1 /*
   2  * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_GC_G1_CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_CONCURRENTMARK_HPP
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/g1/g1RegionToSpaceMapper.hpp"
  30 #include "gc/g1/heapRegionSet.hpp"
  31 #include "gc/shared/taskqueue.hpp"
  32 
  33 class G1CollectedHeap;
  34 class CMBitMap;
  35 class CMTask;
  36 class ConcurrentMark;
  37 typedef GenericTaskQueue<oop, mtGC>            CMTaskQueue;
  38 typedef GenericTaskQueueSet<CMTaskQueue, mtGC> CMTaskQueueSet;
  39 
  40 // Closure used by CM during concurrent reference discovery
  41 // and reference processing (during remarking) to determine
  42 // if a particular object is alive. It is primarily used
  43 // to determine if referents of discovered reference objects
  44 // are alive. An instance is also embedded into the
  45 // reference processor as the _is_alive_non_header field
  46 class G1CMIsAliveClosure: public BoolObjectClosure {
  47   G1CollectedHeap* _g1;
  48  public:
  49   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
  50 
  51   bool do_object_b(oop obj);
  52 };
  53 
  54 // A generic CM bit map.  This is essentially a wrapper around the BitMap
  55 // class, with one bit per (1<<_shifter) HeapWords.
  56 
  57 class CMBitMapRO VALUE_OBJ_CLASS_SPEC {
  58  protected:
  59   HeapWord* _bmStartWord;      // base address of range covered by map
  60   size_t    _bmWordSize;       // map size (in #HeapWords covered)
  61   const int _shifter;          // map to char or bit
  62   BitMap    _bm;               // the bit map itself
  63 
  64  public:
  65   // constructor
  66   CMBitMapRO(int shifter);
  67 
  68   enum { do_yield = true };
  69 
  70   // inquiries
  71   HeapWord* startWord()   const { return _bmStartWord; }
  72   size_t    sizeInWords() const { return _bmWordSize;  }
  73   // the following is one past the last word in space
  74   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
  75 
  76   // read marks
  77 
  78   bool isMarked(HeapWord* addr) const {
  79     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
  80            "outside underlying space?");
  81     return _bm.at(heapWordToOffset(addr));
  82   }
  83 
  84   // iteration
  85   inline bool iterate(BitMapClosure* cl, MemRegion mr);
  86   inline bool iterate(BitMapClosure* cl);
  87 
  88   // Return the address corresponding to the next marked bit at or after
  89   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  90   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  91   HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
  92                                      const HeapWord* limit = NULL) const;
  93   // Return the address corresponding to the next unmarked bit at or after
  94   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
  95   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
  96   HeapWord* getNextUnmarkedWordAddress(const HeapWord* addr,
  97                                        const HeapWord* limit = NULL) const;
  98 
  99   // conversion utilities
 100   HeapWord* offsetToHeapWord(size_t offset) const {
 101     return _bmStartWord + (offset << _shifter);
 102   }
 103   size_t heapWordToOffset(const HeapWord* addr) const {
 104     return pointer_delta(addr, _bmStartWord) >> _shifter;
 105   }
 106   int heapWordDiffToOffsetDiff(size_t diff) const;
 107 
 108   // The argument addr should be the start address of a valid object
 109   HeapWord* nextObject(HeapWord* addr) {
 110     oop obj = (oop) addr;
 111     HeapWord* res =  addr + obj->size();
 112     assert(offsetToHeapWord(heapWordToOffset(res)) == res, "sanity");
 113     return res;
 114   }
 115 
 116   void print_on_error(outputStream* st, const char* prefix) const;
 117 
 118   // debugging
 119   NOT_PRODUCT(bool covers(MemRegion rs) const;)
 120 };
 121 
 122 class CMBitMapMappingChangedListener : public G1MappingChangedListener {
 123  private:
 124   CMBitMap* _bm;
 125  public:
 126   CMBitMapMappingChangedListener() : _bm(NULL) {}
 127 
 128   void set_bitmap(CMBitMap* bm) { _bm = bm; }
 129 
 130   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 131 };
 132 
 133 class CMBitMap : public CMBitMapRO {
 134  private:
 135   CMBitMapMappingChangedListener _listener;
 136 
 137  public:
 138   static size_t compute_size(size_t heap_size);
 139   // Returns the amount of bytes on the heap between two marks in the bitmap.
 140   static size_t mark_distance();
 141   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 142   // mark bitmap corresponds to. This is the same as the mark distance above.
 143   static size_t heap_map_factor() {
 144     return mark_distance();
 145   }
 146 
 147   CMBitMap() : CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 148 
 149   // Initializes the underlying BitMap to cover the given area.
 150   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 151 
 152   // Write marks.
 153   inline void mark(HeapWord* addr);
 154   inline void clear(HeapWord* addr);
 155   inline bool parMark(HeapWord* addr);
 156   inline bool parClear(HeapWord* addr);
 157 
 158   void markRange(MemRegion mr);
 159   void clearRange(MemRegion mr);
 160 
 161   // Starting at the bit corresponding to "addr" (inclusive), find the next
 162   // "1" bit, if any.  This bit starts some run of consecutive "1"'s; find
 163   // the end of this run (stopping at "end_addr").  Return the MemRegion
 164   // covering from the start of the region corresponding to the first bit
 165   // of the run to the end of the region corresponding to the last bit of
 166   // the run.  If there is no "1" bit at or after "addr", return an empty
 167   // MemRegion.
 168   MemRegion getAndClearMarkedRegion(HeapWord* addr, HeapWord* end_addr);
 169 
 170   // Clear the whole mark bitmap.
 171   void clearAll();
 172 };
 173 
 174 // Represents a marking stack used by ConcurrentMarking in the G1 collector.
