From f88bf33b4fea827b124eefc8f97d8e0f2d06a2d4 Mon Sep 17 00:00:00 2001
From: Ben Vanik <ben.vanik@gmail.com>
Date: Tue, 19 May 2015 22:20:49 -0700
Subject: [PATCH] Moving threads to XHostThread and making shutdown not crash.

---
 src/xenia/apu/audio_system.cc            | 40 +++++++-------
 src/xenia/apu/audio_system.h             | 14 ++---
 src/xenia/emulator.cc                    | 51 ++++++++++--------
 src/xenia/emulator.h                     |  9 ----
 src/xenia/gpu/gl4/command_processor.cc   | 23 ++++----
 src/xenia/gpu/gl4/command_processor.h    | 12 +++--
 src/xenia/gpu/gl4/gl4_gpu.cc             |  4 +-
 src/xenia/gpu/gl4/gl4_gpu.h              |  2 +-
 src/xenia/gpu/gl4/gl4_graphics_system.cc |  4 +-
 src/xenia/gpu/gl4/gl4_graphics_system.h  |  2 +-
 src/xenia/gpu/gpu.cc                     |  6 +--
 src/xenia/gpu/gpu.h                      |  4 +-
 src/xenia/gpu/graphics_system.cc         |  5 +-
 src/xenia/gpu/graphics_system.h          |  4 +-
 src/xenia/kernel/kernel_state.cc         | 67 +++++++++++++-----------
 src/xenia/kernel/kernel_state.h          | 13 ++++-
 src/xenia/kernel/objects/xmodule.cc      | 15 ++++++
 src/xenia/kernel/objects/xmodule.h       |  1 +
 src/xenia/kernel/objects/xthread.cc      | 21 ++++----
 src/xenia/kernel/objects/xuser_module.cc |  8 +--
 src/xenia/ui/win32/win32_control.cc      |  1 +
 src/xenia/ui/win32/win32_loop.cc         |  7 ++-
 22 files changed, 175 insertions(+), 138 deletions(-)

diff --git a/src/xenia/apu/audio_system.cc b/src/xenia/apu/audio_system.cc
index da1d8cab2..7b980c8dd 100644
--- a/src/xenia/apu/audio_system.cc
+++ b/src/xenia/apu/audio_system.cc
@@ -14,8 +14,8 @@
 #include "xenia/base/math.h"
 #include "xenia/cpu/processor.h"
 #include "xenia/cpu/thread_state.h"
-#include "xenia/kernel/objects/xthread.h"
 #include "xenia/emulator.h"
+#include "xenia/kernel/objects/xthread.h"
 #include "xenia/profiling.h"
 
 // As with normal Microsoft, there are like twelve different ways to access
@@ -48,7 +48,6 @@ namespace xe {
 namespace apu {
 
 using namespace xe::cpu;
-using namespace xe::kernel;
 
 // Size of a hardware XMA context.
 const uint32_t kXmaContextSize = 64;
@@ -56,7 +55,7 @@ const uint32_t kXmaContextSize = 64;
 const uint32_t kXmaContextCount = 320;
 
 AudioSystem::AudioSystem(Emulator* emulator)
-    : emulator_(emulator), memory_(emulator->memory()), running_(false) {
+    : emulator_(emulator), memory_(emulator->memory()), worker_running_(false) {
   memset(clients_, 0, sizeof(clients_));
   for (size_t i = 0; i < maximum_client_count_; ++i) {
     unused_clients_.push(i);
@@ -91,22 +90,18 @@ X_STATUS AudioSystem::Setup() {
   }
   registers_.next_context = 1;
 
-  // Setup our worker thread
-  std::function<int()> thread_fn = [this]() {
-    this->ThreadStart();
+  worker_running_ = true;
+  worker_thread_ = new kernel::XHostThread(emulator()->kernel_state(),
+                                           128 * 1024, 0, [this]() {
+    this->WorkerThreadMain();
     return 0;
-  };
-
-  running_ = true;
-
-  thread_ = std::make_unique<XHostThread>(emulator()->kernel_state(), 
-                                        128 * 1024, 0, thread_fn);
-  thread_->Create();
+  });
+  worker_thread_->Create();
 
   return X_STATUS_SUCCESS;
 }
 
-void AudioSystem::ThreadStart() {
+void AudioSystem::WorkerThreadMain() {
   xe::threading::set_name("Audio Worker");
 
   // Initialize driver and ringbuffer.
@@ -115,7 +110,7 @@ void AudioSystem::ThreadStart() {
   auto processor = emulator_->processor();
 
