1 /*
   2  * Copyright (c) 2001, 2013, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP
  27 
  28 #include "gc_implementation/g1/heapRegionSet.hpp"
  29 #include "utilities/taskqueue.hpp"
  30 
  31 class G1CollectedHeap;
  32 class CMTask;
  33 typedef GenericTaskQueue<oop, mtGC>            CMTaskQueue;
  34 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
  35 
  36 // Closure used by CM during concurrent reference discovery
  37 // and reference processing (during remarking) to determine
  38 // if a particular object is alive. It is primarily used
  39 // to determine if referents of discovered reference objects
  40 // are alive. An instance is also embedded into the
  41 // reference processor as the _is_alive_non_header field
  42 class G1CMIsAliveClosure: public BoolObjectClosure {
  43   G1CollectedHeap* _g1;
  44  public:
  45   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
  46 
  47   bool do_object_b(oop obj);
  48 };
  49 
  50 // A generic CM bit map.  This is essentially a wrapper around the BitMap
  51 // class, with one bit per (1<<_shifter) HeapWords.
  52 
  53 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
  54  protected:
  55   HeapWord* _bmStartWord;      // base address of range covered by map
  56   size_t    _bmWordSize;       // map size (in #HeapWords covered)
  57   const int _shifter;          // map to char or bit
  58   VirtualSpace _virtual_space; // underlying the bit map
  59   BitMap    _bm;               // the bit map itself
  60 
  61  public:
  62   // constructor
  63   CMBitMapRO(int shifter);
  64 
  65   enum { do_yield = true };
  66 
  67   // inquiries
  68   HeapWord* startWord()   const { return _bmStartWord; }
  69   size_t    sizeInWords() const { return _bmWordSize;  }
  70   // the following is one past the last word in space
  71   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
  72 
  73   // read marks
  74 
  75   bool isMarked(HeapWord* addr) const {
  76     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
  77            "outside underlying space?");
  78     return _bm.at(heapWordToOffset(addr));
  79   }
  80 
  81   // iteration
  82   inline bool iterate(BitMapClosure* cl, MemRegion mr);
  83   inline bool iterate(BitMapClosure* cl);
  84 
  85   // Return the address corresponding to the next marked bit at or after
  86   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  87   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  88   HeapWord* getNextMarkedWordAddress(HeapWord* addr,
  89                                      HeapWord* limit = NULL) const;
  90   // Return the address corresponding to the next unmarked bit at or after
  91   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  92   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  93   HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
  94                                        HeapWord* limit = NULL) const;
  95 
  96   // conversion utilities
  97   HeapWord* offsetToHeapWord(size_t offset) const {
  98     return _bmStartWord + (offset << _shifter);
  99   }
 100   size_t heapWordToOffset(HeapWord* addr) const {
 101     return pointer_delta(addr, _bmStartWord) >> _shifter;
 102   }
 103   int heapWordDiffToOffsetDiff(size_t diff) const;
 104 
 105   // The argument addr should be the start address of a valid object
 106   HeapWord* nextObject(HeapWord* addr) {
 107     oop obj = (oop) addr;
 108     HeapWord* res =  addr + obj->size();
 109     assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
 110     return res;
 111   }
 112 
 113   void print_on_error(outputStream* st, const char* prefix) const;
 114 
 115   // debugging
 116   NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
 117 };
 118 
 119 class CMBitMap : public CMBitMapRO {
 120 
 121  public:
 122   // constructor
 123   CMBitMap(int shifter) :
 124     CMBitMapRO(shifter) {}
 125 
 126   // Allocates the back store for the marking bitmap
 127   bool allocate(ReservedSpace heap_rs);
 128 
 129   // write marks
 130   void mark(HeapWord* addr) {
 131     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 132            "outside underlying space?");
 133     _bm.set_bit(heapWordToOffset(addr));
 134   }
 135   void clear(HeapWord* addr) {
 136     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 137            "outside underlying space?");
 138     _bm.clear_bit(heapWordToOffset(addr));
 139   }
 140   bool parMark(HeapWord* addr) {
 141     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 142            "outside underlying space?");
 143     return _bm.par_set_bit(heapWordToOffset(addr));
 144   }
 145   bool parClear(HeapWord* addr) {
 146     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 147            "outside underlying space?");
 148     return _bm.par_clear_bit(heapWordToOffset(addr));
 149   }
 150   void markRange(MemRegion mr);
 151   void clearAll();
 152   void clearRange(MemRegion mr);
 153 
 154   // Starting at the bit corresponding to "addr" (inclusive), find the next
 155   // "1" bit, if any.  This bit starts some run of consecutive "1"'s; find
 156   // the end of this run (stopping at "end_addr").  Return the MemRegion
 157   // covering from the start of the region corresponding to the first bit
 158   // of the run to the end of the region corresponding to the last bit of
 159   // the run.  If there is no "1" bit at or after "addr", return an empty
 160   // MemRegion.
 161   MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
 162 };
 163 
 164 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
 165 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
 166   VirtualSpace _virtual_space;   // Underlying backing store for actual stack
 167   ConcurrentMark* _cm;
 168   oop* _base;        // bottom of stack
 169   jint _index;       // one more than last occupied index
 170   jint _capacity;    // max #elements
 171   jint _saved_index; // value of _index saved at start of GC
 172   NOT_PRODUCT(jint _max_depth;)   // max depth plumbed during run
 173 
 174   bool  _overflow;
 175   bool  _should_expand;
 176   DEBUG_ONLY(bool _drain_in_progress;)
 177   DEBUG_ONLY(bool _drain_in_progress_yields;)
 178 
 179  public:
 180   CMMarkStack(ConcurrentMark* cm);
 181   ~CMMarkStack();
 182 
 183 #ifndef PRODUCT
 184   jint max_depth() const {
 185     return _max_depth;
 186   }
 187 #endif
 188 
 189   bool allocate(size_t capacity);
 190 
 191   oop pop() {
 192     if (!isEmpty()) {
 193       return _base[--_index] ;
 194     }
 195     return NULL;
 196   }
 197 
 198   // If overflow happens, don't do the push, and record the overflow.
 199   // *Requires* that "ptr" is already marked.
 200   void push(oop ptr) {
 201     if (isFull()) {
 202       // Record overflow.
 203       _overflow = true;
 204       return;
 205     } else {
 206       _base[_index++] = ptr;
 207       NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
 208     }
 209   }
 210   // Non-block impl.  Note: concurrency is allowed only with other
 211   // "par_push" operations, not with "pop" or "drain".  We would need
 212   // parallel versions of them if such concurrency was desired.
 213   void par_push(oop ptr);
 214 
 215   // Pushes the first "n" elements of "ptr_arr" on the stack.
 216   // Non-block impl.  Note: concurrency is allowed only with other
 217   // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
 218   void par_adjoin_arr(oop* ptr_arr, int n);
 219 
 220   // Pushes the first "n" elements of "ptr_arr" on the stack.
 221   // Locking impl: concurrency is allowed only with
 222   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 223   // locking strategy.
 224   void par_push_arr(oop* ptr_arr, int n);
 225 
 226   // If returns false, the array was empty.  Otherwise, removes up to "max"
 227   // elements from the stack, and transfers them to "ptr_arr" in an
 228   // unspecified order.  The actual number transferred is given in "n" ("n
 229   // == 0" is deliberately redundant with the return value.)  Locking impl:
 230   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 231   // operations, which use the same locking strategy.
 232   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 233 
 234   // Drain the mark stack, applying the given closure to all fields of
 235   // objects on the stack.  (That is, continue until the stack is empty,
 236   // even if closure applications add entries to the stack.)  The "bm"
 237   // argument, if non-null, may be used to verify that only marked objects
 238   // are on the mark stack.  If "yield_after" is "true", then the
 239   // concurrent marker performing the drain offers to yield after
 240   // processing each object.  If a yield occurs, stops the drain operation
 241   // and returns false.  Otherwise, returns true.
 242   template<class OopClosureClass>
 243   bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
 244 
 245   bool isEmpty()    { return _index == 0; }
 246   bool isFull()     { return _index == _capacity; }
 247   int  maxElems()   { return _capacity; }
 248 
 249   bool overflow() { return _overflow; }
 250   void clear_overflow() { _overflow = false; }
 251 
 252   bool should_expand() const { return _should_expand; }
 253   void set_should_expand();
 254 
 255   // Expand the stack, typically in response to an overflow condition
 256   void expand();
 257 
 258   int  size() { return _index; }
 259 
 260   void setEmpty()   { _index = 0; clear_overflow(); }
 261 
 262   // Record the current index.
 263   void note_start_of_gc();
 264 
 265   // Make sure that we have not added any entries to the stack during GC.
 266   void note_end_of_gc();
 267 
 268   // iterate over the oops in the mark stack, up to the bound recorded via
 269   // the call above.
 270   void oops_do(OopClosure* f);
 271 };
 272 
 273 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
 274 private:
 275 #ifndef PRODUCT
 276   uintx _num_remaining;
 277   bool _force;
 278 #endif // !defined(PRODUCT)
 279 
 280 public:
 281   void init() PRODUCT_RETURN;
 282   void update() PRODUCT_RETURN;
 283   bool should_force() PRODUCT_RETURN_( return false; );
 284 };
 285 
 286 // this will enable a variety of different statistics per GC task
 287 #define _MARKING_STATS_       0
 288 // this will enable the higher verbose levels
 289 #define _MARKING_VERBOSE_     0
 290 
 291 #if _MARKING_STATS_
 292 #define statsOnly(statement)  \
 293 do {                          \
 294   statement ;                 \
 295 } while (0)
 296 #else // _MARKING_STATS_
 297 #define statsOnly(statement)  \
 298 do {                          \
 299 } while (0)
 300 #endif // _MARKING_STATS_
 301 
 302 typedef enum {
 303   no_verbose  = 0,   // verbose turned off
 304   stats_verbose,     // only prints stats at the end of marking
 305   low_verbose,       // low verbose, mostly per region and per major event
 306   medium_verbose,    // a bit more detailed than low
 307   high_verbose       // per object verbose
 308 } CMVerboseLevel;
 309 
 310 class YoungList;
 311 
 312 // Root Regions are regions that are not empty at the beginning of a
 313 // marking cycle and which we might collect during an evacuation pause
 314 // while the cycle is active. Given that, during evacuation pauses, we
 315 // do not copy objects that are explicitly marked, what we have to do
 316 // for the root regions is to scan them and mark all objects reachable
 317 // from them. According to the SATB assumptions, we only need to visit
 318 // each object once during marking. So, as long as we finish this scan
 319 // before the next evacuation pause, we can copy the objects from the
 320 // root regions without having to mark them or do anything else to them.
 321 //
 322 // Currently, we only support root region scanning once (at the start
 323 // of the marking cycle) and the root regions are all the survivor
 324 // regions populated during the initial-mark pause.
 325 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
 326 private:
 327   YoungList*           _young_list;
 328   ConcurrentMark*      _cm;
 329 
 330   volatile bool        _scan_in_progress;
 331   volatile bool        _should_abort;
 332   HeapRegion* volatile _next_survivor;
 333 
 334 public:
 335   CMRootRegions();
 336   // We actually do most of the initialization in this method.
 337   void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
 338 
 339   // Reset the claiming / scanning of the root regions.
 340   void prepare_for_scan();
 341 
 342   // Forces get_next() to return NULL so that the iteration aborts early.
 343   void abort() { _should_abort = true; }
 344 
 345   // Return true if the CM thread are actively scanning root regions,
 346   // false otherwise.
 347   bool scan_in_progress() { return _scan_in_progress; }
 348 
 349   // Claim the next root region to scan atomically, or return NULL if
 350   // all have been claimed.
 351   HeapRegion* claim_next();
 352 
 353   // Flag that we're done with root region scanning and notify anyone
 354   // who's waiting on it. If aborted is false, assume that all regions
 355   // have been claimed.
 356   void scan_finished();
 357 
 358   // If CM threads are still scanning root regions, wait until they
 359   // are done. Return true if we had to wait, false otherwise.
 360   bool wait_until_scan_finished();
 361 };
 362 
 363 class ConcurrentMarkThread;
 364 
 365 class ConcurrentMark: public CHeapObj<mtGC> {
 366   friend class CMMarkStack;
 367   friend class ConcurrentMarkThread;
 368   friend class CMTask;
 369   friend class CMBitMapClosure;
 370   friend class CMGlobalObjectClosure;
 371   friend class CMRemarkTask;
 372   friend class CMConcurrentMarkingTask;
 373   friend class G1ParNoteEndTask;
 374   friend class CalcLiveObjectsClosure;
 375   friend class G1CMRefProcTaskProxy;
 376   friend class G1CMRefProcTaskExecutor;
 377   friend class G1CMKeepAliveAndDrainClosure;
 378   friend class G1CMDrainMarkingStackClosure;
 379 
 380 protected:
 381   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 382   G1CollectedHeap*      _g1h;        // The heap
 383   uint                  _parallel_marking_threads; // The number of marking
 384                                                    // threads we're using
 385   uint                  _max_parallel_marking_threads; // Max number of marking
 386                                                        // threads we'll ever use
 387   double                _sleep_factor; // How much we have to sleep, with
 388                                        // respect to the work we just did, to
 389                                        // meet the marking overhead goal
 390   double                _marking_task_overhead; // Marking target overhead for
 391                                                 // a single task
 392 
 393   // Same as the two above, but for the cleanup task
 394   double                _cleanup_sleep_factor;
 395   double                _cleanup_task_overhead;
 396 
 397   FreeRegionList        _cleanup_list;
 398 
 399   // Concurrent marking support structures
 400   CMBitMap                _markBitMap1;
 401   CMBitMap                _markBitMap2;
 402   CMBitMapRO*             _prevMarkBitMap; // Completed mark bitmap
 403   CMBitMap*               _nextMarkBitMap; // Under-construction mark bitmap
 404 
 405   BitMap                  _region_bm;
 406   BitMap                  _card_bm;
 407 
 408   // Heap bounds
 409   HeapWord*               _heap_start;
 410   HeapWord*               _heap_end;
 411 
 412   // Root region tracking and claiming
 413   CMRootRegions           _root_regions;
 414 
 415   // For gray objects
 416   CMMarkStack             _markStack; // Grey objects behind global finger
 417   HeapWord* volatile      _finger;  // The global finger, region aligned,
 418                                     // always points to the end of the
 419                                     // last claimed region
 420 
 421   // Marking tasks
 422   uint                    _max_worker_id;// Maximum worker id
 423   uint                    _active_tasks; // Task num currently active
 424   CMTask**                _tasks;        // Task queue array (max_worker_id len)
 425   CMTaskQueueSet*         _task_queues;  // Task queue set
 426   ParallelTaskTerminator  _terminator;   // For termination
 427 
 428   // Two sync barriers that are used to synchronize tasks when an
 429   // overflow occurs. The algorithm is the following. All tasks enter
 430   // the first one to ensure that they have all stopped manipulating
 431   // the global data structures. After they exit it, they re-initialize
 432   // their data structures and task 0 re-initializes the global data
 433   // structures. Then, they enter the second sync barrier. This
 434   // ensure, that no task starts doing work before all data
 435   // structures (local and global) have been re-initialized. When they
 436   // exit it, they are free to start working again.
 437   WorkGangBarrierSync     _first_overflow_barrier_sync;
 438   WorkGangBarrierSync     _second_overflow_barrier_sync;
 439 
 440   // This is set by any task, when an overflow on the global data
 441   // structures is detected
 442   volatile bool           _has_overflown;
 443   // True: marking is concurrent, false: we're in remark
 444   volatile bool           _concurrent;
 445   // Set at the end of a Full GC so that marking aborts
 446   volatile bool           _has_aborted;
 447 
 448   // Used when remark aborts due to an overflow to indicate that
 449   // another concurrent marking phase should start
 450   volatile bool           _restart_for_overflow;
 451 
 452   // This is true from the very start of concurrent marking until the
 453   // point when all the tasks complete their work. It is really used
 454   // to determine the points between the end of concurrent marking and
 455   // time of remark.
 456   volatile bool           _concurrent_marking_in_progress;
 457 
 458   // Verbose level
 459   CMVerboseLevel          _verbose_level;
 460 
 461   // All of these times are in ms
 462   NumberSeq _init_times;
 463   NumberSeq _remark_times;
 464   NumberSeq   _remark_mark_times;
 465   NumberSeq   _remark_weak_ref_times;
 466   NumberSeq _cleanup_times;
 467   double    _total_counting_time;
 468   double    _total_rs_scrub_time;
 469 
 470   double*   _accum_task_vtime;   // Accumulated task vtime
 471 
 472   FlexibleWorkGang* _parallel_workers;
 473 
 474   ForceOverflowSettings _force_overflow_conc;
 475   ForceOverflowSettings _force_overflow_stw;
 476 
 477   void weakRefsWork(bool clear_all_soft_refs);
 478 
 479   void swapMarkBitMaps();
 480 
 481   // It resets the global marking data structures, as well as the
 482   // task local ones; should be called during initial mark.
 483   void reset();
 484 
 485   // Resets all the marking data structures. Called when we have to restart
 486   // marking or when marking completes (via set_non_marking_state below).
 487   void reset_marking_state(bool clear_overflow = true);
 488 
 489   // We do this after we're done with marking so that the marking data
 490   // structures are initialized to a sensible and predictable state.
 491   void set_non_marking_state();
 492 
 493   // Called to indicate how many threads are currently active.
 494   void set_concurrency(uint active_tasks);
 495 
 496   // It should be called to indicate which phase we're in (concurrent
 497   // mark or remark) and how many threads are currently active.
