一个简单的模拟实例说明Task及其调度问题

Task对于.NET的重要性毋庸置疑。通过最近的一些面试经历，发现很多人对与Task及其调度机制，以及线程和线程池之间的关系并没有清晰的认识。本文采用最简单的方式模拟了Task的实现，旨在说明Task是什么？它是如何被调度执行的？源代码从这里下载。

一、Task（Job）
二、TaskScheduler（JobScheduler）
三、基于线程池的调度
四、使用指定线程进行调度
五、异步等待
六、await关键字的运用
七、状态机

一、Task（Job）

Task代表一项具有某种状态的操作，我们使用如下这个Job类型来模拟Task。Job封装的操作体现为一个Action委托，状态则通过JobStatus枚举来表示（对应TaskStatus枚举）。简单起见，我们仅仅定义了四种状态（创建、调度、执行和完成）。Invoke方法负责执行封装的Action委托，并对状态进行相应设置。

public class Job
{
    private readonly Action _work;
    public Job(Action work)=> _work = work;
    public JobStatus Status { get; internal set; }

    internal protected virtual void Invoke()
    {
        Status = JobStatus.Running;
        _work();
        Status = JobStatus.Completed;    
    }
}

public enum JobStatus
{
    Created,
    Scheduled,
    Running,
    Completed
}

二、TaskScheduler（JobScheduler）

Task承载的操作通过调度得以执行，具体的调度策略取决于调度器的选择。Task调度器通过TaskScheduler表示，我们利用如下这个JobScheduler类型对它进行模拟。如下面的代码片段所示，我们只为抽象类JobScheduler定义了唯一的QueueJob方法来调度作为参数的Job对象。静态Current属性表示当前默认实现的调度器。

public abstract class JobScheduler
{
    public abstract void QueueJob(Job job);
    public static JobScheduler Current { get; set; } = new ThreadPoolJobScheduler ();
}

对于开发者来说，执行Task就是将它提交给调度器，这一操作体现在我们为Job类型定义的静态Start方法中。该方法通过参数指定具体的调度器，如果没有显式指定，默认采用JobScheduler的Current静态属性设置的默认调度器。为了方便后面的演示，我们还定义了一个静态的Run方法，该方法会将指定的Action对象封装成Job，并调用Start方法利用默认的调度器进行调度。

public class Job
{
    private readonly Action _work;
    public Job(Action work)=> _work = work;
    public JobStatus Status { get; internal set; }

    internal protected virtual void Invoke()
    {
        Status = JobStatus.Running;
        _work();
        Status = JobStatus.Completed;

    }

    public void Start(JobScheduler? scheduler = null) => (scheduler ?? JobScheduler.Current).QueueJob(this);
    public static Job Run(Action work)
    {
        var job = new Job(work);
        job.Start();
        return job;
    }
}

三、基于线程池的调度

Task如何执行取决于选择怎样的调度器，.NET默认采用基于线程池的调度策略，这一策略体现在ThreadPoolTaskScheduler类型上，我们使用如下这个ThreadPoolJobScheduler 进行模拟。如下面的代码片段所示，重写的QueueJob方法通过调用ThreadPool.QueueUserWorkItem方法执行指定Job对象封装的Action委托。JobScheduler的Current属性设置的默认调度器就是这么一个ThreadPoolJobScheduler 对象。

public class ThreadPoolJobScheduler : JobScheduler
{
    public override void QueueJob(Job job)
    {
        job.Status = JobStatus.Scheduled;
        var executionContext = ExecutionContext.Capture();
        ThreadPool.QueueUserWorkItem(_ => ExecutionContext.Run(executionContext!, _ => job.Invoke(), null));
    }
}

我们按照如下的方式调用Job的静态Run方法创建并执行了三个Job，每个Job封装的Action委托在执行的时候会将当前线程ID打印出来。

_ = Job.Run(() => Console.WriteLine($"Job1 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job2 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job3 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));

Console.ReadLine();

由于采用默认的基于线程池的调度策略，所以三个Job会在三个不同的线程上执行。

四、使用指定线程进行调度

我们知道.NET进程只有一个全局的线程池，对于一些需要长时间运行且具有较高优先级的操作，采用基于线程池的调用未必是好的选择。比如在一个Web应用中，线程池的工作线程会被用来处理请求，对于一个需要持续运行的Job可能会因为可用工作线程的不足而被阻塞。.NET对于这种情况具有不同的处理方式（启动Task的时候选择TaskCreationOptions.LongRunning选项），这里我们使用自定义调度器的方式来解决这个问题。如下这个DedicatedThreadJobScheduler 利用创建的“专有线程”来保证被调用的Job能够“立即”执行。线程的数量通过构造函数的参数指定，线程在无事可做的时候被“挂起”以及有新的Job被调度时被“复苏”通过一个ManualResetEvent对象来完成。

public class DedicatedThreadJobScheduler : JobScheduler
{
    private readonly Queue<Job>[] _queues;
    private readonly Thread[] _threads;
    private readonly ManualResetEvent[] _events;
    public DedicatedThreadJobScheduler (int threadCount)
    {
        _queues = new Queue<Job>[threadCount];
        _threads = new  Thread[threadCount];
        _events = new ManualResetEvent[threadCount];

        for (int index = 0; index < threadCount; index++)
        {
            var queue = _queues[index] = new Queue<Job>();
            var thread  = _threads[index] = new Thread(Invoke);
            _events[index] = new ManualResetEvent(true);
            thread.Start(index);
        }

        void Invoke(object? state)
        {
            var index = (int)state!;
            var @event = _events[index];
            while (true)
            {
                if (@event.WaitOne())
                {
                    while (true)
                    {
                        if (!_queues[index].TryDequeue(out var job))
                        {
                            Suspend(index);
                            break;
                        }
                        job.Invoke();
                    }
                }
            }
        }
    }
    public override void QueueJob(Job job)
    {
        job.Status = JobStatus.Scheduled;
        var (queue, index) =  _queues.Select((queue, index) => (queue, index)).OrderBy(it => it.queue.Count).First();
        queue.Enqueue(job);
        Resume(index);
    }

    public void Suspend(int index) => _events[index].Reset();
    public void Resume(int index) => _events[index].Set();
}

还是上面演示的程序，这次我们将当前调度器设置为上面这个DedicatedThreadJobScheduler ，并将使用的线程数设置为2。

JobScheduler.Current = new DedicatedThreadJobScheduler (2);
_ = Job.Run(() => Console.WriteLine($"Job1 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job2 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job3 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job4 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job5 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));
_ = Job.Run(() => Console.WriteLine($"Job6 is excuted in thread {Thread.CurrentThread.ManagedThreadId}"));

Console.ReadLine();

我们会发现所有的操作只会在两个固定的线程中被执行。

五、异步等待

如果需要在某个Task执行之后接着执行后续的操作，我们可以调用其ContinueWith方法指定待执行的操作，现在我们将这个方法定义Job类型上。Job与Task的ContinueWith有些差异，在这里我们认为ContinueWith指定的也是一个Job，那么多个Job则可以按照预先编排的顺序构成一个链表。当前Job执行后，只需要将后续这个Job交付给调度器就可以了。如下面的代码片段所示，我们利用_continue字段来表示异步等待执行的Job，并利用它维持一个Job链表。ContinueWith方法会将指定的Action委托封装成Job并添加到链表末端。

public class Job
{
    private readonly Action _work;
    private Job? _continue;
    public Job(Action work) => _work = work;
    public JobStatus Status { get; internal set; }
    public void Start(JobScheduler? scheduler = null) => (scheduler ?? JobScheduler.Current).QueueJob(this);
    internal protected virtual void Invoke()
    {
        Status = JobStatus.Running;
        _work();
        Status = JobStatus.Completed;
        _continue?.Start();
    }

    public static Job Run(Action work)
    {
        var job = new Job(work);
        job.Start();
        return job;
    }

    public Job ContinueWith(Action<Job> continuation)
    {
        if (_continue == null)
        {
            var job = new Job(() => continuation(this));
            _continue = job;
        }
        else
        {
            _continue.ContinueWith(continuation);
        }
        return this;
    }
}

利用ContinueWith方法实现异步操作的按序执行体现在如下的程序中。

Job.Run(() =>{
    Thread.Sleep(1000);
    Console.WriteLine("Foo1");
}).ContinueWith(_ =>{
    Thread.Sleep(100);
    Console.WriteLine("Bar1");
}).ContinueWith(_ =>{
    Thread.Sleep(100);
    Console.WriteLine("Baz1");
});


Job.Run(() =>{
    Thread.Sleep(100);
    Console.WriteLine("Foo2");
}).ContinueWith(_ =>{
    Thread.Sleep(10);
    Console.WriteLine("Bar2");
}).ContinueWith(_ =>{
    Thread.Sleep(10);
    Console.WriteLine("Baz2");
});

Console.ReadLine();

输出结果

六、await关键字的运用

虽然ContinueWith方法能够解决“异步等待”的问题，但是我们更喜欢使用await关键字，接下来我们就为Job赋予这个能力。为此我们定义了如下这个实现了ICriticalNotifyCompletion接口的JobAwaiter结构体。顾名思义，该接口用来发送操作完成的通知。一个JobAwaiter对象由一个Job对象构建而成，当它自身执行完成之后，OnCompleted方法会被调用，我们利用它执行后续的操作。

public struct JobAwaiter: ICriticalNotifyCompletion
{
    private readonly Job _job;
    public bool IsCompleted => _job.Status ==  JobStatus.Completed;
    public JobAwaiter(Job job)
    {
        _job = job;
        if (job.Status == JobStatus.Created)
        {
            job.Start();
        }
    }
    public void OnCompleted(Action continuation)
    {
        _job.ContinueWith(_ => continuation());
    }
    public void GetResult() { }
    public void UnsafeOnCompleted(Action continuation)=>OnCompleted(continuation);
}

我们在Job类型上添加这个GetAwaiter方法返回根据自身创建的JobAwaiter对象。

public class Job
{
    private readonly Action _work;
    private Job? _continue;
    public Job(Action work) => _work = work;
    public JobStatus Status { get; internal set; }
    public void Start(JobScheduler? scheduler = null) => (scheduler ?? JobScheduler.Current).QueueJob(this);
    internal protected virtual void Invoke()
    {
        Status = JobStatus.Running;
        _work();
        Status = JobStatus.Completed;
        _continue?.Start();
    }


    public static Job Run(Action work)
    {
        var job = new Job(work);
        job.Start();
        return job;
    }
    public Job ContinueWith(Action<Job> continuation)
    {
        if (_continue == null)
        {
            var job = new Job(() => continuation(this));
            _continue = job;
        }
        else
        {
            _continue.ContinueWith(continuation);
        }
        return this;
    }
    public JobAwaiter GetAwaiter() => new(this);
}

任何一个类型一旦拥有了这样一个GetAwaiter方法，我们就能将await关键词应用在对应的对象上面。

await Foo();
await Bar();
await Baz();

Console.ReadLine();

static Job Foo() =>  new Job(() =>
{
    Thread.Sleep(1000);
    Console.WriteLine("Foo");
});

static Job Bar() => new Job(() =>
{
    Thread.Sleep(100);
    Console.WriteLine("Bar");
});

static Job Baz() => new Job(() =>
{
    Thread.Sleep(10);
    Console.WriteLine("Baz");
});

输出结果：

七、状态机

我想你应该知道await关键字仅仅是编译器提供的语法糖，编译后的代码会利用一个“状态机”实现“异步等待”的功能，上面这段代码最终编译成如下的形式。值得一提的是，Debug和Release模式编译出来的代码是不同的，下面给出的是Release模式下的编译结果，上述的状态机体现为生成的<<Main>$>d__0这个结构体。它的实现其实很简单：如果个方法出现了N个await关键字，它们相当于将整个方法的执行流程切割成N+1段，状态机的状态体现为当前应该执行那段，具体的执行体现在MoveNext方法上。GetAwaiter方法返回的ICriticalNotifyCompletion对象用来确定当前操作是否结束，如果结束则可以直接指定后续操作，否则需要调用AwaitUnsafeOnCompleted对后续操作进行处理。

// Program
using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Runtime.InteropServices;
using System.Threading.Tasks;
using Jobs;

[CompilerGenerated]
internal class Program
{
	[StructLayout(LayoutKind.Auto)]
	[CompilerGenerated]
	private struct <<Main>$>d__0 : IAsyncStateMachine
	{
		public int <>1__state;

		public AsyncTaskMethodBuilder <>t__builder;

		private JobAwaiter <>u__1;

		private void MoveNext()
		{
			int num = <>1__state;
			try
			{
				JobAwaiter awaiter;
				switch (num)
				{
				default:
					awaiter = <<Main>$>g__Foo|0_0().GetAwaiter();
					if (!awaiter.IsCompleted)
					{
						num = (<>1__state = 0);
						<>u__1 = awaiter;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
						return;
					}
					goto IL_006c;
				case 0:
					awaiter = <>u__1;
					<>u__1 = default(JobAwaiter);
					num = (<>1__state = -1);
					goto IL_006c;
				case 1:
					awaiter = <>u__1;
					<>u__1 = default(JobAwaiter);
					num = (<>1__state = -1);
					goto IL_00c6;
				case 2:
					{
						awaiter = <>u__1;
						<>u__1 = default(JobAwaiter);
						num = (<>1__state = -1);
						break;
					}
					IL_00c6:
					awaiter.GetResult();
					awaiter = <<Main>$>g__Baz|0_2().GetAwaiter();
					if (!awaiter.IsCompleted)
					{
						num = (<>1__state = 2);
						<>u__1 = awaiter;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
						return;
					}
					break;
					IL_006c:
					awaiter.GetResult();
					awaiter = <<Main>$>g__Bar|0_1().GetAwaiter();
					if (!awaiter.IsCompleted)
					{
						num = (<>1__state = 1);
						<>u__1 = awaiter;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref this);
						return;
					}
					goto IL_00c6;
				}
				awaiter.GetResult();
				Console.ReadLine();
			}
			catch (Exception exception)
			{
				<>1__state = -2;
				<>t__builder.SetException(exception);
				return;
			}
			<>1__state = -2;
			<>t__builder.SetResult();
		}

		void IAsyncStateMachine.MoveNext()
		{
			//ILSpy generated this explicit interface implementation from .override directive in MoveNext
			this.MoveNext();
		}

		[DebuggerHidden]
		private void SetStateMachine([System.Runtime.CompilerServices.Nullable(1)] IAsyncStateMachine stateMachine)
		{
			<>t__builder.SetStateMachine(stateMachine);
		}

		void IAsyncStateMachine.SetStateMachine([System.Runtime.CompilerServices.Nullable(1)] IAsyncStateMachine stateMachine)
		{
			//ILSpy generated this explicit interface implementation from .override directive in SetStateMachine
			this.SetStateMachine(stateMachine);
		}
	}

	[AsyncStateMachine(typeof(<<Main>$>d__0))]
	private static Task <Main>$(string[] args)
	{
		<<Main>$>d__0 stateMachine = default(<<Main>$>d__0);
		stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
		stateMachine.<>1__state = -1;
		stateMachine.<>t__builder.Start(ref stateMachine);
		return stateMachine.<>t__builder.Task;
	}

	[SpecialName]
	private static void <Main>(string[] args)
	{
		<Main>$(args).GetAwaiter().GetResult();
	}
}

上面提到过，编译器生成的状态机代码在Debug和Release模式是不一样的。在Release模式下状态机是一个结构体，虽然是以接口ICriticalNotifyCompletion的方式使用它，但是由于使用了ref关键字，所以不会涉及装箱，所以不会对GC造成任何影响。但是Debug模式下生成的状态机则是一个类（如下所示），将会涉及针对堆内存的分配和回收。对于遍布await关键字的应用程序，两者之间的性能差异肯定是不同的。实际上针对Task的很多优化策略，比如使用ValueTask，对某些Task<T>对象（比如状态为Completed的Task<bool>对象）的复用，以及使用IValueTaskSource等，都是为了解决内存分配的问题。

// Program
using System;
using System.Diagnostics;
using System.Runtime.CompilerServices;
using System.Threading.Tasks;
using Jobs;

[CompilerGenerated]
internal class Program
{
	[CompilerGenerated]
	private sealed class <<Main>$>d__0 : IAsyncStateMachine
	{
		public int <>1__state;

		public AsyncTaskMethodBuilder <>t__builder;

		public string[] args;

		private JobAwaiter <>u__1;

		private void MoveNext()
		{
			int num = <>1__state;
			try
			{
				JobAwaiter awaiter3;
				JobAwaiter awaiter2;
				JobAwaiter awaiter;
				switch (num)
				{
				default:
					awaiter3 = <<Main>$>g__Foo|0_0().GetAwaiter();
					if (!awaiter3.IsCompleted)
					{
						num = (<>1__state = 0);
						<>u__1 = awaiter3;
						<<Main>$>d__0 stateMachine = this;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter3, ref stateMachine);
						return;
					}
					goto IL_007e;
				case 0:
					awaiter3 = <>u__1;
					<>u__1 = default(JobAwaiter);
					num = (<>1__state = -1);
					goto IL_007e;
				case 1:
					awaiter2 = <>u__1;
					<>u__1 = default(JobAwaiter);
					num = (<>1__state = -1);
					goto IL_00dd;
				case 2:
					{
						awaiter = <>u__1;
						<>u__1 = default(JobAwaiter);
						num = (<>1__state = -1);
						break;
					}
					IL_00dd:
					awaiter2.GetResult();
					awaiter = <<Main>$>g__Baz|0_2().GetAwaiter();
					if (!awaiter.IsCompleted)
					{
						num = (<>1__state = 2);
						<>u__1 = awaiter;
						<<Main>$>d__0 stateMachine = this;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter, ref stateMachine);
						return;
					}
					break;
					IL_007e:
					awaiter3.GetResult();
					awaiter2 = <<Main>$>g__Bar|0_1().GetAwaiter();
					if (!awaiter2.IsCompleted)
					{
						num = (<>1__state = 1);
						<>u__1 = awaiter2;
						<<Main>$>d__0 stateMachine = this;
						<>t__builder.AwaitUnsafeOnCompleted(ref awaiter2, ref stateMachine);
						return;
					}
					goto IL_00dd;
				}
				awaiter.GetResult();
				Console.ReadLine();
			}
			catch (Exception exception)
			{
				<>1__state = -2;
				<>t__builder.SetException(exception);
				return;
			}
			<>1__state = -2;
			<>t__builder.SetResult();
		}

		void IAsyncStateMachine.MoveNext()
		{
			//ILSpy generated this explicit interface implementation from .override directive in MoveNext
			this.MoveNext();
		}

		[DebuggerHidden]
		private void SetStateMachine([System.Runtime.CompilerServices.Nullable(1)] IAsyncStateMachine stateMachine)
		{
		}

		void IAsyncStateMachine.SetStateMachine([System.Runtime.CompilerServices.Nullable(1)] IAsyncStateMachine stateMachine)
		{
			//ILSpy generated this explicit interface implementation from .override directive in SetStateMachine
			this.SetStateMachine(stateMachine);
		}
	}

	[AsyncStateMachine(typeof(<<Main>$>d__0))]
	[DebuggerStepThrough]
	private static Task <Main>$(string[] args)
	{
		<<Main>$>d__0 stateMachine = new <<Main>$>d__0();
		stateMachine.<>t__builder = AsyncTaskMethodBuilder.Create();
		stateMachine.args = args;
		stateMachine.<>1__state = -1;
		stateMachine.<>t__builder.Start(ref stateMachine);
		return stateMachine.<>t__builder.Task;
	}

	[SpecialName]
	[DebuggerStepThrough]
	private static void <Main>(string[] args)
	{
		<Main>$(args).GetAwaiter().GetResult();
	}
}