// Licensed to the .NET Foundation under one or more agreements. // The .NET Foundation licenses this file to you under the MIT license. // See the LICENSE file in the project root for more information. // // THREADS.CPP // // // #include "common.h" #include "frames.h" #include "threads.h" #include "stackwalk.h" #include "excep.h" #include "comsynchronizable.h" #include "log.h" #include "gcheaputilities.h" #include "mscoree.h" #include "dbginterface.h" #include "corprof.h" // profiling #include "eeprofinterfaces.h" #include "eeconfig.h" #include "perfcounters.h" #include "corhost.h" #include "win32threadpool.h" #include "jitinterface.h" #include "eventtrace.h" #include "comutilnative.h" #include "finalizerthread.h" #include "threadsuspend.h" #include "wrappers.h" #include "nativeoverlapped.h" #include "mdaassistants.h" #include "appdomain.inl" #include "vmholder.h" #include "exceptmacros.h" #include "win32threadpool.h" #ifdef FEATURE_COMINTEROP #include "runtimecallablewrapper.h" #include "interoputil.h" #include "interoputil.inl" #endif // FEATURE_COMINTEROP #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT #include "olecontexthelpers.h" #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT #ifdef FEATURE_PERFTRACING #include "eventpipebuffermanager.h" #endif // FEATURE_PERFTRACING SPTR_IMPL(ThreadStore, ThreadStore, s_pThreadStore); CONTEXT *ThreadStore::s_pOSContext = NULL; CLREvent *ThreadStore::s_pWaitForStackCrawlEvent; PTR_ThreadLocalModule ThreadLocalBlock::GetTLMIfExists(ModuleIndex index) { WRAPPER_NO_CONTRACT; SUPPORTS_DAC; if (index.m_dwIndex >= m_TLMTableSize) return NULL; return m_pTLMTable[index.m_dwIndex].pTLM; } PTR_ThreadLocalModule ThreadLocalBlock::GetTLMIfExists(MethodTable* pMT) { WRAPPER_NO_CONTRACT; ModuleIndex index = pMT->GetModuleForStatics()->GetModuleIndex(); return GetTLMIfExists(index); } #ifndef DACCESS_COMPILE BOOL Thread::s_fCleanFinalizedThread = FALSE; Volatile Thread::s_threadPoolCompletionCountOverflow = 0; CrstStatic g_DeadlockAwareCrst; #if defined(_DEBUG) BOOL MatchThreadHandleToOsId ( HANDLE h, DWORD osId ) { #ifndef FEATURE_PAL LIMITED_METHOD_CONTRACT; DWORD id = GetThreadId(h); // OS call GetThreadId may fail, and return 0. In this case we can not // make a decision if the two match or not. Instead, we ignore this check. return id == 0 || id == osId; #else // !FEATURE_PAL return TRUE; #endif // !FEATURE_PAL } #endif // _DEBUG #ifdef _DEBUG_IMPL template<> AutoCleanupGCAssert::AutoCleanupGCAssert() { SCAN_SCOPE_BEGIN; STATIC_CONTRACT_MODE_COOPERATIVE; } template<> AutoCleanupGCAssert::AutoCleanupGCAssert() { SCAN_SCOPE_BEGIN; STATIC_CONTRACT_MODE_PREEMPTIVE; } template<> void GCAssert::BeginGCAssert() { SCAN_SCOPE_BEGIN; STATIC_CONTRACT_MODE_COOPERATIVE; } template<> void GCAssert::BeginGCAssert() { SCAN_SCOPE_BEGIN; STATIC_CONTRACT_MODE_PREEMPTIVE; } #endif // #define NEW_TLS 1 #ifdef _DEBUG void Thread::SetFrame(Frame *pFrame) { CONTRACTL { NOTHROW; GC_NOTRIGGER; DEBUG_ONLY; MODE_COOPERATIVE; // It only makes sense for a Thread to call SetFrame on itself. PRECONDITION(this == GetThread()); PRECONDITION(CheckPointer(pFrame)); } CONTRACTL_END; if (g_pConfig->fAssertOnFailFast()) { Frame *pWalk = m_pFrame; BOOL fExist = FALSE; while (pWalk != (Frame*) -1) { if (pWalk == pFrame) { fExist = TRUE; break; } pWalk = pWalk->m_Next; } pWalk = m_pFrame; while (fExist && pWalk != pFrame && pWalk != (Frame*)-1) { if (pWalk->GetVTablePtr() == ContextTransitionFrame::GetMethodFrameVPtr()) { _ASSERTE (((ContextTransitionFrame *)pWalk)->GetReturnDomain() == m_pDomain); } pWalk = pWalk->m_Next; } } m_pFrame = pFrame; // If stack overrun corruptions are expected, then skip this check // as the Frame chain may have been corrupted. if (g_pConfig->fAssertOnFailFast() == false) return; Frame* espVal = (Frame*)GetCurrentSP(); while (pFrame != (Frame*) -1) { static Frame* stopFrame = 0; if (pFrame == stopFrame) _ASSERTE(!"SetFrame frame == stopFrame"); _ASSERTE(espVal < pFrame); _ASSERTE(pFrame < m_CacheStackBase); _ASSERTE(pFrame->GetFrameType() < Frame::TYPE_COUNT); pFrame = pFrame->m_Next; } } #endif // _DEBUG //************************************************************************ // PRIVATE GLOBALS //************************************************************************ extern unsigned __int64 getTimeStamp(); extern unsigned __int64 getTickFrequency(); unsigned __int64 tgetFrequency() { static unsigned __int64 cachedFreq = (unsigned __int64) -1; if (cachedFreq != (unsigned __int64) -1) return cachedFreq; else { cachedFreq = getTickFrequency(); return cachedFreq; } } #endif // #ifndef DACCESS_COMPILE static StackWalkAction DetectHandleILStubsForDebugger_StackWalkCallback(CrawlFrame *pCF, VOID *pData) { WRAPPER_NO_CONTRACT; // It suffices to wait for the first CrawlFrame with non-NULL function MethodDesc *pMD = pCF->GetFunction(); if (pMD != NULL) { *(bool *)pData = pMD->IsILStub(); return SWA_ABORT; } return SWA_CONTINUE; } // This is really just a heuristic to detect if we are executing in an M2U IL stub or // one of the marshaling methods it calls. It doesn't deal with U2M IL stubs. // We loop through the frame chain looking for an uninitialized TransitionFrame. // If there is one, then we are executing in an M2U IL stub or one of the methods it calls. // On the other hand, if there is an initialized TransitionFrame, then we are not. // Also, if there is an HMF on the stack, then we stop. This could be the case where // an IL stub calls an FCALL which ends up in a managed method, and the debugger wants to // stop in those cases. Some examples are COMException..ctor and custom marshalers. // // X86 IL stubs use InlinedCallFrame and are indistinguishable from ordinary methods with // inlined P/Invoke when judging just from the frame chain. We use stack walk to decide // this case. bool Thread::DetectHandleILStubsForDebugger() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; Frame* pFrame = GetFrame(); if (pFrame != NULL) { while (pFrame != FRAME_TOP) { // Check for HMF's. See the comment at the beginning of this function. if (pFrame->GetVTablePtr() == HelperMethodFrame::GetMethodFrameVPtr()) { break; } // If there is an entry frame (i.e. U2M managed), we should break. else if (pFrame->GetFrameType() == Frame::TYPE_ENTRY) { break; } // Check for M2U transition frames. See the comment at the beginning of this function. else if (pFrame->GetFrameType() == Frame::TYPE_EXIT) { if (pFrame->GetReturnAddress() == NULL) { // If the return address is NULL, then the frame has not been initialized yet. // We may see InlinedCallFrame in ordinary methods as well. Have to do // stack walk to find out if this is really an IL stub. bool fInILStub = false; StackWalkFrames(&DetectHandleILStubsForDebugger_StackWalkCallback, &fInILStub, QUICKUNWIND, dac_cast(pFrame)); if (fInILStub) return true; } else { // The frame is fully initialized. return false; } } pFrame = pFrame->Next(); } } return false; } extern "C" { #ifndef __llvm__ __declspec(thread) #else // !__llvm__ __thread #endif // !__llvm__ ThreadLocalInfo gCurrentThreadInfo = { NULL, // m_pThread NULL, // m_pAppDomain NULL, // m_EETlsData }; } // extern "C" // index into TLS Array. Definition added by compiler EXTERN_C UINT32 _tls_index; #ifndef DACCESS_COMPILE BOOL SetThread(Thread* t) { LIMITED_METHOD_CONTRACT gCurrentThreadInfo.m_pThread = t; return TRUE; } BOOL SetAppDomain(AppDomain* ad) { LIMITED_METHOD_CONTRACT gCurrentThreadInfo.m_pAppDomain = ad; return TRUE; } BOOL Thread::Alert () { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; BOOL fRetVal = FALSE; { HANDLE handle = GetThreadHandle(); if (handle != INVALID_HANDLE_VALUE && handle != SWITCHOUT_HANDLE_VALUE) { fRetVal = ::QueueUserAPC(UserInterruptAPC, handle, APC_Code); } } return fRetVal; } DWORD Thread::Join(DWORD timeout, BOOL alertable) { WRAPPER_NO_CONTRACT; return JoinEx(timeout,alertable?WaitMode_Alertable:WaitMode_None); } DWORD Thread::JoinEx(DWORD timeout, WaitMode mode) { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; BOOL alertable = (mode & WaitMode_Alertable)?TRUE:FALSE; Thread *pCurThread = GetThread(); _ASSERTE(pCurThread || dbgOnly_IsSpecialEEThread()); { // We're not hosted, so WaitMode_InDeadlock is irrelevant. Clear it, so that this wait can be // forwarded to a SynchronizationContext if needed. mode = (WaitMode)(mode & ~WaitMode_InDeadlock); HANDLE handle = GetThreadHandle(); if (handle == INVALID_HANDLE_VALUE || handle == SWITCHOUT_HANDLE_VALUE) { return WAIT_FAILED; } if (pCurThread) { return pCurThread->DoAppropriateWait(1, &handle, FALSE, timeout, mode); } else { return WaitForSingleObjectEx(handle,timeout,alertable); } } } extern INT32 MapFromNTPriority(INT32 NTPriority); BOOL Thread::SetThreadPriority( int nPriority // thread priority level ) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; BOOL fRet; { if (GetThreadHandle() == INVALID_HANDLE_VALUE) { // When the thread starts running, we will set the thread priority. fRet = TRUE; } else fRet = ::SetThreadPriority(GetThreadHandle(), nPriority); } if (fRet) { GCX_COOP(); THREADBASEREF pObject = (THREADBASEREF)ObjectFromHandle(m_ExposedObject); if (pObject != NULL) { // TODO: managed ThreadPriority only supports up to 4. pObject->SetPriority (MapFromNTPriority(nPriority)); } } return fRet; } int Thread::GetThreadPriority() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; int nRetVal = -1; if (GetThreadHandle() == INVALID_HANDLE_VALUE) { nRetVal = FALSE; } else nRetVal = ::GetThreadPriority(GetThreadHandle()); return nRetVal; } void Thread::ChooseThreadCPUGroupAffinity() { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; if (!CPUGroupInfo::CanEnableGCCPUGroups() || !CPUGroupInfo::CanEnableThreadUseAllCpuGroups()) return; //Borrow the ThreadStore Lock here: Lock ThreadStore before distributing threads ThreadStoreLockHolder TSLockHolder(TRUE); // this thread already has CPU group affinity set if (m_pAffinityMask != 0) return; if (GetThreadHandle() == INVALID_HANDLE_VALUE) return; GROUP_AFFINITY groupAffinity; CPUGroupInfo::ChooseCPUGroupAffinity(&groupAffinity); CPUGroupInfo::SetThreadGroupAffinity(GetThreadHandle(), &groupAffinity, NULL); m_wCPUGroup = groupAffinity.Group; m_pAffinityMask = groupAffinity.Mask; } void Thread::ClearThreadCPUGroupAffinity() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; if (!CPUGroupInfo::CanEnableGCCPUGroups() || !CPUGroupInfo::CanEnableThreadUseAllCpuGroups()) return; ThreadStoreLockHolder TSLockHolder(TRUE); // this thread does not have CPU group affinity set if (m_pAffinityMask == 0) return; GROUP_AFFINITY groupAffinity; groupAffinity.Group = m_wCPUGroup; groupAffinity.Mask = m_pAffinityMask; CPUGroupInfo::ClearCPUGroupAffinity(&groupAffinity); m_wCPUGroup = 0; m_pAffinityMask = 0; } DWORD Thread::StartThread() { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_ANY; } CONTRACTL_END; DWORD dwRetVal = (DWORD) -1; #ifdef _DEBUG _ASSERTE (m_Creater.IsCurrentThread()); m_Creater.Clear(); #endif _ASSERTE (GetThreadHandle() != INVALID_HANDLE_VALUE && GetThreadHandle() != SWITCHOUT_HANDLE_VALUE); dwRetVal = ::ResumeThread(GetThreadHandle()); return dwRetVal; } // Class static data: LONG Thread::m_DebugWillSyncCount = -1; LONG Thread::m_DetachCount = 0; LONG Thread::m_ActiveDetachCount = 0; int Thread::m_offset_counter = 0; Volatile Thread::m_threadsAtUnsafePlaces = 0; //------------------------------------------------------------------------- // Public function: SetupThreadNoThrow() // Creates Thread for current thread if not previously created. // Returns NULL for failure (usually due to out-of-memory.) //------------------------------------------------------------------------- Thread* SetupThreadNoThrow(HRESULT *pHR) { CONTRACTL { NOTHROW; SO_TOLERANT; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; HRESULT hr = S_OK; Thread *pThread = GetThread(); if (pThread != NULL) { return pThread; } EX_TRY { pThread = SetupThread(); } EX_CATCH { // We failed SetupThread. GET_EXCEPTION() may depend on Thread object. if (__pException == NULL) { hr = E_OUTOFMEMORY; } else { hr = GET_EXCEPTION()->GetHR(); } } EX_END_CATCH(SwallowAllExceptions); if (pHR) { *pHR = hr; } return pThread; } void DeleteThread(Thread* pThread) { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; //_ASSERTE (pThread == GetThread()); SetThread(NULL); SetAppDomain(NULL); if (pThread->HasThreadStateNC(Thread::TSNC_ExistInThreadStore)) { pThread->DetachThread(FALSE); } else { #ifdef FEATURE_COMINTEROP pThread->RevokeApartmentSpy(); #endif // FEATURE_COMINTEROP FastInterlockOr((ULONG *)&pThread->m_State, Thread::TS_Dead); // ~Thread() calls SafeSetThrowables which has a conditional contract // which says that if you call it with a NULL throwable then it is // MODE_ANY, otherwise MODE_COOPERATIVE. Scan doesn't understand that // and assumes that we're violating the MODE_COOPERATIVE. CONTRACT_VIOLATION(ModeViolation); delete pThread; } } void EnsurePreemptive() { WRAPPER_NO_CONTRACT; Thread *pThread = GetThread(); if (pThread && pThread->PreemptiveGCDisabled()) { pThread->EnablePreemptiveGC(); } } typedef StateHolder EnsurePreemptiveModeIfException; Thread* SetupThread(BOOL fInternal) { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} SO_TOLERANT; } CONTRACTL_END; Thread* pThread; if ((pThread = GetThread()) != NULL) return pThread; #ifdef FEATURE_STACK_PROBE RetailStackProbe(ADJUST_PROBE(DEFAULT_ENTRY_PROBE_AMOUNT), NULL); #endif //FEATURE_STACK_PROBE CONTRACT_VIOLATION(SOToleranceViolation); // For interop debugging, we must mark that we're in a can't-stop region // b.c we may take Crsts here that may block the helper thread. // We're especially fragile here b/c we don't have a Thread object yet CantStopHolder hCantStop; EnsurePreemptiveModeIfException ensurePreemptive; #ifdef _DEBUG CHECK chk; if (g_pConfig->SuppressChecks()) { // EnterAssert will suppress any checks chk.EnterAssert(); } #endif // Normally, HasStarted is called from the thread's entrypoint to introduce it to // the runtime. But sometimes that thread is used for DLL_THREAD_ATTACH notifications // that call into managed code. In that case, a call to SetupThread here must // find the correct Thread object and install it into TLS. if (ThreadStore::s_pThreadStore->m_PendingThreadCount != 0) { DWORD ourOSThreadId = ::GetCurrentThreadId(); { ThreadStoreLockHolder TSLockHolder; _ASSERTE(pThread == NULL); while ((pThread = ThreadStore::s_pThreadStore->GetAllThreadList(pThread, Thread::TS_Unstarted | Thread::TS_FailStarted, Thread::TS_Unstarted)) != NULL) { if (pThread->GetOSThreadId() == ourOSThreadId) { break; } } if (pThread != NULL) { STRESS_LOG2(LF_SYNC, LL_INFO1000, "T::ST - recycling thread 0x%p (state: 0x%x)\n", pThread, pThread->m_State.Load()); } } // It's perfectly reasonable to not find this guy. It's just an unrelated // thread spinning up. if (pThread) { if (IsThreadPoolWorkerSpecialThread()) { FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_TPWorkerThread); pThread->SetBackground(TRUE); } else if (IsThreadPoolIOCompletionSpecialThread()) { FastInterlockOr ((ULONG *) &pThread->m_State, Thread::TS_CompletionPortThread); pThread->SetBackground(TRUE); } else if (IsTimerSpecialThread() || IsWaitSpecialThread()) { FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_TPWorkerThread); pThread->SetBackground(TRUE); } BOOL fStatus = pThread->HasStarted(); ensurePreemptive.SuppressRelease(); return fStatus ? pThread : NULL; } } // First time we've seen this thread in the runtime: pThread = new Thread(); // What state are we in here? COOP??? Holder,DeleteThread> threadHolder(pThread); CExecutionEngine::SetupTLSForThread(pThread); // A host can deny a thread entering runtime by returning a NULL IHostTask. // But we do want threads used by threadpool. if (IsThreadPoolWorkerSpecialThread() || IsThreadPoolIOCompletionSpecialThread() || IsTimerSpecialThread() || IsWaitSpecialThread()) { fInternal = TRUE; } if (!pThread->InitThread(fInternal) || !pThread->PrepareApartmentAndContext()) ThrowOutOfMemory(); // reset any unstarted bits on the thread object FastInterlockAnd((ULONG *) &pThread->m_State, ~Thread::TS_Unstarted); FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_LegalToJoin); ThreadStore::AddThread(pThread); BOOL fOK = SetThread(pThread); _ASSERTE (fOK); fOK = SetAppDomain(pThread->GetDomain()); _ASSERTE (fOK); #ifdef FEATURE_INTEROP_DEBUGGING // Ensure that debugger word slot is allocated UnsafeTlsSetValue(g_debuggerWordTLSIndex, 0); #endif // We now have a Thread object visable to the RS. unmark special status. hCantStop.Release(); pThread->SetupThreadForHost(); threadHolder.SuppressRelease(); FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_FullyInitialized); #ifdef DEBUGGING_SUPPORTED // // If we're debugging, let the debugger know that this // thread is up and running now. // if (CORDebuggerAttached()) { g_pDebugInterface->ThreadCreated(pThread); } else { LOG((LF_CORDB, LL_INFO10000, "ThreadCreated() not called due to CORDebuggerAttached() being FALSE for thread 0x%x\n", pThread->GetThreadId())); } #endif // DEBUGGING_SUPPORTED #ifdef PROFILING_SUPPORTED // If a profiler is present, then notify the profiler that a // thread has been created. if (!IsGCSpecialThread()) { BEGIN_PIN_PROFILER(CORProfilerTrackThreads()); { GCX_PREEMP(); g_profControlBlock.pProfInterface->ThreadCreated( (ThreadID)pThread); } DWORD osThreadId = ::GetCurrentThreadId(); g_profControlBlock.pProfInterface->ThreadAssignedToOSThread( (ThreadID)pThread, osThreadId); END_PIN_PROFILER(); } #endif // PROFILING_SUPPORTED _ASSERTE(!pThread->IsBackground()); // doesn't matter, but worth checking pThread->SetBackground(TRUE); ensurePreemptive.SuppressRelease(); if (IsThreadPoolWorkerSpecialThread()) { FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_TPWorkerThread); } else if (IsThreadPoolIOCompletionSpecialThread()) { FastInterlockOr ((ULONG *) &pThread->m_State, Thread::TS_CompletionPortThread); } else if (IsTimerSpecialThread() || IsWaitSpecialThread()) { FastInterlockOr((ULONG *) &pThread->m_State, Thread::TS_TPWorkerThread); } #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING if (g_fEnableARM) { pThread->QueryThreadProcessorUsage(); } #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING #ifdef FEATURE_EVENT_TRACE ETW::ThreadLog::FireThreadCreated(pThread); #endif // FEATURE_EVENT_TRACE return pThread; } //------------------------------------------------------------------------- // Public function: SetupUnstartedThread() // This sets up a Thread object for an exposed System.Thread that // has not been started yet. This allows us to properly enumerate all threads // in the ThreadStore, so we can report on even unstarted threads. Clearly // there is no physical thread to match, yet. // // When there is, complete the setup with code:Thread::HasStarted() //------------------------------------------------------------------------- Thread* SetupUnstartedThread(BOOL bRequiresTSL) { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; Thread* pThread = new Thread(); FastInterlockOr((ULONG *) &pThread->m_State, (Thread::TS_Unstarted | Thread::TS_WeOwn)); ThreadStore::AddThread(pThread, bRequiresTSL); return pThread; } //------------------------------------------------------------------------- // Public function: DestroyThread() // Destroys the specified Thread object, for a thread which is about to die. //------------------------------------------------------------------------- void DestroyThread(Thread *th) { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE (th == GetThread()); _ASSERTE(g_fEEShutDown || th->m_dwLockCount == 0 || th->m_fRudeAborted); #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING if (g_fEnableARM) { AppDomain* pDomain = th->GetDomain(); pDomain->UpdateProcessorUsage(th->QueryThreadProcessorUsage()); FireEtwThreadTerminated((ULONGLONG)th, (ULONGLONG)pDomain, GetClrInstanceId()); } #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING th->FinishSOWork(); GCX_PREEMP_NO_DTOR(); if (th->IsAbortRequested()) { // Reset trapping count. th->UnmarkThreadForAbort(Thread::TAR_ALL); } // Clear any outstanding stale EH state that maybe still active on the thread. #ifdef WIN64EXCEPTIONS ExceptionTracker::PopTrackers((void*)-1); #else // !WIN64EXCEPTIONS #ifdef _TARGET_X86_ PTR_ThreadExceptionState pExState = th->GetExceptionState(); if (pExState->IsExceptionInProgress()) { GCX_COOP(); pExState->GetCurrentExceptionTracker()->UnwindExInfo((void *)-1); } #else // !_TARGET_X86_ #error Unsupported platform #endif // _TARGET_X86_ #endif // WIN64EXCEPTIONS #ifdef FEATURE_PERFTRACING // Before the thread dies, mark its buffers as no longer owned // so that they can be cleaned up after the thread dies. EventPipeBufferList *pBufferList = th->GetEventPipeBufferList(); if(pBufferList != NULL) { pBufferList->SetOwnedByThread(false); } #endif // FEATURE_PERFTRACING if (g_fEEShutDown == 0) { th->SetThreadState(Thread::TS_ReportDead); th->OnThreadTerminate(FALSE); } } //------------------------------------------------------------------------- // Public function: DetachThread() // Marks the thread as needing to be destroyed, but doesn't destroy it yet. //------------------------------------------------------------------------- HRESULT Thread::DetachThread(BOOL fDLLThreadDetach) { // !!! Can not use contract here. // !!! Contract depends on Thread object for GC_TRIGGERS. // !!! At the end of this function, we call InternalSwitchOut, // !!! and then GetThread()=NULL, and dtor of contract does not work any more. STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_NOTRIGGER; // @todo . We need to probe here, but can't introduce destructors etc. BEGIN_CONTRACT_VIOLATION(SOToleranceViolation); // Clear any outstanding stale EH state that maybe still active on the thread. #ifdef WIN64EXCEPTIONS ExceptionTracker::PopTrackers((void*)-1); #else // !WIN64EXCEPTIONS #ifdef _TARGET_X86_ PTR_ThreadExceptionState pExState = GetExceptionState(); if (pExState->IsExceptionInProgress()) { GCX_COOP(); pExState->GetCurrentExceptionTracker()->UnwindExInfo((void *)-1); } #else // !_TARGET_X86_ #error Unsupported platform #endif // _TARGET_X86_ #endif // WIN64EXCEPTIONS #ifdef FEATURE_COMINTEROP IErrorInfo *pErrorInfo; // Avoid calling GetErrorInfo() if ole32 has already executed the DLL_THREAD_DETACH, // otherwise we'll cause ole32 to re-allocate and leak its TLS data (SOleTlsData). if (ClrTeb::GetOleReservedPtr() != NULL && GetErrorInfo(0, &pErrorInfo) == S_OK) { // if this is our IErrorInfo, release it now - we don't want ole32 to do it later as // part of its DLL_THREAD_DETACH as we won't be able to handle the call at that point if (!ComInterfaceSlotIs(pErrorInfo, 2, Unknown_ReleaseSpecial_IErrorInfo)) { // if it's not our IErrorInfo, put it back SetErrorInfo(0, pErrorInfo); } pErrorInfo->Release(); } // Revoke our IInitializeSpy registration only if we are not in DLL_THREAD_DETACH // (COM will do it or may have already done it automatically in that case). if (!fDLLThreadDetach) { RevokeApartmentSpy(); } #endif // FEATURE_COMINTEROP _ASSERTE(!PreemptiveGCDisabled()); _ASSERTE(g_fEEShutDown || m_dwLockCount == 0 || m_fRudeAborted); _ASSERTE ((m_State & Thread::TS_Detached) == 0); _ASSERTE (this == GetThread()); #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING if (g_fEnableARM && m_pDomain) { m_pDomain->UpdateProcessorUsage(QueryThreadProcessorUsage()); FireEtwThreadTerminated((ULONGLONG)this, (ULONGLONG)m_pDomain, GetClrInstanceId()); } #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING FinishSOWork(); FastInterlockIncrement(&Thread::m_DetachCount); if (IsAbortRequested()) { // Reset trapping count. UnmarkThreadForAbort(Thread::TAR_ALL); } if (!IsBackground()) { FastInterlockIncrement(&Thread::m_ActiveDetachCount); ThreadStore::CheckForEEShutdown(); } END_CONTRACT_VIOLATION; HANDLE hThread = GetThreadHandle(); SetThreadHandle (SWITCHOUT_HANDLE_VALUE); while (m_dwThreadHandleBeingUsed > 0) { // Another thread is using the handle now. #undef Sleep // We can not call __SwitchToThread since we can not go back to host. ::Sleep(10); #define Sleep(a) Dont_Use_Sleep(a) } if (m_WeOwnThreadHandle && m_ThreadHandleForClose == INVALID_HANDLE_VALUE) { m_ThreadHandleForClose = hThread; } // We need to make sure that TLS are touched last here. SetThread(NULL); SetAppDomain(NULL); #ifdef ENABLE_CONTRACTS_DATA m_pClrDebugState = NULL; #endif //ENABLE_CONTRACTS_DATA #ifdef FEATURE_PERFTRACING // Before the thread dies, mark its buffers as no longer owned // so that they can be cleaned up after the thread dies. EventPipeBufferList *pBufferList = m_pEventPipeBufferList.Load(); if(pBufferList != NULL) { pBufferList->SetOwnedByThread(false); } #endif // FEATURE_PERFTRACING FastInterlockOr((ULONG*)&m_State, (int) (Thread::TS_Detached | Thread::TS_ReportDead)); // Do not touch Thread object any more. It may be destroyed. // These detached threads will be cleaned up by finalizer thread. But if the process uses // little managed heap, it will be a while before GC happens, and finalizer thread starts // working on detached thread. So we wake up finalizer thread to clean up resources. // // (It's possible that this is the startup thread, and startup failed, and so the finalization // machinery isn't fully initialized. Hence this check.) if (g_fEEStarted) FinalizerThread::EnableFinalization(); return S_OK; } DWORD GetRuntimeId() { LIMITED_METHOD_CONTRACT; return _tls_index; } //--------------------------------------------------------------------------- // Creates new Thread for reverse p-invoke calls. //--------------------------------------------------------------------------- Thread* WINAPI CreateThreadBlockThrow() { WRAPPER_NO_CONTRACT; // This is a workaround to disable our check for throwing exception in SetupThread. // We want to throw an exception for reverse p-invoke, and our assertion may fire if // a unmanaged caller does not setup an exception handler. CONTRACT_VIOLATION(ThrowsViolation); // WON'T FIX - This enables catastrophic failure exception in reverse P/Invoke - the only way we can communicate an error to legacy code. Thread* pThread = NULL; BEGIN_ENTRYPOINT_THROWS; if (!CanRunManagedCode()) { // CLR is shutting down - someone's DllMain detach event may be calling back into managed code. // It is misleading to use our COM+ exception code, since this is not a managed exception. ULONG_PTR arg = E_PROCESS_SHUTDOWN_REENTRY; RaiseException(EXCEPTION_EXX, 0, 1, &arg); } HRESULT hr = S_OK; pThread = SetupThreadNoThrow(&hr); if (pThread == NULL) { // Creating Thread failed, and we need to throw an exception to report status. // It is misleading to use our COM+ exception code, since this is not a managed exception. ULONG_PTR arg = hr; RaiseException(EXCEPTION_EXX, 0, 1, &arg); } END_ENTRYPOINT_THROWS; return pThread; } #ifdef _DEBUG DWORD_PTR Thread::OBJREF_HASH = OBJREF_TABSIZE; #endif extern "C" void STDCALL JIT_PatchedCodeStart(); extern "C" void STDCALL JIT_PatchedCodeLast(); //--------------------------------------------------------------------------- // One-time initialization. Called during Dll initialization. So // be careful what you do in here! //--------------------------------------------------------------------------- void InitThreadManager() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; InitializeYieldProcessorNormalizedCrst(); // All patched helpers should fit into one page. // If you hit this assert on retail build, there is most likely problem with BBT script. _ASSERTE_ALL_BUILDS("clr/src/VM/threads.cpp", (BYTE*)JIT_PatchedCodeLast - (BYTE*)JIT_PatchedCodeStart < (ptrdiff_t)GetOsPageSize()); // I am using virtual protect to cover the entire range that this code falls in. // // We could reset it to non-writeable inbetween GCs and such, but then we'd have to keep on re-writing back and forth, // so instead we'll leave it writable from here forward. DWORD oldProt; if (!ClrVirtualProtect((void *)JIT_PatchedCodeStart, (BYTE*)JIT_PatchedCodeLast - (BYTE*)JIT_PatchedCodeStart, PAGE_EXECUTE_READWRITE, &oldProt)) { _ASSERTE(!"ClrVirtualProtect of code page failed"); COMPlusThrowWin32(); } #ifndef FEATURE_PAL _ASSERTE(GetThread() == NULL); PTEB Teb = NtCurrentTeb(); BYTE** tlsArray = (BYTE**)Teb->ThreadLocalStoragePointer; BYTE* tlsData = (BYTE*)tlsArray[_tls_index]; size_t offsetOfCurrentThreadInfo = (BYTE*)&gCurrentThreadInfo - tlsData; _ASSERTE(offsetOfCurrentThreadInfo < 0x8000); _ASSERTE(_tls_index < 0x10000); // Save gCurrentThreadInfo location for debugger g_TlsIndex = (DWORD)(_tls_index + (offsetOfCurrentThreadInfo << 16) + 0x80000000); _ASSERTE(g_TrapReturningThreads == 0); #endif // !FEATURE_PAL #ifdef FEATURE_INTEROP_DEBUGGING g_debuggerWordTLSIndex = UnsafeTlsAlloc(); if (g_debuggerWordTLSIndex == TLS_OUT_OF_INDEXES) COMPlusThrowWin32(); #endif __ClrFlsGetBlock = CExecutionEngine::GetTlsData; IfFailThrow(Thread::CLRSetThreadStackGuarantee(Thread::STSGuarantee_Force)); ThreadStore::InitThreadStore(); // NOTE: CRST_UNSAFE_ANYMODE prevents a GC mode switch when entering this crst. // If you remove this flag, we will switch to preemptive mode when entering // g_DeadlockAwareCrst, which means all functions that enter it will become // GC_TRIGGERS. (This includes all uses of CrstHolder.) So be sure // to update the contracts if you remove this flag. g_DeadlockAwareCrst.Init(CrstDeadlockDetection, CRST_UNSAFE_ANYMODE); #ifdef _DEBUG // Randomize OBJREF_HASH to handle hash collision. Thread::OBJREF_HASH = OBJREF_TABSIZE - (DbgGetEXETimeStamp()%10); #endif // _DEBUG ThreadSuspend::Initialize(); } //************************************************************************ // Thread members //************************************************************************ #if defined(_DEBUG) && defined(TRACK_SYNC) // One outstanding synchronization held by this thread: struct Dbg_TrackSyncEntry { UINT_PTR m_caller; AwareLock *m_pAwareLock; BOOL Equiv (UINT_PTR caller, void *pAwareLock) { LIMITED_METHOD_CONTRACT; return (m_caller == caller) && (m_pAwareLock == pAwareLock); } BOOL Equiv (void *pAwareLock) { LIMITED_METHOD_CONTRACT; return (m_pAwareLock == pAwareLock); } }; // Each thread has a stack that tracks all enter and leave requests struct Dbg_TrackSyncStack : public Dbg_TrackSync { enum { MAX_TRACK_SYNC = 20, // adjust stack depth as necessary }; void EnterSync (UINT_PTR caller, void *pAwareLock); void LeaveSync (UINT_PTR caller, void *pAwareLock); Dbg_TrackSyncEntry m_Stack [MAX_TRACK_SYNC]; UINT_PTR m_StackPointer; BOOL m_Active; Dbg_TrackSyncStack() : m_StackPointer(0), m_Active(TRUE) { LIMITED_METHOD_CONTRACT; } }; // ensure that registers are preserved across this call #ifdef _MSC_VER #pragma optimize("", off) #endif // A pain to do all this from ASM, but watch out for trashed registers EXTERN_C void EnterSyncHelper (UINT_PTR caller, void *pAwareLock) { BEGIN_ENTRYPOINT_THROWS; WRAPPER_NO_CONTRACT; GetThread()->m_pTrackSync->EnterSync(caller, pAwareLock); END_ENTRYPOINT_THROWS; } EXTERN_C void LeaveSyncHelper (UINT_PTR caller, void *pAwareLock) { BEGIN_ENTRYPOINT_THROWS; WRAPPER_NO_CONTRACT; GetThread()->m_pTrackSync->LeaveSync(caller, pAwareLock); END_ENTRYPOINT_THROWS; } #ifdef _MSC_VER #pragma optimize("", on) #endif void Dbg_TrackSyncStack::EnterSync(UINT_PTR caller, void *pAwareLock) { LIMITED_METHOD_CONTRACT; STRESS_LOG4(LF_SYNC, LL_INFO100, "Dbg_TrackSyncStack::EnterSync, IP=%p, Recursion=%u, LockState=%x, HoldingThread=%p.\n", caller, ((AwareLock*)pAwareLock)->GetRecursionLevel(), ((AwareLock*)pAwareLock)->GetLockState(), ((AwareLock*)pAwareLock)->GetHoldingThread()); if (m_Active) { if (m_StackPointer >= MAX_TRACK_SYNC) { _ASSERTE(!"Overflowed synchronization stack checking. Disabling"); m_Active = FALSE; return; } } m_Stack[m_StackPointer].m_caller = caller; m_Stack[m_StackPointer].m_pAwareLock = (AwareLock *) pAwareLock; m_StackPointer++; } void Dbg_TrackSyncStack::LeaveSync(UINT_PTR caller, void *pAwareLock) { WRAPPER_NO_CONTRACT; STRESS_LOG4(LF_SYNC, LL_INFO100, "Dbg_TrackSyncStack::LeaveSync, IP=%p, Recursion=%u, LockState=%x, HoldingThread=%p.