4957990: Perm heap bloat in JVM

Treat ProfileData in MDO's as a source of weak, not strong, roots. Fixes the bug for stop-world collection -- the case of concurrent collection will be fixed separately.

Reviewed-by: jcoomes, jmasa, kvn, never

hotspot/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp

@@ -58,9 +58,8 @@ void MarkFromRootsTask::do_it(GCTaskManager* manager, uint which) {
     PrintGCDetails && TraceParallelOldGCTasks, true, gclog_or_tty));
   ParCompactionManager* cm =
     ParCompactionManager::gc_thread_compaction_manager(which);
-  // cm->allocate_stacks();
   assert(cm->stacks_have_been_allocated(),
-    "Stack space has not been allocated");
+         "Stack space has not been allocated");
   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
 
   switch (_root_type) {
@@ -129,9 +128,8 @@ void RefProcTaskProxy::do_it(GCTaskManager* manager, uint which)
     PrintGCDetails && TraceParallelOldGCTasks, true, gclog_or_tty));
   ParCompactionManager* cm =
     ParCompactionManager::gc_thread_compaction_manager(which);
-  // cm->allocate_stacks();
   assert(cm->stacks_have_been_allocated(),
-    "Stack space has not been allocated");
+         "Stack space has not been allocated");
   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
   PSParallelCompact::FollowStackClosure follow_stack_closure(cm);
   _rp_task.work(_work_id, *PSParallelCompact::is_alive_closure(),

hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp

@@ -61,12 +61,16 @@ ParCompactionManager::ParCompactionManager() :
   int size =
     (SystemDictionary::number_of_classes() * 2) * 2 / ParallelGCThreads;
   _revisit_klass_stack = new (ResourceObj::C_HEAP) GrowableArray<Klass*>(size, true);
+  // From some experiments (#klass/k)^2 for k = 10 seems a better fit, but this will
+  // have to do for now until we are able to investigate a more optimal setting.
+  _revisit_mdo_stack = new (ResourceObj::C_HEAP) GrowableArray<DataLayout*>(size*2, true);
 
 }
 
 ParCompactionManager::~ParCompactionManager() {
   delete _overflow_stack;
   delete _revisit_klass_stack;
+  delete _revisit_mdo_stack;
   // _manager_array and _stack_array are statics
   // shared with all instances of ParCompactionManager
   // should not be deallocated.
@@ -195,6 +199,7 @@ ParCompactionManager::gc_thread_compaction_manager(int index) {
 void ParCompactionManager::reset() {
   for(uint i=0; i<ParallelGCThreads+1; i++) {
     manager_array(i)->revisit_klass_stack()->clear();
+    manager_array(i)->revisit_mdo_stack()->clear();
   }
 }
 
@@ -296,6 +301,7 @@ void ParCompactionManager::drain_region_stacks() {
 
 #ifdef ASSERT
 bool ParCompactionManager::stacks_have_been_allocated() {
-  return (revisit_klass_stack()->data_addr() != NULL);
+  return (revisit_klass_stack()->data_addr() != NULL &&
+          revisit_mdo_stack()->data_addr() != NULL);
 }
 #endif

hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp

@@ -93,6 +93,7 @@ class ParCompactionManager : public CHeapObj {
 
 #if 1  // does this happen enough to need a per thread stack?
   GrowableArray<Klass*>*        _revisit_klass_stack;
+  GrowableArray<DataLayout*>*   _revisit_mdo_stack;
 #endif
   static ParMarkBitMap* _mark_bitmap;
 
@@ -154,6 +155,7 @@ class ParCompactionManager : public CHeapObj {
 #if 1
   // Probably stays as a growable array
   GrowableArray<Klass*>* revisit_klass_stack() { return _revisit_klass_stack; }
+  GrowableArray<DataLayout*>* revisit_mdo_stack() { return _revisit_mdo_stack; }
 #endif
 
   // Save oop for later processing.  Must not fail.

hotspot/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp

@@ -482,6 +482,9 @@ void PSMarkSweep::allocate_stacks() {
 
   int size = SystemDictionary::number_of_classes() * 2;
   _revisit_klass_stack = new (ResourceObj::C_HEAP) GrowableArray<Klass*>(size, true);
+  // (#klass/k)^2, for k ~ 10 appears a better setting, but this will have to do for
+  // now until we investigate a more optimal setting.
+  _revisit_mdo_stack   = new (ResourceObj::C_HEAP) GrowableArray<DataLayout*>(size*2, true);
 }
 
 
@@ -495,6 +498,7 @@ void PSMarkSweep::deallocate_stacks() {
 
   delete _marking_stack;
   delete _revisit_klass_stack;
+  delete _revisit_mdo_stack;
 }
 
 void PSMarkSweep::mark_sweep_phase1(bool clear_all_softrefs) {
@@ -540,6 +544,10 @@ void PSMarkSweep::mark_sweep_phase1(bool clear_all_softrefs) {
   follow_weak_klass_links();
   assert(_marking_stack->is_empty(), "just drained");
 
+  // Visit memoized mdo's and clear unmarked weak refs
+  follow_mdo_weak_refs();
+  assert(_marking_stack->is_empty(), "just drained");
+
   // Visit symbol and interned string tables and delete unmarked oops
   SymbolTable::unlink(is_alive_closure());
   StringTable::unlink(is_alive_closure());

hotspot/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp

@@ -2378,7 +2378,10 @@ void PSParallelCompact::marking_phase(ParCompactionManager* cm,
 
   // Update subklass/sibling/implementor links of live klasses
   // revisit_klass_stack is used in follow_weak_klass_links().
-  follow_weak_klass_links(cm);
+  follow_weak_klass_links();
+
+  // Revisit memoized MDO's and clear any unmarked weak refs
+  follow_mdo_weak_refs();
 
   // Visit symbol and interned string tables and delete unmarked oops
   SymbolTable::unlink(is_alive_closure());
@@ -2721,17 +2724,25 @@ void PSParallelCompact::follow_stack(ParCompactionManager* cm) {
 }
 
 void
-PSParallelCompact::follow_weak_klass_links(ParCompactionManager* serial_cm) {
+PSParallelCompact::follow_weak_klass_links() {
   // All klasses on the revisit stack are marked at this point.
   // Update and follow all subklass, sibling and implementor links.
-  for (uint i = 0; i < ParallelGCThreads+1; i++) {
+  if (PrintRevisitStats) {
+    gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes());
+  }
+  for (uint i = 0; i < ParallelGCThreads + 1; i++) {
     ParCompactionManager* cm = ParCompactionManager::manager_array(i);
     KeepAliveClosure keep_alive_closure(cm);
-    for (int i = 0; i < cm->revisit_klass_stack()->length(); i++) {
-      cm->revisit_klass_stack()->at(i)->follow_weak_klass_links(
+    int length = cm->revisit_klass_stack()->length();
+    if (PrintRevisitStats) {
+      gclog_or_tty->print_cr("Revisit klass stack[%d] length = %d", i, length);
+    }
+    for (int j = 0; j < length; j++) {
+      cm->revisit_klass_stack()->at(j)->follow_weak_klass_links(
         is_alive_closure(),
         &keep_alive_closure);
     }
+    // revisit_klass_stack is cleared in reset()
     follow_stack(cm);
   }
 }
@@ -2741,6 +2752,35 @@ PSParallelCompact::revisit_weak_klass_link(ParCompactionManager* cm, Klass* k) {
   cm->revisit_klass_stack()->push(k);
 }
 
+#if ( defined(COMPILER1) || defined(COMPILER2) )
+void PSParallelCompact::revisit_mdo(ParCompactionManager* cm, DataLayout* p) {
+  cm->revisit_mdo_stack()->push(p);
+}
+
+void PSParallelCompact::follow_mdo_weak_refs() {
+  // All strongly reachable oops have been marked at this point;
+  // we can visit and clear any weak references from MDO's which
+  // we memoized during the strong marking phase.
+  if (PrintRevisitStats) {
+    gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes());
+  }
+  for (uint i = 0; i < ParallelGCThreads + 1; i++) {
+    ParCompactionManager* cm = ParCompactionManager::manager_array(i);
+    GrowableArray<DataLayout*>* rms = cm->revisit_mdo_stack();
+    int length = rms->length();
+    if (PrintRevisitStats) {
+      gclog_or_tty->print_cr("Revisit MDO stack[%d] length = %d", i, length);
+    }
+    for (int j = 0; j < length; j++) {
+      rms->at(j)->follow_weak_refs(is_alive_closure());
+    }
+    // revisit_mdo_stack is cleared in reset()
+    follow_stack(cm);
+  }
+}
+#endif //  ( COMPILER1 || COMPILER2 )
+
+
 #ifdef VALIDATE_MARK_SWEEP
 
 void PSParallelCompact::track_adjusted_pointer(void* p, bool isroot) {

hotspot/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.hpp

@@ -901,7 +901,8 @@ class PSParallelCompact : AllStatic {
   static void marking_phase(ParCompactionManager* cm,
                             bool maximum_heap_compaction);
   static void follow_stack(ParCompactionManager* cm);
-  static void follow_weak_klass_links(ParCompactionManager* cm);
+  static void follow_weak_klass_links();
+  static void follow_mdo_weak_refs();
 
   template <class T> static inline void adjust_pointer(T* p, bool is_root);
   static void adjust_root_pointer(oop* p) { adjust_pointer(p, true); }
@@ -1221,6 +1222,9 @@ class PSParallelCompact : AllStatic {
   // Update subklass/sibling/implementor links at end of marking.
   static void revisit_weak_klass_link(ParCompactionManager* cm, Klass* k);
 
+  // Clear unmarked oops in MDOs at the end of marking.
+  static void revisit_mdo(ParCompactionManager* cm, DataLayout* p);
+
 #ifndef PRODUCT
   // Debugging support.
   static const char* space_names[last_space_id];