6786503: Overflow list performance can be improved

Avoid overflow list walk in CMS & ParNew when it is unnecessary. Fix a couple of correctness issues, including a C-heap leak, in ParNew at the intersection of promotion failure, work queue overflow and object array chunking. Add stress testing option and related assertion checking.

Reviewed-by: jmasa

hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp

@@ -8508,7 +8508,7 @@ bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
   size_t i = num;
   oop  cur = _overflow_list;
   const markOop proto = markOopDesc::prototype();
-  NOT_PRODUCT(size_t n = 0;)
+  NOT_PRODUCT(ssize_t n = 0;)
   for (oop next; i > 0 && cur != NULL; cur = next, i--) {
     next = oop(cur->mark());
     cur->set_mark(proto);   // until proven otherwise
@@ -8525,45 +8525,131 @@ bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) {
   return !stack->isEmpty();
 }
 
-// Multi-threaded; use CAS to break off a prefix
+#define BUSY  (oop(0x1aff1aff))
+// (MT-safe) Get a prefix of at most "num" from the list.
+// The overflow list is chained through the mark word of
+// each object in the list. We fetch the entire list,
+// break off a prefix of the right size and return the
+// remainder. If other threads try to take objects from
+// the overflow list at that time, they will wait for
+// some time to see if data becomes available. If (and
+// only if) another thread places one or more object(s)
+// on the global list before we have returned the suffix
+// to the global list, we will walk down our local list
+// to find its end and append the global list to
+// our suffix before returning it. This suffix walk can
+// prove to be expensive (quadratic in the amount of traffic)
+// when there are many objects in the overflow list and
+// there is much producer-consumer contention on the list.
+// *NOTE*: The overflow list manipulation code here and
+// in ParNewGeneration:: are very similar in shape,
+// except that in the ParNew case we use the old (from/eden)
+// copy of the object to thread the list via its klass word.
+// Because of the common code, if you make any changes in
+// the code below, please check the ParNew version to see if
+// 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) {
-  assert(work_q->size() == 0, "That's the current policy");
+  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) {
     return false;
   }
   // Grab the entire list; we'll put back a suffix
-  oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
+  oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
-  if (prefix == NULL) {  // someone grabbed it before we did ...
+  Thread* tid = Thread::current();
-    // ... we could spin for a short while, but for now we don't
+  size_t CMSOverflowSpinCount = (size_t)ParallelGCThreads;
-    return false;
+  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.
+  for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) {
+    os::sleep(tid, sleep_time_millis, false);
+    if (_overflow_list == NULL) {
+      // Nothing left to take
+      return false;
+    } else if (_overflow_list != BUSY) {
+      // Try and grab the prefix
+      prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
+    }
   }
+  // If the list was found to be empty, or we spun long
+  // enough, we give up and return empty-handed. If we leave
+  // the list in the BUSY state below, it must be the case that
+  // some other thread holds the overflow list and will set it
+  // to a non-BUSY state in the future.
+  if (prefix == NULL || prefix == BUSY) {
+     // Nothing to take or waited long enough
+     if (prefix == NULL) {
+       // Write back the NULL in case we overwrote it with BUSY above
+       // and it is still the same value.
+       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
+     }
+     return false;
+  }
+  assert(prefix != NULL && prefix != BUSY, "Error");
   size_t i = num;
   oop cur = prefix;
+  // Walk down the first "num" objects, unless we reach the end.
   for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--);
-  if (cur->mark() != NULL) {
+  if (cur->mark() == NULL) {
+    // We have "num" or fewer elements in the list, so there
+    // is nothing to return to the global list.
+    // Write back the NULL in lieu of the BUSY we wrote
+    // above, if it is still the same value.
+    if (_overflow_list == BUSY) {
+      (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
+    }
+  } else {
+    // Chop off the suffix and rerturn it to the global list.
+    assert(cur->mark() != BUSY, "Error");
     oop suffix_head = cur->mark(); // suffix will be put back on global list
     cur->set_mark(NULL);           // break off suffix
-    // Find tail of suffix so we can prepend suffix to global list
+    // It's possible that the list is still in the empty(busy) state
-    for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
+    // we left it in a short while ago; in that case we may be
-    oop suffix_tail = cur;
+    // able to place back the suffix without incurring the cost
-    assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
+    // of a walk down the list.
-           "Tautology");
     oop observed_overflow_list = _overflow_list;
-    do {
+    oop cur_overflow_list = observed_overflow_list;
-      cur = observed_overflow_list;
+    bool attached = false;
-      suffix_tail->set_mark(markOop(cur));
+    while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
       observed_overflow_list =
-        (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur);
+        (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
-    } while (cur != observed_overflow_list);
+      if (cur_overflow_list == observed_overflow_list) {
+        attached = true;
+        break;
+      } else cur_overflow_list = observed_overflow_list;
+    }
+    if (!attached) {
+      // Too bad, someone else sneaked in (at least) an element; we'll need
+      // to do a splice. Find tail of suffix so we can prepend suffix to global
+      // list.
+      for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark()));
+      oop suffix_tail = cur;
+      assert(suffix_tail != NULL && suffix_tail->mark() == NULL,
+             "Tautology");
+      observed_overflow_list = _overflow_list;
+      do {
+        cur_overflow_list = observed_overflow_list;
+        if (cur_overflow_list != BUSY) {
+          // Do the splice ...
+          suffix_tail->set_mark(markOop(cur_overflow_list));
+        } else { // cur_overflow_list == BUSY
+          suffix_tail->set_mark(NULL);
+        }
+        // ... and try to place spliced list back on overflow_list ...
+        observed_overflow_list =
+          (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list);
+      } while (cur_overflow_list != observed_overflow_list);
+      // ... until we have succeeded in doing so.
+    }
   }
 
