1 /*
   2  * Copyright (c) 2001, 2015, 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_G1_CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_CONCURRENTMARK_HPP
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/g1/g1RegionToSpaceMapper.hpp"
  30 #include "gc/g1/heapRegionSet.hpp"
  31 #include "gc/shared/taskqueue.hpp"
  32 
  33 class G1CollectedHeap;
  34 class CMBitMap;
  35 class CMTask;
  36 class ConcurrentMark;
  37 typedef GenericTaskQueue<oop, mtGC>            CMTaskQueue;
  38 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
  39 
  40 // Closure used by CM during concurrent reference discovery
  41 // and reference processing (during remarking) to determine
  42 // if a particular object is alive. It is primarily used
  43 // to determine if referents of discovered reference objects
  44 // are alive. An instance is also embedded into the
  45 // reference processor as the _is_alive_non_header field
  46 class G1CMIsAliveClosure: public BoolObjectClosure {
  47   G1CollectedHeap* _g1;
  48  public:
  49   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
  50 
  51   bool do_object_b(oop obj);
  52 };
  53 
  54 // A generic CM bit map.  This is essentially a wrapper around the BitMap
  55 // class, with one bit per (1<<_shifter) HeapWords.
  56 
  57 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
  58  protected:
  59   HeapWord* _bmStartWord;      // base address of range covered by map
  60   size_t    _bmWordSize;       // map size (in #HeapWords covered)
  61   const int _shifter;          // map to char or bit
  62   BitMap    _bm;               // the bit map itself
  63 
  64  public:
  65   // constructor
  66   CMBitMapRO(int shifter);
  67 
  68   enum { do_yield = true };
  69 
  70   // inquiries
  71   HeapWord* startWord()   const { return _bmStartWord; }
  72   size_t    sizeInWords() const { return _bmWordSize;  }
  73   // the following is one past the last word in space
  74   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
  75 
  76   // read marks
  77 
  78   bool isMarked(HeapWord* addr) const {
  79     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
  80            "outside underlying space?");
  81     return _bm.at(heapWordToOffset(addr));
  82   }
  83 
  84   // iteration
  85   inline bool iterate(BitMapClosure* cl, MemRegion mr);
  86   inline bool iterate(BitMapClosure* cl);
  87 
  88   // Return the address corresponding to the next marked bit at or after
  89   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  90   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  91   HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
  92                                      const HeapWord* limit = NULL) const;
  93   // Return the address corresponding to the next unmarked bit at or after
  94   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  95   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  96   HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
  97                                        const HeapWord* limit = NULL) const;
  98 
  99   // conversion utilities
 100   HeapWord* offsetToHeapWord(size_t offset) const {
 101     return _bmStartWord + (offset << _shifter);
 102   }
 103   size_t heapWordToOffset(const HeapWord* addr) const {
 104     return pointer_delta(addr, _bmStartWord) >> _shifter;
 105   }
 106   int heapWordDiffToOffsetDiff(size_t diff) const;
 107 
 108   // The argument addr should be the start address of a valid object
 109   HeapWord* nextObject(HeapWord* addr) {
 110     oop obj = (oop) addr;
 111     HeapWord* res =  addr + obj->size();
 112     assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
 113     return res;
 114   }
 115 
 116   void print_on_error(outputStream* st, const char* prefix) const;
 117 
 118   // debugging
 119   NOT_PRODUCT(bool covers(MemRegion rs) const;)
 120 };
 121 
 122 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
 123  private:
 124   CMBitMap* _bm;
 125  public:
 126   CMBitMapMappingChangedListener() : _bm(NULL) {}
 127 
 128   void set_bitmap(CMBitMap* bm) { _bm = bm; }
 129 
 130   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 131 };
 132 
 133 class CMBitMap : public CMBitMapRO {
 134  private:
 135   CMBitMapMappingChangedListener _listener;
 136 
 137  public:
 138   static size_t compute_size(size_t heap_size);
 139   // Returns the amount of bytes on the heap between two marks in the bitmap.
 140   static size_t mark_distance();
 141   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 142   // mark bitmap corresponds to. This is the same as the mark distance above.
 143   static size_t heap_map_factor() {
 144     return mark_distance();
 145   }
 146 
 147   CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 148 
 149   // Initializes the underlying BitMap to cover the given area.
 150   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 151 
 152   // Write marks.
 153   inline void mark(HeapWord* addr);
 154   inline void clear(HeapWord* addr);
 155   inline bool parMark(HeapWord* addr);
 156   inline bool parClear(HeapWord* addr);
 157 
 158   void markRange(MemRegion mr);
 159   void clearRange(MemRegion mr);
 160 
 161   // Starting at the bit corresponding to "addr" (inclusive), find the next
 162   // "1" bit, if any.  This bit starts some run of consecutive "1"'s; find
 163   // the end of this run (stopping at "end_addr").  Return the MemRegion
 164   // covering from the start of the region corresponding to the first bit
 165   // of the run to the end of the region corresponding to the last bit of
 166   // the run.  If there is no "1" bit at or after "addr", return an empty
 167   // MemRegion.
 168   MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
 169 
 170   // Clear the whole mark bitmap.
 171   void clearAll();
 172 };
 173 
 174 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
 175 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
 176   VirtualSpace _virtual_space;   // Underlying backing store for actual stack
 177   ConcurrentMark* _cm;
 178   oop* _base;        // bottom of stack
 179   jint _index;       // one more than last occupied index
 180   jint _capacity;    // max #elements
 181   jint _saved_index; // value of _index saved at start of GC
 182 
 183   bool  _overflow;
 184   bool  _should_expand;
 185 
 186  public:
 187   CMMarkStack(ConcurrentMark* cm);
 188   ~CMMarkStack();
 189 
 190   bool allocate(size_t capacity);
 191 
 192   // Pushes the first "n" elements of "ptr_arr" on the stack.
 193   // Locking impl: concurrency is allowed only with
 194   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 195   // locking strategy.
 196   void par_push_arr(oop* ptr_arr, int n);
 197 
 198   // If returns false, the array was empty.  Otherwise, removes up to "max"
 199   // elements from the stack, and transfers them to "ptr_arr" in an
 200   // unspecified order.  The actual number transferred is given in "n" ("n
 201   // == 0" is deliberately redundant with the return value.)  Locking impl:
 202   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 203   // operations, which use the same locking strategy.
 204   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 205 
 206   bool isEmpty()    { return _index == 0; }
 207   int  maxElems()   { return _capacity; }
 208 
 209   bool overflow() { return _overflow; }
 210   void clear_overflow() { _overflow = false; }
 211 
 212   bool should_expand() const { return _should_expand; }
 213   void set_should_expand();
 214 
 215   // Expand the stack, typically in response to an overflow condition
 216   void expand();
 217 
 218   int  size() { return _index; }
 219 
 220   void setEmpty()   { _index = 0; clear_overflow(); }
 221 
 222   // Record the current index.
 223   void note_start_of_gc();
 224 
 225   // Make sure that we have not added any entries to the stack during GC.
 226   void note_end_of_gc();
 227 
 228   // Apply fn to each oop in the mark stack, up to the bound recorded
 229   // via one of the above "note" functions.  The mark stack must not
 230   // be modified while iterating.
 231   template<typename Fn> void iterate(Fn fn);
 232 };
 233 
 234 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
 235 private:
 236 #ifndef PRODUCT
 237   uintx _num_remaining;
 238   bool _force;
 239 #endif // !defined(PRODUCT)
 240 
 241 public:
 242   void init() PRODUCT_RETURN;
 243   void update() PRODUCT_RETURN;
 244   bool should_force() PRODUCT_RETURN_( return false; );
 245 };
 246 
 247 class YoungList;
 248 
 249 // Root Regions are regions that are not empty at the beginning of a
 250 // marking cycle and which we might collect during an evacuation pause
 251 // while the cycle is active. Given that, during evacuation pauses, we
 252 // do not copy objects that are explicitly marked, what we have to do
 253 // for the root regions is to scan them and mark all objects reachable
 254 // from them. According to the SATB assumptions, we only need to visit
 255 // each object once during marking. So, as long as we finish this scan
 256 // before the next evacuation pause, we can copy the objects from the
 257 // root regions without having to mark them or do anything else to them.
 258 //
 259 // Currently, we only support root region scanning once (at the start
 260 // of the marking cycle) and the root regions are all the survivor
 261 // regions populated during the initial-mark pause.
 262 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
 263 private:
 264   YoungList*           _young_list;
 265   ConcurrentMark*      _cm;
 266 
 267   volatile bool        _scan_in_progress;
 268   volatile bool        _should_abort;
 269   HeapRegion* volatile _next_survivor;
 270 
 271 public:
 272   CMRootRegions();
 273   // We actually do most of the initialization in this method.
 274   void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
 275 
 276   // Reset the claiming / scanning of the root regions.
 277   void prepare_for_scan();
 278 
 279   // Forces get_next() to return NULL so that the iteration aborts early.
 280   void abort() { _should_abort = true; }
 281 
 282   // Return true if the CM thread are actively scanning root regions,
 283   // false otherwise.
 284   bool scan_in_progress() { return _scan_in_progress; }
 285 
 286   // Claim the next root region to scan atomically, or return NULL if
 287   // all have been claimed.
 288   HeapRegion* claim_next();
 289 
 290   // Flag that we're done with root region scanning and notify anyone
 291   // who's waiting on it. If aborted is false, assume that all regions
 292   // have been claimed.
 293   void scan_finished();
 294 
 295   // If CM threads are still scanning root regions, wait until they
 296   // are done. Return true if we had to wait, false otherwise.
