diff --git a/DDrawCompat/Common/Time.cpp b/DDrawCompat/Common/Time.cpp
index 2609803..4c8f628 100644
--- a/DDrawCompat/Common/Time.cpp
+++ b/DDrawCompat/Common/Time.cpp
@@ -5,11 +5,26 @@
 namespace Time
 {
 	long long g_qpcFrequency = 0;
+	HANDLE g_waitableTimer = nullptr;
 
 	void init()
 	{
 		LARGE_INTEGER qpc;
 		QueryPerformanceFrequency(&qpc);
 		g_qpcFrequency = qpc.QuadPart;
+
+		g_waitableTimer = CreateWaitableTimer(nullptr, FALSE, nullptr);
+	}
+
+	void waitForNextTick()
+	{
+		LARGE_INTEGER due = {};
+		due.QuadPart = -1;
+		if (!g_waitableTimer ||
+			!SetWaitableTimer(g_waitableTimer, &due, 0, nullptr, nullptr, FALSE) ||
+			WAIT_OBJECT_0 != WaitForSingleObject(g_waitableTimer, INFINITE))
+		{
+			Sleep(1);
+		}
 	}
 }
diff --git a/DDrawCompat/Common/Time.h b/DDrawCompat/Common/Time.h
index 93e11c9..8bbae39 100644
--- a/DDrawCompat/Common/Time.h
+++ b/DDrawCompat/Common/Time.h
@@ -25,10 +25,5 @@ namespace Time
 		return qpc.QuadPart;
 	}
 
-	inline ULONG64 queryThreadCycleTime()
-	{
-		ULONG64 cycleTime = 0;
-		QueryThreadCycleTime(GetCurrentThread(), &cycleTime);
-		return cycleTime;
-	}
+	void waitForNextTick();
 }
diff --git a/DDrawCompat/Config/Config.h b/DDrawCompat/Config/Config.h
index 920a9d8..6a5bdb8 100644
--- a/DDrawCompat/Config/Config.h
+++ b/DDrawCompat/Config/Config.h
@@ -21,9 +21,6 @@
 
 namespace Config
 {
-	const unsigned delayedFlipModeTimeout = 200;
-	const unsigned maxPaletteUpdatesPerMs = 5;
-
 	extern Settings::AlternatePixelCenter alternatePixelCenter;
 	extern Settings::Antialiasing antialiasing;
 	extern Settings::ConfigHotKey configHotKey;
diff --git a/DDrawCompat/D3dDdi/Device.cpp b/DDrawCompat/D3dDdi/Device.cpp
index 57b2272..c974941 100644
--- a/DDrawCompat/D3dDdi/Device.cpp
+++ b/DDrawCompat/D3dDdi/Device.cpp
@@ -27,6 +27,7 @@ namespace D3dDdi
 		: m_origVtable(CompatVtable<D3DDDI_DEVICEFUNCS>::s_origVtable)
 		, m_adapter(adapter)
 		, m_device(device)
+		, m_eventQuery(nullptr)
 		, m_renderTarget(nullptr)
 		, m_renderTargetSubResourceIndex(0)
 		, m_sharedPrimary(nullptr)
@@ -34,6 +35,10 @@ namespace D3dDdi
 		, m_state(*this)
 		, m_shaderBlitter(*this)
 	{
+		D3DDDIARG_CREATEQUERY createQuery = {};
+		createQuery.QueryType = D3DDDIQUERYTYPE_EVENT;
+		m_origVtable.pfnCreateQuery(m_device, &createQuery);
+		m_eventQuery = createQuery.hQuery;
 	}
 
 	void Device::add(Adapter& adapter, HANDLE device)
@@ -388,6 +393,33 @@ namespace D3dDdi
 		m_state.updateConfig();
 	}
 
+	void Device::waitForIdle()
+	{
+		D3dDdi::ScopedCriticalSection lock;
+		flushPrimitives();
+		D3DDDIARG_ISSUEQUERY issueQuery = {};
+		issueQuery.hQuery = m_eventQuery;
+		issueQuery.Flags.End = 1;
+		m_origVtable.pfnIssueQuery(m_device, &issueQuery);
+
+		if (m_origVtable.pfnFlush1)
+		{
+			m_origVtable.pfnFlush1(m_device, 0);
+		}
+		else
+		{
+			m_origVtable.pfnFlush(m_device);
+		}
+
+		BOOL result = FALSE;
+		D3DDDIARG_GETQUERYDATA getQueryData = {};
+		getQueryData.hQuery = m_eventQuery;
+		getQueryData.pData = &result;
+		while (S_FALSE == m_origVtable.pfnGetQueryData(m_device, &getQueryData))
+		{
+		}
+	}
+
 	std::map<HANDLE, Device> Device::s_devices;
 	bool Device::s_isFlushEnabled = true;
 }
diff --git a/DDrawCompat/D3dDdi/Device.h b/DDrawCompat/D3dDdi/Device.h
index f843f1f..5760e26 100644
--- a/DDrawCompat/D3dDdi/Device.h
+++ b/DDrawCompat/D3dDdi/Device.h
@@ -58,6 +58,7 @@ namespace D3dDdi
 		void prepareForGpuWrite();
 		void setRenderTarget(const D3DDDIARG_SETRENDERTARGET& data);
 		void updateConfig();
+		void waitForIdle();
 
 		static void add(Adapter& adapter, HANDLE device);
 		static Device& get(HANDLE device) { return s_devices.find(device)->second; }
@@ -75,6 +76,7 @@ namespace D3dDdi
 		D3DDDI_DEVICEFUNCS m_origVtable;
 		Adapter& m_adapter;
 		HANDLE m_device;
+		HANDLE m_eventQuery;
 		std::map<HANDLE, std::unique_ptr<Resource>> m_resources;
 		Resource* m_renderTarget;
 		UINT m_renderTargetSubResourceIndex;
diff --git a/DDrawCompat/D3dDdi/KernelModeThunks.cpp b/DDrawCompat/D3dDdi/KernelModeThunks.cpp
index 81846df..424d916 100644
--- a/DDrawCompat/D3dDdi/KernelModeThunks.cpp
+++ b/DDrawCompat/D3dDdi/KernelModeThunks.cpp
@@ -2,7 +2,6 @@
 #include <string>
 
 #include <Windows.h>
-#include <VersionHelpers.h>
 
 #include <Common/Log.h>
 #include <Common/Hook.h>
@@ -107,6 +106,21 @@ namespace
 		return LOG_RESULT(result);
 	}
 
+	NTSTATUS APIENTRY createDevice(D3DKMT_CREATEDEVICE* pData)
+	{
+		LOG_FUNC("D3DKMTCreateDevice", pData);
+		NTSTATUS result = D3DKMTCreateDevice(pData);
+		if (SUCCEEDED(result))
+		{
+			D3DKMT_SETQUEUEDLIMIT limit = {};
+			limit.hDevice = pData->hDevice;
+			limit.Type = D3DKMT_SET_QUEUEDLIMIT_PRESENT;
+			limit.QueuedPresentLimit = 1;
+			D3DKMTSetQueuedLimit(&limit);
+		}
+		return LOG_RESULT(result);
+	}
+
 	HDC WINAPI ddrawCreateDcA(LPCSTR pwszDriver, LPCSTR pwszDevice, LPCSTR pszPort, const DEVMODEA* pdm)
 	{
 		LOG_FUNC("ddrawCreateDCA", pwszDriver, pwszDevice, pszPort, pdm);
@@ -152,7 +166,7 @@ namespace
 	{
 		D3DKMT_GETSCANLINE data = {};
 		getVidPnSource(data.hAdapter, data.VidPnSourceId);
-		if (!data.hAdapter || FAILED(D3DKMTGetScanLine(&data)) || data.InVerticalBlank)
+		if (!data.hAdapter || FAILED(D3DKMTGetScanLine(&data)))
 		{
 			return -1;
 		}
@@ -189,19 +203,6 @@ namespace
 		return LOG_RESULT(result);
 	}
 
-	void pollForVerticalBlank()
-	{
-		int scanLine = getScanLine();
-		int prevScanLine = scanLine;
-		auto qpcStart = Time::queryPerformanceCounter();
-		while (scanLine >= prevScanLine && Time::queryPerformanceCounter() - qpcStart < Time::g_qpcFrequency / 60)
-		{
-			Sleep(1);
-			prevScanLine = scanLine;
-			scanLine = getScanLine();
-		}
-	}
-
 	NTSTATUS APIENTRY present(D3DKMT_PRESENT* pData)
 	{
 		LOG_FUNC("D3DKMTPresent", pData);
@@ -252,6 +253,7 @@ namespace
 	{
 		LOG_FUNC("D3DKMTSetGammaRamp", pData);
 		UINT vsyncCounter = D3dDdi::KernelModeThunks::getVsyncCounter();
+		DDraw::RealPrimarySurface::setUpdateReady();
 		DDraw::RealPrimarySurface::flush();
 		HRESULT result = D3DKMTSetGammaRamp(pData);
 		if (SUCCEEDED(result))
@@ -324,30 +326,24 @@ namespace
 
 	void waitForVerticalBlank()
 	{
-		if (IsWindows8OrGreater())
+		auto qpcStart = Time::queryPerformanceCounter();
+		int scanLine = getScanLine();
+		int prevScanLine = 0;
+		while (scanLine >= prevScanLine)
 		{
-			D3DKMT_WAITFORVERTICALBLANKEVENT data = {};
-
-			{
-				Compat::ScopedSrwLockShared lock(g_adapterInfoSrwLock);
-				data.hAdapter = g_lastOpenAdapterInfo.adapter;
-				data.VidPnSourceId = g_lastOpenAdapterInfo.vidPnSourceId;
-			}
-
-			if (!data.hAdapter)
-			{
-				updateGdiAdapterInfo();
-				data.hAdapter = g_gdiAdapterInfo.adapter;
-				data.VidPnSourceId = g_gdiAdapterInfo.vidPnSourceId;
-			}
-
-			if (data.hAdapter && SUCCEEDED(D3DKMTWaitForVerticalBlankEvent(&data)))
-			{
-				return;
-			}
+			Time::waitForNextTick();
+			prevScanLine = scanLine;
+			scanLine = getScanLine();
 		}
 
-		pollForVerticalBlank();
+		if (scanLine < 0)
+		{
+			auto msElapsed = static_cast<DWORD>(Time::qpcToMs(Time::queryPerformanceCounter() - qpcStart));
+			if (msElapsed < 16)
+			{
+				Sleep(16 - msElapsed);
+			}
+		}
 	}
 }
 
@@ -412,6 +408,7 @@ namespace D3dDdi
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "CreateDCA", ddrawCreateDcA);
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTCloseAdapter", closeAdapter);
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTCreateDCFromMemory", createDcFromMemory);
+			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTCreateDevice", createDevice);
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTOpenAdapterFromHdc", openAdapterFromHdc);
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTPresent", present);
 			Compat::hookIatFunction(Dll::g_origDDrawModule, "D3DKMTQueryAdapterInfo", queryAdapterInfo);
diff --git a/DDrawCompat/D3dDdi/Resource.cpp b/DDrawCompat/D3dDdi/Resource.cpp
index 902477a..850bae4 100644
--- a/DDrawCompat/D3dDdi/Resource.cpp
+++ b/DDrawCompat/D3dDdi/Resource.cpp
@@ -832,6 +832,16 @@ namespace D3dDdi
 		auto rtIndex = rtSurface.resource ? 0 : data.DstSubResourceIndex;
 		auto rtRect = rtSurface.resource ? data.SrcRect : data.DstRect;
 
+		if (D3DDDIPOOL_SYSTEMMEM == srcResource->m_fixedData.Pool)
+		{
+			srcResource = repo.getTempTexture(srcWidth, srcHeight, getPixelFormat(srcResource->m_fixedData.Format)).resource;
+			if (!srcResource)
+			{
+				return E_OUTOFMEMORY;
+			}
+			copySubResourceRegion(*srcResource, 0, data.SrcRect, data.hSrcResource, data.SrcSubResourceIndex, data.SrcRect);
+		}
+
 		if (D3DDDIFMT_P8 == srcResource->m_origData.Format)
 		{
 			auto entries(Gdi::Palette::getHardwarePalette());
diff --git a/DDrawCompat/D3dDdi/Resource.h b/DDrawCompat/D3dDdi/Resource.h
index 2d59e1e..a12213a 100644
--- a/DDrawCompat/D3dDdi/Resource.h
+++ b/DDrawCompat/D3dDdi/Resource.h
@@ -27,6 +27,7 @@ namespace D3dDdi
 
 		operator HANDLE() const { return m_handle; }
 		const Resource* getCustomResource() { return m_msaaSurface.resource ? m_msaaSurface.resource : m_msaaResolvedSurface.resource; }
+		Device& getDevice() const { return m_device; }
 		const D3DDDIARG_CREATERESOURCE2& getFixedDesc() const { return m_fixedData; }
 		const D3DDDIARG_CREATERESOURCE2& getOrigDesc() const { return m_origData; }
 		bool isClampable() const { return m_isClampable; }
diff --git a/DDrawCompat/DDraw/DirectDraw.cpp b/DDrawCompat/DDraw/DirectDraw.cpp
index 4f03783..1d44506 100644
--- a/DDrawCompat/DDraw/DirectDraw.cpp
+++ b/DDrawCompat/DDraw/DirectDraw.cpp
@@ -73,6 +73,7 @@ namespace
 	template <typename TDirectDraw>
 	HRESULT STDMETHODCALLTYPE WaitForVerticalBlank(TDirectDraw* This, DWORD dwFlags, HANDLE hEvent)
 	{
+		DDraw::RealPrimarySurface::setUpdateReady();
 		DDraw::RealPrimarySurface::flush();
 		return getOrigVtable(This).WaitForVerticalBlank(This, dwFlags, hEvent);
 	}
diff --git a/DDrawCompat/DDraw/DirectDrawPalette.cpp b/DDrawCompat/DDraw/DirectDrawPalette.cpp
index c04109b..4bf69d7 100644
--- a/DDrawCompat/DDraw/DirectDrawPalette.cpp
+++ b/DDrawCompat/DDraw/DirectDrawPalette.cpp
@@ -49,7 +49,8 @@ namespace DDraw
 				updatesInLastMs.pop_front();
 			}
 
-			if (updatesInLastMs.size() >= Config::maxPaletteUpdatesPerMs)
+			const unsigned maxPaletteUpdatesPerMs = 5;
+			if (updatesInLastMs.size() >= maxPaletteUpdatesPerMs)
 			{
 				Sleep(1);
 				updatesInLastMs.clear();
diff --git a/DDrawCompat/DDraw/RealPrimarySurface.cpp b/DDrawCompat/DDraw/RealPrimarySurface.cpp
index d5dbe76..046a7e1 100644
--- a/DDrawCompat/DDraw/RealPrimarySurface.cpp
+++ b/DDrawCompat/DDraw/RealPrimarySurface.cpp
@@ -1,10 +1,10 @@
-#include <atomic>
 #include <memory>
 #include <vector>
 
 #include <Common/Comparison.h>
 #include <Common/CompatPtr.h>
 #include <Common/Hook.h>
+#include <Common/ScopedCriticalSection.h>
 #include <Common/Time.h>
 #include <Config/Config.h>
 #include <D3dDdi/Device.h>
@@ -29,6 +29,8 @@
 
 namespace
 {
+	const unsigned DELAYED_FLIP_MODE_TIMEOUT_MS = 200;
+
 	void onRelease();
 
 	CompatWeakPtr<IDirectDrawSurface7> g_frontBuffer;
@@ -40,11 +42,18 @@ namespace
 	bool g_isFullscreen = false;
 	DDraw::Surface* g_lastFlipSurface = nullptr;
 
+	Compat::CriticalSection g_presentCs;
+	bool g_isDelayedFlipPending = false;
 	bool g_isUpdatePending = false;
-	bool g_waitingForPrimaryUnlock = false;
-	std::atomic<long long> g_qpcLastUpdate = 0;
+	bool g_isUpdateReady = false;
+	DWORD g_lastUpdateThreadId = 0;
+	long long g_qpcLastUpdate = 0;
+	long long g_qpcUpdateStart = 0;
+
+	long long g_qpcDelayedFlipEnd = 0;
 	UINT g_flipEndVsyncCount = 0;
 	UINT g_presentEndVsyncCount = 0;
+
 	HWND g_devicePresentationWindow = nullptr;
 	HWND g_deviceWindow = nullptr;
 	HWND* g_deviceWindowPtr = nullptr;
@@ -112,24 +121,14 @@ namespace
 		return 1;
 	}
 
-	bool isFlipPending()
-	{
-		return static_cast<INT>(D3dDdi::KernelModeThunks::getVsyncCounter() - g_flipEndVsyncCount) < 0;
-	}
-
-	bool isPresentPending()
-	{
-		return static_cast<INT>(D3dDdi::KernelModeThunks::getVsyncCounter() - g_presentEndVsyncCount) < 0;
-	}
-
 	void onRelease()
 	{
 		LOG_FUNC("RealPrimarySurface::onRelease");
 
 		g_frontBuffer = nullptr;
+		g_lastFlipSurface = nullptr;
 		g_clipper.release();
 		g_isFullscreen = false;
-		g_waitingForPrimaryUnlock = false;
 		g_surfaceDesc = {};
 