\n", caller, ((AwareLock*)pAwareLock)->GetRecursionLevel(), ((AwareLock*)pAwareLock)->GetLockState(), ((AwareLock*)pAwareLock)->GetHoldingThread()); if (m_Active) { if (m_StackPointer == 0) _ASSERTE(!"Underflow in leaving synchronization"); else if (m_Stack[m_StackPointer - 1].Equiv(pAwareLock)) { m_StackPointer--; } else { for (int i=m_StackPointer - 2; i>=0; i--) { if (m_Stack[i].Equiv(pAwareLock)) { _ASSERTE(!"Locks are released out of order. This might be okay..."); memcpy(&m_Stack[i], &m_Stack[i+1], sizeof(m_Stack[0]) * (m_StackPointer - i - 1)); return; } } _ASSERTE(!"Trying to release a synchronization lock which isn't held"); } } } #endif // TRACK_SYNC static DWORD dwHashCodeSeed = 123456789; #ifdef _DEBUG void CheckADValidity(AppDomain* pDomain, DWORD ADValidityKind) { CONTRACTL { NOTHROW; FORBID_FAULT; GC_NOTRIGGER; MODE_ANY; } CONTRACTL_END; // // Note: this apparently checks if any one of the supplied conditions is satisified, rather // than checking that *all* of them are satisfied. One would have expected it to assert all of the // conditions but it does not. // CONTRACT_VIOLATION(FaultViolation); if (::GetAppDomain()==pDomain) return; if ((ADValidityKind & ADV_DEFAULTAD) && pDomain->IsDefaultDomain()) return; if ((ADValidityKind & ADV_ITERATOR) && pDomain->IsHeldByIterator()) return; if ((ADValidityKind & ADV_CREATING) && pDomain->IsBeingCreated()) return; if ((ADValidityKind & ADV_COMPILATION) && pDomain->IsCompilationDomain()) return; if ((ADValidityKind & ADV_FINALIZER) && IsFinalizerThread()) return; if ((ADValidityKind & ADV_RUNNINGIN) && pDomain->IsRunningIn(GetThread())) return; if ((ADValidityKind & ADV_REFTAKER) && pDomain->IsHeldByRefTaker()) return; _ASSERTE(!"Appdomain* can be invalid"); } #endif //-------------------------------------------------------------------- // Thread construction //-------------------------------------------------------------------- Thread::Thread() { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; m_pFrame = FRAME_TOP; m_fPreemptiveGCDisabled = 0; #ifdef _DEBUG m_ulForbidTypeLoad = 0; m_GCOnTransitionsOK = TRUE; #endif #ifdef ENABLE_CONTRACTS m_pClrDebugState = NULL; m_ulEnablePreemptiveGCCount = 0; #endif m_dwLockCount = 0; m_dwBeginLockCount = 0; #ifdef _DEBUG dbg_m_cSuspendedThreads = 0; dbg_m_cSuspendedThreadsWithoutOSLock = 0; m_Creater.Clear(); m_dwUnbreakableLockCount = 0; #endif m_dwForbidSuspendThread = 0; // Initialize lock state m_pHead = &m_embeddedEntry; m_embeddedEntry.pNext = m_pHead; m_embeddedEntry.pPrev = m_pHead; m_embeddedEntry.dwLLockID = 0; m_embeddedEntry.dwULockID = 0; m_embeddedEntry.wReaderLevel = 0; m_pBlockingLock = NULL; m_alloc_context.init(); m_thAllocContextObj = 0; m_UserInterrupt = 0; m_WaitEventLink.m_Next = NULL; m_WaitEventLink.m_LinkSB.m_pNext = NULL; m_ThreadHandle = INVALID_HANDLE_VALUE; m_ThreadHandleForClose = INVALID_HANDLE_VALUE; m_ThreadHandleForResume = INVALID_HANDLE_VALUE; m_WeOwnThreadHandle = FALSE; #ifdef _DEBUG m_ThreadId = UNINITIALIZED_THREADID; #endif //_DEBUG // Initialize this variable to a very different start value for each thread // Using linear congruential generator from Knuth Vol. 2, p. 102, line 24 dwHashCodeSeed = dwHashCodeSeed * 1566083941 + 1; m_dwHashCodeSeed = dwHashCodeSeed; m_hijackLock = FALSE; m_OSThreadId = 0; m_Priority = INVALID_THREAD_PRIORITY; m_ExternalRefCount = 1; m_UnmanagedRefCount = 0; m_State = TS_Unstarted; m_StateNC = TSNC_Unknown; // It can't be a LongWeakHandle because we zero stuff out of the exposed // object as it is finalized. At that point, calls to GetCurrentThread() // had better get a new one,! m_ExposedObject = CreateGlobalShortWeakHandle(NULL); GlobalShortWeakHandleHolder exposedObjectHolder(m_ExposedObject); m_StrongHndToExposedObject = CreateGlobalStrongHandle(NULL); GlobalStrongHandleHolder strongHndToExposedObjectHolder(m_StrongHndToExposedObject); m_LastThrownObjectHandle = NULL; m_ltoIsUnhandled = FALSE; m_AbortReason = NULL; m_debuggerFilterContext = NULL; m_debuggerCantStop = 0; m_fInteropDebuggingHijacked = FALSE; m_profilerCallbackState = 0; #ifdef FEATURE_PROFAPI_ATTACH_DETACH m_dwProfilerEvacuationCounter = 0; #endif // FEATURE_PROFAPI_ATTACH_DETACH m_pProfilerFilterContext = NULL; m_CacheStackBase = 0; m_CacheStackLimit = 0; m_CacheStackSufficientExecutionLimit = 0; m_LastAllowableStackAddress= 0; m_ProbeLimit = 0; #ifdef _DEBUG m_pCleanedStackBase = NULL; #endif #ifdef STACK_GUARDS_DEBUG m_pCurrentStackGuard = NULL; #endif #ifdef FEATURE_HIJACK m_ppvHJRetAddrPtr = (VOID**) 0xCCCCCCCCCCCCCCCC; m_pvHJRetAddr = (VOID*) 0xCCCCCCCCCCCCCCCC; #ifndef PLATFORM_UNIX X86_ONLY(m_LastRedirectIP = 0); X86_ONLY(m_SpinCount = 0); #endif // PLATFORM_UNIX #endif // FEATURE_HIJACK #if defined(_DEBUG) && defined(TRACK_SYNC) m_pTrackSync = new Dbg_TrackSyncStack; NewHolder trackSyncHolder(static_cast(m_pTrackSync)); #endif // TRACK_SYNC m_RequestedStackSize = 0; m_PreventAsync = 0; m_PreventAbort = 0; m_nNestedMarshalingExceptions = 0; m_pDomain = NULL; #ifdef FEATURE_COMINTEROP m_fDisableComObjectEagerCleanup = false; #endif //FEATURE_COMINTEROP m_fHasDeadThreadBeenConsideredForGCTrigger = false; m_TraceCallCount = 0; m_ThrewControlForThread = 0; m_OSContext = NULL; m_ThreadTasks = (ThreadTasks)0; m_pLoadLimiter= NULL; m_pLoadingFile = NULL; // The state and the tasks must be 32-bit aligned for atomicity to be guaranteed. _ASSERTE((((size_t) &m_State) & 3) == 0); _ASSERTE((((size_t) &m_ThreadTasks) & 3) == 0); // Track perf counter for the logical thread object. COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cCurrentThreadsLogical++); // On all callbacks, call the trap code, which we now have // wired to cause a GC. Thus we will do a GC on all Transition Frame Transitions (and more). if (GCStress::IsEnabled()) { m_State = (ThreadState) (m_State | TS_GCOnTransitions); } m_AbortType = EEPolicy::TA_None; m_AbortInfo = 0; m_AbortEndTime = MAXULONGLONG; m_RudeAbortEndTime = MAXULONGLONG; m_AbortController = 0; m_AbortRequestLock = 0; m_fRudeAbortInitiated = FALSE; m_pIOCompletionContext = NULL; #ifdef _DEBUG m_fRudeAborted = FALSE; m_dwAbortPoint = 0; #endif m_OSContext = new CONTEXT(); NewHolder contextHolder(m_OSContext); m_pSavedRedirectContext = NULL; NewHolder savedRedirectContextHolder(m_pSavedRedirectContext); #ifdef FEATURE_COMINTEROP m_pRCWStack = new RCWStackHeader(); #endif #ifdef _DEBUG m_bGCStressing = FALSE; m_bUniqueStacking = FALSE; #endif m_pPendingTypeLoad = NULL; #ifdef FEATURE_PREJIT m_pIBCInfo = NULL; #endif m_dwAVInRuntimeImplOkayCount = 0; #if defined(HAVE_GCCOVER) && defined(USE_REDIRECT_FOR_GCSTRESS) && !defined(PLATFORM_UNIX) // GCCOVER m_fPreemptiveGCDisabledForGCStress = false; #endif #ifdef _DEBUG m_pHelperMethodFrameCallerList = (HelperMethodFrameCallerList*)-1; #endif m_dwHostTaskRefCount = 0; m_pExceptionDuringStartup = NULL; #ifdef HAVE_GCCOVER m_pbDestCode = NULL; m_pbSrcCode = NULL; #if defined(GCCOVER_TOLERATE_SPURIOUS_AV) m_pLastAVAddress = NULL; #endif // defined(GCCOVER_TOLERATE_SPURIOUS_AV) #endif // HAVE_GCCOVER m_fCompletionPortDrained = FALSE; m_debuggerActivePatchSkipper = NULL; m_dwThreadHandleBeingUsed = 0; SetProfilerCallbacksAllowed(TRUE); m_pCreatingThrowableForException = NULL; #ifdef _DEBUG m_dwDisableAbortCheckCount = 0; #endif // _DEBUG #ifdef WIN64EXCEPTIONS m_dwIndexClauseForCatch = 0; m_sfEstablisherOfActualHandlerFrame.Clear(); #endif // WIN64EXCEPTIONS m_threadPoolCompletionCount = 0; Thread *pThread = GetThread(); InitContext(); if (pThread) { _ASSERTE(pThread->GetDomain()); // Start off the new thread in the default context of // the creating thread's appDomain. This could be changed by SetDelegate SetKickOffDomainId(pThread->GetDomain()->GetId()); } else SetKickOffDomainId((ADID)DefaultADID); // Do not expose thread until it is fully constructed g_pThinLockThreadIdDispenser->NewId(this, this->m_ThreadId); // // DO NOT ADD ADDITIONAL CONSTRUCTION AFTER THIS POINT. // NewId() allows this Thread instance to be accessed via a Thread Id. Do not // add additional construction after this point to prevent the race condition // of accessing a partially constructed Thread via Thread Id lookup. // exposedObjectHolder.SuppressRelease(); strongHndToExposedObjectHolder.SuppressRelease(); #if defined(_DEBUG) && defined(TRACK_SYNC) trackSyncHolder.SuppressRelease(); #endif contextHolder.SuppressRelease(); savedRedirectContextHolder.SuppressRelease(); managedThreadCurrentCulture = NULL; managedThreadCurrentUICulture = NULL; #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING m_ullProcessorUsageBaseline = 0; #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING #ifdef FEATURE_COMINTEROP m_uliInitializeSpyCookie.QuadPart = 0ul; m_fInitializeSpyRegistered = false; m_pLastSTACtxCookie = NULL; #endif // FEATURE_COMINTEROP m_fGCSpecial = FALSE; m_wCPUGroup = 0; m_pAffinityMask = 0; m_pAllLoggedTypes = NULL; #ifdef FEATURE_PERFTRACING m_pEventPipeBufferList = NULL; m_eventWriteInProgress = false; memset(&m_activityId, 0, sizeof(m_activityId)); #endif // FEATURE_PERFTRACING m_HijackReturnKind = RT_Illegal; } //-------------------------------------------------------------------- // Failable initialization occurs here. //-------------------------------------------------------------------- BOOL Thread::InitThread(BOOL fInternal) { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; HANDLE hDup = INVALID_HANDLE_VALUE; BOOL ret = TRUE; // This message actually serves a purpose (which is why it is always run) // The Stress log is run during hijacking, when other threads can be suspended // at arbitrary locations (including when holding a lock that NT uses to serialize // all memory allocations). By sending a message now, we insure that the stress // log will not allocate memory at these critical times an avoid deadlock. STRESS_LOG2(LF_ALWAYS, LL_ALWAYS, "SetupThread managed Thread %p Thread Id = %x\n", this, GetThreadId()); if ((m_State & TS_WeOwn) == 0) { COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cRecognizedThreads++); } else { COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cCurrentThreadsPhysical++); } #ifndef FEATURE_PAL // workaround: Remove this when we flow impersonation token to host. BOOL reverted = FALSE; HANDLE threadToken = INVALID_HANDLE_VALUE; #endif // !FEATURE_PAL if (m_ThreadHandle == INVALID_HANDLE_VALUE) { // For WinCE, all clients have the same handle for a thread. Duplication is // not possible. We make sure we never close this handle unless we created // the thread (TS_WeOwn). // // For Win32, each client has its own handle. This is achieved by duplicating // the pseudo-handle from ::GetCurrentThread(). Unlike WinCE, this service // returns a pseudo-handle which is only useful for duplication. In this case // each client is responsible for closing its own (duplicated) handle. // // We don't bother duplicating if WeOwn, because we created the handle in the // first place. // Thread is created when or after the physical thread started running HANDLE curProcess = ::GetCurrentProcess(); #ifndef FEATURE_PAL // If we're impersonating on NT, then DuplicateHandle(GetCurrentThread()) is going to give us a handle with only // THREAD_TERMINATE, THREAD_QUERY_INFORMATION, and THREAD_SET_INFORMATION. This doesn't include // THREAD_SUSPEND_RESUME nor THREAD_GET_CONTEXT. We need to be able to suspend the thread, and we need to be // able to get its context. Therefore, if we're impersonating, we revert to self, dup the handle, then // re-impersonate before we leave this routine. if (!RevertIfImpersonated(&reverted, &threadToken)) { COMPlusThrowWin32(); } class EnsureResetThreadToken { private: BOOL m_NeedReset; HANDLE m_threadToken; public: EnsureResetThreadToken(HANDLE threadToken, BOOL reverted) { m_threadToken = threadToken; m_NeedReset = reverted; } ~EnsureResetThreadToken() { UndoRevert(m_NeedReset, m_threadToken); if (m_threadToken != INVALID_HANDLE_VALUE) { CloseHandle(m_threadToken); } } }; EnsureResetThreadToken resetToken(threadToken, reverted); #endif // !FEATURE_PAL if (::DuplicateHandle(curProcess, ::GetCurrentThread(), curProcess, &hDup, 0 /*ignored*/, FALSE /*inherit*/, DUPLICATE_SAME_ACCESS)) { _ASSERTE(hDup != INVALID_HANDLE_VALUE); SetThreadHandle(hDup); m_WeOwnThreadHandle = TRUE; } else { COMPlusThrowWin32(); } } if ((m_State & TS_WeOwn) == 0) { if (!AllocHandles()) { ThrowOutOfMemory(); } } _ASSERTE(HasValidThreadHandle()); m_random.Init(); // Set floating point mode to round to nearest #ifndef FEATURE_PAL (void) _controlfp_s( NULL, _RC_NEAR, _RC_CHOP|_RC_UP|_RC_DOWN|_RC_NEAR ); m_pTEB = (struct _NT_TIB*)NtCurrentTeb(); #endif // !FEATURE_PAL if (m_CacheStackBase == 0) { _ASSERTE(m_CacheStackLimit == 0); _ASSERTE(m_LastAllowableStackAddress == 0); _ASSERTE(m_ProbeLimit == 0); ret = SetStackLimits(fAll); if (ret == FALSE) { ThrowOutOfMemory(); } } ret = Thread::AllocateIOCompletionContext(); if (!ret) { ThrowOutOfMemory(); } _ASSERTE(ret); // every failure case for ret should throw. return ret; } // Allocate all the handles. When we are kicking of a new thread, we can call // here before the thread starts running. BOOL Thread::AllocHandles() { WRAPPER_NO_CONTRACT; _ASSERTE(!m_DebugSuspendEvent.IsValid()); _ASSERTE(!m_EventWait.IsValid()); BOOL fOK = TRUE; EX_TRY { // create a manual reset event for getting the thread to a safe point m_DebugSuspendEvent.CreateManualEvent(FALSE); m_EventWait.CreateManualEvent(TRUE); } EX_CATCH { fOK = FALSE; if (!m_DebugSuspendEvent.IsValid()) { m_DebugSuspendEvent.CloseEvent(); } if (!m_EventWait.IsValid()) { m_EventWait.CloseEvent(); } } EX_END_CATCH(RethrowTerminalExceptions); return fOK; } //-------------------------------------------------------------------- // This is the alternate path to SetupThread/InitThread. If we created // an unstarted thread, we have SetupUnstartedThread/HasStarted. //-------------------------------------------------------------------- BOOL Thread::HasStarted(BOOL bRequiresTSL) { CONTRACTL { NOTHROW; DISABLED(GC_NOTRIGGER); SO_TOLERANT; } CONTRACTL_END; // @todo need a probe that tolerates not having a thread setup at all CONTRACT_VIOLATION(SOToleranceViolation); _ASSERTE(!m_fPreemptiveGCDisabled); // can't use PreemptiveGCDisabled() here // This is cheating a little. There is a pathway here from SetupThread, but only // via IJW SystemDomain::RunDllMain. Normally SetupThread returns a thread in // preemptive mode, ready for a transition. But in the IJW case, it can return a // cooperative mode thread. RunDllMain handles this "surprise" correctly. m_fPreemptiveGCDisabled = TRUE; // Normally, HasStarted is called from the thread's entrypoint to introduce it to // the runtime. But sometimes that thread is used for DLL_THREAD_ATTACH notifications // that call into managed code. In that case, the second HasStarted call is // redundant and should be ignored. if (GetThread() == this) return TRUE; _ASSERTE(GetThread() == 0); _ASSERTE(HasValidThreadHandle()); BOOL fKeepTLS = FALSE; BOOL fCanCleanupCOMState = FALSE; BOOL res = TRUE; res = SetStackLimits(fAll); if (res == FALSE) { m_pExceptionDuringStartup = Exception::GetOOMException(); goto FAILURE; } // If any exception happens during HasStarted, we will cache the exception in Thread::m_pExceptionDuringStartup // which will be thrown in Thread.Start as an internal exception EX_TRY { // // Initialization must happen in the following order - hosts like SQL Server depend on this. // CExecutionEngine::SetupTLSForThread(this); fCanCleanupCOMState = TRUE; res = PrepareApartmentAndContext(); if (!res) { ThrowOutOfMemory(); } InitThread(FALSE); if (SetThread(this) == FALSE) { ThrowOutOfMemory(); } if (SetAppDomain(m_pDomain) == FALSE) { ThrowOutOfMemory(); } SetupThreadForHost(); ThreadStore::TransferStartedThread(this, bRequiresTSL); #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING if (g_fEnableARM) { QueryThreadProcessorUsage(); } #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING #ifdef FEATURE_EVENT_TRACE ETW::ThreadLog::FireThreadCreated(this); #endif // FEATURE_EVENT_TRACE } EX_CATCH { if (__pException != NULL) { __pException.SuppressRelease(); m_pExceptionDuringStartup = __pException; } res = FALSE; } EX_END_CATCH(SwallowAllExceptions); FAILURE: if (res == FALSE) { if (m_fPreemptiveGCDisabled) { m_fPreemptiveGCDisabled = FALSE; } _ASSERTE (HasThreadState(TS_Unstarted)); SetThreadState(TS_FailStarted); if (GetThread() != NULL && IsAbortRequested()) UnmarkThreadForAbort(TAR_ALL); if (!fKeepTLS) { #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT // // Undo our call to PrepareApartmentAndContext above, so we don't leak a CoInitialize // If we're keeping TLS, then the host's call to ExitTask will clean this up instead. // if (fCanCleanupCOMState) { // The thread pointer in TLS may not be set yet, if we had a failure before we set it. // So we'll set it up here (we'll unset it a few lines down). if (SetThread(this) != FALSE) { CleanupCOMState(); } } #endif FastInterlockDecrement(&ThreadStore::s_pThreadStore->m_PendingThreadCount); // One of the components of OtherThreadsComplete() has changed, so check whether // we should now exit the EE. ThreadStore::CheckForEEShutdown(); DecExternalCount(/*holdingLock*/ !bRequiresTSL); SetThread(NULL); SetAppDomain(NULL); } } else { FastInterlockOr((ULONG *) &m_State, TS_FullyInitialized); #ifdef DEBUGGING_SUPPORTED // // If we're debugging, let the debugger know that this // thread is up and running now. // if (CORDebuggerAttached()) { g_pDebugInterface->ThreadCreated(this); } else { LOG((LF_CORDB, LL_INFO10000, "ThreadCreated() not called due to CORDebuggerAttached() being FALSE for thread 0x%x\n", GetThreadId())); } #endif // DEBUGGING_SUPPORTED #ifdef PROFILING_SUPPORTED // If a profiler is running, let them know about the new thread. // // The call to IsGCSpecial is crucial to avoid a deadlock. See code:Thread::m_fGCSpecial for more // information if (!IsGCSpecial()) { BEGIN_PIN_PROFILER(CORProfilerTrackThreads()); BOOL gcOnTransition = GC_ON_TRANSITIONS(FALSE); // disable GCStress 2 to avoid the profiler receiving a RuntimeThreadSuspended notification even before the ThreadCreated notification { GCX_PREEMP(); g_profControlBlock.pProfInterface->ThreadCreated((ThreadID) this); } GC_ON_TRANSITIONS(gcOnTransition); DWORD osThreadId = ::GetCurrentThreadId(); g_profControlBlock.pProfInterface->ThreadAssignedToOSThread( (ThreadID) this, osThreadId); END_PIN_PROFILER(); } #endif // PROFILING_SUPPORTED // CoreCLR does not support user-requested thread suspension _ASSERTE(!(m_State & TS_SuspendUnstarted)); } return res; } BOOL Thread::AllocateIOCompletionContext() { WRAPPER_NO_CONTRACT; PIOCompletionContext pIOC = new (nothrow) IOCompletionContext; if(pIOC != NULL) { pIOC->lpOverlapped = NULL; m_pIOCompletionContext = pIOC; return TRUE; } else { return FALSE; } } VOID Thread::FreeIOCompletionContext() { WRAPPER_NO_CONTRACT; if (m_pIOCompletionContext != NULL) { PIOCompletionContext pIOC = (PIOCompletionContext) m_pIOCompletionContext; delete pIOC; m_pIOCompletionContext = NULL; } } void Thread::HandleThreadStartupFailure() { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; _ASSERTE(GetThread() != NULL); struct ProtectArgs { OBJECTREF pThrowable; OBJECTREF pReason; } args; memset(&args, 0, sizeof(ProtectArgs)); GCPROTECT_BEGIN(args); MethodTable *pMT = MscorlibBinder::GetException(kThreadStartException); args.pThrowable = AllocateObject(pMT); MethodDescCallSite exceptionCtor(METHOD__THREAD_START_EXCEPTION__EX_CTOR); if (m_pExceptionDuringStartup) { args.pReason = CLRException::GetThrowableFromException(m_pExceptionDuringStartup); Exception::Delete(m_pExceptionDuringStartup); m_pExceptionDuringStartup = NULL; } ARG_SLOT args1[] = { ObjToArgSlot(args.pThrowable), ObjToArgSlot(args.pReason), }; exceptionCtor.Call(args1); GCPROTECT_END(); //Prot RaiseTheExceptionInternalOnly(args.pThrowable, FALSE); } #ifndef FEATURE_PAL BOOL RevertIfImpersonated(BOOL *bReverted, HANDLE *phToken) { WRAPPER_NO_CONTRACT; BOOL bImpersonated = OpenThreadToken(GetCurrentThread(), // we are assuming that if this call fails, TOKEN_IMPERSONATE, // we are not impersonating. There is no win32 TRUE, // api to figure this out. The only alternative phToken); // is to use NtCurrentTeb->IsImpersonating(). if (bImpersonated) { *bReverted = RevertToSelf(); return *bReverted; } return TRUE; } void UndoRevert(BOOL bReverted, HANDLE hToken) { if (bReverted) { if (!SetThreadToken(NULL, hToken)) { _ASSERT("Undo Revert -> SetThreadToken failed"); STRESS_LOG1(LF_EH, LL_INFO100, "UndoRevert/SetThreadToken failed for hToken = %d\n",hToken); EEPOLICY_HANDLE_FATAL_ERROR(COR_E_SECURITY); } } return; } #endif // !FEATURE_PAL // We don't want ::CreateThread() calls scattered throughout the source. So gather // them all here. BOOL Thread::CreateNewThread(SIZE_T stackSize, LPTHREAD_START_ROUTINE start, void *args, LPCWSTR pName) { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; BOOL bRet; //This assert is here to prevent a bug in the future // CreateTask currently takes a DWORD and we will downcast // if that interface changes to take a SIZE_T this Assert needs to be removed. // _ASSERTE(stackSize <= 0xFFFFFFFF); #ifndef FEATURE_PAL HandleHolder token; BOOL bReverted = FALSE; bRet = RevertIfImpersonated(&bReverted, &token); if (bRet != TRUE) return bRet; #endif // !FEATURE_PAL m_StateNC = (ThreadStateNoConcurrency)((ULONG)m_StateNC | TSNC_CLRCreatedThread); bRet = CreateNewOSThread(stackSize, start, args); #ifndef FEATURE_PAL UndoRevert(bReverted, token); if (pName != NULL) SetThreadName(m_ThreadHandle, pName); #endif // !FEATURE_PAL return bRet; } // This is to avoid the 64KB/1MB aliasing problem present on Pentium 4 processors, // which can significantly impact performance with HyperThreading enabled DWORD WINAPI Thread::intermediateThreadProc(PVOID arg) { WRAPPER_NO_CONTRACT; m_offset_counter++; if (m_offset_counter * offset_multiplier > (int) GetOsPageSize()) m_offset_counter = 0; (void)_alloca(m_offset_counter * offset_multiplier); intermediateThreadParam* param = (intermediateThreadParam*)arg; LPTHREAD_START_ROUTINE ThreadFcnPtr = param->lpThreadFunction; PVOID args = param->lpArg; delete param; return ThreadFcnPtr(args); } HANDLE Thread::CreateUtilityThread(Thread::StackSizeBucket stackSizeBucket, LPTHREAD_START_ROUTINE start, void *args, LPCWSTR pName, DWORD flags, DWORD* pThreadId) { LIMITED_METHOD_CONTRACT; // TODO: we should always use small stacks for most of these threads. For CLR 4, we're being conservative // here because this is a last-minute fix. SIZE_T stackSize; switch (stackSizeBucket) { case StackSize_Small: stackSize = 256 * 1024; break; case StackSize_Medium: stackSize = 512 * 1024; break; default: _ASSERTE(!"Bad stack size bucket"); case StackSize_Large: stackSize = 1024 * 1024; break; } flags |= STACK_SIZE_PARAM_IS_A_RESERVATION; DWORD threadId; HANDLE hThread = CreateThread(NULL, stackSize, start, args, flags, &threadId); #ifndef FEATURE_PAL SetThreadName(hThread, pName); #endif // !FEATURE_PAL if (pThreadId) *pThreadId = threadId; return hThread; } BOOL Thread::GetProcessDefaultStackSize(SIZE_T* reserveSize, SIZE_T* commitSize) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // // Let's get the stack sizes from the PE file that started process. // static SIZE_T ExeSizeOfStackReserve = 0; static SIZE_T ExeSizeOfStackCommit = 0; static BOOL fSizesGot = FALSE; #ifndef FEATURE_PAL if (!fSizesGot) { HINSTANCE hInst = WszGetModuleHandle(NULL); _ASSERTE(hInst); // WszGetModuleHandle should never fail on the module that started the process. EX_TRY { PEDecoder pe(hInst); pe.GetEXEStackSizes(&ExeSizeOfStackReserve, &ExeSizeOfStackCommit); fSizesGot = TRUE; } EX_CATCH { fSizesGot = FALSE; } EX_END_CATCH(SwallowAllExceptions); } #endif // !FEATURE_PAL if (!fSizesGot) { //return some somewhat-reasonable numbers if (NULL != reserveSize) *reserveSize = 256*1024; if (NULL != commitSize) *commitSize = 256*1024; return FALSE; } if (NULL != reserveSize) *reserveSize = ExeSizeOfStackReserve; if (NULL != commitSize) *commitSize = ExeSizeOfStackCommit; return TRUE; } BOOL Thread::CreateNewOSThread(SIZE_T sizeToCommitOrReserve, LPTHREAD_START_ROUTINE start, void *args) { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; DWORD ourId = 0; HANDLE h = NULL; DWORD dwCreationFlags = CREATE_SUSPENDED; dwCreationFlags |= STACK_SIZE_PARAM_IS_A_RESERVATION; #ifndef FEATURE_PAL // the PAL does its own adjustments as necessary if (sizeToCommitOrReserve != 0 && sizeToCommitOrReserve <= GetOsPageSize()) { // On Windows, passing a value that is <= one page size bizarrely causes the OS to use the default stack size instead of // a minimum, which is undesirable. This adjustment fixes that issue to use a minimum stack size (typically 64 KB). sizeToCommitOrReserve = GetOsPageSize() + 1; } #endif // !FEATURE_PAL intermediateThreadParam* lpThreadArgs = new (nothrow) intermediateThreadParam; if (lpThreadArgs == NULL) { return FALSE; } NewHolder argHolder(lpThreadArgs); // Make sure we have all our handles, in case someone tries to suspend us // as we are starting up. if (!AllocHandles()) { // OS is out of handles/memory? return FALSE; } lpThreadArgs->lpThreadFunction = start; lpThreadArgs->lpArg = args; h = ::CreateThread(NULL /*=SECURITY_ATTRIBUTES*/, sizeToCommitOrReserve, intermediateThreadProc, lpThreadArgs, dwCreationFlags, &ourId); if (h == NULL) return FALSE; argHolder.SuppressRelease(); _ASSERTE(!m_fPreemptiveGCDisabled); // leave in preemptive until HasStarted. SetThreadHandle(h); m_WeOwnThreadHandle = TRUE; // Before we do the resume, we need to take note of the new ThreadId. This // is necessary because -- before the thread starts executing at KickofThread -- // it may perform some DllMain DLL_THREAD_ATTACH notifications. These could // call into managed code. During the consequent SetupThread, we need to // perform the Thread::HasStarted call instead of going through the normal // 'new thread' pathway. _ASSERTE(GetOSThreadId() == 0); _ASSERTE(ourId != 0); m_OSThreadId = ourId; FastInterlockIncrement(&ThreadStore::s_pThreadStore->m_PendingThreadCount); #ifdef _DEBUG m_Creater.SetToCurrentThread(); #endif return TRUE; } // // #threadDestruction // // General comments on thread destruction. // // The C++ Thread object can survive beyond the time when the Win32 thread has died. // This is important if an exposed object has been created for this thread. The // exposed object will survive until it is GC'ed. // // A client like an exposed object can place an external reference count on that // object. We also place a reference count on it when we construct it, and we lose // that count when the thread finishes doing useful work (OnThreadTerminate). // // One way OnThreadTerminate() is called is when the thread finishes doing useful // work. This case always happens on the correct thread. // // The other way OnThreadTerminate() is called is during product shutdown. We do // a "best effort" to eliminate all threads except the Main thread before shutdown // happens. But there may be some background threads or external threads still // running. // // When the final reference count disappears, we destruct. Until then, the thread // remains in the ThreadStore, but is marked as "Dead". // // @TODO cwb: for a typical shutdown, only background threads are still around. // Should we interrupt them? What about the non-typical shutdown? int Thread::IncExternalCount() { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; Thread *pCurThread = GetThread(); _ASSERTE(m_ExternalRefCount > 0); int retVal = FastInterlockIncrement((LONG*)&m_ExternalRefCount); // If we have an exposed object and the refcount is greater than one // we must make sure to keep a strong handle to the exposed object // so that we keep it alive even if nobody has a reference to it. if (pCurThread && ((*((void**)m_ExposedObject)) != NULL)) { // The exposed object exists and needs a strong handle so check // to see if it has one. // Only a managed thread can setup StrongHnd. if ((*((void**)m_StrongHndToExposedObject)) == NULL) { GCX_COOP(); // Store the object in the strong handle. StoreObjectInHandle(m_StrongHndToExposedObject, ObjectFromHandle(m_ExposedObject)); } } return retVal; } int Thread::DecExternalCount(BOOL holdingLock) { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; // Note that it's possible to get here with a NULL current thread (during // shutdown of the thread manager). Thread *pCurThread = GetThread(); _ASSERTE (pCurThread == NULL || IsAtProcessExit() || (!holdingLock && !ThreadStore::HoldingThreadStore(pCurThread)) || (holdingLock && ThreadStore::HoldingThreadStore(pCurThread))); BOOL ToggleGC = FALSE; BOOL SelfDelete = FALSE; int retVal; // Must synchronize count and exposed object handle manipulation. We use the // thread lock for this, which implies that we must be in pre-emptive mode // to begin with and avoid any activity that would invoke a GC (this // acquires the thread store lock). if (pCurThread) { // TODO: we would prefer to use a GC Holder here, however it is hard // to get the case where we're deleting this thread correct given // the current macros. We want to supress the release of the holder // here which puts us in Preemptive mode, and also the switch to // Cooperative mode below, but since both holders will be named // the same thing (due to the generic nature of the macro) we can // not use GCX_*_SUPRESS_RELEASE() for 2 holders in the same scope // b/c they will both apply simply to the most narrowly scoped // holder. ToggleGC = pCurThread->PreemptiveGCDisabled(); if (ToggleGC) { pCurThread->EnablePreemptiveGC(); } } GCX_ASSERT_PREEMP(); ThreadStoreLockHolder tsLock(!holdingLock); _ASSERTE(m_ExternalRefCount >= 1); _ASSERTE(!holdingLock || ThreadStore::s_pThreadStore->m_Crst.GetEnterCount() > 0 || IsAtProcessExit()); retVal = FastInterlockDecrement((LONG*)&m_ExternalRefCount); if (retVal == 0) { HANDLE h = GetThreadHandle(); if (h == INVALID_HANDLE_VALUE) { h = m_ThreadHandleForClose; m_ThreadHandleForClose = INVALID_HANDLE_VALUE; } // Can not assert like this. We have already removed the Unstarted bit. //_ASSERTE (IsUnstarted() || h != INVALID_HANDLE_VALUE); if (h != INVALID_HANDLE_VALUE && m_WeOwnThreadHandle) { ::CloseHandle(h); SetThreadHandle(INVALID_HANDLE_VALUE); } // Switch back to cooperative mode to manipulate the thread. if (pCurThread) { // TODO: we would prefer to use GCX_COOP here, see comment above. pCurThread->DisablePreemptiveGC(); } GCX_ASSERT_COOP(); // during process detach the thread might still be in the thread list // if it hasn't seen its DLL_THREAD_DETACH yet. Use the following // tweak to decide if the thread has terminated yet. if (!HasValidThreadHandle()) { SelfDelete = this == pCurThread; m_ExceptionState.FreeAllStackTraces(); if (SelfDelete) { SetThread(NULL); } delete this; } tsLock.Release(); // It only makes sense to restore the GC mode if we didn't just destroy // our own thread object. if (pCurThread && !SelfDelete && !ToggleGC) { pCurThread->EnablePreemptiveGC(); } // Cannot use this here b/c it creates a holder named the same as GCX_ASSERT_COOP // in the same scope above... // // GCX_ASSERT_PREEMP() return retVal; } else if (pCurThread == NULL) { // We're in shutdown, too late to be worrying about having a strong // handle to the exposed thread object, we've already performed our // final GC. tsLock.Release(); return retVal; } else { // Check to see if the external ref count reaches exactly one. If this // is the case and we have an exposed object then it is that exposed object // that is holding a reference to us. To make sure that we are not the // ones keeping the exposed object alive we need to remove the strong // reference we have to it. if ((retVal == 1) && ((*((void**)m_StrongHndToExposedObject)) != NULL)) { // Switch back to cooperative mode to manipulate the object. // Don't want to switch back to COOP until we let go of the lock // however we are allowed to call StoreObjectInHandle here in preemptive // mode because we are