Task.Yield() in library needs ConfigureWait(false)

2019-04-04 17:49发布

It's recommended that one use ConfigureAwait(false) whenever when you can, especially in libraries because it can help avoid deadlocks and improve performance.

I have written a library that makes heavy use of async (accesses web services for a DB). The users of the library were getting a deadlock and after much painful debugging and tinkering I tracked it down to the single use of await Task.Yield(). Everywhere else that I have an await, I use .ConfigureAwait(false), however that is not supported on Task.Yield().

What is the recommended solution for situations where one needs the equivalent of Task.Yield().ConfigureAwait(false)?

I've read about how there was a SwitchTo method that was removed. I can see why that could be dangerous, but why is there no equivalent of Task.Yield().ConfigureAwait(false)?

Edit:

To provide further context for my question, here is some code. I am implementing an open source library for accessing DynamoDB (a distributed database as a service from AWS) that supports async. A number of operations return IAsyncEnumerable<T> as provided by the IX-Async library. That library doesn't provide a good way of generating async enumerables from data sources that provide rows in "chunks" i.e. each async request returns many items. So I have my own generic type for this. The library supports a read ahead option allowing the user to specify how much data should be requested ahead of when it is actually needed by a call to MoveNext().

Basically, how this works is that I make requests for chunks by calling GetMore() and passing along state between these. I put those tasks in a chunks queue and dequeue them and turn them into actual results that I put in a separate queue. The NextChunk() method is the issue here. Depending on the value of ReadAhead I will keeping getting the next chunk as soon as the last one is done (All) or not until a value is needed but not available (None) or only get the next chunk beyond the values that are currently being used (Some). Because of that, getting the next chunk should run in parallel/not block getting the next value. The enumerator code for this is:

private class ChunkedAsyncEnumerator<TState, TResult> : IAsyncEnumerator<TResult>
{
    private readonly ChunkedAsyncEnumerable<TState, TResult> enumerable;
    private readonly ConcurrentQueue<Task<TState>> chunks = new ConcurrentQueue<Task<TState>>();
    private readonly Queue<TResult> results = new Queue<TResult>();
    private CancellationTokenSource cts = new CancellationTokenSource();
    private TState lastState;
    private TResult current;
    private bool complete; // whether we have reached the end

    public ChunkedAsyncEnumerator(ChunkedAsyncEnumerable<TState, TResult> enumerable, TState initialState)
    {
        this.enumerable = enumerable;
        lastState = initialState;
        if(enumerable.ReadAhead != ReadAhead.None)
            chunks.Enqueue(NextChunk(initialState));
    }

    private async Task<TState> NextChunk(TState state, CancellationToken? cancellationToken = null)
    {
        await Task.Yield(); // ** causes deadlock
        var nextState = await enumerable.GetMore(state, cancellationToken ?? cts.Token).ConfigureAwait(false);
        if(enumerable.ReadAhead == ReadAhead.All && !enumerable.IsComplete(nextState))
            chunks.Enqueue(NextChunk(nextState)); // This is a read ahead, so it shouldn't be tied to our token

        return nextState;
    }

    public Task<bool> MoveNext(CancellationToken cancellationToken)
    {
        cancellationToken.ThrowIfCancellationRequested();

        if(results.Count > 0)
        {
            current = results.Dequeue();
            return TaskConstants.True;
        }
        return complete ? TaskConstants.False : MoveNextAsync(cancellationToken);
    }

    private async Task<bool> MoveNextAsync(CancellationToken cancellationToken)
    {
        Task<TState> nextStateTask;
        if(chunks.TryDequeue(out nextStateTask))
            lastState = await nextStateTask.WithCancellation(cancellationToken).ConfigureAwait(false);
        else
            lastState = await NextChunk(lastState, cancellationToken).ConfigureAwait(false);

        complete = enumerable.IsComplete(lastState);
        foreach(var result in enumerable.GetResults(lastState))
            results.Enqueue(result);

        if(!complete && enumerable.ReadAhead == ReadAhead.Some)
            chunks.Enqueue(NextChunk(lastState)); // This is a read ahead, so it shouldn't be tied to our token

        return await MoveNext(cancellationToken).ConfigureAwait(false);
    }

    public TResult Current { get { return current; } }

    // Dispose() implementation omitted
}

I make no claim this code is perfect. Sorry it is so long, wasn't sure how to simplify. The important part is the NextChunk method and the call to Task.Yield(). This functionality is used through a static construction method:

internal static class AsyncEnumerableEx
{
    public static IAsyncEnumerable<TResult> GenerateChunked<TState, TResult>(
        TState initialState,
        Func<TState, CancellationToken, Task<TState>> getMore,
        Func<TState, IEnumerable<TResult>> getResults,
        Func<TState, bool> isComplete,
        ReadAhead readAhead = ReadAhead.None)
    { ... }
}

3条回答
Animai°情兽
2楼-- · 2019-04-04 18:04

I noticed you edited your question after you accepted the existing answer, so perhaps you're interested in more rants on the subject. Here you go :)

It's recommended that one use ConfigureAwait(false) whenever when you can, especially in libraries because it can help avoid deadlocks and improve performance.

It's recommended so, only if you're absolutely sure that any API your calling in your implementation (including Framework APIs) doesn't depend on any properties of synchronization context. That's especially important for a library code, and even more so if the library is suitable for both client-side and server-side use. E.g, CurrentCulture is a common overlook: it would never be an issue for a desktop app, but it well may be for an ASP.NET app.

