[gpu] Adjustments on the frame composer

src/core/hle/service/nvdrv/devices/nvdisp_disp0.cpp

@@ -94,6 +94,10 @@ void nvdisp_disp0::Composite(std::span<const Nvnflinger::HwcLayer> sorted_layers
     Common::ADPF::BeginFrameWork();
 }
 
+void nvdisp_disp0::WaitForComposite() {
+    system.GPU().WaitForComposite();
+}
+
 Kernel::KEvent* nvdisp_disp0::QueryEvent(u32 event_id) {
     LOG_CRITICAL(Service_NVDRV, "Unknown DISP Event {}", event_id);
     return nullptr;

src/core/hle/service/nvdrv/devices/nvdisp_disp0.h

@@ -37,6 +37,8 @@ public:
     /// Performs a screen flip, compositing each buffer.
     void Composite(std::span<const Nvnflinger::HwcLayer> sorted_layers);
 
+    void WaitForComposite();
+
     Kernel::KEvent* QueryEvent(u32 event_id) override;
 
 private:

src/core/hle/service/nvnflinger/hardware_composer.cpp

@@ -57,6 +57,11 @@ u32 HardwareComposer::ComposeLocked(f32* out_speed_scale, Display& display,
     // Set default speed limit to 100%.
     *out_speed_scale = 1.0f;
 
+    // The composition queued last frame reads guest memory on the GPU thread, so its framebuffers
+    // may only go back to the guest once it has run.
+    nvdisp.WaitForComposite();
+    this->ReleaseFramebuffersLocked(display);
+
     // If no layers are available, skip the logic.
     bool any_visible = false;
     for (auto& layer : display.stack.layers) {
@@ -158,55 +163,30 @@ u32 HardwareComposer::ComposeLocked(f32* out_speed_scale, Display& display,
         nvdisp.Composite(composition_stack);
     }
 
-    // Batch framebuffer releases, instead of one-into-one.
-    std::vector<std::pair<Layer*, Framebuffer*>> to_release;
-    for (auto& [layer_id, framebuffer] : m_framebuffers) {
-        if (!framebuffer.is_acquired)
-            continue;
-
-        auto layer = display.stack.FindLayer(layer_id);
-        if (!layer)
-            continue;
-
-        // Overlay layers always release after every compose
-        // Non-overlay layers release based on their swap interval
-        if (layer->is_overlay || framebuffer.release_frame_number <= m_frame_number) {
-            to_release.emplace_back(layer.get(), &framebuffer);
-        }
-    }
-    for (auto& [layer, framebuffer] : to_release) {
-        layer->buffer_item_consumer->ReleaseBuffer(framebuffer->item, android::Fence::NoFence());
-        framebuffer->is_acquired = false;
-    }
-
     // Advance by 1 frame (60 FPS compositing)
     m_frame_number += 1;
 
-    // Release any necessary framebuffers (non-overlay layers only, as overlays are already released above).
+    return 1;
+}
+
+void HardwareComposer::ReleaseFramebuffersLocked(Display& display) {
     for (auto& [layer_id, framebuffer] : m_framebuffers) {
         if (!framebuffer.is_acquired) {
-            // Already released.
             continue;
         }
 
-        if (framebuffer.release_frame_number > m_frame_number) {
+        const auto layer = display.stack.FindLayer(layer_id);
+        if (!layer) {
             continue;
         }
 
-        if (const auto layer = display.stack.FindLayer(layer_id); layer != nullptr) {
-            // Skip overlay layers as they were already released above
-            if (layer->is_overlay) {
-                continue;
-            }
-
-            // TODO: support release fence
-            // This is needed to prevent screen tearing
-            layer->buffer_item_consumer->ReleaseBuffer(framebuffer.item, android::Fence::NoFence());
-            framebuffer.is_acquired = false;
+        if (!layer->is_overlay && framebuffer.release_frame_number > m_frame_number) {
+            continue;
         }
+
+        layer->buffer_item_consumer->ReleaseBuffer(framebuffer.item, android::Fence::NoFence());
+        framebuffer.is_acquired = false;
     }
-
-    return 1;
 }
 
 void HardwareComposer::RemoveLayerLocked(Display& display, ConsumerId consumer_id) {

src/core/hle/service/nvnflinger/hardware_composer.h

@@ -52,6 +52,7 @@ private:
 private:
     bool TryAcquireFramebufferLocked(Layer& layer, Framebuffer& framebuffer);
     CacheStatus CacheFramebufferLocked(Layer& layer, ConsumerId consumer_id);
+    void ReleaseFramebuffersLocked(Display& display);
 };
 
 } // namespace Service::Nvnflinger

src/video_core/gpu.cpp

@@ -10,6 +10,7 @@
 #include <condition_variable>
 #include <list>
 #include <memory>
+#include <utility>
 
 #include "common/assert.h"
 #include "common/settings.h"
@@ -129,7 +130,7 @@ struct GPU::Impl {
     [[nodiscard]] u64 RequestSyncOperation(Func&& action) {
         std::unique_lock lck{sync_request_mutex};
         const u64 fence = ++last_sync_fence;
-        sync_requests.emplace_back(action);
+        sync_requests.emplace_back(std::forward<Func>(action));
         return fence;
     }
 
@@ -232,9 +233,9 @@ struct GPU::Impl {
     }
 
     void RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers, std::vector<Service::Nvidia::NvFence>&& fences) {
-        size_t num_fences{fences.size()};
+        const size_t num_fences{fences.size()};
         size_t current_request_counter{};
-        {
+        if (num_fences != 0) {
             std::unique_lock<std::mutex> lk(request_swap_mutex);
             if (free_swap_counters.empty()) {
                 current_request_counter = request_swap_counters.size();
@@ -245,27 +246,41 @@ struct GPU::Impl {
                 free_swap_counters.pop_front();
             }
         }
-        const auto wait_fence = RequestSyncOperation([this, current_request_counter, &layers, &fences, num_fences] {
-            auto& syncpoint_manager = system.Host1x().GetSyncpointManager();
-            if (num_fences == 0) {
-                renderer->Composite(layers);
-            }
-            const auto executer = [this, current_request_counter, layers_copy = layers]() {
-                {
-                    std::unique_lock<std::mutex> lk(request_swap_mutex);
-                    if (--request_swap_counters[current_request_counter] != 0) {
-                        return;
-                    }
-                    free_swap_counters.push_back(current_request_counter);
+        pending_composite_fence = RequestSyncOperation(
+            [this, current_request_counter, num_fences, layers = std::move(layers),
+             fences = std::move(fences)] {
+                if (num_fences == 0) {
+                    renderer->Composite(layers);
+                    return;
                 }
-                renderer->Composite(layers_copy);
-            };
-            for (size_t i = 0; i < num_fences; i++) {
-                syncpoint_manager.RegisterGuestAction(fences[i].id, fences[i].value, executer);
-            }
-        });
+                auto& syncpoint_manager = system.Host1x().GetSyncpointManager();
+                const auto executer = [this, current_request_counter, layers]() {
+                    {
+                        std::unique_lock<std::mutex> lk(request_swap_mutex);
+                        if (--request_swap_counters[current_request_counter] != 0) {
+                            return;
+                        }
+                        free_swap_counters.push_back(current_request_counter);
+                    }
+                    renderer->Composite(layers);
+                };
+                for (size_t i = 0; i < num_fences; i++) {
+                    syncpoint_manager.RegisterGuestAction(fences[i].id, fences[i].value, executer);
+                }
+            });
         gpu_thread.TickGPU(is_async);
-        WaitForSyncOperation(wait_fence);
+    }
+
+    void WaitForComposite() {
+        const u64 fence = pending_composite_fence;
+        if (fence == 0) {
+            return;
+        }
+        pending_composite_fence = 0;
+        if (shutting_down.load(std::memory_order_relaxed)) {
+            return;
+        }
+        WaitForSyncOperation(fence);
     }
 
     std::vector<u8> GetAppletCaptureBuffer() {
@@ -318,6 +333,7 @@ struct GPU::Impl {
     std::deque<size_t> free_swap_counters;
     std::deque<size_t> request_swap_counters;
     std::mutex request_swap_mutex;
+    u64 pending_composite_fence{};
 };
 
 GPU::GPU(Core::System& system, bool is_async, bool use_nvdec)
@@ -435,6 +451,10 @@ void GPU::RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
     impl->RequestComposite(std::move(layers), std::move(fences));
 }
 
+void GPU::WaitForComposite() {
+    impl->WaitForComposite();
+}
+
 std::vector<u8> GPU::GetAppletCaptureBuffer() {
     return impl->GetAppletCaptureBuffer();
 }

src/video_core/gpu.h

@@ -218,6 +218,8 @@ public:
     void RequestComposite(std::vector<Tegra::FramebufferConfig>&& layers,
                           std::vector<Service::Nvidia::NvFence>&& fences);
 
+    void WaitForComposite();
+
     std::vector<u8> GetAppletCaptureBuffer();
 
     /// Performs any additional setup necessary in order to begin GPU emulation.