fix gx worker exit, wrap and frame end

runtime/include/aurora_events.h

@@ -134,6 +134,7 @@ inline void ApplyPendingMkwDynamicAspectSurface() {
 
 void GxWorkerBeginFrame();
 void GxWorkerEndFrame(AuroraImguiFrame* imgui);
+void SetAuroraFrameGxBegun(bool begun);
 
 inline bool BeginAuroraFrame() {
     if (GxWorker::ShouldDefer()) {
@@ -149,6 +150,7 @@ inline bool BeginAuroraFrame() {
     if (!aurora_begin_frame()) {
         return false;
     }
+    SetAuroraFrameGxBegun(true);
     ApplyPendingMkwDynamicAspectSurface();
     return true;
 }
@@ -163,6 +165,7 @@ inline void EndAuroraFrame() {
         return;
     }
     aurora_end_frame();
+    SetAuroraFrameGxBegun(false);
 }
 
 // Poll Aurora events and update cached window/framebuffer dimensions.

runtime/include/gx_worker.h

@@ -28,6 +28,8 @@ void FlushFifo();
 // Starts the worker unless [video] gx_worker disables it. Call once Aurora is up.
 void Start();
 void Stop();
+// Stops without running queued work; bounded, for fatal exits.
+void Abandon();
 
 inline bool Active() noexcept { return detail::g_active.load(std::memory_order_relaxed); }
 inline bool OnWorker() noexcept { return detail::t_onWorker; }

runtime/src/hle/gx/gx_frame.cpp

@@ -112,26 +112,29 @@ void EnsureAuroraFrameActive() {
 }
 
 namespace {
-// GX worker only.
-bool g_workerFrameBegun = false;
+// Whether the GX half of the current frame has begun, on either path. Inline frame calls only run
+// after a sync, so the ring orders every access.
+bool g_frameGxBegun = false;
 }
 
+void SetAuroraFrameGxBegun(bool begun) { g_frameGxBegun = begun; }
+
 void GxWorkerBeginFrame() {
     // gfx::begin_frame can briefly fail to map a staging buffer; retry like the inline path does.
     const auto deadline = std::chrono::steady_clock::now() + std::chrono::milliseconds(100);
-    while (!(g_workerFrameBegun = aurora_begin_frame_gx()) && std::chrono::steady_clock::now() < deadline) {
+    while (!(g_frameGxBegun = aurora_begin_frame_gx()) && std::chrono::steady_clock::now() < deadline) {
         std::this_thread::sleep_for(std::chrono::milliseconds(1));
     }
-    if (!g_workerFrameBegun) {
+    if (!g_frameGxBegun) {
         RT_LOGF(RT_TAG_GX, "GX worker could not begin an aurora frame\n");
     }
 }
 
 void GxWorkerEndFrame(AuroraImguiFrame* imgui) {
-    if (g_workerFrameBegun) {
+    if (g_frameGxBegun) {
         aurora_end_frame_gx(imgui);
     } else {
         aurora_release_imgui_frame(imgui);
     }
-    g_workerFrameBegun = false;
+    g_frameGxBegun = false;
 }

runtime/src/hle/gx/gx_worker.cpp

@@ -60,6 +60,8 @@ std::mutex g_tailMutex;
 std::condition_variable g_tailCv;
 void (*g_waitCallback)() = nullptr;
 std::atomic<bool> g_stop{false};
+std::atomic<bool> g_abandon{false};
+std::atomic<bool> g_exited{false};
 std::thread* g_thread = nullptr;
 std::vector<uint8_t> g_fifoBytes;
 
@@ -139,6 +141,13 @@ void ReplayFifo(void* payload) {
 
 void WorkerMain() {
     detail::t_onWorker = true;
+    struct SignalExit {
+        ~SignalExit() {
+            std::lock_guard lock{g_tailMutex};
+            g_exited.store(true, std::memory_order_release);
+            g_tailCv.notify_all();
+        }
+    } signalExit;
     AuroraSetGxProducerThread();
 #if defined(_WIN32)
     SetThreadDescription(GetCurrentThread(), L"GX worker");
@@ -173,6 +182,9 @@ void WorkerMain() {
             continue;
         }
         while (tail != head) {
+            if (g_abandon.load(std::memory_order_acquire)) {
+                return;
+            }
             uint8_t* record = g_ring + tail % kCapacity;
             const Header* header = reinterpret_cast<const Header*>(record);
             if (header->run != nullptr) {
@@ -205,15 +217,16 @@ void* Reserve(size_t payloadSize, void (*run)(void*)) {
     size_t offset = static_cast<size_t>(g_write % kCapacity);
     const size_t remaining = kCapacity - offset;
     if (total > remaining) {
-        WaitForSpace(remaining + total);
+        WaitForSpace(remaining);
         auto* wrap = reinterpret_cast<Header*>(g_ring + offset);
         wrap->run = nullptr;
         wrap->size = static_cast<uint32_t>(remaining);
         g_write += remaining;
+        // Publish the skip so the worker can drain past it while we wait for the front.
+        Commit();
         offset = 0;
-    } else {
-        WaitForSpace(total);
     }
+    WaitForSpace(total);
     auto* header = reinterpret_cast<Header*>(g_ring + offset);
     header->run = run;
     header->size = static_cast<uint32_t>(total);
@@ -303,6 +316,8 @@ void Start() {
     }
     g_fifoBytes.reserve(kFifoFlushBytes + 4096);
     g_stop.store(false, std::memory_order_relaxed);
+    g_abandon.store(false, std::memory_order_relaxed);
+    g_exited.store(false, std::memory_order_relaxed);
     g_thread = new std::thread(WorkerMain);
     detail::g_active.store(true, std::memory_order_release);
     RT_LOGF(RT_TAG_GX, "GX worker started\n");
@@ -325,4 +340,32 @@ void Stop() {
     }
 }
 
+void Abandon() {
+    if (!Active() || detail::t_onWorker) {
+        return;
+    }
+    detail::g_active.store(false, std::memory_order_release);
+    g_abandon.store(true, std::memory_order_release);
+    g_stop.store(true, std::memory_order_release);
+    {
+        std::lock_guard lock{g_headMutex};
+        g_headCv.notify_one();
+    }
+    bool exited = false;
+    {
+        std::unique_lock lock{g_tailMutex};
+        exited = g_tailCv.wait_for(lock, std::chrono::seconds{2},
+                                   [] { return g_exited.load(std::memory_order_acquire); });
+    }
+    if (g_thread != nullptr) {
+        if (exited) {
+            g_thread->join();
+            delete g_thread;
+        } else {
+            g_thread->detach();
+        }
+        g_thread = nullptr;
+    }
+}
+
 } // namespace GxWorker

runtime/src/main.cpp

@@ -765,6 +765,8 @@ extern "C" void DumpHostStackTraceForRuntimeHelper() {
 
 void MarkFatalErrorReported() {
     g_fatalErrorReported.store(true, std::memory_order_release);
+    // Every caller exits next; keep the worker out of static teardown.
+    GxWorker::Abandon();
 }
 
 void ShowRuntimeFatalPopup(std::string_view category, std::string_view details) noexcept {