 297   bool wait_until_scan_finished();
 298 };
 299 
 300 class ConcurrentMarkThread;
 301 
 302 class ConcurrentMark: public CHeapObj<mtGC> {
 303   friend class CMMarkStack;
 304   friend class ConcurrentMarkThread;
 305   friend class CMTask;
 306   friend class CMBitMapClosure;
 307   friend class CMRemarkTask;
 308   friend class CMConcurrentMarkingTask;
 309   friend class G1ParNoteEndTask;
 310   friend class CalcLiveObjectsClosure;
 311   friend class G1CMRefProcTaskProxy;
 312   friend class G1CMRefProcTaskExecutor;
 313   friend class G1CMKeepAliveAndDrainClosure;
 314   friend class G1CMDrainMarkingStackClosure;
 315 
 316 protected:
 317   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 318   G1CollectedHeap*      _g1h;        // The heap
 319   uint                  _parallel_marking_threads; // The number of marking
 320                                                    // threads we're using
 321   uint                  _max_parallel_marking_threads; // Max number of marking
 322                                                        // threads we'll ever use
 323   double                _sleep_factor; // How much we have to sleep, with
 324                                        // respect to the work we just did, to
 325                                        // meet the marking overhead goal
 326   double                _marking_task_overhead; // Marking target overhead for
 327                                                 // a single task
 328 
 329   // Same as the two above, but for the cleanup task
 330   double                _cleanup_sleep_factor;
 331   double                _cleanup_task_overhead;
 332 
 333   FreeRegionList        _cleanup_list;
 334 
 335   // Concurrent marking support structures
 336   CMBitMap                _markBitMap1;
 337   CMBitMap                _markBitMap2;
 338   CMBitMapRO*             _prevMarkBitMap; // Completed mark bitmap
 339   CMBitMap*               _nextMarkBitMap; // Under-construction mark bitmap
 340 
 341   BitMap                  _region_bm;
 342   BitMap                  _card_bm;
 343 
 344   // Heap bounds
 345   HeapWord*               _heap_start;
 346   HeapWord*               _heap_end;
 347 
 348   // Root region tracking and claiming
 349   CMRootRegions           _root_regions;
 350 
 351   // For gray objects
 352   CMMarkStack             _markStack; // Grey objects behind global finger
 353   HeapWord* volatile      _finger;  // The global finger, region aligned,
 354                                     // always points to the end of the
 355                                     // last claimed region
 356 
 357   // Marking tasks
 358   uint                    _max_worker_id;// Maximum worker id
 359   uint                    _active_tasks; // Task num currently active
 360   CMTask**                _tasks;        // Task queue array (max_worker_id len)
 361   CMTaskQueueSet*         _task_queues;  // Task queue set
 362   ParallelTaskTerminator  _terminator;   // For termination
 363 
 364   // Two sync barriers that are used to synchronize tasks when an
 365   // overflow occurs. The algorithm is the following. All tasks enter
 366   // the first one to ensure that they have all stopped manipulating
 367   // the global data structures. After they exit it, they re-initialize
 368   // their data structures and task 0 re-initializes the global data
 369   // structures. Then, they enter the second sync barrier. This
 370   // ensure, that no task starts doing work before all data
 371   // structures (local and global) have been re-initialized. When they
 372   // exit it, they are free to start working again.
 373   WorkGangBarrierSync     _first_overflow_barrier_sync;
 374   WorkGangBarrierSync     _second_overflow_barrier_sync;
 375 
 376   // This is set by any task, when an overflow on the global data
 377   // structures is detected
 378   volatile bool           _has_overflown;
 379   // True: marking is concurrent, false: we're in remark
 380   volatile bool           _concurrent;
 381   // Set at the end of a Full GC so that marking aborts
 382   volatile bool           _has_aborted;
 383 
 384   // Used when remark aborts due to an overflow to indicate that
 385   // another concurrent marking phase should start
 386   volatile bool           _restart_for_overflow;
 387 
 388   // This is true from the very start of concurrent marking until the
 389   // point when all the tasks complete their work. It is really used
 390   // to determine the points between the end of concurrent marking and
 391   // time of remark.
 392   volatile bool           _concurrent_marking_in_progress;
 393 
 394   // All of these times are in ms
 395   NumberSeq _init_times;
 396   NumberSeq _remark_times;
 397   NumberSeq _remark_mark_times;
 398   NumberSeq _remark_weak_ref_times;
 399   NumberSeq _cleanup_times;
 400   double    _total_counting_time;
 401   double    _total_rs_scrub_time;
 402 
 403   double*   _accum_task_vtime;   // Accumulated task vtime
 404 
 405   WorkGang* _parallel_workers;
 406 
 407   ForceOverflowSettings _force_overflow_conc;
 408   ForceOverflowSettings _force_overflow_stw;
 409 
 410   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 411   void weakRefsWork(bool clear_all_soft_refs);
 412 
 413   void swapMarkBitMaps();
 414 
 415   // It resets the global marking data structures, as well as the
 416   // task local ones; should be called during initial mark.
 417   void reset();
 418 
 419   // Resets all the marking data structures. Called when we have to restart
 420   // marking or when marking completes (via set_non_marking_state below).
 421   void reset_marking_state(bool clear_overflow = true);
 422 
 423   // We do this after we're done with marking so that the marking data
 424   // structures are initialized to a sensible and predictable state.
 425   void set_non_marking_state();
 426 
 427   // Called to indicate how many threads are currently active.
 428   void set_concurrency(uint active_tasks);
 429 
 430   // It should be called to indicate which phase we're in (concurrent
 431   // mark or remark) and how many threads are currently active.
