feat(capture/windows): hook APIs to avoid output reparenting that bre…

…aks DDA (#3530)

* Revert "feat(ddprobe): allow to manually specify gpu preference (#3521)"

This reverts commit 6a233cb.

* Keep display revert delay input type change from 6a233cb

* Remove ddprobe

* feat(capture/windows): hook APIs to avoid output reparenting that breaks DDA

.codeql-prebuild-cpp-Windows.sh

@@ -11,6 +11,7 @@ dependencies=(
   "mingw-w64-ucrt-x86_64-cmake"
   "mingw-w64-ucrt-x86_64-cppwinrt"
   "mingw-w64-ucrt-x86_64-curl-winssl"
+  "mingw-w64-ucrt-x86_64-MinHook"
   "mingw-w64-ucrt-x86_64-miniupnpc"
   "mingw-w64-ucrt-x86_64-nlohmann-json"
   "mingw-w64-ucrt-x86_64-nodejs"

.github/workflows/CI.yml

@@ -885,6 +885,7 @@ jobs:
             mingw-w64-ucrt-x86_64-cppwinrt
             mingw-w64-ucrt-x86_64-curl-winssl
             mingw-w64-ucrt-x86_64-graphviz
+            mingw-w64-ucrt-x86_64-MinHook
             mingw-w64-ucrt-x86_64-miniupnpc
             mingw-w64-ucrt-x86_64-nlohmann-json
             mingw-w64-ucrt-x86_64-nodejs

cmake/dependencies/common.cmake

@@ -28,7 +28,7 @@ include_directories(SYSTEM ${MINIUPNP_INCLUDE_DIRS})
 # ffmpeg pre-compiled binaries
 if(NOT DEFINED FFMPEG_PREPARED_BINARIES)
     if(WIN32)
-        set(FFMPEG_PLATFORM_LIBRARIES mfplat ole32 strmiids mfuuid vpl)
+        set(FFMPEG_PLATFORM_LIBRARIES mfplat ole32 strmiids mfuuid vpl MinHook)
     elseif(UNIX AND NOT APPLE)
         set(FFMPEG_PLATFORM_LIBRARIES numa va va-drm va-x11 X11)
     endif()

cmake/packaging/windows.cmake

@@ -11,7 +11,6 @@ install(TARGETS dxgi-info RUNTIME DESTINATION "tools" COMPONENT dxgi)
 install(TARGETS audio-info RUNTIME DESTINATION "tools" COMPONENT audio)
 
 # Mandatory tools
-install(TARGETS ddprobe RUNTIME DESTINATION "tools" COMPONENT application)
 install(TARGETS sunshinesvc RUNTIME DESTINATION "tools" COMPONENT application)
 
 # Mandatory scripts

docs/building.md

@@ -90,6 +90,7 @@ dependencies=(
   "mingw-w64-ucrt-x86_64-curl-winssl"
   "mingw-w64-ucrt-x86_64-doxygen"  # Optional, for docs... better to install official Doxygen
   "mingw-w64-ucrt-x86_64-graphviz"  # Optional, for docs
+  "mingw-w64-ucrt-x86_64-MinHook"
   "mingw-w64-ucrt-x86_64-miniupnpc"
   "mingw-w64-ucrt-x86_64-nlohmann-json"
   "mingw-w64-ucrt-x86_64-nodejs"

docs/configuration.md

@@ -848,37 +848,6 @@ editing the `conf` file in a text editor. Use the examples as reference.
     </tr>
 </table>
 
-### gpu_preference
-
-<table>
-    <tr>
-        <td>Description</td>
-        <td colspan="2">
-            Specify the GPU preference for the Sunshine process.
-            <br>
-            <br>
-            If set to negative number (-1 by default), Sunshine will try to detect the best GPU for the streamed display, but if it fails you will get a black screen.
-            <br>
-            Setting it to 0 will allow Windows to try and select the best GPU.
-            <br>
-            Setting it to 1 and above will prioritize the GPU that matches this number (the number has to be guessed, but it starts at 1 and increases).
-            @note{Applies to Windows only.}
-        </td>
-    </tr>
-    <tr>
-        <td>Default</td>
-        <td colspan="2">@code{}
-            -1
-            @endcode</td>
-    </tr>
-    <tr>
-        <td>Example</td>
-        <td colspan="2">@code{}
-            2
-            @endcode</td>
-    </tr>
-</table>
-
 ### output_name
 
 <table>

src/config.cpp

@@ -468,7 +468,6 @@ namespace config {
     {},  // capture
     {},  // encoder
     {},  // adapter_name
-    -1,  // gpu_preference
     {},  // output_name
 
     {
@@ -1122,7 +1121,6 @@ namespace config {
     string_f(vars, "capture", video.capture);
     string_f(vars, "encoder", video.encoder);
     string_f(vars, "adapter_name", video.adapter_name);
-    int_f(vars, "gpu_preference", video.gpu_preference);
     string_f(vars, "output_name", video.output_name);
 
     generic_f(vars, "dd_configuration_option", video.dd.configuration_option, dd::config_option_from_view);

src/config.h

@@ -77,7 +77,6 @@ namespace config {
     std::string capture;
     std::string encoder;
     std::string adapter_name;
-    int gpu_preference;
     std::string output_name;
 
     struct dd_t {

src/platform/windows/display_base.cpp

@@ -9,10 +9,22 @@
 #include <boost/algorithm/string/join.hpp>
 #include <boost/process/v1.hpp>
 
+#include <MinHook.h>
+
 // We have to include boost/process/v1.hpp before display.h due to WinSock.h,
 // but that prevents the definition of NTSTATUS so we must define it ourself.
 typedef long NTSTATUS;
 
+// Definition from the WDK's d3dkmthk.h
+typedef enum _D3DKMT_GPU_PREFERENCE_QUERY_STATE: DWORD {
+  D3DKMT_GPU_PREFERENCE_STATE_UNINITIALIZED,  ///< The GPU preference isn't initialized.
+  D3DKMT_GPU_PREFERENCE_STATE_HIGH_PERFORMANCE,  ///< The highest performing GPU is preferred.
+  D3DKMT_GPU_PREFERENCE_STATE_MINIMUM_POWER,  ///< The minimum-powered GPU is preferred.
+  D3DKMT_GPU_PREFERENCE_STATE_UNSPECIFIED,  ///< A GPU preference isn't specified.
+  D3DKMT_GPU_PREFERENCE_STATE_NOT_FOUND,  ///< A GPU preference isn't found.
+  D3DKMT_GPU_PREFERENCE_STATE_USER_SPECIFIED_GPU  ///< A specific GPU is preferred.
+} D3DKMT_GPU_PREFERENCE_QUERY_STATE;
+
 #include "display.h"
 #include "misc.h"
 #include "src/config.h"
@@ -329,115 +341,6 @@ namespace platf::dxgi {
     return capture_e::ok;
   }
 
