1 /*
   2  * Copyright (c) 2001, 2016, 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_G1CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_G1CONCURRENTMARK_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 G1CMBitMap;
  35 class G1CMTask;
  36 class G1ConcurrentMark;
  37 typedef GenericTaskQueue<oop, mtGC>              G1CMTaskQueue;
  38 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
  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 G1CMBitMapRO 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   G1CMBitMapRO(int shifter);
  67 
  68   // inquiries
  69   HeapWord* startWord()   const { return _bmStartWord; }
  70   // the following is one past the last word in space
  71   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
  72 
  73   // read marks
  74 
  75   bool isMarked(HeapWord* addr) const {
  76     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
  77            "outside underlying space?");
  78     return _bm.at(heapWordToOffset(addr));
  79   }
  80 
  81   // iteration
  82   inline bool iterate(BitMapClosure* cl, MemRegion mr);
  83 
  84   // Return the address corresponding to the next marked bit at or after
  85   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  86   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  87   HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
  88                                      const HeapWord* limit = NULL) const;
  89 
  90   // conversion utilities
  91   HeapWord* offsetToHeapWord(size_t offset) const {
  92     return _bmStartWord + (offset << _shifter);
  93   }
  94   size_t heapWordToOffset(const HeapWord* addr) const {
  95     return pointer_delta(addr, _bmStartWord) >> _shifter;
  96   }
  97 
  98   // The argument addr should be the start address of a valid object
  99   inline HeapWord* nextObject(HeapWord* addr);
 100 
 101   void print_on_error(outputStream* st, const char* prefix) const;
 102 
 103   // debugging
 104   NOT_PRODUCT(bool covers(MemRegion rs) const;)
 105 };
 106 
 107 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
 108  private:
 109   G1CMBitMap* _bm;
 110  public:
 111   G1CMBitMapMappingChangedListener() : _bm(NULL) {}
 112 
 113   void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
 114 
 115   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 116 };
 117 
 118 class G1CMBitMap : public G1CMBitMapRO {
 119  private:
 120   G1CMBitMapMappingChangedListener _listener;
 121 
 122  public:
 123   static size_t compute_size(size_t heap_size);
 124   // Returns the amount of bytes on the heap between two marks in the bitmap.
 125   static size_t mark_distance();
 126   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 127   // mark bitmap corresponds to. This is the same as the mark distance above.
 128   static size_t heap_map_factor() {
 129     return mark_distance();
 130   }
 131 
 132   G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 133 
 134   // Initializes the underlying BitMap to cover the given area.
 135   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 136 
 137   // Write marks.
 138   inline void mark(HeapWord* addr);
 139   inline void clear(HeapWord* addr);
 140   inline bool parMark(HeapWord* addr);
 141 
 142   void clear_range(MemRegion mr);
 143 };
 144 
 145 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
 146 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
 147   VirtualSpace _virtual_space;   // Underlying backing store for actual stack
 148   G1ConcurrentMark* _cm;
 149   oop* _base;        // bottom of stack
 150   jint _index;       // one more than last occupied index
 151   jint _capacity;    // max #elements
 152   jint _saved_index; // value of _index saved at start of GC
 153 
 154   bool  _overflow;
 155   bool  _should_expand;
 156 
 157  public:
 158   G1CMMarkStack(G1ConcurrentMark* cm);
 159   ~G1CMMarkStack();
 160 
 161   bool allocate(size_t capacity);
 162 
 163   // Pushes the first "n" elements of "ptr_arr" on the stack.
 164   // Locking impl: concurrency is allowed only with
 165   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 166   // locking strategy.
 167   void par_push_arr(oop* ptr_arr, int n);
 168 
 169   // If returns false, the array was empty.  Otherwise, removes up to "max"
 170   // elements from the stack, and transfers them to "ptr_arr" in an
 171   // unspecified order.  The actual number transferred is given in "n" ("n
 172   // == 0" is deliberately redundant with the return value.)  Locking impl:
 173   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 174   // operations, which use the same locking strategy.
 175   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 176 
 177   bool isEmpty()    { return _index == 0; }
 178   int  maxElems()   { return _capacity; }
 179 
 180   bool overflow() { return _overflow; }
 181   void clear_overflow() { _overflow = false; }
 182 
 183   bool should_expand() const { return _should_expand; }
 184   void set_should_expand();
 185 
 186   // Expand the stack, typically in response to an overflow condition
 187   void expand();
 188 
 189   int  size() { return _index; }
 190 
 191   void setEmpty()   { _index = 0; clear_overflow(); }
 192 
 193   // Record the current index.
 194   void note_start_of_gc();
 195 
 196   // Make sure that we have not added any entries to the stack during GC.
 197   void note_end_of_gc();
 198 
 199   // Apply fn to each oop in the mark stack, up to the bound recorded
 200   // via one of the above "note" functions.  The mark stack must not
 201   // be modified while iterating.
