Core/TaskScheduler: Fix work ending condition

Use remaining task count instead of idle worker count, this avoids a race condition where a worker signals idle after being tasked with a new job

include/Nazara/Core/TaskScheduler.hpp

@@ -40,11 +40,10 @@ namespace Nz
 			friend Worker;
 
 			Worker& GetWorker(unsigned int workerIndex);
-			void NotifyWorkerActive();
-			void NotifyWorkerIdle();
+			void NotifyTaskCompletion();
 
 			std::atomic_bool m_idle;
-			std::atomic_uint m_idleWorkerCount;
+			std::atomic_uint m_remainingTasks;
 			std::size_t m_nextWorkerIndex;
 			std::vector<Worker> m_workers;
 			MemoryPool<Task> m_tasks;

src/Nazara/Core/TaskScheduler.cpp

@@ -116,6 +116,7 @@ namespace Nz
 			{
 				// Wait until task scheduler started
 				m_notifier.wait(false);
+				m_notifier.clear();
 
 				StackArray<unsigned int> randomWorkerIndices = NazaraStackArrayNoInit(unsigned int, m_owner.GetWorkerCount() - 1);
 				{
@@ -158,20 +159,13 @@ namespace Nz
 
 						(*task)();
 
-#ifdef NAZARA_WITH_TSAN
-						// Workaround for TSan false-positive
-						__tsan_release(task);
-#endif
+						m_owner.NotifyTaskCompletion();
 					}
 					else
 					{
 						// Wait for tasks if we don't have any right now
-						m_owner.NotifyWorkerIdle();
-
 						m_notifier.wait(false);
 						m_notifier.clear();
-
-						m_owner.NotifyWorkerActive();
 					}
 				}
 				while (m_running.load(std::memory_order_relaxed));
@@ -210,7 +204,7 @@ namespace Nz
 
 	TaskScheduler::TaskScheduler(unsigned int workerCount) :
 	m_idle(false),
-	m_idleWorkerCount(0),
+	m_remainingTasks(0),
 	m_nextWorkerIndex(0),
 	m_tasks(256 * sizeof(Task)),
 	m_workerCount(workerCount)
@@ -224,9 +218,6 @@ namespace Nz
 
 		for (unsigned int i = 0; i < m_workerCount; ++i)
 			m_workers[i].WakeUp();
-
-		// Wait until all worked started
-		m_idle.wait(false);
 	}
 
 	TaskScheduler::~TaskScheduler()
@@ -246,6 +237,8 @@ namespace Nz
 		__tsan_release(taskPtr);
 #endif
 
+		m_remainingTasks++;
+
 		Worker& worker = m_workers[m_nextWorkerIndex++];
 		worker.AddTask(taskPtr);
 
@@ -256,13 +249,6 @@ namespace Nz
 	void TaskScheduler::WaitForTasks()
 	{
 		m_idle.wait(false);
-
-#ifdef NAZARA_WITH_TSAN
-		// Workaround for TSan false-positive
-		for (Task& task : m_tasks)
-			__tsan_acquire(&task);
-#endif
-
 		m_tasks.Clear();
 	}
 
@@ -271,14 +257,9 @@ namespace Nz
 		return m_workers[workerIndex];
 	}
 
-	void TaskScheduler::NotifyWorkerActive()
+	void TaskScheduler::NotifyTaskCompletion()
 	{
-		m_idleWorkerCount--;
-	}
-
-	void TaskScheduler::NotifyWorkerIdle()
-	{
-		if (++m_idleWorkerCount == m_workers.size())
+		if (--m_remainingTasks == 0)
 		{
 			m_idle = true;
 			m_idle.notify_one();

tests/UnitTests/Engine/Core/TaskSchedulerTests.cpp

@@ -1,14 +1,17 @@
+#include <Nazara/Core/Clock.hpp>
 #include <Nazara/Core/TaskScheduler.hpp>
 #include <catch2/catch_test_macros.hpp>
 #include <atomic>
+#include <chrono>
+#include <thread>
 
 SCENARIO("TaskScheduler", "[CORE][TaskScheduler]")
 {
-	for (std::size_t workerCount : { 0, 1, 2, 4 })
+	for (std::size_t workerCount : { 0, 1, 2, 4, 8 })
 	{
 		GIVEN("A task scheduler with " << workerCount << " workers")
 		{
-			Nz::TaskScheduler scheduler(4);
+			Nz::TaskScheduler scheduler(workerCount);
 
 			WHEN("We add a single task and wait for it")
 			{
@@ -19,6 +22,26 @@ SCENARIO("TaskScheduler", "[CORE][TaskScheduler]")
 				CHECK(executed);
 			}
 
+			WHEN("We add time-consuming tasks, they are split between workers")
+			{
+				std::atomic_uint count = 0;
+
+				Nz::HighPrecisionClock clock;
+				for (unsigned int i = 0; i < scheduler.GetWorkerCount(); ++i)
+				{
+					scheduler.AddTask([&]
+					{
+						std::this_thread::sleep_for(std::chrono::milliseconds(100));
+						count++;
+					});
+				}
+				scheduler.WaitForTasks();
+				Nz::Time elapsedTime = clock.GetElapsedTime();
+
+				CHECK(count == scheduler.GetWorkerCount());
+				CHECK(elapsedTime < Nz::Time::Milliseconds(120));
+			}
+
 			WHEN("We add a lot of tasks and wait for all of them")
 			{
 				constexpr std::size_t taskCount = 512;