 432   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 433 
 434   // Prints all gathered CM-related statistics
 435   void print_stats();
 436 
 437   bool cleanup_list_is_empty() {
 438     return _cleanup_list.is_empty();
 439   }
 440 
 441   // Accessor methods
 442   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 443   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 444   double sleep_factor()                     { return _sleep_factor; }
 445   double marking_task_overhead()            { return _marking_task_overhead;}
 446   double cleanup_sleep_factor()             { return _cleanup_sleep_factor; }
 447   double cleanup_task_overhead()            { return _cleanup_task_overhead;}
 448 
 449   HeapWord*               finger()          { return _finger;   }
 450   bool                    concurrent()      { return _concurrent; }
 451   uint                    active_tasks()    { return _active_tasks; }
 452   ParallelTaskTerminator* terminator()      { return &_terminator; }
 453 
 454   // It claims the next available region to be scanned by a marking
 455   // task/thread. It might return NULL if the next region is empty or
 456   // we have run out of regions. In the latter case, out_of_regions()
 457   // determines whether we've really run out of regions or the task
 458   // should call claim_region() again. This might seem a bit
 459   // awkward. Originally, the code was written so that claim_region()
 460   // either successfully returned with a non-empty region or there
 461   // were no more regions to be claimed. The problem with this was
 462   // that, in certain circumstances, it iterated over large chunks of
 463   // the heap finding only empty regions and, while it was working, it
 464   // was preventing the calling task to call its regular clock
 465   // method. So, this way, each task will spend very little time in
 466   // claim_region() and is allowed to call the regular clock method
 467   // frequently.
 468   HeapRegion* claim_region(uint worker_id);
 469 
 470   // It determines whether we've run out of regions to scan. Note that
 471   // the finger can point past the heap end in case the heap was expanded
 472   // to satisfy an allocation without doing a GC. This is fine, because all
 473   // objects in those regions will be considered live anyway because of
 474   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 475   bool        out_of_regions() { return _finger >= _heap_end; }
 476 
 477   // Returns the task with the given id
 478   CMTask* task(int id) {
 479     assert(0 <= id && id < (int) _active_tasks,
 480            "task id not within active bounds");
 481     return _tasks[id];
 482   }
 483 
 484   // Returns the task queue with the given id
 485   CMTaskQueue* task_queue(int id) {
 486     assert(0 <= id && id < (int) _active_tasks,
 487            "task queue id not within active bounds");
 488     return (CMTaskQueue*) _task_queues->queue(id);
 489   }
 490 
 491   // Returns the task queue set
 492   CMTaskQueueSet* task_queues()  { return _task_queues; }
 493 
 494   // Access / manipulation of the overflow flag which is set to
 495   // indicate that the global stack has overflown
 496   bool has_overflown()           { return _has_overflown; }
 497   void set_has_overflown()       { _has_overflown = true; }
 498   void clear_has_overflown()     { _has_overflown = false; }
 499   bool restart_for_overflow()    { return _restart_for_overflow; }
 500 
 501   // Methods to enter the two overflow sync barriers
 502   void enter_first_sync_barrier(uint worker_id);
 503   void enter_second_sync_barrier(uint worker_id);
 504 
 505   ForceOverflowSettings* force_overflow_conc() {
 506     return &_force_overflow_conc;
 507   }
 508 
 509   ForceOverflowSettings* force_overflow_stw() {
 510     return &_force_overflow_stw;
 511   }
 512 
 513   ForceOverflowSettings* force_overflow() {
 514     if (concurrent()) {
 515       return force_overflow_conc();
 516     } else {
 517       return force_overflow_stw();
 518     }
 519   }
 520 
 521   // Live Data Counting data structures...
 522   // These data structures are initialized at the start of
 523   // marking. They are written to while marking is active.
 524   // They are aggregated during remark; the aggregated values
 525   // are then used to populate the _region_bm, _card_bm, and
 526   // the total live bytes, which are then subsequently updated
 527   // during cleanup.
 528 
 529   // An array of bitmaps (one bit map per task). Each bitmap
 530   // is used to record the cards spanned by the live objects
 531   // marked by that task/worker.
 532   BitMap*  _count_card_bitmaps;
 533 
 534   // Used to record the number of marked live bytes
 535   // (for each region, by worker thread).
 536   size_t** _count_marked_bytes;
 537 
 538   // Card index of the bottom of the G1 heap. Used for biasing indices into
 539   // the card bitmaps.
 540   intptr_t _heap_bottom_card_num;
 541 
 542   // Set to true when initialization is complete
 543   bool _completed_initialization;
 544 
 545 public:
 546   // Manipulation of the global mark stack.
 547   // The push and pop operations are used by tasks for transfers
 548   // between task-local queues and the global mark stack, and use
 549   // locking for concurrency safety.