 498   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 499 
 500   // Prints all gathered CM-related statistics
 501   void print_stats();
 502 
 503   bool cleanup_list_is_empty() {
 504     return _cleanup_list.is_empty();
 505   }
 506 
 507   // Accessor methods
 508   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 509   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 510   double sleep_factor()                     { return _sleep_factor; }
 511   double marking_task_overhead()            { return _marking_task_overhead;}
 512   double cleanup_sleep_factor()             { return _cleanup_sleep_factor; }
 513   double cleanup_task_overhead()            { return _cleanup_task_overhead;}
 514 
 515   bool use_parallel_marking_threads() const {
 516     assert(parallel_marking_threads() <=
 517            max_parallel_marking_threads(), "sanity");
 518     assert((_parallel_workers == NULL && parallel_marking_threads() == 0) ||
 519            parallel_marking_threads() > 0,
 520            "parallel workers not set up correctly");
 521     return _parallel_workers != NULL;
 522   }
 523 
 524   HeapWord*               finger()          { return _finger;   }
 525   bool                    concurrent()      { return _concurrent; }
 526   uint                    active_tasks()    { return _active_tasks; }
 527   ParallelTaskTerminator* terminator()      { return &_terminator; }
 528 
 529   // It claims the next available region to be scanned by a marking
 530   // task/thread. It might return NULL if the next region is empty or
 531   // we have run out of regions. In the latter case, out_of_regions()
 532   // determines whether we've really run out of regions or the task
 533   // should call claim_region() again. This might seem a bit
 534   // awkward. Originally, the code was written so that claim_region()
 535   // either successfully returned with a non-empty region or there
 536   // were no more regions to be claimed. The problem with this was
 537   // that, in certain circumstances, it iterated over large chunks of
 538   // the heap finding only empty regions and, while it was working, it
 539   // was preventing the calling task to call its regular clock
 540   // method. So, this way, each task will spend very little time in
 541   // claim_region() and is allowed to call the regular clock method
 542   // frequently.
 543   HeapRegion* claim_region(uint worker_id);
 544 
 545   // It determines whether we've run out of regions to scan. Note that
 546   // the finger can point past the heap end in case the heap was expanded
 547   // to satisfy an allocation without doing a GC. This is fine, because all
 548   // objects in those regions will be considered live anyway because of
 549   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 550   bool        out_of_regions() { return _finger >= _heap_end; }
 551 
 552   // Returns the task with the given id
 553   CMTask* task(int id) {
 554     assert(0 <= id && id < (int) _active_tasks,
 555            "task id not within active bounds");
 556     return _tasks[id];
 557   }
 558 
 559   // Returns the task queue with the given id
 560   CMTaskQueue* task_queue(int id) {
 561     assert(0 <= id && id < (int) _active_tasks,
 562            "task queue id not within active bounds");
 563     return (CMTaskQueue*) _task_queues->queue(id);
 564   }
 565 
 566   // Returns the task queue set
 567   CMTaskQueueSet* task_queues()  { return _task_queues; }
 568 
 569   // Access / manipulation of the overflow flag which is set to
 570   // indicate that the global stack has overflown
 571   bool has_overflown()           { return _has_overflown; }
 572   void set_has_overflown()       { _has_overflown = true; }
 573   void clear_has_overflown()     { _has_overflown = false; }
 574   bool restart_for_overflow()    { return _restart_for_overflow; }
 575 
 576   // Methods to enter the two overflow sync barriers
 577   void enter_first_sync_barrier(uint worker_id);
 578   void enter_second_sync_barrier(uint worker_id);
 579 
 580   ForceOverflowSettings* force_overflow_conc() {
 581     return &_force_overflow_conc;
 582   }
 583 
 584   ForceOverflowSettings* force_overflow_stw() {
 585     return &_force_overflow_stw;
 586   }
 587 
 588   ForceOverflowSettings* force_overflow() {
 589     if (concurrent()) {
 590       return force_overflow_conc();
 591     } else {
 592       return force_overflow_stw();
 593     }
 594   }
 595 
 596   // Live Data Counting data structures...
 597   // These data structures are initialized at the start of
 598   // marking. They are written to while marking is active.
 599   // They are aggregated during remark; the aggregated values
 600   // are then used to populate the _region_bm, _card_bm, and
 601   // the total live bytes, which are then subsequently updated
 602   // during cleanup.
 603 
 604   // An array of bitmaps (one bit map per task). Each bitmap
 605   // is used to record the cards spanned by the live objects
 606   // marked by that task/worker.
 607   BitMap*  _count_card_bitmaps;
 608 
 609   // Used to record the number of marked live bytes
 610   // (for each region, by worker thread).
 611   size_t** _count_marked_bytes;
 612 
 613   // Card index of the bottom of the G1 heap. Used for biasing indices into
 614   // the card bitmaps.
 615   intptr_t _heap_bottom_card_num;
 616 
 617   // Set to true when initialization is complete
 618   bool _completed_initialization;
 619 
 620 public:
 621   // Manipulation of the global mark stack.
 622   // Notice that the first mark_stack_push is CAS-based, whereas the
 623   // two below are Mutex-based. This is OK since the first one is only
 624   // called during evacuation pauses and doesn't compete with the
 625   // other two (which are called by the marking tasks during
 626   // concurrent marking or remark).
 627   bool mark_stack_push(oop p) {
 628     _markStack.par_push(p);
 629     if (_markStack.overflow()) {
 630       set_has_overflown();
 631       return false;
 632     }
 633     return true;
 634   }
 635   bool mark_stack_push(oop* arr, int n) {
 636     _markStack.par_push_arr(arr, n);
 637     if (_markStack.overflow()) {
 638       set_has_overflown();
 639       return false;
 640     }
 641     return true;
 642   }
 643   void mark_stack_pop(oop* arr, int max, int* n) {
 644     _markStack.par_pop_arr(arr, max, n);
 645   }
 646   size_t mark_stack_size()                { return _markStack.size(); }
 647   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 648   bool mark_stack_overflow()              { return _markStack.overflow(); }
 649   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 650 
 651   CMRootRegions* root_regions() { return &_root_regions; }
 652 
 653   bool concurrent_marking_in_progress() {
 654     return _concurrent_marking_in_progress;
 655   }
 656   void set_concurrent_marking_in_progress() {
 657     _concurrent_marking_in_progress = true;
 658   }
 659   void clear_concurrent_marking_in_progress() {
 660     _concurrent_marking_in_progress = false;
 661   }
 662 
 663   void update_accum_task_vtime(int i, double vtime) {
 664     _accum_task_vtime[i] += vtime;
 665   }
 666 
 667   double all_task_accum_vtime() {
 668     double ret = 0.0;
 669     for (uint i = 0; i < _max_worker_id; ++i)
 670       ret += _accum_task_vtime[i];
 671     return ret;
 672   }
 673 
 674   // Attempts to steal an object from the task queues of other tasks
 675   bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
 676     return _task_queues->steal(worker_id, hash_seed, obj);
 677   }
 678 
 679   ConcurrentMark(G1CollectedHeap* g1h, ReservedSpace heap_rs);
 680   ~ConcurrentMark();
 681 
 682   ConcurrentMarkThread* cmThread() { return _cmThread; }
 683 
 684   CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 685   CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 686 
 687   // Returns the number of GC threads to be used in a concurrent
 688   // phase based on the number of GC threads being used in a STW
 689   // phase.
 690   uint scale_parallel_threads(uint n_par_threads);
 691 
 692   // Calculates the number of GC threads to be used in a concurrent phase.
 693   uint calc_parallel_marking_threads();
 694 
 695   // The following three are interaction between CM and
 696   // G1CollectedHeap
 697 
 698   // This notifies CM that a root during initial-mark needs to be
 699   // grayed. It is MT-safe. word_size is the size of the object in
 700   // words. It is passed explicitly as sometimes we cannot calculate
 701   // it from the given object because it might be in an inconsistent
 702   // state (e.g., in to-space and being copied). So the caller is
 703   // responsible for dealing with this issue (e.g., get the size from
 704   // the from-space image when the to-space image might be
 705   // inconsistent) and always passing the size. hr is the region that
 706   // contains the object and it's passed optionally from callers who
 707   // might already have it (no point in recalculating it).
 708   inline void grayRoot(oop obj, size_t word_size,
 709                        uint worker_id, HeapRegion* hr = NULL);
 710 
 711   // It iterates over the heap and for each object it comes across it
 712   // will dump the contents of its reference fields, as well as
 713   // liveness information for the object and its referents. The dump
 714   // will be written to a file with the following name:
 715   // G1PrintReachableBaseFile + "." + str.
 716   // vo decides whether the prev (vo == UsePrevMarking), the next
 717   // (vo == UseNextMarking) marking information, or the mark word
 718   // (vo == UseMarkWord) will be used to determine the liveness of
 719   // each object / referent.
 720   // If all is true, all objects in the heap will be dumped, otherwise
 721   // only the live ones. In the dump the following symbols / breviations
 722   // are used:
 723   //   M : an explicitly live object (its bitmap bit is set)
 724   //   > : an implicitly live object (over tams)
 725   //   O : an object outside the G1 heap (typically: in the perm gen)
 726   //   NOT : a reference field whose referent is not live
 727   //   AND MARKED : indicates that an object is both explicitly and
 728   //   implicitly live (it should be one or the other, not both)
 729   void print_reachable(const char* str,
 730                        VerifyOption vo, bool all) PRODUCT_RETURN;
 731 
 732   // Clear the next marking bitmap (will be called concurrently).
 733   void clearNextBitmap();
 734 
 735   // These two do the work that needs to be done before and after the
 736   // initial root checkpoint. Since this checkpoint can be done at two
 737   // different points (i.e. an explicit pause or piggy-backed on a
 738   // young collection), then it's nice to be able to easily share the
 739   // pre/post code. It might be the case that we can put everything in
 740   // the post method. TP
 741   void checkpointRootsInitialPre();
 742   void checkpointRootsInitialPost();
 743 
 744   // Scan all the root regions and mark everything reachable from
 745   // them.
 746   void scanRootRegions();
 747 
 748   // Scan a single root region and mark everything reachable from it.
 749   void scanRootRegion(HeapRegion* hr, uint worker_id);
 750 
 751   // Do concurrent phase of marking, to a tentative transitive closure.
 752   void markFromRoots();
 753 
 754   void checkpointRootsFinal(bool clear_all_soft_refs);
 755   void checkpointRootsFinalWork();
 756   void cleanup();
 757   void completeCleanup();
 758 
 759   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 760   // this carefully!
 761   inline void markPrev(oop p);
 762 
 763   // Clears marks for all objects in the given range, for the prev,
 764   // next, or both bitmaps.  NB: the previous bitmap is usually
 765   // read-only, so use this carefully!
 766   void clearRangePrevBitmap(MemRegion mr);
 767   void clearRangeNextBitmap(MemRegion mr);
 768   void clearRangeBothBitmaps(MemRegion mr);
 769 
 770   // Notify data structures that a GC has started.
 771   void note_start_of_gc() {
 772     _markStack.note_start_of_gc();
 773   }
 774 
 775   // Notify data structures that a GC is finished.
 776   void note_end_of_gc() {
 777     _markStack.note_end_of_gc();
 778   }
 779 
 780   // Verify that there are no CSet oops on the stacks (taskqueues /
 781   // global mark stack), enqueued SATB buffers, per-thread SATB
 782   // buffers, and fingers (global / per-task). The boolean parameters
 783   // decide which of the above data structures to verify. If marking
 784   // is not in progress, it's a no-op.
 785   void verify_no_cset_oops(bool verify_stacks,
 786                            bool verify_enqueued_buffers,
 787                            bool verify_thread_buffers,
 788                            bool verify_fingers) PRODUCT_RETURN;
 789 
 790   // It is called at the end of an evacuation pause during marking so
 791   // that CM is notified of where the new end of the heap is. It
 792   // doesn't do anything if concurrent_marking_in_progress() is false,
 793   // unless the force parameter is true.
 794   void update_g1_committed(bool force = false);
 795 
 796   bool isMarked(oop p) const {
 797     assert(p != NULL && p->is_oop(), "expected an oop");
 798     HeapWord* addr = (HeapWord*)p;
 799     assert(addr >= _nextMarkBitMap->startWord() ||
 800            addr < _nextMarkBitMap->endWord(), "in a region");
 801 
 802     return _nextMarkBitMap->isMarked(addr);
 803   }
 804 
 805   inline bool not_yet_marked(oop p) const;
 806 
 807   // XXX Debug code
 808   bool containing_card_is_marked(void* p);
 809   bool containing_cards_are_marked(void* start, void* last);
 810 
 811   bool isPrevMarked(oop p) const {
 812     assert(p != NULL && p->is_oop(), "expected an oop");
 813     HeapWord* addr = (HeapWord*)p;
 814     assert(addr >= _prevMarkBitMap->startWord() ||
 815            addr < _prevMarkBitMap->endWord(), "in a region");
 816 
 817     return _prevMarkBitMap->isMarked(addr);
 818   }
 819 
 820   inline bool do_yield_check(uint worker_i = 0);
 821 
 822   // Called to abort the marking cycle after a Full GC takes place.
 823   void abort();
 824 
 825   bool has_aborted()       { return _has_aborted; }
 826   void set_has_aborted()   { _has_aborted = true; }
 827   void clear_has_aborted() { _has_aborted = false; }
 828 
 829   // This prints the global/local fingers. It is used for debugging.
 830   NOT_PRODUCT(void print_finger();)
 831 
 832   void print_summary_info();
 833 
 834   void print_worker_threads_on(outputStream* st) const;
 835 
 836   void print_on_error(outputStream* st) const;
 837 
 838   // The following indicate whether a given verbose level has been
 839   // set. Notice that anything above stats is conditional to
 840   // _MARKING_VERBOSE_ having been set to 1
 841   bool verbose_stats() {
 842     return _verbose_level >= stats_verbose;
 843   }
 844   bool verbose_low() {
 845     return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
 846   }
 847   bool verbose_medium() {
 848     return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
 849   }
 850   bool verbose_high() {
 851     return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
 852   }
 853 
 854   // Liveness counting
 855 
 856   // Utility routine to set an exclusive range of cards on the given
 857   // card liveness bitmap
 858   inline void set_card_bitmap_range(BitMap* card_bm,
 859                                     BitMap::idx_t start_idx,
 860                                     BitMap::idx_t end_idx,
 861                                     bool is_par);
 862 
 863   // Returns the card number of the bottom of the G1 heap.
 864   // Used in biasing indices into accounting card bitmaps.
 865   intptr_t heap_bottom_card_num() const {
 866     return _heap_bottom_card_num;
 867   }
 868 
 869   // Returns the card bitmap for a given task or worker id.
 870   BitMap* count_card_bitmap_for(uint worker_id) {
 871     assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
 872     assert(_count_card_bitmaps != NULL, "uninitialized");
 873     BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
 874     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 875     return task_card_bm;
 876   }
 877 
 878   // Returns the array containing the marked bytes for each region,
 879   // for the given worker or task id.
 880   size_t* count_marked_bytes_array_for(uint worker_id) {
 881     assert(0 <= worker_id && worker_id < _max_worker_id, "oob");
 882     assert(_count_marked_bytes != NULL, "uninitialized");
 883     size_t* marked_bytes_array = _count_marked_bytes[worker_id];
 884     assert(marked_bytes_array != NULL, "uninitialized");
 885     return marked_bytes_array;
 886   }
 887 
 888   // Returns the index in the liveness accounting card table bitmap
 889   // for the given address
 890   inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
 891 
 892   // Counts the size of the given memory region in the the given
 893   // marked_bytes array slot for the given HeapRegion.
 894   // Sets the bits in the given card bitmap that are associated with the
 895   // cards that are spanned by the memory region.
 896   inline void count_region(MemRegion mr, HeapRegion* hr,
 897                            size_t* marked_bytes_array,
 898                            BitMap* task_card_bm);
 899 
 900   // Counts the given memory region in the task/worker counting
 901   // data structures for the given worker id.
 902   inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
 903 
 904   // Counts the given memory region in the task/worker counting
 905   // data structures for the given worker id.
 906   inline void count_region(MemRegion mr, uint worker_id);
 907 
 908   // Counts the given object in the given task/worker counting
 909   // data structures.
 910   inline void count_object(oop obj, HeapRegion* hr,
 911                            size_t* marked_bytes_array,
 912                            BitMap* task_card_bm);
 913 
 914   // Counts the given object in the task/worker counting data
 915   // structures for the given worker id.
 916   inline void count_object(oop obj, HeapRegion* hr, uint worker_id);
 917 
 918   // Attempts to mark the given object and, if successful, counts
 919   // the object in the given task/worker counting structures.
 920   inline bool par_mark_and_count(oop obj, HeapRegion* hr,
 921                                  size_t* marked_bytes_array,
 922                                  BitMap* task_card_bm);
 923 
 924   // Attempts to mark the given object and, if successful, counts
 925   // the object in the task/worker counting structures for the
 926   // given worker id.
 927   inline bool par_mark_and_count(oop obj, size_t word_size,
 928                                  HeapRegion* hr, uint worker_id);
 929 
 930   // Attempts to mark the given object and, if successful, counts
 931   // the object in the task/worker counting structures for the
 932   // given worker id.
 933   inline bool par_mark_and_count(oop obj, HeapRegion* hr, uint worker_id);
 934 
 935   // Similar to the above routine but we don't know the heap region that
 936   // contains the object to be marked/counted, which this routine looks up.
 937   inline bool par_mark_and_count(oop obj, uint worker_id);
 938 
 939   // Similar to the above routine but there are times when we cannot
 940   // safely calculate the size of obj due to races and we, therefore,
 941   // pass the size in as a parameter. It is the caller's responsibility
 942   // to ensure that the size passed in for obj is valid.
 943   inline bool par_mark_and_count(oop obj, size_t word_size, uint worker_id);
 944 
 945   // Unconditionally mark the given object, and unconditionally count
 946   // the object in the counting structures for worker id 0.
 947   // Should *not* be called from parallel code.
 948   inline bool mark_and_count(oop obj, HeapRegion* hr);
 949 
 950   // Similar to the above routine but we don't know the heap region that
 951   // contains the object to be marked/counted, which this routine looks up.
 952   // Should *not* be called from parallel code.
 953   inline bool mark_and_count(oop obj);
 954 
 955   // Returns true if initialization was successfully completed.