   // Main run loop.
-  while (running_) {
+  while (worker_running_) {
     auto result =
         WaitForMultipleObjectsEx(DWORD(xe::countof(client_wait_handles_)),
                                  client_wait_handles_, FALSE, INFINITE, FALSE);
@@ -135,8 +130,8 @@ void AudioSystem::ThreadStart() {
         lock_.unlock();
         if (client_callback) {
           uint64_t args[] = {client_callback_arg};
-          processor->Execute(thread_->thread_state(), client_callback, args,
-                             xe::countof(args));
+          processor->Execute(worker_thread_->thread_state(), client_callback,
+                             args, xe::countof(args));
         }
         pumped++;
         index++;
@@ -145,7 +140,7 @@ void AudioSystem::ThreadStart() {
                    WAIT_OBJECT_0);
     }
 
-    if (!running_) {
+    if (!worker_running_) {
       break;
     }
 
@@ -154,7 +149,7 @@ void AudioSystem::ThreadStart() {
       Sleep(500);
     }
   }
-  running_ = false;
+  worker_running_ = false;
 
   // TODO(benvanik): call module API to kill?
 }
@@ -162,9 +157,10 @@ void AudioSystem::ThreadStart() {
 void AudioSystem::Initialize() {}
 
 void AudioSystem::Shutdown() {
-  running_ = false;
-  ResetEvent(client_wait_handles_[maximum_client_count_]);
-  thread_->Wait(0, 0, 0, NULL);
+  worker_running_ = false;
+  SetEvent(client_wait_handles_[maximum_client_count_]);
+  worker_thread_->Wait(0, 0, 0, nullptr);
+  worker_thread_->Release();
 
   memory()->SystemHeapFree(registers_.xma_context_array_ptr);
 }
diff --git a/src/xenia/apu/audio_system.h b/src/xenia/apu/audio_system.h
index fc2f61381..125402c1f 100644
--- a/src/xenia/apu/audio_system.h
+++ b/src/xenia/apu/audio_system.h
@@ -13,15 +13,17 @@
 #include <atomic>
 #include <mutex>
 #include <queue>
-#include <thread>
 
 #include "xenia/emulator.h"
 #include "xenia/xbox.h"
 
 namespace xe {
+namespace kernel {
+class XHostThread;
+}  // namespace kernel
+}  // namespace xe
 
-namespace kernel { class XHostThread; }
-
+namespace xe {
 namespace apu {
 
 class AudioDriver;
@@ -59,7 +61,7 @@ class AudioSystem {
   virtual void Initialize();
 
  private:
-  void ThreadStart();
+  void WorkerThreadMain();
 
   static uint64_t MMIOReadRegisterThunk(AudioSystem* as, uint32_t addr) {
     return as->ReadRegister(addr);
@@ -76,8 +78,8 @@ class AudioSystem {
   Memory* memory_;
   cpu::Processor* processor_;
 
-  std::unique_ptr<kernel::XHostThread> thread_;
-  std::atomic<bool> running_;
+  std::atomic<bool> worker_running_;
+  kernel::XHostThread* worker_thread_;
 
   std::mutex lock_;
 
diff --git a/src/xenia/emulator.cc b/src/xenia/emulator.cc
index a88760980..c2d2df23b 100644
--- a/src/xenia/emulator.cc
+++ b/src/xenia/emulator.cc
@@ -28,7 +28,6 @@ using namespace xe::apu;
 using namespace xe::cpu;
 using namespace xe::gpu;
 using namespace xe::hid;
-using namespace xe::kernel;
 using namespace xe::kernel::fs;
 using namespace xe::ui;
 
@@ -38,17 +37,14 @@ Emulator::Emulator(const std::wstring& command_line)
 Emulator::~Emulator() {
   // Note that we delete things in the reverse order they were initialized.
 
-  xam_.reset();
-  xboxkrnl_.reset();
-  kernel_state_.reset();
-
-  file_system_.reset();
-
-  input_system_.reset();
-
   // Give the systems time to shutdown before we delete them.
   graphics_system_->Shutdown();
   audio_system_->Shutdown();
+
+  kernel_state_.reset();
+  file_system_.reset();
+
+  input_system_.reset();
   graphics_system_.reset();
   audio_system_.reset();
 
@@ -95,7 +91,7 @@ X_STATUS Emulator::Setup() {
   }
 
   // Initialize the GPU.
-  graphics_system_ = std::move(xe::gpu::Create());
+  graphics_system_ = std::move(xe::gpu::Create(this));
   if (!graphics_system_) {
     return X_STATUS_NOT_IMPLEMENTED;
   }
@@ -106,15 +102,6 @@ X_STATUS Emulator::Setup() {
     return X_STATUS_NOT_IMPLEMENTED;
   }
 
