src/hotspot/share/gc/g1/g1DirtyCardQueue.cpp

rev 56899 : 8087198: G1 card refinement: batching, sorting
Reviewed-by: tschatzl

@@ -224,25 +224,109 @@
   _completed_buffers_tail = NULL;
   _num_cards = 0;
   return result;
 }
 
+class G1RefineBufferedCards : public StackObj {
+  BufferNode* const _node;
+  CardTable::CardValue** const _node_buffer;
+  const size_t _node_buffer_size;
+  size_t* _total_refined_cards;
+  G1RemSet* const _g1rs;
+  DEBUG_ONLY(KVHashtable<CardTable::CardValue* COMMA HeapWord* COMMA mtGC> _card_top_map;)
+
+  static inline int compare_card(const void* p1,
+                                 const void* p2) {
+    return *(CardTable::CardValue**)p2 - *(CardTable::CardValue**)p1;
+  }
+
+  // Sorts the cards from start_index to _node_buffer_size in *decreasing*
+  // address order. This order improves performance of processing the cards
+  // later starting from start_index.
+  void sort_cards(size_t start_index) {
+    qsort(_node_buffer + start_index,
+          _node_buffer_size - start_index,
+          sizeof(CardTable::CardValue*),
+          compare_card);
+  }
+
+  // Returns the index to the first clean card in the buffer.
+  size_t clean_cards() {
+    const size_t start = _node->index();
+    assert(start <= _node_buffer_size, "invariant");
+    size_t first_clean = _node_buffer_size;
+    // We don't check for SuspendibleThreadSet::should_yield(), because
+    // cleaning and redirtying the cards is fast.
+    for (int i = _node_buffer_size - 1; i >= static_cast<int>(start); --i) {
+      CardTable::CardValue* cp = _node_buffer[i];
+      if (_g1rs->clean_card_before_refine(cp
+                                          DEBUG_ONLY(COMMA _card_top_map))) {
+        first_clean--;
+        _node_buffer[first_clean] = cp;
+      }
+    }
+    assert(first_clean >= start && first_clean <= _node_buffer_size, "invariant");
+    // Skipped cards are considered as refined.
+    *_total_refined_cards += first_clean - start;
+    return first_clean;
+  }
+
+  bool refine_cleaned_cards(uint worker_id, size_t start_index) {
+    for (size_t i = start_index; i < _node_buffer_size; ++i) {
+      if (SuspendibleThreadSet::should_yield()) {
+        redirty_unrefined_cards(i);
+        _node->set_index(i);
+        return false;
+      }
+      _g1rs->refine_card_concurrently(_node_buffer[i], worker_id
+                                      DEBUG_ONLY(COMMA _card_top_map));
+      *_total_refined_cards += 1;
+    }
+    _node->set_index(_node_buffer_size);
+    return true;
+  }
+
+  void redirty_unrefined_cards(size_t start) {
+    for ( ; start < _node_buffer_size; ++start) {
+      *_node_buffer[start] = G1CardTable::dirty_card_val();
+    }
+  }
+
+public:
+  G1RefineBufferedCards(BufferNode* node,
+                        size_t node_buffer_size,
+                        size_t* total_refined_cards) :
+    _node(node),
+    _node_buffer(reinterpret_cast<CardTable::CardValue**>(BufferNode::make_buffer_from_node(node))),
+    _node_buffer_size(node_buffer_size),
+    _total_refined_cards(total_refined_cards),
+    _g1rs(G1CollectedHeap::heap()->rem_set())
+    DEBUG_ONLY(COMMA _card_top_map(node_buffer_size)) {}
+
+  bool refine(uint worker_id) {
+    size_t first_clean_index = clean_cards();
+    // This fence serves two purposes. First, the cards must be cleaned
+    // before processing the contents. Second, we can't proceed with
+    // processing a region until after the read of the region's top in
+    // collect_and_clean_cards(), for synchronization with possibly concurrent
+    // humongous object allocation (see comment at the StoreStore fence before
+    // setting the regions' tops in humongous allocation path).
+    // It's okay that reading region's top and reading region's type were racy
+    // wrto each other. We need both set, in any order, to proceed.
+    OrderAccess::fence();
+    sort_cards(first_clean_index);
+    return refine_cleaned_cards(worker_id, first_clean_index);
+  }
+};
+
 bool G1DirtyCardQueueSet::refine_buffer(BufferNode* node,
                                         uint worker_id,
                                         size_t* total_refined_cards) {
-  G1RemSet* rem_set = G1CollectedHeap::heap()->rem_set();
-  size_t size = buffer_size();
-  void** buffer = BufferNode::make_buffer_from_node(node);
-  size_t i = node->index();
-  assert(i <= size, "invariant");
-  for ( ; (i < size) && !SuspendibleThreadSet::should_yield(); ++i) {
-    CardTable::CardValue* cp = static_cast<CardTable::CardValue*>(buffer[i]);
-    rem_set->refine_card_concurrently(cp, worker_id);
-  }
-  *total_refined_cards += (i - node->index());
-  node->set_index(i);
-  return i == size;
+  G1RefineBufferedCards buffered_cards(node,
+                                       buffer_size(),
+                                       total_refined_cards);
+  return buffered_cards.refine(worker_id);
 }
 
 #ifndef ASSERT
 #define assert_fully_consumed(node, buffer_size)
 #else