 550   bool mark_stack_push(oop* arr, int n) {
 551     _markStack.par_push_arr(arr, n);
 552     if (_markStack.overflow()) {
 553       set_has_overflown();
 554       return false;
 555     }
 556     return true;
 557   }
 558   void mark_stack_pop(oop* arr, int max, int* n) {
 559     _markStack.par_pop_arr(arr, max, n);
 560   }
 561   size_t mark_stack_size()                { return _markStack.size(); }
 562   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 563   bool mark_stack_overflow()              { return _markStack.overflow(); }
 564   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 565 
 566   CMRootRegions* root_regions() { return &_root_regions; }
 567 
 568   bool concurrent_marking_in_progress() {
 569     return _concurrent_marking_in_progress;
 570   }
 571   void set_concurrent_marking_in_progress() {
 572     _concurrent_marking_in_progress = true;
 573   }
 574   void clear_concurrent_marking_in_progress() {
 575     _concurrent_marking_in_progress = false;
 576   }
 577 
 578   void update_accum_task_vtime(int i, double vtime) {
 579     _accum_task_vtime[i] += vtime;
 580   }
 581 
 582   double all_task_accum_vtime() {
 583     double ret = 0.0;
 584     for (uint i = 0; i < _max_worker_id; ++i)
 585       ret += _accum_task_vtime[i];
 586     return ret;
 587   }
 588 
 589   // Attempts to steal an object from the task queues of other tasks
 590   bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
 591 
 592   ConcurrentMark(G1CollectedHeap* g1h,
 593                  G1RegionToSpaceMapper* prev_bitmap_storage,
 594                  G1RegionToSpaceMapper* next_bitmap_storage);
 595   ~ConcurrentMark();
 596 
 597   ConcurrentMarkThread* cmThread() { return _cmThread; }
 598 
 599   CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 600   CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 601 
 602   // Returns the number of GC threads to be used in a concurrent
 603   // phase based on the number of GC threads being used in a STW
 604   // phase.
 605   uint scale_parallel_threads(uint n_par_threads);
 606 
 607   // Calculates the number of GC threads to be used in a concurrent phase.
 608   uint calc_parallel_marking_threads();
 609 
 610   // The following three are interaction between CM and
 611   // G1CollectedHeap
 612 
 613   // This notifies CM that a root during initial-mark needs to be
 614   // grayed. It is MT-safe. word_size is the size of the object in
 615   // words. It is passed explicitly as sometimes we cannot calculate
 616   // it from the given object because it might be in an inconsistent
 617   // state (e.g., in to-space and being copied). So the caller is
 618   // responsible for dealing with this issue (e.g., get the size from
 619   // the from-space image when the to-space image might be
 620   // inconsistent) and always passing the size. hr is the region that
 621   // contains the object and it's passed optionally from callers who
 622   // might already have it (no point in recalculating it).
 623   inline void grayRoot(oop obj,
 624                        size_t word_size,
 625                        uint worker_id,
 626                        HeapRegion* hr = NULL);
 627 
 628   // It iterates over the heap and for each object it comes across it
 629   // will dump the contents of its reference fields, as well as
 630   // liveness information for the object and its referents. The dump
 631   // will be written to a file with the following name:
 632   // G1PrintReachableBaseFile + "." + str.
 633   // vo decides whether the prev (vo == UsePrevMarking), the next
 634   // (vo == UseNextMarking) marking information, or the mark word
 635   // (vo == UseMarkWord) will be used to determine the liveness of
 636   // each object / referent.
 637   // If all is true, all objects in the heap will be dumped, otherwise
 638   // only the live ones. In the dump the following symbols / breviations
 639   // are used:
 640   //   M : an explicitly live object (its bitmap bit is set)
 641   //   > : an implicitly live object (over tams)
 642   //   O : an object outside the G1 heap (typically: in the perm gen)
 643   //   NOT : a reference field whose referent is not live
 644   //   AND MARKED : indicates that an object is both explicitly and
 645   //   implicitly live (it should be one or the other, not both)
 646   void print_reachable(const char* str,
 647                        VerifyOption vo,
 648                        bool all) PRODUCT_RETURN;
 649 
 650   // Clear the next marking bitmap (will be called concurrently).
 651   void clearNextBitmap();
 652 
 653   // Return whether the next mark bitmap has no marks set. To be used for assertions
 654   // only. Will not yield to pause requests.
 655   bool nextMarkBitmapIsClear();
 656 
 657   // These two do the work that needs to be done before and after the
 658   // initial root checkpoint. Since this checkpoint can be done at two
 659   // different points (i.e. an explicit pause or piggy-backed on a
 660   // young collection), then it's nice to be able to easily share the
 661   // pre/post code. It might be the case that we can put everything in
 662   // the post method. TP
 663   void checkpointRootsInitialPre();
 664   void checkpointRootsInitialPost();
 665 
 666   // Scan all the root regions and mark everything reachable from
 667   // them.
 668   void scanRootRegions();
 669 
 670   // Scan a single root region and mark everything reachable from it.
 671   void scanRootRegion(HeapRegion* hr, uint worker_id);
 672 
 673   // Do concurrent phase of marking, to a tentative transitive closure.
 674   void markFromRoots();
 675 
 676   void checkpointRootsFinal(bool clear_all_soft_refs);
 677   void checkpointRootsFinalWork();
 678   void cleanup();
 679   void completeCleanup();
 680 
 681   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 682   // this carefully!
 683   inline void markPrev(oop p);
 684 
 685   // Clears marks for all objects in the given range, for the prev or
 686   // next bitmaps.  NB: the previous bitmap is usually
 687   // read-only, so use this carefully!
 688   void clearRangePrevBitmap(MemRegion mr);
 689   void clearRangeNextBitmap(MemRegion mr);
 690 
 691   // Notify data structures that a GC has started.
 692   void note_start_of_gc() {
 693     _markStack.note_start_of_gc();
 694   }
 695 
 696   // Notify data structures that a GC is finished.
 697   void note_end_of_gc() {
 698     _markStack.note_end_of_gc();
 699   }
 700 
 701   // Verify that there are no CSet oops on the stacks (taskqueues /
 702   // global mark stack) and fingers (global / per-task).
 703   // If marking is not in progress, it's a no-op.
 704   void verify_no_cset_oops() PRODUCT_RETURN;
 705 
 706   bool isPrevMarked(oop p) const {
 707     assert(p != NULL && p->is_oop(), "expected an oop");
 708     HeapWord* addr = (HeapWord*)p;
 709     assert(addr >= _prevMarkBitMap->startWord() ||
 710            addr < _prevMarkBitMap->endWord(), "in a region");
 711 
 712     return _prevMarkBitMap->isMarked(addr);
 713   }
 714 
 715   inline bool do_yield_check(uint worker_i = 0);
 716 
 717   // Called to abort the marking cycle after a Full GC takes place.
 718   void abort();
 719 
 720   bool has_aborted()      { return _has_aborted; }
 721 
 722   void print_summary_info();
 723 
 724   void print_worker_threads_on(outputStream* st) const;
 725 
 726   void print_on_error(outputStream* st) const;
 727 
 728   // Liveness counting
 729 
 730   // Utility routine to set an exclusive range of cards on the given
 731   // card liveness bitmap
 732   inline void set_card_bitmap_range(BitMap* card_bm,
 733                                     BitMap::idx_t start_idx,
 734                                     BitMap::idx_t end_idx,
 735                                     bool is_par);
 736 
 737   // Returns the card number of the bottom of the G1 heap.
 738   // Used in biasing indices into accounting card bitmaps.
 739   intptr_t heap_bottom_card_num() const {
 740     return _heap_bottom_card_num;
 741   }
 742 
 743   // Returns the card bitmap for a given task or worker id.
 744   BitMap* count_card_bitmap_for(uint worker_id) {
 745     assert(worker_id < _max_worker_id, "oob");
 746     assert(_count_card_bitmaps != NULL, "uninitialized");
 747     BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
 748     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 749     return task_card_bm;
 750   }
 751 
 752   // Returns the array containing the marked bytes for each region,
 753   // for the given worker or task id.
 754   size_t* count_marked_bytes_array_for(uint worker_id) {
 755     assert(worker_id < _max_worker_id, "oob");
 756     assert(_count_marked_bytes != NULL, "uninitialized");
 757     size_t* marked_bytes_array = _count_marked_bytes[worker_id];
 758     assert(marked_bytes_array != NULL, "uninitialized");
 759     return marked_bytes_array;
 760   }
 761 
 762   // Returns the index in the liveness accounting card table bitmap
 763   // for the given address
 764   inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
 765 
 766   // Counts the size of the given memory region in the the given
 767   // marked_bytes array slot for the given HeapRegion.
 768   // Sets the bits in the given card bitmap that are associated with the
 769   // cards that are spanned by the memory region.
 770   inline void count_region(MemRegion mr,
 771                            HeapRegion* hr,
 772                            size_t* marked_bytes_array,
 773                            BitMap* task_card_bm);
 774 
 775   // Counts the given memory region in the task/worker counting
 776   // data structures for the given worker id.
 777   inline void count_region(MemRegion mr, HeapRegion* hr, uint worker_id);
 778 
 779   // Counts the given object in the given task/worker counting
 780   // data structures.
 781   inline void count_object(oop obj,
 782                            HeapRegion* hr,
 783                            size_t* marked_bytes_array,
 784                            BitMap* task_card_bm);
 785 
 786   // Attempts to mark the given object and, if successful, counts
 787   // the object in the given task/worker counting structures.
 788   inline bool par_mark_and_count(oop obj,
 789                                  HeapRegion* hr,
 790                                  size_t* marked_bytes_array,
 791                                  BitMap* task_card_bm);
 792 
 793   // Attempts to mark the given object and, if successful, counts
 794   // the object in the task/worker counting structures for the
 795   // given worker id.
 796   inline bool par_mark_and_count(oop obj,
 797                                  size_t word_size,
 798                                  HeapRegion* hr,
 799                                  uint worker_id);
 800 
 801   // Returns true if initialization was successfully completed.
 802   bool completed_initialization() const {
 803     return _completed_initialization;
 804   }
 805 
 806 protected:
 807   // Clear all the per-task bitmaps and arrays used to store the
 808   // counting data.
 809   void clear_all_count_data();
 810 
 811   // Aggregates the counting data for each worker/task
 812   // that was constructed while marking. Also sets
 813   // the amount of marked bytes for each region and
 814   // the top at concurrent mark count.
 815   void aggregate_count_data();
 816 
 817   // Verification routine
 818   void verify_count_data();
 819 };
 820 
 821 // A class representing a marking task.
 822 class CMTask : public TerminatorTerminator {
 823 private:
 824   enum PrivateConstants {
 825     // the regular clock call is called once the scanned words reaches
 826     // this limit
 827     words_scanned_period          = 12*1024,
 828     // the regular clock call is called once the number of visited
 829     // references reaches this limit
 830     refs_reached_period           = 384,
 831     // initial value for the hash seed, used in the work stealing code
 832     init_hash_seed                = 17,
 833     // how many entries will be transferred between global stack and
 834     // local queues
 835     global_stack_transfer_size    = 16
 836   };
 837 
 838   uint                        _worker_id;
 839   G1CollectedHeap*            _g1h;
 840   ConcurrentMark*             _cm;
 841   CMBitMap*                   _nextMarkBitMap;
 842   // the task queue of this task
 843   CMTaskQueue*                _task_queue;
 844 private:
 845   // the task queue set---needed for stealing
 846   CMTaskQueueSet*             _task_queues;
 847   // indicates whether the task has been claimed---this is only  for
 848   // debugging purposes
 849   bool                        _claimed;
 850 
 851   // number of calls to this task
 852   int                         _calls;
 853 
 854   // when the virtual timer reaches this time, the marking step should
 855   // exit
 856   double                      _time_target_ms;
 857   // the start time of the current marking step
 858   double                      _start_time_ms;
 859 
 860   // the oop closure used for iterations over oops
 861   G1CMOopClosure*             _cm_oop_closure;
 862 
 863   // the region this task is scanning, NULL if we're not scanning any
 864   HeapRegion*                 _curr_region;
 865   // the local finger of this task, NULL if we're not scanning a region
 866   HeapWord*                   _finger;
 867   // limit of the region this task is scanning, NULL if we're not scanning one
 868   HeapWord*                   _region_limit;
 869 
 870   // the number of words this task has scanned
 871   size_t                      _words_scanned;
 872   // When _words_scanned reaches this limit, the regular clock is
 873   // called. Notice that this might be decreased under certain
 874   // circumstances (i.e. when we believe that we did an expensive
 875   // operation).
 876   size_t                      _words_scanned_limit;
 877   // the initial value of _words_scanned_limit (i.e. what it was
 878   // before it was decreased).
 879   size_t                      _real_words_scanned_limit;
 880 
 881   // the number of references this task has visited
 882   size_t                      _refs_reached;
 883   // When _refs_reached reaches this limit, the regular clock is
 884   // called. Notice this this might be decreased under certain
 885   // circumstances (i.e. when we believe that we did an expensive
 886   // operation).
 887   size_t                      _refs_reached_limit;
 888   // the initial value of _refs_reached_limit (i.e. what it was before
 889   // it was decreased).
 890   size_t                      _real_refs_reached_limit;
 891 
 892   // used by the work stealing stuff
 893   int                         _hash_seed;
 894   // if this is true, then the task has aborted for some reason
 895   bool                        _has_aborted;
 896   // set when the task aborts because it has met its time quota
 897   bool                        _has_timed_out;
 898   // true when we're draining SATB buffers; this avoids the task
 899   // aborting due to SATB buffers being available (as we're already
 900   // dealing with them)
 901   bool                        _draining_satb_buffers;
 902 
 903   // number sequence of past step times
 904   NumberSeq                   _step_times_ms;
 905   // elapsed time of this task
 906   double                      _elapsed_time_ms;
 907   // termination time of this task
 908   double                      _termination_time_ms;
 909   // when this task got into the termination protocol
 910   double                      _termination_start_time_ms;
 911 
 912   // true when the task is during a concurrent phase, false when it is
 913   // in the remark phase (so, in the latter case, we do not have to
 914   // check all the things that we have to check during the concurrent
 915   // phase, i.e. SATB buffer availability...)
 916   bool                        _concurrent;
 917 
 918   TruncatedSeq                _marking_step_diffs_ms;
 919 
 920   // Counting data structures. Embedding the task's marked_bytes_array
 921   // and card bitmap into the actual task saves having to go through
 922   // the ConcurrentMark object.
 923   size_t*                     _marked_bytes_array;
 924   BitMap*                     _card_bm;
 925 
 926   // it updates the local fields after this task has claimed
 927   // a new region to scan
 928   void setup_for_region(HeapRegion* hr);
 929   // it brings up-to-date the limit of the region
 930   void update_region_limit();
 931 
 932   // called when either the words scanned or the refs visited limit
 933   // has been reached
 934   void reached_limit();
 935   // recalculates the words scanned and refs visited limits
 936   void recalculate_limits();
 937   // decreases the words scanned and refs visited limits when we reach
 938   // an expensive operation
 939   void decrease_limits();
 940   // it checks whether the words scanned or refs visited reached their
 941   // respective limit and calls reached_limit() if they have
 942   void check_limits() {
 943     if (_words_scanned >= _words_scanned_limit ||
 944         _refs_reached >= _refs_reached_limit) {
 945       reached_limit();
 946     }
 947   }
 948   // this is supposed to be called regularly during a marking step as
 949   // it checks a bunch of conditions that might cause the marking step
 950   // to abort
 951   void regular_clock_call();
 952   bool concurrent() { return _concurrent; }
 953 
 954   // Test whether obj might have already been passed over by the
 955   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 956   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 957 
 958   template<bool scan> void process_grey_object(oop obj);
 959 
 960 public:
 961   // It resets the task; it should be called right at the beginning of
 962   // a marking phase.
 963   void reset(CMBitMap* _nextMarkBitMap);
 964   // it clears all the fields that correspond to a claimed region.
 965   void clear_region_fields();
 966 
 967   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
 968 
 969   // The main method of this class which performs a marking step
 970   // trying not to exceed the given duration. However, it might exit
 971   // prematurely, according to some conditions (i.e. SATB buffers are
 972   // available for processing).
 973   void do_marking_step(double target_ms,
 974                        bool do_termination,
 975                        bool is_serial);
 976 
 977   // These two calls start and stop the timer
 978   void record_start_time() {
 979     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 980   }
 981   void record_end_time() {
 982     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 983   }
 984 
 985   // returns the worker ID associated with this task.
 986   uint worker_id() { return _worker_id; }
 987 
 988   // From TerminatorTerminator. It determines whether this task should
 989   // exit the termination protocol after it's entered it.
 990   virtual bool should_exit_termination();
 991 
 992   // Resets the local region fields after a task has finished scanning a
 993   // region; or when they have become stale as a result of the region
 994   // being evacuated.
 995   void giveup_current_region();
 996 
 997   HeapWord* finger()            { return _finger; }
 998 
 999   bool has_aborted()            { return _has_aborted; }
1000   void set_has_aborted()        { _has_aborted = true; }
1001   void clear_has_aborted()      { _has_aborted = false; }
1002   bool has_timed_out()          { return _has_timed_out; }
1003   bool claimed()                { return _claimed; }
1004 
1005   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1006 
1007   // Increment the number of references this task has visited.
1008   void increment_refs_reached() { ++_refs_reached; }
1009 
1010   // Grey the object by marking it.  If not already marked, push it on
1011   // the local queue if below the finger.
1012   // Precondition: obj is in region.
1013   // Precondition: obj is below region's NTAMS.
1014   inline void make_reference_grey(oop obj, HeapRegion* region);
1015 
1016   // Grey the object (by calling make_grey_reference) if required,
1017   // e.g. obj is below its containing region's NTAMS.
1018   // Precondition: obj is a valid heap object.
1019   inline void deal_with_reference(oop obj);
1020 
1021   // It scans an object and visits its children.
1022   inline void scan_object(oop obj);
1023 
1024   // It pushes an object on the local queue.
1025   inline void push(oop obj);
1026 
1027   // These two move entries to/from the global stack.
1028   void move_entries_to_global_stack();
1029   void get_entries_from_global_stack();
1030 
1031   // It pops and scans objects from the local queue. If partially is
1032   // true, then it stops when the queue size is of a given limit. If
1033   // partially is false, then it stops when the queue is empty.
1034   void drain_local_queue(bool partially);
1035   // It moves entries from the global stack to the local queue and
1036   // drains the local queue. If partially is true, then it stops when
1037   // both the global stack and the local queue reach a given size. If
1038   // partially if false, it tries to empty them totally.
1039   void drain_global_stack(bool partially);
1040   // It keeps picking SATB buffers and processing them until no SATB
1041   // buffers are available.
1042   void drain_satb_buffers();
1043 
1044   // moves the local finger to a new location
1045   inline void move_finger_to(HeapWord* new_finger) {
1046     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1047     _finger = new_finger;
1048   }
1049 
1050   CMTask(uint worker_id,
1051          ConcurrentMark *cm,
1052          size_t* marked_bytes,
1053          BitMap* card_bm,
1054          CMTaskQueue* task_queue,
1055          CMTaskQueueSet* task_queues);
1056 
1057   // it prints statistics associated with this task
1058   void print_stats();
1059 };
1060 
1061 // Class that's used to to print out per-region liveness
1062 // information. It's currently used at the end of marking and also
1063 // after we sort the old regions at the end of the cleanup operation.
1064 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1065 private:
1066   // Accumulators for these values.
1067   size_t _total_used_bytes;
1068   size_t _total_capacity_bytes;
1069   size_t _total_prev_live_bytes;
1070   size_t _total_next_live_bytes;
1071 
1072   // These are set up when we come across a "stars humongous" region
1073   // (as this is where most of this information is stored, not in the
1074   // subsequent "continues humongous" regions). After that, for every
1075   // region in a given humongous region series we deduce the right
1076   // values for it by simply subtracting the appropriate amount from
1077   // these fields. All these values should reach 0 after we've visited
1078   // the last region in the series.
1079   size_t _hum_used_bytes;
1080   size_t _hum_capacity_bytes;
1081   size_t _hum_prev_live_bytes;
1082   size_t _hum_next_live_bytes;
1083 
1084   // Accumulator for the remembered set size
1085   size_t _total_remset_bytes;
1086 
1087   // Accumulator for strong code roots memory size
1088   size_t _total_strong_code_roots_bytes;
1089 
1090   static double perc(size_t val, size_t total) {
1091     if (total == 0) {
1092       return 0.0;
1093     } else {
1094       return 100.0 * ((double) val / (double) total);
1095     }
1096   }
1097 
1098   static double bytes_to_mb(size_t val) {
1099     return (double) val / (double) M;
1100   }
1101 
1102   // See the .cpp file.
1103   size_t get_hum_bytes(size_t* hum_bytes);
1104   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1105                      size_t* prev_live_bytes, size_t* next_live_bytes);
1106 
1107 public:
1108   // The header and footer are printed in the constructor and
1109   // destructor respectively.
1110   G1PrintRegionLivenessInfoClosure(const char* phase_name);
1111   virtual bool doHeapRegion(HeapRegion* r);
1112   ~G1PrintRegionLivenessInfoClosure();
1113 };
1114 
1115 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP