1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_GC_G1_CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_CONCURRENTMARK_HPP
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/g1/g1RegionToSpaceMapper.hpp"
  30 #include "gc/g1/heapRegionSet.hpp"
  31 #include "gc/shared/taskqueue.hpp"
  32 
  33 class G1CollectedHeap;
  34 class CMBitMap;
  35 class CMTask;
  36 class ConcurrentMark;
  37 typedef GenericTaskQueue<oop, mtGC>            CMTaskQueue;
  38 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
  39 
  40 // Closure used by CM during concurrent reference discovery
  41 // and reference processing (during remarking) to determine
  42 // if a particular object is alive. It is primarily used
  43 // to determine if referents of discovered reference objects
  44 // are alive. An instance is also embedded into the
  45 // reference processor as the _is_alive_non_header field
  46 class G1CMIsAliveClosure: public BoolObjectClosure {
  47   G1CollectedHeap* _g1;
  48  public:
  49   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
  50 
  51   bool do_object_b(oop obj);
  52 };
  53 
  54 // A generic CM bit map.  This is essentially a wrapper around the BitMap
  55 // class, with one bit per (1<<_shifter) HeapWords.
  56 
  57 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
  58  protected:
  59   HeapWord* _bmStartWord;      // base address of range covered by map
  60   size_t    _bmWordSize;       // map size (in #HeapWords covered)
  61   const int _shifter;          // map to char or bit
  62   BitMap    _bm;               // the bit map itself
  63 
  64  public:
  65   // constructor
  66   CMBitMapRO(int shifter);
  67 
  68   // 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   HeapWord* nextObject(HeapWord* addr) {
 100     oop obj = (oop) addr;
 101     HeapWord* res =  addr + obj->size();
 102     assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
 103     return res;
 104   }
 105 
 106   void print_on_error(outputStream* st, const char* prefix) const;
 107 
 108   // debugging
 109   NOT_PRODUCT(bool covers(MemRegion rs) const;)
 110 };
 111 
 112 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
 113  private:
 114   CMBitMap* _bm;
 115  public:
 116   CMBitMapMappingChangedListener() : _bm(NULL) {}
 117 
 118   void set_bitmap(CMBitMap* bm) { _bm = bm; }
 119 
 120   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 121 };
 122 
 123 class CMBitMap : public CMBitMapRO {
 124  private:
 125   CMBitMapMappingChangedListener _listener;
 126 
 127  public:
 128   static size_t compute_size(size_t heap_size);
 129   // Returns the amount of bytes on the heap between two marks in the bitmap.
 130   static size_t mark_distance();
 131   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 132   // mark bitmap corresponds to. This is the same as the mark distance above.
 133   static size_t heap_map_factor() {
 134     return mark_distance();
 135   }
 136 
 137   CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 138 
 139   // Initializes the underlying BitMap to cover the given area.
 140   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 141 
 142   // Write marks.
 143   inline void mark(HeapWord* addr);
 144   inline void clear(HeapWord* addr);
 145   inline bool parMark(HeapWord* addr);
 146 
 147   void clearRange(MemRegion mr);
 148 
 149   // Clear the whole mark bitmap.
 150   void clearAll();
 151 };
 152 
 153 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
 154 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
 155   VirtualSpace _virtual_space;   // Underlying backing store for actual stack
 156   ConcurrentMark* _cm;
 157   oop* _base;        // bottom of stack
 158   jint _index;       // one more than last occupied index
 159   jint _capacity;    // max #elements
 160   jint _saved_index; // value of _index saved at start of GC
 161 
 162   bool  _overflow;
 163   bool  _should_expand;
 164 
 165  public:
 166   CMMarkStack(ConcurrentMark* cm);
 167   ~CMMarkStack();
 168 
 169   bool allocate(size_t capacity);
 170 
 171   // Pushes the first "n" elements of "ptr_arr" on the stack.
 172   // Locking impl: concurrency is allowed only with
 173   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 174   // locking strategy.
 175   void par_push_arr(oop* ptr_arr, int n);
 176 
 177   // If returns false, the array was empty.  Otherwise, removes up to "max"
 178   // elements from the stack, and transfers them to "ptr_arr" in an
 179   // unspecified order.  The actual number transferred is given in "n" ("n
 180   // == 0" is deliberately redundant with the return value.)  Locking impl:
 181   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 182   // operations, which use the same locking strategy.
 183   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 184 
 185   bool isEmpty()    { return _index == 0; }
 186   int  maxElems()   { return _capacity; }
 187 
 188   bool overflow() { return _overflow; }
 189   void clear_overflow() { _overflow = false; }
 190 
 191   bool should_expand() const { return _should_expand; }
 192   void set_should_expand();
 193 
 194   // Expand the stack, typically in response to an overflow condition
 195   void expand();
 196 
 197   int  size() { return _index; }
 198 
 199   void setEmpty()   { _index = 0; clear_overflow(); }
 200 
 201   // Record the current index.
