changeset 1843:0715f0cf171d

Merge
author jcoomes
date Fri, 08 Oct 2010 09:29:09 -0700
parents 22e4420d19f7 (current diff) 6e0aac35bfa9 (diff)
children b98784e85f71 c32059ef4dc0 75b0735b4d04
files src/share/vm/includeDB_core src/share/vm/memory/referenceProcessor.hpp src/share/vm/runtime/globals.hpp src/share/vm/runtime/thread.cpp src/share/vm/runtime/thread.hpp
diffstat 65 files changed, 1599 insertions(+), 716 deletions(-) [+]
line wrap: on
line diff
--- a/make/solaris/makefiles/sparcWorks.make	Wed Oct 06 14:18:32 2010 -0700
+++ b/make/solaris/makefiles/sparcWorks.make	Fri Oct 08 09:29:09 2010 -0700
@@ -51,9 +51,9 @@
   VALIDATED_COMPILER_REVS   := 5.8
   VALIDATED_C_COMPILER_REVS := 5.8
 else
-  # Validated compilers for JDK7 are SS12 (5.9) or SS12 update 1 (5.10)
-  VALIDATED_COMPILER_REVS   := 5.9 5.10
-  VALIDATED_C_COMPILER_REVS := 5.9 5.10
+  # Validated compiler for JDK7 is SS12 update 1 + patches (5.10)
+  VALIDATED_COMPILER_REVS   := 5.10
+  VALIDATED_C_COMPILER_REVS := 5.10
 endif
 
 # Warning messages about not using the above validated versions
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/cmsCollectorPolicy.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2007, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -39,7 +39,7 @@
   if (_generations == NULL)
     vm_exit_during_initialization("Unable to allocate gen spec");
 
-  if (UseParNewGC && ParallelGCThreads > 0) {
+  if (ParNewGeneration::in_use()) {
     if (UseAdaptiveSizePolicy) {
       _generations[0] = new GenerationSpec(Generation::ASParNew,
                                            _initial_gen0_size, _max_gen0_size);
@@ -79,7 +79,7 @@
 
 void ConcurrentMarkSweepPolicy::initialize_gc_policy_counters() {
   // initialize the policy counters - 2 collectors, 3 generations
-  if (UseParNewGC && ParallelGCThreads > 0) {
+  if (ParNewGeneration::in_use()) {
     _gc_policy_counters = new GCPolicyCounters("ParNew:CMS", 2, 3);
   }
   else {
@@ -102,7 +102,7 @@
 
   assert(size_policy() != NULL, "A size policy is required");
   // initialize the policy counters - 2 collectors, 3 generations
-  if (UseParNewGC && ParallelGCThreads > 0) {
+  if (ParNewGeneration::in_use()) {
     _gc_policy_counters = new CMSGCAdaptivePolicyCounters("ParNew:CMS", 2, 3,
       size_policy());
   }
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/cmsPermGen.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/cmsPermGen.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2007, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -50,6 +50,18 @@
   }
 }
 
+HeapWord* CMSPermGen::request_expand_and_allocate(Generation* gen,
+                                                  size_t size,
+                                                  GCCause::Cause prev_cause /* ignored */) {
+  HeapWord* obj = gen->expand_and_allocate(size, false);
+  if (gen->capacity() >= _capacity_expansion_limit) {
+    set_capacity_expansion_limit(gen->capacity() + MaxPermHeapExpansion);
+    assert(((ConcurrentMarkSweepGeneration*)gen)->should_concurrent_collect(),
+           "Should kick off a collection if one not in progress");
+  }
+  return obj;
+}
+
 void CMSPermGen::compute_new_size() {
   _gen->compute_new_size();
 }
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/cmsPermGen.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/cmsPermGen.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -33,6 +33,10 @@
   // The "generation" view.
   ConcurrentMarkSweepGeneration* _gen;
 
+  // Override default implementation from PermGen
+  virtual HeapWord* request_expand_and_allocate(Generation* gen, size_t size,
+                                                GCCause::Cause prev_cause);
+
  public:
   CMSPermGen(ReservedSpace rs, size_t initial_byte_size,
              CardTableRS* ct, FreeBlockDictionary::DictionaryChoice);
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/compactibleFreeListSpace.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -124,7 +124,8 @@
   checkFreeListConsistency();
 
   // Initialize locks for parallel case.
-  if (ParallelGCThreads > 0) {
+
+  if (CollectedHeap::use_parallel_gc_threads()) {
     for (size_t i = IndexSetStart; i < IndexSetSize; i += IndexSetStride) {
       _indexedFreeListParLocks[i] = new Mutex(Mutex::leaf - 1, // == ExpandHeap_lock - 1
                                               "a freelist par lock",
@@ -1071,7 +1072,8 @@
   // at address below "p" in finding the object that contains "p"
   // and those objects (if garbage) may have been modified to hold
   // live range information.
-  // assert(ParallelGCThreads > 0 || _bt.block_start(p) == p, "Should be a block boundary");
+  // assert(CollectedHeap::use_parallel_gc_threads() || _bt.block_start(p) == p,
+  //        "Should be a block boundary");
   if (FreeChunk::indicatesFreeChunk(p)) return false;
   klassOop k = oop(p)->klass_or_null();
   if (k != NULL) {
@@ -2932,7 +2934,9 @@
          "n_tasks calculation incorrect");
   SequentialSubTasksDone* pst = conc_par_seq_tasks();
   assert(!pst->valid(), "Clobbering existing data?");
-  pst->set_par_threads(n_threads);
+  // Sets the condition for completion of the subtask (how many threads
+  // need to finish in order to be done).
+  pst->set_n_threads(n_threads);
   pst->set_n_tasks((int)n_tasks);
 }
 
@@ -2972,6 +2976,8 @@
          "n_tasks calculation incorrect");
   SequentialSubTasksDone* pst = conc_par_seq_tasks();
   assert(!pst->valid(), "Clobbering existing data?");
-  pst->set_par_threads(n_threads);
+  // Sets the condition for completion of the subtask (how many threads
+  // need to finish in order to be done).
+  pst->set_n_threads(n_threads);
   pst->set_n_tasks((int)n_tasks);
 }
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -195,7 +195,7 @@
            "Offset of FreeChunk::_prev within FreeChunk must match"
            "  that of OopDesc::_klass within OopDesc");
   )
-  if (ParallelGCThreads > 0) {
+  if (CollectedHeap::use_parallel_gc_threads()) {
     typedef CMSParGCThreadState* CMSParGCThreadStatePtr;
     _par_gc_thread_states =
       NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads);
@@ -540,8 +540,6 @@
   _is_alive_closure(_span, &_markBitMap),
   _restart_addr(NULL),
   _overflow_list(NULL),
-  _preserved_oop_stack(NULL),
-  _preserved_mark_stack(NULL),
   _stats(cmsGen),
   _eden_chunk_array(NULL),     // may be set in ctor body
   _eden_chunk_capacity(0),     // -- ditto --
@@ -616,7 +614,7 @@
   }
 
   // Support for multi-threaded concurrent phases
-  if (ParallelGCThreads > 0 && CMSConcurrentMTEnabled) {
+  if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) {
     if (FLAG_IS_DEFAULT(ConcGCThreads)) {
       // just for now
       FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4);
@@ -628,6 +626,8 @@
         warning("GC/CMS: _conc_workers allocation failure: "
               "forcing -CMSConcurrentMTEnabled");
         CMSConcurrentMTEnabled = false;
+      } else {
+        _conc_workers->initialize_workers();
       }
     } else {
       CMSConcurrentMTEnabled = false;
@@ -936,7 +936,7 @@
   // along with all the other pointers into the heap but
   // compaction is expected to be a rare event with
   // a heap using cms so don't do it without seeing the need.
-  if (ParallelGCThreads > 0) {
+  if (CollectedHeap::use_parallel_gc_threads()) {
     for (uint i = 0; i < ParallelGCThreads; i++) {
       _par_gc_thread_states[i]->promo.reset();
     }
@@ -2630,7 +2630,8 @@
   // Should call gc_prologue_work() for all cms gens we are responsible for
   bool registerClosure =    _collectorState >= Marking
                          && _collectorState < Sweeping;
-  ModUnionClosure* muc = ParallelGCThreads > 0 ? &_modUnionClosurePar
+  ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ?
+                                               &_modUnionClosurePar
                                                : &_modUnionClosure;
   _cmsGen->gc_prologue_work(full, registerClosure, muc);
   _permGen->gc_prologue_work(full, registerClosure, muc);
@@ -2731,7 +2732,7 @@
   collector()->gc_epilogue(full);
 
   // Also reset promotion tracking in par gc thread states.
-  if (ParallelGCThreads > 0) {
+  if (CollectedHeap::use_parallel_gc_threads()) {
     for (uint i = 0; i < ParallelGCThreads; i++) {
       _par_gc_thread_states[i]->promo.stopTrackingPromotions(i);
     }
@@ -3263,6 +3264,7 @@
 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size,
                                                    bool   tlab,
                                                    bool   parallel) {
+  CMSSynchronousYieldRequest yr;
   assert(!tlab, "Can't deal with TLAB allocation");
   MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag);
   expand(word_size*HeapWordSize, MinHeapDeltaBytes,
@@ -3709,20 +3711,26 @@
 class CMSConcMarkingTerminator: public ParallelTaskTerminator {
   CMSCollector*       _collector;
   CMSConcMarkingTask* _task;
-  bool _yield;
- protected:
+ public:
   virtual void yield();
- public:
+
   // "n_threads" is the number of threads to be terminated.
   // "queue_set" is a set of work queues of other threads.
   // "collector" is the CMS collector associated with this task terminator.
   // "yield" indicates whether we need the gang as a whole to yield.
-  CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set,
-                           CMSCollector* collector, bool yield) :
+  CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) :
     ParallelTaskTerminator(n_threads, queue_set),
-    _collector(collector),
-    _yield(yield) { }
-
+    _collector(collector) { }
+
+  void set_task(CMSConcMarkingTask* task) {
+    _task = task;
+  }
+};
+
+class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator {
+  CMSConcMarkingTask* _task;
+ public:
+  bool should_exit_termination();
   void set_task(CMSConcMarkingTask* task) {
     _task = task;
   }
@@ -3731,13 +3739,14 @@
 // MT Concurrent Marking Task
 class CMSConcMarkingTask: public YieldingFlexibleGangTask {
   CMSCollector* _collector;
-  YieldingFlexibleWorkGang* _workers;        // the whole gang
   int           _n_workers;                  // requested/desired # workers
   bool          _asynch;
   bool          _result;
   CompactibleFreeListSpace*  _cms_space;
   CompactibleFreeListSpace* _perm_space;
-  HeapWord*     _global_finger;
+  char          _pad_front[64];   // padding to ...
+  HeapWord*     _global_finger;   // ... avoid sharing cache line
+  char          _pad_back[64];
   HeapWord*     _restart_addr;
 
   //  Exposed here for yielding support
@@ -3745,28 +3754,30 @@
 
   // The per thread work queues, available here for stealing
   OopTaskQueueSet*  _task_queues;
+
+  // Termination (and yielding) support
   CMSConcMarkingTerminator _term;
+  CMSConcMarkingTerminatorTerminator _term_term;
 
  public:
   CMSConcMarkingTask(CMSCollector* collector,
                  CompactibleFreeListSpace* cms_space,
                  CompactibleFreeListSpace* perm_space,
-                 bool asynch, int n_workers,
+                 bool asynch,
                  YieldingFlexibleWorkGang* workers,
                  OopTaskQueueSet* task_queues):
     YieldingFlexibleGangTask("Concurrent marking done multi-threaded"),
     _collector(collector),
     _cms_space(cms_space),
     _perm_space(perm_space),
-    _asynch(asynch), _n_workers(n_workers), _result(true),
-    _workers(workers), _task_queues(task_queues),
-    _term(n_workers, task_queues, _collector, asynch),
+    _asynch(asynch), _n_workers(0), _result(true),
+    _task_queues(task_queues),
+    _term(_n_workers, task_queues, _collector),
     _bit_map_lock(collector->bitMapLock())
   {
-    assert(n_workers <= workers->total_workers(),
-           "Else termination won't work correctly today"); // XXX FIX ME!
-    _requested_size = n_workers;
+    _requested_size = _n_workers;
     _term.set_task(this);
+    _term_term.set_task(this);
     assert(_cms_space->bottom() < _perm_space->bottom(),
            "Finger incorrectly initialized below");
     _restart_addr = _global_finger = _cms_space->bottom();
@@ -3781,7 +3792,16 @@
 
   CMSConcMarkingTerminator* terminator() { return &_term; }
 
+  virtual void set_for_termination(int active_workers) {
+    terminator()->reset_for_reuse(active_workers);
+  }
+
   void work(int i);
+  bool should_yield() {
+    return    ConcurrentMarkSweepThread::should_yield()
+           && !_collector->foregroundGCIsActive()
+           && _asynch;
+  }
 
   virtual void coordinator_yield();  // stuff done by coordinator
   bool result() { return _result; }
@@ -3803,10 +3823,17 @@
   void bump_global_finger(HeapWord* f);
 };
 
+bool CMSConcMarkingTerminatorTerminator::should_exit_termination() {
+  assert(_task != NULL, "Error");
+  return _task->yielding();
+  // Note that we do not need the disjunct || _task->should_yield() above
+  // because we want terminating threads to yield only if the task
+  // is already in the midst of yielding, which happens only after at least one
+  // thread has yielded.
+}
+
 void CMSConcMarkingTerminator::yield() {
-  if (ConcurrentMarkSweepThread::should_yield() &&
-      !_collector->foregroundGCIsActive() &&
-      _yield) {
+  if (_task->should_yield()) {
     _task->yield();
   } else {
     ParallelTaskTerminator::yield();
@@ -4031,6 +4058,7 @@
 
 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure {
  private:
+  CMSConcMarkingTask* _task;
   MemRegion     _span;
   CMSBitMap*    _bit_map;
   CMSMarkStack* _overflow_stack;
@@ -4038,11 +4066,12 @@
  protected:
   DO_OOP_WORK_DEFN
  public:
-  Par_ConcMarkingClosure(CMSCollector* collector, OopTaskQueue* work_queue,
+  Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue,
                          CMSBitMap* bit_map, CMSMarkStack* overflow_stack,
                          CMSMarkStack* revisit_stack):
     Par_KlassRememberingOopClosure(collector, NULL, revisit_stack),
-    _span(_collector->_span),
+    _task(task),
+    _span(collector->_span),
     _work_queue(work_queue),
     _bit_map(bit_map),
     _overflow_stack(overflow_stack)
@@ -4051,6 +4080,11 @@
   virtual void do_oop(narrowOop* p);
   void trim_queue(size_t max);
   void handle_stack_overflow(HeapWord* lost);
+  void do_yield_check() {
+    if (_task->should_yield()) {
+      _task->yield();
+    }
+  }
 };
 
 // Grey object scanning during work stealing phase --
@@ -4094,6 +4128,7 @@
         handle_stack_overflow(addr);
       }
     } // Else, some other thread got there first
+    do_yield_check();
   }
 }
 
@@ -4109,6 +4144,7 @@
       assert(_span.contains((HeapWord*)new_oop), "Not in span");
       assert(new_oop->is_parsable(), "Should be parsable");
       new_oop->oop_iterate(this);  // do_oop() above
+      do_yield_check();
     }
   }
 }
@@ -4136,7 +4172,7 @@
   CMSMarkStack* ovflw = &(_collector->_markStack);
   CMSMarkStack* revisit = &(_collector->_revisitStack);
   int* seed = _collector->hash_seed(i);
-  Par_ConcMarkingClosure cl(_collector, work_q, bm, ovflw, revisit);
+  Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit);
   while (true) {
     cl.trim_queue(0);
     assert(work_q->size() == 0, "Should have been emptied above");
@@ -4149,9 +4185,11 @@
       assert(obj_to_scan->is_oop(), "Should be an oop");
       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
       obj_to_scan->oop_iterate(&cl);
-    } else if (terminator()->offer_termination()) {
+    } else if (terminator()->offer_termination(&_term_term)) {
       assert(work_q->size() == 0, "Impossible!");
       break;
+    } else if (yielding() || should_yield()) {
+      yield();
     }
   }
 }
@@ -4220,9 +4258,12 @@
   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
   CompactibleFreeListSpace* perm_space = _permGen->cmsSpace();
 
-  CMSConcMarkingTask tsk(this, cms_space, perm_space,
-                         asynch, num_workers /* number requested XXX */,
-                         conc_workers(), task_queues());
+  CMSConcMarkingTask tsk(this,
+                         cms_space,
+                         perm_space,
+                         asynch,
+                         conc_workers(),
+                         task_queues());
 
   // Since the actual number of workers we get may be different
   // from the number we requested above, do we need to do anything different
@@ -4326,6 +4367,10 @@
   verify_overflow_empty();
   _abort_preclean = false;
   if (CMSPrecleaningEnabled) {
+    // Precleaning is currently not MT but the reference processor
+    // may be set for MT.  Disable it temporarily here.
+    ReferenceProcessor* rp = ref_processor();
+    ReferenceProcessorMTProcMutator z(rp, false);
     _eden_chunk_index = 0;
     size_t used = get_eden_used();
     size_t capacity = get_eden_capacity();
@@ -4918,7 +4963,7 @@
       // dirtied since the first checkpoint in this GC cycle and prior to
       // the most recent young generation GC, minus those cleaned up by the
       // concurrent precleaning.
-      if (CMSParallelRemarkEnabled && ParallelGCThreads > 0) {
+      if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) {
         TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty);
         do_remark_parallel();
       } else {
@@ -5012,7 +5057,6 @@
 // Parallel remark task
 class CMSParRemarkTask: public AbstractGangTask {
   CMSCollector* _collector;
-  WorkGang*     _workers;
   int           _n_workers;
   CompactibleFreeListSpace* _cms_space;
   CompactibleFreeListSpace* _perm_space;
@@ -5025,21 +5069,21 @@
   CMSParRemarkTask(CMSCollector* collector,
                    CompactibleFreeListSpace* cms_space,
                    CompactibleFreeListSpace* perm_space,
-                   int n_workers, WorkGang* workers,
+                   int n_workers, FlexibleWorkGang* workers,
                    OopTaskQueueSet* task_queues):
     AbstractGangTask("Rescan roots and grey objects in parallel"),
     _collector(collector),
     _cms_space(cms_space), _perm_space(perm_space),
     _n_workers(n_workers),
-    _workers(workers),
     _task_queues(task_queues),
-    _term(workers->total_workers(), task_queues) { }
+    _term(n_workers, task_queues) { }
 
   OopTaskQueueSet* task_queues() { return _task_queues; }
 
   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
 
   ParallelTaskTerminator* terminator() { return &_term; }
+  int n_workers() { return _n_workers; }
 
   void work(int i);
 
@@ -5057,6 +5101,11 @@
   void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
 };
 
+// work_queue(i) is passed to the closure
+// Par_MarkRefsIntoAndScanClosure.  The "i" parameter
+// also is passed to do_dirty_card_rescan_tasks() and to
+// do_work_steal() to select the i-th task_queue.
+
 void CMSParRemarkTask::work(int i) {
   elapsedTimer _timer;
   ResourceMark rm;
@@ -5128,6 +5177,7 @@
 
   // Do the rescan tasks for each of the two spaces
   // (cms_space and perm_space) in turn.
+  // "i" is passed to select the "i-th" task_queue
   do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl);
   do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl);
   _timer.stop();
@@ -5150,6 +5200,7 @@
   }
 }
 
+// Note that parameter "i" is not used.
 void
 CMSParRemarkTask::do_young_space_rescan(int i,
   Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space,
@@ -5309,8 +5360,13 @@
     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
                                          (size_t)ParGCDesiredObjsFromOverflowList);
     // Now check if there's any work in the overflow list
+    // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
+    // only affects the number of attempts made to get work from the
+    // overflow list and does not affect the number of workers.  Just
+    // pass ParallelGCThreads so this behavior is unchanged.
     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
-                                                work_q)) {
+                                                work_q,
+                                                ParallelGCThreads)) {
       // found something in global overflow list;
       // not yet ready to go stealing work from others.
       // We'd like to assert(work_q->size() != 0, ...)
@@ -5367,11 +5423,12 @@
 // Merge the per-thread plab arrays into the global survivor chunk
 // array which will provide the partitioning of the survivor space
 // for CMS rescan.
-void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv) {
+void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv,
+                                              int no_of_gc_threads) {
   assert(_survivor_plab_array  != NULL, "Error");
   assert(_survivor_chunk_array != NULL, "Error");
   assert(_collectorState == FinalMarking, "Error");
-  for (uint j = 0; j < ParallelGCThreads; j++) {
+  for (int j = 0; j < no_of_gc_threads; j++) {
     _cursor[j] = 0;
   }
   HeapWord* top = surv->top();
@@ -5379,7 +5436,7 @@
   for (i = 0; i < _survivor_chunk_capacity; i++) {  // all sca entries
     HeapWord* min_val = top;          // Higher than any PLAB address
     uint      min_tid = 0;            // position of min_val this round
-    for (uint j = 0; j < ParallelGCThreads; j++) {
+    for (int j = 0; j < no_of_gc_threads; j++) {
       ChunkArray* cur_sca = &_survivor_plab_array[j];
       if (_cursor[j] == cur_sca->end()) {
         continue;
@@ -5413,7 +5470,7 @@
   // Verify that we used up all the recorded entries
   #ifdef ASSERT
     size_t total = 0;
-    for (uint j = 0; j < ParallelGCThreads; j++) {
+    for (int j = 0; j < no_of_gc_threads; j++) {
       assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant");
       total += _cursor[j];
     }
@@ -5448,13 +5505,15 @@
     // Each valid entry in [0, _eden_chunk_index) represents a task.
     size_t n_tasks = _eden_chunk_index + 1;
     assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error");
-    pst->set_par_threads(n_threads);
+    // Sets the condition for completion of the subtask (how many threads
+    // need to finish in order to be done).
+    pst->set_n_threads(n_threads);
     pst->set_n_tasks((int)n_tasks);
   }
 
   // Merge the survivor plab arrays into _survivor_chunk_array
   if (_survivor_plab_array != NULL) {
-    merge_survivor_plab_arrays(dng->from());
+    merge_survivor_plab_arrays(dng->from(), n_threads);
   } else {
     assert(_survivor_chunk_index == 0, "Error");
   }
@@ -5463,7 +5522,9 @@
   {
     SequentialSubTasksDone* pst = dng->to()->par_seq_tasks();
     assert(!pst->valid(), "Clobbering existing data?");
-    pst->set_par_threads(n_threads);
+    // Sets the condition for completion of the subtask (how many threads
+    // need to finish in order to be done).
+    pst->set_n_threads(n_threads);
     pst->set_n_tasks(1);
     assert(pst->valid(), "Error");
   }
@@ -5474,7 +5535,9 @@
     assert(!pst->valid(), "Clobbering existing data?");
     size_t n_tasks = _survivor_chunk_index + 1;
     assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error");
-    pst->set_par_threads(n_threads);
+    // Sets the condition for completion of the subtask (how many threads
+    // need to finish in order to be done).
+    pst->set_n_threads(n_threads);
     pst->set_n_tasks((int)n_tasks);
     assert(pst->valid(), "Error");
   }
@@ -5483,7 +5546,7 @@
 // Parallel version of remark
 void CMSCollector::do_remark_parallel() {
   GenCollectedHeap* gch = GenCollectedHeap::heap();
-  WorkGang* workers = gch->workers();
+  FlexibleWorkGang* workers = gch->workers();
   assert(workers != NULL, "Need parallel worker threads.");
   int n_workers = workers->total_workers();
   CompactibleFreeListSpace* cms_space  = _cmsGen->cmsSpace();
@@ -5636,13 +5699,11 @@
 ////////////////////////////////////////////////////////
 // Parallel Reference Processing Task Proxy Class
 ////////////////////////////////////////////////////////
-class CMSRefProcTaskProxy: public AbstractGangTask {
+class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues {
   typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask;
   CMSCollector*          _collector;
   CMSBitMap*             _mark_bit_map;
   const MemRegion        _span;
-  OopTaskQueueSet*       _task_queues;
-  ParallelTaskTerminator _term;
   ProcessTask&           _task;
 
 public:
@@ -5650,24 +5711,21 @@
                       CMSCollector*    collector,
                       const MemRegion& span,
                       CMSBitMap*       mark_bit_map,
-                      int              total_workers,
+                      AbstractWorkGang* workers,
                       OopTaskQueueSet* task_queues):
-    AbstractGangTask("Process referents by policy in parallel"),
+    AbstractGangTaskWOopQueues("Process referents by policy in parallel",
+      task_queues),
     _task(task),
-    _collector(collector), _span(span), _mark_bit_map(mark_bit_map),
-    _task_queues(task_queues),
-    _term(total_workers, task_queues)
+    _collector(collector), _span(span), _mark_bit_map(mark_bit_map)
     {
       assert(_collector->_span.equals(_span) && !_span.is_empty(),
              "Inconsistency in _span");
     }
 
-  OopTaskQueueSet* task_queues() { return _task_queues; }
+  OopTaskQueueSet* task_queues() { return queues(); }
 
   OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); }
 
-  ParallelTaskTerminator* terminator() { return &_term; }
-
   void do_work_steal(int i,
                      CMSParDrainMarkingStackClosure* drain,
                      CMSParKeepAliveClosure* keep_alive,
@@ -5739,8 +5797,13 @@
     size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
                                          (size_t)ParGCDesiredObjsFromOverflowList);
     // Now check if there's any work in the overflow list
+    // Passing ParallelGCThreads as the third parameter, no_of_gc_threads,
+    // only affects the number of attempts made to get work from the
+    // overflow list and does not affect the number of workers.  Just
+    // pass ParallelGCThreads so this behavior is unchanged.
     if (_collector->par_take_from_overflow_list(num_from_overflow_list,
-                                                work_q)) {
+                                                work_q,
+                                                ParallelGCThreads)) {
       // Found something in global overflow list;
       // not yet ready to go stealing work from others.
       // We'd like to assert(work_q->size() != 0, ...)
@@ -5773,13 +5836,12 @@
 void CMSRefProcTaskExecutor::execute(ProcessTask& task)
 {
   GenCollectedHeap* gch = GenCollectedHeap::heap();
-  WorkGang* workers = gch->workers();
+  FlexibleWorkGang* workers = gch->workers();
   assert(workers != NULL, "Need parallel worker threads.");
-  int n_workers = workers->total_workers();
   CMSRefProcTaskProxy rp_task(task, &_collector,
                               _collector.ref_processor()->span(),
                               _collector.markBitMap(),
-                              n_workers, _collector.task_queues());
+                              workers, _collector.task_queues());
   workers->run_task(&rp_task);
 }
 
@@ -5787,7 +5849,7 @@
 {
 
   GenCollectedHeap* gch = GenCollectedHeap::heap();
-  WorkGang* workers = gch->workers();
+  FlexibleWorkGang* workers = gch->workers();
   assert(workers != NULL, "Need parallel worker threads.");
   CMSRefEnqueueTaskProxy enq_task(task);
   workers->run_task(&enq_task);
@@ -5814,6 +5876,14 @@
   {
     TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty);
     if (rp->processing_is_mt()) {
+      // Set the degree of MT here.  If the discovery is done MT, there
+      // may have been a different number of threads doing the discovery
+      // and a different number of discovered lists may have Ref objects.
+      // That is OK as long as the Reference lists are balanced (see
+      // balance_all_queues() and balance_queues()).
+
+
+      rp->set_mt_degree(ParallelGCThreads);
       CMSRefProcTaskExecutor task_executor(*this);
       rp->process_discovered_references(&_is_alive_closure,
                                         &cmsKeepAliveClosure,
@@ -5874,6 +5944,7 @@
 
   rp->set_enqueuing_is_done(true);
   if (rp->processing_is_mt()) {
+    rp->balance_all_queues();
     CMSRefProcTaskExecutor task_executor(*this);
     rp->enqueue_discovered_references(&task_executor);
   } else {
@@ -8708,7 +8779,8 @@
 // similar changes might be needed.
 // CR 6797058 has been filed to consolidate the common code.
 bool CMSCollector::par_take_from_overflow_list(size_t num,
-                                               OopTaskQueue* work_q) {
+                                               OopTaskQueue* work_q,
+                                               int no_of_gc_threads) {
   assert(work_q->size() == 0, "First empty local work queue");
   assert(num < work_q->max_elems(), "Can't bite more than we can chew");
   if (_overflow_list == NULL) {
@@ -8717,7 +8789,9 @@
   // Grab the entire list; we'll put back a suffix
   oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
   Thread* tid = Thread::current();
-  size_t CMSOverflowSpinCount = (size_t)ParallelGCThreads;
+  // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was
+  // set to ParallelGCThreads.
+  size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads;
   size_t sleep_time_millis = MAX2((size_t)1, num/100);
   // If the list is busy, we spin for a short while,
   // sleeping between attempts to get the list.
@@ -8867,23 +8941,10 @@
 // failures where possible, thus, incrementally hardening the VM
 // in such low resource situations.
 void CMSCollector::preserve_mark_work(oop p, markOop m) {
-  if (_preserved_oop_stack == NULL) {
-    assert(_preserved_mark_stack == NULL,
-           "bijection with preserved_oop_stack");
-    // Allocate the stacks
-    _preserved_oop_stack  = new (ResourceObj::C_HEAP)
-      GrowableArray<oop>(PreserveMarkStackSize, true);
-    _preserved_mark_stack = new (ResourceObj::C_HEAP)
-      GrowableArray<markOop>(PreserveMarkStackSize, true);
-    if (_preserved_oop_stack == NULL || _preserved_mark_stack == NULL) {
-      vm_exit_out_of_memory(2* PreserveMarkStackSize * sizeof(oop) /* punt */,
-                            "Preserved Mark/Oop Stack for CMS (C-heap)");
-    }
-  }
-  _preserved_oop_stack->push(p);
-  _preserved_mark_stack->push(m);
+  _preserved_oop_stack.push(p);
+  _preserved_mark_stack.push(m);
   assert(m == p->mark(), "Mark word changed");
-  assert(_preserved_oop_stack->length() == _preserved_mark_stack->length(),
+  assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
          "bijection");
 }
 
@@ -8925,42 +8986,30 @@
 // effect on performance so great that this will
 // likely just be in the noise anyway.
 void CMSCollector::restore_preserved_marks_if_any() {
-  if (_preserved_oop_stack == NULL) {
-    assert(_preserved_mark_stack == NULL,
-           "bijection with preserved_oop_stack");
-    return;
-  }
-
   assert(SafepointSynchronize::is_at_safepoint(),
          "world should be stopped");
   assert(Thread::current()->is_ConcurrentGC_thread() ||
          Thread::current()->is_VM_thread(),
          "should be single-threaded");
-
-  int length = _preserved_oop_stack->length();
-  assert(_preserved_mark_stack->length() == length, "bijection");
-  for (int i = 0; i < length; i++) {
-    oop p = _preserved_oop_stack->at(i);
+  assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
+         "bijection");
+
+  while (!_preserved_oop_stack.is_empty()) {
+    oop p = _preserved_oop_stack.pop();
     assert(p->is_oop(), "Should be an oop");
     assert(_span.contains(p), "oop should be in _span");
     assert(p->mark() == markOopDesc::prototype(),
            "Set when taken from overflow list");
-    markOop m = _preserved_mark_stack->at(i);
+    markOop m = _preserved_mark_stack.pop();
     p->set_mark(m);
   }
-  _preserved_mark_stack->clear();
-  _preserved_oop_stack->clear();
-  assert(_preserved_mark_stack->is_empty() &&
-         _preserved_oop_stack->is_empty(),
+  assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(),
          "stacks were cleared above");
 }
 
 #ifndef PRODUCT
 bool CMSCollector::no_preserved_marks() const {
-  return (   (   _preserved_mark_stack == NULL
-              && _preserved_oop_stack == NULL)
-          || (   _preserved_mark_stack->is_empty()
-              && _preserved_oop_stack->is_empty()));
+  return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty();
 }
 #endif
 
@@ -9256,4 +9305,3 @@
              true /* recordGCEndTime */,
              true /* countCollection */ );
 }
-
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -537,8 +537,8 @@
   // The following array-pair keeps track of mark words
   // displaced for accomodating overflow list above.
   // This code will likely be revisited under RFE#4922830.
-  GrowableArray<oop>*     _preserved_oop_stack;
-  GrowableArray<markOop>* _preserved_mark_stack;
+  Stack<oop>     _preserved_oop_stack;
+  Stack<markOop> _preserved_mark_stack;
 
   int*             _hash_seed;
 
@@ -729,7 +729,9 @@
 
   // Support for marking stack overflow handling
   bool take_from_overflow_list(size_t num, CMSMarkStack* to_stack);
-  bool par_take_from_overflow_list(size_t num, OopTaskQueue* to_work_q);
+  bool par_take_from_overflow_list(size_t num,
+                                   OopTaskQueue* to_work_q,
+                                   int no_of_gc_threads);
   void push_on_overflow_list(oop p);
   void par_push_on_overflow_list(oop p);
   // the following is, obviously, not, in general, "MT-stable"
@@ -768,7 +770,7 @@
   void abortable_preclean(); // Preclean while looking for possible abort
   void initialize_sequential_subtasks_for_young_gen_rescan(int i);
   // Helper function for above; merge-sorts the per-thread plab samples
-  void merge_survivor_plab_arrays(ContiguousSpace* surv);
+  void merge_survivor_plab_arrays(ContiguousSpace* surv, int no_of_gc_threads);
   // Resets (i.e. clears) the per-thread plab sample vectors
   void reset_survivor_plab_arrays();
 
--- a/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepThread.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2006, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -123,24 +123,44 @@
   // or given timeout, whichever is earlier.
   void    wait_on_cms_lock(long t); // milliseconds
 
-  // The CMS thread will yield during the work portion of it's cycle
+  // The CMS thread will yield during the work portion of its cycle
   // only when requested to.  Both synchronous and asychronous requests
-  // are provided.  A synchronous request is used for young gen
-  // collections and direct allocations.  The requesting thread increments
-  // pending_yields at the beginning of an operation, and decrements it when
-  // the operation is completed.  The CMS thread yields when pending_yields
-  // is positive.  An asynchronous request is used by iCMS in the stop_icms()
-  // operation. A single yield satisfies the outstanding asynch yield requests.
-  // The requesting thread increments both pending_yields and pending_decrements.
-  // After yielding, the CMS thread decrements both by the amount in
-  // pending_decrements.
+  // are provided:
+  // (1) A synchronous request is used for young gen collections and
+  //     for direct allocations.  The requesting thread increments
+  //     _pending_yields at the beginning of an operation, and decrements
+  //     _pending_yields when that operation is completed.
+  //     In turn, the CMS thread yields when _pending_yields is positive,
+  //     and continues to yield until the value reverts to 0.
+  // (2) An asynchronous request, on the other hand, is used by iCMS
+  //     for the stop_icms() operation. A single yield satisfies all of
+  //     the outstanding asynch yield requests, of which there may
+  //     occasionally be several in close succession. To accomplish
+  //     this, an asynch-requesting thread atomically increments both
+  //     _pending_yields and _pending_decrements. An asynchr requesting
+  //     thread does not wait and "acknowledge" completion of an operation
+  //     and deregister the request, like the synchronous version described
+  //     above does. In turn, after yielding, the CMS thread decrements both
+  //     _pending_yields and _pending_decrements by the value seen in
+  //     _pending_decrements before the decrement.
+  //  NOTE: The above scheme is isomorphic to having two request counters,
+  //  one for async requests and one for sync requests, and for the CMS thread
+  //  to check the sum of the two counters to decide whether it should yield
+  //  and to clear only the async counter when it yields. However, it turns out
+  //  to be more efficient for CMS code to just check a single counter
+  //  _pending_yields that holds the sum (of both sync and async requests), and
+  //  a second counter _pending_decrements that only holds the async requests,
+  //  for greater efficiency, since in a typical CMS run, there are many more
+  //  pontential (i.e. static) yield points than there are actual
+  //  (i.e. dynamic) yields because of requests, which are few and far between.
+  //
   // Note that, while "_pending_yields >= _pending_decrements" is an invariant,
   // we cannot easily test that invariant, since the counters are manipulated via
   // atomic instructions without explicit locking and we cannot read
   // the two counters atomically together: one suggestion is to
   // use (for example) 16-bit counters so as to be able to read the
   // two counters atomically even on 32-bit platforms. Notice that
-  // the second assert in acknowledge_yield_request() does indeed
+  // the second assert in acknowledge_yield_request() below does indeed
   // check a form of the above invariant, albeit indirectly.
 
   static void increment_pending_yields()   {
@@ -152,6 +172,7 @@
     assert(_pending_yields >= 0, "can't be negative");
   }
   static void asynchronous_yield_request() {
+    assert(CMSIncrementalMode, "Currently only used w/iCMS");
     increment_pending_yields();
     Atomic::inc(&_pending_decrements);
     assert(_pending_decrements >= 0, "can't be negative");
@@ -159,6 +180,7 @@
   static void acknowledge_yield_request() {
     jint decrement = _pending_decrements;
     if (decrement > 0) {
+      assert(CMSIncrementalMode, "Currently only used w/iCMS");
       // Order important to preserve: _pending_yields >= _pending_decrements
       Atomic::add(-decrement, &_pending_decrements);
       Atomic::add(-decrement, &_pending_yields);
@@ -195,7 +217,7 @@
   }
 }
 
-// For scoped increment/decrement of yield requests
+// For scoped increment/decrement of (synchronous) yield requests
 class CMSSynchronousYieldRequest: public StackObj {
  public:
   CMSSynchronousYieldRequest() {
--- a/src/share/vm/gc_implementation/g1/concurrentMark.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/concurrentMark.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -278,15 +278,16 @@
   if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base);
 }
 
-void CMRegionStack::push(MemRegion mr) {
+void CMRegionStack::push_lock_free(MemRegion mr) {
   assert(mr.word_size() > 0, "Precondition");
   while (true) {
-    if (isFull()) {
+    jint index = _index;
+
+    if (index >= _capacity) {
       _overflow = true;
       return;
     }
     // Otherwise...
-    jint index = _index;
     jint next_index = index+1;
     jint res = Atomic::cmpxchg(next_index, &_index, index);
     if (res == index) {
@@ -297,19 +298,17 @@
   }
 }
 
-// Currently we do not call this at all. Normally we would call it
-// during the concurrent marking / remark phases but we now call
-// the lock-based version instead. But we might want to resurrect this
-// code in the future. So, we'll leave it here commented out.
-#if 0
-MemRegion CMRegionStack::pop() {
+// Lock-free pop of the region stack. Called during the concurrent
+// marking / remark phases. Should only be called in tandem with
+// other lock-free pops.
+MemRegion CMRegionStack::pop_lock_free() {
   while (true) {
-    // Otherwise...
     jint index = _index;
 
     if (index == 0) {
       return MemRegion();
     }
+    // Otherwise...
     jint next_index = index-1;
     jint res = Atomic::cmpxchg(next_index, &_index, index);
     if (res == index) {
@@ -326,7 +325,11 @@
     // Otherwise, we need to try again.
   }
 }
-#endif // 0
+
+#if 0
+// The routines that manipulate the region stack with a lock are
+// not currently used. They should be retained, however, as a
+// diagnostic aid.
 
 void CMRegionStack::push_with_lock(MemRegion mr) {
   assert(mr.word_size() > 0, "Precondition");
@@ -361,6 +364,7 @@
     }
   }
 }
+#endif
 
 bool CMRegionStack::invalidate_entries_into_cset() {
   bool result = false;
@@ -583,10 +587,13 @@
 #endif
 
     guarantee(parallel_marking_threads() > 0, "peace of mind");
-    _parallel_workers = new WorkGang("G1 Parallel Marking Threads",
-                                     (int) parallel_marking_threads(), false, true);
-    if (_parallel_workers == NULL)
+    _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads",
+         (int) _parallel_marking_threads, false, true);
+    if (_parallel_workers == NULL) {
       vm_exit_during_initialization("Failed necessary allocation.");
+    } else {
+      _parallel_workers->initialize_workers();
+    }
   }
 
   // so that the call below can read a sensible value
@@ -645,8 +652,9 @@
   // We do reset all of them, since different phases will use
   // different number of active threads. So, it's easiest to have all
   // of them ready.
-  for (int i = 0; i < (int) _max_task_num; ++i)
+  for (int i = 0; i < (int) _max_task_num; ++i) {
     _tasks[i]->reset(_nextMarkBitMap);
+  }
 
   // we need this to make sure that the flag is on during the evac
   // pause with initial mark piggy-backed
@@ -985,7 +993,7 @@
                              "below the finger, pushing it",
                              mr.start(), mr.end());
 
-    if (!region_stack_push(mr)) {
+    if (!region_stack_push_lock_free(mr)) {
       if (verbose_low())
         gclog_or_tty->print_cr("[global] region stack has overflown.");
     }
@@ -1451,7 +1459,7 @@
                                   _bm, _g1h->concurrent_mark(),
                                   _region_bm, _card_bm);
     calccl.no_yield();
-    if (ParallelGCThreads > 0) {
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       _g1h->heap_region_par_iterate_chunked(&calccl, i,
                                             HeapRegion::FinalCountClaimValue);
     } else {
@@ -1531,7 +1539,7 @@
     G1NoteEndOfConcMarkClosure g1_note_end(_g1h,
                                            &_par_cleanup_thread_state[i]->list,
                                            i);
-    if (ParallelGCThreads > 0) {
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       _g1h->heap_region_par_iterate_chunked(&g1_note_end, i,
                                             HeapRegion::NoteEndClaimValue);
     } else {
@@ -1575,7 +1583,7 @@
   {}
 
   void work(int i) {
-    if (ParallelGCThreads > 0) {
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       _g1rs->scrub_par(_region_bm, _card_bm, i,
                        HeapRegion::ScrubRemSetClaimValue);
     } else {
@@ -1647,7 +1655,7 @@
   // Do counting once more with the world stopped for good measure.
   G1ParFinalCountTask g1_par_count_task(g1h, nextMarkBitMap(),
                                         &_region_bm, &_card_bm);
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     assert(g1h->check_heap_region_claim_values(
                                                HeapRegion::InitialClaimValue),
            "sanity check");
@@ -1695,7 +1703,7 @@
   // Note end of marking in all heap regions.
   double note_end_start = os::elapsedTime();
   G1ParNoteEndTask g1_par_note_end_task(g1h, _par_cleanup_thread_state);
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     int n_workers = g1h->workers()->total_workers();
     g1h->set_par_threads(n_workers);
     g1h->workers()->run_task(&g1_par_note_end_task);
@@ -1720,7 +1728,7 @@
   if (G1ScrubRemSets) {
     double rs_scrub_start = os::elapsedTime();
     G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm);
-    if (ParallelGCThreads > 0) {
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       int n_workers = g1h->workers()->total_workers();
       g1h->set_par_threads(n_workers);
       g1h->workers()->run_task(&g1_par_scrub_rs_task);
@@ -1934,7 +1942,7 @@
 
   g1h->ensure_parsability(false);
 
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     G1CollectedHeap::StrongRootsScope srs(g1h);
     // this is remark, so we'll use up all available threads
     int active_workers = ParallelGCThreads;
@@ -2330,6 +2338,39 @@
   return NULL;
 }
 
+bool ConcurrentMark::invalidate_aborted_regions_in_cset() {
+  bool result = false;
+  for (int i = 0; i < (int)_max_task_num; ++i) {
+    CMTask* the_task = _tasks[i];
+    MemRegion mr = the_task->aborted_region();
+    if (mr.start() != NULL) {
+      assert(mr.end() != NULL, "invariant");
+      assert(mr.word_size() > 0, "invariant");
+      HeapRegion* hr = _g1h->heap_region_containing(mr.start());
+      assert(hr != NULL, "invariant");
+      if (hr->in_collection_set()) {
+        // The region points into the collection set
+        the_task->set_aborted_region(MemRegion());
+        result = true;
+      }
+    }
+  }
+  return result;
+}
+
+bool ConcurrentMark::has_aborted_regions() {
+  for (int i = 0; i < (int)_max_task_num; ++i) {
+    CMTask* the_task = _tasks[i];
+    MemRegion mr = the_task->aborted_region();
+    if (mr.start() != NULL) {
+      assert(mr.end() != NULL, "invariant");
+      assert(mr.word_size() > 0, "invariant");
+      return true;
+    }
+  }
+  return false;
+}
+
 void ConcurrentMark::oops_do(OopClosure* cl) {
   if (_markStack.size() > 0 && verbose_low())
     gclog_or_tty->print_cr("[global] scanning the global marking stack, "
@@ -2348,13 +2389,22 @@
     queue->oops_do(cl);
   }
 
-  // finally, invalidate any entries that in the region stack that
+  // Invalidate any entries, that are in the region stack, that
   // point into the collection set
   if (_regionStack.invalidate_entries_into_cset()) {
     // otherwise, any gray objects copied during the evacuation pause
     // might not be visited.
     assert(_should_gray_objects, "invariant");
   }
+
+  // Invalidate any aborted regions, recorded in the individual CM
+  // tasks, that point into the collection set.
+  if (invalidate_aborted_regions_in_cset()) {
+    // otherwise, any gray objects copied during the evacuation pause
+    // might not be visited.
+    assert(_should_gray_objects, "invariant");
+  }
+
 }
 
 void ConcurrentMark::clear_marking_state() {
@@ -2635,7 +2685,7 @@
   // irrespective whether all collection set regions are below the
   // finger, if the region stack is not empty. This is expected to be
   // a rare case, so I don't think it's necessary to be smarted about it.
-  if (!region_stack_empty())
+  if (!region_stack_empty() || has_aborted_regions())
     _should_gray_objects = true;
 }
 
@@ -2654,8 +2704,10 @@
   _nextMarkBitMap->clearAll();
   // Empty mark stack
   clear_marking_state();
-  for (int i = 0; i < (int)_max_task_num; ++i)
+  for (int i = 0; i < (int)_max_task_num; ++i) {
     _tasks[i]->clear_region_fields();
+    _tasks[i]->clear_aborted_region();
+  }
   _has_aborted = true;
 
   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
@@ -2933,6 +2985,7 @@
 
   _nextMarkBitMap                = nextMarkBitMap;
   clear_region_fields();
+  clear_aborted_region();
 
   _calls                         = 0;
   _elapsed_time_ms               = 0.0;
@@ -3369,14 +3422,14 @@
 
   CMObjectClosure oc(this);
   SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set();
-  if (ParallelGCThreads > 0)
+  if (G1CollectedHeap::use_parallel_gc_threads())
     satb_mq_set.set_par_closure(_task_id, &oc);
   else
     satb_mq_set.set_closure(&oc);
 
   // This keeps claiming and applying the closure to completed buffers
   // until we run out of buffers or we need to abort.
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     while (!has_aborted() &&
            satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) {
       if (_cm->verbose_medium())
@@ -3396,7 +3449,7 @@
 
   if (!concurrent() && !has_aborted()) {
     // We should only do this during remark.
-    if (ParallelGCThreads > 0)
+    if (G1CollectedHeap::use_parallel_gc_threads())
       satb_mq_set.par_iterate_closure_all_threads(_task_id);
     else
       satb_mq_set.iterate_closure_all_threads();
@@ -3408,7 +3461,7 @@
          concurrent() ||
          satb_mq_set.completed_buffers_num() == 0, "invariant");
 
-  if (ParallelGCThreads > 0)
+  if (G1CollectedHeap::use_parallel_gc_threads())
     satb_mq_set.set_par_closure(_task_id, NULL);
   else
     satb_mq_set.set_closure(NULL);
@@ -3425,20 +3478,32 @@
   assert(_region_finger == NULL,
          "it should be NULL when we're not scanning a region");
 
-  if (!_cm->region_stack_empty()) {
+  if (!_cm->region_stack_empty() || !_aborted_region.is_empty()) {
     if (_cm->verbose_low())
       gclog_or_tty->print_cr("[%d] draining region stack, size = %d",
                              _task_id, _cm->region_stack_size());
 
-    MemRegion mr = _cm->region_stack_pop_with_lock();
-    // it returns MemRegion() if the pop fails
-    statsOnly(if (mr.start() != NULL) ++_region_stack_pops );
+    MemRegion mr;
+
+    if (!_aborted_region.is_empty()) {
+      mr = _aborted_region;
+      _aborted_region = MemRegion();
+
+      if (_cm->verbose_low())
+        gclog_or_tty->print_cr("[%d] scanning aborted region [ " PTR_FORMAT ", " PTR_FORMAT " )",
+                             _task_id, mr.start(), mr.end());
+    } else {
+      mr = _cm->region_stack_pop_lock_free();
+      // it returns MemRegion() if the pop fails
+      statsOnly(if (mr.start() != NULL) ++_region_stack_pops );
+    }
 
     while (mr.start() != NULL) {
       if (_cm->verbose_medium())
         gclog_or_tty->print_cr("[%d] we are scanning region "
                                "["PTR_FORMAT", "PTR_FORMAT")",
                                _task_id, mr.start(), mr.end());
+
       assert(mr.end() <= _cm->finger(),
              "otherwise the region shouldn't be on the stack");
       assert(!mr.is_empty(), "Only non-empty regions live on the region stack");
@@ -3451,7 +3516,7 @@
         if (has_aborted())
           mr = MemRegion();
         else {
-          mr = _cm->region_stack_pop_with_lock();
+          mr = _cm->region_stack_pop_lock_free();
           // it returns MemRegion() if the pop fails
           statsOnly(if (mr.start() != NULL) ++_region_stack_pops );
         }
@@ -3465,6 +3530,10 @@
         // have definitely set _region_finger to something non-null.
         assert(_region_finger != NULL, "invariant");
 
+        // Make sure that any previously aborted region has been
+        // cleared.
+        assert(_aborted_region.is_empty(), "aborted region not cleared");
+
         // The iteration was actually aborted. So now _region_finger
         // points to the address of the object we last scanned. If we
         // leave it there, when we restart this task, we will rescan
@@ -3477,14 +3546,14 @@
 
         if (!newRegion.is_empty()) {
           if (_cm->verbose_low()) {
-            gclog_or_tty->print_cr("[%d] pushing unscanned region"
-                                   "[" PTR_FORMAT "," PTR_FORMAT ") on region stack",
+            gclog_or_tty->print_cr("[%d] recording unscanned region"
+                                   "[" PTR_FORMAT "," PTR_FORMAT ") in CMTask",
                                    _task_id,
                                    newRegion.start(), newRegion.end());
           }
-          // Now push the part of the region we didn't scan on the
-          // region stack to make sure a task scans it later.
-          _cm->region_stack_push_with_lock(newRegion);
+          // Now record the part of the region we didn't scan to
+          // make sure this task scans it later.
+          _aborted_region = newRegion;
         }
         // break from while
         mr = MemRegion();
@@ -3654,6 +3723,8 @@
 
   assert(concurrent() || _cm->region_stack_empty(),
          "the region stack should have been cleared before remark");
+  assert(concurrent() || !_cm->has_aborted_regions(),
+         "aborted regions should have been cleared before remark");
   assert(_region_finger == NULL,
          "this should be non-null only when a region is being scanned");
 
@@ -3943,6 +4014,7 @@
       // that, if a condition is false, we can immediately find out
       // which one.
       guarantee(_cm->out_of_regions(), "only way to reach here");
+      guarantee(_aborted_region.is_empty(), "only way to reach here");
       guarantee(_cm->region_stack_empty(), "only way to reach here");
       guarantee(_cm->mark_stack_empty(), "only way to reach here");
       guarantee(_task_queue->size() == 0, "only way to reach here");
@@ -4042,7 +4114,8 @@
     _nextMarkBitMap(NULL), _hash_seed(17),
     _task_queue(task_queue),
     _task_queues(task_queues),
-    _oop_closure(NULL) {
+    _oop_closure(NULL),
+    _aborted_region(MemRegion()) {
   guarantee(task_queue != NULL, "invariant");
   guarantee(task_queues != NULL, "invariant");
 
--- a/src/share/vm/gc_implementation/g1/concurrentMark.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/concurrentMark.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -250,21 +250,23 @@
 
   // This is lock-free; assumes that it will only be called in parallel
   // with other "push" operations (no pops).
-  void push(MemRegion mr);
-
-#if 0
-  // This is currently not used. See the comment in the .cpp file.
+  void push_lock_free(MemRegion mr);
 
   // Lock-free; assumes that it will only be called in parallel
   // with other "pop" operations (no pushes).
-  MemRegion pop();
-#endif // 0
+  MemRegion pop_lock_free();
+
+#if 0
+  // The routines that manipulate the region stack with a lock are
+  // not currently used. They should be retained, however, as a
+  // diagnostic aid.
 
   // These two are the implementations that use a lock. They can be
   // called concurrently with each other but they should not be called
   // concurrently with the lock-free versions (push() / pop()).
   void push_with_lock(MemRegion mr);
   MemRegion pop_with_lock();
+#endif
 
   bool isEmpty()    { return _index == 0; }
   bool isFull()     { return _index == _capacity; }
@@ -398,6 +400,7 @@
   volatile bool           _concurrent;
   // set at the end of a Full GC so that marking aborts
   volatile bool           _has_aborted;
+
   // used when remark aborts due to an overflow to indicate that
   // another concurrent marking phase should start
   volatile bool           _restart_for_overflow;
@@ -548,23 +551,30 @@
   bool mark_stack_overflow()            { return _markStack.overflow(); }
   bool mark_stack_empty()               { return _markStack.isEmpty(); }
 
-  // Manipulation of the region stack
-  bool region_stack_push(MemRegion mr) {
+  // (Lock-free) Manipulation of the region stack
+  bool region_stack_push_lock_free(MemRegion mr) {
     // Currently we only call the lock-free version during evacuation
     // pauses.
     assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
 
-    _regionStack.push(mr);
+    _regionStack.push_lock_free(mr);
     if (_regionStack.overflow()) {
       set_has_overflown();
       return false;
     }
     return true;
   }
+
+  // Lock-free version of region-stack pop. Should only be
+  // called in tandem with other lock-free pops.
+  MemRegion region_stack_pop_lock_free() {
+    return _regionStack.pop_lock_free();
+  }
+
 #if 0
-  // Currently this is not used. See the comment in the .cpp file.
-  MemRegion region_stack_pop() { return _regionStack.pop(); }
-#endif // 0
+  // The routines that manipulate the region stack with a lock are
+  // not currently used. They should be retained, however, as a
+  // diagnostic aid.
 
   bool region_stack_push_with_lock(MemRegion mr) {
     // Currently we only call the lock-based version during either
@@ -579,6 +589,7 @@
     }
     return true;
   }
+
   MemRegion region_stack_pop_with_lock() {
     // Currently we only call the lock-based version during either
     // concurrent marking or remark.
@@ -587,11 +598,21 @@
 
     return _regionStack.pop_with_lock();
   }
+#endif
 
   int region_stack_size()               { return _regionStack.size(); }
   bool region_stack_overflow()          { return _regionStack.overflow(); }
   bool region_stack_empty()             { return _regionStack.isEmpty(); }
 
+  // Iterate over any regions that were aborted while draining the
+  // region stack (any such regions are saved in the corresponding
+  // CMTask) and invalidate (i.e. assign to the empty MemRegion())
+  // any regions that point into the collection set.
+  bool invalidate_aborted_regions_in_cset();
+
+  // Returns true if there are any aborted memory regions.
+  bool has_aborted_regions();
+
   bool concurrent_marking_in_progress() {
     return _concurrent_marking_in_progress;
   }
@@ -856,6 +877,15 @@
   // stack.
   HeapWord*                   _region_finger;
 
+  // If we abort while scanning a region we record the remaining
+  // unscanned portion and check this field when marking restarts.
+  // This avoids having to push on the region stack while other
+  // marking threads may still be popping regions.
+  // If we were to push the unscanned portion directly to the
+  // region stack then we would need to using locking versions
+  // of the push and pop operations.
+  MemRegion                   _aborted_region;
+
   // the number of words this task has scanned
   size_t                      _words_scanned;
   // When _words_scanned reaches this limit, the regular clock is
@@ -1012,6 +1042,15 @@
   void clear_has_aborted()      { _has_aborted = false; }
   bool claimed() { return _claimed; }
 
+  // Support routines for the partially scanned region that may be
+  // recorded as a result of aborting while draining the CMRegionStack
+  MemRegion aborted_region()    { return _aborted_region; }
+  void set_aborted_region(MemRegion mr)
+                                { _aborted_region = mr; }
+
+  // Clears any recorded partially scanned region
+  void clear_aborted_region()   { set_aborted_region(MemRegion()); }
+
   void set_oop_closure(OopClosure* oop_closure) {
     _oop_closure = oop_closure;
   }
--- a/src/share/vm/gc_implementation/g1/concurrentMarkThread.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/concurrentMarkThread.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -303,9 +303,10 @@
 }
 
 void ConcurrentMarkThread::sleepBeforeNextCycle() {
-  clear_in_progress();
   // We join here because we don't want to do the "shouldConcurrentMark()"
   // below while the world is otherwise stopped.
+  assert(!in_progress(), "should have been cleared");
+
   MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag);
   while (!started()) {
     CGC_lock->wait(Mutex::_no_safepoint_check_flag);
--- a/src/share/vm/gc_implementation/g1/concurrentMarkThread.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/concurrentMarkThread.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -69,12 +69,12 @@
 
   ConcurrentMark* cm()     { return _cm; }
 
-  void set_started()       { _started = true;  }
-  void clear_started()     { _started = false; }
+  void set_started()       { assert(!_in_progress, "cycle in progress"); _started = true;  }
+  void clear_started()     { assert(_in_progress, "must be starting a cycle"); _started = false; }
   bool started()           { return _started;  }
 
-  void set_in_progress()   { _in_progress = true;  }
-  void clear_in_progress() { _in_progress = false; }
+  void set_in_progress()   { assert(_started, "must be starting a cycle"); _in_progress = true;  }
+  void clear_in_progress() { assert(!_started, "must not be starting a new cycle"); _in_progress = false; }
   bool in_progress()       { return _in_progress;  }
 
   // This flag returns true from the moment a marking cycle is
--- a/src/share/vm/gc_implementation/g1/dirtyCardQueue.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/dirtyCardQueue.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -37,11 +37,10 @@
 class DirtyCardQueue: public PtrQueue {
 public:
   DirtyCardQueue(PtrQueueSet* qset_, bool perm = false) :
-    PtrQueue(qset_, perm)
-  {
-    // Dirty card queues are always active.
-    _active = true;
-  }
+    // Dirty card queues are always active, so we create them with their
+    // active field set to true.
+    PtrQueue(qset_, perm, true /* active */) { }
+
   // Apply the closure to all elements, and reset the index to make the
   // buffer empty.  If a closure application returns "false", return
   // "false" immediately, halting the iteration.  If "consume" is true,
--- a/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -961,7 +961,8 @@
     }
 
     // Rebuild remembered sets of all regions.
-    if (ParallelGCThreads > 0) {
+
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       ParRebuildRSTask rebuild_rs_task(this);
       assert(check_heap_region_claim_values(
              HeapRegion::InitialClaimValue), "sanity check");
@@ -1784,6 +1785,14 @@
 
   _full_collections_completed += 1;
 
+  // We need to clear the "in_progress" flag in the CM thread before
+  // we wake up any waiters (especially when ExplicitInvokesConcurrent
+  // is set) so that if a waiter requests another System.gc() it doesn't
+  // incorrectly see that a marking cyle is still in progress.
+  if (outer) {
+    _cmThread->clear_in_progress();
+  }
+
   // This notify_all() will ensure that a thread that called
   // System.gc() with (with ExplicitGCInvokesConcurrent set or not)
   // and it's waiting for a full GC to finish will be woken up. It is
@@ -1960,7 +1969,7 @@
                                                  int worker,
                                                  jint claim_value) {
   const size_t regions = n_regions();
-  const size_t worker_num = (ParallelGCThreads > 0 ? ParallelGCThreads : 1);
+  const size_t worker_num = (G1CollectedHeap::use_parallel_gc_threads() ? ParallelGCThreads : 1);
   // try to spread out the starting points of the workers
   const size_t start_index = regions / worker_num * (size_t) worker;
 
@@ -2527,7 +2536,7 @@
 }
 
 void G1CollectedHeap::print_gc_threads_on(outputStream* st) const {
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     workers()->print_worker_threads_on(st);
   }
 
@@ -2543,7 +2552,7 @@
 }
 
 void G1CollectedHeap::gc_threads_do(ThreadClosure* tc) const {
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     workers()->threads_do(tc);
   }
   tc->do_thread(_cmThread);
@@ -3083,7 +3092,7 @@
   if (r != NULL) {
     r_used = r->used();
 
-    if (ParallelGCThreads > 0) {
+    if (G1CollectedHeap::use_parallel_gc_threads()) {
       // need to take the lock to guard against two threads calling
       // get_gc_alloc_region concurrently (very unlikely but...)
       MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag);
@@ -4182,6 +4191,8 @@
 
 // *** Common G1 Evacuation Stuff
 
+// This method is run in a GC worker.
+
 void
 G1CollectedHeap::
 g1_process_strong_roots(bool collecting_perm_gen,
@@ -4259,7 +4270,7 @@
 };
 
 void G1CollectedHeap::save_marks() {
-  if (ParallelGCThreads == 0) {
+  if (!CollectedHeap::use_parallel_gc_threads()) {
     SaveMarksClosure sm;
     heap_region_iterate(&sm);
   }
@@ -4284,7 +4295,7 @@
 
   assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
   double start_par = os::elapsedTime();
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     // The individual threads will set their evac-failure closures.
     StrongRootsScope srs(this);
     if (ParallelGCVerbose) G1ParScanThreadState::print_termination_stats_hdr();
--- a/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -656,6 +656,9 @@
   bool _unclean_regions_coming;
 
 public:
+
+  SubTasksDone* process_strong_tasks() { return _process_strong_tasks; }
+
   void set_refine_cte_cl_concurrency(bool concurrent);
 
   RefToScanQueue *task_queue(int i) const;
@@ -684,7 +687,7 @@
 
   void set_par_threads(int t) {
     SharedHeap::set_par_threads(t);
-    _process_strong_tasks->set_par_threads(t);
+    _process_strong_tasks->set_n_threads(t);
   }
 
   virtual CollectedHeap::Name kind() const {
@@ -1688,8 +1691,8 @@
     ref = new_ref;
   }
 
-  int refs_to_scan()            { return refs()->size(); }
-  int overflowed_refs_to_scan() { return refs()->overflow_stack()->length(); }
+  int refs_to_scan()            { return (int)refs()->size(); }
+  int overflowed_refs_to_scan() { return (int)refs()->overflow_stack()->size(); }
 
   template <class T> void update_rs(HeapRegion* from, T* p, int tid) {
     if (G1DeferredRSUpdate) {
--- a/src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -72,7 +72,10 @@
 // </NEW PREDICTION>
 
 G1CollectorPolicy::G1CollectorPolicy() :
-  _parallel_gc_threads((ParallelGCThreads > 0) ? ParallelGCThreads : 1),
+  _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
+    ? ParallelGCThreads : 1),
+
+
   _n_pauses(0),
   _recent_CH_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
   _recent_G1_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
@@ -1073,7 +1076,7 @@
 }
 
 double G1CollectorPolicy::avg_value (double* data) {
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     double ret = 0.0;
     for (uint i = 0; i < ParallelGCThreads; ++i)
       ret += data[i];
@@ -1084,7 +1087,7 @@
 }
 
 double G1CollectorPolicy::max_value (double* data) {
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     double ret = data[0];
     for (uint i = 1; i < ParallelGCThreads; ++i)
       if (data[i] > ret)
@@ -1096,7 +1099,7 @@
 }
 
 double G1CollectorPolicy::sum_of_values (double* data) {
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     double sum = 0.0;
     for (uint i = 0; i < ParallelGCThreads; i++)
       sum += data[i];
@@ -1110,7 +1113,7 @@
                                    double* data2) {
   double ret = data1[0] + data2[0];
 
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     for (uint i = 1; i < ParallelGCThreads; ++i) {
       double data = data1[i] + data2[i];
       if (data > ret)
@@ -1126,7 +1129,7 @@
 void G1CollectorPolicy::record_collection_pause_end() {
   double end_time_sec = os::elapsedTime();
   double elapsed_ms = _last_pause_time_ms;
-  bool parallel = ParallelGCThreads > 0;
+  bool parallel = G1CollectedHeap::use_parallel_gc_threads();
   double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0;
   size_t rs_size =
     _cur_collection_pause_used_regions_at_start - collection_set_size();
@@ -1941,7 +1944,7 @@
       // Further, we're now always doing parallel collection.  But I'm still
       // leaving this here as a placeholder for a more precise assertion later.
       // (DLD, 10/05.)
-      assert((true || ParallelGCThreads > 0) ||
+      assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
              _g1->evacuation_failed() ||
              recent_survival_rate <= 1.0, "Or bad frac");
       return recent_survival_rate;
@@ -1961,7 +1964,7 @@
     // Further, we're now always doing parallel collection.  But I'm still
     // leaving this here as a placeholder for a more precise assertion later.
     // (DLD, 10/05.)
-    assert((true || ParallelGCThreads > 0) ||
+    assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
            last_survival_rate <= 1.0, "Or bad frac");
     return last_survival_rate;
   } else {
@@ -2121,7 +2124,7 @@
 }
 
 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
-  bool parallel = ParallelGCThreads > 0;
+  bool parallel = G1CollectedHeap::use_parallel_gc_threads();
   MainBodySummary*    body_summary = summary->main_body_summary();
   if (summary->get_total_seq()->num() > 0) {
     print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
@@ -2559,7 +2562,7 @@
     gclog_or_tty->print_cr("  clear marked regions + work1: %8.3f ms.",
                   (clear_marked_end - start)*1000.0);
   }
-  if (ParallelGCThreads > 0) {
+  if (G1CollectedHeap::use_parallel_gc_threads()) {
     const size_t OverpartitionFactor = 4;
     const size_t MinWorkUnit = 8;
     const size_t WorkUnit =
--- a/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/g1MarkSweep.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -101,22 +101,6 @@
   GenMarkSweep::_preserved_count_max = 0;
   GenMarkSweep::_preserved_marks = NULL;
   GenMarkSweep::_preserved_count = 0;
-  GenMarkSweep::_preserved_mark_stack = NULL;
-  GenMarkSweep::_preserved_oop_stack = NULL;
-
-  GenMarkSweep::_marking_stack =
-    new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
-  GenMarkSweep::_objarray_stack =
-    new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
-
-  int size = SystemDictionary::number_of_classes() * 2;
-  GenMarkSweep::_revisit_klass_stack =
-    new (ResourceObj::C_HEAP) GrowableArray<Klass*>(size, true);
-  // (#klass/k)^2 for k ~ 10 appears a better fit, but this will have to do
-  // for now until we have a chance to work out a more optimal setting.
-  GenMarkSweep::_revisit_mdo_stack =
-    new (ResourceObj::C_HEAP) GrowableArray<DataLayout*>(size*2, true);
-
 }
 
 void G1MarkSweep::mark_sweep_phase1(bool& marked_for_unloading,
@@ -145,7 +129,7 @@
 
   // Follow system dictionary roots and unload classes
   bool purged_class = SystemDictionary::do_unloading(&GenMarkSweep::is_alive);
-  assert(GenMarkSweep::_marking_stack->is_empty(),
+  assert(GenMarkSweep::_marking_stack.is_empty(),
          "stack should be empty by now");
 
   // Follow code cache roots (has to be done after system dictionary,
@@ -157,19 +141,19 @@
 
   // Update subklass/sibling/implementor links of live klasses
   GenMarkSweep::follow_weak_klass_links();
-  assert(GenMarkSweep::_marking_stack->is_empty(),
+  assert(GenMarkSweep::_marking_stack.is_empty(),
          "stack should be empty by now");
 
   // Visit memoized MDO's and clear any unmarked weak refs
   GenMarkSweep::follow_mdo_weak_refs();
-  assert(GenMarkSweep::_marking_stack->is_empty(), "just drained");
+  assert(GenMarkSweep::_marking_stack.is_empty(), "just drained");
 
 
   // Visit symbol and interned string tables and delete unmarked oops
   SymbolTable::unlink(&GenMarkSweep::is_alive);
   StringTable::unlink(&GenMarkSweep::is_alive);
 
-  assert(GenMarkSweep::_marking_stack->is_empty(),
+  assert(GenMarkSweep::_marking_stack.is_empty(),
          "stack should be empty by now");
 }
 
--- a/src/share/vm/gc_implementation/g1/g1RemSet.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/g1RemSet.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -523,7 +523,7 @@
   assert(!_traversal_in_progress, "Invariant between iterations.");
   set_traversal(true);
   if (ParallelGCThreads > 0) {
-    _seq_task->set_par_threads((int)n_workers());
+    _seq_task->set_n_threads((int)n_workers());
   }
   guarantee( _cards_scanned == NULL, "invariant" );
   _cards_scanned = NEW_C_HEAP_ARRAY(size_t, n_workers());
--- a/src/share/vm/gc_implementation/g1/ptrQueue.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/ptrQueue.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -89,6 +89,10 @@
     return _buf == NULL ? 0 : _sz - _index;
   }
 
+  bool is_empty() {
+    return _buf == NULL || _sz == _index;
+  }
+
   // Set the "active" property of the queue to "b".  An enqueue to an
   // inactive thread is a no-op.  Setting a queue to inactive resets its
   // log to the empty state.
--- a/src/share/vm/gc_implementation/g1/satbQueue.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/g1/satbQueue.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -29,7 +29,12 @@
 class ObjPtrQueue: public PtrQueue {
 public:
   ObjPtrQueue(PtrQueueSet* qset_, bool perm = false) :
-    PtrQueue(qset_, perm, qset_->is_active()) { }
+    // SATB queues are only active during marking cycles. We create
+    // them with their active field set to false. If a thread is
+    // created during a cycle and its SATB queue needs to be activated
+    // before the thread starts running, we'll need to set its active
+    // field to true. This is done in JavaThread::initialize_queues().
+    PtrQueue(qset_, perm, false /* active */) { }
   // Apply the closure to all elements, and reset the index to make the
   // buffer empty.
   void apply_closure(ObjectClosure* cl);
--- a/src/share/vm/gc_implementation/includeDB_gc_concurrentMarkSweep	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/includeDB_gc_concurrentMarkSweep	Fri Oct 08 09:29:09 2010 -0700
@@ -171,6 +171,7 @@
 concurrentMarkSweepGeneration.hpp       generationCounters.hpp
 concurrentMarkSweepGeneration.hpp       memoryService.hpp
 concurrentMarkSweepGeneration.hpp       mutexLocker.hpp
+concurrentMarkSweepGeneration.hpp       stack.inline.hpp
 concurrentMarkSweepGeneration.hpp       taskqueue.hpp
 concurrentMarkSweepGeneration.hpp       virtualspace.hpp
 concurrentMarkSweepGeneration.hpp       yieldingWorkgroup.hpp
--- a/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/includeDB_gc_parallelScavenge	Fri Oct 08 09:29:09 2010 -0700
@@ -187,9 +187,11 @@
 psCompactionManager.cpp                 psParallelCompact.hpp
 psCompactionManager.cpp                 psCompactionManager.hpp
 psCompactionManager.cpp                 psOldGen.hpp
+psCompactionManager.cpp                 stack.inline.hpp
 psCompactionManager.cpp                 systemDictionary.hpp
 
 psCompactionManager.hpp                 allocation.hpp
+psCompactionManager.hpp                 stack.hpp
 psCompactionManager.hpp                 taskqueue.hpp
 
 psCompactionManager.inline.hpp		psCompactionManager.hpp
@@ -233,12 +235,14 @@
 psMarkSweep.cpp                         referenceProcessor.hpp
 psMarkSweep.cpp                         safepoint.hpp
 psMarkSweep.cpp                         spaceDecorator.hpp
+psMarkSweep.cpp                         stack.inline.hpp
 psMarkSweep.cpp                         symbolTable.hpp
 psMarkSweep.cpp                         systemDictionary.hpp
 psMarkSweep.cpp                         vmThread.hpp
 
 psMarkSweep.hpp                         markSweep.inline.hpp
 psMarkSweep.hpp                         collectorCounters.hpp
+psMarkSweep.hpp                         stack.hpp
 
 psMarkSweepDecorator.cpp                liveRange.hpp
 psMarkSweepDecorator.cpp                markSweep.inline.hpp
@@ -280,6 +284,7 @@
 psParallelCompact.cpp			referencePolicy.hpp
 psParallelCompact.cpp			referenceProcessor.hpp
 psParallelCompact.cpp			safepoint.hpp
+psParallelCompact.cpp			stack.inline.hpp
 psParallelCompact.cpp			symbolTable.hpp
 psParallelCompact.cpp			systemDictionary.hpp
 psParallelCompact.cpp			vmThread.hpp
@@ -367,6 +372,7 @@
 psScavenge.cpp                          referenceProcessor.hpp
 psScavenge.cpp                          resourceArea.hpp
 psScavenge.cpp                          spaceDecorator.hpp
+psScavenge.cpp                          stack.inline.hpp
 psScavenge.cpp                          threadCritical.hpp
 psScavenge.cpp                          vmThread.hpp
 psScavenge.cpp                          vm_operations.hpp
@@ -376,6 +382,7 @@
 psScavenge.hpp                          collectorCounters.hpp
 psScavenge.hpp                          oop.hpp
 psScavenge.hpp                          psVirtualspace.hpp
+psScavenge.hpp                          stack.hpp
 
 psScavenge.inline.hpp                   cardTableExtension.hpp
 psScavenge.inline.hpp                   parallelScavengeHeap.hpp
--- a/src/share/vm/gc_implementation/includeDB_gc_serial	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/includeDB_gc_serial	Fri Oct 08 09:29:09 2010 -0700
@@ -93,11 +93,13 @@
 markSweep.hpp                           growableArray.hpp
 markSweep.hpp                           markOop.hpp
 markSweep.hpp                           oop.hpp
+markSweep.hpp                           stack.hpp
 markSweep.hpp                           timer.hpp
 markSweep.hpp                           universe.hpp
 
 markSweep.inline.hpp                    collectedHeap.hpp
 markSweep.inline.hpp                    markSweep.hpp
+markSweep.inline.hpp                    stack.inline.hpp
 
 mutableSpace.hpp                        immutableSpace.hpp
 mutableSpace.hpp                        memRegion.hpp
--- a/src/share/vm/gc_implementation/parNew/parCardTableModRefBS.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parNew/parCardTableModRefBS.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2007, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2007, 2010 Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -44,7 +44,7 @@
 
     int n_strides = n_threads * StridesPerThread;
     SequentialSubTasksDone* pst = sp->par_seq_tasks();
-    pst->set_par_threads(n_threads);
+    pst->set_n_threads(n_threads);
     pst->set_n_tasks(n_strides);
 
     int stride = 0;
--- a/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -34,12 +34,12 @@
                                        Generation* old_gen_,
                                        int thread_num_,
                                        ObjToScanQueueSet* work_queue_set_,
-                                       GrowableArray<oop>**  overflow_stack_set_,
+                                       Stack<oop>* overflow_stacks_,
                                        size_t desired_plab_sz_,
                                        ParallelTaskTerminator& term_) :
   _to_space(to_space_), _old_gen(old_gen_), _young_gen(gen_), _thread_num(thread_num_),
   _work_queue(work_queue_set_->queue(thread_num_)), _to_space_full(false),
-  _overflow_stack(overflow_stack_set_[thread_num_]),
+  _overflow_stack(overflow_stacks_ ? overflow_stacks_ + thread_num_ : NULL),
   _ageTable(false), // false ==> not the global age table, no perf data.
   _to_space_alloc_buffer(desired_plab_sz_),
   _to_space_closure(gen_, this), _old_gen_closure(gen_, this),
@@ -159,11 +159,12 @@
   assert(ParGCUseLocalOverflow, "Else should not call");
   assert(young_gen()->overflow_list() == NULL, "Error");
   ObjToScanQueue* queue = work_queue();
-  GrowableArray<oop>* of_stack = overflow_stack();
-  uint num_overflow_elems = of_stack->length();
-  uint num_take_elems     = MIN2(MIN2((queue->max_elems() - queue->size())/4,
-                                      (juint)ParGCDesiredObjsFromOverflowList),
-                                 num_overflow_elems);
+  Stack<oop>* const of_stack = overflow_stack();
+  const size_t num_overflow_elems = of_stack->size();
+  const size_t space_available = queue->max_elems() - queue->size();
+  const size_t num_take_elems = MIN3(space_available / 4,
+                                     ParGCDesiredObjsFromOverflowList,
+                                     num_overflow_elems);
   // Transfer the most recent num_take_elems from the overflow
   // stack to our work queue.
   for (size_t i = 0; i != num_take_elems; i++) {
@@ -271,7 +272,7 @@
                         ParNewGeneration&       gen,
                         Generation&             old_gen,
                         ObjToScanQueueSet&      queue_set,
-                        GrowableArray<oop>**    overflow_stacks_,
+                        Stack<oop>*             overflow_stacks_,
                         size_t                  desired_plab_sz,
                         ParallelTaskTerminator& term);
 
@@ -302,17 +303,19 @@
 ParScanThreadStateSet::ParScanThreadStateSet(
   int num_threads, Space& to_space, ParNewGeneration& gen,
   Generation& old_gen, ObjToScanQueueSet& queue_set,
-  GrowableArray<oop>** overflow_stack_set_,
+  Stack<oop>* overflow_stacks,
   size_t desired_plab_sz, ParallelTaskTerminator& term)
   : ResourceArray(sizeof(ParScanThreadState), num_threads),
     _gen(gen), _next_gen(old_gen), _term(term)
 {
   assert(num_threads > 0, "sanity check!");
+  assert(ParGCUseLocalOverflow == (overflow_stacks != NULL),
+         "overflow_stack allocation mismatch");
   // Initialize states.
   for (int i = 0; i < num_threads; ++i) {
     new ((ParScanThreadState*)_data + i)
         ParScanThreadState(&to_space, &gen, &old_gen, i, &queue_set,
-                           overflow_stack_set_, desired_plab_sz, term);
+                           overflow_stacks, desired_plab_sz, term);
   }
 }
 
@@ -596,14 +599,11 @@
   for (uint i2 = 0; i2 < ParallelGCThreads; i2++)
     _task_queues->queue(i2)->initialize();
 
-  _overflow_stacks = NEW_C_HEAP_ARRAY(GrowableArray<oop>*, ParallelGCThreads);
-  guarantee(_overflow_stacks != NULL, "Overflow stack set allocation failure");
-  for (uint i = 0; i < ParallelGCThreads; i++) {
-    if (ParGCUseLocalOverflow) {
-      _overflow_stacks[i] = new (ResourceObj::C_HEAP) GrowableArray<oop>(512, true);
-      guarantee(_overflow_stacks[i] != NULL, "Overflow Stack allocation failure.");
-    } else {
-      _overflow_stacks[i] = NULL;
+  _overflow_stacks = NULL;
+  if (ParGCUseLocalOverflow) {
+    _overflow_stacks = NEW_C_HEAP_ARRAY(Stack<oop>, ParallelGCThreads);
+    for (size_t i = 0; i < ParallelGCThreads; ++i) {
+      new (_overflow_stacks + i) Stack<oop>();
     }
   }
 
@@ -937,12 +937,9 @@
   } else {
     assert(HandlePromotionFailure,
       "Should only be here if promotion failure handling is on");
-    if (_promo_failure_scan_stack != NULL) {
-      // Can be non-null because of reference processing.
-      // Free stack with its elements.
-      delete _promo_failure_scan_stack;
-      _promo_failure_scan_stack = NULL;
-    }
+    assert(_promo_failure_scan_stack.is_empty(), "post condition");
+    _promo_failure_scan_stack.clear(true); // Clear cached segments.
+
     remove_forwarding_pointers();
     if (PrintGCDetails) {
       gclog_or_tty->print(" (promotion failed)");
@@ -1397,8 +1394,8 @@
   size_t objsFromOverflow = MIN2((size_t)(work_q->max_elems() - work_q->size())/4,
                                  (size_t)ParGCDesiredObjsFromOverflowList);
 
+  assert(!UseCompressedOops, "Error");
   assert(par_scan_state->overflow_stack() == NULL, "Error");
-  assert(!UseCompressedOops, "Error");
   if (_overflow_list == NULL) return false;
 
   // Otherwise, there was something there; try claiming the list.
@@ -1533,3 +1530,7 @@
 const char* ParNewGeneration::name() const {
   return "par new generation";
 }
+
+bool ParNewGeneration::in_use() {
+  return UseParNewGC && ParallelGCThreads > 0;
+}
--- a/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -52,7 +52,7 @@
   friend class ParScanThreadStateSet;
  private:
   ObjToScanQueue *_work_queue;
-  GrowableArray<oop>* _overflow_stack;
+  Stack<oop>* const _overflow_stack;
 
   ParGCAllocBuffer _to_space_alloc_buffer;
 
@@ -120,7 +120,7 @@
   ParScanThreadState(Space* to_space_, ParNewGeneration* gen_,
                      Generation* old_gen_, int thread_num_,
                      ObjToScanQueueSet* work_queue_set_,
-                     GrowableArray<oop>** overflow_stack_set_,
+                     Stack<oop>* overflow_stacks_,
                      size_t desired_plab_sz_,
                      ParallelTaskTerminator& term_);
 
@@ -144,7 +144,7 @@
   void trim_queues(int max_size);
 
   // Private overflow stack usage
-  GrowableArray<oop>* overflow_stack() { return _overflow_stack; }
+  Stack<oop>* overflow_stack() { return _overflow_stack; }
   bool take_from_overflow_stack();
   void push_on_overflow_stack(oop p);
 
@@ -301,7 +301,7 @@
   ObjToScanQueueSet* _task_queues;
 
   // Per-worker-thread local overflow stacks
-  GrowableArray<oop>** _overflow_stacks;
+  Stack<oop>* _overflow_stacks;
 
   // Desired size of survivor space plab's
   PLABStats _plab_stats;
@@ -350,6 +350,8 @@
     delete _task_queues;
   }
 
+  static bool in_use();
+
   virtual void ref_processor_init();
   virtual Generation::Name kind()        { return Generation::ParNew; }
   virtual const char* name() const;
--- a/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -59,8 +59,6 @@
     PrintGCDetails && TraceParallelOldGCTasks, true, gclog_or_tty));
   ParCompactionManager* cm =
     ParCompactionManager::gc_thread_compaction_manager(which);
-  assert(cm->stacks_have_been_allocated(),
-         "Stack space has not been allocated");
   PSParallelCompact::MarkAndPushClosure mark_and_push_closure(cm);
 
   switch (_root_type) {
@@ -119,7 +117,6 @@
 
   // Do the real work
   cm->follow_marking_stacks();
-  // cm->deallocate_stacks();
 }
 
 
@@ -135,8 +132,6 @@
     PrintGCDetails && TraceParallelOldGCTasks, true, gclog_or_tty));
   ParCompactionManager* cm =
     ParCompactionManager::gc_thread_compaction_manager(which);
-  assert(cm->stacks_have_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(),
--- a/src/share/vm/gc_implementation/parallelScavenge/pcTasks.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/pcTasks.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2005, 2008, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -242,7 +242,11 @@
 //
 
 class DrainStacksCompactionTask : public GCTask {
+ uint _stack_index;
+ uint stack_index() { return _stack_index; }
  public:
+  DrainStacksCompactionTask(uint stack_index) : GCTask(),
+                                                _stack_index(stack_index) {};
   char* name() { return (char *)"drain-region-task"; }
   virtual void do_it(GCTaskManager* manager, uint which);
 };
--- a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -46,23 +46,6 @@
   marking_stack()->initialize();
   _objarray_stack.initialize();
   region_stack()->initialize();
-
-  // Note that _revisit_klass_stack is allocated out of the
-  // C heap (as opposed to out of ResourceArena).
-  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 _revisit_klass_stack;
-  delete _revisit_mdo_stack;
-  // _manager_array and _stack_array are statics
-  // shared with all instances of ParCompactionManager
-  // should not be deallocated.
 }
 
 void ParCompactionManager::initialize(ParMarkBitMap* mbm) {
@@ -134,9 +117,9 @@
 }
 
 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();
+  for(uint i = 0; i < ParallelGCThreads + 1; i++) {
+    assert(manager_array(i)->revisit_klass_stack()->is_empty(), "sanity");
+    assert(manager_array(i)->revisit_mdo_stack()->is_empty(), "sanity");
   }
 }
 
@@ -178,10 +161,3 @@
     }
   } while (!region_stack()->is_empty());
 }
-
-#ifdef ASSERT
-bool ParCompactionManager::stacks_have_been_allocated() {
-  return (revisit_klass_stack()->data_addr() != NULL &&
-          revisit_mdo_stack()->data_addr() != NULL);
-}
-#endif
--- a/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -80,10 +80,9 @@
   // type of TaskQueue.
   RegionTaskQueue               _region_stack;
 
-#if 1  // does this happen enough to need a per thread stack?
-  GrowableArray<Klass*>*        _revisit_klass_stack;
-  GrowableArray<DataLayout*>*   _revisit_mdo_stack;
-#endif
+  Stack<Klass*>                 _revisit_klass_stack;
+  Stack<DataLayout*>            _revisit_mdo_stack;
+
   static ParMarkBitMap* _mark_bitmap;
 
   Action _action;
@@ -113,10 +112,7 @@
   inline static ParCompactionManager* manager_array(int index);
 
   ParCompactionManager();
-  ~ParCompactionManager();
 
-  void allocate_stacks();
-  void deallocate_stacks();
   ParMarkBitMap* mark_bitmap() { return _mark_bitmap; }
 
   // Take actions in preparation for a compaction.
@@ -129,11 +125,8 @@
   bool should_verify_only();
   bool should_reset_only();
 
-#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
+  Stack<Klass*>* revisit_klass_stack() { return &_revisit_klass_stack; }
+  Stack<DataLayout*>* revisit_mdo_stack() { return &_revisit_mdo_stack; }
 
   // Save for later processing.  Must not fail.
   inline void push(oop obj) { _marking_stack.push(obj); }
@@ -162,10 +155,6 @@
   // Process tasks remaining on any stack
   void drain_region_stacks();
 
-  // Debugging support
-#ifdef ASSERT
-  bool stacks_have_been_allocated();
-#endif
 };
 
 inline ParCompactionManager* ParCompactionManager::manager_array(int index) {
--- a/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psMarkSweep.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -466,33 +466,16 @@
   _preserved_count_max  = pointer_delta(to_space->end(), to_space->top(), sizeof(jbyte));
   // Now divide by the size of a PreservedMark
   _preserved_count_max /= sizeof(PreservedMark);
-
-  _preserved_mark_stack = NULL;
-  _preserved_oop_stack = NULL;
-
-  _marking_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
-  _objarray_stack = new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
-
-  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);
 }
 
 
 void PSMarkSweep::deallocate_stacks() {
-  if (_preserved_oop_stack) {
-    delete _preserved_mark_stack;
-    _preserved_mark_stack = NULL;
-    delete _preserved_oop_stack;
-    _preserved_oop_stack = NULL;
-  }
-
-  delete _marking_stack;
-  delete _objarray_stack;
-  delete _revisit_klass_stack;
-  delete _revisit_mdo_stack;
+  _preserved_mark_stack.clear(true);
+  _preserved_oop_stack.clear(true);
+  _marking_stack.clear();
+  _objarray_stack.clear(true);
+  _revisit_klass_stack.clear(true);
+  _revisit_mdo_stack.clear(true);
 }
 
 void PSMarkSweep::mark_sweep_phase1(bool clear_all_softrefs) {
@@ -542,17 +525,17 @@
 
   // Update subklass/sibling/implementor links of live klasses
   follow_weak_klass_links();
-  assert(_marking_stack->is_empty(), "just drained");
+  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");
+  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());
 
-  assert(_marking_stack->is_empty(), "stack should be empty by now");
+  assert(_marking_stack.is_empty(), "stack should be empty by now");
 }
 
 
--- a/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psParallelCompact.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -2170,6 +2170,16 @@
     heap->update_counters();
   }
 
+#ifdef ASSERT
+  for (size_t i = 0; i < ParallelGCThreads + 1; ++i) {
+    ParCompactionManager* const cm =
+      ParCompactionManager::manager_array(int(i));
+    assert(cm->marking_stack()->is_empty(),       "should be empty");
+    assert(cm->region_stack()->is_empty(),        "should be empty");
+    assert(cm->revisit_klass_stack()->is_empty(), "should be empty");
+  }
+#endif // ASSERT
+
   if (VerifyAfterGC && heap->total_collections() >= VerifyGCStartAt) {
     HandleMark hm;  // Discard invalid handles created during verification
     gclog_or_tty->print(" VerifyAfterGC:");
@@ -2449,7 +2459,7 @@
 
   const unsigned int task_count = MAX2(parallel_gc_threads, 1U);
   for (unsigned int j = 0; j < task_count; j++) {
-    q->enqueue(new DrainStacksCompactionTask());
+    q->enqueue(new DrainStacksCompactionTask(j));
   }
 
   // Find all regions that are available (can be filled immediately) and
@@ -2711,21 +2721,22 @@
   // All klasses on the revisit stack are marked at this point.
   // Update and follow all subklass, sibling and implementor links.
   if (PrintRevisitStats) {
-    gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes());
+    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);
-    int length = cm->revisit_klass_stack()->length();
+    Stack<Klass*>* const rks = cm->revisit_klass_stack();
     if (PrintRevisitStats) {
-      gclog_or_tty->print_cr("Revisit klass stack[%d] length = %d", i, length);
+      gclog_or_tty->print_cr("Revisit klass stack[%u] length = " SIZE_FORMAT,
+                             i, rks->size());
     }
-    for (int j = 0; j < length; j++) {
-      cm->revisit_klass_stack()->at(j)->follow_weak_klass_links(
-        is_alive_closure(),
-        &keep_alive_closure);
+    while (!rks->is_empty()) {
+      Klass* const k = rks->pop();
+      k->follow_weak_klass_links(is_alive_closure(), &keep_alive_closure);
     }
-    // revisit_klass_stack is cleared in reset()
+
     cm->follow_marking_stacks();
   }
 }
@@ -2744,19 +2755,20 @@
   // 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());
+    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();
+    Stack<DataLayout*>* rms = cm->revisit_mdo_stack();
     if (PrintRevisitStats) {
-      gclog_or_tty->print_cr("Revisit MDO stack[%d] length = %d", i, length);
+      gclog_or_tty->print_cr("Revisit MDO stack[%u] size = " SIZE_FORMAT,
+                             i, rms->size());
     }
-    for (int j = 0; j < length; j++) {
-      rms->at(j)->follow_weak_refs(is_alive_closure());
+    while (!rms->is_empty()) {
+      rms->pop()->follow_weak_refs(is_alive_closure());
     }
-    // revisit_mdo_stack is cleared in reset()
+
     cm->follow_marking_stacks();
   }
 }
--- a/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -185,7 +185,6 @@
 
 
 void PSPromotionManager::drain_stacks_depth(bool totally_drain) {
-  assert(claimed_stack_depth()->overflow_stack() != NULL, "invariant");
   totally_drain = totally_drain || _totally_drain;
 
 #ifdef ASSERT
--- a/src/share/vm/gc_implementation/parallelScavenge/psScavenge.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psScavenge.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -34,9 +34,10 @@
 int                        PSScavenge::_tenuring_threshold = 0;
 HeapWord*                  PSScavenge::_young_generation_boundary = NULL;
 elapsedTimer               PSScavenge::_accumulated_time;
-GrowableArray<markOop>*    PSScavenge::_preserved_mark_stack = NULL;
-GrowableArray<oop>*        PSScavenge::_preserved_oop_stack = NULL;
+Stack<markOop>             PSScavenge::_preserved_mark_stack;
+Stack<oop>                 PSScavenge::_preserved_oop_stack;
 CollectorCounters*         PSScavenge::_counters = NULL;
+bool                       PSScavenge::_promotion_failed = false;
 
 // Define before use
 class PSIsAliveClosure: public BoolObjectClosure {
@@ -223,6 +224,9 @@
   assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint");
   assert(Thread::current() == (Thread*)VMThread::vm_thread(), "should be in vm thread");
 
+  assert(_preserved_mark_stack.is_empty(), "should be empty");
+  assert(_preserved_oop_stack.is_empty(), "should be empty");
+
   TimeStamp scavenge_entry;
   TimeStamp scavenge_midpoint;
   TimeStamp scavenge_exit;
@@ -636,24 +640,20 @@
     young_gen->object_iterate(&unforward_closure);
 
     if (PrintGC && Verbose) {
-      gclog_or_tty->print_cr("Restoring %d marks",
-                              _preserved_oop_stack->length());
+      gclog_or_tty->print_cr("Restoring %d marks", _preserved_oop_stack.size());
     }
 
     // Restore any saved marks.
-    for (int i=0; i < _preserved_oop_stack->length(); i++) {
-      oop obj       = _preserved_oop_stack->at(i);
-      markOop mark  = _preserved_mark_stack->at(i);
+    while (!_preserved_oop_stack.is_empty()) {
+      oop obj      = _preserved_oop_stack.pop();
+      markOop mark = _preserved_mark_stack.pop();
       obj->set_mark(mark);
     }
 
-    // Deallocate the preserved mark and oop stacks.
-    // The stacks were allocated as CHeap objects, so
-    // we must call delete to prevent mem leaks.
-    delete _preserved_mark_stack;
-    _preserved_mark_stack = NULL;
-    delete _preserved_oop_stack;
-    _preserved_oop_stack = NULL;
+    // Clear the preserved mark and oop stack caches.
+    _preserved_mark_stack.clear(true);
+    _preserved_oop_stack.clear(true);
+    _promotion_failed = false;
   }
 
   // Reset the PromotionFailureALot counters.
@@ -661,27 +661,16 @@
 }
 
 // This method is called whenever an attempt to promote an object
-// fails. Some markOops will need preserving, some will not. Note
+// fails. Some markOops will need preservation, some will not. Note
 // that the entire eden is traversed after a failed promotion, with
 // all forwarded headers replaced by the default markOop. This means
 // it is not neccessary to preserve most markOops.
 void PSScavenge::oop_promotion_failed(oop obj, markOop obj_mark) {
-  if (_preserved_mark_stack == NULL) {
-    ThreadCritical tc; // Lock and retest
-    if (_preserved_mark_stack == NULL) {
-      assert(_preserved_oop_stack == NULL, "Sanity");
-      _preserved_mark_stack = new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
-      _preserved_oop_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
-    }
-  }
-
-  // Because we must hold the ThreadCritical lock before using
-  // the stacks, we should be safe from observing partial allocations,
-  // which are also guarded by the ThreadCritical lock.
+  _promotion_failed = true;
   if (obj_mark->must_be_preserved_for_promotion_failure(obj)) {
     ThreadCritical tc;
-    _preserved_oop_stack->push(obj);
-    _preserved_mark_stack->push(obj_mark);
+    _preserved_oop_stack.push(obj);
+    _preserved_mark_stack.push(obj_mark);
   }
 }
 
--- a/src/share/vm/gc_implementation/parallelScavenge/psScavenge.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/parallelScavenge/psScavenge.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -61,9 +61,10 @@
   static HeapWord*           _young_generation_boundary; // The lowest address possible for the young_gen.
                                                          // This is used to decide if an oop should be scavenged,
                                                          // cards should be marked, etc.
-  static GrowableArray<markOop>* _preserved_mark_stack; // List of marks to be restored after failed promotion
-  static GrowableArray<oop>*     _preserved_oop_stack;  // List of oops that need their mark restored.
+  static Stack<markOop>          _preserved_mark_stack; // List of marks to be restored after failed promotion
+  static Stack<oop>              _preserved_oop_stack;  // List of oops that need their mark restored.
   static CollectorCounters*      _counters;         // collector performance counters
+  static bool                    _promotion_failed;
 
   static void clean_up_failed_promotion();
 
@@ -79,8 +80,7 @@
   // Accessors
   static int              tenuring_threshold()  { return _tenuring_threshold; }
   static elapsedTimer*    accumulated_time()    { return &_accumulated_time; }
-  static bool             promotion_failed()
-    { return _preserved_mark_stack != NULL; }
+  static bool             promotion_failed()    { return _promotion_failed; }
   static int              consecutive_skipped_scavenges()
     { return _consecutive_skipped_scavenges; }
 
--- a/src/share/vm/gc_implementation/shared/concurrentGCThread.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/shared/concurrentGCThread.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -185,7 +185,7 @@
   instanceKlassHandle klass (THREAD, k);
   instanceHandle thread_oop = klass->allocate_instance_handle(CHECK_NULL);
 
-  const char thread_name[] = "Surrogate Locker Thread (CMS)";
+  const char thread_name[] = "Surrogate Locker Thread (Concurrent GC)";
   Handle string = java_lang_String::create_from_str(thread_name, CHECK_NULL);
 
   // Initialize thread_oop to put it into the system threadGroup
--- a/src/share/vm/gc_implementation/shared/markSweep.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/shared/markSweep.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -25,13 +25,13 @@
 #include "incls/_precompiled.incl"
 #include "incls/_markSweep.cpp.incl"
 
-GrowableArray<oop>*          MarkSweep::_marking_stack = NULL;
-GrowableArray<ObjArrayTask>* MarkSweep::_objarray_stack = NULL;
-GrowableArray<Klass*>*       MarkSweep::_revisit_klass_stack = NULL;
-GrowableArray<DataLayout*>*  MarkSweep::_revisit_mdo_stack = NULL;
+Stack<oop>              MarkSweep::_marking_stack;
+Stack<DataLayout*>      MarkSweep::_revisit_mdo_stack;
+Stack<Klass*>           MarkSweep::_revisit_klass_stack;
+Stack<ObjArrayTask>     MarkSweep::_objarray_stack;
 
-GrowableArray<oop>*     MarkSweep::_preserved_oop_stack = NULL;
-GrowableArray<markOop>* MarkSweep::_preserved_mark_stack= NULL;
+Stack<oop>              MarkSweep::_preserved_oop_stack;
+Stack<markOop>          MarkSweep::_preserved_mark_stack;
 size_t                  MarkSweep::_preserved_count = 0;
 size_t                  MarkSweep::_preserved_count_max = 0;
 PreservedMark*          MarkSweep::_preserved_marks = NULL;
@@ -58,37 +58,42 @@
 #endif
 
 void MarkSweep::revisit_weak_klass_link(Klass* k) {
-  _revisit_klass_stack->push(k);
+  _revisit_klass_stack.push(k);
 }
 
 void MarkSweep::follow_weak_klass_links() {
   // All klasses on the revisit stack are marked at this point.
   // Update and follow all subklass, sibling and implementor links.
   if (PrintRevisitStats) {
-    gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes());
-    gclog_or_tty->print_cr("Revisit klass stack length = %d", _revisit_klass_stack->length());
+    gclog_or_tty->print_cr("#classes in system dictionary = %d",
+                           SystemDictionary::number_of_classes());
+    gclog_or_tty->print_cr("Revisit klass stack size = " SIZE_FORMAT,
+                           _revisit_klass_stack.size());
   }
-  for (int i = 0; i < _revisit_klass_stack->length(); i++) {
-    _revisit_klass_stack->at(i)->follow_weak_klass_links(&is_alive,&keep_alive);
+  while (!_revisit_klass_stack.is_empty()) {
+    Klass* const k = _revisit_klass_stack.pop();
+    k->follow_weak_klass_links(&is_alive, &keep_alive);
   }
   follow_stack();
 }
 
 void MarkSweep::revisit_mdo(DataLayout* p) {
-  _revisit_mdo_stack->push(p);
+  _revisit_mdo_stack.push(p);
 }
 
 void MarkSweep::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.
-  assert(_marking_stack->is_empty(), "Marking stack should be empty");
+  assert(_marking_stack.is_empty(), "Marking stack should be empty");
   if (PrintRevisitStats) {
-    gclog_or_tty->print_cr("#classes in system dictionary = %d", SystemDictionary::number_of_classes());
-    gclog_or_tty->print_cr("Revisit MDO stack length = %d", _revisit_mdo_stack->length());
+    gclog_or_tty->print_cr("#classes in system dictionary = %d",
+                           SystemDictionary::number_of_classes());
+    gclog_or_tty->print_cr("Revisit MDO stack size = " SIZE_FORMAT,
+                           _revisit_mdo_stack.size());
   }
-  for (int i = 0; i < _revisit_mdo_stack->length(); i++) {
-    _revisit_mdo_stack->at(i)->follow_weak_refs(&is_alive);
+  while (!_revisit_mdo_stack.is_empty()) {
+    _revisit_mdo_stack.pop()->follow_weak_refs(&is_alive);
   }
   follow_stack();
 }
@@ -106,41 +111,37 @@
 
 void MarkSweep::follow_stack() {
   do {
-    while (!_marking_stack->is_empty()) {
-      oop obj = _marking_stack->pop();
+    while (!_marking_stack.is_empty()) {
+      oop obj = _marking_stack.pop();
       assert (obj->is_gc_marked(), "p must be marked");
       obj->follow_contents();
     }
     // Process ObjArrays one at a time to avoid marking stack bloat.
-    if (!_objarray_stack->is_empty()) {
-      ObjArrayTask task = _objarray_stack->pop();
+    if (!_objarray_stack.is_empty()) {
+      ObjArrayTask task = _objarray_stack.pop();
       objArrayKlass* const k = (objArrayKlass*)task.obj()->blueprint();
       k->oop_follow_contents(task.obj(), task.index());
     }
-  } while (!_marking_stack->is_empty() || !_objarray_stack->is_empty());
+  } while (!_marking_stack.is_empty() || !_objarray_stack.is_empty());
 }
 
 MarkSweep::FollowStackClosure MarkSweep::follow_stack_closure;
 
 void MarkSweep::FollowStackClosure::do_void() { follow_stack(); }
 
-// We preserve the mark which should be replaced at the end and the location that it
-// will go.  Note that the object that this markOop belongs to isn't currently at that
-// address but it will be after phase4
+// We preserve the mark which should be replaced at the end and the location
+// that it will go.  Note that the object that this markOop belongs to isn't
+// currently at that address but it will be after phase4
 void MarkSweep::preserve_mark(oop obj, markOop mark) {
-  // we try to store preserved marks in the to space of the new generation since this
-  // is storage which should be available.  Most of the time this should be sufficient
-  // space for the marks we need to preserve but if it isn't we fall back in using
-  // GrowableArrays to keep track of the overflow.
+  // We try to store preserved marks in the to space of the new generation since
+  // this is storage which should be available.  Most of the time this should be
+  // sufficient space for the marks we need to preserve but if it isn't we fall
+  // back to using Stacks to keep track of the overflow.
   if (_preserved_count < _preserved_count_max) {
     _preserved_marks[_preserved_count++].init(obj, mark);
   } else {
-    if (_preserved_mark_stack == NULL) {
-      _preserved_mark_stack = new (ResourceObj::C_HEAP) GrowableArray<markOop>(40, true);
-      _preserved_oop_stack = new (ResourceObj::C_HEAP) GrowableArray<oop>(40, true);
-    }
-    _preserved_mark_stack->push(mark);
-    _preserved_oop_stack->push(obj);
+    _preserved_mark_stack.push(mark);
+    _preserved_oop_stack.push(obj);
   }
 }
 
@@ -151,8 +152,7 @@
 void MarkSweep::AdjustPointerClosure::do_oop(narrowOop* p) { adjust_pointer(p, _is_root); }
 
 void MarkSweep::adjust_marks() {
-  assert(_preserved_oop_stack == NULL ||
-         _preserved_oop_stack->length() == _preserved_mark_stack->length(),
+  assert( _preserved_oop_stack.size() == _preserved_mark_stack.size(),
          "inconsistent preserved oop stacks");
 
   // adjust the oops we saved earlier
@@ -161,21 +161,19 @@
   }
 
   // deal with the overflow stack
-  if (_preserved_oop_stack) {
-    for (int i = 0; i < _preserved_oop_stack->length(); i++) {
-      oop* p = _preserved_oop_stack->adr_at(i);
-      adjust_pointer(p);
-    }
+  StackIterator<oop> iter(_preserved_oop_stack);
+  while (!iter.is_empty()) {
+    oop* p = iter.next_addr();
+    adjust_pointer(p);
   }
 }
 
 void MarkSweep::restore_marks() {
-  assert(_preserved_oop_stack == NULL ||
-         _preserved_oop_stack->length() == _preserved_mark_stack->length(),
+  assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(),
          "inconsistent preserved oop stacks");
   if (PrintGC && Verbose) {
-    gclog_or_tty->print_cr("Restoring %d marks", _preserved_count +
-                  (_preserved_oop_stack ? _preserved_oop_stack->length() : 0));
+    gclog_or_tty->print_cr("Restoring %d marks",
+                           _preserved_count + _preserved_oop_stack.size());
   }
 
   // restore the marks we saved earlier
@@ -184,12 +182,10 @@
   }
 
   // deal with the overflow
-  if (_preserved_oop_stack) {
-    for (int i = 0; i < _preserved_oop_stack->length(); i++) {
-      oop obj       = _preserved_oop_stack->at(i);
-      markOop mark  = _preserved_mark_stack->at(i);
-      obj->set_mark(mark);
-    }
+  while (!_preserved_oop_stack.is_empty()) {
+    oop obj       = _preserved_oop_stack.pop();
+    markOop mark  = _preserved_mark_stack.pop();
+    obj->set_mark(mark);
   }
 }
 
--- a/src/share/vm/gc_implementation/shared/markSweep.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/shared/markSweep.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -104,23 +104,22 @@
   friend class KeepAliveClosure;
   friend class VM_MarkSweep;
   friend void marksweep_init();
-  friend class DataLayout;
 
   //
   // Vars
   //
  protected:
   // Traversal stacks used during phase1
-  static GrowableArray<oop>*             _marking_stack;
-  static GrowableArray<ObjArrayTask>*    _objarray_stack;
+  static Stack<oop>                      _marking_stack;
+  static Stack<ObjArrayTask>             _objarray_stack;
   // Stack for live klasses to revisit at end of marking phase
-  static GrowableArray<Klass*>*          _revisit_klass_stack;
+  static Stack<Klass*>                   _revisit_klass_stack;
   // Set (stack) of MDO's to revisit at end of marking phase
-  static GrowableArray<DataLayout*>*    _revisit_mdo_stack;
+  static Stack<DataLayout*>              _revisit_mdo_stack;
 
   // Space for storing/restoring mark word
-  static GrowableArray<markOop>*         _preserved_mark_stack;
-  static GrowableArray<oop>*             _preserved_oop_stack;
+  static Stack<markOop>                  _preserved_mark_stack;
+  static Stack<oop>                      _preserved_oop_stack;
   static size_t                          _preserved_count;
   static size_t                          _preserved_count_max;
   static PreservedMark*                  _preserved_marks;
--- a/src/share/vm/gc_implementation/shared/markSweep.inline.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_implementation/shared/markSweep.inline.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -72,7 +72,7 @@
     oop obj = oopDesc::decode_heap_oop_not_null(heap_oop);
     if (!obj->mark()->is_marked()) {
       mark_object(obj);
-      _marking_stack->push(obj);
+      _marking_stack.push(obj);
     }
   }
 }
@@ -80,7 +80,7 @@
 void MarkSweep::push_objarray(oop obj, size_t index) {
   ObjArrayTask task(obj, index);
   assert(task.is_valid(), "bad ObjArrayTask");
-  _objarray_stack->push(task);
+  _objarray_stack.push(task);
 }
 
 template <class T> inline void MarkSweep::adjust_pointer(T* p, bool isroot) {
--- a/src/share/vm/gc_interface/collectedHeap.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_interface/collectedHeap.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -34,7 +34,9 @@
 
 // Memory state functions.
 
-CollectedHeap::CollectedHeap()
+
+CollectedHeap::CollectedHeap() : _n_par_threads(0)
+
 {
   const size_t max_len = size_t(arrayOopDesc::max_array_length(T_INT));
   const size_t elements_per_word = HeapWordSize / sizeof(jint);
--- a/src/share/vm/gc_interface/collectedHeap.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/gc_interface/collectedHeap.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -59,6 +59,8 @@
   MemRegion _reserved;
   BarrierSet* _barrier_set;
   bool _is_gc_active;
+  int _n_par_threads;
+
   unsigned int _total_collections;          // ... started
   unsigned int _total_full_collections;     // ... started
   NOT_PRODUCT(volatile size_t _promotion_failure_alot_count;)
@@ -293,6 +295,12 @@
   }
   GCCause::Cause gc_cause() { return _gc_cause; }
 
+  // Number of threads currently working on GC tasks.
+  int n_par_threads() { return _n_par_threads; }
+
+  // May be overridden to set additional parallelism.
+  virtual void set_par_threads(int t) { _n_par_threads = t; };
+
   // Preload classes into the shared portion of the heap, and then dump
   // that data to a file so that it can be loaded directly by another
   // VM (then terminate).
@@ -606,6 +614,14 @@
     return (CIFireOOMAt > 1 && _fire_out_of_memory_count >= CIFireOOMAt);
   }
 #endif
+
+ public:
+  // This is a convenience method that is used in cases where
+  // the actual number of GC worker threads is not pertinent but
+  // only whether there more than 0.  Use of this method helps
+  // reduce the occurrence of ParallelGCThreads to uses where the
+  // actual number may be germane.
+  static bool use_parallel_gc_threads() { return ParallelGCThreads > 0; }
 };
 
 // Class to set and reset the GC cause for a CollectedHeap.
--- a/src/share/vm/includeDB_core	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/includeDB_core	Fri Oct 08 09:29:09 2010 -0700
@@ -1437,12 +1437,14 @@
 defNewGeneration.cpp                    referencePolicy.hpp
 defNewGeneration.cpp                    space.inline.hpp
 defNewGeneration.cpp                    spaceDecorator.hpp
+defNewGeneration.cpp                    stack.inline.hpp
 defNewGeneration.cpp                    thread_<os_family>.inline.hpp
 
 defNewGeneration.hpp                    ageTable.hpp
 defNewGeneration.hpp                    cSpaceCounters.hpp
 defNewGeneration.hpp                    generation.inline.hpp
 defNewGeneration.hpp                    generationCounters.hpp
+defNewGeneration.hpp                    stack.hpp
 
 defNewGeneration.inline.hpp             cardTableRS.hpp
 defNewGeneration.inline.hpp             defNewGeneration.hpp
@@ -3461,6 +3463,7 @@
 permGen.hpp                             generation.hpp
 permGen.hpp                             handles.hpp
 permGen.hpp                             iterator.hpp
+permGen.hpp                             mutexLocker.hpp
 permGen.hpp                             virtualspace.hpp
 
 placeholders.cpp                        fieldType.hpp
@@ -3871,6 +3874,10 @@
 
 specialized_oop_closures.hpp            atomic.hpp
 
+stack.hpp                               allocation.inline.hpp
+
+stack.inline.hpp                        stack.hpp
+
 stackMapFrame.cpp                       globalDefinitions.hpp
 stackMapFrame.cpp                       handles.inline.hpp
 stackMapFrame.cpp                       oop.inline.hpp
@@ -4115,6 +4122,7 @@
 taskqueue.cpp                           debug.hpp
 taskqueue.cpp				oop.inline.hpp
 taskqueue.cpp                           os.hpp
+taskqueue.cpp                           stack.inline.hpp
 taskqueue.cpp                           taskqueue.hpp
 taskqueue.cpp                           thread_<os_family>.inline.hpp
 
@@ -4122,6 +4130,7 @@
 taskqueue.hpp                           allocation.inline.hpp
 taskqueue.hpp                           mutex.hpp
 taskqueue.hpp                           orderAccess_<os_arch>.inline.hpp
+taskqueue.hpp				stack.hpp
 
 templateInterpreter.cpp                 interpreter.hpp
 templateInterpreter.cpp                 interpreterGenerator.hpp
@@ -4741,6 +4750,7 @@
 workgroup.cpp                           os.hpp
 workgroup.cpp                           workgroup.hpp
 
+workgroup.hpp                           taskqueue.hpp
 workgroup.hpp                           thread_<os_family>.inline.hpp
 
 xmlstream.cpp                           allocation.hpp
--- a/src/share/vm/memory/allocation.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/allocation.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -289,16 +289,17 @@
 
 // One of the following macros must be used when allocating
 // an array or object from an arena
-#define NEW_ARENA_ARRAY(arena, type, size)\
-  (type*) arena->Amalloc((size) * sizeof(type))
+#define NEW_ARENA_ARRAY(arena, type, size) \
+  (type*) (arena)->Amalloc((size) * sizeof(type))
 
-#define REALLOC_ARENA_ARRAY(arena, type, old, old_size, new_size)\
-  (type*) arena->Arealloc((char*)(old), (old_size) * sizeof(type), (new_size) * sizeof(type) )
+#define REALLOC_ARENA_ARRAY(arena, type, old, old_size, new_size)    \
+  (type*) (arena)->Arealloc((char*)(old), (old_size) * sizeof(type), \
+                            (new_size) * sizeof(type) )
 
-#define FREE_ARENA_ARRAY(arena, type, old, size)\
-  arena->Afree((char*)(old), (size) * sizeof(type))
+#define FREE_ARENA_ARRAY(arena, type, old, size) \
+  (arena)->Afree((char*)(old), (size) * sizeof(type))
 
-#define NEW_ARENA_OBJ(arena, type)\
+#define NEW_ARENA_OBJ(arena, type) \
   NEW_ARENA_ARRAY(arena, type, 1)
 
 
--- a/src/share/vm/memory/defNewGeneration.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/defNewGeneration.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -87,9 +87,7 @@
     _gch->oop_since_save_marks_iterate(_level, _scan_cur_or_nonheap,
                                        _scan_older);
   } while (!_gch->no_allocs_since_save_marks(_level));
-  guarantee(_gen->promo_failure_scan_stack() == NULL
-            || _gen->promo_failure_scan_stack()->length() == 0,
-            "Failed to finish scan");
+  guarantee(_gen->promo_failure_scan_is_complete(), "Failed to finish scan");
 }
 
 ScanClosure::ScanClosure(DefNewGeneration* g, bool gc_barrier) :
@@ -130,9 +128,6 @@
                                    int level,
                                    const char* policy)
   : Generation(rs, initial_size, level),
-    _objs_with_preserved_marks(NULL),
-    _preserved_marks_of_objs(NULL),
-    _promo_failure_scan_stack(NULL),
     _promo_failure_drain_in_progress(false),
     _should_allocate_from_space(false)
 {
@@ -604,12 +599,8 @@
   } else {
     assert(HandlePromotionFailure,
       "Should not be here unless promotion failure handling is on");
-    assert(_promo_failure_scan_stack != NULL &&
-      _promo_failure_scan_stack->length() == 0, "post condition");
-
-    // deallocate stack and it's elements
-    delete _promo_failure_scan_stack;
-    _promo_failure_scan_stack = NULL;
+    assert(_promo_failure_scan_stack.is_empty(), "post condition");
+    _promo_failure_scan_stack.clear(true); // Clear cached segments.
 
     remove_forwarding_pointers();
     if (PrintGCDetails) {
@@ -620,7 +611,7 @@
     // case there can be live objects in to-space
     // as a result of a partial evacuation of eden
     // and from-space.
-    swap_spaces();   // For the sake of uniformity wrt ParNewGeneration::collect().
+    swap_spaces();   // For uniformity wrt ParNewGeneration.
     from()->set_next_compaction_space(to());
     gch->set_incremental_collection_will_fail();
 
@@ -653,34 +644,23 @@
   RemoveForwardPointerClosure rspc;
   eden()->object_iterate(&rspc);
   from()->object_iterate(&rspc);
+
   // Now restore saved marks, if any.
-  if (_objs_with_preserved_marks != NULL) {
-    assert(_preserved_marks_of_objs != NULL, "Both or none.");
-    assert(_objs_with_preserved_marks->length() ==
-           _preserved_marks_of_objs->length(), "Both or none.");
-    for (int i = 0; i < _objs_with_preserved_marks->length(); i++) {
-      oop obj   = _objs_with_preserved_marks->at(i);
-      markOop m = _preserved_marks_of_objs->at(i);
-      obj->set_mark(m);
-    }
-    delete _objs_with_preserved_marks;
-    delete _preserved_marks_of_objs;
-    _objs_with_preserved_marks = NULL;
-    _preserved_marks_of_objs = NULL;
+  assert(_objs_with_preserved_marks.size() == _preserved_marks_of_objs.size(),
+         "should be the same");
+  while (!_objs_with_preserved_marks.is_empty()) {
+    oop obj   = _objs_with_preserved_marks.pop();
+    markOop m = _preserved_marks_of_objs.pop();
+    obj->set_mark(m);
   }
+  _objs_with_preserved_marks.clear(true);
+  _preserved_marks_of_objs.clear(true);
 }
 
 void DefNewGeneration::preserve_mark_if_necessary(oop obj, markOop m) {
   if (m->must_be_preserved_for_promotion_failure(obj)) {
-    if (_objs_with_preserved_marks == NULL) {
-      assert(_preserved_marks_of_objs == NULL, "Both or none.");
-      _objs_with_preserved_marks = new (ResourceObj::C_HEAP)
-        GrowableArray<oop>(PreserveMarkStackSize, true);
-      _preserved_marks_of_objs = new (ResourceObj::C_HEAP)
-        GrowableArray<markOop>(PreserveMarkStackSize, true);
-    }
-    _objs_with_preserved_marks->push(obj);
-    _preserved_marks_of_objs->push(m);
+    _objs_with_preserved_marks.push(obj);
+    _preserved_marks_of_objs.push(m);
   }
 }
 
@@ -695,7 +675,7 @@
   old->forward_to(old);
   _promotion_failed = true;
 
-  push_on_promo_failure_scan_stack(old);
+  _promo_failure_scan_stack.push(old);
 
   if (!_promo_failure_drain_in_progress) {
     // prevent recursion in copy_to_survivor_space()
@@ -748,20 +728,9 @@
   return obj;
 }
 
-void DefNewGeneration::push_on_promo_failure_scan_stack(oop obj) {
-  if (_promo_failure_scan_stack == NULL) {
-    _promo_failure_scan_stack = new (ResourceObj::C_HEAP)
-                                    GrowableArray<oop>(40, true);
-  }
-
-  _promo_failure_scan_stack->push(obj);
-}
-
 void DefNewGeneration::drain_promo_failure_scan_stack() {
-  assert(_promo_failure_scan_stack != NULL, "precondition");
-
-  while (_promo_failure_scan_stack->length() > 0) {
-     oop obj = _promo_failure_scan_stack->pop();
+  while (!_promo_failure_scan_stack.is_empty()) {
+     oop obj = _promo_failure_scan_stack.pop();
      obj->oop_iterate(_promo_failure_scan_stack_closure);
   }
 }
--- a/src/share/vm/memory/defNewGeneration.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/defNewGeneration.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -77,10 +77,10 @@
   // word being overwritten with a self-forwarding-pointer.
   void   preserve_mark_if_necessary(oop obj, markOop m);
 
-  // When one is non-null, so is the other.  Together, they each pair is
-  // an object with a preserved mark, and its mark value.
-  GrowableArray<oop>*     _objs_with_preserved_marks;
-  GrowableArray<markOop>* _preserved_marks_of_objs;
+  // Together, these keep <object with a preserved mark, mark value> pairs.
+  // They should always contain the same number of elements.
+  Stack<oop>     _objs_with_preserved_marks;
+  Stack<markOop> _preserved_marks_of_objs;
 
   // Returns true if the collection can be safely attempted.
   // If this method returns false, a collection is not
@@ -94,11 +94,7 @@
     _promo_failure_scan_stack_closure = scan_stack_closure;
   }
 
-  GrowableArray<oop>* _promo_failure_scan_stack;
-  GrowableArray<oop>* promo_failure_scan_stack() const {
-    return _promo_failure_scan_stack;
-  }
-  void push_on_promo_failure_scan_stack(oop);
+  Stack<oop> _promo_failure_scan_stack;
   void drain_promo_failure_scan_stack(void);
   bool _promo_failure_drain_in_progress;
 
@@ -184,8 +180,6 @@
     void do_void();
   };
 
-  class FastEvacuateFollowersClosure;
-  friend class FastEvacuateFollowersClosure;
   class FastEvacuateFollowersClosure: public VoidClosure {
     GenCollectedHeap* _gch;
     int _level;
@@ -336,6 +330,10 @@
 
   void verify(bool allow_dirty);
 
+  bool promo_failure_scan_is_complete() const {
+    return _promo_failure_scan_stack.is_empty();
+  }
+
  protected:
   // If clear_space is true, clear the survivor spaces.  Eden is
   // cleared if the minimum size of eden is 0.  If mangle_space
--- a/src/share/vm/memory/genCollectedHeap.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/genCollectedHeap.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -676,7 +676,7 @@
 
 void GenCollectedHeap::set_par_threads(int t) {
   SharedHeap::set_par_threads(t);
-  _gen_process_strong_tasks->set_par_threads(t);
+  _gen_process_strong_tasks->set_n_threads(t);
 }
 
 class AssertIsPermClosure: public OopClosure {
--- a/src/share/vm/memory/genCollectedHeap.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/genCollectedHeap.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -74,6 +74,7 @@
   // Data structure for claiming the (potentially) parallel tasks in
   // (gen-specific) strong roots processing.
   SubTasksDone* _gen_process_strong_tasks;
+  SubTasksDone* gen_process_strong_tasks() { return _gen_process_strong_tasks; }
 
   // In block contents verification, the number of header words to skip
   NOT_PRODUCT(static size_t _skip_header_HeapWords;)
--- a/src/share/vm/memory/genMarkSweep.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/genMarkSweep.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -161,17 +161,6 @@
 
   _preserved_marks = (PreservedMark*)scratch;
   _preserved_count = 0;
-  _preserved_mark_stack = NULL;
-  _preserved_oop_stack = NULL;
-
-  _marking_stack       = new (ResourceObj::C_HEAP) GrowableArray<oop>(4000, true);
-  _objarray_stack      = new (ResourceObj::C_HEAP) GrowableArray<ObjArrayTask>(50, true);
-
-  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 to be a better fit, but this will have to do for
-  // now until we have had a chance to investigate a more optimal setting.
-  _revisit_mdo_stack   = new (ResourceObj::C_HEAP) GrowableArray<DataLayout*>(2*size, true);
 
 #ifdef VALIDATE_MARK_SWEEP
   if (ValidateMarkSweep) {
@@ -206,17 +195,12 @@
     gch->release_scratch();
   }
 
-  if (_preserved_oop_stack) {
-    delete _preserved_mark_stack;
-    _preserved_mark_stack = NULL;
-    delete _preserved_oop_stack;
-    _preserved_oop_stack = NULL;
-  }
-
-  delete _marking_stack;
-  delete _objarray_stack;
-  delete _revisit_klass_stack;
-  delete _revisit_mdo_stack;
+  _preserved_mark_stack.clear(true);
+  _preserved_oop_stack.clear(true);
+  _marking_stack.clear();
+  _objarray_stack.clear(true);
+  _revisit_klass_stack.clear(true);
+  _revisit_mdo_stack.clear(true);
 
 #ifdef VALIDATE_MARK_SWEEP
   if (ValidateMarkSweep) {
@@ -274,17 +258,17 @@
 
   // Update subklass/sibling/implementor links of live klasses
   follow_weak_klass_links();
-  assert(_marking_stack->is_empty(), "just drained");
+  assert(_marking_stack.is_empty(), "just drained");
 
   // Visit memoized MDO's and clear any unmarked weak refs
   follow_mdo_weak_refs();
-  assert(_marking_stack->is_empty(), "just drained");
+  assert(_marking_stack.is_empty(), "just drained");
 
   // Visit symbol and interned string tables and delete unmarked oops
   SymbolTable::unlink(&is_alive);
   StringTable::unlink(&is_alive);
 
-  assert(_marking_stack->is_empty(), "stack should be empty by now");
+  assert(_marking_stack.is_empty(), "stack should be empty by now");
 }
 
 
--- a/src/share/vm/memory/permGen.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/permGen.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -25,6 +25,17 @@
 #include "incls/_precompiled.incl"
 #include "incls/_permGen.cpp.incl"
 
+HeapWord* PermGen::request_expand_and_allocate(Generation* gen, size_t size,
+                                               GCCause::Cause prev_cause) {
+  if (gen->capacity() < _capacity_expansion_limit ||
+      prev_cause != GCCause::_no_gc || UseG1GC) {  // last disjunct is a temporary hack for G1
+    return gen->expand_and_allocate(size, false);
+  }
+  // We have reached the limit of capacity expansion where
+  // we will not expand further until a GC is done; request denied.
+  return NULL;
+}
+
 HeapWord* PermGen::mem_allocate_in_gen(size_t size, Generation* gen) {
   GCCause::Cause next_cause = GCCause::_permanent_generation_full;
   GCCause::Cause prev_cause = GCCause::_no_gc;
@@ -37,10 +48,14 @@
       if ((obj = gen->allocate(size, false)) != NULL) {
         return obj;
       }
-      if (gen->capacity() < _capacity_expansion_limit ||
-          prev_cause != GCCause::_no_gc) {
-        obj = gen->expand_and_allocate(size, false);
-      }
+      // Attempt to expand and allocate the requested space:
+      // specific subtypes may use specific policy to either expand
+      // or not. The default policy (see above) is to expand until
+      // _capacity_expansion_limit, and no further unless a GC is done.
+      // Concurrent collectors may decide to kick off a concurrent
+      // collection under appropriate conditions.
+      obj = request_expand_and_allocate(gen, size, prev_cause);
+
       if (obj != NULL || prev_cause == GCCause::_last_ditch_collection) {
         return obj;
       }
@@ -119,5 +134,5 @@
   if (_gen->capacity() > desired_capacity) {
     _gen->shrink(_gen->capacity() - desired_capacity);
   }
-  _capacity_expansion_limit = _gen->capacity() + MaxPermHeapExpansion;
+  set_capacity_expansion_limit(_gen->capacity() + MaxPermHeapExpansion);
 }
--- a/src/share/vm/memory/permGen.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/permGen.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2008, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -30,15 +30,26 @@
 class GenRemSet;
 class CSpaceCounters;
 
-// PermGen models the part of the heap
+// PermGen models the part of the heap used to allocate class meta-data.
 
 class PermGen : public CHeapObj {
   friend class VMStructs;
  protected:
   size_t _capacity_expansion_limit;  // maximum expansion allowed without a
                                      // full gc occurring
+  void set_capacity_expansion_limit(size_t limit) {
+    assert_locked_or_safepoint(Heap_lock);
+    _capacity_expansion_limit = limit;
+  }
 
   HeapWord* mem_allocate_in_gen(size_t size, Generation* gen);
+  // Along with mem_allocate_in_gen() above, implements policy for
+  // "scheduling" allocation/expansion/collection of the perm gen.
+  // The virtual method request_...() below can be overridden by
+  // subtypes that want to implement a different expansion/collection
+  // policy from the default provided.
+  virtual HeapWord* request_expand_and_allocate(Generation* gen, size_t size,
+                                                GCCause::Cause prev_cause);
 
  public:
   enum Name {
--- a/src/share/vm/memory/referenceProcessor.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/referenceProcessor.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -137,16 +137,17 @@
   _discovery_is_atomic = atomic_discovery;
   _discovery_is_mt     = mt_discovery;
   _num_q               = mt_degree;
-  _discoveredSoftRefs  = NEW_C_HEAP_ARRAY(DiscoveredList, _num_q * subclasses_of_ref);
+  _max_num_q           = mt_degree;
+  _discoveredSoftRefs  = NEW_C_HEAP_ARRAY(DiscoveredList, _max_num_q * subclasses_of_ref);
   if (_discoveredSoftRefs == NULL) {
     vm_exit_during_initialization("Could not allocated RefProc Array");
   }
-  _discoveredWeakRefs    = &_discoveredSoftRefs[_num_q];
-  _discoveredFinalRefs   = &_discoveredWeakRefs[_num_q];
-  _discoveredPhantomRefs = &_discoveredFinalRefs[_num_q];
+  _discoveredWeakRefs    = &_discoveredSoftRefs[_max_num_q];
+  _discoveredFinalRefs   = &_discoveredWeakRefs[_max_num_q];
+  _discoveredPhantomRefs = &_discoveredFinalRefs[_max_num_q];
   assert(sentinel_ref() != NULL, "_sentinelRef is NULL");
   // Initialized all entries to _sentinelRef
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
         _discoveredSoftRefs[i].set_head(sentinel_ref());
     _discoveredSoftRefs[i].set_length(0);
   }
@@ -159,7 +160,7 @@
 #ifndef PRODUCT
 void ReferenceProcessor::verify_no_references_recorded() {
   guarantee(!_discovering_refs, "Discovering refs?");
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
     guarantee(_discoveredSoftRefs[i].empty(),
               "Found non-empty discovered list");
   }
@@ -167,7 +168,11 @@
 #endif
 
 void ReferenceProcessor::weak_oops_do(OopClosure* f) {
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+  // Should this instead be
+  // for (int i = 0; i < subclasses_of_ref; i++_ {
+  //   for (int j = 0; j < _num_q; j++) {
+  //     int index = i * _max_num_q + j;
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
     if (UseCompressedOops) {
       f->do_oop((narrowOop*)_discoveredSoftRefs[i].adr_head());
     } else {
@@ -395,7 +400,15 @@
     assert(work_id < (unsigned int)_ref_processor.num_q(), "Index out-of-bounds");
     // Simplest first cut: static partitioning.
     int index = work_id;
-    for (int j = 0; j < subclasses_of_ref; j++, index += _n_queues) {
+    // The increment on "index" must correspond to the maximum number of queues
+    // (n_queues) with which that ReferenceProcessor was created.  That
+    // is because of the "clever" way the discovered references lists were
+    // allocated and are indexed into.  That number is ParallelGCThreads
+    // currently.  Assert that.
+    assert(_n_queues == (int) ParallelGCThreads, "Different number not expected");
+    for (int j = 0;
+         j < subclasses_of_ref;
+         j++, index += _n_queues) {
       _ref_processor.enqueue_discovered_reflist(
         _refs_lists[index], _pending_list_addr);
       _refs_lists[index].set_head(_sentinel_ref);
@@ -410,11 +423,11 @@
   if (_processing_is_mt && task_executor != NULL) {
     // Parallel code
     RefProcEnqueueTask tsk(*this, _discoveredSoftRefs,
-                           pending_list_addr, sentinel_ref(), _num_q);
+                           pending_list_addr, sentinel_ref(), _max_num_q);
     task_executor->execute(tsk);
   } else {
     // Serial code: call the parent class's implementation
-    for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+    for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
       enqueue_discovered_reflist(_discoveredSoftRefs[i], pending_list_addr);
       _discoveredSoftRefs[i].set_head(sentinel_ref());
       _discoveredSoftRefs[i].set_length(0);
@@ -614,8 +627,9 @@
   complete_gc->do_void();
   NOT_PRODUCT(
     if (PrintGCDetails && TraceReferenceGC) {
-      gclog_or_tty->print(" Dropped %d dead Refs out of %d "
-        "discovered Refs by policy ", iter.removed(), iter.processed());
+      gclog_or_tty->print_cr(" Dropped %d dead Refs out of %d "
+        "discovered Refs by policy  list " INTPTR_FORMAT,
+        iter.removed(), iter.processed(), (address)refs_list.head());
     }
   )
 }
@@ -651,8 +665,9 @@
   }
   NOT_PRODUCT(
     if (PrintGCDetails && TraceReferenceGC) {
-      gclog_or_tty->print(" Dropped %d active Refs out of %d "
-        "Refs in discovered list ", iter.removed(), iter.processed());
+      gclog_or_tty->print_cr(" Dropped %d active Refs out of %d "
+        "Refs in discovered list " INTPTR_FORMAT,
+        iter.removed(), iter.processed(), (address)refs_list.head());
     }
   )
 }
@@ -689,8 +704,9 @@
   complete_gc->do_void();
   NOT_PRODUCT(
     if (PrintGCDetails && TraceReferenceGC) {
-      gclog_or_tty->print(" Dropped %d active Refs out of %d "
-        "Refs in discovered list ", iter.removed(), iter.processed());
+      gclog_or_tty->print_cr(" Dropped %d active Refs out of %d "
+        "Refs in discovered list " INTPTR_FORMAT,
+        iter.removed(), iter.processed(), (address)refs_list.head());
     }
   )
 }
@@ -704,6 +720,7 @@
                                    BoolObjectClosure* is_alive,
                                    OopClosure*        keep_alive,
                                    VoidClosure*       complete_gc) {
+  ResourceMark rm;
   DiscoveredListIterator iter(refs_list, keep_alive, is_alive);
   while (iter.has_next()) {
     iter.update_discovered();
@@ -743,8 +760,8 @@
 
 void ReferenceProcessor::abandon_partial_discovery() {
   // loop over the lists
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
-    if (TraceReferenceGC && PrintGCDetails && ((i % _num_q) == 0)) {
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
+    if (TraceReferenceGC && PrintGCDetails && ((i % _max_num_q) == 0)) {
       gclog_or_tty->print_cr(
         "\nAbandoning %s discovered list",
         list_name(i));
@@ -766,7 +783,9 @@
                     OopClosure& keep_alive,
                     VoidClosure& complete_gc)
   {
-    _ref_processor.process_phase1(_refs_lists[i], _policy,
+    Thread* thr = Thread::current();
+    int refs_list_index = ((WorkerThread*)thr)->id();
+    _ref_processor.process_phase1(_refs_lists[refs_list_index], _policy,
                                   &is_alive, &keep_alive, &complete_gc);
   }
 private:
@@ -802,6 +821,11 @@
                     OopClosure& keep_alive,
                     VoidClosure& complete_gc)
   {
+    // Don't use "refs_list_index" calculated in this way because
+    // balance_queues() has moved the Ref's into the first n queues.
+    // Thread* thr = Thread::current();
+    // int refs_list_index = ((WorkerThread*)thr)->id();
+    // _ref_processor.process_phase3(_refs_lists[refs_list_index], _clear_referent,
     _ref_processor.process_phase3(_refs_lists[i], _clear_referent,
                                   &is_alive, &keep_alive, &complete_gc);
   }
@@ -810,23 +834,47 @@
 };
 
 // Balances reference queues.
+// Move entries from all queues[0, 1, ..., _max_num_q-1] to
+// queues[0, 1, ..., _num_q-1] because only the first _num_q
+// corresponding to the active workers will be processed.
 void ReferenceProcessor::balance_queues(DiscoveredList ref_lists[])
 {
   // calculate total length
   size_t total_refs = 0;
-  for (int i = 0; i < _num_q; ++i) {
+  if (TraceReferenceGC && PrintGCDetails) {
+    gclog_or_tty->print_cr("\nBalance ref_lists ");
+  }
+
+  for (int i = 0; i < _max_num_q; ++i) {
     total_refs += ref_lists[i].length();
+    if (TraceReferenceGC && PrintGCDetails) {
+      gclog_or_tty->print("%d ", ref_lists[i].length());
+    }
+  }
+  if (TraceReferenceGC && PrintGCDetails) {
+    gclog_or_tty->print_cr(" = %d", total_refs);
   }
   size_t avg_refs = total_refs / _num_q + 1;
   int to_idx = 0;
-  for (int from_idx = 0; from_idx < _num_q; from_idx++) {
-    while (ref_lists[from_idx].length() > avg_refs) {
+  for (int from_idx = 0; from_idx < _max_num_q; from_idx++) {
+    bool move_all = false;
+    if (from_idx >= _num_q) {
+      move_all = ref_lists[from_idx].length() > 0;
+    }
+    while ((ref_lists[from_idx].length() > avg_refs) ||
+           move_all) {
       assert(to_idx < _num_q, "Sanity Check!");
       if (ref_lists[to_idx].length() < avg_refs) {
         // move superfluous refs
-        size_t refs_to_move =
-          MIN2(ref_lists[from_idx].length() - avg_refs,
-               avg_refs - ref_lists[to_idx].length());
+        size_t refs_to_move;
+        // Move all the Ref's if the from queue will not be processed.
+        if (move_all) {
+          refs_to_move = MIN2(ref_lists[from_idx].length(),
+                              avg_refs - ref_lists[to_idx].length());
+        } else {
+          refs_to_move = MIN2(ref_lists[from_idx].length() - avg_refs,
+                              avg_refs - ref_lists[to_idx].length());
+        }
         oop move_head = ref_lists[from_idx].head();
         oop move_tail = move_head;
         oop new_head  = move_head;
@@ -840,11 +888,35 @@
         ref_lists[to_idx].inc_length(refs_to_move);
         ref_lists[from_idx].set_head(new_head);
         ref_lists[from_idx].dec_length(refs_to_move);
+        if (ref_lists[from_idx].length() == 0) {
+          break;
+        }
       } else {
-        ++to_idx;
+        to_idx = (to_idx + 1) % _num_q;
       }
     }
   }
+#ifdef ASSERT
+  size_t balanced_total_refs = 0;
+  for (int i = 0; i < _max_num_q; ++i) {
+    balanced_total_refs += ref_lists[i].length();
+    if (TraceReferenceGC && PrintGCDetails) {
+      gclog_or_tty->print("%d ", ref_lists[i].length());
+    }
+  }
+  if (TraceReferenceGC && PrintGCDetails) {
+    gclog_or_tty->print_cr(" = %d", balanced_total_refs);
+    gclog_or_tty->flush();
+  }
+  assert(total_refs == balanced_total_refs, "Balancing was incomplete");
+#endif
+}
+
+void ReferenceProcessor::balance_all_queues() {
+  balance_queues(_discoveredSoftRefs);
+  balance_queues(_discoveredWeakRefs);
+  balance_queues(_discoveredFinalRefs);
+  balance_queues(_discoveredPhantomRefs);
 }
 
 void
@@ -857,8 +929,17 @@
   VoidClosure*                 complete_gc,
   AbstractRefProcTaskExecutor* task_executor)
 {
-  bool mt = task_executor != NULL && _processing_is_mt;
-  if (mt && ParallelRefProcBalancingEnabled) {
+  bool mt_processing = task_executor != NULL && _processing_is_mt;
+  // If discovery used MT and a dynamic number of GC threads, then
+  // the queues must be balanced for correctness if fewer than the
+  // maximum number of queues were used.  The number of queue used
+  // during discovery may be different than the number to be used
+  // for processing so don't depend of _num_q < _max_num_q as part
+  // of the test.
+  bool must_balance = _discovery_is_mt;
+
+  if ((mt_processing && ParallelRefProcBalancingEnabled) ||
+      must_balance) {
     balance_queues(refs_lists);
   }
   if (PrintReferenceGC && PrintGCDetails) {
@@ -875,7 +956,7 @@
   //   policy reasons. Keep alive the transitive closure of all
   //   such referents.
   if (policy != NULL) {
-    if (mt) {
+    if (mt_processing) {
       RefProcPhase1Task phase1(*this, refs_lists, policy, true /*marks_oops_alive*/);
       task_executor->execute(phase1);
     } else {
@@ -891,7 +972,7 @@
 
   // Phase 2:
   // . Traverse the list and remove any refs whose referents are alive.
-  if (mt) {
+  if (mt_processing) {
     RefProcPhase2Task phase2(*this, refs_lists, !discovery_is_atomic() /*marks_oops_alive*/);
     task_executor->execute(phase2);
   } else {
@@ -902,7 +983,7 @@
 
   // Phase 3:
   // . Traverse the list and process referents as appropriate.
-  if (mt) {
+  if (mt_processing) {
     RefProcPhase3Task phase3(*this, refs_lists, clear_referent, true /*marks_oops_alive*/);
     task_executor->execute(phase3);
   } else {
@@ -915,7 +996,11 @@
 
 void ReferenceProcessor::clean_up_discovered_references() {
   // loop over the lists
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+  // Should this instead be
+  // for (int i = 0; i < subclasses_of_ref; i++_ {
+  //   for (int j = 0; j < _num_q; j++) {
+  //     int index = i * _max_num_q + j;
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
     if (TraceReferenceGC && PrintGCDetails && ((i % _num_q) == 0)) {
       gclog_or_tty->print_cr(
         "\nScrubbing %s discovered list of Null referents",
@@ -976,7 +1061,7 @@
       id = next_id();
     }
   }
-  assert(0 <= id && id < _num_q, "Id is out-of-bounds (call Freud?)");
+  assert(0 <= id && id < _max_num_q, "Id is out-of-bounds (call Freud?)");
 
   // Get the discovered queue to which we will add
   DiscoveredList* list = NULL;
@@ -1001,6 +1086,10 @@
     default:
       ShouldNotReachHere();
   }
+  if (TraceReferenceGC && PrintGCDetails) {
+    gclog_or_tty->print_cr("Thread %d gets list " INTPTR_FORMAT,
+      id, list);
+  }
   return list;
 }
 
@@ -1243,7 +1332,7 @@
   {
     TraceTime tt("Preclean SoftReferences", PrintGCDetails && PrintReferenceGC,
               false, gclog_or_tty);
-    for (int i = 0; i < _num_q; i++) {
+    for (int i = 0; i < _max_num_q; i++) {
       if (yield->should_return()) {
         return;
       }
@@ -1340,15 +1429,16 @@
 
   NOT_PRODUCT(
     if (PrintGCDetails && PrintReferenceGC) {
-      gclog_or_tty->print(" Dropped %d Refs out of %d "
-        "Refs in discovered list ", iter.removed(), iter.processed());
+      gclog_or_tty->print_cr(" Dropped %d Refs out of %d "
+        "Refs in discovered list " INTPTR_FORMAT,
+        iter.removed(), iter.processed(), (address)refs_list.head());
     }
   )
 }
 
 const char* ReferenceProcessor::list_name(int i) {
-   assert(i >= 0 && i <= _num_q * subclasses_of_ref, "Out of bounds index");
-   int j = i / _num_q;
+   assert(i >= 0 && i <= _max_num_q * subclasses_of_ref, "Out of bounds index");
+   int j = i / _max_num_q;
    switch (j) {
      case 0: return "SoftRef";
      case 1: return "WeakRef";
@@ -1372,7 +1462,7 @@
 #ifndef PRODUCT
 void ReferenceProcessor::clear_discovered_references() {
   guarantee(!_discovering_refs, "Discovering refs?");
-  for (int i = 0; i < _num_q * subclasses_of_ref; i++) {
+  for (int i = 0; i < _max_num_q * subclasses_of_ref; i++) {
     oop obj = _discoveredSoftRefs[i].head();
     while (obj != sentinel_ref()) {
       oop next = java_lang_ref_Reference::discovered(obj);
--- a/src/share/vm/memory/referenceProcessor.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/referenceProcessor.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -85,8 +85,10 @@
 
   // The discovered ref lists themselves
 
-  // The MT'ness degree of the queues below
+  // The active MT'ness degree of the queues below
   int             _num_q;
+  // The maximum MT'ness degree of the queues below
+  int             _max_num_q;
   // Arrays of lists of oops, one per thread
   DiscoveredList* _discoveredSoftRefs;
   DiscoveredList* _discoveredWeakRefs;
@@ -95,6 +97,7 @@
 
  public:
   int num_q()                            { return _num_q; }
+  void set_mt_degree(int v)              { _num_q = v; }
   DiscoveredList* discovered_soft_refs() { return _discoveredSoftRefs; }
   static oop  sentinel_ref()             { return _sentinelRef; }
   static oop* adr_sentinel_ref()         { return &_sentinelRef; }
@@ -244,6 +247,7 @@
     _bs(NULL),
     _is_alive_non_header(NULL),
     _num_q(0),
+    _max_num_q(0),
     _processing_is_mt(false),
     _next_id(0)
   {}
@@ -312,6 +316,9 @@
   void weak_oops_do(OopClosure* f);       // weak roots
   static void oops_do(OopClosure* f);     // strong root(s)
 
+  // Balance each of the discovered lists.
+  void balance_all_queues();
+
   // Discover a Reference object, using appropriate discovery criteria
   bool discover_reference(oop obj, ReferenceType rt);
 
--- a/src/share/vm/memory/sharedHeap.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/sharedHeap.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -50,7 +50,8 @@
   _perm_gen(NULL), _rem_set(NULL),
   _strong_roots_parity(0),
   _process_strong_tasks(new SubTasksDone(SH_PS_NumElements)),
-  _workers(NULL), _n_par_threads(0)
+  _n_par_threads(0),
+  _workers(NULL)
 {
   if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
     vm_exit_during_initialization("Failed necessary allocation.");
@@ -60,11 +61,13 @@
       (UseConcMarkSweepGC && CMSParallelRemarkEnabled) ||
        UseG1GC) &&
       ParallelGCThreads > 0) {
-    _workers = new WorkGang("Parallel GC Threads", ParallelGCThreads,
+    _workers = new FlexibleWorkGang("Parallel GC Threads", ParallelGCThreads,
                             /* are_GC_task_threads */true,
                             /* are_ConcurrentGC_threads */false);
     if (_workers == NULL) {
       vm_exit_during_initialization("Failed necessary allocation.");
+    } else {
+      _workers->initialize_workers();
     }
   }
 }
@@ -77,8 +80,9 @@
 }
 
 void SharedHeap::set_par_threads(int t) {
+  assert(t == 0 || !UseSerialGC, "Cannot have parallel threads");
   _n_par_threads = t;
-  _process_strong_tasks->set_par_threads(t);
+  _process_strong_tasks->set_n_threads(t);
 }
 
 class AssertIsPermClosure: public OopClosure {
--- a/src/share/vm/memory/sharedHeap.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/memory/sharedHeap.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -38,6 +38,7 @@
 class ObjectClosure;
 class SubTasksDone;
 class WorkGang;
+class FlexibleWorkGang;
 class CollectorPolicy;
 class KlassHandle;
 
@@ -74,7 +75,7 @@
   int _strong_roots_parity;
 
   // If we're doing parallel GC, use this gang of threads.
-  WorkGang* _workers;
+  FlexibleWorkGang* _workers;
 
   // Number of parallel threads currently working on GC tasks.
   // O indicates use sequential code; 1 means use parallel code even with
@@ -189,7 +190,7 @@
     SO_CodeCache           = 0x10
   };
 
-  WorkGang* workers() const { return _workers; }
+  FlexibleWorkGang* workers() const { return _workers; }
 
   // Sets the number of parallel threads that will be doing tasks
   // (such as process strong roots) subsequently.
--- a/src/share/vm/runtime/globals.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/runtime/globals.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -273,6 +273,10 @@
 //    UnlockExperimentalVMOptions flag, which allows the control and
 //    modification of the experimental flags.
 //
+// Nota bene: neither diagnostic nor experimental options should be used casually,
+//    and they are not supported on production loads, except under explicit
+//    direction from support engineers.
+//
 // manageable flags are writeable external product flags.
 //    They are dynamically writeable through the JDK management interface
 //    (com.sun.management.HotSpotDiagnosticMXBean API) and also through JConsole.
@@ -634,6 +638,9 @@
   develop(bool, ZapJNIHandleArea, trueInDebug,                              \
           "Zap freed JNI handle space with 0xFEFEFEFE")                     \
                                                                             \
+  notproduct(bool, ZapStackSegments, trueInDebug,                           \
+             "Zap allocated/freed Stack segments with 0xFADFADED")          \
+                                                                            \
   develop(bool, ZapUnusedHeapArea, trueInDebug,                             \
           "Zap unused heap space with 0xBAADBABE")                          \
                                                                             \
@@ -1799,17 +1806,17 @@
   develop(uintx, PromotionFailureALotInterval, 5,                           \
           "Total collections between promotion failures alot")              \
                                                                             \
-  develop(intx, WorkStealingSleepMillis, 1,                                 \
+  experimental(intx, WorkStealingSleepMillis, 1,                            \
           "Sleep time when sleep is used for yields")                       \
                                                                             \
-  develop(uintx, WorkStealingYieldsBeforeSleep, 1000,                       \
+  experimental(uintx, WorkStealingYieldsBeforeSleep, 1000,                  \
           "Number of yields before a sleep is done during workstealing")    \
                                                                             \
-  develop(uintx, WorkStealingHardSpins, 4096,                               \
+  experimental(uintx, WorkStealingHardSpins, 4096,                          \
           "Number of iterations in a spin loop between checks on "          \
           "time out of hard spin")                                          \
                                                                             \
-  develop(uintx, WorkStealingSpinToYieldRatio, 10,                          \
+  experimental(uintx, WorkStealingSpinToYieldRatio, 10,                     \
           "Ratio of hard spins to calls to yield")                          \
                                                                             \
   product(uintx, PreserveMarkStackSize, 1024,                               \
--- a/src/share/vm/runtime/thread.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/runtime/thread.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1645,7 +1645,29 @@
   satb_mark_queue().flush();
   dirty_card_queue().flush();
 }
-#endif
+
+void JavaThread::initialize_queues() {
+  assert(!SafepointSynchronize::is_at_safepoint(),
+         "we should not be at a safepoint");
+
+  ObjPtrQueue& satb_queue = satb_mark_queue();
+  SATBMarkQueueSet& satb_queue_set = satb_mark_queue_set();
+  // The SATB queue should have been constructed with its active
+  // field set to false.
+  assert(!satb_queue.is_active(), "SATB queue should not be active");
+  assert(satb_queue.is_empty(), "SATB queue should be empty");
+  // If we are creating the thread during a marking cycle, we should
+  // set the active field of the SATB queue to true.
+  if (satb_queue_set.is_active()) {
+    satb_queue.set_active(true);
+  }
+
+  DirtyCardQueue& dirty_queue = dirty_card_queue();
+  // The dirty card queue should have been constructed with its
+  // active field set to true.
+  assert(dirty_queue.is_active(), "dirty card queue should be active");
+}
+#endif // !SERIALGC
 
 void JavaThread::cleanup_failed_attach_current_thread() {
   if (get_thread_profiler() != NULL) {
@@ -3627,6 +3649,10 @@
 void Threads::add(JavaThread* p, bool force_daemon) {
   // The threads lock must be owned at this point
   assert_locked_or_safepoint(Threads_lock);
+
+  // See the comment for this method in thread.hpp for its purpose and
+  // why it is called here.
+  p->initialize_queues();
   p->set_next(_thread_list);
   _thread_list = p;
   _number_of_threads++;
--- a/src/share/vm/runtime/thread.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/runtime/thread.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1490,6 +1490,29 @@
   }
 #endif // !SERIALGC
 
+  // This method initializes the SATB and dirty card queues before a
+  // JavaThread is added to the Java thread list. Right now, we don't
+  // have to do anything to the dirty card queue (it should have been
+  // activated when the thread was created), but we have to activate
+  // the SATB queue if the thread is created while a marking cycle is
+  // in progress. The activation / de-activation of the SATB queues at
+  // the beginning / end of a marking cycle is done during safepoints
+  // so we have to make sure this method is called outside one to be
+  // able to safely read the active field of the SATB queue set. Right
+  // now, it is called just before the thread is added to the Java
+  // thread list in the Threads::add() method. That method is holding
+  // the Threads_lock which ensures we are outside a safepoint. We
+  // cannot do the obvious and set the active field of the SATB queue
+  // when the thread is created given that, in some cases, safepoints
+  // might happen between the JavaThread constructor being called and the
+  // thread being added to the Java thread list (an example of this is
+  // when the structure for the DestroyJavaVM thread is created).
+#ifndef SERIALGC
+  void initialize_queues();
+#else // !SERIALGC
+  void initialize_queues() { }
+#endif // !SERIALGC
+
   // Machine dependent stuff
   #include "incls/_thread_pd.hpp.incl"
 
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/share/vm/utilities/stack.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -0,0 +1,204 @@
+/*
+ * Copyright 2009 Sun Microsystems, Inc.  All Rights Reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ */
+
+// Class Stack (below) grows and shrinks by linking together "segments" which
+// are allocated on demand.  Segments are arrays of the element type (E) plus an
+// extra pointer-sized field to store the segment link.  Recently emptied
+// segments are kept in a cache and reused.
+//
+// Notes/caveats:
+//
+// The size of an element must either evenly divide the size of a pointer or be
+// a multiple of the size of a pointer.
+//
+// Destructors are not called for elements popped off the stack, so element
+// types which rely on destructors for things like reference counting will not
+// work properly.
+//
+// Class Stack allocates segments from the C heap.  However, two protected
+// virtual methods are used to alloc/free memory which subclasses can override:
+//
+//      virtual void* alloc(size_t bytes);
+//      virtual void  free(void* addr, size_t bytes);
+//
+// The alloc() method must return storage aligned for any use.  The
+// implementation in class Stack assumes that alloc() will terminate the process
+// if the allocation fails.
+
+template <class E> class StackIterator;
+
+// StackBase holds common data/methods that don't depend on the element type,
+// factored out to reduce template code duplication.
+class StackBase
+{
+public:
+  size_t segment_size()   const { return _seg_size; } // Elements per segment.
+  size_t max_size()       const { return _max_size; } // Max elements allowed.
+  size_t max_cache_size() const { return _max_cache_size; } // Max segments
+                                                            // allowed in cache.
+
+  size_t cache_size() const { return _cache_size; }   // Segments in the cache.
+
+protected:
+  // The ctor arguments correspond to the like-named functions above.
+  // segment_size:    number of items per segment
+  // max_cache_size:  maxmium number of *segments* to cache
+  // max_size:        maximum number of items allowed, rounded to a multiple of
+  //                  the segment size (0 == unlimited)
+  inline StackBase(size_t segment_size, size_t max_cache_size, size_t max_size);
+
+  // Round max_size to a multiple of the segment size.  Treat 0 as unlimited.
+  static inline size_t adjust_max_size(size_t max_size, size_t seg_size);
+
+protected:
+  const size_t _seg_size;       // Number of items per segment.
+  const size_t _max_size;       // Maximum number of items allowed in the stack.
+  const size_t _max_cache_size; // Maximum number of segments to cache.
+  size_t       _cur_seg_size;   // Number of items in the current segment.
+  size_t       _full_seg_size;  // Number of items in already-filled segments.
+  size_t       _cache_size;     // Number of segments in the cache.
+};
+
+#ifdef __GNUC__
+#define inline
+#endif // __GNUC__
+
+template <class E>
+class Stack:  public StackBase
+{
+public:
+  friend class StackIterator<E>;
+
+  // segment_size:    number of items per segment
+  // max_cache_size:  maxmium number of *segments* to cache
+  // max_size:        maximum number of items allowed, rounded to a multiple of
+  //                  the segment size (0 == unlimited)
+  inline Stack(size_t segment_size = default_segment_size(),
+               size_t max_cache_size = 4, size_t max_size = 0);
+  inline ~Stack() { clear(true); }
+
+  inline bool is_empty() const { return _cur_seg == NULL; }
+  inline bool is_full()  const { return _full_seg_size >= max_size(); }
+
+  // Performance sensitive code should use is_empty() instead of size() == 0 and
+  // is_full() instead of size() == max_size().  Using a conditional here allows
+  // just one var to be updated when pushing/popping elements instead of two;
+  // _full_seg_size is updated only when pushing/popping segments.
+  inline size_t size() const {
+    return is_empty() ? 0 : _full_seg_size + _cur_seg_size;
+  }
+
+  inline void push(E elem);
+  inline E    pop();
+
+  // Clear everything from the stack, releasing the associated memory.  If
+  // clear_cache is true, also release any cached segments.
+  void clear(bool clear_cache = false);
+
+  static inline size_t default_segment_size();
+
+protected:
+  // Each segment includes space for _seg_size elements followed by a link
+  // (pointer) to the previous segment; the space is allocated as a single block
+  // of size segment_bytes().  _seg_size is rounded up if necessary so the link
+  // is properly aligned.  The C struct for the layout would be:
+  //
+  // struct segment {
+  //   E     elements[_seg_size];
+  //   E*    link;
+  // };
+
+  // Round up seg_size to keep the link field aligned.
+  static inline size_t adjust_segment_size(size_t seg_size);
+
+  // Methods for allocation size and getting/setting the link.
+  inline size_t link_offset() const;              // Byte offset of link field.
+  inline size_t segment_bytes() const;            // Segment size in bytes.
+  inline E**    link_addr(E* seg) const;          // Address of the link field.
+  inline E*     get_link(E* seg) const;           // Extract the link from seg.
+  inline E*     set_link(E* new_seg, E* old_seg); // new_seg.link = old_seg.
+
+  virtual E*    alloc(size_t bytes);
+  virtual void  free(E* addr, size_t bytes);
+
+  void push_segment();
+  void pop_segment();
+
+  void free_segments(E* seg);          // Free all segments in the list.
+  inline void reset(bool reset_cache); // Reset all data fields.
+
+  DEBUG_ONLY(void verify(bool at_empty_transition) const;)
+  DEBUG_ONLY(void zap_segment(E* seg, bool zap_link_field) const;)
+
+private:
+  E* _cur_seg;    // Current segment.
+  E* _cache;      // Segment cache to avoid ping-ponging.
+};
+
+template <class E> class ResourceStack:  public Stack<E>, public ResourceObj
+{
+public:
+  // If this class becomes widely used, it may make sense to save the Thread
+  // and use it when allocating segments.
+  ResourceStack(size_t segment_size = Stack<E>::default_segment_size()):
+    Stack<E>(segment_size, max_uintx)
+    { }
+
+  // Set the segment pointers to NULL so the parent dtor does not free them;
+  // that must be done by the ResourceMark code.
+  ~ResourceStack() { Stack<E>::reset(true); }
+
+protected:
+  virtual E*   alloc(size_t bytes);
+  virtual void free(E* addr, size_t bytes);
+
+private:
+  void clear(bool clear_cache = false);
+};
+
+template <class E>
+class StackIterator: public StackObj
+{
+public:
+  StackIterator(Stack<E>& stack): _stack(stack) { sync(); }
+
+  Stack<E>& stack() const { return _stack; }
+
+  bool is_empty() const { return _cur_seg == NULL; }
+
+  E  next() { return *next_addr(); }
+  E* next_addr();
+
+  void sync(); // Sync the iterator's state to the stack's current state.
+
+private:
+  Stack<E>& _stack;
+  size_t    _cur_seg_size;
+  E*        _cur_seg;
+  size_t    _full_seg_size;
+};
+
+#ifdef __GNUC__
+#undef inline
+#endif // __GNUC__
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/src/share/vm/utilities/stack.inline.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -0,0 +1,273 @@
+/*
+ * Copyright 2009 Sun Microsystems, Inc.  All Rights Reserved.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
+ *
+ * This code is free software; you can redistribute it and/or modify it
+ * under the terms of the GNU General Public License version 2 only, as
+ * published by the Free Software Foundation.
+ *
+ * This code is distributed in the hope that it will be useful, but WITHOUT
+ * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
+ * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
+ * version 2 for more details (a copy is included in the LICENSE file that
+ * accompanied this code).
+ *
+ * You should have received a copy of the GNU General Public License version
+ * 2 along with this work; if not, write to the Free Software Foundation,
+ * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ */
+
+StackBase::StackBase(size_t segment_size, size_t max_cache_size,
+                     size_t max_size):
+  _seg_size(segment_size),
+  _max_cache_size(max_cache_size),
+  _max_size(adjust_max_size(max_size, segment_size))
+{
+  assert(_max_size % _seg_size == 0, "not a multiple");
+}
+
+size_t StackBase::adjust_max_size(size_t max_size, size_t seg_size)
+{
+  assert(seg_size > 0, "cannot be 0");
+  assert(max_size >= seg_size || max_size == 0, "max_size too small");
+  const size_t limit = max_uintx - (seg_size - 1);
+  if (max_size == 0 || max_size > limit) {
+    max_size = limit;
+  }
+  return (max_size + seg_size - 1) / seg_size * seg_size;
+}
+
+template <class E>
+Stack<E>::Stack(size_t segment_size, size_t max_cache_size, size_t max_size):
+  StackBase(adjust_segment_size(segment_size), max_cache_size, max_size)
+{
+  reset(true);
+}
+
+template <class E>
+void Stack<E>::push(E item)
+{
+  assert(!is_full(), "pushing onto a full stack");
+  if (_cur_seg_size == _seg_size) {
+    push_segment();
+  }
+  _cur_seg[_cur_seg_size] = item;
+  ++_cur_seg_size;
+}
+
+template <class E>
+E Stack<E>::pop()
+{
+  assert(!is_empty(), "popping from an empty stack");
+  if (_cur_seg_size == 1) {
+    E tmp = _cur_seg[--_cur_seg_size];
+    pop_segment();
+    return tmp;
+  }
+  return _cur_seg[--_cur_seg_size];
+}
+
+template <class E>
+void Stack<E>::clear(bool clear_cache)
+{
+  free_segments(_cur_seg);
+  if (clear_cache) free_segments(_cache);
+  reset(clear_cache);
+}
+
+template <class E>
+size_t Stack<E>::default_segment_size()
+{
+  // Number of elements that fit in 4K bytes minus the size of two pointers
+  // (link field and malloc header).
+  return (4096 - 2 * sizeof(E*)) / sizeof(E);
+}
+
+template <class E>
+size_t Stack<E>::adjust_segment_size(size_t seg_size)
+{
+  const size_t elem_sz = sizeof(E);
+  const size_t ptr_sz = sizeof(E*);
+  assert(elem_sz % ptr_sz == 0 || ptr_sz % elem_sz == 0, "bad element size");
+  if (elem_sz < ptr_sz) {
+    return align_size_up(seg_size * elem_sz, ptr_sz) / elem_sz;
+  }
+  return seg_size;
+}
+
+template <class E>
+size_t Stack<E>::link_offset() const
+{
+  return align_size_up(_seg_size * sizeof(E), sizeof(E*));
+}
+
+template <class E>
+size_t Stack<E>::segment_bytes() const
+{
+  return link_offset() + sizeof(E*);
+}
+
+template <class E>
+E** Stack<E>::link_addr(E* seg) const
+{
+  return (E**) ((char*)seg + link_offset());
+}
+
+template <class E>
+E* Stack<E>::get_link(E* seg) const
+{
+  return *link_addr(seg);
+}
+
+template <class E>
+E* Stack<E>::set_link(E* new_seg, E* old_seg)
+{
+  *link_addr(new_seg) = old_seg;
+  return new_seg;
+}
+
+template <class E>
+E* Stack<E>::alloc(size_t bytes)
+{
+  return (E*) NEW_C_HEAP_ARRAY(char, bytes);
+}
+
+template <class E>
+void Stack<E>::free(E* addr, size_t bytes)
+{
+  FREE_C_HEAP_ARRAY(char, (char*) addr);
+}
+
+template <class E>
+void Stack<E>::push_segment()
+{
+  assert(_cur_seg_size == _seg_size, "current segment is not full");
+  E* next;
+  if (_cache_size > 0) {
+    // Use a cached segment.
+    next = _cache;
+    _cache = get_link(_cache);
+    --_cache_size;
+  } else {
+    next = alloc(segment_bytes());
+    DEBUG_ONLY(zap_segment(next, true);)
+  }
+  const bool at_empty_transition = is_empty();
+  _cur_seg = set_link(next, _cur_seg);
+  _cur_seg_size = 0;
+  _full_seg_size += at_empty_transition ? 0 : _seg_size;
+  DEBUG_ONLY(verify(at_empty_transition);)
+}
+
+template <class E>
+void Stack<E>::pop_segment()
+{
+  assert(_cur_seg_size == 0, "current segment is not empty");
+  E* const prev = get_link(_cur_seg);
+  if (_cache_size < _max_cache_size) {
+    // Add the current segment to the cache.
+    DEBUG_ONLY(zap_segment(_cur_seg, false);)
+    _cache = set_link(_cur_seg, _cache);
+    ++_cache_size;
+  } else {
+    DEBUG_ONLY(zap_segment(_cur_seg, true);)
+    free(_cur_seg, segment_bytes());
+  }
+  const bool at_empty_transition = prev == NULL;
+  _cur_seg = prev;
+  _cur_seg_size = _seg_size;
+  _full_seg_size -= at_empty_transition ? 0 : _seg_size;
+  DEBUG_ONLY(verify(at_empty_transition);)
+}
+
+template <class E>
+void Stack<E>::free_segments(E* seg)
+{
+  const size_t bytes = segment_bytes();
+  while (seg != NULL) {
+    E* const prev = get_link(seg);
+    free(seg, bytes);
+    seg = prev;
+  }
+}
+
+template <class E>
+void Stack<E>::reset(bool reset_cache)
+{
+  _cur_seg_size = _seg_size; // So push() will alloc a new segment.
+  _full_seg_size = 0;
+  _cur_seg = NULL;
+  if (reset_cache) {
+    _cache_size = 0;
+    _cache = NULL;
+  }
+}
+
+#ifdef ASSERT
+template <class E>
+void Stack<E>::verify(bool at_empty_transition) const
+{
+  assert(size() <= max_size(), "stack exceeded bounds");
+  assert(cache_size() <= max_cache_size(), "cache exceeded bounds");
+  assert(_cur_seg_size <= segment_size(), "segment index exceeded bounds");
+
+  assert(_full_seg_size % _seg_size == 0, "not a multiple");
+  assert(at_empty_transition || is_empty() == (size() == 0), "mismatch");
+  assert((_cache == NULL) == (cache_size() == 0), "mismatch");
+
+  if (is_empty()) {
+    assert(_cur_seg_size == segment_size(), "sanity");
+  }
+}
+
+template <class E>
+void Stack<E>::zap_segment(E* seg, bool zap_link_field) const
+{
+  if (!ZapStackSegments) return;
+  const size_t zap_bytes = segment_bytes() - (zap_link_field ? 0 : sizeof(E*));
+  uint32_t* cur = (uint32_t*)seg;
+  const uint32_t* end = cur + zap_bytes / sizeof(uint32_t);
+  while (cur < end) {
+    *cur++ = 0xfadfaded;
+  }
+}
+#endif
+
+template <class E>
+E* ResourceStack<E>::alloc(size_t bytes)
+{
+  return (E*) resource_allocate_bytes(bytes);
+}
+
+template <class E>
+void ResourceStack<E>::free(E* addr, size_t bytes)
+{
+  resource_free_bytes((char*) addr, bytes);
+}
+
+template <class E>
+void StackIterator<E>::sync()
+{
+  _full_seg_size = _stack._full_seg_size;
+  _cur_seg_size = _stack._cur_seg_size;
+  _cur_seg = _stack._cur_seg;
+}
+
+template <class E>
+E* StackIterator<E>::next_addr()
+{
+  assert(!is_empty(), "no items left");
+  if (_cur_seg_size == 1) {
+    E* addr = _cur_seg;
+    _cur_seg = _stack.get_link(_cur_seg);
+    _cur_seg_size = _stack.segment_size();
+    _full_seg_size -= _stack.segment_size();
+    return addr;
+  }
+  return _cur_seg + --_cur_seg_size;
+}
--- a/src/share/vm/utilities/taskqueue.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/taskqueue.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -144,6 +144,7 @@
 
 bool
 ParallelTaskTerminator::offer_termination(TerminatorTerminator* terminator) {
+  assert(_n_threads > 0, "Initialization is incorrect");
   assert(_offered_termination < _n_threads, "Invariant");
   Atomic::inc(&_offered_termination);
 
@@ -255,3 +256,9 @@
     _index < objArrayOop(_obj)->length();
 }
 #endif // ASSERT
+
+void ParallelTaskTerminator::reset_for_reuse(int n_threads) {
+  reset_for_reuse();
+  _n_threads = n_threads;
+}
+
--- a/src/share/vm/utilities/taskqueue.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/taskqueue.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -305,6 +305,12 @@
   return false;
 }
 
+// pop_local_slow() is done by the owning thread and is trying to
+// get the last task in the queue.  It will compete with pop_global()
+// that will be used by other threads.  The tag age is incremented
+// whenever the queue goes empty which it will do here if this thread
+// gets the last task or in pop_global() if the queue wraps (top == 0
+// and pop_global() succeeds, see pop_global()).
 template<class E, unsigned int N>
 bool GenericTaskQueue<E, N>::pop_local_slow(uint localBot, Age oldAge) {
   // This queue was observed to contain exactly one element; either this
@@ -366,75 +372,47 @@
 // OverflowTaskQueue is a TaskQueue that also includes an overflow stack for
 // elements that do not fit in the TaskQueue.
 //
-// Three methods from super classes are overridden:
+// This class hides two methods from super classes:
 //
-// initialize() - initialize the super classes and create the overflow stack
 // push() - push onto the task queue or, if that fails, onto the overflow stack
 // is_empty() - return true if both the TaskQueue and overflow stack are empty
 //
-// Note that size() is not overridden--it returns the number of elements in the
+// Note that size() is not hidden--it returns the number of elements in the
 // TaskQueue, and does not include the size of the overflow stack.  This
 // simplifies replacement of GenericTaskQueues with OverflowTaskQueues.
 template<class E, unsigned int N = TASKQUEUE_SIZE>
 class OverflowTaskQueue: public GenericTaskQueue<E, N>
 {
 public:
-  typedef GrowableArray<E>       overflow_t;
+  typedef Stack<E>               overflow_t;
   typedef GenericTaskQueue<E, N> taskqueue_t;
 
   TASKQUEUE_STATS_ONLY(using taskqueue_t::stats;)
 
-  OverflowTaskQueue();
-  ~OverflowTaskQueue();
-  void initialize();
-
-  inline overflow_t* overflow_stack() const { return _overflow_stack; }
-
   // Push task t onto the queue or onto the overflow stack.  Return true.
   inline bool push(E t);
 
   // Attempt to pop from the overflow stack; return true if anything was popped.
   inline bool pop_overflow(E& t);
 
+  inline overflow_t* overflow_stack() { return &_overflow_stack; }
+
   inline bool taskqueue_empty() const { return taskqueue_t::is_empty(); }
-  inline bool overflow_empty()  const { return overflow_stack()->is_empty(); }
+  inline bool overflow_empty()  const { return _overflow_stack.is_empty(); }
   inline bool is_empty()        const {
     return taskqueue_empty() && overflow_empty();
   }
 
 private:
-  overflow_t* _overflow_stack;
+  overflow_t _overflow_stack;
 };
 
 template <class E, unsigned int N>
-OverflowTaskQueue<E, N>::OverflowTaskQueue()
-{
-  _overflow_stack = NULL;
-}
-
-template <class E, unsigned int N>
-OverflowTaskQueue<E, N>::~OverflowTaskQueue()
-{
-  if (_overflow_stack != NULL) {
-    delete _overflow_stack;
-    _overflow_stack = NULL;
-  }
-}
-
-template <class E, unsigned int N>
-void OverflowTaskQueue<E, N>::initialize()
-{
-  taskqueue_t::initialize();
-  assert(_overflow_stack == NULL, "memory leak");
-  _overflow_stack = new (ResourceObj::C_HEAP) GrowableArray<E>(10, true);
-}
-
-template <class E, unsigned int N>
 bool OverflowTaskQueue<E, N>::push(E t)
 {
   if (!taskqueue_t::push(t)) {
     overflow_stack()->push(t);
-    TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->length()));
+    TASKQUEUE_STATS_ONLY(stats.record_overflow(overflow_stack()->size()));
   }
   return true;
 }
@@ -637,6 +615,9 @@
   // in an MT-safe manner, once the previous round of use of
   // the terminator is finished.
   void reset_for_reuse();
+  // Same as above but the number of parallel threads is set to the
+  // given number.
+  void reset_for_reuse(int n_threads);
 
 #ifdef TRACESPINNING
   static uint total_yields() { return _total_yields; }
@@ -782,3 +763,4 @@
 
 typedef OverflowTaskQueue<size_t>             RegionTaskQueue;
 typedef GenericTaskQueueSet<RegionTaskQueue>  RegionTaskQueueSet;
+
--- a/src/share/vm/utilities/workgroup.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/workgroup.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2007, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -53,28 +53,52 @@
                    int         workers,
                    bool        are_GC_task_threads,
                    bool        are_ConcurrentGC_threads) :
-  AbstractWorkGang(name, are_GC_task_threads, are_ConcurrentGC_threads)
-{
+  AbstractWorkGang(name, are_GC_task_threads, are_ConcurrentGC_threads) {
   // Save arguments.
   _total_workers = workers;
+}
+
+GangWorker* WorkGang::allocate_worker(int which) {
+  GangWorker* new_worker = new GangWorker(this, which);
+  return new_worker;
+}
+
+// The current implementation will exit if the allocation
+// of any worker fails.  Still, return a boolean so that
+// a future implementation can possibly do a partial
+// initialization of the workers and report such to the
+// caller.
+bool WorkGang::initialize_workers() {
 
   if (TraceWorkGang) {
-    tty->print_cr("Constructing work gang %s with %d threads", name, workers);
+    tty->print_cr("Constructing work gang %s with %d threads",
+                  name(),
+                  total_workers());
   }
-  _gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, workers);
+  _gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, total_workers());
   if (gang_workers() == NULL) {
     vm_exit_out_of_memory(0, "Cannot create GangWorker array.");
+    return false;
+  }
+  os::ThreadType worker_type;
+  if (are_ConcurrentGC_threads()) {
+    worker_type = os::cgc_thread;
+  } else {
+    worker_type = os::pgc_thread;
   }
   for (int worker = 0; worker < total_workers(); worker += 1) {
-    GangWorker* new_worker = new GangWorker(this, worker);
+    GangWorker* new_worker = allocate_worker(worker);
     assert(new_worker != NULL, "Failed to allocate GangWorker");
     _gang_workers[worker] = new_worker;
-    if (new_worker == NULL || !os::create_thread(new_worker, os::pgc_thread))
+    if (new_worker == NULL || !os::create_thread(new_worker, worker_type)) {
       vm_exit_out_of_memory(0, "Cannot create worker GC thread. Out of system resources.");
+      return false;
+    }
     if (!DisableStartThread) {
       os::start_thread(new_worker);
     }
   }
+  return true;
 }
 
 AbstractWorkGang::~AbstractWorkGang() {
@@ -383,7 +407,7 @@
   return _tasks != NULL;
 }
 
-void SubTasksDone::set_par_threads(int t) {
+void SubTasksDone::set_n_threads(int t) {
 #ifdef ASSERT
   assert(_claimed == 0 || _threads_completed == _n_threads,
          "should not be called while tasks are being processed!");
--- a/src/share/vm/utilities/workgroup.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/workgroup.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2002, 2009, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -29,6 +29,7 @@
 class YieldingFlexibleGangWorker;
 class YieldingFlexibleGangTask;
 class WorkData;
+class AbstractWorkGang;
 
 // An abstract task to be worked on by a gang.
 // You subclass this to supply your own work() method
@@ -38,6 +39,13 @@
   // The argument tells you which member of the gang you are.
   virtual void work(int i) = 0;
 
+  // This method configures the task for proper termination.
+  // Some tasks do not have any requirements on termination
+  // and may inherit this method that does nothing.  Some
+  // tasks do some coordination on termination and override
+  // this method to implement that coordination.
+  virtual void set_for_termination(int active_workers) {};
+
   // Debugging accessor for the name.
   const char* name() const PRODUCT_RETURN_(return NULL;);
   int counter() { return _counter; }
@@ -64,6 +72,18 @@
   virtual ~AbstractGangTask() { }
 };
 
+class AbstractGangTaskWOopQueues : public AbstractGangTask {
+  OopTaskQueueSet*       _queues;
+  ParallelTaskTerminator _terminator;
+ public:
+  AbstractGangTaskWOopQueues(const char* name, OopTaskQueueSet* queues) :
+    AbstractGangTask(name), _queues(queues), _terminator(0, _queues) {}
+  ParallelTaskTerminator* terminator() { return &_terminator; }
+  virtual void set_for_termination(int active_workers) {
+    terminator()->reset_for_reuse(active_workers);
+  }
+  OopTaskQueueSet* queues() { return _queues; }
+};
 
 // Class AbstractWorkGang:
 // An abstract class representing a gang of workers.
@@ -114,6 +134,9 @@
   int total_workers() const {
     return _total_workers;
   }
+  virtual int active_workers() const {
+    return _total_workers;
+  }
   bool terminate() const {
     return _terminate;
   }
@@ -199,6 +222,13 @@
            bool are_GC_task_threads, bool are_ConcurrentGC_threads);
   // Run a task, returns when the task is done (or terminated).
   virtual void run_task(AbstractGangTask* task);
+  void run_task(AbstractGangTask* task, uint no_of_parallel_workers);
+  // Allocate a worker and return a pointer to it.
+  virtual GangWorker* allocate_worker(int which);
+  // Initialize workers in the gang.  Return true if initialization
+  // succeeded. The type of the worker can be overridden in a derived
+  // class with the appropriate implementation of allocate_worker().
+  bool initialize_workers();
 };
 
 // Class GangWorker:
@@ -226,6 +256,34 @@
   AbstractWorkGang* gang() const { return _gang; }
 };
 
+class FlexibleWorkGang: public WorkGang {
+ protected:
+  int _active_workers;
+ public:
+  // Constructor and destructor.
+  FlexibleWorkGang(const char* name, int workers,
+                   bool are_GC_task_threads,
+                   bool  are_ConcurrentGC_threads) :
+    WorkGang(name, workers, are_GC_task_threads, are_ConcurrentGC_threads) {
+    _active_workers = ParallelGCThreads;
+  };
+  // Accessors for fields
+  virtual int active_workers() const { return _active_workers; }
+  void set_active_workers(int v) { _active_workers = v; }
+};
+
+// Work gangs in garbage collectors: 2009-06-10
+//
+// SharedHeap - work gang for stop-the-world parallel collection.
+//   Used by
+//     ParNewGeneration
+//     CMSParRemarkTask
+//     CMSRefProcTaskExecutor
+//     G1CollectedHeap
+//     G1ParFinalCountTask
+// ConcurrentMark
+// CMSCollector
+
 // A class that acts as a synchronisation barrier. Workers enter
 // the barrier and must wait until all other workers have entered
 // before any of them may leave.
@@ -271,7 +329,7 @@
   int _n_threads;
   jint _threads_completed;
 #ifdef ASSERT
-  jint _claimed;
+  volatile jint _claimed;
 #endif
 
   // Set all tasks to unclaimed.
@@ -286,9 +344,10 @@
   // True iff the object is in a valid state.
   bool valid();
 
-  // Set the number of parallel threads doing the tasks to "t".  Can only
+  // Get/set the number of parallel threads doing the tasks to "t".  Can only
   // be called before tasks start or after they are complete.
-  void set_par_threads(int t);
+  int n_threads() { return _n_threads; }
+  void set_n_threads(int t);
 
   // Returns "false" if the task "t" is unclaimed, and ensures that task is
   // claimed.  The task "t" is required to be within the range of "this".
@@ -315,13 +374,17 @@
 protected:
   jint _n_tasks;     // Total number of tasks available.
   jint _n_claimed;   // Number of tasks claimed.
+  // _n_threads is used to determine when a sub task is done.
+  // See comments on SubTasksDone::_n_threads
   jint _n_threads;   // Total number of parallel threads.
   jint _n_completed; // Number of completed threads.
 
   void clear();
 
 public:
-  SequentialSubTasksDone() { clear(); }
+  SequentialSubTasksDone() {
+    clear();
+  }
   ~SequentialSubTasksDone() {}
 
   // True iff the object is in a valid state.
@@ -330,11 +393,12 @@
   // number of tasks
   jint n_tasks() const { return _n_tasks; }
 
-  // Set the number of parallel threads doing the tasks to t.
+  // Get/set the number of parallel threads doing the tasks to t.
   // Should be called before the task starts but it is safe
   // to call this once a task is running provided that all
   // threads agree on the number of threads.
-  void set_par_threads(int t) { _n_threads = t; }
+  int n_threads() { return _n_threads; }
+  void set_n_threads(int t) { _n_threads = t; }
 
   // Set the number of tasks to be claimed to t. As above,
   // should be called before the tasks start but it is safe
--- a/src/share/vm/utilities/yieldingWorkgroup.cpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/yieldingWorkgroup.cpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2005, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2005, 2010 Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -32,29 +32,13 @@
 
 YieldingFlexibleWorkGang::YieldingFlexibleWorkGang(
   const char* name, int workers, bool are_GC_task_threads) :
-  AbstractWorkGang(name, are_GC_task_threads, false) {
-  // Save arguments.
-  _total_workers = workers;
-  assert(_total_workers > 0, "Must have more than 1 worker");
-
-  _yielded_workers = 0;
+  FlexibleWorkGang(name, workers, are_GC_task_threads, false),
+    _yielded_workers(0) {}
 
-  if (TraceWorkGang) {
-    tty->print_cr("Constructing work gang %s with %d threads", name, workers);
-  }
-  _gang_workers = NEW_C_HEAP_ARRAY(GangWorker*, workers);
-  assert(gang_workers() != NULL, "Failed to allocate gang workers");
-  for (int worker = 0; worker < total_workers(); worker += 1) {
-    YieldingFlexibleGangWorker* new_worker =
-      new YieldingFlexibleGangWorker(this, worker);
-    assert(new_worker != NULL, "Failed to allocate YieldingFlexibleGangWorker");
-    _gang_workers[worker] = new_worker;
-    if (new_worker == NULL || !os::create_thread(new_worker, os::pgc_thread))
-      vm_exit_out_of_memory(0, "Cannot create worker GC thread. Out of system resources.");
-    if (!DisableStartThread) {
-      os::start_thread(new_worker);
-    }
-  }
+GangWorker* YieldingFlexibleWorkGang::allocate_worker(int which) {
+  YieldingFlexibleGangWorker* new_member =
+      new YieldingFlexibleGangWorker(this, which);
+  return (YieldingFlexibleGangWorker*) new_member;
 }
 
 // Run a task; returns when the task is done, or the workers yield,
@@ -142,6 +126,7 @@
     _active_workers = total_workers();
   }
   new_task->set_actual_size(_active_workers);
+  new_task->set_for_termination(_active_workers);
 
   assert(_started_workers == 0, "Tabula rasa non");
   assert(_finished_workers == 0, "Tabula rasa non");
@@ -161,22 +146,22 @@
   for (Status status = yielding_task()->status();
        status != COMPLETED && status != YIELDED && status != ABORTED;
        status = yielding_task()->status()) {
-    assert(started_workers() <= active_workers(), "invariant");
-    assert(finished_workers() <= active_workers(), "invariant");
-    assert(yielded_workers() <= active_workers(), "invariant");
+    assert(started_workers() <= total_workers(), "invariant");
+    assert(finished_workers() <= total_workers(), "invariant");
+    assert(yielded_workers() <= total_workers(), "invariant");
     monitor()->wait(Mutex::_no_safepoint_check_flag);
   }
   switch (yielding_task()->status()) {
     case COMPLETED:
     case ABORTED: {
-      assert(finished_workers() == active_workers(), "Inconsistent status");
+      assert(finished_workers() == total_workers(), "Inconsistent status");
       assert(yielded_workers() == 0, "Invariant");
       reset();   // for next task; gang<->task binding released
       break;
     }
     case YIELDED: {
       assert(yielded_workers() > 0, "Invariant");
-      assert(yielded_workers() + finished_workers() == active_workers(),
+      assert(yielded_workers() + finished_workers() == total_workers(),
              "Inconsistent counts");
       break;
     }
@@ -208,7 +193,6 @@
 void YieldingFlexibleWorkGang::reset() {
   _started_workers  = 0;
   _finished_workers = 0;
-  _active_workers   = 0;
   yielding_task()->set_gang(NULL);
   _task = NULL;    // unbind gang from task
 }
@@ -216,7 +200,7 @@
 void YieldingFlexibleWorkGang::yield() {
   assert(task() != NULL, "Inconsistency; should have task binding");
   MutexLockerEx ml(monitor(), Mutex::_no_safepoint_check_flag);
-  assert(yielded_workers() < active_workers(), "Consistency check");
+  assert(yielded_workers() < total_workers(), "Consistency check");
   if (yielding_task()->status() == ABORTING) {
     // Do not yield; we need to abort as soon as possible
     // XXX NOTE: This can cause a performance pathology in the
@@ -227,7 +211,7 @@
     // us to return at each potential yield point.
     return;
   }
-  if (++_yielded_workers + finished_workers() == active_workers()) {
+  if (++_yielded_workers + finished_workers() == total_workers()) {
     yielding_task()->set_status(YIELDED);
     monitor()->notify_all();
   } else {
--- a/src/share/vm/utilities/yieldingWorkgroup.hpp	Wed Oct 06 14:18:32 2010 -0700
+++ b/src/share/vm/utilities/yieldingWorkgroup.hpp	Fri Oct 08 09:29:09 2010 -0700
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2005, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -54,6 +54,25 @@
   virtual void loop();
 };
 
+class FlexibleGangTask: public AbstractGangTask {
+  int _actual_size;                      // size of gang obtained
+protected:
+  int _requested_size;                   // size of gang requested
+public:
+ FlexibleGangTask(const char* name): AbstractGangTask(name),
+    _requested_size(0) {}
+
+  // The abstract work method.
+  // The argument tells you which member of the gang you are.
+  virtual void work(int i) = 0;
+
+  int requested_size() const { return _requested_size; }
+  int actual_size()    const { return _actual_size; }
+
+  void set_requested_size(int sz) { _requested_size = sz; }
+  void set_actual_size(int sz)    { _actual_size    = sz; }
+};
+
 // An abstract task to be worked on by a flexible work gang,
 // and where the workers will periodically yield, usually
 // in response to some condition that is signalled by means
@@ -70,19 +89,15 @@
 // maximum) in response to task requests at certain points.
 // The last part (the flexible part) has not yet been fully
 // fleshed out and is a work in progress.
-class YieldingFlexibleGangTask: public AbstractGangTask {
+class YieldingFlexibleGangTask: public FlexibleGangTask {
   Status _status;
   YieldingFlexibleWorkGang* _gang;
-  int _actual_size;                      // size of gang obtained
 
 protected:
-  int _requested_size;                   // size of gang requested
-
   // Constructor and desctructor: only construct subclasses.
-  YieldingFlexibleGangTask(const char* name): AbstractGangTask(name),
+  YieldingFlexibleGangTask(const char* name): FlexibleGangTask(name),
     _status(INACTIVE),
-    _gang(NULL),
-    _requested_size(0) { }
+    _gang(NULL) { }
 
   virtual ~YieldingFlexibleGangTask() { }
 
@@ -122,24 +137,18 @@
   virtual void abort();
 
   Status status()  const { return _status; }
+  bool yielding()  const { return _status == YIELDING; }
   bool yielded()   const { return _status == YIELDED; }
   bool completed() const { return _status == COMPLETED; }
   bool aborted()   const { return _status == ABORTED; }
   bool active()    const { return _status == ACTIVE; }
-
-  int requested_size() const { return _requested_size; }
-  int actual_size()    const { return _actual_size; }
-
-  void set_requested_size(int sz) { _requested_size = sz; }
-  void set_actual_size(int sz)    { _actual_size    = sz; }
 };
-
 // Class YieldingWorkGang: A subclass of WorkGang.
 // In particular, a YieldingWorkGang is made up of
 // YieldingGangWorkers, and provides infrastructure
 // supporting yielding to the "GangOverseer",
 // being the thread that orchestrates the WorkGang via run_task().
-class YieldingFlexibleWorkGang: public AbstractWorkGang {
+class YieldingFlexibleWorkGang: public FlexibleWorkGang {
   // Here's the public interface to this class.
 public:
   // Constructor and destructor.
@@ -151,6 +160,9 @@
            "Incorrect cast");
     return (YieldingFlexibleGangTask*)task();
   }
+  // Allocate a worker and return a pointer to it.
+  GangWorker* allocate_worker(int which);
+
   // Run a task; returns when the task is done, or the workers yield,
   // or the task is aborted, or the work gang is terminated via stop().
   // A task that has been yielded can be continued via this same interface
@@ -180,10 +192,6 @@
   void abort();
 
 private:
-  // The currently active workers in this gang.
-  // This is a number that is dynamically adjusted by
-  // the run_task() method at each subsequent invocation,
-  // using data in the YieldingFlexibleGangTask.
   int _active_workers;
   int _yielded_workers;
   void wait_for_gang();
@@ -194,6 +202,7 @@
     return _active_workers;
   }
 
+  // Accessors for fields
   int yielded_workers() const {
     return _yielded_workers;
   }