Add support or optional spin on park

jcl/src/java.base/share/classes/java/lang/Thread.java

@@ -790,9 +790,9 @@ public final boolean isDaemon() {
  * @see			Thread#interrupted
  */
 public boolean isInterrupted() {
-	synchronized(lock) {
+	//synchronized(lock) {
 		return isInterruptedImpl();
-	}
+	//}
 }
 
 private native boolean isInterruptedImpl();
@@ -1496,15 +1496,15 @@ public static enum State {
  * @see State
  */
 public State getState() {
-	synchronized(lock) {
+	//synchronized(lock) {
 		if (threadRef == NO_REF) {
 			if (isDead()) {
 				return State.TERMINATED;
 			}
 			return State.NEW;
 		}
 		return State.values()[getStateImpl(threadRef)];
-	}
+	//}
 }
 
 private native int getStateImpl(long threadRef);

runtime/oti/j9nonbuilder.h

@@ -5645,6 +5645,11 @@ typedef struct J9VMThread {
 #if defined(J9VM_OPT_JFR)
 	J9ThreadJFRState threadJfrState;
 #endif /* defined(J9VM_OPT_JFR) */
+	U_64 parkWaitTime;
+	U_64 parkCount;
+	U_64 parkWaitSlidingWindowIndex;
+	U_64 *parkWaitSlidingWindow;
+	volatile UDATA prePark;
 } J9VMThread;
 
 #define J9VMTHREAD_ALIGNMENT  0x100
@@ -5927,6 +5932,15 @@ typedef struct J9JavaVM {
 	UDATA thrNestedSpinning;
 	UDATA thrTryEnterNestedSpinning;
 	UDATA thrDeflationPolicy;
+	UDATA thrParkSpinCount1;
+	UDATA thrParkSpinCount2;
+	U_64 parkSpinWaitThreshold;
+	U_64 parkWaitSlidingWindowSize;
+	double parkSpinRatioOfAvgWait;
+	UDATA parkYield;
+	UDATA parkLock;
+	U_64 parkSleepMultiplier;
+	U_64 yieldUsleepMultiplier;
 	UDATA gcOptions;
 	UDATA  ( *unhookVMEvent)(struct J9JavaVM *javaVM, UDATA eventNumber, void * currentHandler, void * oldHandler) ;
 	UDATA classLoadingMaxStack;

runtime/vm/FastJNI_java_lang_Thread.cpp

@@ -65,12 +65,12 @@ Fast_java_lang_Thread_interruptedImpl(J9VMThread *currentThread)
 	jboolean rc = JNI_FALSE;
 #if defined(WIN32)
 	/* Synchronize on the thread lock around interrupted() on windows */
-	j9object_t threadLock = J9VMJAVALANGTHREAD_LOCK(currentThread, currentThread->threadObject);
-	threadLock = (j9object_t)objectMonitorEnter(currentThread, threadLock);
-	if (J9_OBJECT_MONITOR_ENTER_FAILED(threadLock)) {
-		setNativeOutOfMemoryError(currentThread, J9NLS_VM_FAILED_TO_ALLOCATE_MONITOR);
-		goto done;
-	}
+	// j9object_t threadLock = J9VMJAVALANGTHREAD_LOCK(currentThread, currentThread->threadObject);
+	// threadLock = (j9object_t)objectMonitorEnter(currentThread, threadLock);
+	// if (J9_OBJECT_MONITOR_ENTER_FAILED(threadLock)) {
+	// 	setNativeOutOfMemoryError(currentThread, J9NLS_VM_FAILED_TO_ALLOCATE_MONITOR);
+	// 	goto done;
+	// }
 	/* This field is only initialized on windows */
 	J9JavaVM * const vm = currentThread->javaVM;
 	if (NULL != vm->sidecarClearInterruptFunction) {
@@ -81,8 +81,8 @@ Fast_java_lang_Thread_interruptedImpl(J9VMThread *currentThread)
 		rc = JNI_TRUE;
 	}
 #if defined(WIN32)
-	objectMonitorExit(currentThread, threadLock);
-done:
+	// objectMonitorExit(currentThread, threadLock);
+// done:
 #endif /* WIN32 */
 	return rc;
 }

runtime/vm/j9vm.tdf

@@ -1012,3 +1012,7 @@ TraceEntry=Trc_VM_snapshot_subAllocateSnapshotMemory_Entry NoEnv Overhead=1 Leve
 TraceExit=Trc_VM_snapshot_subAllocateSnapshotMemory_Exit NoEnv Overhead=1 Level=1 Template="subAllocateMemory() Memory allocated = %p."
 TraceEvent=Trc_VM_snapshot_loadWarmClassFromSnapshot_ClassLoadHookFailed Overhead=1 Level=1 Template="loadWarmClassFromSnapshot() Warm class load hook failed class=%p, %s"
 TraceEvent=Trc_VM_snapshot_loadWarmClassFromSnapshot_ClassInfo Overhead=1 Level=1 Template="loadWarmClassFromSnapshot() LoadClass clazz=%p, %s"
+
+TraceEvent=Trc_VM_ThreadHelp_timeCompensationHelper_parkWait Env Overhead=1 Level=5 Template="Park early break average=%zu spinTime=%zu osYield=%zu cpuYield=%zu data=%zu parkWait=%zu vmthread=%p"
+TraceEvent=Trc_VM_ThreadHelp_timeCompensationHelper_parkWaitSpinElapsed Env Overhead=1 Level=5 Template="Park spin elapsed average=%zu spinTime=%zu parkWait=%zu osYield=%zu cpuYield=%zu vmthread=%p"
+TraceEvent=Trc_VM_ThreadHelp_timeCompensationHelper_parkWaitNoSpin Env Overhead=1 Level=5 Template="After park no spin average=%zu average=%zu parkWait=%zu vmthread=%p"

runtime/vm/threadhelp.cpp

@@ -427,9 +427,9 @@ IDATA
 timeCompensationHelper(J9VMThread *vmThread, U_8 threadHelperType, omrthread_monitor_t monitor, I_64 millis, I_32 nanos)
 {
 	IDATA rc = 0;
+	J9JavaVM *vm = vmThread->javaVM;
 
 #if defined(J9VM_OPT_CRIU_SUPPORT)
-	J9JavaVM *vm = vmThread->javaVM;
 	/* Time compensation only supports CRIURestoreNonPortableMode which is default mode. */
 	bool waitTimed = (millis > 0) || (nanos > 0);
 	bool compensationMightBeRequired = waitTimed && !J9_IS_CRIU_RESTORED(vm);
@@ -452,8 +452,78 @@ timeCompensationHelper(J9VMThread *vmThread, U_8 threadHelperType, omrthread_mon
 		rc = omrthread_monitor_wait_timed(monitor, millis, nanos);
 		break;
 	case HELPER_TYPE_THREAD_PARK:
-		rc = omrthread_park(millis, nanos);
-		break;
+		{
+			PORT_ACCESS_FROM_VMC(vmThread);
+			U_64 parkStart = j9time_nano_time();
+			U_64 slidingWindowSize = vm->parkWaitSlidingWindowSize;
+			U_64 avgWait = 0;
+			U_64 cpuYields = 0;
+			U_64 osYields = 0;
+			bool noSpin = true;
+			bool shouldYield = vm->parkYield > 0;
+			UDATA data = 0;
+			if (0 != slidingWindowSize) {
+				if (vmThread->parkWaitSlidingWindowIndex > slidingWindowSize) {
+					U_64 totalWait = 0;
+					vmThread->prePark = 1;
+
+					for (U_64 i = 0; i < slidingWindowSize; i++) {
+						totalWait += vmThread->parkWaitSlidingWindow[i];
+
+					}
+					avgWait = totalWait/slidingWindowSize;
+					bool spinElapsed = true;
+					//Trc_VM_ThreadHelp_timeCompensationHelper_parkWait(vmThread, avgWait, vmThread->parkCount, vmThread);
+
+					if (avgWait < vm->parkSpinWaitThreshold) {
+						noSpin = false;
+						for (IDATA spinCount2 = vm->thrParkSpinCount2; spinCount2 > 0; --spinCount2) {
+							U_64 spinTime = j9time_nano_time() - parkStart;
+
+							if ((vmThread->prePark == 0) || (spinTime > vm->parkSpinWaitThreshold)) {
+								data++;
+								Trc_VM_ThreadHelp_timeCompensationHelper_parkWait(vmThread, avgWait, spinTime, osYields, cpuYields, data, vmThread->prePark, vmThread);
+								spinElapsed = false;
+								break;
+							}
+
+							for (IDATA spinCount1 = vm->thrParkSpinCount1; spinCount1 > 0; --spinCount1) {
+								VM_AtomicSupport::yieldCPU();
+								cpuYields++;
+								if (vmThread->prePark == 0) {
+									data++;
+									spinElapsed = false;
+									break;
+								}
+							}
+							if (shouldYield) {
+								if (vmThread->prePark != 0) {
+									usleep(((osYields * vm->parkSleepMultiplier) + 1) * vm->yieldUsleepMultiplier);
+									//omrthread_yield_new(vm->thrParkSpinCount2 - spinCount2);
+									osYields++;
+								}
+							}
+						}
+
+						if (spinElapsed) {
+							Trc_VM_ThreadHelp_timeCompensationHelper_parkWaitSpinElapsed(vmThread, avgWait, j9time_nano_time() - parkStart, vmThread->prePark, osYields, cpuYields, vmThread);
+						}
+					}
+				}
+				rc = omrthread_park(millis, nanos);
+				U_64 parkEnd = j9time_nano_time();
+				vmThread->parkWaitSlidingWindow[vmThread->parkWaitSlidingWindowIndex % slidingWindowSize] = parkEnd - parkStart;
+				vmThread->parkWaitSlidingWindowIndex += 1;
+				vmThread->prePark = 0;
+				if (noSpin) {
+					Trc_VM_ThreadHelp_timeCompensationHelper_parkWaitNoSpin(vmThread, avgWait, parkEnd - parkStart, vmThread->prePark, vmThread);
+				}
+
+			} else {
+				rc = omrthread_park(millis, nanos);
+			}
+			break;
+		}
 	case HELPER_TYPE_THREAD_SLEEP:
 		/* Returns 0 when the timeout specified passed.
 		 * A timeout of 0 (0ms, 0ns) indicates an immediate return.

runtime/vm/threadpark.cpp

@@ -28,6 +28,7 @@
 #include "omrthread.h"
 #include "ut_j9vm.h"
 
+#include "AtomicSupport.hpp"
 #include "VMHelpers.hpp"
 
 #include <string.h>
@@ -127,6 +128,7 @@ void
 threadUnparkImpl(J9VMThread *vmThread, j9object_t threadObject)
 {
 	j9object_t threadLock = J9VMJAVALANGTHREAD_LOCK(vmThread, threadObject);
+	J9JavaVM *vm = vmThread->javaVM;
 
 	if (NULL == threadLock) {
 		/* thread not fully set up yet so we cannot really need to unpark, just return */
@@ -136,8 +138,10 @@ threadUnparkImpl(J9VMThread *vmThread, j9object_t threadObject)
 		 * the enter and the gc could move the targetThreadObject we have to push/pop on a special frame so that
 		 * we will have the updated value if the object is moved */
 		PUSH_OBJECT_IN_SPECIAL_FRAME(vmThread, threadObject);
-		threadLock = (j9object_t)objectMonitorEnter(vmThread, threadLock);
-		if (J9_OBJECT_MONITOR_ENTER_FAILED(threadLock)) {
+		if (vm->parkLock == 1) {
+			threadLock = (j9object_t)objectMonitorEnter(vmThread, threadLock);
+		}
+		if ((vm->parkLock == 1) && J9_OBJECT_MONITOR_ENTER_FAILED(threadLock)) {
 #if defined(J9VM_OPT_CRIU_SUPPORT)
 			if (J9_OBJECT_MONITOR_CRIU_SINGLE_THREAD_MODE_THROW == (UDATA)threadLock) {
 				setCRIUSingleThreadModeJVMCRIUException(vmThread, 0, 0);
@@ -157,8 +161,12 @@ threadUnparkImpl(J9VMThread *vmThread, j9object_t threadObject)
 			if (NULL != otherVmThread) {
 				/* in this case the thread is already dead so we don't need to unpark */
 				omrthread_unpark(otherVmThread->osThread);
+				VM_AtomicSupport::writeBarrier();
+				otherVmThread->prePark = 0;
+			}
+			if (vm->parkLock == 1) {
+				objectMonitorExit(vmThread, threadLock);
 			}
-			objectMonitorExit(vmThread, threadLock);
 			/*Trc_JCL_unpark_Exit(vmThread);*/
 		}
 	}

runtime/vm/vmthinit.c

@@ -122,6 +122,8 @@ void freeVMThread(J9JavaVM *vm, J9VMThread *vmThread)
 		/* Free J9RIParameters. */
 		j9mem_free_memory(vmThread->riParameters);
 	}
+	j9mem_free_memory(vmThread->parkWaitSlidingWindow);
+
 #endif /* defined(J9VM_PORT_RUNTIME_INSTRUMENTATION) */
 	if (J9JAVAVM_COMPRESS_OBJECT_REFERENCES(vm)) {
 		j9mem_free_memory32(vmThread->startOfMemoryBlock);

runtime/vm/vmthread.cpp

@@ -301,6 +301,10 @@ allocateVMThread(J9JavaVM *vm, omrthread_t osThread, UDATA privateFlags, void *m
 #if defined(J9VM_OPT_JFR)
 	newThread->threadJfrState.prevTimestamp = -1;
 #endif /* defined(J9VM_OPT_JFR) */
+	newThread->parkWaitSlidingWindow = (U_64*) j9mem_allocate_memory(vm->parkWaitSlidingWindowSize * sizeof(U_64), OMRMEM_CATEGORY_VM);
+	if (NULL == newThread->parkWaitSlidingWindow) {
+		goto fail;
+	}
 
 	/* If an exclusive access request is in progress, mark this thread */
 
@@ -571,6 +575,15 @@ threadParseArguments(J9JavaVM *vm, char *optArg)
 	vm->thrNestedSpinning = 1;
 	vm->thrTryEnterNestedSpinning = 1;
 	vm->thrDeflationPolicy = J9VM_DEFLATION_POLICY_ASAP;
+	vm->thrParkSpinCount1 = 0;
+	vm->thrParkSpinCount2 = 0;
+	vm->parkSpinWaitThreshold = 0;
+	vm->parkWaitSlidingWindowSize = 0;
+	vm->parkSpinRatioOfAvgWait = 0.95;
+	vm->parkLock = 0;
+	vm->parkYield = 0;
+	vm->parkSleepMultiplier = 0;
+	vm->yieldUsleepMultiplier = 0;
 
 	if (cpus > 1) {
 #if (defined(LINUXPPC)) && !defined(J9VM_ENV_LITTLE_ENDIAN)
@@ -828,6 +841,55 @@ threadParseArguments(J9JavaVM *vm, char *optArg)
 			continue;
 		}
 
+		if (try_scan(&scan_start, "parkSpinCount1=")) {
+			if (scan_udata(&scan_start, &vm->thrParkSpinCount1)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkSpinCount2=")) {
+			if (scan_udata(&scan_start, &vm->thrParkSpinCount2)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkSpinWaitThreshold=")) {
+			if (scan_udata(&scan_start, &vm->parkSpinWaitThreshold)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkWaitSlidingWindowSize=")) {
+			if (scan_udata(&scan_start, &vm->parkWaitSlidingWindowSize)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkSpinRatioOfAvgWait=")) {
+			if (scan_double(&scan_start, &vm->parkSpinRatioOfAvgWait)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkYield=")) {
+			if (scan_udata(&scan_start, &vm->parkYield)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkLock=")) {
+			if (scan_udata(&scan_start, &vm->parkLock)) {
+				goto _error;
+			}
+			continue;
+		}
+		if (try_scan(&scan_start, "parkSleepMultiplier=")) {
+			if (scan_udata(&scan_start, &vm->parkSleepMultiplier)) {
+				goto _error;
+			}
+			continue;
+		}
+
 #if !defined(WIN32) && defined(OMR_NOTIFY_POLICY_CONTROL)
 		if (try_scan(&scan_start, "notifyPolicy=")) {
 			char *oldScanStart = scan_start;
@@ -1209,6 +1271,8 @@ threadParseArguments(J9JavaVM *vm, char *optArg)
 				goto _error;
 			}
 			**(UDATA**)omrthread_global((char*)"yieldUsleepMultiplier") = yieldUsleepMultiplier;
+
+			vm->yieldUsleepMultiplier = yieldUsleepMultiplier;
 			continue;
 		}