 		DDraw::RealPrimarySurface::updateDevicePresentationWindowPos();
@@ -152,16 +151,20 @@ namespace
 		const bool isFlippable = 0 != (desc.ddsCaps.dwCaps & DDSCAPS_FLIP);
 		g_surfaceDesc = desc;
 		g_isFullscreen = isFlippable;
-		g_isUpdatePending = true;
-		g_qpcLastUpdate = Time::queryPerformanceCounter() - Time::msToQpc(Config::delayedFlipModeTimeout);
 
 		if (isFlippable)
 		{
 			g_frontBuffer->Flip(g_frontBuffer, getLastSurface(), DDFLIP_WAIT);
-			g_flipEndVsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter() + 1;
-			g_presentEndVsyncCount = g_flipEndVsyncCount;
-			D3dDdi::KernelModeThunks::waitForVsyncCounter(g_flipEndVsyncCount);
+			D3dDdi::KernelModeThunks::waitForVsyncCounter(D3dDdi::KernelModeThunks::getVsyncCounter() + 1);
 		}
+
+		Compat::ScopedCriticalSection lock(g_presentCs);
+		g_isUpdatePending = false;
+		g_isUpdateReady = false;
+		g_qpcLastUpdate = Time::queryPerformanceCounter() - Time::msToQpc(DELAYED_FLIP_MODE_TIMEOUT_MS);
+		g_qpcUpdateStart = g_qpcLastUpdate;
+		g_presentEndVsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter();
+		g_flipEndVsyncCount = g_presentEndVsyncCount;
 	}
 
 	void presentToPrimaryChain(CompatWeakPtr<IDirectDrawSurface7> src)
@@ -202,11 +205,15 @@ namespace
 		D3dDdi::KernelModeThunks::setDcPaletteOverride(nullptr);
 	}
 
-	void updateNow(CompatWeakPtr<IDirectDrawSurface7> src, UINT flipInterval)
+	void updateNow(CompatWeakPtr<IDirectDrawSurface7> src)
 	{
+		{
+			Compat::ScopedCriticalSection lock(g_presentCs);
+			g_isUpdatePending = false;
+			g_isUpdateReady = false;
+		}
+
 		presentToPrimaryChain(src);
-		g_isUpdatePending = false;
-		g_waitingForPrimaryUnlock = false;
 
 		if (g_isFullscreen && g_devicePresentationWindow)
 		{
@@ -214,60 +221,26 @@ namespace
 			g_frontBuffer->Flip(g_frontBuffer, getBackBuffer(), DDFLIP_WAIT);
 			*g_deviceWindowPtr = g_deviceWindow;
 		}
-		g_presentEndVsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter() + max(flipInterval, 1);
-	}
-
-	void updateNowIfNotBusy()
-	{
-		auto primary(DDraw::PrimarySurface::getPrimary());
-		RECT emptyRect = {};
-		HRESULT result = primary ? primary->BltFast(primary, 0, 0, primary, &emptyRect, DDBLTFAST_WAIT) : DD_OK;
-		g_waitingForPrimaryUnlock = DDERR_SURFACEBUSY == result || DDERR_LOCKEDSURFACES == result;
-
-		if (!g_waitingForPrimaryUnlock)
-		{
-			const auto msSinceLastUpdate = Time::qpcToMs(Time::queryPerformanceCounter() - g_qpcLastUpdate);
-			updateNow(primary, msSinceLastUpdate > Config::delayedFlipModeTimeout ? 0 : 1);
-		}
+		g_presentEndVsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter() + 1;
 	}
 
 	unsigned WINAPI updateThreadProc(LPVOID /*lpParameter*/)
 	{
-		bool skipWaitForVsync = false;
-
+		int msUntilUpdateReady = 0;
 		while (true)
 		{
-			if (!skipWaitForVsync)
+			if (msUntilUpdateReady > 0)
+			{
+				Sleep(1);
+			}
+			else
 			{
 				D3dDdi::KernelModeThunks::waitForVsyncCounter(D3dDdi::KernelModeThunks::getVsyncCounter() + 1);
 			}
-			skipWaitForVsync = false;
-			Sleep(1);
 
 			DDraw::ScopedThreadLock lock;
-			Gdi::Caret::blink();
-			if (Gdi::Cursor::update())
-			{
-				g_isUpdatePending = true;
-			}
-
-			if (g_isUpdatePending && !isPresentPending())
-			{
-				auto qpcNow = Time::queryPerformanceCounter();
-				auto qpcLastVsync = D3dDdi::KernelModeThunks::getQpcLastVsync();
-				if (Time::qpcToMs(qpcNow - qpcLastVsync) < 1 ||
-					Time::qpcToMs(qpcNow - g_qpcLastUpdate) < 1 && Time::qpcToMs(qpcNow - qpcLastVsync) <= 3)
-				{
-					skipWaitForVsync = true;
-				}
-				else
-				{
-					updateNowIfNotBusy();
-				}
-			}
+			msUntilUpdateReady = DDraw::RealPrimarySurface::flush();
 		}
-
-		return 0;
 	}
 }
 
@@ -325,27 +298,19 @@ namespace DDraw
 	HRESULT RealPrimarySurface::flip(CompatPtr<IDirectDrawSurface7> surfaceTargetOverride, DWORD flags)
 	{
 		const DWORD flipInterval = getFlipInterval(flags);
-		if (0 == flipInterval)
+		if (0 == flipInterval ||
+			Time::qpcToMs(Time::queryPerformanceCounter() - g_qpcLastUpdate) < DELAYED_FLIP_MODE_TIMEOUT_MS)
 		{
-			g_isUpdatePending = true;
-			return DD_OK;
-		}
-
-		const auto msSinceLastUpdate = Time::qpcToMs(Time::queryPerformanceCounter() - g_qpcLastUpdate);
-		const bool isFlipDelayed = msSinceLastUpdate >= 0 && msSinceLastUpdate <= Config::delayedFlipModeTimeout;
-		if (isFlipDelayed)
-		{
-			if (!isPresentPending())
-			{
-				CompatPtr<IDirectDrawSurface7> prevPrimarySurface(
-					surfaceTargetOverride ? surfaceTargetOverride : PrimarySurface::getLastSurface());
-				updateNow(prevPrimarySurface, 0);
-			}
-			g_isUpdatePending = true;
+			PrimarySurface::waitForIdle();
+			Compat::ScopedCriticalSection lock(g_presentCs);
+			g_isDelayedFlipPending = true;
+			g_isUpdatePending = false;
+			g_isUpdateReady = false;
+			g_lastUpdateThreadId = GetCurrentThreadId();
 		}
 		else
 		{
-			updateNow(PrimarySurface::getPrimary(), flipInterval);
+			updateNow(PrimarySurface::getPrimary());
 		}
 		g_flipEndVsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter() + flipInterval;
 
