SWDEV-431879 - Introduce IsHandlerPending back.

It seems that due to removal of vdev()->isHandlerPending(),
Marker queued to ensure finish is not enqueued and that cause
hung at waiting event for kernel enqueue command.

Change-Id: I364abb2dcb4897b11a7eb61b5d85013b69292792

rocclr/device/device.hpp

@@ -1287,6 +1287,9 @@ class VirtualDevice : public amd::HeapObject {
   //! Returns true if device has active wait setting
   bool ActiveWait() const;
 
+  //! Returns the status of queue handler callback
+  virtual bool isHandlerPending() const = 0;
+
   //! Returns fence state of the VirtualGPU
   virtual bool isFenceDirty() const = 0;
 

rocclr/device/pal/palvirtual.hpp

@@ -341,6 +341,8 @@ class VirtualGPU : public device::VirtualDevice {
 
   void profilerAttach(bool enable = false) {}
 
+  bool isHandlerPending() const { return false; }
+
   bool isFenceDirty() const { return false; }
 
   inline bool dispatchAqlPacket(uint8_t* aqlpacket, amd::AccumulateCommand* vcmd = nullptr) {

rocclr/device/rocm/rocvirtual.cpp

@@ -468,10 +468,12 @@ hsa_signal_t VirtualGPU::HwQueueTracker::ActiveSignal(
           ClPrint(amd::LOG_INFO, amd::LOG_SIG, "Set Handler: handle(0x%lx), timestamp(%p)",
             prof_signal->signal_.handle, prof_signal);
         }
+        SetHandlerPending(false);
         // Update the current command/marker with HW event
         prof_signal->retain();
         ts->command().SetHwEvent(prof_signal);
       } else if (ts->command().profilingInfo().marker_ts_) {
+        SetHandlerPending(true);
         // Update the current command/marker with HW event
         prof_signal->retain();
         ts->command().SetHwEvent(prof_signal);

rocclr/device/rocm/rocvirtual.hpp

@@ -223,7 +223,7 @@ class VirtualGPU : public device::VirtualDevice {
 
   class HwQueueTracker : public amd::EmbeddedObject {
    public:
-    HwQueueTracker(const VirtualGPU& gpu): gpu_(gpu) {}
+    HwQueueTracker(const VirtualGPU& gpu): gpu_(gpu), handlerPending_(false) {}
 
     ~HwQueueTracker();
 
@@ -265,6 +265,12 @@ class VirtualGPU : public device::VirtualDevice {
     //! Empty check for external signals
     bool IsExternalSignalListEmpty() const { return external_signals_.empty(); }
 
+    //! Set the status to indicate a pending handler
+    void SetHandlerPending(bool pending) { handlerPending_ = pending; }
+
+    //! Check if callback has been queued
+    bool IsHandlerPending() const { return handlerPending_; }
+
     //! Get/Set SDMA profiling
     bool GetSDMAProfiling() { return sdma_profiling_; }
     void SetSDMAProfiling(bool profile) {
@@ -289,6 +295,7 @@ class VirtualGPU : public device::VirtualDevice {
     const VirtualGPU& gpu_;       //!< VirtualGPU, associated with this tracker
     std::vector<ProfilingSignal*> external_signals_; //!< External signals for a wait in this queue
     std::vector<hsa_signal_t> waiting_signals_;   //!< Current waiting signals in this queue
+    bool handlerPending_;         //!< This indicates if we have queued a callback handler
   };
 
   VirtualGPU(Device& device, bool profiling = false, bool cooperative = false,
@@ -406,6 +413,10 @@ class VirtualGPU : public device::VirtualDevice {
 
   Timestamp* timestamp() const { return timestamp_; }
 
+  //! Indicates the status of the callback handler. The callback would process the commands
+  //! and would collect profiling data, update refcounts
+  bool isHandlerPending() const { return barriers_.IsHandlerPending(); }
+
   void* allocKernArg(size_t size, size_t alignment);
   bool isFenceDirty() const { return fence_dirty_; }
   void resetFenceDirty() { fence_dirty_ = false; }

rocclr/platform/commandqueue.cpp

@@ -131,7 +131,8 @@ void HostQueue::finish(bool cpu_wait) {
       (command->NotifyEvent() != nullptr) ? command->NotifyEvent()->HwEvent() : command->HwEvent();
     force_marker = (hw_event == nullptr);
   }
-  if (nullptr == command || force_marker || vdev()->isFenceDirty()) {
+  if (nullptr == command || force_marker ||
+      vdev()->isHandlerPending() || vdev()->isFenceDirty()) {
     if (nullptr != command) {
       command->release();
     }