# HG changeset patch
# User iveresov
# Date 1242672766 25200
# Node ID d508a8bac491641000776e9fd1b2b8e52a2df262
# Parent  74a1337e4acc3daa2459417a62da8a55bd350280
6841831: G1: assert(contains_reference(from),"We just added it!") fires
Summary: During parallel rset updating we have to make sure that the worker ids of the refinement threads do not intersect with the worker ids that can be claimed by the mutator threads.
Reviewed-by: tonyp

diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.cpp
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.cpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.cpp	Mon May 18 11:52:46 2009 -0700
@@ -33,12 +33,13 @@
   _threads(NULL), _n_threads(0)
 {
   if (G1ConcRefine) {
-    _n_threads = (G1ParallelRSetThreads > 0) ? G1ParallelRSetThreads : ParallelGCThreads;
+    _n_threads = (int)thread_num();
     if (_n_threads > 0) {
       _threads = NEW_C_HEAP_ARRAY(ConcurrentG1RefineThread*, _n_threads);
+      int worker_id_offset = (int)DirtyCardQueueSet::num_par_ids();
       ConcurrentG1RefineThread *next = NULL;
       for (int i = _n_threads - 1; i >= 0; i--) {
-        ConcurrentG1RefineThread* t = new ConcurrentG1RefineThread(this, next, i);
+        ConcurrentG1RefineThread* t = new ConcurrentG1RefineThread(this, next, worker_id_offset, i);
         assert(t != NULL, "Conc refine should have been created");
         assert(t->cg1r() == this, "Conc refine thread should refer to this");
         _threads[i] = t;
@@ -48,6 +49,13 @@
   }
 }
 
+size_t ConcurrentG1Refine::thread_num() {
+  if (G1ConcRefine) {
+    return (G1ParallelRSetThreads > 0) ? G1ParallelRSetThreads : ParallelGCThreads;
+  }
+  return 0;
+}
+
 void ConcurrentG1Refine::init() {
   if (G1ConcRSLogCacheSize > 0 || G1ConcRSCountTraversals) {
     G1CollectedHeap* g1h = G1CollectedHeap::heap();
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.hpp
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.hpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentG1Refine.hpp	Mon May 18 11:52:46 2009 -0700
@@ -29,7 +29,6 @@
 class ConcurrentG1Refine: public CHeapObj {
   ConcurrentG1RefineThread** _threads;
   int _n_threads;
-
   // The cache for card refinement.
   bool     _use_cache;
   bool     _def_use_cache;
@@ -86,4 +85,6 @@
 
   void clear_and_record_card_counts();
   void print_final_card_counts();
+
+  static size_t thread_num();
 };
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.cpp
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.cpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.cpp	Mon May 18 11:52:46 2009 -0700
@@ -30,8 +30,10 @@
 // The CM thread is created when the G1 garbage collector is used
 
 ConcurrentG1RefineThread::
-ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread *next, int worker_id) :
+ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread *next,
+                         int worker_id_offset, int worker_id) :
   ConcurrentGCThread(),
+  _worker_id_offset(worker_id_offset),
   _worker_id(worker_id),
   _active(false),
   _next(next),
@@ -114,7 +116,7 @@
     } else {
       lower_limit = DCQBarrierProcessCompletedThreshold / 4; // For now.
     }
-    while (dcqs.apply_closure_to_completed_buffer(_worker_id, lower_limit)) {
+    while (dcqs.apply_closure_to_completed_buffer(_worker_id + _worker_id_offset, lower_limit)) {
       double end_vtime_sec;
       double elapsed_vtime_sec;
       int elapsed_vtime_ms;
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.hpp
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.hpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentG1RefineThread.hpp	Mon May 18 11:52:46 2009 -0700
@@ -34,6 +34,7 @@
   double _vtime_start;  // Initial virtual time.
   double _vtime_accum;  // Initial virtual time.
   int _worker_id;
+  int _worker_id_offset;
 
   // The refinement threads collection is linked list. A predecessor can activate a successor
   // when the number of the rset update buffer crosses a certain threshold. A successor
@@ -73,7 +74,8 @@
 
  public:
   // Constructor
-  ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread* next, int worker_id);
+  ConcurrentG1RefineThread(ConcurrentG1Refine* cg1r, ConcurrentG1RefineThread* next,
+                           int worker_id_offset, int worker_id);
 
   // Printing
   void print();
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/dirtyCardQueue.cpp
--- a/hotspot/src/share/vm/gc_implementation/g1/dirtyCardQueue.cpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/dirtyCardQueue.cpp	Mon May 18 11:52:46 2009 -0700
@@ -71,11 +71,11 @@
   _all_active = true;
 }
 
+// Determines how many mutator threads can process the buffers in parallel.
 size_t DirtyCardQueueSet::num_par_ids() {
-  return MAX2(ParallelGCThreads, (size_t)2);
+  return os::processor_count();
 }
 
-
 void DirtyCardQueueSet::initialize(Monitor* cbl_mon, Mutex* fl_lock,
                                    int max_completed_queue,
                                    Mutex* lock, PtrQueueSet* fl_owner) {
@@ -85,8 +85,6 @@
 
   _shared_dirty_card_queue.set_lock(lock);
   _free_ids = new FreeIdSet((int) num_par_ids(), _cbl_mon);
-  bool b = _free_ids->claim_perm_id(0);
-  guarantee(b, "Must reserve id zero for concurrent refinement thread.");
 }
 
 void DirtyCardQueueSet::handle_zero_index_for_thread(JavaThread* t) {
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/g1/heapRegionRemSet.cpp
--- a/hotspot/src/share/vm/gc_implementation/g1/heapRegionRemSet.cpp	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/g1/heapRegionRemSet.cpp	Mon May 18 11:52:46 2009 -0700
@@ -1052,10 +1052,11 @@
 
 }
 
+// Determines how many threads can add records to an rset in parallel.
+// This can be done by either mutator threads together with the
+// concurrent refinement threads or GC threads.
 int HeapRegionRemSet::num_par_rem_sets() {
-  // We always have at least two, so that a mutator thread can claim an
-  // id and add to a rem set.
-  return (int) MAX2(ParallelGCThreads, (size_t)2);
+  return (int)MAX2(DirtyCardQueueSet::num_par_ids() + ConcurrentG1Refine::thread_num(), ParallelGCThreads);
 }
 
 HeapRegionRemSet::HeapRegionRemSet(G1BlockOffsetSharedArray* bosa,
diff -r 74a1337e4acc -r d508a8bac491 hotspot/src/share/vm/gc_implementation/includeDB_gc_g1
--- a/hotspot/src/share/vm/gc_implementation/includeDB_gc_g1	Mon May 11 16:30:56 2009 -0700
+++ b/hotspot/src/share/vm/gc_implementation/includeDB_gc_g1	Mon May 18 11:52:46 2009 -0700
@@ -282,6 +282,7 @@
 
 heapRegionRemSet.cpp                    allocation.hpp
 heapRegionRemSet.cpp                    bitMap.inline.hpp
+heapRegionRemSet.cpp                    concurrentG1Refine.hpp
 heapRegionRemSet.cpp                    g1BlockOffsetTable.inline.hpp
 heapRegionRemSet.cpp                    g1CollectedHeap.inline.hpp
 heapRegionRemSet.cpp                    heapRegionRemSet.hpp