   // Push the prefix elements on work_q
   assert(prefix != NULL, "control point invariant");
   const markOop proto = markOopDesc::prototype();
   oop next;
-  NOT_PRODUCT(size_t n = 0;)
+  NOT_PRODUCT(ssize_t n = 0;)
   for (cur = prefix; cur != NULL; cur = next) {
     next = oop(cur->mark());
     cur->set_mark(proto);   // until proven otherwise
@@ -8597,11 +8683,16 @@ void CMSCollector::par_push_on_overflow_list(oop p) {
   oop cur_overflow_list;
   do {
     cur_overflow_list = observed_overflow_list;
-    p->set_mark(markOop(cur_overflow_list));
+    if (cur_overflow_list != BUSY) {
+      p->set_mark(markOop(cur_overflow_list));
+    } else {
+      p->set_mark(NULL);
+    }
     observed_overflow_list =
       (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list);
   } while (cur_overflow_list != observed_overflow_list);
 }
+#undef BUSY
 
 // Single threaded
 // General Note on GrowableArray: pushes may silently fail
@@ -8610,7 +8701,7 @@ void CMSCollector::par_push_on_overflow_list(oop p) {
 // a lot of code in the JVM. The prudent thing for GrowableArray
 // to do (for now) is to exit with an error. However, that may
 // be too draconian in some cases because the caller may be
-// able to recover without much harm. For suych cases, we
+// able to recover without much harm. For such cases, we
 // should probably introduce a "soft_push" method which returns
 // an indication of success or failure with the assumption that
 // the caller may be able to recover from a failure; code in
@@ -8618,8 +8709,6 @@ void CMSCollector::par_push_on_overflow_list(oop p) {
 // failures where possible, thus, incrementally hardening the VM
 // in such low resource situations.
 void CMSCollector::preserve_mark_work(oop p, markOop m) {
-  int PreserveMarkStackSize = 128;
-
   if (_preserved_oop_stack == NULL) {
     assert(_preserved_mark_stack == NULL,
            "bijection with preserved_oop_stack");

hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp

@@ -595,7 +595,7 @@ class CMSCollector: public CHeapObj {
   size_t        _ser_kac_preclean_ovflw;
   size_t        _ser_kac_ovflw;
   size_t        _par_kac_ovflw;
-  NOT_PRODUCT(size_t _num_par_pushes;)
+  NOT_PRODUCT(ssize_t _num_par_pushes;)
 
   // ("Weak") Reference processing support
   ReferenceProcessor*            _ref_processor;

hotspot/src/share/vm/gc_implementation/includeDB_gc_parNew

@@ -77,6 +77,7 @@ parNewGeneration.cpp                    resourceArea.hpp
 parNewGeneration.cpp                    sharedHeap.hpp
 parNewGeneration.cpp                    space.hpp
 parNewGeneration.cpp                    spaceDecorator.hpp
+parNewGeneration.cpp                    thread.hpp
 parNewGeneration.cpp                    workgroup.hpp
 
 parNewGeneration.hpp                    defNewGeneration.hpp

hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp

@@ -404,6 +404,8 @@ void ParEvacuateFollowersClosure::do_void() {
     if (terminator()->offer_termination()) break;
     par_scan_state()->end_term_time();
   }
+  assert(par_gen()->_overflow_list == NULL && par_gen()->_num_par_pushes == 0,
+         "Broken overflow list?");
   // Finish the last termination pause.
   par_scan_state()->end_term_time();
 }
@@ -456,6 +458,8 @@ ParNewGeneration(ReservedSpace rs, size_t initial_byte_size, int level)
   _is_alive_closure(this),
   _plab_stats(YoungPLABSize, PLABWeight)
 {
+  NOT_PRODUCT(_overflow_counter = ParGCWorkQueueOverflowInterval;)
+  NOT_PRODUCT(_num_par_pushes = 0;)
   _task_queues = new ObjToScanQueueSet(ParallelGCThreads);
   guarantee(_task_queues != NULL, "task_queues allocation failure.");
 
@@ -993,12 +997,19 @@ oop ParNewGeneration::copy_to_survivor_space_avoiding_promotion_undo(
              "push forwarded object");
     }
     // Push it on one of the queues of to-be-scanned objects.
-    if (!par_scan_state->work_queue()->push(obj_to_push)) {
+    bool simulate_overflow = false;
+    NOT_PRODUCT(
+      if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
+        // simulate a stack overflow
+        simulate_overflow = true;
+      }
+    )
+    if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
       // Add stats for overflow pushes.
       if (Verbose && PrintGCDetails) {
         gclog_or_tty->print("queue overflow!\n");
       }
-      push_on_overflow_list(old);
+      push_on_overflow_list(old, par_scan_state);
       par_scan_state->note_overflow_push();
     }
     par_scan_state->note_push();
@@ -1110,9 +1121,16 @@ oop ParNewGeneration::copy_to_survivor_space_with_undo(
              "push forwarded object");
     }
     // Push it on one of the queues of to-be-scanned objects.
-    if (!par_scan_state->work_queue()->push(obj_to_push)) {
+    bool simulate_overflow = false;
+    NOT_PRODUCT(
+      if (ParGCWorkQueueOverflowALot && should_simulate_overflow()) {
+        // simulate a stack overflow
+        simulate_overflow = true;
+      }
+    )
+    if (simulate_overflow || !par_scan_state->work_queue()->push(obj_to_push)) {
       // Add stats for overflow pushes.
-      push_on_overflow_list(old);
+      push_on_overflow_list(old, par_scan_state);
       par_scan_state->note_overflow_push();
     }
     par_scan_state->note_push();
@@ -1135,89 +1153,190 @@ oop ParNewGeneration::copy_to_survivor_space_with_undo(
   return forward_ptr;
 }
 
-void ParNewGeneration::push_on_overflow_list(oop from_space_obj) {
+#ifndef PRODUCT
-  oop cur_overflow_list = _overflow_list;
+// It's OK to call this multi-threaded;  the worst thing
+// that can happen is that we'll get a bunch of closely
+// spaced simulated oveflows, but that's OK, in fact
+// probably good as it would exercise the overflow code
+// under contention.
+bool ParNewGeneration::should_simulate_overflow() {
+  if (_overflow_counter-- <= 0) { // just being defensive
+    _overflow_counter = ParGCWorkQueueOverflowInterval;
+    return true;
+  } else {
+    return false;
+  }
+}
+#endif
+
+#define BUSY (oop(0x1aff1aff))
+void ParNewGeneration::push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state) {
   // if the object has been forwarded to itself, then we cannot
   // use the klass pointer for the linked list.  Instead we have
   // to allocate an oopDesc in the C-Heap and use that for the linked list.
+  // XXX This is horribly inefficient when a promotion failure occurs
+  // and should be fixed. XXX FIX ME !!!
+#ifndef PRODUCT
+  Atomic::inc_ptr(&_num_par_pushes);
+  assert(_num_par_pushes > 0, "Tautology");
+#endif
   if (from_space_obj->forwardee() == from_space_obj) {
     oopDesc* listhead = NEW_C_HEAP_ARRAY(oopDesc, 1);
     listhead->forward_to(from_space_obj);
     from_space_obj = listhead;
   }
-  while (true) {
+  oop observed_overflow_list = _overflow_list;
-    from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
+  oop cur_overflow_list;
-    oop observed_overflow_list =
+  do {
-      (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
-    if (observed_overflow_list == cur_overflow_list) break;
-    // Otherwise...
     cur_overflow_list = observed_overflow_list;
-  }
+    if (cur_overflow_list != BUSY) {
+      from_space_obj->set_klass_to_list_ptr(cur_overflow_list);
+    } else {
+      from_space_obj->set_klass_to_list_ptr(NULL);
+    }
+    observed_overflow_list =
+      (oop)Atomic::cmpxchg_ptr(from_space_obj, &_overflow_list, cur_overflow_list);
+  } while (cur_overflow_list != observed_overflow_list);
 }
 
+// *NOTE*: The overflow list manipulation code here and
+// in CMSCollector:: are very similar in shape,
+// except that in the CMS case we thread the objects
+// directly into the list via their mark word, and do
+// not need to deal with special cases below related
+// to chunking of object arrays and promotion failure
+// handling.
+// CR 6797058 has been filed to attempt consolidation of
+// the common code.
+// Because of the common code, if you make any changes in
+// the code below, please check the CMS version to see if
+// similar changes might be needed.
+// See CMSCollector::par_take_from_overflow_list() for
+// more extensive documentation comments.
 bool
 ParNewGeneration::take_from_overflow_list(ParScanThreadState* par_scan_state) {
   ObjToScanQueue* work_q = par_scan_state->work_queue();
+  assert(work_q->size() == 0, "Should first empty local work queue");
   // How many to take?
-  int objsFromOverflow = MIN2(work_q->max_elems()/4,
+  size_t objsFromOverflow = MIN2((size_t)work_q->max_elems()/4,
-                              (juint)ParGCDesiredObjsFromOverflowList);
+                                 (size_t)ParGCDesiredObjsFromOverflowList);
 
   if (_overflow_list == NULL) return false;
 
   // Otherwise, there was something there; try claiming the list.
-  oop prefix = (oop)Atomic::xchg_ptr(NULL, &_overflow_list);
+  oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
-
-  if (prefix == NULL) {
-    return false;
-  }
   // Trim off a prefix of at most objsFromOverflow items
-  int i = 1;
+  Thread* tid = Thread::current();
+  size_t spin_count = (size_t)ParallelGCThreads;
+  size_t sleep_time_millis = MAX2((size_t)1, objsFromOverflow/100);
+  for (size_t spin = 0; prefix == BUSY && spin < spin_count; spin++) {
+    // someone grabbed it before we did ...
+    // ... we spin for a short while...
+    os::sleep(tid, sleep_time_millis, false);
+    if (_overflow_list == NULL) {
+      // nothing left to take
+      return false;
+    } else if (_overflow_list != BUSY) {
+     // try and grab the prefix
+     prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list);
+    }
+  }
+  if (prefix == NULL || prefix == BUSY) {
+     // Nothing to take or waited long enough
+     if (prefix == NULL) {
+       // Write back the NULL in case we overwrote it with BUSY above
+       // and it is still the same value.
+       (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
+     }
+     return false;
+  }
+  assert(prefix != NULL && prefix != BUSY, "Error");
+  size_t i = 1;
   oop cur = prefix;
   while (i < objsFromOverflow && cur->klass_or_null() != NULL) {
     i++; cur = oop(cur->klass());
   }
 
   // Reattach remaining (suffix) to overflow list
-  if (cur->klass_or_null() != NULL) {
+  if (cur->klass_or_null() == NULL) {
-    oop suffix = oop(cur->klass());
+    // Write back the NULL in lieu of the BUSY we wrote
-    cur->set_klass_to_list_ptr(NULL);
+    // above and it is still the same value.
-
+    if (_overflow_list == BUSY) {
-    // Find last item of suffix list
+      (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY);
-    oop last = suffix;
-    while (last->klass_or_null() != NULL) {
-      last = oop(last->klass());
     }
-    // Atomically prepend suffix to current overflow list
+  } else {
-    oop cur_overflow_list = _overflow_list;
+    assert(cur->klass_or_null() != BUSY, "Error");
-    while (true) {
+    oop suffix = oop(cur->klass());       // suffix will be put back on global list
-      last->set_klass_to_list_ptr(cur_overflow_list);
+    cur->set_klass_to_list_ptr(NULL);     // break off suffix
-      oop observed_overflow_list =
+    // It's possible that the list is still in the empty(busy) state
-        (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
+    // we left it in a short while ago; in that case we may be
-      if (observed_overflow_list == cur_overflow_list) break;
+    // able to place back the suffix.
-      // Otherwise...
+    oop observed_overflow_list = _overflow_list;
-      cur_overflow_list = observed_overflow_list;
+    oop cur_overflow_list = observed_overflow_list;
+    bool attached = false;
+    while (observed_overflow_list == BUSY || observed_overflow_list == NULL) {
+      observed_overflow_list =
+        (oop) Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
+      if (cur_overflow_list == observed_overflow_list) {
+        attached = true;
+        break;
+      } else cur_overflow_list = observed_overflow_list;
+    }
+    if (!attached) {
+      // Too bad, someone else got in in between; we'll need to do a splice.
+      // Find the last item of suffix list
+      oop last = suffix;
+      while (last->klass_or_null() != NULL) {
+        last = oop(last->klass());
+      }
+      // Atomically prepend suffix to current overflow list
+      observed_overflow_list = _overflow_list;
+      do {
+        cur_overflow_list = observed_overflow_list;
+        if (cur_overflow_list != BUSY) {
+          // Do the splice ...
+          last->set_klass_to_list_ptr(cur_overflow_list);
+        } else { // cur_overflow_list == BUSY
+          last->set_klass_to_list_ptr(NULL);
+        }
+        observed_overflow_list =
+          (oop)Atomic::cmpxchg_ptr(suffix, &_overflow_list, cur_overflow_list);
+      } while (cur_overflow_list != observed_overflow_list);
     }
   }
 
   // Push objects on prefix list onto this thread's work queue
-  assert(cur != NULL, "program logic");
+  assert(prefix != NULL && prefix != BUSY, "program logic");
   cur = prefix;
-  int n = 0;
+  ssize_t n = 0;
   while (cur != NULL) {
     oop obj_to_push = cur->forwardee();
     oop next        = oop(cur->klass_or_null());
     cur->set_klass(obj_to_push->klass());
-    if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
+    // This may be an array object that is self-forwarded. In that case, the list pointer
-      obj_to_push = cur;
+    // space, cur, is not in the Java heap, but rather in the C-heap and should be freed.
+    if (!is_in_reserved(cur)) {
+      // This can become a scaling bottleneck when there is work queue overflow coincident
+      // with promotion failure.
+      oopDesc* f = cur;
+      FREE_C_HEAP_ARRAY(oopDesc, f);
+    } else if (par_scan_state->should_be_partially_scanned(obj_to_push, cur)) {
       assert(arrayOop(cur)->length() == 0, "entire array remaining to be scanned");
+      obj_to_push = cur;
     }
-    work_q->push(obj_to_push);
+    bool ok = work_q->push(obj_to_push);
+    assert(ok, "Should have succeeded");
     cur = next;
     n++;
   }
   par_scan_state->note_overflow_refill(n);
+#ifndef PRODUCT
+  assert(_num_par_pushes >= n, "Too many pops?");
+  Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes);
+#endif
   return true;
 }
+#undef BUSY
 
 void ParNewGeneration::ref_processor_init()
 {

hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp

@@ -278,6 +278,7 @@ class ParNewGeneration: public DefNewGeneration {
   friend class ParNewRefProcTask;
   friend class ParNewRefProcTaskExecutor;
   friend class ParScanThreadStateSet;
+  friend class ParEvacuateFollowersClosure;
 
  private:
   // XXX use a global constant instead of 64!
@@ -296,6 +297,7 @@ class ParNewGeneration: public DefNewGeneration {
   // klass-pointers (klass information already copied to the forwarded
   // image.)  Manipulated with CAS.
   oop _overflow_list;
+  NOT_PRODUCT(ssize_t _num_par_pushes;)
 
   // If true, older generation does not support promotion undo, so avoid.
   static bool _avoid_promotion_undo;
@@ -372,8 +374,12 @@ class ParNewGeneration: public DefNewGeneration {
   oop copy_to_survivor_space_with_undo(ParScanThreadState* par_scan_state,
                              oop obj, size_t obj_sz, markOop m);
 
+  // in support of testing overflow code
+  NOT_PRODUCT(int _overflow_counter;)
+  NOT_PRODUCT(bool should_simulate_overflow();)
+
   // Push the given (from-space) object on the global overflow list.
-  void push_on_overflow_list(oop from_space_obj);
+  void push_on_overflow_list(oop from_space_obj, ParScanThreadState* par_scan_state);
 
   // If the global overflow list is non-empty, move some tasks from it
   // onto "work_q" (which must be empty).  No more than 1/4 of the

hotspot/src/share/vm/memory/referenceProcessor.cpp

@@ -721,12 +721,6 @@ ReferenceProcessor::process_phase3(DiscoveredList&    refs_list,
                              iter.obj(), iter.obj()->blueprint()->internal_name());
     }
     assert(iter.obj()->is_oop(UseConcMarkSweepGC), "Adding a bad reference");
-    // If discovery is concurrent, we may have objects with null referents,
-    // being those that were concurrently cleared after they were discovered
-    // (and not subsequently precleaned).
-    assert(   (discovery_is_atomic() && iter.referent()->is_oop())
-           || (!discovery_is_atomic() && iter.referent()->is_oop_or_null(UseConcMarkSweepGC)),
-           "Adding a bad referent");
     iter.next();
   }
   // Remember to keep sentinel pointer around

hotspot/src/share/vm/runtime/globals.hpp

@@ -1307,7 +1307,14 @@ class CommandLineFlags {
   product(intx, ParGCArrayScanChunk, 50,                                    \
           "Scan a subset and push remainder, if array is bigger than this") \
                                                                             \
-  product(intx, ParGCDesiredObjsFromOverflowList, 20,                       \
+  notproduct(bool, ParGCWorkQueueOverflowALot, false,                       \
+          "Whether we should simulate work queue overflow in ParNew")       \
+                                                                            \
+  notproduct(uintx, ParGCWorkQueueOverflowInterval, 1000,                   \
+          "An `interval' counter that determines how frequently"            \
+          " we simulate overflow; a smaller number increases frequency")    \
+                                                                            \
+  product(uintx, ParGCDesiredObjsFromOverflowList, 20,                      \
           "The desired number of objects to claim from the overflow list")  \
                                                                             \
   product(uintx, CMSParPromoteBlocksToClaim, 50,                            \
@@ -1429,8 +1436,8 @@ class CommandLineFlags {
           "Whether we should simulate frequent marking stack / work queue"  \
           " overflow")                                                      \
                                                                             \
-  notproduct(intx, CMSMarkStackOverflowInterval, 1000,                      \
+  notproduct(uintx, CMSMarkStackOverflowInterval, 1000,                     \
-          "A per-thread `interval' counter that determines how frequently"  \
+          "An `interval' counter that determines how frequently"            \
           " we simulate overflow; a smaller number increases frequency")    \
                                                                             \
   product(uintx, CMSMaxAbortablePrecleanLoops, 0,                           \
@@ -1648,7 +1655,7 @@ class CommandLineFlags {
   develop(uintx, WorkStealingYieldsBeforeSleep, 1000,                       \
           "Number of yields before a sleep is done during workstealing")    \
                                                                             \
-  product(uintx, PreserveMarkStackSize, 40,                                 \
+  product(uintx, PreserveMarkStackSize, 1024,                               \
            "Size for stack used in promotion failure handling")             \
                                                                             \
   product_pd(bool, UseTLAB, "Use thread-local object allocation")           \