-  // Setup the core components.
-  if (!processor_->Setup()) {
-    return result;
-  }
-  result = graphics_system_->Setup(processor_.get(), main_window_->loop(),
-                                   main_window_.get());
-  if (result) {
-    return result;
-  }
   result = input_system_->Setup();
   if (result) {
     return result;
@@ -124,7 +111,17 @@ X_STATUS Emulator::Setup() {
   file_system_ = std::make_unique<FileSystem>();
 
   // Shared kernel state.
-  kernel_state_ = std::make_unique<KernelState>(this);
+  kernel_state_ = std::make_unique<kernel::KernelState>(this);
+
+  // Setup the core components.
+  if (!processor_->Setup()) {
+    return result;
+  }
+  result = graphics_system_->Setup(processor_.get(), main_window_->loop(),
+                                   main_window_.get());
+  if (result) {
+    return result;
+  }
 
   result = audio_system_->Setup();
   if (result) {
@@ -132,8 +129,8 @@ X_STATUS Emulator::Setup() {
   }
 
   // HLE kernel modules.
-  xboxkrnl_ = std::make_unique<XboxkrnlModule>(this, kernel_state_.get());
-  xam_ = std::make_unique<XamModule>(this, kernel_state_.get());
+  kernel_state_->LoadKernelModule<kernel::XboxkrnlModule>();
+  kernel_state_->LoadKernelModule<kernel::XamModule>();
 
   return result;
 }
@@ -192,7 +189,15 @@ X_STATUS Emulator::LaunchSTFSTitle(const std::wstring& path) {
 
 X_STATUS Emulator::CompleteLaunch(const std::wstring& path,
                                   const std::string& module_path) {
-  return xboxkrnl_->LaunchModule(module_path.c_str());
+  auto xboxkrnl = static_cast<kernel::XboxkrnlModule*>(
+      kernel_state_->GetModule("xboxkrnl.exe"));
+  int result = xboxkrnl->LaunchModule(module_path.c_str());
+  xboxkrnl->Release();
+  if (result == 0) {
+    return X_STATUS_SUCCESS;
+  } else {
+    return X_STATUS_UNSUCCESSFUL;
+  }
 }
 
 }  // namespace xe
diff --git a/src/xenia/emulator.h b/src/xenia/emulator.h
index 6d389a7f5..ab26d3db2 100644
--- a/src/xenia/emulator.h
+++ b/src/xenia/emulator.h
@@ -32,10 +32,6 @@ class GraphicsSystem;
 namespace hid {
 class InputSystem;
 }  // namespace hid
-namespace kernel {
-class XamModule;
-class XboxkrnlModule;
-}  // namespace kernel
 namespace ui {
 class MainWindow;
 }  // namespace ui
@@ -68,9 +64,6 @@ class Emulator {
 
   kernel::KernelState* kernel_state() const { return kernel_state_.get(); }
 
-  kernel::XboxkrnlModule* xboxkrnl() const { return xboxkrnl_.get(); }
-  kernel::XamModule* xam() const { return xam_.get(); }
-
   X_STATUS Setup();
 
   // TODO(benvanik): raw binary.
@@ -97,8 +90,6 @@ class Emulator {
   std::unique_ptr<kernel::fs::FileSystem> file_system_;
 
   std::unique_ptr<kernel::KernelState> kernel_state_;
-  std::unique_ptr<kernel::XamModule> xam_;
-  std::unique_ptr<kernel::XboxkrnlModule> xboxkrnl_;
 };
 
 }  // namespace xe
diff --git a/src/xenia/gpu/gl4/command_processor.cc b/src/xenia/gpu/gl4/command_processor.cc
index 7308b35e9..4ac0206b4 100644
--- a/src/xenia/gpu/gl4/command_processor.cc
+++ b/src/xenia/gpu/gl4/command_processor.cc
@@ -19,6 +19,8 @@
 #include "xenia/gpu/sampler_info.h"
 #include "xenia/gpu/texture_info.h"
 #include "xenia/gpu/xenos.h"
+#include "xenia/emulator.h"
+#include "xenia/kernel/objects/xthread.h"
 #include "xenia/profiling.h"
 
