fixed some additional pthread_{get/set}affinity_np flakiness

src/libcore/thread.cpp

@@ -324,21 +324,38 @@ void Thread::setCoreAffinity(int coreID) {
 	if (getenv("VALGRIND_OPTS") != NULL)
 		return;
 
-	int nCores = sysconf(_SC_NPROCESSORS_CONF);
+	const pthread_t threadID = d->thread.native_handle();
-	size_t size = CPU_ALLOC_SIZE(nCores);
+
-	cpu_set_t *cpuset = CPU_ALLOC(nCores);
+	int nCores = sysconf(_SC_NPROCESSORS_CONF),
-	CPU_ZERO_S(size, cpuset);
+		nLogicalCores = nCores;
-	if (cpuset == NULL) {
+
-		Log(EWarn, "Thread::setCoreAffinity(): could not allocate cpu_set_t");
+	size_t size = 0;
-		return;
+	cpu_set_t *cpuset = NULL;
+	int retval = 0;
+
+	/* The kernel may expected a larger cpu_set_t than would
+	   be warranted by the physical core count. Keep querying
+	   with increasingly larger buffers if the
+	   pthread_getaffinity_np operation fails */
+	for (int i = 0; i<6; ++i) { 
+		size = CPU_ALLOC_SIZE(nLogicalCores);
+		cpuset = CPU_ALLOC(nLogicalCores);
+		if (!cpuset) {
+			Log(EWarn, "Thread::setCoreAffinity(): could not allocate cpu_set_t");
+			return;
+		}
+		CPU_ZERO_S(size, cpuset);
+
+		int retval = pthread_getaffinity_np(threadID, size, cpuset);
+		if (retval == 0)
+			break;
+		CPU_FREE(cpuset);
+		nLogicalCores *= 2;
 	}
 
-	const pthread_t threadID = d->thread.native_handle();
-	int retval = pthread_getaffinity_np(threadID, size, cpuset);
 	if (retval) {
 		Log(EWarn, "Thread::setCoreAffinity(): pthread_getaffinity_np(): could "
 			"not read thread affinity map: %s", strerror(retval));
-		CPU_FREE(cpuset);
 		return;
 	}
 

src/libcore/util.cpp

@@ -166,27 +166,51 @@ int getCoreCount() {
 	/* Determine the number of present cores */
 	int nCores = sysconf(_SC_NPROCESSORS_CONF);
 
-	/* Some of the cores may not be available to the user
+	/* Don't try to query CPU affinity if running inside Valgrind */
-	   (e.g. on certain cluster nodes) -- determine the number
+	if (getenv("VALGRIND_OPTS") == NULL) {
-	   of actual available cores here. */
+		/* Some of the cores may not be available to the user
-	size_t size = CPU_ALLOC_SIZE(nCores);
+		   (e.g. on certain cluster nodes) -- determine the number
-	cpu_set_t *cpuset = CPU_ALLOC(nCores);
+		   of actual available cores here. */
-	CPU_ZERO_S(size, cpuset);
+		int nLogicalCores = nCores;
-	int retval = pthread_getaffinity_np(pthread_self(), size, cpuset);
+		size_t size = 0;
-	if (retval) {
+		cpu_set_t *cpuset = NULL;
-		SLog(EWarn, "getCoreCount(): pthread_getaffinity_np(): could "
+		int retval = 0;
-			"not read thread affinity map: %s", strerror(retval));
+
-		__cached_core_count = nCores;
+		/* The kernel may expected a larger cpu_set_t than would
+		   be warranted by the physical core count. Keep querying
+		   with increasingly larger buffers if the
+		   pthread_getaffinity_np operation fails */
+		for (int i = 0; i<6; ++i) { 
+			size = CPU_ALLOC_SIZE(nLogicalCores);
+			cpuset = CPU_ALLOC(nLogicalCores);
+			if (!cpuset) {
+				SLog(EWarn, "getCoreCount(): could not allocate cpu_set_t");
+				goto done;
+			}
+			CPU_ZERO_S(size, cpuset);
+
+			int retval = pthread_getaffinity_np(pthread_self(), size, cpuset);
+			if (retval == 0)
+				break;
+			CPU_FREE(cpuset);
+			nLogicalCores *= 2;
+		}
+
+		if (retval) {
+			SLog(EWarn, "getCoreCount(): pthread_getaffinity_np(): could "
+				"not read thread affinity map: %s", strerror(retval));
+			goto done;
+		}
+
+		int availableCores = 0;
+		for (int i=0; i<nCores; ++i)
+			availableCores += CPU_ISSET_S(i, size, cpuset) ? 1 : 0;
 		CPU_FREE(cpuset);
-		return nCores;
 	}
 
-	int availableCores = 0;
+done:
-	for (int i=0; i<nCores; ++i)
+	__cached_core_count = nCores;
-		availableCores += CPU_ISSET_S(i, size, cpuset) ? 1 : 0;
+	return nCores;
-	CPU_FREE(cpuset);
-	__cached_core_count = availableCores;
-	return availableCores;
 #endif
 }
 

src/mtsgui/mainwindow.cpp

@@ -1358,7 +1358,7 @@ void MainWindow::on_actionSettings_triggered() {
 			ref<Scheduler> sched = Scheduler::getInstance();
 			sched->pause();
 
-			if (localWorkers.size() != d.getLocalWorkerCount()) {
+			if ((int) localWorkers.size() != d.getLocalWorkerCount()) {
 				/* Completely remove old workers so that CPU affinities can be reassigned */
 				while (!localWorkers.empty()) {
 					Worker *worker = localWorkers.back();