setting the value to NULL. CONTRACT_VIOLATION(ModeViolation); // Clear the handle and leave the lock. // We do not have to to DisablePreemptiveGC here, because // we just want to put NULL into a handle. StoreObjectInHandle(m_StrongHndToExposedObject, NULL); tsLock.Release(); // Switch back to the initial GC mode. if (ToggleGC) { pCurThread->DisablePreemptiveGC(); } GCX_ASSERT_COOP(); return retVal; } } tsLock.Release(); // Switch back to the initial GC mode. if (ToggleGC) { pCurThread->DisablePreemptiveGC(); } return retVal; } //-------------------------------------------------------------------- // Destruction. This occurs after the associated native thread // has died. //-------------------------------------------------------------------- Thread::~Thread() { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; // TODO: enable this //_ASSERTE(GetThread() != this); _ASSERTE(m_ThrewControlForThread == 0); // AbortRequest is coupled with TrapReturningThread. // We should have unmarked the thread for abort. // !!! Can not assert here. If a thread has no managed code on stack // !!! we leave the g_TrapReturningThread set so that the thread will be // !!! aborted if it enters managed code. //_ASSERTE(!IsAbortRequested()); // We should not have the Thread marked for abort. But if we have // we need to unmark it so that g_TrapReturningThreads is decremented. if (IsAbortRequested()) { UnmarkThreadForAbort(TAR_ALL); } #if defined(_DEBUG) && defined(TRACK_SYNC) _ASSERTE(IsAtProcessExit() || ((Dbg_TrackSyncStack *) m_pTrackSync)->m_StackPointer == 0); delete m_pTrackSync; #endif // TRACK_SYNC _ASSERTE(IsDead() || IsUnstarted() || IsAtProcessExit()); if (m_WaitEventLink.m_Next != NULL && !IsAtProcessExit()) { WaitEventLink *walk = &m_WaitEventLink; while (walk->m_Next) { ThreadQueue::RemoveThread(this, (SyncBlock*)((DWORD_PTR)walk->m_Next->m_WaitSB & ~1)); StoreEventToEventStore (walk->m_Next->m_EventWait); } m_WaitEventLink.m_Next = NULL; } if (m_StateNC & TSNC_ExistInThreadStore) { BOOL ret; ret = ThreadStore::RemoveThread(this); _ASSERTE(ret); } #ifdef _DEBUG m_pFrame = (Frame *)POISONC; #endif // Update Perfmon counters. COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cCurrentThreadsLogical--); // Current recognized threads are non-runtime threads that are alive and ran under the // runtime. Check whether this Thread was one of them. if ((m_State & TS_WeOwn) == 0) { COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cRecognizedThreads--); } else { COUNTER_ONLY(GetPerfCounters().m_LocksAndThreads.cCurrentThreadsPhysical--); } // Normally we shouldn't get here with a valid thread handle; however if SetupThread // failed (due to an OOM for example) then we need to CloseHandle the thread // handle if we own it. if (m_WeOwnThreadHandle && (GetThreadHandle() != INVALID_HANDLE_VALUE)) { CloseHandle(GetThreadHandle()); } if (m_DebugSuspendEvent.IsValid()) { m_DebugSuspendEvent.CloseEvent(); } if (m_EventWait.IsValid()) { m_EventWait.CloseEvent(); } FreeIOCompletionContext(); if (m_OSContext) delete m_OSContext; if (GetSavedRedirectContext()) { delete GetSavedRedirectContext(); SetSavedRedirectContext(NULL); } #ifdef FEATURE_COMINTEROP if (m_pRCWStack) delete m_pRCWStack; #endif if (m_pExceptionDuringStartup) { Exception::Delete (m_pExceptionDuringStartup); } ClearContext(); if (!IsAtProcessExit()) { // Destroy any handles that we're using to hold onto exception objects SafeSetThrowables(NULL); DestroyShortWeakHandle(m_ExposedObject); DestroyStrongHandle(m_StrongHndToExposedObject); } g_pThinLockThreadIdDispenser->DisposeId(GetThreadId()); #ifdef FEATURE_PREJIT if (m_pIBCInfo) { delete m_pIBCInfo; } #endif #ifdef FEATURE_EVENT_TRACE // Destruct the thread local type cache for allocation sampling if(m_pAllLoggedTypes) { ETW::TypeSystemLog::DeleteTypeHashNoLock(&m_pAllLoggedTypes); } #endif // FEATURE_EVENT_TRACE // Wait for another thread to leave its loop in DeadlockAwareLock::TryBeginEnterLock CrstHolder lock(&g_DeadlockAwareCrst); } #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT void Thread::BaseCoUninitialize() { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_SO_INTOLERANT; STATIC_CONTRACT_MODE_PREEMPTIVE; _ASSERTE(GetThread() == this); BEGIN_SO_TOLERANT_CODE(this); // BEGIN_SO_TOLERANT_CODE wraps a __try/__except around this call, so if the OS were to allow // an exception to leak through to us, we'll catch it. ::CoUninitialize(); END_SO_TOLERANT_CODE; }// BaseCoUninitialize #ifdef FEATURE_COMINTEROP void Thread::BaseWinRTUninitialize() { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_SO_INTOLERANT; STATIC_CONTRACT_MODE_PREEMPTIVE; _ASSERTE(WinRTSupported()); _ASSERTE(GetThread() == this); _ASSERTE(IsWinRTInitialized()); BEGIN_SO_TOLERANT_CODE(this); RoUninitialize(); END_SO_TOLERANT_CODE; } #endif // FEATURE_COMINTEROP void Thread::CoUninitialize() { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; // Running threads might have performed a CoInitialize which must // now be balanced. BOOL needsUninitialize = IsCoInitialized() #ifdef FEATURE_COMINTEROP || IsWinRTInitialized() #endif // FEATURE_COMINTEROP ; if (!IsAtProcessExit() && needsUninitialize) { GCX_PREEMP(); CONTRACT_VIOLATION(ThrowsViolation); if (IsCoInitialized()) { BaseCoUninitialize(); FastInterlockAnd((ULONG *)&m_State, ~TS_CoInitialized); } #ifdef FEATURE_COMINTEROP if (IsWinRTInitialized()) { _ASSERTE(WinRTSupported()); BaseWinRTUninitialize(); ResetWinRTInitialized(); } #endif // FEATURE_COMNITEROP } } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT void Thread::CleanupDetachedThreads() { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE(!ThreadStore::HoldingThreadStore()); ThreadStoreLockHolder threadStoreLockHolder; Thread *thread = ThreadStore::GetAllThreadList(NULL, 0, 0); STRESS_LOG0(LF_SYNC, LL_INFO1000, "T::CDT called\n"); while (thread != NULL) { Thread *next = ThreadStore::GetAllThreadList(thread, 0, 0); if (thread->IsDetached() && thread->m_UnmanagedRefCount == 0) { STRESS_LOG1(LF_SYNC, LL_INFO1000, "T::CDT - detaching thread 0x%p\n", thread); // Unmark that the thread is detached while we have the // thread store lock. This will ensure that no other // thread will race in here and try to delete it, too. FastInterlockAnd((ULONG*)&(thread->m_State), ~TS_Detached); FastInterlockDecrement(&m_DetachCount); if (!thread->IsBackground()) FastInterlockDecrement(&m_ActiveDetachCount); // If the debugger is attached, then we need to unlock the // thread store before calling OnThreadTerminate. That // way, we won't be holding the thread store lock if we // need to block sending a detach thread event. BOOL debuggerAttached = #ifdef DEBUGGING_SUPPORTED CORDebuggerAttached(); #else // !DEBUGGING_SUPPORTED FALSE; #endif // !DEBUGGING_SUPPORTED if (debuggerAttached) ThreadStore::UnlockThreadStore(); thread->OnThreadTerminate(debuggerAttached ? FALSE : TRUE); #ifdef DEBUGGING_SUPPORTED if (debuggerAttached) { ThreadSuspend::LockThreadStore(ThreadSuspend::SUSPEND_OTHER); // We remember the next Thread in the thread store // list before deleting the current one. But we can't // use that Thread pointer now that we release the // thread store lock in the middle of the loop. We // have to start from the beginning of the list every // time. If two threads T1 and T2 race into // CleanupDetachedThreads, then T1 will grab the first // Thread on the list marked for deletion and release // the lock. T2 will grab the second one on the // list. T2 may complete destruction of its Thread, // then T1 might re-acquire the thread store lock and // try to use the next Thread in the thread store. But // T2 just deleted that next Thread. thread = ThreadStore::GetAllThreadList(NULL, 0, 0); } else #endif // DEBUGGING_SUPPORTED { thread = next; } } else if (thread->HasThreadState(TS_Finalized)) { STRESS_LOG1(LF_SYNC, LL_INFO1000, "T::CDT - finalized thread 0x%p\n", thread); thread->ResetThreadState(TS_Finalized); // We have finalized the managed Thread object. Now it is time to clean up the unmanaged part thread->DecExternalCount(TRUE); thread = next; } else { thread = next; } } s_fCleanFinalizedThread = FALSE; } #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT void Thread::CleanupCOMState() { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; #ifdef FEATURE_COMINTEROP if (GetFinalApartment() == Thread::AS_InSTA) ReleaseRCWsInCachesNoThrow(GetCurrentCtxCookie()); #endif // FEATURE_COMINTEROP // Running threads might have performed a CoInitialize which must // now be balanced. However only the thread that called COInitialize can // call CoUninitialize. BOOL needsUninitialize = IsCoInitialized() #ifdef FEATURE_COMINTEROP || IsWinRTInitialized() #endif // FEATURE_COMINTEROP ; if (needsUninitialize) { GCX_PREEMP(); CONTRACT_VIOLATION(ThrowsViolation); if (IsCoInitialized()) { BaseCoUninitialize(); ResetCoInitialized(); } #ifdef FEATURE_COMINTEROP if (IsWinRTInitialized()) { _ASSERTE(WinRTSupported()); BaseWinRTUninitialize(); ResetWinRTInitialized(); } #endif // FEATURE_COMINTEROP } } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT // See general comments on thread destruction (code:#threadDestruction) above. void Thread::OnThreadTerminate(BOOL holdingLock) { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; // #ReportDeadOnThreadTerminate // Caller should have put the TS_ReportDead bit on by now. // We don't want any windows after the exit event but before the thread is marked dead. // If a debugger attached during such a window (or even took a dump at the exit event), // then it may not realize the thread is dead. // So ensure we mark the thread as dead before we send the tool notifications. // The TS_ReportDead bit will cause the debugger to view this as TS_Dead. _ASSERTE(HasThreadState(TS_ReportDead)); // Should not use OSThreadId: // OSThreadId may change for the current thread is the thread is blocked and rescheduled // by host. Thread *pCurrentThread = GetThread(); DWORD CurrentThreadID = pCurrentThread?pCurrentThread->GetThreadId():0; DWORD ThisThreadID = GetThreadId(); #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT // If the currently running thread is the thread that died and it is an STA thread, then we // need to release all the RCW's in the current context. However, we cannot do this if we // are in the middle of process detach. if (!IsAtProcessExit() && this == GetThread()) { CleanupCOMState(); } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT if (g_fEEShutDown != 0) { // We have started shutdown. Not safe to touch CLR state. return; } // We took a count during construction, and we rely on the count being // non-zero as we terminate the thread here. _ASSERTE(m_ExternalRefCount > 0); // The thread is no longer running. It's important that we zero any general OBJECTHANDLE's // on this Thread object. That's because we need the managed Thread object to be subject to // GC and yet any HANDLE is opaque to the GC when it comes to collecting cycles. If e.g. the // Thread's AbortReason (which is an arbitrary object) contains transitively a reference back // to the Thread, then we have an uncollectible cycle. When the thread is executing, nothing // can be collected anyway. But now that we stop running the cycle concerns us. // // It's important that we only use OBJECTHANDLE's that are retrievable while the thread is // still running. That's what allows us to zero them here with impunity: { // No handles to clean up in the m_ExceptionState _ASSERTE(!m_ExceptionState.IsExceptionInProgress()); GCX_COOP(); // Destroy the LastThrown handle (and anything that violates the above assert). SafeSetThrowables(NULL); // Cleaning up the AbortReason is tricky, since the handle is only valid if the ADID is valid // ...and we can only perform this operation if other threads aren't racing to update these // values on our thread asynchronously. ClearAbortReason(); // Free all structures related to thread statics for this thread DeleteThreadStaticData(); } if (GCHeapUtilities::IsGCHeapInitialized()) { // Guaranteed to NOT be a shutdown case, because we tear down the heap before // we tear down any threads during shutdown. if (ThisThreadID == CurrentThreadID) { GCX_COOP(); GCHeapUtilities::GetGCHeap()->FixAllocContext(&m_alloc_context, false, NULL, NULL); m_alloc_context.init(); } } // We switch a thread to dead when it has finished doing useful work. But it // remains in the thread store so long as someone keeps it alive. An exposed // object will do this (it releases the refcount in its finalizer). If the // thread is never released, we have another look during product shutdown and // account for the unreleased refcount of the uncollected exposed object: if (IsDead()) { GCX_COOP(); _ASSERTE(IsAtProcessExit()); ClearContext(); if (m_ExposedObject != NULL) DecExternalCount(holdingLock); // may destruct now } else { #ifdef DEBUGGING_SUPPORTED // // If we're debugging, let the debugger know that this thread is // gone. // // There is a race here where the debugger could have attached after // we checked (and thus didn't release the lock). In this case, // we can't call out to the debugger or we risk a deadlock. // if (!holdingLock && CORDebuggerAttached()) { g_pDebugInterface->DetachThread(this); } #endif // DEBUGGING_SUPPORTED #ifdef PROFILING_SUPPORTED // If a profiler is present, then notify the profiler of thread destroy { BEGIN_PIN_PROFILER(CORProfilerTrackThreads()); GCX_PREEMP(); g_profControlBlock.pProfInterface->ThreadDestroyed((ThreadID) this); END_PIN_PROFILER(); } #endif // PROFILING_SUPPORTED if (!holdingLock) { LOG((LF_SYNC, INFO3, "OnThreadTerminate obtain lock\n")); ThreadSuspend::LockThreadStore(ThreadSuspend::SUSPEND_OTHER); } if (GCHeapUtilities::IsGCHeapInitialized() && ThisThreadID != CurrentThreadID) { // We must be holding the ThreadStore lock in order to clean up alloc context. // We should never call FixAllocContext during GC. GCHeapUtilities::GetGCHeap()->FixAllocContext(&m_alloc_context, false, NULL, NULL); m_alloc_context.init(); } FastInterlockOr((ULONG *) &m_State, TS_Dead); ThreadStore::s_pThreadStore->m_DeadThreadCount++; ThreadStore::s_pThreadStore->IncrementDeadThreadCountForGCTrigger(); if (IsUnstarted()) ThreadStore::s_pThreadStore->m_UnstartedThreadCount--; else { if (IsBackground()) ThreadStore::s_pThreadStore->m_BackgroundThreadCount--; } FastInterlockAnd((ULONG *) &m_State, ~(TS_Unstarted | TS_Background)); // // If this thread was told to trip for debugging between the // sending of the detach event above and the locking of the // thread store lock, then remove the flag and decrement the // global trap returning threads count. // if (!IsAtProcessExit()) { // A thread can't die during a GCPending, because the thread store's // lock is held by the GC thread. if (m_State & TS_DebugSuspendPending) UnmarkForSuspension(~TS_DebugSuspendPending); // CoreCLR does not support user-requested thread suspension _ASSERTE(!(m_State & TS_UserSuspendPending)); if (CurrentThreadID == ThisThreadID && IsAbortRequested()) { UnmarkThreadForAbort(Thread::TAR_ALL); } } if (GetThreadHandle() != INVALID_HANDLE_VALUE) { if (m_ThreadHandleForClose == INVALID_HANDLE_VALUE) { m_ThreadHandleForClose = GetThreadHandle(); } SetThreadHandle (INVALID_HANDLE_VALUE); } m_OSThreadId = 0; // If nobody else is holding onto the thread, we may destruct it here: ULONG oldCount = DecExternalCount(TRUE); // If we are shutting down the process, we only have one thread active in the // system. So we can disregard all the reasons that hold this thread alive -- // TLS is about to be reclaimed anyway. if (IsAtProcessExit()) while (oldCount > 0) { oldCount = DecExternalCount(TRUE); } // ASSUME THAT THE THREAD IS DELETED, FROM HERE ON _ASSERTE(ThreadStore::s_pThreadStore->m_ThreadCount >= 0); _ASSERTE(ThreadStore::s_pThreadStore->m_BackgroundThreadCount >= 0); _ASSERTE(ThreadStore::s_pThreadStore->m_ThreadCount >= ThreadStore::s_pThreadStore->m_BackgroundThreadCount); _ASSERTE(ThreadStore::s_pThreadStore->m_ThreadCount >= ThreadStore::s_pThreadStore->m_UnstartedThreadCount); _ASSERTE(ThreadStore::s_pThreadStore->m_ThreadCount >= ThreadStore::s_pThreadStore->m_DeadThreadCount); // One of the components of OtherThreadsComplete() has changed, so check whether // we should now exit the EE. ThreadStore::CheckForEEShutdown(); if (ThisThreadID == CurrentThreadID) { // NULL out the thread block in the tls. We can't do this if we aren't on the // right thread. But this will only happen during a shutdown. And we've made // a "best effort" to reduce to a single thread before we begin the shutdown. SetThread(NULL); SetAppDomain(NULL); } if (!holdingLock) { LOG((LF_SYNC, INFO3, "OnThreadTerminate releasing lock\n")); ThreadSuspend::UnlockThreadStore(ThisThreadID == CurrentThreadID); } } } // Helper functions to check for duplicate handles. we only do this check if // a waitfor multiple fails. int __cdecl compareHandles( const void *arg1, const void *arg2 ) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; HANDLE h1 = *(HANDLE*)arg1; HANDLE h2 = *(HANDLE*)arg2; return (h1 == h2) ? 0 : ((h1 < h2) ? -1 : 1); } BOOL CheckForDuplicateHandles(int countHandles, HANDLE *handles) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; qsort(handles,countHandles,sizeof(HANDLE),compareHandles); for (int i=1; i < countHandles; i++) { if (handles[i-1] == handles[i]) return TRUE; } return FALSE; } //-------------------------------------------------------------------- // Based on whether this thread has a message pump, do the appropriate // style of Wait. //-------------------------------------------------------------------- DWORD Thread::DoAppropriateWait(int countHandles, HANDLE *handles, BOOL waitAll, DWORD millis, WaitMode mode, PendingSync *syncState) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; INDEBUG(BOOL alertable = (mode & WaitMode_Alertable) != 0;); _ASSERTE(alertable || syncState == 0); struct Param { Thread *pThis; int countHandles; HANDLE *handles; BOOL waitAll; DWORD millis; WaitMode mode; DWORD dwRet; } param; param.pThis = this; param.countHandles = countHandles; param.handles = handles; param.waitAll = waitAll; param.millis = millis; param.mode = mode; param.dwRet = (DWORD) -1; EE_TRY_FOR_FINALLY(Param *, pParam, ¶m) { pParam->dwRet = pParam->pThis->DoAppropriateWaitWorker(pParam->countHandles, pParam->handles, pParam->waitAll, pParam->millis, pParam->mode); } EE_FINALLY { if (syncState) { if (!GOT_EXCEPTION() && param.dwRet >= WAIT_OBJECT_0 && param.dwRet < (DWORD)(WAIT_OBJECT_0 + countHandles)) { // This thread has been removed from syncblk waiting list by the signalling thread syncState->Restore(FALSE); } else syncState->Restore(TRUE); } _ASSERTE (param.dwRet != WAIT_IO_COMPLETION); } EE_END_FINALLY; return(param.dwRet); } DWORD Thread::DoAppropriateWait(AppropriateWaitFunc func, void *args, DWORD millis, WaitMode mode, PendingSync *syncState) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; INDEBUG(BOOL alertable = (mode & WaitMode_Alertable) != 0;); _ASSERTE(alertable || syncState == 0); struct Param { Thread *pThis; AppropriateWaitFunc func; void *args; DWORD millis; WaitMode mode; DWORD dwRet; } param; param.pThis = this; param.func = func; param.args = args; param.millis = millis; param.mode = mode; param.dwRet = (DWORD) -1; EE_TRY_FOR_FINALLY(Param *, pParam, ¶m) { pParam->dwRet = pParam->pThis->DoAppropriateWaitWorker(pParam->func, pParam->args, pParam->millis, pParam->mode); } EE_FINALLY { if (syncState) { if (!GOT_EXCEPTION() && WAIT_OBJECT_0 == param.dwRet) { // This thread has been removed from syncblk waiting list by the signalling thread syncState->Restore(FALSE); } else syncState->Restore(TRUE); } _ASSERTE (WAIT_IO_COMPLETION != param.dwRet); } EE_END_FINALLY; return(param.dwRet); } #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT //-------------------------------------------------------------------- // helper to do message wait //-------------------------------------------------------------------- DWORD MsgWaitHelper(int numWaiters, HANDLE* phEvent, BOOL bWaitAll, DWORD millis, BOOL bAlertable) { STATIC_CONTRACT_THROWS; // The true contract for GC trigger should be the following. But this puts a very strong restriction // on contract for functions that call EnablePreemptiveGC. //if (GetThread() && !ThreadStore::HoldingThreadStore(GetThread())) {GC_TRIGGERS;} else {GC_NOTRIGGER;} STATIC_CONTRACT_SO_INTOLERANT; STATIC_CONTRACT_GC_TRIGGERS; DWORD flags = 0; DWORD dwReturn=WAIT_ABANDONED; Thread* pThread = GetThread(); // If pThread is NULL, we'd better shut down. if (pThread == NULL) _ASSERTE (g_fEEShutDown); DWORD lastError = 0; BEGIN_SO_TOLERANT_CODE(pThread); // If we're going to pump, we cannot use WAIT_ALL. That's because the wait would // only be satisfied if a message arrives while the handles are signalled. If we // want true WAIT_ALL, we need to fire up a different thread in the MTA and wait // on his result. This isn't implemented yet. // // A change was added to WaitHandleNative::CorWaitMultipleNative to disable WaitAll // in an STA with more than one handle. if (bWaitAll) { if (numWaiters == 1) bWaitAll = FALSE; // The check that's supposed to prevent this condition from occuring, in WaitHandleNative::CorWaitMultipleNative, // is unfortunately behind FEATURE_COMINTEROP instead of FEATURE_COMINTEROP_APARTMENT_SUPPORT. // So on CoreCLR (where FEATURE_COMINTEROP is not currently defined) we can actually reach this point. // We can't fix this, because it's a breaking change, so we just won't assert here. // The result is that WaitAll on an STA thread in CoreCLR will behave stragely, as described above. } if (bWaitAll) flags |= COWAIT_WAITALL; if (bAlertable) flags |= COWAIT_ALERTABLE; HRESULT hr = S_OK; hr = CoWaitForMultipleHandles(flags, millis, numWaiters, phEvent, &dwReturn); if (hr == RPC_S_CALLPENDING) { dwReturn = WAIT_TIMEOUT; } else if (FAILED(hr)) { // The service behaves differently on an STA vs. MTA in how much // error information it propagates back, and in which form. We currently // only get here in the STA case, so bias this logic that way. dwReturn = WAIT_FAILED; } else { dwReturn += WAIT_OBJECT_0; // success -- bias back } lastError = ::GetLastError(); END_SO_TOLERANT_CODE; // END_SO_TOLERANT_CODE overwrites lasterror. Let's reset it. ::SetLastError(lastError); return dwReturn; } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT DWORD WaitForMultipleObjectsEx_SO_TOLERANT (DWORD nCount, HANDLE *lpHandles, BOOL bWaitAll,DWORD dwMilliseconds, BOOL bAlertable) { STATIC_CONTRACT_SO_INTOLERANT; DWORD dwRet = WAIT_FAILED; DWORD lastError = 0; BEGIN_SO_TOLERANT_CODE (GetThread ()); dwRet = ::WaitForMultipleObjectsEx (nCount, lpHandles, bWaitAll, dwMilliseconds, bAlertable); lastError = ::GetLastError(); END_SO_TOLERANT_CODE; // END_SO_TOLERANT_CODE overwrites lasterror. Let's reset it. ::SetLastError(lastError); return dwRet; } //-------------------------------------------------------------------- // Do appropriate wait based on apartment state (STA or MTA) DWORD Thread::DoAppropriateAptStateWait(int numWaiters, HANDLE* pHandles, BOOL bWaitAll, DWORD timeout, WaitMode mode) { CONTRACTL { THROWS; GC_TRIGGERS; SO_INTOLERANT; } CONTRACTL_END; BOOL alertable = (mode & WaitMode_Alertable) != 0; #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT if (alertable && !GetDomain()->MustForceTrivialWaitOperations()) { ApartmentState as = GetFinalApartment(); if (AS_InMTA != as) { return MsgWaitHelper(numWaiters, pHandles, bWaitAll, timeout, alertable); } } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT return WaitForMultipleObjectsEx_SO_TOLERANT(numWaiters, pHandles, bWaitAll, timeout, alertable); } // A helper called by our two flavors of DoAppropriateWaitWorker void Thread::DoAppropriateWaitWorkerAlertableHelper(WaitMode mode) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; // If thread abort is prevented, we do not want this thread to see thread abort and thread interrupt exception. if (IsAbortPrevented()) { return; } // A word about ordering for Interrupt. If someone tries to interrupt a thread // that's in the interruptible state, we queue an APC. But if they try to interrupt // a thread that's not in the interruptible state, we just record that fact. So // we have to set TS_Interruptible before we test to see whether someone wants to // interrupt us or else we have a race condition that causes us to skip the APC. FastInterlockOr((ULONG *) &m_State, TS_Interruptible); if (HasThreadStateNC(TSNC_InRestoringSyncBlock)) { // The thread is restoring SyncBlock for Object.Wait. ResetThreadStateNC(TSNC_InRestoringSyncBlock); } else { HandleThreadInterrupt((mode & WaitMode_ADUnload) != 0); // Safe to clear the interrupted state, no APC could have fired since we // reset m_UserInterrupt (which inhibits our APC callback from doing // anything). FastInterlockAnd((ULONG *) &m_State, ~TS_Interrupted); } } void MarkOSAlertableWait() { LIMITED_METHOD_CONTRACT; GetThread()->SetThreadStateNC (Thread::TSNC_OSAlertableWait); } void UnMarkOSAlertableWait() { LIMITED_METHOD_CONTRACT; GetThread()->ResetThreadStateNC (Thread::TSNC_OSAlertableWait); } //-------------------------------------------------------------------- // Based on whether this thread has a message pump, do the appropriate // style of Wait. //-------------------------------------------------------------------- DWORD Thread::DoAppropriateWaitWorker(int countHandles, HANDLE *handles, BOOL waitAll, DWORD millis, WaitMode mode) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; DWORD ret = 0; BOOL alertable = (mode & WaitMode_Alertable) != 0; // Waits from SynchronizationContext.WaitHelper are always just WaitMode_IgnoreSyncCtx. // So if we defer to a sync ctx, we will lose any extra bits. We must therefore not // defer to a sync ctx if doing any non-default wait. // If you're doing a default wait, but want to ignore sync ctx, specify WaitMode_IgnoreSyncCtx // which will make mode != WaitMode_Alertable. BOOL ignoreSyncCtx = (mode != WaitMode_Alertable); if (GetDomain()->MustForceTrivialWaitOperations()) ignoreSyncCtx = TRUE; // Unless the ignoreSyncCtx flag is set, first check to see if there is a synchronization // context on the current thread and if there is, dispatch to it to do the wait. // If the wait is non alertable we cannot forward the call to the sync context // since fundamental parts of the system (such as the GC) rely on non alertable // waits not running any managed code. Also if we are past the point in shutdown were we // are allowed to run managed code then we can't forward the call to the sync context. if (!ignoreSyncCtx && alertable && CanRunManagedCode(LoaderLockCheck::None) && !HasThreadStateNC(Thread::TSNC_BlockedForShutdown)) { GCX_COOP(); BOOL fSyncCtxPresent = FALSE; OBJECTREF SyncCtxObj = NULL; GCPROTECT_BEGIN(SyncCtxObj) { GetSynchronizationContext(&SyncCtxObj); if (SyncCtxObj != NULL) { SYNCHRONIZATIONCONTEXTREF syncRef = (SYNCHRONIZATIONCONTEXTREF)SyncCtxObj; if (syncRef->IsWaitNotificationRequired()) { fSyncCtxPresent = TRUE; ret = DoSyncContextWait(&SyncCtxObj, countHandles, handles, waitAll, millis); } } } GCPROTECT_END(); if (fSyncCtxPresent) return ret; } // Before going to pre-emptive mode the thread needs to be flagged as waiting for // the debugger. This used to be accomplished by the TS_Interruptible flag but that // doesn't work reliably, see DevDiv Bugs 699245. Some methods call in here already in // COOP mode so we set the bit before the transition. For the calls that are already // in pre-emptive mode those are still buggy. This is only a partial fix. BOOL isCoop = PreemptiveGCDisabled(); ThreadStateNCStackHolder tsNC(isCoop && alertable, TSNC_DebuggerSleepWaitJoin); GCX_PREEMP(); if (alertable) { DoAppropriateWaitWorkerAlertableHelper(mode); } StateHolder OSAlertableWait(alertable); ThreadStateHolder tsh(alertable, TS_Interruptible | TS_Interrupted); ULONGLONG dwStart = 0, dwEnd; retry: if (millis != INFINITE) { dwStart = CLRGetTickCount64(); } ret = DoAppropriateAptStateWait(countHandles, handles, waitAll, millis, mode); if (ret == WAIT_IO_COMPLETION) { _ASSERTE (alertable); if (m_State & TS_Interrupted) { HandleThreadInterrupt(mode & WaitMode_ADUnload); } // We could be woken by some spurious APC or an EE APC queued to // interrupt us. In the latter case the TS_Interrupted bit will be set // in the thread state bits. Otherwise we just go back to sleep again. if (millis != INFINITE) { dwEnd = CLRGetTickCount64(); if (dwEnd >= dwStart + millis) { ret = WAIT_TIMEOUT; goto WaitCompleted; } else { millis -= (DWORD)(dwEnd - dwStart); } } goto retry; } _ASSERTE((ret >= WAIT_OBJECT_0 && ret < (WAIT_OBJECT_0 + (DWORD)countHandles)) || (ret >= WAIT_ABANDONED && ret < (WAIT_ABANDONED + (DWORD)countHandles)) || (ret == WAIT_TIMEOUT) || (ret == WAIT_FAILED)); // countHandles is used as an unsigned -- it should never be negative. _ASSERTE(countHandles >= 0); // We support precisely one WAIT_FAILED case, where we attempt to wait on a // thread handle and the thread is in the process of dying we might get a // invalid handle substatus. Turn this into a successful wait. // There are three cases to consider: // 1) Only waiting on one handle: return success right away. // 2) Waiting for all handles to be signalled: retry the wait without the // affected handle. // 3) Waiting for one of multiple handles to be signalled: return with the // first handle that is either signalled or has become invalid. if (ret == WAIT_FAILED) { DWORD errorCode = ::GetLastError(); if (errorCode == ERROR_INVALID_PARAMETER) { if (CheckForDuplicateHandles(countHandles, handles)) COMPlusThrow(kDuplicateWaitObjectException); else COMPlusThrowHR(HRESULT_FROM_WIN32(errorCode)); } else if (errorCode == ERROR_ACCESS_DENIED) { // A Win32 ACL could prevent us from waiting on the handle. COMPlusThrow(kUnauthorizedAccessException); } else if (errorCode == ERROR_NOT_ENOUGH_MEMORY) { ThrowOutOfMemory(); } #ifdef FEATURE_PAL else if (errorCode == ERROR_NOT_SUPPORTED) { // "Wait for any" and "wait for all" operations on multiple wait handles are not supported when a cross-process sync // object is included in the array COMPlusThrow(kPlatformNotSupportedException, W("PlatformNotSupported_NamedSyncObjectWaitAnyWaitAll")); } #endif else if (errorCode != ERROR_INVALID_HANDLE) { ThrowWin32(errorCode); } if (countHandles == 1) ret = WAIT_OBJECT_0; else if (waitAll) { // Probe all handles with a timeout of zero. When we find one that's // invalid, move it out of the list and retry the wait. for (int i = 0; i < countHandles; i++) { // WaitForSingleObject won't pump memssage; we already probe enough space // before calling this function and we don't want to fail here, so we don't // do a transition to tolerant code here DWORD subRet = WaitForSingleObject (handles[i], 0); if (subRet != WAIT_FAILED) continue; _ASSERTE(::GetLastError() == ERROR_INVALID_HANDLE); if ((countHandles - i - 1) > 0) memmove(&handles[i], &handles[i+1], (countHandles - i - 1) * sizeof(HANDLE)); countHandles--; break; } // Compute the new timeout value by assume that the timeout // is not large enough for more than one wrap dwEnd = CLRGetTickCount64(); if (millis != INFINITE) { if (dwEnd >= dwStart + millis) { ret = WAIT_TIMEOUT; goto WaitCompleted; } else { millis -= (DWORD)(dwEnd - dwStart); } } goto retry; } else { // Probe all handles with a timeout as zero, succeed with the first // handle that doesn't timeout. ret = WAIT_OBJECT_0; int i; for (i = 0; i < countHandles; i++) { TryAgain: // WaitForSingleObject won't pump memssage; we already probe enough space // before calling this function and we don't want to fail here, so we don't // do a transition to tolerant code here DWORD subRet = WaitForSingleObject (handles[i], 0); if ((subRet == WAIT_OBJECT_0) || (subRet == WAIT_FAILED)) break; if (subRet == WAIT_ABANDONED) { ret = (ret - WAIT_OBJECT_0) + WAIT_ABANDONED; break; } // If we get alerted it just masks the real state of the current // handle, so retry the wait. if (subRet == WAIT_IO_COMPLETION) goto TryAgain; _ASSERTE(subRet == WAIT_TIMEOUT); ret++; } } } WaitCompleted: _ASSERTE((ret != WAIT_TIMEOUT) || (millis != INFINITE)); return ret; } DWORD Thread::DoAppropriateWaitWorker(AppropriateWaitFunc func, void *args, DWORD millis, WaitMode mode) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; BOOL alertable = (mode & WaitMode_Alertable)!=0; // Before going to pre-emptive mode the thread needs to be flagged as waiting for // the debugger. This used to be accomplished by the TS_Interruptible flag but that // doesn't work reliably, see DevDiv Bugs 699245. Some methods call in here already in // COOP mode so we set the bit before the transition. For the calls that are already // in pre-emptive mode those are still buggy. This is only a partial fix. BOOL isCoop = PreemptiveGCDisabled(); ThreadStateNCStackHolder tsNC(isCoop && alertable, TSNC_DebuggerSleepWaitJoin); GCX_PREEMP(); // // @TODO cwb: we don't know whether a thread has a message pump or // how to pump its messages, currently. // @TODO cwb: WinCE isn't going to support Thread.Interrupt() correctly until // we get alertable waits on that platform. DWORD ret; if(alertable) { DoAppropriateWaitWorkerAlertableHelper(mode); } DWORD option; if (alertable) { option = WAIT_ALERTABLE; #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT ApartmentState as = GetFinalApartment(); if ((AS_InMTA != as) && !GetDomain()->MustForceTrivialWaitOperations()) { option |= WAIT_MSGPUMP; } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT } else { option = 0; } ThreadStateHolder tsh(alertable, TS_Interruptible | TS_Interrupted); ULONGLONG dwStart = 0; ULONGLONG dwEnd; retry: if (millis != INFINITE) { dwStart = CLRGetTickCount64(); } ret = func(args, millis, option); if (ret == WAIT_IO_COMPLETION) { _ASSERTE (alertable); if ((m_State & TS_Interrupted)) { HandleThreadInterrupt(mode & WaitMode_ADUnload); } if (millis != INFINITE) { dwEnd = CLRGetTickCount64(); if (dwEnd >= dwStart + millis) { ret = WAIT_TIMEOUT; goto WaitCompleted; } else { millis -= (DWORD)(dwEnd - dwStart); } } goto retry; } WaitCompleted: _ASSERTE(ret == WAIT_OBJECT_0 || ret == WAIT_ABANDONED || ret == WAIT_TIMEOUT || ret == WAIT_FAILED); _ASSERTE((ret != WAIT_TIMEOUT) || (millis != INFINITE)); return ret; } //-------------------------------------------------------------------- // Only one style of wait for DoSignalAndWait since we don't support this on STA Threads //-------------------------------------------------------------------- DWORD Thread::DoSignalAndWait(HANDLE *handles, DWORD millis, BOOL alertable, PendingSync *syncState) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; _ASSERTE(alertable || syncState == 0); struct Param { Thread *pThis; HANDLE *handles; DWORD millis; BOOL alertable; DWORD dwRet; } param; param.pThis = this; param.handles = handles; param.millis = millis; param.alertable = alertable; param.dwRet = (DWORD) -1; EE_TRY_FOR_FINALLY(Param *, pParam, ¶m) { pParam->dwRet = pParam->pThis->DoSignalAndWaitWorker(pParam->handles, pParam->millis, pParam->alertable); } EE_FINALLY { if (syncState) { if (!GOT_EXCEPTION() && WAIT_OBJECT_0 == param.dwRet) { // This thread has been removed from syncblk waiting list by the signalling thread syncState->Restore(FALSE); } else syncState->Restore(TRUE); } _ASSERTE (WAIT_IO_COMPLETION != param.dwRet); } EE_END_FINALLY; return(param.dwRet); } DWORD Thread::DoSignalAndWaitWorker(HANDLE* pHandles, DWORD millis,BOOL alertable) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; DWORD ret = 0; GCX_PREEMP(); if(alertable) { DoAppropriateWaitWorkerAlertableHelper(WaitMode_None); } StateHolder OSAlertableWait(alertable); ThreadStateHolder tsh(alertable, TS_Interruptible | TS_Interrupted); ULONGLONG dwStart = 0, dwEnd; if (INFINITE != millis) { dwStart = CLRGetTickCount64(); } ret = SignalObjectAndWait(pHandles[0], pHandles[1], millis, alertable); retry: if (WAIT_IO_COMPLETION == ret) { _ASSERTE (alertable); // We could be woken by some spurious APC or an EE APC queued to // interrupt us. In the latter case the TS_Interrupted bit will be set // in the thread state bits. Otherwise we just go back to sleep again. if ((m_State & TS_Interrupted)) { HandleThreadInterrupt(FALSE); } if (INFINITE != millis) { dwEnd = CLRGetTickCount64(); if (dwStart + millis <= dwEnd) { ret = WAIT_TIMEOUT; goto WaitCompleted; } else { millis -= (DWORD)(dwEnd - dwStart); } dwStart = CLRGetTickCount64(); } //Retry case we don't want to signal again so only do the wait... ret = WaitForSingleObjectEx(pHandles[1],millis,TRUE); goto retry; } if (WAIT_FAILED == ret) { DWORD errorCode = ::GetLastError(); //If the handle to signal is a mutex and // the calling thread is not the owner, errorCode is ERROR_NOT_OWNER switch(errorCode) { case ERROR_INVALID_HANDLE: case ERROR_NOT_OWNER: case ERROR_ACCESS_DENIED: COMPlusThrowWin32(); break; case ERROR_TOO_MANY_POSTS: ret = ERROR_TOO_MANY_POSTS; break; default: CONSISTENCY_CHECK_MSGF(0, ("This errorCode is not understood '(%d)''\n", errorCode)); COMPlusThrowWin32(); break; } } WaitCompleted: //Check that the return state is valid _ASSERTE(WAIT_OBJECT_0 == ret || WAIT_ABANDONED == ret || WAIT_TIMEOUT == ret || WAIT_FAILED == ret || ERROR_TOO_MANY_POSTS == ret); //Wrong to time out if the wait was infinite _ASSERTE((WAIT_TIMEOUT != ret) || (INFINITE != millis)); return ret; } DWORD Thread::DoSyncContextWait(OBJECTREF *pSyncCtxObj, int countHandles, HANDLE *handles, BOOL waitAll, DWORD millis) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; PRECONDITION(CheckPointer(handles)); PRECONDITION(IsProtectedByGCFrame (pSyncCtxObj)); } CONTRACTL_END; MethodDescCallSite invokeWaitMethodHelper(METHOD__SYNCHRONIZATION_CONTEXT__INVOKE_WAIT_METHOD_HELPER); BASEARRAYREF handleArrayObj = (BASEARRAYREF)AllocatePrimitiveArray(ELEMENT_TYPE_I, countHandles); memcpyNoGCRefs(handleArrayObj->GetDataPtr(), handles, countHandles * sizeof(HANDLE)); ARG_SLOT args[6] = { ObjToArgSlot(*pSyncCtxObj), ObjToArgSlot(handleArrayObj), BoolToArgSlot(waitAll), (ARG_SLOT)millis, }; // Needed by TriggerGCForMDAInternal to avoid infinite recursion ThreadStateNCStackHolder holder(TRUE, TSNC_InsideSyncContextWait); return invokeWaitMethodHelper.Call_RetI4(args); } // Called out of SyncBlock::Wait() to block this thread until the Notify occurs. BOOL Thread::Block(INT32 timeOut, PendingSync *syncState) { WRAPPER_NO_CONTRACT; _ASSERTE(this == GetThread()); // Before calling Block, the SyncBlock queued us onto it's list of waiting threads. // However, before calling Block the SyncBlock temporarily left the synchronized // region. This allowed threads to enter the region and call Notify, in which // case we may have been signalled before we entered the Wait. So we aren't in the // m_WaitSB list any longer. Not a problem: the following Wait will return // immediately. But it means we cannot enforce the following assertion: // _ASSERTE(m_WaitSB != NULL); return (Wait(syncState->m_WaitEventLink->m_Next->m_EventWait, timeOut, syncState) != WAIT_OBJECT_0); } // Return whether or not a timeout occurred. TRUE=>we waited successfully DWORD Thread::Wait(HANDLE *objs, int cntObjs, INT32 timeOut, PendingSync *syncInfo) { WRAPPER_NO_CONTRACT; DWORD dwResult; DWORD dwTimeOut32; _ASSERTE(timeOut >= 0 || timeOut == INFINITE_TIMEOUT); dwTimeOut32 = (timeOut == INFINITE_TIMEOUT ? INFINITE : (DWORD) timeOut); dwResult = DoAppropriateWait(cntObjs, objs, FALSE /*=waitAll*/, dwTimeOut32, WaitMode_Alertable /*alertable*/, syncInfo); // Either we succeeded in the wait, or we timed out _ASSERTE((dwResult >= WAIT_OBJECT_0 && dwResult < (DWORD)(WAIT_OBJECT_0 + cntObjs)) || (dwResult == WAIT_TIMEOUT)); return dwResult; } // Return whether or not a timeout occurred. TRUE=>we waited successfully DWORD Thread::Wait(CLREvent *pEvent, INT32 timeOut, PendingSync *syncInfo) { WRAPPER_NO_CONTRACT; DWORD dwResult; DWORD dwTimeOut32; _ASSERTE(timeOut >= 0 || timeOut == INFINITE_TIMEOUT); dwTimeOut32 = (timeOut == INFINITE_TIMEOUT ? INFINITE : (DWORD) timeOut); dwResult = pEvent->Wait(dwTimeOut32, TRUE /*alertable*/, syncInfo); // Either we succeeded in the wait, or we timed out _ASSERTE((dwResult == WAIT_OBJECT_0) || (dwResult == WAIT_TIMEOUT)); return dwResult; } void Thread::Wake(SyncBlock *psb) { WRAPPER_NO_CONTRACT; CLREvent* hEvent = NULL; WaitEventLink *walk = &m_WaitEventLink; while (walk->m_Next) { if (walk->m_Next->m_WaitSB == psb) { hEvent = walk->m_Next->m_EventWait; // We are guaranteed that only one thread can change walk->m_Next->m_WaitSB // since the thread is helding the syncblock. walk->m_Next->m_WaitSB = (SyncBlock*)((DWORD_PTR)walk->m_Next->m_WaitSB | 1); break; } #ifdef _DEBUG else if ((SyncBlock*)((DWORD_PTR)walk->m_Next & ~1) == psb) { _ASSERTE (!"Can not wake a thread on the same SyncBlock more than once"); } #endif } PREFIX_ASSUME (hEvent != NULL); hEvent->Set(); } #define WAIT_INTERRUPT_THREADABORT 0x1 #define WAIT_INTERRUPT_INTERRUPT 0x2 #define WAIT_INTERRUPT_OTHEREXCEPTION 0x4 // When we restore DWORD EnterMonitorForRestore(SyncBlock *pSB) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; DWORD state = 0; EX_TRY { pSB->EnterMonitor(); } EX_CATCH { // Assume it is a normal exception unless proven. state = WAIT_INTERRUPT_OTHEREXCEPTION; Thread *pThread = GetThread(); if (pThread->IsAbortInitiated()) { state = WAIT_INTERRUPT_THREADABORT; } else if (__pException != NULL) { if (__pException->GetHR() == COR_E_THREADINTERRUPTED) { state = WAIT_INTERRUPT_INTERRUPT; } } } EX_END_CATCH(SwallowAllExceptions); return state; } // This is the service that backs us out of a wait that we interrupted. We must // re-enter the monitor to the same extent the SyncBlock would, if we returned // through it (instead of throwing through it). And we need to cancel the wait, // if it didn't get notified away while we are processing the interrupt. void PendingSync::Restore(BOOL bRemoveFromSB) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE(m_EnterCount); Thread *pCurThread = GetThread(); _ASSERTE (pCurThread == m_OwnerThread); WaitEventLink *pRealWaitEventLink = m_WaitEventLink->m_Next; pRealWaitEventLink->m_RefCount --; if (pRealWaitEventLink->m_RefCount == 0) { if (bRemoveFromSB) { ThreadQueue::RemoveThread(pCurThread, pRealWaitEventLink->m_WaitSB); } if (pRealWaitEventLink->m_EventWait != &pCurThread->m_EventWait) { // Put the event back to the pool. StoreEventToEventStore(pRealWaitEventLink->m_EventWait); } // Remove from the link. m_WaitEventLink->m_Next = m_WaitEventLink->m_Next->m_Next; } // Someone up the stack is responsible for keeping the syncblock alive by protecting // the object that owns it. But this relies on assertions that EnterMonitor is only // called in cooperative mode. Even though we are safe in preemptive, do the // switch. GCX_COOP_THREAD_EXISTS(pCurThread); // We need to make sure that EnterMonitor succeeds. We may have code like // lock (a) // { // a.Wait // } // We need to make sure that the finally from lock is excuted with the lock owned. DWORD state = 0; SyncBlock *psb = (SyncBlock*)((DWORD_PTR)pRealWaitEventLink->m_WaitSB & ~1); for (LONG i=0; i < m_EnterCount;) { if ((state & (WAIT_INTERRUPT_THREADABORT | WAIT_INTERRUPT_INTERRUPT)) != 0) { // If the thread has been interrupted by Thread.Interrupt or Thread.Abort, // disable the check at the beginning of DoAppropriateWait pCurThread->SetThreadStateNC(Thread::TSNC_InRestoringSyncBlock); } DWORD result = EnterMonitorForRestore(psb); if (result == 0) { i++; } else { // We block the thread until the thread acquires the lock. // This is to make sure that when catch/finally is executed, the thread has the lock. // We do not want thread to run its catch/finally if the lock is not taken. state |= result; // If the thread is being rudely aborted, and the thread has // no Cer on stack, we will not run managed code to release the // lock, so we can terminate the loop. if (pCurThread->IsRudeAbortInitiated() && !pCurThread->IsExecutingWithinCer()) { break; } } } pCurThread->ResetThreadStateNC(Thread::TSNC_InRestoringSyncBlock); if ((state & WAIT_INTERRUPT_THREADABORT) != 0) { pCurThread->HandleThreadAbort(); } else if ((state & WAIT_INTERRUPT_INTERRUPT) != 0) { COMPlusThrow(kThreadInterruptedException); } } // This is the callback from the OS, when we queue an APC to interrupt a waiting thread. // The callback occurs on the thread we wish to interrupt. It is a STATIC method. void WINAPI Thread::UserInterruptAPC(ULONG_PTR data) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; _ASSERTE(data == APC_Code); Thread *pCurThread = GetThread(); if (pCurThread) { // We should only take action if an interrupt is currently being // requested (our synchronization does not guarantee that we won't fire // spuriously). It's safe to check the m_UserInterrupt field and then // set TS_Interrupted in a non-atomic fashion because m_UserInterrupt is // only cleared in this thread's context (though it may be set from any // context). if (pCurThread->IsUserInterrupted()) { // Set bit to indicate this routine was called (as opposed to other // generic APCs). FastInterlockOr((ULONG *) &pCurThread->m_State, TS_Interrupted); } } } // This is the workhorse for Thread.Interrupt(). void Thread::UserInterrupt(ThreadInterruptMode mode) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; FastInterlockOr((DWORD*)&m_UserInterrupt, mode); if (HasValidThreadHandle() && HasThreadState (TS_Interruptible)) { Alert(); } } // Implementation of Thread.Sleep(). void Thread::UserSleep(INT32 time) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; INCONTRACT(_ASSERTE(!GetThread()->GCNoTrigger())); DWORD res; // Before going to pre-emptive mode the thread needs to be flagged as waiting for // the debugger. This used to be accomplished by the TS_Interruptible flag but that // doesn't work reliably, see DevDiv Bugs 699245. ThreadStateNCStackHolder tsNC(TRUE, TSNC_DebuggerSleepWaitJoin); GCX_PREEMP(); // A word about ordering for Interrupt. If someone tries to interrupt a thread // that's in the interruptible state, we queue an APC. But if they try to interrupt // a thread that's not in the interruptible state, we just record that fact. So // we have to set TS_Interruptible before we test to see whether someone wants to // interrupt us or else we have a race condition that causes us to skip the APC. FastInterlockOr((ULONG *) &m_State, TS_Interruptible); // If someone has interrupted us, we should not enter the wait. if (IsUserInterrupted()) { HandleThreadInterrupt(FALSE); } ThreadStateHolder tsh(TRUE, TS_Interruptible | TS_Interrupted); FastInterlockAnd((ULONG *) &m_State, ~TS_Interrupted); DWORD dwTime = (DWORD)time; retry: ULONGLONG start = CLRGetTickCount64(); res = ClrSleepEx (dwTime, TRUE); if (res == WAIT_IO_COMPLETION) { // We could be woken by some spurious APC or an EE APC queued to // interrupt us. In the latter case the TS_Interrupted bit will be set // in the thread state bits. Otherwise we just go back to sleep again. if ((m_State & TS_Interrupted)) { HandleThreadInterrupt(FALSE); } if (dwTime == INFINITE) { goto retry; } else { ULONGLONG actDuration = CLRGetTickCount64() - start; if (dwTime > actDuration) { dwTime -= (DWORD)actDuration; goto retry; } else { res = WAIT_TIMEOUT; } } } _ASSERTE(res == WAIT_TIMEOUT || res == WAIT_OBJECT_0); } // Correspondence between an EE Thread and an exposed System.Thread: OBJECTREF Thread::GetExposedObject() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; TRIGGERSGC(); Thread *pCurThread = GetThread(); _ASSERTE (!(pCurThread == NULL || IsAtProcessExit())); _ASSERTE(pCurThread->PreemptiveGCDisabled()); if (ObjectFromHandle(m_ExposedObject) == NULL) { // Allocate the exposed thread object. THREADBASEREF attempt = (THREADBASEREF) AllocateObject(g_pThreadClass); GCPROTECT_BEGIN(attempt); // The exposed object keeps us alive until it is GC'ed. This // doesn't mean the physical thread continues to run, of course. // We have to set this outside of the ThreadStore lock, because this might trigger a GC. attempt->SetInternal(this); BOOL fNeedThreadStore = (! ThreadStore::HoldingThreadStore(pCurThread)); // Take a lock to make sure that only one thread creates the object. ThreadStoreLockHolder tsHolder(fNeedThreadStore); // Check to see if another thread has not already created the exposed object. if (ObjectFromHandle(m_ExposedObject) == NULL) { // Keep a weak reference to the exposed object. StoreObjectInHandle(m_ExposedObject, (OBJECTREF) attempt); ObjectInHandleHolder exposedHolder(m_ExposedObject); // Increase the external ref count. We can't call IncExternalCount because we // already hold the thread lock and IncExternalCount won't be able to take it. ULONG retVal = FastInterlockIncrement ((LONG*)&m_ExternalRefCount); // Check to see if we need to store a strong pointer to the object. if (retVal > 1) StoreObjectInHandle(m_StrongHndToExposedObject, (OBJECTREF) attempt); ObjectInHandleHolder strongHolder(m_StrongHndToExposedObject); attempt->SetManagedThreadId(GetThreadId()); // Note that we are NOT calling the constructor on the Thread. That's // because this is an internal create where we don't want a Start // address. And we don't want to expose such a constructor for our // customers to accidentally call. The following is in lieu of a true // constructor: attempt->InitExisting(); exposedHolder.SuppressRelease(); strongHolder.SuppressRelease(); } else { attempt->ClearInternal(); } GCPROTECT_END(); } return ObjectFromHandle(m_ExposedObject); } // We only set non NULL exposed objects for unstarted threads that haven't exited // their constructor yet. So there are no race conditions. void Thread::SetExposedObject(OBJECTREF exposed) { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; if (exposed != NULL) { _ASSERTE (GetThread() != this); _ASSERTE(IsUnstarted()); _ASSERTE(ObjectFromHandle(m_ExposedObject) == NULL); // The exposed object keeps us alive until it is GC'ed. This doesn't mean the // physical thread continues to run, of course. StoreObjectInHandle(m_ExposedObject, exposed); // This makes sure the contexts on the backing thread // and the managed thread start off in sync with each other. // BEWARE: the IncExternalCount call below may cause GC to happen. // IncExternalCount will store exposed in m_StrongHndToExposedObject which is in default domain. // If the creating thread is killed before the target thread is killed in Thread.Start, Thread object // will be kept alive forever. // Instead, IncExternalCount should be called after the target thread has been started in Thread.Start. // IncExternalCount(); } else { // Simply set both of the handles to NULL. The GC of the old exposed thread // object will take care of decrementing the external ref count. StoreObjectInHandle(m_ExposedObject, NULL); StoreObjectInHandle(m_StrongHndToExposedObject, NULL); } } void Thread::SetLastThrownObject(OBJECTREF throwable, BOOL isUnhandled) { CONTRACTL { if ((throwable == NULL) || CLRException::IsPreallocatedExceptionObject(throwable)) NOTHROW; else THROWS; // From CreateHandle GC_NOTRIGGER; if (throwable == NULL) MODE_ANY; else MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; STRESS_LOG_COND1(LF_EH, LL_INFO100, OBJECTREFToObject(throwable) != NULL, "in Thread::SetLastThrownObject: obj = %p\n", OBJECTREFToObject(throwable)); // you can't have a NULL unhandled exception _ASSERTE(!(throwable == NULL && isUnhandled)); if (m_LastThrownObjectHandle != NULL) { // We'll somtimes use a handle for a preallocated exception object. We should never, ever destroy one of // these handles... they'll be destroyed when the Runtime shuts down. if (!CLRException::IsPreallocatedExceptionHandle(m_LastThrownObjectHandle)) { DestroyHandle(m_LastThrownObjectHandle); } m_LastThrownObjectHandle = NULL; // Make sure to set this to NULL here just in case we throw trying to make // a new handle below. } if (throwable != NULL) { _ASSERTE(this == GetThread()); // Non-compliant exceptions are always wrapped. // The use of the ExceptionNative:: helper here (rather than the global ::IsException helper) // is hokey, but we need a GC_NOTRIGGER version and it's only for an ASSERT. _ASSERTE(IsException(throwable->GetMethodTable())); // If we're tracking one of the preallocated exception objects, then just use the global handle that // matches it rather than creating a new one. if (CLRException::IsPreallocatedExceptionObject(throwable)) { m_LastThrownObjectHandle = CLRException::GetPreallocatedHandleForObject(throwable); } else { BEGIN_SO_INTOLERANT_CODE(GetThread()); { m_LastThrownObjectHandle = GetDomain()->CreateHandle(throwable); } END_SO_INTOLERANT_CODE; } _ASSERTE(m_LastThrownObjectHandle != NULL); m_ltoIsUnhandled = isUnhandled; } else { m_ltoIsUnhandled = FALSE; } } void Thread::SetSOForLastThrownObject() { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; SO_TOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; // If we are saving stack overflow exception, we can just null out the current handle. // The current domain is going to be unloaded or the process is going to be killed, so // we will not leak a handle. m_LastThrownObjectHandle = CLRException::GetPreallocatedStackOverflowExceptionHandle(); } // // This is a nice wrapper for SetLastThrownObject which catches any exceptions caused by not being able to create // the handle for the throwable, and setting the last thrown object to the preallocated out of memory exception // instead. // OBJECTREF Thread::SafeSetLastThrownObject(OBJECTREF throwable) { CONTRACTL { NOTHROW; GC_NOTRIGGER; if (throwable == NULL) MODE_ANY; else MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; // We return the original throwable if nothing goes wrong. OBJECTREF ret = throwable; EX_TRY { // Try to set the throwable. SetLastThrownObject(throwable); } EX_CATCH { // If it didn't work, then set the last thrown object to the preallocated OOM exception, and return that // object instead of the original throwable. ret = CLRException::GetPreallocatedOutOfMemoryException(); SetLastThrownObject(ret); } EX_END_CATCH(SwallowAllExceptions); return ret; } // // This is a nice wrapper for SetThrowable and SetLastThrownObject, which catches any exceptions caused by not // being able to create the handle for the throwable, and sets the throwable to the preallocated out of memory // exception instead. It also updates the last thrown object, which is always updated when the throwable is // updated. // OBJECTREF Thread::SafeSetThrowables(OBJECTREF throwable DEBUG_ARG(ThreadExceptionState::SetThrowableErrorChecking stecFlags), BOOL isUnhandled) { CONTRACTL { NOTHROW; GC_NOTRIGGER; if (throwable == NULL) MODE_ANY; else MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; // We return the original throwable if nothing goes wrong. OBJECTREF ret = throwable; EX_TRY { // Try to set the throwable. SetThrowable(throwable DEBUG_ARG(stecFlags)); // Now, if the last thrown object is different, go ahead and update it. This makes sure that we re-throw // the right object when we rethrow. if (LastThrownObject() != throwable) { SetLastThrownObject(throwable); } if (isUnhandled) { MarkLastThrownObjectUnhandled(); } } EX_CATCH { // If either set didn't work, then set both throwables to the preallocated OOM exception, and return that // object instead of the original throwable. ret = CLRException::GetPreallocatedOutOfMemoryException(); // Neither of these will throw because we're setting with a preallocated exception. SetThrowable(ret DEBUG_ARG(stecFlags)); SetLastThrownObject(ret, isUnhandled); } EX_END_CATCH(SwallowAllExceptions); return ret; } // This method will sync the managed exception state to be in sync with the topmost active exception // for a given thread void Thread::SyncManagedExceptionState(bool fIsDebuggerThread) { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_ANY; } CONTRACTL_END; { GCX_COOP(); // Syncup the LastThrownObject on the managed thread SafeUpdateLastThrownObject(); } #ifdef FEATURE_CORRUPTING_EXCEPTIONS // Since the catch clause has successfully executed and we are exiting it, reset the corruption severity // in the ThreadExceptionState for the last active exception. This will ensure that when the next exception // gets thrown/raised, EH tracker wont pick up an invalid value. if (!fIsDebuggerThread) { CEHelper::ResetLastActiveCorruptionSeverityPostCatchHandler(this); } #endif // FEATURE_CORRUPTING_EXCEPTIONS } void Thread::SetLastThrownObjectHandle(OBJECTHANDLE h) { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; if (m_LastThrownObjectHandle != NULL && !CLRException::IsPreallocatedExceptionHandle(m_LastThrownObjectHandle)) { DestroyHandle(m_LastThrownObjectHandle); } m_LastThrownObjectHandle = h; } // // Create a duplicate handle of the current throwable and set the last thrown object to that. This ensures that the // last thrown object and the current throwable have handles that are in the same app domain. // void Thread::SafeUpdateLastThrownObject(void) { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; SO_INTOLERANT; } CONTRACTL_END; OBJECTHANDLE hThrowable = GetThrowableAsHandle(); if (hThrowable != NULL) { EX_TRY { IGCHandleManager *pHandleTable = GCHandleUtilities::GetGCHandleManager(); // Creating a duplicate handle here ensures that the AD of the last thrown object // matches the domain of the current throwable. OBJECTHANDLE duplicateHandle = pHandleTable->CreateDuplicateHandle(hThrowable); SetLastThrownObjectHandle(duplicateHandle); } EX_CATCH { // If we can't create a duplicate handle, we set both throwables to the preallocated OOM exception. SafeSetThrowables(CLRException::GetPreallocatedOutOfMemoryException()); } EX_END_CATCH(SwallowAllExceptions); } } // Background threads must be counted, because the EE should shut down when the // last non-background thread terminates. But we only count running ones. void Thread::SetBackground(BOOL isBack, BOOL bRequiresTSL) { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; // booleanize IsBackground() which just returns bits if (isBack == !!IsBackground()) return; LOG((LF_SYNC, INFO3, "SetBackground obtain lock\n")); ThreadStoreLockHolder TSLockHolder(FALSE); if (bRequiresTSL) { TSLockHolder.Acquire(); } if (IsDead()) { // This can only happen in a race condition, where the correct thing to do // is ignore it. If it happens without the race condition, we throw an // exception. } else if (isBack) { if (!IsBackground()) { FastInterlockOr((ULONG *) &m_State, TS_Background); // unstarted threads don't contribute to the background count if (!IsUnstarted()) ThreadStore::s_pThreadStore->m_BackgroundThreadCount++; // If we put the main thread into a wait, until only background threads exist, // then we make that // main thread a background thread. This cleanly handles the case where it // may or may not be one as it enters the wait. // One of the components of OtherThreadsComplete() has changed, so check whether // we should now exit the EE. ThreadStore::CheckForEEShutdown(); } } else { if (IsBackground()) { FastInterlockAnd((ULONG *) &m_State, ~TS_Background); // unstarted threads don't contribute to the background count if (!IsUnstarted()) ThreadStore::s_pThreadStore->m_BackgroundThreadCount--; _ASSERTE(ThreadStore::s_pThreadStore->m_BackgroundThreadCount >= 0); _ASSERTE(ThreadStore::s_pThreadStore->m_BackgroundThreadCount <= ThreadStore::s_pThreadStore->m_ThreadCount); } } if (bRequiresTSL) { TSLockHolder.Release(); } } #ifdef FEATURE_COMINTEROP class ApartmentSpyImpl : public IUnknownCommon { public: HRESULT STDMETHODCALLTYPE PreInitialize(DWORD dwCoInit, DWORD dwCurThreadAptRefs) { LIMITED_METHOD_CONTRACT; return S_OK; } HRESULT STDMETHODCALLTYPE PostInitialize(HRESULT hrCoInit, DWORD dwCoInit, DWORD dwNewThreadAptRefs) { LIMITED_METHOD_CONTRACT; return hrCoInit; // this HRESULT will be returned from CoInitialize(Ex) } HRESULT STDMETHODCALLTYPE PreUninitialize(DWORD dwCurThreadAptRefs) { // Don't assume that Thread exists and do not create it. STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_MODE_PREEMPTIVE; HRESULT hr = S_OK; if (dwCurThreadAptRefs == 1 && !g_fEEShutDown) { // This is the last CoUninitialize on this thread and the CLR is still running. If it's an STA // we take the opportunity to perform COM/WinRT cleanup now, when the apartment is still alive. Thread *pThread = GetThreadNULLOk(); if (pThread != NULL) { BEGIN_EXTERNAL_ENTRYPOINT(&hr) { if (pThread->GetFinalApartment() == Thread::AS_InSTA) { // This will release RCWs and purge the WinRT factory cache on all AppDomains. It // will also synchronize with the finalizer thread which ensures that the RCWs // that were already in the global RCW cleanup list will be cleaned up as well. // ReleaseRCWsInCachesNoThrow(GetCurrentCtxCookie()); } } END_EXTERNAL_ENTRYPOINT; } } return hr; } HRESULT STDMETHODCALLTYPE PostUninitialize(DWORD dwNewThreadAptRefs) { LIMITED_METHOD_CONTRACT; return S_OK; } }; #endif // FEATURE_COMINTEROP // When the thread starts running, make sure it is running in the correct apartment // and context. BOOL Thread::PrepareApartmentAndContext() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; m_OSThreadId = ::GetCurrentThreadId(); #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT // Be very careful in here because we haven't set up e.g. TLS yet. if (m_State & (TS_InSTA | TS_InMTA)) { // Make sure TS_InSTA and TS_InMTA aren't both set. _ASSERTE(!((m_State & TS_InSTA) && (m_State & TS_InMTA))); // Determine the apartment state to set based on the requested state. ApartmentState aState = m_State & TS_InSTA ? AS_InSTA : AS_InMTA; // Clear the requested apartment state from the thread. This is requested since // the thread might actually be a fiber that has already been initialized to // a different apartment state than the requested one. If we didn't clear // the requested apartment state, then we could end up with both TS_InSTA and // TS_InMTA set at the same time. FastInterlockAnd ((ULONG *) &m_State, ~TS_InSTA & ~TS_InMTA); // Attempt to set the requested apartment state. SetApartment(aState, FALSE); } // In the case where we own the thread and we have switched it to a different // starting context, it is the responsibility of the caller (KickOffThread()) // to notice that the context changed, and to adjust the delegate that it will // dispatch on, as appropriate. #endif //FEATURE_COMINTEROP_APARTMENT_SUPPORT #ifdef FEATURE_COMINTEROP // Our IInitializeSpy will be registered in AppX always, in classic processes // only if the internal config switch is on. if (AppX::IsAppXProcess() || g_pConfig->EnableRCWCleanupOnSTAShutdown()) { NewHolder pSpyImpl = new ApartmentSpyImpl(); IfFailThrow(CoRegisterInitializeSpy(pSpyImpl, &m_uliInitializeSpyCookie)); pSpyImpl.SuppressRelease(); m_fInitializeSpyRegistered = true; } #endif // FEATURE_COMINTEROP return TRUE; } #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT // TS_InSTA (0x00004000) -> AS_InSTA (0) // TS_InMTA (0x00008000) -> AS_InMTA (1) #define TS_TO_AS(ts) \ (Thread::ApartmentState)((((DWORD)ts) >> 14) - 1) \ // Retrieve the apartment state of the current thread. There are three possible // states: thread hosts an STA, thread is part of the MTA or thread state is // undecided. The last state may indicate that the apartment has not been set at // all (nobody has called CoInitializeEx) or that the EE does not know the // current state (EE has not called CoInitializeEx). Thread::ApartmentState Thread::GetApartment() { CONTRACTL { NOTHROW; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; ApartmentState as = AS_Unknown; ThreadState maskedTs = (ThreadState)(((DWORD)m_State) & (TS_InSTA|TS_InMTA)); if (maskedTs) { _ASSERTE((maskedTs == TS_InSTA) || (maskedTs == TS_InMTA)); static_assert_no_msg(TS_TO_AS(TS_InSTA) == AS_InSTA); static_assert_no_msg(TS_TO_AS(TS_InMTA) == AS_InMTA); as = TS_TO_AS(maskedTs); } if ( #ifdef MDA_SUPPORTED (NULL == MDA_GET_ASSISTANT(InvalidApartmentStateChange)) && #endif (as != AS_Unknown)) { return as; } return GetApartmentRare(as); } Thread::ApartmentState Thread::GetApartmentRare(Thread::ApartmentState as) { CONTRACTL { NOTHROW; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; if (this == GetThread()) { THDTYPE type; HRESULT hr = S_OK; #ifdef MDA_SUPPORTED MdaInvalidApartmentStateChange* pProbe = MDA_GET_ASSISTANT(InvalidApartmentStateChange); if (pProbe) { // Without notifications from OLE32, we cannot know when the apartment state of a // thread changes. But we have cached this fact and depend on it for all our // blocking and COM Interop behavior to work correctly. Using the CDH, log that it // is not changing underneath us, on those platforms where it is relatively cheap for // us to do so. if (as != AS_Unknown) { hr = GetCurrentThreadTypeNT5(&type); if (hr == S_OK) { if (type == THDTYPE_PROCESSMESSAGES && as == AS_InMTA) { pProbe->ReportViolation(this, as, FALSE); } else if (type == THDTYPE_BLOCKMESSAGES && as == AS_InSTA) { pProbe->ReportViolation(this, as, FALSE); } } } } #endif if (as == AS_Unknown) { hr = GetCurrentThreadTypeNT5(&type); if (hr == S_OK) { as = (type == THDTYPE_PROCESSMESSAGES) ? AS_InSTA : AS_InMTA; // If we get back THDTYPE_PROCESSMESSAGES, we are guaranteed to // be an STA thread. If not, we are an MTA thread, however // we can't know if the thread has been explicitly set to MTA // (via a call to CoInitializeEx) or if it has been implicitly // made MTA (if it hasn't been CoInitializeEx'd but CoInitialize // has already been called on some other thread in the process. if (as == AS_InSTA) FastInterlockOr((ULONG *) &m_State, AS_InSTA); } } } return as; } // Retrieve the explicit apartment state of the current thread. There are three possible // states: thread hosts an STA, thread is part of the MTA or thread state is // undecided. The last state may indicate that the apartment has not been set at // all (nobody has called CoInitializeEx), the EE does not know the // current state (EE has not called CoInitializeEx), or the thread is implicitly in // the MTA. Thread::ApartmentState Thread::GetExplicitApartment() { CONTRACTL { NOTHROW; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; _ASSERTE(!((m_State & TS_InSTA) && (m_State & TS_InMTA))); // Initialize m_State by calling GetApartment. GetApartment(); ApartmentState as = (m_State & TS_InSTA) ? AS_InSTA : (m_State & TS_InMTA) ? AS_InMTA : AS_Unknown; return as; } Thread::ApartmentState Thread::GetFinalApartment() { CONTRACTL { NOTHROW; GC_TRIGGERS; MODE_ANY; SO_TOLERANT; } CONTRACTL_END; _ASSERTE(this == GetThread()); ApartmentState as = AS_Unknown; if (g_fEEShutDown) { // On shutdown, do not use cached value. Someone might have called // CoUninitialize. FastInterlockAnd ((ULONG *) &m_State, ~TS_InSTA & ~TS_InMTA); } as = GetApartment(); if (as == AS_Unknown) { // On Win2k and above, GetApartment will only return AS_Unknown if CoInitialize // hasn't been called in the process. In that case we can simply assume MTA. However we // cannot cache this value in the Thread because if a CoInitialize does occur, then the // thread state might change. as = AS_InMTA; } return as; } // when we get apartment tear-down notification, // we want reset the apartment state we cache on the thread VOID Thread::ResetApartment() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // reset the TS_InSTA bit and TS_InMTA bit ThreadState t_State = (ThreadState)(~(TS_InSTA | TS_InMTA)); FastInterlockAnd((ULONG *) &m_State, t_State); } // Attempt to set current thread's apartment state. The actual apartment state // achieved is returned and may differ from the input state if someone managed // to call CoInitializeEx on this thread first (note that calls to SetApartment // made before the thread has started are guaranteed to succeed). // The fFireMDAOnMismatch indicates if we should fire the apartment state probe // on an apartment state mismatch. Thread::ApartmentState Thread::SetApartment(ApartmentState state, BOOL fFireMDAOnMismatch) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_ANY; INJECT_FAULT(COMPlusThrowOM();); } CONTRACTL_END; // Reset any bits that request for CoInitialize ResetRequiresCoInitialize(); // Setting the state to AS_Unknown indicates we should CoUninitialize // the thread. if (state == AS_Unknown) { BOOL needUninitialize = (m_State & TS_CoInitialized) #ifdef FEATURE_COMINTEROP || IsWinRTInitialized() #endif // FEATURE_COMINTEROP ; if (needUninitialize) { GCX_PREEMP(); // If we haven't CoInitialized the thread, then we don't have anything to do. if (m_State & TS_CoInitialized) { // We should never be attempting to CoUninitialize another thread than // the currently running thread. _ASSERTE(m_OSThreadId == ::GetCurrentThreadId()); // CoUninitialize the thread and reset the STA/MTA/CoInitialized state bits. ::CoUninitialize(); ThreadState uninitialized = static_cast(TS_InSTA | TS_InMTA | TS_CoInitialized); FastInterlockAnd((ULONG *) &m_State, ~uninitialized); } #ifdef FEATURE_COMINTEROP if (IsWinRTInitialized()) { _ASSERTE(WinRTSupported()); BaseWinRTUninitialize(); ResetWinRTInitialized(); } #endif // FEATURE_COMINTEROP } return GetApartment(); } // Call GetApartment to initialize the current apartment state. // // Important note: For Win2k and above this can return AS_InMTA even if the current // thread has never been CoInitialized. Because of this we MUST NOT look at the // return value of GetApartment here. We can however look at the m_State flags // since these will only be set to TS_InMTA if we know for a fact the the // current thread has explicitly been made MTA (via a call to CoInitializeEx). GetApartment(); // If the current thread is STA, then it is impossible to change it to // MTA. if (m_State & TS_InSTA) { #ifdef MDA_SUPPORTED if (state == AS_InMTA && fFireMDAOnMismatch) { MDA_TRIGGER_ASSISTANT(InvalidApartmentStateChange, ReportViolation(this, state, TRUE)); } #endif return AS_InSTA; } // If the current thread is EXPLICITLY MTA, then it is impossible to change it to // STA. if (m_State & TS_InMTA) { #ifdef MDA_SUPPORTED if (state == AS_InSTA && fFireMDAOnMismatch) { MDA_TRIGGER_ASSISTANT(InvalidApartmentStateChange, ReportViolation(this, state, TRUE)); } #endif return AS_InMTA; } // If the thread isn't even started yet, we mark the state bits without // calling CoInitializeEx (since we're obviously not in the correct thread // context yet). We'll retry this call when the thread is started. // Don't use the TS_Unstarted state bit to check for this, it's cleared far // too late in the day for us. Instead check whether we're in the correct // thread context. if (m_OSThreadId != ::GetCurrentThreadId()) { FastInterlockOr((ULONG *) &m_State, (state == AS_InSTA) ? TS_InSTA : TS_InMTA); return state; } HRESULT hr; { GCX_PREEMP(); // Attempt to set apartment by calling CoInitializeEx. This may fail if // another caller (outside EE) beat us to it. // // Important note: When calling CoInitializeEx(COINIT_MULTITHREADED) on a // thread that has never been CoInitialized, the return value will always // be S_OK, even if another thread in the process has already been // CoInitialized to MTA. However if the current thread has already been // CoInitialized to MTA, then S_FALSE will be returned. hr = ::CoInitializeEx(NULL, (state == AS_InSTA) ? COINIT_APARTMENTTHREADED : COINIT_MULTITHREADED); } if (SUCCEEDED(hr)) { ThreadState t_State = (state == AS_InSTA) ? TS_InSTA : TS_InMTA; if (hr == S_OK) { // The thread has never been CoInitialized. t_State = (ThreadState)(t_State | TS_CoInitialized); } else { _ASSERTE(hr == S_FALSE); // If the thread has already been CoInitialized to the proper mode, then // we don't want to leave an outstanding CoInit so we CoUninit. { GCX_PREEMP(); ::CoUninitialize(); } } // We succeeded in setting the apartment state to the requested state. FastInterlockOr((ULONG *) &m_State, t_State); } else if (hr == RPC_E_CHANGED_MODE) { // We didn't manage to enforce the requested apartment state, but at least // we can work out what the state is now. No need to actually do the CoInit -- // obviously someone else already took care of that. FastInterlockOr((ULONG *) &m_State, ((state == AS_InSTA) ? TS_InMTA : TS_InSTA)); #ifdef MDA_SUPPORTED if (fFireMDAOnMismatch) { // Report via the customer debug helper that we failed to set the apartment type // to the specified type. MDA_TRIGGER_ASSISTANT(InvalidApartmentStateChange, ReportViolation(this, state, TRUE)); } #endif } else if (hr == E_OUTOFMEMORY) { COMPlusThrowOM(); } else { _ASSERTE(!"Unexpected HRESULT returned from CoInitializeEx!"); } #ifdef FEATURE_COMINTEROP // If WinRT is supported on this OS, also initialize it at the same time. Since WinRT sits on top of COM // we need to make sure that it is initialized in the same threading mode as we just started COM itself // with (or that we detected COM had already been started with). if (WinRTSupported() && !IsWinRTInitialized()) { GCX_PREEMP(); BOOL isSTA = m_State & TS_InSTA; _ASSERTE(isSTA || (m_State & TS_InMTA)); HRESULT hrWinRT = RoInitialize(isSTA ? RO_INIT_SINGLETHREADED : RO_INIT_MULTITHREADED); if (SUCCEEDED(hrWinRT)) { if (hrWinRT == S_OK) { SetThreadStateNC(TSNC_WinRTInitialized); } else { _ASSERTE(hrWinRT == S_FALSE); // If the thread has already been initialized, back it out. We may not // always be able to call RoUninitialize on shutdown so if there's // a way to avoid having to, we should take advantage of that. RoUninitialize(); } } else if (hrWinRT == E_OUTOFMEMORY) { COMPlusThrowOM(); } else { // We don't check for RPC_E_CHANGEDMODE, since we're using the mode that was read in by // initializing COM above. COM and WinRT need to always be in the same mode, so we should never // see that return code at this point. _ASSERTE(!"Unexpected HRESULT From RoInitialize"); } } // Since we've just called CoInitialize, COM has effectively been started up. // To ensure the CLR is aware of this, we need to call EnsureComStarted. EnsureComStarted(FALSE); #endif // FEATURE_COMINTEROP return GetApartment(); } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT //---------------------------------------------------------------------------- // // ThreadStore Implementation // //---------------------------------------------------------------------------- ThreadStore::ThreadStore() : m_Crst(CrstThreadStore, (CrstFlags) (CRST_UNSAFE_ANYMODE | CRST_DEBUGGER_THREAD)), m_ThreadCount(0), m_MaxThreadCount(0), m_UnstartedThreadCount(0), m_BackgroundThreadCount(0), m_PendingThreadCount(0), m_DeadThreadCount(0), m_DeadThreadCountForGCTrigger(0), m_TriggerGCForDeadThreads(false), m_GuidCreated(FALSE), m_HoldingThread(0) { CONTRACTL { THROWS; GC_NOTRIGGER; } CONTRACTL_END; m_TerminationEvent.CreateManualEvent(FALSE); _ASSERTE(m_TerminationEvent.IsValid()); } void ThreadStore::InitThreadStore() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; s_pThreadStore = new ThreadStore; g_pThinLockThreadIdDispenser = new IdDispenser(); ThreadSuspend::g_pGCSuspendEvent = new CLREvent(); ThreadSuspend::g_pGCSuspendEvent->CreateManualEvent(FALSE); s_pWaitForStackCrawlEvent = new CLREvent(); s_pWaitForStackCrawlEvent->CreateManualEvent(FALSE); s_DeadThreadCountThresholdForGCTrigger = static_cast(CLRConfig::GetConfigValue(CLRConfig::INTERNAL_Thread_DeadThreadCountThresholdForGCTrigger)); if (s_DeadThreadCountThresholdForGCTrigger < 0) { s_DeadThreadCountThresholdForGCTrigger = 0; } s_DeadThreadGCTriggerPeriodMilliseconds = CLRConfig::GetConfigValue(CLRConfig::INTERNAL_Thread_DeadThreadGCTriggerPeriodMilliseconds); s_DeadThreadGenerationCounts = nullptr; } // Enter and leave the critical section around the thread store. Clients should // use LockThreadStore and UnlockThreadStore because ThreadStore lock has // additional semantics well beyond a normal lock. DEBUG_NOINLINE void ThreadStore::Enter() { WRAPPER_NO_CONTRACT; ANNOTATION_SPECIAL_HOLDER_CALLER_NEEDS_DYNAMIC_CONTRACT; CHECK_ONE_STORE(); m_Crst.Enter(); // Threadstore needs special shutdown handling. if (g_fSuspendOnShutdown) { m_Crst.ReleaseAndBlockForShutdownIfNotSpecialThread(); } } DEBUG_NOINLINE void ThreadStore::Leave() { WRAPPER_NO_CONTRACT; ANNOTATION_SPECIAL_HOLDER_CALLER_NEEDS_DYNAMIC_CONTRACT; CHECK_ONE_STORE(); m_Crst.Leave(); } void ThreadStore::LockThreadStore() { WRAPPER_NO_CONTRACT; // The actual implementation is in ThreadSuspend class since it is coupled // with thread suspension logic ThreadSuspend::LockThreadStore(ThreadSuspend::SUSPEND_OTHER); } void ThreadStore::UnlockThreadStore() { WRAPPER_NO_CONTRACT; // The actual implementation is in ThreadSuspend class since it is coupled // with thread suspension logic ThreadSuspend::UnlockThreadStore(FALSE, ThreadSuspend::SUSPEND_OTHER); } // AddThread adds 'newThread' to m_ThreadList void ThreadStore::AddThread(Thread *newThread, BOOL bRequiresTSL) { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; LOG((LF_SYNC, INFO3, "AddThread obtain lock\n")); ThreadStoreLockHolder TSLockHolder(FALSE); if (bRequiresTSL) { TSLockHolder.Acquire(); } s_pThreadStore->m_ThreadList.InsertTail(newThread); s_pThreadStore->m_ThreadCount++; if (s_pThreadStore->m_MaxThreadCount < s_pThreadStore->m_ThreadCount) s_pThreadStore->m_MaxThreadCount = s_pThreadStore->m_ThreadCount; if (newThread->IsUnstarted()) s_pThreadStore->m_UnstartedThreadCount++; newThread->SetThreadStateNC(Thread::TSNC_ExistInThreadStore); _ASSERTE(!newThread->IsBackground()); _ASSERTE(!newThread->IsDead()); if (bRequiresTSL) { TSLockHolder.Release(); } } // this function is just desgined to avoid deadlocks during abnormal process termination, and should not be used for any other purpose BOOL ThreadStore::CanAcquireLock() { WRAPPER_NO_CONTRACT; { return (s_pThreadStore->m_Crst.m_criticalsection.LockCount == -1 || (size_t)s_pThreadStore->m_Crst.m_criticalsection.OwningThread == (size_t)GetCurrentThreadId()); } } // Whenever one of the components of OtherThreadsComplete() has changed in the // correct direction, see whether we can now shutdown the EE because only background // threads are running. void ThreadStore::CheckForEEShutdown() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; if (g_fWeControlLifetime && s_pThreadStore->OtherThreadsComplete()) { BOOL bRet; bRet = s_pThreadStore->m_TerminationEvent.Set(); _ASSERTE(bRet); } } BOOL ThreadStore::RemoveThread(Thread *target) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; BOOL found; Thread *ret; #if 0 // This assert is not valid when failing to create background GC thread. // Main GC thread holds the TS lock. _ASSERTE (ThreadStore::HoldingThreadStore()); #endif _ASSERTE(s_pThreadStore->m_Crst.GetEnterCount() > 0 || IsAtProcessExit()); _ASSERTE(s_pThreadStore->DbgFindThread(target)); ret = s_pThreadStore->m_ThreadList.FindAndRemove(target); _ASSERTE(ret && ret == target); found = (ret != NULL); if (found) { target->ResetThreadStateNC(Thread::TSNC_ExistInThreadStore); s_pThreadStore->m_ThreadCount--; if (target->IsDead()) { s_pThreadStore->m_DeadThreadCount--; s_pThreadStore->DecrementDeadThreadCountForGCTrigger(); } // Unstarted threads are not in the Background count: if (target->IsUnstarted()) s_pThreadStore->m_UnstartedThreadCount--; else if (target->IsBackground()) s_pThreadStore->m_BackgroundThreadCount--; FastInterlockExchangeAdd( &Thread::s_threadPoolCompletionCountOverflow, target->m_threadPoolCompletionCount); _ASSERTE(s_pThreadStore->m_ThreadCount >= 0); _ASSERTE(s_pThreadStore->m_BackgroundThreadCount >= 0); _ASSERTE(s_pThreadStore->m_ThreadCount >= s_pThreadStore->m_BackgroundThreadCount); _ASSERTE(s_pThreadStore->m_ThreadCount >= s_pThreadStore->m_UnstartedThreadCount); _ASSERTE(s_pThreadStore->m_ThreadCount >= s_pThreadStore->m_DeadThreadCount); // One of the components of OtherThreadsComplete() has changed, so check whether // we should now exit the EE. CheckForEEShutdown(); } return found; } // When a thread is created as unstarted. Later it may get started, in which case // someone calls Thread::HasStarted() on that physical thread. This completes // the Setup and calls here. void ThreadStore::TransferStartedThread(Thread *thread, BOOL bRequiresTSL) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE(GetThread() == thread); LOG((LF_SYNC, INFO3, "TransferUnstartedThread obtain lock\n")); ThreadStoreLockHolder TSLockHolder(FALSE); if (bRequiresTSL) { TSLockHolder.Acquire(); } _ASSERTE(s_pThreadStore->DbgFindThread(thread)); _ASSERTE(thread->HasValidThreadHandle()); _ASSERTE(thread->m_State & Thread::TS_WeOwn); _ASSERTE(thread->IsUnstarted()); _ASSERTE(!thread->IsDead()); if (thread->m_State & Thread::TS_AbortRequested) { PAL_CPP_THROW(EEException *, new EEException(COR_E_THREADABORTED)); } // Of course, m_ThreadCount is already correct since it includes started and // unstarted threads. s_pThreadStore->m_UnstartedThreadCount--; // We only count background threads that have been started if (thread->IsBackground()) s_pThreadStore->m_BackgroundThreadCount++; _ASSERTE(s_pThreadStore->m_PendingThreadCount > 0); FastInterlockDecrement(&s_pThreadStore->m_PendingThreadCount); // As soon as we erase this bit, the thread becomes eligible for suspension, // stopping, interruption, etc. FastInterlockAnd((ULONG *) &thread->m_State, ~Thread::TS_Unstarted); FastInterlockOr((ULONG *) &thread->m_State, Thread::TS_LegalToJoin); // release ThreadStore Crst to avoid Crst Violation when calling HandleThreadAbort later if (bRequiresTSL) { TSLockHolder.Release(); } // One of the components of OtherThreadsComplete() has changed, so check whether // we should now exit the EE. CheckForEEShutdown(); } LONG ThreadStore::s_DeadThreadCountThresholdForGCTrigger = 0; DWORD ThreadStore::s_DeadThreadGCTriggerPeriodMilliseconds = 0; SIZE_T *ThreadStore::s_DeadThreadGenerationCounts = nullptr; void ThreadStore::IncrementDeadThreadCountForGCTrigger() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // Although all increments and decrements are usually done inside a lock, that is not sufficient to synchronize with a // background GC thread resetting this value, hence the interlocked operation. Ignore overflow; overflow would likely never // occur, the count is treated as unsigned, and nothing bad would happen if it were to overflow. SIZE_T count = static_cast(FastInterlockIncrement(&m_DeadThreadCountForGCTrigger)); SIZE_T countThreshold = static_cast(s_DeadThreadCountThresholdForGCTrigger); if (count < countThreshold || countThreshold == 0) { return; } IGCHeap *gcHeap = GCHeapUtilities::GetGCHeap(); if (gcHeap == nullptr) { return; } SIZE_T gcLastMilliseconds = gcHeap->GetLastGCStartTime(gcHeap->GetMaxGeneration()); SIZE_T gcNowMilliseconds = gcHeap->GetNow(); if (gcNowMilliseconds - gcLastMilliseconds < s_DeadThreadGCTriggerPeriodMilliseconds) { return; } if (!g_fEEStarted) // required for FinalizerThread::EnableFinalization() below { return; } // The GC is triggered on the finalizer thread since it's not safe to trigger it on DLL_THREAD_DETACH. // TriggerGCForDeadThreadsIfNecessary() will determine which generation of GC to trigger, and may not actually trigger a GC. // If a GC is triggered, since there would be a delay before the dead thread count is updated, clear the count and wait for // it to reach the threshold again. If a GC would not be triggered, the count is still cleared here to prevent waking up the // finalizer thread to do the work in TriggerGCForDeadThreadsIfNecessary() for every dead thread. m_DeadThreadCountForGCTrigger = 0; m_TriggerGCForDeadThreads = true; FinalizerThread::EnableFinalization(); } void ThreadStore::DecrementDeadThreadCountForGCTrigger() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // Although all increments and decrements are usually done inside a lock, that is not sufficient to synchronize with a // background GC thread resetting this value, hence the interlocked operation. if (FastInterlockDecrement(&m_DeadThreadCountForGCTrigger) < 0) { m_DeadThreadCountForGCTrigger = 0; } } void ThreadStore::OnMaxGenerationGCStarted() { LIMITED_METHOD_CONTRACT; // A dead thread may contribute to triggering a GC at most once. After a max-generation GC occurs, if some dead thread // objects are still reachable due to references to the thread objects, they will not contribute to triggering a GC again. // Synchronize the store with increment/decrement operations occurring on different threads, and make the change visible to // other threads in order to prevent unnecessary GC triggers. FastInterlockExchange(&m_DeadThreadCountForGCTrigger, 0); } bool ThreadStore::ShouldTriggerGCForDeadThreads() { LIMITED_METHOD_CONTRACT; return m_TriggerGCForDeadThreads; } void ThreadStore::TriggerGCForDeadThreadsIfNecessary() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; if (!m_TriggerGCForDeadThreads) { return; } m_TriggerGCForDeadThreads = false; if (g_fEEShutDown) { // Not safe to touch CLR state return; } unsigned gcGenerationToTrigger = 0; IGCHeap *gcHeap = GCHeapUtilities::GetGCHeap(); _ASSERTE(gcHeap != nullptr); SIZE_T generationCountThreshold = static_cast(s_DeadThreadCountThresholdForGCTrigger) / 2; unsigned maxGeneration = gcHeap->GetMaxGeneration(); if (!s_DeadThreadGenerationCounts) { // initialize this field on first use with an entry for every table. s_DeadThreadGenerationCounts = new (nothrow) SIZE_T[maxGeneration + 1]; if (!s_DeadThreadGenerationCounts) { return; } } memset(s_DeadThreadGenerationCounts, 0, sizeof(SIZE_T) * (maxGeneration + 1)); { ThreadStoreLockHolder threadStoreLockHolder; GCX_COOP(); // Determine the generation for which to trigger a GC. Iterate over all dead threads that have not yet been considered // for triggering a GC and see how many are in which generations. for (Thread *thread = ThreadStore::GetAllThreadList(NULL, Thread::TS_Dead, Thread::TS_Dead); thread != nullptr; thread = ThreadStore::GetAllThreadList(thread, Thread::TS_Dead, Thread::TS_Dead)) { if (thread->HasDeadThreadBeenConsideredForGCTrigger()) { continue; } Object *exposedObject = OBJECTREFToObject(thread->GetExposedObjectRaw()); if (exposedObject == nullptr) { continue; } unsigned exposedObjectGeneration = gcHeap->WhichGeneration(exposedObject); SIZE_T newDeadThreadGenerationCount = ++s_DeadThreadGenerationCounts[exposedObjectGeneration]; if (exposedObjectGeneration > gcGenerationToTrigger && newDeadThreadGenerationCount >= generationCountThreshold) { gcGenerationToTrigger = exposedObjectGeneration; if (gcGenerationToTrigger >= maxGeneration) { break; } } } // Make sure that enough time has elapsed since the last GC of the desired generation. We don't want to trigger GCs // based on this heuristic too often. Give it some time to let the memory pressure trigger GCs automatically, and only // if it doesn't in the given time, this heuristic may kick in to trigger a GC. SIZE_T gcLastMilliseconds = gcHeap->GetLastGCStartTime(gcGenerationToTrigger); SIZE_T gcNowMilliseconds = gcHeap->GetNow(); if (gcNowMilliseconds - gcLastMilliseconds < s_DeadThreadGCTriggerPeriodMilliseconds) { return; } // For threads whose exposed objects are in the generation of GC that will be triggered or in a lower GC generation, // mark them as having contributed to a GC trigger to prevent redundant GC triggers for (Thread *thread = ThreadStore::GetAllThreadList(NULL, Thread::TS_Dead, Thread::TS_Dead); thread != nullptr; thread = ThreadStore::GetAllThreadList(thread, Thread::TS_Dead, Thread::TS_Dead)) { if (thread->HasDeadThreadBeenConsideredForGCTrigger()) { continue; } Object *exposedObject = OBJECTREFToObject(thread->GetExposedObjectRaw()); if (exposedObject == nullptr) { continue; } if (gcGenerationToTrigger < maxGeneration && gcHeap->WhichGeneration(exposedObject) > gcGenerationToTrigger) { continue; } thread->SetHasDeadThreadBeenConsideredForGCTrigger(); } } // ThreadStoreLockHolder, GCX_COOP() GCHeapUtilities::GetGCHeap()->GarbageCollect(gcGenerationToTrigger, FALSE, collection_non_blocking); } #endif // #ifndef DACCESS_COMPILE // Access the list of threads. You must be inside a critical section, otherwise // the "cursor" thread might disappear underneath you. Pass in NULL for the // cursor to begin at the start of the list. Thread *ThreadStore::GetAllThreadList(Thread *cursor, ULONG mask, ULONG bits) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; SUPPORTS_DAC; #ifndef DACCESS_COMPILE _ASSERTE((s_pThreadStore->m_Crst.GetEnterCount() > 0) || IsAtProcessExit()); #endif while (TRUE) { cursor = (cursor ? s_pThreadStore->m_ThreadList.GetNext(cursor) : s_pThreadStore->m_ThreadList.GetHead()); if (cursor == NULL) break; if ((cursor->m_State & mask) == bits) return cursor; } return NULL; } // Iterate over the threads that have been started Thread *ThreadStore::GetThreadList(Thread *cursor) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; SUPPORTS_DAC; return GetAllThreadList(cursor, (Thread::TS_Unstarted | Thread::TS_Dead), 0); } //--------------------------------------------------------------------------------------- // // Grab a consistent snapshot of the thread's state, for reporting purposes only. // // Return Value: // the current state of the thread // Thread::ThreadState Thread::GetSnapshotState() { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; SUPPORTS_DAC; } CONTRACTL_END; ThreadState res = m_State; if (res & TS_ReportDead) { res = (ThreadState) (res | TS_Dead); } return res; } #ifndef DACCESS_COMPILE BOOL CLREventWaitWithTry(CLREventBase *pEvent, DWORD timeout, BOOL fAlertable, DWORD *pStatus) { CONTRACTL { NOTHROW; WRAPPER(GC_TRIGGERS); } CONTRACTL_END; BOOL fLoop = TRUE; EX_TRY { *pStatus = pEvent->Wait(timeout, fAlertable); fLoop = FALSE; } EX_CATCH { } EX_END_CATCH(SwallowAllExceptions); return fLoop; } // We shut down the EE only when all the non-background threads have terminated // (unless this is an exceptional termination). So the main thread calls here to // wait before tearing down the EE. void ThreadStore::WaitForOtherThreads() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; CHECK_ONE_STORE(); Thread *pCurThread = GetThread(); // Regardless of whether the main thread is a background thread or not, force // it to be one. This simplifies our rules for counting non-background threads. pCurThread->SetBackground(TRUE); LOG((LF_SYNC, INFO3, "WaitForOtherThreads obtain lock\n")); ThreadStoreLockHolder TSLockHolder(TRUE); if (!OtherThreadsComplete()) { TSLockHolder.Release(); FastInterlockOr((ULONG *) &pCurThread->m_State, Thread::TS_ReportDead); DWORD ret = WAIT_OBJECT_0; while (CLREventWaitWithTry(&m_TerminationEvent, INFINITE, TRUE, &ret)) { } _ASSERTE(ret == WAIT_OBJECT_0); } } // Every EE process can lazily create a GUID that uniquely identifies it (for // purposes of remoting). const GUID &ThreadStore::GetUniqueEEId() { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; if (!m_GuidCreated) { ThreadStoreLockHolder TSLockHolder(TRUE); if (!m_GuidCreated) { HRESULT hr = ::CoCreateGuid(&m_EEGuid); _ASSERTE(SUCCEEDED(hr)); if (SUCCEEDED(hr)) m_GuidCreated = TRUE; } if (!m_GuidCreated) return IID_NULL; } return m_EEGuid; } #ifdef _DEBUG BOOL ThreadStore::DbgFindThread(Thread *target) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; CHECK_ONE_STORE(); // Cache the current change stamp for g_TrapReturningThreads LONG chgStamp = g_trtChgStamp; STRESS_LOG3(LF_STORE, LL_INFO100, "ThreadStore::DbgFindThread - [thread=%p]. trt=%d. chgStamp=%d\n", GetThread(), g_TrapReturningThreads.Load(), chgStamp); #if 0 // g_TrapReturningThreads debug code. int iRetry = 0; Retry: #endif // g_TrapReturningThreads debug code. BOOL found = FALSE; Thread *cur = NULL; LONG cnt = 0; LONG cntBack = 0; LONG cntUnstart = 0; LONG cntDead = 0; LONG cntReturn = 0; while ((cur = GetAllThreadList(cur, 0, 0)) != NULL) { cnt++; if (cur->IsDead()) cntDead++; // Unstarted threads do not contribute to the count of background threads if (cur->IsUnstarted()) cntUnstart++; else if (cur->IsBackground()) cntBack++; if (cur == target) found = TRUE; // Note that (DebugSuspendPending | SuspendPending) implies a count of 2. // We don't count GCPending because a single trap is held for the entire // GC, instead of counting each interesting thread. if (cur->m_State & Thread::TS_DebugSuspendPending) cntReturn++; // CoreCLR does not support user-requested thread suspension _ASSERTE(!(cur->m_State & Thread::TS_UserSuspendPending)); if (cur->m_TraceCallCount > 0) cntReturn++; if (cur->IsAbortRequested()) cntReturn++; } _ASSERTE(cnt == m_ThreadCount); _ASSERTE(cntUnstart == m_UnstartedThreadCount); _ASSERTE(cntBack == m_BackgroundThreadCount); _ASSERTE(cntDead == m_DeadThreadCount); _ASSERTE(0 <= m_PendingThreadCount); #if 0 // g_TrapReturningThreads debug code. if (cntReturn != g_TrapReturningThreads /*&& !g_fEEShutDown*/) { // If count is off, try again, to account for multiple threads. if (iRetry < 4) { // printf("Retry %d. cntReturn:%d, gReturn:%d\n", iRetry, cntReturn, g_TrapReturningThreads); ++iRetry; goto Retry; } printf("cnt:%d, Un:%d, Back:%d, Dead:%d, cntReturn:%d, TrapReturn:%d, eeShutdown:%d, threadShutdown:%d\n", cnt,cntUnstart,cntBack,cntDead,cntReturn,g_TrapReturningThreads, g_fEEShutDown, Thread::IsAtProcessExit()); LOG((LF_CORDB, LL_INFO1000, "SUSPEND: cnt:%d, Un:%d, Back:%d, Dead:%d, cntReturn:%d, TrapReturn:%d, eeShutdown:%d, threadShutdown:%d\n", cnt,cntUnstart,cntBack,cntDead,cntReturn,g_TrapReturningThreads, g_fEEShutDown, Thread::IsAtProcessExit()) ); //_ASSERTE(cntReturn + 2 >= g_TrapReturningThreads); } if (iRetry > 0 && iRetry < 4) { printf("%d retries to re-sync counted TrapReturn with global TrapReturn.\n", iRetry); } #endif // g_TrapReturningThreads debug code. STRESS_LOG4(LF_STORE, LL_INFO100, "ThreadStore::DbgFindThread - [thread=%p]. trt=%d. chg=%d. cnt=%d\n", GetThread(), g_TrapReturningThreads.Load(), g_trtChgStamp.Load(), cntReturn); // Because of race conditions and the fact that the GC places its // own count, I can't assert this precisely. But I do want to be // sure that this count isn't wandering ever higher -- with a // nasty impact on the performance of GC mode changes and method // call chaining! // // We don't bother asserting this during process exit, because // during a shutdown we will quietly terminate threads that are // being waited on. (If we aren't shutting down, we carefully // decrement our counts and alert anyone waiting for us to // return). // // Note: we don't actually assert this if // ThreadStore::TrapReturningThreads() updated g_TrapReturningThreads // between the beginning of this function and the moment of the assert. // *** The order of evaluation in the if condition is important *** _ASSERTE( (g_trtChgInFlight != 0 || (cntReturn + 2 >= g_TrapReturningThreads) || chgStamp != g_trtChgStamp) || g_fEEShutDown); return found; } #endif // _DEBUG void Thread::HandleThreadInterrupt (BOOL fWaitForADUnload) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_SO_TOLERANT; // If we're waiting for shutdown, we don't want to abort/interrupt this thread if (HasThreadStateNC(Thread::TSNC_BlockedForShutdown)) return; BEGIN_SO_INTOLERANT_CODE(this); if ((m_UserInterrupt & TI_Abort) != 0) { // If the thread is waiting for AD unload to finish, and the thread is interrupted, // we can start aborting. HandleThreadAbort(fWaitForADUnload); } if ((m_UserInterrupt & TI_Interrupt) != 0) { ResetThreadState ((ThreadState)(TS_Interrupted | TS_Interruptible)); FastInterlockAnd ((DWORD*)&m_UserInterrupt, ~TI_Interrupt); COMPlusThrow(kThreadInterruptedException); } END_SO_INTOLERANT_CODE; } #ifdef _DEBUG #define MAXSTACKBYTES (2 * GetOsPageSize()) void CleanStackForFastGCStress () { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; PVOID StackLimit = ClrTeb::GetStackLimit(); size_t nBytes = (size_t)&nBytes - (size_t)StackLimit; nBytes &= ~sizeof (size_t); if (nBytes > MAXSTACKBYTES) { nBytes = MAXSTACKBYTES; } size_t* buffer = (size_t*) _alloca (nBytes); memset(buffer, 0, nBytes); GetThread()->m_pCleanedStackBase = &nBytes; } void Thread::ObjectRefFlush(Thread* thread) { BEGIN_PRESERVE_LAST_ERROR; // The constructor and destructor of AutoCleanupSONotMainlineHolder (allocated by SO_NOT_MAINLINE_FUNCTION below) // may trash the last error, so we need to save and restore last error here. Also, we need to add a scope here // because we can't let the destructor run after we call SetLastError(). { // this is debug only code, so no need to validate STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_NOTRIGGER; STATIC_CONTRACT_ENTRY_POINT; _ASSERTE(thread->PreemptiveGCDisabled()); // Should have been in managed code memset(thread->dangerousObjRefs, 0, sizeof(thread->dangerousObjRefs)); thread->m_allObjRefEntriesBad = FALSE; CLEANSTACKFORFASTGCSTRESS (); } END_PRESERVE_LAST_ERROR; } #endif #if defined(STRESS_HEAP) PtrHashMap *g_pUniqueStackMap = NULL; Crst *g_pUniqueStackCrst = NULL; #define UniqueStackDepth 8 BOOL StackCompare (UPTR val1, UPTR val2) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; size_t *p1 = (size_t *)(val1 << 1); size_t *p2 = (size_t *)val2; if (p1[0] != p2[0]) { return FALSE; } size_t nElem = p1[0]; if (nElem >= UniqueStackDepth) { nElem = UniqueStackDepth; } p1 ++; p2 ++; for (size_t n = 0; n < nElem; n ++) { if (p1[n] != p2[n]) { return FALSE; } } return TRUE; } void UniqueStackSetupMap() { WRAPPER_NO_CONTRACT; if (g_pUniqueStackCrst == NULL) { Crst *Attempt = new Crst ( CrstUniqueStack, CrstFlags(CRST_REENTRANCY | CRST_UNSAFE_ANYMODE)); if (FastInterlockCompareExchangePointer(&g_pUniqueStackCrst, Attempt, NULL) != NULL) { // We lost the race delete Attempt; } } // Now we have a Crst we can use to synchronize the remainder of the init. if (g_pUniqueStackMap == NULL) { CrstHolder ch(g_pUniqueStackCrst); if (g_pUniqueStackMap == NULL) { PtrHashMap *map = new (SystemDomain::GetGlobalLoaderAllocator()->GetLowFrequencyHeap()) PtrHashMap (); LockOwner lock = {g_pUniqueStackCrst, IsOwnerOfCrst}; map->Init (256, StackCompare, TRUE, &lock); g_pUniqueStackMap = map; } } } BOOL StartUniqueStackMapHelper() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; BOOL fOK = TRUE; EX_TRY { if (g_pUniqueStackMap == NULL) { UniqueStackSetupMap(); } } EX_CATCH { fOK = FALSE; } EX_END_CATCH(SwallowAllExceptions); return fOK; } BOOL StartUniqueStackMap () { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; return StartUniqueStackMapHelper(); } #ifndef FEATURE_PAL size_t UpdateStackHash(size_t hash, size_t retAddr) { return ((hash << 3) + hash) ^ retAddr; } /***********************************************************************/ size_t getStackHash(size_t* stackTrace, size_t* stackTop, size_t* stackStop, size_t stackBase, size_t stackLimit) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // return a hash of every return address found between 'stackTop' (the lowest address) // and 'stackStop' (the highest address) size_t hash = 0; int idx = 0; #ifdef _TARGET_X86_ static size_t moduleBase = (size_t) -1; static size_t moduleTop = (size_t) -1; if (moduleTop == (size_t) -1) { MEMORY_BASIC_INFORMATION mbi; if (ClrVirtualQuery(getStackHash, &mbi, sizeof(mbi))) { moduleBase = (size_t)mbi.AllocationBase; moduleTop = (size_t)mbi.BaseAddress + mbi.RegionSize; } else { // way bad error, probably just assert and exit _ASSERTE (!"ClrVirtualQuery failed"); moduleBase = 0; moduleTop = 0; } } while (stackTop < stackStop) { // Clean out things that point to stack, as those can't be return addresses if (*stackTop > moduleBase && *stackTop < moduleTop) { TADDR dummy; if (isRetAddr((TADDR)*stackTop, &dummy)) { hash = UpdateStackHash(hash, *stackTop); // If there is no jitted code on the stack, then just use the // top 16 frames as the context. idx++; if (idx <= UniqueStackDepth) { stackTrace [idx] = *stackTop; } } } stackTop++; } #else // _TARGET_X86_ CONTEXT ctx; ClrCaptureContext(&ctx); UINT_PTR uControlPc = (UINT_PTR)GetIP(&ctx); UINT_PTR uImageBase; UINT_PTR uPrevControlPc = uControlPc; for (;;) { RtlLookupFunctionEntry(uControlPc, ARM_ONLY((DWORD*))(&uImageBase), NULL ); if (((UINT_PTR)g_pMSCorEE) != uImageBase) { break; } uControlPc = Thread::VirtualUnwindCallFrame(&ctx); UINT_PTR uRetAddrForHash = uControlPc; if (uPrevControlPc == uControlPc) { // This is a special case when we fail to acquire the loader lock // in RtlLookupFunctionEntry(), which then returns false. The end // result is that we cannot go any further on the stack and // we will loop infinitely (because the owner of the loader lock // is blocked on us). hash = 0; break; } else { uPrevControlPc = uControlPc; } hash = UpdateStackHash(hash, uRetAddrForHash); // If there is no jitted code on the stack, then just use the // top 16 frames as the context. idx++; if (idx <= UniqueStackDepth) { stackTrace [idx] = uRetAddrForHash; } } #endif // _TARGET_X86_ stackTrace [0] = idx; return(hash); } void UniqueStackHelper(size_t stackTraceHash, size_t *stackTrace) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; EX_TRY { size_t nElem = stackTrace[0]; if (nElem >= UniqueStackDepth) { nElem = UniqueStackDepth; } AllocMemHolder stackTraceInMap = SystemDomain::GetGlobalLoaderAllocator()->GetLowFrequencyHeap()->AllocMem(S_SIZE_T(sizeof(size_t *)) * (S_SIZE_T(nElem) + S_SIZE_T(1))); memcpy (stackTraceInMap, stackTrace, sizeof(size_t *) * (nElem + 1)); g_pUniqueStackMap->InsertValue(stackTraceHash, stackTraceInMap); stackTraceInMap.SuppressRelease(); } EX_CATCH { } EX_END_CATCH(SwallowAllExceptions); } /***********************************************************************/ /* returns true if this stack has not been seen before, useful for running tests only once per stack trace. */ BOOL Thread::UniqueStack(void* stackStart) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_NOT_MAINLINE; } CONTRACTL_END; // If we where not told where to start, start at the caller of UniqueStack if (stackStart == 0) { stackStart = &stackStart; } if (g_pUniqueStackMap == NULL) { if (!StartUniqueStackMap ()) { // We fail to initialize unique stack map due to OOM. // Let's say the stack is unique. return TRUE; } } size_t stackTrace[UniqueStackDepth+1] = {0}; // stackTraceHash represents a hash of entire stack at the time we make the call, // We insure at least GC per unique stackTrace. What information is contained in // 'stackTrace' is somewhat arbitrary. We choose it to mean all functions live // on the stack up to the first jitted function. size_t stackTraceHash; Thread* pThread = GetThread(); void* stopPoint = pThread->m_CacheStackBase; #ifdef _TARGET_X86_ // Find the stop point (most jitted function) Frame* pFrame = pThread->GetFrame(); for(;;) { // skip GC frames if (pFrame == 0 || pFrame == (Frame*) -1) break; pFrame->GetFunction(); // This insures that helper frames are inited if (pFrame->GetReturnAddress() != 0) { stopPoint = pFrame; break; } pFrame = pFrame->Next(); } #endif // _TARGET_X86_ // Get hash of all return addresses between here an the top most jitted function stackTraceHash = getStackHash (stackTrace, (size_t*) stackStart, (size_t*) stopPoint, size_t(pThread->m_CacheStackBase), size_t(pThread->m_CacheStackLimit)); if (stackTraceHash == 0 || g_pUniqueStackMap->LookupValue (stackTraceHash, stackTrace) != (LPVOID)INVALIDENTRY) { return FALSE; } BOOL fUnique = FALSE; { CrstHolder ch(g_pUniqueStackCrst); #ifdef _DEBUG if (GetThread ()) GetThread ()->m_bUniqueStacking = TRUE; #endif if (g_pUniqueStackMap->LookupValue (stackTraceHash, stackTrace) != (LPVOID)INVALIDENTRY) { fUnique = FALSE; } else { fUnique = TRUE; FAULT_NOT_FATAL(); UniqueStackHelper(stackTraceHash, stackTrace); } #ifdef _DEBUG if (GetThread ()) GetThread ()->m_bUniqueStacking = FALSE; #endif } #ifdef _DEBUG static int fCheckStack = -1; if (fCheckStack == -1) { fCheckStack = CLRConfig::GetConfigValue(CLRConfig::INTERNAL_FastGCCheckStack); } if (fCheckStack && pThread->m_pCleanedStackBase > stackTrace && pThread->m_pCleanedStackBase - stackTrace > (int) MAXSTACKBYTES) { _ASSERTE (!"Garbage on stack"); } #endif return fUnique; } #else // !FEATURE_PAL BOOL Thread::UniqueStack(void* stackStart) { return FALSE; } #endif // !FEATURE_PAL #endif // STRESS_HEAP /* * GetStackLowerBound * * Returns the lower bound of the stack space. Note -- the practical bound is some number of pages greater than * this value -- those pages are reserved for a stack overflow exception processing. * * Parameters: * None * * Returns: * address of the lower bound of the threads's stack. */ void * Thread::GetStackLowerBound() { // Called during fiber switch. Can not have non-static contract. STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_NOTRIGGER; STATIC_CONTRACT_SO_TOLERANT; #ifndef FEATURE_PAL MEMORY_BASIC_INFORMATION lowerBoundMemInfo; SIZE_T dwRes; dwRes = ClrVirtualQuery((const void *)&lowerBoundMemInfo, &lowerBoundMemInfo, sizeof(MEMORY_BASIC_INFORMATION)); if (sizeof(MEMORY_BASIC_INFORMATION) == dwRes) { return (void *)(lowerBoundMemInfo.AllocationBase); } else { return NULL; } #else // !FEATURE_PAL return PAL_GetStackLimit(); #endif // !FEATURE_PAL } /* * GetStackUpperBound * * Return the upper bound of the thread's stack space. * * Parameters: * None * * Returns: * address of the base of the threads's stack. */ void *Thread::GetStackUpperBound() { // Called during fiber switch. Can not have non-static contract. STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_NOTRIGGER; STATIC_CONTRACT_SO_TOLERANT; return ClrTeb::GetStackBase(); } BOOL Thread::SetStackLimits(SetStackLimitScope scope) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; if (scope == fAll) { m_CacheStackBase = GetStackUpperBound(); m_CacheStackLimit = GetStackLowerBound(); if (m_CacheStackLimit == NULL) { _ASSERTE(!"Failed to set stack limits"); return FALSE; } // Compute the limit used by EnsureSufficientExecutionStack and cache it on the thread. This minimum stack size should // be sufficient to allow a typical non-recursive call chain to execute, including potential exception handling and // garbage collection. Used for probing for available stack space through RuntimeImports.EnsureSufficientExecutionStack, // among other things. #ifdef BIT64 const UINT_PTR MinExecutionStackSize = 128 * 1024; #else // !BIT64 const UINT_PTR MinExecutionStackSize = 64 * 1024; #endif // BIT64 _ASSERTE(m_CacheStackBase >= m_CacheStackLimit); if ((reinterpret_cast(m_CacheStackBase) - reinterpret_cast(m_CacheStackLimit)) > MinExecutionStackSize) { m_CacheStackSufficientExecutionLimit = reinterpret_cast(m_CacheStackLimit) + MinExecutionStackSize; } else { m_CacheStackSufficientExecutionLimit = reinterpret_cast(m_CacheStackBase); } } // Ensure that we've setup the stack guarantee properly before we cache the stack limits // as they depend upon the stack guarantee. if (FAILED(CLRSetThreadStackGuarantee())) return FALSE; // Cache the last stack addresses that we are allowed to touch. We throw a stack overflow // if we cross that line. Note that we ignore any subsequent calls to STSG for Whidbey until // we see an exception and recache the values. We use the LastAllowableAddresses to // determine if we've taken a hard SO and the ProbeLimits on the probes themselves. m_LastAllowableStackAddress = GetLastNormalStackAddress(); if (g_pConfig->ProbeForStackOverflow()) { m_ProbeLimit = m_LastAllowableStackAddress; } else { // If we have stack probing disabled, set the probeLimit to 0 so that all probes will pass. This // way we don't have to do an extra check in the probe code. m_ProbeLimit = 0; } return TRUE; } //--------------------------------------------------------------------------------------------- // Routines we use to managed a thread's stack, for fiber switching or stack overflow purposes. //--------------------------------------------------------------------------------------------- HRESULT Thread::CLRSetThreadStackGuarantee(SetThreadStackGuaranteeScope fScope) { CONTRACTL { WRAPPER(NOTHROW); GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; #ifndef FEATURE_PAL // TODO: we need to measure what the stack usage needs are at the limits in the hosted scenario for host callbacks if (Thread::IsSetThreadStackGuaranteeInUse(fScope)) { // Tune this as needed ULONG uGuardSize = SIZEOF_DEFAULT_STACK_GUARANTEE; int EXTRA_PAGES = 0; #if defined(_WIN64) // Free Build EH Stack Stats: // -------------------------------- // currently the maximum stack usage we'll face while handling a SO includes: // 4.3k for the OS (kernel32!RaiseException, Rtl EH dispatch code, RtlUnwindEx [second pass]) // 1.2k for the CLR EH setup (NakedThrowHelper*) // 4.5k for other heavy CLR stack creations (2x CONTEXT, 1x REGDISPLAY) // ~1.0k for other misc CLR stack allocations // ----- // 11.0k --> ~2.75 pages for CLR SO EH dispatch // // -plus we might need some more for debugger EH dispatch, Watson, etc... // -also need to take into account that we can lose up to 1 page of the guard region // -additionally, we need to provide some region to hosts to allow for lock acquisition in a hosted scenario // EXTRA_PAGES = 3; INDEBUG(EXTRA_PAGES += 1); int ThreadGuardPages = CLRConfig::GetConfigValue(CLRConfig::EXTERNAL_ThreadGuardPages); if (ThreadGuardPages == 0) { uGuardSize += (EXTRA_PAGES * GetOsPageSize()); } else { uGuardSize += (ThreadGuardPages * GetOsPageSize()); } #else // _WIN64 #ifdef _DEBUG uGuardSize += (1 * GetOsPageSize()); // one extra page for debug infrastructure #endif // _DEBUG #endif // _WIN64 LOG((LF_EH, LL_INFO10000, "STACKOVERFLOW: setting thread stack guarantee to 0x%x\n", uGuardSize)); if (!::SetThreadStackGuarantee(&uGuardSize)) { return HRESULT_FROM_GetLastErrorNA(); } } #endif // !FEATURE_PAL return S_OK; } /* * GetLastNormalStackAddress * * GetLastNormalStackAddress returns the last stack address before the guard * region of a thread. This is the last address that one could write to before * a stack overflow occurs. * * Parameters: * StackLimit - the base of the stack allocation * * Returns: * Address of the first page of the guard region. */ UINT_PTR Thread::GetLastNormalStackAddress(UINT_PTR StackLimit) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; UINT_PTR cbStackGuarantee = GetStackGuarantee(); // Here we take the "hard guard region size", the "stack guarantee" and the "fault page" and add them // all together. Note that the "fault page" is the reason for the extra GetOsPageSize() below. The OS // will guarantee us a certain amount of stack remaining after a stack overflow. This is called the // "stack guarantee". But to do this, it has to fault on the page before that region as the app is // allowed to fault at the very end of that page. So, as a result, the last normal stack address is // one page sooner. return StackLimit + (cbStackGuarantee #ifndef FEATURE_PAL + GetOsPageSize() #endif // !FEATURE_PAL + HARD_GUARD_REGION_SIZE); } #ifdef _DEBUG static void DebugLogMBIFlags(UINT uState, UINT uProtect) { CONTRACTL { NOTHROW; GC_NOTRIGGER; CANNOT_TAKE_LOCK; } CONTRACTL_END; #ifndef FEATURE_PAL #define LOG_FLAG(flags, name) \ if (flags & name) \ { \ LOG((LF_EH, LL_INFO1000, "" #name " ")); \ } \ if (uState) { LOG((LF_EH, LL_INFO1000, "State: ")); LOG_FLAG(uState, MEM_COMMIT); LOG_FLAG(uState, MEM_RESERVE); LOG_FLAG(uState, MEM_DECOMMIT); LOG_FLAG(uState, MEM_RELEASE); LOG_FLAG(uState, MEM_FREE); LOG_FLAG(uState, MEM_PRIVATE); LOG_FLAG(uState, MEM_MAPPED); LOG_FLAG(uState, MEM_RESET); LOG_FLAG(uState, MEM_TOP_DOWN); LOG_FLAG(uState, MEM_WRITE_WATCH); LOG_FLAG(uState, MEM_PHYSICAL); LOG_FLAG(uState, MEM_LARGE_PAGES); LOG_FLAG(uState, MEM_4MB_PAGES); } if (uProtect) { LOG((LF_EH, LL_INFO1000, "Protect: ")); LOG_FLAG(uProtect, PAGE_NOACCESS); LOG_FLAG(uProtect, PAGE_READONLY); LOG_FLAG(uProtect, PAGE_READWRITE); LOG_FLAG(uProtect, PAGE_WRITECOPY); LOG_FLAG(uProtect, PAGE_EXECUTE); LOG_FLAG(uProtect, PAGE_EXECUTE_READ); LOG_FLAG(uProtect, PAGE_EXECUTE_READWRITE); LOG_FLAG(uProtect, PAGE_EXECUTE_WRITECOPY); LOG_FLAG(uProtect, PAGE_GUARD); LOG_FLAG(uProtect, PAGE_NOCACHE); LOG_FLAG(uProtect, PAGE_WRITECOMBINE); } #undef LOG_FLAG #endif // !FEATURE_PAL } static void DebugLogStackRegionMBIs(UINT_PTR uLowAddress, UINT_PTR uHighAddress) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_INTOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; MEMORY_BASIC_INFORMATION meminfo; UINT_PTR uStartOfThisRegion = uLowAddress; LOG((LF_EH, LL_INFO1000, "----------------------------------------------------------------------\n")); while (uStartOfThisRegion < uHighAddress) { SIZE_T res = ClrVirtualQuery((const void *)uStartOfThisRegion, &meminfo, sizeof(meminfo)); if (sizeof(meminfo) != res) { LOG((LF_EH, LL_INFO1000, "VirtualQuery failed on %p\n", uStartOfThisRegion)); break; } UINT_PTR uStartOfNextRegion = uStartOfThisRegion + meminfo.RegionSize; if (uStartOfNextRegion > uHighAddress) { uStartOfNextRegion = uHighAddress; } UINT_PTR uRegionSize = uStartOfNextRegion - uStartOfThisRegion; LOG((LF_EH, LL_INFO1000, "0x%p -> 0x%p (%d pg) ", uStartOfThisRegion, uStartOfNextRegion - 1, uRegionSize / GetOsPageSize())); DebugLogMBIFlags(meminfo.State, meminfo.Protect); LOG((LF_EH, LL_INFO1000, "\n")); uStartOfThisRegion = uStartOfNextRegion; } LOG((LF_EH, LL_INFO1000, "----------------------------------------------------------------------\n")); } // static void Thread::DebugLogStackMBIs() { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_INTOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; Thread* pThread = GetThread(); // N.B. this can be NULL! UINT_PTR uStackLimit = (UINT_PTR)GetStackLowerBound(); UINT_PTR uStackBase = (UINT_PTR)GetStackUpperBound(); if (pThread) { uStackLimit = (UINT_PTR)pThread->GetCachedStackLimit(); uStackBase = (UINT_PTR)pThread->GetCachedStackBase(); } else { uStackLimit = (UINT_PTR)GetStackLowerBound(); uStackBase = (UINT_PTR)GetStackUpperBound(); } UINT_PTR uStackSize = uStackBase - uStackLimit; LOG((LF_EH, LL_INFO1000, "----------------------------------------------------------------------\n")); LOG((LF_EH, LL_INFO1000, "Stack Snapshot 0x%p -> 0x%p (%d pg)\n", uStackLimit, uStackBase, uStackSize / GetOsPageSize())); if (pThread) { LOG((LF_EH, LL_INFO1000, "Last normal addr: 0x%p\n", pThread->GetLastNormalStackAddress())); } DebugLogStackRegionMBIs(uStackLimit, uStackBase); } #endif // _DEBUG // // IsSPBeyondLimit // // Determines if the stack pointer is beyond the stack limit, in which case // we can assume we've taken a hard SO. // // Parameters: none // // Returns: bool indicating if SP is beyond the limit or not // BOOL Thread::IsSPBeyondLimit() { WRAPPER_NO_CONTRACT; // Reset the stack limits if necessary. // @todo . Add a vectored handler for X86 so that we reset the stack limits // there, as anything that supports SetThreadStackGuarantee will support vectored handlers. // Then we can always assume during EH processing that our stack limits are good and we // don't have to call ResetStackLimits. ResetStackLimits(); char *approxSP = (char *)GetCurrentSP(); if (approxSP < (char *)(GetLastAllowableStackAddress())) { return TRUE; } return FALSE; } __declspec(noinline) void AllocateSomeStack(){ LIMITED_METHOD_CONTRACT; #ifdef _TARGET_X86_ const size_t size = 0x200; #else //_TARGET_X86_ const size_t size = 0x400; #endif //_TARGET_X86_ INT8* mem = (INT8*)_alloca(size); // Actually touch the memory we just allocated so the compiler can't // optimize it away completely. // NOTE: this assumes the stack grows down (towards 0). VolatileStore(mem, 0); } #ifndef FEATURE_PAL // static // private BOOL Thread::DoesRegionContainGuardPage(UINT_PTR uLowAddress, UINT_PTR uHighAddress) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; SIZE_T dwRes; MEMORY_BASIC_INFORMATION meminfo; UINT_PTR uStartOfCurrentRegion = uLowAddress; while (uStartOfCurrentRegion < uHighAddress) { #undef VirtualQuery // This code can run below YieldTask, which means that it must not call back into the host. // The reason is that YieldTask is invoked by the host, and the host needs not be reentrant. dwRes = VirtualQuery((const void *)uStartOfCurrentRegion, &meminfo, sizeof(meminfo)); #define VirtualQuery(lpAddress, lpBuffer, dwLength) Dont_Use_VirtualQuery(lpAddress, lpBuffer, dwLength) // If the query fails then assume we have no guard page. if (sizeof(meminfo) != dwRes) { return FALSE; } if (meminfo.Protect & PAGE_GUARD) { return TRUE; } uStartOfCurrentRegion += meminfo.RegionSize; } return FALSE; } #endif // !FEATURE_PAL /* * DetermineIfGuardPagePresent * * DetermineIfGuardPagePresent returns TRUE if the thread's stack contains a proper guard page. This function makes * a physical check of the stack, rather than relying on whether or not the CLR is currently processing a stack * overflow exception. * * It seems reasonable to want to check just the 3rd page for !MEM_COMMIT or PAGE_GUARD, but that's no good in a * world where a) one can extend the guard region arbitrarily with SetThreadStackGuarantee(), b) a thread's stack * could be pre-committed, and c) another lib might reset the guard page very high up on the stack, much as we * do. In that world, we have to do VirtualQuery from the lower bound up until we find a region with PAGE_GUARD on * it. If we've never SO'd, then that's two calls to VirtualQuery. * * Parameters: * None * * Returns: * TRUE if the thread has a guard page, FALSE otherwise. */ BOOL Thread::DetermineIfGuardPagePresent() { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; #ifndef FEATURE_PAL BOOL bStackGuarded = FALSE; UINT_PTR uStackBase = (UINT_PTR)GetCachedStackBase(); UINT_PTR uStackLimit = (UINT_PTR)GetCachedStackLimit(); // Note: we start our queries after the hard guard page (one page up from the base of the stack.) We know the // very last region of the stack is never the guard page (its always the uncomitted "hard" guard page) so there's // no need to waste a query on it. bStackGuarded = DoesRegionContainGuardPage(uStackLimit + HARD_GUARD_REGION_SIZE, uStackBase); LOG((LF_EH, LL_INFO10000, "Thread::DetermineIfGuardPagePresent: stack guard page: %s\n", bStackGuarded ? "PRESENT" : "MISSING")); return bStackGuarded; #else // !FEATURE_PAL return TRUE; #endif // !FEATURE_PAL } /* * GetLastNormalStackAddress * * GetLastNormalStackAddress returns the last stack address before the guard * region of this thread. This is the last address that one could write to * before a stack overflow occurs. * * Parameters: * None * * Returns: * Address of the first page of the guard region. */ UINT_PTR Thread::GetLastNormalStackAddress() { WRAPPER_NO_CONTRACT; return GetLastNormalStackAddress((UINT_PTR)m_CacheStackLimit); } #ifdef FEATURE_STACK_PROBE /* * CanResetStackTo * * Given a target stack pointer, this function will tell us whether or not we could restore the guard page if we * unwound the stack that far. * * Parameters: * stackPointer -- stack pointer that we want to try to reset the thread's stack up to. * * Returns: * TRUE if there's enough room to reset the stack, false otherwise. */ BOOL Thread::CanResetStackTo(LPCVOID stackPointer) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; // How much space between the given stack pointer and the first guard page? // // This must be signed since the stack pointer might be in the guard region, // which is at a lower address than GetLastNormalStackAddress will return. INT_PTR iStackSpaceLeft = (INT_PTR)stackPointer - GetLastNormalStackAddress(); // We need to have enough space to call back into the EE from the handler, so we use the twice the entry point amount. // We need enough to do work and enough that partway through that work we won't probe and COMPlusThrowSO. const INT_PTR iStackSizeThreshold = (ADJUST_PROBE(DEFAULT_ENTRY_PROBE_AMOUNT * 2) * GetOsPageSize()); if (iStackSpaceLeft > iStackSizeThreshold) { return TRUE; } else { return FALSE; } } /* * IsStackSpaceAvailable * * Given a number of stack pages, this function will tell us whether or not we have that much space * before the top of the stack. If we are in the guard region we must be already handling an SO, * so we report how much space is left in the guard region * * Parameters: * numPages -- the number of pages that we need. This can be a fractional amount. * * Returns: * TRUE if there's that many pages of stack available */ BOOL Thread::IsStackSpaceAvailable(float numPages) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; // How much space between the current stack pointer and the first guard page? // // This must be signed since the stack pointer might be in the guard region, // which is at a lower address than GetLastNormalStackAddress will return. float iStackSpaceLeft = static_cast((INT_PTR)GetCurrentSP() - (INT_PTR)GetLastNormalStackAddress()); // If we have access to the stack guarantee (either in the guard region or we've tripped the guard page), then // use that. if ((iStackSpaceLeft/GetOsPageSize()) < numPages && !DetermineIfGuardPagePresent()) { UINT_PTR stackGuarantee = GetStackGuarantee(); // GetLastNormalStackAddress actually returns the 2nd to last stack page on the stack. We'll add that to our available // amount of stack, in addition to any sort of stack guarantee we might have. // // All these values are OS supplied, and will never overflow. (If they do, that means the stack is on the order // over GB, which isn't possible. iStackSpaceLeft += stackGuarantee + GetOsPageSize(); } if ((iStackSpaceLeft/GetOsPageSize()) < numPages) { return FALSE; } return TRUE; } #endif // FEATURE_STACK_PROBE /* * GetStackGuarantee * * Returns the amount of stack guaranteed after an SO but before the OS rips the process. * * Parameters: * none * * Returns: * The stack guarantee in OS pages. */ UINT_PTR Thread::GetStackGuarantee() { WRAPPER_NO_CONTRACT; #ifndef FEATURE_PAL // There is a new API available on new OS's called SetThreadStackGuarantee. It allows you to change the size of // the guard region on a per-thread basis. If we're running on an OS that supports the API, then we must query // it to see if someone has changed the size of the guard region for this thread. if (!IsSetThreadStackGuaranteeInUse()) { return SIZEOF_DEFAULT_STACK_GUARANTEE; } ULONG cbNewStackGuarantee = 0; // Passing in a value of 0 means that we're querying, and the value is changed with the new guard region // size. if (::SetThreadStackGuarantee(&cbNewStackGuarantee) && (cbNewStackGuarantee != 0)) { return cbNewStackGuarantee; } #endif // FEATURE_PAL return SIZEOF_DEFAULT_STACK_GUARANTEE; } #ifndef FEATURE_PAL // // MarkPageAsGuard // // Given a page base address, try to turn it into a guard page and then requery to determine success. // // static // private BOOL Thread::MarkPageAsGuard(UINT_PTR uGuardPageBase) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; DWORD flOldProtect; ClrVirtualProtect((LPVOID)uGuardPageBase, 1, (PAGE_READWRITE | PAGE_GUARD), &flOldProtect); // Intentionally ignore return value -- if it failed, we'll find out below // and keep moving up the stack until we either succeed or we hit the guard // region. If we don't succeed before we hit the guard region, we'll end up // with a fatal error. // Now, make sure the guard page is really there. If its not, then VirtualProtect most likely failed // because our stack had grown onto the page we were trying to protect by the time we made it into // VirtualProtect. So try the next page down. MEMORY_BASIC_INFORMATION meminfo; SIZE_T dwRes; dwRes = ClrVirtualQuery((const void *)uGuardPageBase, &meminfo, sizeof(meminfo)); return ((sizeof(meminfo) == dwRes) && (meminfo.Protect & PAGE_GUARD)); } /* * RestoreGuardPage * * RestoreGuardPage will replace the guard page on this thread's stack. The assumption is that it was removed by * the OS due to a stack overflow exception. This function requires that you know that you have enough stack space * to restore the guard page, so make sure you know what you're doing when you decide to call this. * * Parameters: * None * * Returns: * Nothing */ VOID Thread::RestoreGuardPage() { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; CANNOT_TAKE_LOCK; } CONTRACTL_END; // Need a hard SO probe here. CONTRACT_VIOLATION(SOToleranceViolation); BOOL bStackGuarded = DetermineIfGuardPagePresent(); // If the guard page is still there, then just return. if (bStackGuarded) { LOG((LF_EH, LL_INFO100, "Thread::RestoreGuardPage: no need to restore... guard page is already there.\n")); return; } UINT_PTR approxStackPointer; UINT_PTR guardPageBase; UINT_PTR guardRegionThreshold; BOOL pageMissing; if (!bStackGuarded) { // The normal guard page is the 3rd page from the base. The first page is the "hard" guard, the second one is // reserve, and the 3rd one is marked as a guard page. However, since there is now an API (on some platforms) // to change the size of the guard region, we'll just go ahead and protect the next page down from where we are // now. The guard page will get pushed forward again, just like normal, until the next stack overflow. approxStackPointer = (UINT_PTR)GetCurrentSP(); guardPageBase = (UINT_PTR)ALIGN_DOWN(approxStackPointer, GetOsPageSize()) - GetOsPageSize(); // OS uses soft guard page to update the stack info in TEB. If our guard page is not beyond the current stack, the TEB // will not be updated, and then OS's check of stack during exception will fail. if (approxStackPointer >= guardPageBase) { guardPageBase -= GetOsPageSize(); } // If we're currently "too close" to the page we want to mark as a guard then the call to VirtualProtect to set // PAGE_GUARD will fail, but it won't return an error. Therefore, we protect the page, then query it to make // sure it worked. If it didn't, we try the next page down. We'll either find a page to protect, or run into // the guard region and rip the process down with EEPOLICY_HANDLE_FATAL_ERROR below. guardRegionThreshold = GetLastNormalStackAddress(); pageMissing = TRUE; while (pageMissing) { LOG((LF_EH, LL_INFO10000, "Thread::RestoreGuardPage: restoring guard page @ 0x%p, approxStackPointer=0x%p, " "last normal stack address=0x%p\n", guardPageBase, approxStackPointer, guardRegionThreshold)); // Make sure we set the guard page above the guard region. if (guardPageBase < guardRegionThreshold) { goto lFatalError; } if (MarkPageAsGuard(guardPageBase)) { // The current GuardPage should be beyond the current SP. _ASSERTE (guardPageBase < approxStackPointer); pageMissing = FALSE; } else { guardPageBase -= GetOsPageSize(); } } } FinishSOWork(); INDEBUG(DebugLogStackMBIs()); return; lFatalError: STRESS_LOG2(LF_EH, LL_ALWAYS, "Thread::RestoreGuardPage: too close to the guard region (0x%p) to restore guard page @0x%p\n", guardRegionThreshold, guardPageBase); _ASSERTE(!"Too close to the guard page to reset it!"); EEPOLICY_HANDLE_FATAL_ERROR(COR_E_STACKOVERFLOW); } #endif // !FEATURE_PAL #endif // #ifndef DACCESS_COMPILE // // InitRegDisplay: initializes a REGDISPLAY for a thread. If validContext // is false, pRD is filled from the current context of the thread. The // thread's current context is also filled in pctx. If validContext is true, // pctx should point to a valid context and pRD is filled from that. // bool Thread::InitRegDisplay(const PREGDISPLAY pRD, PT_CONTEXT pctx, bool validContext) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; if (!validContext) { if (GetFilterContext()!= NULL) { pctx = GetFilterContext(); } else { #ifdef DACCESS_COMPILE DacNotImpl(); #else pctx->ContextFlags = CONTEXT_FULL; _ASSERTE(this != GetThread()); // do not call GetThreadContext on the active thread BOOL ret = EEGetThreadContext(this, pctx); if (!ret) { SetIP(pctx, 0); #ifdef _TARGET_X86_ pRD->ControlPC = pctx->Eip; pRD->PCTAddr = (TADDR)&(pctx->Eip); #elif defined(_TARGET_AMD64_) // nothing more to do here, on Win64 setting the IP to 0 is enough. #elif defined(_TARGET_ARM_) // nothing more to do here, on Win64 setting the IP to 0 is enough. #else PORTABILITY_ASSERT("NYI for platform Thread::InitRegDisplay"); #endif return false; } #endif // DACCESS_COMPILE } } FillRegDisplay( pRD, pctx ); return true; } void Thread::FillRegDisplay(const PREGDISPLAY pRD, PT_CONTEXT pctx) { WRAPPER_NO_CONTRACT; SUPPORTS_DAC; ::FillRegDisplay(pRD, pctx); #if defined(DEBUG_REGDISPLAY) && !defined(_TARGET_X86_) CONSISTENCY_CHECK(!pRD->_pThread || pRD->_pThread == this); pRD->_pThread = this; CheckRegDisplaySP(pRD); #endif // defined(DEBUG_REGDISPLAY) && !defined(_TARGET_X86_) } #ifdef DEBUG_REGDISPLAY void CheckRegDisplaySP (REGDISPLAY *pRD) { if (pRD->SP && pRD->_pThread) { #ifndef NO_FIXED_STACK_LIMIT _ASSERTE(PTR_VOID(pRD->SP) >= pRD->_pThread->GetCachedStackLimit()); #endif // NO_FIXED_STACK_LIMIT _ASSERTE(PTR_VOID(pRD->SP) < pRD->_pThread->GetCachedStackBase()); } } #endif // DEBUG_REGDISPLAY // Trip Functions // ============== // When a thread reaches a safe place, it will rendezvous back with us, via one of // the following trip functions: void CommonTripThread() { #ifndef DACCESS_COMPILE CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; Thread *thread = GetThread(); thread->HandleThreadAbort (); if (thread->CatchAtSafePoint()) { _ASSERTE(!ThreadStore::HoldingThreadStore(thread)); #ifdef FEATURE_HIJACK thread->UnhijackThread(); #endif // FEATURE_HIJACK // Trap thread->PulseGCMode(); } #else DacNotImpl(); #endif // #ifndef DACCESS_COMPILE } #ifndef DACCESS_COMPILE void Thread::SetFilterContext(CONTEXT *pContext) { // SetFilterContext is like pushing a Frame onto the Frame chain. CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; // Absolutely must be in coop to coordinate w/ Runtime suspension. PRECONDITION(GetThread() == this); // must be on current thread. } CONTRACTL_END; m_debuggerFilterContext = pContext; } #endif // #ifndef DACCESS_COMPILE T_CONTEXT *Thread::GetFilterContext(void) { LIMITED_METHOD_DAC_CONTRACT; return m_debuggerFilterContext; } #ifndef DACCESS_COMPILE // @todo - eventually complete remove the CantStop count on the thread and use // the one in the PreDef block. For now, we increment both our thread counter, // and the FLS counter. Eventually we can remove our thread counter and only use // the FLS counter. void Thread::SetDebugCantStop(bool fCantStop) { LIMITED_METHOD_CONTRACT; if (fCantStop) { IncCantStopCount(); m_debuggerCantStop++; } else { DecCantStopCount(); m_debuggerCantStop--; } } // @todo - remove this, we only read this from oop. bool Thread::GetDebugCantStop(void) { LIMITED_METHOD_CONTRACT; return m_debuggerCantStop != 0; } //----------------------------------------------------------------------------- // Call w/a wrapper. // We've already transitioned AppDomains here. This just places a 1st-pass filter to sniff // for catch-handler found callbacks for the debugger. //----------------------------------------------------------------------------- void MakeADCallDebuggerWrapper( FPAPPDOMAINCALLBACK fpCallback, CtxTransitionBaseArgs * args, ContextTransitionFrame* pFrame) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_MODE_ANY; BYTE * pCatcherStackAddr = (BYTE*) pFrame; struct Param : NotifyOfCHFFilterWrapperParam { FPAPPDOMAINCALLBACK fpCallback; CtxTransitionBaseArgs *args; } param; param.pFrame = pCatcherStackAddr; param.fpCallback = fpCallback; param.args = args; PAL_TRY(Param *, pParam, ¶m) { pParam->fpCallback(pParam->args); } PAL_EXCEPT_FILTER(AppDomainTransitionExceptionFilter) { // Should never reach here b/c handler should always continue search. _ASSERTE(false); } PAL_ENDTRY } // Invoke a callback in another appdomain. // Caller should have checked that we're actually transitioning domains here. void MakeCallWithAppDomainTransition( ADID TargetDomain, FPAPPDOMAINCALLBACK fpCallback, CtxTransitionBaseArgs * args) { DEBUG_ASSURE_NO_RETURN_BEGIN(MAKECALL) Thread* _ctx_trans_pThread = GetThread(); TESTHOOKCALL(EnteringAppDomain((TargetDomain.m_dwId))); AppDomain* pTargetDomain = SystemDomain::GetAppDomainFromId(TargetDomain, ADV_CURRENTAD); _ASSERTE(_ctx_trans_pThread != NULL); _ASSERTE(_ctx_trans_pThread->GetDomain()->GetId()!= TargetDomain); bool _ctx_trans_fRaiseNeeded = false; Exception* _ctx_trans_pTargetDomainException=NULL; \ FrameWithCookie _ctx_trans_Frame; ContextTransitionFrame* _ctx_trans_pFrame = &_ctx_trans_Frame; args->pCtxFrame = _ctx_trans_pFrame; TESTHOOKCALL(EnteredAppDomain((TargetDomain.m_dwId))); /* work around unreachable code warning */ EX_TRY { // Invoke the callback if (CORDebuggerAttached()) { // If a debugger is attached, do it through a wrapper that will sniff for CHF callbacks. MakeADCallDebuggerWrapper(fpCallback, args, GET_CTX_TRANSITION_FRAME()); } else { // If no debugger is attached, call directly. fpCallback(args); } } EX_CATCH { LOG((LF_EH|LF_APPDOMAIN, LL_INFO1000, "ENTER_DOMAIN(%s, %s, %d): exception in flight\n", __FUNCTION__, __FILE__, __LINE__)); _ctx_trans_pTargetDomainException=EXTRACT_EXCEPTION(); _ctx_trans_fRaiseNeeded = true; } /* SwallowAllExceptions is fine because we don't get to this point */ /* unless fRaiseNeeded = true or no exception was thrown */ EX_END_CATCH(SwallowAllExceptions); TESTHOOKCALL(LeavingAppDomain((TargetDomain.m_dwId))); if (_ctx_trans_fRaiseNeeded) { LOG((LF_EH, LL_INFO1000, "RaiseCrossContextException(%s, %s, %d)\n", __FUNCTION__, __FILE__, __LINE__)); _ctx_trans_pThread->RaiseCrossContextException(_ctx_trans_pTargetDomainException,_ctx_trans_pFrame); } LOG((LF_APPDOMAIN, LL_INFO1000, "LEAVE_DOMAIN(%s, %s, %d)\n", __FUNCTION__, __FILE__, __LINE__)); #ifdef FEATURE_TESTHOOKS TESTHOOKCALL(LeftAppDomain(TargetDomain.m_dwId)); #endif DEBUG_ASSURE_NO_RETURN_END(MAKECALL) } void Thread::InitContext() { CONTRACTL { THROWS; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; // this should only be called when initializing a thread _ASSERTE(m_pDomain == NULL); GCX_COOP_NO_THREAD_BROKEN(); m_pDomain = SystemDomain::System()->DefaultDomain(); _ASSERTE(m_pDomain); m_pDomain->ThreadEnter(this, NULL); } void Thread::ClearContext() { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} } CONTRACTL_END; if (!m_pDomain) return; m_pDomain->ThreadExit(this, NULL); // must set exposed context to null first otherwise object verification // checks will fail AV when m_Context is null m_pDomain = NULL; #ifdef FEATURE_COMINTEROP m_fDisableComObjectEagerCleanup = false; #endif //FEATURE_COMINTEROP } void DECLSPEC_NORETURN Thread::RaiseCrossContextException(Exception* pExOrig, ContextTransitionFrame* pFrame) { CONTRACTL { THROWS; WRAPPER(GC_TRIGGERS); } CONTRACTL_END; // pEx is NULL means that the exception is CLRLastThrownObjectException CLRLastThrownObjectException lastThrown; Exception* pException = pExOrig ? pExOrig : &lastThrown; COMPlusThrow(CLRException::GetThrowableFromException(pException)); } struct FindADCallbackType { AppDomain *pSearchDomain; AppDomain *pPrevDomain; Frame *pFrame; int count; enum TargetTransition {fFirstTransitionInto, fMostRecentTransitionInto} fTargetTransition; FindADCallbackType() : pSearchDomain(NULL), pPrevDomain(NULL), pFrame(NULL) { LIMITED_METHOD_CONTRACT; } }; StackWalkAction StackWalkCallback_FindAD(CrawlFrame* pCF, void* data) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; FindADCallbackType *pData = (FindADCallbackType *)data; Frame *pFrame = pCF->GetFrame(); if (!pFrame) return SWA_CONTINUE; AppDomain *pReturnDomain = pFrame->GetReturnDomain(); if (!pReturnDomain || pReturnDomain == pData->pPrevDomain) return SWA_CONTINUE; LOG((LF_APPDOMAIN, LL_INFO100, "StackWalkCallback_FindAD transition frame %8.8x into AD [%d]\n", pFrame, pReturnDomain->GetId().m_dwId)); if (pData->pPrevDomain == pData->pSearchDomain) { ++pData->count; // this is a transition into the domain we are unloading, so save it in case it is the first pData->pFrame = pFrame; if (pData->fTargetTransition == FindADCallbackType::fMostRecentTransitionInto) return SWA_ABORT; // only need to find last transition, so bail now } pData->pPrevDomain = pReturnDomain; return SWA_CONTINUE; } // This determines if a thread is running in the given domain at any point on the stack Frame *Thread::IsRunningIn(AppDomain *pDomain, int *count) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; FindADCallbackType fct; fct.pSearchDomain = pDomain; if (!fct.pSearchDomain) return FALSE; // set prev to current so if are currently running in the target domain, // we will detect the transition fct.pPrevDomain = m_pDomain; fct.fTargetTransition = FindADCallbackType::fMostRecentTransitionInto; fct.count = 0; // when this returns, if there is a transition into the AD, it will be in pFirstFrame StackWalkAction res; res = StackWalkFrames(StackWalkCallback_FindAD, (void*) &fct, ALLOW_ASYNC_STACK_WALK); if (count) *count = fct.count; return fct.pFrame; } // This finds the very first frame on the stack where the thread transitioned into the given domain Frame *Thread::GetFirstTransitionInto(AppDomain *pDomain, int *count) { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; FindADCallbackType fct; fct.pSearchDomain = pDomain; // set prev to current so if are currently running in the target domain, // we will detect the transition fct.pPrevDomain = m_pDomain; fct.fTargetTransition = FindADCallbackType::fFirstTransitionInto; fct.count = 0; // when this returns, if there is a transition into the AD, it will be in pFirstFrame StackWalkAction res; res = StackWalkFrames(StackWalkCallback_FindAD, (void*) &fct, ALLOW_ASYNC_STACK_WALK); if (count) *count = fct.count; return fct.pFrame; } BOOL Thread::HaveExtraWorkForFinalizer() { LIMITED_METHOD_CONTRACT; return m_ThreadTasks || ThreadpoolMgr::HaveTimerInfosToFlush() || ExecutionManager::IsCacheCleanupRequired() || Thread::CleanupNeededForFinalizedThread() || (m_DetachCount > 0) || SystemDomain::System()->RequireAppDomainCleanup() || ThreadStore::s_pThreadStore->ShouldTriggerGCForDeadThreads(); } void Thread::DoExtraWorkForFinalizer() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE(GetThread() == this); _ASSERTE(this == FinalizerThread::GetFinalizerThread()); #ifdef FEATURE_COMINTEROP_APARTMENT_SUPPORT if (RequiresCoInitialize()) { SetApartment(AS_InMTA, FALSE); } #endif // FEATURE_COMINTEROP_APARTMENT_SUPPORT if (RequireSyncBlockCleanup()) { #ifndef FEATURE_PAL InteropSyncBlockInfo::FlushStandbyList(); #endif // !FEATURE_PAL #ifdef FEATURE_COMINTEROP RCW::FlushStandbyList(); #endif // FEATURE_COMINTEROP SyncBlockCache::GetSyncBlockCache()->CleanupSyncBlocks(); } if (SystemDomain::System()->RequireAppDomainCleanup()) { SystemDomain::System()->ProcessDelayedUnloadLoaderAllocators(); } if(m_DetachCount > 0 || Thread::CleanupNeededForFinalizedThread()) { Thread::CleanupDetachedThreads(); } if(ExecutionManager::IsCacheCleanupRequired() && GCHeapUtilities::GetGCHeap()->GetCondemnedGeneration()>=1) { ExecutionManager::ClearCaches(); } // If there were any TimerInfos waiting to be released, they'll get flushed now ThreadpoolMgr::FlushQueueOfTimerInfos(); ThreadStore::s_pThreadStore->TriggerGCForDeadThreadsIfNecessary(); } // HELPERS FOR THE BASE OF A MANAGED THREAD, INCLUDING AD TRANSITION SUPPORT // We have numerous places where we start up a managed thread. This includes several places in the // ThreadPool, the 'new Thread(...).Start()' case, and the Finalizer. Try to factor the code so our // base exception handling behavior is consistent across those places. The resulting code is convoluted, // but it's better than the prior situation of each thread being on a different plan. // We need Middle & Outer methods for the usual problem of combining C++ & SEH. /* The effect of all this is that we get: Base of thread -- OS unhandled exception filter that we hook SEH handler from DispatchOuter C++ handler from DispatchMiddle And if there is an AppDomain transition before we call back to user code, we additionally get: AppDomain transition -- contains its own handlers to terminate the first pass and marshal the exception. SEH handler from DispatchOuter C++ handler from DispatchMiddle Regardless of whether or not there is an AppDomain transition, we then have: User code that obviously can throw. So if we don't have an AD transition, or we take a fault before we successfully transition the AppDomain, then the base-most DispatchOuter/Middle will deal with the exception. This may involve swallowing exceptions or it may involve Watson & debugger attach. It will always involve notifications to any AppDomain.UnhandledException event listeners. But if we did transition the AppDomain, then any Watson, debugger attach and UnhandledException events will occur in that AppDomain in the initial first pass. So we get a good debugging experience and we get notifications to the host that show which AppDomain is allowing exceptions to go unhandled (so perhaps it can be unloaded or otherwise dealt with). The trick is that if the exception goes unhandled at the process level, we would normally try to fire AppDomain events and display the faulting exception on the console from two more places. These are the base-most DispatchOuter/Middle pair and the hook of the OS unhandled exception handler at the base of the thread. This is redundant and messy. (There's no concern with getting a 2nd Watson because we only do one of these per process anyway). The solution for the base-most DispatchOuter/Middle is to use the ManagedThreadCallState.flags to control whether the exception has already been dealt with or not. These flags cause the ThreadBaseRedirectingFilter to either do normal "base of the thread" exception handling, or to ignore the exception because it has already been reported in the AppDomain we transitioned to. But turning off the reporting in the OS unhandled exception filter is harder. We don't want to flip a bit on the Thread to disable this, unless we can be sure we are only disabling something we already reported, and that this thread will never recover from that situation and start executing code again. Here's the normal nightmare scenario with SEH: 1) exception of type A is thrown 2) All the filters in the 1st pass say they don't want an A 3) The exception gets all the way out and is considered unhandled. We report this "fact". 4) Imagine we then set a bit that says this thread shouldn't report unhandled exceptions. 5) The 2nd pass starts. 6) Inside a finally, someone throws an exception of type B. 7) A new 1st pass starts from the point of the throw, with a type B. 8) Now a filter says "Yes, I will swallow exception B." 9) We no longer have an unhandled exception, and execution continues merrily. This is an unavoidable consequence of the 2-pass model. If you report unhandled exceptions in the 1st pass (for good debugging), you might find that this was premature and you don't have an unhandled exception when you get to the 2nd pass. But it would not be optimal if in step 4 we set a bit that says we should suppress normal notifications and reporting on this thread, believing that the process will terminate. The solution is to recognize that the base OS unhandled exception filter runs in two modes. In the first mode, it operates as today and serves as our backstop. In the second mode it is fully redundant with the handlers pushed after the AppDomain transition, which are completely containing the exception to the AD that it occurred in (for purposes of reporting). So we just need a flag on the thread that says whether or not that set of handlers are pushed and functioning. That flag enables / disables the base exception reporting and is called TSNC_AppDomainContainUnhandled */ enum ManagedThreadCallStateFlags { MTCSF_NormalBase, MTCSF_ContainToAppDomain, MTCSF_SuppressDuplicate, }; struct ManagedThreadCallState { ADID pAppDomainId; AppDomain* pUnsafeAppDomain; BOOL bDomainIsAsID; ADCallBackFcnType pTarget; LPVOID args; UnhandledExceptionLocation filterType; ManagedThreadCallStateFlags flags; BOOL IsAppDomainEqual(AppDomain* pApp) { LIMITED_METHOD_CONTRACT; return bDomainIsAsID?(pApp->GetId()==pAppDomainId):(pUnsafeAppDomain==pApp); } ManagedThreadCallState(ADID AppDomainId,ADCallBackFcnType Target,LPVOID Args, UnhandledExceptionLocation FilterType, ManagedThreadCallStateFlags Flags): pAppDomainId(AppDomainId), pUnsafeAppDomain(NULL), bDomainIsAsID(TRUE), pTarget(Target), args(Args), filterType(FilterType), flags(Flags) { LIMITED_METHOD_CONTRACT; }; protected: ManagedThreadCallState(AppDomain* AppDomain,ADCallBackFcnType Target,LPVOID Args, UnhandledExceptionLocation FilterType, ManagedThreadCallStateFlags Flags): pAppDomainId(ADID(0)), pUnsafeAppDomain(AppDomain), bDomainIsAsID(FALSE), pTarget(Target), args(Args), filterType(FilterType), flags(Flags) { LIMITED_METHOD_CONTRACT; }; void InitForFinalizer(AppDomain* AppDomain,ADCallBackFcnType Target,LPVOID Args) { LIMITED_METHOD_CONTRACT; filterType=FinalizerThread; pUnsafeAppDomain=AppDomain; pTarget=Target; args=Args; }; friend void ManagedThreadBase_NoADTransition(ADCallBackFcnType pTarget, UnhandledExceptionLocation filterType); friend void ManagedThreadBase::FinalizerAppDomain(AppDomain* pAppDomain, ADCallBackFcnType pTarget, LPVOID args, ManagedThreadCallState *pTurnAround); }; // The following static helpers are outside of the ManagedThreadBase struct because I // don't want to change threads.h whenever I change the mechanism for how unhandled // exceptions works. The ManagedThreadBase struct is for the public exposure of the // API only. static void ManagedThreadBase_DispatchOuter(ManagedThreadCallState *pCallState); static void ManagedThreadBase_DispatchInner(ManagedThreadCallState *pCallState) { CONTRACTL { GC_TRIGGERS; THROWS; MODE_COOPERATIVE; } CONTRACTL_END; // Go ahead and dispatch the call. (*pCallState->pTarget) (pCallState->args); } static void ManagedThreadBase_DispatchMiddle(ManagedThreadCallState *pCallState) { STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_THROWS; STATIC_CONTRACT_MODE_COOPERATIVE; STATIC_CONTRACT_SO_TOLERANT; // We have the probe outside the EX_TRY below since corresponding EX_CATCH // also invokes SO_INTOLERANT code. BEGIN_SO_INTOLERANT_CODE(GetThread()); EX_TRY_CPP_ONLY { // During an unwind, we have some cleanup: // // 1) We should no longer suppress any unhandled exception reporting at the base // of the thread, because any handler that contained the exception to the AppDomain // where it occurred is now being removed from the stack. // // 2) We need to unwind the Frame chain. We cannot do it when we get to the __except clause // because at this point we are in the 2nd phase and the stack has been popped. Any // stack crawling from another thread will see a frame chain in a popped region of stack. // Nor can we pop it in a filter, since this would destroy all the stack-walking information // we need to perform the 2nd pass. So doing it in a C++ destructor will ensure it happens // during the 2nd pass but before the stack is actually popped. class Cleanup { Frame *m_pEntryFrame; Thread *m_pThread; public: Cleanup(Thread* pThread) { m_pThread = pThread; m_pEntryFrame = pThread->m_pFrame; } ~Cleanup() { GCX_COOP(); m_pThread->SetFrame(m_pEntryFrame); m_pThread->ResetThreadStateNC(Thread::TSNC_AppDomainContainUnhandled); } }; Cleanup cleanup(GetThread()); ManagedThreadBase_DispatchInner(pCallState); } EX_CATCH_CPP_ONLY { GCX_COOP(); Exception *pException = GET_EXCEPTION(); // RudeThreadAbort is a pre-allocated instance of ThreadAbort. So the following is sufficient. // For Whidbey, by default only swallow certain exceptions. If reverting back to Everett's // behavior (swallowing all unhandled exception), then swallow all unhandled exception. // if (SwallowUnhandledExceptions() || IsExceptionOfType(kThreadAbortException, pException) || IsExceptionOfType(kAppDomainUnloadedException, pException)) { // Do nothing to swallow the exception } else { // Setting up the unwind_and_continue_handler ensures that C++ exceptions do not leak out. // An example is when Thread1 in Default AppDomain creates AppDomain2, enters it, creates // another thread T2 and T2 throws OOM exception (that goes unhandled). At the transition // boundary, END_DOMAIN_TRANSITION will catch it and invoke RaiseCrossContextException // that will rethrow the OOM as a C++ exception. // // Without unwind_and_continue_handler below, the exception will fly up the stack to // this point, where it will be rethrown and thus leak out. INSTALL_UNWIND_AND_CONTINUE_HANDLER; EX_RETHROW; UNINSTALL_UNWIND_AND_CONTINUE_HANDLER; } } EX_END_CATCH(SwallowAllExceptions); END_SO_INTOLERANT_CODE; } /* typedef struct Param { ManagedThreadCallState * m_pCallState; Frame * m_pFrame; Param(ManagedThreadCallState * pCallState, Frame * pFrame): m_pCallState(pCallState), m_pFrame(pFrame) {} } TryParam; */ typedef struct Param: public NotifyOfCHFFilterWrapperParam { ManagedThreadCallState * m_pCallState; Param(ManagedThreadCallState * pCallState): m_pCallState(pCallState) {} } TryParam; // Dispatch to the appropriate filter, based on the active CallState. static LONG ThreadBaseRedirectingFilter(PEXCEPTION_POINTERS pExceptionInfo, LPVOID pParam) { STATIC_CONTRACT_THROWS; STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_MODE_ANY; LONG (*ptrFilter) (PEXCEPTION_POINTERS, PVOID); TryParam * pRealParam = reinterpret_cast(pParam); ManagedThreadCallState * _pCallState = pRealParam->m_pCallState; ManagedThreadCallStateFlags flags = _pCallState->flags; if (flags == MTCSF_SuppressDuplicate) { LOG((LF_EH, LL_INFO100, "ThreadBaseRedirectingFilter: setting TSNC_AppDomainContainUnhandled\n")); GetThread()->SetThreadStateNC(Thread::TSNC_AppDomainContainUnhandled); return EXCEPTION_CONTINUE_SEARCH; } LONG ret = -1; BEGIN_SO_INTOLERANT_CODE_NO_THROW_CHECK_THREAD(return EXCEPTION_CONTINUE_SEARCH;); // This will invoke the swallowing filter. If that returns EXCEPTION_CONTINUE_SEARCH, // it will trigger unhandled exception processing. ptrFilter = ThreadBaseExceptionAppDomainFilter; // WARNING - ptrFilter may not return // This occurs when the debugger decides to intercept an exception and catch it in a frame closer // to the leaf than the one executing this filter ret = (*ptrFilter) (pExceptionInfo, _pCallState); // Although EXCEPTION_EXECUTE_HANDLER can also be returned in cases corresponding to // unhandled exceptions, all of those cases have already notified the debugger of an unhandled // exception which prevents a second notification indicating the exception was caught if (ret == EXCEPTION_EXECUTE_HANDLER) { // WARNING - NotifyOfCHFFilterWrapper may not return // This occurs when the debugger decides to intercept an exception and catch it in a frame closer // to the leaf than the one executing this filter NotifyOfCHFFilterWrapper(pExceptionInfo, pRealParam); } // If we are containing unhandled exceptions to the AppDomain we transitioned into, and the // exception is coming out, then this exception is going unhandled. We have already done // Watson and managed events, so suppress all filters below us. Otherwise we are swallowing // it and returning out of the AppDomain. if (flags == MTCSF_ContainToAppDomain) { if(ret == EXCEPTION_CONTINUE_SEARCH) { _pCallState->flags = MTCSF_SuppressDuplicate; } else if(ret == EXCEPTION_EXECUTE_HANDLER) { _pCallState->flags = MTCSF_NormalBase; } // else if( EXCEPTION_CONTINUE_EXECUTION ) do nothing } // Get the reference to the current thread.. Thread *pCurThread = GetThread(); _ASSERTE(pCurThread); if (flags == MTCSF_ContainToAppDomain) { if (((ManagedThreadCallState *) _pCallState)->flags == MTCSF_SuppressDuplicate) { // Set the flag that we have done unhandled exception processing // for this managed thread that started in a non-default domain LOG((LF_EH, LL_INFO100, "ThreadBaseRedirectingFilter: setting TSNC_AppDomainContainUnhandled\n")); pCurThread->SetThreadStateNC(Thread::TSNC_AppDomainContainUnhandled); } } else { _ASSERTE(flags == MTCSF_NormalBase); LOG((LF_EH, LL_INFO100, "ThreadBaseRedirectingFilter: setting TSNC_ProcessedUnhandledException\n")); // // In the default domain, when an exception goes unhandled on a managed thread whose threadbase is in the VM (e.g. explicitly spawned threads, // ThreadPool threads, finalizer thread, etc), CLR can end up in the unhandled exception processing path twice. // // The first attempt to perform UE processing happens at the managed thread base (via this function). When it completes, // we will set TSNC_ProcessedUnhandledException state against the thread to indicate that we have perform the unhandled exception processing. // // On the desktop CLR, after the first attempt, we will return back to the OS with EXCEPTION_CONTINUE_SEARCH as unhandled exceptions cannot be swallowed. When the exception reaches // the native threadbase in the OS kernel, the OS will invoke the UEF registered for the process. This can result in CLR's UEF (COMUnhandledExceptionFilter) // getting invoked that will attempt to perform UE processing yet again for the same thread. To avoid this duplicate processing, we check the presence of // TSNC_ProcessedUnhandledException state on the thread and if present, we simply return back to the OS. // // On desktop CoreCLR, we will only do UE processing once (at the managed threadbase) since no thread is created in default domain - all are created and executed in non-default domain. // As a result, we go via completely different codepath that prevents duplication of UE processing from happening, especially since desktop CoreCLR is targetted for SL and SL // always passes us a flag to swallow unhandled exceptions. // // On CoreSys CoreCLR, the host can ask CoreCLR to run all code in the default domain. As a result, when we return from the first attempt to perform UE // processing, the call could return back with EXCEPTION_EXECUTE_HANDLER since, like desktop CoreCLR is instructed by SL host to swallow all unhandled exceptions, // CoreSys CoreCLR can also be instructed by its Phone host to swallow all unhandled exceptions. As a result, the exception dispatch will never continue to go upstack // to the native threadbase in the OS kernel and thus, there will never be a second attempt to perform UE processing. Hence, we dont, and shouldnt, need to set // TSNC_ProcessedUnhandledException state against the thread if we are in SingleAppDomain mode and have been asked to swallow the exception. // // If we continue to set TSNC_ProcessedUnhandledException and a ThreadPool Thread A has an exception go unhandled, we will swallow it correctly for the first time. // The next time Thread A has an exception go unhandled, our UEF will see TSNC_ProcessedUnhandledException set and assume (incorrectly) UE processing has happened and // will fail to honor the host policy (e.g. swallow unhandled exception). Thus, the 2nd unhandled exception may end up crashing the app when it should not. // if (ret != EXCEPTION_EXECUTE_HANDLER) { // Since we have already done unhandled exception processing for it, we dont want it // to happen again if our UEF gets invoked upon returning back to the OS. // // Set the flag to indicate so. pCurThread->SetThreadStateNC(Thread::TSNC_ProcessedUnhandledException); } } END_SO_INTOLERANT_CODE; return ret; } static void ManagedThreadBase_DispatchOuter(ManagedThreadCallState *pCallState) { STATIC_CONTRACT_GC_TRIGGERS; STATIC_CONTRACT_THROWS; STATIC_CONTRACT_MODE_COOPERATIVE; // HasStarted() must have already been performed by our caller _ASSERTE(GetThread() != NULL); Thread *pThread = GetThread(); #ifdef WIN64EXCEPTIONS Frame *pFrame = pThread->m_pFrame; #endif // WIN64EXCEPTIONS // The sole purpose of having this frame is to tell the debugger that we have a catch handler here // which may swallow managed exceptions. The debugger needs this in order to send a // CatchHandlerFound (CHF) notification. FrameWithCookie catchFrame; TryParam param(pCallState); param.pFrame = &catchFrame; struct TryArgs { TryParam *pTryParam; Thread *pThread; BOOL *pfHadException; #ifdef WIN64EXCEPTIONS Frame *pFrame; #endif // WIN64EXCEPTIONS }args; args.pTryParam = ¶m; args.pThread = pThread; BOOL fHadException = TRUE; args.pfHadException = &fHadException; #ifdef WIN64EXCEPTIONS args.pFrame = pFrame; #endif // WIN64EXCEPTIONS PAL_TRY(TryArgs *, pArgs, &args) { PAL_TRY(TryParam *, pParam, pArgs->pTryParam) { ManagedThreadBase_DispatchMiddle(pParam->m_pCallState); } PAL_EXCEPT_FILTER(ThreadBaseRedirectingFilter) { // Note: one of our C++ exceptions will never reach this filter because they're always caught by // the EX_CATCH in ManagedThreadBase_DispatchMiddle(). // // If eCLRDeterminedPolicy, we only swallow for TA, RTA, and ADU exception. // For eHostDeterminedPolicy, we will swallow all the managed exception. #ifdef WIN64EXCEPTIONS // this must be done after the second pass has run, it does not // reference anything on the stack, so it is safe to run in an // SEH __except clause as well as a C++ catch clause. ExceptionTracker::PopTrackers(pArgs->pFrame); #endif // WIN64EXCEPTIONS // Fortunately, ThreadAbortExceptions are always if (pArgs->pThread->IsAbortRequested()) pArgs->pThread->EEResetAbort(Thread::TAR_Thread); } PAL_ENDTRY; *(pArgs->pfHadException) = FALSE; } PAL_FINALLY { catchFrame.Pop(); } PAL_ENDTRY; } // For the implementation, there are three variants of work possible: // 1. Establish the base of a managed thread, and switch to the correct AppDomain. static void ManagedThreadBase_FullTransitionWithAD(ADID pAppDomain, ADCallBackFcnType pTarget, LPVOID args, UnhandledExceptionLocation filterType) { CONTRACTL { GC_TRIGGERS; THROWS; MODE_COOPERATIVE; } CONTRACTL_END; ManagedThreadCallState CallState(pAppDomain, pTarget, args, filterType, MTCSF_NormalBase); ManagedThreadBase_DispatchOuter(&CallState); } // 2. Establish the base of a managed thread, but the AppDomain transition must be // deferred until later. void ManagedThreadBase_NoADTransition(ADCallBackFcnType pTarget, UnhandledExceptionLocation filterType) { CONTRACTL { GC_TRIGGERS; THROWS; MODE_COOPERATIVE; } CONTRACTL_END; AppDomain *pAppDomain = GetAppDomain(); ManagedThreadCallState CallState(pAppDomain, pTarget, NULL, filterType, MTCSF_NormalBase); // self-describing, to create a pTurnAround data for eventual delivery to a subsequent AppDomain // transition. CallState.args = &CallState; ManagedThreadBase_DispatchOuter(&CallState); } // And here are the various exposed entrypoints for base thread behavior // The 'new Thread(...).Start()' case from COMSynchronizable kickoff thread worker void ManagedThreadBase::KickOff(ADID pAppDomain, ADCallBackFcnType pTarget, LPVOID args) { WRAPPER_NO_CONTRACT; ManagedThreadBase_FullTransitionWithAD(pAppDomain, pTarget, args, ManagedThread); } // The IOCompletion, QueueUserWorkItem, AddTimer, RegisterWaitForSingleObject cases in the ThreadPool void ManagedThreadBase::ThreadPool(ADID pAppDomain, ADCallBackFcnType pTarget, LPVOID args) { WRAPPER_NO_CONTRACT; ManagedThreadBase_FullTransitionWithAD(pAppDomain, pTarget, args, ThreadPoolThread); } // The Finalizer thread establishes exception handling at its base, but defers all the AppDomain // transitions. void ManagedThreadBase::FinalizerBase(ADCallBackFcnType pTarget) { WRAPPER_NO_CONTRACT; ManagedThreadBase_NoADTransition(pTarget, FinalizerThread); } void ManagedThreadBase::FinalizerAppDomain(AppDomain *pAppDomain, ADCallBackFcnType pTarget, LPVOID args, ManagedThreadCallState *pTurnAround) { WRAPPER_NO_CONTRACT; pTurnAround->InitForFinalizer(pAppDomain,pTarget,args); _ASSERTE(pTurnAround->flags == MTCSF_NormalBase); ManagedThreadBase_DispatchInner(pTurnAround); } //+---------------------------------------------------------------------------- // // Method: Thread::GetStaticFieldAddress private // // Synopsis: Get the address of the field relative to the current thread. // If an address has not been assigned yet then create one. // //+---------------------------------------------------------------------------- LPVOID Thread::GetStaticFieldAddress(FieldDesc *pFD) { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; _ASSERTE(pFD != NULL); _ASSERTE(pFD->IsThreadStatic()); _ASSERTE(!pFD->IsRVA()); // for static field the MethodTable is exact even for generic classes MethodTable *pMT = pFD->GetEnclosingMethodTable(); // We need to make sure that the class has been allocated, however // we should not call the class constructor ThreadStatics::GetTLM(pMT)->EnsureClassAllocated(pMT); PTR_BYTE base = NULL; if (pFD->GetFieldType() == ELEMENT_TYPE_CLASS || pFD->GetFieldType() == ELEMENT_TYPE_VALUETYPE) { base = pMT->GetGCThreadStaticsBasePointer(); } else { base = pMT->GetNonGCThreadStaticsBasePointer(); } _ASSERTE(base != NULL); DWORD offset = pFD->GetOffset(); _ASSERTE(offset <= FIELD_OFFSET_LAST_REAL_OFFSET); LPVOID result = (LPVOID)((PTR_BYTE)base + (DWORD)offset); // For value classes, the handle points at an OBJECTREF // which holds the boxed value