閱讀目錄：
基礎
核心鎖：基於核心物件構造的鎖機制，就是通常說的核心構造模式。使用者模式構造和核心模式構造
優點：cpu利用最大化。它發現資源被鎖住，請求就排隊等候。執行緒切換到別處幹活，直到接受到可用訊號，執行緒再切回來繼續處理請求。
缺點：託管程式碼->使用者模式程式碼->核心程式碼損耗、執行緒上下文切換損耗。
在鎖的時間比較短時，系統頻繁忙於休眠、切換，是個很大的效能損耗。
自旋鎖：原子操作+自迴圈。通常說的使用者構造模式。 執行緒不休眠，一直迴圈嘗試對資源訪問，直到可用。
優點：完美解決核心鎖的缺點。
缺點：長時間一直迴圈會導致cpu的白白浪費，高併發競爭下、CPU的消耗特別嚴重。
混合鎖：核心鎖+自旋鎖。 混合鎖是先自旋鎖一段時間或自旋多少次，再轉成核心鎖。
優點：核心鎖和自旋鎖的折中方案，利用前二者優點，避免出現極端情況（自旋時間過長，核心鎖時間過短）。
缺點： 自旋多少時間、自旋多少次，這些策略很難把控。
在作業系統及net框架層，這塊演算法策略做的已經非常優了，有些API函式也提供了時間及次數可配置項，讓使用者根據需求自行判斷。
自旋鎖示例
來看下我們自己簡單實現的自旋鎖：
int signal = 0; var li = new List<int>(); Parallel.For(0, 1000 * 10000, r => { while (Interlocked.Exchange(ref signal, 1) != 0)//加自旋鎖 { //黑魔法 } li.Add(r); Interlocked.Exchange(ref signal, 0); //釋放鎖 }); Console.WriteLine(li.Count); //輸出：10000000
上面就是自旋鎖：Interlocked.Exchange+while
1：定義signal 0可用，1不可用。
2：Parallel模擬併發競爭，原子更改signal狀態。 後續執行緒自旋訪問signal，是否可用。
3：A執行緒使用完後，更改signal為0。 剩餘執行緒競爭訪問資源，B執行緒勝利後，更改signal為1，失敗執行緒繼續自旋，直到可用。
SpinLock
SpinLock是net4.0後Net提供的自旋鎖類庫，內部做了優化。
簡單看下例項：
var li = new List<int>(); var sl = new SpinLock(); Parallel.For(0, 1000 * 10000, r => { bool gotLock = false; //釋放成功 sl.Enter(ref gotLock); //進入鎖 li.Add(r); if (gotLock) sl.Exit(); //釋放 }); Console.WriteLine(li.Count); //輸出：10000000
繼續SpinLock
new SpinLock(false) 這個建構函式主要用來檢查死鎖用，true是開啟。
在開啟狀態下，一旦發生死鎖會直接拋異常的。
SpinLock實現的部分原始碼：
public void Enter(ref bool lockTaken) { if (lockTaken) { lockTaken = false; throw new System.ArgumentException(Environment.GetResourceString("SpinLock_TryReliableEnter_ArgumentException")); } // Fast path to acquire the lock if the lock is released // If the thread tracking enabled set the new owner to the current thread id // Id not, set the anonymous bit lock int observedOwner = m_owner; int newOwner = 0; bool threadTrackingEnabled = (m_owner & LOCK_ID_DISABLE_MASK) == 0; if (threadTrackingEnabled) { if (observedOwner == LOCK_UNOWNED) newOwner = Thread.CurrentThread.ManagedThreadId; } else if ((observedOwner & LOCK_ANONYMOUS_OWNED) == LOCK_UNOWNED) { newOwner = observedOwner | LOCK_ANONYMOUS_OWNED; // set the lock bit } if (newOwner != 0) { #if !FEATURE_CORECLR Thread.BeginCriticalRegion(); #endif #if PFX_LEGACY_3_5 if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner) == observedOwner) { lockTaken = true; return; } #else if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner, ref lockTaken) == observedOwner) { // Fast path succeeded return; } #endif #if !FEATURE_CORECLR Thread.EndCriticalRegion(); #endif } //Fast path failed, try slow path ContinueTryEnter(Timeout.Infinite, ref lockTaken); } private void ContinueTryEnter(int millisecondsTimeout, ref bool lockTaken) { long startTicks = 0; if (millisecondsTimeout != Timeout.Infinite && millisecondsTimeout != 0) { startTicks = DateTime.UtcNow.Ticks; } #if !FEATURE_PAL && !FEATURE_CORECLR // PAL doesn't support eventing, and we don't compile CDS providers for Coreclr if (CdsSyncEtwBCLProvider.Log.IsEnabled()) { CdsSyncEtwBCLProvider.Log.SpinLock_FastPathFailed(m_owner); } #endif if (IsThreadOwnerTrackingEnabled) { // Slow path for enabled thread tracking mode ContinueTryEnterWithThreadTracking(millisecondsTimeout, startTicks, ref lockTaken); return; } // then thread tracking is disabled // In this case there are three ways to acquire the lock // 1- the first way the thread either tries to get the lock if it's free or updates the waiters, if the turn >= the processors count then go to 3 else go to 2 // 2- In this step the waiter threads spins and tries to acquire the lock, the number of spin iterations and spin count is dependent on the thread turn // the late the thread arrives the more it spins and less frequent it check the lock avilability // Also the spins count is increaes each iteration // If the spins iterations finished and failed to acquire the lock, go to step 3 // 3- This is the yielding step, there are two ways of yielding Thread.Yield and Sleep(1) // If the timeout is expired in after step 1, we need to decrement the waiters count before returning int observedOwner; //***Step 1, take the lock or update the waiters // try to acquire the lock directly if possoble or update the waiters count SpinWait spinner = new SpinWait(); while (true) { observedOwner = m_owner; if ((observedOwner & LOCK_ANONYMOUS_OWNED) == LOCK_UNOWNED) { #if !FEATURE_CORECLR Thread.BeginCriticalRegion(); #endif #if PFX_LEGACY_3_5 if (Interlocked.CompareExchange(ref m_owner, observedOwner | 1, observedOwner) == observedOwner) { lockTaken = true; return; } #else if (Interlocked.CompareExchange(ref m_owner, observedOwner | 1, observedOwner, ref lockTaken) == observedOwner) { return; } #endif #if !FEATURE_CORECLR Thread.EndCriticalRegion(); #endif } else //failed to acquire the lock,then try to update the waiters. If the waiters count reached the maximum, jsut break the loop to avoid overflow if ((observedOwner & WAITERS_MASK) == MAXIMUM_WAITERS || Interlocked.CompareExchange(ref m_owner, observedOwner + 2, observedOwner) == observedOwner) break; spinner.SpinOnce(); } // Check the timeout. if (millisecondsTimeout == 0 || (millisecondsTimeout != Timeout.Infinite && TimeoutExpired(startTicks, millisecondsTimeout))) { DecrementWaiters(); return; } //***Step 2. Spinning //lock acquired failed and waiters updated int turn = ((observedOwner + 2) & WAITERS_MASK) / 2; int processorCount = PlatformHelper.ProcessorCount; if (turn < processorCount) { int processFactor = 1; for (int i = 1; i <= turn * SPINNING_FACTOR; i++) { Thread.SpinWait((turn + i) * SPINNING_FACTOR * processFactor); if (processFactor < processorCount) processFactor++; observedOwner = m_owner; if ((observedOwner & LOCK_ANONYMOUS_OWNED) == LOCK_UNOWNED) { #if !FEATURE_CORECLR Thread.BeginCriticalRegion(); #endif int newOwner = (observedOwner & WAITERS_MASK) == 0 ? // Gets the number of waiters, if zero observedOwner | 1 // don't decrement it. just set the lock bit, it is zzero because a previous call of Exit(false) ehich corrupted the waiters : (observedOwner - 2) | 1; // otherwise decrement the waiters and set the lock bit Contract.Assert((newOwner & WAITERS_MASK) >= 0); #if PFX_LEGACY_3_5 if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner) == observedOwner) { lockTaken = true; return; } #else if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner, ref lockTaken) == observedOwner) { return; } #endif #if !FEATURE_CORECLR Thread.EndCriticalRegion(); #endif } } } // Check the timeout. if (millisecondsTimeout != Timeout.Infinite && TimeoutExpired(startTicks, millisecondsTimeout)) { DecrementWaiters(); return; } //*** Step 3, Yielding //Sleep(1) every 50 yields int yieldsoFar = 0; while (true) { observedOwner = m_owner; if ((observedOwner & LOCK_ANONYMOUS_OWNED) == LOCK_UNOWNED) { #if !FEATURE_CORECLR Thread.BeginCriticalRegion(); #endif int newOwner = (observedOwner & WAITERS_MASK) == 0 ? // Gets the number of waiters, if zero observedOwner | 1 // don't decrement it. just set the lock bit, it is zzero because a previous call of Exit(false) ehich corrupted the waiters : (observedOwner - 2) | 1; // otherwise decrement the waiters and set the lock bit Contract.Assert((newOwner & WAITERS_MASK) >= 0); #if PFX_LEGACY_3_5 if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner) == observedOwner) { lockTaken = true; return; } #else if (Interlocked.CompareExchange(ref m_owner, newOwner, observedOwner, ref lockTaken) == observedOwner) { return; } #endif #if !FEATURE_CORECLR Thread.EndCriticalRegion(); #endif } if (yieldsoFar % SLEEP_ONE_FREQUENCY == 0) { Thread.Sleep(1); } else if (yieldsoFar % SLEEP_ZERO_FREQUENCY == 0) { Thread.Sleep(0); } else { #if PFX_LEGACY_3_5 Platform.Yield(); #else Thread.Yield(); #endif } if (yieldsoFar % TIMEOUT_CHECK_FREQUENCY == 0) { //Check the timeout. if (millisecondsTimeout != Timeout.Infinite && TimeoutExpired(startTicks, millisecondsTimeout)) { DecrementWaiters(); return; } } yieldsoFar++; } } /// <summary> /// decrements the waiters, in case of the timeout is expired /// </summary> private void DecrementWaiters() { SpinWait spinner = new SpinWait(); while (true) { int observedOwner = m_owner; if ((observedOwner & WAITERS_MASK) == 0) return; // don't decrement the waiters if it's corrupted by previous call of Exit(false) if (Interlocked.CompareExchange(ref m_owner, observedOwner - 2, observedOwner) == observedOwner) { Contract.Assert(!IsThreadOwnerTrackingEnabled); // Make sure the waiters never be negative which will cause the thread tracking bit to be flipped break; } spinner.SpinOnce(); } }
從程式碼中發現SpinLock並不是簡單的實現那樣一直自旋，其內部做了很多優化。
1：內部使用了Interlocked.CompareExchange保持原子操作， m_owner 0可用，1不可用。
2：第一次獲得鎖失敗後，繼續呼叫ContinueTryEnter，ContinueTryEnter有三種獲得鎖的情況。
3：ContinueTryEnter函式第一種獲得鎖的方式，使用了while+SpinWait。
4：第一種方式達到最大等待者數量後，命中走第二種。 繼續自旋 turn * 100次。100這個值是處理器核數(4, 8 ,16)下最好的。
5：第二種如果還不能獲得鎖，走第三種。這種就帶有混合構造的意思了，如下：
if (yieldsoFar % 40 == 0) Thread.Sleep(1); else if (yieldsoFar % 10 == 0) Thread.Sleep(0); else Thread.Yield();
Thread.Sleep(1) ： 終止當前執行緒，放棄剩下時間片 休眠1毫秒， 退出跟其他執行緒搶佔cpu。當然這個一般會更多，系統無法保證這麼細的時間粒度。
Thread.Sleep(0)： 終止當前執行緒，放棄剩下時間片。 但立馬還會跟其他執行緒搶cpu，能不能搶到跟執行緒優先順序有關。
Thread.Yeild()： 結束當前執行緒，讓出CPU給其他準備好的執行緒。其他執行緒ok後或沒有還沒有準備好，繼續執行當前，Thread.Yeild()會返回個bool值，表示CPU是否讓出成功。
從原始碼中可以學到不少程式設計技巧，比如可以借鑑自旋+Thread.Yeild() 或 while+Thread.Yeild()等組合使用方式。
總結
本章介紹了自旋鎖的基礎及樓主的經驗。 關於SpinLock類原始碼這塊，只簡單理解了下並沒有深究。
測試了下SpinLock和自己實現的自旋鎖效能對比（並行新增1000w List<int>()），SpinLock是單純的自旋鎖效能2倍以上。
另外測試了lock的效能，是系統SpinLock效能的3倍以上，可見lock內部自旋的效率更高，CLR暫沒開源，看不到CLR具體實現的程式碼。
參考http://www.projky.com/dotnet/4.0/System/Threading/SpinLock.cs.html