 175 class CMMarkStack VALUE_OBJ_CLASS_SPEC {
 176   VirtualSpace _virtual_space;   // Underlying backing store for actual stack
 177   ConcurrentMark* _cm;
 178   oop* _base;        // bottom of stack
 179   jint _index;       // one more than last occupied index
 180   jint _capacity;    // max #elements
 181   jint _saved_index; // value of _index saved at start of GC
 182 
 183   bool  _overflow;
 184   bool  _should_expand;
 185 
 186  public:
 187   CMMarkStack(ConcurrentMark* cm);
 188   ~CMMarkStack();
 189 
 190   bool allocate(size_t capacity);
 191 
 192   // Pushes the first "n" elements of "ptr_arr" on the stack.
 193   // Locking impl: concurrency is allowed only with
 194   // "par_push_arr" and/or "par_pop_arr" operations, which use the same
 195   // locking strategy.
 196   void par_push_arr(oop* ptr_arr, int n);
 197 
 198   // If returns false, the array was empty.  Otherwise, removes up to "max"
 199   // elements from the stack, and transfers them to "ptr_arr" in an
 200   // unspecified order.  The actual number transferred is given in "n" ("n
 201   // == 0" is deliberately redundant with the return value.)  Locking impl:
 202   // concurrency is allowed only with "par_push_arr" and/or "par_pop_arr"
 203   // operations, which use the same locking strategy.
 204   bool par_pop_arr(oop* ptr_arr, int max, int* n);
 205 
 206   bool isEmpty()    { return _index == 0; }
 207   int  maxElems()   { return _capacity; }
 208 
 209   bool overflow() { return _overflow; }
 210   void clear_overflow() { _overflow = false; }
 211 
 212   bool should_expand() const { return _should_expand; }
 213   void set_should_expand();
 214 
 215   // Expand the stack, typically in response to an overflow condition
 216   void expand();
 217 
 218   int  size() { return _index; }
 219 
 220   void setEmpty()   { _index = 0; clear_overflow(); }
 221 
 222   // Record the current index.
 223   void note_start_of_gc();
 224 
 225   // Make sure that we have not added any entries to the stack during GC.
 226   void note_end_of_gc();
 227 
 228   // Apply fn to each oop in the mark stack, up to the bound recorded
 229   // via one of the above "note" functions.  The mark stack must not
 230   // be modified while iterating.
 231   template<typename Fn> void iterate(Fn fn);
 232 };
 233 
 234 class ForceOverflowSettings VALUE_OBJ_CLASS_SPEC {
 235 private:
 236 #ifndef PRODUCT
 237   uintx _num_remaining;
 238   bool _force;
 239 #endif // !defined(PRODUCT)
 240 
 241 public:
 242   void init() PRODUCT_RETURN;
 243   void update() PRODUCT_RETURN;
 244   bool should_force() PRODUCT_RETURN_( return false; );
 245 };
 246 
 247 // this will enable a variety of different statistics per GC task
 248 #define _MARKING_STATS_       0
 249 // this will enable the higher verbose levels
 250 #define _MARKING_VERBOSE_     0
 251 
 252 #if _MARKING_STATS_
 253 #define statsOnly(statement)  \
 254 do {                          \
 255   statement ;                 \
 256 } while (0)
 257 #else // _MARKING_STATS_
 258 #define statsOnly(statement)  \
 259 do {                          \
 260 } while (0)
 261 #endif // _MARKING_STATS_
 262 
 263 typedef enum {
 264   no_verbose  = 0,   // verbose turned off
 265   stats_verbose,     // only prints stats at the end of marking
 266   low_verbose,       // low verbose, mostly per region and per major event
 267   medium_verbose,    // a bit more detailed than low
 268   high_verbose       // per object verbose
 269 } CMVerboseLevel;
 270 
 271 class YoungList;
 272 
 273 // Root Regions are regions that are not empty at the beginning of a
 274 // marking cycle and which we might collect during an evacuation pause
 275 // while the cycle is active. Given that, during evacuation pauses, we
 276 // do not copy objects that are explicitly marked, what we have to do
 277 // for the root regions is to scan them and mark all objects reachable
 278 // from them. According to the SATB assumptions, we only need to visit
 279 // each object once during marking. So, as long as we finish this scan
 280 // before the next evacuation pause, we can copy the objects from the
 281 // root regions without having to mark them or do anything else to them.
 282 //
 283 // Currently, we only support root region scanning once (at the start
 284 // of the marking cycle) and the root regions are all the survivor
 285 // regions populated during the initial-mark pause.
 286 class CMRootRegions VALUE_OBJ_CLASS_SPEC {
 287 private:
 288   YoungList*           _young_list;
 289   ConcurrentMark*      _cm;
 290 
 291   volatile bool        _scan_in_progress;
 292   volatile bool        _should_abort;
 293   HeapRegion* volatile _next_survivor;
 294 
 295 public:
 296   CMRootRegions();
 297   // We actually do most of the initialization in this method.
 298   void init(G1CollectedHeap* g1h, ConcurrentMark* cm);
 299 
 300   // Reset the claiming / scanning of the root regions.
 301   void prepare_for_scan();
 302 
 303   // Forces get_next() to return NULL so that the iteration aborts early.
 304   void abort() { _should_abort = true; }
 305 
 306   // Return true if the CM thread are actively scanning root regions,
 307   // false otherwise.
 308   bool scan_in_progress() { return _scan_in_progress; }
 309 
 310   // Claim the next root region to scan atomically, or return NULL if
 311   // all have been claimed.
 312   HeapRegion* claim_next();
 313 
 314   // Flag that we're done with root region scanning and notify anyone
 315   // who's waiting on it. If aborted is false, assume that all regions
 316   // have been claimed.
 317   void scan_finished();
 318 
 319   // If CM threads are still scanning root regions, wait until they
 320   // are done. Return true if we had to wait, false otherwise.
 321   bool wait_until_scan_finished();
 322 };
 323 
 324 class ConcurrentMarkThread;
 325 
 326 class ConcurrentMark: public CHeapObj<mtGC> {
 327   friend class CMMarkStack;
 328   friend class ConcurrentMarkThread;
 329   friend class CMTask;
 330   friend class CMBitMapClosure;
 331   friend class CMRemarkTask;
 332   friend class CMConcurrentMarkingTask;
 333   friend class G1ParNoteEndTask;
 334   friend class CalcLiveObjectsClosure;
 335   friend class G1CMRefProcTaskProxy;
 336   friend class G1CMRefProcTaskExecutor;
 337   friend class G1CMKeepAliveAndDrainClosure;
 338   friend class G1CMDrainMarkingStackClosure;
 339 
 340 protected:
 341   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 342   G1CollectedHeap*      _g1h;        // The heap
 343   uint                  _parallel_marking_threads; // The number of marking
 344                                                    // threads we're using
 345   uint                  _max_parallel_marking_threads; // Max number of marking
 346                                                        // threads we'll ever use
 347   double                _sleep_factor; // How much we have to sleep, with
 348                                        // respect to the work we just did, to
 349                                        // meet the marking overhead goal
 350   double                _marking_task_overhead; // Marking target overhead for
 351                                                 // a single task
 352 
 353   // Same as the two above, but for the cleanup task
 354   double                _cleanup_sleep_factor;
 355   double                _cleanup_task_overhead;
 356 
 357   FreeRegionList        _cleanup_list;
 358 
 359   // Concurrent marking support structures
 360   CMBitMap                _markBitMap1;
 361   CMBitMap                _markBitMap2;
 362   CMBitMapRO*             _prevMarkBitMap; // Completed mark bitmap
 363   CMBitMap*               _nextMarkBitMap; // Under-construction mark bitmap
 364 
 365   BitMap                  _region_bm;
 366   BitMap                  _card_bm;
 367 
 368   // Heap bounds
 369   HeapWord*               _heap_start;
 370   HeapWord*               _heap_end;
 371 
 372   // Root region tracking and claiming
 373   CMRootRegions           _root_regions;
 374 
 375   // For gray objects
 376   CMMarkStack             _markStack; // Grey objects behind global finger
 377   HeapWord* volatile      _finger;  // The global finger, region aligned,
 378                                     // always points to the end of the
 379                                     // last claimed region
 380 
 381   // Marking tasks
 382   uint                    _max_worker_id;// Maximum worker id
 383   uint                    _active_tasks; // Task num currently active
 384   CMTask**                _tasks;        // Task queue array (max_worker_id len)
 385   CMTaskQueueSet*         _task_queues;  // Task queue set
 386   ParallelTaskTerminator  _terminator;   // For termination
 387 
 388   // Two sync barriers that are used to synchronize tasks when an
 389   // overflow occurs. The algorithm is the following. All tasks enter
 390   // the first one to ensure that they have all stopped manipulating
 391   // the global data structures. After they exit it, they re-initialize
 392   // their data structures and task 0 re-initializes the global data
 393   // structures. Then, they enter the second sync barrier. This
 394   // ensure, that no task starts doing work before all data
 395   // structures (local and global) have been re-initialized. When they
 396   // exit it, they are free to start working again.
 397   WorkGangBarrierSync     _first_overflow_barrier_sync;
 398   WorkGangBarrierSync     _second_overflow_barrier_sync;
 399 
 400   // This is set by any task, when an overflow on the global data
 401   // structures is detected
 402   volatile bool           _has_overflown;
 403   // True: marking is concurrent, false: we're in remark
 404   volatile bool           _concurrent;
 405   // Set at the end of a Full GC so that marking aborts
 406   volatile bool           _has_aborted;
 407 
 408   // Used when remark aborts due to an overflow to indicate that
 409   // another concurrent marking phase should start
 410   volatile bool           _restart_for_overflow;
 411 
 412   // This is true from the very start of concurrent marking until the
 413   // point when all the tasks complete their work. It is really used
 414   // to determine the points between the end of concurrent marking and
 415   // time of remark.
 416   volatile bool           _concurrent_marking_in_progress;
 417 
 418   // Verbose level
 419   CMVerboseLevel          _verbose_level;
 420 
 421   // All of these times are in ms
 422   NumberSeq _init_times;
 423   NumberSeq _remark_times;
 424   NumberSeq _remark_mark_times;
 425   NumberSeq _remark_weak_ref_times;
 426   NumberSeq _cleanup_times;
 427   double    _total_counting_time;
 428   double    _total_rs_scrub_time;
 429 
 430   double*   _accum_task_vtime;   // Accumulated task vtime
 431 
 432   WorkGang* _parallel_workers;
 433 
 434   ForceOverflowSettings _force_overflow_conc;
 435   ForceOverflowSettings _force_overflow_stw;
 436 
 437   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 438   void weakRefsWork(bool clear_all_soft_refs);
 439 
 440   void swapMarkBitMaps();
 441 
 442   // It resets the global marking data structures, as well as the
 443   // task local ones; should be called during initial mark.
 444   void reset();
 445 
 446   // Resets all the marking data structures. Called when we have to restart
 447   // marking or when marking completes (via set_non_marking_state below).
 448   void reset_marking_state(bool clear_overflow = true);
 449 
 450   // We do this after we're done with marking so that the marking data
 451   // structures are initialized to a sensible and predictable state.
 452   void set_non_marking_state();
 453 
 454   // Called to indicate how many threads are currently active.
 455   void set_concurrency(uint active_tasks);
 456 
 457   // It should be called to indicate which phase we're in (concurrent
 458   // mark or remark) and how many threads are currently active.
 459   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 460 
 461   // Prints all gathered CM-related statistics
 462   void print_stats();
 463 
 464   bool cleanup_list_is_empty() {
 465     return _cleanup_list.is_empty();
 466   }
 467 
 468   // Accessor methods
 469   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 470   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 471   double sleep_factor()                     { return _sleep_factor; }
 472   double marking_task_overhead()            { return _marking_task_overhead;}
 473   double cleanup_sleep_factor()             { return _cleanup_sleep_factor; }
 474   double cleanup_task_overhead()            { return _cleanup_task_overhead;}
 475 
 476   HeapWord*               finger()          { return _finger;   }
 477   bool                    concurrent()      { return _concurrent; }
 478   uint                    active_tasks()    { return _active_tasks; }
 479   ParallelTaskTerminator* terminator()      { return &_terminator; }
 480 
 481   // It claims the next available region to be scanned by a marking
 482   // task/thread. It might return NULL if the next region is empty or
 483   // we have run out of regions. In the latter case, out_of_regions()
 484   // determines whether we've really run out of regions or the task
 485   // should call claim_region() again. This might seem a bit
 486   // awkward. Originally, the code was written so that claim_region()
 487   // either successfully returned with a non-empty region or there
 488   // were no more regions to be claimed. The problem with this was
 489   // that, in certain circumstances, it iterated over large chunks of
 490   // the heap finding only empty regions and, while it was working, it
 491   // was preventing the calling task to call its regular clock
 492   // method. So, this way, each task will spend very little time in
 493   // claim_region() and is allowed to call the regular clock method
 494   // frequently.
 495   HeapRegion* claim_region(uint worker_id);
 496 
 497   // It determines whether we've run out of regions to scan. Note that
 498   // the finger can point past the heap end in case the heap was expanded
 499   // to satisfy an allocation without doing a GC. This is fine, because all
 500   // objects in those regions will be considered live anyway because of
 501   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 502   bool        out_of_regions() { return _finger >= _heap_end; }
 503 
 504   // Returns the task with the given id
 505   CMTask* task(int id) {
 506     assert(0 <= id && id < (int) _active_tasks,
 507            "task id not within active bounds");
 508     return _tasks[id];
 509   }
 510 
 511   // Returns the task queue with the given id
 512   CMTaskQueue* task_queue(int id) {
 513     assert(0 <= id && id < (int) _active_tasks,
 514            "task queue id not within active bounds");
 515     return (CMTaskQueue*) _task_queues->queue(id);
 516   }
 517 
 518   // Returns the task queue set
 519   CMTaskQueueSet* task_queues()  { return _task_queues; }
 520 
 521   // Access / manipulation of the overflow flag which is set to
 522   // indicate that the global stack has overflown
 523   bool has_overflown()           { return _has_overflown; }
 524   void set_has_overflown()       { _has_overflown = true; }
 525   void clear_has_overflown()     { _has_overflown = false; }
 526   bool restart_for_overflow()    { return _restart_for_overflow; }
 527 
 528   // Methods to enter the two overflow sync barriers
 529   void enter_first_sync_barrier(uint worker_id);
 530   void enter_second_sync_barrier(uint worker_id);
 531 
 532   ForceOverflowSettings* force_overflow_conc() {
 533     return &_force_overflow_conc;
 534   }
 535 
 536   ForceOverflowSettings* force_overflow_stw() {
 537     return &_force_overflow_stw;
 538   }
 539 
 540   ForceOverflowSettings* force_overflow() {
 541     if (concurrent()) {
 542       return force_overflow_conc();
 543     } else {
 544       return force_overflow_stw();
 545     }
 546   }
 547 
 548   // Live Data Counting data structures...
 549   // These data structures are initialized at the start of
 550   // marking. They are written to while marking is active.
 551   // They are aggregated during remark; the aggregated values
 552   // are then used to populate the _region_bm, _card_bm, and
 553   // the total live bytes, which are then subsequently updated
 554   // during cleanup.
 555 
 556   // An array of bitmaps (one bit map per task). Each bitmap
 557   // is used to record the cards spanned by the live objects
 558   // marked by that task/worker.
 559   BitMap*  _count_card_bitmaps;
 560 
 561   // Used to record the number of marked live bytes
 562   // (for each region, by worker thread).
 563   size_t** _count_marked_bytes;
 564 
 565   // Card index of the bottom of the G1 heap. Used for biasing indices into
 566   // the card bitmaps.
 567   intptr_t _heap_bottom_card_num;
 568 
 569   // Set to true when initialization is complete
 570   bool _completed_initialization;
 571 
 572 public:
 573   // Manipulation of the global mark stack.
 574   // The push and pop operations are used by tasks for transfers
 575   // between task-local queues and the global mark stack, and use
 576   // locking for concurrency safety.
 577   bool mark_stack_push(oop* arr, int n) {
 578     _markStack.par_push_arr(arr, n);
 579     if (_markStack.overflow()) {
 580       set_has_overflown();
 581       return false;
 582     }
 583     return true;
 584   }
 585   void mark_stack_pop(oop* arr, int max, int* n) {
 586     _markStack.par_pop_arr(arr, max, n);
 587   }
 588   size_t mark_stack_size()                { return _markStack.size(); }
 589   size_t partial_mark_stack_size_target() { return _markStack.maxElems()/3; }
 590   bool mark_stack_overflow()              { return _markStack.overflow(); }
 591   bool mark_stack_empty()                 { return _markStack.isEmpty(); }
 592 
 593   CMRootRegions* root_regions() { return &_root_regions; }
 594 
 595   bool concurrent_marking_in_progress() {
 596     return _concurrent_marking_in_progress;
 597   }
 598   void set_concurrent_marking_in_progress() {
 599     _concurrent_marking_in_progress = true;
 600   }
 601   void clear_concurrent_marking_in_progress() {
 602     _concurrent_marking_in_progress = false;
 603   }
 604 
 605   void update_accum_task_vtime(int i, double vtime) {
 606     _accum_task_vtime[i] += vtime;
 607   }
 608 
 609   double all_task_accum_vtime() {
 610     double ret = 0.0;
 611     for (uint i = 0; i < _max_worker_id; ++i)
 612       ret += _accum_task_vtime[i];
 613     return ret;
 614   }
 615 
 616   // Attempts to steal an object from the task queues of other tasks
 617   bool try_stealing(uint worker_id, int* hash_seed, oop& obj);
 618 
 619   ConcurrentMark(G1CollectedHeap* g1h,
 620                  G1RegionToSpaceMapper* prev_bitmap_storage,
 621                  G1RegionToSpaceMapper* next_bitmap_storage);
 622   ~ConcurrentMark();
 623 
 624   ConcurrentMarkThread* cmThread() { return _cmThread; }
 625 
 626   CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 627   CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 628 
 629   // Returns the number of GC threads to be used in a concurrent
 630   // phase based on the number of GC threads being used in a STW
 631   // phase.
 632   uint scale_parallel_threads(uint n_par_threads);
 633 
 634   // Calculates the number of GC threads to be used in a concurrent phase.
 635   uint calc_parallel_marking_threads();
 636 
 637   // The following three are interaction between CM and
 638   // G1CollectedHeap
 639 
 640   // This notifies CM that a root during initial-mark needs to be
 641   // grayed. It is MT-safe. word_size is the size of the object in
 642   // words. It is passed explicitly as sometimes we cannot calculate
 643   // it from the given object because it might be in an inconsistent
 644   // state (e.g., in to-space and being copied). So the caller is
 645   // responsible for dealing with this issue (e.g., get the size from
 646   // the from-space image when the to-space image might be
 647   // inconsistent) and always passing the size. hr is the region that
 648   // contains the object and it's passed optionally from callers who
 649   // might already have it (no point in recalculating it).
 650   inline void grayRoot(oop obj,
 651                        size_t word_size,
 652                        uint worker_id,
 653                        HeapRegion* hr = NULL);
 654 
 655   // It iterates over the heap and for each object it comes across it
 656   // will dump the contents of its reference fields, as well as
 657   // liveness information for the object and its referents. The dump
 658   // will be written to a file with the following name:
 659   // G1PrintReachableBaseFile + "." + str.
 660   // vo decides whether the prev (vo == UsePrevMarking), the next
 661   // (vo == UseNextMarking) marking information, or the mark word
 662   // (vo == UseMarkWord) will be used to determine the liveness of
 663   // each object / referent.
 664   // If all is true, all objects in the heap will be dumped, otherwise
 665   // only the live ones. In the dump the following symbols / breviations
 666   // are used:
 667   //   M : an explicitly live object (its bitmap bit is set)
 668   //   > : an implicitly live object (over tams)
 669   //   O : an object outside the G1 heap (typically: in the perm gen)
 670   //   NOT : a reference field whose referent is not live
 671   //   AND MARKED : indicates that an object is both explicitly and
 672   //   implicitly live (it should be one or the other, not both)
 673   void print_reachable(const char* str,
 674                        VerifyOption vo,
 675                        bool all) PRODUCT_RETURN;
 676 
 677   // Clear the next marking bitmap (will be called concurrently).
 678   void clearNextBitmap();
 679 
 680   // Return whether the next mark bitmap has no marks set. To be used for assertions
 681   // only. Will not yield to pause requests.
 682   bool nextMarkBitmapIsClear();
 683 
 684   // These two do the work that needs to be done before and after the
 685   // initial root checkpoint. Since this checkpoint can be done at two
 686   // different points (i.e. an explicit pause or piggy-backed on a
 687   // young collection), then it's nice to be able to easily share the
 688   // pre/post code. It might be the case that we can put everything in
 689   // the post method. TP
 690   void checkpointRootsInitialPre();
 691   void checkpointRootsInitialPost();
 692 
 693   // Scan all the root regions and mark everything reachable from
 694   // them.
 695   void scanRootRegions();
 696 
 697   // Scan a single root region and mark everything reachable from it.
 698   void scanRootRegion(HeapRegion* hr, uint worker_id);
 699 
 700   // Do concurrent phase of marking, to a tentative transitive closure.
 701   void markFromRoots();
 702 
 703   void checkpointRootsFinal(bool clear_all_soft_refs);
 704   void checkpointRootsFinalWork();
 705   void cleanup();
 706   void completeCleanup();
 707 
 708   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 709   // this carefully!
 710   inline void markPrev(oop p);
 711 
 712   // Clears marks for all objects in the given range, for the prev or
 713   // next bitmaps.  NB: the previous bitmap is usually
 714   // read-only, so use this carefully!
 715   void clearRangePrevBitmap(MemRegion mr);
 716   void clearRangeNextBitmap(MemRegion mr);
 717 
 718   // Notify data structures that a GC has started.
 719   void note_start_of_gc() {
 720     _markStack.note_start_of_gc();
 721   }
 722 
 723   // Notify data structures that a GC is finished.
 724   void note_end_of_gc() {
 725     _markStack.note_end_of_gc();
 726   }
 727 
 728   // Verify that there are no CSet oops on the stacks (taskqueues /
 729   // global mark stack) and fingers (global / per-task).
 730   // If marking is not in progress, it's a no-op.
 731   void verify_no_cset_oops() PRODUCT_RETURN;
 732 
 733   bool isPrevMarked(oop p) const {
 734     assert(p != NULL && p->is_oop(), "expected an oop");
 735     HeapWord* addr = (HeapWord*)p;
 736     assert(addr >= _prevMarkBitMap->startWord() ||
 737            addr < _prevMarkBitMap->endWord(), "in a region");
 738 
 739     return _prevMarkBitMap->isMarked(addr);
 740   }
 741 
 742   inline bool do_yield_check(uint worker_i = 0);
 743 
 744   // Called to abort the marking cycle after a Full GC takes place.
 745   void abort();
 746 
 747   bool has_aborted()      { return _has_aborted; }
 748 
 749   // This prints the global/local fingers. It is used for debugging.
 750   NOT_PRODUCT(void print_finger();)
 751 
 752   void print_summary_info();
 753 
 754   void print_worker_threads_on(outputStream* st) const;
 755 
 756   void print_on_error(outputStream* st) const;
 757 
 758   // The following indicate whether a given verbose level has been
 759   // set. Notice that anything above stats is conditional to
 760   // _MARKING_VERBOSE_ having been set to 1
 761   bool verbose_stats() {
 762     return _verbose_level >= stats_verbose;
 763   }
 764   bool verbose_low() {
 765     return _MARKING_VERBOSE_ && _verbose_level >= low_verbose;
 766   }
 767   bool verbose_medium() {
 768     return _MARKING_VERBOSE_ && _verbose_level >= medium_verbose;
 769   }
 770   bool verbose_high() {
 771     return _MARKING_VERBOSE_ && _verbose_level >= high_verbose;
 772   }
 773 
 774   // Liveness counting
 775 
 776   // Utility routine to set an exclusive range of cards on the given
 777   // card liveness bitmap
 778   inline void set_card_bitmap_range(BitMap* card_bm,
 779                                     BitMap::idx_t start_idx,
 780                                     BitMap::idx_t end_idx,
 781                                     bool is_par);
 782 
 783   // Returns the card number of the bottom of the G1 heap.
 784   // Used in biasing indices into accounting card bitmaps.
 785   intptr_t heap_bottom_card_num() const {
 786     return _heap_bottom_card_num;
 787   }
 788 
 789   // Returns the card bitmap for a given task or worker id.
 790   BitMap* count_card_bitmap_for(uint worker_id) {
 791     assert(worker_id < _max_worker_id, "oob");
 792     assert(_count_card_bitmaps != NULL, "uninitialized");
 793     BitMap* task_card_bm = &_count_card_bitmaps[worker_id];
 794     assert(task_card_bm->size() == _card_bm.size(), "size mismatch");
 795     return task_card_bm;
 796   }
 797 
 798   // Returns the array containing the marked bytes for each region,
 799   // for the given worker or task id.
 800   size_t* count_marked_bytes_array_for(uint worker_id) {
 801     assert(worker_id < _max_worker_id, "oob");
 802     assert(_count_marked_bytes != NULL, "uninitialized");
 803     size_t* marked_bytes_array = _count_marked_bytes[worker_id];
 804     assert(marked_bytes_array != NULL, "uninitialized");
 805     return marked_bytes_array;
 806   }
 807 
 808   // Returns the index in the liveness accounting card table bitmap
 809   // for the given address
 810   inline BitMap::idx_t card_bitmap_index_for(HeapWord* addr);
 811 
 812   // Counts the size of the given memory region in the the given
 813   // marked_bytes array slot for the given HeapRegion.
 814   // Sets the bits in the given card bitmap that are associated with the
 815   // cards that are spanned by the memory region.
 816   inline void count_region(MemRegion mr,
 817                            HeapRegion* hr,
 818                            size_t* marked_bytes_array,
 819                            BitMap* task_card_bm);
 820 
 821   // Counts the given object in the given task/worker counting
 822   // data structures.
 823   inline void count_object(oop obj,
 824                            HeapRegion* hr,
 825                            size_t* marked_bytes_array,
 826                            BitMap* task_card_bm,
 827                            size_t word_size);
 828 
 829   // Attempts to mark the given object and, if successful, counts
 830   // the object in the given task/worker counting structures.
 831   inline bool par_mark_and_count(oop obj,
 832                                  HeapRegion* hr,
 833                                  size_t* marked_bytes_array,
 834                                  BitMap* task_card_bm);
 835 
 836   // Attempts to mark the given object and, if successful, counts
 837   // the object in the task/worker counting structures for the
 838   // given worker id.
 839   inline bool par_mark_and_count(oop obj,
 840                                  size_t word_size,
 841                                  HeapRegion* hr,
 842                                  uint worker_id);
 843 
 844   // Returns true if initialization was successfully completed.
 845   bool completed_initialization() const {
 846     return _completed_initialization;
 847   }
 848 
 849 protected:
 850   // Clear all the per-task bitmaps and arrays used to store the
 851   // counting data.
 852   void clear_all_count_data();
 853 
 854   // Aggregates the counting data for each worker/task
 855   // that was constructed while marking. Also sets
 856   // the amount of marked bytes for each region and
 857   // the top at concurrent mark count.
 858   void aggregate_count_data();
 859 
 860   // Verification routine
 861   void verify_count_data();
 862 };
 863 
 864 // A class representing a marking task.
 865 class CMTask : public TerminatorTerminator {
 866 private:
 867   enum PrivateConstants {
 868     // the regular clock call is called once the scanned words reaches
 869     // this limit
 870     words_scanned_period          = 12*1024,
 871     // the regular clock call is called once the number of visited
 872     // references reaches this limit
 873     refs_reached_period           = 384,
 874     // initial value for the hash seed, used in the work stealing code
 875     init_hash_seed                = 17,
 876     // how many entries will be transferred between global stack and
 877     // local queues
 878     global_stack_transfer_size    = 16
 879   };
 880 
 881   uint                        _worker_id;
 882   G1CollectedHeap*            _g1h;
 883   ConcurrentMark*             _cm;
 884   CMBitMap*                   _nextMarkBitMap;
 885   // the task queue of this task
 886   CMTaskQueue*                _task_queue;
 887 private:
 888   // the task queue set---needed for stealing
 889   CMTaskQueueSet*             _task_queues;
 890   // indicates whether the task has been claimed---this is only  for
 891   // debugging purposes
 892   bool                        _claimed;
 893 
 894   // number of calls to this task
 895   int                         _calls;
 896 
 897   // when the virtual timer reaches this time, the marking step should
 898   // exit
 899   double                      _time_target_ms;
 900   // the start time of the current marking step
 901   double                      _start_time_ms;
 902 
 903   // the oop closure used for iterations over oops
 904   G1CMOopClosure*             _cm_oop_closure;
 905 
 906   // the region this task is scanning, NULL if we're not scanning any
 907   HeapRegion*                 _curr_region;
 908   // the local finger of this task, NULL if we're not scanning a region
 909   HeapWord*                   _finger;
 910   // limit of the region this task is scanning, NULL if we're not scanning one
 911   HeapWord*                   _region_limit;
 912 
 913   // the number of words this task has scanned
 914   size_t                      _words_scanned;
 915   // When _words_scanned reaches this limit, the regular clock is
 916   // called. Notice that this might be decreased under certain
 917   // circumstances (i.e. when we believe that we did an expensive
 918   // operation).
 919   size_t                      _words_scanned_limit;
 920   // the initial value of _words_scanned_limit (i.e. what it was
 921   // before it was decreased).
 922   size_t                      _real_words_scanned_limit;
 923 
 924   // the number of references this task has visited
 925   size_t                      _refs_reached;
 926   // When _refs_reached reaches this limit, the regular clock is
 927   // called. Notice this this might be decreased under certain
 928   // circumstances (i.e. when we believe that we did an expensive
 929   // operation).
 930   size_t                      _refs_reached_limit;
 931   // the initial value of _refs_reached_limit (i.e. what it was before
 932   // it was decreased).
 933   size_t                      _real_refs_reached_limit;
 934 
 935   // used by the work stealing stuff
 936   int                         _hash_seed;
 937   // if this is true, then the task has aborted for some reason
 938   bool                        _has_aborted;
 939   // set when the task aborts because it has met its time quota
 940   bool                        _has_timed_out;
 941   // true when we're draining SATB buffers; this avoids the task
 942   // aborting due to SATB buffers being available (as we're already
 943   // dealing with them)
 944   bool                        _draining_satb_buffers;
 945 
 946   // number sequence of past step times
 947   NumberSeq                   _step_times_ms;
 948   // elapsed time of this task
 949   double                      _elapsed_time_ms;
 950   // termination time of this task
 951   double                      _termination_time_ms;
 952   // when this task got into the termination protocol
 953   double                      _termination_start_time_ms;
 954 
 955   // true when the task is during a concurrent phase, false when it is
 956   // in the remark phase (so, in the latter case, we do not have to
 957   // check all the things that we have to check during the concurrent
 958   // phase, i.e. SATB buffer availability...)
 959   bool                        _concurrent;
 960 
 961   TruncatedSeq                _marking_step_diffs_ms;
 962 
 963   // Counting data structures. Embedding the task's marked_bytes_array
 964   // and card bitmap into the actual task saves having to go through
 965   // the ConcurrentMark object.
 966   size_t*                     _marked_bytes_array;
 967   BitMap*                     _card_bm;
 968 
 969   // LOTS of statistics related with this task
 970 #if _MARKING_STATS_
 971   NumberSeq                   _all_clock_intervals_ms;
 972   double                      _interval_start_time_ms;
 973 
 974   size_t                      _aborted;
 975   size_t                      _aborted_overflow;
 976   size_t                      _aborted_cm_aborted;
 977   size_t                      _aborted_yield;
 978   size_t                      _aborted_timed_out;
 979   size_t                      _aborted_satb;
 980   size_t                      _aborted_termination;
 981 
 982   size_t                      _steal_attempts;
 983   size_t                      _steals;
 984 
 985   size_t                      _clock_due_to_marking;
 986   size_t                      _clock_due_to_scanning;
 987 
 988   size_t                      _local_pushes;
 989   size_t                      _local_pops;
 990   size_t                      _local_max_size;
 991   size_t                      _objs_scanned;
 992 
 993   size_t                      _global_pushes;
 994   size_t                      _global_pops;
 995   size_t                      _global_max_size;
 996 
 997   size_t                      _global_transfers_to;
 998   size_t                      _global_transfers_from;
 999 
1000   size_t                      _regions_claimed;
1001   size_t                      _objs_found_on_bitmap;
1002 
1003   size_t                      _satb_buffers_processed;
1004 #endif // _MARKING_STATS_
1005 
1006   // it updates the local fields after this task has claimed
1007   // a new region to scan
1008   void setup_for_region(HeapRegion* hr);
1009   // it brings up-to-date the limit of the region
1010   void update_region_limit();
1011 
1012   // called when either the words scanned or the refs visited limit
1013   // has been reached
1014   void reached_limit();
1015   // recalculates the words scanned and refs visited limits
1016   void recalculate_limits();
1017   // decreases the words scanned and refs visited limits when we reach
1018   // an expensive operation
1019   void decrease_limits();
1020   // it checks whether the words scanned or refs visited reached their
1021   // respective limit and calls reached_limit() if they have
1022   void check_limits() {
1023     if (_words_scanned >= _words_scanned_limit ||
1024         _refs_reached >= _refs_reached_limit) {
1025       reached_limit();
1026     }
1027   }
1028   // this is supposed to be called regularly during a marking step as
1029   // it checks a bunch of conditions that might cause the marking step
1030   // to abort
1031   void regular_clock_call();
1032   bool concurrent() { return _concurrent; }
1033 
1034   // Test whether obj might have already been passed over by the
1035   // mark bitmap scan, and so needs to be pushed onto the mark stack.
1036   bool is_below_finger(oop obj, HeapWord* global_finger) const;
1037 
1038   template<bool scan> void process_grey_object(oop obj);
1039 
1040 public:
1041   // It resets the task; it should be called right at the beginning of
1042   // a marking phase.
1043   void reset(CMBitMap* _nextMarkBitMap);
1044   // it clears all the fields that correspond to a claimed region.
1045   void clear_region_fields();
1046 
1047   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
1048 
1049   // The main method of this class which performs a marking step
1050   // trying not to exceed the given duration. However, it might exit
1051   // prematurely, according to some conditions (i.e. SATB buffers are
1052   // available for processing).
1053   void do_marking_step(double target_ms,
1054                        bool do_termination,
1055                        bool is_serial);
1056 
1057   // These two calls start and stop the timer
1058   void record_start_time() {
1059     _elapsed_time_ms = os::elapsedTime() * 1000.0;
1060   }
1061   void record_end_time() {
1062     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
1063   }
1064 
1065   // returns the worker ID associated with this task.
1066   uint worker_id() { return _worker_id; }
1067 
1068   // From TerminatorTerminator. It determines whether this task should
1069   // exit the termination protocol after it's entered it.
1070   virtual bool should_exit_termination();
1071 
1072   // Resets the local region fields after a task has finished scanning a
1073   // region; or when they have become stale as a result of the region
1074   // being evacuated.
1075   void giveup_current_region();
1076 
1077   HeapWord* finger()            { return _finger; }
1078 
1079   bool has_aborted()            { return _has_aborted; }
1080   void set_has_aborted()        { _has_aborted = true; }
1081   void clear_has_aborted()      { _has_aborted = false; }
1082   bool has_timed_out()          { return _has_timed_out; }
1083   bool claimed()                { return _claimed; }
1084 
1085   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
1086 
1087   // Increment the number of references this task has visited.
1088   void increment_refs_reached() { ++_refs_reached; }
1089 
1090   // Grey the object by marking it.  If not already marked, push it on
1091   // the local queue if below the finger.
1092   // Precondition: obj is in region.
1093   // Precondition: obj is below region's NTAMS.
1094   inline void make_reference_grey(oop obj, HeapRegion* region);
1095 
1096   // Grey the object (by calling make_grey_reference) if required,
1097   // e.g. obj is below its containing region's NTAMS.
1098   // Precondition: obj is a valid heap object.
1099   inline void deal_with_reference(oop obj);
1100 
1101   // It scans an object and visits its children.
1102   inline void scan_object(oop obj);
1103 
1104   // It pushes an object on the local queue.
1105   inline void push(oop obj);
1106 
1107   // These two move entries to/from the global stack.
1108   void move_entries_to_global_stack();
1109   void get_entries_from_global_stack();
1110 
1111   // It pops and scans objects from the local queue. If partially is
1112   // true, then it stops when the queue size is of a given limit. If
1113   // partially is false, then it stops when the queue is empty.
1114   void drain_local_queue(bool partially);
1115   // It moves entries from the global stack to the local queue and
1116   // drains the local queue. If partially is true, then it stops when
1117   // both the global stack and the local queue reach a given size. If
1118   // partially if false, it tries to empty them totally.
1119   void drain_global_stack(bool partially);
1120   // It keeps picking SATB buffers and processing them until no SATB
1121   // buffers are available.
1122   void drain_satb_buffers();
1123 
1124   // moves the local finger to a new location
1125   inline void move_finger_to(HeapWord* new_finger) {
1126     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
1127     _finger = new_finger;
1128   }
1129 
1130   CMTask(uint worker_id,
1131          ConcurrentMark *cm,
1132          size_t* marked_bytes,
1133          BitMap* card_bm,
1134          CMTaskQueue* task_queue,
1135          CMTaskQueueSet* task_queues);
1136 
1137   // it prints statistics associated with this task
1138   void print_stats();
1139 
1140 #if _MARKING_STATS_
1141   void increase_objs_found_on_bitmap() { ++_objs_found_on_bitmap; }
1142 #endif // _MARKING_STATS_
1143 };
1144 
1145 // Class that's used to to print out per-region liveness
1146 // information. It's currently used at the end of marking and also
1147 // after we sort the old regions at the end of the cleanup operation.
1148 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
1149 private:
1150   outputStream* _out;
1151 
1152   // Accumulators for these values.
1153   size_t _total_used_bytes;
1154   size_t _total_capacity_bytes;
1155   size_t _total_prev_live_bytes;
1156   size_t _total_next_live_bytes;
1157 
1158   // These are set up when we come across a "stars humongous" region
1159   // (as this is where most of this information is stored, not in the
1160   // subsequent "continues humongous" regions). After that, for every
1161   // region in a given humongous region series we deduce the right
1162   // values for it by simply subtracting the appropriate amount from
1163   // these fields. All these values should reach 0 after we've visited
1164   // the last region in the series.
1165   size_t _hum_used_bytes;
1166   size_t _hum_capacity_bytes;
1167   size_t _hum_prev_live_bytes;
1168   size_t _hum_next_live_bytes;
1169 
1170   // Accumulator for the remembered set size
1171   size_t _total_remset_bytes;
1172 
1173   // Accumulator for strong code roots memory size
1174   size_t _total_strong_code_roots_bytes;
1175 
1176   static double perc(size_t val, size_t total) {
1177     if (total == 0) {
1178       return 0.0;
1179     } else {
1180       return 100.0 * ((double) val / (double) total);
1181     }
1182   }
1183 
1184   static double bytes_to_mb(size_t val) {
1185     return (double) val / (double) M;
1186   }
1187 
1188   // See the .cpp file.
1189   size_t get_hum_bytes(size_t* hum_bytes);
1190   void get_hum_bytes(size_t* used_bytes, size_t* capacity_bytes,
1191                      size_t* prev_live_bytes, size_t* next_live_bytes);
1192 
1193 public:
1194   // The header and footer are printed in the constructor and
1195   // destructor respectively.
1196   G1PrintRegionLivenessInfoClosure(outputStream* out, const char* phase_name);
1197   virtual bool doHeapRegion(HeapRegion* r);
1198   ~G1PrintRegionLivenessInfoClosure();
1199 };
1200 
1201 #endif // SHARE_VM_GC_G1_CONCURRENTMARK_HPP