CpuBoundWork#CpuBoundWork(): don't spin on atomic int to acquire slot

This is inefficient and involves unfair scheduling. The latter implies
possible bad surprises regarding waiting durations on busy nodes. Instead,
use AsioConditionVariable#Wait() if there are no free slots. It's notified
by others' CpuBoundWork#~CpuBoundWork() once finished.

Author: Al2Klimov

lib/base/io-engine.cpp

@@ -20,24 +20,39 @@ CpuBoundWork::CpuBoundWork(boost::asio::yield_context yc)
 	: m_Done(false)
 {
 	auto& ioEngine (IoEngine::Get());
+	auto& sem (ioEngine.m_CpuBoundSemaphore);
+	std::unique_lock<std::mutex> lock (sem.Mutex);
 
-	for (;;) {
-		auto availableSlots (ioEngine.m_CpuBoundSemaphore.fetch_sub(1));
+	if (sem.FreeSlots) {
+		--sem.FreeSlots;
+		return;
+	}
 
-		if (availableSlots < 1) {
-			ioEngine.m_CpuBoundSemaphore.fetch_add(1);
-			IoEngine::YieldCurrentCoroutine(yc);
-			continue;
-		}
+	auto cv (Shared<AsioConditionVariable>::Make(ioEngine.GetIoContext()));
+
+	sem.Waiting.emplace(cv);
+	lock.unlock();
 
-		break;
+	try {
+		cv->Wait(yc);
+	} catch (...) {
+		Done();
+		throw;
 	}
 }
 
 void CpuBoundWork::Done()
 {
 	if (!m_Done) {
-		IoEngine::Get().m_CpuBoundSemaphore.fetch_add(1);
+		auto& sem (IoEngine::Get().m_CpuBoundSemaphore);
+		std::unique_lock<std::mutex> lock (sem.Mutex);
+
+		if (sem.Waiting.empty()) {
+			++sem.FreeSlots;
+		} else {
+			sem.Waiting.front()->Set();
+			sem.Waiting.pop();
+		}
 
 		m_Done = true;
 	}
@@ -58,7 +73,11 @@ boost::asio::io_context& IoEngine::GetIoContext()
 IoEngine::IoEngine() : m_IoContext(), m_KeepAlive(boost::asio::make_work_guard(m_IoContext)), m_Threads(decltype(m_Threads)::size_type(Configuration::Concurrency * 2u)), m_AlreadyExpiredTimer(m_IoContext)
 {
 	m_AlreadyExpiredTimer.expires_at(boost::posix_time::neg_infin);
-	m_CpuBoundSemaphore.store(Configuration::Concurrency * 3u / 2u);
+
+	{
+		std::unique_lock<std::mutex> lock (m_CpuBoundSemaphore.Mutex);
+		m_CpuBoundSemaphore.FreeSlots = Configuration::Concurrency * 3u / 2u;
+	}
 
 	for (auto& thread : m_Threads) {
 		thread = std::thread(&IoEngine::RunEventLoop, this);

lib/base/io-engine.hpp

@@ -6,10 +6,14 @@
 #include "base/exception.hpp"
 #include "base/lazy-init.hpp"
 #include "base/logger.hpp"
+#include "base/shared.hpp"
 #include "base/shared-object.hpp"
 #include <atomic>
+#include <cstdint>
 #include <exception>
 #include <memory>
+#include <mutex>
+#include <queue>
 #include <thread>
 #include <utility>
 #include <vector>
@@ -47,6 +51,24 @@ class CpuBoundWork
 	bool m_Done;
 };
 
+/**
+ * Condition variable which doesn't block I/O threads
+ *
+ * @ingroup base
+ */
+class AsioConditionVariable
+{
+public:
+	AsioConditionVariable(boost::asio::io_context& io, bool init = false);
+
+	void Set();
+	void Clear();
+	void Wait(boost::asio::yield_context yc);
+
+private:
+	boost::asio::deadline_timer m_Timer;
+};
+
 /**
  * Async I/O engine
  *
@@ -120,29 +142,16 @@ class IoEngine
 	boost::asio::executor_work_guard<boost::asio::io_context::executor_type> m_KeepAlive;
 	std::vector<std::thread> m_Threads;
 	boost::asio::deadline_timer m_AlreadyExpiredTimer;
-	std::atomic_int_fast32_t m_CpuBoundSemaphore;
-};
 
-class TerminateIoThread : public std::exception
-{
+	struct {
+		std::mutex Mutex;
+		uint_fast32_t FreeSlots;
+		std::queue<Shared<AsioConditionVariable>::Ptr> Waiting;
+	} m_CpuBoundSemaphore;
 };
 
-/**
- * Condition variable which doesn't block I/O threads
- *
- * @ingroup base
- */
-class AsioConditionVariable
+class TerminateIoThread : public std::exception
 {
-public:
-	AsioConditionVariable(boost::asio::io_context& io, bool init = false);
-
-	void Set();
-	void Clear();
-	void Wait(boost::asio::yield_context yc);
-
-private:
-	boost::asio::deadline_timer m_Timer;
 };
 
 /**