 #include "third_party/xxhash/xxhash.h"
@@ -99,12 +101,12 @@ bool CommandProcessor::Initialize(std::unique_ptr<GLContext> context) {
   context_ = std::move(context);
 
   worker_running_ = true;
-  worker_thread_ = std::thread([this]() {
-    xe::threading::set_name("GL4 Worker");
-    xe::Profiler::ThreadEnter("GL4 Worker");
-    WorkerMain();
-    xe::Profiler::ThreadExit();
-  });
+  worker_thread_ = new kernel::XHostThread(
+      graphics_system_->emulator()->kernel_state(), 128 * 1024, 0, [this]() {
+        WorkerThreadMain();
+        return 0;
+      });
+  worker_thread_->Create();
 
   return true;
 }
@@ -114,7 +116,8 @@ void CommandProcessor::Shutdown() {
 
   worker_running_ = false;
   SetEvent(write_ptr_index_event_);
-  worker_thread_.join();
+  worker_thread_->Wait(0, 0, 0, nullptr);
+  worker_thread_->Release();
 
   all_pipelines_.clear();
   all_shaders_.clear();
@@ -160,14 +163,16 @@ void CommandProcessor::EndTracing() {
 
 void CommandProcessor::CallInThread(std::function<void()> fn) {
   if (pending_fns_.empty() &&
-      worker_thread_.get_id() == std::this_thread::get_id()) {
+      worker_thread_ == kernel::XThread::GetCurrentThread()) {
     fn();
   } else {
     pending_fns_.push(std::move(fn));
   }
 }
 
-void CommandProcessor::WorkerMain() {
+void CommandProcessor::WorkerThreadMain() {
+  xe::threading::set_name("GL4 Worker");
+
   context_->MakeCurrent();
   if (!SetupGL()) {
     XEFATAL("Unable to setup command processor GL state");
diff --git a/src/xenia/gpu/gl4/command_processor.h b/src/xenia/gpu/gl4/command_processor.h
index a919401ba..583fc317b 100644
--- a/src/xenia/gpu/gl4/command_processor.h
+++ b/src/xenia/gpu/gl4/command_processor.h
@@ -14,7 +14,6 @@
 #include <functional>
 #include <memory>
 #include <queue>
-#include <thread>
 #include <unordered_map>
 #include <vector>
 
@@ -28,6 +27,12 @@
 #include "xenia/gpu/xenos.h"
 #include "xenia/memory.h"
 
+namespace xe {
+namespace kernel {
+class XHostThread;
+}  // namespace kernel
+}  // namespace xe
+
 namespace xe {
 namespace gpu {
 namespace gl4 {
@@ -134,7 +139,7 @@ class CommandProcessor {
     } handles;
   };
 
-  void WorkerMain();
+  void WorkerThreadMain();
   bool SetupGL();
   void ShutdownGL();
   GLuint CreateGeometryProgram(const std::string& source);
@@ -226,8 +231,9 @@ class CommandProcessor {
   TraceState trace_state_;
   std::wstring trace_frame_path_;
 
-  std::thread worker_thread_;
   std::atomic<bool> worker_running_;
+  kernel::XHostThread* worker_thread_;
+
   std::unique_ptr<GLContext> context_;
   SwapHandler swap_handler_;
   std::queue<std::function<void()>> pending_fns_;
diff --git a/src/xenia/gpu/gl4/gl4_gpu.cc b/src/xenia/gpu/gl4/gl4_gpu.cc
index c21ba0bed..955a9e17d 100644
--- a/src/xenia/gpu/gl4/gl4_gpu.cc
+++ b/src/xenia/gpu/gl4/gl4_gpu.cc
@@ -47,9 +47,9 @@ void InitializeIfNeeded() {
 
 void CleanupOnShutdown() {}
 
-std::unique_ptr<GraphicsSystem> Create() {
+std::unique_ptr<GraphicsSystem> Create(Emulator* emulator) {
   InitializeIfNeeded();
-  return std::make_unique<GL4GraphicsSystem>();
+  return std::make_unique<GL4GraphicsSystem>(emulator);
 }
 
 }  // namespace gl4
diff --git a/src/xenia/gpu/gl4/gl4_gpu.h b/src/xenia/gpu/gl4/gl4_gpu.h
index 761cc3b2d..a36429a21 100644
--- a/src/xenia/gpu/gl4/gl4_gpu.h
+++ b/src/xenia/gpu/gl4/gl4_gpu.h
@@ -18,7 +18,7 @@ namespace xe {
 namespace gpu {
 namespace gl4 {
 
-std::unique_ptr<GraphicsSystem> Create();
+std::unique_ptr<GraphicsSystem> Create(Emulator* emulator);
 
 }  // namespace gl4
 }  // namespace gpu
diff --git a/src/xenia/gpu/gl4/gl4_graphics_system.cc b/src/xenia/gpu/gl4/gl4_graphics_system.cc
index bd2ee1f87..f29d186d8 100644
--- a/src/xenia/gpu/gl4/gl4_graphics_system.cc
+++ b/src/xenia/gpu/gl4/gl4_graphics_system.cc
@@ -23,8 +23,8 @@ namespace gl4 {
 
 extern "C" GLEWContext* glewGetContext();
 
-GL4GraphicsSystem::GL4GraphicsSystem()
-    : GraphicsSystem(), timer_queue_(nullptr), vsync_timer_(nullptr) {}
+GL4GraphicsSystem::GL4GraphicsSystem(Emulator* emulator)
+    : GraphicsSystem(emulator), timer_queue_(nullptr), vsync_timer_(nullptr) {}
 
 GL4GraphicsSystem::~GL4GraphicsSystem() = default;
 
diff --git a/src/xenia/gpu/gl4/gl4_graphics_system.h b/src/xenia/gpu/gl4/gl4_graphics_system.h
index 558af2c1c..6eeec1310 100644
--- a/src/xenia/gpu/gl4/gl4_graphics_system.h
+++ b/src/xenia/gpu/gl4/gl4_graphics_system.h
@@ -23,7 +23,7 @@ namespace gl4 {
 
 class GL4GraphicsSystem : public GraphicsSystem {
  public:
-  GL4GraphicsSystem();
+  GL4GraphicsSystem(Emulator* emulator);
   ~GL4GraphicsSystem() override;
 
   X_STATUS Setup(cpu::Processor* processor, ui::PlatformLoop* target_loop,
diff --git a/src/xenia/gpu/gpu.cc b/src/xenia/gpu/gpu.cc
index 867a302a9..b2ebd0a11 100644
--- a/src/xenia/gpu/gpu.cc
+++ b/src/xenia/gpu/gpu.cc
@@ -27,14 +27,14 @@ DEFINE_bool(vsync, true, "Enable VSYNC.");
 namespace xe {
 namespace gpu {
 
-std::unique_ptr<GraphicsSystem> Create() {
+std::unique_ptr<GraphicsSystem> Create(Emulator* emulator) {
   if (FLAGS_gpu.compare("gl4") == 0) {
-    return xe::gpu::gl4::Create();
+    return xe::gpu::gl4::Create(emulator);
   } else {
     // Create best available.
     std::unique_ptr<GraphicsSystem> best;
 
-    best = xe::gpu::gl4::Create();
+    best = xe::gpu::gl4::Create(emulator);
     if (best) {
       return best;
     }
diff --git a/src/xenia/gpu/gpu.h b/src/xenia/gpu/gpu.h
index 912e05406..b9965229d 100644
--- a/src/xenia/gpu/gpu.h
+++ b/src/xenia/gpu/gpu.h
@@ -21,9 +21,7 @@ class Emulator;
 namespace xe {
 namespace gpu {
 
-std::unique_ptr<GraphicsSystem> Create();
-
-std::unique_ptr<GraphicsSystem> CreateGL4();
+std::unique_ptr<GraphicsSystem> Create(Emulator* emulator);
 
 }  // namespace gpu
 }  // namespace xe
diff --git a/src/xenia/gpu/graphics_system.cc b/src/xenia/gpu/graphics_system.cc
index 22600cc5d..09d21838d 100644
--- a/src/xenia/gpu/graphics_system.cc
+++ b/src/xenia/gpu/graphics_system.cc
@@ -17,8 +17,9 @@
 namespace xe {
 namespace gpu {
 
-GraphicsSystem::GraphicsSystem()
-    : memory_(nullptr),
+GraphicsSystem::GraphicsSystem(Emulator* emulator)
+    : emulator_(emulator),
+      memory_(nullptr),
       processor_(nullptr),
       target_loop_(nullptr),
       target_window_(nullptr),
diff --git a/src/xenia/gpu/graphics_system.h b/src/xenia/gpu/graphics_system.h
index 8d0a74b82..3fc7a4424 100644
--- a/src/xenia/gpu/graphics_system.h
+++ b/src/xenia/gpu/graphics_system.h
@@ -25,6 +25,7 @@ class GraphicsSystem {
  public:
   virtual ~GraphicsSystem();
 
+  Emulator* emulator() const { return emulator_; }
   Memory* memory() const { return memory_; }
   cpu::Processor* processor() const { return processor_; }
 
@@ -54,8 +55,9 @@ class GraphicsSystem {
   virtual void ClearCaches() {}
 
  protected:
-  GraphicsSystem();
+  GraphicsSystem(Emulator* emulator);
 
+  Emulator* emulator_;
   Memory* memory_;
   cpu::Processor* processor_;
   ui::PlatformLoop* target_loop_;
diff --git a/src/xenia/kernel/kernel_state.cc b/src/xenia/kernel/kernel_state.cc
index 4d64d7f2e..c6033ba7d 100644
--- a/src/xenia/kernel/kernel_state.cc
+++ b/src/xenia/kernel/kernel_state.cc
@@ -70,6 +70,15 @@ KernelState::KernelState(Emulator* emulator)
 KernelState::~KernelState() {
   SetExecutableModule(nullptr);
 