class, so derefernce and unbox. if (pFD->GetFieldType() == ELEMENT_TYPE_VALUETYPE) { OBJECTREF obj = ObjectToOBJECTREF(*(Object**) result); result = obj->GetData(); } return result; } #endif // #ifndef DACCESS_COMPILE //+---------------------------------------------------------------------------- // // Method: Thread::GetStaticFieldAddrNoCreate private // // Synopsis: Get the address of the field relative to the thread. // If an address has not been assigned, return NULL. // No creating is allowed. // //+---------------------------------------------------------------------------- TADDR Thread::GetStaticFieldAddrNoCreate(FieldDesc *pFD) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SUPPORTS_DAC; } CONTRACTL_END; _ASSERTE(pFD != NULL); _ASSERTE(pFD->IsThreadStatic()); // for static field the MethodTable is exact even for generic classes PTR_MethodTable pMT = pFD->GetEnclosingMethodTable(); PTR_BYTE base = NULL; if (pFD->GetFieldType() == ELEMENT_TYPE_CLASS || pFD->GetFieldType() == ELEMENT_TYPE_VALUETYPE) { base = pMT->GetGCThreadStaticsBasePointer(dac_cast(this)); } else { base = pMT->GetNonGCThreadStaticsBasePointer(dac_cast(this)); } if (base == NULL) return NULL; DWORD offset = pFD->GetOffset(); _ASSERTE(offset <= FIELD_OFFSET_LAST_REAL_OFFSET); TADDR result = dac_cast(base) + (DWORD)offset; // For value classes, the handle points at an OBJECTREF // which holds the boxed value class, so derefernce and unbox. if (pFD->IsByValue()) { _ASSERTE(result != NULL); PTR_Object obj = *PTR_UNCHECKED_OBJECTREF(result); if (obj == NULL) return NULL; result = dac_cast(obj->GetData()); } return result; } #ifndef DACCESS_COMPILE // // NotifyFrameChainOfExceptionUnwind // ----------------------------------------------------------- // This method will walk the Frame chain from pStartFrame to // the last frame that is below pvLimitSP and will call each // frame's ExceptionUnwind method. It will return the first // Frame that is above pvLimitSP. // Frame * Thread::NotifyFrameChainOfExceptionUnwind(Frame* pStartFrame, LPVOID pvLimitSP) { CONTRACTL { NOTHROW; DISABLED(GC_TRIGGERS); // due to UnwindFrameChain from NOTRIGGER areas MODE_COOPERATIVE; PRECONDITION(CheckPointer(pStartFrame)); PRECONDITION(CheckPointer(pvLimitSP)); } CONTRACTL_END; Frame * pFrame; #ifdef _DEBUG // // assert that the specified Thread's Frame chain actually // contains the start Frame. // pFrame = m_pFrame; while ((pFrame != pStartFrame) && (pFrame != FRAME_TOP)) { pFrame = pFrame->Next(); } CONSISTENCY_CHECK_MSG(pFrame == pStartFrame, "pStartFrame is not on pThread's Frame chain!"); #endif // _DEBUG pFrame = pStartFrame; while (pFrame < pvLimitSP) { CONSISTENCY_CHECK(pFrame != PTR_NULL); CONSISTENCY_CHECK((pFrame) > static_cast((LPVOID)GetCurrentSP())); pFrame->ExceptionUnwind(); pFrame = pFrame->Next(); } // return the frame after the last one notified of the unwind return pFrame; } //+---------------------------------------------------------------------------- // // Method: Thread::DeleteThreadStaticData private // // Synopsis: Delete the static data for each appdomain that this thread // visited. // // //+---------------------------------------------------------------------------- void Thread::DeleteThreadStaticData() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; m_ThreadLocalBlock.FreeTable(); } //+---------------------------------------------------------------------------- // // Method: Thread::DeleteThreadStaticData public // // Synopsis: Delete the static data for the given module. This is called // when the AssemblyLoadContext unloads. // // //+---------------------------------------------------------------------------- void Thread::DeleteThreadStaticData(ModuleIndex index) { m_ThreadLocalBlock.FreeTLM(index.m_dwIndex, FALSE /* isThreadShuttingDown */); } void Thread::InitCultureAccessors() { CONTRACTL { THROWS; GC_TRIGGERS; } CONTRACTL_END; OBJECTREF *pCurrentCulture = NULL; Thread *pThread = GetThread(); GCX_COOP(); if (managedThreadCurrentCulture == NULL) { managedThreadCurrentCulture = MscorlibBinder::GetField(FIELD__THREAD__CULTURE); pCurrentCulture = (OBJECTREF*)pThread->GetStaticFieldAddress(managedThreadCurrentCulture); } if (managedThreadCurrentUICulture == NULL) { managedThreadCurrentUICulture = MscorlibBinder::GetField(FIELD__THREAD__UI_CULTURE); pCurrentCulture = (OBJECTREF*)pThread->GetStaticFieldAddress(managedThreadCurrentUICulture); } } ARG_SLOT Thread::CallPropertyGet(BinderMethodID id, OBJECTREF pObject) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; if (!pObject) { return 0; } ARG_SLOT retVal; GCPROTECT_BEGIN(pObject); MethodDescCallSite propGet(id, &pObject); // Set up the Stack. ARG_SLOT pNewArgs = ObjToArgSlot(pObject); // Make the actual call. retVal = propGet.Call_RetArgSlot(&pNewArgs); GCPROTECT_END(); return retVal; } ARG_SLOT Thread::CallPropertySet(BinderMethodID id, OBJECTREF pObject, OBJECTREF pValue) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; if (!pObject) { return 0; } ARG_SLOT retVal; GCPROTECT_BEGIN(pObject); GCPROTECT_BEGIN(pValue); MethodDescCallSite propSet(id, &pObject); // Set up the Stack. ARG_SLOT pNewArgs[] = { ObjToArgSlot(pObject), ObjToArgSlot(pValue) }; // Make the actual call. retVal = propSet.Call_RetArgSlot(pNewArgs); GCPROTECT_END(); GCPROTECT_END(); return retVal; } OBJECTREF Thread::GetCulture(BOOL bUICulture) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; FieldDesc * pFD; _ASSERTE(PreemptiveGCDisabled()); // This is the case when we're building mscorlib and haven't yet created // the system assembly. if (SystemDomain::System()->SystemAssembly()==NULL || g_fForbidEnterEE) { return NULL; } // Get the actual thread culture. OBJECTREF pCurThreadObject = GetExposedObject(); _ASSERTE(pCurThreadObject!=NULL); THREADBASEREF pThreadBase = (THREADBASEREF)(pCurThreadObject); OBJECTREF pCurrentCulture = bUICulture ? pThreadBase->GetCurrentUICulture() : pThreadBase->GetCurrentUserCulture(); if (pCurrentCulture==NULL) { GCPROTECT_BEGIN(pThreadBase); if (bUICulture) { // Call the Getter for the CurrentUICulture. This will cause it to populate the field. ARG_SLOT retVal = CallPropertyGet(METHOD__THREAD__GET_UI_CULTURE, (OBJECTREF)pThreadBase); pCurrentCulture = ArgSlotToObj(retVal); } else { //This is faster than calling the property, because this is what the call does anyway. pFD = MscorlibBinder::GetField(FIELD__CULTURE_INFO__CURRENT_CULTURE); _ASSERTE(pFD); pFD->CheckRunClassInitThrowing(); pCurrentCulture = pFD->GetStaticOBJECTREF(); _ASSERTE(pCurrentCulture!=NULL); } GCPROTECT_END(); } return pCurrentCulture; } // copy culture name into szBuffer and return length int Thread::GetParentCultureName(__out_ecount(length) LPWSTR szBuffer, int length, BOOL bUICulture) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; // This is the case when we're building mscorlib and haven't yet created // the system assembly. if (SystemDomain::System()->SystemAssembly()==NULL) { const WCHAR *tempName = W("en"); INT32 tempLength = (INT32)wcslen(tempName); _ASSERTE(length>=tempLength); memcpy(szBuffer, tempName, tempLength*sizeof(WCHAR)); return tempLength; } ARG_SLOT Result = 0; INT32 retVal=0; WCHAR *buffer=NULL; INT32 bufferLength=0; STRINGREF cultureName = NULL; GCX_COOP(); struct _gc { OBJECTREF pCurrentCulture; OBJECTREF pParentCulture; } gc; ZeroMemory(&gc, sizeof(gc)); GCPROTECT_BEGIN(gc); gc.pCurrentCulture = GetCulture(bUICulture); if (gc.pCurrentCulture != NULL) { Result = CallPropertyGet(METHOD__CULTURE_INFO__GET_PARENT, gc.pCurrentCulture); } if (Result) { gc.pParentCulture = (OBJECTREF)(ArgSlotToObj(Result)); if (gc.pParentCulture != NULL) { Result = 0; Result = CallPropertyGet(METHOD__CULTURE_INFO__GET_NAME, gc.pParentCulture); } } GCPROTECT_END(); if (Result==0) { return 0; } // Extract the data out of the String. cultureName = (STRINGREF)(ArgSlotToObj(Result)); cultureName->RefInterpretGetStringValuesDangerousForGC((WCHAR**)&buffer, &bufferLength); if (bufferLengthSystemAssembly()==NULL || g_fForbidEnterEE) { const WCHAR *tempName = W("en-US"); INT32 tempLength = (INT32)wcslen(tempName); _ASSERTE(length>=tempLength); memcpy(szBuffer, tempName, tempLength*sizeof(WCHAR)); return tempLength; } ARG_SLOT Result = 0; INT32 retVal=0; WCHAR *buffer=NULL; INT32 bufferLength=0; STRINGREF cultureName = NULL; GCX_COOP (); OBJECTREF pCurrentCulture = NULL; GCPROTECT_BEGIN(pCurrentCulture) { pCurrentCulture = GetCulture(bUICulture); if (pCurrentCulture != NULL) Result = CallPropertyGet(METHOD__CULTURE_INFO__GET_NAME, pCurrentCulture); } GCPROTECT_END(); if (Result==0) { return 0; } // Extract the data out of the String. cultureName = (STRINGREF)(ArgSlotToObj(Result)); cultureName->RefInterpretGetStringValuesDangerousForGC((WCHAR**)&buffer, &bufferLength); if (bufferLengthGetCultureId(bUICulture); } EX_CATCH { } EX_END_CATCH (SwallowAllExceptions); return (INT32)Result; } // Return a language identifier. LCID Thread::GetCultureId(BOOL bUICulture) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; // This is the case when we're building mscorlib and haven't yet created // the system assembly. if (SystemDomain::System()->SystemAssembly()==NULL || g_fForbidEnterEE) { return (LCID) -1; } LCID Result = (LCID) -1; #ifdef FEATURE_USE_LCID GCX_COOP(); OBJECTREF pCurrentCulture = NULL; GCPROTECT_BEGIN(pCurrentCulture) { pCurrentCulture = GetCulture(bUICulture); if (pCurrentCulture != NULL) Result = (LCID)CallPropertyGet(METHOD__CULTURE_INFO__GET_ID, pCurrentCulture); } GCPROTECT_END(); #endif return Result; } void Thread::SetCultureId(LCID lcid, BOOL bUICulture) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_ANY; } CONTRACTL_END; GCX_COOP(); OBJECTREF CultureObj = NULL; GCPROTECT_BEGIN(CultureObj) { // Convert the LCID into a CultureInfo. GetCultureInfoForLCID(lcid, &CultureObj); // Set the newly created culture as the thread's culture. SetCulture(&CultureObj, bUICulture); } GCPROTECT_END(); } void Thread::SetCulture(OBJECTREF *CultureObj, BOOL bUICulture) { CONTRACTL { THROWS; GC_TRIGGERS; MODE_COOPERATIVE; } CONTRACTL_END; // Retrieve the exposed thread object. OBJECTREF pCurThreadObject = GetExposedObject(); _ASSERTE(pCurThreadObject!=NULL); // Set the culture property on the thread. THREADBASEREF pThreadBase = (THREADBASEREF)(pCurThreadObject); CallPropertySet(bUICulture ? METHOD__THREAD__SET_UI_CULTURE : METHOD__THREAD__SET_CULTURE, (OBJECTREF)pThreadBase, *CultureObj); } void Thread::SetHasPromotedBytes () { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; m_fPromoted = TRUE; _ASSERTE(GCHeapUtilities::IsGCInProgress() && IsGCThread ()); if (!m_fPreemptiveGCDisabled) { if (FRAME_TOP == GetFrame()) m_fPromoted = FALSE; } } BOOL ThreadStore::HoldingThreadStore(Thread *pThread) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; if (pThread) { return (pThread == s_pThreadStore->m_HoldingThread); } else { return (s_pThreadStore->m_holderthreadid.IsCurrentThread()); } } LONG Thread::GetTotalThreadPoolCompletionCount() { CONTRACTL { NOTHROW; MODE_ANY; } CONTRACTL_END; LONG total; if (g_fEEStarted) //make sure we actually have a thread store { // make sure up-to-date thread-local counts are visible to us ::FlushProcessWriteBuffers(); // enumerate all threads, summing their local counts. ThreadStoreLockHolder tsl; total = s_threadPoolCompletionCountOverflow.Load(); Thread *pThread = NULL; while ((pThread = ThreadStore::GetAllThreadList(pThread, 0, 0)) != NULL) { total += pThread->m_threadPoolCompletionCount; } } else { total = s_threadPoolCompletionCountOverflow.Load(); } return total; } INT32 Thread::ResetManagedThreadObject(INT32 nPriority) { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; GCX_COOP(); return ResetManagedThreadObjectInCoopMode(nPriority); } INT32 Thread::ResetManagedThreadObjectInCoopMode(INT32 nPriority) { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; THREADBASEREF pObject = (THREADBASEREF)ObjectFromHandle(m_ExposedObject); if (pObject != NULL) { pObject->ResetCulture(); pObject->ResetName(); nPriority = pObject->GetPriority(); } return nPriority; } void Thread::FullResetThread() { CONTRACTL { NOTHROW; GC_TRIGGERS; } CONTRACTL_END; GCX_COOP(); // We need to put this thread in COOPERATIVE GC first to solve race between AppDomain::Unload // and Thread::Reset. AppDomain::Unload does a full GC to collect all roots in one AppDomain. // ThreadStaticData used to be coupled with a managed array of objects in the managed Thread // object, however this is no longer the case. // TODO: Do we still need to put this thread into COOP mode? GCX_FORBID(); DeleteThreadStaticData(); m_alloc_context.alloc_bytes = 0; m_fPromoted = FALSE; } BOOL Thread::IsRealThreadPoolResetNeeded() { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_COOPERATIVE; SO_TOLERANT; } CONTRACTL_END; if(!IsBackground()) return TRUE; THREADBASEREF pObject = (THREADBASEREF)ObjectFromHandle(m_ExposedObject); if(pObject != NULL) { INT32 nPriority = pObject->GetPriority(); if(nPriority != ThreadNative::PRIORITY_NORMAL) return TRUE; } return FALSE; } void Thread::InternalReset(BOOL fFull, BOOL fNotFinalizerThread, BOOL fThreadObjectResetNeeded, BOOL fResetAbort) { CONTRACTL { NOTHROW; if(!fNotFinalizerThread || fThreadObjectResetNeeded) {GC_TRIGGERS;SO_INTOLERANT;} else {GC_NOTRIGGER;SO_TOLERANT;} } CONTRACTL_END; _ASSERTE (this == GetThread()); FinishSOWork(); INT32 nPriority = ThreadNative::PRIORITY_NORMAL; if (!fNotFinalizerThread && this == FinalizerThread::GetFinalizerThread()) { nPriority = ThreadNative::PRIORITY_HIGHEST; } if(fThreadObjectResetNeeded) { nPriority = ResetManagedThreadObject(nPriority); } if (fFull) { FullResetThread(); } //m_MarshalAlloc.Collapse(NULL); if (fResetAbort && IsAbortRequested()) { UnmarkThreadForAbort(TAR_ALL); } if (fResetAbort && IsAborted()) ClearAborted(); if (IsThreadPoolThread() && fThreadObjectResetNeeded) { SetBackground(TRUE); if (nPriority != ThreadNative::PRIORITY_NORMAL) { SetThreadPriority(THREAD_PRIORITY_NORMAL); } } else if (!fNotFinalizerThread && this == FinalizerThread::GetFinalizerThread()) { SetBackground(TRUE); if (nPriority != ThreadNative::PRIORITY_HIGHEST) { SetThreadPriority(THREAD_PRIORITY_HIGHEST); } } } HRESULT Thread::Abort () { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} SO_TOLERANT; } CONTRACTL_END; BEGIN_SO_INTOLERANT_CODE_NO_THROW_CHECK_THREAD(return COR_E_STACKOVERFLOW;); EX_TRY { UserAbort(TAR_Thread, EEPolicy::TA_Safe, INFINITE, Thread::UAC_Host); } EX_CATCH { } EX_END_CATCH(SwallowAllExceptions); END_SO_INTOLERANT_CODE; return S_OK; } HRESULT Thread::RudeAbort() { CONTRACTL { NOTHROW; if (GetThread()) {GC_TRIGGERS;} else {DISABLED(GC_NOTRIGGER);} SO_TOLERANT; } CONTRACTL_END; BEGIN_SO_INTOLERANT_CODE_NO_THROW_CHECK_THREAD(return COR_E_STACKOVERFLOW); EX_TRY { UserAbort(TAR_Thread, EEPolicy::TA_Rude, INFINITE, Thread::UAC_Host); } EX_CATCH { } EX_END_CATCH(SwallowAllExceptions); END_SO_INTOLERANT_CODE; return S_OK; } HRESULT Thread::NeedsPriorityScheduling(BOOL *pbNeedsPriorityScheduling) { CONTRACTL { NOTHROW; GC_NOTRIGGER; SO_TOLERANT; } CONTRACTL_END; *pbNeedsPriorityScheduling = (m_fPreemptiveGCDisabled || (g_fEEStarted && this == FinalizerThread::GetFinalizerThread())); return S_OK; } HRESULT Thread::LocksHeld(SIZE_T *pLockCount) { LIMITED_METHOD_CONTRACT; *pLockCount = m_dwLockCount; return S_OK; } HRESULT Thread::BeginPreventAsyncAbort() { WRAPPER_NO_CONTRACT; #ifdef _DEBUG int count = #endif FastInterlockIncrement((LONG*)&m_PreventAbort); #ifdef _DEBUG ASSERT(count > 0); FastInterlockIncrement((LONG*)&m_dwDisableAbortCheckCount); #endif return S_OK; } HRESULT Thread::EndPreventAsyncAbort() { WRAPPER_NO_CONTRACT; #ifdef _DEBUG int count = #endif FastInterlockDecrement((LONG*)&m_PreventAbort); #ifdef _DEBUG ASSERT(count >= 0); FastInterlockDecrement((LONG*)&m_dwDisableAbortCheckCount); #endif return S_OK; } ULONG Thread::AddRef() { WRAPPER_NO_CONTRACT; _ASSERTE(m_ExternalRefCount > 0); _ASSERTE (m_UnmanagedRefCount != (DWORD) -1); ULONG ref = FastInterlockIncrement((LONG*)&m_UnmanagedRefCount); return ref; } ULONG Thread::Release() { WRAPPER_NO_CONTRACT; SUPPORTS_DAC_HOST_ONLY; _ASSERTE (m_ExternalRefCount > 0); _ASSERTE (m_UnmanagedRefCount > 0); ULONG ref = FastInterlockDecrement((LONG*)&m_UnmanagedRefCount); return ref; } HRESULT Thread::QueryInterface(REFIID riid, void **ppUnk) { LIMITED_METHOD_CONTRACT; return E_NOINTERFACE; } void Thread::SetupThreadForHost() { CONTRACTL { THROWS; GC_TRIGGERS; SO_TOLERANT; } CONTRACTL_END; _ASSERTE (GetThread() == this); CONTRACT_VIOLATION(SOToleranceViolation); } ETaskType GetCurrentTaskType() { STATIC_CONTRACT_NOTHROW; STATIC_CONTRACT_GC_NOTRIGGER; STATIC_CONTRACT_SO_TOLERANT; ETaskType TaskType = TT_UNKNOWN; size_t type = (size_t)ClrFlsGetValue (TlsIdx_ThreadType); if (type & ThreadType_DbgHelper) { TaskType = TT_DEBUGGERHELPER; } else if (type & ThreadType_GC) { TaskType = TT_GC; } else if (type & ThreadType_Finalizer) { TaskType = TT_FINALIZER; } else if (type & ThreadType_Timer) { TaskType = TT_THREADPOOL_TIMER; } else if (type & ThreadType_Gate) { TaskType = TT_THREADPOOL_GATE; } else if (type & ThreadType_Wait) { TaskType = TT_THREADPOOL_WAIT; } else if (type & ThreadType_Threadpool_IOCompletion) { TaskType = TT_THREADPOOL_IOCOMPLETION; } else if (type & ThreadType_Threadpool_Worker) { TaskType = TT_THREADPOOL_WORKER; } else { Thread *pThread = GetThread(); if (pThread) { TaskType = TT_USER; } } return TaskType; } DeadlockAwareLock::DeadlockAwareLock(const char *description) : m_pHoldingThread(NULL) #ifdef _DEBUG , m_description(description) #endif { LIMITED_METHOD_CONTRACT; } DeadlockAwareLock::~DeadlockAwareLock() { CONTRACTL { NOTHROW; GC_NOTRIGGER; MODE_ANY; CAN_TAKE_LOCK; } CONTRACTL_END; // Wait for another thread to leave its loop in DeadlockAwareLock::TryBeginEnterLock CrstHolder lock(&g_DeadlockAwareCrst); } CHECK DeadlockAwareLock::CheckDeadlock(Thread *pThread) { CONTRACTL { PRECONDITION(g_DeadlockAwareCrst.OwnedByCurrentThread()); NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; // Note that this check is recursive in order to produce descriptive check failure messages. Thread *pHoldingThread = m_pHoldingThread.Load(); if (pThread == pHoldingThread) { CHECK_FAILF(("Lock %p (%s) is held by thread %d", this, m_description, pThread)); } if (pHoldingThread != NULL) { DeadlockAwareLock *pBlockingLock = pHoldingThread->m_pBlockingLock.Load(); if (pBlockingLock != NULL) { CHECK_MSGF(pBlockingLock->CheckDeadlock(pThread), ("Deadlock: Lock %p (%s) is held by thread %d", this, m_description, pHoldingThread)); } } CHECK_OK; } BOOL DeadlockAwareLock::CanEnterLock() { Thread * pThread = GetThread(); CONSISTENCY_CHECK_MSG(pThread != NULL, "Cannot do deadlock detection on non-EE thread"); CONSISTENCY_CHECK_MSG(pThread->m_pBlockingLock.Load() == NULL, "Cannot block on two locks at once"); { CrstHolder lock(&g_DeadlockAwareCrst); // Look for deadlocks DeadlockAwareLock *pLock = this; while (TRUE) { Thread * holdingThread = pLock->m_pHoldingThread; if (holdingThread == pThread) { // Deadlock! return FALSE; } if (holdingThread == NULL) { // Lock is unheld break; } pLock = holdingThread->m_pBlockingLock; if (pLock == NULL) { // Thread is running free break; } } return TRUE; } } BOOL DeadlockAwareLock::TryBeginEnterLock() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; Thread * pThread = GetThread(); CONSISTENCY_CHECK_MSG(pThread != NULL, "Cannot do deadlock detection on non-EE thread"); CONSISTENCY_CHECK_MSG(pThread->m_pBlockingLock.Load() == NULL, "Cannot block on two locks at once"); { CrstHolder lock(&g_DeadlockAwareCrst); // Look for deadlocks DeadlockAwareLock *pLock = this; while (TRUE) { Thread * holdingThread = pLock->m_pHoldingThread; if (holdingThread == pThread) { // Deadlock! return FALSE; } if (holdingThread == NULL) { // Lock is unheld break; } pLock = holdingThread->m_pBlockingLock; if (pLock == NULL) { // Thread is running free break; } } pThread->m_pBlockingLock = this; } return TRUE; }; void DeadlockAwareLock::BeginEnterLock() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; Thread * pThread = GetThread(); CONSISTENCY_CHECK_MSG(pThread != NULL, "Cannot do deadlock detection on non-EE thread"); CONSISTENCY_CHECK_MSG(pThread->m_pBlockingLock.Load() == NULL, "Cannot block on two locks at once"); { CrstHolder lock(&g_DeadlockAwareCrst); // Look for deadlock loop CONSISTENCY_CHECK_MSG(CheckDeadlock(pThread), "Deadlock detected!"); pThread->m_pBlockingLock = this; } }; void DeadlockAwareLock::EndEnterLock() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; Thread * pThread = GetThread(); CONSISTENCY_CHECK(m_pHoldingThread.Load() == NULL || m_pHoldingThread.Load() == pThread); CONSISTENCY_CHECK(pThread->m_pBlockingLock.Load() == this); // No need to take a lock when going from blocking to holding. This // transition implies the lack of a deadlock that other threads can see. // (If they would see a deadlock after the transition, they would see // one before as well.) m_pHoldingThread = pThread; } void DeadlockAwareLock::LeaveLock() { CONTRACTL { NOTHROW; GC_NOTRIGGER; } CONTRACTL_END; CONSISTENCY_CHECK(m_pHoldingThread == GetThread()); CONSISTENCY_CHECK(GetThread()->m_pBlockingLock.Load() == NULL); m_pHoldingThread = NULL; } #ifdef _DEBUG // Normally, any thread we operate on has a Thread block in its TLS. But there are // a few special threads we don't normally execute managed code on. // // There is a scenario where we run managed code on such a thread, which is when the // DLL_THREAD_ATTACH notification of an (IJW?) module calls into managed code. This // is incredibly dangerous. If a GC is provoked, the system may have trouble performing // the GC because its threads aren't available yet. static DWORD SpecialEEThreads[10]; static LONG cnt_SpecialEEThreads = 0; void dbgOnly_IdentifySpecialEEThread() { WRAPPER_NO_CONTRACT; LONG ourCount = FastInterlockIncrement(&cnt_SpecialEEThreads); _ASSERTE(ourCount < (LONG) NumItems(SpecialEEThreads)); SpecialEEThreads[ourCount-1] = ::GetCurrentThreadId(); } BOOL dbgOnly_IsSpecialEEThread() { WRAPPER_NO_CONTRACT; DWORD ourId = ::GetCurrentThreadId(); for (LONG i=0; iWe probably should use Thread::GetThreadId DWORD helperID = g_pDebugInterface->GetHelperThreadID(); if (helperID == ourId) return TRUE; } #endif //Clean this up if (GetThread() == NULL) return TRUE; return FALSE; } #endif // _DEBUG // There is an MDA which can detect illegal reentrancy into the CLR. For instance, if you call managed // code from a native vectored exception handler, this might cause a reverse PInvoke to occur. But if the // exception was triggered from code that was executing in cooperative GC mode, we now have GC holes and // general corruption. #ifdef MDA_SUPPORTED NOINLINE BOOL HasIllegalReentrancyRare() { CONTRACTL { NOTHROW; GC_TRIGGERS; ENTRY_POINT; MODE_ANY; } CONTRACTL_END; Thread *pThread = GetThread(); if (pThread == NULL || !pThread->PreemptiveGCDisabled()) return FALSE; BEGIN_ENTRYPOINT_VOIDRET; MDA_TRIGGER_ASSISTANT(Reentrancy, ReportViolation()); END_ENTRYPOINT_VOIDRET; return TRUE; } #endif // Actually fire the Reentrancy probe, if warranted. BOOL HasIllegalReentrancy() { CONTRACTL { NOTHROW; GC_TRIGGERS; ENTRY_POINT; MODE_ANY; } CONTRACTL_END; #ifdef MDA_SUPPORTED if (NULL == MDA_GET_ASSISTANT(Reentrancy)) return FALSE; return HasIllegalReentrancyRare(); #else return FALSE; #endif // MDA_SUPPORTED } #endif // #ifndef DACCESS_COMPILE #ifdef DACCESS_COMPILE void STATIC_DATA::EnumMemoryRegions(CLRDataEnumMemoryFlags flags) { WRAPPER_NO_CONTRACT; DAC_ENUM_STHIS(STATIC_DATA); } void Thread::EnumMemoryRegions(CLRDataEnumMemoryFlags flags) { WRAPPER_NO_CONTRACT; DAC_ENUM_VTHIS(); if (flags != CLRDATA_ENUM_MEM_MINI && flags != CLRDATA_ENUM_MEM_TRIAGE) { if (m_pDomain.IsValid()) { m_pDomain->EnumMemoryRegions(flags, true); } } if (m_debuggerFilterContext.IsValid()) { m_debuggerFilterContext.EnumMem(); } OBJECTHANDLE_EnumMemoryRegions(m_LastThrownObjectHandle); m_ExceptionState.EnumChainMemoryRegions(flags); m_ThreadLocalBlock.EnumMemoryRegions(flags); if (flags != CLRDATA_ENUM_MEM_MINI && flags != CLRDATA_ENUM_MEM_TRIAGE) { // // Allow all of the frames on the stack to enumerate // their memory. // PTR_Frame frame = m_pFrame; while (frame.IsValid() && frame.GetAddr() != dac_cast(FRAME_TOP)) { frame->EnumMemoryRegions(flags); frame = frame->m_Next; } } // // Try and do a stack trace and save information // for each part of the stack. This is very vulnerable // to memory problems so ignore all exceptions here. // CATCH_ALL_EXCEPT_RETHROW_COR_E_OPERATIONCANCELLED ( EnumMemoryRegionsWorker(flags); ); } void Thread::EnumMemoryRegionsWorker(CLRDataEnumMemoryFlags flags) { WRAPPER_NO_CONTRACT; if (IsUnstarted()) { return; } T_CONTEXT context; BOOL DacGetThreadContext(Thread* thread, T_CONTEXT* context); REGDISPLAY regDisp; StackFrameIterator frameIter; TADDR previousSP = 0; //start at zero; this allows first check to always succeed. TADDR currentSP; // Init value. The Limit itself is not legal, so move one target pointer size to the smallest-magnitude // legal address. currentSP = dac_cast(m_CacheStackLimit) + sizeof(TADDR); if (GetFilterContext()) { context = *GetFilterContext(); } else { DacGetThreadContext(this, &context); } FillRegDisplay(®Disp, &context); frameIter.Init(this, NULL, ®Disp, 0); while (frameIter.IsValid()) { // // There are identical stack pointer checking semantics in code:ClrDataAccess::EnumMemWalkStackHelper // You ***MUST*** maintain identical semantics for both checks! // // Before we continue, we should check to be sure we have a valid // stack pointer. This is to prevent stacks that are not walked // properly due to // a) stack corruption bugs // b) bad stack walks // from continuing on indefinitely. // // We will force SP to strictly increase. // this check can only happen for real stack frames (i.e. not for explicit frames that don't update the RegDisplay) // for ia64, SP may be equal, but in this case BSP must strictly decrease. // We will force SP to be properly aligned. // We will force SP to be in the correct range. // if (frameIter.GetFrameState() == StackFrameIterator::SFITER_FRAMELESS_METHOD) { // This check cannot be applied to explicit frames; they may not move the SP at all. // Also, a single function can push several on the stack at a time with no guarantees about // ordering so we can't check that the addresses of the explicit frames are monotonically increasing. // There is the potential that the walk will not terminate if a set of explicit frames reference // each other circularly. While we could choose a limit for the number of explicit frames allowed // in a row like the total stack size/pointer size, we have no known problems with this scenario. // Thus for now we ignore it. currentSP = (TADDR)GetRegdisplaySP(®Disp); if (currentSP <= previousSP) { _ASSERTE(!"Target stack has been corrupted, SP for current frame must be larger than previous frame."); break; } } // On windows desktop, the stack pointer should be a multiple // of pointer-size-aligned in the target address space if (currentSP % sizeof(TADDR) != 0) { _ASSERTE(!"Target stack has been corrupted, SP must be aligned."); break; } if (!IsAddressInStack(currentSP)) { _ASSERTE(!"Target stack has been corrupted, SP must in in the stack range."); break; } // Enumerate the code around the call site to help debugger stack walking heuristics PCODE callEnd = GetControlPC(®Disp); DacEnumCodeForStackwalk(callEnd); if (flags != CLRDATA_ENUM_MEM_MINI && flags != CLRDATA_ENUM_MEM_TRIAGE) { if (frameIter.m_crawl.GetAppDomain()) { frameIter.m_crawl.GetAppDomain()->EnumMemoryRegions(flags, true); } } // To stackwalk through funceval frames, we need to be sure to preserve the // DebuggerModule's m_pRuntimeDomainFile. This is the only case that doesn't use the current // vmDomainFile in code:DacDbiInterfaceImpl::EnumerateInternalFrames. The following // code mimics that function. // Allow failure, since we want to continue attempting to walk the stack regardless of the outcome. EX_TRY { if ((frameIter.GetFrameState() == StackFrameIterator::SFITER_FRAME_FUNCTION) || (frameIter.GetFrameState() == StackFrameIterator::SFITER_SKIPPED_FRAME_FUNCTION)) { Frame * pFrame = frameIter.m_crawl.GetFrame(); g_pDebugInterface->EnumMemoryRegionsIfFuncEvalFrame(flags, pFrame); } } EX_CATCH_RETHROW_ONLY_COR_E_OPERATIONCANCELLED MethodDesc* pMD = frameIter.m_crawl.GetFunction(); if (pMD != NULL) { pMD->EnumMemoryRegions(flags); #if defined(WIN64EXCEPTIONS) && defined(FEATURE_PREJIT) // Enumerate unwind info // Note that we don't do this based on the MethodDesc because in theory there isn't a 1:1 correspondence // between MethodDesc and code (and so unwind info, and even debug info). Eg., EnC creates new versions // of the code, but the MethodDesc always points at the latest version (which isn't necessarily // the one on the stack). In practice this is unlikely to be a problem since wanting a minidump // and making EnC edits are usually mutually exclusive. if (frameIter.m_crawl.IsFrameless()) { frameIter.m_crawl.GetJitManager()->EnumMemoryRegionsForMethodUnwindInfo(flags, frameIter.m_crawl.GetCodeInfo()); } #endif // defined(WIN64EXCEPTIONS) && defined(FEATURE_PREJIT) } previousSP = currentSP; if (frameIter.Next() != SWA_CONTINUE) { break; } } } void ThreadStore::EnumMemoryRegions(CLRDataEnumMemoryFlags flags) { SUPPORTS_DAC; WRAPPER_NO_CONTRACT; // This will write out the context of the s_pThreadStore. ie // just the pointer // s_pThreadStore.EnumMem(); if (s_pThreadStore.IsValid()) { // write out the whole ThreadStore structure DacEnumHostDPtrMem(s_pThreadStore); // The thread list may be corrupt, so just // ignore exceptions during enumeration. EX_TRY { Thread* thread = s_pThreadStore->m_ThreadList.GetHead(); LONG dwNumThreads = s_pThreadStore->m_ThreadCount; for (LONG i = 0; (i < dwNumThreads) && (thread != NULL); i++) { // Even if this thread is totally broken and we can't enum it, struggle on. // If we do not, we will leave this loop and not enum stack memory for any further threads. CATCH_ALL_EXCEPT_RETHROW_COR_E_OPERATIONCANCELLED( thread->EnumMemoryRegions(flags); ); thread = s_pThreadStore->m_ThreadList.GetNext(thread); } } EX_CATCH_RETHROW_ONLY_COR_E_OPERATIONCANCELLED } } #endif // #ifdef DACCESS_COMPILE #ifdef FEATURE_APPDOMAIN_RESOURCE_MONITORING // For the purposes of tracking resource usage we implement a simple cpu resource usage counter on each // thread. Every time QueryThreadProcessorUsage() is invoked it returns the amount of cpu time (a combination // of user and kernel mode time) used since the last call to QueryThreadProcessorUsage(). The result is in 100 // nanosecond units. ULONGLONG Thread::QueryThreadProcessorUsage() { LIMITED_METHOD_CONTRACT; // Get current values for the amount of kernel and user time used by this thread over its entire lifetime. FILETIME sCreationTime, sExitTime, sKernelTime, sUserTime; HANDLE hThread = GetThreadHandle(); BOOL fResult = GetThreadTimes(hThread, &sCreationTime, &sExitTime, &sKernelTime, &sUserTime); if (!fResult) { #ifdef _DEBUG ULONG error = GetLastError(); printf("GetThreadTimes failed: %d; handle is %p\n", error, hThread); _ASSERTE(FALSE); #endif return 0; } // Combine the user and kernel times into a single value (FILETIME is just a structure representing an // unsigned int64 in two 32-bit pieces). _ASSERTE(sizeof(FILETIME) == sizeof(UINT64)); ULONGLONG ullCurrentUsage = *(ULONGLONG*)&sKernelTime + *(ULONGLONG*)&sUserTime; // Store the current processor usage as the new baseline, and retrieve the previous usage. ULONGLONG ullPreviousUsage = VolatileLoad(&m_ullProcessorUsageBaseline); if (ullPreviousUsage >= ullCurrentUsage || ullPreviousUsage != (ULONGLONG)InterlockedCompareExchange64( (LONGLONG*)&m_ullProcessorUsageBaseline, (LONGLONG)ullCurrentUsage, (LONGLONG)ullPreviousUsage)) { // another thread beat us to it, and already reported this usage. return 0; } // The result is the difference between this value and the previous usage value. return ullCurrentUsage - ullPreviousUsage; } #endif // FEATURE_APPDOMAIN_RESOURCE_MONITORING