6819065: G1: eliminate high serial card table clearing time

Reviewed-by: iveresov, tonyp

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

@@ -446,6 +446,59 @@ void YoungList::print() {
   gclog_or_tty->print_cr("");
 }
 
+void G1CollectedHeap::push_dirty_cards_region(HeapRegion* hr)
+{
+  // Claim the right to put the region on the dirty cards region list
+  // by installing a self pointer.
+  HeapRegion* next = hr->get_next_dirty_cards_region();
+  if (next == NULL) {
+    HeapRegion* res = (HeapRegion*)
+      Atomic::cmpxchg_ptr(hr, hr->next_dirty_cards_region_addr(),
+                          NULL);
+    if (res == NULL) {
+      HeapRegion* head;
+      do {
+        // Put the region to the dirty cards region list.
+        head = _dirty_cards_region_list;
+        next = (HeapRegion*)
+          Atomic::cmpxchg_ptr(hr, &_dirty_cards_region_list, head);
+        if (next == head) {
+          assert(hr->get_next_dirty_cards_region() == hr,
+                 "hr->get_next_dirty_cards_region() != hr");
+          if (next == NULL) {
+            // The last region in the list points to itself.
+            hr->set_next_dirty_cards_region(hr);
+          } else {
+            hr->set_next_dirty_cards_region(next);
+          }
+        }
+      } while (next != head);
+    }
+  }
+}
+
+HeapRegion* G1CollectedHeap::pop_dirty_cards_region()
+{
+  HeapRegion* head;
+  HeapRegion* hr;
+  do {
+    head = _dirty_cards_region_list;
+    if (head == NULL) {
+      return NULL;
+    }
+    HeapRegion* new_head = head->get_next_dirty_cards_region();
+    if (head == new_head) {
+      // The last region.
+      new_head = NULL;
+    }
+    hr = (HeapRegion*)Atomic::cmpxchg_ptr(new_head, &_dirty_cards_region_list,
+                                          head);
+  } while (hr != head);
+  assert(hr != NULL, "invariant");
+  hr->set_next_dirty_cards_region(NULL);
+  return hr;
+}
+
 void G1CollectedHeap::stop_conc_gc_threads() {
   _cg1r->stop();
   _czft->stop();
@@ -1329,7 +1382,8 @@ G1CollectedHeap::G1CollectedHeap(G1CollectorPolicy* policy_) :
   _gc_time_stamp(0),
   _surviving_young_words(NULL),
   _in_cset_fast_test(NULL),
-  _in_cset_fast_test_base(NULL) {
+  _in_cset_fast_test_base(NULL),
+  _dirty_cards_region_list(NULL) {
   _g1h = this; // To catch bugs.
   if (_process_strong_tasks == NULL || !_process_strong_tasks->valid()) {
     vm_exit_during_initialization("Failed necessary allocation.");
@@ -4691,15 +4745,58 @@ void G1CollectedHeap::dirtyCardsForYoungRegions(CardTableModRefBS* ct_bs, HeapRe
   }
 }
 
+
+class G1ParCleanupCTTask : public AbstractGangTask {
+  CardTableModRefBS* _ct_bs;
+  G1CollectedHeap* _g1h;
+public:
+  G1ParCleanupCTTask(CardTableModRefBS* ct_bs,
+                     G1CollectedHeap* g1h) :
+    AbstractGangTask("G1 Par Cleanup CT Task"),
+    _ct_bs(ct_bs),
+    _g1h(g1h)
+  { }
+
+  void work(int i) {
+    HeapRegion* r;
+    while (r = _g1h->pop_dirty_cards_region()) {
+      clear_cards(r);
+    }
+  }
+  void clear_cards(HeapRegion* r) {
+    // Cards for Survivor and Scan-Only regions will be dirtied later.
+    if (!r->is_scan_only() && !r->is_survivor()) {
+      _ct_bs->clear(MemRegion(r->bottom(), r->end()));
+    }
+  }
+};
+
+
 void G1CollectedHeap::cleanUpCardTable() {
   CardTableModRefBS* ct_bs = (CardTableModRefBS*) (barrier_set());
   double start = os::elapsedTime();
 