 202   template<typename Fn> void iterate(Fn fn);
 203 };
 204 
 205 class YoungList;
 206 
 207 // Root Regions are regions that are not empty at the beginning of a
 208 // marking cycle and which we might collect during an evacuation pause
 209 // while the cycle is active. Given that, during evacuation pauses, we
 210 // do not copy objects that are explicitly marked, what we have to do
 211 // for the root regions is to scan them and mark all objects reachable
 212 // from them. According to the SATB assumptions, we only need to visit
 213 // each object once during marking. So, as long as we finish this scan
 214 // before the next evacuation pause, we can copy the objects from the
 215 // root regions without having to mark them or do anything else to them.
 216 //
 217 // Currently, we only support root region scanning once (at the start
 218 // of the marking cycle) and the root regions are all the survivor
 219 // regions populated during the initial-mark pause.
 220 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
 221 private:
 222   YoungList*           _young_list;
 223   G1ConcurrentMark*    _cm;
 224 
 225   volatile bool        _scan_in_progress;
 226   volatile bool        _should_abort;
 227   HeapRegion* volatile _next_survivor;
 228 
 229   void notify_scan_done();
 230 
 231 public:
 232   G1CMRootRegions();
 233   // We actually do most of the initialization in this method.
 234   void init(G1CollectedHeap* g1h, G1ConcurrentMark* cm);
 235 
 236   // Reset the claiming / scanning of the root regions.
 237   void prepare_for_scan();
 238 
 239   // Forces get_next() to return NULL so that the iteration aborts early.
 240   void abort() { _should_abort = true; }
 241 
 242   // Return true if the CM thread are actively scanning root regions,
 243   // false otherwise.
 244   bool scan_in_progress() { return _scan_in_progress; }
 245 
 246   // Claim the next root region to scan atomically, or return NULL if
 247   // all have been claimed.
 248   HeapRegion* claim_next();
 249 
 250   void cancel_scan();
 251 
 252   // Flag that we're done with root region scanning and notify anyone
 253   // who's waiting on it. If aborted is false, assume that all regions
 254   // have been claimed.
 255   void scan_finished();
 256 
 257   // If CM threads are still scanning root regions, wait until they
 258   // are done. Return true if we had to wait, false otherwise.
 259   bool wait_until_scan_finished();
 260 };
 261 
 262 class ConcurrentMarkThread;
 263 
 264 class G1ConcurrentMark: public CHeapObj<mtGC> {
 265   friend class ConcurrentMarkThread;
 266   friend class G1ParNoteEndTask;
 267   friend class CalcLiveObjectsClosure;
 268   friend class G1CMRefProcTaskProxy;
 269   friend class G1CMRefProcTaskExecutor;
 270   friend class G1CMKeepAliveAndDrainClosure;
 271   friend class G1CMDrainMarkingStackClosure;
 272   friend class G1CMBitMapClosure;
 273   friend class G1CMConcurrentMarkingTask;
 274   friend class G1CMMarkStack;
 275   friend class G1CMRemarkTask;
 276   friend class G1CMTask;
 277 
 278 protected:
 279   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 280   G1CollectedHeap*      _g1h;        // The heap
 281   uint                  _parallel_marking_threads; // The number of marking
 282                                                    // threads we're using
 283   uint                  _max_parallel_marking_threads; // Max number of marking
 284                                                        // threads we'll ever use
 285   double                _sleep_factor; // How much we have to sleep, with
 286                                        // respect to the work we just did, to
 287                                        // meet the marking overhead goal
 288   double                _marking_task_overhead; // Marking target overhead for
 289                                                 // a single task
 290 
 291   FreeRegionList        _cleanup_list;
 292 
 293   // Concurrent marking support structures
 294   G1CMBitMap              _markBitMap1;
 295   G1CMBitMap              _markBitMap2;
 296   G1CMBitMapRO*           _prevMarkBitMap; // Completed mark bitmap
 297   G1CMBitMap*             _nextMarkBitMap; // Under-construction mark bitmap
 298 
 299   BitMap                  _region_bm;
 300   BitMap                  _card_bm;
 301 
 302   // Heap bounds
 303   HeapWord*               _heap_start;
 304   HeapWord*               _heap_end;
 305 
 306   // Root region tracking and claiming
 307   G1CMRootRegions         _root_regions;
 308 
 309   // For gray objects
 310   G1CMMarkStack           _markStack; // Grey objects behind global finger
 311   HeapWord* volatile      _finger;  // The global finger, region aligned,
 312                                     // always points to the end of the
 313                                     // last claimed region
 314 
 315   // Marking tasks
 316   uint                    _max_worker_id;// Maximum worker id
 317   uint                    _active_tasks; // Task num currently active
 318   G1CMTask**              _tasks;        // Task queue array (max_worker_id len)
 319   G1CMTaskQueueSet*       _task_queues;  // Task queue set
 320   ParallelTaskTerminator  _terminator;   // For termination
 321 
 322   // Two sync barriers that are used to synchronize tasks when an
 323   // overflow occurs. The algorithm is the following. All tasks enter
 324   // the first one to ensure that they have all stopped manipulating
 325   // the global data structures. After they exit it, they re-initialize
 326   // their data structures and task 0 re-initializes the global data
 327   // structures. Then, they enter the second sync barrier. This
 328   // ensure, that no task starts doing work before all data
 329   // structures (local and global) have been re-initialized. When they
 330   // exit it, they are free to start working again.
 331   WorkGangBarrierSync     _first_overflow_barrier_sync;
 332   WorkGangBarrierSync     _second_overflow_barrier_sync;
 333 
 334   // This is set by any task, when an overflow on the global data
 335   // structures is detected
 336   volatile bool           _has_overflown;
 337   // True: marking is concurrent, false: we're in remark
 338   volatile bool           _concurrent;
 339   // Set at the end of a Full GC so that marking aborts
 340   volatile bool           _has_aborted;
 341 
 342   // Used when remark aborts due to an overflow to indicate that
 343   // another concurrent marking phase should start
 344   volatile bool           _restart_for_overflow;
 345 
 346   // This is true from the very start of concurrent marking until the
 347   // point when all the tasks complete their work. It is really used
 348   // to determine the points between the end of concurrent marking and
 349   // time of remark.
 350   volatile bool           _concurrent_marking_in_progress;
 351 
 352   // There would be a race between ConcurrentMarkThread and VMThread(ConcurrentMark::abort())
 353   // to call ConcurrentGCTimer::register_gc_concurrent_end().
 354   // And this variable is used to keep track of concurrent phase.
 355   volatile uint           _concurrent_phase_status;
 356   // Concurrent phase is not yet started.
 357   static const uint       ConcPhaseNotStarted = 0;
 358   // Concurrent phase is started.
 359   static const uint       ConcPhaseStarted = 1;
 360   // Caller thread of ConcurrentGCTimer::register_gc_concurrent_end() is ending concurrent phase.
 361   // So other thread should wait until the status to be changed to ConcPhaseNotStarted.
 362   static const uint       ConcPhaseStopping = 2;
 363 
 364   // All of these times are in ms
 365   NumberSeq _init_times;
 366   NumberSeq _remark_times;
 367   NumberSeq _remark_mark_times;
 368   NumberSeq _remark_weak_ref_times;
 369   NumberSeq _cleanup_times;
 370   double    _total_counting_time;
 371   double    _total_rs_scrub_time;
 372 
 373   double*   _accum_task_vtime;   // Accumulated task vtime
 374 
 375   WorkGang* _parallel_workers;
 376 
 377   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 378   void weakRefsWork(bool clear_all_soft_refs);
 379 
 380   void swapMarkBitMaps();
 381 
 382   // It resets the global marking data structures, as well as the
 383   // task local ones; should be called during initial mark.
 384   void reset();
 385 
 386   // Resets all the marking data structures. Called when we have to restart
 387   // marking or when marking completes (via set_non_marking_state below).
 388   void reset_marking_state(bool clear_overflow = true);
 389 
 390   // We do this after we're done with marking so that the marking data
 391   // structures are initialized to a sensible and predictable state.
 392   void set_non_marking_state();
 393 
 394   // Called to indicate how many threads are currently active.
 395   void set_concurrency(uint active_tasks);
 396 
 397   // It should be called to indicate which phase we're in (concurrent
 398   // mark or remark) and how many threads are currently active.
 399   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 400 
 401   // Prints all gathered CM-related statistics
 402   void print_stats();
 403 
 404   bool cleanup_list_is_empty() {
 405     return _cleanup_list.is_empty();
 406   }
 407 
 408   // Accessor methods
 409   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 410   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 411   double sleep_factor()                     { return _sleep_factor; }
 412   double marking_task_overhead()            { return _marking_task_overhead;}
 413 
 414   HeapWord*               finger()          { return _finger;   }
 415   bool                    concurrent()      { return _concurrent; }
 416   uint                    active_tasks()    { return _active_tasks; }
 417   ParallelTaskTerminator* terminator()      { return &_terminator; }
 418 
 419   // It claims the next available region to be scanned by a marking
 420   // task/thread. It might return NULL if the next region is empty or
 421   // we have run out of regions. In the latter case, out_of_regions()
 422   // determines whether we've really run out of regions or the task
 423   // should call claim_region() again. This might seem a bit
 424   // awkward. Originally, the code was written so that claim_region()
 425   // either successfully returned with a non-empty region or there
 426   // were no more regions to be claimed. The problem with this was
 427   // that, in certain circumstances, it iterated over large chunks of
 428   // the heap finding only empty regions and, while it was working, it
 429   // was preventing the calling task to call its regular clock
 430   // method. So, this way, each task will spend very little time in
 431   // claim_region() and is allowed to call the regular clock method
 432   // frequently.
 433   HeapRegion* claim_region(uint worker_id);
 434 
 435   // It determines whether we've run out of regions to scan. Note that
 436   // the finger can point past the heap end in case the heap was expanded
 437   // to satisfy an allocation without doing a GC. This is fine, because all
 438   // objects in those regions will be considered live anyway because of
 439   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 440   bool        out_of_regions() { return _finger >= _heap_end; }
 441 
 442   // Returns the task with the given id
 443   G1CMTask* task(int id) {
 444     assert(0 <= id && id < (int) _active_tasks,
 445            "task id not within active bounds");
 446     return _tasks[id];
 447   }
 448 
 449   // Returns the task queue with the given id
 450   G1CMTaskQueue* task_queue(int id) {
 451     assert(0 <= id && id < (int) _active_tasks,
 452            "task queue id not within active bounds");
 453     return (G1CMTaskQueue*) _task_queues->queue(id);
 454   }
 455 
 456   // Returns the task queue set
 457   G1CMTaskQueueSet* task_queues()  { return _task_queues; }
 458 
 459   // Access / manipulation of the overflow flag which is set to
 460   // indicate that the global stack has overflown
 461   bool has_overflown()           { return _has_overflown; }
 462   void set_has_overflown()       { _has_overflown = true; }
 463   void clear_has_overflown()     { _has_overflown = false; }
 464   bool restart_for_overflow()    { return _restart_for_overflow; }
 465 
 466   // Methods to enter the two overflow sync barriers
 467   void enter_first_sync_barrier(uint worker_id);
 468   void enter_second_sync_barrier(uint worker_id);
 469 
 470   // Live Data Counting data structures...
 471   // These data structures are initialized at the start of
 472   // marking. They are written to while marking is active.
 473   // They are aggregated during remark; the aggregated values
 474   // are then used to populate the _region_bm, _card_bm, and
 475   // the total live bytes, which are then subsequently updated
 476   // during cleanup.
 477 
 478   // An array of bitmaps (one bit map per task). Each bitmap
 479   // is used to record the cards spanned by the live objects
 480   // marked by that task/worker.
 481   BitMap*  _count_card_bitmaps;
 482 
 483   // Used to record the number of marked live bytes
 484   // (for each region, by worker thread).
 485   size_t** _count_marked_bytes;
 486 
 487   // Card index of the bottom of the G1 heap. Used for biasing indices into
 488   // the card bitmaps.
 489   intptr_t _heap_bottom_card_num;
 490 
 491   // Set to true when initialization is complete
 492   bool _completed_initialization;
 493 
 494   // end_timer, true to end gc timer after ending concurrent phase.
 495   void register_concurrent_phase_end_common(bool end_timer);
 496 
 497   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 498   // true, periodically insert checks to see if this method should exit prematurely.
 499   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 500 public:
 501   // Manipulation of the global mark stack.
 502   // The push and pop operations are used by tasks for transfers
 503   // between task-local queues and the global mark stack, and use
 504   // locking for concurrency safety.
 505   bool mark_stack_push(oop* arr, int n) {
 506     _markStack.par_push_arr(arr, n);
 507     if (_markStack.overflow()) {
 508       set_has_overflown();
 509       return false;
 510     }
 511     return true;
 512   }
 513   void mark_stack_pop(oop* arr, int max, int* n) {
 514     _markStack.par_pop_arr(arr, max, n);
 515   }
 516   size_t mark_stack_size()                { return _markStack.size(); }
 517   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 518   bool mark_stack_overflow()              { return _markStack.overflow(); }
 519   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 520 
 521   G1CMRootRegions* root_regions() { return &_root_regions; }
 522 
 523   bool concurrent_marking_in_progress() {
 524     return _concurrent_marking_in_progress;
 525   }
 526   void set_concurrent_marking_in_progress() {
 527     _concurrent_marking_in_progress = true;
 528   }
 529   void clear_concurrent_marking_in_progress() {
 530     _concurrent_marking_in_progress = false;
 531   }
 532 
 533   void register_concurrent_phase_start(const char* title);
 534   void register_concurrent_phase_end();
 535   // Ends both concurrent phase and timer.
 536   void register_concurrent_gc_end_and_stop_timer();
 537 
 538   void update_accum_task_vtime(int i, double vtime) {
 539     _accum_task_vtime[i] += vtime;
 540   }
 541 
 542   double all_task_accum_vtime() {
 543     double ret = 0.0;
 544     for (uint i = 0; i < _max_worker_id; ++i)
 545       ret += _accum_task_vtime[i];
 546     return ret;
 547   }
 548 
 549   // Attempts to steal an object from the task queues of other tasks
 550   bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
 551 
 552   G1ConcurrentMark(G1CollectedHeap* g1h,
 553                    G1RegionToSpaceMapper* prev_bitmap_storage,
 554                    G1RegionToSpaceMapper* next_bitmap_storage);
 555   ~G1ConcurrentMark();
 556 
 557   ConcurrentMarkThread* cmThread() { return _cmThread; }
 558 
 559   G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 560   G1CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 561 
 562   // Returns the number of GC threads to be used in a concurrent
 563   // phase based on the number of GC threads being used in a STW
 564   // phase.
 565   uint scale_parallel_threads(uint n_par_threads);
 566 
 567   // Calculates the number of GC threads to be used in a concurrent phase.
 568   uint calc_parallel_marking_threads();
 569 
 570   // The following three are interaction between CM and
 571   // G1CollectedHeap
 572 
 573   // This notifies CM that a root during initial-mark needs to be
 574   // grayed. It is MT-safe. word_size is the size of the object in
 575   // words. It is passed explicitly as sometimes we cannot calculate
 576   // it from the given object because it might be in an inconsistent
 577   // state (e.g., in to-space and being copied). So the caller is
 578   // responsible for dealing with this issue (e.g., get the size from
 579   // the from-space image when the to-space image might be
 580   // inconsistent) and always passing the size. hr is the region that
 581   // contains the object and it's passed optionally from callers who
 582   // might already have it (no point in recalculating it).
 583   inline void grayRoot(oop obj,
 584                        size_t word_size,
 585                        uint worker_id,
 586                        HeapRegion* hr = NULL);
 587 
 588   // Prepare internal data structures for the next mark cycle. This includes clearing
 589   // the next mark bitmap and some internal data structures. This method is intended
 590   // to be called concurrently to the mutator. It will yield to safepoint requests.
 591   void cleanup_for_next_mark();
 592 
 593   // Clear the previous marking bitmap during safepoint.
 594   void clear_prev_bitmap(WorkGang* workers);
 595 
 596   // Return whether the next mark bitmap has no marks set. To be used for assertions
 597   // only. Will not yield to pause requests.
 598   bool nextMarkBitmapIsClear();
 599 
 600   // These two do the work that needs to be done before and after the
 601   // initial root checkpoint. Since this checkpoint can be done at two
 602   // different points (i.e. an explicit pause or piggy-backed on a
 603   // young collection), then it's nice to be able to easily share the
 604   // pre/post code. It might be the case that we can put everything in
 605   // the post method. TP
 606   void checkpointRootsInitialPre();
 607   void checkpointRootsInitialPost();
 608 
 609   // Scan all the root regions and mark everything reachable from
 610   // them.
 611   void scan_root_regions();
 612 
 613   // Scan a single root region and mark everything reachable from it.
 614   void scanRootRegion(HeapRegion* hr, uint worker_id);
 615 
 616   // Do concurrent phase of marking, to a tentative transitive closure.
 617   void mark_from_roots();
 618 
 619   void checkpointRootsFinal(bool clear_all_soft_refs);
 620   void checkpointRootsFinalWork();
 621   void cleanup();
 622   void complete_cleanup();
 623 
 624   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 625   // this carefully!
 626   inline void markPrev(oop p);
 627 
 628   // Clears marks for all objects in the given range, for the prev or
 629   // next bitmaps.  NB: the previous bitmap is usually
 630   // read-only, so use this carefully!
 631   void clearRangePrevBitmap(MemRegion mr);
 632 
 633   // Notify data structures that a GC has started.
 634   void note_start_of_gc() {
 635     _markStack.note_start_of_gc();
 636   }
 637 
 638   // Notify data structures that a GC is finished.
 639   void note_end_of_gc() {
 640     _markStack.note_end_of_gc();
 641   }
 642 
 643   // Verify that there are no CSet oops on the stacks (taskqueues /
 644   // global mark stack) and fingers (global / per-task).
 645   // If marking is not in progress, it's a no-op.
 646   void verify_no_cset_oops() PRODUCT_RETURN;
 647 
 648   inline bool isPrevMarked(oop p) const;
 649 
 650   inline bool do_yield_check(uint worker_i = 0);
 651 
 652   // Called to abort the marking cycle after a Full GC takes place.
 653   void abort();
 654 
 655   bool has_aborted()      { return _has_aborted; }
 656 
 657   void print_summary_info();
 658 
 659   void print_worker_threads_on(outputStream* st) const;
 660 
 661   void print_on_error(outputStream* st) const;
 662 
 663   // Liveness counting
 664 
 665   // Utility routine to set an exclusive range of cards on the given
 666   // card liveness bitmap
 667   inline void set_card_bitmap_range(BitMap* card_bm,
 668                                     BitMap::idx_t start_idx,
 669                                     BitMap::idx_t end_idx,
 670                                     bool is_par);
 671 
 672   // Returns the card number of the bottom of the G1 heap.
 673   // Used in biasing indices into accounting card bitmaps.
 674   intptr_t heap_bottom_card_num() const {
 675     return _heap_bottom_card_num;
 676   }
 677 
 678   // Returns the card bitmap for a given task or worker id.
 679   BitMap* count_card_bitmap_for(uint worker_id) {
 680     assert(worker_id < _max_worker_id, "oob");
 681     assert(_count_card_bitmaps != NULL, "uninitialized");
 682     BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
 683     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 684     return task_card_bm;
 685   }
 686 
 687   // Returns the array containing the marked bytes for each region,
 688   // for the given worker or task id.
 689   size_t* count_marked_bytes_array_for(uint worker_id) {
 690     assert(worker_id < _max_worker_id, "oob");
 691     assert(_count_marked_bytes != NULL, "uninitialized");
 692     size_t* marked_bytes_array = _count_marked_bytes[worker_id];
 693     assert(marked_bytes_array != NULL, "uninitialized");
 694     return marked_bytes_array;
 695   }
 696 
 697   // Returns the index in the liveness accounting card table bitmap
 698   // for the given address
 699   inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
 700 
 701   // Counts the size of the given memory region in the the given
 702   // marked_bytes array slot for the given HeapRegion.
 703   // Sets the bits in the given card bitmap that are associated with the
 704   // cards that are spanned by the memory region.
 705   inline void count_region(MemRegion mr,
 706                            HeapRegion* hr,
 707                            size_t* marked_bytes_array,
 708                            BitMap* task_card_bm);
 709 
 710   // Counts the given object in the given task/worker counting
 711   // data structures.
 712   inline void count_object(oop obj,
 713                            HeapRegion* hr,
 714                            size_t* marked_bytes_array,
 715                            BitMap* task_card_bm,
 716                            size_t word_size);
 717 
 718   // Attempts to mark the given object and, if successful, counts
 719   // the object in the given task/worker counting structures.
 720   inline bool par_mark_and_count(oop obj,
 721                                  HeapRegion* hr,
 722                                  size_t* marked_bytes_array,
 723                                  BitMap* task_card_bm);
 724 
 725   // Attempts to mark the given object and, if successful, counts
 726   // the object in the task/worker counting structures for the
 727   // given worker id.
 728   inline bool par_mark_and_count(oop obj,
 729                                  size_t word_size,
 730                                  HeapRegion* hr,
 731                                  uint worker_id);
 732 
 733   // Returns true if initialization was successfully completed.
 734   bool completed_initialization() const {
 735     return _completed_initialization;
 736   }
 737 
 738 protected:
 739   // Clear all the per-task bitmaps and arrays used to store the
 740   // counting data.
 741   void clear_all_count_data();
 742 
 743   // Aggregates the counting data for each worker/task
 744   // that was constructed while marking. Also sets
 745   // the amount of marked bytes for each region and
 746   // the top at concurrent mark count.
 747   void aggregate_count_data();
 748 
 749   // Verification routine
 750   void verify_count_data();
 751 };
 752 
 753 // A class representing a marking task.
 754 class G1CMTask : public TerminatorTerminator {
 755 private:
 756   enum PrivateConstants {
 757     // the regular clock call is called once the scanned words reaches
 758     // this limit
 759     words_scanned_period          = 12*1024,
 760     // the regular clock call is called once the number of visited
 761     // references reaches this limit
 762     refs_reached_period           = 384,
 763     // initial value for the hash seed, used in the work stealing code
 764     init_hash_seed                = 17,
 765     // how many entries will be transferred between global stack and
 766     // local queues
 767     global_stack_transfer_size    = 16
 768   };
 769 
 770   uint                        _worker_id;
 771   G1CollectedHeap*            _g1h;
 772   G1ConcurrentMark*           _cm;
 773   G1CMBitMap*                 _nextMarkBitMap;
 774   // the task queue of this task
 775   G1CMTaskQueue*              _task_queue;
 776 private:
 777   // the task queue set---needed for stealing
 778   G1CMTaskQueueSet*           _task_queues;
 779   // indicates whether the task has been claimed---this is only  for
 780   // debugging purposes
 781   bool                        _claimed;
 782 
 783   // number of calls to this task
 784   int                         _calls;
 785 
 786   // when the virtual timer reaches this time, the marking step should
 787   // exit
 788   double                      _time_target_ms;
 789   // the start time of the current marking step
 790   double                      _start_time_ms;
 791 
 792   // the oop closure used for iterations over oops
 793   G1CMOopClosure*             _cm_oop_closure;
 794 
 795   // the region this task is scanning, NULL if we're not scanning any
 796   HeapRegion*                 _curr_region;
 797   // the local finger of this task, NULL if we're not scanning a region
 798   HeapWord*                   _finger;
 799   // limit of the region this task is scanning, NULL if we're not scanning one
 800   HeapWord*                   _region_limit;
 801 
 802   // the number of words this task has scanned
 803   size_t                      _words_scanned;
 804   // When _words_scanned reaches this limit, the regular clock is
 805   // called. Notice that this might be decreased under certain
 806   // circumstances (i.e. when we believe that we did an expensive
 807   // operation).
 808   size_t                      _words_scanned_limit;
 809   // the initial value of _words_scanned_limit (i.e. what it was
 810   // before it was decreased).
 811   size_t                      _real_words_scanned_limit;
 812 
 813   // the number of references this task has visited
 814   size_t                      _refs_reached;
 815   // When _refs_reached reaches this limit, the regular clock is
 816   // called. Notice this this might be decreased under certain
 817   // circumstances (i.e. when we believe that we did an expensive
 818   // operation).
 819   size_t                      _refs_reached_limit;
 820   // the initial value of _refs_reached_limit (i.e. what it was before
 821   // it was decreased).
 822   size_t                      _real_refs_reached_limit;
 823 
 824   // used by the work stealing stuff
 825   int                         _hash_seed;
 826   // if this is true, then the task has aborted for some reason
 827   bool                        _has_aborted;
 828   // set when the task aborts because it has met its time quota
 829   bool                        _has_timed_out;
 830   // true when we're draining SATB buffers; this avoids the task
 831   // aborting due to SATB buffers being available (as we're already
 832   // dealing with them)
 833   bool                        _draining_satb_buffers;
 834 
 835   // number sequence of past step times
 836   NumberSeq                   _step_times_ms;
 837   // elapsed time of this task
 838   double                      _elapsed_time_ms;
 839   // termination time of this task
 840   double                      _termination_time_ms;
 841   // when this task got into the termination protocol
 842   double                      _termination_start_time_ms;
 843 
 844   // true when the task is during a concurrent phase, false when it is
 845   // in the remark phase (so, in the latter case, we do not have to
 846   // check all the things that we have to check during the concurrent
 847   // phase, i.e. SATB buffer availability...)
 848   bool                        _concurrent;
 849 
 850   TruncatedSeq                _marking_step_diffs_ms;
 851 
 852   // Counting data structures. Embedding the task's marked_bytes_array
 853   // and card bitmap into the actual task saves having to go through
 854   // the ConcurrentMark object.
 855   size_t*                     _marked_bytes_array;
 856   BitMap*                     _card_bm;
 857 
 858   // it updates the local fields after this task has claimed
 859   // a new region to scan
 860   void setup_for_region(HeapRegion* hr);
 861   // it brings up-to-date the limit of the region
 862   void update_region_limit();
 863 
 864   // called when either the words scanned or the refs visited limit
 865   // has been reached
 866   void reached_limit();
 867   // recalculates the words scanned and refs visited limits
 868   void recalculate_limits();
 869   // decreases the words scanned and refs visited limits when we reach
 870   // an expensive operation
 871   void decrease_limits();
 872   // it checks whether the words scanned or refs visited reached their
 873   // respective limit and calls reached_limit() if they have
 874   void check_limits() {
 875     if (_words_scanned >= _words_scanned_limit ||
 876         _refs_reached >= _refs_reached_limit) {
 877       reached_limit();
 878     }
 879   }
 880   // this is supposed to be called regularly during a marking step as
 881   // it checks a bunch of conditions that might cause the marking step
 882   // to abort
 883   void regular_clock_call();
 884   bool concurrent() { return _concurrent; }
 885 
 886   // Test whether obj might have already been passed over by the
 887   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 888   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 889 
 890   template<bool scan> void process_grey_object(oop obj);
 891 
 892 public:
 893   // It resets the task; it should be called right at the beginning of
 894   // a marking phase.
 895   void reset(G1CMBitMap* _nextMarkBitMap);
 896   // it clears all the fields that correspond to a claimed region.
 897   void clear_region_fields();
 898 
 899   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
 900 
 901   // The main method of this class which performs a marking step
 902   // trying not to exceed the given duration. However, it might exit
 903   // prematurely, according to some conditions (i.e. SATB buffers are
 904   // available for processing).
 905   void do_marking_step(double target_ms,
 906                        bool do_termination,
 907                        bool is_serial);
 908 
 909   // These two calls start and stop the timer
 910   void record_start_time() {
 911     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 912   }
 913   void record_end_time() {
 914     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 915   }
 916 
 917   // returns the worker ID associated with this task.
 918   uint worker_id() { return _worker_id; }
 919 
 920   // From TerminatorTerminator. It determines whether this task should
 921   // exit the termination protocol after it's entered it.
 922   virtual bool should_exit_termination();
 923 
 924   // Resets the local region fields after a task has finished scanning a
 925   // region; or when they have become stale as a result of the region
 926   // being evacuated.
 927   void giveup_current_region();
 928 
 929   HeapWord* finger()            { return _finger; }
 930 
 931   bool has_aborted()            { return _has_aborted; }
 932   void set_has_aborted()        { _has_aborted = true; }
 933   void clear_has_aborted()      { _has_aborted = false; }
 934   bool has_timed_out()          { return _has_timed_out; }
 935   bool claimed()                { return _claimed; }
 936 
 937   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 938 
 939   // Increment the number of references this task has visited.
 940   void increment_refs_reached() { ++_refs_reached; }
 941 
 942   // Grey the object by marking it.  If not already marked, push it on
 943   // the local queue if below the finger.
 944   // Precondition: obj is in region.
 945   // Precondition: obj is below region's NTAMS.
 946   inline void make_reference_grey(oop obj, HeapRegion* region);
 947 
 948   // Grey the object (by calling make_grey_reference) if required,
 949   // e.g. obj is below its containing region's NTAMS.
 950   // Precondition: obj is a valid heap object.
 951   inline void deal_with_reference(oop obj);
 952 
 953   // It scans an object and visits its children.
 954   inline void scan_object(oop obj);
 955 
 956   // It pushes an object on the local queue.
 957   inline void push(oop obj);
 958 
 959   // These two move entries to/from the global stack.
 960   void move_entries_to_global_stack();
 961   void get_entries_from_global_stack();
 962 
 963   // It pops and scans objects from the local queue. If partially is
 964   // true, then it stops when the queue size is of a given limit. If
 965   // partially is false, then it stops when the queue is empty.
 966   void drain_local_queue(bool partially);
 967   // It moves entries from the global stack to the local queue and
 968   // drains the local queue. If partially is true, then it stops when
 969   // both the global stack and the local queue reach a given size. If
 970   // partially if false, it tries to empty them totally.
 971   void drain_global_stack(bool partially);
 972   // It keeps picking SATB buffers and processing them until no SATB
 973   // buffers are available.
 974   void drain_satb_buffers();
 975 
 976   // moves the local finger to a new location
 977   inline void move_finger_to(HeapWord* new_finger) {
 978     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 979     _finger = new_finger;
 980   }
 981 
 982   G1CMTask(uint worker_id,
 983            G1ConcurrentMark *cm,
 984            size_t* marked_bytes,
 985            BitMap* card_bm,
 986            G1CMTaskQueue* task_queue,
 987            G1CMTaskQueueSet* task_queues);
 988 
 989   // it prints statistics associated with this task
 990   void print_stats();
 991 };
 992 
 993 // Class that's used to to print out per-region liveness
 994 // information. It's currently used at the end of marking and also
 995 // after we sort the old regions at the end of the cleanup operation.
 996 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
 997 private:
 998   // Accumulators for these values.
 999   size_t _total_used_bytes;
1000   size_t _total_capacity_bytes;
1001   size_t _total_prev_live_bytes;
1002   size_t _total_next_live_bytes;
1003 
1004   // These are set up when we come across a "stars humongous" region
1005   // (as this is where most of this information is stored, not in the
1006   // subsequent "continues humongous" regions). After that, for every
1007   // region in a given humongous region series we deduce the right
1008   // values for it by simply subtracting the appropriate amount from
1009   // these fields. All these values should reach 0 after we've visited
1010   // the last region in the series.
1011   size_t _hum_used_bytes;
1012   size_t _hum_capacity_bytes;
1013   size_t _hum_prev_live_bytes;
1014   size_t _hum_next_live_bytes;
1015 
1016   // Accumulator for the remembered set size
1017   size_t _total_remset_bytes;
1018 
1019   // Accumulator for strong code roots memory size
1020   size_t _total_strong_code_roots_bytes;
1021 
1022   static double perc(size_t val, size_t total) {
1023     if (total == 0) {
1024       return 0.0;
1025     } else {
1026       return 100.0 * ((double) val / (double) total);
1027     }
1028   }
1029 
1030   static double bytes_to_mb(size_t val) {
1031     return (double) val / (double) M;
1032   }
1033 
1034   // See the .cpp file.
1035   size_t get_hum_bytes(size_t* hum_bytes);
1036   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1037                      size_t* prev_live_bytes, size_t* next_live_bytes);
1038 
1039 public:
1040   // The header and footer are printed in the constructor and
1041   // destructor respectively.
1042   G1PrintRegionLivenessInfoClosure(const char* phase_name);
1043   virtual bool doHeapRegion(HeapRegion* r);
1044   ~G1PrintRegionLivenessInfoClosure();
1045 };
1046 
1047 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP