diff --git a/libpolyml/processes.cpp b/libpolyml/processes.cpp index 2d637b1c..2102e01a 100644 --- a/libpolyml/processes.cpp +++ b/libpolyml/processes.cpp @@ -1,2262 +1,2250 @@ /* Title: Thread functions Author: David C.J. Matthews - Copyright (c) 2007,2008,2013-15, 2017 David C.J. Matthews + Copyright (c) 2007,2008,2013-15, 2017, 2019 David C.J. Matthews This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License version 2.1 as published by the Free Software Foundation. This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */ #ifdef HAVE_CONFIG_H #include "config.h" #elif defined(_WIN32) #include "winconfig.h" #else #error "No configuration file" #endif #ifdef HAVE_STDIO_H #include #endif #ifdef HAVE_ERRNO_H #include #endif #ifdef HAVE_STDLIB_H #include #endif #ifdef HAVE_STRING_H #include #endif #ifdef HAVE_LIMITS_H #include #endif #ifdef HAVE_ASSERT_H #include #define ASSERT(x) assert(x) #else #define ASSERT(x) #endif #ifdef HAVE_PROCESS_H #include #endif #ifdef HAVE_SYS_TYPES_H #include #endif #ifdef HAVE_SYS_STAT_H #include #endif #ifdef HAVE_SYS_TIME_H #include #endif #ifdef HAVE_UNISTD_H #include // Want unistd for _SC_NPROCESSORS_ONLN at least #endif #ifdef HAVE_SYS_SELECT_H #include #endif #ifdef HAVE_WINDOWS_H #include #endif #if ((!defined(_WIN32) || defined(__CYGWIN__)) && defined(HAVE_LIBPTHREAD) && defined(HAVE_PTHREAD_H)) #define HAVE_PTHREAD 1 #include #endif #ifdef HAVE_SYS_SYSCTL_H // Used determine number of processors in Mac OS X. #include #endif #if (defined(_WIN32) && ! defined(__CYGWIN__)) #include #endif #include #include /************************************************************************ * * Include runtime headers * ************************************************************************/ #include "globals.h" #include "gc.h" #include "mpoly.h" #include "arb.h" #include "machine_dep.h" #include "diagnostics.h" #include "processes.h" #include "run_time.h" #include "sys.h" #include "sighandler.h" #include "scanaddrs.h" #include "save_vec.h" #include "rts_module.h" #include "noreturn.h" #include "memmgr.h" #include "locking.h" #include "profiling.h" #include "sharedata.h" #include "exporter.h" #include "statistics.h" #include "rtsentry.h" extern "C" { POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadGeneral(PolyObject *threadId, PolyWord code, PolyWord arg); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadKillSelf(PolyObject *threadId); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadMutexBlock(PolyObject *threadId, PolyWord arg); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadMutexUnlock(PolyObject *threadId, PolyWord arg); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadCondVarWait(PolyObject *threadId, PolyWord arg); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadCondVarWaitUntil(PolyObject *threadId, PolyWord lockArg, PolyWord timeArg); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadCondVarWake(PolyWord targetThread); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadForkThread(PolyObject *threadId, PolyWord function, PolyWord attrs, PolyWord stack); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadIsActive(PolyWord targetThread); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadInterruptThread(PolyWord targetThread); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadKillThread(PolyWord targetThread); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadBroadcastInterrupt(PolyObject *threadId); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadTestInterrupt(PolyObject *threadId); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadNumProcessors(); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadNumPhysicalProcessors(); POLYEXTERNALSYMBOL POLYUNSIGNED PolyThreadMaxStackSize(PolyObject *threadId, PolyWord newSize); } #define SAVE(x) taskData->saveVec.push(x) #define SIZEOF(x) (sizeof(x)/sizeof(PolyWord)) // These values are stored in the second word of thread id object as // a tagged integer. They may be set and read by the thread in the ML // code. #define PFLAG_BROADCAST 1 // If set, accepts a broadcast // How to handle interrrupts #define PFLAG_IGNORE 0 // Ignore interrupts completely #define PFLAG_SYNCH 2 // Handle synchronously #define PFLAG_ASYNCH 4 // Handle asynchronously #define PFLAG_ASYNCH_ONCE 6 // First handle asynchronously then switch to synch. #define PFLAG_INTMASK 6 // Mask of the above bits struct _entrypts processesEPT[] = { { "PolyThreadGeneral", (polyRTSFunction)&PolyThreadGeneral}, { "PolyThreadKillSelf", (polyRTSFunction)&PolyThreadKillSelf}, { "PolyThreadMutexBlock", (polyRTSFunction)&PolyThreadMutexBlock}, { "PolyThreadMutexUnlock", (polyRTSFunction)&PolyThreadMutexUnlock}, { "PolyThreadCondVarWait", (polyRTSFunction)&PolyThreadCondVarWait}, { "PolyThreadCondVarWaitUntil", (polyRTSFunction)&PolyThreadCondVarWaitUntil}, { "PolyThreadCondVarWake", (polyRTSFunction)&PolyThreadCondVarWake}, { "PolyThreadForkThread", (polyRTSFunction)&PolyThreadForkThread}, { "PolyThreadIsActive", (polyRTSFunction)&PolyThreadIsActive}, { "PolyThreadInterruptThread", (polyRTSFunction)&PolyThreadInterruptThread}, { "PolyThreadKillThread", (polyRTSFunction)&PolyThreadKillThread}, { "PolyThreadBroadcastInterrupt", (polyRTSFunction)&PolyThreadBroadcastInterrupt}, { "PolyThreadTestInterrupt", (polyRTSFunction)&PolyThreadTestInterrupt}, { "PolyThreadNumProcessors", (polyRTSFunction)&PolyThreadNumProcessors}, { "PolyThreadNumPhysicalProcessors",(polyRTSFunction)&PolyThreadNumPhysicalProcessors}, { "PolyThreadMaxStackSize", (polyRTSFunction)&PolyThreadMaxStackSize}, { NULL, NULL} // End of list. }; class Processes: public ProcessExternal, public RtsModule { public: Processes(); virtual void Init(void); virtual void Stop(void); void GarbageCollect(ScanAddress *process); public: void BroadcastInterrupt(void); void BeginRootThread(PolyObject *rootFunction); void RequestProcessExit(int n); // Request all ML threads to exit and set the process result code. // Called when a thread has completed - doesn't return. virtual NORETURNFN(void ThreadExit(TaskData *taskData)); // Called when a thread may block. Returns some time later when perhaps // the input is available. virtual void ThreadPauseForIO(TaskData *taskData, Waiter *pWait); // Return the task data for the current thread. virtual TaskData *GetTaskDataForThread(void); // Create a new task data object for the current thread. virtual TaskData *CreateNewTaskData(Handle threadId, Handle threadFunction, Handle args, PolyWord flags); // ForkFromRTS. Creates a new thread from within the RTS. virtual bool ForkFromRTS(TaskData *taskData, Handle proc, Handle arg); // Create a new thread. The "args" argument is only used for threads // created in the RTS by the signal handler. Handle ForkThread(TaskData *taskData, Handle threadFunction, Handle args, PolyWord flags, PolyWord stacksize); // Process general RTS requests from ML. Handle ThreadDispatch(TaskData *taskData, Handle args, Handle code); virtual void ThreadUseMLMemory(TaskData *taskData); virtual void ThreadReleaseMLMemory(TaskData *taskData); virtual poly_exn* GetInterrupt(void) { return interrupt_exn; } // If the schedule lock is already held we need to use these functions. void ThreadUseMLMemoryWithSchedLock(TaskData *taskData); void ThreadReleaseMLMemoryWithSchedLock(TaskData *taskData); // Requests from the threads for actions that need to be performed by // the root thread. Make the request and wait until it has completed. virtual void MakeRootRequest(TaskData *taskData, MainThreadRequest *request); // Deal with any interrupt or kill requests. virtual bool ProcessAsynchRequests(TaskData *taskData); // Process an interrupt request synchronously. virtual void TestSynchronousRequests(TaskData *taskData); // Process any events, synchronous or asynchronous. virtual void TestAnyEvents(TaskData *taskData); // Set a thread to be interrupted or killed. Wakes up the // thread if necessary. MUST be called with schedLock held. void MakeRequest(TaskData *p, ThreadRequests request); // Profiling control. virtual void StartProfiling(void); virtual void StopProfiling(void); #ifdef HAVE_WINDOWS_H // Windows: Called every millisecond while profiling is on. void ProfileInterrupt(void); #else // Unix: Start a profile timer for a thread. void StartProfilingTimer(void); #endif // Memory allocation. Tries to allocate space. If the allocation succeeds it // may update the allocation values in the taskData object. If the heap is exhausted // it may set this thread (or other threads) to raise an exception. PolyWord *FindAllocationSpace(TaskData *taskData, POLYUNSIGNED words, bool alwaysInSeg); - // Find a task that matches the specified identifier and returns - // it if it exists. MUST be called with schedLock held. - TaskData *TaskForIdentifier(PolyObject *taskId); + // Get the task data value from the task reference. + // The task data reference is a volatile ref containing the + // address of the C++ task data. + // N.B. This is updated when the thread exits and the TaskData object + // is deleted. + TaskData *TaskForIdentifier(PolyObject *taskId) { + return *(TaskData**)(((ThreadObject*)taskId)->threadRef.AsObjPtr()); + } // Signal handling support. The ML signal handler thread blocks until it is // woken up by the signal detection thread. virtual bool WaitForSignal(TaskData *taskData, PLock *sigLock); virtual void SignalArrived(void); virtual void SetSingleThreaded(void) { singleThreaded = true; } // Operations on mutexes void MutexBlock(TaskData *taskData, Handle hMutex); void MutexUnlock(TaskData *taskData, Handle hMutex); // Operations on condition variables. void WaitInfinite(TaskData *taskData, Handle hMutex); void WaitUntilTime(TaskData *taskData, Handle hMutex, Handle hTime); bool WakeThread(PolyObject *targetThread); // Generally, the system runs with multiple threads. After a // fork, though, there is only one thread. bool singleThreaded; // Each thread has an entry in this vector. std::vector taskArray; /* schedLock: This lock must be held when making scheduling decisions. It must also be held before adding items to taskArray, removing them or scanning the vector. It must also be held before deleting a TaskData object or using it in a thread other than the "owner" */ PLock schedLock; #ifdef HAVE_PTHREAD pthread_key_t tlsId; #elif defined(HAVE_WINDOWS_H) DWORD tlsId; #endif // We make an exception packet for Interrupt and store it here. // This exception can be raised if we run out of store so we need to // make sure we have the packet before we do. poly_exn *interrupt_exn; /* initialThreadWait: The initial thread waits on this for wake-ups from the ML threads requesting actions such as GC or close-down. */ PCondVar initialThreadWait; // A requesting thread sets this to indicate the request. This value // is only reset once the request has been satisfied. MainThreadRequest *threadRequest; PCondVar mlThreadWait; // All the threads block on here until the request has completed. int exitResult; bool exitRequest; #ifdef HAVE_WINDOWS_H // Used in profiling HANDLE hStopEvent; /* Signalled to stop all threads. */ HANDLE profilingHd; HANDLE mainThreadHandle; // Handle for main thread LONGLONG lastCPUTime; // CPU used by main thread. #endif TaskData *sigTask; // Pointer to current signal task. }; // Global process data. static Processes processesModule; ProcessExternal *processes = &processesModule; Processes::Processes(): singleThreaded(false), schedLock("Scheduler"), interrupt_exn(0), threadRequest(0), exitResult(0), exitRequest(false), sigTask(0) { #ifdef HAVE_WINDOWS_H hStopEvent = NULL; profilingHd = NULL; lastCPUTime = 0; mainThreadHandle = NULL; #endif } enum _mainThreadPhase mainThreadPhase = MTP_USER_CODE; // Get the attribute flags. static POLYUNSIGNED ThreadAttrs(TaskData *taskData) { return UNTAGGED_UNSIGNED(taskData->threadObject->flags); } // General interface to thread. Ideally the various cases will be made into // separate functions. POLYUNSIGNED PolyThreadGeneral(PolyObject *threadId, PolyWord code, PolyWord arg) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedCode = taskData->saveVec.push(code); Handle pushedArg = taskData->saveVec.push(arg); Handle result = 0; try { result = processesModule.ThreadDispatch(taskData, pushedArg, pushedCode); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); if (result == 0) return TAGGED(0).AsUnsigned(); else return result->Word().AsUnsigned(); } POLYUNSIGNED PolyThreadMutexBlock(PolyObject *threadId, PolyWord arg) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedArg = taskData->saveVec.push(arg); if (profileMode == kProfileMutexContention) taskData->addProfileCount(1); try { processesModule.MutexBlock(taskData, pushedArg); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } POLYUNSIGNED PolyThreadMutexUnlock(PolyObject *threadId, PolyWord arg) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedArg = taskData->saveVec.push(arg); try { processesModule.MutexUnlock(taskData, pushedArg); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } /* A mutex was locked i.e. the count was ~1 or less. We will have set it to ~1. This code blocks if the count is still ~1. It does actually return if another thread tries to lock the mutex and hasn't yet set the value to ~1 but that doesn't matter since whenever we return we simply try to get the lock again. */ void Processes::MutexBlock(TaskData *taskData, Handle hMutex) { schedLock.Lock(); // We have to check the value again with schedLock held rather than // simply waiting because otherwise the unlocking thread could have // set the variable back to 1 (unlocked) and signalled any waiters // before we actually got to wait. if (UNTAGGED(DEREFHANDLE(hMutex)->Get(0)) < 0) { // Set this so we can see what we're blocked on. taskData->blockMutex = DEREFHANDLE(hMutex); // Now release the ML memory. A GC can start. ThreadReleaseMLMemoryWithSchedLock(taskData); // Wait until we're woken up. We mustn't block if we have been // interrupted, and are processing interrupts asynchronously, or // we've been killed. switch (taskData->requests) { case kRequestKill: // We've been killed. Handle this later. break; case kRequestInterrupt: { // We've been interrupted. POLYUNSIGNED attrs = ThreadAttrs(taskData) & PFLAG_INTMASK; if (attrs == PFLAG_ASYNCH || attrs == PFLAG_ASYNCH_ONCE) break; // If we're ignoring interrupts or handling them synchronously // we don't do anything here. } case kRequestNone: globalStats.incCount(PSC_THREADS_WAIT_MUTEX); taskData->threadLock.Wait(&schedLock); globalStats.decCount(PSC_THREADS_WAIT_MUTEX); } taskData->blockMutex = 0; // No longer blocked. ThreadUseMLMemoryWithSchedLock(taskData); } // Return and try and get the lock again. schedLock.Unlock(); // Test to see if we have been interrupted and if this thread // processes interrupts asynchronously we should raise an exception // immediately. Perhaps we do that whenever we exit from the RTS. } /* Unlock a mutex. Called after incrementing the count and discovering that at least one other thread has tried to lock it. We may need to wake up threads that are blocked. */ void Processes::MutexUnlock(TaskData *taskData, Handle hMutex) { // The caller has already set the variable to 1 (unlocked). // We need to acquire schedLock so that we can // be sure that any thread that is trying to lock sees either // the updated value (and so doesn't wait) or has successfully // waited on its threadLock (and so will be woken up). schedLock.Lock(); // Unlock any waiters. for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; // If the thread is blocked