-  ct_bs->clear(_g1_committed);
-
+  // Iterate over the dirty cards region list.
+  G1ParCleanupCTTask cleanup_task(ct_bs, this);
+  if (ParallelGCThreads > 0) {
+    set_par_threads(workers()->total_workers());
+    workers()->run_task(&cleanup_task);
+    set_par_threads(0);
+  } else {
+    while (_dirty_cards_region_list) {
+      HeapRegion* r = _dirty_cards_region_list;
+      cleanup_task.clear_cards(r);
+      _dirty_cards_region_list = r->get_next_dirty_cards_region();
+      if (_dirty_cards_region_list == r) {
+        // The last region.
+        _dirty_cards_region_list = NULL;
+      }
+      r->set_next_dirty_cards_region(NULL);
+    }
+  }
   // now, redirty the cards of the scan-only and survivor regions
   // (it seemed faster to do it this way, instead of iterating over
-  // all regions and then clearing / dirtying as approprite)
+  // all regions and then clearing / dirtying as appropriate)
   dirtyCardsForYoungRegions(ct_bs, _young_list->first_scan_only_region());
   dirtyCardsForYoungRegions(ct_bs, _young_list->first_survivor_region());
 

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

@@ -158,6 +158,7 @@ class G1CollectedHeap : public SharedHeap {
   friend class RegionSorter;
   friend class CountRCClosure;
   friend class EvacPopObjClosure;
+  friend class G1ParCleanupCTTask;
 
   // Other related classes.
   friend class G1MarkSweep;
@@ -1191,6 +1192,16 @@ public:
   ConcurrentMark* concurrent_mark() const { return _cm; }
   ConcurrentG1Refine* concurrent_g1_refine() const { return _cg1r; }
 
+  // The dirty cards region list is used to record a subset of regions
+  // whose cards need clearing. The list if populated during the
+  // remembered set scanning and drained during the card table
+  // cleanup. Although the methods are reentrant, population/draining
+  // phases must not overlap. For synchronization purposes the last
+  // element on the list points to itself.
+  HeapRegion* _dirty_cards_region_list;
+  void push_dirty_cards_region(HeapRegion* hr);
+  HeapRegion* pop_dirty_cards_region();
+
 public:
   void stop_conc_gc_threads();
 

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

@@ -219,6 +219,7 @@ public:
     HeapRegionRemSet* hrrs = r->rem_set();
     if (hrrs->iter_is_complete()) return false; // All done.
     if (!_try_claimed && !hrrs->claim_iter()) return false;
+    _g1h->push_dirty_cards_region(r);
     // If we didn't return above, then
     //   _try_claimed || r->claim_iter()
     // is true: either we're supposed to work on claimed-but-not-complete
@@ -242,6 +243,10 @@ public:
       assert(card_region != NULL, "Yielding cards not in the heap?");
       _cards++;
 
+      if (!card_region->is_on_dirty_cards_region_list()) {
+        _g1h->push_dirty_cards_region(card_region);
+      }
+
        // If the card is dirty, then we will scan it during updateRS.
       if (!card_region->in_collection_set() && !_ct_bs->is_card_dirty(card_index)) {
           if (!_ct_bs->is_card_claimed(card_index) && _ct_bs->claim_card(card_index)) {

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

@@ -351,6 +351,7 @@ HeapRegion(G1BlockOffsetSharedArray* sharedOffsetArray,
     _claimed(InitialClaimValue), _evacuation_failed(false),
     _prev_marked_bytes(0), _next_marked_bytes(0), _sort_index(-1),
     _young_type(NotYoung), _next_young_region(NULL),
+    _next_dirty_cards_region(NULL),
     _young_index_in_cset(-1), _surv_rate_group(NULL), _age_index(-1),
     _rem_set(NULL), _zfs(NotZeroFilled)
 {

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

@@ -227,6 +227,9 @@ class HeapRegion: public G1OffsetTableContigSpace {
   // next region in the young "generation" region set
   HeapRegion* _next_young_region;
 
+  // Next region whose cards need cleaning
+  HeapRegion* _next_dirty_cards_region;
+
   // For parallel heapRegion traversal.
   jint _claimed;
 
@@ -468,6 +471,11 @@ class HeapRegion: public G1OffsetTableContigSpace {
     _next_young_region = hr;
   }
 
+  HeapRegion* get_next_dirty_cards_region() const { return _next_dirty_cards_region; }
+  HeapRegion** next_dirty_cards_region_addr() { return &_next_dirty_cards_region; }
+  void set_next_dirty_cards_region(HeapRegion* hr) { _next_dirty_cards_region = hr; }
+  bool is_on_dirty_cards_region_list() const { return get_next_dirty_cards_region() != NULL; }
+
   // Allows logical separation between objects allocated before and after.
   void save_marks();