1 /*
   2  * Copyright (c) 2001, 2011, 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/heapRegionSets.hpp"
  29 #include "utilities/taskqueue.hpp"
  30 
  31 class G1CollectedHeap;
  32 class CMTask;
  33 typedef GenericTaskQueue<oop>            CMTaskQueue;
  34 typedef GenericTaskQueueSet<CMTaskQueue> 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) :
  46     _g1(g1)
  47   {}
  48 
  49   void do_object(oop obj) {
  50     ShouldNotCallThis();
  51   }
  52   bool do_object_b(oop obj);
  53 };
  54 
  55 // A generic CM bit map.  This is essentially a wrapper around the BitMap
  56 // class, with one bit per (1<<_shifter) HeapWords.
  57 
  58 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
  59  protected:
  60   HeapWord* _bmStartWord;      // base address of range covered by map
  61   size_t    _bmWordSize;       // map size (in #HeapWords covered)
  62   const int _shifter;          // map to char or bit
  63   VirtualSpace _virtual_space; // underlying the bit map
  64   BitMap    _bm;               // the bit map itself
  65 
  66  public:
  67   // constructor
  68   CMBitMapRO(ReservedSpace rs, int shifter);
  69 
  70   enum { do_yield = true };
  71 
  72   // inquiries
  73   HeapWord* startWord()   const { return _bmStartWord; }
  74   size_t    sizeInWords() const { return _bmWordSize;  }
  75   // the following is one past the last word in space
  76   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
  77 
  78   // read marks
  79 
  80   bool isMarked(HeapWord* addr) const {
  81     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
  82            "outside underlying space?");
  83     return _bm.at(heapWordToOffset(addr));
  84   }
  85 
  86   // iteration
  87   bool iterate(BitMapClosure* cl) { return _bm.iterate(cl); }
  88   bool iterate(BitMapClosure* cl, MemRegion mr);
  89 
  90   // Return the address corresponding to the next marked 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* getNextMarkedWordAddress(HeapWord* addr,
  94                                      HeapWord* limit = NULL) const;
  95   // Return the address corresponding to the next unmarked bit at or after
  96   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  97   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  98   HeapWord* getNextUnmarkedWordAddress(HeapWord* addr,
  99                                        HeapWord* limit = NULL) const;
 100 
 101   // conversion utilities
 102   // XXX Fix these so that offsets are size_t's...
 103   HeapWord* offsetToHeapWord(size_t offset) const {
 104     return _bmStartWord + (offset << _shifter);
 105   }
 106   size_t heapWordToOffset(HeapWord* addr) const {
 107     return pointer_delta(addr, _bmStartWord) >> _shifter;
 108   }
 109   int heapWordDiffToOffsetDiff(size_t diff) const;
 110   HeapWord* nextWord(HeapWord* addr) {
 111     return offsetToHeapWord(heapWordToOffset(addr) + 1);
 112   }
 113 
 114   void mostly_disjoint_range_union(BitMap*   from_bitmap,
 115                                    size_t    from_start_index,
 116                                    HeapWord* to_start_word,
 117                                    size_t    word_num);
 118 
 119   // debugging
 120   NOT_PRODUCT(bool covers(ReservedSpace rs) const;)
 121 };
 122 
 123 class CMBitMap : public CMBitMapRO {
 124 
 125  public:
 126   // constructor
 127   CMBitMap(ReservedSpace rs, int shifter) :
 128     CMBitMapRO(rs, shifter) {}
 129 
 130   // write marks
 131   void mark(HeapWord* addr) {
 132     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 133            "outside underlying space?");
 134     _bm.set_bit(heapWordToOffset(addr));
 135   }
 136   void clear(HeapWord* addr) {
 137     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 138            "outside underlying space?");
 139     _bm.clear_bit(heapWordToOffset(addr));
 140   }
 141   bool parMark(HeapWord* addr) {
 142     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 143            "outside underlying space?");
 144     return _bm.par_set_bit(heapWordToOffset(addr));
 145   }
 146   bool parClear(HeapWord* addr) {
 147     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 148            "outside underlying space?");
 149     return _bm.par_clear_bit(heapWordToOffset(addr));
 150   }
 151   void markRange(MemRegion mr);
 152   void clearAll();
 153   void clearRange(MemRegion mr);
 154 
 155   // Starting at the bit corresponding to "addr" (inclusive), find the next
 156   // "1" bit, if any.  This bit starts some run of consecutive "1"'s; find
 157   // the end of this run (stopping at "end_addr").  Return the MemRegion
 158   // covering from the start of the region corresponding to the first bit
 159   // of the run to the end of the region corresponding to the last bit of
 160   // the run.  If there is no "1" bit at or after "addr", return an empty
 161   // MemRegion.
 162   MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
 163 };
 164 
 165 // Represents a marking stack used by the CM collector.
 166 // Ideally this should be GrowableArray<> just like MSC's marking stack(s).
 167 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
 168   ConcurrentMark* _cm;
 169   oop*   _base;      // bottom of stack
 170   jint   _index;     // one more than last occupied index
 171   jint   _capacity;  // max #elements
 172   jint   _oops_do_bound;  // Number of elements to include in next iteration.
 173   NOT_PRODUCT(jint _max_depth;)  // max depth plumbed during run
 174 
 175   bool   _overflow;
 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   void allocate(size_t size);
 184 
 185   oop pop() {
 186     if (!isEmpty()) {
 187       return _base[--_index] ;
 188     }
 189     return NULL;
 190   }
 191 
 192   // If overflow happens, don't do the push, and record the overflow.
 193   // *Requires* that "ptr" is already marked.
 194   void push(oop ptr) {
 195     if (isFull()) {
 196       // Record overflow.
 197       _overflow = true;
 198       return;
 199     } else {
 200       _base[_index++] = ptr;
 201       NOT_PRODUCT(_max_depth = MAX2(_max_depth, _index));
 202     }
 203   }
 204   // Non-block impl.  Note: concurrency is allowed only with other
 205   // "par_push" operations, not with "pop" or "drain".  We would need
 206   // parallel versions of them if such concurrency was desired.
 207   void par_push(oop ptr);
 208 
 209   // Pushes the first "n" elements of "ptr_arr" on the stack.
 210   // Non-block impl.  Note: concurrency is allowed only with other
 211   // "par_adjoin_arr" or "push" operations, not with "pop" or "drain".
 212   void par_adjoin_arr(oop* ptr_arr, int n);
 213 
 214   // Pushes the first "n" elements of "ptr_arr" on the stack.
 215   // Locking impl: concurrency is allowed only with
 216   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 217   // locking strategy.
 218   void par_push_arr(oop* ptr_arr, int n);
 219 
 220   // If returns false, the array was empty.  Otherwise, removes up to "max"
 221   // elements from the stack, and transfers them to "ptr_arr" in an
 222   // unspecified order.  The actual number transferred is given in "n" ("n
 223   // == 0" is deliberately redundant with the return value.)  Locking impl:
 224   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 225   // operations, which use the same locking strategy.
 226   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 227 
 228   // Drain the mark stack, applying the given closure to all fields of
 229   // objects on the stack.  (That is, continue until the stack is empty,
 230   // even if closure applications add entries to the stack.)  The "bm"
 231   // argument, if non-null, may be used to verify that only marked objects
 232   // are on the mark stack.  If "yield_after" is "true", then the
 233   // concurrent marker performing the drain offers to yield after
 234   // processing each object.  If a yield occurs, stops the drain operation
 235   // and returns false.  Otherwise, returns true.
 236   template<class OopClosureClass>
 237   bool drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after = false);
 238 
 239   bool isEmpty()    { return _index == 0; }
 240   bool isFull()     { return _index == _capacity; }
 241   int maxElems()    { return _capacity; }
 242 
 243   bool overflow() { return _overflow; }
 244   void clear_overflow() { _overflow = false; }
 245 
 246   int  size() { return _index; }
 247 
 248   void setEmpty()   { _index = 0; clear_overflow(); }
 249 
 250   // Record the current size; a subsequent "oops_do" will iterate only over
 251   // indices valid at the time of this call.
 252   void set_oops_do_bound(jint bound = -1) {
 253     if (bound == -1) {
 254       _oops_do_bound = _index;
 255     } else {
 256       _oops_do_bound = bound;
 257     }
 258   }
 259   jint oops_do_bound() { return _oops_do_bound; }
 260   // iterate over the oops in the mark stack, up to the bound recorded via
 261   // the call above.
 262   void oops_do(OopClosure* f);
 263 };
 264 
 265 class CMRegionStack VALUE_OBJ_CLASS_SPEC {
 266   MemRegion* _base;
 267   jint _capacity;
 268   jint _index;
 269   jint _oops_do_bound;
 270   bool _overflow;
 271 public:
 272   CMRegionStack();
 273   ~CMRegionStack();
 274   void allocate(size_t size);
 275 
 276   // This is lock-free; assumes that it will only be called in parallel
 277   // with other "push" operations (no pops).
 278   void push_lock_free(MemRegion mr);
 279 
 280   // Lock-free; assumes that it will only be called in parallel
 281   // with other "pop" operations (no pushes).
 282   MemRegion pop_lock_free();
 283 
 284 #if 0
 285   // The routines that manipulate the region stack with a lock are
 286   // not currently used. They should be retained, however, as a
 287   // diagnostic aid.
 288 
 289   // These two are the implementations that use a lock. They can be
 290   // called concurrently with each other but they should not be called
 291   // concurrently with the lock-free versions (push() / pop()).
 292   void push_with_lock(MemRegion mr);
 293   MemRegion pop_with_lock();
 294 #endif
 295 
 296   bool isEmpty()    { return _index == 0; }
 297   bool isFull()     { return _index == _capacity; }
 298 
 299   bool overflow() { return _overflow; }
 300   void clear_overflow() { _overflow = false; }
 301 
 302   int  size() { return _index; }
 303 
 304   // It iterates over the entries in the region stack and it
 305   // invalidates (i.e. assigns MemRegion()) the ones that point to
 306   // regions in the collection set.
 307   bool invalidate_entries_into_cset();
 308 
 309   // This gives an upper bound up to which the iteration in
 310   // invalidate_entries_into_cset() will reach. This prevents
 311   // newly-added entries to be unnecessarily scanned.
 312   void set_oops_do_bound() {
 313     _oops_do_bound = _index;
 314   }
 315 
 316   void setEmpty()   { _index = 0; clear_overflow(); }
 317 };
 318 
 319 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
 320 private:
 321 #ifndef PRODUCT
 322   uintx _num_remaining;
 323   bool _force;
 324 #endif // !defined(PRODUCT)
 325 
 326 public:
 327   void init() PRODUCT_RETURN;
 328   void update() PRODUCT_RETURN;
 329   bool should_force() PRODUCT_RETURN_( return false; );
 330 };
 331 
 332 // this will enable a variety of different statistics per GC task
 333 #define _MARKING_STATS_       0
 334 // this will enable the higher verbose levels
 335 #define _MARKING_VERBOSE_     0
 336 
 337 #if _MARKING_STATS_
 338 #define statsOnly(statement)  \
 339 do {                          \
 340   statement ;                 \
 341 } while (0)
 342 #else // _MARKING_STATS_
 343 #define statsOnly(statement)  \
 344 do {                          \
 345 } while (0)
 346 #endif // _MARKING_STATS_
 347 
 348 typedef enum {
 349   no_verbose  = 0,   // verbose turned off
 350   stats_verbose,     // only prints stats at the end of marking
 351   low_verbose,       // low verbose, mostly per region and per major event
 352   medium_verbose,    // a bit more detailed than low
 353   high_verbose       // per object verbose
 354 } CMVerboseLevel;
 355 
 356 
 357 class ConcurrentMarkThread;
 358 
 359 class ConcurrentMark: public CHeapObj {
 360   friend class ConcurrentMarkThread;
 361   friend class CMTask;
 362   friend class CMBitMapClosure;
 363   friend class CSetMarkOopClosure;
 364   friend class CMGlobalObjectClosure;
 365   friend class CMRemarkTask;
 366   friend class CMConcurrentMarkingTask;
 367   friend class G1ParNoteEndTask;
 368   friend class CalcLiveObjectsClosure;
 369   friend class G1CMRefProcTaskProxy;
 370   friend class G1CMRefProcTaskExecutor;
 371   friend class G1CMParKeepAliveAndDrainClosure;
 372   friend class G1CMParDrainMarkingStackClosure;
 373 
 374 protected:
 375   ConcurrentMarkThread* _cmThread;   // the thread doing the work
 376   G1CollectedHeap*      _g1h;        // the heap.
 377   size_t                _parallel_marking_threads; // the number of marking
 378                                                    // threads we're use
 379   size_t                _max_parallel_marking_threads; // max number of marking
 380                                                    // threads we'll ever use
 381   double                _sleep_factor; // how much we have to sleep, with
 382                                        // respect to the work we just did, to
 383                                        // meet the marking overhead goal
 384   double                _marking_task_overhead; // marking target overhead for
 385                                                 // a single task
 386 
 387   // same as the two above, but for the cleanup task
 388   double                _cleanup_sleep_factor;
 389   double                _cleanup_task_overhead;
 390 
 391   FreeRegionList        _cleanup_list;
 392 
 393   // CMS marking support structures
 394   CMBitMap                _markBitMap1;
 395   CMBitMap                _markBitMap2;
 396   CMBitMapRO*             _prevMarkBitMap; // completed mark bitmap
 397   CMBitMap*               _nextMarkBitMap; // under-construction mark bitmap
 398   bool                    _at_least_one_mark_complete;
 399 
 400   BitMap                  _region_bm;
 401   BitMap                  _card_bm;
 402 
 403   // Heap bounds
 404   HeapWord*               _heap_start;
 405   HeapWord*               _heap_end;
 406 
 407   // For gray objects
 408   CMMarkStack             _markStack; // Grey objects behind global finger.
 409   CMRegionStack           _regionStack; // Grey regions behind global finger.
 410   HeapWord* volatile      _finger;  // the global finger, region aligned,
 411                                     // always points to the end of the
 412                                     // last claimed region
 413 
 414   // marking tasks
 415   size_t                  _max_task_num; // maximum task number
 416   size_t                  _active_tasks; // task num currently active
 417   CMTask**                _tasks;        // task queue array (max_task_num len)
 418   CMTaskQueueSet*         _task_queues;  // task queue set
 419   ParallelTaskTerminator  _terminator;   // for termination
 420 
 421   // Two sync barriers that are used to synchronise tasks when an
 422   // overflow occurs. The algorithm is the following. All tasks enter
 423   // the first one to ensure that they have all stopped manipulating
 424   // the global data structures. After they exit it, they re-initialise
 425   // their data structures and task 0 re-initialises the global data
 426   // structures. Then, they enter the second sync barrier. This
 427   // ensure, that no task starts doing work before all data
 428   // structures (local and global) have been re-initialised. When they
 429   // exit it, they are free to start working again.
 430   WorkGangBarrierSync     _first_overflow_barrier_sync;
 431   WorkGangBarrierSync     _second_overflow_barrier_sync;
 432 
 433   // this is set by any task, when an overflow on the global data
 434   // structures is detected.
 435   volatile bool           _has_overflown;
 436   // true: marking is concurrent, false: we're in remark
 437   volatile bool           _concurrent;
 438   // set at the end of a Full GC so that marking aborts
 439   volatile bool           _has_aborted;
 440 
 441   // used when remark aborts due to an overflow to indicate that
 442   // another concurrent marking phase should start
 443   volatile bool           _restart_for_overflow;
 444 
 445   // This is true from the very start of concurrent marking until the
 446   // point when all the tasks complete their work. It is really used
 447   // to determine the points between the end of concurrent marking and
 448   // time of remark.
 449   volatile bool           _concurrent_marking_in_progress;
 450 
 451   // verbose level
 452   CMVerboseLevel          _verbose_level;
 453 
 454   // These two fields are used to implement the optimisation that
 455   // avoids pushing objects on the global/region stack if there are
 456   // no collection set regions above the lowest finger.
 457 
 458   // This is the lowest finger (among the global and local fingers),
 459   // which is calculated before a new collection set is chosen.
 460   HeapWord* _min_finger;
 461   // If this flag is true, objects/regions that are marked below the
 462   // finger should be pushed on the stack(s). If this is flag is
 463   // false, it is safe not to push them on the stack(s).
 464   bool      _should_gray_objects;
 465 
 466   // All of these times are in ms.
 467   NumberSeq _init_times;
 468   NumberSeq _remark_times;
 469   NumberSeq   _remark_mark_times;
 470   NumberSeq   _remark_weak_ref_times;
 471   NumberSeq _cleanup_times;
 472   double    _total_counting_time;
 473   double    _total_rs_scrub_time;
 474 
 475   double*   _accum_task_vtime;   // accumulated task vtime
 476 
 477   FlexibleWorkGang* _parallel_workers;
 478 
 479   ForceOverflowSettings _force_overflow_conc;
 480   ForceOverflowSettings _force_overflow_stw;
 481 
 482   void weakRefsWork(bool clear_all_soft_refs);
 483 
 484   void swapMarkBitMaps();
 485 
 486   // It resets the global marking data structures, as well as the
 487   // task local ones; should be called during initial mark.
 488   void reset();
 489   // It resets all the marking data structures.
 490   void clear_marking_state(bool clear_overflow = true);
 491 
 492   // It should be called to indicate which phase we're in (concurrent
 493   // mark or remark) and how many threads are currently active.
 494   void set_phase(size_t active_tasks, bool concurrent);
 495   // We do this after we're done with marking so that the marking data
 496   // structures are initialised to a sensible and predictable state.
 497   void set_non_marking_state();
 498 
 499   // prints all gathered CM-related statistics
 500   void print_stats();
 501 
 502   bool cleanup_list_is_empty() {
 503     return _cleanup_list.is_empty();
 504   }
 505 
 506   // accessor methods
 507   size_t parallel_marking_threads() { return _parallel_marking_threads; }
 508   size_t max_parallel_marking_threads() { return _max_parallel_marking_threads;}
 509   double sleep_factor()             { return _sleep_factor; }
 510   double marking_task_overhead()    { return _marking_task_overhead;}
 511   double cleanup_sleep_factor()     { return _cleanup_sleep_factor; }
 512   double cleanup_task_overhead()    { return _cleanup_task_overhead;}
 513 
 514   HeapWord*               finger()        { return _finger;   }
 515   bool                    concurrent()    { return _concurrent; }
 516   size_t                  active_tasks()  { return _active_tasks; }
 517   ParallelTaskTerminator* terminator()    { return &_terminator; }
 518 
 519   // It claims the next available region to be scanned by a marking
 520   // task. It might return NULL if the next region is empty or we have
 521   // run out of regions. In the latter case, out_of_regions()
 522   // determines whether we've really run out of regions or the task
 523   // should call claim_region() again.  This might seem a bit
 524   // awkward. Originally, the code was written so that claim_region()
 525   // either successfully returned with a non-empty region or there
 526   // were no more regions to be claimed. The problem with this was
 527   // that, in certain circumstances, it iterated over large chunks of
 528   // the heap finding only empty regions and, while it was working, it
 529   // was preventing the calling task to call its regular clock
 530   // method. So, this way, each task will spend very little time in
 531   // claim_region() and is allowed to call the regular clock method
 532   // frequently.
 533   HeapRegion* claim_region(int task);
 534 
 535   // It determines whether we've run out of regions to scan.
 536   bool        out_of_regions() { return _finger == _heap_end; }
 537 
 538   // Returns the task with the given id
 539   CMTask* task(int id) {
 540     assert(0 <= id && id < (int) _active_tasks,
 541            "task id not within active bounds");
 542     return _tasks[id];
 543   }
 544 
 545   // Returns the task queue with the given id
 546   CMTaskQueue* task_queue(int id) {
 547     assert(0 <= id && id < (int) _active_tasks,
 548            "task queue id not within active bounds");
 549     return (CMTaskQueue*) _task_queues->queue(id);
 550   }
 551 
 552   // Returns the task queue set
 553   CMTaskQueueSet* task_queues()  { return _task_queues; }
 554 
 555   // Access / manipulation of the overflow flag which is set to
 556   // indicate that the global stack or region stack has overflown
 557   bool has_overflown()           { return _has_overflown; }
 558   void set_has_overflown()       { _has_overflown = true; }
 559   void clear_has_overflown()     { _has_overflown = false; }
 560 
 561   bool has_aborted()             { return _has_aborted; }
 562   bool restart_for_overflow()    { return _restart_for_overflow; }
 563 
 564   // Methods to enter the two overflow sync barriers
 565   void enter_first_sync_barrier(int task_num);
 566   void enter_second_sync_barrier(int task_num);
 567 
 568   ForceOverflowSettings* force_overflow_conc() {
 569     return &_force_overflow_conc;
 570   }
 571 
 572   ForceOverflowSettings* force_overflow_stw() {
 573     return &_force_overflow_stw;
 574   }
 575 
 576   ForceOverflowSettings* force_overflow() {
 577     if (concurrent()) {
 578       return force_overflow_conc();
 579     } else {
 580       return force_overflow_stw();
 581     }
 582   }
 583 
 584   // Live Data Counting data structures...
 585   // These data structures are initialized at the start of
 586   // marking. They are written to while marking is active.
 587   // They are aggregated during remark; the aggregated values
 588   // are then used to populate the _region_bm, _card_bm, and
 589   // the total live bytes, which are then subsequently updated
 590   // during cleanup.
 591 
 592   // An array of bitmaps (one bit map per task). Each bitmap
 593   // is used to record the cards spanned by the live objects
 594   // marked by that task/worker.
 595   BitMap*  _count_card_bitmaps;
 596 
 597   // Used to record the number of marked live bytes
 598   // (for each region, by worker thread).
 599   size_t** _count_marked_bytes;
 600 
 601   // Card index of the bottom of the G1 heap. Used for biasing indices into
 602   // the card bitmaps.
 603   intptr_t _heap_bottom_card_num;
 604 
 605 public:
 606   // Manipulation of the global mark stack.
 607   // Notice that the first mark_stack_push is CAS-based, whereas the
 608   // two below are Mutex-based. This is OK since the first one is only
 609   // called during evacuation pauses and doesn't compete with the
 610   // other two (which are called by the marking tasks during
 611   // concurrent marking or remark).
 612   bool mark_stack_push(oop p) {
 613     _markStack.par_push(p);
 614     if (_markStack.overflow()) {
 615       set_has_overflown();
 616       return false;
 617     }
 618     return true;
 619   }
 620   bool mark_stack_push(oop* arr, int n) {
 621     _markStack.par_push_arr(arr, n);
 622     if (_markStack.overflow()) {
 623       set_has_overflown();
 624       return false;
 625     }
 626     return true;
 627   }
 628   void mark_stack_pop(oop* arr, int max, int* n) {
 629     _markStack.par_pop_arr(arr, max, n);
 630   }
 631   size_t mark_stack_size()                { return _markStack.size(); }
 632   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 633   bool mark_stack_overflow()              { return _markStack.overflow(); }
 634   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 635 
 636   // (Lock-free) Manipulation of the region stack
 637   bool region_stack_push_lock_free(MemRegion mr) {
 638     // Currently we only call the lock-free version during evacuation
 639     // pauses.
 640     assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped");
 641 
 642     _regionStack.push_lock_free(mr);
 643     if (_regionStack.overflow()) {
 644       set_has_overflown();
 645       return false;
 646     }
 647     return true;
 648   }
 649 
 650   // Lock-free version of region-stack pop. Should only be
 651   // called in tandem with other lock-free pops.
 652   MemRegion region_stack_pop_lock_free() {
 653     return _regionStack.pop_lock_free();
 654   }
 655 
 656 #if 0
 657   // The routines that manipulate the region stack with a lock are
 658   // not currently used. They should be retained, however, as a
 659   // diagnostic aid.
 660 
 661   bool region_stack_push_with_lock(MemRegion mr) {
 662     // Currently we only call the lock-based version during either
 663     // concurrent marking or remark.
 664     assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
 665            "if we are at a safepoint it should be the remark safepoint");
 666 
 667     _regionStack.push_with_lock(mr);
 668     if (_regionStack.overflow()) {
 669       set_has_overflown();
 670       return false;
 671     }
 672     return true;
 673   }
 674 
 675   MemRegion region_stack_pop_with_lock() {
 676     // Currently we only call the lock-based version during either
 677     // concurrent marking or remark.
 678     assert(!SafepointSynchronize::is_at_safepoint() || !concurrent(),
 679            "if we are at a safepoint it should be the remark safepoint");
 680 
 681     return _regionStack.pop_with_lock();
 682   }
 683 #endif
 684 
 685   int region_stack_size()               { return _regionStack.size(); }
 686   bool region_stack_overflow()          { return _regionStack.overflow(); }
 687   bool region_stack_empty()             { return _regionStack.isEmpty(); }
 688 
 689   // Iterate over any regions that were aborted while draining the
 690   // region stack (any such regions are saved in the corresponding
 691   // CMTask) and invalidate (i.e. assign to the empty MemRegion())
 692   // any regions that point into the collection set.
 693   bool invalidate_aborted_regions_in_cset();
 694 
 695   // Returns true if there are any aborted memory regions.
 696   bool has_aborted_regions();
 697 
 698   bool concurrent_marking_in_progress() {
 699     return _concurrent_marking_in_progress;
 700   }
 701   void set_concurrent_marking_in_progress() {
 702     _concurrent_marking_in_progress = true;
 703   }
 704   void clear_concurrent_marking_in_progress() {
 705     _concurrent_marking_in_progress = false;
 706   }
 707 
 708   void update_accum_task_vtime(int i, double vtime) {
 709     _accum_task_vtime[i] += vtime;
 710   }
 711 
 712   double all_task_accum_vtime() {
 713     double ret = 0.0;
 714     for (int i = 0; i < (int)_max_task_num; ++i)
 715       ret += _accum_task_vtime[i];
 716     return ret;
 717   }
 718 
 719   // Attempts to steal an object from the task queues of other tasks
 720   bool try_stealing(int task_num, int* hash_seed, oop& obj) {
 721     return _task_queues->steal(task_num, hash_seed, obj);
 722   }
 723 
 724   // It grays an object by first marking it. Then, if it's behind the
 725   // global finger, it also pushes it on the global stack.
 726   void deal_with_reference(oop obj, int worker_i);
 727 
 728   ConcurrentMark(ReservedSpace rs, int max_regions);
 729 
 730   ConcurrentMarkThread* cmThread() { return _cmThread; }
 731 
 732   CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 733   CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 734 
 735   // Returns the number of GC threads to be used in a concurrent
 736   // phase based on the number of GC threads being used in a STW
 737   // phase.
 738   size_t scale_parallel_threads(size_t n_par_threads);
 739 
 740   // Calculates the number of GC threads to be used in a concurrent phase.
 741   size_t calc_parallel_marking_threads();
 742 
 743   // The following three are interaction between CM and
 744   // G1CollectedHeap
 745 
 746   // This notifies CM that a root during initial-mark needs to be
 747   // grayed and it's MT-safe. Currently, we just mark it. But, in the
 748   // future, we can experiment with pushing it on the stack and we can
 749   // do this without changing G1CollectedHeap.
 750   void grayRoot(oop p, int worker_i);
 751 
 752   // It's used during evacuation pauses to gray a region, if
 753   // necessary, and it's MT-safe. It assumes that the caller has
 754   // marked any objects on that region. If _should_gray_objects is
 755   // true and we're still doing concurrent marking, the region is
 756   // pushed on the region stack, if it is located below the global
 757   // finger, otherwise we do nothing.
 758   void grayRegionIfNecessary(MemRegion mr);
 759 
 760   // It's used during evacuation pauses to mark and, if necessary,
 761   // gray a single object and it's MT-safe. It assumes the caller did
 762   // not mark the object. If _should_gray_objects is true and we're
 763   // still doing concurrent marking, the objects is pushed on the
 764   // global stack, if it is located below the global finger, otherwise
 765   // we do nothing.
 766   void markAndGrayObjectIfNecessary(oop p, int worker_i);
 767 
 768   // It iterates over the heap and for each object it comes across it
 769   // will dump the contents of its reference fields, as well as
 770   // liveness information for the object and its referents. The dump
 771   // will be written to a file with the following name:
 772   // G1PrintReachableBaseFile + "." + str.
 773   // vo decides whether the prev (vo == UsePrevMarking), the next
 774   // (vo == UseNextMarking) marking information, or the mark word
 775   // (vo == UseMarkWord) will be used to determine the liveness of
 776   // each object / referent.
 777   // If all is true, all objects in the heap will be dumped, otherwise
 778   // only the live ones. In the dump the following symbols / breviations
 779   // are used:
 780   //   M : an explicitly live object (its bitmap bit is set)
 781   //   > : an implicitly live object (over tams)
 782   //   O : an object outside the G1 heap (typically: in the perm gen)
 783   //   NOT : a reference field whose referent is not live
 784   //   AND MARKED : indicates that an object is both explicitly and
 785   //   implicitly live (it should be one or the other, not both)
 786   void print_reachable(const char* str,
 787                        VerifyOption vo, bool all) PRODUCT_RETURN;
 788 
 789   // Clear the next marking bitmap (will be called concurrently).
 790   void clearNextBitmap();
 791 
 792   // These two do the work that needs to be done before and after the
 793   // initial root checkpoint. Since this checkpoint can be done at two
 794   // different points (i.e. an explicit pause or piggy-backed on a
 795   // young collection), then it's nice to be able to easily share the
 796   // pre/post code. It might be the case that we can put everything in
 797   // the post method. TP
 798   void checkpointRootsInitialPre();
 799   void checkpointRootsInitialPost();
 800 
 801   // Do concurrent phase of marking, to a tentative transitive closure.
 802   void markFromRoots();
 803 
 804   // Process all unprocessed SATB buffers. It is called at the
 805   // beginning of an evacuation pause.
 806   void drainAllSATBBuffers();
 807 
 808   void checkpointRootsFinal(bool clear_all_soft_refs);
 809   void checkpointRootsFinalWork();
 810   void cleanup();
 811   void completeCleanup();
 812 
 813   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 814   // this carefully!
 815   void markPrev(oop p);
 816 
 817   // Clears the mark in the next bitmap for the given object.
 818   void clear_mark(oop p);
 819 
 820   // Clears marks for all objects in the given range, for both prev and
 821   // next bitmaps.  NB: the previous bitmap is usually read-only, so use
 822   // this carefully!
 823   void clearRangeBothMaps(MemRegion mr);
 824 
 825   // Record the current top of the mark and region stacks; a
 826   // subsequent oops_do() on the mark stack and
 827   // invalidate_entries_into_cset() on the region stack will iterate
 828   // only over indices valid at the time of this call.
 829   void set_oops_do_bound() {
 830     _markStack.set_oops_do_bound();
 831     _regionStack.set_oops_do_bound();
 832   }
 833   // Iterate over the oops in the mark stack and all local queues. It
 834   // also calls invalidate_entries_into_cset() on the region stack.
 835   void oops_do(OopClosure* f);
 836   // It is called at the end of an evacuation pause during marking so
 837   // that CM is notified of where the new end of the heap is. It
 838   // doesn't do anything if concurrent_marking_in_progress() is false,
 839   // unless the force parameter is true.
 840   void update_g1_committed(bool force = false);
 841 
 842   void complete_marking_in_collection_set();
 843 
 844   // It indicates that a new collection set is being chosen.
 845   void newCSet();
 846 
 847   // It registers a collection set heap region with CM. This is used
 848   // to determine whether any heap regions are located above the finger.
 849   void registerCSetRegion(HeapRegion* hr);
 850 
 851   // Resets the region fields of any active CMTask whose region fields
 852   // are in the collection set (i.e. the region currently claimed by
 853   // the CMTask will be evacuated and may be used, subsequently, as
 854   // an alloc region). When this happens the region fields in the CMTask
 855   // are stale and, hence, should be cleared causing the worker thread
 856   // to claim a new region.
 857   void reset_active_task_region_fields_in_cset();
 858 
 859   // Registers the maximum region-end associated with a set of
 860   // regions with CM. Again this is used to determine whether any
 861   // heap regions are located above the finger.
 862   void register_collection_set_finger(HeapWord* max_finger) {
 863     // max_finger is the highest heap region end of the regions currently
 864     // contained in the collection set. If this value is larger than
 865     // _min_finger then we need to gray objects.
 866     // This routine is like registerCSetRegion but for an entire
 867     // collection of regions.
 868     if (max_finger > _min_finger) {
 869       _should_gray_objects = true;
 870     }
 871   }
 872 
 873   // Returns "true" if at least one mark has been completed.
 874   bool at_least_one_mark_complete() { return _at_least_one_mark_complete; }
 875 
 876   bool isMarked(oop p) const {
 877     assert(p != NULL && p->is_oop(), "expected an oop");
 878     HeapWord* addr = (HeapWord*)p;
 879     assert(addr >= _nextMarkBitMap->startWord() ||
 880            addr < _nextMarkBitMap->endWord(), "in a region");
 881 
 882     return _nextMarkBitMap->isMarked(addr);
 883   }
 884 
 885   inline bool not_yet_marked(oop p) const;
 886 
 887   // XXX Debug code
 888   bool containing_card_is_marked(void* p);
 889   bool containing_cards_are_marked(void* start, void* last);
 890 
 891   bool isPrevMarked(oop p) const {
 892     assert(p != NULL && p->is_oop(), "expected an oop");
 893     HeapWord* addr = (HeapWord*)p;
 894     assert(addr >= _prevMarkBitMap->startWord() ||
 895            addr < _prevMarkBitMap->endWord(), "in a region");
 896 
 897     return _prevMarkBitMap->isMarked(addr);
 898   }
 899 
 900   inline bool do_yield_check(int worker_i = 0);
 901   inline bool should_yield();
 902 
 903   // Called to abort the marking cycle after a Full GC takes palce.
 904   void abort();
 905 
 906   // This prints the global/local fingers. It is used for debugging.
 907   NOT_PRODUCT(void print_finger();)
 908 
 909   void print_summary_info();
 910 
 911   void print_worker_threads_on(outputStream* st) const;
 912 
 913   // The following indicate whether a given verbose level has been
 914   // set. Notice that anything above stats is conditional to
 915   // _MARKING_VERBOSE_ having been set to 1
 916   bool verbose_stats() {
 917     return _verbose_level >= stats_verbose;
 918   }
 919   bool verbose_low() {
 920     return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
 921   }
 922   bool verbose_medium() {
 923     return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
 924   }
 925   bool verbose_high() {
 926     return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
 927   }
 928 
 929   // Counting data structure accessors
 930 
 931   // Returns the card number of the bottom of the G1 heap.
 932   // Used in biasing indices into accounting card bitmaps.
 933   intptr_t heap_bottom_card_num() const {
 934     return _heap_bottom_card_num;
 935   }
 936 
 937   // Returns the card bitmap for a given task or worker id.
 938   BitMap* count_card_bitmap_for(int worker_i) {
 939     assert(0 <= worker_i && (size_t) worker_i < _max_task_num, "oob");
 940     assert(_count_card_bitmaps != NULL, "uninitialized");
 941     BitMap* task_card_bm = &_count_card_bitmaps[worker_i];
 942     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 943     return task_card_bm;
 944   }
 945 
 946   // Returns the array containing the marked bytes for each region,
 947   // for the given worker or task id.
 948   size_t* count_marked_bytes_array_for(int worker_i) {
 949     assert(0 <= worker_i && (size_t) worker_i < _max_task_num, "oob");
 950     assert(_count_marked_bytes != NULL, "uninitialized");
 951     size_t* marked_bytes_array = _count_marked_bytes[worker_i];
 952     assert(marked_bytes_array != NULL, "uninitialized");
 953     return marked_bytes_array;
 954   }
 955 
 956   // Counts the size of the given memory region in the the given
 957   // marked_bytes array slot for the given HeapRegion.
 958   // Sets the bits in the given card bitmap that are associated with the
 959   // cards that are spanned by the memory region.
 960   inline void count_region(MemRegion mr, HeapRegion* hr,
 961                            size_t* marked_bytes_array,
 962                            BitMap* task_card_bm);
 963   
 964   // Counts the given memory region in the ask/worker counting
 965   // data structures for the given worker id.
 966   inline void count_region(MemRegion mr, int worker_i);
 967   
 968   // Counts the given object in the given task/worker counting
 969   // data structures.
 970   inline void count_object(oop obj, HeapRegion* hr,
 971                            size_t* marked_bytes_array,
 972                            BitMap* task_card_bm);
 973   
 974   // Counts the given object in the task/worker counting data
 975   // structures for the given worker id.
 976   inline void count_object(oop obj, HeapRegion* hr, int worker_i);
 977   
 978   // Attempts to mark the given object and, if successful, counts
 979   // the object in the given task/worker counting structures.
 980   inline bool par_mark_and_count(oop obj, HeapRegion* hr,
 981                                  size_t* marked_bytes_array,
 982                                  BitMap* task_card_bm);
 983 
 984   // Attempts to mark the given object and, if successful, counts
 985   // the object in the task/worker counting structures for the
 986   // given worker id.
 987   inline bool par_mark_and_count(oop obj, HeapRegion* hr, int worker_i);
 988 
 989   // Similar to the above routine but we don't know the heap region that
 990   // contains the object to be marked/counted, which this routine looks up.
 991   inline bool par_mark_and_count(oop obj, int worker_i);
 992 
 993   // Unconditionally mark the given object, and unconditinally count
 994   // the object in the counting structures for worker id 0.
 995   // Should *not* be called from parallel code.
 996   inline bool mark_and_count(oop obj, HeapRegion* hr);
 997  
 998   // Similar to the above routine but we don't know the heap region that
 999   // contains the object to be marked/counted, which this routine looks up.
1000   // Should *not* be called from parallel code.
1001   inline bool mark_and_count(oop obj);
1002 
1003   // Clears the count data for the given region from _all_ of
1004   // the per-task counting data structures.
1005   void clear_count_data_for_heap_region(HeapRegion* hr);
1006 
1007 protected:
1008   // Clear all the per-task bitmaps and arrays used to store the
1009   // counting data.
1010   void clear_all_count_data();
1011 
1012   // Aggregates the counting data for each worker/task
1013   // that was constructed while marking. Also sets
1014   // the amount of marked bytes for each region and
1015   // the top at concurrent mark count.
1016   void aggregate_and_clear_count_data();
1017 
1018   // Verification routine
1019   void verify_count_data();
1020 };
1021 
1022 // A class representing a marking task.
1023 class CMTask : public TerminatorTerminator {
1024 private:
1025   enum PrivateConstants {
1026     // the regular clock call is called once the scanned words reaches
1027     // this limit
1028     words_scanned_period          = 12*1024,
1029     // the regular clock call is called once the number of visited
1030     // references reaches this limit
1031     refs_reached_period           = 384,
1032     // initial value for the hash seed, used in the work stealing code
1033     init_hash_seed                = 17,
1034     // how many entries will be transferred between global stack and
1035     // local queues
1036     global_stack_transfer_size    = 16
1037   };
1038 
1039   int                         _task_id;
1040   G1CollectedHeap*            _g1h;
1041   ConcurrentMark*             _cm;
1042   CMBitMap*                   _nextMarkBitMap;
1043   // the task queue of this task
1044   CMTaskQueue*                _task_queue;
1045 private:
1046   // the task queue set---needed for stealing
1047   CMTaskQueueSet*             _task_queues;
1048   // indicates whether the task has been claimed---this is only  for
1049   // debugging purposes
1050   bool                        _claimed;
1051 
1052   // number of calls to this task
1053   int                         _calls;
1054 
1055   // when the virtual timer reaches this time, the marking step should
1056   // exit
1057   double                      _time_target_ms;
1058   // the start time of the current marking step
1059   double                      _start_time_ms;
1060 
1061   // the oop closure used for iterations over oops
1062   G1CMOopClosure*             _cm_oop_closure;
1063 
1064   // the region this task is scanning, NULL if we're not scanning any
1065   HeapRegion*                 _curr_region;
1066   // the local finger of this task, NULL if we're not scanning a region
1067   HeapWord*                   _finger;
1068   // limit of the region this task is scanning, NULL if we're not scanning one
1069   HeapWord*                   _region_limit;
1070 
1071   // This is used only when we scan regions popped from the region
1072   // stack. It records what the last object on such a region we
1073   // scanned was. It is used to ensure that, if we abort region
1074   // iteration, we do not rescan the first part of the region. This
1075   // should be NULL when we're not scanning a region from the region
1076   // stack.
1077   HeapWord*                   _region_finger;
1078 
1079   // If we abort while scanning a region we record the remaining
1080   // unscanned portion and check this field when marking restarts.
1081   // This avoids having to push on the region stack while other
1082   // marking threads may still be popping regions.
1083   // If we were to push the unscanned portion directly to the
1084   // region stack then we would need to using locking versions
1085   // of the push and pop operations.
1086   MemRegion                   _aborted_region;
1087 
1088   // the number of words this task has scanned
1089   size_t                      _words_scanned;
1090   // When _words_scanned reaches this limit, the regular clock is
1091   // called. Notice that this might be decreased under certain
1092   // circumstances (i.e. when we believe that we did an expensive
1093   // operation).
1094   size_t                      _words_scanned_limit;
1095   // the initial value of _words_scanned_limit (i.e. what it was
1096   // before it was decreased).
1097   size_t                      _real_words_scanned_limit;
1098 
1099   // the number of references this task has visited
1100   size_t                      _refs_reached;
1101   // When _refs_reached reaches this limit, the regular clock is
1102   // called. Notice this this might be decreased under certain
1103   // circumstances (i.e. when we believe that we did an expensive
1104   // operation).
1105   size_t                      _refs_reached_limit;
1106   // the initial value of _refs_reached_limit (i.e. what it was before
1107   // it was decreased).
1108   size_t                      _real_refs_reached_limit;
1109 
1110   // used by the work stealing stuff
1111   int                         _hash_seed;
1112   // if this is true, then the task has aborted for some reason
1113   bool                        _has_aborted;
1114   // set when the task aborts because it has met its time quota
1115   bool                        _has_timed_out;
1116   // true when we're draining SATB buffers; this avoids the task
1117   // aborting due to SATB buffers being available (as we're already
1118   // dealing with them)
1119   bool                        _draining_satb_buffers;
1120 
1121   // number sequence of past step times
1122   NumberSeq                   _step_times_ms;
1123   // elapsed time of this task
1124   double                      _elapsed_time_ms;
1125   // termination time of this task
1126   double                      _termination_time_ms;
1127   // when this task got into the termination protocol
1128   double                      _termination_start_time_ms;
1129 
1130   // true when the task is during a concurrent phase, false when it is
1131   // in the remark phase (so, in the latter case, we do not have to
1132   // check all the things that we have to check during the concurrent
1133   // phase, i.e. SATB buffer availability...)
1134   bool                        _concurrent;
1135 
1136   TruncatedSeq                _marking_step_diffs_ms;
1137 
1138   // Counting data structures. Embedding the task's marked_bytes_array
1139   // and card bitmap into the actual task saves having to go through
1140   // the ConcurrentMark object.
1141   size_t*                     _marked_bytes_array;
1142   BitMap*                     _card_bm;
1143 
1144   // LOTS of statistics related with this task
1145 #if _MARKING_STATS_
1146   NumberSeq                   _all_clock_intervals_ms;
1147   double                      _interval_start_time_ms;
1148 
1149   int                         _aborted;
1150   int                         _aborted_overflow;
1151   int                         _aborted_cm_aborted;
1152   int                         _aborted_yield;
1153   int                         _aborted_timed_out;
1154   int                         _aborted_satb;
1155   int                         _aborted_termination;
1156 
1157   int                         _steal_attempts;
1158   int                         _steals;
1159 
1160   int                         _clock_due_to_marking;
1161   int                         _clock_due_to_scanning;
1162 
1163   int                         _local_pushes;
1164   int                         _local_pops;
1165   int                         _local_max_size;
1166   int                         _objs_scanned;
1167 
1168   int                         _global_pushes;
1169   int                         _global_pops;
1170   int                         _global_max_size;
1171 
1172   int                         _global_transfers_to;
1173   int                         _global_transfers_from;
1174 
1175   int                         _region_stack_pops;
1176 
1177   int                         _regions_claimed;
1178   int                         _objs_found_on_bitmap;
1179 
1180   int                         _satb_buffers_processed;
1181 #endif // _MARKING_STATS_
1182 
1183   // it updates the local fields after this task has claimed
1184   // a new region to scan
1185   void setup_for_region(HeapRegion* hr);
1186   // it brings up-to-date the limit of the region
1187   void update_region_limit();
1188 
1189   // called when either the words scanned or the refs visited limit
1190   // has been reached
1191   void reached_limit();
1192   // recalculates the words scanned and refs visited limits
1193   void recalculate_limits();
1194   // decreases the words scanned and refs visited limits when we reach
1195   // an expensive operation
1196   void decrease_limits();
1197   // it checks whether the words scanned or refs visited reached their
1198   // respective limit and calls reached_limit() if they have
1199   void check_limits() {
1200     if (_words_scanned >= _words_scanned_limit ||
1201         _refs_reached >= _refs_reached_limit) {
1202       reached_limit();
1203     }
1204   }
1205   // this is supposed to be called regularly during a marking step as
1206   // it checks a bunch of conditions that might cause the marking step
1207   // to abort
1208   void regular_clock_call();
1209   bool concurrent() { return _concurrent; }
1210 
1211 public:
1212   // It resets the task; it should be called right at the beginning of
1213   // a marking phase.
1214   void reset(CMBitMap* _nextMarkBitMap);
1215   // it clears all the fields that correspond to a claimed region.
1216   void clear_region_fields();
1217 
1218   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1219 
1220   // The main method of this class which performs a marking step
1221   // trying not to exceed the given duration. However, it might exit
1222   // prematurely, according to some conditions (i.e. SATB buffers are
1223   // available for processing).
1224   void do_marking_step(double target_ms, bool do_stealing, bool do_termination);
1225 
1226   // These two calls start and stop the timer
1227   void record_start_time() {
1228     _elapsed_time_ms = os::elapsedTime() * 1000.0;
1229   }
1230   void record_end_time() {
1231     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1232   }
1233 
1234   // returns the task ID
1235   int task_id() { return _task_id; }
1236 
1237   // From TerminatorTerminator. It determines whether this task should
1238   // exit the termination protocol after it's entered it.
1239   virtual bool should_exit_termination();
1240 
1241   // Resets the local region fields after a task has finished scanning a
1242   // region; or when they have become stale as a result of the region
1243   // being evacuated.
1244   void giveup_current_region();
1245 
1246   HeapWord* finger()            { return _finger; }
1247 
1248   bool has_aborted()            { return _has_aborted; }
1249   void set_has_aborted()        { _has_aborted = true; }
1250   void clear_has_aborted()      { _has_aborted = false; }
1251   bool has_timed_out()          { return _has_timed_out; }
1252   bool claimed()                { return _claimed; }
1253 
1254   // Support routines for the partially scanned region that may be
1255   // recorded as a result of aborting while draining the CMRegionStack
1256   MemRegion aborted_region()    { return _aborted_region; }
1257   void set_aborted_region(MemRegion mr)
1258                                 { _aborted_region = mr; }
1259 
1260   // Clears any recorded partially scanned region
1261   void clear_aborted_region()   { set_aborted_region(MemRegion()); }
1262 
1263   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1264 
1265   // It grays the object by marking it and, if necessary, pushing it
1266   // on the local queue
1267   inline void deal_with_reference(oop obj);
1268 
1269   // It scans an object and visits its children.
1270   void scan_object(oop obj);
1271 
1272   // It pushes an object on the local queue.
1273   inline void push(oop obj);
1274 
1275   // These two move entries to/from the global stack.
1276   void move_entries_to_global_stack();
1277   void get_entries_from_global_stack();
1278 
1279   // It pops and scans objects from the local queue. If partially is
1280   // true, then it stops when the queue size is of a given limit. If
1281   // partially is false, then it stops when the queue is empty.
1282   void drain_local_queue(bool partially);
1283   // It moves entries from the global stack to the local queue and
1284   // drains the local queue. If partially is true, then it stops when
1285   // both the global stack and the local queue reach a given size. If
1286   // partially if false, it tries to empty them totally.
1287   void drain_global_stack(bool partially);
1288   // It keeps picking SATB buffers and processing them until no SATB
1289   // buffers are available.
1290   void drain_satb_buffers();
1291   // It keeps popping regions from the region stack and processing
1292   // them until the region stack is empty.
1293   void drain_region_stack(BitMapClosure* closure);
1294 
1295   // moves the local finger to a new location
1296   inline void move_finger_to(HeapWord* new_finger) {
1297     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1298     _finger = new_finger;
1299   }
1300 
1301   // moves the region finger to a new location
1302   inline void move_region_finger_to(HeapWord* new_finger) {
1303     assert(new_finger < _cm->finger(), "invariant");
1304     _region_finger = new_finger;
1305   }
1306 
1307   CMTask(int task_num, ConcurrentMark *cm,
1308          size_t* marked_bytes, BitMap* card_bm,
1309          CMTaskQueue* task_queue, CMTaskQueueSet* task_queues);
1310 
1311   // it prints statistics associated with this task
1312   void print_stats();
1313 
1314 #if _MARKING_STATS_
1315   void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1316 #endif // _MARKING_STATS_
1317 };
1318 
1319 // Class that's used to to print out per-region liveness
1320 // information. It's currently used at the end of marking and also
1321 // after we sort the old regions at the end of the cleanup operation.
1322 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1323 private:
1324   outputStream* _out;
1325 
1326   // Accumulators for these values.
1327   size_t _total_used_bytes;
1328   size_t _total_capacity_bytes;
1329   size_t _total_prev_live_bytes;
1330   size_t _total_next_live_bytes;
1331 
1332   // These are set up when we come across a "stars humongous" region
1333   // (as this is where most of this information is stored, not in the
1334   // subsequent "continues humongous" regions). After that, for every
1335   // region in a given humongous region series we deduce the right
1336   // values for it by simply subtracting the appropriate amount from
1337   // these fields. All these values should reach 0 after we've visited
1338   // the last region in the series.
1339   size_t _hum_used_bytes;
1340   size_t _hum_capacity_bytes;
1341   size_t _hum_prev_live_bytes;
1342   size_t _hum_next_live_bytes;
1343 
1344   static double perc(size_t val, size_t total) {
1345     if (total == 0) {
1346       return 0.0;
1347     } else {
1348       return 100.0 * ((double) val / (double) total);
1349     }
1350   }
1351 
1352   static double bytes_to_mb(size_t val) {
1353     return (double) val / (double) M;
1354   }
1355 
1356   // See the .cpp file.
1357   size_t get_hum_bytes(size_t* hum_bytes);
1358   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1359                      size_t* prev_live_bytes, size_t* next_live_bytes);
1360 
1361 public:
1362   // The header and footer are printed in the constructor and
1363   // destructor respectively.
1364   G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1365   virtual bool doHeapRegion(HeapRegion* r);
1366   ~G1PrintRegionLivenessInfoClosure();
1367 };
1368 
1369 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_CONCURRENTMARK_HPP