 956   bool completed_initialization() const {
 957     return _completed_initialization;
 958   }
 959 
 960 protected:
 961   // Clear all the per-task bitmaps and arrays used to store the
 962   // counting data.
 963   void clear_all_count_data();
 964 
 965   // Aggregates the counting data for each worker/task
 966   // that was constructed while marking. Also sets
 967   // the amount of marked bytes for each region and
 968   // the top at concurrent mark count.
 969   void aggregate_count_data();
 970 
 971   // Verification routine
 972   void verify_count_data();
 973 };
 974 
 975 // A class representing a marking task.
 976 class CMTask : public TerminatorTerminator {
 977 private:
 978   enum PrivateConstants {
 979     // the regular clock call is called once the scanned words reaches
 980     // this limit
 981     words_scanned_period          = 12*1024,
 982     // the regular clock call is called once the number of visited
 983     // references reaches this limit
 984     refs_reached_period           = 384,
 985     // initial value for the hash seed, used in the work stealing code
 986     init_hash_seed                = 17,
 987     // how many entries will be transferred between global stack and
 988     // local queues
 989     global_stack_transfer_size    = 16
 990   };
 991 
 992   uint                        _worker_id;
 993   G1CollectedHeap*            _g1h;
 994   ConcurrentMark*             _cm;
 995   CMBitMap*                   _nextMarkBitMap;
 996   // the task queue of this task
 997   CMTaskQueue*                _task_queue;
 998 private:
 999   // the task queue set---needed for stealing
1000   CMTaskQueueSet*             _task_queues;
1001   // indicates whether the task has been claimed---this is only  for
1002   // debugging purposes
1003   bool                        _claimed;
1004 
1005   // number of calls to this task
1006   int                         _calls;
1007 
1008   // when the virtual timer reaches this time, the marking step should
1009   // exit
1010   double                      _time_target_ms;
1011   // the start time of the current marking step
1012   double                      _start_time_ms;
1013 
1014   // the oop closure used for iterations over oops
1015   G1CMOopClosure*             _cm_oop_closure;
1016 
1017   // the region this task is scanning, NULL if we're not scanning any
1018   HeapRegion*                 _curr_region;
1019   // the local finger of this task, NULL if we're not scanning a region
1020   HeapWord*                   _finger;
1021   // limit of the region this task is scanning, NULL if we're not scanning one
1022   HeapWord*                   _region_limit;
1023 
1024   // the number of words this task has scanned
1025   size_t                      _words_scanned;
1026   // When _words_scanned reaches this limit, the regular clock is
1027   // called. Notice that this might be decreased under certain
1028   // circumstances (i.e. when we believe that we did an expensive
1029   // operation).
1030   size_t                      _words_scanned_limit;
1031   // the initial value of _words_scanned_limit (i.e. what it was
1032   // before it was decreased).
1033   size_t                      _real_words_scanned_limit;
1034 
1035   // the number of references this task has visited
1036   size_t                      _refs_reached;
1037   // When _refs_reached reaches this limit, the regular clock is
1038   // called. Notice this this might be decreased under certain
1039   // circumstances (i.e. when we believe that we did an expensive
1040   // operation).
1041   size_t                      _refs_reached_limit;
1042   // the initial value of _refs_reached_limit (i.e. what it was before
1043   // it was decreased).
1044   size_t                      _real_refs_reached_limit;
1045 
1046   // used by the work stealing stuff
1047   int                         _hash_seed;
1048   // if this is true, then the task has aborted for some reason
1049   bool                        _has_aborted;
1050   // set when the task aborts because it has met its time quota
1051   bool                        _has_timed_out;
1052   // true when we're draining SATB buffers; this avoids the task
1053   // aborting due to SATB buffers being available (as we're already
1054   // dealing with them)
1055   bool                        _draining_satb_buffers;
1056 
1057   // number sequence of past step times
1058   NumberSeq                   _step_times_ms;
1059   // elapsed time of this task
1060   double                      _elapsed_time_ms;
1061   // termination time of this task
1062   double                      _termination_time_ms;
1063   // when this task got into the termination protocol
1064   double                      _termination_start_time_ms;
1065 
1066   // true when the task is during a concurrent phase, false when it is
1067   // in the remark phase (so, in the latter case, we do not have to
1068   // check all the things that we have to check during the concurrent
1069   // phase, i.e. SATB buffer availability...)
1070   bool                        _concurrent;
1071 
1072   TruncatedSeq                _marking_step_diffs_ms;
1073 
1074   // Counting data structures. Embedding the task's marked_bytes_array
1075   // and card bitmap into the actual task saves having to go through
1076   // the ConcurrentMark object.
1077   size_t*                     _marked_bytes_array;
1078   BitMap*                     _card_bm;
1079 
1080   // LOTS of statistics related with this task
1081 #if _MARKING_STATS_
1082   NumberSeq                   _all_clock_intervals_ms;
1083   double                      _interval_start_time_ms;
1084 
1085   int                         _aborted;
1086   int                         _aborted_overflow;
1087   int                         _aborted_cm_aborted;
1088   int                         _aborted_yield;
1089   int                         _aborted_timed_out;
1090   int                         _aborted_satb;
1091   int                         _aborted_termination;
1092 
1093   int                         _steal_attempts;
1094   int                         _steals;
1095 
1096   int                         _clock_due_to_marking;
1097   int                         _clock_due_to_scanning;
1098 
1099   int                         _local_pushes;
1100   int                         _local_pops;
1101   int                         _local_max_size;
1102   int                         _objs_scanned;
1103 
1104   int                         _global_pushes;
1105   int                         _global_pops;
1106   int                         _global_max_size;
1107 
1108   int                         _global_transfers_to;
1109   int                         _global_transfers_from;
1110 
1111   int                         _regions_claimed;
1112   int                         _objs_found_on_bitmap;
1113 
1114   int                         _satb_buffers_processed;
1115 #endif // _MARKING_STATS_
1116 
1117   // it updates the local fields after this task has claimed
1118   // a new region to scan
1119   void setup_for_region(HeapRegion* hr);
1120   // it brings up-to-date the limit of the region
1121   void update_region_limit();
1122 
1123   // called when either the words scanned or the refs visited limit
1124   // has been reached
1125   void reached_limit();
1126   // recalculates the words scanned and refs visited limits
1127   void recalculate_limits();
1128   // decreases the words scanned and refs visited limits when we reach
1129   // an expensive operation
1130   void decrease_limits();
1131   // it checks whether the words scanned or refs visited reached their
1132   // respective limit and calls reached_limit() if they have
1133   void check_limits() {
1134     if (_words_scanned >= _words_scanned_limit ||
1135         _refs_reached >= _refs_reached_limit) {
1136       reached_limit();
1137     }
1138   }
1139   // this is supposed to be called regularly during a marking step as
1140   // it checks a bunch of conditions that might cause the marking step
1141   // to abort
1142   void regular_clock_call();
1143   bool concurrent() { return _concurrent; }
1144 
1145 public:
1146   // It resets the task; it should be called right at the beginning of
1147   // a marking phase.
1148   void reset(CMBitMap* _nextMarkBitMap);
1149   // it clears all the fields that correspond to a claimed region.
1150   void clear_region_fields();
1151 
1152   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1153 
1154   // The main method of this class which performs a marking step
1155   // trying not to exceed the given duration. However, it might exit
1156   // prematurely, according to some conditions (i.e. SATB buffers are
1157   // available for processing).
1158   void do_marking_step(double target_ms,
1159                        bool do_termination,
1160                        bool is_serial);
1161 
1162   // These two calls start and stop the timer
1163   void record_start_time() {
1164     _elapsed_time_ms = os::elapsedTime() * 1000.0;
1165   }
1166   void record_end_time() {
1167     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1168   }
1169 
1170   // returns the worker ID associated with this task.
1171   uint worker_id() { return _worker_id; }
1172 
1173   // From TerminatorTerminator. It determines whether this task should
1174   // exit the termination protocol after it's entered it.
1175   virtual bool should_exit_termination();
1176 
1177   // Resets the local region fields after a task has finished scanning a
1178   // region; or when they have become stale as a result of the region
1179   // being evacuated.
1180   void giveup_current_region();
1181 
1182   HeapWord* finger()            { return _finger; }
1183 
1184   bool has_aborted()            { return _has_aborted; }
1185   void set_has_aborted()        { _has_aborted = true; }
1186   void clear_has_aborted()      { _has_aborted = false; }
1187   bool has_timed_out()          { return _has_timed_out; }
1188   bool claimed()                { return _claimed; }
1189 
1190   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1191 
1192   // It grays the object by marking it and, if necessary, pushing it
1193   // on the local queue
1194   inline void deal_with_reference(oop obj);
1195 
1196   // It scans an object and visits its children.
1197   void scan_object(oop obj);
1198 
1199   // It pushes an object on the local queue.
1200   inline void push(oop obj);
1201 
1202   // These two move entries to/from the global stack.
1203   void move_entries_to_global_stack();
1204   void get_entries_from_global_stack();
1205 
1206   // It pops and scans objects from the local queue. If partially is
1207   // true, then it stops when the queue size is of a given limit. If
1208   // partially is false, then it stops when the queue is empty.
1209   void drain_local_queue(bool partially);
1210   // It moves entries from the global stack to the local queue and
1211   // drains the local queue. If partially is true, then it stops when
1212   // both the global stack and the local queue reach a given size. If
1213   // partially if false, it tries to empty them totally.
1214   void drain_global_stack(bool partially);
1215   // It keeps picking SATB buffers and processing them until no SATB
1216   // buffers are available.
1217   void drain_satb_buffers();
1218 
1219   // moves the local finger to a new location
1220   inline void move_finger_to(HeapWord* new_finger) {
1221     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1222     _finger = new_finger;
1223   }
1224 
1225   CMTask(uint worker_id, ConcurrentMark *cm,
1226          size_t* marked_bytes, BitMap* card_bm,
1227          CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1228 
1229   // it prints statistics associated with this task
1230   void print_stats();
1231 
1232 #if _MARKING_STATS_
1233   void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1234 #endif // _MARKING_STATS_
1235 };
1236 
1237 // Class that's used to to print out per-region liveness
1238 // information. It's currently used at the end of marking and also
1239 // after we sort the old regions at the end of the cleanup operation.
1240 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1241 private:
1242   outputStream* _out;
1243 
1244   // Accumulators for these values.
1245   size_t _total_used_bytes;
1246   size_t _total_capacity_bytes;
1247   size_t _total_prev_live_bytes;
1248   size_t _total_next_live_bytes;
1249 
1250   // These are set up when we come across a "stars humongous" region
1251   // (as this is where most of this information is stored, not in the
1252   // subsequent "continues humongous" regions). After that, for every
1253   // region in a given humongous region series we deduce the right
1254   // values for it by simply subtracting the appropriate amount from
1255   // these fields. All these values should reach 0 after we've visited
1256   // the last region in the series.
1257   size_t _hum_used_bytes;
1258   size_t _hum_capacity_bytes;
1259   size_t _hum_prev_live_bytes;
1260   size_t _hum_next_live_bytes;
1261 
1262   // Accumulator for the remembered set size
1263   size_t _total_remset_bytes;
1264 
1265   // Accumulator for strong code roots memory size
1266   size_t _total_strong_code_roots_bytes;
1267 
1268   static double perc(size_t val, size_t total) {
1269     if (total == 0) {
1270       return 0.0;
1271     } else {
1272       return 100.0 * ((double) val / (double) total);
1273     }
1274   }
1275 
1276   static double bytes_to_mb(size_t val) {
1277     return (double) val / (double) M;
1278   }
1279 
1280   // See the .cpp file.
1281   size_t get_hum_bytes(size_t* hum_bytes);
1282   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1283                      size_t* prev_live_bytes, size_t* next_live_bytes);
1284 
1285 public:
1286   // The header and footer are printed in the constructor and
1287   // destructor respectively.
1288   G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1289   virtual bool doHeapRegion(HeapRegion* r);
1290   ~G1PrintRegionLivenessInfoClosure();
1291 };
1292 
1293 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP