buffer_cache: Make ReadMemory thread-safe

src/video_core/amdgpu/liverpool.cpp

@@ -73,6 +73,20 @@ Liverpool::~Liverpool() {
     process_thread.join();
 }
 
+void Liverpool::ProcessCommands() {
+    // Process incoming commands with high priority
+    while (num_commands) {
+        Common::UniqueFunction<void> callback{};
+        {
+            std::unique_lock lk{submit_mutex};
+            callback = std::move(command_queue.front());
+            command_queue.pop();
+            --num_commands;
+        }
+        callback();
+    }
+}
+
 void Liverpool::Process(std::stop_token stoken) {
     Common::SetCurrentThreadName("shadPS4:GpuCommandProcessor");
 
@@ -91,18 +105,7 @@ void Liverpool::Process(std::stop_token stoken) {
         curr_qid = -1;
 
         while (num_submits || num_commands) {
-
-            // Process incoming commands with high priority
-            while (num_commands) {
-                Common::UniqueFunction<void> callback{};
-                {
-                    std::unique_lock lk{submit_mutex};
-                    callback = std::move(command_queue.front());
-                    command_queue.pop();
-                    --num_commands;
-                }
-                callback();
-            }
+            ProcessCommands();
 
             curr_qid = (curr_qid + 1) % num_mapped_queues;
 
@@ -148,6 +151,8 @@ Liverpool::Task Liverpool::ProcessCeUpdate(std::span<const u32> ccb) {
     FIBER_ENTER(ccb_task_name);
 
     while (!ccb.empty()) {
+        ProcessCommands();
+
         const auto* header = reinterpret_cast<const PM4Header*>(ccb.data());
         const u32 type = header->type;
         if (type != 3) {
@@ -228,6 +233,8 @@ Liverpool::Task Liverpool::ProcessGraphics(std::span<const u32> dcb, std::span<c
         const auto* header = reinterpret_cast<const PM4Header*>(dcb.data());
         const u32 type = header->type;
 
+        ProcessCommands();
+
         switch (type) {
         default:
             UNREACHABLE_MSG("Wrong PM4 type {}", type);
@@ -815,6 +822,8 @@ Liverpool::Task Liverpool::ProcessCompute(const u32* acb, u32 acb_dwords, u32 vq
 
     auto base_addr = reinterpret_cast<VAddr>(acb);
     while (acb_dwords > 0) {
+        ProcessCommands();
+
         auto* header = reinterpret_cast<const PM4Header*>(acb);
         u32 next_dw_off = header->type3.NumWords() + 1;
 

src/video_core/amdgpu/liverpool.h

@@ -1587,6 +1587,7 @@ private:
     template <bool is_indirect = false>
     Task ProcessCompute(const u32* acb, u32 acb_dwords, u32 vqid);
 
+    void ProcessCommands();
     void Process(std::stop_token stoken);
 
     struct GpuQueue {

src/video_core/buffer_cache/buffer_cache.cpp

@@ -140,14 +140,22 @@ void BufferCache::InvalidateMemory(VAddr device_addr, u64 size) {
 }
 
 void BufferCache::ReadMemory(VAddr device_addr, u64 size) {
-    ForEachBufferInRange(device_addr, size, [this, device_addr, size](BufferId buffer_id, Buffer& buffer) {
-        const VAddr buffer_start = buffer.CpuAddr();
-        const VAddr buffer_end = buffer_start + buffer.SizeBytes();
-        const VAddr download_start = std::max(buffer_start, device_addr);
-        const VAddr download_end = std::min(buffer_end, device_addr + size);
-        const u64 download_size = download_end - download_start;
-        DownloadBufferMemory(buffer, download_start, download_size);
+    if (!memory_tracker.IsRegionGpuModified(device_addr, size)) {
+        return;
+    }
+    std::binary_semaphore sem{0};
+    liverpool->SendCommand([this, &sem, device_addr, size] {
+        ForEachBufferInRange(device_addr, size, [this, device_addr, size](BufferId buffer_id, Buffer& buffer) {
+            const VAddr buffer_start = buffer.CpuAddr();
+            const VAddr buffer_end = buffer_start + buffer.SizeBytes();
+            const VAddr download_start = std::max(buffer_start, device_addr);
+            const VAddr download_end = std::min(buffer_end, device_addr + size);
+            const u64 download_size = download_end - download_start;
+            DownloadBufferMemory(buffer, download_start, download_size);
+        });
+        sem.release();
     });
+    sem.acquire();
 }
 
 void BufferCache::DownloadBufferMemory(const Buffer& buffer, VAddr device_addr, u64 size) {

src/video_core/renderer_vulkan/vk_rasterizer.cpp

@@ -488,7 +488,7 @@ bool Rasterizer::BindResources(const Pipeline* pipeline) {
         uses_dma |= stage->dma_types != Shader::IR::Type::Void;
     }
 
-    if (uses_dma && !fault_process_pending) {
+    if (uses_dma) {
         // We only use fault buffer for DMA right now.
         {
             // TODO: GPU might have written to memory (for example with EVENT_WRITE_EOP)