 202   void note_start_of_gc();
 203 
 204   // Make sure that we have not added any entries to the stack during GC.
 205   void note_end_of_gc();
 206 
 207   // Apply fn to each oop in the mark stack, up to the bound recorded
 208   // via one of the above "note" functions.  The mark stack must not
 209   // be modified while iterating.
 210   template<typename Fn> void iterate(Fn fn);
 211 };
 212 
 213 class YoungList;
 214 
 215 // Root Regions are regions that are not empty at the beginning of a
 216 // marking cycle and which we might collect during an evacuation pause
 217 // while the cycle is active. Given that, during evacuation pauses, we
 218 // do not copy objects that are explicitly marked, what we have to do
 219 // for the root regions is to scan them and mark all objects reachable
 220 // from them. According to the SATB assumptions, we only need to visit
 221 // each object once during marking. So, as long as we finish this scan
 222 // before the next evacuation pause, we can copy the objects from the
 223 // root regions without having to mark them or do anything else to them.
 224 //
 225 // Currently, we only support root region scanning once (at the start
 226 // of the marking cycle) and the root regions are all the survivor
 227 // regions populated during the initial-mark pause.
 228 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
 229 private:
 230   YoungList*           _young_list;
 231   ConcurrentMark*      _cm;
 232 
 233   volatile bool        _scan_in_progress;
 234   volatile bool        _should_abort;
 235   HeapRegion* volatile _next_survivor;
 236 
 237 public:
 238   CMRootRegions();
 239   // We actually do most of the initialization in this method.
 240   void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
 241 
 242   // Reset the claiming / scanning of the root regions.
 243   void prepare_for_scan();
 244 
 245   // Forces get_next() to return NULL so that the iteration aborts early.
 246   void abort() { _should_abort = true; }
 247 
 248   // Return true if the CM thread are actively scanning root regions,
 249   // false otherwise.
 250   bool scan_in_progress() { return _scan_in_progress; }
 251 
 252   // Claim the next root region to scan atomically, or return NULL if
 253   // all have been claimed.
 254   HeapRegion* claim_next();
 255 
 256   // Flag that we're done with root region scanning and notify anyone
 257   // who's waiting on it. If aborted is false, assume that all regions
 258   // have been claimed.
 259   void scan_finished();
 260 
 261   // If CM threads are still scanning root regions, wait until they
 262   // are done. Return true if we had to wait, false otherwise.
 263   bool wait_until_scan_finished();
 264 };
 265 
 266 class ConcurrentMarkThread;
 267 
 268 class ConcurrentMark: public CHeapObj<mtGC> {
 269   friend class CMMarkStack;
 270   friend class ConcurrentMarkThread;
 271   friend class CMTask;
 272   friend class CMBitMapClosure;
 273   friend class CMRemarkTask;
 274   friend class CMConcurrentMarkingTask;
 275   friend class G1ParNoteEndTask;
 276   friend class CalcLiveObjectsClosure;
 277   friend class G1CMRefProcTaskProxy;
 278   friend class G1CMRefProcTaskExecutor;
 279   friend class G1CMKeepAliveAndDrainClosure;
 280   friend class G1CMDrainMarkingStackClosure;
 281 
 282 protected:
 283   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 284   G1CollectedHeap*      _g1h;        // The heap
 285   uint                  _parallel_marking_threads; // The number of marking
 286                                                    // threads we're using
 287   uint                  _max_parallel_marking_threads; // Max number of marking
 288                                                        // threads we'll ever use
 289   double                _sleep_factor; // How much we have to sleep, with
 290                                        // respect to the work we just did, to
 291                                        // meet the marking overhead goal
 292   double                _marking_task_overhead; // Marking target overhead for
 293                                                 // a single task
 294 
 295   FreeRegionList        _cleanup_list;
 296 
 297   // Concurrent marking support structures
 298   CMBitMap                _markBitMap1;
 299   CMBitMap                _markBitMap2;
 300   CMBitMapRO*             _prevMarkBitMap; // Completed mark bitmap
 301   CMBitMap*               _nextMarkBitMap; // Under-construction mark bitmap
 302 
 303   BitMap                  _region_bm;
 304   BitMap                  _card_bm;
 305 
 306   // Heap bounds
 307   HeapWord*               _heap_start;
 308   HeapWord*               _heap_end;
 309 
 310   // Root region tracking and claiming
 311   CMRootRegions           _root_regions;
 312 
 313   // For gray objects
 314   CMMarkStack             _markStack; // Grey objects behind global finger
 315   HeapWord* volatile      _finger;  // The global finger, region aligned,
 316                                     // always points to the end of the
 317                                     // last claimed region
 318 
 319   // Marking tasks
 320   uint                    _max_worker_id;// Maximum worker id
 321   uint                    _active_tasks; // Task num currently active
 322   CMTask**                _tasks;        // Task queue array (max_worker_id len)
 323   CMTaskQueueSet*         _task_queues;  // Task queue set
 324   ParallelTaskTerminator  _terminator;   // For termination
 325 
 326   // Two sync barriers that are used to synchronize tasks when an
 327   // overflow occurs. The algorithm is the following. All tasks enter
 328   // the first one to ensure that they have all stopped manipulating
 329   // the global data structures. After they exit it, they re-initialize
 330   // their data structures and task 0 re-initializes the global data
 331   // structures. Then, they enter the second sync barrier. This
 332   // ensure, that no task starts doing work before all data
 333   // structures (local and global) have been re-initialized. When they
 334   // exit it, they are free to start working again.
 335   WorkGangBarrierSync     _first_overflow_barrier_sync;
 336   WorkGangBarrierSync     _second_overflow_barrier_sync;
 337 
 338   // This is set by any task, when an overflow on the global data
 339   // structures is detected
 340   volatile bool           _has_overflown;
 341   // True: marking is concurrent, false: we're in remark
 342   volatile bool           _concurrent;
 343   // Set at the end of a Full GC so that marking aborts
 344   volatile bool           _has_aborted;
 345 
 346   // Used when remark aborts due to an overflow to indicate that
 347   // another concurrent marking phase should start
 348   volatile bool           _restart_for_overflow;
 349 
 350   // This is true from the very start of concurrent marking until the
 351   // point when all the tasks complete their work. It is really used
 352   // to determine the points between the end of concurrent marking and
 353   // time of remark.
 354   volatile bool           _concurrent_marking_in_progress;
 355 
 356   // All of these times are in ms
 357   NumberSeq _init_times;
 358   NumberSeq _remark_times;
 359   NumberSeq _remark_mark_times;
 360   NumberSeq _remark_weak_ref_times;
 361   NumberSeq _cleanup_times;
 362   double    _total_counting_time;
 363   double    _total_rs_scrub_time;
 364 
 365   double*   _accum_task_vtime;   // Accumulated task vtime
 366 
 367   WorkGang* _parallel_workers;
 368 
 369   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 370   void weakRefsWork(bool clear_all_soft_refs);
 371 
 372   void swapMarkBitMaps();
 373 
 374   // It resets the global marking data structures, as well as the
 375   // task local ones; should be called during initial mark.
 376   void reset();
 377 
 378   // Resets all the marking data structures. Called when we have to restart
 379   // marking or when marking completes (via set_non_marking_state below).
 380   void reset_marking_state(bool clear_overflow = true);
 381 
 382   // We do this after we're done with marking so that the marking data
 383   // structures are initialized to a sensible and predictable state.
 384   void set_non_marking_state();
 385 
 386   // Called to indicate how many threads are currently active.
 387   void set_concurrency(uint active_tasks);
 388 
 389   // It should be called to indicate which phase we're in (concurrent
 390   // mark or remark) and how many threads are currently active.
 391   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 392 
 393   // Prints all gathered CM-related statistics
 394   void print_stats();
 395 
 396   bool cleanup_list_is_empty() {
 397     return _cleanup_list.is_empty();
 398   }
 399 
 400   // Accessor methods
 401   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 402   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 403   double sleep_factor()                     { return _sleep_factor; }
 404   double marking_task_overhead()            { return _marking_task_overhead;}
 405 
 406   HeapWord*               finger()          { return _finger;   }
 407   bool                    concurrent()      { return _concurrent; }
 408   uint                    active_tasks()    { return _active_tasks; }
 409   ParallelTaskTerminator* terminator()      { return &_terminator; }
 410 
 411   // It claims the next available region to be scanned by a marking
 412   // task/thread. It might return NULL if the next region is empty or
 413   // we have run out of regions. In the latter case, out_of_regions()
 414   // determines whether we've really run out of regions or the task
 415   // should call claim_region() again. This might seem a bit
 416   // awkward. Originally, the code was written so that claim_region()
 417   // either successfully returned with a non-empty region or there
 418   // were no more regions to be claimed. The problem with this was
 419   // that, in certain circumstances, it iterated over large chunks of
 420   // the heap finding only empty regions and, while it was working, it
 421   // was preventing the calling task to call its regular clock
 422   // method. So, this way, each task will spend very little time in
 423   // claim_region() and is allowed to call the regular clock method
 424   // frequently.
 425   HeapRegion* claim_region(uint worker_id);
 426 
 427   // It determines whether we've run out of regions to scan. Note that
 428   // the finger can point past the heap end in case the heap was expanded
 429   // to satisfy an allocation without doing a GC. This is fine, because all
 430   // objects in those regions will be considered live anyway because of
 431   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 432   bool        out_of_regions() { return _finger >= _heap_end; }
 433 
 434   // Returns the task with the given id
 435   CMTask* task(int id) {
 436     assert(0 <= id && id < (int) _active_tasks,
 437            "task id not within active bounds");
 438     return _tasks[id];
 439   }
 440 
 441   // Returns the task queue with the given id
 442   CMTaskQueue* task_queue(int id) {
 443     assert(0 <= id && id < (int) _active_tasks,
 444            "task queue id not within active bounds");
 445     return (CMTaskQueue*) _task_queues->queue(id);
 446   }
 447 
 448   // Returns the task queue set
 449   CMTaskQueueSet* task_queues()  { return _task_queues; }
 450 
 451   // Access / manipulation of the overflow flag which is set to
 452   // indicate that the global stack has overflown
 453   bool has_overflown()           { return _has_overflown; }
 454   void set_has_overflown()       { _has_overflown = true; }
 455   void clear_has_overflown()     { _has_overflown = false; }
 456   bool restart_for_overflow()    { return _restart_for_overflow; }
 457 
 458   // Methods to enter the two overflow sync barriers
 459   void enter_first_sync_barrier(uint worker_id);
 460   void enter_second_sync_barrier(uint worker_id);
 461 
 462   // Live Data Counting data structures...
 463   // These data structures are initialized at the start of
 464   // marking. They are written to while marking is active.
 465   // They are aggregated during remark; the aggregated values
 466   // are then used to populate the _region_bm, _card_bm, and
 467   // the total live bytes, which are then subsequently updated
 468   // during cleanup.
 469 
 470   // An array of bitmaps (one bit map per task). Each bitmap
 471   // is used to record the cards spanned by the live objects
 472   // marked by that task/worker.
 473   BitMap*  _count_card_bitmaps;
 474 
 475   // Used to record the number of marked live bytes
 476   // (for each region, by worker thread).
 477   size_t** _count_marked_bytes;
 478 
 479   // Card index of the bottom of the G1 heap. Used for biasing indices into
 480   // the card bitmaps.
 481   intptr_t _heap_bottom_card_num;
 482 
 483   // Set to true when initialization is complete
 484   bool _completed_initialization;
 485 
 486 public:
 487   // Manipulation of the global mark stack.
 488   // The push and pop operations are used by tasks for transfers
 489   // between task-local queues and the global mark stack, and use
 490   // locking for concurrency safety.
 491   bool mark_stack_push(oop* arr, int n) {
 492     _markStack.par_push_arr(arr, n);
 493     if (_markStack.overflow()) {
 494       set_has_overflown();
 495       return false;
 496     }
 497     return true;
 498   }
 499   void mark_stack_pop(oop* arr, int max, int* n) {
 500     _markStack.par_pop_arr(arr, max, n);
 501   }
 502   size_t mark_stack_size()                { return _markStack.size(); }
 503   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 504   bool mark_stack_overflow()              { return _markStack.overflow(); }
 505   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 506 
 507   CMRootRegions* root_regions() { return &_root_regions; }
 508 
 509   bool concurrent_marking_in_progress() {
 510     return _concurrent_marking_in_progress;
 511   }
 512   void set_concurrent_marking_in_progress() {
 513     _concurrent_marking_in_progress = true;
 514   }
 515   void clear_concurrent_marking_in_progress() {
 516     _concurrent_marking_in_progress = false;
 517   }
 518 
 519   void update_accum_task_vtime(int i, double vtime) {
 520     _accum_task_vtime[i] += vtime;
 521   }
 522 
 523   double all_task_accum_vtime() {
 524     double ret = 0.0;
 525     for (uint i = 0; i < _max_worker_id; ++i)
 526       ret += _accum_task_vtime[i];
 527     return ret;
 528   }
 529 
 530   // Attempts to steal an object from the task queues of other tasks
 531   bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
 532 
 533   ConcurrentMark(G1CollectedHeap* g1h,
 534                  G1RegionToSpaceMapper* prev_bitmap_storage,
 535                  G1RegionToSpaceMapper* next_bitmap_storage);
 536   ~ConcurrentMark();
 537 
 538   ConcurrentMarkThread* cmThread() { return _cmThread; }
 539 
 540   CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 541   CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 542 
 543   // Returns the number of GC threads to be used in a concurrent
 544   // phase based on the number of GC threads being used in a STW
 545   // phase.
