6720309: G1: don't synchronously update RSet during evacuation pauses

6720334: G1: don't update RSets of collection set regions during an evacuation pause
Summary: Introduced a deferred update mechanism for delaying the rset updates during the collection pause
Reviewed-by: apetrusenko, tonyp

src/share/vm/gc_implementation/g1/concurrentG1RefineThread.cpp

@@ -133,14 +133,12 @@ void ConcurrentG1RefineThread::queueBasedRefinement() {
       _co_tracker.update(false);
 
       if (G1SmoothConcRefine) {
-        start_vtime_sec = os::elapsedVTime();
         prev_buffer_num = curr_buffer_num;
-
         _sts.leave();
         os::sleep(Thread::current(), (jlong) _interval_ms, false);
         _sts.join();
+        start_vtime_sec = os::elapsedVTime();
       }
-
       n_logs++;
     }
     // Make sure we harvest the PYA, if any.

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

@@ -78,8 +78,8 @@ size_t DirtyCardQueueSet::num_par_ids() {
 
 void DirtyCardQueueSet::initialize(Monitor* cbl_mon, Mutex* fl_lock,
                                    int max_completed_queue,
-                                   Mutex* lock) {
-  PtrQueueSet::initialize(cbl_mon, fl_lock, max_completed_queue);
+                                   Mutex* lock, PtrQueueSet* fl_owner) {
+  PtrQueueSet::initialize(cbl_mon, fl_lock, max_completed_queue, fl_owner);
   set_buffer_size(DCQBarrierQueueBufferSize);
   set_process_completed_threshold(DCQBarrierProcessCompletedThreshold);
 

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

@@ -88,7 +88,7 @@ public:
 
   void initialize(Monitor* cbl_mon, Mutex* fl_lock,
                   int max_completed_queue = 0,
-                  Mutex* lock = NULL);
+                  Mutex* lock = NULL, PtrQueueSet* fl_owner = NULL);
 
   // The number of parallel ids that can be claimed to allow collector or
   // mutator threads to do card-processing work.

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

@@ -136,6 +136,14 @@ public:
   int calls() { return _calls; }
 };
 
+class RedirtyLoggedCardTableEntryFastClosure : public CardTableEntryClosure {
+public:
+  bool do_card_ptr(jbyte* card_ptr, int worker_i) {
+    *card_ptr = CardTableModRefBS::dirty_card_val();
+    return true;
+  }
+};
+
 YoungList::YoungList(G1CollectedHeap* g1h)
   : _g1h(g1h), _head(NULL),
     _scan_only_head(NULL), _scan_only_tail(NULL), _curr_scan_only(NULL),
@@ -961,7 +969,8 @@ void G1CollectedHeap::do_collection(bool full, bool clear_all_soft_refs,
     // dirtied, so this should abandon those logs, and set "do_traversal"
     // to true.
     concurrent_g1_refine()->set_pya_restart();
-
+    assert(!G1DeferredRSUpdate
+           || (G1DeferredRSUpdate && (dirty_card_queue_set().completed_buffers_num() == 0)), "Should not be any");
     assert(regions_accounted_for(), "Region leakage!");
   }
 
@@ -1466,6 +1475,13 @@ jint G1CollectedHeap::initialize() {
                                                   G1DirtyCardQueueMax,
                                                   Shared_DirtyCardQ_lock);
   }
+  if (G1DeferredRSUpdate) {
+    dirty_card_queue_set().initialize(DirtyCardQ_CBL_mon,
+                                      DirtyCardQ_FL_lock,
+                                      0,
+                                      Shared_DirtyCardQ_lock,
+                                      &JavaThread::dirty_card_queue_set());
+  }
   // In case we're keeping closure specialization stats, initialize those
   // counts and that mechanism.
   SpecializationStats::clear();
@@ -2918,27 +2934,51 @@ public:
   }
 };
 
-class RecreateRSetEntriesClosure: public OopClosure {
+class UpdateRSetImmediate : public OopsInHeapRegionClosure {
 private:
   G1CollectedHeap* _g1;
   G1RemSet* _g1_rem_set;
-  HeapRegion* _from;
 public:
-  RecreateRSetEntriesClosure(G1CollectedHeap* g1, HeapRegion* from) :
-    _g1(g1), _g1_rem_set(g1->g1_rem_set()), _from(from)
-  {}
+  UpdateRSetImmediate(G1CollectedHeap* g1) :
+    _g1(g1), _g1_rem_set(g1->g1_rem_set()) {}
+
+  void do_oop(narrowOop* p) {
+    guarantee(false, "NYI");
+  }
+  void do_oop(oop* p) {
+    assert(_from->is_in_reserved(p), "paranoia");
+    if (*p != NULL && !_from->is_survivor()) {
+      _g1_rem_set->par_write_ref(_from, p, 0);
+    }
+  }
+};
+
+class UpdateRSetDeferred : public OopsInHeapRegionClosure {
+private:
+  G1CollectedHeap* _g1;
+  DirtyCardQueue *_dcq;
+  CardTableModRefBS* _ct_bs;
+
+public:
+  UpdateRSetDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
+    _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) {}
 
   void do_oop(narrowOop* p) {
     guarantee(false, "NYI");
   }
   void do_oop(oop* p) {
     assert(_from->is_in_reserved(p), "paranoia");
-    if (*p != NULL) {
-      _g1_rem_set->write_ref(_from, p);
+    if (!_from->is_in_reserved(*p) && !_from->is_survivor()) {
+      size_t card_index = _ct_bs->index_for(p);
+      if (_ct_bs->mark_card_deferred(card_index)) {
+        _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
+      }
     }
   }
 };
 
+
+
 class RemoveSelfPointerClosure: public ObjectClosure {
 private:
   G1CollectedHeap* _g1;
@@ -2946,11 +2986,11 @@ private:
   HeapRegion* _hr;
   size_t _prev_marked_bytes;
   size_t _next_marked_bytes;
+  OopsInHeapRegionClosure *_cl;
 public:
-  RemoveSelfPointerClosure(G1CollectedHeap* g1, HeapRegion* hr) :
-    _g1(g1), _cm(_g1->concurrent_mark()), _hr(hr),
-    _prev_marked_bytes(0), _next_marked_bytes(0)
-  {}
+  RemoveSelfPointerClosure(G1CollectedHeap* g1, OopsInHeapRegionClosure* cl) :
+    _g1(g1), _cm(_g1->concurrent_mark()),  _prev_marked_bytes(0),
+    _next_marked_bytes(0), _cl(cl) {}
 
   size_t prev_marked_bytes() { return _prev_marked_bytes; }
   size_t next_marked_bytes() { return _next_marked_bytes; }
@@ -2988,8 +3028,7 @@ public:
       // that, if evacuation fails, we might have remembered set
       // entries missing given that we skipped cards on the
       // collection set. So, we'll recreate such entries now.
-      RecreateRSetEntriesClosure cl(_g1, _hr);
-      obj->oop_iterate(&cl);
+      obj->oop_iterate(_cl);
       assert(_cm->isPrevMarked(obj), "Should be marked!");
     } else {
       // The object has been either evacuated or is dead. Fill it with a
@@ -3002,14 +3041,23 @@ public:
 };
 
 void G1CollectedHeap::remove_self_forwarding_pointers() {
+  UpdateRSetImmediate immediate_update(_g1h);
+  DirtyCardQueue dcq(&_g1h->dirty_card_queue_set());
+  UpdateRSetDeferred deferred_update(_g1h, &dcq);
+  OopsInHeapRegionClosure *cl;
+  if (G1DeferredRSUpdate) {
+    cl = &deferred_update;
+  } else {
+    cl = &immediate_update;
+  }
   HeapRegion* cur = g1_policy()->collection_set();
-
   while (cur != NULL) {
     assert(g1_policy()->assertMarkedBytesDataOK(), "Should be!");
 
+    RemoveSelfPointerClosure rspc(_g1h, cl);
     if (cur->evacuation_failed()) {
-      RemoveSelfPointerClosure rspc(_g1h, cur);
       assert(cur->in_collection_set(), "bad CS");
+      cl->set_region(cur);
       cur->object_iterate(&rspc);
 
       // A number of manipulations to make the TAMS be the current top,
@@ -3518,6 +3566,9 @@ class G1ParScanThreadState : public StackObj {
 protected:
   G1CollectedHeap* _g1h;
   RefToScanQueue*  _refs;
+  DirtyCardQueue   _dcq;
+  CardTableModRefBS* _ct_bs;
+  G1RemSet* _g1_rem;
 
   typedef GrowableArray<oop*> OverflowQueue;
   OverflowQueue* _overflowed_refs;
@@ -3559,10 +3610,32 @@ protected:
 
   void   add_to_undo_waste(size_t waste)         { _undo_waste += waste; }
 
+  DirtyCardQueue& dirty_card_queue()             { return _dcq;  }
+  CardTableModRefBS* ctbs()                      { return _ct_bs; }
+
+  void immediate_rs_update(HeapRegion* from, oop* p, int tid) {
+    _g1_rem->par_write_ref(from, p, tid);
+  }
+
+  void deferred_rs_update(HeapRegion* from, oop* p, int tid) {
+    // If the new value of the field points to the same region or
+    // is the to-space, we don't need to include it in the Rset updates.
+    if (!from->is_in_reserved(*p) && !from->is_survivor()) {
+      size_t card_index = ctbs()->index_for(p);
+      // If the card hasn't been added to the buffer, do it.
+      if (ctbs()->mark_card_deferred(card_index)) {
+        dirty_card_queue().enqueue((jbyte*)ctbs()->byte_for_index(card_index));
+      }
+    }
+  }
+
 public:
   G1ParScanThreadState(G1CollectedHeap* g1h, int queue_num)
     : _g1h(g1h),
       _refs(g1h->task_queue(queue_num)),
+      _dcq(&g1h->dirty_card_queue_set()),
+      _ct_bs((CardTableModRefBS*)_g1h->barrier_set()),
+      _g1_rem(g1h->g1_rem_set()),
       _hash_seed(17), _queue_num(queue_num),
       _term_attempts(0),
       _age_table(false),
@@ -3640,6 +3713,14 @@ public:
   int refs_to_scan()                             { return refs()->size();                 }
   int overflowed_refs_to_scan()                  { return overflowed_refs()->length();    }
 
