Anthony Steele


Resynchronising async code

After writing “Avoiding simple mistakes in async await” and revising it many times, it becomes inescapable that some of the uses and abuses of async code are not simple, particularly about breaking out of async and preventing deadlocks.

In order to understand async deadlocks, you need to understand the Synchronisation Context and how it differs in the different kinds of application. If your code is running in an application model that sets a synchronisation context which only allows one delegate to run at a time, your code can deadlock if you try to synchronously wait for code which is asynchronous to complete. Application models that have this property are: Windows desktop GUI applications (Windows forms and WPF), and in ASP web applications; but it is false in console applications, in windows services or in work which has been explicitly queued on a thread pool thread (e.g. via Task.Run(...)), and is false in ASP.NET Core.

Application Model Has an exclusive synch. context
WinForms Yes
Queued on the thread pool via Task.Run No
Default (Console apps, NUnit tests) No

The options to avoid deadlock are:

Best to stay async

The best option is to stay async from top to bottom, and don’t resynchronise. Don’t throw away the benefits of async. This may be a fair amount of work, adding await and async Task on many methods, but it should always be the first choice.

In general async code contains one or more await statements, but also lots of synchronous statements that are not awaited. Doing synchronous things in async code is generally safe. Doing asynchronous things in synchronous code and not awaiting it is generally dangerous, and should be avoided. Code that tries to embed async code within synchronous code often has synchronisation-related problems. Code that is never async or code that is always async tends to not have these.

Know when you do need to re-sync

There is a short list of times when re-syncing is not avoidable.

Know how to re-synchronise

There are a few ways to re-synchronise:

Just Wait

This is group of properties and methods calls, .Result, .GetAwaiter().GetResult() and .Wait().

This is very risky and problematic. If you are running somewhere with a synchronisation context (Windows forms, WFP & ASP) or inside a library that can be used in those kinds of apps, you should absolutely avoid this at all costs. In other words, code like this in a library can deadlock the calling app.

This is where deadlocks can occur as you will prevent any continuations inside the Task returning function from being able to be continued on the current (and now blocked) synchronisation context.

Status: Avoid entirely except in very specific cases where the application model is known, e.g. in the Main method of a console app.

Launch a task with Task.Run.

How does this avoid deadlocks? Task.Run executes on the threadpool, which can change the SynchronizationContext, at the heavy cost of a second thread.

Status: Workable but not performant. Use techniques below instead.

Denial of context

Set the current SynchronizationContext to null, so the code that you call is denied access to it.

var context = SynchronizationContext.Current;

	// do something async

This has the advantage over Task.Run of not using an additional thread for the initial synchronous invocations, and applying this technique at multiple levels comes at no extra cost, whereas each Task.Run with cost yet another thread.

But rather than just set the SynchronizationContext to null, we prefer the technique below, using a custom SynchronizationContext in a library.

Status: Workable. Use techniques below instead.

Run your code in in a single threaded fashion using JoinableTaskFactory

Rather than just mitigating from deadlocks, you can run your code using a JoinableTaskFactory from vs-threading. JoinableTaskFactory internally sets a custom SynchronizationContext that will pass the synchronous continuations back to the blocking thread that was used to initially wait with. This results in a single threaded execution, which is more economical with your precious threads.

Here is an example of it in action:

var context = new JoinableTaskContext();
var jtf = new JoinableTaskFactory(context);
var result = jtf.Run(() => DoSomethingAsync());

Status: Best practice in many cases.

Console application example

Using C# 7.1 or later, you can declare the program entry point as e.g. public static async Task<int> Main(string[] args) and the code to wait for the task will be generated for you. On earlier versions, you can just wait yourself, as is discussed here:

class Program
	static void Main(string[] args)
		var task = AsyncMain();

	private static async Task AsyncMain()
	  // await the rest of the code

It is not a concern that we are blocking on async here: as we are in a console app or windows services where there is no exclusive synchronisation context, there will be no deadlocks, and an additional thread being used is acceptable.

.GetAwaiter().GetResult() is a little nicer than .Result in that it behaves the same, but you will get the first exceptions thrown, instead of an AggregateException.

Thanks to Stuart Lang for correcting mistakes and explaining complexities.