 546   uint scale_parallel_threads(uint n_par_threads);
 547 
 548   // Calculates the number of GC threads to be used in a concurrent phase.
 549   uint calc_parallel_marking_threads();
 550 
 551   // The following three are interaction between CM and
 552   // G1CollectedHeap
 553 
 554   // This notifies CM that a root during initial-mark needs to be
 555   // grayed. It is MT-safe. word_size is the size of the object in
 556   // words. It is passed explicitly as sometimes we cannot calculate
 557   // it from the given object because it might be in an inconsistent
 558   // state (e.g., in to-space and being copied). So the caller is
 559   // responsible for dealing with this issue (e.g., get the size from
 560   // the from-space image when the to-space image might be
 561   // inconsistent) and always passing the size. hr is the region that
 562   // contains the object and it's passed optionally from callers who
 563   // might already have it (no point in recalculating it).
 564   inline void grayRoot(oop obj,
 565                        size_t word_size,
 566                        uint worker_id,
 567                        HeapRegion* hr = NULL);
 568 
 569   // Clear the next marking bitmap (will be called concurrently).
 570   void clearNextBitmap();
 571 
 572   // Return whether the next mark bitmap has no marks set. To be used for assertions
 573   // only. Will not yield to pause requests.
 574   bool nextMarkBitmapIsClear();
 575 
 576   // These two do the work that needs to be done before and after the
 577   // initial root checkpoint. Since this checkpoint can be done at two
 578   // different points (i.e. an explicit pause or piggy-backed on a
 579   // young collection), then it's nice to be able to easily share the
 580   // pre/post code. It might be the case that we can put everything in
 581   // the post method. TP
 582   void checkpointRootsInitialPre();
 583   void checkpointRootsInitialPost();
 584 
 585   // Scan all the root regions and mark everything reachable from
 586   // them.
 587   void scanRootRegions();
 588 
 589   // Scan a single root region and mark everything reachable from it.
 590   void scanRootRegion(HeapRegion* hr, uint worker_id);
 591 
 592   // Do concurrent phase of marking, to a tentative transitive closure.
 593   void markFromRoots();
 594 
 595   void checkpointRootsFinal(bool clear_all_soft_refs);
 596   void checkpointRootsFinalWork();
 597   void cleanup();
 598   void completeCleanup();
 599 
 600   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 601   // this carefully!
 602   inline void markPrev(oop p);
 603 
 604   // Clears marks for all objects in the given range, for the prev or
 605   // next bitmaps.  NB: the previous bitmap is usually
 606   // read-only, so use this carefully!
 607   void clearRangePrevBitmap(MemRegion mr);
 608 
 609   // Notify data structures that a GC has started.
 610   void note_start_of_gc() {
 611     _markStack.note_start_of_gc();
 612   }
 613 
 614   // Notify data structures that a GC is finished.
 615   void note_end_of_gc() {
 616     _markStack.note_end_of_gc();
 617   }
 618 
 619   // Verify that there are no CSet oops on the stacks (taskqueues /
 620   // global mark stack) and fingers (global / per-task).
 621   // If marking is not in progress, it's a no-op.
 622   void verify_no_cset_oops() PRODUCT_RETURN;
 623 
 624   bool isPrevMarked(oop p) const {
 625     assert(p != NULL && p->is_oop(), "expected an oop");
 626     HeapWord* addr = (HeapWord*)p;
 627     assert(addr >= _prevMarkBitMap->startWord() ||
 628            addr < _prevMarkBitMap->endWord(), "in a region");
 629 
 630     return _prevMarkBitMap->isMarked(addr);
 631   }
 632 
 633   inline bool do_yield_check(uint worker_i = 0);
 634 
 635   // Called to abort the marking cycle after a Full GC takes place.
 636   void abort();
 637 
 638   bool has_aborted()      { return _has_aborted; }
 639 
 640   void print_summary_info();
 641 
 642   void print_worker_threads_on(outputStream* st) const;
 643 
 644   void print_on_error(outputStream* st) const;
 645 
 646   // Liveness counting
 647 
 648   // Utility routine to set an exclusive range of cards on the given
 649   // card liveness bitmap
 650   inline void set_card_bitmap_range(BitMap* card_bm,
 651                                     BitMap::idx_t start_idx,
 652                                     BitMap::idx_t end_idx,
 653                                     bool is_par);
 654 
 655   // Returns the card number of the bottom of the G1 heap.
 656   // Used in biasing indices into accounting card bitmaps.
 657   intptr_t heap_bottom_card_num() const {
 658     return _heap_bottom_card_num;
 659   }
 660 
 661   // Returns the card bitmap for a given task or worker id.
 662   BitMap* count_card_bitmap_for(uint worker_id) {
 663     assert(worker_id < _max_worker_id, "oob");
 664     assert(_count_card_bitmaps != NULL, "uninitialized");
 665     BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
 666     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 667     return task_card_bm;
 668   }
 669 
 670   // Returns the array containing the marked bytes for each region,
 671   // for the given worker or task id.
 672   size_t* count_marked_bytes_array_for(uint worker_id) {
 673     assert(worker_id < _max_worker_id, "oob");
 674     assert(_count_marked_bytes != NULL, "uninitialized");
 675     size_t* marked_bytes_array = _count_marked_bytes[worker_id];
 676     assert(marked_bytes_array != NULL, "uninitialized");
 677     return marked_bytes_array;
 678   }
 679 
 680   // Returns the index in the liveness accounting card table bitmap
 681   // for the given address
 682   inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
 683 
 684   // Counts the size of the given memory region in the the given
 685   // marked_bytes array slot for the given HeapRegion.
 686   // Sets the bits in the given card bitmap that are associated with the
 687   // cards that are spanned by the memory region.
 688   inline void count_region(MemRegion mr,
 689                            HeapRegion* hr,
 690                            size_t* marked_bytes_array,
 691                            BitMap* task_card_bm);
 692 
 693   // Counts the given object in the given task/worker counting
 694   // data structures.
 695   inline void count_object(oop obj,
 696                            HeapRegion* hr,
 697                            size_t* marked_bytes_array,
 698                            BitMap* task_card_bm,
 699                            size_t word_size);
 700 
 701   // Attempts to mark the given object and, if successful, counts
 702   // the object in the given task/worker counting structures.
 703   inline bool par_mark_and_count(oop obj,
 704                                  HeapRegion* hr,
 705                                  size_t* marked_bytes_array,
 706                                  BitMap* task_card_bm);
 707 
 708   // Attempts to mark the given object and, if successful, counts
 709   // the object in the task/worker counting structures for the
 710   // given worker id.
 711   inline bool par_mark_and_count(oop obj,
 712                                  size_t word_size,
 713                                  HeapRegion* hr,
 714                                  uint worker_id);
 715 
 716   // Returns true if initialization was successfully completed.
 717   bool completed_initialization() const {
 718     return _completed_initialization;
 719   }
 720 
 721 protected:
 722   // Clear all the per-task bitmaps and arrays used to store the
 723   // counting data.
 724   void clear_all_count_data();
 725 
 726   // Aggregates the counting data for each worker/task
 727   // that was constructed while marking. Also sets
 728   // the amount of marked bytes for each region and
 729   // the top at concurrent mark count.
 730   void aggregate_count_data();
 731 
 732   // Verification routine
 733   void verify_count_data();
 734 };
 735 
 736 // A class representing a marking task.
 737 class CMTask : public TerminatorTerminator {
 738 private:
 739   enum PrivateConstants {
 740     // the regular clock call is called once the scanned words reaches
 741     // this limit
 742     words_scanned_period          = 12*1024,
 743     // the regular clock call is called once the number of visited
 744     // references reaches this limit
 745     refs_reached_period           = 384,
 746     // initial value for the hash seed, used in the work stealing code
 747     init_hash_seed                = 17,
 748     // how many entries will be transferred between global stack and
 749     // local queues
 750     global_stack_transfer_size    = 16
 751   };
 752 
 753   uint                        _worker_id;
 754   G1CollectedHeap*            _g1h;
 755   ConcurrentMark*             _cm;
 756   CMBitMap*                   _nextMarkBitMap;
 757   // the task queue of this task
 758   CMTaskQueue*                _task_queue;
 759 private:
 760   // the task queue set---needed for stealing
 761   CMTaskQueueSet*             _task_queues;
 762   // indicates whether the task has been claimed---this is only  for
 763   // debugging purposes
 764   bool                        _claimed;
 765 
 766   // number of calls to this task
 767   int                         _calls;
 768 
 769   // when the virtual timer reaches this time, the marking step should
 770   // exit
 771   double                      _time_target_ms;
 772   // the start time of the current marking step
 773   double                      _start_time_ms;
 774 
 775   // the oop closure used for iterations over oops
 776   G1CMOopClosure*             _cm_oop_closure;
 777 
 778   // the region this task is scanning, NULL if we're not scanning any
 779   HeapRegion*                 _curr_region;
 780   // the local finger of this task, NULL if we're not scanning a region
 781   HeapWord*                   _finger;
 782   // limit of the region this task is scanning, NULL if we're not scanning one
 783   HeapWord*                   _region_limit;
 784 
 785   // the number of words this task has scanned
 786   size_t                      _words_scanned;
 787   // When _words_scanned reaches this limit, the regular clock is
 788   // called. Notice that this might be decreased under certain
 789   // circumstances (i.e. when we believe that we did an expensive
 790   // operation).
 791   size_t                      _words_scanned_limit;
 792   // the initial value of _words_scanned_limit (i.e. what it was
 793   // before it was decreased).
 794   size_t                      _real_words_scanned_limit;
 795 
 796   // the number of references this task has visited
 797   size_t                      _refs_reached;
 798   // When _refs_reached reaches this limit, the regular clock is
 799   // called. Notice this this might be decreased under certain
 800   // circumstances (i.e. when we believe that we did an expensive
 801   // operation).
 802   size_t                      _refs_reached_limit;
 803   // the initial value of _refs_reached_limit (i.e. what it was before
 804   // it was decreased).
 805   size_t                      _real_refs_reached_limit;
 806 
 807   // used by the work stealing stuff
 808   int                         _hash_seed;
 809   // if this is true, then the task has aborted for some reason
 810   bool                        _has_aborted;
 811   // set when the task aborts because it has met its time quota
 812   bool                        _has_timed_out;
 813   // true when we're draining SATB buffers; this avoids the task
 814   // aborting due to SATB buffers being available (as we're already
 815   // dealing with them)
 816   bool                        _draining_satb_buffers;
 817 
 818   // number sequence of past step times
 819   NumberSeq                   _step_times_ms;
 820   // elapsed time of this task
 821   double                      _elapsed_time_ms;
 822   // termination time of this task
 823   double                      _termination_time_ms;
 824   // when this task got into the termination protocol
 825   double                      _termination_start_time_ms;
 826 
 827   // true when the task is during a concurrent phase, false when it is
 828   // in the remark phase (so, in the latter case, we do not have to
 829   // check all the things that we have to check during the concurrent
 830   // phase, i.e. SATB buffer availability...)
 831   bool                        _concurrent;
 832 
 833   TruncatedSeq                _marking_step_diffs_ms;
 834 
 835   // Counting data structures. Embedding the task's marked_bytes_array
 836   // and card bitmap into the actual task saves having to go through
 837   // the ConcurrentMark object.
 838   size_t*                     _marked_bytes_array;
 839   BitMap*                     _card_bm;
 840 
 841   // it updates the local fields after this task has claimed
 842   // a new region to scan
 843   void setup_for_region(HeapRegion* hr);
 844   // it brings up-to-date the limit of the region
 845   void update_region_limit();
 846 
 847   // called when either the words scanned or the refs visited limit
 848   // has been reached
 849   void reached_limit();
 850   // recalculates the words scanned and refs visited limits
 851   void recalculate_limits();
 852   // decreases the words scanned and refs visited limits when we reach
 853   // an expensive operation
 854   void decrease_limits();
 855   // it checks whether the words scanned or refs visited reached their
 856   // respective limit and calls reached_limit() if they have
 857   void check_limits() {
 858     if (_words_scanned >= _words_scanned_limit ||
 859         _refs_reached >= _refs_reached_limit) {
 860       reached_limit();
 861     }
 862   }
 863   // this is supposed to be called regularly during a marking step as
 864   // it checks a bunch of conditions that might cause the marking step
 865   // to abort
 866   void regular_clock_call();
 867   bool concurrent() { return _concurrent; }
 868 
 869   // Test whether obj might have already been passed over by the
 870   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 871   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 872 
 873   template<bool scan> void process_grey_object(oop obj);
 874 
 875 public:
 876   // It resets the task; it should be called right at the beginning of
 877   // a marking phase.
 878   void reset(CMBitMap* _nextMarkBitMap);
 879   // it clears all the fields that correspond to a claimed region.
 880   void clear_region_fields();
 881 
 882   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
 883 
 884   // The main method of this class which performs a marking step
 885   // trying not to exceed the given duration. However, it might exit
 886   // prematurely, according to some conditions (i.e. SATB buffers are
 887   // available for processing).
 888   void do_marking_step(double target_ms,
 889                        bool do_termination,
 890                        bool is_serial);
 891 
 892   // These two calls start and stop the timer
 893   void record_start_time() {
 894     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 895   }
 896   void record_end_time() {
 897     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 898   }
 899 
 900   // returns the worker ID associated with this task.
 901   uint worker_id() { return _worker_id; }
 902 
 903   // From TerminatorTerminator. It determines whether this task should
 904   // exit the termination protocol after it's entered it.