+  for (auto user_module : user_modules_) {
+    user_module->Release();
+  }
+  user_modules_.clear();
+  for (auto kernel_module : kernel_modules_) {
+    kernel_module->Release();
+  }
+  kernel_modules_.clear();
+
   // Delete all objects.
   delete object_table_;
 
@@ -80,10 +89,6 @@ KernelState::~KernelState() {
 
   assert_true(shared_kernel_state_ == this);
   shared_kernel_state_ = nullptr;
-
-  for (XUserModule* mod : user_modules_) {
-    mod->Release();
-  }
 }
 
 KernelState* KernelState::shared() { return shared_kernel_state_; }
@@ -99,14 +104,15 @@ void KernelState::UnregisterModule(XModule* module) {}
 
 bool KernelState::IsKernelModule(const char* name) {
   if (!name) {
-    // executing module isn't a kernel module
+    // Executing module isn't a kernel module.
     return false;
-  } else if (strcasecmp(name, "xam.xex") == 0) {
-    return true;
-  } else if (strcasecmp(name, "xboxkrnl.exe") == 0) {
-    return true;
   }
-
+  std::lock_guard<std::recursive_mutex> lock(object_mutex_);
+  for (auto kernel_module : kernel_modules_) {
+    if (kernel_module->Matches(name)) {
+      return true;
+    }
+  }
   return false;
 }
 
@@ -115,32 +121,24 @@ XModule* KernelState::GetModule(const char* name) {
     // NULL name = self.
     // TODO(benvanik): lookup module from caller address.
     return GetExecutableModule();
-  } else if (strcasecmp(name, "xam.xex") == 0) {
-    auto module = emulator_->xam();
-    module->Retain();
-    return module;
-  } else if (strcasecmp(name, "xboxkrnl.exe") == 0) {
-    auto module = emulator_->xboxkrnl();
-    module->Retain();
-    return module;
   } else if (strcasecmp(name, "kernel32.dll") == 0) {
     // Some games request this, for some reason. wtf.
-    return nullptr;
-  } else {
-    std::lock_guard<std::recursive_mutex> lock(object_mutex_);
-
-    for (XUserModule* module : user_modules_) {
-      if ((strcasecmp(xe::find_name_from_path(module->path()).c_str(), name) ==
-           0) ||
-          (strcasecmp(module->name().c_str(), name) == 0) ||
-          (strcasecmp(module->path().c_str(), name) == 0)) {
-        module->Retain();
-        return module;
-      }
-    }
-
     return nullptr;
   }
+  std::lock_guard<std::recursive_mutex> lock(object_mutex_);
+  for (auto kernel_module : kernel_modules_) {
+    if (kernel_module->Matches(name)) {
+      kernel_module->Retain();
+      return kernel_module;
+    }
+  }
+  for (auto user_module : user_modules_) {
+    if (user_module->Matches(name)) {
+      user_module->Retain();
+      return user_module;
+    }
+  }
+  return nullptr;
 }
 
 XUserModule* KernelState::GetExecutableModule() {
@@ -166,6 +164,11 @@ void KernelState::SetExecutableModule(XUserModule* module) {
   }
 }
 
+void KernelState::LoadKernelModule(XKernelModule* kernel_module) {
+  std::lock_guard<std::recursive_mutex> lock(object_mutex_);
+  kernel_modules_.push_back(kernel_module);
+}
+
 XUserModule* KernelState::LoadUserModule(const char* raw_name) {
   // Some games try to load relative to launch module, others specify full path.
   std::string name = xe::find_name_from_path(raw_name);
diff --git a/src/xenia/kernel/kernel_state.h b/src/xenia/kernel/kernel_state.h
index f70f123e6..eab953d2c 100644
--- a/src/xenia/kernel/kernel_state.h
+++ b/src/xenia/kernel/kernel_state.h
@@ -37,6 +37,7 @@ namespace xe {
 namespace kernel {
 
 class Dispatcher;
+class XKernelModule;
 class XModule;
 class XNotifyListener;
 class XThread;
@@ -74,7 +75,13 @@ class KernelState {
   XModule* GetModule(const char* name);
   XUserModule* GetExecutableModule();
   void SetExecutableModule(XUserModule* module);
-  XUserModule* LoadUserModule(const char *name);
+  template <typename T>
+  XKernelModule* LoadKernelModule() {
+    auto kernel_module = std::make_unique<T>(emulator_, this);
+    LoadKernelModule(kernel_module.get());
+    return kernel_module.release();
+  }
+  XUserModule* LoadUserModule(const char* name);
 
   void RegisterThread(XThread* thread);
   void UnregisterThread(XThread* thread);
@@ -94,6 +101,8 @@ class KernelState {
                                      uint32_t extended_error, uint32_t length);
 
  private:
+  void LoadKernelModule(XKernelModule* kernel_module);
+
   Emulator* emulator_;
   Memory* memory_;
   cpu::Processor* processor_;
@@ -113,7 +122,7 @@ class KernelState {
 
   uint32_t process_type_;
   XUserModule* executable_module_;
-
+  std::vector<XKernelModule*> kernel_modules_;
   std::vector<XUserModule*> user_modules_;
 
   friend class XObject;
diff --git a/src/xenia/kernel/objects/xmodule.cc b/src/xenia/kernel/objects/xmodule.cc
index 192ae6525..fc1e4247d 100644
--- a/src/xenia/kernel/objects/xmodule.cc
+++ b/src/xenia/kernel/objects/xmodule.cc
@@ -9,6 +9,8 @@
 
 #include "xenia/kernel/objects/xmodule.h"
 
+#include "xenia/base/string.h"
+
 namespace xe {
 namespace kernel {
 
@@ -31,6 +33,19 @@ XModule::XModule(KernelState* kernel_state, const std::string& path)
 
 XModule::~XModule() { kernel_state_->UnregisterModule(this); }
 
+bool XModule::Matches(const std::string& name) const {
+  if (strcasecmp(xe::find_name_from_path(path_).c_str(), name.c_str()) == 0) {
+    return true;
+  }
+  if (strcasecmp(name_.c_str(), name.c_str()) == 0) {
+    return true;
+  }
+  if (strcasecmp(path_.c_str(), name.c_str()) == 0) {
+    return true;
+  }
+  return false;
+}
+
 void XModule::OnLoad() { kernel_state_->RegisterModule(this); }
 
 X_STATUS XModule::GetSection(const char* name, uint32_t* out_section_data,
diff --git a/src/xenia/kernel/objects/xmodule.h b/src/xenia/kernel/objects/xmodule.h
index 51218e8ac..123ace447 100644
--- a/src/xenia/kernel/objects/xmodule.h
+++ b/src/xenia/kernel/objects/xmodule.h
@@ -25,6 +25,7 @@ class XModule : public XObject {
 
   const std::string& path() const { return path_; }
   const std::string& name() const { return name_; }
+  bool Matches(const std::string& name) const;
 
   virtual uint32_t GetProcAddressByOrdinal(uint16_t ordinal) = 0;
   virtual uint32_t GetProcAddressByName(const char* name) = 0;
diff --git a/src/xenia/kernel/objects/xthread.cc b/src/xenia/kernel/objects/xthread.cc
index 623787f04..11d6d74e9 100644
--- a/src/xenia/kernel/objects/xthread.cc
+++ b/src/xenia/kernel/objects/xthread.cc
@@ -156,7 +156,7 @@ X_STATUS XThread::Create() {
   scratch_address_ = memory()->SystemHeapAlloc(scratch_size_);
 
   // Allocate TLS block.
-  uint32_t tls_size = 32; // Default 32 (is this OK?)
+  uint32_t tls_size = 32;  // Default 32 (is this OK?)
   if (module && module->xex_header()) {
     const xe_xex2_header_t* header = module->xex_header();
     tls_size = header->tls_info.slot_count * header->tls_info.data_size;
@@ -194,15 +194,15 @@ X_STATUS XThread::Create() {
          thread_state_->stack_base());
 
   uint8_t* pcr = memory()->TranslateVirtual(pcr_address_);
-  std::memset(pcr, 0x0, 0x2D8 + 0xAB0); // Zero the PCR
+  std::memset(pcr, 0x0, 0x2D8 + 0xAB0);  // Zero the PCR
   xe::store_and_swap<uint32_t>(pcr + 0x000, tls_address_);
   xe::store_and_swap<uint32_t>(pcr + 0x030, pcr_address_);
   xe::store_and_swap<uint32_t>(pcr + 0x070, thread_state_->stack_address() +
                                                 thread_state_->stack_size());
   xe::store_and_swap<uint32_t>(pcr + 0x074, thread_state_->stack_address());
   xe::store_and_swap<uint32_t>(pcr + 0x100, thread_state_address_);
-  xe::store_and_swap<uint8_t> (pcr + 0x10C, 1); // Current CPU(?)
-  xe::store_and_swap<uint32_t>(pcr + 0x150, 0); // DPC active bool?
+  xe::store_and_swap<uint8_t>(pcr + 0x10C, 1);   // Current CPU(?)
+  xe::store_and_swap<uint32_t>(pcr + 0x150, 0);  // DPC active bool?
 
