From b7a00e32ecfa501b48e4f6ad3f02f0abb78ef3a3 Mon Sep 17 00:00:00 2001
From: Philip Rebohle <philip.rebohle@tu-dortmund.de>
Date: Sun, 21 Jan 2018 00:49:07 +0100
Subject: [PATCH] [dxvk] Limit size of the CS command queue

Prevents memory leaks and fixes stuttering in Heaven.
---
 src/d3d11/d3d11_context.cpp     |  3 +++
 src/d3d11/d3d11_context_imm.cpp |  5 +++++
 src/dxvk/dxvk_cs.cpp            | 14 ++++++++------
 src/dxvk/dxvk_cs.h              |  4 +++-
 4 files changed, 19 insertions(+), 7 deletions(-)

diff --git a/src/d3d11/d3d11_context.cpp b/src/d3d11/d3d11_context.cpp
index 1e0e07ca..6d689da2 100644
--- a/src/d3d11/d3d11_context.cpp
+++ b/src/d3d11/d3d11_context.cpp
@@ -565,6 +565,8 @@ namespace dxvk {
       if (((size == bufferSlice.length())
        && (bufferSlice.buffer()->memFlags() & VK_MEMORY_PROPERTY_HOST_VISIBLE_BIT))) {
         auto physicalSlice = bufferSlice.buffer()->allocPhysicalSlice();
+        physicalSlice.resource()->acquire();
+        
         std::memcpy(physicalSlice.mapPtr(0), pSrcData, size);
         
         EmitCs([
@@ -572,6 +574,7 @@ namespace dxvk {
           cPhysicalSlice = std::move(physicalSlice)
         ] (DxvkContext* ctx) {
           ctx->invalidateBuffer(cDstBuffer, cPhysicalSlice);
+          cPhysicalSlice.resource()->release();
         });
       } else {
         EmitCs([
diff --git a/src/d3d11/d3d11_context_imm.cpp b/src/d3d11/d3d11_context_imm.cpp
index 943fa137..4f4617af 100644
--- a/src/d3d11/d3d11_context_imm.cpp
+++ b/src/d3d11/d3d11_context_imm.cpp
@@ -98,6 +98,8 @@ namespace dxvk {
         // it as the 'new' mapped slice. This assumes that the
         // only way to invalidate a buffer is by mapping it.
         auto physicalSlice = buffer->allocPhysicalSlice();
+        physicalSlice.resource()->acquire();
+        
         resource->GetBufferInfo()->mappedSlice = physicalSlice;
         
         EmitCs([
@@ -105,6 +107,7 @@ namespace dxvk {
           cPhysicalSlice = physicalSlice
         ] (DxvkContext* ctx) {
           ctx->invalidateBuffer(cBuffer, cPhysicalSlice);
+          cPhysicalSlice.resource()->release();
         });
       } else if (MapType != D3D11_MAP_WRITE_NO_OVERWRITE) {
         // Synchronize with CS thread so that we know whether
@@ -167,12 +170,14 @@ namespace dxvk {
       // to be preserved, copy the image's contents into the buffer.
       if (MapType == D3D11_MAP_WRITE_DISCARD) {
         physicalSlice = textureInfo->imageBuffer->allocPhysicalSlice();
+        physicalSlice.resource()->acquire();
         
         EmitCs([
           cImageBuffer   = textureInfo->imageBuffer,
           cPhysicalSlice = physicalSlice
         ] (DxvkContext* ctx) {
           ctx->invalidateBuffer(cImageBuffer, cPhysicalSlice);
+          cPhysicalSlice.resource()->release();
         });
       } else {
         const VkImageSubresourceLayers subresourceLayers = {
diff --git a/src/dxvk/dxvk_cs.cpp b/src/dxvk/dxvk_cs.cpp
index aabcaea0..b3308d95 100644
--- a/src/dxvk/dxvk_cs.cpp
+++ b/src/dxvk/dxvk_cs.cpp
@@ -44,6 +44,10 @@ namespace dxvk {
     { std::unique_lock<std::mutex> lock(m_mutex);
       m_chunks.push(std::move(chunk));
       m_chunksPending += 1;
+      
+      m_condOnSync.wait(lock, [this] {
+        return m_stopped.load() || (m_chunksPending < MaxChunksInFlight);
+      });
     }
     
     m_condOnAdd.notify_one();
@@ -78,13 +82,11 @@ namespace dxvk {
       if (chunk != nullptr) {
         chunk->executeAll(m_context.ptr());
         
-        const bool doNotify = [this] {
-          std::unique_lock<std::mutex> lock(m_mutex);
-          return --m_chunksPending == 0;
-        }();
+        { std::unique_lock<std::mutex> lock(m_mutex);
+          m_chunksPending -= 1;
+        }
         
-        if (doNotify)
-          m_condOnSync.notify_one();
+        m_condOnSync.notify_one();
       }
     }
   }
diff --git a/src/dxvk/dxvk_cs.h b/src/dxvk/dxvk_cs.h
index 7e0d36e8..f7e1b641 100644
--- a/src/dxvk/dxvk_cs.h
+++ b/src/dxvk/dxvk_cs.h
@@ -138,7 +138,9 @@ namespace dxvk {
    * commands on a DXVK context. 
    */
   class DxvkCsThread {
-    
+    // Limit the number of chunks in the queue
+    // to prevent memory leaks, stuttering etc.
+    constexpr static uint32_t MaxChunksInFlight = 128;
   public:
     
     DxvkCsThread(const Rc<DxvkContext>& context);