1 /*
   2  * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #ifndef SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
  26 #define SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP
  27 
  28 #include "classfile/javaClasses.hpp"
  29 #include "gc/g1/g1ConcurrentMarkObjArrayProcessor.hpp"
  30 #include "gc/g1/g1RegionToSpaceMapper.hpp"
  31 #include "gc/g1/heapRegionSet.hpp"
  32 #include "gc/shared/taskqueue.hpp"
  33 
  34 class G1CollectedHeap;
  35 class G1CMBitMap;
  36 class G1CMTask;
  37 class G1ConcurrentMark;
  38 class ConcurrentGCTimer;
  39 class G1OldTracer;
  40 class G1SurvivorRegions;
  41 
  42 // This is a container class for either an oop or a continuation address for
  43 // mark stack entries. Both are pushed onto the mark stack.
  44 class G1TaskQueueEntry VALUE_OBJ_CLASS_SPEC {
  45 private:
  46   void* _holder;
  47 
  48   static const uintptr_t ArraySliceBit = 1;
  49 
  50   G1TaskQueueEntry(oop obj) : _holder(obj) {
  51     assert(_holder != NULL, "Not allowed to set NULL task queue element");
  52   }
  53   G1TaskQueueEntry(HeapWord* addr) : _holder((void*)((uintptr_t)addr | ArraySliceBit)) { }
  54 public:
  55   G1TaskQueueEntry(const G1TaskQueueEntry& other) { _holder = other._holder; }
  56   G1TaskQueueEntry() : _holder(NULL) { }
  57 
  58   static G1TaskQueueEntry from_slice(HeapWord* what) { return G1TaskQueueEntry(what); }
  59   static G1TaskQueueEntry from_oop(oop obj) { return G1TaskQueueEntry(obj); }
  60 
  61   void assign(const G1TaskQueueEntry& t) {
  62     _holder = t._holder;
  63   }
  64 
  65   volatile G1TaskQueueEntry& operator=(const volatile G1TaskQueueEntry& t) volatile {
  66     _holder = t._holder;
  67     return *this;
  68   }
  69 
  70   oop obj() const {
  71     assert(!is_array_slice(), "Trying to read array slice " PTR_FORMAT " as oop", p2i(_holder));
  72     return (oop)_holder;
  73   }
  74 
  75   HeapWord* slice() const {
  76     assert(is_array_slice(), "Trying to read oop " PTR_FORMAT " as array slice", p2i(_holder));
  77     return (HeapWord*)((uintptr_t)_holder & ~ArraySliceBit);
  78   }
  79 
  80   bool is_oop() const { return !is_array_slice(); }
  81   bool is_array_slice() const { return ((uintptr_t)_holder & ArraySliceBit) != 0; }
  82   bool is_null() const { return _holder == NULL; }
  83 };
  84 
  85 typedef GenericTaskQueue<G1TaskQueueEntry, mtGC> G1CMTaskQueue;
  86 typedef GenericTaskQueueSet<G1CMTaskQueue, mtGC> G1CMTaskQueueSet;
  87 
  88 // Closure used by CM during concurrent reference discovery
  89 // and reference processing (during remarking) to determine
  90 // if a particular object is alive. It is primarily used
  91 // to determine if referents of discovered reference objects
  92 // are alive. An instance is also embedded into the
  93 // reference processor as the _is_alive_non_header field
  94 class G1CMIsAliveClosure: public BoolObjectClosure {
  95   G1CollectedHeap* _g1;
  96  public:
  97   G1CMIsAliveClosure(G1CollectedHeap* g1) : _g1(g1) { }
  98 
  99   bool do_object_b(oop obj);
 100 };
 101 
 102 // A generic CM bit map.  This is essentially a wrapper around the BitMap
 103 // class, with one bit per (1<<_shifter) HeapWords.
 104 
 105 class G1CMBitMapRO VALUE_OBJ_CLASS_SPEC {
 106  protected:
 107   HeapWord*  _bmStartWord; // base address of range covered by map
 108   size_t     _bmWordSize;  // map size (in #HeapWords covered)
 109   const int  _shifter;     // map to char or bit
 110   BitMapView _bm;          // the bit map itself
 111 
 112  public:
 113   // constructor
 114   G1CMBitMapRO(int shifter);
 115 
 116   // inquiries
 117   HeapWord* startWord()   const { return _bmStartWord; }
 118   // the following is one past the last word in space
 119   HeapWord* endWord()     const { return _bmStartWord + _bmWordSize; }
 120 
 121   // read marks
 122 
 123   bool isMarked(HeapWord* addr) const {
 124     assert(_bmStartWord <= addr && addr < (_bmStartWord + _bmWordSize),
 125            "outside underlying space?");
 126     return _bm.at(heapWordToOffset(addr));
 127   }
 128 
 129   // iteration
 130   inline bool iterate(BitMapClosure* cl, MemRegion mr);
 131 
 132   // Return the address corresponding to the next marked bit at or after
 133   // "addr", and before "limit", if "limit" is non-NULL.  If there is no
 134   // such bit, returns "limit" if that is non-NULL, or else "endWord()".
 135   HeapWord* getNextMarkedWordAddress(const HeapWord* addr,
 136                                      const HeapWord* limit = NULL) const;
 137 
 138   // conversion utilities
 139   HeapWord* offsetToHeapWord(size_t offset) const {
 140     return _bmStartWord + (offset << _shifter);
 141   }
 142   size_t heapWordToOffset(const HeapWord* addr) const {
 143     return pointer_delta(addr, _bmStartWord) >> _shifter;
 144   }
 145 
 146   // The argument addr should be the start address of a valid object
 147   inline HeapWord* nextObject(HeapWord* addr);
 148 
 149   void print_on_error(outputStream* st, const char* prefix) const;
 150 
 151   // debugging
 152   NOT_PRODUCT(bool covers(MemRegion rs) const;)
 153 };
 154 
 155 class G1CMBitMapMappingChangedListener : public G1MappingChangedListener {
 156  private:
 157   G1CMBitMap* _bm;
 158  public:
 159   G1CMBitMapMappingChangedListener() : _bm(NULL) {}
 160 
 161   void set_bitmap(G1CMBitMap* bm) { _bm = bm; }
 162 
 163   virtual void on_commit(uint start_idx, size_t num_regions, bool zero_filled);
 164 };
 165 
 166 class G1CMBitMap : public G1CMBitMapRO {
 167  private:
 168   G1CMBitMapMappingChangedListener _listener;
 169 
 170  public:
 171   static size_t compute_size(size_t heap_size);
 172   // Returns the amount of bytes on the heap between two marks in the bitmap.
 173   static size_t mark_distance();
 174   // Returns how many bytes (or bits) of the heap a single byte (or bit) of the
 175   // mark bitmap corresponds to. This is the same as the mark distance above.
 176   static size_t heap_map_factor() {
 177     return mark_distance();
 178   }
 179 
 180   G1CMBitMap() : G1CMBitMapRO(LogMinObjAlignment), _listener() { _listener.set_bitmap(this); }
 181 
 182   // Initializes the underlying BitMap to cover the given area.
 183   void initialize(MemRegion heap, G1RegionToSpaceMapper* storage);
 184 
 185   // Write marks.
 186   inline void mark(HeapWord* addr);
 187   inline void clear(HeapWord* addr);
 188   inline bool parMark(HeapWord* addr);
 189 
 190   void clear_range(MemRegion mr);
 191 };
 192 
 193 // Represents the overflow mark stack used by concurrent marking.
 194 //
 195 // Stores oops in a huge buffer in virtual memory that is always fully committed.
 196 // Resizing may only happen during a STW pause when the stack is empty.
 197 //
 198 // Memory is allocated on a "chunk" basis, i.e. a set of oops. For this, the mark
 199 // stack memory is split into evenly sized chunks of oops. Users can only
 200 // add or remove entries on that basis.
 201 // Chunks are filled in increasing address order. Not completely filled chunks
 202 // have a NULL element as a terminating element.
 203 //
 204 // Every chunk has a header containing a single pointer element used for memory
 205 // management. This wastes some space, but is negligible (< .1% with current sizing).
 206 //
 207 // Memory management is done using a mix of tracking a high water-mark indicating
 208 // that all chunks at a lower address are valid chunks, and a singly linked free
 209 // list connecting all empty chunks.
 210 class G1CMMarkStack VALUE_OBJ_CLASS_SPEC {
 211 public:
 212   // Number of oops that can fit in a single chunk.
 213   static const size_t EntriesPerChunk = 1024 - 1 /* One reference for the next pointer */;
 214 private:
 215   struct TaskQueueEntryChunk {
 216     TaskQueueEntryChunk* next;
 217     G1TaskQueueEntry data[EntriesPerChunk];
 218   };
 219 
 220   size_t _max_chunk_capacity;    // Maximum number of OopChunk elements on the stack.
 221 
 222   TaskQueueEntryChunk* _base;               // Bottom address of allocated memory area.
 223   size_t _chunk_capacity;        // Current maximum number of OopChunk elements.
 224 
 225   char _pad0[DEFAULT_CACHE_LINE_SIZE];
 226   TaskQueueEntryChunk* volatile _free_list;  // Linked list of free chunks that can be allocated by users.
 227   char _pad1[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*)];
 228   TaskQueueEntryChunk* volatile _chunk_list; // List of chunks currently containing data.
 229   volatile size_t _chunks_in_chunk_list;
 230   char _pad2[DEFAULT_CACHE_LINE_SIZE - sizeof(TaskQueueEntryChunk*) - sizeof(size_t)];
 231 
 232   volatile size_t _hwm;          // High water mark within the reserved space.
 233   char _pad4[DEFAULT_CACHE_LINE_SIZE - sizeof(size_t)];
 234 
 235   // Allocate a new chunk from the reserved memory, using the high water mark. Returns
 236   // NULL if out of memory.
 237   TaskQueueEntryChunk* allocate_new_chunk();
 238 
 239   volatile bool _out_of_memory;
 240 
 241   // Atomically add the given chunk to the list.
 242   void add_chunk_to_list(TaskQueueEntryChunk* volatile* list, TaskQueueEntryChunk* elem);
 243   // Atomically remove and return a chunk from the given list. Returns NULL if the
 244   // list is empty.
 245   TaskQueueEntryChunk* remove_chunk_from_list(TaskQueueEntryChunk* volatile* list);
 246 
 247   void add_chunk_to_chunk_list(TaskQueueEntryChunk* elem);
 248   void add_chunk_to_free_list(TaskQueueEntryChunk* elem);
 249 
 250   TaskQueueEntryChunk* remove_chunk_from_chunk_list();
 251   TaskQueueEntryChunk* remove_chunk_from_free_list();
 252 
 253   bool  _should_expand;
 254 
 255   // Resizes the mark stack to the given new capacity. Releases any previous
 256   // memory if successful.
 257   bool resize(size_t new_capacity);
 258 
 259  public:
 260   G1CMMarkStack();
 261   ~G1CMMarkStack();
 262 
 263   // Alignment and minimum capacity of this mark stack in number of oops.
 264   static size_t capacity_alignment();
 265 
 266   // Allocate and initialize the mark stack with the given number of oops.
 267   bool initialize(size_t initial_capacity, size_t max_capacity);
 268 
 269   // Pushes the given buffer containing at most EntriesPerChunk elements on the mark
 270   // stack. If less than EntriesPerChunk elements are to be pushed, the array must
 271   // be terminated with a NULL.
 272   // Returns whether the buffer contents were successfully pushed to the global mark
 273   // stack.
 274   bool par_push_chunk(G1TaskQueueEntry* buffer);
 275 
 276   // Pops a chunk from this mark stack, copying them into the given buffer. This
 277   // chunk may contain up to EntriesPerChunk elements. If there are less, the last
 278   // element in the array is a NULL pointer.
 279   bool par_pop_chunk(G1TaskQueueEntry* buffer);
 280 
 281   // Return whether the chunk list is empty. Racy due to unsynchronized access to
 282   // _chunk_list.
 283   bool is_empty() const { return _chunk_list == NULL; }
 284 
 285   size_t capacity() const  { return _chunk_capacity; }
 286 
 287   bool is_out_of_memory() const { return _out_of_memory; }
 288   void clear_out_of_memory() { _out_of_memory = false; }
 289 
 290   bool should_expand() const { return _should_expand; }
 291   void set_should_expand(bool value) { _should_expand = value; }
 292 
 293   // Expand the stack, typically in response to an overflow condition
 294   void expand();
 295 
 296   // Return the approximate number of oops on this mark stack. Racy due to
 297   // unsynchronized access to _chunks_in_chunk_list.
 298   size_t size() const { return _chunks_in_chunk_list * EntriesPerChunk; }
 299 
 300   void set_empty();
 301 
 302   // Apply Fn to every oop on the mark stack. The mark stack must not
 303   // be modified while iterating.
 304   template<typename Fn> void iterate(Fn fn) const PRODUCT_RETURN;
 305 };
 306 
 307 // Root Regions are regions that are not empty at the beginning of a
 308 // marking cycle and which we might collect during an evacuation pause
 309 // while the cycle is active. Given that, during evacuation pauses, we
 310 // do not copy objects that are explicitly marked, what we have to do
 311 // for the root regions is to scan them and mark all objects reachable
 312 // from them. According to the SATB assumptions, we only need to visit
 313 // each object once during marking. So, as long as we finish this scan
 314 // before the next evacuation pause, we can copy the objects from the
 315 // root regions without having to mark them or do anything else to them.
 316 //
 317 // Currently, we only support root region scanning once (at the start
 318 // of the marking cycle) and the root regions are all the survivor
 319 // regions populated during the initial-mark pause.
 320 class G1CMRootRegions VALUE_OBJ_CLASS_SPEC {
 321 private:
 322   const G1SurvivorRegions* _survivors;
 323   G1ConcurrentMark*        _cm;
 324 
 325   volatile bool            _scan_in_progress;
 326   volatile bool            _should_abort;
 327   volatile int             _claimed_survivor_index;
 328 
 329   void notify_scan_done();
 330 
 331 public:
 332   G1CMRootRegions();
 333   // We actually do most of the initialization in this method.
 334   void init(const G1SurvivorRegions* survivors, G1ConcurrentMark* cm);
 335 
 336   // Reset the claiming / scanning of the root regions.
 337   void prepare_for_scan();
 338 
 339   // Forces get_next() to return NULL so that the iteration aborts early.
 340   void abort() { _should_abort = true; }
 341 
 342   // Return true if the CM thread are actively scanning root regions,
 343   // false otherwise.
 344   bool scan_in_progress() { return _scan_in_progress; }
 345 
 346   // Claim the next root region to scan atomically, or return NULL if
 347   // all have been claimed.
 348   HeapRegion* claim_next();
 349 
 350   // The number of root regions to scan.
 351   uint num_root_regions() const;
 352 
 353   void cancel_scan();
 354 
 355   // Flag that we're done with root region scanning and notify anyone
 356   // who's waiting on it. If aborted is false, assume that all regions
 357   // have been claimed.
 358   void scan_finished();
 359 
 360   // If CM threads are still scanning root regions, wait until they
 361   // are done. Return true if we had to wait, false otherwise.
 362   bool wait_until_scan_finished();
 363 };
 364 
 365 class ConcurrentMarkThread;
 366 
 367 class G1ConcurrentMark: public CHeapObj<mtGC> {
 368   friend class ConcurrentMarkThread;
 369   friend class G1ParNoteEndTask;
 370   friend class G1VerifyLiveDataClosure;
 371   friend class G1CMRefProcTaskProxy;
 372   friend class G1CMRefProcTaskExecutor;
 373   friend class G1CMKeepAliveAndDrainClosure;
 374   friend class G1CMDrainMarkingStackClosure;
 375   friend class G1CMBitMapClosure;
 376   friend class G1CMConcurrentMarkingTask;
 377   friend class G1CMRemarkTask;
 378   friend class G1CMTask;
 379 
 380 protected:
 381   ConcurrentMarkThread* _cmThread;   // The thread doing the work
 382   G1CollectedHeap*      _g1h;        // The heap
 383   uint                  _parallel_marking_threads; // The number of marking
 384                                                    // threads we're using
 385   uint                  _max_parallel_marking_threads; // Max number of marking
 386                                                        // threads we'll ever use
 387   double                _sleep_factor; // How much we have to sleep, with
 388                                        // respect to the work we just did, to
 389                                        // meet the marking overhead goal
 390   double                _marking_task_overhead; // Marking target overhead for
 391                                                 // a single task
 392 
 393   FreeRegionList        _cleanup_list;
 394 
 395   // Concurrent marking support structures
 396   G1CMBitMap              _markBitMap1;
 397   G1CMBitMap              _markBitMap2;
 398   G1CMBitMapRO*           _prevMarkBitMap; // Completed mark bitmap
 399   G1CMBitMap*             _nextMarkBitMap; // Under-construction mark bitmap
 400 
 401   // Heap bounds
 402   HeapWord*               _heap_start;
 403   HeapWord*               _heap_end;
 404 
 405   // Root region tracking and claiming
 406   G1CMRootRegions         _root_regions;
 407 
 408   // For gray objects
 409   G1CMMarkStack           _global_mark_stack; // Grey objects behind global finger
 410   HeapWord* volatile      _finger;  // The global finger, region aligned,
 411                                     // always points to the end of the
 412                                     // last claimed region
 413 
 414   // Marking tasks
 415   uint                    _max_worker_id;// Maximum worker id
 416   uint                    _active_tasks; // Task num currently active
 417   G1CMTask**              _tasks;        // Task queue array (max_worker_id len)
 418   G1CMTaskQueueSet*       _task_queues;  // Task queue set
 419   ParallelTaskTerminator  _terminator;   // For termination
 420 
 421   // Two sync barriers that are used to synchronize tasks when an
 422   // overflow occurs. The algorithm is the following. All tasks enter
 423   // the first one to ensure that they have all stopped manipulating
 424   // the global data structures. After they exit it, they re-initialize
 425   // their data structures and task 0 re-initializes the global data
 426   // structures. Then, they enter the second sync barrier. This
 427   // ensure, that no task starts doing work before all data
 428   // structures (local and global) have been re-initialized. When they
 429   // exit it, they are free to start working again.
 430   WorkGangBarrierSync     _first_overflow_barrier_sync;
 431   WorkGangBarrierSync     _second_overflow_barrier_sync;
 432 
 433   // This is set by any task, when an overflow on the global data
 434   // structures is detected
 435   volatile bool           _has_overflown;
 436   // True: marking is concurrent, false: we're in remark
 437   volatile bool           _concurrent;
 438   // Set at the end of a Full GC so that marking aborts
 439   volatile bool           _has_aborted;
 440 
 441   // Used when remark aborts due to an overflow to indicate that
 442   // another concurrent marking phase should start
 443   volatile bool           _restart_for_overflow;
 444 
 445   // This is true from the very start of concurrent marking until the
 446   // point when all the tasks complete their work. It is really used
 447   // to determine the points between the end of concurrent marking and
 448   // time of remark.
 449   volatile bool           _concurrent_marking_in_progress;
 450 
 451   ConcurrentGCTimer*      _gc_timer_cm;
 452 
 453   G1OldTracer*            _gc_tracer_cm;
 454 
 455   // All of these times are in ms
 456   NumberSeq _init_times;
 457   NumberSeq _remark_times;
 458   NumberSeq _remark_mark_times;
 459   NumberSeq _remark_weak_ref_times;
 460   NumberSeq _cleanup_times;
 461   double    _total_counting_time;
 462   double    _total_rs_scrub_time;
 463 
 464   double*   _accum_task_vtime;   // Accumulated task vtime
 465 
 466   WorkGang* _parallel_workers;
 467 
 468   void weakRefsWorkParallelPart(BoolObjectClosure* is_alive, bool purged_classes);
 469   void weakRefsWork(bool clear_all_soft_refs);
 470 
 471   void swapMarkBitMaps();
 472 
 473   // It resets the global marking data structures, as well as the
 474   // task local ones; should be called during initial mark.
 475   void reset();
 476 
 477   // Resets all the marking data structures. Called when we have to restart
 478   // marking or when marking completes (via set_non_marking_state below).
 479   void reset_marking_state(bool clear_overflow = true);
 480 
 481   // We do this after we're done with marking so that the marking data
 482   // structures are initialized to a sensible and predictable state.
 483   void set_non_marking_state();
 484 
 485   // Called to indicate how many threads are currently active.
 486   void set_concurrency(uint active_tasks);
 487 
 488   // It should be called to indicate which phase we're in (concurrent
 489   // mark or remark) and how many threads are currently active.
 490   void set_concurrency_and_phase(uint active_tasks, bool concurrent);
 491 
 492   // Prints all gathered CM-related statistics
 493   void print_stats();
 494 
 495   bool cleanup_list_is_empty() {
 496     return _cleanup_list.is_empty();
 497   }
 498 
 499   // Accessor methods
 500   uint parallel_marking_threads() const     { return _parallel_marking_threads; }
 501   uint max_parallel_marking_threads() const { return _max_parallel_marking_threads;}
 502   double sleep_factor()                     { return _sleep_factor; }
 503   double marking_task_overhead()            { return _marking_task_overhead;}
 504 
 505   HeapWord*               finger()          { return _finger;   }
 506   bool                    concurrent()      { return _concurrent; }
 507   uint                    active_tasks()    { return _active_tasks; }
 508   ParallelTaskTerminator* terminator()      { return &_terminator; }
 509 
 510   // It claims the next available region to be scanned by a marking
 511   // task/thread. It might return NULL if the next region is empty or
 512   // we have run out of regions. In the latter case, out_of_regions()
 513   // determines whether we've really run out of regions or the task
 514   // should call claim_region() again. This might seem a bit
 515   // awkward. Originally, the code was written so that claim_region()
 516   // either successfully returned with a non-empty region or there
 517   // were no more regions to be claimed. The problem with this was
 518   // that, in certain circumstances, it iterated over large chunks of
 519   // the heap finding only empty regions and, while it was working, it
 520   // was preventing the calling task to call its regular clock
 521   // method. So, this way, each task will spend very little time in
 522   // claim_region() and is allowed to call the regular clock method
 523   // frequently.
 524   HeapRegion* claim_region(uint worker_id);
 525 
 526   // It determines whether we've run out of regions to scan. Note that
 527   // the finger can point past the heap end in case the heap was expanded
 528   // to satisfy an allocation without doing a GC. This is fine, because all
 529   // objects in those regions will be considered live anyway because of
 530   // SATB guarantees (i.e. their TAMS will be equal to bottom).
 531   bool        out_of_regions() { return _finger >= _heap_end; }
 532 
 533   // Returns the task with the given id
 534   G1CMTask* task(int id) {
 535     assert(0 <= id && id < (int) _active_tasks,
 536            "task id not within active bounds");
 537     return _tasks[id];
 538   }
 539 
 540   // Returns the task queue with the given id
 541   G1CMTaskQueue* task_queue(int id) {
 542     assert(0 <= id && id < (int) _active_tasks,
 543            "task queue id not within active bounds");
 544     return (G1CMTaskQueue*) _task_queues->queue(id);
 545   }
 546 
 547   // Returns the task queue set
 548   G1CMTaskQueueSet* task_queues()  { return _task_queues; }
 549 
 550   // Access / manipulation of the overflow flag which is set to
 551   // indicate that the global stack has overflown
 552   bool has_overflown()           { return _has_overflown; }
 553   void set_has_overflown()       { _has_overflown = true; }
 554   void clear_has_overflown()     { _has_overflown = false; }
 555   bool restart_for_overflow()    { return _restart_for_overflow; }
 556 
 557   // Methods to enter the two overflow sync barriers
 558   void enter_first_sync_barrier(uint worker_id);
 559   void enter_second_sync_barrier(uint worker_id);
 560 
 561   // Card index of the bottom of the G1 heap. Used for biasing indices into
 562   // the card bitmaps.
 563   intptr_t _heap_bottom_card_num;
 564 
 565   // Set to true when initialization is complete
 566   bool _completed_initialization;
 567 
 568   // end_timer, true to end gc timer after ending concurrent phase.
 569   void register_concurrent_phase_end_common(bool end_timer);
 570 
 571   // Clear the given bitmap in parallel using the given WorkGang. If may_yield is
 572   // true, periodically insert checks to see if this method should exit prematurely.
 573   void clear_bitmap(G1CMBitMap* bitmap, WorkGang* workers, bool may_yield);
 574 public:
 575   // Manipulation of the global mark stack.
 576   // The push and pop operations are used by tasks for transfers
 577   // between task-local queues and the global mark stack.
 578   bool mark_stack_push(G1TaskQueueEntry* arr) {
 579     if (!_global_mark_stack.par_push_chunk(arr)) {
 580       set_has_overflown();
 581       return false;
 582     }
 583     return true;
 584   }
 585   bool mark_stack_pop(G1TaskQueueEntry* arr) {
 586     return _global_mark_stack.par_pop_chunk(arr);
 587   }
 588   size_t mark_stack_size()                { return _global_mark_stack.size(); }
 589   size_t partial_mark_stack_size_target() { return _global_mark_stack.capacity()/3; }
 590   bool mark_stack_overflow()              { return _global_mark_stack.is_out_of_memory(); }
 591   bool mark_stack_empty()                 { return _global_mark_stack.is_empty(); }
 592 
 593   G1CMRootRegions* 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 concurrent_cycle_start();
 606   void concurrent_cycle_end();
 607 
 608   void update_accum_task_vtime(int i, double vtime) {
 609     _accum_task_vtime[i] += vtime;
 610   }
 611 
 612   double all_task_accum_vtime() {
 613     double ret = 0.0;
 614     for (uint i = 0; i < _max_worker_id; ++i)
 615       ret += _accum_task_vtime[i];
 616     return ret;
 617   }
 618 
 619   // Attempts to steal an object from the task queues of other tasks
 620   bool try_stealing(uint worker_id, int* hash_seed, G1TaskQueueEntry& task_entry);
 621 
 622   G1ConcurrentMark(G1CollectedHeap* g1h,
 623                    G1RegionToSpaceMapper* prev_bitmap_storage,
 624                    G1RegionToSpaceMapper* next_bitmap_storage);
 625   ~G1ConcurrentMark();
 626 
 627   ConcurrentMarkThread* cmThread() { return _cmThread; }
 628 
 629   G1CMBitMapRO* prevMarkBitMap() const { return _prevMarkBitMap; }
 630   G1CMBitMap*   nextMarkBitMap() const { return _nextMarkBitMap; }
 631 
 632   // Returns the number of GC threads to be used in a concurrent
 633   // phase based on the number of GC threads being used in a STW
 634   // phase.
 635   uint scale_parallel_threads(uint n_par_threads);
 636 
 637   // Calculates the number of GC threads to be used in a concurrent phase.
 638   uint calc_parallel_marking_threads();
 639 
 640   // The following three are interaction between CM and
 641   // G1CollectedHeap
 642 
 643   // This notifies CM that a root during initial-mark needs to be
 644   // grayed. It is MT-safe. hr is the region that
 645   // contains the object and it's passed optionally from callers who
 646   // might already have it (no point in recalculating it).
 647   inline void grayRoot(oop obj,
 648                        HeapRegion* hr = NULL);
 649 
 650   // Prepare internal data structures for the next mark cycle. This includes clearing
 651   // the next mark bitmap and some internal data structures. This method is intended
 652   // to be called concurrently to the mutator. It will yield to safepoint requests.
 653   void cleanup_for_next_mark();
 654 
 655   // Clear the previous marking bitmap during safepoint.
 656   void clear_prev_bitmap(WorkGang* workers);
 657 
 658   // Return whether the next mark bitmap has no marks set. To be used for assertions
 659   // only. Will not yield to pause requests.
 660   bool nextMarkBitmapIsClear();
 661 
 662   // These two do the work that needs to be done before and after the
 663   // initial root checkpoint. Since this checkpoint can be done at two
 664   // different points (i.e. an explicit pause or piggy-backed on a
 665   // young collection), then it's nice to be able to easily share the
 666   // pre/post code. It might be the case that we can put everything in
 667   // the post method. TP
 668   void checkpointRootsInitialPre();
 669   void checkpointRootsInitialPost();
 670 
 671   // Scan all the root regions and mark everything reachable from
 672   // them.
 673   void scan_root_regions();
 674 
 675   // Scan a single root region and mark everything reachable from it.
 676   void scanRootRegion(HeapRegion* hr);
 677 
 678   // Do concurrent phase of marking, to a tentative transitive closure.
 679   void mark_from_roots();
 680 
 681   void checkpointRootsFinal(bool clear_all_soft_refs);
 682   void checkpointRootsFinalWork();
 683   void cleanup();
 684   void complete_cleanup();
 685 
 686   // Mark in the previous bitmap.  NB: this is usually read-only, so use
 687   // this carefully!
 688   inline void markPrev(oop p);
 689 
 690   // Clears marks for all objects in the given range, for the prev or
 691   // next bitmaps.  NB: the previous bitmap is usually
 692   // read-only, so use this carefully!
 693   void clearRangePrevBitmap(MemRegion mr);
 694 
 695   // Verify that there are no CSet oops on the stacks (taskqueues /
 696   // global mark stack) and fingers (global / per-task).
 697   // If marking is not in progress, it's a no-op.
 698   void verify_no_cset_oops() PRODUCT_RETURN;
 699 
 700   inline bool isPrevMarked(oop p) const;
 701 
 702   inline bool do_yield_check();
 703 
 704   // Abandon current marking iteration due to a Full GC.
 705   void abort();
 706 
 707   bool has_aborted()      { return _has_aborted; }
 708 
 709   void print_summary_info();
 710 
 711   void print_worker_threads_on(outputStream* st) const;
 712   void threads_do(ThreadClosure* tc) const;
 713 
 714   void print_on_error(outputStream* st) const;
 715 
 716   // Attempts to mark the given object on the next mark bitmap.
 717   inline bool par_mark(oop obj);
 718 
 719   // Returns true if initialization was successfully completed.
 720   bool completed_initialization() const {
 721     return _completed_initialization;
 722   }
 723 
 724   ConcurrentGCTimer* gc_timer_cm() const { return _gc_timer_cm; }
 725   G1OldTracer* gc_tracer_cm() const { return _gc_tracer_cm; }
 726 
 727 private:
 728   // Clear (Reset) all liveness count data.
 729   void clear_live_data(WorkGang* workers);
 730 
 731 #ifdef ASSERT
 732   // Verify all of the above data structures that they are in initial state.
 733   void verify_live_data_clear();
 734 #endif
 735 
 736   // Aggregates the per-card liveness data based on the current marking. Also sets
 737   // the amount of marked bytes for each region.
 738   void create_live_data();
 739 
 740   void finalize_live_data();
 741 
 742   void verify_live_data();
 743 };
 744 
 745 // A class representing a marking task.
 746 class G1CMTask : public TerminatorTerminator {
 747 private:
 748   enum PrivateConstants {
 749     // The regular clock call is called once the scanned words reaches
 750     // this limit
 751     words_scanned_period          = 12*1024,
 752     // The regular clock call is called once the number of visited
 753     // references reaches this limit
 754     refs_reached_period           = 1024,
 755     // Initial value for the hash seed, used in the work stealing code
 756     init_hash_seed                = 17
 757   };
 758 
 759   G1CMObjArrayProcessor       _objArray_processor;
 760 
 761   uint                        _worker_id;
 762   G1CollectedHeap*            _g1h;
 763   G1ConcurrentMark*           _cm;
 764   G1CMBitMap*                 _nextMarkBitMap;
 765   // the task queue of this task
 766   G1CMTaskQueue*              _task_queue;
 767 private:
 768   // the task queue set---needed for stealing
 769   G1CMTaskQueueSet*           _task_queues;
 770   // indicates whether the task has been claimed---this is only  for
 771   // debugging purposes
 772   bool                        _claimed;
 773 
 774   // number of calls to this task
 775   int                         _calls;
 776 
 777   // when the virtual timer reaches this time, the marking step should
 778   // exit
 779   double                      _time_target_ms;
 780   // the start time of the current marking step
 781   double                      _start_time_ms;
 782 
 783   // the oop closure used for iterations over oops
 784   G1CMOopClosure*             _cm_oop_closure;
 785 
 786   // the region this task is scanning, NULL if we're not scanning any
 787   HeapRegion*                 _curr_region;
 788   // the local finger of this task, NULL if we're not scanning a region
 789   HeapWord*                   _finger;
 790   // limit of the region this task is scanning, NULL if we're not scanning one
 791   HeapWord*                   _region_limit;
 792 
 793   // the number of words this task has scanned
 794   size_t                      _words_scanned;
 795   // When _words_scanned reaches this limit, the regular clock is
 796   // called. Notice that this might be decreased under certain
 797   // circumstances (i.e. when we believe that we did an expensive
 798   // operation).
 799   size_t                      _words_scanned_limit;
 800   // the initial value of _words_scanned_limit (i.e. what it was
 801   // before it was decreased).
 802   size_t                      _real_words_scanned_limit;
 803 
 804   // the number of references this task has visited
 805   size_t                      _refs_reached;
 806   // When _refs_reached reaches this limit, the regular clock is
 807   // called. Notice this this might be decreased under certain
 808   // circumstances (i.e. when we believe that we did an expensive
 809   // operation).
 810   size_t                      _refs_reached_limit;
 811   // the initial value of _refs_reached_limit (i.e. what it was before
 812   // it was decreased).
 813   size_t                      _real_refs_reached_limit;
 814 
 815   // used by the work stealing stuff
 816   int                         _hash_seed;
 817   // if this is true, then the task has aborted for some reason
 818   bool                        _has_aborted;
 819   // set when the task aborts because it has met its time quota
 820   bool                        _has_timed_out;
 821   // true when we're draining SATB buffers; this avoids the task
 822   // aborting due to SATB buffers being available (as we're already
 823   // dealing with them)
 824   bool                        _draining_satb_buffers;
 825 
 826   // number sequence of past step times
 827   NumberSeq                   _step_times_ms;
 828   // elapsed time of this task
 829   double                      _elapsed_time_ms;
 830   // termination time of this task
 831   double                      _termination_time_ms;
 832   // when this task got into the termination protocol
 833   double                      _termination_start_time_ms;
 834 
 835   // true when the task is during a concurrent phase, false when it is
 836   // in the remark phase (so, in the latter case, we do not have to
 837   // check all the things that we have to check during the concurrent
 838   // phase, i.e. SATB buffer availability...)
 839   bool                        _concurrent;
 840 
 841   TruncatedSeq                _marking_step_diffs_ms;
 842 
 843   // it updates the local fields after this task has claimed
 844   // a new region to scan
 845   void setup_for_region(HeapRegion* hr);
 846   // it brings up-to-date the limit of the region
 847   void update_region_limit();
 848 
 849   // called when either the words scanned or the refs visited limit
 850   // has been reached
 851   void reached_limit();
 852   // recalculates the words scanned and refs visited limits
 853   void recalculate_limits();
 854   // decreases the words scanned and refs visited limits when we reach
 855   // an expensive operation
 856   void decrease_limits();
 857   // it checks whether the words scanned or refs visited reached their
 858   // respective limit and calls reached_limit() if they have
 859   void check_limits() {
 860     if (_words_scanned >= _words_scanned_limit ||
 861         _refs_reached >= _refs_reached_limit) {
 862       reached_limit();
 863     }
 864   }
 865   // this is supposed to be called regularly during a marking step as
 866   // it checks a bunch of conditions that might cause the marking step
 867   // to abort
 868   void regular_clock_call();
 869   bool concurrent() { return _concurrent; }
 870 
 871   // Test whether obj might have already been passed over by the
 872   // mark bitmap scan, and so needs to be pushed onto the mark stack.
 873   bool is_below_finger(oop obj, HeapWord* global_finger) const;
 874 
 875   template<bool scan> void process_grey_task_entry(G1TaskQueueEntry task_entry);
 876 public:
 877   // Apply the closure on the given area of the objArray. Return the number of words
 878   // scanned.
 879   inline size_t scan_objArray(objArrayOop obj, MemRegion mr);
 880   // It resets the task; it should be called right at the beginning of
 881   // a marking phase.
 882   void reset(G1CMBitMap* _nextMarkBitMap);
 883   // it clears all the fields that correspond to a claimed region.
 884   void clear_region_fields();
 885 
 886   void set_concurrent(bool concurrent) { _concurrent = concurrent; }
 887 
 888   // The main method of this class which performs a marking step
 889   // trying not to exceed the given duration. However, it might exit
 890   // prematurely, according to some conditions (i.e. SATB buffers are
 891   // available for processing).
 892   void do_marking_step(double target_ms,
 893                        bool do_termination,
 894                        bool is_serial);
 895 
 896   // These two calls start and stop the timer
 897   void record_start_time() {
 898     _elapsed_time_ms = os::elapsedTime() * 1000.0;
 899   }
 900   void record_end_time() {
 901     _elapsed_time_ms = os::elapsedTime() * 1000.0 - _elapsed_time_ms;
 902   }
 903 
 904   // returns the worker ID associated with this task.
 905   uint worker_id() { return _worker_id; }
 906 
 907   // From TerminatorTerminator. It determines whether this task should
 908   // exit the termination protocol after it's entered it.
 909   virtual bool should_exit_termination();
 910 
 911   // Resets the local region fields after a task has finished scanning a
 912   // region; or when they have become stale as a result of the region
 913   // being evacuated.
 914   void giveup_current_region();
 915 
 916   HeapWord* finger()            { return _finger; }
 917 
 918   bool has_aborted()            { return _has_aborted; }
 919   void set_has_aborted()        { _has_aborted = true; }
 920   void clear_has_aborted()      { _has_aborted = false; }
 921   bool has_timed_out()          { return _has_timed_out; }
 922   bool claimed()                { return _claimed; }
 923 
 924   void set_cm_oop_closure(G1CMOopClosure* cm_oop_closure);
 925 
 926   // Increment the number of references this task has visited.
 927   void increment_refs_reached() { ++_refs_reached; }
 928 
 929   // Grey the object by marking it.  If not already marked, push it on
 930   // the local queue if below the finger.
 931   // obj is below its region's NTAMS.
 932   inline void make_reference_grey(oop obj);
 933 
 934   // Grey the object (by calling make_grey_reference) if required,
 935   // e.g. obj is below its containing region's NTAMS.
 936   // Precondition: obj is a valid heap object.
 937   inline void deal_with_reference(oop obj);
 938 
 939   // It scans an object and visits its children.
 940   inline void scan_task_entry(G1TaskQueueEntry task_entry);
 941 
 942   // It pushes an object on the local queue.
 943   inline void push(G1TaskQueueEntry task_entry);
 944 
 945   // Move entries to the global stack.
 946   void move_entries_to_global_stack();
 947   // Move entries from the global stack, return true if we were successful to do so.
 948   bool get_entries_from_global_stack();
 949 
 950   // It pops and scans objects from the local queue. If partially is
 951   // true, then it stops when the queue size is of a given limit. If
 952   // partially is false, then it stops when the queue is empty.
 953   void drain_local_queue(bool partially);
 954   // It moves entries from the global stack to the local queue and
 955   // drains the local queue. If partially is true, then it stops when
 956   // both the global stack and the local queue reach a given size. If
 957   // partially if false, it tries to empty them totally.
 958   void drain_global_stack(bool partially);
 959   // It keeps picking SATB buffers and processing them until no SATB
 960   // buffers are available.
 961   void drain_satb_buffers();
 962 
 963   // moves the local finger to a new location
 964   inline void move_finger_to(HeapWord* new_finger) {
 965     assert(new_finger >= _finger && new_finger < _region_limit, "invariant");
 966     _finger = new_finger;
 967   }
 968 
 969   G1CMTask(uint worker_id,
 970            G1ConcurrentMark *cm,
 971            G1CMTaskQueue* task_queue,
 972            G1CMTaskQueueSet* task_queues);
 973 
 974   // it prints statistics associated with this task
 975   void print_stats();
 976 };
 977 
 978 // Class that's used to to print out per-region liveness
 979 // information. It's currently used at the end of marking and also
 980 // after we sort the old regions at the end of the cleanup operation.
 981 class G1PrintRegionLivenessInfoClosure: public HeapRegionClosure {
 982 private:
 983   // Accumulators for these values.
 984   size_t _total_used_bytes;
 985   size_t _total_capacity_bytes;
 986   size_t _total_prev_live_bytes;
 987   size_t _total_next_live_bytes;
 988 
 989   // Accumulator for the remembered set size
 990   size_t _total_remset_bytes;
 991 
 992   // Accumulator for strong code roots memory size
 993   size_t _total_strong_code_roots_bytes;
 994 
 995   static double perc(size_t val, size_t total) {
 996     if (total == 0) {
 997       return 0.0;
 998     } else {
 999       return 100.0 * ((double) val / (double) total);
1000     }
1001   }
1002 
1003   static double bytes_to_mb(size_t val) {
1004     return (double) val / (double) M;
1005   }
1006 
1007 public:
1008   // The header and footer are printed in the constructor and
1009   // destructor respectively.
1010   G1PrintRegionLivenessInfoClosure(const char* phase_name);
1011   virtual bool doHeapRegion(HeapRegion* r);
1012   ~G1PrintRegionLivenessInfoClosure();
1013 };
1014 
1015 #endif // SHARE_VM_GC_G1_G1CONCURRENTMARK_HPP