   // Setup the thread state block (last error/etc).
   uint8_t* p = memory()->TranslateVirtual(thread_state_address_);
@@ -622,15 +622,14 @@ X_STATUS XThread::Delay(uint32_t processor_mode, uint32_t alertable,
 void* XThread::GetWaitHandle() { return event_->GetWaitHandle(); }
 
 XHostThread::XHostThread(KernelState* kernel_state, uint32_t stack_size,
-              uint32_t creation_flags, std::function<int()> host_fn):
-              XThread(kernel_state, stack_size, 0, 0, 0, creation_flags),
-              host_fn_(host_fn) {
-}
+                         uint32_t creation_flags, std::function<int()> host_fn)
+    : XThread(kernel_state, stack_size, 0, 0, 0, creation_flags),
+      host_fn_(host_fn) {}
 
 void XHostThread::Execute() {
-  XELOGKERNEL("XThread::Execute thid %d (handle=%.8X, '%s', native=%.8X, <host>)",
-              thread_id_, handle(), name_.c_str(),
-              xe::threading::current_thread_id());
+  XELOGKERNEL(
+      "XThread::Execute thid %d (handle=%.8X, '%s', native=%.8X, <host>)",
+      thread_id_, handle(), name_.c_str(), xe::threading::current_thread_id());
 
   // Let the kernel know we are starting.
   kernel_state()->OnThreadExecute(this);
diff --git a/src/xenia/kernel/objects/xuser_module.cc b/src/xenia/kernel/objects/xuser_module.cc
index 230627477..c05cd9435 100644
--- a/src/xenia/kernel/objects/xuser_module.cc
+++ b/src/xenia/kernel/objects/xuser_module.cc
@@ -354,7 +354,7 @@ void XUserModule::Dump() {
 
         if (kernel_state_->IsKernelModule(library->name)) {
           KernelExport* kernel_export =
-            export_resolver->GetExportByOrdinal(library->name, info->ordinal);
+              export_resolver->GetExportByOrdinal(library->name, info->ordinal);
           if (kernel_export) {
             known_count++;
             if (kernel_export->is_implemented) {
@@ -371,7 +371,7 @@ void XUserModule::Dump() {
           XModule* module = kernel_state_->GetModule(library->name);
           if (module) {
             uint32_t export_addr =
-                      module->GetProcAddressByOrdinal(info->ordinal);
+                module->GetProcAddressByOrdinal(info->ordinal);
             if (export_addr) {
               impl_count++;
               known_count++;
@@ -400,10 +400,10 @@ void XUserModule::Dump() {
         const char* name = "UNKNOWN";
         bool implemented = false;
 
-        KernelExport* kernel_export;
+        KernelExport* kernel_export = nullptr;
         if (kernel_state_->IsKernelModule(library->name)) {
           kernel_export =
-            export_resolver->GetExportByOrdinal(library->name, info->ordinal);
+              export_resolver->GetExportByOrdinal(library->name, info->ordinal);
           if (kernel_export) {
             name = kernel_export->name;
             implemented = kernel_export->is_implemented;
diff --git a/src/xenia/ui/win32/win32_control.cc b/src/xenia/ui/win32/win32_control.cc
index 7ba359edf..e67644fc2 100644
--- a/src/xenia/ui/win32/win32_control.cc
+++ b/src/xenia/ui/win32/win32_control.cc
@@ -17,6 +17,7 @@ Win32Control::Win32Control(uint32_t flags) : Control(flags), hwnd_(nullptr) {}
 
 Win32Control::~Win32Control() {
   if (hwnd_) {
+    SetWindowLongPtr(hwnd_, GWLP_USERDATA, 0);
     CloseWindow(hwnd_);
     hwnd_ = nullptr;
   }
diff --git a/src/xenia/ui/win32/win32_loop.cc b/src/xenia/ui/win32/win32_loop.cc
index f85221db8..38f325500 100644
--- a/src/xenia/ui/win32/win32_loop.cc
+++ b/src/xenia/ui/win32/win32_loop.cc
@@ -29,7 +29,7 @@ class PostedFn {
 
 Win32Loop::Win32Loop() : thread_id_(0) {
   xe::threading::Fence init_fence;
-  thread_ = std::thread([&]() {
+  thread_ = std::thread([&init_fence, this]() {
     xe::threading::set_name("Win32 Loop");
     thread_id_ = GetCurrentThreadId();
 
@@ -46,7 +46,10 @@ Win32Loop::Win32Loop() : thread_id_(0) {
   init_fence.Wait();
 }
 
-Win32Loop::~Win32Loop() = default;
+Win32Loop::~Win32Loop() {
+  Quit();
+  thread_.join();
+}
 
 void Win32Loop::ThreadMain() {
   MSG msg;