on this mutex we can signal the thread. if (p && p->blockMutex == DEREFHANDLE(hMutex)) p->threadLock.Signal(); } schedLock.Unlock(); } POLYUNSIGNED PolyThreadCondVarWait(PolyObject *threadId, PolyWord arg) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedArg = taskData->saveVec.push(arg); try { processesModule.WaitInfinite(taskData, pushedArg); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } POLYUNSIGNED PolyThreadCondVarWaitUntil(PolyObject *threadId, PolyWord lockArg, PolyWord timeArg) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedLockArg = taskData->saveVec.push(lockArg); Handle pushedTimeArg = taskData->saveVec.push(timeArg); try { processesModule.WaitUntilTime(taskData, pushedLockArg, pushedTimeArg); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } // Atomically drop a mutex and wait for a wake up. // It WILL NOT RAISE AN EXCEPTION unless it is set to handle exceptions // asynchronously (which it shouldn't do if the ML caller code is correct). // It may return as a result of any of the following: // an explicit wake up. // an interrupt, either direct or broadcast // a trap i.e. a request to handle an asynchronous event. void Processes::WaitInfinite(TaskData *taskData, Handle hMutex) { schedLock.Lock(); // Atomically release the mutex. This is atomic because we hold schedLock // so no other thread can call signal or broadcast. Handle decrResult = taskData->AtomicIncrement(hMutex); if (UNTAGGED(decrResult->Word()) != 1) { taskData->AtomicReset(hMutex); // The mutex was locked so we have to release any waiters. // Unlock any waiters. for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; // If the thread is blocked on this mutex we can signal the thread. if (p && p->blockMutex == DEREFHANDLE(hMutex)) p->threadLock.Signal(); } } // Wait until we're woken up. Don't block if we have been interrupted // or killed. if (taskData->requests == kRequestNone) { // Now release the ML memory. A GC can start. ThreadReleaseMLMemoryWithSchedLock(taskData); globalStats.incCount(PSC_THREADS_WAIT_CONDVAR); taskData->threadLock.Wait(&schedLock); globalStats.decCount(PSC_THREADS_WAIT_CONDVAR); // We want to use the memory again. ThreadUseMLMemoryWithSchedLock(taskData); } schedLock.Unlock(); } // Atomically drop a mutex and wait for a wake up or a time to wake up void Processes::WaitUntilTime(TaskData *taskData, Handle hMutex, Handle hWakeTime) { // Convert the time into the correct format for WaitUntil before acquiring // schedLock. div_longc could do a GC which requires schedLock. #if (defined(_WIN32) && ! defined(__CYGWIN__)) // On Windows it is the number of 100ns units since the epoch FILETIME tWake; getFileTimeFromArb(taskData, hWakeTime, &tWake); #else // Unix style times. struct timespec tWake; // On Unix we represent times as a number of microseconds. Handle hMillion = Make_arbitrary_precision(taskData, 1000000); tWake.tv_sec = get_C_ulong(taskData, DEREFWORD(div_longc(taskData, hMillion, hWakeTime))); tWake.tv_nsec = 1000*get_C_ulong(taskData, DEREFWORD(rem_longc(taskData, hMillion, hWakeTime))); #endif schedLock.Lock(); // Atomically release the mutex. This is atomic because we hold schedLock // so no other thread can call signal or broadcast. Handle decrResult = taskData->AtomicIncrement(hMutex); if (UNTAGGED(decrResult->Word()) != 1) { taskData->AtomicReset(hMutex); // The mutex was locked so we have to release any waiters. // Unlock any waiters. for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; // If the thread is blocked on this mutex we can signal the thread. if (p && p->blockMutex == DEREFHANDLE(hMutex)) p->threadLock.Signal(); } } // Wait until we're woken up. Don't block if we have been interrupted // or killed. if (taskData->requests == kRequestNone) { // Now release the ML memory. A GC can start. ThreadReleaseMLMemoryWithSchedLock(taskData); globalStats.incCount(PSC_THREADS_WAIT_CONDVAR); (void)taskData->threadLock.WaitUntil(&schedLock, &tWake); globalStats.decCount(PSC_THREADS_WAIT_CONDVAR); // We want to use the memory again. ThreadUseMLMemoryWithSchedLock(taskData); } schedLock.Unlock(); } bool Processes::WakeThread(PolyObject *targetThread) { bool result = false; // Default to failed. // Acquire the schedLock first. This ensures that this is // atomic with respect to waiting. schedLock.Lock(); TaskData *p = TaskForIdentifier(targetThread); if (p && p->threadObject == targetThread) { POLYUNSIGNED attrs = ThreadAttrs(p) & PFLAG_INTMASK; if (p->requests == kRequestNone || (p->requests == kRequestInterrupt && attrs == PFLAG_IGNORE)) { p->threadLock.Signal(); result = true; } } schedLock.Unlock(); return result; } POLYUNSIGNED PolyThreadCondVarWake(PolyWord targetThread) { if (processesModule.WakeThread(targetThread.AsObjPtr())) return TAGGED(1).AsUnsigned(); else return TAGGED(0).AsUnsigned(); } // Test if a thread is active. POLYUNSIGNED PolyThreadIsActive(PolyWord targetThread) { - processesModule.schedLock.Lock(); + // There's a race here: the thread may be exiting but since we're not doing + // anything with the TaskData object we don't need a lock. TaskData *p = processesModule.TaskForIdentifier(targetThread.AsObjPtr()); - processesModule.schedLock.Unlock(); if (p != 0) return TAGGED(1).AsUnsigned(); else return TAGGED(0).AsUnsigned(); } // Send an interrupt to a specific thread POLYUNSIGNED PolyThreadInterruptThread(PolyWord targetThread) { + // Must lock here because the thread may be exiting. processesModule.schedLock.Lock(); TaskData *p = processesModule.TaskForIdentifier(targetThread.AsObjPtr()); if (p) processesModule.MakeRequest(p, kRequestInterrupt); processesModule.schedLock.Unlock(); // If the thread cannot be identified return false. // The caller can then raise an exception if (p == 0) return TAGGED(0).AsUnsigned(); else return TAGGED(1).AsUnsigned(); } // Kill a specific thread POLYUNSIGNED PolyThreadKillThread(PolyWord targetThread) { processesModule.schedLock.Lock(); TaskData *p = processesModule.TaskForIdentifier(targetThread.AsObjPtr()); if (p) processesModule.MakeRequest(p, kRequestKill); processesModule.schedLock.Unlock(); // If the thread cannot be identified return false. // The caller can then raise an exception if (p == 0) return TAGGED(0).AsUnsigned(); else return TAGGED(1).AsUnsigned(); } POLYUNSIGNED PolyThreadBroadcastInterrupt(PolyObject * /*threadId*/) { processesModule.BroadcastInterrupt(); return TAGGED(0).AsUnsigned(); } POLYUNSIGNED PolyThreadTestInterrupt(PolyObject *threadId) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); try { processesModule.TestSynchronousRequests(taskData); // Also process any asynchronous requests that may be pending. // These will be handled "soon" but if we have just switched from deferring // interrupts this guarantees that any deferred interrupts will be handled now. if (processesModule.ProcessAsynchRequests(taskData)) throw IOException(); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } // Return the number of processors. // Returns 1 if there is any problem. POLYUNSIGNED PolyThreadNumProcessors(void) { return TAGGED(NumberOfProcessors()).AsUnsigned(); } // Return the number of physical processors. // Returns 0 if there is any problem. POLYUNSIGNED PolyThreadNumPhysicalProcessors(void) { return TAGGED(NumberOfPhysicalProcessors()).AsUnsigned(); } // Set the maximum stack size. POLYUNSIGNED PolyThreadMaxStackSize(PolyObject *threadId, PolyWord newSize) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); try { taskData->threadObject->mlStackSize = newSize; if (newSize != TAGGED(0)) { uintptr_t current = taskData->currentStackSpace(); // Current size in words uintptr_t newWords = getPolyUnsigned(taskData, newSize); if (current > newWords) raise_exception0(taskData, EXC_interrupt); } } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); return TAGGED(0).AsUnsigned(); } // Old dispatch function. This is only required because the pre-built compiler // may use some of these e.g. fork. Handle Processes::ThreadDispatch(TaskData *taskData, Handle args, Handle code) { unsigned c = get_C_unsigned(taskData, code->Word()); TaskData *ptaskData = taskData; switch (c) { case 1: MutexBlock(taskData, args); return SAVE(TAGGED(0)); case 2: MutexUnlock(taskData, args); return SAVE(TAGGED(0)); case 7: // Fork a new thread. The arguments are the function to run and the attributes. return ForkThread(ptaskData, SAVE(args->WordP()->Get(0)), (Handle)0, args->WordP()->Get(1), // For backwards compatibility we check the length here args->WordP()->Length() <= 2 ? TAGGED(0) : args->WordP()->Get(2)); case 10: // Broadcast an interrupt to all threads that are interested. BroadcastInterrupt(); return SAVE(TAGGED(0)); default: { char msg[100]; sprintf(msg, "Unknown thread function: %u", c); raise_fail(taskData, msg); return 0; } } } // Fill unused allocation space with a dummy object to preserve the invariant // that memory is always valid. void TaskData::FillUnusedSpace(void) { if (allocPointer > allocLimit) gMem.FillUnusedSpace(allocLimit, allocPointer-allocLimit); } TaskData::TaskData(): allocPointer(0), allocLimit(0), allocSize(MIN_HEAP_SIZE), allocCount(0), stack(0), threadObject(0), signalStack(0), foreignStack(TAGGED(0)), inML(false), requests(kRequestNone), blockMutex(0), inMLHeap(false), runningProfileTimer(false) { #ifdef HAVE_WINDOWS_H lastCPUTime = 0; #endif #ifdef HAVE_WINDOWS_H threadHandle = 0; #endif threadExited = false; } TaskData::~TaskData() { if (signalStack) free(signalStack); if (stack) gMem.DeleteStackSpace(stack); #ifdef HAVE_WINDOWS_H if (threadHandle) CloseHandle(threadHandle); #endif } - -// Find a task that matches the specified identifier and returns -// it if it exists. MUST be called with schedLock held. -TaskData *Processes::TaskForIdentifier(PolyObject *taskId) -{ - // The index is in the first word of the thread object. - unsigned index = (unsigned)(UNTAGGED_UNSIGNED(taskId->Get(0))); - // Check the index is valid and matches the object stored in the table. - if (index < taskArray.size()) - { - TaskData *p = taskArray[index]; - if (p && p->threadObject == taskId) - return p; - } - return 0; -} - -// Return the task data for a task id. -TaskData *TaskData::FindTaskForId(PolyObject *taskId) -{ - PLocker lock(&processesModule.schedLock); - return processesModule.TaskForIdentifier(taskId); -} - // Broadcast an interrupt to all relevant threads. void Processes::BroadcastInterrupt(void) { // If a thread is set to accept broadcast interrupts set it to // "interrupted". schedLock.Lock(); for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; if (p) { POLYUNSIGNED attrs = ThreadAttrs(p); if (attrs & PFLAG_BROADCAST) MakeRequest(p, kRequestInterrupt); } } schedLock.Unlock(); } // Set the asynchronous request variable for the thread. Must be called // with the schedLock held. Tries to wake the thread up if possible. void Processes::MakeRequest(TaskData *p, ThreadRequests request) { // We don't override a request to kill by an interrupt request. if (p->requests < request) { p->requests = request; p->InterruptCode(); p->threadLock.Signal(); // Set the value in the ML object as well so the ML code can see it p->threadObject->requestCopy = TAGGED(request); } } void Processes::ThreadExit(TaskData *taskData) { if (debugOptions & DEBUG_THREADS) Log("THREAD: Thread %p exiting\n", taskData); #ifdef HAVE_PTHREAD // Block any profile interrupt from now on. We're deleting the ML stack for this thread. sigset_t block_sigs; sigemptyset(&block_sigs); sigaddset(&block_sigs, SIGVTALRM); pthread_sigmask(SIG_BLOCK, &block_sigs, NULL); // Remove the thread-specific data since it's no // longer valid. pthread_setspecific(tlsId, 0); #endif if (singleThreaded) finish(0); schedLock.Lock(); ThreadReleaseMLMemoryWithSchedLock(taskData); // Allow a GC if it was waiting for us. taskData->threadExited = true; initialThreadWait.Signal(); // Tell it we've finished. schedLock.Unlock(); #ifdef HAVE_PTHREAD pthread_exit(0); #elif defined(HAVE_WINDOWS_H) ExitThread(0); #endif } // These two functions are used for calls from outside where // the lock has not yet been acquired. void Processes::ThreadUseMLMemory(TaskData *taskData) { // Trying to acquire the lock here may block if a GC is in progress schedLock.Lock(); ThreadUseMLMemoryWithSchedLock(taskData); schedLock.Unlock(); } void Processes::ThreadReleaseMLMemory(TaskData *taskData) { schedLock.Lock(); ThreadReleaseMLMemoryWithSchedLock(taskData); schedLock.Unlock(); } // Called when a thread wants to resume using the ML heap. That could // be after a wait for some reason or after executing some foreign code. // Since there could be a GC in progress already at this point we may either // be blocked waiting to acquire schedLock or we may need to wait until // we are woken up at the end of the GC. void Processes::ThreadUseMLMemoryWithSchedLock(TaskData *taskData) { TaskData *ptaskData = taskData; // If there is a request outstanding we have to wait for it to // complete. We notify the root thread and wait for it. while (threadRequest != 0) { initialThreadWait.Signal(); // Wait for the GC to happen mlThreadWait.Wait(&schedLock); } ASSERT(! ptaskData->inMLHeap); ptaskData->inMLHeap = true; } // Called to indicate that the thread has temporarily finished with the // ML memory either because it is going to wait for something or because // it is going to run foreign code. If there is an outstanding GC request // that can proceed. void Processes::ThreadReleaseMLMemoryWithSchedLock(TaskData *taskData) { TaskData *ptaskData = taskData; ASSERT(ptaskData->inMLHeap); ptaskData->inMLHeap = false; // Put a dummy object in any unused space. This maintains the // invariant that the allocated area is filled with valid objects. ptaskData->FillUnusedSpace(); // if (threadRequest != 0) initialThreadWait.Signal(); } // Make a request to the root thread. void Processes::MakeRootRequest(TaskData *taskData, MainThreadRequest *request) { if (singleThreaded) { mainThreadPhase = request->mtp; ThreadReleaseMLMemoryWithSchedLock(taskData); // Primarily to call FillUnusedSpace request->Perform(); ThreadUseMLMemoryWithSchedLock(taskData); mainThreadPhase = MTP_USER_CODE; } else { PLocker locker(&schedLock); // Wait for any other requests. while (threadRequest != 0) { // Deal with any pending requests. ThreadReleaseMLMemoryWithSchedLock(taskData); ThreadUseMLMemoryWithSchedLock(taskData); // Drops schedLock while waiting. } // Now the other requests have been dealt with (and we have schedLock). request->completed = false; threadRequest = request; // Wait for it to complete. while (! request->completed) { ThreadReleaseMLMemoryWithSchedLock(taskData); ThreadUseMLMemoryWithSchedLock(taskData); // Drops schedLock while waiting. } } } // Find space for an object. Returns a pointer to the start. "words" must include // the length word and the result points at where the length word will go. PolyWord *Processes::FindAllocationSpace(TaskData *taskData, POLYUNSIGNED words, bool alwaysInSeg) { bool triedInterrupt = false; #ifdef POLYML32IN64 if (words & 1) words++; // Must always be an even number of words. #endif while (1) { // After a GC allocPointer and allocLimit are zero and when allocating the // heap segment we request a minimum of zero words. if (taskData->allocPointer != 0 && taskData->allocPointer >= taskData->allocLimit + words) { // There's space in the current segment, taskData->allocPointer -= words; #ifdef POLYML32IN64 // Zero the last word. If we've rounded up an odd number the caller won't set it. if (words != 0) taskData->allocPointer[words-1] = PolyWord::FromUnsigned(0); ASSERT((uintptr_t)taskData->allocPointer & 4); // Must be odd-word aligned #endif return taskData->allocPointer; } else // Insufficient space in this area. { if (words > taskData->allocSize && ! alwaysInSeg) { // If the object we want is larger than the heap segment size // we allocate it separately rather than in the segment. PolyWord *foundSpace = gMem.AllocHeapSpace(words); if (foundSpace) return foundSpace; } else { // Fill in any unused space in the existing segment taskData->FillUnusedSpace(); // Get another heap segment with enough space for this object. uintptr_t requestSpace = taskData->allocSize+words; uintptr_t spaceSize = requestSpace; // Get the space and update spaceSize with the actual size. PolyWord *space = gMem.AllocHeapSpace(words, spaceSize); if (space) { // Double the allocation size for the next time if // we succeeded in allocating the whole space. taskData->allocCount++; if (spaceSize == requestSpace) taskData->allocSize = taskData->allocSize*2; taskData->allocLimit = space; taskData->allocPointer = space+spaceSize; // Actually allocate the object taskData->allocPointer -= words; #ifdef POLYML32IN64 ASSERT((uintptr_t)taskData->allocPointer & 4); // Must be odd-word aligned #endif return taskData->allocPointer; } } // It's possible that another thread has requested a GC in which case // we will have memory when that happens. We don't want to start // another GC. if (! singleThreaded) { PLocker locker(&schedLock); if (threadRequest != 0) { ThreadReleaseMLMemoryWithSchedLock(taskData); ThreadUseMLMemoryWithSchedLock(taskData); continue; // Try again } } // Try garbage-collecting. If this failed return 0. if (! QuickGC(taskData, words)) { extern FILE *polyStderr; if (! triedInterrupt) { triedInterrupt = true; fprintf(polyStderr,"Run out of store - interrupting threads\n"); if (debugOptions & DEBUG_THREADS) Log("THREAD: Run out of store, interrupting threads\n"); BroadcastInterrupt(); try { if (ProcessAsynchRequests(taskData)) return 0; // Has been interrupted. } catch(KillException &) { // The thread may have been killed. ThreadExit(taskData); } // Not interrupted: pause this thread to allow for other // interrupted threads to free something. #if defined(_WIN32) Sleep(5000); #else sleep(5); #endif // Try again. } else { // That didn't work. Exit. fprintf(polyStderr,"Failed to recover - exiting\n"); RequestProcessExit(1); // Begins the shutdown process ThreadExit(taskData); // And terminate this thread. } } // Try again. There should be space now. } } } #ifdef _MSC_VER // Don't tell me that exitThread has a non-void type. #pragma warning(disable:4646) #endif Handle exitThread(TaskData *taskData) /* A call to this is put on the stack of a new thread so when the thread function returns the thread goes away. */ { processesModule.ThreadExit(taskData); } // Terminate the current thread. Never returns. POLYUNSIGNED PolyThreadKillSelf(PolyObject *threadId) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); // Possibly not needed since we never return processesModule.ThreadExit(taskData); return 0; } /* Called when a thread is about to block, usually because of IO. If this is interruptable (currently only used for Posix functions) the process will be set to raise an exception if any signal is handled. It may also raise an exception if another thread has called broadcastInterrupt. */ void Processes::ThreadPauseForIO(TaskData *taskData, Waiter *pWait) { TestAnyEvents(taskData); // Consider this a blocking call that may raise Interrupt ThreadReleaseMLMemory(taskData); globalStats.incCount(PSC_THREADS_WAIT_IO); pWait->Wait(1000); // Wait up to a second globalStats.decCount(PSC_THREADS_WAIT_IO); ThreadUseMLMemory(taskData); TestAnyEvents(taskData); // Check if we've been interrupted. } // Default waiter: simply wait for the time. In Unix it may be woken // up by a signal. void Waiter::Wait(unsigned maxMillisecs) { // Since this is used only when we can't monitor the source directly // we set this to 10ms so that we're not waiting too long. if (maxMillisecs > 10) maxMillisecs = 10; #if (defined(_WIN32) && ! defined(__CYGWIN__)) Sleep(maxMillisecs); #else // Unix fd_set read_fds, write_fds, except_fds; struct timeval toWait = { 0, 0 }; toWait.tv_sec = maxMillisecs / 1000; toWait.tv_usec = (maxMillisecs % 1000) * 1000; FD_ZERO(&read_fds); FD_ZERO(&write_fds); FD_ZERO(&except_fds); select(FD_SETSIZE, &read_fds, &write_fds, &except_fds, &toWait); #endif } static Waiter defWait; Waiter *Waiter::defaultWaiter = &defWait; #ifdef HAVE_WINDOWS_H // Wait for the specified handle to be signalled. void WaitHandle::Wait(unsigned maxMillisecs) { // Wait until we get input or we're woken up. if (m_Handle == NULL) Sleep(maxMillisecs); else WaitForSingleObject(m_Handle, maxMillisecs); } #endif #if (!defined(_WIN32) || defined(__CYGWIN__)) // Unix and Cygwin: Wait for a file descriptor on input. void WaitInputFD::Wait(unsigned maxMillisecs) { fd_set read_fds, write_fds, except_fds; struct timeval toWait = { 0, 0 }; toWait.tv_sec = maxMillisecs / 1000; toWait.tv_usec = (maxMillisecs % 1000) * 1000; FD_ZERO(&read_fds); if (m_waitFD >= 0) FD_SET(m_waitFD, &read_fds); FD_ZERO(&write_fds); FD_ZERO(&except_fds); select(FD_SETSIZE, &read_fds, &write_fds, &except_fds, &toWait); } #endif // Get the task data for the current thread. This is held in // thread-local storage. Normally this is passed in taskData but // in a few cases this isn't available. TaskData *Processes::GetTaskDataForThread(void) { #ifdef HAVE_PTHREAD return (TaskData *)pthread_getspecific(tlsId); #elif defined(HAVE_WINDOWS_H) return (TaskData *)TlsGetValue(tlsId); #else // If there's no threading. return taskArray[0]; #endif } // Called to create a task data object in the current thread. // This is currently only used if a thread created in foreign code calls // a callback. TaskData *Processes::CreateNewTaskData(Handle threadId, Handle threadFunction, Handle args, PolyWord flags) { TaskData *taskData = machineDependent->CreateTaskData(); #if defined(HAVE_WINDOWS_H) HANDLE thisProcess = GetCurrentProcess(); DuplicateHandle(thisProcess, GetCurrentThread(), thisProcess, &(taskData->threadHandle), THREAD_ALL_ACCESS, FALSE, 0); #endif unsigned thrdIndex; { PLocker lock(&schedLock); // See if there's a spare entry in the array. for (thrdIndex = 0; thrdIndex < taskArray.size() && taskArray[thrdIndex] != 0; thrdIndex++); if (thrdIndex == taskArray.size()) // Need to expand the array { try { taskArray.push_back(taskData); } catch (std::bad_alloc&) { delete(taskData); throw MemoryException(); } } else { taskArray[thrdIndex] = taskData; } } taskData->stack = gMem.NewStackSpace(machineDependent->InitialStackSize()); if (taskData->stack == 0) { delete(taskData); throw MemoryException(); } // TODO: Check that there isn't a problem if we try to allocate // memory here and result in a GC. taskData->InitStackFrame(taskData, threadFunction, args); ThreadUseMLMemory(taskData); // If the forking thread has created an ML thread object use that // otherwise create a new one in the current context. if (threadId != 0) taskData->threadObject = (ThreadObject*)threadId->WordP(); else { + // Make a thread reference to point to this taskData object. + Handle threadRef = MakeVolatileWord(taskData, taskData); + // Make a thread object. Since it's in the thread table it can't be garbage collected. taskData->threadObject = (ThreadObject*)alloc(taskData, sizeof(ThreadObject)/sizeof(PolyWord), F_MUTABLE_BIT); - taskData->threadObject->index = TAGGED(thrdIndex); // Set to the index + taskData->threadObject->threadRef = threadRef->Word(); taskData->threadObject->flags = flags != TAGGED(0) ? TAGGED(PFLAG_SYNCH): flags; taskData->threadObject->threadLocal = TAGGED(0); // Empty thread-local store taskData->threadObject->requestCopy = TAGGED(0); // Cleared interrupt state taskData->threadObject->mlStackSize = TAGGED(0); // Unlimited stack size for (unsigned i = 0; i < sizeof(taskData->threadObject->debuggerSlots)/sizeof(PolyWord); i++) taskData->threadObject->debuggerSlots[i] = TAGGED(0); } #ifdef HAVE_PTHREAD initThreadSignals(taskData); pthread_setspecific(tlsId, taskData); #elif defined(HAVE_WINDOWS_H) TlsSetValue(tlsId, taskData); #endif globalStats.incCount(PSC_THREADS); return taskData; } // This function is run when a new thread has been forked. The // parameter is the taskData value for the new thread. This function // is also called directly for the main thread. #ifdef HAVE_PTHREAD static void *NewThreadFunction(void *parameter) { TaskData *taskData = (TaskData *)parameter; #ifdef HAVE_WINDOWS_H // Cygwin: Get the Windows thread handle in case it's needed for profiling. HANDLE thisProcess = GetCurrentProcess(); DuplicateHandle(thisProcess, GetCurrentThread(), thisProcess, &(taskData->threadHandle), THREAD_ALL_ACCESS, FALSE, 0); #endif initThreadSignals(taskData); pthread_setspecific(processesModule.tlsId, taskData); taskData->saveVec.init(); // Remove initial data globalStats.incCount(PSC_THREADS); processes->ThreadUseMLMemory(taskData); try { (void)taskData->EnterPolyCode(); // Will normally (always?) call ExitThread. } catch (KillException &) { processesModule.ThreadExit(taskData); } return 0; } #elif defined(HAVE_WINDOWS_H) static DWORD WINAPI NewThreadFunction(void *parameter) { TaskData *taskData = (TaskData *)parameter; TlsSetValue(processesModule.tlsId, taskData); taskData->saveVec.init(); // Removal initial data globalStats.incCount(PSC_THREADS); processes->ThreadUseMLMemory(taskData); try { (void)taskData->EnterPolyCode(); } catch (KillException &) { processesModule.ThreadExit(taskData); } return 0; } #else static void NewThreadFunction(void *parameter) { TaskData *taskData = (TaskData *)parameter; initThreadSignals(taskData); taskData->saveVec.init(); // Removal initial data globalStats.incCount(PSC_THREADS); processes->ThreadUseMLMemory(taskData); try { (void)taskData->EnterPolyCode(); } catch (KillException &) { processesModule.ThreadExit(taskData); } } #endif // Sets up the initial thread from the root function. This is run on // the initial thread of the process so it will work if we don't // have pthreads. // When multithreading this thread also deals with all garbage-collection // and similar operations and the ML threads send it requests to deal with // that. These require all the threads to pause until the operation is complete // since they affect all memory but they are also sometimes highly recursive. // On Mac OS X and on Linux if the stack limit is set to unlimited only the // initial thread has a large stack and newly created threads have smaller // stacks. We need to make sure that any significant stack usage occurs only // on the inital thread. void Processes::BeginRootThread(PolyObject *rootFunction) { int exitLoopCount = 100; // Maximum 100 * 400 ms. if (taskArray.size() < 1) { try { taskArray.push_back(0); } catch (std::bad_alloc&) { ::Exit("Unable to create the initial thread - insufficient memory"); } } try { // We can't use ForkThread because we don't have a taskData object before we start TaskData *taskData = machineDependent->CreateTaskData(); + Handle threadRef = MakeVolatileWord(taskData, taskData); taskData->threadObject = (ThreadObject*)alloc(taskData, sizeof(ThreadObject) / sizeof(PolyWord), F_MUTABLE_BIT); - taskData->threadObject->index = TAGGED(0); // Index 0 + taskData->threadObject->threadRef = threadRef->Word(); // The initial thread is set to accept broadcast interrupt requests // and handle them synchronously. This is for backwards compatibility. taskData->threadObject->flags = TAGGED(PFLAG_BROADCAST|PFLAG_ASYNCH); // Flags taskData->threadObject->threadLocal = TAGGED(0); // Empty thread-local store taskData->threadObject->requestCopy = TAGGED(0); // Cleared interrupt state taskData->threadObject->mlStackSize = TAGGED(0); // Unlimited stack size for (unsigned i = 0; i < sizeof(taskData->threadObject->debuggerSlots)/sizeof(PolyWord); i++) taskData->threadObject->debuggerSlots[i] = TAGGED(0); #if defined(HAVE_WINDOWS_H) taskData->threadHandle = mainThreadHandle; #endif taskArray[0] = taskData; taskData->stack = gMem.NewStackSpace(machineDependent->InitialStackSize()); if (taskData->stack == 0) ::Exit("Unable to create the initial thread - insufficient memory"); taskData->InitStackFrame(taskData, taskData->saveVec.push(rootFunction), (Handle)0); // Create a packet for the Interrupt exception once so that we don't have to // allocate when we need to raise it. // We can only do this once the taskData object has been created. if (interrupt_exn == 0) interrupt_exn = makeExceptionPacket(taskData, EXC_interrupt); if (singleThreaded) { // If we don't have threading enter the code as if this were a new thread. // This will call finish so will never return. NewThreadFunction(taskData); } schedLock.Lock(); int errorCode = 0; #ifdef HAVE_PTHREAD if (pthread_create(&taskData->threadId, NULL, NewThreadFunction, taskData) != 0) errorCode = errno; #elif defined(HAVE_WINDOWS_H) taskData->threadHandle = CreateThread(NULL, 0, NewThreadFunction, taskData, 0, NULL); if (taskData->threadHandle == NULL) errorCode = GetLastError(); #endif if (errorCode != 0) { // Thread creation failed. taskArray[0] = 0; delete(taskData); ExitWithError("Unable to create initial thread:", errorCode); } if (debugOptions & DEBUG_THREADS) Log("THREAD: Forked initial root thread %p\n", taskData); } catch (std::bad_alloc &) { ::Exit("Unable to create the initial thread - insufficient memory"); } // Wait until the threads terminate or make a request. // We only release schedLock while waiting. while (1) { // Look at the threads to see if they are running. bool allStopped = true; bool noUserThreads = true; bool signalThreadRunning = false; for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; if (p) { if (p == sigTask) signalThreadRunning = true; else if (! p->threadExited) noUserThreads = false; if (p->inMLHeap) { allStopped = false; // It must be running - interrupt it if we are waiting. if (threadRequest != 0) p->InterruptCode(); } else if (p->threadExited) // Has the thread terminated? { // Wait for it to actually stop then delete the task data. #ifdef HAVE_PTHREAD pthread_join(p->threadId, NULL); #elif defined(HAVE_WINDOWS_H) WaitForSingleObject(p->threadHandle, INFINITE); #endif - delete(p); + // The thread ref is no longer valid. + *(TaskData**)(p->threadObject->threadRef.AsObjPtr()) = 0; + delete(p); // Delete the task Data *i = 0; globalStats.decCount(PSC_THREADS); } } } if (noUserThreads) { // If all threads apart from the signal thread have exited then // we can finish but we must make sure that the signal thread has // exited before we finally finish and deallocate the memory. if (signalThreadRunning) exitRequest = true; else break; // Really no threads. } if (allStopped && threadRequest != 0) { mainThreadPhase = threadRequest->mtp; gMem.ProtectImmutable(false); // GC, sharing and export may all write to the immutable area threadRequest->Perform(); gMem.ProtectImmutable(true); mainThreadPhase = MTP_USER_CODE; threadRequest->completed = true; threadRequest = 0; // Allow a new request. mlThreadWait.Signal(); } // Have we had a request to stop? This may have happened while in the GC. if (exitRequest) { // Set this to kill the threads. for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *taskData = *i; if (taskData && taskData->requests != kRequestKill) MakeRequest(taskData, kRequestKill); } // Leave exitRequest set so that if we're in the process of // creating a new thread we will request it to stop when the // taskData object has been added to the table. } // Now release schedLock and wait for a thread // to wake us up or for the timer to expire to update the statistics. if (! initialThreadWait.WaitFor(&schedLock, 400)) { // We didn't receive a request in the last 400ms if (exitRequest) { if (--exitLoopCount < 0) { // The loop count has expired and there is at least one thread that hasn't exited. // Assume we've deadlocked. #if defined(HAVE_WINDOWS_H) ExitProcess(1); #else _exit(1); // Something is stuck. Get out without calling destructors. #endif } } } // Update the periodic stats. // Calculate the free memory. We have to be careful here because although // we have the schedLock we don't have any lock that prevents a thread // from allocating a new segment. Since these statistics are only // very rough it doesn't matter if there's a glitch. // One possibility would be see if the value of // gMem.GetFreeAllocSpace() has changed from what it was at the // start and recalculate if it has. // We also count the number of threads in ML code. Taking the // lock in EnterPolyCode on every RTS call turned out to be // expensive. uintptr_t freeSpace = 0; unsigned threadsInML = 0; for (std::vector::iterator j = taskArray.begin(); j != taskArray.end(); j++) { TaskData *taskData = *j; if (taskData) { // This gets the values last time it was in the RTS. PolyWord *limit = taskData->allocLimit, *ptr = taskData->allocPointer; if (limit < ptr && (uintptr_t)(ptr-limit) < taskData->allocSize) freeSpace += ptr-limit; if (taskData->inML) threadsInML++; } } // Add the space in the allocation areas after calculating the sizes for the // threads in case a thread has allocated some more. freeSpace += gMem.GetFreeAllocSpace(); globalStats.updatePeriodicStats(freeSpace, threadsInML); } schedLock.Unlock(); finish(exitResult); // Close everything down and exit. } // Create a new thread. Returns the ML thread identifier object if it succeeds. // May raise an exception. Handle Processes::ForkThread(TaskData *taskData, Handle threadFunction, Handle args, PolyWord flags, PolyWord stacksize) { if (singleThreaded) raise_exception_string(taskData, EXC_thread, "Threads not available"); try { // Create a taskData object for the new thread TaskData *newTaskData = machineDependent->CreateTaskData(); // We allocate the thread object in the PARENT's space + Handle threadRef = MakeVolatileWord(taskData, newTaskData); Handle threadId = alloc_and_save(taskData, sizeof(ThreadObject) / sizeof(PolyWord), F_MUTABLE_BIT); newTaskData->threadObject = (ThreadObject*)DEREFHANDLE(threadId); - newTaskData->threadObject->index = TAGGED(0); + newTaskData->threadObject->threadRef = threadRef->Word(); newTaskData->threadObject->flags = flags; // Flags newTaskData->threadObject->threadLocal = TAGGED(0); // Empty thread-local store newTaskData->threadObject->requestCopy = TAGGED(0); // Cleared interrupt state newTaskData->threadObject->mlStackSize = stacksize; for (unsigned i = 0; i < sizeof(newTaskData->threadObject->debuggerSlots)/sizeof(PolyWord); i++) newTaskData->threadObject->debuggerSlots[i] = TAGGED(0); unsigned thrdIndex; schedLock.Lock(); // Before forking a new thread check to see whether we have been asked // to exit. Processes::Exit sets the current set of threads to exit but won't // see a new thread. if (taskData->requests == kRequestKill) { schedLock.Unlock(); // Raise an exception although the thread may exit before we get there. raise_exception_string(taskData, EXC_thread, "Thread is exiting"); } // See if there's a spare entry in the array. for (thrdIndex = 0; thrdIndex < taskArray.size() && taskArray[thrdIndex] != 0; thrdIndex++); if (thrdIndex == taskArray.size()) // Need to expand the array { try { taskArray.push_back(newTaskData); } catch (std::bad_alloc&) { delete(newTaskData); schedLock.Unlock(); raise_exception_string(taskData, EXC_thread, "Too many threads"); } } else { taskArray[thrdIndex] = newTaskData; } - newTaskData->threadObject->index = TAGGED(thrdIndex); // Set to the index schedLock.Unlock(); newTaskData->stack = gMem.NewStackSpace(machineDependent->InitialStackSize()); if (newTaskData->stack == 0) { delete(newTaskData); raise_exception_string(taskData, EXC_thread, "Unable to allocate thread stack"); } // Allocate anything needed for the new stack in the parent's heap. // The child still has inMLHeap set so mustn't GC. newTaskData->InitStackFrame(taskData, threadFunction, args); // Now actually fork the thread. bool success = false; schedLock.Lock(); #ifdef HAVE_PTHREAD success = pthread_create(&newTaskData->threadId, NULL, NewThreadFunction, newTaskData) == 0; #elif defined(HAVE_WINDOWS_H) newTaskData->threadHandle = CreateThread(NULL, 0, NewThreadFunction, newTaskData, 0, NULL); success = newTaskData->threadHandle != NULL; #endif if (success) { schedLock.Unlock(); if (debugOptions & DEBUG_THREADS) Log("THREAD: Forking new thread %p from thread %p\n", newTaskData, taskData); return threadId; } // Thread creation failed. taskArray[thrdIndex] = 0; delete(newTaskData); schedLock.Unlock(); if (debugOptions & DEBUG_THREADS) Log("THREAD: Fork from thread %p failed\n", taskData); raise_exception_string(taskData, EXC_thread, "Thread creation failed"); } catch (std::bad_alloc &) { raise_exception_string(taskData, EXC_thread, "Insufficient memory"); } } // ForkFromRTS. Creates a new thread from within the RTS. This is currently used // only to run a signal function. bool Processes::ForkFromRTS(TaskData *taskData, Handle proc, Handle arg) { try { (void)ForkThread(taskData, proc, arg, TAGGED(PFLAG_SYNCH), TAGGED(0)); return true; } catch (IOException &) { // If it failed return false; } } POLYUNSIGNED PolyThreadForkThread(PolyObject *threadId, PolyWord function, PolyWord attrs, PolyWord stack) { TaskData *taskData = TaskData::FindTaskForId(threadId); ASSERT(taskData != 0); taskData->PreRTSCall(); Handle reset = taskData->saveVec.mark(); Handle pushedFunction = taskData->saveVec.push(function); Handle result = 0; try { result = processesModule.ForkThread(taskData, pushedFunction, (Handle)0, attrs, stack); } catch (KillException &) { processes->ThreadExit(taskData); // TestSynchronousRequests may test for kill } catch (...) { } // If an ML exception is raised taskData->saveVec.reset(reset); taskData->PostRTSCall(); if (result == 0) return TAGGED(0).AsUnsigned(); else return result->Word().AsUnsigned(); } // Deal with any interrupt or kill requests. bool Processes::ProcessAsynchRequests(TaskData *taskData) { bool wasInterrupted = false; TaskData *ptaskData = taskData; schedLock.Lock(); switch (ptaskData->requests) { case kRequestNone: schedLock.Unlock(); break; case kRequestInterrupt: { // Handle asynchronous interrupts only. // We've been interrupted. POLYUNSIGNED attrs = ThreadAttrs(ptaskData); POLYUNSIGNED intBits = attrs & PFLAG_INTMASK; if (intBits == PFLAG_ASYNCH || intBits == PFLAG_ASYNCH_ONCE) { if (intBits == PFLAG_ASYNCH_ONCE) { // Set this so from now on it's synchronous. // This word is only ever set by the thread itself so // we don't need to synchronise. attrs = (attrs & (~PFLAG_INTMASK)) | PFLAG_SYNCH; ptaskData->threadObject->flags = TAGGED(attrs); } ptaskData->requests = kRequestNone; // Clear this ptaskData->threadObject->requestCopy = TAGGED(0); // And in the ML copy schedLock.Unlock(); // Don't actually throw the exception here. taskData->SetException(interrupt_exn); wasInterrupted = true; } else schedLock.Unlock(); } break; case kRequestKill: // The thread has been asked to stop. schedLock.Unlock(); throw KillException(); // Doesn't return. } #ifndef HAVE_WINDOWS_H // Start the profile timer if needed. if (profileMode == kProfileTime) { if (! ptaskData->runningProfileTimer) { ptaskData->runningProfileTimer = true; StartProfilingTimer(); } } else ptaskData->runningProfileTimer = false; // The timer will be stopped next time it goes off. #endif return wasInterrupted; } // If this thread is processing interrupts synchronously and has been // interrupted clear the interrupt and raise the exception. This is // called from IO routines which may block. void Processes::TestSynchronousRequests(TaskData *taskData) { TaskData *ptaskData = taskData; schedLock.Lock(); switch (ptaskData->requests) { case kRequestNone: schedLock.Unlock(); break; case kRequestInterrupt: { // Handle synchronous interrupts only. // We've been interrupted. POLYUNSIGNED attrs = ThreadAttrs(ptaskData); POLYUNSIGNED intBits = attrs & PFLAG_INTMASK; if (intBits == PFLAG_SYNCH) { ptaskData->requests = kRequestNone; // Clear this ptaskData->threadObject->requestCopy = TAGGED(0); schedLock.Unlock(); taskData->SetException(interrupt_exn); throw IOException(); } else schedLock.Unlock(); } break; case kRequestKill: // The thread has been asked to stop. schedLock.Unlock(); throw KillException(); // Doesn't return. } } // Check for asynchronous or synchronous events void Processes::TestAnyEvents(TaskData *taskData) { TestSynchronousRequests(taskData); if (ProcessAsynchRequests(taskData)) throw IOException(); } // Request that the process should exit. // This will usually be called from an ML thread as a result of // a call to OS.Process.exit but on Windows it can be called from the GUI thread. void Processes::RequestProcessExit(int n) { if (singleThreaded) finish(n); exitResult = n; exitRequest = true; PLocker lock(&schedLock); // Lock so we know the main thread is waiting initialThreadWait.Signal(); // Wake it if it's sleeping. } /******************************************************************************/ /* */ /* catchVTALRM - handler for alarm-clock signal */ /* */ /******************************************************************************/ #if !defined(HAVE_WINDOWS_H) // N.B. This may be called either by an ML thread or by the main thread. // On the main thread taskData will be null. static void catchVTALRM(SIG_HANDLER_ARGS(sig, context)) { ASSERT(sig == SIGVTALRM); if (profileMode != kProfileTime) { // We stop the timer for this thread on the next signal after we end profile static struct itimerval stoptime = {{0, 0}, {0, 0}}; /* Stop the timer */ setitimer(ITIMER_VIRTUAL, & stoptime, NULL); } else { TaskData *taskData = processes->GetTaskDataForThread(); handleProfileTrap(taskData, (SIGNALCONTEXT*)context); } } #else /* Windows including Cygwin */ // This runs as a separate thread. Every millisecond it checks the CPU time used // by each ML thread and increments the count for each thread that has used a // millisecond of CPU time. static bool testCPUtime(HANDLE hThread, LONGLONG &lastCPUTime) { FILETIME cTime, eTime, kTime, uTime; // Try to get the thread CPU time if possible. This isn't supported // in Windows 95/98 so if it fails we just include this thread anyway. if (GetThreadTimes(hThread, &cTime, &eTime, &kTime, &uTime)) { LONGLONG totalTime = 0; LARGE_INTEGER li; li.LowPart = kTime.dwLowDateTime; li.HighPart = kTime.dwHighDateTime; totalTime += li.QuadPart; li.LowPart = uTime.dwLowDateTime; li.HighPart = uTime.dwHighDateTime; totalTime += li.QuadPart; if (totalTime - lastCPUTime >= 10000) { lastCPUTime = totalTime; return true; } return false; } else return true; // Failed to get thread time, maybe Win95. } void Processes::ProfileInterrupt(void) { // Wait for millisecond or until the stop event is signalled. while (WaitForSingleObject(hStopEvent, 1) == WAIT_TIMEOUT) { // We need to hold schedLock to examine the taskArray but // that is held during garbage collection. if (schedLock.Trylock()) { for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *p = *i; if (p && p->threadHandle) { if (testCPUtime(p->threadHandle, p->lastCPUTime)) { CONTEXT context; SuspendThread(p->threadHandle); context.ContextFlags = CONTEXT_CONTROL; /* Get Eip and Esp */ if (GetThreadContext(p->threadHandle, &context)) { handleProfileTrap(p, &context); } ResumeThread(p->threadHandle); } } } schedLock.Unlock(); } // Check the CPU time used by the main thread. This is used for GC // so we need to check that as well. if (testCPUtime(mainThreadHandle, lastCPUTime)) handleProfileTrap(NULL, NULL); } } DWORD WINAPI ProfilingTimer(LPVOID parm) { processesModule.ProfileInterrupt(); return 0; } #endif // Profiling control. Called by the root thread. void Processes::StartProfiling(void) { #ifdef HAVE_WINDOWS_H DWORD threadId; extern FILE *polyStdout; if (profilingHd) return; ResetEvent(hStopEvent); profilingHd = CreateThread(NULL, 0, ProfilingTimer, NULL, 0, &threadId); if (profilingHd == NULL) fputs("Creating ProfilingTimer thread failed.\n", polyStdout); /* Give this a higher than normal priority so it pre-empts the main thread. Without this it will tend only to be run when the main thread blocks for some reason. */ SetThreadPriority(profilingHd, THREAD_PRIORITY_ABOVE_NORMAL); #else // In Linux, at least, we need to run a timer in each thread. // We request each to enter the RTS so that it will start the timer. // Since this is being run by the main thread while all the ML threads // are paused this may not actually be necessary. for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { TaskData *taskData = *i; if (taskData) { taskData->InterruptCode(); } } StartProfilingTimer(); // Start the timer in the root thread. #endif } void Processes::StopProfiling(void) { #ifdef HAVE_WINDOWS_H if (hStopEvent) SetEvent(hStopEvent); // Wait for the thread to stop if (profilingHd) WaitForSingleObject(profilingHd, 10000); CloseHandle(profilingHd); profilingHd = NULL; #endif } // Called by the ML signal handling thread. It blocks until a signal // arrives. There should only be a single thread waiting here. bool Processes::WaitForSignal(TaskData *taskData, PLock *sigLock) { TaskData *ptaskData = taskData; // We need to hold the signal lock until we have acquired schedLock. schedLock.Lock(); sigLock->Unlock(); if (sigTask != 0) { schedLock.Unlock(); return false; } sigTask = ptaskData; if (ptaskData->requests == kRequestNone) { // Now release the ML memory. A GC can start. ThreadReleaseMLMemoryWithSchedLock(ptaskData); globalStats.incCount(PSC_THREADS_WAIT_SIGNAL); ptaskData->threadLock.Wait(&schedLock); globalStats.decCount(PSC_THREADS_WAIT_SIGNAL); // We want to use the memory again. ThreadUseMLMemoryWithSchedLock(ptaskData); } sigTask = 0; schedLock.Unlock(); return true; } // Called by the signal detection thread to wake up the signal handler // thread. Must be called AFTER releasing sigLock. void Processes::SignalArrived(void) { PLocker locker(&schedLock); if (sigTask) sigTask->threadLock.Signal(); } #ifdef HAVE_PTHREAD // This is called when the thread exits in foreign code and // ThreadExit has not been called. static void threaddata_destructor(void *p) { TaskData *pt = (TaskData *)p; pt->threadExited = true; // This doesn't actually wake the main thread and relies on the // regular check to release the task data. } #endif void Processes::Init(void) { #ifdef HAVE_PTHREAD pthread_key_create(&tlsId, threaddata_destructor); #elif defined(HAVE_WINDOWS_H) tlsId = TlsAlloc(); #else singleThreaded = true; #endif #if defined(HAVE_WINDOWS_H) /* Windows including Cygwin. */ // Create stop event for time profiling. hStopEvent = CreateEvent(NULL, TRUE, FALSE, NULL); // Get the thread handle for this thread. HANDLE thisProcess = GetCurrentProcess(); DuplicateHandle(thisProcess, GetCurrentThread(), thisProcess, &mainThreadHandle, THREAD_ALL_ACCESS, FALSE, 0); #else // Set up a signal handler. This will be the same for all threads. markSignalInuse(SIGVTALRM); setSignalHandler(SIGVTALRM, catchVTALRM); #endif } #ifndef HAVE_WINDOWS_H // On Linux, at least, each thread needs to run this. void Processes::StartProfilingTimer(void) { // set virtual timer to go off every millisecond struct itimerval starttime; starttime.it_interval.tv_sec = starttime.it_value.tv_sec = 0; starttime.it_interval.tv_usec = starttime.it_value.tv_usec = 1000; setitimer(ITIMER_VIRTUAL,&starttime,NULL); } #endif void Processes::Stop(void) { #ifdef HAVE_PTHREAD pthread_key_delete(tlsId); #elif defined(HAVE_WINDOWS_H) TlsFree(tlsId); #endif #if defined(HAVE_WINDOWS_H) /* Stop the timer and profiling threads. */ if (hStopEvent) SetEvent(hStopEvent); if (profilingHd) { WaitForSingleObject(profilingHd, 10000); CloseHandle(profilingHd); profilingHd = NULL; } if (hStopEvent) CloseHandle(hStopEvent); hStopEvent = NULL; if (mainThreadHandle) CloseHandle(mainThreadHandle); mainThreadHandle = NULL; #else profileMode = kProfileOff; // Make sure the timer is not running struct itimerval stoptime; memset(&stoptime, 0, sizeof(stoptime)); setitimer(ITIMER_VIRTUAL, &stoptime, NULL); #endif } void Processes::GarbageCollect(ScanAddress *process) /* Ensures that all the objects are retained and their addresses updated. */ { /* The interrupt exn */ if (interrupt_exn != 0) { PolyObject *p = interrupt_exn; process->ScanRuntimeAddress(&p, ScanAddress::STRENGTH_STRONG); interrupt_exn = (PolyException*)p; } for (std::vector::iterator i = taskArray.begin(); i != taskArray.end(); i++) { if (*i) (*i)->GarbageCollect(process); } } void TaskData::GarbageCollect(ScanAddress *process) { saveVec.gcScan(process); if (threadObject != 0) { PolyObject *p = threadObject; process->ScanRuntimeAddress(&p, ScanAddress::STRENGTH_STRONG); threadObject = (ThreadObject*)p; } if (blockMutex != 0) process->ScanRuntimeAddress(&blockMutex, ScanAddress::STRENGTH_STRONG); // The allocation spaces are no longer valid. allocPointer = 0; allocLimit = 0; // Divide the allocation size by four. If we have made a single allocation // since the last GC the size will have been doubled after the allocation. // On average for each thread, apart from the one that ran out of space // and requested the GC, half of the space will be unused so reducing by // four should give a good estimate for next time. if (allocCount != 0) { // Do this only once for each GC. allocCount = 0; allocSize = allocSize/4; if (allocSize < MIN_HEAP_SIZE) allocSize = MIN_HEAP_SIZE; } process->ScanRuntimeWord(&foreignStack); } // Return the number of processors. extern unsigned NumberOfProcessors(void) { #if (defined(_WIN32) && ! defined(__CYGWIN__)) SYSTEM_INFO info; memset(&info, 0, sizeof(info)); GetSystemInfo(&info); if (info.dwNumberOfProcessors == 0) // Just in case info.dwNumberOfProcessors = 1; return info.dwNumberOfProcessors; #elif(defined(_SC_NPROCESSORS_ONLN)) long res = sysconf(_SC_NPROCESSORS_ONLN); if (res <= 0) res = 1; return res; #elif(defined(HAVE_SYSCTL) && defined(CTL_HW) && defined(HW_NCPU)) static int mib[2] = { CTL_HW, HW_NCPU }; int nCPU = 1; size_t len = sizeof(nCPU); if (sysctl(mib, 2, &nCPU, &len, NULL, 0) == 0 && len == sizeof(nCPU)) return nCPU; else return 1; #else // Can't determine. return 1; #endif } // Return the number of physical processors. If hyperthreading is // enabled this returns less than NumberOfProcessors. Returns zero if // it cannot be determined. // This can be used in Cygwin as well as native Windows. #if (defined(HAVE_SYSTEM_LOGICAL_PROCESSOR_INFORMATION)) typedef BOOL (WINAPI *GETP)(SYSTEM_LOGICAL_PROCESSOR_INFORMATION*, PDWORD); // Windows - use GetLogicalProcessorInformation if it's available. static unsigned WinNumPhysicalProcessors(void) { GETP getProcInfo = (GETP) GetProcAddress(GetModuleHandle(_T("kernel32")), "GetLogicalProcessorInformation"); if (getProcInfo == 0) return 0; // It's there - use it. SYSTEM_LOGICAL_PROCESSOR_INFORMATION *buff = 0; DWORD space = 0; while (getProcInfo(buff, &space) == FALSE) { if (GetLastError() != ERROR_INSUFFICIENT_BUFFER) { free(buff); return 0; } free(buff); buff = (PSYSTEM_LOGICAL_PROCESSOR_INFORMATION)malloc(space); if (buff == 0) return 0; } // Calculate the number of full entries in case it's truncated. unsigned nItems = space / sizeof(SYSTEM_LOGICAL_PROCESSOR_INFORMATION); unsigned numProcs = 0; for (unsigned i = 0; i < nItems; i++) { if (buff[i].Relationship == RelationProcessorCore) numProcs++; } free(buff); return numProcs; } #endif // Read and parse /proc/cpuinfo static unsigned LinuxNumPhysicalProcessors(void) { // Find out the total. This should be the maximum. unsigned nProcs = NumberOfProcessors(); // If there's only one we don't need to check further. if (nProcs <= 1) return nProcs; long *cpus = (long*)calloc(nProcs, sizeof(long)); if (cpus == 0) return 0; FILE *cpuInfo = fopen("/proc/cpuinfo", "r"); if (cpuInfo == NULL) { free(cpus); return 0; } char line[40]; unsigned count = 0; while (fgets(line, sizeof(line), cpuInfo) != NULL) { if (strncmp(line, "core id\t\t:", 10) == 0) { long n = strtol(line+10, NULL, 10); unsigned i = 0; // Skip this id if we've seen it already while (i < count && cpus[i] != n) i++; if (i == count) cpus[count++] = n; } if (strchr(line, '\n') == 0) { int ch; do { ch = getc(cpuInfo); } while (ch != '\n' && ch != EOF); } } fclose(cpuInfo); free(cpus); return count; } extern unsigned NumberOfPhysicalProcessors(void) { unsigned numProcs = 0; #if (defined(HAVE_SYSTEM_LOGICAL_PROCESSOR_INFORMATION)) numProcs = WinNumPhysicalProcessors(); if (numProcs != 0) return numProcs; #endif #if (defined(HAVE_SYSCTLBYNAME) && defined(HAVE_SYS_SYSCTL_H)) // Mac OS X int nCores; size_t len = sizeof(nCores); if (sysctlbyname("hw.physicalcpu", &nCores, &len, NULL, 0) == 0) return (unsigned)nCores; #endif numProcs = LinuxNumPhysicalProcessors(); if (numProcs != 0) return numProcs; // Any other cases? return numProcs; } diff --git a/libpolyml/processes.h b/libpolyml/processes.h index fd19b337..bdc7533c 100644 --- a/libpolyml/processes.h +++ b/libpolyml/processes.h @@ -1,357 +1,362 @@ /* Title: Lightweight process library Author: David C.J. Matthews - Copyright (c) 2007-8, 2012, 2015, 2017 David C.J. Matthews + Copyright (c) 2007-8, 2012, 2015, 2017, 2019 David C.J. Matthews This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License version 2.1 as published by the Free Software Foundation. This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */ #ifndef _PROCESSES_H_ #define _PROCESSES_H_ #ifdef HAVE_CONFIG_H #include "config.h" #elif defined(_WIN32) #include "winconfig.h" #else #error "No configuration file" #endif #include "globals.h" #include "rts_module.h" #include "save_vec.h" #include "noreturn.h" #include "locking.h" class SaveVecEntry; typedef SaveVecEntry *Handle; class StackSpace; class PolyWord; class ScanAddress; class MDTaskData; class Exporter; class StackObject; #ifdef HAVE_WINDOWS_H typedef void *HANDLE; #endif #ifdef HAVE_SIGNAL_H #include #endif #ifdef HAVE_UCONTEXT_H #include #endif #ifdef HAVE_PTHREAD_H #include #endif // SIGNALCONTEXT is the argument type that is passed to GetPCandSPFromContext // to get the actual PC and SP in a profiling trap. #if defined(HAVE_WINDOWS_H) // First because it's used in both native Windows and Cygwin. #include #define SIGNALCONTEXT CONTEXT // This is the thread context. #elif defined(HAVE_UCONTEXT_T) #define SIGNALCONTEXT ucontext_t #elif defined(HAVE_STRUCT_SIGCONTEXT) #define SIGNALCONTEXT struct sigcontext #else #define SIGNALCONTEXT void #endif #define MIN_HEAP_SIZE 4096 // Minimum and initial heap segment size (words) // This is the ML "thread identifier" object. The fields // are read and set by the ML code. class ThreadObject: public PolyObject { public: - PolyWord index; // Tagged integer with the index into the taskArray - // Not used by ML + PolyWord threadRef; // Weak ref containing the address of the thread data. Not used by ML PolyWord flags; // Tagged integer containing flags indicating how interrupts // are handled. Set by ML but only by the thread itself PolyWord threadLocal; // Head of a list of thread-local store items. // Handled entirely by ML but only by the thread. PolyWord requestCopy; // A tagged integer copy of the "requests" field. // This is provided so that ML can easily test if there // is an interrupt pending. PolyWord mlStackSize; // A tagged integer with the maximum ML stack size in bytes PolyWord debuggerSlots[4]; // These are used by the debugger. }; // Other threads may make requests to a thread. typedef enum { kRequestNone = 0, // Increasing severity kRequestInterrupt = 1, kRequestKill = 2 } ThreadRequests; // Per-thread data. This is subclassed for each architecture. class TaskData { public: TaskData(); virtual ~TaskData(); void FillUnusedSpace(void); virtual void GarbageCollect(ScanAddress *process); virtual Handle EnterPolyCode() = 0; // Start running ML virtual void InterruptCode() = 0; virtual bool AddTimeProfileCount(SIGNALCONTEXT *context) = 0; // Initialise the stack for a new thread. The parent task object is passed in because any // allocation that needs to be made must be made in the parent. virtual void InitStackFrame(TaskData *parentTask, Handle proc, Handle arg) = 0; virtual void SetException(poly_exn *exc) = 0; // If a foreign function calls back to ML we need to set up the call to the // ML callback function. virtual Handle EnterCallbackFunction(Handle func, Handle args) = 0; // The scheduler needs versions of atomic increment and atomic reset that // work in exactly the same way as the code-generated versions (if any). // Atomic decrement isn't needed since it only ever releases a mutex. virtual Handle AtomicIncrement(Handle mutexp) = 0; // Reset a mutex to one. This needs to be atomic with respect to the // atomic increment and decrement instructions. virtual void AtomicReset(Handle mutexp) = 0; virtual void CopyStackFrame(StackObject *old_stack, uintptr_t old_length, StackObject *new_stack, uintptr_t new_length) = 0; virtual uintptr_t currentStackSpace(void) const = 0; // Add a count to the local function if we are using store profiling. virtual void addProfileCount(POLYUNSIGNED words) = 0; // Functions called before and after an RTS call. virtual void PreRTSCall(void) { inML = false; } virtual void PostRTSCall(void) { inML = true; } SaveVec saveVec; PolyWord *allocPointer; // Allocation pointer - decremented towards... PolyWord *allocLimit; // ... lower limit of allocation - POLYUNSIGNED allocSize; // The preferred heap segment size + uintptr_t allocSize; // The preferred heap segment size unsigned allocCount; // The number of allocations since the last GC StackSpace *stack; ThreadObject *threadObject; // Pointer to the thread object. int lastError; // Last error from foreign code. void *signalStack; // Stack to handle interrupts (Unix only) PolyWord foreignStack; // Stack of saved data used in call_sym_and_convert bool inML; // True when this is in ML, false in the RTS // Get a TaskData pointer given the ML taskId. - static TaskData *FindTaskForId(PolyObject *taskId); + // This is called at the start of every RTS function that may allocate memory. + // It is can be called safely to get the thread's own TaskData object without + // a lock but any call to get the TaskData for another thread must take the + // schedLock first in case the thread is exiting. + static TaskData *FindTaskForId(PolyObject *taskId) { + return *(TaskData**)(((ThreadObject*)taskId)->threadRef.AsObjPtr()); + } private: // If a thread has to block it will block on this. PCondVar threadLock; // External requests made are stored here until they // can be actioned. ThreadRequests requests; // Pointer to the mutex when blocked. Set to NULL when it doesn't apply. PolyObject *blockMutex; // This is set to false when a thread blocks or enters foreign code, // While it is true the thread can manipulate ML memory so no other // thread can garbage collect. bool inMLHeap; // In Linux, at least, we need to run a separate timer in each thread bool runningProfileTimer; #ifdef HAVE_WINDOWS_H LONGLONG lastCPUTime; // Used for profiling #endif public: bool threadExited; private: #ifdef HAVE_PTHREAD_H pthread_t threadId; #endif #ifdef HAVE_WINDOWS_H public: // Because, on Cygwin, it's used in NewThreadFunction HANDLE threadHandle; private: #endif friend class Processes; }; NORETURNFN(extern Handle exitThread(TaskData *mdTaskData)); class ScanAddress; // Indicate what the main thread is doing if the profile // timer goes off. extern enum _mainThreadPhase { MTP_USER_CODE=0, MTP_GCPHASESHARING, MTP_GCPHASEMARK, MTP_GCPHASECOMPACT, MTP_GCPHASEUPDATE, MTP_GCQUICK, MTP_SHARING, MTP_EXPORTING, MTP_SAVESTATE, MTP_LOADSTATE, MTP_PROFILING, MTP_SIGHANDLER, MTP_CYGWINSPAWN, MTP_STOREMODULE, MTP_LOADMODULE, MTP_MAXENTRY } mainThreadPhase; // Data structure used for requests from a thread to the root // thread. These are GCs or similar. class MainThreadRequest { public: MainThreadRequest (enum _mainThreadPhase phase): mtp(phase), completed(false) {} virtual ~MainThreadRequest () {} // Suppress silly GCC warning const enum _mainThreadPhase mtp; bool completed; virtual void Perform() = 0; }; class PLock; // Class to wait for a given time or for an event, whichever comes first. // // A pointer to this class or a subclass is passed to ThreadPauseForIO. // Because a thread may be interrupted or killed by another ML thread we // don't allow any thread to block indefinitely. Instead whenever a // thread wants to do an operation that may block we have it enter a // loop that polls for the desired condition and if it is not ready it // calls ThreadPauseForIO. The default action is to block for a short // period and then return so that the caller can poll again. That can // limit performance when, for example, reading from a pipe so where possible // we use a sub-class that waits until either input is available or it times // out, whichever comes first, using "select" in Unix or MsgWaitForMultipleObjects // in Windows. // During a call to Waiter::Wait the thread is set as "not using ML memory" // so a GC can happen while this thread is blocked. class Waiter { public: Waiter() {} virtual ~Waiter() {} virtual void Wait(unsigned maxMillisecs); static Waiter *defaultWaiter; }; #ifdef HAVE_WINDOWS_H class WaitHandle: public Waiter { public: WaitHandle(HANDLE h): m_Handle(h) {} virtual void Wait(unsigned maxMillisecs); private: HANDLE m_Handle; }; #endif #if (! defined(_WIN32) || defined(__CYGWIN__)) // Unix: Wait until a file descriptor is available for input class WaitInputFD: public Waiter { public: WaitInputFD(int fd): m_waitFD(fd) {} virtual void Wait(unsigned maxMillisecs); private: int m_waitFD; }; #endif // External interface to the Process module. These functions are all implemented // by the Processes class. class ProcessExternal { public: virtual ~ProcessExternal() {} // Defined to suppress a warning from GCC virtual TaskData *GetTaskDataForThread(void) = 0; virtual TaskData *CreateNewTaskData(Handle threadId, Handle threadFunction, Handle args, PolyWord flags) = 0; // Request all ML threads to exit and set the result code. Does not cause // the calling thread itself to exit since this may be called on the GUI thread. virtual void RequestProcessExit(int n) = 0; // Exit from this thread. virtual NORETURNFN(void ThreadExit(TaskData *taskData)) = 0; virtual void BroadcastInterrupt(void) = 0; virtual void BeginRootThread(PolyObject *rootFunction) = 0; // Called when a thread may block. Returns some time later when perhaps // the input is available. virtual void ThreadPauseForIO(TaskData *taskData, Waiter *pWait) = 0; // As ThreadPauseForIO but when there is no stream virtual void ThreadPause(TaskData *taskData) { ThreadPauseForIO(taskData, Waiter::defaultWaiter); } // If a thread is blocking for some time it should release its use // of the ML memory. That allows a GC. ThreadUseMLMemory returns true if // a GC was in progress. virtual void ThreadUseMLMemory(TaskData *taskData) = 0; virtual void ThreadReleaseMLMemory(TaskData *taskData) = 0; // Requests from the threads for actions that need to be performed by // the root thread. virtual void MakeRootRequest(TaskData *taskData, MainThreadRequest *request) = 0; // Deal with any interrupt or kill requests. virtual bool ProcessAsynchRequests(TaskData *taskData) = 0; // Process an interrupt request synchronously. virtual void TestSynchronousRequests(TaskData *taskData) = 0; // Process any events, synchronous or asynchronous. virtual void TestAnyEvents(TaskData *taskData) = 0; // ForkFromRTS. Creates a new thread from within the RTS. virtual bool ForkFromRTS(TaskData *taskData, Handle proc, Handle arg) = 0; // Profiling control. virtual void StartProfiling(void) = 0; virtual void StopProfiling(void) = 0; // Find space for an object. Returns a pointer to the start. "words" must include // the length word and the result points at where the length word will go. // If the allocation succeeds it may update the allocation values in the taskData object. // If the heap is exhausted it may set this thread (or other threads) to raise an exception. virtual PolyWord *FindAllocationSpace(TaskData *taskData, POLYUNSIGNED words, bool alwaysInSeg) = 0; // Signal handling support. The ML signal handler thread blocks until it is // woken up by the signal detection thread. virtual bool WaitForSignal(TaskData *taskData, PLock *sigLock) = 0; virtual void SignalArrived(void) = 0; // After a Unix fork we only have a single thread in the new process. virtual void SetSingleThreaded(void) = 0; virtual poly_exn* GetInterrupt(void) = 0; }; // Return the number of processors. Used when configuring multi-threaded GC. extern unsigned NumberOfProcessors(void); extern unsigned NumberOfPhysicalProcessors(void); extern ProcessExternal *processes; extern struct _entrypts processesEPT[]; #endif diff --git a/libpolyml/x86_dep.cpp b/libpolyml/x86_dep.cpp index e11064d8..9b0b17a7 100644 --- a/libpolyml/x86_dep.cpp +++ b/libpolyml/x86_dep.cpp @@ -1,1386 +1,1386 @@ /* Title: Machine dependent code for i386 and X64 under Windows and Unix Copyright (c) 2000-7 Cambridge University Technical Services Limited Further work copyright David C. J. Matthews 2011-19 This library is free software; you can redistribute it and/or modify it under the terms of the GNU Lesser General Public License version 2.1 as published by the Free Software Foundation. This library is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details. You should have received a copy of the GNU Lesser General Public License along with this library; if not, write to the Free Software Foundation, Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA */ #ifdef HAVE_CONFIG_H #include "config.h" #elif defined(_WIN32) #include "winconfig.h" #else #error "No configuration file" #endif #ifdef HAVE_STDLIB_H #include #endif #include #ifdef HAVE_SIGNAL_H #include #endif #ifdef HAVE_ASSERT_H #include #define ASSERT(x) assert(x) #else #define ASSERT(x) #endif #ifdef HAVE_STRING_H #include #endif #ifdef HAVE_ERRNO_H #include #endif #if (defined(_WIN32)) #include #include #endif #include "globals.h" #include "run_time.h" #include "diagnostics.h" #include "processes.h" #include "profiling.h" #include "machine_dep.h" #include "scanaddrs.h" #include "memmgr.h" #include "rtsentry.h" #include "sys.h" // Temporary /********************************************************************** * * Register usage: * * %Reax: First argument to function. Result of function call. * %Rebx: Second argument to function. * %Recx: General register * %Redx: Closure pointer in call. * %Rebp: Points to memory used for extra registers * %Resi: General register. * %Redi: General register. * %Resp: Stack pointer. * The following apply only on the X64 * %R8: Third argument to function * %R9: Fourth argument to function * %R10: Fifth argument to function * %R11: General register * %R12: General register * %R13: General register * %R14: General register * %R15: Memory allocation pointer * **********************************************************************/ #ifdef HOSTARCHITECTURE_X86_64 struct fpSaveArea { double fpregister[7]; // Save area for xmm0-6 }; #else // Structure of floating point save area. // This is dictated by the hardware. typedef byte fpregister[10]; struct fpSaveArea { unsigned short cw; unsigned short _unused0; unsigned short sw; unsigned short _unused1; unsigned short tw; unsigned short _unused2; unsigned fip; unsigned short fcs0; unsigned short _unused3; unsigned foo; unsigned short fcs1; unsigned short _unused4; fpregister registers[8]; }; #endif /* the amount of ML stack space to reserve for registers, C exception handling etc. The compiler requires us to reserve 2 stack-frames worth (2 * 20 words). We actually reserve slightly more than this. */ #if (!defined(_WIN32) && !defined(HAVE_SIGALTSTACK)) // If we can't handle signals on a separate stack make sure there's space // on the Poly stack. #define OVERFLOW_STACK_SIZE (50+1024) #else #define OVERFLOW_STACK_SIZE 50 #endif union stackItem { /* #ifndef POLYML32IN64 stackItem(PolyWord v) { words[0] = v.AsUnsigned(); }; stackItem() { words[0] = TAGGED(0).AsUnsigned(); } POLYUNSIGNED words[1]; #else // In 32-in-64 we need to clear the second PolyWord. This assumes little-endian. stackItem(PolyWord v) { words[0] = v.AsUnsigned(); words[1] = 0; }; stackItem() { words[0] = TAGGED(0).AsUnsigned(); words[1] = 0; } POLYUNSIGNED words[2]; #endif */ stackItem(PolyWord v) { argValue = v.AsUnsigned(); } stackItem() { argValue = TAGGED(0).AsUnsigned(); } // These return the low order word. PolyWord w()const { return PolyWord::FromUnsigned((POLYUNSIGNED)argValue); } operator PolyWord () { return PolyWord::FromUnsigned((POLYUNSIGNED)argValue); } POLYCODEPTR codeAddr; // Return addresses stackItem *stackAddr; // Stack addresses uintptr_t argValue; // Treat an address as an int }; class X86TaskData; // This is passed as the argument vector to X86AsmSwitchToPoly. // The offsets are built into the assembly code and the code-generator. // localMpointer and stackPtr are updated before control returns to C. typedef struct _AssemblyArgs { public: PolyWord *localMpointer; // Allocation ptr + 1 word stackItem *handlerRegister; // Current exception handler PolyWord *localMbottom; // Base of memory + 1 word stackItem *stackLimit; // Lower limit of stack stackItem exceptionPacket; // Set if there is an exception byte unusedRequestCode; // No longer used. byte unusedFlag; // No longer used byte returnReason; // Reason for returning from ML. byte unusedRestore; // No longer used. uintptr_t saveCStack; // Saved C stack frame. PolyWord threadId; // My thread id. Saves having to call into RTS for it. stackItem *stackPtr; // Current stack pointer byte *noLongerUsed; // Now removed byte *heapOverFlowCall; // These are filled in with the functions. byte *stackOverFlowCall; byte *stackOverFlowCallEx; // Saved registers, where applicable. stackItem p_rax; stackItem p_rbx; stackItem p_rcx; stackItem p_rdx; stackItem p_rsi; stackItem p_rdi; #ifdef HOSTARCHITECTURE_X86_64 stackItem p_r8; stackItem p_r9; stackItem p_r10; stackItem p_r11; stackItem p_r12; stackItem p_r13; stackItem p_r14; #endif struct fpSaveArea p_fp; } AssemblyArgs; // These next few are temporarily added for the interpreter // This duplicates some code in reals.cpp but is now updated. #define DOUBLESIZE (sizeof(double)/sizeof(POLYUNSIGNED)) union realdb { double dble; POLYUNSIGNED puns[DOUBLESIZE]; }; #define LGWORDSIZE (sizeof(uintptr_t) / sizeof(PolyWord)) class X86TaskData: public TaskData { public: X86TaskData(); unsigned allocReg; // The register to take the allocated space. POLYUNSIGNED allocWords; // The words to allocate. Handle callBackResult; AssemblyArgs assemblyInterface; int saveRegisterMask; // Registers that need to be updated by a GC. virtual void GarbageCollect(ScanAddress *process); void ScanStackAddress(ScanAddress *process, stackItem &val, StackSpace *stack); virtual Handle EnterPolyCode(); // Start running ML virtual void InterruptCode(); virtual bool AddTimeProfileCount(SIGNALCONTEXT *context); virtual void InitStackFrame(TaskData *parentTask, Handle proc, Handle arg); virtual void SetException(poly_exn *exc); // Release a mutex in exactly the same way as compiler code virtual Handle AtomicIncrement(Handle mutexp); virtual void AtomicReset(Handle mutexp); // Return the minimum space occupied by the stack. Used when setting a limit. // N.B. This is PolyWords not native words. virtual uintptr_t currentStackSpace(void) const { return (this->stack->top - (PolyWord*)assemblyInterface.stackPtr) + OVERFLOW_STACK_SIZE*sizeof(uintptr_t)/sizeof(PolyWord); } // Increment the profile count for an allocation. Also now used for mutex contention. virtual void addProfileCount(POLYUNSIGNED words) { add_count(this, assemblyInterface.stackPtr[0].codeAddr, words); } // PreRTSCall: After calling from ML to the RTS we need to save the current heap pointer virtual void PreRTSCall(void) { TaskData::PreRTSCall(); SaveMemRegisters(); } // PostRTSCall: Before returning we need to restore the heap pointer. // If there has been a GC in the RTS call we need to create a new heap area. virtual void PostRTSCall(void) { SetMemRegisters(); TaskData::PostRTSCall(); } virtual void CopyStackFrame(StackObject *old_stack, uintptr_t old_length, StackObject *new_stack, uintptr_t new_length); virtual Handle EnterCallbackFunction(Handle func, Handle args); int SwitchToPoly(); void HeapOverflowTrap(byte *pcPtr); void SetMemRegisters(); void SaveMemRegisters(); void SetRegisterMask(); PLock interruptLock; stackItem *get_reg(int n); stackItem *®SP() { return assemblyInterface.stackPtr; } stackItem ®AX() { return assemblyInterface.p_rax; } stackItem ®BX() { return assemblyInterface.p_rbx; } stackItem ®CX() { return assemblyInterface.p_rcx; } stackItem ®DX() { return assemblyInterface.p_rdx; } stackItem ®SI() { return assemblyInterface.p_rsi; } stackItem ®DI() { return assemblyInterface.p_rdi; } #ifdef HOSTARCHITECTURE_X86_64 stackItem ®8() { return assemblyInterface.p_r8; } stackItem ®9() { return assemblyInterface.p_r9; } stackItem ®10() { return assemblyInterface.p_r10; } stackItem ®11() { return assemblyInterface.p_r11; } stackItem ®12() { return assemblyInterface.p_r12; } stackItem ®13() { return assemblyInterface.p_r13; } stackItem ®14() { return assemblyInterface.p_r14; } #endif #if (defined(_WIN32) && !defined(__CYGWIN__)) DWORD savedErrno; #else int savedErrno; #endif }; class X86Dependent: public MachineDependent { public: X86Dependent() {} // Create a task data object. virtual TaskData *CreateTaskData(void) { return new X86TaskData(); } // Initial size of stack in PolyWords virtual unsigned InitialStackSize(void) { return (128+OVERFLOW_STACK_SIZE) * sizeof(uintptr_t) / sizeof(PolyWord); } virtual void ScanConstantsWithinCode(PolyObject *addr, PolyObject *oldAddr, POLYUNSIGNED length, ScanAddress *process); virtual Architectures MachineArchitecture(void) #ifndef HOSTARCHITECTURE_X86_64 { return MA_I386; } #elif defined(POLYML32IN64) { return MA_X86_64_32; } #else { return MA_X86_64; } #endif }; // Values for the returnReason byte enum RETURN_REASON { RETURN_IO_CALL_NOW_UNUSED = 0, RETURN_HEAP_OVERFLOW = 1, RETURN_STACK_OVERFLOW = 2, RETURN_STACK_OVERFLOWEX = 3, RETURN_CALLBACK_RETURN = 6, RETURN_CALLBACK_EXCEPTION = 7, RETURN_KILL_SELF = 9 }; extern "C" { // These are declared in the assembly code segment. void X86AsmSwitchToPoly(void *); extern int X86AsmKillSelf(void); extern int X86AsmCallbackReturn(void); extern int X86AsmCallbackException(void); extern int X86AsmPopArgAndClosure(void); extern int X86AsmRaiseException(void); extern int X86AsmCallExtraRETURN_HEAP_OVERFLOW(void); extern int X86AsmCallExtraRETURN_STACK_OVERFLOW(void); extern int X86AsmCallExtraRETURN_STACK_OVERFLOWEX(void); POLYUNSIGNED X86AsmAtomicIncrement(PolyObject*); POLYUNSIGNED X86AsmAtomicDecrement(PolyObject*); }; X86TaskData::X86TaskData(): allocReg(0), allocWords(0), saveRegisterMask(0) { assemblyInterface.heapOverFlowCall = (byte*)X86AsmCallExtraRETURN_HEAP_OVERFLOW; assemblyInterface.stackOverFlowCall = (byte*)X86AsmCallExtraRETURN_STACK_OVERFLOW; assemblyInterface.stackOverFlowCallEx = (byte*)X86AsmCallExtraRETURN_STACK_OVERFLOWEX; savedErrno = 0; } void X86TaskData::GarbageCollect(ScanAddress *process) { TaskData::GarbageCollect(process); // Process the parent first assemblyInterface.threadId = threadObject; if (stack != 0) { // Now the values on the stack. for (stackItem *q = assemblyInterface.stackPtr; q < (stackItem*)stack->top; q++) ScanStackAddress(process, *q, stack); } // Register mask for (int i = 0; i < 16; i++) { if (saveRegisterMask & (1 << i)) ScanStackAddress(process, *get_reg(i), stack); } } // Process a value within the stack. void X86TaskData::ScanStackAddress(ScanAddress *process, stackItem &stackItem, StackSpace *stack) { // We may have return addresses on the stack which could look like // tagged values. Check whether the value is in the code area before // checking whether it is untagged. #ifdef POLYML32IN64 // In 32-in-64 return addresses always have the top 32 bits non-zero. if (stackItem.argValue < ((uintptr_t)1 << 32)) { // It's either a tagged integer or an object pointer. if (stackItem.w().IsDataPtr()) { PolyWord val = process->ScanObjectAddress(stackItem.w().AsObjPtr()); stackItem = val; } } else { // Could be a code address or a stack address. MemSpace *space = gMem.SpaceForAddress(stackItem.codeAddr - 1); if (space == 0 || space->spaceType != ST_CODE) return; PolyObject *obj = gMem.FindCodeObject(stackItem.codeAddr); ASSERT(obj != 0); // Process the address of the start. Don't update anything. process->ScanObjectAddress(obj); } #else // The -1 here is because we may have a zero-sized cell in the last // word of a space. MemSpace *space = gMem.SpaceForAddress(stackItem.codeAddr-1); if (space == 0) return; if (space->spaceType == ST_CODE) { PolyObject *obj = gMem.FindCodeObject(stackItem.codeAddr); // If it is actually an integer it might be outside a valid code object. if (obj == 0) { ASSERT(stackItem.w().IsTagged()); // It must be an integer } else // Process the address of the start. Don't update anything. process->ScanObjectAddress(obj); } else if (space->spaceType == ST_LOCAL && stackItem.w().IsDataPtr()) // Local values must be word addresses. { PolyWord val = process->ScanObjectAddress(stackItem.w().AsObjPtr()); stackItem = val; } #endif } // Copy a stack void X86TaskData::CopyStackFrame(StackObject *old_stack, uintptr_t old_length, StackObject *new_stack, uintptr_t new_length) { /* Moves a stack, updating all references within the stack */ #ifdef POLYML32IN64 old_length = old_length / 2; new_length = new_length / 2; #endif stackItem *old_base = (stackItem *)old_stack; stackItem *new_base = (stackItem*)new_stack; stackItem *old_top = old_base + old_length; /* Calculate the offset of the new stack from the old. If the frame is being extended objects in the new frame will be further up the stack than in the old one. */ uintptr_t offset = new_base - old_base + new_length - old_length; stackItem *oldStackPtr = assemblyInterface.stackPtr; // Adjust the stack pointer and handler pointer since these point into the stack. assemblyInterface.stackPtr = assemblyInterface.stackPtr + offset; assemblyInterface.handlerRegister = assemblyInterface.handlerRegister + offset; // We need to adjust any values on the stack that are pointers within the stack. // Skip the unused part of the stack. size_t i = oldStackPtr - old_base; ASSERT (i <= old_length); i = old_length - i; stackItem *old = oldStackPtr; stackItem *newp = assemblyInterface.stackPtr; while (i--) { stackItem old_word = *old++; if (old_word.w().IsDataPtr() && old_word.stackAddr >= old_base && old_word.stackAddr <= old_top) old_word.stackAddr = old_word.stackAddr + offset; else if (old_word.w().IsDataPtr() && IsHeapAddress(old_word.stackAddr)) { stackItem *addr = (stackItem*)old_word.w().AsStackAddr(); if (addr >= old_base && addr <= old_top) { addr += offset; old_word = PolyWord::FromStackAddr((PolyWord*)addr); } } *newp++ = old_word; } ASSERT(old == ((stackItem*)old_stack)+old_length); ASSERT(newp == ((stackItem*)new_stack)+new_length); // And change any registers that pointed into the old stack for (int j = 0; j < 16; j++) { if (saveRegisterMask & (1 << j)) { stackItem *regAddr = get_reg(j); stackItem old_word = *regAddr; if (old_word.w().IsDataPtr() && old_word.stackAddr >= old_base && old_word.stackAddr <= old_top) old_word.stackAddr = old_word.stackAddr + offset; else if (old_word.w().IsDataPtr() && IsHeapAddress(old_word.stackAddr)) { stackItem *addr = (stackItem*)old_word.w().AsStackAddr(); if (addr >= old_base && addr <= old_top) { addr += offset; old_word = PolyWord::FromStackAddr((PolyWord*)addr); } } *regAddr = old_word; } } } Handle X86TaskData::EnterPolyCode() /* Called from "main" to enter the code. */ { Handle hOriginal = this->saveVec.mark(); // Set this up for the IO calls. while (1) { this->saveVec.reset(hOriginal); // Remove old RTS arguments and results. // Run the ML code and return with the function to call. this->inML = true; int ioFunction = SwitchToPoly(); this->inML = false; try { switch (ioFunction) { case -1: // We've been interrupted. This usually involves simulating a // stack overflow so we could come here because of a genuine // stack overflow. // Previously this code was executed on every RTS call but there // were problems on Mac OS X at least with contention on schedLock. // Process any asynchronous events i.e. interrupts or kill processes->ProcessAsynchRequests(this); // Release and re-acquire use of the ML memory to allow another thread // to GC. processes->ThreadReleaseMLMemory(this); processes->ThreadUseMLMemory(this); break; case -2: // A callback has returned. return callBackResult; // Return the saved value. Not used in the new interface. default: Crash("Unknown io operation %d\n", ioFunction); } } catch (IOException &) { } } } // Run the current ML process. X86AsmSwitchToPoly saves the C state so that // whenever the ML requires assistance from the rest of the RTS it simply // returns to C with the appropriate values set in assemblyInterface.requestCode and // int X86TaskData::SwitchToPoly() // (Re)-enter the Poly code from C. Returns with the io function to call or // -1 if we are responding to an interrupt. { Handle mark = this->saveVec.mark(); do { this->saveVec.reset(mark); // Remove old data e.g. from arbitrary precision. SetMemRegisters(); // We need to save the C stack entry across this call in case // we're making a callback and the previous C stack entry is // for the original call. uintptr_t savedCStack = this->assemblyInterface.saveCStack; // Restore the saved error state. #if (defined(_WIN32) && !defined(__CYGWIN__)) SetLastError(savedErrno); #else errno = savedErrno; #endif // Enter the ML code. X86AsmSwitchToPoly(&this->assemblyInterface); this->assemblyInterface.saveCStack = savedCStack; // Save the error codes. We may have made an RTS/FFI call that // has set these and we don't want to do anything to change them. #if (defined(_WIN32) && !defined(__CYGWIN__)) savedErrno = GetLastError(); #else savedErrno = errno; #endif SaveMemRegisters(); // Update globals from the memory registers. // Handle any heap/stack overflows or arbitrary precision traps. switch (this->assemblyInterface.returnReason) { case RETURN_HEAP_OVERFLOW: // The heap has overflowed. SetRegisterMask(); this->HeapOverflowTrap(assemblyInterface.stackPtr[0].codeAddr); // Computes a value for allocWords only break; case RETURN_STACK_OVERFLOW: case RETURN_STACK_OVERFLOWEX: { SetRegisterMask(); uintptr_t min_size; // Size in PolyWords if (assemblyInterface.returnReason == RETURN_STACK_OVERFLOW) { min_size = (this->stack->top - (PolyWord*)assemblyInterface.stackPtr) + OVERFLOW_STACK_SIZE * sizeof(uintptr_t) / sizeof(PolyWord); } else { // Stack limit overflow. If the required stack space is larger than // the fixed overflow size the code will calculate the limit in %EDI. stackItem *stackP = regDI().stackAddr; min_size = (this->stack->top - (PolyWord*)stackP) + OVERFLOW_STACK_SIZE * sizeof(uintptr_t) / sizeof(PolyWord); } try { // The stack check has failed. This may either be because we really have // overflowed the stack or because the stack limit value has been adjusted // to result in a call here. CheckAndGrowStack(this, min_size); } catch (IOException &) { // We may get an exception while handling this if we run out of store } { PLocker l(&interruptLock); // Set the stack limit. This clears any interrupt and also sets the // correct value if we've grown the stack. this->assemblyInterface.stackLimit = (stackItem*)this->stack->bottom + OVERFLOW_STACK_SIZE; } return -1; // We're in a safe state to handle any interrupts. } case RETURN_CALLBACK_RETURN: // regSP has been set by the assembly code. N.B. This may not be the same value as when // EnterCallbackFunction was called because the callback may have grown and moved the stack. // Remove the extra exception handler we created in EnterCallbackFunction ASSERT(assemblyInterface.handlerRegister == regSP()); regSP() += 1; assemblyInterface.handlerRegister = (*(regSP()++)).stackAddr; // Restore the previous handler. this->callBackResult = this->saveVec.push(regAX()); // Argument to return is in RAX. return -2; case RETURN_CALLBACK_EXCEPTION: // An ML callback has raised an exception. // It isn't possible to do anything here except abort. Crash("An ML function called from foreign code raised an exception. Unable to continue."); case RETURN_KILL_SELF: exitThread(this); default: Crash("Unknown return reason code %u", this->assemblyInterface.returnReason); } } while (1); } void X86TaskData::InitStackFrame(TaskData *parentTaskData, Handle proc, Handle arg) /* Initialise stack frame. */ { StackSpace *space = this->stack; StackObject * newStack = space->stack(); uintptr_t stack_size = space->spaceSize() * sizeof(PolyWord) / sizeof(stackItem); uintptr_t topStack = stack_size-6; stackItem *stackTop = (stackItem*)newStack + topStack; assemblyInterface.stackPtr = stackTop; assemblyInterface.stackLimit = (stackItem*)space->bottom + OVERFLOW_STACK_SIZE; assemblyInterface.handlerRegister = (stackItem*)newStack+topStack+4; // Floating point save area. memset(&assemblyInterface.p_fp, 0, sizeof(struct fpSaveArea)); #ifndef HOSTARCHITECTURE_X86_64 // Set the control word for 64-bit precision otherwise we get inconsistent results. assemblyInterface.p_fp.cw = 0x027f ; // Control word assemblyInterface.p_fp.tw = 0xffff; // Tag registers - all unused #endif // Initial entry point - on the stack. stackTop[0].codeAddr = (byte*)&X86AsmPopArgAndClosure; // Push the argument and the closure on the stack. We can't put them into the registers // yet because we might get a GC before we actually start the code. stackTop[1] = proc->Word(); // Closure stackTop[2] = (arg == 0) ? TAGGED(0) : DEREFWORD(arg); // Argument /* We initialise the end of the stack with a sequence that will jump to kill_self whether the process ends with a normal return or by raising an exception. A bit of this was added to fix a bug when stacks were objects on the heap and could be scanned by the GC. */ stackTop[5] = TAGGED(0); // Probably no longer needed // Set the default handler and return address to point to this code. // PolyWord killJump(PolyWord::FromCodePtr((byte*)&X86AsmKillSelf)); // Exception handler. stackTop[4].codeAddr = (byte*)&X86AsmKillSelf; // Normal return address. We need a separate entry on the stack from // the exception handler because it is possible that the code we are entering // may replace this entry with an argument. The code-generator optimises tail-recursive // calls to functions with more args than the called function. stackTop[3].codeAddr = (byte*)&X86AsmKillSelf; #ifdef POLYML32IN64 // In 32-in-64 RBX always contains the heap base address. assemblyInterface.p_rbx.stackAddr = (stackItem*)globalHeapBase; #endif } // In Solaris-x86 the registers are named EIP and ESP. #if (!defined(REG_EIP) && defined(EIP)) #define REG_EIP EIP #endif #if (!defined(REG_ESP) && defined(ESP)) #define REG_ESP ESP #endif // Get the PC and SP(stack) from a signal context. This is needed for profiling. // This version gets the actual sp and pc if we are in ML. bool X86TaskData::AddTimeProfileCount(SIGNALCONTEXT *context) { stackItem * sp = 0; POLYCODEPTR pc = 0; if (context != 0) { // The tests for HAVE_UCONTEXT_T, HAVE_STRUCT_SIGCONTEXT and HAVE_WINDOWS_H need // to follow the tests in processes.h. #if defined(HAVE_WINDOWS_H) #ifdef _WIN64 sp = (stackItem *)context->Rsp; pc = (POLYCODEPTR)context->Rip; #else // Windows 32 including cygwin. sp = (stackItem *)context->Esp; pc = (POLYCODEPTR)context->Eip; #endif #elif defined(HAVE_UCONTEXT_T) #ifdef HAVE_MCONTEXT_T_GREGS // Linux #ifndef HOSTARCHITECTURE_X86_64 pc = (byte*)context->uc_mcontext.gregs[REG_EIP]; sp = (stackItem*)context->uc_mcontext.gregs[REG_ESP]; #else /* HOSTARCHITECTURE_X86_64 */ pc = (byte*)context->uc_mcontext.gregs[REG_RIP]; sp = (stackItem*)context->uc_mcontext.gregs[REG_RSP]; #endif /* HOSTARCHITECTURE_X86_64 */ #elif defined(HAVE_MCONTEXT_T_MC_ESP) // FreeBSD #ifndef HOSTARCHITECTURE_X86_64 pc = (byte*)context->uc_mcontext.mc_eip; sp = (stackItem*)context->uc_mcontext.mc_esp; #else /* HOSTARCHITECTURE_X86_64 */ pc = (byte*)context->uc_mcontext.mc_rip; sp = (stackItem*)context->uc_mcontext.mc_rsp; #endif /* HOSTARCHITECTURE_X86_64 */ #else // Mac OS X #ifndef HOSTARCHITECTURE_X86_64 #if(defined(HAVE_STRUCT_MCONTEXT_SS)||defined(HAVE_STRUCT___DARWIN_MCONTEXT32_SS)) pc = (byte*)context->uc_mcontext->ss.eip; sp = (PolyWord*)context->uc_mcontext->ss.esp; #elif(defined(HAVE_STRUCT___DARWIN_MCONTEXT32___SS)) pc = (byte*)context->uc_mcontext->__ss.__eip; sp = (PolyWord*)context->uc_mcontext->__ss.__esp; #endif #else /* HOSTARCHITECTURE_X86_64 */ #if(defined(HAVE_STRUCT_MCONTEXT_SS)||defined(HAVE_STRUCT___DARWIN_MCONTEXT64_SS)) pc = (byte*)context->uc_mcontext->ss.rip; sp = (PolyWord*)context->uc_mcontext->ss.rsp; #elif(defined(HAVE_STRUCT___DARWIN_MCONTEXT64___SS)) pc = (byte*)context->uc_mcontext->__ss.__rip; sp = (PolyWord*)context->uc_mcontext->__ss.__rsp; #endif #endif /* HOSTARCHITECTURE_X86_64 */ #endif #elif defined(HAVE_STRUCT_SIGCONTEXT) #if defined(HOSTARCHITECTURE_X86_64) && defined(__OpenBSD__) // CPP defines missing in amd64/signal.h in OpenBSD pc = (byte*)context->sc_rip; sp = (PolyWord*)context->sc_rsp; #else // !HOSTARCHITEXTURE_X86_64 || !defined(__OpenBSD__) pc = (byte*)context->sc_pc; sp = (PolyWord*)context->sc_sp; #endif #endif } if (pc != 0) { // See if the PC we've got is an ML code address. MemSpace *space = gMem.SpaceForAddress(pc); if (space != 0 && (space->spaceType == ST_CODE || space->spaceType == ST_PERMANENT)) { add_count(this, pc, 1); return true; } } // See if the sp value is in the current stack. if (sp >= (stackItem*)this->stack->bottom && sp < (stackItem*)this->stack->top) { // We may be in the assembly code. The top of the stack will be a return address. pc = sp[0].w().AsCodePtr(); MemSpace *space = gMem.SpaceForAddress(pc); if (space != 0 && (space->spaceType == ST_CODE || space->spaceType == ST_PERMANENT)) { add_count(this, pc, 1); return true; } } // See if the value of regSP is a valid stack pointer. // This works if we happen to be in an RTS call using a "Full" call. // It doesn't work if we've used a "Fast" call because that doesn't save the SP. sp = assemblyInterface.stackPtr; if (sp >= (stackItem*)this->stack->bottom && sp < (stackItem*)this->stack->top) { // We may be in the run-time system. pc = sp[0].w().AsCodePtr(); MemSpace *space = gMem.SpaceForAddress(pc); if (space != 0 && (space->spaceType == ST_CODE || space->spaceType == ST_PERMANENT)) { add_count(this, pc, 1); return true; } } // None of those worked return false; } // This is called from a different thread so we have to be careful. void X86TaskData::InterruptCode() { PLocker l(&interruptLock); // Set the stack limit pointer to the top of the stack to cause // a trap when we next check for stack overflow. // We use a lock here to ensure that we always use the current value of the // stack. The thread we're interrupting could be growing the stack at this point. if (this->stack != 0) this->assemblyInterface.stackLimit = (stackItem*)(this->stack->top-1); } // This is called from SwitchToPoly before we enter the ML code. void X86TaskData::SetMemRegisters() { // Copy the current store limits into variables before we go into the assembly code. // If we haven't yet set the allocation area or we don't have enough we need // to create one (or a new one). if (this->allocPointer <= this->allocLimit + this->allocWords) { if (this->allocPointer < this->allocLimit) Crash ("Bad length in heap overflow trap"); // Find some space to allocate in. Updates taskData->allocPointer and // returns a pointer to the newly allocated space (if allocWords != 0) PolyWord *space = processes->FindAllocationSpace(this, this->allocWords, true); if (space == 0) { // We will now raise an exception instead of returning. // Set allocWords to zero so we don't set the allocation register // since that could be holding the exception packet. this->allocWords = 0; } // Undo the allocation just now. this->allocPointer += this->allocWords; } if (this->allocWords != 0) { // If we have had a heap trap we actually do the allocation here. // We will have already garbage collected and recovered sufficient space. // This also happens if we have just trapped because of store profiling. this->allocPointer -= this->allocWords; // Now allocate // Set the allocation register to this area. N.B. This is an absolute address. if (this->allocReg < 15) get_reg(this->allocReg)[0].codeAddr = (POLYCODEPTR)(this->allocPointer + 1); /* remember: it's off-by-one */ this->allocWords = 0; } // If we have run out of store, either just above or while allocating in the RTS, // allocPointer and allocLimit will have been set to zero as part of the GC. We will // now be raising an exception which may free some store but we need to come back here // before we allocate anything. The compiled code uses unsigned arithmetic to check for // heap overflow but only after subtracting the space required. We need to make sure // that the values are still non-negative after substracting any object size. if (this->allocPointer == 0) this->allocPointer += MAX_OBJECT_SIZE; if (this->allocLimit == 0) this->allocLimit += MAX_OBJECT_SIZE; this->assemblyInterface.localMbottom = this->allocLimit + 1; this->assemblyInterface.localMpointer = this->allocPointer + 1; // If we are profiling store allocation we set mem_hl so that a trap // will be generated. if (profileMode == kProfileStoreAllocation) this->assemblyInterface.localMbottom = this->assemblyInterface.localMpointer; this->assemblyInterface.returnReason = RETURN_IO_CALL_NOW_UNUSED; this->assemblyInterface.threadId = this->threadObject; } // This is called whenever we have returned from ML to C. void X86TaskData::SaveMemRegisters() { this->allocPointer = this->assemblyInterface.localMpointer - 1; this->allocWords = 0; this->assemblyInterface.exceptionPacket = TAGGED(0); this->saveRegisterMask = 0; } // Called on a GC or stack overflow trap. The register mask // is in the bytes after the trap call. void X86TaskData::SetRegisterMask() { byte *pc = assemblyInterface.stackPtr[0].codeAddr; if (*pc == 0xcd) // CD - INT n is used for a single byte { pc++; saveRegisterMask = *pc++; } else if (*pc == 0xca) // CA - FAR RETURN is used for a two byte mask { pc++; saveRegisterMask = pc[0] | (pc[1] << 8); pc += 2; } assemblyInterface.stackPtr[0].codeAddr = pc; } stackItem *X86TaskData::get_reg(int n) /* Returns a pointer to the register given by n. */ { switch (n) { case 0: return &assemblyInterface.p_rax; case 1: return &assemblyInterface.p_rcx; case 2: return &assemblyInterface.p_rdx; case 3: return &assemblyInterface.p_rbx; // Should not have rsp or rbp. case 6: return &assemblyInterface.p_rsi; case 7: return &assemblyInterface.p_rdi; #ifdef HOSTARCHITECTURE_X86_64 case 8: return &assemblyInterface.p_r8; case 9: return &assemblyInterface.p_r9; case 10: return &assemblyInterface.p_r10; case 11: return &assemblyInterface.p_r11; case 12: return &assemblyInterface.p_r12; case 13: return &assemblyInterface.p_r13; case 14: return &assemblyInterface.p_r14; // R15 is the heap pointer so shouldn't occur here. #endif /* HOSTARCHITECTURE_X86_64 */ default: Crash("Unknown register %d\n", n); } } // Called as a result of a heap overflow trap void X86TaskData::HeapOverflowTrap(byte *pcPtr) { X86TaskData *mdTask = this; POLYUNSIGNED wordsNeeded = 0; // The next instruction, after any branches round forwarding pointers or pop // instructions, will be a store of register containing the adjusted heap pointer. // We need to find that register and the value in it in order to find out how big // the area we actually wanted is. N.B. The code-generator and assembly code // must generate the correct instruction sequence. // byte *pcPtr = assemblyInterface.programCtr; while (true) { if (pcPtr[0] == 0xeb) { // Forwarding pointer if (pcPtr[1] >= 128) pcPtr += 256 - pcPtr[1] + 2; else pcPtr += pcPtr[1] + 2; } else if ((pcPtr[0] & 0xf8) == 0x58) // Pop instruction. pcPtr++; else if (pcPtr[0] == 0x41 && ((pcPtr[1] & 0xf8) == 0x58)) // Pop with Rex prefix pcPtr += 2; else break; } #ifndef HOSTARCHITECTURE_X86_64 // This should be movl REG,0[%ebp]. ASSERT(pcPtr[0] == 0x89); mdTask->allocReg = (pcPtr[1] >> 3) & 7; // Remember this until we allocate the memory stackItem *reg = get_reg(mdTask->allocReg); stackItem reg_val = *reg; // The space we need is the difference between this register // and the current value of newptr. // The +1 here is because assemblyInterface.localMpointer is A.M.pointer +1. The reason // is that after the allocation we have the register pointing at the address we will // actually use. wordsNeeded = (this->allocPointer - (PolyWord*)reg_val.stackAddr) + 1; *reg = TAGGED(0); // Clear this - it's not a valid address. /* length in words, including length word */ ASSERT (wordsNeeded <= (1<<24)); /* Max object size including length/flag word is 2^24 words. */ #else /* HOSTARCHITECTURE_X86_64 */ ASSERT(pcPtr[1] == 0x89 || pcPtr[1] == 0x8b); if (pcPtr[1] == 0x89) { // New (5.4) format. This should be movq REG,%r15 ASSERT(pcPtr[0] == 0x49 || pcPtr[0] == 0x4d); mdTask->allocReg = (pcPtr[2] >> 3) & 7; // Remember this until we allocate the memory if (pcPtr[0] & 0x4) mdTask->allocReg += 8; } else { // Alternative form of movq REG,%r15 ASSERT(pcPtr[0] == 0x4c || pcPtr[0] == 0x4d); mdTask->allocReg = pcPtr[2] & 7; // Remember this until we allocate the memory if (pcPtr[0] & 0x1) mdTask->allocReg += 8; } stackItem *reg = get_reg(this->allocReg); stackItem reg_val = *reg; wordsNeeded = (POLYUNSIGNED)((this->allocPointer - (PolyWord*)reg_val.stackAddr) + 1); *reg = TAGGED(0); // Clear this - it's not a valid address. #endif /* HOSTARCHITECTURE_X86_64 */ if (profileMode == kProfileStoreAllocation) addProfileCount(wordsNeeded); mdTask->allocWords = wordsNeeded; // The actual allocation is done in SetMemRegisters. } void X86TaskData::SetException(poly_exn *exc) // Set up the stack to raise an exception. { // Do we need to set the PC value any longer? It may be necessary if // we have taken a trap because another thread has sent a broadcast interrupt. (--assemblyInterface.stackPtr)->codeAddr = (byte*)X86AsmRaiseException; regAX() = (PolyWord)exc; /* put exception data into eax */ assemblyInterface.exceptionPacket = (PolyWord)exc; // Set for direct calls. } // Sets up a callback function on the current stack. The present state is that // the ML code has made a call in to foreign_dispatch. We need to set the stack // up so that we will enter the callback (as with CallCodeTupled) but when we return // the result we enter callback_return. Handle X86TaskData::EnterCallbackFunction(Handle func, Handle args) { // If we ever implement a light version of the FFI that allows a call to C // code without saving enough to allow allocation in C code we need to ensure // that this code doesn't do any allocation. Essentially we need the values // in localMpointer and localMbottom to be valid across a call to C. If we do // a callback the ML callback function would pick up the values saved in the // originating call. // However, it is essential that the light version still saves the stack pointer // and reloads it afterwards. // Set up an exception handler so we will enter callBackException if there is an exception. (--regSP())->stackAddr = assemblyInterface.handlerRegister; // Create a special handler entry - *(--regSP()) = PolyWord::FromCodePtr((byte*)&X86AsmCallbackException); + (--regSP())->codeAddr = (byte*)&X86AsmCallbackException; assemblyInterface.handlerRegister = regSP(); // Push the call to callBackReturn onto the stack as the return address. - *(--regSP()) = PolyWord::FromCodePtr((byte*)&X86AsmCallbackReturn); + (--regSP())->codeAddr = (byte*)&X86AsmCallbackReturn; // Set up the entry point of the callback. PolyObject *functToCall = func->WordP(); regDX() = (PolyWord)functToCall; // Closure address regAX() = args->Word(); // Push entry point address - *(--regSP()) = functToCall->Get(0); // First word of closure is entry pt. + (--regSP())->codeAddr = *(POLYCODEPTR*)functToCall; // First word of closure is entry pt. return EnterPolyCode(); } // Decode and process an effective address. There may // be a constant address in here but in any case we need // to decode it to work out where the next instruction starts. // If this is an lea instruction any addresses are just constants // so must not be treated as addresses. static void skipea(PolyObject *base, byte **pt, ScanAddress *process, bool lea) { unsigned int modrm = *((*pt)++); unsigned int md = modrm >> 6; unsigned int rm = modrm & 7; if (md == 3) { } /* Register. */ else if (rm == 4) { /* s-i-b present. */ unsigned int sib = *((*pt)++); if (md == 0) { if ((sib & 7) == 5) { if (! lea) { #ifndef HOSTARCHITECTURE_X86_64 process->ScanConstant(base, *pt, PROCESS_RELOC_DIRECT); #endif /* HOSTARCHITECTURE_X86_64 */ } (*pt) += 4; } } else if (md == 1) (*pt)++; else if (md == 2) (*pt) += 4; } else if (md == 0 && rm == 5) { if (!lea) { #ifndef HOSTARCHITECTURE_X86_64 /* Absolute address. */ process->ScanConstant(base, *pt, PROCESS_RELOC_DIRECT); #endif /* HOSTARCHITECTURE_X86_64 */ } *pt += 4; } else { if (md == 1) *pt += 1; else if (md == 2) *pt += 4; } } /* Added to deal with constants within the code rather than in the constant area. The constant area is still needed for the function name. DCJM 2/1/2001 */ void X86Dependent::ScanConstantsWithinCode(PolyObject *addr, PolyObject *old, POLYUNSIGNED length, ScanAddress *process) { byte *pt = (byte*)addr; PolyWord *end = addr->Offset(length - 1); #ifdef POLYML32IN64 // If this begins with enter-int it's interpreted code - ignore if (pt[0] == 0xff && pt[1] == 0x55 && pt[2] == 0x48) return; #endif while (true) { // Escape prefixes come before any Rex byte if (*pt == 0xf2 || *pt == 0xf3 || *pt == 0x66) pt++; #ifdef HOSTARCHITECTURE_X86_64 // REX prefixes. Set this first. byte lastRex; if (*pt >= 0x40 && *pt <= 0x4f) lastRex = *pt++; else lastRex = 0; //printf("pt=%p *pt=%x\n", pt, *pt); #endif /* HOSTARCHITECTURE_X86_64 */ switch (*pt) { case 0x00: return; // This is actually the first byte of the old "marker" word. case 0xf4: return; // Halt - now used as a marker. case 0x50: case 0x51: case 0x52: case 0x53: case 0x54: case 0x55: case 0x56: case 0x57: /* Push */ case 0x58: case 0x59: case 0x5a: case 0x5b: case 0x5c: case 0x5d: case 0x5e: case 0x5f: /* Pop */ case 0x90: /* nop */ case 0xc3: /* ret */ case 0xf9: /* stc */ case 0xce: /* into */ case 0xf0: /* lock. */ case 0xf3: /* rep/repe */ case 0xa4: case 0xa5: case 0xaa: case 0xab: /* movs/stos */ case 0xa6: /* cmpsb */ case 0x9e: /* sahf */ case 0x99: /* cqo/cdq */ pt++; break; case 0x70: case 0x71: case 0x72: case 0x73: case 0x74: case 0x75: case 0x76: case 0x77: case 0x78: case 0x79: case 0x7a: case 0x7b: case 0x7c: case 0x7d: case 0x7e: case 0x7f: case 0xeb: /* short jumps. */ case 0xcd: /* INT - now used for a register mask */ case 0xa8: /* TEST_ACC8 */ case 0x6a: /* PUSH_8 */ pt += 2; break; case 0xc2: /* RET_16 */ case 0xca: /* FAR RET 16 - used for a register mask */ pt += 3; break; case 0x8d: /* leal. */ pt++; skipea(addr, &pt, process, true); break; case 0x03: case 0x0b: case 0x13: case 0x1b: case 0x23: case 0x2b: case 0x33: case 0x3b: /* Add r,ea etc. */ case 0x88: /* MOVB_R_A */ case 0x89: /* MOVL_R_A */ case 0x8b: /* MOVL_A_R */ case 0x62: /* BOUNDL */ case 0xff: /* Group5 */ case 0xd1: /* Group2_1_A */ case 0x8f: /* POP_A */ case 0xd3: /* Group2_CL_A */ case 0x87: // XCHNG case 0x63: // MOVSXD pt++; skipea(addr, &pt, process, false); break; case 0xf6: /* Group3_a */ { int isTest = 0; pt++; /* The test instruction has an immediate operand. */ if ((*pt & 0x38) == 0) isTest = 1; skipea(addr, &pt, process, false); if (isTest) pt++; break; } case 0xf7: /* Group3_A */ { int isTest = 0; pt++; /* The test instruction has an immediate operand. */ if ((*pt & 0x38) == 0) isTest = 1; skipea(addr, &pt, process, false); if (isTest) pt += 4; break; } case 0xc1: /* Group2_8_A */ case 0xc6: /* MOVB_8_A */ case 0x83: /* Group1_8_A */ case 0x80: /* Group1_8_a */ case 0x6b: // IMUL Ev,Ib pt++; skipea(addr, &pt, process, false); pt++; break; case 0x69: // IMUL Ev,Iv pt++; skipea(addr, &pt, process, false); pt += 4; break; case 0x81: /* Group1_32_A */ { pt ++; #ifndef HOSTARCHITECTURE_X86_64 unsigned opCode = *pt; #endif skipea(addr, &pt, process, false); // Only check the 32 bit constant if this is a comparison. // For other operations this may be untagged and shouldn't be an address. #ifndef HOSTARCHITECTURE_X86_64 if ((opCode & 0x38) == 0x38) process->ScanConstant(addr, pt, PROCESS_RELOC_DIRECT); #endif pt += 4; break; } case 0xe8: case 0xe9: // Long jump and call. These are used to call constant (known) functions // and also long jumps within the function. { pt++; POLYSIGNED disp = (pt[3] & 0x80) ? -1 : 0; // Set the sign just in case. for(unsigned i = 4; i > 0; i--) disp = (disp << 8) | pt[i-1]; byte *absAddr = pt + disp + 4; // The address is relative to AFTER the constant // If the new address is within the current piece of code we don't do anything if (absAddr >= (byte*)addr && absAddr < (byte*)end) {} else { #ifdef HOSTARCHITECTURE_X86_64 ASSERT(sizeof(PolyWord) == 4); // Should only be used internally on x64 #endif /* HOSTARCHITECTURE_X86_64 */ if (addr != old) { // The old value of the displacement was relative to the old address before // we copied this code segment. // We have to correct it back to the original address. absAddr = absAddr - (byte*)addr + (byte*)old; // We have to correct the displacement for the new location and store // that away before we call ScanConstant. size_t newDisp = absAddr - pt - 4; for (unsigned i = 0; i < 4; i++) { pt[i] = (byte)(newDisp & 0xff); newDisp >>= 8; } } process->ScanConstant(addr, pt, PROCESS_RELOC_I386RELATIVE); } pt += 4; break; } case 0xc7:/* MOVL_32_A */ { pt++; if ((*pt & 0xc0) == 0x40 /* Byte offset or sib present */ && ((*pt & 7) != 4) /* But not sib present */ && pt[1] == 256-sizeof(PolyWord)) { /* We may use a move instruction to set the length word on a new segment. We mustn't try to treat this as a constant. */ pt += 6; /* Skip the modrm byte, the offset and the constant. */ } else { skipea(addr, &pt, process, false); #ifndef HOSTARCHITECTURE_X86_64 // This isn't used for addresses even in 32-in-64 process->ScanConstant(addr, pt, PROCESS_RELOC_DIRECT); #endif /* HOSTARCHITECTURE_X86_64 */ pt += 4; } break; } case 0xb8: case 0xb9: case 0xba: case 0xbb: case 0xbc: case 0xbd: case 0xbe: case 0xbf: /* MOVL_32_64_R */ pt ++; #ifdef HOSTARCHITECTURE_X86_64 if ((lastRex & 8) == 0) pt += 4; // 32-bit mode on 64-bits else #endif /* HOSTARCHITECTURE_X86_64 */ { // This is no longer generated in 64-bit mode but needs to // be retained in native 64-bit for backwards compatibility. #ifndef POLYML32IN64 // 32 bits in 32-bit mode, 64-bits in 64-bit mode. process->ScanConstant(addr, pt, PROCESS_RELOC_DIRECT); #endif pt += sizeof(PolyWord); } break; case 0x68: /* PUSH_32 */ pt ++; #if (!defined(HOSTARCHITECTURE_X86_64) || defined(POLYML32IN64)) // Currently the only inline constant in 32-in-64. process->ScanConstant(addr, pt, PROCESS_RELOC_DIRECT); #endif pt += 4; break; case 0x0f: /* ESCAPE */ { pt++; switch (*pt) { case 0xb6: /* movzl */ case 0xb7: // movzw case 0xc1: /* xaddl */ case 0xae: // ldmxcsr/stmxcsr case 0xaf: // imul case 0x40: case 0x41: case 0x42: case 0x43: case 0x44: case 0x45: case 0x46: case 0x47: case 0x48: case 0x49: case 0x4a: case 0x4b: case 0x4c: case 0x4d: case 0x4e: case 0x4f: // cmov pt++; skipea(addr, &pt, process, false); break; case 0x80: case 0x81: case 0x82: case 0x83: case 0x84: case 0x85: case 0x86: case 0x87: case 0x88: case 0x89: case 0x8a: case 0x8b: case 0x8c: case 0x8d: case 0x8e: case 0x8f: /* Conditional branches with 32-bit displacement. */ pt += 5; break; case 0x90: case 0x91: case 0x92: case 0x93: case 0x94: case 0x95: case 0x96: case 0x97: case 0x98: case 0x99: case 0x9a: case 0x9b: case 0x9c: case 0x9d: case 0x9e: case 0x9f: /* SetCC. */ pt++; skipea(addr, &pt, process, false); break; // These are SSE2 instructions case 0x10: case 0x11: case 0x58: case 0x5c: case 0x59: case 0x5e: case 0x2e: case 0x2a: case 0x54: case 0x57: case 0x5a: case 0x6e: case 0x7e: case 0x2c: case 0x2d: pt++; skipea(addr, &pt, process, false); break; case 0x73: // PSRLDQ - EA,imm pt++; skipea(addr, &pt, process, false); pt++; break; default: Crash("Unknown opcode %d at %p\n", *pt, pt); } break; } case 0xd8: case 0xd9: case 0xda: case 0xdb: case 0xdc: case 0xdd: case 0xde: case 0xdf: // Floating point escape instructions { pt++; if ((*pt & 0xe0) == 0xe0) pt++; else skipea(addr, &pt, process, false); break; } default: Crash("Unknown opcode %d at %p\n", *pt, pt); } } } // Increment the value contained in the first word of the mutex. Handle X86TaskData::AtomicIncrement(Handle mutexp) { PolyObject *p = DEREFHANDLE(mutexp); POLYUNSIGNED result = X86AsmAtomicIncrement(p); return this->saveVec.push(PolyWord::FromUnsigned(result)); } // Release a mutex. Because the atomic increment and decrement // use the hardware LOCK prefix we can simply set this to one. void X86TaskData::AtomicReset(Handle mutexp) { DEREFHANDLE(mutexp)->Set(0, TAGGED(1)); } static X86Dependent x86Dependent; MachineDependent *machineDependent = &x86Dependent;