8069367: Eagerly reclaimed humongous objects left on mark stack

Summary: Prevent eager reclaim of objects that might be on mark stack.
Reviewed-by: brutisso, tschatzl

src/share/vm/gc_implementation/g1/concurrentMark.cpp

@@ -3510,22 +3510,29 @@ void ConcurrentMark::print_finger() {
 }
 #endif
 
-void CMTask::scan_object(oop obj) {
+template<bool scan>
+inline void CMTask::process_grey_object(oop obj) {
+  assert(scan || obj->is_typeArray(), "Skipping scan of grey non-typeArray");
   assert(_nextMarkBitMap->isMarked((HeapWord*) obj), "invariant");
 
   if (_cm->verbose_high()) {
-    gclog_or_tty->print_cr("[%u] we're scanning object "PTR_FORMAT,
+    gclog_or_tty->print_cr("[%u] processing grey object " PTR_FORMAT,
                            _worker_id, p2i((void*) obj));
   }
 
   size_t obj_size = obj->size();
   _words_scanned += obj_size;
 
-  obj->oop_iterate(_cm_oop_closure);
+  if (scan) {
+    obj->oop_iterate(_cm_oop_closure);
+  }
   statsOnly( ++_objs_scanned );
   check_limits();
 }
 
+template void CMTask::process_grey_object<true>(oop);
+template void CMTask::process_grey_object<false>(oop);
+
 // Closure for iteration over bitmaps
 class CMBitMapClosure : public BitMapClosure {
 private:

src/share/vm/gc_implementation/g1/concurrentMark.hpp

@@ -1112,6 +1112,8 @@ private:
   // mark bitmap scan, and so needs to be pushed onto the mark stack.
   bool is_below_finger(HeapWord* objAddr, HeapWord* global_finger) const;
 
+  template<bool scan> void process_grey_object(oop obj);
+
 public:
   // It resets the task; it should be called right at the beginning of
   // a marking phase.
@@ -1164,7 +1166,7 @@ public:
   inline void deal_with_reference(oop obj);
 
   // It scans an object and visits its children.
-  void scan_object(oop obj);
+  void scan_object(oop obj) { process_grey_object<true>(obj); }
 
   // It pushes an object on the local queue.
   inline void push(oop obj);

src/share/vm/gc_implementation/g1/concurrentMark.inline.hpp

@@ -332,14 +332,28 @@ inline void CMTask::deal_with_reference(oop obj) {
           // be pushed on the stack. So, some duplicate work, but no
           // correctness problems.
           if (is_below_finger(objAddr, global_finger)) {
-            if (_cm->verbose_high()) {
-              gclog_or_tty->print_cr("[%u] below a finger (local: " PTR_FORMAT
-                                     ", global: " PTR_FORMAT ") pushing "
-                                     PTR_FORMAT " on mark stack",
-                                     _worker_id, p2i(_finger),
-                                     p2i(global_finger), p2i(objAddr));
+            if (obj->is_typeArray()) {
+              // Immediately process arrays of primitive types, rather
+              // than pushing on the mark stack.  This keeps us from
+              // adding humongous objects to the mark stack that might
+              // be reclaimed before the entry is processed - see
+              // selection of candidates for eager reclaim of humongous
+              // objects.  The cost of the additional type test is
+              // mitigated by avoiding a trip through the mark stack,
+              // by only doing a bookkeeping update and avoiding the
+              // actual scan of the object - a typeArray contains no
+              // references, and the metadata is built-in.
+              process_grey_object<false>(obj);
+            } else {
+              if (_cm->verbose_high()) {
+                gclog_or_tty->print_cr("[%u] below a finger (local: " PTR_FORMAT
+                                       ", global: " PTR_FORMAT ") pushing "
+                                       PTR_FORMAT " on mark stack",
+                                       _worker_id, p2i(_finger),
+                                       p2i(global_finger), p2i(objAddr));
+              }
+              push(obj);
             }
-            push(obj);
           }
         }
       }

src/share/vm/gc_implementation/g1/g1CollectedHeap.cpp

@@ -1853,7 +1853,7 @@ G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
   _secondary_free_list("Secondary Free List", new SecondaryFreeRegionListMtSafeChecker()),
   _old_set("Old Set", false /* humongous */, new OldRegionSetMtSafeChecker()),
   _humongous_set("Master Humongous Set", true /* humongous */, new HumongousRegionSetMtSafeChecker()),
-  _humongous_is_live(),
+  _humongous_reclaim_candidates(),
   _has_humongous_reclaim_candidates(false),
   _free_regions_coming(false),
   _young_list(new YoungList(this)),
@@ -2048,8 +2048,14 @@ jint G1CollectedHeap::initialize() {
 