@@ -358,16 +323,60 @@ namespace DDraw
 		{
 			g_lastFlipSurface = nullptr;
 		}
+
+		g_qpcDelayedFlipEnd = Time::queryPerformanceCounter();
 		return DD_OK;
 	}
 
-	void RealPrimarySurface::flush()
+	int RealPrimarySurface::flush()
 	{
-		DDraw::ScopedThreadLock lock;
-		if (g_isUpdatePending && !isPresentPending())
+		auto vsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter();
+		if (static_cast<int>(vsyncCount - g_presentEndVsyncCount) < 0)
 		{
-			updateNowIfNotBusy();
+			return -1;
 		}
+
+		{
+			Compat::ScopedCriticalSection lock(g_presentCs);
+			if (!g_isUpdateReady)
+			{
+				if (g_isUpdatePending)
+				{
+					auto msSinceUpdateStart = Time::qpcToMs(Time::queryPerformanceCounter() - g_qpcUpdateStart);
+					if (msSinceUpdateStart < 10)
+					{
+						return 10 - static_cast<int>(msSinceUpdateStart);
+					}
+					g_isUpdateReady = true;
+				}
+				else if (g_isDelayedFlipPending)
+				{
+					auto msSinceDelayedFlipEnd = Time::qpcToMs(Time::queryPerformanceCounter() - g_qpcDelayedFlipEnd);
+					if (msSinceDelayedFlipEnd < 3)
+					{
+						return 3 - static_cast<int>(msSinceDelayedFlipEnd);
+					}
+					g_isDelayedFlipPending = false;
+					g_isUpdateReady = true;
+				}
+			}
+
+			if (!g_isUpdateReady)
+			{
+				return -1;
+			}
+		}
+
+		auto src(g_isDelayedFlipPending ? g_lastFlipSurface->getDDS() : DDraw::PrimarySurface::getPrimary());
+		RECT emptyRect = {};
+		HRESULT result = src ? src->BltFast(src, 0, 0, src, &emptyRect, DDBLTFAST_WAIT) : DD_OK;
+		if (DDERR_SURFACEBUSY == result || DDERR_LOCKEDSURFACES == result)
+		{
+			return 1;
+		}
+
+		updateNow(src);
+		return 0;
 	}
 
 	HWND RealPrimarySurface::getDevicePresentationWindow()
@@ -441,8 +450,16 @@ namespace DDraw
 
 	void RealPrimarySurface::scheduleUpdate()
 	{
+		Compat::ScopedCriticalSection lock(g_presentCs);
 		g_qpcLastUpdate = Time::queryPerformanceCounter();
-		g_isUpdatePending = true;
+		if (!g_isUpdatePending)
+		{
+			g_qpcUpdateStart = g_qpcLastUpdate;
+			g_isUpdatePending = true;
+			g_isDelayedFlipPending = false;
+			g_lastUpdateThreadId = GetCurrentThreadId();
+		}
+		g_isUpdateReady = false;
 	}
 
 	HRESULT RealPrimarySurface::setGammaRamp(DDGAMMARAMP* rampData)
@@ -457,14 +474,13 @@ namespace DDraw
 		return gammaControl->SetGammaRamp(gammaControl, 0, rampData);
 	}
 
-	void RealPrimarySurface::update()
+	void RealPrimarySurface::setUpdateReady()
 	{
-		DDraw::ScopedThreadLock lock;
-		g_qpcLastUpdate = Time::queryPerformanceCounter();
-		g_isUpdatePending = true;
-		if (g_waitingForPrimaryUnlock)
+		Compat::ScopedCriticalSection lock(g_presentCs);
+		if ((g_isUpdatePending || g_isDelayedFlipPending) && GetCurrentThreadId() == g_lastUpdateThreadId)
 		{
-			updateNowIfNotBusy();
+			g_isUpdateReady = true;
+			g_isDelayedFlipPending = false;
 		}
 	}
 
@@ -490,17 +506,23 @@ namespace DDraw
 			});
 	}
 
-	bool RealPrimarySurface::waitForFlip(Surface* surface)
+	bool RealPrimarySurface::waitForFlip(CompatWeakPtr<IDirectDrawSurface7> surface)
 	{
 		auto primary(DDraw::PrimarySurface::getPrimary());
-		if (!surface || !primary ||
-			surface != g_lastFlipSurface &&
-			surface != Surface::getSurface(*DDraw::PrimarySurface::getPrimary()))
+		if (!surface || !primary || !g_lastFlipSurface ||
+			surface != primary && surface != g_lastFlipSurface->getDDS())
 		{
 			return true;
 		}
 
-		D3dDdi::KernelModeThunks::waitForVsyncCounter(g_flipEndVsyncCount);
+		auto vsyncCount = D3dDdi::KernelModeThunks::getVsyncCounter();
+		while (static_cast<int>(vsyncCount - g_flipEndVsyncCount) < 0)
+		{
+			flush();
+			++vsyncCount;
+			D3dDdi::KernelModeThunks::waitForVsyncCounter(vsyncCount);
+			g_qpcDelayedFlipEnd = Time::queryPerformanceCounter();
+		}
 		return true;
 	}
 }
diff --git a/DDrawCompat/DDraw/RealPrimarySurface.h b/DDrawCompat/DDraw/RealPrimarySurface.h
index 09174a5..9ddd7a2 100644
--- a/DDrawCompat/DDraw/RealPrimarySurface.h
+++ b/DDrawCompat/DDraw/RealPrimarySurface.h
@@ -16,7 +16,7 @@ namespace DDraw
 		static HRESULT create(CompatRef<DirectDraw> dd);
 
 		static HRESULT flip(CompatPtr<IDirectDrawSurface7> surfaceTargetOverride, DWORD flags);
-		static void flush();
+		static int flush();
 		static HWND getDevicePresentationWindow();
 		static HRESULT getGammaRamp(DDGAMMARAMP* rampData);
 		static RECT getMonitorRect();
@@ -29,8 +29,8 @@ namespace DDraw
 		static HRESULT restore();
 		static void scheduleUpdate();
 		static HRESULT setGammaRamp(DDGAMMARAMP* rampData);
-		static void update();
+		static void setUpdateReady();
 		static void updateDevicePresentationWindowPos();
-		static bool waitForFlip(Surface* surface);
+		static bool waitForFlip(CompatWeakPtr<IDirectDrawSurface7> surface);
 	};
 }
diff --git a/DDrawCompat/DDraw/Surfaces/PrimarySurface.cpp b/DDrawCompat/DDraw/Surfaces/PrimarySurface.cpp
index 02f37bb..7a5f49f 100644
--- a/DDrawCompat/DDraw/Surfaces/PrimarySurface.cpp
+++ b/DDrawCompat/DDraw/Surfaces/PrimarySurface.cpp
@@ -14,6 +14,7 @@
 namespace
 {
 	CompatWeakPtr<IDirectDrawSurface7> g_primarySurface;
+	D3dDdi::Device* g_device = nullptr;
 	HANDLE g_gdiResourceHandle = nullptr;
 	HANDLE g_frontResource = nullptr;
 	DWORD g_origCaps = 0;
@@ -28,6 +29,7 @@ namespace DDraw
 	{
 		LOG_FUNC("PrimarySurface::~PrimarySurface");
 
+		g_device = nullptr;
 		g_gdiResourceHandle = nullptr;
 		g_frontResource = nullptr;
 		g_primarySurface = nullptr;
@@ -91,6 +93,7 @@ namespace DDraw
 			ResizePalette(g_palette, 256);
 		}
 
+		g_device = D3dDdi::Device::findDeviceByResource(DirectDrawSurface::getDriverResourceHandle(*surface));
 		data->restore();
 		D3dDdi::Device::updateAllConfig();
 		return DD_OK;
@@ -265,7 +268,15 @@ namespace DDraw
 			ReleaseDC(g_deviceWindow, dc);
 		}
 
-		RealPrimarySurface::update();
+		RealPrimarySurface::scheduleUpdate();
+	}
+
+	void PrimarySurface::waitForIdle()
+	{
+		if (g_device)
+		{
+			g_device->waitForIdle();
+		}
 	}
 
 	CompatWeakPtr<IDirectDrawPalette> PrimarySurface::s_palette;
diff --git a/DDrawCompat/DDraw/Surfaces/PrimarySurface.h b/DDrawCompat/DDraw/Surfaces/PrimarySurface.h
index 1f77a09..a2d8aa0 100644
--- a/DDrawCompat/DDraw/Surfaces/PrimarySurface.h
+++ b/DDrawCompat/DDraw/Surfaces/PrimarySurface.h
@@ -31,6 +31,7 @@ namespace DDraw
 		static bool isGdiSurface(TSurface* surface);
 
 		static void updateFrontResource();
+		static void waitForIdle();
 
 		virtual void restore();
 
diff --git a/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.cpp b/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.cpp
index 210da24..e8d6f57 100644
--- a/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.cpp
+++ b/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.cpp
@@ -91,11 +91,13 @@ namespace DDraw
 			return DDERR_SURFACELOST;
 		}
 
+		RealPrimarySurface::flush();
 		HRESULT result = SurfaceImpl::Blt(This, lpDestRect, lpDDSrcSurface, lpSrcRect, dwFlags, lpDDBltFx);
 		if (SUCCEEDED(result))
 		{
 			bltToGdi(This, lpDestRect, lpDDSrcSurface, lpSrcRect, dwFlags, lpDDBltFx);
-			RealPrimarySurface::update();
+			RealPrimarySurface::scheduleUpdate();
+			PrimarySurface::waitForIdle();
 		}
 		return result;
 	}
@@ -109,10 +111,12 @@ namespace DDraw
 			return DDERR_SURFACELOST;
 		}
 
+		RealPrimarySurface::flush();
 		HRESULT result = SurfaceImpl::BltFast(This, dwX, dwY, lpDDSrcSurface, lpSrcRect, dwTrans);
 		if (SUCCEEDED(result))
 		{
-			RealPrimarySurface::update();
+			RealPrimarySurface::scheduleUpdate();
+			PrimarySurface::waitForIdle();
 		}
 		return result;
 	}
@@ -120,7 +124,9 @@ namespace DDraw
 	template <typename TSurface>
 	HRESULT PrimarySurfaceImpl<TSurface>::Flip(TSurface* This, TSurface* lpDDSurfaceTargetOverride, DWORD dwFlags)
 	{
-		DDraw::RealPrimarySurface::waitForFlip(m_data);
+		RealPrimarySurface::setUpdateReady();
+		RealPrimarySurface::flush();
+		RealPrimarySurface::waitForFlip(m_data->getDDS());
 		auto surfaceTargetOverride(CompatPtr<TSurface>::from(lpDDSurfaceTargetOverride));
 		const bool isFlipEmulated = 0 != (PrimarySurface::getOrigCaps() & DDSCAPS_SYSTEMMEMORY);
 		if (isFlipEmulated)
@@ -155,6 +161,13 @@ namespace DDraw
 		return result;
 	}
 
+	template <typename TSurface>
+	HRESULT PrimarySurfaceImpl<TSurface>::GetDC(TSurface* This, HDC* lphDC)
+	{
+		RealPrimarySurface::flush();
+		return SurfaceImpl::GetDC(This, lphDC);
+	}
+
 	template <typename TSurface>
 	HRESULT PrimarySurfaceImpl<TSurface>::GetSurfaceDesc(TSurface* This, TSurfaceDesc* lpDDSurfaceDesc)
 	{
@@ -187,6 +200,7 @@ namespace DDraw
 			return DDERR_SURFACELOST;
 		}
 
+		RealPrimarySurface::flush();
 		HRESULT result = SurfaceImpl::Lock(This, lpDestRect, lpDDSurfaceDesc, dwFlags, hEvent);
 		if (SUCCEEDED(result))
 		{
@@ -201,7 +215,7 @@ namespace DDraw
 		HRESULT result = SurfaceImpl::ReleaseDC(This, hDC);
 		if (SUCCEEDED(result))
 		{
-			RealPrimarySurface::update();
+			RealPrimarySurface::scheduleUpdate();
 		}
 		return result;
 	}
@@ -244,7 +258,7 @@ namespace DDraw
 		HRESULT result = SurfaceImpl::Unlock(This, lpRect);
 		if (SUCCEEDED(result))
 		{
-			RealPrimarySurface::update();
+			RealPrimarySurface::scheduleUpdate();
 		}
 		return result;
 	}
diff --git a/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.h b/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.h
index 5edccce..9ea2b86 100644
--- a/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.h
+++ b/DDrawCompat/DDraw/Surfaces/PrimarySurfaceImpl.h
@@ -21,6 +21,7 @@ namespace DDraw
 			TSurface* lpDDSrcSurface, LPRECT lpSrcRect, DWORD dwTrans) override;
 		virtual HRESULT Flip(TSurface* This, TSurface* lpDDSurfaceTargetOverride, DWORD dwFlags) override;
 		virtual HRESULT GetCaps(TSurface* This, TDdsCaps* lpDDSCaps) override;
+		virtual HRESULT GetDC(TSurface* This, HDC* lphDC);
 		virtual HRESULT GetSurfaceDesc(TSurface* This, TSurfaceDesc* lpDDSurfaceDesc) override;
 		virtual HRESULT IsLost(TSurface* This) override;
 		virtual HRESULT Lock(TSurface* This, LPRECT lpDestRect, TSurfaceDesc* lpDDSurfaceDesc,
diff --git a/DDrawCompat/DDraw/Surfaces/Surface.h b/DDrawCompat/DDraw/Surfaces/Surface.h
index d27985c..34f57ae 100644
--- a/DDrawCompat/DDraw/Surfaces/Surface.h
+++ b/DDrawCompat/DDraw/Surfaces/Surface.h
@@ -30,6 +30,8 @@ namespace DDraw
 		template <typename TSurface>
 		static Surface* getSurface(TSurface& dds);
 
+		CompatWeakPtr<IDirectDrawSurface7> getDDS() const { return m_surface; };
+
 		template <typename TSurface>
 		SurfaceImpl<TSurface>* getImpl() const;
 
diff --git a/DDrawCompat/DDraw/Surfaces/SurfaceImpl.cpp b/DDrawCompat/DDraw/Surfaces/SurfaceImpl.cpp
index 07ea92e..b9b3280 100644
--- a/DDrawCompat/DDraw/Surfaces/SurfaceImpl.cpp
+++ b/DDrawCompat/DDraw/Surfaces/SurfaceImpl.cpp
@@ -109,7 +109,7 @@ namespace DDraw
 		TSurface* This, LPRECT lpDestRect, TSurface* lpDDSrcSurface, LPRECT lpSrcRect,
 		DWORD dwFlags, LPDDBLTFX lpDDBltFx)
 	{
-		RealPrimarySurface::waitForFlip(m_data);
+		RealPrimarySurface::waitForFlip(m_data->getDDS());
 		DirectDrawClipper::update();
 		return blt(This, lpDDSrcSurface, lpSrcRect, [=](TSurface* This, TSurface* lpDDSrcSurface, LPRECT lpSrcRect)
 			{ return getOrigVtable(This).Blt(This, lpDestRect, lpDDSrcSurface, lpSrcRect, dwFlags, lpDDBltFx); });
@@ -119,7 +119,7 @@ namespace DDraw
 	HRESULT SurfaceImpl<TSurface>::BltFast(
 		TSurface* This, DWORD dwX, DWORD dwY, TSurface* lpDDSrcSurface, LPRECT lpSrcRect, DWORD dwTrans)
 	{
-		RealPrimarySurface::waitForFlip(m_data);
+		RealPrimarySurface::waitForFlip(m_data->getDDS());
 		return blt(This, lpDDSrcSurface, lpSrcRect, [=](TSurface* This, TSurface* lpDDSrcSurface, LPRECT lpSrcRect)
 			{ return getOrigVtable(This).BltFast(This, dwX, dwY, lpDDSrcSurface, lpSrcRect, dwTrans); });
 	}
@@ -144,7 +144,7 @@ namespace DDraw
 	template <typename TSurface>
 	HRESULT SurfaceImpl<TSurface>::GetDC(TSurface* This, HDC* lphDC)
 	{
-		RealPrimarySurface::waitForFlip(m_data);
+		RealPrimarySurface::waitForFlip(m_data->getDDS());
 		HRESULT result = getOrigVtable(This).GetDC(This, lphDC);
 		if (SUCCEEDED(result))
 		{
@@ -181,7 +181,7 @@ namespace DDraw
 		TSurface* This, LPRECT lpDestRect, TSurfaceDesc* lpDDSurfaceDesc,
 		DWORD dwFlags, HANDLE hEvent)
 	{
-		RealPrimarySurface::waitForFlip(m_data);
+		RealPrimarySurface::waitForFlip(m_data->getDDS());
 		HRESULT result = getOrigVtable(This).Lock(This, lpDestRect, lpDDSurfaceDesc, dwFlags, hEvent);
 		if (SUCCEEDED(result))
 		{
diff --git a/DDrawCompat/Gdi/Cursor.cpp b/DDrawCompat/Gdi/Cursor.cpp
index f914677..10f8712 100644
--- a/DDrawCompat/Gdi/Cursor.cpp
+++ b/DDrawCompat/Gdi/Cursor.cpp
@@ -2,6 +2,7 @@
 
 #include <Common/Hook.h>
 #include <Common/Log.h>
+#include <DDraw/RealPrimarySurface.h>
 #include <DDraw/Surfaces/PrimarySurface.h>
 #include <Gdi/Cursor.h>
 #include <Win32/DisplayMode.h>
@@ -222,7 +223,7 @@ namespace Gdi
 			}
 		}
 
-		bool update()
+		void update()
 		{
 			Compat::ScopedCriticalSection lock(g_cs);
 			if (!IsRectEmpty(&g_monitorClipRect))
@@ -238,11 +239,9 @@ namespace Gdi
 					(cursorInfo.hCursor != g_prevCursorInfo.hCursor || cursorInfo.ptScreenPos != g_prevCursorInfo.ptScreenPos))
 				{
 					g_prevCursorInfo = cursorInfo;
-					return true;
+					DDraw::RealPrimarySurface::scheduleUpdate();
 				}
 			}
-
-			return false;
 		}
 	}
 }
diff --git a/DDrawCompat/Gdi/Cursor.h b/DDrawCompat/Gdi/Cursor.h
index d26987a..620bf0a 100644
--- a/DDrawCompat/Gdi/Cursor.h
+++ b/DDrawCompat/Gdi/Cursor.h
@@ -12,6 +12,6 @@ namespace Gdi
 		HCURSOR setCursor(HCURSOR cursor);
 		void setMonitorClipRect(const RECT& rect);
 		void setEmulated(bool isEmulated);
-		bool update();
+		void update();
 	}
 }
diff --git a/DDrawCompat/Gdi/GuiThread.cpp b/DDrawCompat/Gdi/GuiThread.cpp
index 4f8631b..a024389 100644
--- a/DDrawCompat/Gdi/GuiThread.cpp
+++ b/DDrawCompat/Gdi/GuiThread.cpp
@@ -4,6 +4,9 @@
 #include <Common/Hook.h>
 #include <D3dDdi/ScopedCriticalSection.h>
 #include <Dll/Dll.h>
+#include <DDraw/RealPrimarySurface.h>
+#include <Gdi/Caret.h>
+#include <Gdi/Cursor.h>
 #include <Gdi/GuiThread.h>
 #include <Gdi/Region.h>
 #include <Gdi/WinProc.h>
@@ -83,7 +86,7 @@ namespace
 		}
 
 		MSG msg = {};
-		while (GetMessage(&msg, nullptr, 0, 0))
+		while (CALL_ORIG_FUNC(GetMessageA)(&msg, nullptr, 0, 0))
 		{
 			TranslateMessage(&msg);
 			DispatchMessage(&msg);
@@ -91,6 +94,16 @@ namespace
 
 		return 0;
 	}
+
+	unsigned WINAPI updateThreadProc(LPVOID /*lpParameter*/)
+	{
+		while (true)
+		{
+			Sleep(5);
+			Gdi::Caret::blink();
+			Gdi::Cursor::update();
+		}
+	}
 }
 
 namespace Gdi
@@ -154,6 +167,7 @@ namespace Gdi
 		void start()
 		{
 			Dll::createThread(messageWindowThreadProc, &g_threadId, THREAD_PRIORITY_TIME_CRITICAL, 0);
+			Dll::createThread(updateThreadProc, nullptr, THREAD_PRIORITY_TIME_CRITICAL, 0);
 		}
 	}
 }
diff --git a/DDrawCompat/Gdi/WinProc.cpp b/DDrawCompat/Gdi/WinProc.cpp
index 4560f59..40f6a12 100644
--- a/DDrawCompat/Gdi/WinProc.cpp
+++ b/DDrawCompat/Gdi/WinProc.cpp
@@ -147,6 +147,23 @@ namespace
 		return ddcWindowProc(hwnd, uMsg, wParam, lParam, CallWindowProcW, getWindowProc(hwnd).wndProcW);
 	}
 
+	BOOL WINAPI getMessage(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax,
+		decltype(&GetMessageA) origGetMessage)
+	{
+		DDraw::RealPrimarySurface::setUpdateReady();
+		return origGetMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax);
+	}
+
+	BOOL WINAPI getMessageA(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax)
+	{
+		return getMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax, CALL_ORIG_FUNC(GetMessageA));
+	}
+
+	BOOL WINAPI getMessageW(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax)
+	{
+		return getMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax, CALL_ORIG_FUNC(GetMessageW));
+	}
+
 	LONG getWindowLong(HWND hWnd, int nIndex,
 		decltype(&GetWindowLongA) origGetWindowLong, WNDPROC(WindowProc::* wndProc))
 	{
@@ -290,6 +307,23 @@ namespace
 		}
 	}
 
+	BOOL peekMessage(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax, UINT wRemoveMsg,
+		decltype(&PeekMessageA) origPeekMessage)
+	{
+		DDraw::RealPrimarySurface::setUpdateReady();
+		return origPeekMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax, wRemoveMsg);
+	}
+
+	BOOL WINAPI peekMessageA(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax, UINT wRemoveMsg)
+	{
+		return peekMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax, wRemoveMsg, CALL_ORIG_FUNC(PeekMessageA));
+	}
+
+	BOOL WINAPI peekMessageW(LPMSG lpMsg, HWND hWnd, UINT wMsgFilterMin, UINT wMsgFilterMax, UINT wRemoveMsg)
+	{
+		return peekMessage(lpMsg, hWnd, wMsgFilterMin, wMsgFilterMax, wRemoveMsg, CALL_ORIG_FUNC(PeekMessageW));
+	}
+
 	BOOL WINAPI setLayeredWindowAttributes(HWND hwnd, COLORREF crKey, BYTE bAlpha, DWORD dwFlags)
 	{
 		LOG_FUNC("SetLayeredWindowAttributes", hwnd, crKey, bAlpha, dwFlags);
@@ -455,8 +489,12 @@ namespace Gdi
 
 		void installHooks()
 		{
+			HOOK_FUNCTION(user32, GetMessageA, getMessageA);
+			HOOK_FUNCTION(user32, GetMessageW, getMessageW);
 			HOOK_FUNCTION(user32, GetWindowLongA, getWindowLongA);
 			HOOK_FUNCTION(user32, GetWindowLongW, getWindowLongW);
+			HOOK_FUNCTION(user32, PeekMessageA, peekMessageA);
+			HOOK_FUNCTION(user32, PeekMessageW, peekMessageW);
 			HOOK_FUNCTION(user32, SetLayeredWindowAttributes, setLayeredWindowAttributes);
 			HOOK_FUNCTION(user32, SetWindowLongA, setWindowLongA);
 			HOOK_FUNCTION(user32, SetWindowLongW, setWindowLongW);