Task为.NET提供了基于任务的异步模式，它不是线程，它运行在线程池的线程上。本着开源的精神， 本文以解读基于.NET4.5 Task源码的方式来揭秘Task的实现原理。
Task的创建
Task的创建方式主要有2种：Task.Run 和Task.Factory.StartNew，各自有不同的overload，这里只解读其中的一种方式，其他有兴趣的请自行解读。
先来看看Task.Run源码：
public static Task Run(Action action, CancellationToken cancellationToken)
{
StackCrawlMark stackMark = StackCrawlMark.LookForMyCaller;
return Task.InternalStartNew((Task) null, (Delegate) action, (object) null, cancellationToken, TaskScheduler.Default, TaskCreationOptions.DenyChildAttach, InternalTaskOptions.None, ref stackMark);
}
调用了Task.InternalStartNew，第一个参数为null，并传入TaskScheduler.Default和TaskCreationOptions.DenyChildAttach.
再来看看Task.Factory.StartNew源码：
public Task StartNew(Action<object> action, object state, CancellationToken cancellationToken)
{
StackCrawlMark stackMark = StackCrawlMark.LookForMyCaller;
Task internalCurrent = Task.InternalCurrent;
return Task.InternalStartNew(internalCurrent, (Delegate) action, state, cancellationToken, this.GetDefaultScheduler(internalCurrent), this.m_defaultCreationOptions, InternalTaskOptions.None, ref stackMark);
}
也是调用Task.InternalStartNew，第一个参数为internalCurrent，当前为null，并传入GetDefaultScheduler(internalCurrent)和m_defaultCreationOptions。
private TaskScheduler GetDefaultScheduler(Task currTask)
{
if (this.m_defaultScheduler != null)
return this.m_defaultScheduler;
if (currTask != null && (currTask.CreationOptions & TaskCreationOptions.HideScheduler) == TaskCreationOptions.None)
return currTask.ExecutingTaskScheduler;
return TaskScheduler.Default;
}
如果internalCurrent不为空而且options是TaskCreationOptions.HideScheduler，那么启用internalCurrent的TaskScheduler。可惜internalCurrent为null，所以启用默认的TaskScheduler，跟入代码发现默认的TaskScheduler是ThreadPoolTaskScheduler，看名字就知道用的是线程池的任务调度，跟“黑盒”传说的一样的。m_defaultCreationOptions在Task.Factory的默认无参构造函数里被赋值TaskCreationOptions.None。
public abstract class TaskScheduler
{
private static readonly ConditionalWeakTable<TaskScheduler, object> s_activeTaskSchedulers = new ConditionalWeakTable<TaskScheduler, object>();
private static readonly TaskScheduler s_defaultTaskScheduler = (TaskScheduler) new ThreadPoolTaskScheduler();
...
}
目前来看两个方法最大的区别在于TaskCreationOption的不同，一个是DenyChildAttach，另一个是None。
接着往下看InternalStartNew：
internal static Task InternalStartNew(Task creatingTask, Delegate action, object state, CancellationToken cancellationToken, TaskScheduler scheduler, TaskCreationOptions options, InternalTaskOptions internalOptions, ref StackCrawlMark stackMark)
{
if (scheduler == null)
throw new ArgumentNullException("scheduler");
Task task = new Task(action, state, creatingTask, cancellationToken, options, internalOptions | InternalTaskOptions.QueuedByRuntime, scheduler);
task.PossiblyCaptureContext(ref stackMark);
task.ScheduleAndStart(false);
return task;
}
首先实例化一个Task：
internal Task(Delegate action, object state, Task parent, CancellationToken cancellationToken, TaskCreationOptions creationOptions, InternalTaskOptions internalOptions, TaskScheduler scheduler)
{
if (action == null)
throw new ArgumentNullException("action");
if ((creationOptions & TaskCreationOptions.AttachedToParent) != TaskCreationOptions.None || (internalOptions & InternalTaskOptions.SelfReplicating) != InternalTaskOptions.None)
this.m_parent = parent;
this.TaskConstructorCore((object) action, state, cancellationToken, creationOptions, internalOptions, scheduler);
}
如果option是AttachToParent，那么internalCurrent就赋值给m_parent，目前为null，SelfReplicating是用来做并行计算的，会在TPL里详解。随后调用TaskConstructorCore。
internal void TaskConstructorCore(object action, object state, CancellationToken cancellationToken, TaskCreationOptions creationOptions, InternalTaskOptions internalOptions, TaskScheduler scheduler)
{
this.m_action = action;
this.m_stateObject = state;
this.m_taskScheduler = scheduler;
if ((creationOptions & ~(TaskCreationOptions.PreferFairness | TaskCreationOptions.LongRunning | TaskCreationOptions.AttachedToParent | TaskCreationOptions.DenyChildAttach | TaskCreationOptions.HideScheduler | TaskCreationOptions.RunContinuationsAsynchronously)) != TaskCreationOptions.None)
throw new ArgumentOutOfRangeException("creationOptions");
if ((creationOptions & TaskCreationOptions.LongRunning) != TaskCreationOptions.None && (internalOptions & InternalTaskOptions.SelfReplicating) != InternalTaskOptions.None)
throw new InvalidOperationException(Environment.GetResourceString("Task_ctor_LRandSR"));
int num = (int) (creationOptions | (TaskCreationOptions) internalOptions);
if (this.m_action == null || (internalOptions & InternalTaskOptions.ContinuationTask) != InternalTaskOptions.None)
num |= ;
this.m_stateFlags = num;
if (this.m_parent != null && (creationOptions & TaskCreationOptions.AttachedToParent) != TaskCreationOptions.None && (this.m_parent.CreationOptions & TaskCreationOptions.DenyChildAttach) == TaskCreationOptions.None)
this.m_parent.AddNewChild();
if (!cancellationToken.CanBeCanceled)
return;
this.AssignCancellationToken(cancellationToken, (Task) null, (TaskContinuation) null);
}
如果options不为DenyChildAttach而且m_parent不为空，则把当前task作为child添加到m_parent。也就是说Task.Run不允许把要执行的task作为当前task的child。
Task已创建，接着调用PossiblyCaptureContext来获取execution context。
internal static ExecutionContext Capture(ref StackCrawlMark stackMark, ExecutionContext.CaptureOptions options)
{
ExecutionContext.Reader executionContextReader = Thread.CurrentThread.GetExecutionContextReader();
if (executionContextReader.IsFlowSuppressed)
return (ExecutionContext) null;
SecurityContext securityContext = SecurityContext.Capture(executionContextReader, ref stackMark);
HostExecutionContext executionContext1 = HostExecutionContextManager.CaptureHostExecutionContext();
SynchronizationContext synchronizationContext = (SynchronizationContext) null;
LogicalCallContext logicalCallContext = (LogicalCallContext) null;
if (!executionContextReader.IsNull)
{
if ((options & ExecutionContext.CaptureOptions.IgnoreSyncCtx) == ExecutionContext.CaptureOptions.None)
synchronizationContext = executionContextReader.SynchronizationContext == null ? (SynchronizationContext) null : executionContextReader.SynchronizationContext.CreateCopy();
if (executionContextReader.LogicalCallContext.HasInfo)
logicalCallContext = executionContextReader.LogicalCallContext.Clone();
}
Dictionary<IAsyncLocal, object> dictionary = (Dictionary<IAsyncLocal, object>) null;
List<IAsyncLocal> asyncLocalList = (List<IAsyncLocal>) null;
if (!executionContextReader.IsNull)
{
dictionary = executionContextReader.DangerousGetRawExecutionContext()._localValues;
asyncLocalList = executionContextReader.DangerousGetRawExecutionContext()._localChangeNotifications;
}
if ((options & ExecutionContext.CaptureOptions.OptimizeDefaultCase) != ExecutionContext.CaptureOptions.None && securityContext == null && (executionContext1 == null && synchronizationContext == null) && ((logicalCallContext == null || !logicalCallContext.HasInfo) && (dictionary == null && asyncLocalList == null)))
return ExecutionContext.s_dummyDefaultEC;
ExecutionContext executionContext2 = new ExecutionContext();
executionContext2.SecurityContext = securityContext;
if (executionContext2.SecurityContext != null)
executionContext2.SecurityContext.ExecutionContext = executionContext2;
executionContext2._hostExecutionContext = executionContext1;
executionContext2._syncContext = synchronizationContext;
executionContext2.LogicalCallContext = logicalCallContext;
executionContext2._localValues = dictionary;
executionContext2._localChangeNotifications = asyncLocalList;
executionContext2.isNewCapture = true;
return executionContext2;
}
ExecutionContext包含了SecurityContext，SynchronizationContext以及LogicalCallContext，其中SynchronizationContext需要做CreateCopy，LogicalCallContext需要做clone，所有这一切都是用户态的，不涉及内核，性能棒棒哒！
接着调用ScheduleAndStart:
internal void ScheduleAndStart(bool needsProtection)
{
if (needsProtection)
{
if (!this.MarkStarted())
return;
}
else
this.m_stateFlags = this.m_stateFlags | ;
if (Task.s_asyncDebuggingEnabled)
Task.AddToActiveTasks(this);
if (AsyncCausalityTracer.LoggingOn && (this.Options & (TaskCreationOptions) ) == TaskCreationOptions.None)
AsyncCausalityTracer.TraceOperationCreation(CausalityTraceLevel.Required, this.Id, "Task: " + ((Delegate) this.m_action).Method.Name, 0UL);
try
{
this.m_taskScheduler.InternalQueueTask(this);
}
catch (ThreadAbortException ex)
{
this.AddException((object) ex);
this.FinishThreadAbortedTask(true, false);
}
catch (System.Exception ex)
{
TaskSchedulerException schedulerException = new TaskSchedulerException(ex);
this.AddException((object) schedulerException);
this.Finish(false);
if ((this.Options & (TaskCreationOptions) ) == TaskCreationOptions.None)
this.m_contingentProperties.m_exceptionsHolder.MarkAsHandled(false);
throw schedulerException;
}
} internal void InternalQueueTask(Task task)
{
task.FireTaskScheduledIfNeeded(this);
this.QueueTask(task);
}
FireTaskScheduledIfNeeded判断是否开启EWT Trace，接着调用ThreadPoolTaskScheduler.QueueTask。
private static readonly ParameterizedThreadStart s_longRunningThreadWork = new ParameterizedThreadStart(ThreadPoolTaskScheduler.LongRunningThreadWork);
private static void LongRunningThreadWork(object obj)
{
(obj as Task).ExecuteEntry(false);
}
protected internal override void QueueTask(Task task)
{
if ((task.Options & TaskCreationOptions.LongRunning) != TaskCreationOptions.None)
{
new Thread(ThreadPoolTaskScheduler.s_longRunningThreadWork)
{
IsBackground = true
}.Start((object) task);
}
else
{
bool forceGlobal = (uint) (task.Options & TaskCreationOptions.PreferFairness) > 0U;
ThreadPool.UnsafeQueueCustomWorkItem((IThreadPoolWorkItem) task, forceGlobal);
}
}
如果options是LongRunning，那么单独创建一个线程执行该任务(ExecuteEntry)，否则就调用ThreadPool.UnsafeQueueCustomWorkItem，这个方法我们熟，还记得在.net线程池内幕里有讲到的global work queue和local work queue吗？给ThreadPool添加一个任务实际上是在global work queue添加一个任务，而task就是往local work queue里添加任务。
ThreadPoolWorkQueue源码：
public void Enqueue(IThreadPoolWorkItem callback, bool forceGlobal)
{
ThreadPoolWorkQueueThreadLocals queueThreadLocals = (ThreadPoolWorkQueueThreadLocals) null;
if (!forceGlobal)
queueThreadLocals = ThreadPoolWorkQueueThreadLocals.threadLocals;
if (this.loggingEnabled)
FrameworkEventSource.Log.ThreadPoolEnqueueWorkObject((object) callback);
if (queueThreadLocals != null)
{
queueThreadLocals.workStealingQueue.LocalPush(callback);
}
else
{
ThreadPoolWorkQueue.QueueSegment comparand = this.queueHead;
while (!comparand.TryEnqueue(callback))
{
Interlocked.CompareExchange<ThreadPoolWorkQueue.QueueSegment>(ref comparand.Next, new ThreadPoolWorkQueue.QueueSegment(), (ThreadPoolWorkQueue.QueueSegment) null);
for (; comparand.Next != null; comparand = this.queueHead)
Interlocked.CompareExchange<ThreadPoolWorkQueue.QueueSegment>(ref this.queueHead, comparand.Next, comparand);
}
}
this.EnsureThreadRequested();
}
由于线程已经执行过任务（global的也有可能是local的），所以代码会走到queueThreadLocals.workStealingQueue.LocalPush(callback)。
internal volatile IThreadPoolWorkItem[] m_array = new IThreadPoolWorkItem[];
private SpinLock m_foreignLock = new SpinLock(false);
public void LocalPush(IThreadPoolWorkItem obj)
{
int num1 = this.m_tailIndex;
if (num1 == int.MaxValue)
{
bool lockTaken = false;
try
{
this.m_foreignLock.Enter(ref lockTaken);
if (this.m_tailIndex == int.MaxValue)
{
this.m_headIndex = this.m_headIndex & this.m_mask;
this.m_tailIndex = num1 = this.m_tailIndex & this.m_mask;
}
}
finally
{
if (lockTaken)
this.m_foreignLock.Exit(true);
}
}
if (num1 < this.m_headIndex + this.m_mask)
{
Volatile.Write<IThreadPoolWorkItem>(ref this.m_array[num1 & this.m_mask], obj);
this.m_tailIndex = num1 + ;
}
else
{
bool lockTaken = false;
try
{
this.m_foreignLock.Enter(ref lockTaken);
int num2 = this.m_headIndex;
int num3 = this.m_tailIndex - this.m_headIndex;
if (num3 >= this.m_mask)
{
IThreadPoolWorkItem[] threadPoolWorkItemArray = new IThreadPoolWorkItem[this.m_array.Length << ];
for (int index = ; index < this.m_array.Length; ++index)
threadPoolWorkItemArray[index] = this.m_array[index + num2 & this.m_mask];
this.m_array = threadPoolWorkItemArray;
this.m_headIndex = ;
this.m_tailIndex = num1 = num3;
this.m_mask = this.m_mask << | ;
}
Volatile.Write<IThreadPoolWorkItem>(ref this.m_array[num1 & this.m_mask], obj);
this.m_tailIndex = num1 + ;
}
finally
{
if (lockTaken)
this.m_foreignLock.Exit(false);
}
}
}
Local work queue(m_array)首先被限死为32，如果queue超过最大数了，则扩大为原来的2倍，以此类推。这里也使用了自旋锁和内存写屏障来代替同步锁提高性能。
至此，task已被创建好，并加入到了ThreadPool的local work queue。那么task是如何被调度的呢？为什么LongRunning就要单独起一个线程去做？请听下回分解!