 905   virtual bool should_exit_termination();
 906 
 907   // Resets the local region fields after a task has finished scanning a
 908   // region; or when they have become stale as a result of the region
 909   // being evacuated.
 910   void giveup_current_region();
 911 
 912   HeapWord* finger()            { return _finger; }
 913 
 914   bool has_aborted()            { return _has_aborted; }
 915   void set_has_aborted()        { _has_aborted = true; }
 916   void clear_has_aborted()      { _has_aborted = false; }
 917   bool has_timed_out()          { return _has_timed_out; }
 918   bool claimed()                { return _claimed; }
 919 
 920   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 921 
 922   // Increment the number of references this task has visited.
 923   void increment_refs_reached() { ++_refs_reached; }
 924 
 925   // Grey the object by marking it.  If not already marked, push it on
 926   // the local queue if below the finger.
 927   // Precondition: obj is in region.
 928   // Precondition: obj is below region's NTAMS.
 929   inline void make_reference_grey(oop obj, HeapRegion* region);
 930 
 931   // Grey the object (by calling make_grey_reference) if required,
 932   // e.g. obj is below its containing region's NTAMS.
 933   // Precondition: obj is a valid heap object.
 934   inline void deal_with_reference(oop obj);
 935 
 936   // It scans an object and visits its children.
 937   inline void scan_object(oop obj);
 938 
 939   // It pushes an object on the local queue.
 940   inline void push(oop obj);
 941 
 942   // These two move entries to/from the global stack.
 943   void move_entries_to_global_stack();
 944   void get_entries_from_global_stack();
 945 
 946   // It pops and scans objects from the local queue. If partially is
 947   // true, then it stops when the queue size is of a given limit. If
 948   // partially is false, then it stops when the queue is empty.
 949   void drain_local_queue(bool partially);
 950   // It moves entries from the global stack to the local queue and
 951   // drains the local queue. If partially is true, then it stops when
 952   // both the global stack and the local queue reach a given size. If
 953   // partially if false, it tries to empty them totally.
 954   void drain_global_stack(bool partially);
 955   // It keeps picking SATB buffers and processing them until no SATB
 956   // buffers are available.
 957   void drain_satb_buffers();
 958 
 959   // moves the local finger to a new location
 960   inline void move_finger_to(HeapWord* new_finger) {
 961     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 962     _finger = new_finger;
 963   }
 964 
 965   CMTask(uint worker_id,
 966          ConcurrentMark *cm,
 967          size_t* marked_bytes,
 968          BitMap* card_bm,
 969          CMTaskQueue* task_queue,
 970          CMTaskQueueSet* task_queues);
 971 
 972   // it prints statistics associated with this task
 973   void print_stats();
 974 };
 975 
 976 // Class that's used to to print out per-region liveness
 977 // information. It's currently used at the end of marking and also
 978 // after we sort the old regions at the end of the cleanup operation.
 979 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
 980 private:
 981   // Accumulators for these values.
 982   size_t _total_used_bytes;
 983   size_t _total_capacity_bytes;
 984   size_t _total_prev_live_bytes;
 985   size_t _total_next_live_bytes;
 986 
 987   // These are set up when we come across a "stars humongous" region
 988   // (as this is where most of this information is stored, not in the
 989   // subsequent "continues humongous" regions). After that, for every
 990   // region in a given humongous region series we deduce the right
 991   // values for it by simply subtracting the appropriate amount from
 992   // these fields. All these values should reach 0 after we've visited
 993   // the last region in the series.
 994   size_t _hum_used_bytes;
 995   size_t _hum_capacity_bytes;
 996   size_t _hum_prev_live_bytes;
 997   size_t _hum_next_live_bytes;
 998 
 999   // Accumulator for the remembered set size
1000   size_t _total_remset_bytes;
1001 
1002   // Accumulator for strong code roots memory size
1003   size_t _total_strong_code_roots_bytes;
1004 
1005   static double perc(size_t val, size_t total) {
1006     if (total == 0) {
1007       return 0.0;
1008     } else {
1009       return 100.0 * ((double) val / (double) total);
1010     }
1011   }
1012 
1013   static double bytes_to_mb(size_t val) {
1014     return (double) val / (double) M;
1015   }
1016 
1017   // See the .cpp file.
1018   size_t get_hum_bytes(size_t* hum_bytes);
1019   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1020                      size_t* prev_live_bytes, size_t* next_live_bytes);
1021 
1022 public:
1023   // The header and footer are printed in the constructor and
1024   // destructor respectively.
1025   G1PrintRegionLivenessInfoClosure(const char* phase_name);
1026   virtual bool doHeapRegion(HeapRegion* r);
1027   ~G1PrintRegionLivenessInfoClosure();
1028 };
1029 
1030 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP