Fix l0 teardown and sycl windows teardown

sycl/doc/design/GlobalObjectsInRuntime.md

@@ -55,9 +55,65 @@ Deinitialization is platform-specific. Upon application shutdown, the DPC++
 runtime frees memory pointed by `GlobalHandler` global pointer, which triggers
 destruction of nested `std::unique_ptr`s.
 
+### Shutdown Tasks and Challenges
+
+As the user's app ends, SYCL's primary goal is to release any UR adapters that 
+have been gotten, and teardown the plugins/adapters themselves.  Additionally, 
+we need to stop deferring any new buffer releases and clean up any memory 
+whose release was deferred. 
+
+To this end, the shutdown occurs in two phases: early and late. The purpose
+for early shutdown is primarily to stop any further deferring of memory release.
+This is because the deferred memory release is based on threads and on Windows
+the threads will be abandoned. So as soon as possible we want to stop deferring
+memory and try to let go any that has been deferred.  The purpose for late 
+shutdown is to hold onto the handles and adapters longer than the user's
+application. We don't want to initiate late shutdown until after all the users 
+static and thread local vars have been destroyed, in case those destructors are 
+calling SYCL. 
+
+In the early shutdown we stop deferring, tell the scheduler to prepare for release, and
+try releasing the memory that has been deferred so far.  Following this, if 
+the user has any global or static handles to sycl objects, they'll be destroyed.
+Finally, the late shutdown routine is called the last of the UR handles and 
+adapters are let go, as is the GlobalHandler itself.
+
+
+#### Threads
+The deferred memory marshalling is built on a thread pool, but there is a
+challenge here in that on Windows, once the end of the users main() is reached
+and their app is shutting down, the Windows OS will abandon all remaining 
+in-flight threads. These threads can be .join() but they simply return instantly,
+the threads are not completed. Further any thread specific variables
+(or thread_local static vars) will NOT have their destructors called.  Note
+that the standard while-loop-over-condition-var pattern will cause a hang - 
+we cannot "wait" on abandoned threads. 
+On Windows, short of adding some user called API to signal this, there is 
+no way to detect or avoid this. None of the "end-of-library" lifecycle events
+occurs before the threads are abandoned.  ( not std::atexit(), not globals or 
+static, or static thread_local var destruction, not DllMain(DLL_PROCESS_DETACH) )
+This means that on Windows, once we arrive at shutdown_early we cannot wait on
+host events or the thread pool. 
+
+For the deferred memory itself, there is no issue here. The Windows OS will
+reclaim the memory for us. The issue of which we must be wary is placing UR 
+handles (and similar) in host threads. The RAII mechanism of unique and 
+shared pointers will not work in any thread that is abandoned on Windows. 
+
+One last note about threads. It is entirely the OS's discretion when to
+start or schedule a thread. If the main process is very busy then it is 
+possible that threads the SYCL library creates (host_tasks/thread_pool)
+won't even be started until AFTER the host application main() function is done. 
+This is not a normal occurrence, but it can happen if there is no call to queue.wait()
+
+
 ### Linux
 
-On Linux DPC++ runtime uses `__attribute__((destructor))` property with low
+On Linux, the "early_shutdown()" is begun by the destruction of a static
+StaticVarShutdownHandler object, which is initialized by 
+platform::get_platforms().
+
+late_shutdown() timing uses `__attribute__((destructor))` property with low
 priority value 110. This approach does not guarantee, that `GlobalHandler`
 destructor is the last thing to run, as user code may contain a similar function
 with the same priority value. At the same time, users may specify priorities
@@ -72,10 +128,14 @@ times, the memory leak may impact code performance.
 
 ### Windows
 
-To identify shutdown moment on Windows, DPC++ runtime uses default `DllMain`
-function with `DLL_PROCESS_DETACH` reason. This guarantees, that global objects
-deinitialization happens right before `sycl.dll` is unloaded from process
-address space.
+Differing from Linux, on Windows the "early_shutdown()" is begun by 
+DllMain(PROCESS_DETACH), unless statically linked. 
+
+The "late_shutdown()" is begun by the destruction of a 
+static StaticVarShutdownHandler object, which is initialized by 
+platform::get_platforms().  ( On linux, this is when we do "early_shutdown()". 
+Go figure.)  This is as late as we can manage, but it is later than any user 
+application global, static, or thread_local variable destruction.
 
 ### Recommendations for DPC++ runtime developers
 
@@ -109,8 +169,7 @@ for (adapter in initializedAdapters) {
 urLoaderTearDown();
 ```
 
-Which in turn is called by either `shutdown_late()` or `shutdown_win()`
-depending on platform.
+Which in turn is called by `shutdown_late()`.
 
 ![](images/adapter-lifetime.jpg)
 

sycl/source/detail/global_handler.cpp

@@ -37,9 +37,11 @@ using LockGuard = std::lock_guard<SpinLock>;
 SpinLock GlobalHandler::MSyclGlobalHandlerProtector{};
 
 // forward decl
-void shutdown_win(); // TODO: win variant will go away soon
 void shutdown_early();
 void shutdown_late();
+#ifdef _WIN32
+BOOL isLinkedStatically();
+#endif
 
 // Utility class to track references on object.
 // Used for GlobalHandler now and created as thread_local object on the first
@@ -60,10 +62,6 @@ class ObjectUsageCounter {
 
       LockGuard Guard(GlobalHandler::MSyclGlobalHandlerProtector);
       MCounter--;
-      GlobalHandler *RTGlobalObjHandler = GlobalHandler::getInstancePtr();
-      if (RTGlobalObjHandler) {
-        RTGlobalObjHandler->prepareSchedulerToRelease(!MCounter);
-      }
     } catch (std::exception &e) {
       __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in ~ObjectUsageCounter", e);
     }
@@ -254,24 +252,37 @@ void GlobalHandler::releaseDefaultContexts() {
   MPlatformToDefaultContextCache.Inst.reset(nullptr);
 }
 
-struct EarlyShutdownHandler {
-  ~EarlyShutdownHandler() {
+// Shutdown is split into two parts. shutdown_early() stops any more
+// objects from being deferred and takes an initial pass at freeing them.
+// shutdown_late() finishes and releases the adapters and the GlobalHandler.
+// For Windows, early shutdown is typically called from DllMain,
+// and late shutdown is here.
+// For Linux, early shutdown is here, and late shutdown is called from
+// a low priority destructor.
+struct StaticVarShutdownHandler {
+
+  ~StaticVarShutdownHandler() {
     try {
 #ifdef _WIN32
-      // on Windows we keep to the existing shutdown procedure
-      GlobalHandler::instance().releaseDefaultContexts();
+      // If statically linked, DllMain will not be called. So we do its work
+      // here.
+      if (isLinkedStatically()) {
+        shutdown_early();
+      }
+
+      shutdown_late();
 #else
       shutdown_early();
 #endif
     } catch (std::exception &e) {
-      __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in ~EarlyShutdownHandler",
-                                        e);
+      __SYCL_REPORT_EXCEPTION_TO_STREAM(
+          "exception in ~StaticVarShutdownHandler", e);
     }
   }
 };
 
-void GlobalHandler::registerEarlyShutdownHandler() {
-  static EarlyShutdownHandler handler{};
+void GlobalHandler::registerStaticVarShutdownHandler() {
+  static StaticVarShutdownHandler handler{};
 }
 
 bool GlobalHandler::isOkToDefer() const { return OkToDefer; }
@@ -304,35 +315,29 @@ void GlobalHandler::prepareSchedulerToRelease(bool Blocking) {
 #ifndef _WIN32
   if (Blocking)
     drainThreadPool();
+#endif
   if (MScheduler.Inst)
     MScheduler.Inst->releaseResources(Blocking ? BlockingT::BLOCKING
                                                : BlockingT::NON_BLOCKING);
-#endif
 }
 
 void GlobalHandler::drainThreadPool() {
   if (MHostTaskThreadPool.Inst)
     MHostTaskThreadPool.Inst->drain();
 }
 
-#ifdef _WIN32
-// because of something not-yet-understood on Windows
-// threads may be shutdown once the end of main() is reached
-// making an orderly shutdown difficult. Fortunately, Windows
-// itself is very aggressive about reclaiming memory. Thus,
-// we focus solely on unloading the adapters, so as to not
-// accidentally retain device handles. etc
-void shutdown_win() {
-  GlobalHandler *&Handler = GlobalHandler::getInstancePtr();
-  Handler->unloadAdapters();
-}
-#else
 void shutdown_early() {
   const LockGuard Lock{GlobalHandler::MSyclGlobalHandlerProtector};
   GlobalHandler *&Handler = GlobalHandler::getInstancePtr();
   if (!Handler)
     return;
 
+#if defined(XPTI_ENABLE_INSTRUMENTATION) && defined(_WIN32)
+  if (xptiTraceEnabled())
+    return; // When doing xpti tracing, we can't safely shutdown on Win.
+            // TODO: figure out why XPTI prevents release.
+#endif
+
   // Now that we are shutting down, we will no longer defer MemObj releases.
   Handler->endDeferredRelease();
 
@@ -354,6 +359,12 @@ void shutdown_late() {
   if (!Handler)
     return;
 
+#if defined(XPTI_ENABLE_INSTRUMENTATION) && defined(_WIN32)
+  if (xptiTraceEnabled())
+    return; // When doing xpti tracing, we can't safely shutdown on Win.
+            // TODO: figure out why XPTI prevents release.
+#endif
+
   // First, release resources, that may access adapters.
   Handler->MPlatformCache.Inst.reset(nullptr);
   Handler->MScheduler.Inst.reset(nullptr);
@@ -370,7 +381,6 @@ void shutdown_late() {
   delete Handler;
   Handler = nullptr;
 }
-#endif
 
 #ifdef _WIN32
 extern "C" __SYCL_EXPORT BOOL WINAPI DllMain(HINSTANCE hinstDLL,
@@ -391,23 +401,18 @@ extern "C" __SYCL_EXPORT BOOL WINAPI DllMain(HINSTANCE hinstDLL,
     if (PrintUrTrace)
       std::cout << "---> DLL_PROCESS_DETACH syclx.dll\n" << std::endl;
 
-#ifdef XPTI_ENABLE_INSTRUMENTATION
-    if (xptiTraceEnabled())
-      return TRUE; // When doing xpti tracing, we can't safely call shutdown.
-                   // TODO: figure out what XPTI is doing that prevents
-                   // release.
-#endif
-
     try {
-      shutdown_win();
+      shutdown_early();
     } catch (std::exception &e) {
-      __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in shutdown_win", e);
+      __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in DLL_PROCESS_DETACH", e);
       return FALSE;
     }
+
     break;
   case DLL_PROCESS_ATTACH:
     if (PrintUrTrace)
       std::cout << "---> DLL_PROCESS_ATTACH syclx.dll\n" << std::endl;
+
     break;
   case DLL_THREAD_ATTACH:
     break;
@@ -416,6 +421,29 @@ extern "C" __SYCL_EXPORT BOOL WINAPI DllMain(HINSTANCE hinstDLL,
   }
   return TRUE; // Successful DLL_PROCESS_ATTACH.
 }
+BOOL isLinkedStatically() {
+  // If the exePath is the same as the dllPath,
+  // or if the module handle for DllMain is not retrievable,
+  // then we are linked statically
+  // Otherwise we are dynamically linked or loaded.
+  HMODULE hModule = nullptr;
+  auto LpModuleAddr = reinterpret_cast<LPCSTR>(&DllMain);
+  if (!GetModuleHandleExA(GET_MODULE_HANDLE_EX_FLAG_FROM_ADDRESS, LpModuleAddr,
+                          &hModule)) {
+    return true; // not retrievable, therefore statically linked
+  } else {
+    char dllPath[MAX_PATH];
+    if (GetModuleFileNameA(hModule, dllPath, MAX_PATH)) {
+      char exePath[MAX_PATH];
+      if (GetModuleFileNameA(NULL, exePath, MAX_PATH)) {
+        if (std::string(dllPath) == std::string(exePath)) {
+          return true; // paths identical, therefore statically linked
+        }
+      }
+    }
+  }
+  return false; // Otherwise dynamically linked or loaded
+}
 #else
 // Setting low priority on destructor ensures it runs after all other global
 // destructors. Priorities 0-100 are reserved by the compiler. The priority

sycl/source/detail/global_handler.hpp

@@ -73,7 +73,7 @@ class GlobalHandler {
   XPTIRegistry &getXPTIRegistry();
   ThreadPool &getHostTaskThreadPool();
 
-  static void registerEarlyShutdownHandler();
+  static void registerStaticVarShutdownHandler();
 
   bool isOkToDefer() const;
   void endDeferredRelease();
@@ -95,7 +95,6 @@ class GlobalHandler {
 
   bool OkToDefer = true;
 
-  friend void shutdown_win();
   friend void shutdown_early();
   friend void shutdown_late();
   friend class ObjectUsageCounter;

sycl/source/detail/host_task.hpp

@@ -13,6 +13,7 @@
 #pragma once
 
 #include <detail/cg.hpp>
+#include <detail/global_handler.hpp>
 #include <sycl/detail/cg_types.hpp>
 #include <sycl/handler.hpp>
 #include <sycl/interop_handle.hpp>
@@ -33,6 +34,10 @@ class HostTask {
   bool isInteropTask() const { return !!MInteropTask; }
 
   void call(HostProfilingInfo *HPI) {
+    if (!GlobalHandler::instance().isOkToDefer()) {
+      return;
+    }
+
     if (HPI)
       HPI->start();
     MHostTask();
@@ -41,6 +46,10 @@ class HostTask {
   }
 
   void call(HostProfilingInfo *HPI, interop_handle handle) {
+    if (!GlobalHandler::instance().isOkToDefer()) {
+      return;
+    }
+
     if (HPI)
       HPI->start();
     MInteropTask(handle);

sycl/source/detail/platform_impl.cpp

@@ -166,7 +166,7 @@ std::vector<platform> platform_impl::get_platforms() {
 
   // This initializes a function-local variable whose destructor is invoked as
   // the SYCL shared library is first being unloaded.
-  GlobalHandler::registerEarlyShutdownHandler();
+  GlobalHandler::registerStaticVarShutdownHandler();
 
   return Platforms;
 }

sycl/source/detail/program_manager/program_manager.cpp

@@ -3846,7 +3846,9 @@ extern "C" void __sycl_register_lib(sycl_device_binaries desc) {
 // Executed as a part of current module's (.exe, .dll) static initialization
 extern "C" void __sycl_unregister_lib(sycl_device_binaries desc) {
   // Partial cleanup is not necessary at shutdown
+#ifndef _WIN32
   if (!sycl::detail::GlobalHandler::instance().isOkToDefer())
     return;
   sycl::detail::ProgramManager::getInstance().removeImages(desc);
+#endif
 }

sycl/source/detail/queue_impl.hpp

@@ -264,7 +264,17 @@ class queue_impl {
       destructorNotification();
 #endif
       throw_asynchronous();
-      getAdapter()->call<UrApiKind::urQueueRelease>(MQueue);
+      auto status =
+          getAdapter()->call_nocheck<UrApiKind::urQueueRelease>(MQueue);
+      // If loader is already closed, it'll return a not-initialized status
+      // which the UR should convert to SUCCESS code. But that isn't always
+      // working on Windows. This is a temporary workaround until that is fixed.
+      // TODO: Remove this workaround when UR is fixed, and restore
+      // ->call<>() instead of ->call_nocheck<>() above.
+      if (status != UR_RESULT_SUCCESS &&
+          status != UR_RESULT_ERROR_UNINITIALIZED) {
+        __SYCL_CHECK_UR_CODE_NO_EXC(status);
+      }
     } catch (std::exception &e) {
       __SYCL_REPORT_EXCEPTION_TO_STREAM("exception in ~queue_impl", e);
     }