gpu: Use std::stop_token in WaitFence for VSync thread

Fixes a hang that may occur when stopping emulation and the VSync thread is blocked on the syncpoint condition variable.

src/core/hle/service/nvflinger/nvflinger.cpp

@@ -266,10 +266,11 @@ void NVFlinger::Compose() {
 
         auto& gpu = system.GPU();
         const auto& multi_fence = buffer->get().multi_fence;
+        const auto stop_token = vsync_thread.get_stop_token();
         guard->unlock();
         for (u32 fence_id = 0; fence_id < multi_fence.num_fences; fence_id++) {
             const auto& fence = multi_fence.fences[fence_id];
-            gpu.WaitFence(fence.id, fence.value);
+            gpu.WaitFence(fence.id, fence.value, stop_token);
         }
         guard->lock();
 

src/video_core/gpu.cpp

@@ -206,7 +206,7 @@ struct GPU::Impl {
     }
 
     /// Allows the CPU/NvFlinger to wait on the GPU before presenting a frame.
-    void WaitFence(u32 syncpoint_id, u32 value) {
+    void WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token = {}) {
         // Synced GPU, is always in sync
         if (!is_async) {
             return;
@@ -218,13 +218,8 @@ struct GPU::Impl {
         }
         MICROPROFILE_SCOPE(GPU_wait);
         std::unique_lock lock{sync_mutex};
-        sync_cv.wait(lock, [=, this] {
-            if (shutting_down.load(std::memory_order_relaxed)) {
-                // We're shutting down, ensure no threads continue to wait for the next syncpoint
-                return true;
-            }
-            return syncpoints.at(syncpoint_id).load() >= value;
-        });
+        sync_cv.wait(lock, stop_token,
+                     [=, this] { return syncpoints.at(syncpoint_id).load() >= value; });
     }
 
     void IncrementSyncPoint(u32 syncpoint_id) {
@@ -670,8 +665,6 @@ struct GPU::Impl {
     std::unique_ptr<Engines::KeplerMemory> kepler_memory;
     /// Shader build notifier
     std::unique_ptr<VideoCore::ShaderNotify> shader_notify;
-    /// When true, we are about to shut down emulation session, so terminate outstanding tasks
-    std::atomic_bool shutting_down{};
 
     std::array<std::atomic<u32>, Service::Nvidia::MaxSyncPoints> syncpoints{};
 
@@ -680,7 +673,7 @@ struct GPU::Impl {
     std::mutex sync_mutex;
     std::mutex device_mutex;
 
-    std::condition_variable sync_cv;
+    std::condition_variable_any sync_cv;
 
     struct FlushRequest {
         explicit FlushRequest(u64 fence_, VAddr addr_, std::size_t size_)
@@ -819,8 +812,8 @@ const VideoCore::ShaderNotify& GPU::ShaderNotify() const {
     return impl->ShaderNotify();
 }
 
-void GPU::WaitFence(u32 syncpoint_id, u32 value) {
-    impl->WaitFence(syncpoint_id, value);
+void GPU::WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token) {
+    impl->WaitFence(syncpoint_id, value, stop_token);
 }
 
 void GPU::IncrementSyncPoint(u32 syncpoint_id) {

src/video_core/gpu.h

@@ -5,6 +5,7 @@
 #pragma once
 
 #include <memory>
+#include <stop_token>
 
 #include "common/bit_field.h"
 #include "common/common_types.h"
@@ -209,7 +210,7 @@ public:
     [[nodiscard]] const VideoCore::ShaderNotify& ShaderNotify() const;
 
     /// Allows the CPU/NvFlinger to wait on the GPU before presenting a frame.
-    void WaitFence(u32 syncpoint_id, u32 value);
+    void WaitFence(u32 syncpoint_id, u32 value, std::stop_token stop_token = {});
 
     void IncrementSyncPoint(u32 syncpoint_id);