+  void update_rs(HeapRegion* from, oop* p, int tid) {
+    if (G1DeferredRSUpdate) {
+      deferred_rs_update(from, p, tid);
+    } else {
+      immediate_rs_update(from, p, tid);
+    }
+  }
+
   HeapWord* allocate_slow(GCAllocPurpose purpose, size_t word_sz) {
 
     HeapWord* obj = NULL;
@@ -3808,7 +3889,6 @@ public:
   }
 };
 
-
 G1ParClosureSuper::G1ParClosureSuper(G1CollectedHeap* g1, G1ParScanThreadState* par_scan_state) :
   _g1(g1), _g1_rem(_g1->g1_rem_set()), _cm(_g1->concurrent_mark()),
   _par_scan_state(par_scan_state) { }
@@ -3834,7 +3914,7 @@ void G1ParScanClosure::do_oop_nv(oop* p) {
       assert(obj == *p, "the value of *p should not have changed");
       _par_scan_state->push_on_queue(p);
     } else {
-      _g1_rem->par_write_ref(_from, p, _par_scan_state->queue_num());
+      _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
     }
   }
 }
@@ -3972,13 +4052,13 @@ void G1ParCopyClosure<do_gen_barrier, barrier,
     }
     // When scanning the RS, we only care about objs in CS.
     if (barrier == G1BarrierRS) {
-      _g1_rem->par_write_ref(_from, p, _par_scan_state->queue_num());
+      _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
     }
   }
 
   // When scanning moved objs, must look at all oops.
   if (barrier == G1BarrierEvac && obj != NULL) {
-    _g1_rem->par_write_ref(_from, p, _par_scan_state->queue_num());
+    _par_scan_state->update_rs(_from, p, _par_scan_state->queue_num());
   }
 
   if (do_gen_barrier && obj != NULL) {
@@ -4127,6 +4207,7 @@ public:
     G1ParScanExtRootClosure         only_scan_root_cl(_g1h, &pss);
     G1ParScanPermClosure            only_scan_perm_cl(_g1h, &pss);
     G1ParScanHeapRSClosure          only_scan_heap_rs_cl(_g1h, &pss);
+
     G1ParScanAndMarkExtRootClosure  scan_mark_root_cl(_g1h, &pss);
     G1ParScanAndMarkPermClosure     scan_mark_perm_cl(_g1h, &pss);
     G1ParScanAndMarkHeapRSClosure   scan_mark_heap_rs_cl(_g1h, &pss);
@@ -4382,7 +4463,6 @@ void G1CollectedHeap::evacuate_collection_set() {
   g1_rem_set()->prepare_for_oops_into_collection_set_do();
   concurrent_g1_refine()->set_use_cache(false);
   int n_workers = (ParallelGCThreads > 0 ? workers()->total_workers() : 1);
-
   set_par_threads(n_workers);
   G1ParTask g1_par_task(this, n_workers, _task_queues);
 
@@ -4390,8 +4470,9 @@ void G1CollectedHeap::evacuate_collection_set() {
 
   change_strong_roots_parity();  // In preparation for parallel strong roots.
   rem_set()->prepare_for_younger_refs_iterate(true);
-  double start_par = os::elapsedTime();
 
+  assert(dirty_card_queue_set().completed_buffers_num() == 0, "Should be empty");
+  double start_par = os::elapsedTime();
   if (ParallelGCThreads > 0) {
     // The individual threads will set their evac-failure closures.
     workers()->run_task(&g1_par_task);
@@ -4411,8 +4492,8 @@ void G1CollectedHeap::evacuate_collection_set() {
     G1KeepAliveClosure keep_alive(this);
     JNIHandles::weak_oops_do(&is_alive, &keep_alive);
   }
-
   g1_rem_set()->cleanup_after_oops_into_collection_set_do();
+
   concurrent_g1_refine()->set_use_cache(true);
 
   finalize_for_evac_failure();
@@ -4423,7 +4504,6 @@ void G1CollectedHeap::evacuate_collection_set() {
 
   if (evacuation_failed()) {
     remove_self_forwarding_pointers();
-
     if (PrintGCDetails) {
       gclog_or_tty->print(" (evacuation failed)");
     } else if (PrintGC) {
@@ -4431,6 +4511,14 @@ void G1CollectedHeap::evacuate_collection_set() {
     }
   }
 
+  if (G1DeferredRSUpdate) {
+    RedirtyLoggedCardTableEntryFastClosure redirty;
+    dirty_card_queue_set().set_closure(&redirty);
+    dirty_card_queue_set().apply_closure_to_all_completed_buffers();
+    JavaThread::dirty_card_queue_set().merge_bufferlists(&dirty_card_queue_set());
+    assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
+  }
+
   COMPILER2_PRESENT(DerivedPointerTable::update_pointers());
 }
 

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

@@ -457,6 +457,10 @@ protected:
   // And it's mod ref barrier set, used to track updates for the above.
   ModRefBarrierSet* _mr_bs;
 
+  // A set of cards that cover the objects for which the Rsets should be updated
+  // concurrently after the collection.
+  DirtyCardQueueSet _dirty_card_queue_set;
+
   // The Heap Region Rem Set Iterator.
   HeapRegionRemSetIterator** _rem_set_iterator;
 
@@ -666,6 +670,9 @@ public:
 
   RefToScanQueue *task_queue(int i);
 
+  // A set of cards where updates happened during the GC
+  DirtyCardQueueSet& dirty_card_queue_set() { return _dirty_card_queue_set; }
+
   // Create a G1CollectedHeap with the specified policy.
   // Must call the initialize method afterwards.
   // May not return if something goes wrong.

src/share/vm/gc_implementation/g1/g1RemSet.cpp

@@ -177,11 +177,19 @@ HRInto_G1RemSet::HRInto_G1RemSet(G1CollectedHeap* g1, CardTableModRefBS* ct_bs)
     _cards_scanned(NULL), _total_cards_scanned(0)
 {
   _seq_task = new SubTasksDone(NumSeqTasks);
-  _new_refs = NEW_C_HEAP_ARRAY(GrowableArray<oop*>*, ParallelGCThreads);
+  guarantee(n_workers() > 0, "There should be some workers");
+  _new_refs = NEW_C_HEAP_ARRAY(GrowableArray<oop*>*, n_workers());
+  for (uint i = 0; i < n_workers(); i++) {
+    _new_refs[i] = new (ResourceObj::C_HEAP) GrowableArray<oop*>(8192,true);
+  }
 }
 
 HRInto_G1RemSet::~HRInto_G1RemSet() {
   delete _seq_task;
+  for (uint i = 0; i < n_workers(); i++) {
+    delete _new_refs[i];
+  }
+  FREE_C_HEAP_ARRAY(GrowableArray<oop*>*, _new_refs);
 }
 
 void CountNonCleanMemRegionClosure::do_MemRegion(MemRegion mr) {
@@ -281,8 +289,9 @@ public:
         if (!_ct_bs->is_card_claimed(card_index) &&
             !_ct_bs->is_card_dirty(card_index)) {
           assert(_ct_bs->is_card_clean(card_index) ||
-                 _ct_bs->is_card_claimed(card_index),
-                 "Card is either dirty, clean, or claimed");
+                 _ct_bs->is_card_claimed(card_index) ||
+                 _ct_bs->is_card_deferred(card_index),
+                 "Card is either clean, claimed or deferred");
           if (_ct_bs->claim_card(card_index))
             scanCard(card_index, card_region);
         }
@@ -338,14 +347,12 @@ void HRInto_G1RemSet::scanRS(OopsInHeapRegionClosure* oc, int worker_i) {
 
   _g1p->record_scan_rs_start_time(worker_i, rs_time_start * 1000.0);
   _g1p->record_scan_rs_time(worker_i, scan_rs_time_sec * 1000.0);
-  if (ParallelGCThreads > 0) {
-    // In this case, we called scanNewRefsRS and recorded the corresponding
-    // time.
-    double scan_new_refs_time_ms = _g1p->get_scan_new_refs_time(worker_i);
-    if (scan_new_refs_time_ms > 0.0) {
-      closure_app_time_ms += scan_new_refs_time_ms;
-    }
+
+  double scan_new_refs_time_ms = _g1p->get_scan_new_refs_time(worker_i);
+  if (scan_new_refs_time_ms > 0.0) {
+    closure_app_time_ms += scan_new_refs_time_ms;
   }
+
   _g1p->record_obj_copy_time(worker_i, closure_app_time_ms);
 }
 
@@ -469,8 +476,8 @@ HRInto_G1RemSet::scanNewRefsRS(OopsInHeapRegionClosure* oc,
   double scan_new_refs_start_sec = os::elapsedTime();
   G1CollectedHeap* g1h = G1CollectedHeap::heap();
   CardTableModRefBS* ct_bs = (CardTableModRefBS*) (g1h->barrier_set());
-  while (_new_refs[worker_i]->is_nonempty()) {
-    oop* p = _new_refs[worker_i]->pop();
+  for (int i = 0; i < _new_refs[worker_i]->length(); i++) {
+    oop* p = _new_refs[worker_i]->at(i);
     oop obj = *p;
     // *p was in the collection set when p was pushed on "_new_refs", but
     // another thread may have processed this location from an RS, so it
@@ -480,10 +487,6 @@ HRInto_G1RemSet::scanNewRefsRS(OopsInHeapRegionClosure* oc,
       HeapRegion* r = g1h->heap_region_containing(p);
 
       DEBUG_ONLY(HeapRegion* to = g1h->heap_region_containing(obj));
-      assert(ParallelGCThreads > 1
-             || to->rem_set()->contains_reference(p),
-             "Invariant: pushed after being added."
-             "(Not reliable in parallel code.)");
       oc->set_region(r);
       // If "p" has already been processed concurrently, this is
       // idempotent.
@@ -538,8 +541,8 @@ HRInto_G1RemSet::oops_into_collection_set_do(OopsInHeapRegionClosure* oc,
     }
   } else {
     assert(worker_i == 0, "invariant");
-
     updateRS(0);
+    scanNewRefsRS(oc, 0);
     scanRS(oc, 0);
   }
 }
@@ -559,11 +562,7 @@ prepare_for_oops_into_collection_set_do() {
   assert(!_par_traversal_in_progress, "Invariant between iterations.");
   if (ParallelGCThreads > 0) {
     set_par_traversal(true);
-    int n_workers = _g1->workers()->total_workers();
-    _seq_task->set_par_threads(n_workers);
-    for (uint i = 0; i < ParallelGCThreads; i++)
-      _new_refs[i] = new (ResourceObj::C_HEAP) GrowableArray<oop*>(8192,true);
-
+    _seq_task->set_par_threads((int)n_workers());
     if (cg1r->do_traversal()) {
       updateRS(0);
       // Have to do this again after updaters
@@ -587,6 +586,53 @@ class cleanUpIteratorsClosure : public HeapRegionClosure {
   }
 };
 
+class UpdateRSetOopsIntoCSImmediate : public OopClosure {
+  G1CollectedHeap* _g1;
+public:
+  UpdateRSetOopsIntoCSImmediate(G1CollectedHeap* g1) : _g1(g1) { }
+  virtual void do_oop(narrowOop* p) {
+    guarantee(false, "NYI");
+  }
+  virtual void do_oop(oop* p) {
+    HeapRegion* to = _g1->heap_region_containing(*p);
+    if (to->in_collection_set()) {
+      if (to->rem_set()->add_reference(p, 0)) {
+        _g1->schedule_popular_region_evac(to);
+      }
+    }
+  }
+};
+
+class UpdateRSetOopsIntoCSDeferred : public OopClosure {
+  G1CollectedHeap* _g1;
+  CardTableModRefBS* _ct_bs;
+  DirtyCardQueue* _dcq;
+public:
+  UpdateRSetOopsIntoCSDeferred(G1CollectedHeap* g1, DirtyCardQueue* dcq) :
+    _g1(g1), _ct_bs((CardTableModRefBS*)_g1->barrier_set()), _dcq(dcq) { }
+  virtual void do_oop(narrowOop* p) {
+    guarantee(false, "NYI");
+  }
+  virtual void do_oop(oop* p) {
+    oop obj = *p;
+    if (_g1->obj_in_cs(obj)) {
+      size_t card_index = _ct_bs->index_for(p);
+      if (_ct_bs->mark_card_deferred(card_index)) {
+        _dcq->enqueue((jbyte*)_ct_bs->byte_for_index(card_index));
+      }
+    }
+  }
+};
+
+void HRInto_G1RemSet::new_refs_iterate(OopClosure* cl) {
+  for (size_t i = 0; i < n_workers(); i++) {
+    for (int j = 0; j < _new_refs[i]->length(); j++) {
+      oop* p = _new_refs[i]->at(j);
+      cl->do_oop(p);
+    }
+  }
+}
+
 void HRInto_G1RemSet::cleanup_after_oops_into_collection_set_do() {
   guarantee( _cards_scanned != NULL, "invariant" );
   _total_cards_scanned = 0;
@@ -609,11 +655,25 @@ void HRInto_G1RemSet::cleanup_after_oops_into_collection_set_do() {
     if (cg1r->do_traversal()) {
       cg1r->cg1rThread()->set_do_traversal(false);
     }
-    for (uint i = 0; i < ParallelGCThreads; i++) {
-      delete _new_refs[i];
-    }
     set_par_traversal(false);
   }
+
+  if (_g1->evacuation_failed()) {
+    // Restore remembered sets for the regions pointing into
+    // the collection set.
+    if (G1DeferredRSUpdate) {
+      DirtyCardQueue dcq(&_g1->dirty_card_queue_set());
+      UpdateRSetOopsIntoCSDeferred deferred_update(_g1, &dcq);
+      new_refs_iterate(&deferred_update);
+    } else {
+      UpdateRSetOopsIntoCSImmediate immediate_update(_g1);
+      new_refs_iterate(&immediate_update);
+    }
+  }
+  for (uint i = 0; i < n_workers(); i++) {
+    _new_refs[i]->clear();
+  }
+
   assert(!_par_traversal_in_progress, "Invariant between iterations.");
 }
 
@@ -683,7 +743,8 @@ public:
   bool doHeapRegion(HeapRegion* r) {
     if (!r->in_collection_set() &&
         !r->continuesHumongous() &&
-        !r->is_young()) {
+        !r->is_young() &&
+        !r->is_survivor()) {
       _update_rs_oop_cl.set_from(r);
       UpdateRSObjectClosure update_rs_obj_cl(&_update_rs_oop_cl);
 
@@ -820,7 +881,7 @@ void HRInto_G1RemSet::concurrentRefineOneCard(jbyte* card_ptr, int worker_i) {
   // before all the cards on the region are dirtied. This is unlikely,
   // and it doesn't happen often, but it can happen. So, the extra
   // check below filters out those cards.
-  if (r->is_young()) {
+  if (r->is_young() || r->is_survivor()) {
     return;
   }
   // While we are processing RSet buffers during the collection, we

src/share/vm/gc_implementation/g1/g1RemSet.hpp

@@ -155,6 +155,7 @@ protected:
   bool _par_traversal_in_progress;
   void set_par_traversal(bool b);
   GrowableArray<oop*>** _new_refs;
+  void new_refs_iterate(OopClosure* cl);
 
 public:
   // This is called to reset dual hash tables after the gc pause

src/share/vm/gc_implementation/g1/g1RemSet.inline.hpp

@@ -31,24 +31,7 @@ inline size_t G1RemSet::n_workers() {
 }
 
 inline void HRInto_G1RemSet::write_ref_nv(HeapRegion* from, oop* p) {
-  oop obj = *p;
-  assert(from != NULL && from->is_in_reserved(p),
-         "p is not in a from");
-  HeapRegion* to = _g1->heap_region_containing(obj);
-  if (from != to && to != NULL) {
-    if (!to->popular() && !from->is_survivor()) {
-#if G1_REM_SET_LOGGING
-      gclog_or_tty->print_cr("Adding " PTR_FORMAT " (" PTR_FORMAT ") to RS"
-                             " for region [" PTR_FORMAT ", " PTR_FORMAT ")",
-                             p, obj,
-                             to->bottom(), to->end());
-#endif
-      assert(to->rem_set() != NULL, "Need per-region 'into' remsets.");
-      if (to->rem_set()->add_reference(p)) {
-        _g1->schedule_popular_region_evac(to);
-      }
-    }
-  }
+  par_write_ref(from, p, 0);
 }
 
 inline void HRInto_G1RemSet::write_ref(HeapRegion* from, oop* p) {
@@ -82,7 +65,22 @@ inline void HRInto_G1RemSet::par_write_ref(HeapRegion* from, oop* p, int tid) {
   HeapRegion* to = _g1->heap_region_containing(obj);
   // The test below could be optimized by applying a bit op to to and from.
   if (to != NULL && from != NULL && from != to) {
-    if (!to->popular() && !from->is_survivor()) {
+    bool update_delayed = false;
+    // There is a tricky infinite loop if we keep pushing
+    // self forwarding pointers onto our _new_refs list.
+    // The _par_traversal_in_progress flag is true during the collection pause,
+    // false during the evacuation failure handing.
+    if (_par_traversal_in_progress &&
+        to->in_collection_set() && !self_forwarded(obj)) {
+      _new_refs[tid]->push(p);
+      // Deferred updates to the Cset are either discarded (in the normal case),
+      // or processed (if an evacuation failure occurs) at the end
+      // of the collection.
+      // See HRInto_G1RemSet::cleanup_after_oops_into_collection_set_do().
+      update_delayed = true;
+    }
+
+    if (!to->popular() && !update_delayed) {
 #if G1_REM_SET_LOGGING
       gclog_or_tty->print_cr("Adding " PTR_FORMAT " (" PTR_FORMAT ") to RS"
                              " for region [" PTR_FORMAT ", " PTR_FORMAT ")",
@@ -94,11 +92,5 @@ inline void HRInto_G1RemSet::par_write_ref(HeapRegion* from, oop* p, int tid) {
         _g1->schedule_popular_region_evac(to);
       }
     }
-    // There is a tricky infinite loop if we keep pushing
-    // self forwarding pointers onto our _new_refs list.
-    if (_par_traversal_in_progress &&
-        to->in_collection_set() && !self_forwarded(obj)) {
-      _new_refs[tid]->push(p);
-    }
   }
 }

src/share/vm/gc_implementation/g1/g1_globals.hpp

@@ -172,6 +172,9 @@
   develop(bool, G1RSBarrierUseQueue, true,                                  \
           "If true, use queueing RS barrier")                               \
                                                                             \
+  develop(bool, G1DeferredRSUpdate, true,                                   \
+          "If true, use deferred RS updates")                               \
+                                                                            \
   develop(bool, G1RSLogCheckCardTable, false,                               \
           "If true, verify that no dirty cards remain after RS log "        \
           "processing.")                                                    \

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

@@ -91,15 +91,17 @@ PtrQueueSet::PtrQueueSet(bool notify_when_complete) :
   _n_completed_buffers(0),
   _process_completed_threshold(0), _process_completed(false),
   _buf_free_list(NULL), _buf_free_list_sz(0)
-{}
+{
+  _fl_owner = this;
+}
 
 void** PtrQueueSet::allocate_buffer() {
   assert(_sz > 0, "Didn't set a buffer size.");
-  MutexLockerEx x(_fl_lock, Mutex::_no_safepoint_check_flag);
-  if (_buf_free_list != NULL) {
-    void** res = _buf_free_list;
-    _buf_free_list = (void**)_buf_free_list[0];
-    _buf_free_list_sz--;
+  MutexLockerEx x(_fl_owner->_fl_lock, Mutex::_no_safepoint_check_flag);
+  if (_fl_owner->_buf_free_list != NULL) {
+    void** res = _fl_owner->_buf_free_list;
+    _fl_owner->_buf_free_list = (void**)_fl_owner->_buf_free_list[0];
+    _fl_owner->_buf_free_list_sz--;
     // Just override the next pointer with NULL, just in case we scan this part
     // of the buffer.
     res[0] = NULL;
@@ -111,10 +113,10 @@ void** PtrQueueSet::allocate_buffer() {
 
 void PtrQueueSet::deallocate_buffer(void** buf) {
   assert(_sz > 0, "Didn't set a buffer size.");
-  MutexLockerEx x(_fl_lock, Mutex::_no_safepoint_check_flag);
-  buf[0] = (void*)_buf_free_list;
-  _buf_free_list = buf;
-  _buf_free_list_sz++;
+  MutexLockerEx x(_fl_owner->_fl_lock, Mutex::_no_safepoint_check_flag);
+  buf[0] = (void*)_fl_owner->_buf_free_list;
+  _fl_owner->_buf_free_list = buf;
+  _fl_owner->_buf_free_list_sz++;
 }
 
 void PtrQueueSet::reduce_free_list() {
@@ -207,3 +209,58 @@ void PtrQueueSet::set_buffer_size(size_t sz) {
 void PtrQueueSet::set_process_completed_threshold(size_t sz) {
   _process_completed_threshold = sz;
 }
+
+// Merge lists of buffers. Notify waiting threads if the length of the list
+// exceeds threshold. The source queue is emptied as a result. The queues
+// must share the monitor.
+void PtrQueueSet::merge_bufferlists(PtrQueueSet *src) {
+  assert(_cbl_mon == src->_cbl_mon, "Should share the same lock");
+  MutexLockerEx x(_cbl_mon, Mutex::_no_safepoint_check_flag);
+  if (_completed_buffers_tail == NULL) {
+    assert(_completed_buffers_head == NULL, "Well-formedness");
+    _completed_buffers_head = src->_completed_buffers_head;
+    _completed_buffers_tail = src->_completed_buffers_tail;
+  } else {
+    assert(_completed_buffers_head != NULL, "Well formedness");
+    if (src->_completed_buffers_head != NULL) {
+      _completed_buffers_tail->next = src->_completed_buffers_head;
+      _completed_buffers_tail = src->_completed_buffers_tail;
+    }
+  }
+  _n_completed_buffers += src->_n_completed_buffers;
+
+  src->_n_completed_buffers = 0;
+  src->_completed_buffers_head = NULL;
+  src->_completed_buffers_tail = NULL;
+
+  assert(_completed_buffers_head == NULL && _completed_buffers_tail == NULL ||
+         _completed_buffers_head != NULL && _completed_buffers_tail != NULL,
+         "Sanity");
+
+  if (!_process_completed &&
+      _n_completed_buffers >= _process_completed_threshold) {
+    _process_completed = true;
+    if (_notify_when_complete)
+      _cbl_mon->notify_all();
+  }
+}
+
+// Merge free lists of the two queues. The free list of the source
+// queue is emptied as a result. The queues must share the same
+// mutex that guards free lists.
+void PtrQueueSet::merge_freelists(PtrQueueSet* src) {
+  assert(_fl_lock == src->_fl_lock, "Should share the same lock");
+  MutexLockerEx x(_fl_lock, Mutex::_no_safepoint_check_flag);
+  if (_buf_free_list != NULL) {
+    void **p = _buf_free_list;
+    while (*p != NULL) {
+      p = (void**)*p;
+    }
+    *p = src->_buf_free_list;
+  } else {
+    _buf_free_list = src->_buf_free_list;
+  }
+  _buf_free_list_sz += src->_buf_free_list_sz;
+  src->_buf_free_list = NULL;
+  src->_buf_free_list_sz = 0;
+}

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

@@ -155,6 +155,9 @@ protected:
   Mutex* _fl_lock;
   void** _buf_free_list;
   size_t _buf_free_list_sz;
+  // Queue set can share a freelist. The _fl_owner variable
+  // specifies the owner. It is set to "this" by default.
+  PtrQueueSet* _fl_owner;
 
   // The size of all buffers in the set.
   size_t _sz;
@@ -188,10 +191,13 @@ public:
   // Because of init-order concerns, we can't pass these as constructor
   // arguments.
   void initialize(Monitor* cbl_mon, Mutex* fl_lock,
-                  int max_completed_queue = 0) {
+                  int max_completed_queue = 0,
+                  PtrQueueSet *fl_owner = NULL) {
     _max_completed_queue = max_completed_queue;
     assert(cbl_mon != NULL && fl_lock != NULL, "Init order issue?");
-    _cbl_mon = cbl_mon; _fl_lock = fl_lock;
+    _cbl_mon = cbl_mon;
+    _fl_lock = fl_lock;
+    _fl_owner = (fl_owner != NULL) ? fl_owner : this;
   }
 
   // Return an empty oop array of size _sz (required to be non-zero).
@@ -228,4 +234,7 @@ public:
   void reduce_free_list();
 
   size_t completed_buffers_num() { return _n_completed_buffers; }
+
+  void merge_bufferlists(PtrQueueSet* src);
+  void merge_freelists(PtrQueueSet* src);
 };

src/share/vm/memory/cardTableModRefBS.cpp

@@ -356,18 +356,62 @@ void CardTableModRefBS::write_ref_field_work(void* field, oop newVal) {
   inline_write_ref_field(field, newVal);
 }
 
+/*
+   Claimed and deferred bits are used together in G1 during the evacuation
+   pause. These bits can have the following state transitions:
+   1. The claimed bit can be put over any other card state. Except that
+      the "dirty -> dirty and claimed" transition is checked for in
+      G1 code and is not used.
+   2. Deferred bit can be set only if the previous state of the card
+      was either clean or claimed. mark_card_deferred() is wait-free.
+      We do not care if the operation is be successful because if
+      it does not it will only result in duplicate entry in the update
+      buffer because of the "cache-miss". So it's not worth spinning.
+ */
+
 
 bool CardTableModRefBS::claim_card(size_t card_index) {
   jbyte val = _byte_map[card_index];
-  if (val != claimed_card_val()) {
-    jbyte res = Atomic::cmpxchg((jbyte) claimed_card_val(), &_byte_map[card_index], val);
-    if (res == val)
+  assert(val != dirty_card_val(), "Shouldn't claim a dirty card");
+  while (val == clean_card_val() ||
+         (val & (clean_card_mask_val() | claimed_card_val())) != claimed_card_val()) {
+    jbyte new_val = val;
+    if (val == clean_card_val()) {
+      new_val = (jbyte)claimed_card_val();
+    } else {
+      new_val = val | (jbyte)claimed_card_val();
+    }
+    jbyte res = Atomic::cmpxchg(new_val, &_byte_map[card_index], val);
+    if (res == val) {
       return true;
-    else return false;
+    }
+    val = res;
   }
   return false;
 }
 
+bool CardTableModRefBS::mark_card_deferred(size_t card_index) {
+  jbyte val = _byte_map[card_index];
+  // It's already processed
+  if ((val & (clean_card_mask_val() | deferred_card_val())) == deferred_card_val()) {
+    return false;
+  }
+  // Cached bit can be installed either on a clean card or on a claimed card.
+  jbyte new_val = val;
+  if (val == clean_card_val()) {
+    new_val = (jbyte)deferred_card_val();
+  } else {
+    if (val & claimed_card_val()) {
+      new_val = val | (jbyte)deferred_card_val();
+    }
+  }
+  if (new_val != val) {
+    Atomic::cmpxchg(new_val, &_byte_map[card_index], val);
+  }
+  return true;
+}
+
+
 void CardTableModRefBS::non_clean_card_iterate(Space* sp,
                                                MemRegion mr,
                                                DirtyCardToOopClosure* dcto_cl,

src/share/vm/memory/cardTableModRefBS.hpp

@@ -52,11 +52,15 @@ class CardTableModRefBS: public ModRefBarrierSet {
 
   enum CardValues {
     clean_card                  = -1,
+    // The mask contains zeros in places for all other values.
+    clean_card_mask             = clean_card - 31,
+
     dirty_card                  =  0,
     precleaned_card             =  1,
-    claimed_card                =  3,
-    last_card                   =  4,
-    CT_MR_BS_last_reserved      = 10
+    claimed_card                =  2,
+    deferred_card               =  4,
+    last_card                   =  8,
+    CT_MR_BS_last_reserved      = 16
   };
 
   // dirty and precleaned are equivalent wrt younger_refs_iter.
@@ -254,9 +258,11 @@ public:
   };
 
   static int clean_card_val()      { return clean_card; }
+  static int clean_card_mask_val() { return clean_card_mask; }
   static int dirty_card_val()      { return dirty_card; }
   static int claimed_card_val()    { return claimed_card; }
   static int precleaned_card_val() { return precleaned_card; }
+  static int deferred_card_val()   { return deferred_card; }
 
   // For RTTI simulation.
   bool is_a(BarrierSet::Name bsn) {
@@ -329,7 +335,8 @@ public:
   }
 
   bool is_card_claimed(size_t card_index) {
-    return _byte_map[card_index] == claimed_card_val();
+    jbyte val = _byte_map[card_index];
+    return (val & (clean_card_mask_val() | claimed_card_val())) == claimed_card_val();
   }
 
   bool claim_card(size_t card_index);
@@ -338,6 +345,13 @@ public:
     return _byte_map[card_index] == clean_card_val();
   }
 
+  bool is_card_deferred(size_t card_index) {
+    jbyte val = _byte_map[card_index];
+    return (val & (clean_card_mask_val() | deferred_card_val())) == deferred_card_val();
+  }
+
+  bool mark_card_deferred(size_t card_index);
+
   // Card marking array base (adjusted for heap low boundary)
   // This would be the 0th element of _byte_map, if the heap started at 0x0.
   // But since the heap starts at some higher address, this points to somewhere
@@ -434,6 +448,10 @@ public:
     return byte_for(p) - _byte_map;
   }
 
+  const jbyte* byte_for_index(const size_t card_index) const {
+    return _byte_map + card_index;
+  }
+
   void verify();
   void verify_guard();