   _g1h = this;
 
-  _in_cset_fast_test.initialize(_hrm.reserved().start(), _hrm.reserved().end(), HeapRegion::GrainBytes);
-  _humongous_is_live.initialize(_hrm.reserved().start(), _hrm.reserved().end(), HeapRegion::GrainBytes);
+  {
+    HeapWord* start = _hrm.reserved().start();
+    HeapWord* end = _hrm.reserved().end();
+    size_t granularity = HeapRegion::GrainBytes;
+
+    _in_cset_fast_test.initialize(start, end, granularity);
+    _humongous_reclaim_candidates.initialize(start, end, granularity);
+  }
 
   // Create the ConcurrentMark data structure and thread.
   // (Must do this late, so that "max_regions" is defined.)
@@ -2141,11 +2147,6 @@ void G1CollectedHeap::stop() {
   }
 }
 
-void G1CollectedHeap::clear_humongous_is_live_table() {
-  guarantee(G1EagerReclaimHumongousObjects, "Should only be called if true");
-  _humongous_is_live.clear();
-}
-
 size_t G1CollectedHeap::conservative_max_heap_alignment() {
   return HeapRegion::max_region_size();
 }
@@ -3666,12 +3667,6 @@ size_t G1CollectedHeap::cards_scanned() {
   return g1_rem_set()->cardsScanned();
 }
 
-bool G1CollectedHeap::humongous_region_is_always_live(uint index) {
-  HeapRegion* region = region_at(index);
-  assert(region->startsHumongous(), "Must start a humongous object");
-  return oop(region->bottom())->is_objArray() || !region->rem_set()->is_empty();
-}
-
 class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
  private:
   size_t _total_humongous;
@@ -3679,14 +3674,59 @@ class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
 
   DirtyCardQueue _dcq;
 
-  bool humongous_region_is_candidate(uint index) {
-    HeapRegion* region = G1CollectedHeap::heap()->region_at(index);
-    assert(region->startsHumongous(), "Must start a humongous object");
+  // We don't nominate objects with many remembered set entries, on
+  // the assumption that such objects are likely still live.
+  bool is_remset_small(HeapRegion* region) const {
     HeapRegionRemSet* const rset = region->rem_set();
-    bool const allow_stale_refs = G1EagerReclaimHumongousObjectsWithStaleRefs;
-    return !oop(region->bottom())->is_objArray() &&
-           ((allow_stale_refs && rset->occupancy_less_or_equal_than(G1RSetSparseRegionEntries)) ||
-            (!allow_stale_refs && rset->is_empty()));
+    return G1EagerReclaimHumongousObjectsWithStaleRefs
+      ? rset->occupancy_less_or_equal_than(G1RSetSparseRegionEntries)
+      : rset->is_empty();
+  }
+
+  bool is_typeArray_region(HeapRegion* region) const {
+    return oop(region->bottom())->is_typeArray();
+  }
+
+  bool humongous_region_is_candidate(G1CollectedHeap* heap, HeapRegion* region) const {
+    assert(region->startsHumongous(), "Must start a humongous object");
+
+    // Candidate selection must satisfy the following constraints
+    // while concurrent marking is in progress:
+    //
+    // * In order to maintain SATB invariants, an object must not be
+    // reclaimed if it was allocated before the start of marking and
+    // has not had its references scanned.  Such an object must have
+    // its references (including type metadata) scanned to ensure no
+    // live objects are missed by the marking process.  Objects
+    // allocated after the start of concurrent marking don't need to
+    // be scanned.
+    //
+    // * An object must not be reclaimed if it is on the concurrent
+    // mark stack.  Objects allocated after the start of concurrent
+    // marking are never pushed on the mark stack.
+    //
+    // Nominating only objects allocated after the start of concurrent
+    // marking is sufficient to meet both constraints.  This may miss
+    // some objects that satisfy the constraints, but the marking data
+    // structures don't support efficiently performing the needed
+    // additional tests or scrubbing of the mark stack.
+    //
+    // However, we presently only nominate is_typeArray() objects.
+    // A humongous object containing references induces remembered
+    // set entries on other regions.  In order to reclaim such an
+    // object, those remembered sets would need to be cleaned up.
+    //
+    // We also treat is_typeArray() objects specially, allowing them
+    // to be reclaimed even if allocated before the start of
+    // concurrent mark.  For this we rely on mark stack insertion to
+    // exclude is_typeArray() objects, preventing reclaiming an object
+    // that is in the mark stack.  We also rely on the metadata for
+    // such objects to be built-in and so ensured to be kept live.
+    // Frequent allocation and drop of large binary blobs is an
+    // important use case for eager reclaim, and this special handling
+    // may reduce needed headroom.
+
+    return is_typeArray_region(region) && is_remset_small(region);
   }
 
  public:
@@ -3702,14 +3742,17 @@ class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
     }
     G1CollectedHeap* g1h = G1CollectedHeap::heap();
 
-    uint region_idx = r->hrm_index();
-    bool is_candidate = humongous_region_is_candidate(region_idx);
-    // Is_candidate already filters out humongous object with large remembered sets.
-    // If we have a humongous object with a few remembered sets, we simply flush these
-    // remembered set entries into the DCQS. That will result in automatic
-    // re-evaluation of their remembered set entries during the following evacuation
-    // phase.
+    bool is_candidate = humongous_region_is_candidate(g1h, r);
+    uint rindex = r->hrm_index();
+    g1h->set_humongous_reclaim_candidate(rindex, is_candidate);
     if (is_candidate) {
+      _candidate_humongous++;
+      g1h->register_humongous_region_with_in_cset_fast_test(rindex);
+      // Is_candidate already filters out humongous object with large remembered sets.
+      // If we have a humongous object with a few remembered sets, we simply flush these
+      // remembered set entries into the DCQS. That will result in automatic
+      // re-evaluation of their remembered set entries during the following evacuation
+      // phase.
       if (!r->rem_set()->is_empty()) {
         guarantee(r->rem_set()->occupancy_less_or_equal_than(G1RSetSparseRegionEntries),
                   "Found a not-small remembered set here. This is inconsistent with previous assumptions.");
@@ -3726,8 +3769,6 @@ class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
         r->rem_set()->clear_locked();
       }
       assert(r->rem_set()->is_empty(), "At this point any humongous candidate remembered set must be empty.");
-      g1h->register_humongous_region_with_in_cset_fast_test(region_idx);
-      _candidate_humongous++;
     }
     _total_humongous++;
 
@@ -3747,6 +3788,7 @@ void G1CollectedHeap::register_humongous_regions_with_in_cset_fast_test() {
   }
   double time = os::elapsed_counter();
 
+  // Collect reclaim candidate information and register candidates with cset.
   RegisterHumongousWithInCSetFastTestClosure cl;
   heap_region_iterate(&cl);
 
@@ -3756,10 +3798,6 @@ void G1CollectedHeap::register_humongous_regions_with_in_cset_fast_test() {
                                                                   cl.candidate_humongous());
   _has_humongous_reclaim_candidates = cl.candidate_humongous() > 0;
 
-  if (_has_humongous_reclaim_candidates || G1TraceEagerReclaimHumongousObjects) {
-    clear_humongous_is_live_table();
-  }
-
   // Finally flush all remembered set entries to re-check into the global DCQS.
   cl.flush_rem_set_entries();
 }
@@ -6321,11 +6359,11 @@ class G1FreeHumongousRegionClosure : public HeapRegionClosure {
     // required because stale remembered sets might reference locations that
     // are currently allocated into.
     uint region_idx = r->hrm_index();
-    if (g1h->humongous_is_live(region_idx) ||
-        g1h->humongous_region_is_always_live(region_idx)) {
+    if (!g1h->is_humongous_reclaim_candidate(region_idx) ||
+        !r->rem_set()->is_empty()) {
 
       if (G1TraceEagerReclaimHumongousObjects) {
-        gclog_or_tty->print_cr("Live humongous region %u size "SIZE_FORMAT" start "PTR_FORMAT" length "UINT32_FORMAT" with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d live-other %d obj array %d",
+        gclog_or_tty->print_cr("Live humongous region %u size "SIZE_FORMAT" start "PTR_FORMAT" length "UINT32_FORMAT" with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d reclaim candidate %d type array %d",
                                region_idx,
                                obj->size()*HeapWordSize,
                                r->bottom(),
@@ -6333,20 +6371,21 @@ class G1FreeHumongousRegionClosure : public HeapRegionClosure {
                                r->rem_set()->occupied(),
                                r->rem_set()->strong_code_roots_list_length(),
                                next_bitmap->isMarked(r->bottom()),
-                               g1h->humongous_is_live(region_idx),
-                               obj->is_objArray()
+                               g1h->is_humongous_reclaim_candidate(region_idx),
+                               obj->is_typeArray()
                               );
       }
 
       return false;
     }
 
-    guarantee(!obj->is_objArray(),
-              err_msg("Eagerly reclaiming object arrays is not supported, but the object "PTR_FORMAT" is.",
+    guarantee(obj->is_typeArray(),
+              err_msg("Only eagerly reclaiming type arrays is supported, but the object "
+                      PTR_FORMAT " is not.",
                       r->bottom()));
 
     if (G1TraceEagerReclaimHumongousObjects) {
-      gclog_or_tty->print_cr("Dead humongous region %u size "SIZE_FORMAT" start "PTR_FORMAT" length "UINT32_FORMAT" with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d live-other %d obj array %d",
+      gclog_or_tty->print_cr("Dead humongous region %u size "SIZE_FORMAT" start "PTR_FORMAT" length "UINT32_FORMAT" with remset "SIZE_FORMAT" code roots "SIZE_FORMAT" is marked %d reclaim candidate %d type array %d",
                              region_idx,
                              obj->size()*HeapWordSize,
                              r->bottom(),
@@ -6354,8 +6393,8 @@ class G1FreeHumongousRegionClosure : public HeapRegionClosure {
                              r->rem_set()->occupied(),
                              r->rem_set()->strong_code_roots_list_length(),
                              next_bitmap->isMarked(r->bottom()),
-                             g1h->humongous_is_live(region_idx),
-                             obj->is_objArray()
+                             g1h->is_humongous_reclaim_candidate(region_idx),
+                             obj->is_typeArray()
                             );
     }
     // Need to clear mark bit of the humongous object if already set.

src/share/vm/gc_implementation/g1/g1CollectedHeap.hpp

@@ -233,7 +233,6 @@ private:
   // It keeps track of the humongous regions.
   HeapRegionSet _humongous_set;
 
-  void clear_humongous_is_live_table();
   void eagerly_reclaim_humongous_regions();
 
   // The number of regions we could create by expansion.
@@ -303,22 +302,26 @@ private:
   // Helper for monitoring and management support.
   G1MonitoringSupport* _g1mm;
 
-  // Records whether the region at the given index is kept live by roots or
-  // references from the young generation.
-  class HumongousIsLiveBiasedMappedArray : public G1BiasedMappedArray<bool> {
+  // Records whether the region at the given index is (still) a
+  // candidate for eager reclaim.  Only valid for humongous start
+  // regions; other regions have unspecified values.  Humongous start
+  // regions are initialized at start of collection pause, with
+  // candidates removed from the set as they are found reachable from
+  // roots or the young generation.
+  class HumongousReclaimCandidates : public G1BiasedMappedArray<bool> {
    protected:
     bool default_value() const { return false; }
    public:
     void clear() { G1BiasedMappedArray<bool>::clear(); }
-    void set_live(uint region) {
-      set_by_index(region, true);
+    void set_candidate(uint region, bool value) {
+      set_by_index(region, value);
     }
-    bool is_live(uint region) {
+    bool is_candidate(uint region) {
       return get_by_index(region);
     }
   };
 
-  HumongousIsLiveBiasedMappedArray _humongous_is_live;
+  HumongousReclaimCandidates _humongous_reclaim_candidates;
   // Stores whether during humongous object registration we found candidate regions.
   // If not, we can skip a few steps.
   bool _has_humongous_reclaim_candidates;
@@ -655,18 +658,15 @@ public:
   virtual void gc_prologue(bool full);
   virtual void gc_epilogue(bool full);
 
-  inline void set_humongous_is_live(oop obj);
+  // Modify the reclaim candidate set and test for presence.
+  // These are only valid for starts_humongous regions.
+  inline void set_humongous_reclaim_candidate(uint region, bool value);
+  inline bool is_humongous_reclaim_candidate(uint region);
 
-  bool humongous_is_live(uint region) {
-    return _humongous_is_live.is_live(region);
-  }
+  // Remove from the reclaim candidate set.  Also remove from the
+  // collection set so that later encounters avoid the slow path.
+  inline void set_humongous_is_live(oop obj);
 
-  // Returns whether the given region (which must be a humongous (start) region)
-  // is to be considered conservatively live regardless of any other conditions.
-  bool humongous_region_is_always_live(uint index);
-  // Returns whether the given region (which must be a humongous (start) region)
-  // is considered a candidate for eager reclamation.
-  bool humongous_region_is_candidate(uint index);
   // Register the given region to be part of the collection set.
   inline void register_humongous_region_with_in_cset_fast_test(uint index);
   // Register regions with humongous objects (actually on the start region) in

src/share/vm/gc_implementation/g1/g1CollectedHeap.inline.hpp

@@ -348,20 +348,30 @@ inline bool G1CollectedHeap::is_obj_ill(const oop obj) const {
   return is_obj_ill(obj, heap_region_containing(obj));
 }
 
+inline void G1CollectedHeap::set_humongous_reclaim_candidate(uint region, bool value) {
+  assert(_hrm.at(region)->startsHumongous(), "Must start a humongous object");
+  _humongous_reclaim_candidates.set_candidate(region, value);
+}
+
+inline bool G1CollectedHeap::is_humongous_reclaim_candidate(uint region) {
+  assert(_hrm.at(region)->startsHumongous(), "Must start a humongous object");
+  return _humongous_reclaim_candidates.is_candidate(region);
+}
+
 inline void G1CollectedHeap::set_humongous_is_live(oop obj) {
   uint region = addr_to_region((HeapWord*)obj);
-  // We not only set the "live" flag in the humongous_is_live table, but also
+  // Clear the flag in the humongous_reclaim_candidates table.  Also
   // reset the entry in the _in_cset_fast_test table so that subsequent references
   // to the same humongous object do not go into the slow path again.
   // This is racy, as multiple threads may at the same time enter here, but this
   // is benign.
-  // During collection we only ever set the "live" flag, and only ever clear the
+  // During collection we only ever clear the "candidate" flag, and only ever clear the
   // entry in the in_cset_fast_table.
   // We only ever evaluate the contents of these tables (in the VM thread) after
   // having synchronized the worker threads with the VM thread, or in the same
   // thread (i.e. within the VM thread).
-  if (!_humongous_is_live.is_live(region)) {
-    _humongous_is_live.set_live(region);
+  if (is_humongous_reclaim_candidate(region)) {
+    set_humongous_reclaim_candidate(region, false);
     _in_cset_fast_test.clear_humongous(region);
   }
 }

test/TEST.groups

@@ -132,7 +132,8 @@ hotspot_compiler = \
   sanity/ExecuteInternalVMTests.java
 
 hotspot_gc = \
-  sanity/ExecuteInternalVMTests.java
+  sanity/ExecuteInternalVMTests.java \
+  -gc/g1/TestGreyReclaimedHumongousObjects.java
 
 hotspot_runtime = \
   sanity/ExecuteInternalVMTests.java

test/gc/g1/TestGreyReclaimedHumongousObjects.java

+/*
+ * Copyright (c) 2015, 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
+ * 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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
+ * or visit www.oracle.com if you need additional information or have any
+ * questions.
+ */
+
+/*
+ * @test TestGreyReclaimedHumongousObjects.java
+ * @bug 8069367
+ * @requires vm.gc == "G1" | vm.gc == "null"
+ * @summary Test handling of marked but unscanned reclaimed humongous objects.
+ * @key gc
+ * @run main/othervm -XX:+UseG1GC -Xss32m -Xmx128m -XX:G1HeapRegionSize=1m
+ *      -XX:+UnlockExperimentalVMOptions
+ *          -XX:+G1EagerReclaimHumongousObjects
+ *          -XX:+G1EagerReclaimHumongousObjectsWithStaleRefs
+ *      TestGreyReclaimedHumongousObjects 1048576 90
+ */
+
+// This test spawns a bunch of threads, each of them rapidly
+// allocating large objects and storing them into a circular buffer
+// associated with the thread.  The circular buffer results in these
+// objects becoming dead in fairly short order.
+//
+// The situation we're trying to provoke is
+//
+// (1) A humongous object H is marked and added to the mark stack.
+//
+// (2) An evacuation pause determines H is no longer live, and
+// reclaims it.  This occurs before concurrent marking has gotten
+// around to processing the mark stack entry for H.
+//
+// (3) Concurrent marking processes the mark stack entry for H.  The
+// bug is that it would attempt to scan the now dead object.
+//
+// Unfortunately, this test is *very* sensitive to configuration.
+// Among the parameters that affect whether / how often we'll get into
+// the desired situation within a reasonable amount of time are:
+//
+// - THREAD_COUNT: The number of allocating threads.
+//
+// - OLD_COUNT: The number of objects each thread keeps.
+//
+// - MAX_MEMORY: The maximum heap size.
+//
+// - G1HeapRegionSize
+//
+// - The size of the objects being allocated.
+//
+// The parameter values specified here:
+//
+// - THREAD_COUNT = 12
+// - OLD_COUNT == 4
+// - MAX_MEMORY == 128m
+// - G1HeapRegionSize = 1m
+// - Object size = 1048576 (2 regions after header overhead and roundup)
+//
+// seems to work well at provoking the desired state fairly quickly.
+// Even relatively small perturbations may change that.  The key
+// factors seem to be keeping the heap mostly full of live objects but
+// having them become dead fairly quickly.
+
+import java.util.Date;
+import java.util.concurrent.ExecutorService;
+import java.util.concurrent.Executors;
+import java.util.concurrent.ThreadFactory;
+import java.util.concurrent.TimeUnit;
+import sun.management.ManagementFactoryHelper;
+import com.sun.management.HotSpotDiagnosticMXBean;
+import com.sun.management.VMOption;
+
+public class TestGreyReclaimedHumongousObjects {
+
+    static class NamedThreadFactory implements ThreadFactory {
+       private int threadNum = 0;
+
+       @Override
+       public Thread newThread(Runnable r) {
+         return new Thread(r, THREAD_NAME + (threadNum++));
+       }
+    }
+
+    static class Runner extends Thread {
+        private final Date startDate = new Date();
+        private final int obj_size;
+        private final Object[] old_garbage;
+        private int old_index = 0;
+
+        public Runner(int obj_size) {
+            this.obj_size = obj_size;
+            old_garbage = new Object[OLD_COUNT];
+        }
+
+        private void allocate_garbage() {
+            byte[] garbage = new byte[obj_size];
+            old_garbage[Math.abs(++old_index % OLD_COUNT)] = garbage;
+        }
+
+        @Override
+        public void run() {
+            try {
+                while (!isInterrupted()) {
+                    allocate_garbage();
+                    Thread.sleep(0); // Yield, to ensure interruptable.
+                }
+            } catch (InterruptedException e) {
+                System.out.println("Aborted after "
+                                   + (new Date().getTime() - startDate.getTime())
+                                   + " ms");
+                interrupt();
+            }
+        }
+    }
+
+    public static void main(String[] args) throws Exception {
+        HotSpotDiagnosticMXBean diagnostic = ManagementFactoryHelper.getDiagnosticMXBean();
+
+        System.out.println("Max memory= " + MAX_MEMORY + " bytes");
+
+        int obj_size = 0;
+        long seconds_to_run = 0;
+        if (args.length != 2) {
+            throw new RuntimeException("Object size argument must be supplied");
+        } else {
+            obj_size = Integer.parseInt(args[0]);
+            seconds_to_run = Integer.parseInt(args[1]);
+        }
+        System.out.println("Objects size= " + obj_size + " bytes");
+        System.out.println("Seconds to run=" + seconds_to_run);
+
+        int region_size =
+            Integer.parseInt(diagnostic.getVMOption("G1HeapRegionSize").getValue());
+        if (obj_size < (region_size / 2)) {
+            throw new RuntimeException("Object size " + obj_size +
+                                       " is not humongous with region size " + region_size);
+        }
+
+        ExecutorService executor =
+            Executors.newFixedThreadPool(THREAD_COUNT, new NamedThreadFactory());
+        System.out.println("Starting " + THREAD_COUNT + " threads");
+
+        for (int i = 0; i < THREAD_COUNT; i++) {
+            executor.execute(new Runner(obj_size));
+        }
+
+        Thread.sleep(seconds_to_run * 1000);
+        executor.shutdownNow();
+
+        if (!executor.awaitTermination(10, TimeUnit.SECONDS)) {
+            System.err.println("Thread pool did not terminate after 10 seconds after shutdown");
+        }
+    }
+
+    private static final long MAX_MEMORY = Runtime.getRuntime().maxMemory();
+    private static final int OLD_COUNT = 4;
+    private static final int THREAD_COUNT = 12;
+    private static final String THREAD_NAME = "TestGreyRH-";
+}
+