--- old/src/hotspot/share/gc/g1/g1ConcurrentRefine.hpp	2017-11-03 16:09:33.960017679 +0100
+++ new/src/hotspot/share/gc/g1/g1ConcurrentRefine.hpp	2017-11-03 16:09:33.545004889 +0100
@@ -30,30 +30,59 @@
 
 // Forward decl
 class CardTableEntryClosure;
+class G1ConcurrentRefine;
 class G1ConcurrentRefineThread;
 class outputStream;
 class ThreadClosure;
 
-class G1ConcurrentRefine : public CHeapObj<mtGC> {
+// Helper class for refinement thread management. Used to start, stop and
+// iterate over them.
+class G1ConcurrentRefineThreadControl VALUE_OBJ_CLASS_SPEC {
+  G1ConcurrentRefine* _cg1r;
+
   G1ConcurrentRefineThread** _threads;
-  uint _n_worker_threads;
- /*
-  * The value of the update buffer queue length falls into one of 3 zones:
-  * green, yellow, red. If the value is in [0, green) nothing is
-  * done, the buffers are left unprocessed to enable the caching effect of the
-  * dirtied cards. In the yellow zone [green, yellow) the concurrent refinement
-  * threads are gradually activated. In [yellow, red) all threads are
-  * running. If the length becomes red (max queue length) the mutators start
-  * processing the buffers.
-  *
-  * There are some interesting cases (when G1UseAdaptiveConcRefinement
-  * is turned off):
-  * 1) green = yellow = red = 0. In this case the mutator will process all
-  *    buffers. Except for those that are created by the deferred updates
-  *    machinery during a collection.
-  * 2) green = 0. Means no caching. Can be a good way to minimize the
-  *    amount of time spent updating rsets during a collection.
-  */
+  uint _num_max_threads;
+public:
+  G1ConcurrentRefineThreadControl();
+  ~G1ConcurrentRefineThreadControl();
+
+  void initialize(G1ConcurrentRefine* cg1r, uint num_max_threads);
+
+  // If there is a "successor" thread that can be activated given the current id,
+  // activate it.
+  void maybe_activate_next(uint cur_worker_id);
+
+  void print_on(outputStream* st) const;
+  void worker_threads_do(ThreadClosure* tc);
+  void stop();
+};
+
+// Controls refinement threads and their activation based on the number of completed
+// buffers currently available in the global dirty card queue.
+// Refinement threads pick work from the queue based on these thresholds. They are activated
+// gradually based on the amount of work to do.
+// Refinement thread n activates thread n+1 if the instance of this class determines there
+// is enough work available. Threads deactivate themselves if the current amount of
+// completed buffers falls below their individual threshold.
+class G1ConcurrentRefine : public CHeapObj<mtGC> {
+  G1ConcurrentRefineThreadControl _thread_control;
+  /*
+   * The value of the completed dirty card queue length falls into one of 3 zones:
+   * green, yellow, red. If the value is in [0, green) nothing is
+   * done, the buffers are left unprocessed to enable the caching effect of the
+   * dirtied cards. In the yellow zone [green, yellow) the concurrent refinement
+   * threads are gradually activated. In [yellow, red) all threads are
+   * running. If the length becomes red (max queue length) the mutators start
+   * processing the buffers.
+   *
+   * There are some interesting cases (when G1UseAdaptiveConcRefinement
+   * is turned off):
+   * 1) green = yellow = red = 0. In this case the mutator will process all
+   *    buffers. Except for those that are created by the deferred updates
+   *    machinery during a collection.
+   * 2) green = 0. Means no caching. Can be a good way to minimize the
+   *    amount of time spent updating remembered sets during a collection.
+   */
   size_t _green_zone;
   size_t _yellow_zone;
   size_t _red_zone;
@@ -69,24 +98,31 @@
                     size_t update_rs_processed_buffers,
                     double goal_ms);
 
-  // Update thread thresholds to account for updated zone values.
-  void update_thread_thresholds();
+  static uint worker_id_offset();
+  void maybe_activate_more_threads(uint worker_id, size_t num_cur_buffers);
 
- public:
+public:
   ~G1ConcurrentRefine();
 
-  // Returns a G1ConcurrentRefine instance if succeeded to create/initialize G1ConcurrentRefine and G1ConcurrentRefineThreads.
-  // Otherwise, returns NULL with error code.
+  // Returns a G1ConcurrentRefine instance if succeeded to create/initialize the
+  // G1ConcurrentRefine instance. Otherwise, returns NULL with error code.
   static G1ConcurrentRefine* create(jint* ecode);
 
   void stop();
 
+  // Adjust refinement thresholds based on work done during the pause and the goal time.
   void adjust(double update_rs_time, size_t update_rs_processed_buffers, double goal_ms);
 
+  size_t activation_threshold(uint worker_id) const;
+  size_t deactivation_threshold(uint worker_id) const;
+  // Perform a single refinement step. Called by the refinement threads when woken up.
+  bool do_refinement_step(uint worker_id);
+
   // Iterate over all concurrent refinement threads applying the given closure.
   void threads_do(ThreadClosure *tc);
 
-  static uint thread_num();
+  // Maximum number of refinement threads.
+  static uint max_num_threads();
 
   void print_threads_on(outputStream* st) const;