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