nx2: rearm clean read areas when the GPU writes them again

Port the coherency changes from alarok/Eglantine 77fc365f2 without its diagnostic framework or shutdown changes. Only GPU-dirty CPU reads become permanently preflushable, clean areas can be trapped on the next write, non-core readers avoid indexing core arrays, and collected CPU writes precede readback.

Assisted-by: Codex (source review, selective integration and verification)
(cherry picked from commit b64e211957180aa218ec6d9a3bcb020332279c3f)

3 files changed, 29 insertions, 3 deletions

diff --git a/src/core/memory.cpp b/src/core/memory.cpp
index cfc39916db..1f7315f3c1 100644
--- a/src/core/memory.cpp
+++ b/src/core/memory.cpp
@@ -549,6 +549,8 @@ struct Memory::Impl {
     }
 
     void RasterizerArmReadTraps(u64 vaddr, u64 size) {
+        // Clean cached read areas also need invalidation when reactive flushing is enabled.
+        read_trap_generation.fetch_add(1, std::memory_order_release);
         if (vaddr == 0 || size == 0 || !AddressSpaceContains(*current_page_table, vaddr, size) ||
             !ReadTrapsEnabled()) {
             return;
@@ -781,6 +783,19 @@ struct Memory::Impl {
             gpu_device_memory = &system.Host1x().MemoryManager();
         }
         const size_t core = system.GetCurrentHostThreadID();
+        if (core >= Core::Hardware::NUM_CPU_CORES) {
+            // A service thread has no slot in the arrays below, and it does not use a slot of a
+            // core. It keeps no area, so it asks on each read.
+            if (static std::atomic_flag logged; !logged.test_and_set(std::memory_order_relaxed)) {
+                LOG_INFO(HW_Memory, "Read of cached guest memory from a thread that is not a core");
+            }
+            Common::ScratchBuffer<u32> scratch;
+            gpu_device_memory->ApplyOpOnPointer(p, scratch, [&](DAddr address) {
+                DisarmReadTraps(v_address, size);
+                (void)system.GPU().OnCPURead(address, size);
+            });
+            return;
+        }
         auto& current_area = rasterizer_read_areas[core];
         if (const u64 generation = read_trap_generation.load(std::memory_order_acquire);
             read_area_generations[core] != generation) {
diff --git a/src/video_core/buffer_cache/buffer_cache.h b/src/video_core/buffer_cache/buffer_cache.h
index 481b2f0f32..f6d394ce81 100644
--- a/src/video_core/buffer_cache/buffer_cache.h
+++ b/src/video_core/buffer_cache/buffer_cache.h
@@ -234,8 +234,17 @@ std::optional<VideoCore::RasterizerDownloadArea> BufferCache<P>::GetFlushArea(DA
     };
     area->preemtive = !IsRegionGpuModified(device_addr_start_aligned,
                                            device_addr_end_aligned - device_addr_start_aligned);
-    memory_tracker.MarkRegionAsPreflushable(device_addr_start_aligned,
-                                            device_addr_end_aligned - device_addr_start_aligned);
+    // Mark only data that the GPU wrote. A mark stays for the session and sends later GPU writes
+    // in the range to a download at each fence.
+    if (!area->preemtive) {
+        memory_tracker.MarkRegionAsPreflushable(device_addr_start_aligned,
+                                                device_addr_end_aligned - device_addr_start_aligned);
+    } else {
+        // The core keeps this area and has no mark. The next GPU write here must arm the range
+        // again, which ends the area and sets the trap.
+        read_trapped_ranges.Subtract(device_addr_start_aligned,
+                                     device_addr_end_aligned - device_addr_start_aligned);
+    }
     return area;
 }
 
diff --git a/src/video_core/gpu.cpp b/src/video_core/gpu.cpp
index 53f8f063f7..681a4a232d 100644
--- a/src/video_core/gpu.cpp
+++ b/src/video_core/gpu.cpp
@@ -228,7 +228,9 @@ struct GPU::Impl {
                      raster_area.start_address, raster_area.end_address,
                      reads.load(std::memory_order_relaxed));
         }
-        const u64 fence = RequestSyncOperation([this, &raster_area]() {
+        const u64 fence = RequestSyncOperation([this, raster_area]() {
+            // Apply collected guest writes before downloading GPU data over guest memory.
+            InvalidateGPUCache();
             renderer->ReadRasterizer()->FlushRegion(raster_area.start_address, raster_area.end_address - raster_area.start_address);
         });
         gpu_thread.TickGPU(is_async);