Back to your code:

private async Task<TState> NextChunk(...)
{
    await Task.Yield(); // ** causes deadlock
    var nextState = await enumerable.GetMore(...);
    // ...
    return nextState;
}

Most likely, the deadlock is caused by the client of your library, because they use Task.Result (or Task.Wait, Task.WaitAll, Task.IAsyncResult.AsyncWaitHandle etc, let them search) somewhere in the outer frame of the call chain. Albeit Task.Yield() is redundant here, this is not your problem in the first place, but rather theirs: they shouldn't be blocking on the asynchronous APIs and should be using "Async All the Way", as also explained in the Stephen Cleary's article you linked.

Removing Task.Yield() may or may not solve this problem, because enumerable.GetMore() can also use some await SomeApiAsync() without ConfigureAwait(false), thus posting the continuation back to the caller's synchronization context. Moreover, "SomeApiAsync" can happen to be a well established Framework API which is still vulnerable to a deadlock, like SendMailAsync, we'll get back to it later.

Overall, you should only be using Task.Yield() if for some reason you want to return to the caller immediately ("yield" the execution control back to the caller), and then continue asynchronously, at the mercy of the SynchronizationContext installed on the calling thread (or ThreadPool, if SynchronizationContext.Current == null). The continuation well may be executed on the same thread upon the next iteration of the app's core message loop. Some more details can be found here:

So, the right thing would be to avoid blocking code all the way. However, say, you still want to make your code deadlock-proof, you don't care about synchronization context and you're sure the same is true about any system or 3rd party API you use in your implementation.

Then, instead of reinventing ThreadPoolEx.SwitchTo (which was removed for a good reason), you could just use Task.Run, as suggested in the comments:

private Task<TState> NextChunk(...)
{
    // jump to a pool thread without SC to avoid deadlocks
    return Task.Run(async() => 
    {
        var nextState = await enumerable.GetMore(...);
        // ...
        return nextState;
    });
}

IMO, this is still a hack, with the same net effect, although a much more readable one than using a variation of ThreadPoolEx.SwitchTo(). Same as SwitchTo, it still has an associated cost: a redundant thread switch which may hurt ASP.NET performance.

There is another (IMO better) hack, which I proposed here to address the deadlock with aforementioned SendMailAsync. It doesn't incur an extra thread switch:

private Task<TState> NextChunk(...)
{
    return TaskExt.WithNoContext(async() => 
    {
        var nextState = await enumerable.GetMore(...);
        // ...
        return nextState;
    });
}

public static class TaskExt
{
    public static Task<TResult> WithNoContext<TResult>(Func<Task<TResult>> func)
    {
        Task<TResult> task;
        var sc = SynchronizationContext.Current;
        try
        {
            SynchronizationContext.SetSynchronizationContext(null);
            task = func(); // do not await here
        }
        finally
        {
            SynchronizationContext.SetSynchronizationContext(sc);
        }
        return task;
    }
}

This hack works in the way it temporarily removes the synchronization context for the synchronous scope of the original NextChunk method, so it won't be captured for the 1st await continuation inside the async lambda, effectively solving the deadlock problem.

Stephen has provided a slightly different implementation while answering the same question. His IgnoreSynchronizationContext restores the original synchronization context on whatever happens to be the continuation's thread after await (which could be a completely different, random pool thread). I'd rather not restore it after await at all, as long as I don't care about it.

查看更多
Ridiculous、
3楼-- · 2019-04-04 18:04

Inasmuch as the useful and legit API you're looking for is missing, I filed this request proposing its addition to .NET.

I also added it to vs-threading so that the next release of the Microsoft.VisualStudio.Threading NuGet package will include this API. Note that this library is not VS-specific, so you can use it in your app.

查看更多
贪生不怕死
4楼-- · 2019-04-04 18:11

The exact equivalent of Task.Yield().ConfigureAwait(false) (which doesn't exist since ConfigureAwait is a method on Task and Task.Yield returns a custom awaitable) is simply using Task.Factory.StartNew with CancellationToken.None, TaskCreationOptions.PreferFairness and TaskScheduler.Current. In most cases however, Task.Run (which uses the default TaskScheduler) is close enough.

You can verify that by looking at the source for YieldAwaiter and see that it uses ThreadPool.QueueUserWorkItem/ThreadPool.UnsafeQueueUserWorkItem when TaskScheduler.Current is the default one (i.e. thread pool) and Task.Factory.StartNew when it isn't.

You can however create your own awaitable (as I did) that mimics YieldAwaitable but disregards the SynchronizationContext:

async Task Run(int input)
{
    await new NoContextYieldAwaitable();
    // executed on a ThreadPool thread
}

public struct NoContextYieldAwaitable
{
    public NoContextYieldAwaiter GetAwaiter() { return new NoContextYieldAwaiter(); }
    public struct NoContextYieldAwaiter : INotifyCompletion
    {
        public bool IsCompleted { get { return false; } }
        public void OnCompleted(Action continuation)
        {
            var scheduler = TaskScheduler.Current;
            if (scheduler == TaskScheduler.Default)
            {
                ThreadPool.QueueUserWorkItem(RunAction, continuation);
            }
            else
            {
                Task.Factory.StartNew(continuation, CancellationToken.None, TaskCreationOptions.PreferFairness, scheduler);
            }
        }

        public void GetResult() { }
        private static void RunAction(object state) { ((Action)state)(); }
    }
}

Note: I don't recommend actually using NoContextYieldAwaitable, it's just an answer to your question. You should be using Task.Run (or Task.Factory.StartNew with a specific TaskScheduler)

查看更多
登录 后发表回答