[idd] driver: fix deadlock caused by command queue completion callback

The callback runs in a random thread, we can't call directx methods
safely from it, so move reset so it's called automatically when a free
copy list is obtained.

idd/LGIdd/CD3D12CommandQueue.cpp

@@ -1,7 +1,7 @@
 #include "CD3D12CommandQueue.h"
 #include "CDebug.h"
 
-bool CD3D12CommandQueue::Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE type, const WCHAR* name)
+bool CD3D12CommandQueue::Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE type, const WCHAR* name, CallbackMode callbackMode)
 {
   HRESULT hr;
   D3D12_COMMAND_QUEUE_DESC queueDesc = {};
@@ -54,6 +54,11 @@ bool CD3D12CommandQueue::Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE ty
     return false;
   }
 
+  if (callbackMode != DISABLED)
+  {
+    ULONG flags = (callbackMode == FAST) ?
+      WT_EXECUTEINWAITTHREAD : WT_EXECUTEINPERSISTENTTHREAD;
+  
     RegisterWaitForSingleObject(
       &m_waitHandle,
       m_event.Get(),
@@ -63,7 +68,8 @@ bool CD3D12CommandQueue::Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE ty
       },
       this,
       INFINITE,
-    WT_EXECUTEINPERSISTENTTHREAD);
+      flags);
+  }
 
   m_name       = name;
   m_fenceValue = 0;
@@ -92,15 +98,18 @@ bool CD3D12CommandQueue::Execute()
   ID3D12CommandList * lists[] = { m_cmdList.Get() };
   m_queue->ExecuteCommandLists(1, lists);
 
-  m_queue->Signal(m_fence.Get(), ++m_fenceValue);
+  m_pending    = true;
+  m_needsReset = true;
+  ++m_fenceValue;
+
+  m_fence->SetEventOnCompletion(m_fenceValue, m_event.Get());
+  m_queue->Signal(m_fence.Get(), m_fenceValue);
   if (FAILED(hr))
   {
     DEBUG_ERROR_HR(hr, "Failed to set the fence signal (%ls)", m_name);
     return false;
   }
 
-  m_fence->SetEventOnCompletion(m_fenceValue, m_event.Get());
-  m_pending = true;
   return true;
 }
 
@@ -121,6 +130,9 @@ void CD3D12CommandQueue::Wait()
 
 bool CD3D12CommandQueue::Reset()
 {
+  if (!m_needsReset)
+    return true;
+
   HRESULT hr;
   
   hr = m_allocator->Reset();
@@ -137,6 +149,7 @@ bool CD3D12CommandQueue::Reset()
     return false;
   }
 
+  m_needsReset = false;
   return true;
 }
 

idd/LGIdd/CD3D12CommandQueue.h

@@ -4,7 +4,6 @@
 #include <wdf.h>
 #include <wrl.h>
 #include <d3d12.h>
-#include <functional>
 
 using namespace Microsoft::WRL;
 using namespace Microsoft::WRL::Wrappers;
@@ -25,27 +24,35 @@ class CD3D12CommandQueue
     HandleT<HANDLENullTraits> m_event;
     HANDLE m_waitHandle = INVALID_HANDLE_VALUE;
     UINT64 m_fenceValue = 0;
+    bool m_needsReset = false;
 
-    typedef std::function<void(CD3D12CommandQueue * queue, void * param1, void * param2)> CompletionFunction;
+    typedef void (*CompletionFunction)(CD3D12CommandQueue * queue, void * param1, void * param2);
     CompletionFunction   m_completionCallback = nullptr;
     void               * m_completionParams[2];
 
-    bool Reset();
     void OnCompletion()
     {
-      Reset();
-      m_pending = false;
       if (m_completionCallback)
         m_completionCallback(
           this,
           m_completionParams[0],
           m_completionParams[1]);
+      m_pending = false;
     }
 
   public:
     ~CD3D12CommandQueue() { DeInit(); }
 
-    bool Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE type, const WCHAR * name);
+    enum CallbackMode
+    {
+      DISABLED, // no callbacks
+      FAST,     // callback is expected to return almost immediately
+      NORMAL    // normal callback
+    };
+
+    bool Init(ID3D12Device3 * device, D3D12_COMMAND_LIST_TYPE type, const WCHAR * name,
+      CallbackMode callbackMode = DISABLED);
+
     void DeInit();
 
     void SetCompletionCallback(CompletionFunction fn, void * param1, void * param2)
@@ -55,6 +62,7 @@ class CD3D12CommandQueue
       m_completionParams[1] = param2;
     }
 
+    bool Reset();
     bool Execute();
 
     //void Wait();

idd/LGIdd/CD3D12Device.cpp

@@ -108,7 +108,8 @@ CD3D12Device::InitResult CD3D12Device::Init(CIVSHMEM &ivshmem, UINT64 &alignSize
   }
 
   for(int i = 0; i < ARRAYSIZE(m_copyQueue); ++i)
-    if (!m_copyQueue[i].Init(m_device.Get(), D3D12_COMMAND_LIST_TYPE_COPY, L"Copy"))
+    if (!m_copyQueue[i].Init(m_device.Get(), D3D12_COMMAND_LIST_TYPE_COPY, L"Copy",
+        m_indirectCopy ? CD3D12CommandQueue::NORMAL : CD3D12CommandQueue::FAST))
       return InitResult::FAILURE;
 
   //if (!m_computeQueue.Init(m_device.Get(), D3D12_COMMAND_LIST_TYPE_COMPUTE, L"Compute"))
@@ -176,7 +177,10 @@ CD3D12CommandQueue * CD3D12Device::GetCopyQueue()
     for (int i = 0; i < ARRAYSIZE(m_copyQueue); ++i)
     {
       auto& queue = m_copyQueue[i];
-      if (queue.IsReady())
+      if (!queue.IsReady())
+        continue;
+
+      queue.Reset();
       return &queue;
     }
     Sleep(1);

idd/LGIdd/CSwapChainProcessor.cpp

@@ -145,6 +145,21 @@ void CSwapChainProcessor::SwapChainThreadCore()
   }
 }
 
+void CSwapChainProcessor::CompletionFunction(CD3D12CommandQueue * queue, void * param1, void * param2)
+{
+  UNREFERENCED_PARAMETER(queue);
+
+  auto sc    = (CSwapChainProcessor *)param1;
+  auto fbRes = (CFrameBufferResource*)param2;
+
+  if (sc->m_dx12Device->IsIndirectCopy())
+    sc->m_devContext->WriteFrameBuffer(
+      fbRes->GetFrameIndex(),
+      fbRes->GetMap(), 0, fbRes->GetFrameSize(), true);
+  else
+    sc->m_devContext->FinalizeFrameBuffer(fbRes->GetFrameIndex());
+}
+
 bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer)
 {
   ComPtr<ID3D11Texture2D> texture;  
@@ -212,23 +227,7 @@ bool CSwapChainProcessor::SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer
     return false;
   }
 
-  copyQueue->SetCompletionCallback(
-    [](CD3D12CommandQueue * copyQueue, void * param1, void * param2)
-    {
-      UNREFERENCED_PARAMETER(copyQueue);
-
-      auto sc    = (CSwapChainProcessor  *)param1;
-      auto fbRes = (CFrameBufferResource *)param2;
-
-      if (sc->m_dx12Device->IsIndirectCopy())
-        sc->m_devContext->WriteFrameBuffer(
-          fbRes->GetFrameIndex(),
-          fbRes->GetMap(), 0, fbRes->GetFrameSize(), true);
-      else
-        sc->m_devContext->FinalizeFrameBuffer(fbRes->GetFrameIndex());
-    },
-    this,
-    fbRes);
+  copyQueue->SetCompletionCallback(&CompletionFunction, this, fbRes);
 
   srcRes->Sync(*copyQueue);
   copyQueue->GetGfxList()->CopyTextureRegion(

idd/LGIdd/CSwapChainProcessor.h

@@ -54,6 +54,8 @@ private:
   
   void SwapChainThread();
   void SwapChainThreadCore();
+
+  static void CompletionFunction(CD3D12CommandQueue * queue, void * param1, void * param2);
   bool SwapChainNewFrame(ComPtr<IDXGIResource> acquiredBuffer);
 
 public: