() and Task WaitAsync(CancellationToken).

The unconditional wait uses the same obvious method: Task WaitAsync() (which is shorthand for WaitAsync(CancellationToken.None)). The other overload Task WaitAsync(CancellationToken) handles everything else.

To do a timed wait, create a cancellation token to cancel the wait (e.g., using CancellationTokenSource.CancellationTokenSource(TimeSpan) or CancellationTokenHelpers.Timeout(TimeSpan)) and pass it to WaitAsync. When the timer expires, the wait is canceled.

Of course, a cancelable wait can just pass in its CancellationToken. Cancelable timed waits can use chained cancellation tokens (via CancellationTokenSource.CreateLinkedTokenSource or CancellationTokenHelpers.Normalize), and they can follow up with CancellationToken.IsCancellationRequested if it's important to distinguish why the wait was canceled.

That just leaves the atomic wait, and this is where the "interesting design" comes in. If you pass in a token that is already canceled, then WaitAsync will always return synchronously: a successfully completed task if the semaphore is available, otherwise a canceled task. This gives you the "TryLock" kind of behavior, but this also means that if you have a regular cancellation token, it will not result in a canceled task if the semaphore is available.

This special behavior is supported for every kind of "wait" in AsyncEx where atomic waits make sense and that modifies state: AsyncSemaphore.WaitAsync, AsyncLock.LockAsync, AsyncAutoResetEvent.WaitAsync, AsyncMonitor.EnterAsync, AsyncReaderWriterLock.WriterLockAsync, AsyncReaderWriterLock.ReaderLockAsync, AsyncReaderWriterLock.UpgradeableReaderLockAsync, and AsyncReaderWriterLock.UpgradeableReaderKey.UpgradeAsync.

No comments:

Post a Comment