--- old/src/hotspot/share/runtime/heapMonitoring.cpp	2017-10-06 14:50:10.842641126 -0700
+++ new/src/hotspot/share/runtime/heapMonitoring.cpp	2017-10-06 14:50:10.458642416 -0700
@@ -29,7 +29,7 @@
 #include "runtime/heapMonitoring.hpp"
 #include "runtime/vframe.hpp"
 
-const int MaxStackDepth = 1024;
+static const int MaxStackDepth = 1024;
 
 // Internal data structure representing traces.
 struct StackTraceData : CHeapObj<mtInternal> {
@@ -204,32 +204,21 @@
   StackTraceStorage();
 
   static StackTraceStorage* storage() {
-    if (internal_storage == NULL) {
-      internal_storage = new StackTraceStorage();
-    }
-    return internal_storage;
-  }
-
-  static void reset_stack_trace_storage() {
-    delete internal_storage;
-    internal_storage = NULL;
+    static StackTraceStorage internal_storage;
+    return &internal_storage;
   }
 
-  bool is_initialized() {
-    return _initialized;
+  void initialize(int max_storage) {
+    MutexLocker mu(HeapMonitor_lock);
+    free_storage();
+    allocate_storage(max_storage);
+    memset(&_stats, 0, sizeof(_stats));
   }
 
   const jvmtiHeapSamplingStats& get_heap_sampling_stats() const {
     return _stats;
   }
 
-  // Static method to set the storage in place at initialization.
-  static void initialize_stack_trace_storage(int max_storage) {
-    reset_stack_trace_storage();
-    StackTraceStorage *storage = StackTraceStorage::storage();
-    storage->initialize_storage(max_storage);
-  }
-
   void accumulate_sample_rate(size_t rate) {
     _stats.sample_rate_accumulation += rate;
     _stats.sample_rate_count++;
@@ -239,6 +228,9 @@
   volatile bool *initialized_address() { return &_initialized; }
 
  private:
+  // Protects the traces currently sampled (below).
+  volatile intptr_t _stack_storage_lock[1];
+
   // The traces currently sampled.
   GrowableArray<StackTraceData> *_allocated_traces;
 
@@ -288,9 +280,6 @@
     int _size;
   };
 
-  // Instance initialization.
-  void initialize_storage(int max_storage);
-
   // Copies from StackTraceData to jvmtiStackTrace.
   bool deep_copy(jvmtiStackTrace *to, const StackTraceData *from);
 
@@ -301,6 +290,8 @@
   void store_garbage_trace(const StackTraceData &trace);
 
   void free_garbage();
+  void free_storage();
+  void allocate_storage(int max_gc_storage);
 };
 
 StackTraceStorage* StackTraceStorage::internal_storage;
@@ -320,7 +311,7 @@
   _frequent_garbage_traces(NULL),
   _max_gc_storage(0),
   _initialized(false) {
-    memset(&_stats, 0, sizeof(_stats));
+    _stack_storage_lock[0] = 0;
 }
 
 void StackTraceStorage::free_garbage() {
@@ -366,7 +357,7 @@
   }
 }
 
-StackTraceStorage::~StackTraceStorage() {
+void StackTraceStorage::free_storage() {
   delete _allocated_traces;
 
   free_garbage();
@@ -375,7 +366,11 @@
   _initialized = false;
 }
 
-void StackTraceStorage::initialize_storage(int max_gc_storage) {
+StackTraceStorage::~StackTraceStorage() {
+  free_storage();
+}
+
+void StackTraceStorage::allocate_storage(int max_gc_storage) {
   // In case multiple threads got locked and then 1 by 1 got through.
   if (_initialized) {
     return;
@@ -392,6 +387,7 @@
 }
 
 void StackTraceStorage::add_trace(jvmtiStackTrace *trace, oop o) {
+  MutexLocker mu(HeapMonitor_lock);
   StackTraceData new_data(trace, o);
   _stats.sample_count++;
   _stats.stack_depth_accumulation += trace->frame_count;
@@ -400,8 +396,9 @@
 
 size_t StackTraceStorage::weak_oops_do(BoolObjectClosure *is_alive,
                                        OopClosure *f) {
+  MutexLocker mu(HeapMonitor_lock);
   size_t count = 0;
-  if (is_initialized()) {
+  if (initialized()) {
     int len = _allocated_traces->length();
 
     // Compact the oop traces.  Moves the live oops to the beginning of the
@@ -484,6 +481,7 @@
 
 void StackTraceStorage::copy_stack_traces(const StackTraceDataCopier &copier,
                                           jvmtiStackTraces *traces) {
+  MutexLocker mu(HeapMonitor_lock);
   int len = copier.size();
 
   // Create a new array to store the StackTraceData objects.
@@ -592,9 +590,6 @@
 
   _monitoring_rate = monitoring_rate;
 
-  // Initalize and reset.
-  StackTraceStorage::initialize_stack_trace_storage(max_gc_storage);
-
   // Populate the lookup table for fast_log2.
   // This approximates the log2 curve with a step function.
   // Steps have height equal to log2 of the mid-point of the step.
@@ -608,6 +603,8 @@
   if (_rnd == 0) {
     _rnd = 1;
   }
+
+  StackTraceStorage::storage()->initialize(max_gc_storage);
   _enabled = true;
 }
 
@@ -654,7 +651,7 @@
 void HeapMonitoring::object_alloc_do_sample(Thread *t, oopDesc *o, intx byte_size) {
 #if defined(X86) || defined(PPC)
   JavaThread *thread = static_cast<JavaThread *>(t);
-  if (StackTraceStorage::storage()->is_initialized()) {
+  if (StackTraceStorage::storage()->initialized()) {
     assert(t->is_Java_thread(), "non-Java thread passed to do_sample");
     JavaThread *thread = static_cast<JavaThread *>(t);
 
@@ -699,12 +696,6 @@
     // Failure!
     FREE_C_HEAP_ARRAY(jvmtiFrameInfo, trace->frames);
     FREE_C_HEAP_OBJ(trace);
-    return;
-  } else {
-    // There is something like 64K worth of allocation before the VM
-    // initializes.  This is just in the interests of not slowing down
-    // startup.
-    assert(t->is_Java_thread(), "non-Java thread passed to do_sample");
   }
 #else
   Unimplemented();