-  bool
-  set_gpu_preference_on_self(int preference) {
-    // The GPU preferences key uses app path as the value name.
-    WCHAR sunshine_path[MAX_PATH];
-    GetModuleFileNameW(NULL, sunshine_path, ARRAYSIZE(sunshine_path));
-
-    WCHAR value_data[128];
-    swprintf_s(value_data, L"GpuPreference=%d;", preference);
-
-    auto status = RegSetKeyValueW(HKEY_CURRENT_USER,
-      L"Software\\Microsoft\\DirectX\\UserGpuPreferences",
-      sunshine_path,
-      REG_SZ,
-      value_data,
-      (wcslen(value_data) + 1) * sizeof(WCHAR));
-    if (status != ERROR_SUCCESS) {
-      BOOST_LOG(error) << "Failed to set GPU preference: "sv << status;
-      return false;
-    }
-
-    BOOST_LOG(info) << "Set GPU preference: "sv << preference;
-    return true;
-  }
-
-  bool
-  validate_and_test_gpu_preference(const std::string &display_name, bool verify_frame_capture) {
-    std::string cmd = "tools\\ddprobe.exe";
-
-    // We start at 1 because 0 is automatic selection which can be overridden by
-    // the GPU driver control panel options. Since ddprobe.exe can have different
-    // GPU driver overrides than Sunshine.exe, we want to avoid a scenario where
-    // autoselection might work for ddprobe.exe but not for us.
-    for (int i = 1; i < 5; i++) {
-      // Run the probe tool. It returns the status of DuplicateOutput().
-      //
-      // Arg format: [GPU preference] [Display name] [--verify-frame-capture]
-      HRESULT result;
-      std::vector<std::string> args = { std::to_string(i), display_name };
-      try {
-        if (verify_frame_capture) {
-          args.emplace_back("--verify-frame-capture");
-        }
-        result = bp::system(cmd, bp::args(args), bp::std_out > bp::null, bp::std_err > bp::null);
-      }
-      catch (bp::process_error &e) {
-        BOOST_LOG(error) << "Failed to start ddprobe.exe: "sv << e.what();
-        return false;
-      }
-
-      BOOST_LOG(info) << "ddprobe.exe " << boost::algorithm::join(args, " ") << " returned 0x"
-                      << util::hex(result).to_string_view();
-
-      // E_ACCESSDENIED can happen at the login screen. If we get this error,
-      // we know capture would have been supported, because DXGI_ERROR_UNSUPPORTED
-      // would have been raised first if it wasn't.
-      if (result == S_OK || result == E_ACCESSDENIED) {
-        // We found a working GPU preference, so set ourselves to use that.
-        return set_gpu_preference_on_self(i);
-      }
-    }
-
-    // If no valid configuration was found, return false
-    return false;
-  }
-
-  // On hybrid graphics systems, Windows will change the order of GPUs reported by
-  // DXGI in accordance with the user's GPU preference. If the selected GPU is a
-  // render-only device with no displays, DXGI will add virtual outputs to the
-  // that device to avoid confusing applications. While this works properly for most
-  // applications, it breaks the Desktop Duplication API because DXGI doesn't proxy
-  // the virtual DXGIOutput to the real GPU it is attached to. When trying to call
-  // DuplicateOutput() on one of these virtual outputs, it fails with DXGI_ERROR_UNSUPPORTED
-  // (even if you try sneaky stuff like passing the ID3D11Device for the iGPU and the
-  // virtual DXGIOutput from the dGPU). Because the GPU preference is once-per-process,
-  // we spawn a helper tool to probe for us before we set our own GPU preference.
-  bool
-  probe_for_gpu_preference(const std::string &display_name) {
-    static bool set_gpu_preference = false;
-
-    // If we've already been through here, there's nothing to do this time.
-    if (set_gpu_preference) {
-      return true;
-    }
-
-    // If the GPU preference was manually specified, we can skip the probe.
-    if (config::video.gpu_preference >= 0) {
-      if (set_gpu_preference_on_self(config::video.gpu_preference)) {
-        set_gpu_preference = true;
-        return true;
-      }
-    }
-    else {
-      // Try probing with different GPU preferences and verify_frame_capture flag
-      if (validate_and_test_gpu_preference(display_name, true)) {
-        set_gpu_preference = true;
-        return true;
-      }
-
-      // If no valid configuration was found, try again with verify_frame_capture == false
-      if (validate_and_test_gpu_preference(display_name, false)) {
-        set_gpu_preference = true;
-        return true;
-      }
-    }
-
-    // If neither worked, return false
-    return false;
-  }
-
   /**
    * @brief Tests to determine if the Desktop Duplication API can capture the given output.
    * @details When testing for enumeration only, we avoid resyncing the thread desktop.
@@ -510,6 +413,27 @@ namespace platf::dxgi {
     return false;
   }
 
+  /**
+   * @brief Hook for NtGdiDdDDIGetCachedHybridQueryValue() from win32u.dll.
+   * @param gpuPreference A pointer to the location where the preference will be written.
+   * @return Always STATUS_SUCCESS if valid arguments are provided.
+   */
+  NTSTATUS
+  __stdcall NtGdiDdDDIGetCachedHybridQueryValueHook(D3DKMT_GPU_PREFERENCE_QUERY_STATE *gpuPreference) {
+    // By faking a cached GPU preference state of D3DKMT_GPU_PREFERENCE_STATE_UNSPECIFIED, this will
+    // prevent DXGI from performing the normal GPU preference resolution that looks at the registry,
+    // power settings, and the hybrid adapter DDI interface to pick a GPU. Instead, we will not be
+    // bound to any specific GPU. This will prevent DXGI from performing output reparenting (moving
+    // outputs from their true location to the render GPU), which breaks DDA.
+    if (gpuPreference) {
+      *gpuPreference = D3DKMT_GPU_PREFERENCE_STATE_UNSPECIFIED;
+      return 0;  // STATUS_SUCCESS
+    }
+    else {
+      return STATUS_INVALID_PARAMETER;
+    }
+  }
+
   int
   display_base_t::init(const ::video::config_t &config, const std::string &display_name) {
     std::once_flag windows_cpp_once_flag;
@@ -519,13 +443,22 @@ namespace platf::dxgi {
 
       typedef BOOL (*User32_SetProcessDpiAwarenessContext)(DPI_AWARENESS_CONTEXT value);
 
-      auto user32 = LoadLibraryA("user32.dll");
-      auto f = (User32_SetProcessDpiAwarenessContext) GetProcAddress(user32, "SetProcessDpiAwarenessContext");
-      if (f) {
-        f(DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2);
+      {
+        auto user32 = LoadLibraryA("user32.dll");
+        auto f = (User32_SetProcessDpiAwarenessContext) GetProcAddress(user32, "SetProcessDpiAwarenessContext");
+        if (f) {
+          f(DPI_AWARENESS_CONTEXT_PER_MONITOR_AWARE_V2);
+        }
+
+        FreeLibrary(user32);
       }
 
-      FreeLibrary(user32);
+      {
+        // We aren't calling MH_Uninitialize(), but that's okay because this hook lasts for the life of the process
+        MH_Initialize();
+        MH_CreateHookApi(L"win32u.dll", "NtGdiDdDDIGetCachedHybridQueryValue", (void *) NtGdiDdDDIGetCachedHybridQueryValueHook, nullptr);
+        MH_EnableHook(MH_ALL_HOOKS);
+      }
     });
 
     // Get rectangle of full desktop for absolute mouse coordinates
@@ -534,11 +467,6 @@ namespace platf::dxgi {
 
     HRESULT status;
 
-    // We must set the GPU preference before calling any DXGI APIs!
-    if (!probe_for_gpu_preference(display_name)) {
-      BOOST_LOG(warning) << "Failed to set GPU preference. Capture may not work!"sv;
-    }
-
     status = CreateDXGIFactory1(IID_IDXGIFactory1, (void **) &factory);
     if (FAILED(status)) {
       BOOST_LOG(error) << "Failed to create DXGIFactory1 [0x"sv << util::hex(status).to_string_view() << ']';
@@ -1105,12 +1033,6 @@ namespace platf {
 
     BOOST_LOG(debug) << "Detecting monitors..."sv;
 
-    // We must set the GPU preference before calling any DXGI APIs!
-    const auto output_name { display_device::map_output_name(config::video.output_name) };
-    if (!dxgi::probe_for_gpu_preference(output_name)) {
-      BOOST_LOG(warning) << "Failed to set GPU preference. Capture may not work!"sv;
-    }
-
     // We sync the thread desktop once before we start the enumeration process
     // to ensure test_dxgi_duplication() returns consistent results for all GPUs
     // even if the current desktop changes during our enumeration process.

src_assets/common/assets/web/config.html

@@ -167,7 +167,6 @@ <h1 class="my-4">{{ $t('config.configuration') }}</h1>
               "virtual_sink": "",
               "install_steam_audio_drivers": "enabled",
               "adapter_name": "",
-              "gpu_preference": -1,
               "output_name": "",
               "dd_configuration_option": "verify_only",
               "dd_resolution_option": "auto",

src_assets/common/assets/web/configs/tabs/AudioVideo.vue

@@ -68,18 +68,6 @@ const config = ref(props.config)
         :config="config"
     />
 
-    <PlatformLayout :platform="platform">
-      <template #windows>
-        <!-- GPU Preference -->
-        <div class="mb-3">
-          <label for="gpu_preference" class="form-label">{{ $t('config.gpu_preference') }}</label>
-          <input type="number" class="form-control" id="gpu_preference" placeholder="-1" min="-1"
-                 v-model="config.gpu_preference" />
-          <div class="form-text">{{ $t('config.gpu_preference_desc') }}</div>
-        </div>
-      </template>
-    </PlatformLayout>
-
     <DisplayOutputSelector
       :platform="platform"
       :config="config"

src_assets/common/assets/web/public/assets/locale/en.json

@@ -199,8 +199,6 @@
     "fec_percentage": "FEC Percentage",
     "fec_percentage_desc": "Percentage of error correcting packets per data packet in each video frame. Higher values can correct for more network packet loss, but at the cost of increasing bandwidth usage.",
     "ffmpeg_auto": "auto -- let ffmpeg decide (default)",
-    "gpu_preference": "GPU Preference",
-    "gpu_preference_desc": "Specify the GPU preference for the Sunshine process. If set to negative number (-1 by default), Sunshine will try to detect the best GPU for the streamed display, but if it fails you will get a black screen. Setting it to 0 will allow Windows to try and select the best GPU. Setting it to 1 and above will prioritize the GPU that matches this number (the number has to be guessed, but it starts at 1 and increases).",
     "file_apps": "Apps File",
     "file_apps_desc": "The file where current apps of Sunshine are stored.",
     "file_state": "State File",

tools/CMakeLists.txt

@@ -27,12 +27,3 @@ target_link_libraries(sunshinesvc
         wtsapi32
         ${PLATFORM_LIBRARIES})
 target_compile_options(sunshinesvc PRIVATE ${SUNSHINE_COMPILE_OPTIONS})
-
-add_executable(ddprobe ddprobe.cpp)
-set_target_properties(ddprobe PROPERTIES CXX_STANDARD 20)
-target_link_libraries(ddprobe
-        ${CMAKE_THREAD_LIBS_INIT}
-        dxgi
-        d3d11
-        ${PLATFORM_LIBRARIES})
-target_compile_options(ddprobe PRIVATE ${SUNSHINE_COMPILE_OPTIONS})